content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\subsection{A brief review of results on sweeping processes}
The aim of this paper is to extend the study of so-called {\it sweeping process} in the setting of a (possibly infinite dimensional) Riemannian manifold.
Sweeping processes are specific differential inclusions of first order involving proximal normal cones to a moving set. They were extensively studied for the last years in the Euclidean space and more generally in a Hilbert space.
More precisely, consider a Hilbert space $H$ and a moving set $t\to C(t)$ on a time-interval
$I:=[0,T]$ and assume that the set-valued map $C:I \rightrightarrows H$ takes nonempty closed values. A function $u:I\to H$ is a solution of a problem of perturbed sweeping process if it satisfies the following differential inclusion:
\begin{equation}
\left\{
\begin{array}{l}
\displaystyle \frac{du(t)}{dt} + \mathrm{N}(C(t),u(t)) - F(t,u(t)) \ni 0 \vspace{6pt} \\
u(t)\in C(t) \vspace{6pt} \\
u(0)=u_0\ ,
\end{array}
\right. \label{sys1}
\end{equation}
with an initial data $u_0\in C(0)$ and $F: I \times H \rightrightarrows H$ a set-valued map taking nonempty compact values. Here, $\mathrm{N}(C(t),u(t))$ stands for the proximal normal cone to $C(t)$ at the point $u(t)$. This differential inclusion can be thought as following: the point $u(t)$, submitted to the field $F(t,u(t))$, has to live in the set $C(t)$ and so follows its time-evolution.
\bigskip\noindent Let us first give a brief overview of the study for such problems. The sweeping processes have been introduced by
J.J. Moreau in 70's (see \cite{Moreausweep}). He considered the following problem: a point $u(t)$ has to be inside a moving convex set $C(t)$ included in a Hilbert space. When this point is catched-up by the boundary of $C(t)$, it moves in the opposite of the outward normal direction of the boundary, as if it was pushed by the physical boundary in order to stay inside the convex set $C(t)$. Then the position $u(t)$ of this point is described by the following differential inclusion
\begin{equation}
\label{Moreq}
-\dot{u}(t) \in \partial I_{C(t)}(u(t)).
\end{equation}
Here we write $\partial I_{C}$ for the subdifferential of the characteristic function of a convex set $C$. In this work, the sets $C(t)$ are assumed to be convex and so $\partial I_{C(t)}$ is a maximal monotone operator depending on time.
To solve this problem, J.J. Moreau brings a new important idea in proposing a \textit{catching-up} algorithm.
To prove the existence of solutions, he builds discretized solutions in dividing the time interval $I$ into sub-intervals where the convex set $C$ has a little variation. Then by compactness arguments, he shows that a limit mapping can be constructed (when the length of subintervals tends to $0$) which satisfies the desired differential inclusion. \\
Thus it was the first result
concerning sweeping process (with no perturbation $F=0$) because $\partial I_{C(t)}(x) = \mathrm{N}(C(t),x)$ in case of convex sets $C(t)$.
\bigskip\noindent Since then, important improvements have been developed by weakening the assumptions in order to
obtain the most general result of existence for sweeping
process. There are several directions: one can want to add a perturbation
$F$ as written in (\ref{sys1}), one may require a weaker assumption than the convexity of the sets $C(t)$, one would like to obtain results in Banach spaces (and not only in Hilbert spaces).
\bigskip\noindent In \cite{V}, M.~Valadier dealt with sweeping process by sets $C(t)=\mathbb R^n\setminus \textrm{int}(K(t))$ where $K(t)$ are closed and convex sets.
Then in \cite{CXDV}, C.~Castaing, T.X. D\'uc H\={a} and M. Valadier have studied the perturbed problem in finite dimension with convex sets $C(t)$ (or complements of convex sets). In this framework, they proved existence of solutions for (\ref{sys1}) with a convex compact valued perturbation $F$ and a Lipschitzean multifunction $C$.
Later in \cite{Castaingper}, C.~Castaing and M.D.P.~Monteiro Marques have considered similar problems in assuming upper semicontinuity for $F$ and a ``linear compact growth'':
\begin{equation}
F(t,x)\subset \beta(t) (1+|x|)\overline{B(0,1)} \, , \,\, \forall (t,x) \in I \times \mathbb R^n.
\label{hypF}
\end{equation}
Moreover the set-valued map $C$ was supposed to be Hausdorff continuous and satisfying an ``interior ball condition'':
\begin{equation}
\exists r>0 \, , \,\, B(0,r) \subset C(t) \, , \,\, \forall t \in I.
\label{hypC}
\end{equation}
\bigskip\noindent Then the main concept, which appeared to get around the convexity of sets $C(t)$, is the notion of ``uniform prox-regularity''. This property is very well-adapted to the resolution of (\ref{sys1}): a set $C$ is said to be {\it $\eta$-prox-regular} if the projection on $C$ is single-valued and continuous at any point whose the distance to $C$ is smaller than $\eta$. \\
Numerous works have been devoted to study sweeping processes under this assumption of prox-regularity. The case without perturbation ($F={0}$) was firstly treated by G.~Colombo, V.V.~Goncharov in \cite{Colombo}, by H.~Benabdellah in \cite{Benab} and later by L.~Thibault in \cite{Thibsweep} and by G. Colombo, M.D.P.~Monteiro Marques in \cite{Monteiro}.
In \cite{Thibsweep}, the considered problem is
\begin{equation}
\left \{
\begin{array}{l}
-du \in \mathrm{N}(C(t), u(t)) \vspace{6pt} \\
u(0)=u_0\ ,
\end{array}
\right.
\label{eq:mesdiffssm}
\end{equation}
where $du$ is the differential measure of $u$. The existence and uniqueness of solutions of (\ref{eq:mesdiffssm}) are proved with similar assumptions as previously. \\
In an infinite dimensional Hilbert space, the perturbed problem is studied by M.~Bounkhel, J.F.~Edmond and L.~Thibault in \cite{Thibnonconv, Thibsweep, Thibrelax, Thibbv}.
For example in \cite{Thibbv}, the authors show the well-posedness of
\begin{equation}
\left \{
\begin{array}{l}
-du \in \mathrm{N}(C(t), u(t)) + F(t,u(t)) dt \vspace{6pt} \\
u(0)=u_0\ ,
\end{array}
\right.
\label{eq:mesdiffasm}
\end{equation}
with a set-valued map $C$ taking $\eta$-prox regular values (for some $\eta>0$) such that
\begin{equation}
|d_{C(t)}(y) - d_{C(s)}(y)| \leq \mu(]s,t]) \, , \,\, \forall y\in H,\ \forall \ s, t \in I \, , \,\, s\leq t
\label{varadon}
\end{equation}
where $\mu$ is a nonnegative measure satisfying
\begin{equation}
\sup_{s \in I} \mu(\{s\}) <\frac{\eta}{2}.
\label{charge_singleton}
\end{equation}
The proof uses the algorithm developed by J.J.~Moreau with additional arguments to deal with the perturbation $F$ and the prox-regularity assumption.
\bigskip\noindent Indeed the main difficulty of this problem is the weak smoothness of the proximal normal cone. For a fixed closed subset $C$, the set-valued map $x\rightarrow \mathrm{N}(C,x)$ is not in general upper semicontinuous and the uniform prox-regularity assumption permits us to obtain this required smoothness. To finish we mention that the sweeping processes are also studied in the framework of Banach spaces (see \cite{bena2} and \cite{BV}).
\subsection{Sweeping processes on Riemannian manifolds}
As previously explained, sweeping processes have been extensively studied in the context of a linear ambient space (Euclidean space, Hilbert space, Banach space). However, the proof relies on the \textit{catching-up} algorithm (which only requires a projection operator) and the notion of uniform prox-regularity (which is equivalent of an hypomonotonicity property for the proximal normal cone). These observations make us want to extend the study of such differential inclusions in the framework of Riemannian manifolds, where the metric allows us to define a projection and where the Hilbertian structure of the tangent spaces is necessary to have a kind of hypomonotonicity property.
In the Hilbertian situation, the finite dimension is not useful and so we aim to work on a (possibly infinite dimensional) Riemannian Hilbert manifold (see Definition \ref{def:manifold} for details), which is locally isomorphic to a separable Hilbert space. Let $M$ be such a manifold, we consider a sweeping process on $M$, given by a set-valued map $C(\cdot)$ and a single-valued perturbation $f$
\begin{equation}
\left\{
\begin{array}{l}
\displaystyle \frac{dx(t)}{dt} + \mathrm{N}(C(t),x(t)) - f(t,x(t)) \ni 0 \vspace{6pt} \\
x(t)\in C(t) \vspace{6pt} \\
x(0)=x_0\ ,
\end{array}
\right. \label{sys2}
\end{equation}
with an initial data $x_0\in C(0)$.
Note that the first differential inclusion takes place in the fiber-tangent space $T_{x(t)}M$.
To study the well-posedness of this problem, we will generalize the notion of a prox-regular set (Definition \ref{def:prox}) which will be shown to be equivalent to a kind of hypomonotonicity property (see Theorem \ref{thm:hypo}). Some difficulties arise from the framework of manifolds. Indeed in a Hilbertian space $H$, a nonempty closed set $C \subset H $ is said $\eta$-prox-regular or prox-regular with constant $\eta$ if for every point $z$ satisfying $d_C(z)= \underset {c \in C } \inf d(z,c) < \eta$ there exists a unique point $c_z \in C$ such that $d(c_z, z)= d_C(z)$ and the map $z \mapsto c_z$ is continuous. However in infinite-dimensional Riemannian manifolds, geodesics minimizing the distance between two given points do not generally exist. Thus we will make some assumptions so that they do always exist locally. That's why we will give and use a definition of a locally prox-regular set which will take into account the geometry of the ambient manifold even if a definition of uniformly prox-regular sets will be given in Remark \ref{rem:unifprox} in specific manifolds.
In addition we will recover an important property of prox-regular sets. Indeed we will prove that the metric projection onto a locally prox-regular set is single-valued in its neighborhood.
The notions of proximal normal cones and prox-regularity at a point are already introduced in \cite{HP} in the setting of a Riemannian manifold. However we would like to propose another definition of the proximal normal cone (related to the metric projection and better adapted to deal with problems of sweeping processes). Obviously we will show their equivalence. Moreover the prox-regularity at a point is not sufficient to treat our differential inclusions : we need a "larger'' concept of prox-regularity which we will call local prox-regularity as already specified.
Once again these different notions are compatible. Indeed we will prove that a locally prox-regular set in our sense is prox-regular at any point in the sense of \cite{HP}. \\
Finally we prove the main result:
\begin{thm} \label{thm:general} Under some assumptions about the Riemannian Hilbert manifold $M$ (Assumption (\ref{ass:ic})), we consider a Lipschitz bounded mapping $f$ and a Lipschitz set-valued map $C(\cdot)$ taking nonempty and locally prox-regular values (Assumption \ref{ass:C}). Then for all $x_0\in C(0)$, the system (\ref{sys2}) has a unique solution $x\in W^{1,\infty}(I,M)$.
\end{thm}
To our knowledge, sweeping processes on (finite- or infinite-dimensional) manifolds have not been previously studied. However other differential inclusions on Riemannian manifolds have already been treated. For example, in \cite{Grammel}, the author is interested in controllability properties of autonomous differential inclusions: $$ \dot{x}(t) \in F(x(t))$$ in compact finite-dimensional Riemannian manifolds. The map $F$ is supposed to be Lipschitz continuous and takes nonempty convex and compact values.
Then in \cite{Doma}, the authors prove that there exists at least one solution of the problem:
$$\dot{x}(t) \in F(t, x(t)),$$
with an initial condition in a finite-dimensional Riemannian manifold. In this work, $F$ is an integrable bounded field, taking nonempty and closed values. In addition this map is supposed to be lower semicontinuous or to satisfy the Carath\'eodory conditions. Note that compactness arguments are used in their proof so the finite-dimensional assumption is crucial and by such a way, it is not possible to get uniqueness results.
\subsection{Motivation from unilaterally constrained problems}
Let us briefly describe one of our motivations, coming from problems involving unilateral constraints.
Even in the Euclidean space $\mathbb R^d$, several applications for modelling are concerned with sweeping processes involving an admissible set
$$ C(t):=\left\{x\in\mathbb R^d,\ \forall i\in\mathcal I,\ g_i(t,x)\geq 0 \right\}, $$
where $(g_i)_{i\in\mathcal I}$ is a finite collection of inequality constraints.
To apply the existence results to the sweeping process with this moving set, we have to check that $C(t)$ is uniformly prox-regular (with a time-independent constant $\eta$), which may be very difficult. However a criterion was given in \cite[Proposition 2.9]{Juliette} and was used in several recent works (\cite{BV2,BV3,BL}). It would require to check
a kind of Mangasarian-Fromowitz constraint qualification (MFCQ) of the active gradients (we recall that the active gradients correspond to mappings $g_j$ satisfying $g_j(t,x)=0$). Let us specify that this condition describes a positive-linearly independence of the active gradients.
Of course, if the admissible set is also given by equality constraints,
$$ C(t):=\left\{x\in\mathbb R^d,\ \forall i\in \mathcal I_1 \ g_i(t,x)=0 \quad \textrm{and} \quad \forall i\in\mathcal I_2,\ g_i(t,x)\geq 0 \right\}, $$ the same criterion cannot be verified by turning equality constraints into two inequality ones (that is to say by writing for $i\in\mathcal I_1$, $g_i(t,x)=0$ as $g_i(t,x)\geq 0$ and $-g_i(t,x)\geq 0$).
Indeed a gradient and its opposite are clearly not positively independent.
So it is important to have another approach for dealing with such admissible sets (involving inequalities and equalities constraints). In the case where the equality constraints are time-independent, the new point of view is to consider the manifold given by the equality constraints
$$ M:=\{x\in\mathbb R^d,\ \forall i\in\mathcal I_1,\ g_i(x)=0\}, $$ and the admissible set becomes $\tilde{C}(t):=\{x\in M,\ \forall i\in\mathcal I_2, g_i(t,x)\geq 0\}$.
Under some conditions, it is doable to check that the manifold $M$ is a smooth submanifold of $\mathbb R^d$ and so a finite dimensional Riemannian manifold. Consequently it suffices to check that the admissible set $\tilde{C}(\cdot)$ that is only defined with inequalities, is locally prox-regular.
\bigskip
The paper is organized as follows. Section \ref{sec:pre} is first devoted to some definitions and reminders about Riemannian geometry. Then in Section \ref{sec:pre2}, we extend some notions of convex analysis in the setting of a Riemannian Hilbert manifold (proximal normal cone, prox-regular set, ...) and we focus on establishing a characterisation of the prox-regularity property in terms of hypomonotonicity for the proximal normal cone (see Theorem \ref{thm:hypo}). Then we prove Theorem \ref{thm:general} in Section \ref{sec:sweep}, using the catching-up algorithm and studying the convergence of discretized solutions.
\section{Preliminaries and definitions, related to Hilbert manifolds} \label{sec:pre}
For a complete survey on Riemannian geometry and the extension to infinite dimensional framework, we refer the reader to \cite{Klingenberg, Lang}. Let us just recall the main concepts.
\begin{df}[Hilbert manifold] \label{def:manifold}
$M$ is a Hilbert manifold if there exists a separable Hilbert space $H$ and a smooth atlas, defining local charts from $M$ to $H$.
Then it follows that for all $x\in M$, the tangent space $T_xM$ has a Hilbert structure. Its inner product is denoted by $\langle \, ,\, \rangle$, $\langle \, , \,\rangle_x$ or $\langle \, , \,\rangle_{T_xM}$ and the associated norm is represented by $|\ | $ or $|\ |_{T_xM}$.
Such a manifold is called a Riemannian Hilbert manifold if it is endowed with the following Riemannian metric: for two different points $x,y\in M$
$$ d(x,y) :=\inf \left\{ \int_a^b |\dot{\gamma}(s) |ds : \gamma \textmd{ is a $C^1$-curve joining } x = \gamma(a) \textmd{ and } y= \gamma(b)\right\}. $$
The bundle tangent space is denoted by $TM$ and we recall that $$TM = \underset { x \in M} \bigcup \{ x\} \times T_x M .$$
We denote by $\nabla^\star$ the Levi-Civit\'a connection that is the unique connection consistent with the metric $d$.
\end{df}
In the sequel, $M$ will be a connected Riemannian Hilbert manifold and $\nabla^*$ its natural Levi-Civita connection.
Using this connection, we can define the parallel transport:
\begin{df}[Parallel transport along a curve] For a $C^1$-curve $\gamma : [0, T] \rightarrow M$ and $v \in T_{\gamma(0)}M$, the parallel transport of $v$ along $\gamma$ is the unique vector field $t \mapsto V(t)$ satisfying
$$ V(t) \in T_{\gamma(t)}M \textmd{ and } \left\{
\begin{array}{l}
{\nabla^\star}_{\dot \gamma(t)} V(t)=0 \\
V(0)=v.
\end{array} \right. $$
\end{df}
\begin{df}[Geodesics] The geodesics are defined to be (locally) the shortest path between two points. As a consequence, a geodesic $\gamma$ is also a curve such that parallel transport along it preserves the tangent vector to the curve, so
$$ \nabla^\star_{\dot\gamma} \dot\gamma= 0.$$
\end{df}
For any point $x\in M$ and for any vector $v\in T_xM$, there exists a unique geodesic $\gamma_{x,v} \, : I \rightarrow M$ such that
$$ \left\{\begin{array}{ll}
\gamma_{x,v}(0) & = x \\
\dot\gamma_{x,v}(0) & = v,
\end{array}
\right. $$
where $I$ is a maximal open interval in $\mathbb R$ containing $0$ (depending on $x$ and $v$).
\begin{df} The Riemannian Hilbert manifold $M$ is said geodesically complete if
every maximal geodesic is defined on $\mathbb R$ (i.e. $I=\mathbb R$).
\end{df}
From now on, we assume that the connected Riemannian Hilbert manifold $M$ is geodesically complete.
\begin{df}[Exponential map]
If $M$ is geodesically complete, we recall that the exponential map at $x$, denoted by $\exp_x$, is defined in the whole tangent space $T_xM$ and maps each vector $v$ of the tangent space
$T_xM$ to the end of the geodesic segment $[x,y]$ in $M$ issuing from
$x$ in the direction $v$ and having length $d(x,y) = |v |$.
\label{def:expmap}
\end{df}
\begin{df}[Injectivity radius] We define the injectivity radius of
$M$ at $x$, $i_M(x)$ as the supremum of the numbers $r>0$ of all balls
$B(0_x,r)$ in $T_xM$ for which $\exp_x$ is a diffeomorphism from
$B(0_x,r)$ onto $B(x,r)$. Then, we denote the global injectivity
radius of $M$ by
$$ i(M):= \inf_{x\in M} i_M(x).$$
\end{df}
\begin{df}[Convexity radius] The convexity radius $c_M(x)$ of a point $x\in M$ is the supremum of the numbers $r>0$ such that the ball $B(x,r)$ is convex (in the sense that given two points $p,q\in B(x,r)$, there exists a unique geodesic in $B(x,r)$ joining them and of length $d(p,q)$).
We denote the global convexity radius by
$$ c(M):= \inf_{x\in M} c_M(x).$$
\end{df}
\noindent Similarly we define for a bounded subset $\mathcal{B}$ of $M$, $$i(\mathcal{B}):= \inf_{x\in \mathcal{B}} i_M(x) \quad \textrm{and} \quad c(\mathcal{B}) : = \inf_{x\in\mathcal{B} } c_M(x).$$
\noindent Note that for $x \in \mathcal{B}$, the open ball $B(x, c(\mathcal{B})) $ is not in general included in ${\mathcal{B}}$. \\
\noindent Let ${\mathcal{B}}$ a bounded subset of $M$ and $x$, $y \in {\mathcal{B}}$ satisfying $d(x,y) < c({\mathcal{B}}) $, we define $\Gamma_{x,y}\in T_xM$ as the unique vector $w\in T_xM$ satisfying
$$ \gamma_{x,w}(1)=y.$$
We can note that its norm is equal to the distance between
$x$ and $y$: \be{normgam} |\Gamma_{x,y}| = d(x,y).\end{equation}
According to Definition \ref{def:expmap}, we have $y= \exp_x(\Gamma_{x,y})$. In other words, the vector $\Gamma_{x,y}$ could be also expressed
as follows: $$ \Gamma_{x,y} =
\exp_x^{-1}(y).$$
\begin{df}[Sectional curvature] Let $x \in M $ and $\sigma_x \subset T_{x}M$ a two-dimensional plane, the sectional curvature $K(\sigma_x)$ is the Gaussian curvature of the section of $M$ whose tangent space at $x$ is $\sigma_x$
(in other words, it is the Gaussian curvature of the range of $\sigma_x$ under the exponential map at $x$).
\end{df}
Since the interesting notions have been recalled, we can now specify the hypotheses about the manifold.
\begin{ass} \label{ass:ic}
We assume that $M$ is a connected Riemannian Hilbert manifold which is metrically complete and geodesically complete and $M$ has a locally bounded geometry : for every bounded subset $\mathcal{B} \subset M$,
\begin{itemize}
\item $i(\mathcal{B} ) >0$ and $c(\mathcal{B})>0$
\item the sectional curvatures are bounded in $\mathcal{B} $: $ |K(\mathcal{B})| < + \infty $
$$ K(\mathcal{B}) : = \underset{ \begin{array}{c}x\in \mathcal{B} \\ \sigma_x \subset T_xM \end{array}}{\sup}K( \sigma_x) $$
\end{itemize}
Thus we define for a bounded subset $\mathcal{B} \subset M$,
\be{rho0} \rho(\mathcal{B}):= \min\{i(\mathcal{B}),c(\mathcal{B}), \pi/(2\sqrt{|K(\mathcal{B})|})\}. \end{equation}
Moreover the exponential map $\exp$ is supposed to be locally smooth on $TM$ and the inverse map $D\exp^{-1}_x:TM \rightarrow T_xM $ is supposed to be locally Lipschitz: for every bounded subset $\mathcal{B} \subset M$ and for $k \in\{0, 1,2\}$, there exist $C_e({\mathcal B }) >0 $ and $L(\mathcal{B}) >0 $ such that
\begin{equation} \|\exp\|_{C^k({\mathcal B})} \leq L(\mathcal{B}) \label{eq:exp} \end{equation}
and for all
$ x,y,z\in{\mathcal B}$ satisfying $ d(x, y) < \rho({\mathcal{B}})$ and $ d(x, z) < \rho({\mathcal{B}})$, for all $v\in T_yM, w\in T_zM $
$$
\begin{array}{l}
| \exp^{-1}_x (y) -\exp^{-1}_x(z) |_{T_xM} \leq C_e({\mathcal B })d(y,z), \vspace{6pt} \\
| D\exp^{-1}_x(y)[v] -D\exp^{-1}_x(z)[w] |_{T_xM} \leq C_e({\mathcal B })d_{TM}((y,v),(z,w)),
\end{array}
$$
where $d_{TM}$ represents the distance on the tangent bundle:
$$ d_{TM}((y,v),(z,w)) := d(y,z)+ |L_{y\rightarrow z} v - w|_{T_zM}$$
and $L_{y\rightarrow z}$ is the tangential parallel transport from $y$ to $z$ (see Definition \ref{def:partrans}).
\end{ass}
\medskip
\begin{rem}
In the finite dimensional case, it suffices to assume that the connected Riemannian manifold $M$ is geodesically complete because in that case Hopf-Rinow's Theorem asserts that $M$ is automatically metrically complete and the closed and bounded subsets of $M$ are compact. Since the functions $i_M$, $c_M$ are continuous, the quantities $i(\mathcal{B})$ and $c(\mathcal{B})$ are necessarily strictly positive. In the same way, $|K(\mathcal{B})| < +\infty $ is also satisfied.
\end{rem}
We aim to finish this section with the following definition:
\begin{df}[Local parallel transport]
Let $\mathcal{B} \subset M$ be a bounded subset and let $x,y\in \mathcal{B} $ two different points satisfying $d(x,y) \leq\rho (\mathcal{B}) \leq c({\mathcal{B}})$, then there exists a unique geodesic curve $\gamma$ with $\gamma(0)=x$ and $\gamma(1)=y$ and so $ \Gamma_{x,y} $ is well-defined.
Moreover, we denote by $L_{x\rightarrow y}:T_xM \rightarrow T_yM$ the tangential parallel transport defined as follows: for all $v\in T_xM$, $L_{x\rightarrow y}(v)=V(1)$ where $V$ is the unique vector field satisfying $V(0)=v$ and $ {\nabla^\star}_{\dot \gamma(t)} V(t)=0$. The map $L_{x\rightarrow y}$ also defines a linear isometry between $T_xM$ and $T_yM$ with $L_{x\rightarrow y}^{-1}=L_{y\rightarrow x}$.
\noindent Moreover it satisfies
\be{eq:symmetry} L_{x\rightarrow y}\left(\Gamma_{x,y}\right) = -\Gamma_{y,x}. \end{equation}
\label{def:partrans}
\end{df}
\section{Technical preliminaries about convex analysis and the squared distance} \label{sec:pre2}
In this section, we aim to extend the notion of {\it prox-regular subset} to the setting of Riemannian Hilbert manifold and to understand how the main property of {\it hypomonotonicity} (of the proximal normal cone) is modified (see Theorem \ref{thm:hypo}).
In all the sequel we consider a manifold $M$ which satisfies Assumption \ref{ass:ic}.
We first recall some properties about the smoothness of the squared distance:
\begin{prop}[Proposition 2.2 \cite{AF}] For every bounded subset $\mathcal{B}$ of $M$, there exists $C_{\rho(\mathcal{B})} >0$ such that
for all $p,q \in \mathcal{B} $ satisfying $d(p,q)\leq \rho(\mathcal{B}) $:
\be{hyp:H} \|H_ x d^2(p,q) \|_p \leq C_{\rho(\mathcal{B})}, \end{equation}
where $$H_ x d^2(p,q) := D^2 \phi(p)$$ with $$\phi(p):= d^2(p,q) \textmd{ and } \| \Psi \|_p := \underset {v \in T_pM, |v|=1} {\sup} |\Psi [v,v]|.$$
\label{propH}
\end{prop}
\begin{lem} \label{lem:geodesic} For all bounded set $ \mathcal{B}$ included in
$M$, for all $ \tau \in [0,1]$ and $x,y\in \mathcal{B}$ satisfying $d(x,y) < \rho(\mathcal{B}) $, we have
\be{hyp:geod} |\ddot{\gamma}_{x,\Gamma_{x,y} }(\tau) | \leq L(\mathcal{B})
|\Gamma_{x,y}|^2 = L(\mathcal{B}) d(x,y)^2, \end{equation}
where $ {\gamma}_{x,\Gamma_{x,y} }(\tau) = \exp_x(\tau \Gamma_{x,y}) = \exp_x(\tau \, \exp_x^{-1} ( y))$.
\end{lem}
\begin{proof} Let $x,y\in \mathcal{B}$ and consider for $\tau \in[0,1]$
$$ \gamma_{x,\Gamma_{x,y}}(\tau) = \exp_x(\tau \Gamma_{x,y}).$$
By differentiating in $\tau$, we get
$$ \ddot \gamma_{x,\Gamma_{x,y}}(\tau) = D^2 \exp_x (\tau \Gamma_{x,y}) [\Gamma_{x,y},\Gamma_{x,y}].$$
Then the assumption \eqref{eq:exp} gives the expected result.
\end{proof}
As a particular case of \cite[Theorem 1.2]{Karcher} (with a constant function $f$), we recall the link between the exponential map and the gradient of the squared distance.
\begin{lem} \label{lemb} For every bounded set ${\mathcal{B}} \subset M $, for all $x,y\in {\mathcal{B}}$ satisfying $d(x,y) < \rho({\mathcal{B}})$, we have
\begin{equation}
\label{eq:gamma} \nabla_x d^2(x,y) = 2 d(x,y) \nabla_x d(x,y)= -2 \Gamma_{x,y}=-2 \exp_x^{-1}(y) .
\end{equation}
\end{lem}
\noindent For the sake of readability, we give a short proof.
\begin{proof} By definition of $\rho({\mathcal{B}})$, there exists a unique geodesic $\gamma_{x,\Gamma_{x,y}}$ such that $\gamma_{x,\Gamma_{x,y}}(1)=y$. For all $t\in[0,1)$, the constant velocity along the geodesic implies that
$$ d(\gamma_{x,\Gamma_{x,y}}(t),y) = d(x,y) -t|\Gamma_{x,y}| = (1-t) d(x,y).$$
By differentiating at $t=0$, we get
$$\langle \nabla_x d(\gamma_{x,\Gamma_{x,y}}(0),y), \dot \gamma_{x,\Gamma_{x,y}}(0) \rangle = \langle \nabla_x d(x,y), \Gamma_{x,y}\rangle = -d(x,y).$$
From $|\nabla_x d(x,y)|=1$ and $|\Gamma_{x,y}|=d(x,y)$, we deduce (\ref{eq:gamma}).
\end{proof}
\begin{rem}
With the notations of Proposition \ref{propH}, $\nabla_x d^2(p,q)= \nabla \phi (p)$. We deduce from this proposition and from (\ref{eq:gamma}) that the map $p \mapsto \exp_p^{-1}(q) $ is smooth on ${\mathcal{B}}$.
\label{rem:smoothexp}
\end{rem}
\begin{rem}
In the same way, $$\nabla_y d^2(x,y) = 2 d(x,y) \nabla_y d(x,y)= -2 \Gamma_{y,x}=-2 \exp_y^{-1}(x).$$
\label{rem:nablay}
\end{rem}
\begin{df} Let $C$ be a closed subset of $M$ and $x\in C$. The proximal normal cone to $C$ at $x$ is defined as follows:
$$ \mathrm{N}(C,x):=\left\{ v\in T_xM,\ \exists \varepsilon>0,\ x\in \mathrm P_C(\gamma_{x,v}(\varepsilon))\right\},$$
where $\mathrm P_C$ is the projection onto $C$
$$ \mathrm P_C(y):= \left\{z\in C,\ d(y,z)=d_C(y)=\inf_{c\in C} d(y,c) \right\}.$$
\label{def:cone}
\end{df}
\noindent Note that $\mathrm{N}(C,x) $ is a cone because for all $\alpha >0$, for all $
v \in \mathrm{N}(C,x) $, $\alpha v $ belongs to $\mathrm{N}(C,x) $ due to the
equality $\gamma_{x,v}(\varepsilon) = \gamma_{x,\alpha v}\left(\frac{\varepsilon}{\alpha}\right) $.
\begin{df} Consider a nonempty closed subset $C\subset M$, $ C$ is said locally prox-regular if for all bounded set $\mathcal{B} \subset M $, there exists $ 0 <\eta_{\mathcal{B}} \leq \rho({\mathcal{B}})$ such that for every $t\in[0,\eta_{\mathcal{B}})$, $x\in C \cap {\mathcal{B}}$ and $v\in \mathrm{N}(C,x) \setminus \{0\}$
$$ x\in \mathrm P_C\left(\gamma_{x,\frac{v}{|v|}}(t) \right).$$
\label{def:prox}
\end{df}
\begin{rem}
Note that we can define a uniformly prox-regular set : a closed subset $ C\subset M $ will be called $\eta
$-prox-regular or uniformly prox-regular with constant $\eta$ if for all $t\in[0,\eta)$, $x\in C$ and $v\in \mathrm{N}(C,x) \setminus \{0\}$,
$$ x\in \mathrm P_C\left(\gamma_{x,\frac{v}{|v|}}(t) \right).$$
However this definition is meaningful if $ i(M) > 0$, $c(M)> 0 $ and $|K(M)| < + \infty$ (for example il the framework of compact manifolds).
\label{rem:unifprox}
\end{rem}
\noindent Now we want to prove that the local prox-regularity provides a kind of hypomonotonicity property. This result is the purpose of the following lemma.
\begin{lem} \label{lem:hypo} Let $C$ be a closed subset of $M$ and assume that $C$ is locally prox-regular.
Then for all bounded subset $\mathcal{B}$ of
$M$, there exists a constant $E_\mathcal{B} >0$ such that for all $x,y\in C
\cap \mathcal{B}$ satisfying $d(x,y) \leq \rho({\mathcal{B}})$ and $v\in \mathrm{N}(C,x)$
$$ \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq E_\mathcal{B} |v| d(x,y)^2.$$
\end{lem}
\begin{proof} Let $x,y\in C \cap \mathcal{B}$ and $v\in\mathrm{N}(C,x)$. We can assume $|v|=1$ without loss of generality. We define $\tilde{{\mathcal{B}}}:=\{m \in M, d(m, {\mathcal{B}})\leq \eta_{\mathcal{B}}\}$ (where $\eta_{\mathcal{B}}$ is introduced in Definition \ref{def:prox}) and we set $r:=\min(\eta_{\mathcal{B}} ,\rho({\tilde{{\mathcal{B}}}}) )$.
If $d(x,y)\geq r/2 $, the result is easily proved by (\ref{normgam}) with $E_\mathcal{B}=2/r$. \\
Consider also $x,y$ with $d(x,y) < r/2$. Since $C$ is locally prox-regular, for all $t\in(0,r/2)$
$$ x\in \mathrm P_C\left(\gamma_{x, v}(t) \right)$$
which implies
$$ t^2 = d(x, \gamma_{x, v}(t))^2 \leq d(y, \gamma_{x, v}(t))^2.$$
By a second order expansion and setting $m:=\gamma_{x, v}(t)$, we get thanks to (\ref{eq:gamma})
\begin{align*}
d(y,m)^2 & = d(x,m)^2 + \int_0^1 \left[ \frac{d}{d\tilde s} d ^2 (\gamma_{x, \Gamma_{x,y}}(\tilde s),m) \right] _{|\tilde s = s}ds \\
& = d(x,m)^2 + 2 d(x,m)\langle \nabla_x d(x,m),\Gamma_{x,y}\rangle \\
& \quad + \int_0^1 \int_0^s \left[\frac{d^2}{d\tilde s^2} d ^2 (\gamma_{x, \Gamma_{x,y}}(\tilde s),m)\right] _{|\tilde s = \tau} d\tau ds.
\end{align*}
Consequently,
\begin{align*}
-2 d(x,m)\langle \nabla_x d(x,m,\Gamma_{x,y}\rangle & \leq \int_0^1 \int_0^s \left[\frac{d^2}{d\tilde s^2} d ^2 (\gamma_{x, \Gamma_{x,y}}(\tilde s),m)\right] _{|\tilde s = \tau} d\tau ds \\
& \leq \int_0^1 \int_0^s \left[ H_x d^2(\gamma_{x, \Gamma_{x,y}}(\tau),m)[\dot \gamma_{x,\Gamma_{x,y}}(\tau),\dot \gamma_{x,\Gamma_{x,y}}(\tau)] \right. \\
& \qquad \qquad \left. + \langle \nabla_x d^2(\gamma_{x,
\Gamma_{x,y}}(\tau),m), \ddot \gamma_{x,\Gamma_{x,y}}(\tau)
\rangle \right] d\tau ds.
\end{align*}
Since $d(x,\gamma_{x,\Gamma_{x,y}}(\tau)) \leq d(x,y) < r/2 $ and $d(x, \gamma_{x,v}(t)) \leq t\leq r/2$, it is clear that $\gamma_{x,\Gamma_{x,y}}(\tau) $ and $\gamma_{x,v}(t)$ belong to $B(x,r/2)\subset \tilde{{\mathcal{B}}} $.
Moreover\be{eq:di} d(\gamma_{x, \Gamma_{x,y}}(\tau),m)=
d(\gamma_{x, \Gamma_{x,y}}(\tau),\gamma_{x,v}(t))<r/2+ r/2< r \leq \rho(\tilde{{\mathcal{B}}}). \end{equation} Thus by
Proposition \ref{propH}, $$| H_x d^2(\gamma_{x, \Gamma_{x,y}}(\tau),m)[\dot
\gamma_{x,\Gamma_{x,y}}(\tau),\dot \gamma_{x,\Gamma_{x,y}}(\tau)] |\leq
C_{\rho(\tilde{{\mathcal{B}}})} |\Gamma_{x,y}|^2 .$$
Furthermore Lemma \ref{lem:geodesic} with (\ref{eq:di}) gives $$\langle \nabla_x d^2(\gamma_{x,
\Gamma_{x,y}}(\tau),m), \ddot \gamma_{x,\Gamma_{x,y}}(\tau)
\rangle \leq 2 d(\gamma_{x,
\Gamma_{x,y}}(\tau),m) L(\tilde{{\mathcal{B}}}) |\Gamma_{x,y}|^2\leq 2 r L(\tilde{{\mathcal{B}}})
|\Gamma_{x,y}|^2. $$
Thus by adding the previous inequalities we obtain
\begin{align*}
-2 d(x,m)\langle \nabla_x d(x,m),\Gamma_{x,y}\rangle & \leq C_{\rho(\tilde{{\mathcal{B}}})}
|\Gamma_{x,y}|^2 + 2 r L(\tilde{{\mathcal{B}}}) |\Gamma_{x,y}|^2.
\end{align*}
So we deduce from Lemma \ref{lemb} and (\ref{normgam}) that
$$ \langle \exp_x^{-1}( \gamma_{x,v}(t) ), \Gamma_{x,y}\rangle \leq
\frac{C_{\rho(\tilde{{\mathcal{B}}})}+2 r L(\tilde{{\mathcal{B}}})}{2} d(x,y)^2.$$
As $\exp_x^{-1}( \gamma_{x,v}(t) ) = t v $, it comes
$$ \langle v, \Gamma_{x,y}\rangle \leq
\frac{C_{\rho(\tilde{{\mathcal{B}}})}+2 r L(\tilde{{\mathcal{B}}}) }{ 2t} d(x,y)^2.$$
This inequality holds for all $t \in (0 , \ r/2)$ which ends the proof by
setting $E_\mathcal{B}:= \max(C_{\rho(\tilde{{\mathcal{B}}})}/r+2 L(\tilde{{\mathcal{B}}}), 2/r) .$
\end{proof}
\begin{rem}
Note that the constants $E_{\mathcal{B}}$ depend on the geometrical features of $M$ but also on $C$ through the constant $\eta_{\mathcal{B}}$.
\label{rem:constantE}
\end{rem}
\begin{rem} \label{rem:hypo} Let $C$ be a closed subset of $M$ and $x \in C$. If $v \in \mathrm{N}(C,x)$ then there exists $\alpha>0$ and $\Lambda>0$ such that for all $y\in C
\cap B(x,\alpha)$
$$ \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq \Lambda |v | d(x,y)^2.$$
Indeed, assume that $|v|= 1$ (without loss of generality) since $v \in \mathrm{N}(C,x)$, there exists $\varepsilon >0$, such that $x \in \mathrm P_C(\gamma_{x, v}(\varepsilon))$ by Definition \ref{def:cone}. Now it suffices to fix ${\mathcal{B}}= B(x, \varepsilon)$ and take $\alpha := \min(\varepsilon, \rho({\mathcal{B}})/2)$. Thus for all $t \leq \alpha $, $ x \in \mathrm P_C(\gamma_{x, v}(t)) $. By following the previous proof, we can choose $\Lambda:= \frac{C_{\rho({\mathcal{B}})}}{2 \alpha} + 2 L({\mathcal{B}}) $. Here $\alpha$ and $\Lambda$ depend on $v$ and may degenerate. The important property of a locally prox-regular set is that these two quantities may be assumed to be independent on $v$, as proved in Lemma \ref{lem:hypo}.
\end{rem}
The following statement will be useful in Lemma \ref{lem:hypo2} to check that the hypomonotonicity property (which is detailed in Lemma \ref{lem:hypo}) implies the local prox-regularity. Moreover its corollary will allow us to show that the projection onto a locally prox-regular set is single-valued in its neighborhood.
\begin{lem}[{\cite[Theorem 1.2]{Karcher}}] \label{lem:curvature} Consider a manifold $M$ satisfying Assumption \ref{ass:ic} and $\mathcal{B}$ a bounded subset of $M$.
\begin{itemize}
\item If $K(\mathcal{B}) \leq 0$
then for every $x\in {\mathcal{B}}$ and along any geodesic $\gamma$ taking
values in $B(x, r) \subset {\mathcal{B}} $ with $r \leq c(\mathcal{B})$,
$$ \frac{d^2}{dt^2} d^2(\gamma(t),z) \geq 2 |\dot \gamma(t)|^2.$$
\item If $ 0< K(\mathcal{B}) \leq \delta $,
with $S(t):= t \cot (t)$ and \hbox{$r \in \left] 0, \rho(\mathcal{B}) \right]$}, we have for all
$x\in {\mathcal{B}}$ and along any geodesic $\gamma$ taking values in $B(x, r) \subset {\mathcal{B}} $,
$$ \frac{d^2}{dt^2} d^2(\gamma(t),z) \geq 2 S(r \delta^{1/2} ) |\dot \gamma(t)|^2.$$
\end{itemize}
\end{lem}
\begin{cor} \label{cor:new}
Let $\mathcal{B}$ a bounded subset of $M$ containing $x $, $ z_1$ and $z_2$. Suppose that
\begin{itemize}
\item $z_1,z_2 \in B(x, r) \subset {\mathcal{B}}$ with $r \leq c(\mathcal{B})$, if $K(\mathcal{B}) \leq 0$;
\item $z_1,z_2 \in B(x, r) \subset {\mathcal{B}} $ with $r < \rho(\mathcal{B})$ if $0 <K(\mathcal{B}) \leq \delta$ .
\end{itemize}
then there exists $A:=A({\mathcal{B}}) >0 $ such that
$$ \langle \Gamma_{z_2,x} - L_{z_1 \to z_2} (\Gamma_{z_1,x}) , \Gamma_{z_2,z_1} \rangle \geq A d(z_1,z_2)^2.$$
\end{cor}
\begin{proof} We set $\varphi(z):= d^2(z,x)$. Then with a second order expansion of the function $d^2$ and \eqref{eq:gamma},
we have
$$ \varphi(z_1)-\varphi(z_2) = \langle \nabla \varphi(z_2), \exp^{-1}_{z_2}(z_1) \rangle + \int_0^1 \frac{d^2}{dt^2} \varphi(\gamma(t)) dt$$
where $\gamma(t)$ is the geodesic between $\gamma(0)=z_2$ and $\gamma(1)=z_1$.
With $A=2$ (in the case of negative curvature) or $A= 2S(r \delta^{1/2} )$ (if the curvature is at most $\delta$), Lemma \ref{lem:curvature} implies
$$ \varphi(z_1)-\varphi(z_2) \geq \langle \nabla \varphi(z_2), \exp^{-1}_{z_2}(z_1) \rangle + A d(z_1,z_2)^2,$$
which can be written as
$$ \varphi(z_1)-\varphi(z_2) \geq -2 \langle \Gamma_{z_2,x}, \Gamma_{z_2,z_1} \rangle + A d(z_1,z_2)^2.$$
By symmetry, we have
$$ \varphi(z_2)-\varphi(z_1) \geq -2 \langle \Gamma_{z_1,x}, \Gamma_{z_1,z_2} \rangle + A d(z_1,z_2)^2.$$
By summing these two inequalities, it comes
$$ \langle \Gamma_{z_2,x}, \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} + \langle \Gamma_{z_1,x}, \Gamma_{z_1,z_2} \rangle_{T_{z_1} M} \geq A d(z_1,z_2)^2.$$
Using parallel transport along the geodesic from $z_1$ to $z_2$,
$$ \langle \Gamma_{z_1,x}, \Gamma_{z_1,z_2} \rangle_{T_{z_1} M} = -\langle L_{z_1 \to z_2} \Gamma_{z_1,x}, \Gamma_{z_2,z_1} \rangle_{T_{z_2} M},$$
which is the expected result.
\end{proof}
\begin{lem} \label{lem:hypo2} Let $C$ be a closed subset of $M$. Assume that for every bounded subset ${\mathcal{B}}$ of $M$, there exists a constant $\kappa_{\mathcal{B}} >0$ such that for all $x,y\in C \cap {\mathcal{B}}$ satisfying $d(x,y) < \rho({\mathcal{B}})$ and every $v\in \mathrm{N}(C,x)$
\be{eq:hypo} \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq \kappa_{\mathcal{B}} |v| d(x,y)^2.\end{equation}
Then there exists $\theta^{\mathcal{B}} \in ]0, \rho({\mathcal{B}})[ $ such that for all $t <\theta^{\mathcal{B}} $ and for all $v \in \mathrm{N}(C,x) \setminus \{ 0 \}$,
$$ x \in \mathrm P_C\left(\gamma_{x, \frac{v}{|v|}} (t)\right). $$
In other words, $C$ is locally prox-regular.
\end{lem}
\begin{proof} Let ${\mathcal{B}}$ be a bounded subset of $M$, let us fix $x\in C\cap {\mathcal{B}}$ and $v\in \mathrm{N}(C,x)$. Without loss of generality assume that $|v|=1$.
We define $\tilde{{\mathcal{B}}} :=\{ m\in M, d(m,{\mathcal{B}}) < \rho({\mathcal{B}})\}$ and we recall that $ \rho(\tilde{{\mathcal{B}}})\leq\rho({\mathcal{B}})$.
Moreover we set $\theta^{\mathcal{B}} := \min\left(\frac{1}{2\kappa_{\tilde{{\mathcal{B}}} }} ,\frac{\rho(\tilde{{\mathcal{B}}} )}{3} \right)$ if $ K(\tilde{{\mathcal{B}}} ) \leq 0$ and $ \theta^{\mathcal{B}} := \min\left(\frac{\rho(\tilde{{\mathcal{B}}} )}{3}, \frac{1}{3\delta^{1/2}} \arctan(\frac{3 \delta^{1/2}}{ 2\kappa_{\tilde{{\mathcal{B}}} }})\right) $ if $0 <K(\tilde{{\mathcal{B}}} ) \leq \delta$.
We are going to check that for $t < \theta^{\mathcal{B}}< \rho(\tilde{{\mathcal{B}}})/3$
\be{eq:distamontrer} d_C(\gamma_{x,v}(t)) = t. \end{equation}
We assume that (\ref{eq:distamontrer}) does not hold and then there exists $z \in C$ ($z \neq x$) satisfying
$$ d(z,\gamma_{x,v}(t))^2<t^2$$
and in using a second order expansion, we get
\begin{align*}
t^2 > d(z,\gamma_{x,v}(t))^2=t^2+\langle \nabla_x
d^2(x,\gamma_{x,v}(t)), \Gamma_{x,z} \rangle \\ + \int_0^1 \int_0^\sigma
\left[ \frac{d^2}{ds^2}
d^2(\gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t))\right ]_{|s=\tau}
d\tau d\sigma.
\end{align*}
Note that $\Gamma_{x,z}$ is well-defined since $d(x,z) < d(x, \gamma_{x,v}(t) ) + d(z,
\gamma_{x,v}(t) ) <2t < \rho(\tilde{{\mathcal{B}}})$.
As a consequence, $z\in \tilde{{\mathcal{B}}} $ and $$\begin{array}{l}
\displaystyle \int_0^1 \int_0^\sigma
\left[\frac{d^2}{ds^2} d^2(\gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t))
\right ]_{|s=\tau} d\tau d\sigma \vspace{6pt} \\
\hspace{1cm} \leq \displaystyle \langle - \nabla_x
d^2(x,\gamma_{x,v}(t)), \Gamma_{x,z} \rangle = \langle 2
\exp_x^{-1}(\gamma_{x,v}(t)), \Gamma_{x,z} \rangle= \langle 2 t v,
\Gamma_{x,z} \rangle,
\end{array}$$
thanks to Lemma \ref{lemb}. Thus with assumption (\ref{eq:hypo}), it comes
$$ \int_0^1 \int_0^\sigma
\left[\frac{d^2}{ds^2} d^2(\gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t))
\right ]_{|s=\tau} d\tau d\sigma \leq 2 t \kappa_{\tilde{{\mathcal{B}}}} d(x,z)^2.$$
Moreover
\begin{align*}
d( \gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t) ) & \leq d(x,
\gamma_{x,v}(t) ) + d(x, \gamma_{x,\Gamma_{x,z}}(s)) \leq t |v |+ s
|\Gamma_{x,z} | \\
& \leq t + d(x,z) \\ &\leq 3t \leq \rho(\tilde{{\mathcal{B}}}) \leq c(\tilde{{\mathcal{B}}}) .
\end{align*}
If $ K(\tilde{{\mathcal{B}}})\leq 0$ everywhere, Lemma
\ref{lem:curvature} yields for $t \leq \frac{\rho(\tilde{{\mathcal{B}}})}{3} \leq \frac{c(\tilde{{\mathcal{B}}})}{3}$
$$ \left[ \frac{d^2}{ds^2} d^2(\gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t)) \right ]_{|s=\tau} \geq 2|\dot \gamma_{x,\Gamma_{x,z}}(\tau) |^2 = 2|\Gamma_{x,z}|^2 = 2d(x,z)^2.$$
Consequently, we have for $t \leq \frac{\rho(\tilde{{\mathcal{B}}})}{3} \leq \frac{c(\tilde{{\mathcal{B}}})}{3}$
$$ d(x,z)^2 \leq 2t \kappa_{\tilde{{\mathcal{B}}}} d(x,z)^2$$
which is impossible for $t<\frac{1}{2\kappa_{\tilde{{\mathcal{B}}}}}$.
Thus for $t <\theta^{\mathcal{B}}$, $d_C(\gamma_{x,v}(t))=t $ so $x\in \mathrm P_C(\gamma_{x,v}(t)) $.
If $0 <K(\tilde{{\mathcal{B}}}) \leq \delta$, Lemma
\ref{lem:curvature} yields for $t \in \left]0, \frac{\rho(\tilde{{\mathcal{B}}})}{3} \right]$
$$ \left[ \frac{d^2}{ds^2}
d^2(\gamma_{x,\Gamma_{x,z}}(s),\gamma_{x,v}(t)) \right ]_{|s=\tau}
\geq 2 S(3t \delta^{1/2})d(x,z)^2 ,$$
for every $\tau\in [0,s]$ and every $s \in [0,1]$.
Hence, we get
$$ d(x,z)^2 \leq \frac{ 2\kappa_{\tilde{{\mathcal{B}}}} t }{ S(3t \delta^{1/2})}d(x,z)^2
=\frac{2 \kappa_{\tilde{{\mathcal{B}}}} \tan(3t \delta^{1/2})}{3 \delta^{1/2}} d(x,z)^2$$
which is impossible as soon as $t < \theta^{\mathcal{B}}$.
\end{proof}
Combining Lemmas \ref{lem:hypo} and \ref{lem:hypo2}, we obtain the following characterization of locally prox-regular subsets:
\begin{thm}[Hypomonotonicity property] \label{thm:hypo} Let $C$ be a closed subset of $M$. Then $C$ is locally prox-regular if and only if for all bounded subset ${\mathcal{B}}$ of
$M$, there exists a constant $E_{\mathcal{B}} >0$ such that for all $x,y\in C
\cap {\mathcal{B}}$ with $d(x,y) \leq \rho({\mathcal{B}})$ and $v\in \mathrm{N}(C,x)$
$$ \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq E_{\mathcal{B}} |v| d(x,y)^2.$$
\end{thm}
As mentioned in the introduction, the notions of proximal normal cone and of prox-regularity (at a point) have been already appeared in \cite{HP}. Let us recall them.
\begin{df}[\cite{HP}]
For $C$ a closed subset of $M$ and $x\in C$ the proximal normal cone to $C$ at $x$ is defined by
$$ N^P_C(x):= N^P_{\exp^{-1}_x (U\cap C)} (0_x)$$
with $U := \exp_x(V)$, where $V$ is an open neighborhood of $0_x$. \\
Moreover the limiting normal cone to $C$ at $x$ is expressed as follows:
$$N^L_C(x):=\left\{ \lim_{i \to \infty} v_i : v_i \in N^P_C(x_i) \textmd{ and } x_i \to x \textmd{ with } x_i \in C\right\}. $$
\end{df}
\begin{df}\cite[Definition 3.3]{HP}
A closed subset $C\subset M$ is said to be prox-regular at $\bar{x} \in C$ if there exist $\epsilon>0$ and $\alpha>0$ such that $B(\bar{x}, \epsilon)$ is convex and for every $x\in C\cap B(\bar{x}, \epsilon)$, $y\in C \cap B(\bar{x},\epsilon)$ and $v\in N^L_C(x)$ with $|v|\leq \epsilon$ we have
$$ \langle v , \exp^{-1}_x(y) \rangle \leq \frac{\alpha}{2} d(x,y)^2.$$
\label{def:proxHP}
\end{df}
First we would like to show that the two notions of proximal normal cones are equivalent. This result is based on the following proposition.
\begin{prop} \label{prop:hypomo} Let $C$ be a closed subset of $M$ and ${\mathcal{B}}$ a bounded subset of $M$. Consider $x\in C \cap {\mathcal{B}}$ and $v\in T_xM$. Then $v\in \mathrm{N}(C,x)$ if and only if there exist $ \theta \in (0, \rho({\mathcal{B}}))$ and $\Lambda >0$ such that for all $y\in C \cap {\mathcal{B}}$ with $d(x,y)\leq \theta$
\be{eq:hypobis} \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq \Lambda d(x,y)^2.\end{equation}
\end{prop}
\begin{proof} Let $v \neq 0$ satisfying (\ref{eq:hypobis}). For all $ y \in C \cap {\mathcal{B}}$ with $d(x,y)\leq \theta$, we have $$ \langle v, \Gamma_{x,y} \rangle_{T_xM} \leq \kappa |v| d(x,y)^2 ,$$ where $\kappa:= \frac{\Lambda}{|v|}$. Then it suffices us to check that for some small enough $t>0$,
$$ x\in \mathrm P_{C}(\gamma_{x,v}(t)),$$
which is equivalent to
$$ d_C(\gamma_{x,v}(t)) = t|v|.$$
Let assume that it does not hold: $d_C(\gamma_{x,v}(t)) < t|v|$ and choose $z$ satisfying $ d(z,\gamma_{x,v}(t) ) < t|v|$. Following the proof of Lemma \ref{lem:hypo2}, it comes a contradiction for small enough $t$ ($t<\theta^{\mathcal{B}}$). In conclusion, $v \in \mathrm{N}(C,x)$. \\
Furthermore, if $v\in \mathrm{N}(C,x)$ then Remark \ref{rem:hypo} yields that (\ref{eq:hypobis}) is verified.
\end{proof}
\begin{cor}
For $C$ a closed subset of $M$ and $x\in C$, then $$N^P_C(x) = \mathrm{N}(C,x).$$
\label{cor:eqcones}
\end{cor}
\begin{proof}
Indeed, \cite[Lemma 2.1]{HP} states that $v\in N^P_C(x)$ if and only if there is $\alpha>0$ such that
$$ \langle v , \exp^{-1}_x(y) \rangle \leq \frac{\alpha}{2} d(x,y)^2, $$
for every $y$ in a neighborhood of $x$. This property is equivalent to (\ref{eq:hypobis}). We conclude to the equality of these cones thanks to Proposition \ref{prop:hypomo}.
\end{proof}
Now let us specify the link between the different definitions of prox-regularity.
\begin{prop} \label{pro}
Let $C$ be a closed subset of $M$. If $C$ is locally prox-regular (according to Definition \ref{def:prox} ) then $C$ is prox-regular at any point $\bar{x} \in C$ (in the sense of Definition \ref{def:proxHP} ).
\end{prop}
\begin{proof}
Let $\bar{x} \in C $ and ${\mathcal{B}}_0 = B(\bar{x}, 1) $, then $\bar{x} $ belongs to the convex ball $ {\mathcal{B}}= B(\bar{x}, \varepsilon) $, with $\varepsilon= \min(1, \rho({\mathcal{B}}_0)/2) $. Moreover let $x, y \in C\cap {\mathcal{B}}$, it comes $$ d(x,y)\leq 2\varepsilon \leq \rho({\mathcal{B}}_0) \leq \rho({\mathcal{B}}), $$
since ${\mathcal{B}} \subset {\mathcal{B}}_0 $.
Thanks to Lemma \ref{lem:hypo}, for every $v \in N(C,x)$, we have $$ \langle v , \exp^{-1}_x(y) \rangle \leq E_{\mathcal{B}} |v| d(x,y)^2 . $$
Let $w \in N^L_C(x)$ satisfying $|w| \leq \varepsilon$ and $w= \lim w_i$, where $w_i \in N^P_C(x_i)$ with $x_i \in C$ and $x_i \to x$. For $i$ large enough, $x_i \in {\mathcal{B}}$ and $ |w_i| \leq 2 \varepsilon$, hence $$ \langle w_i , \exp^{-1}_{x_i}(y) \rangle_{x_i} \leq 2 E_{\mathcal{B}} \varepsilon d(x_i,y)^2 . $$
Furthermore,
\begin{eqnarray*}
& & \langle w_i , \exp^{-1}_{x_i}(y) \rangle_{x_i} - \langle w, \exp^{-1}_{x}(y) \rangle_{x} \\
& & \hspace{2cm} = \langle L_{x_i\to x} (w_i), L_{x_i\to x}(\exp^{-1}_{x_i}(y)) \rangle_{x} - \langle w, \exp^{-1}_{x}(y) \rangle_{x}\\
& & \hspace{2cm} = \langle L_{x_i\to x} (w_i)- w, \exp^{-1}_{x}(y) \rangle_{x} \\
& & \hspace{4cm} +\langle L_{x_i\to x} (w_i), L_{x_i\to x}(\exp^{-1}_{x_i}(y))- \exp^{-1}_{x}(y) \rangle_{x}\\
& & \hspace{2cm} \to 0.
\end{eqnarray*}
Indeed the second term tends to zero since $ x \mapsto \exp^{-1}_{x}(y)$ is smooth on ${\mathcal{B}}$ thanks to Remark \ref{rem:smoothexp}.
Thus by passing to the limit, the previous inequality becomes:
$$\langle w , \exp^{-1}_{x}(y) \rangle_{x} \leq 2 E_{\mathcal{B}} \varepsilon d(x,y)^2 .$$
So we conclude that $C$ is prox-regular at $\bar{x}$.
\end{proof}
\begin{rem}
In fact, a closed subset $C\subset M$ is prox-regular at $\bar{x} \in C$ according to \cite{HP} if and only if it satisfies a hypomonotonicity property in a convex neighborhood of $\bar{x}$ for every vector of $\mathrm{N}(C,\cdot)=N^P_C$. \\
Moreover, if $C\subset M$ is prox-regular at $x \in C$, it is proved in \cite[Lemma 3.7]{HP} that the limiting cone, the Clarke cone (that is the closed convex hull of the limiting cone) and the proximal normal cone are equal : $N^L_C(x) =N^C_C(x) = N^P_{C}(x) = \mathrm{N}(C,x)$ (where the last equality comes from Corollary \ref{cor:eqcones}).
\end{rem}
Then we aim to end this section by proving the following property (which is also well-known in the Euclidean space): close enough to a locally prox-regular subset, every point has a unique projection onto it.
\begin{thm} \label{thm:projection}
Let $C \subset M$ be a locally prox-regular set. Then for every bounded set ${\mathcal B} \subset M$ satisfying ${\mathcal{B}} \cap C \neq \emptyset $, there exists $\ell({\mathcal{B}}) >0$ such that for any $x\in {\mathcal B}$ with $d_C(x) < \ell({\mathcal{B}})$, the projection set is a singleton:
$$ \mathrm P_C(x) =\{ x^\star\}, $$
with $ x^\star \in C.$
\end{thm}
\begin{proof}
Let ${\mathcal B}_1:= \{ m \in M, d(m, {\mathcal B})\leq 1\}$ and $\ell({\mathcal{B}}) := \min(\rho({\mathcal B}_1)/2, \tau({\mathcal{B}}_1), 1)$ where $\tau({\mathcal{B}}_1) := \frac{A({\mathcal{B}}_1)}{8 E_{{\mathcal{B}}_1}} $, (we recall that these constants are defined in Corollary \ref{cor:new} and in Lemma \ref{lem:hypo}).
Consider $x\in{\mathcal B}$ with $d_C(x) < \ell({\mathcal{B}})$, then for every $\epsilon >0$ satisfying $$d_C(x)+\epsilon < \ell({\mathcal{B}}) \leq \rho({\mathcal{B}}_1)/2$$ we can find $y\in C$ such that
$$ d(x,y)^2 \leq d_C(x)^2+\epsilon^2.$$
Let us consider the map, defined on $M$ by
$ F:= d(x,\cdot)^2.$
Then by Ekeland's variational principle, we can find $z:= z(\epsilon)\in C$ such that
\begin{itemize}
\item $d(y,z) \leq 1$,
\item $d(x,z)^2 \leq d(x,y)^2 \leq d_C(x)^2+\epsilon^2$,
\item and $z$ is the minimizer over $C$ of the function $\psi:= F + \epsilon d(\cdot,z)$.
\end{itemize}
Note that $d(x,z) \leq d_C(x)+\epsilon \leq \ell({\mathcal{B}}) \leq 1$ so $d(z, {\mathcal{B}}) \leq 1 $ and $z \in {\mathcal{B}}_1$.
We denote $\Pi(x,\epsilon)$ the set of points $z\in C$ satisfying the three last conditions. Clearly, for $\epsilon_1> \epsilon_2$, we have $\Pi(x,\epsilon_2) \subset \Pi(x,\epsilon_1)$.
Now consider $\epsilon>0$ and $z_1,z_2\in\Pi(x,\epsilon)$. Then $z_1,z_2\in C \cap {\mathcal{B}}_1$ and by first-order optimality conditions, we are going to prove that for $i=1,2$, there exists $v_i \in T_{z_i} M$ with $ |v_i|_{T_{z_i} M} \leq c_{{\mathcal B}_1} \epsilon$ (for some constant $c_{{\mathcal B}_1} >0$) such that
\begin{equation} w_i := \Gamma_{z_i,x} +v_i \in \mathrm{N}(C, z_i). \label{eq:j} \end{equation}
\medskip\noindent
{\bf Step 1 :} Proof of (\ref{eq:j}). \\
Let $\phi := U \subset M \rightarrow H$ be a local chart around $z_i$.
By definition, we locally have
\begin{align*}
z_i & := \argmin{p\in C} d(x,p)^2 + \epsilon d(p,z_i) \\
& = \phi^{-1} \left[\argmin{P\in \phi(C \cap U)} d(\phi^{-1}(X),\phi^{-1}(P))^2 + \epsilon d(\phi^{-1}(P),\phi^{-1}(Z_i))\right] \\
& = \phi^{-1} \left[\argmin{P\in \phi(U)} d(\phi^{-1}(X),\phi^{-1}(P))^2 + \epsilon d(\phi^{-1}(P),\phi^{-1}(Z_i)) + I_{\phi(C)}(P)\right]
\end{align*}
where $X=\phi(x) $, $P=\phi(p)$, $Z_i =\phi(z_i)$ and $I_{\phi(C)} $ represents the indicatrix function of $\phi(C) $. \\
Since $\phi(C) $ is closed, the characteristic function $I_{\phi(C)}$ is lower semicontinuous.
So we have
$$ 0 \in \nabla d^2(\phi^{-1}(X),\phi^{-1}(\cdot)) (Z_i) + \epsilon \partial^C (d(\phi^{-1}(\cdot),\phi^{-1}(Z_i))) (Z_i) + \lambda, $$ for some $\lambda \in \mathrm{N}_{\phi(C)}^C(Z_i)$, where $\partial^C $ and $\mathrm{N}^C$ represent the Clarke subdifferential and the Clarke normal cone. We refer the reader to \cite[Section 3.1.1]{BS} for this result.
As a consequence, there exists $U_i\in H$ with $|U_i|_H\leq 1$ satisfying
$$ 0 = - 2 [D \phi^{-1}(Z_i)]^* (\Gamma_{z_i,x} ) + \epsilon U_i + \lambda,$$
according to Remark \ref{rem:nablay} .
That implies
$$ [D \phi^{-1}(Z_i)]^* (\Gamma_{z_i,x}) - \frac{\epsilon}{2} U_i \in \mathrm{N}_{\phi(C)}^C( Z_i).$$
By composing with $([D \phi^{-1}(Z_i)]^*)^{-1}$ we have
$$ \Gamma_{z_i,x} - \frac{\epsilon}{2} ([D \phi^{-1}(Z_i)]^* )^{-1}(U_i) \in ([D \phi^{-1}(Z_i)]^*) ^{-1}[\mathrm{N}_{\phi(C)}^C( Z_i)].$$
Moreover,
\begin{align*}
w \in ([D \phi^{-1}(Z_i)]^*)^ {-1}[\mathrm{N}_{\phi(C)}^C(Z_i)] & \Longleftrightarrow [D \phi^{-1}(Z_i)]^{*}(w) \in \mathrm{N}_{\phi(C)}^C(Z_i) \\
& \Longleftrightarrow \forall \tau \in T^C_{\phi(C)}(Z_i), \langle [D \phi^{-1}(Z_i)]^{*}(w) , \tau \rangle \leq 0
\end{align*}
since the Clarke tangent cone $ T^C_{\phi(C)}(Z_i)$ to $\phi(C)$ at $Z_i$ is the polar cone of $\mathrm{N}^C_{\phi(C)}(Z_i)$.
As a consequence,
\begin{align*}
w \in ([D \phi^{-1}(Z_i)]^*)^ {-1}[\mathrm{N}_{\phi(C)}^C(Z_i)] & \Longleftrightarrow \forall \tau \in T^C_{\phi(C)}(Z_i), \langle w , D \phi^{-1}(Z_i)[\tau] \rangle \leq 0 \\
& \Longleftrightarrow w \in \left[D \phi^{-1}(Z_i)(T^C_{\phi(C)}(Z_i))\right]^\circ,
\end{align*}
by definition of a polar cone.
Thus, using the change of charts, we obtain that $$ \left[D \phi^{-1}(Z_i)(T^C_{\phi(C)}(Z_i))\right]^\circ=\left[T^C_{C}(z_i)\right]^\circ = \mathrm{N}^C_C(z_i),.$$
where we have used the definition of the Clarke tangent cone in \cite{HP} and the fact that the Clarke normal cone is exactly its polar cone.
Since the set is locally prox-regular, we know that $w \in \mathrm{N}(C,z_i) $ thanks to Proposition \ref{pro}.
We conclude the proof by choosing $v_i=-\frac{\epsilon}{2} ([D \phi^{-1}(Z_i)]^*)^ {-1}(U_i)$. In that case, $w_i := \Gamma_{z_i,x} +v_i \in \mathrm{N}(C, z_i)$ and we have $|v_i| \leq c_{{\mathcal{B}}_1} \epsilon$ (the constant $c_{{\mathcal{B}}_1}$ depending on $\phi$ and ${\mathcal{B}}_1$).
\medskip\noindent
{\bf Step 2 :} End of the proof. \\
Since $|w_i| \leq d(z_i, x) +|v_i| \leq d_C(x) + \epsilon + c_{{\mathcal{B}}_1} \epsilon \leq \ell({\mathcal{B}})+c_{{\mathcal B}_1}\epsilon$,
and $ d(z_1,z_2) \leq d(z_1, x) + d(z_2, x) \leq \rho({\mathcal{B}}_1)$ from Lemma \ref{lem:hypo}, it comes
$$ \langle w_1, \Gamma_{z_1,z_2} \rangle_{T_{z_1}M} + \langle w_2, \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} \leq 2 E_{{\mathcal{B}}_1} ( \ell({\mathcal{B}}) +c_{{\mathcal{B}}_1} \epsilon ) d(z_1,z_2)^2, $$
where $E_{{\mathcal{B}}_1} >0$ .
Using the parallel transport,
$$\langle w_1, \Gamma_{z_1,z_2} \rangle_{T_{z_1}M} = \langle L_{z_1\to z_2} (w_1), -\Gamma_{z_2,z_1} \rangle_{T_{z_2}M},$$
hence we conclude to
$$ \langle w_2-L_{z_1\to z_2} (w_1), \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} \leq 2 E_{{\mathcal{B}}_1} ( \ell({\mathcal{B}})+ c_{{\mathcal{B}}_1} \epsilon ) d(z_1,z_2)^2.$$
By definition of the $w_i$, it comes
$$ \langle \Gamma_{z_2,x} +v_2-L_{z_1\to z_2} (\Gamma_{z_1,x} +v_1), \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} \leq 2 E_{{\mathcal{B}}_1} ( \ell({\mathcal{B}})+ c_{{\mathcal{B}}_1}\epsilon ) d(z_1,z_2)^2$$
which yields
$$ \langle \Gamma_{z_2,x} -L_{z_1\to z_2}(\Gamma_{z_1,x}), \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} \leq 2 E_{{\mathcal{B}}_1} ( \ell({\mathcal{B}})+ c_{{\mathcal B}_1}\epsilon ) d(z_1,z_2)^2 + 2 c_{{\mathcal{B}}_1} \epsilon d(z_1,z_2).$$
As $z_1, z_2 \in B(x, \ell({\mathcal{B}}))$ with $ \ell({\mathcal{B}})< \rho({\mathcal{B}}_1)$, we deduce from Corollary \ref{cor:new} that $$ \langle \Gamma_{z_2,x} -L_{z_1\to z_2}(\Gamma_{z_1,x}), \Gamma_{z_2,z_1} \rangle_{T_{z_2}M} \geq A({\mathcal{B}}_1) d(z_1,z_2)^2.$$
Hence we observe that
$$\left(A({\mathcal{B}}_1)- 2 E_{{\mathcal{B}}_1} \ell({\mathcal{B}})-2 E_{{\mathcal{B}}_1} c_{{\mathcal{B}}_1} \epsilon\right) d(z_1,z_2) \leq 2 c_{{\mathcal{B}}_1}\epsilon$$
If $\epsilon < \frac{\tau({\mathcal{B}}_1)}{ c_{{\mathcal{B}}_1} }$, then $A({\mathcal{B}}_1)- 2 E_{{\mathcal{B}}_1} \ell({\mathcal{B}})-2 E_{{\mathcal{B}}_1} c_{{\mathcal{B}}_1} \epsilon \geq A({\mathcal{B}}_1)/2 > 0 $ since $ \ell({\mathcal{B}}) \leq \tau({\mathcal{B}}_1) = \frac{A({\mathcal{B}}_1)}{8 E_{{\mathcal{B}}_1} }$.
As a consequence
\begin{equation}
d(z_1,z_2) \leq \frac{2 c_{{\mathcal{B}}_1}}{\left(A({\mathcal{B}}_1)- 2 E_{{\mathcal{B}}_1} \ell({\mathcal{B}})-2 E_{{\mathcal{B}}_1} c_{{\mathcal{B}}_1}\epsilon\right)} \epsilon := \kappa \epsilon.
\label{eq:dz1z2}
\end{equation}
Let $\epsilon_n \downarrow 0 $ and $z_n \in \Pi(x,\epsilon_n) $. It then follows from that for $m\geq n$, $$ d(z_n, z_m) \leq \kappa \epsilon_n, $$ hence $(z_n)$ is a Cauchy sequence that converges to a point $z^\star \in C$ ($M$ is metrically complete). We obtain $z^\star \in \mathrm P_C(x)$ since $d(x,z^\star) = d_C(x)$, consequently $ \mathrm P_C(x) \neq \emptyset $. Furthermore, taking $\epsilon =0$ in \eqref{eq:dz1z2}, we can specify that
$ \mathrm P_C(x) =\{ z^\star\}$.
\end{proof}
\begin{rem}
Note that the constant $\ell({\mathcal{B}})$ depends on the geometrical properties of $M$ but also on $C$ through the term $E_{\mathcal{B}}$ (which is determined by the quantities $\eta_{\mathcal{B}}$ see Remark \ref{rem:constantE} ).
\label{rem:constantell}
\end{rem}
\section{Sweeping process in a Hilbert manifold} \label{sec:sweep}
In this section, we consider a Riemannian Hilbert manifold $M$ satisfying Assumption \ref{ass:ic}.
\begin{ass} \label{ass:C} Let $C:[0,T ] \rightrightarrows M $ be a multivalued map satisfying that
for all $t \in [0,T ], $ the set $C(t)$ is nonempty and locally prox-regular. More precisely, for all bounded set ${\mathcal{B}} \subset M$, there exists $0 < \eta_{\mathcal{B}} \leq \rho({\mathcal{B}})$ such that for every $t \in [0,T ]$, $s \in [0, \eta_{\mathcal{B}})$, $x \in C(t)$ and $v\in N(C(t), x) \setminus \{0\}$, $$ x \in \mathrm P_{C(t)}\left(\gamma_{x, \frac{v}{|v|}}(s)\right).$$ Furthermore, $C$ is supposed to be a Lipschitz map: for all $t,s \in [0,T ]$, $d_H(C(t),C(s)) \leq K_L |t-s|$.
\end{ass}
\begin{ass} \label{ass:f}
We assume that $f :(t,x) \in [0,T] \times M \mapsto f(t,x) \in T_xM $ is bounded and satisfies a Lipschitz regularity
in the following sense:
$$ g : \left\{ \begin{array}{ccc}
[0,T] \times M & \rightarrow & TM \\
(t,x) & \mapsto & (x,f(t,x))
\end{array} \right.$$
is Lipschitz with constant $L_f$. More precisely, for all bounded set ${\mathcal{B}}$ of $M$, for all $(t_1, t_2)
\in [0,T]^2$ and $ (x_1, x_2)
\in {\mathcal{B}}^2$ verifying $d(x_1, x_2) < \rho({\mathcal{B}})$ ,
\begin{align*}
d_{TM}(g(t_1,x_1), g(t_2,x_2)) &:=d(x_1,x_2) + | L_{x_1 \to x_2} (f(t_1,x_1)) - f(t_2,x_2)| \\
&\leq L_f (|t_1-t_2| + d(x_1,x_2) ).
\end{align*}
\end{ass}
\begin{thm} Under Assumptions \ref{ass:ic}, \ref{ass:C} and \ref{ass:f}, for every $x_0\in M$, for all $T>0$, there exists a unique solution $x \in W^{1,\infty}([0,T], M)$ such that
\be{eq:sys}
\left\{ \begin{array}{ll}
\dot{x}(t)+\mathrm{N}(C(t),x(t)) \ni f(t,x(t)) \textmd{ for a.e. } t \in [0,T] \vspace{6pt} \\
x(0)=x_0.
\end{array} \right. \end{equation}
\end{thm}
\medskip\noindent The proof is a mixture of the classical one in a Hilbert space (see the papers cited in the introduction) and of the previous results extending arguments in a Riemannian context.
\begin{proof}
\medskip\noindent {\bf First step: } Construction of discrete solutions. \\
We set $\mathcal{B}_0:=M\cap
B(x_0, 2 T\|f\|_\infty + K_L T )$. We can choose $\bar T \leq T$ such that $ 2 \bar T \|f\|_\infty + K_L \bar T< \min(\eta_{{\mathcal{B}}_0}/2, \ell({\mathcal{B}}_0) )$, where $\ell({\mathcal{B}}_0)$ is defined in Theorem \ref{thm:projection} by replacing $C$ with $C(0)$.
Note that the quantities $\eta_{{\mathcal{B}}_0}$, $\ell({\mathcal{B}}_0) $ are the same for all the sets $C(t), t \in [0,T]$ thanks to Remark \ref{rem:constantell} and Assumption \ref{ass:C}. \\
Then we define $\mathcal{B}:=M\cap
B(x_0, 2 \bar T\|f\|_\infty + K_L \bar T ) \subset \mathcal{B}_0$, hence $\rho(\mathcal{B})/2 \geq \rho(\mathcal{B}_0)/2 $. As a consequence, it yields $$2 \bar T \|f\|_\infty + K_L \bar T< \eta_{{\mathcal{B}}_0}/2 \leq \rho({\mathcal{B}}_0)/2 \leq \rho(\mathcal{B})/2.$$
We fix a time-step $h=\bar T/n$ which obviously satisfies
\be{eq:1} h \|f\|_\infty \leq \rho(\mathcal{B})/2 \end{equation}
So we consider a subdivision of the time-interval $ J=[0,\bar T]$ defined by $t^n_i = ih$ for $i\in\{0,..,n\}$. We build $(x^n_i)_{0\leq i\leq n}$ as follows:
\be{scheme} \left\{ \begin{array}{l}
x^n_0=x_0 \vspace{6pt} \\
x^n_{i+1}\in \mathrm P_{C(t^n_{i+1})}\left[\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)\right].
\end{array} \right. \end{equation}
First let us check that this scheme is well-defined.
\begin{lem} \label{lem:schemewp}
For all $i \in \{0, \dots n\}$, $ x_i^n$ and for all $i \in \{0, \dots n-1\}$ $\gamma_{x^n_i,f(t^n_i,x^n_i)}(h) $ are well-defined and belong to ${\mathcal{B}}$. Furthermore for $i \in \{0, \dots n-1\} $ , $d(x^n_i, x_{i+1}^n) \leq \rho({\mathcal{B}})/2$ and $ d( x_{i+1}^n, \gamma_{x^n_i,f(t^n_i,x^n_i)}(h) ) \leq \rho({\mathcal{B}})/2$.
\end{lem}
\begin{proof}
First $x^n_0= x_0 \in {\mathcal{B}}$ and since $h \|f\|_\infty \leq \rho({\mathcal{B}})$, $\gamma_{x^n_0,f(t^n_0,x^n_0)}(h) $ exists.
Moreover $ d(x_0, \gamma_{x^n_0,f(t^n_0,x^n_0)}(h) ) \leq h \|f\|_\infty $ hence $\gamma_{x^n_0,f(t^n_0,x^n_0)}(h) \in {\mathcal{B}}$.
Assume that for some $i<n$, $ x_i^n$ and $\gamma_{x^n_i,f(t^n_i,x^n_i)}(h) $ are well-defined and belong to ${\mathcal{B}}$.
Let us show that $x_{i+1}^n $ and (if $i<n-1$) $\gamma_{x^n_{i+1},f(t^n_{i+1},x^n_{i+1})}(h) $ satisfy the same properties.
Since $ d(x_i^n, \gamma_{x^n_i,f(t^n_i,x^n_i)}(h) ) \leq h \|f\|_\infty $ and $d_{ C(t_{i+1}^{n})}(x_i^n) \leq d_H(C(t_i^{n}), C(t_{i+1}^{n}) \leq K_Lh $ then
\begin{align*}
d_{C(t_{i+1}^{n})}(\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)) &\leq d(x_i^n, \gamma_{x^n_i,f(t^n_i,x^n_i)}(h)) + d_{C(t_{i+1}^{n})}(x_i^n) \\
&\leq h(\|f\|_\infty +K_L) \leq \ell({\mathcal{B}}_0) .
\end{align*}
In addition $ {\mathcal{B}}_0 \cap C(t_{i+1}^{n}) \neq \emptyset $. Otherwise, for all $z \in C(t^n_{i+1})$, $d(x_0, z) \geq 2 T \|f\|_\infty + K_L T$ and so $d_{C(t^n_{i+1})} (x_0) \geq 2 T \|f\|_\infty + K_L T $. Yet $d_{C(t^n_{i+1})} (x_0) \leq d_H(C(0),C(t^n_{i+1}) ) \leq K_L T $.
Consequently $\mathrm P_{C(t^n_{i+1})}\left[\gamma_{x^n_i,f(t^n_i,x^n_i)}(h) \right ] \neq \emptyset $ according Theorem \ref{thm:projection}, so $ x_{i+1}^n$ exists.
Moreover for $ 0 \leq j \leq i$,
\begin{align*}
d(x_j^n, x_{j+1}^n)& \leq d(x_j^n, \gamma_{x^n_j,f(t^n_j,x^n_j)}(h)) + d(x_{j+1}^n, \gamma_{x^n_j,f(t^n_j,x^n_j)}(h)) \\
&\leq h \|f\|_\infty + d_{C(t_{j+1}^{n})}(\gamma_{x^n_j,f(t^n_j,x^n_j)}(h)) \\
&\leq h(2 \|f\|_\infty +K_L) \leq \rho({\mathcal{B}})/2.
\end{align*}
Thus $d(x_0, x_{i+1}^n) \leq (i+1)h (2 \|f\|_\infty +K_L) \leq \bar{T}(2 \|f\|_\infty +K_L), $ hence $ x_{i+1}^n$ belongs to ${\mathcal{B}}$ and $$d(x_{i+1}^n, \gamma_{x^n_{i},f(t^n_{i},x^n_{i})}(h) ) = d_{C(t_{i+1}^{n})}(\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)) \leq h(\|f\|_\infty +K_L) \leq \rho({\mathcal{B}})/2.$$
Since $ h \|f\|_\infty \leq \rho({\mathcal{B}})/2$, $ \gamma_{x^n_{i+1},f(t^n_{i+1},x^n_{i+1})}(h) $ is well-defined.
Finally if $i<n-1$,
$$\begin{array}{lcl}
d(x_0, \gamma_{x^n_{i+1},f(t^n_{i+1},x^n_{i+1})}(h)) & \leq & d(x_0, x_{i+1}^n) + d(x_{i+1}^n, \gamma_{x^n_{i+1},f(t^n_{i+1},x^n_{i+1})}(h)) \vspace{6pt} \\
&\leq & (i+2)h (2 \|f\|_\infty +K_L) , \vspace{6pt} \\
&\leq & nh (2 \|f\|_\infty +K_L) \leq \bar{T}(2 \|f\|_\infty +K_L),
\end{array}$$
so $ \gamma_{x^n_{i+1},f(t^n_{i+1},x^n_{i+1})} \in {\mathcal{B}}$. The proof is also ended.
\end{proof}
Now thanks to the points $(x^n_i)_{0\leq i\leq n}$ we define two
piecewise maps $x^n$ and $f^n$ on $J$ in the following way:
\be{eqf} \forall t \in J_i:=[ih, (i+1)h[ \, , \,\, f^n(t):=f (t^n_i ,x^n_i) \in T_{x^n_i}M, f^n(\bar T):=f (t^n_n ,x^n_n) \end{equation}
and
\begin{align}
\forall t \in J_i:=[ih, (i+1)h[ \, , \,\, x^n(t) &:=
\gamma_{x^n_i,\Gamma_{x^n_i,x^n_{i+1}}}\left(\frac{t-ih}{h}\right), x^n(\bar T ):=x^n_n.
\label{eqxn}
\end{align}
The function $x^n$ is continuous on $J=[0,\bar T]$.
Note that $\Gamma_{x^n_i,x^n_{i+1}} $ is well-defined since $d(x^n_{i+1},x^n_i ) <\rho({\mathcal{B}})/2 $ (according to Lemma \ref{lem:schemewp}).
Moreover we define two other
piecewise maps $\tau^n$ and $\theta^n$ on $J$ in the following way:
$$\forall t \in J_i:=[ih, (i+1)h[ \, , \,\, \tau^n(t):=t^n_i, \tau^n(\bar T):=\bar T \textmd{ and
} \theta^n(t):=t^n_{i+1}, \theta^n(\bar T):= \bar T . $$
\medskip\noindent {\bf Second step: } Boundedness of the discretized velocities.\\
We claim that the discretized velocities are uniformly bounded. Indeed
for $n$, $i\in\{0,...,n-1\}$ and $t\in J_i$, the scheme (\ref{scheme})
with (\ref{eqxn}) gives
\begin{align}
|\dot x^n(t)| & = \frac{1}{h} \left|\dot \gamma_{x_i^n,\Gamma_{x_i^n,x_{i+1}^n}}\left(\frac{t-ih}{h}\right)\right| \nonumber \\
& \leq \frac{1}{h} |\Gamma_{x_i^n,x_{i+1}^n}| = \frac{d(x_i^n,x_{i+1}^n)}{h} \nonumber \\
& \leq 2 \|f\|_{\infty} + K_L. \label{eq:vit}
\end{align}
So, let
\be{eq:V} V:= \sup_{i,n} h^{-1} d(x^n_{i+1},x^n_i) \leq 2 \|f\|_\infty +K_L <\infty. \end{equation}
Thus we observe
that \be{eq:borne} \forall n \ \textrm{and} \ \forall t, \quad x^n(t) \in \mathcal{B}. \end{equation}
\medskip\noindent {\bf Third step: } Differential inclusion for the discrete solutions. \\
We are now looking for a discrete differential inclusion satisfied by the function $x^n$. More precisely, we want to check that for almost every $i\in\{0,...,n-1\}$, we have
\be{inclu:amontrer} \Gamma_{x^n_{i+1},x^n_i} + h f(t^n_{i+1},x^n_{i+1}) +\mathrm R(h^2)\in \mathrm{N}(C(t^n_{i+1}),x^n_{i+1}),\end{equation}
with $\mathrm R(h^2) \in T_{x^n_{i+1}} M $ and $ |\mathrm R(h^2)| =\mathcal{O}(h^2) $.
By definition of the scheme (\ref{scheme}), we know that (taking $\epsilon=1$ in the definition of proximal normal vectors and writing $ \gamma_{x^n_i,f(t^n_i,x^n_i)}(h) = \gamma_{x^n_i,v}(1) $ where $v= \Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)} $ )
\be{aa} \Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)} \in \mathrm{N}(C(t^n_{i+1}),x^n_{i+1}).\end{equation}
Note that $\Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)}$ is well-defined since $d(x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)) \leq \rho({\mathcal{B}})/2$ (according to Lemma \ref{lem:schemewp}).
Thanks to the smoothness of the exponential map and of the geodesics, let us check the following equality:
\begin{lem}
\be{eq:r1} \Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)} = \Gamma_{x^n_{i+1},x^n_i} + h D \exp_{x^n_{i+1}}^{-1}(x^n_i) [f(t^n_i,x^n_i)] + \mathrm R_1 (h^2),\end{equation}
\end{lem}
\begin{proof}
By setting $y^n_i:=\gamma_{x^n_i,f(t^n_i,x^n_i)}(h) $, we recall that
$$\Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)} = \exp_{x^n_{i+1}}^{-1}(y^n_i).$$
Now we define for $ s \in [0,h]$, $\gamma(s):= \exp_{x^n_{i}}(s f(t^n_i,x^n_i) )$, so by using a first order expansion, it comes
$$ \exp_{x^n_{i+1}}^{-1}(y^n_i) = \exp_{x^n_{i+1}}^{-1}(x^n_i) + \int_{0}^h D\exp_{x^n_{i+1}}^{-1}(\gamma(s))[\dot{\gamma}(s)] ds. $$
Furthermore
\begin{align*}
&\left |\int_{0}^h D\exp_{x^n_{i+1}}^{-1}(\gamma(s))[\dot{\gamma}(s)] ds - h D \exp_{x^n_{i+1}}^{-1}(x^n_i) [f(t^n_i,x^n_i)] \right |& \\
\quad &= \left |\int_{0}^h D\exp_{x^n_{i+1}}^{-1}(\gamma(s))[\dot{\gamma}(s)] ds - \int_{0}^h D \exp_{x^n_{i+1}}^{-1}(\gamma(0)) [\dot{\gamma}(0)] ds\right | &\\
\quad & \leq\int_{0}^h\left | D\exp_{x^n_{i+1}}^{-1}(\gamma(s))[\dot{\gamma}(s)] ds - D \exp_{x^n_{i+1}}^{-1}(\gamma(0)) [\dot{\gamma}(0)] \right | ds &\\
\quad & \leq \int_{0}^h C_e({\mathcal{B}}) d_{TM}( (\gamma(s),\dot{\gamma}(s) ) , (\gamma(0), \dot{\gamma}(0)) ) ds \textmd{ according to Assumption \ref{ass:ic} } &\\
\quad & \leq C_e({\mathcal{B}}) \int_{0}^h (d(\gamma(0),\gamma(s) ) +| L_{\gamma(s) \to\gamma(0) }( \dot{\gamma}(s)) - \dot{\gamma}(0)| )ds &\\
\quad & \leq C_e({\mathcal{B}}) \int_{0}^h ( s |f(t^n_i,x^n_i) | + L({\mathcal{B}}) d(\gamma(0),\gamma(s) ) ^2) ds \textmd{ using Lemma \ref{lem:geodesic} } &\\
\quad & \leq C_e({\mathcal{B}}) (\|f\|_{\infty} h^2 + L({\mathcal{B}}) \|f\|_{\infty}^2 h^3 ) .
\end{align*}
Thus fixing $\mathrm R_1 (h^2) := \displaystyle \int_{0}^h D\exp_{x^n_{i+1}}^{-1}(\gamma(s))[\dot{\gamma}(s)] ds - h D \exp_{x^n_{i+1}}^{-1}(x^n_i) [f(t^n_i,x^n_i)] $, we can write
$$\exp_{x^n_{i+1}}^{-1}(y^n_i) = \exp_{x^n_{i+1}}^{-1}(x^n_i) + h D \exp_{x^n_{i+1}}^{-1}(x^n_i) [f(t^n_i,x^n_i)] + \mathrm R_1 (h^2) , $$
which is the desired result.
\end{proof}
With Assumption \ref{ass:ic}, the map $D\exp^{-1}_{x^n_{i+1}} $ is Lipschitz and so it comes (using the notations introduced in this assumption)
\begin{align*}
& \left| D \exp_{x^n_{i+1}}^{-1}(x^n_i) [f(t^n_i,x^n_i)] - D
\exp_{x^n_{i+1}}^{-1}(x^n_{i+1}) [f(t^n_{i+1},x^n_{i+1})] \right| \\
& \quad \leq C_e(\mathcal{B}) d_{TM}\left(g(t^n_i,x_n^i),g(t^n_{i+1},x^n_{i+1}) \right) \\
& \quad \leq C_e(\mathcal{B}) L_f (h+d(x^n_i,x^n_{i+1})) \\
&\quad \leq C_e(\mathcal{B}) L_f (1+V)h := C_1 h ,
\end{align*}
for some numerical constant $C_1 >0$, where we used Assumption (\ref{ass:f})
(on $g$ introduced there) and (\ref{eq:V}). Thus (\ref{eq:r1}) becomes
\begin{align*}
\Gamma_{x^n_{i+1},\gamma_{x^n_i,f(t^n_i,x^n_i)}(h)} & = \Gamma_{x^n_{i+1},x^n_i} + h D \exp_{x^n_{i+1}}^{-1}(x^n_{i+1}) [f(t^n_{i+1},x^n_{i+1})] + \mathrm R (h^2) \\
& = \Gamma_{x^n_{i+1},x^n_i} + h f(t^n_{i+1},x^n_{i+1}) + \mathrm R (h^2).
\end{align*}
Since for every $x\in M$, $D \exp_{x}^{-1} (x) = [D
\exp_x(0)]^{-1} = Id_{T_xM}$. From this and (\ref{aa}), we deduce
(\ref{inclu:amontrer}).
\medskip\noindent {\bf Fourth step: } Cauchy sequence. \\
Let $n, m$ two integers, by (\ref{eq:borne}) for all $t \in J$,
$x^n(t)$ and $x^m(t)$ belong to $\mathcal{B} $ and so $ d(x^n(t),x^m(t)) < d(x^n(t),x_0) + d(x_0,x^m(t))<\rho(\mathcal{B})/2 + \rho(\mathcal{B})/2 < \rho(\mathcal{B})$.
We consider $t \notin \{ \tau^n(t), \theta^n(t), \tau^m(t), \theta^m(t) \}$.
From (\ref{inclu:amontrer}) and Lemma \ref{lem:hypo}, we get for
$y=x^m(t) \in \mathcal{B}$,
\begin{align*}
& \langle \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} + h f(\theta^n(t),x^n(\theta^n(t)),\Gamma_{x^n(\theta^n(t)),x^m(t)} \rangle_{x^n(\theta^n(t))} \\
& \hspace{2cm} \leq {E_\mathcal{B}} |\Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} + h f(\theta^n(t),x^n(\theta^n(t))) | d(x^n(\theta^n(t)),x^m(t))^2 \\
& \hspace{2.2cm} + \mathcal{O}(h^2)d(x^n(\theta^n(t),x^m(t))) + \mathcal{O}(h^2)d(x^n(\theta^n(t),x^m(t))^2. \end{align*}
Since
\be{ineq} d(x^n(\theta^n(t)),x^m(t)) \leq 1 + d(x^n(\theta^n(t)),x^m(t))^2 ,\end{equation}
it comes
\begin{align*}
\langle \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} + h f(\theta^n(t),x^n(\theta^n(t))),\Gamma_{x^n(\theta^n(t)),x^m(t)} \rangle_{x^n(\theta^n(t))} \nonumber \\
\leq {E_\mathcal{B}}h (V+\|f\|_{\infty}+\mathcal{O}(h)) d(x^n(\theta^n(t)),x^m(t))^2 + \mathcal{O}(h^2).
\end{align*}
As \begin{align*}
d(x^n(\theta^n(t)),x^m(t)) &\leq d(x^n(\theta^n(t)),x^n(t)) +
d(x^n(t),x^m(t)) \\ & \leq V h + d(x^n(t),x^m(t))
\end{align*}
we have $$ d(x^n(\theta^n(t)),x^m(t))^2 \leq 2 d(x^n(t),x^m(t))^2 + \mathcal{O}(h^2)
$$ and as a consequence
\begin{align}
\langle \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} + h f(\theta^n(t),x^n(\theta^n(t))),\Gamma_{x^n(\theta^n(t)),x^m(t)} \rangle_{x^n(\theta^n(t))} \nonumber \\
\leq {2E_\mathcal{B}}h (V+\|f\|_{\infty}+\mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h^2). \label{eq:i}
\end{align}
Let us split the left term in $I_n(t)+II_n(t)$ with $$I_n(t):= \langle \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))},\Gamma_{x^n(\theta^n(t)),x^m(t)} \rangle_{x^n(\theta^n(t))} $$ and $$II_n(t):= h\langle f(\theta^n(t),x^n(\theta^n(t))),\Gamma_{x^n(\theta^n(t)),x^m(t)} \rangle_{x^n(\theta^n(t))}.$$
By symmetry, with changing the role of $x^n$ and $x^m$, we obtain
\begin{align*}
I_m(t)+II_m(t) & \leq {2E_\mathcal{B}}h (V+\|f\|_{\infty}+\mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h^2).
\end{align*}
Thus \be{eq:inm} \frac{ I_n(t)+I_m(t)+II_n(t)+II_m(t)}{h} \leq {4E_\mathcal{B}}(V+\|f\|_{\infty}+\mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h). \end{equation}
Let us study all the quantities $I_n,I_m,II_n$ and $II_m$.
\medskip\noindent
For $I_n$, we use the parallel transport from $x^n(\theta^n(t))$ to $x^n(t)$ to obtain
\begin{align}
I_n(t) = \langle
\Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))},\Gamma_{x^n(\theta^n(t)),x^m(t)}
\rangle_{x^n(\theta^n(t))} \nonumber \\
=- h\langle \dot x^n(t),L_{x^n(\theta^n(t)) \to
x^n(t)}\left[\Gamma_{x^n(\theta^n(t)),x^m(t)}\right]\rangle_{x^n(t)}, \label{eq:in}
\end{align}
since
\begin{align} L_{x^n(\theta^n(t)) \to x^n(t)}\left[\Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))}\right] &=\frac{d(x^n(\theta^n(t)),x^n(\tau^n(t)))}{d(x^n(t),x^n(\tau^n(t)))}\Gamma_{x^n(t),x^n(\tau^n(t))}
\nonumber \\
&= - h\dot x^n(t). \label{eq:der}
\end{align}
Indeed the curve $x^n(t)$ is exactly the geodesic between $x^n(\tau^n(t))$ and $x^n(\theta^n(t))$ with velocity $d(x^n(\theta^n(t)),x^n(\tau^n(t))) /h$.
Similarly for $m$, we have
\be{eq:im} I_m(t) =- h \langle \dot x^m(t),L_{x^m(\theta^m(t)) \to x^m(t)}\left[\Gamma_{x^m(\theta^m(t)),x^n(t)}\right]\rangle_{x^m(t)}.\end{equation}
\medskip\noindent
Concerning $II_n$ and $II_m$, we split them in two parts $II_n = A_n+B_n$ with
\begin{align*}
%
A_n & := h \langle L_{x^n(\theta^n(t)) \to x^n(t)} \left[ f(\theta^n(t),x^n(\theta^n(t))) \right] , \\
& \hspace{2cm} L_{x^n(\theta^n(t)) \to x^n(t)}[\Gamma_{x^n(\theta^n(t)),x^m(t)}]-\Gamma_{x^n(t),x^m(t)} \rangle_{x^n(t)}
\end{align*}
and
$$ B_n := \ h \langle L_{x^n(\theta^n(t)) \to x^n(t)} \left[ f(\theta^n(t),x^n(\theta^n(t))) \right], \Gamma_{x^n(t),x^m(t)} \rangle_{x^n(t)}.$$
The first term $A_n$ is estimated as follows:
\begin{align}
|A_n| & \leq h \|f\|_{\infty} |\Gamma_{x^n(t),x^m(t)}-L_{x^n(\theta^n(t)) \to x^n(t)}[\Gamma_{x^n(\theta^n(t)),x^m(t)}]|_{T_{x^n(t)}M} \nonumber \\
& \leq h \|f\|_{\infty} |\exp_{x^n(t)}^{-1}(x^m(t)) - L_{x^n(\theta^n(t)) \to x^n(t)}[\exp_{x^n(\theta^n(t))}^{-1}(x^m(t))] |_{T_{x^n(t)}M} \nonumber \\
& \leq \frac{1}{2} h \|f\|_{\infty}|\nabla_x d^2(x^n(t), x^m(t)) - L_{x^n(\theta^n(t)) \to x^n(t)}[\nabla_x d^2(x^n(\theta^n(t)), x^m(t)) ] |_{T_{x^n(t)}M} \nonumber \\
& \quad \quad \textmd{ according to Lemma \ref{lemb} } \nonumber \\
& \leq \frac{1}{2} h \|f\|_{\infty} \sup_{s \in [t, \theta^n(t)]} \|H_x d^2(x^n(s), x^m(t))\|_{x^n(s)} d(x^n(\theta^n(t)),x^n(t)) \nonumber \\
& \leq \frac{1}{2} C_{\rho(\mathcal{B})} h \|f\|_{\infty} d(x^n(\theta^n(t)),x^n(t)) \leq \frac{1}{2} C_{\rho(\mathcal{B})} h^2 \|f\|_{\infty} V ,
\label{eq:calculus}
\end{align}
where we used Proposition \ref{propH} and (\ref{eq:V}).
Similarly for $A_m$, we obtain
\begin{align*}
| A_m |\leq \frac{1}{2} C_{\rho(\mathcal{B})} h^2 \|f\|_{\infty} V.
\end{align*}
Finally with (\ref{eq:inm}),(\ref{eq:in}) and (\ref{eq:im}), we obtain
\begin{align}
& -\langle \dot x^n(t),L_{x^n(\theta^n(t)) \to x^n(t)}\left[\Gamma_{x^n(\theta^n(t)),x^m(t)}\right]\rangle_{x^n(t)} \\ & \hspace{2cm}- \langle \dot x^m(t),L_{x^m(\theta^m(t)) \to x^m(t)}\left[\Gamma_{x^m(\theta^m(t)),x^n(t)}\right]\rangle_{x^m(t)} \nonumber \\
& \hspace{3cm} \leq \frac{|B_n+B_m|}{h} + {4E_{\mathcal{B}}}(V+\|f\|_{\infty}+\mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h) \label{eq2}
\end{align}
It remains us to estimate $|B_n+B_m|$. With a parallel transport, we get
\begin{align*}
\frac{B_n+B_m}{h} & = \langle L_{x^n(\theta^n(t)) \to x^n(t)}
\left[ f(\theta^n(t),x^n(\theta^n(t))) \right], \Gamma_{x^n(t),x^m(t)}
\rangle_{x^n(t)} \\
& \hspace{1.5cm} + \langle L_{x^m(\theta^m(t)) \to x^m(t)} \left[
f(\theta^m(t),x^m(\theta^m(t))) \right], \Gamma_{x^m(t),x^n(t)}
\rangle_{x^m(t)} \\
&= \langle L_{x^n(t) \to x^m(t)} L_{x^n(\theta^n(t)) \to x^n(t)}
\left[ f(\theta^n(t),x^n(\theta^n(t))) \right],- \Gamma_{x^m(t),x^n(t)}
\rangle_{x^m(t)} \\
& \hspace{1.5cm} + \langle L_{x^m(\theta^m(t)) \to x^m(t)} \left[
f(\theta^m(t),x^m(\theta^m(t))) \right], \Gamma_{x^m(t),x^n(t)}
\rangle_{x^m(t)},
\end{align*}
since $ L_{x^n(t) \to x^m(t)} (\Gamma_{x^n(t),x^m(t)})
=-\Gamma_{x^m(t),x^n(t)}. $
Consequently,
\begin{align*}
\frac{B_n+B_m}{h} & = \langle L_{x^n(t) \to x^m(t)} L_{x^n(\theta^n(t)) \to x^n(t)}
\left[ f(\theta^n(t),x^n(\theta^n(t))) \right] \\
& \hspace{1.5cm} - L_{x^m(\theta^m(t)) \to x^m(t)} \left[
f(\theta^m(t),x^m(\theta^m(t))) \right] , -\Gamma_{x^m(t),x^n(t)}
\rangle_{x^m(t)} \\
&\leq | L_{x^n(t) \to x^m(t)} L_{x^n(\theta^n(t)) \to x^n(t)} \left[
f(\theta^n(t),x^n(\theta^n(t))) \right] \\
& \hspace{1.5cm} - L_{x^m(\theta^m(t)) \to x^m(t)} \left[
f(\theta^m(t),x^m(\theta^m(t))) \right]|_{T_{x^m(t)}M} \
d(x^m(t),x^n(t)).
\end{align*}
We estimate the last norm by making appear intermediate
points. Indeed, using the Lipschitz regularity on $f$ (Assumption \ref{ass:f}) and (\ref{eq:V}), we have
\begin{align*}
\left| f(t, x^m(t)) - L_{x^m(\theta^m(t)) \to x^m(t)} \left[ f(\theta^m(t),x^m(\theta^m(t)) \right] \right| \\
\leq L_f (h + d( x^m(t),x^m(\theta^m(t)) )) \\
\leq L_f (1+V) h \lesssim h.
\end{align*}
Similarly,
$$ \left| L_{x^n(t) \to x^m(t)} \left(L_{x^n(\theta^n(t)) \to x^n(t)} \left[ f(\theta^n(t),x^n(\theta^n(t))) \right] - f(t,x^n(t)) \right)\, \right| \lesssim h$$
and
$$ \left| L_{x^n(t) \to x^m(t)} \left[f(t,x^n(t))\right] - f(t, x^m(t))\right| \lesssim d(x^n(t),x^m(t)).$$
So we can conclude that
$$ \begin{array}{lll}
\dfrac{|B_n+B_m|}{h} &\lesssim & \left(h+ d(x^n(t),x^m(t)) \right) d(x^m(t),x^n(t))\vspace{6pt} \\
&\lesssim & h d(x^n(t),x^m(t)) + d(x^m(t),x^n(t))^2 \vspace{6pt} \\
&\lesssim & h (1 +d(x^m(t),x^n(t))^2) + d(x^m(t),x^n(t))^2 \textmd{ with} (\ref{ineq}). \vspace{6pt} \\
\end{array}
$$
Finally, (\ref{eq2}) becomes
\begin{align}
& - \langle \dot x^n(t), L_{x^n(\theta^n(t)) \to x^n(t)} \left[ \Gamma_{x^n(\theta^n(t)),x^m(t)}\right] \rangle_{x^n(t)} \nonumber
\\ & \hspace{2cm} - \langle \dot x^m(t), L_{x^m(\theta^m(t)) \to x^m(t)} \left[ \Gamma_{x^m(\theta^m(t)),x^n(t)}\right] \rangle_{x^m(t)} \nonumber \\
& \hspace{5cm} \leq (C_2+ \mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h) \label{eq3},
\end{align}
for some numerical constant $C_2$ (not depending on $n$ and $m$).
For the first terms, since the discretized velocities are uniformly
bounded, it follows that
$$ \begin{array}{l}
|L_{x^n(\theta^n(t)) \to x^n(t)}\left[\Gamma_{x^n(\theta^n(t)),x^m(t)}\right] -
\Gamma_{x^n(t),x^m(t)} | \vspace{6pt} \\ \hspace{0.5cm} = |L_{x^n(\theta^n(t)) \to x^n(t)}\left[\exp^{-1}_{x^n(\theta^n(t))}(x^m(t))\right] -
\exp^{-1}_{x^n(t)}(x^m(t))| \vspace{6pt} \\ \hspace{0.5cm} \leq \frac{1}{2} C_{\rho({\mathcal{B}})} d(x^n(\theta^n(t)), x^n(t))
\lesssim h \textmd{ (same arguments as in the proof of \eqref{eq:calculus})}.
\end{array}$$
Hence
\begin{align*}
-\langle \dot x^n(t),L_{x^n(\theta^n(t)) \to x^n(t)}\left[\Gamma_{x^n(\theta^n(t)),x^m(t)}\right]\rangle_{x^n(t)}
\\ = -\langle \dot x^n(t),\Gamma_{x^n(t),x^m(t)}\rangle_{x^n(t)} +\mathcal{O}(h).
\end{align*}
Producing similar reasoning for the symmetrical quantity, it comes
\begin{align}
-\langle \dot x^n(t),\Gamma_{x^n(t),x^m(t)}\rangle_{x^n(t)} - \langle \dot x^m(t),\Gamma_{x^m(t),x^n(t)} \rangle_{x^m(t)} \nonumber \\
\leq (C_2+ \mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h) \label{eq4},
\end{align}
which gives (according to Lemma \ref{lemb})
\begin{align*}
\langle \dot x^n(t),\nabla_x d^2(x^n(t),x^m(t))\rangle_{x^n(t)} + \langle \dot x^m(t),\nabla_x d^2(x^m(t),x^n(t)) \rangle_{x^m(t)} \nonumber \\
\leq 2(C_2+ \mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h).
\end{align*}
Since $ \nabla_x d^2(x^m(t),x^n(t)) = \nabla_y d^2(x^n(t),x^m(t)) $, we conclude that
$$ \frac{d}{dt} d(x^n(t) ,x^m(t))^2 \leq 2(C_2+ \mathcal{O}(h)) d(x^n(t),x^m(t))^2 + \mathcal{O}(h).$$
As usual, Gronwall's Lemma implies that
$$ \sup_{t\in [0,\bar{T}]} d(x^n(t),x^{m}(t))^2 = \mathcal{O}(h).$$
and so $(x^n)_{n}$ is a Cauchy sequence of $C^0([0,T],M)$. Since $M$ is supposed to be metrically complete, the sequence $(x^n)_{n}$ also strongly converges to a function $x\in C^0([0,\bar{T}],M)$.
Furthermore \be{eq:lim} \sup_{t\in [0,\bar{T}]} d(x^n(t),x(t)) = \mathcal{O}(h^\frac{1}{2})\end{equation}
\medskip\noindent {\bf Fifth step: } The limit function $x$ is solution of (\ref{eq:sys}). \\
Let $t_0 \in [0, \bar T]$, there exists $\Phi:U_0
\rightarrow H$ a chart where $U_0$ is an bounded open set of $M$ containing $x(t_0)$. As $U_0$ is an open set, $$ \exists r < \rho(U_0), \, U:= B(x(t_0),r) \subset U_0. $$ Obviously, $r < \rho(U_0) \leq \rho(U)$.
Since $x$ is continuous on $[0, T]$, there is an open interval $\tilde{J} \subset [0, \bar T]$ containing $t_0$ such that
\begin{align}
\forall t \in \tilde{J}, \, d(x(t),x(t_0)) < \frac{r}{4} .
\label{hypo:J}
\end{align}
Moreover as $x^n$ uniformly converges to $x$ on $[0, \bar T]$, there exists $h_0$ such that for all $h<h_0$, for all $t \in J$
\begin{align}
&d(x^n(t),x(t)) < \frac{r}{4} , \quad
V h < \frac{r}{4}, \quad
K_L h < \frac{r}{4}\quad and \quad
K_L h < \ell(U)
\label{hypo:h0}
\end{align}
where $\ell(U)$ is defined in Theorem \ref{thm:projection} by replacing $C$ with $C(t_0)$.
Note that for all $h<h_0$ and for all $t \in \tilde{J}$, $x^n(t) \in U$.
Recall that $x^n$ satisfies equation (\ref{inclu:amontrer}) on $J$, which is
$$ h^{-1} \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} + f(\theta^n(t),x^n(\theta^n(t))) +\mathrm R(h)\in \mathrm{N}(C(\theta^n(t)),x^n(\theta^n(t))).$$
By Lemma \ref{lem:hypo}
and thanks to the boundedness of $f$ and of the discretized velocities
(\ref{eq:vit}), there exist constants $\beta$ and $ E_U$ such that for all $c^n\in C(\theta^n(t)) \cap U$ satisfying $d(c^n, x^n(\theta^n(t))) \leq \rho(U)$,
\begin{align*}
& \langle h^{-1} \Gamma_{x^n(\theta^n(t)),x^n(\tau^n(t))} +
f(\theta^n(t),x^n(\theta^n(t)))+ \textrm{R}(h),
\Gamma_{x^n(\theta^n(t)),c^n} \rangle_{T_{x^n(\theta^n(t))}M} \\
& \hspace{5cm} \leq \beta{E_U } d(x^n(\theta^n(t)),c^n)^2.
\end{align*}
So using a parallel transport with (\ref{eq:der}),
\begin{align}
&\langle -\dot x^n(t) + L_{x^n(\theta^n(t)) \to x^n(t)}\left[ f(\theta^n(t),x^n(\theta^n(t)))\right]+ \mathrm R_2(h), \nonumber \\
& \hspace{1cm} L_{x^n(\theta^n(t)) \to x^n(t)} \Gamma_{x^n(\theta^n(t)),c^n} \rangle_{T_{x^n(t)}M} \leq \beta {E_U } d(x^n(\theta^n(t)),c^n)^2,
\label{maj}
\end{align}
where $ | \mathrm R_2(h) | = \mathcal{O}(h)$.
Now in order to apply usual arguments in the Hilbertian context, we use the chart $\Phi$ to work in $H$. So we define
$y^n = \Phi(x^n)$. By this way, $\dot y^n$ and $y^n$ are bounded
sequences and $y^n$ strongly converges to $y:=\Phi(x)$. Then we know
that up to a subsequence, we can assume that $\dot y^n$ $*-$weakly
converges to $\dot y$ in $L^\infty(\tilde{J},H)$ which implies that $\dot
y^n$ weakly converges to $\dot y$ in $L^1(\tilde{J},H)$ since $\tilde{J}$ is bounded.
Consequently by Mazur's Lemma, there exists a subsequence $z^n \in L^1(\tilde{J},H)$ satisfying for almost every $t\in \tilde{J}$
\be{eq:zn} z^n(t) \in \textmd{Conv} \left( \dot y^k (t), \ k \geq n \right) \xrightarrow[n \to \infty ]{} \dot y(t) \end{equation}
where \textrm{Conv} stands for the convex combinations and
\be{eq:zn2} z^n \xrightarrow[n \to \infty ]{} \dot y \textmd{ in } L^{1} (\tilde{J},H). \end{equation}
We now look for proving for a.e. $t$ (when (\ref{eq:zn}) holds) and all $c \in C(t)\cap U$ satisfying $ d(x(t),c) <r/4$:
$$ \langle -\dot x(t) +f(t,x(t)) , \Gamma_{x(t),c} \rangle_{T_{x(t)}M} \leq \beta {E_U } d(x(t),c)^2.$$
Let $h < h_0$, $t \in \tilde{J}$ (such that (\ref{eq:zn}) holds), $c \in C(t) \cap U$ satisfying $ d(x(t),c) <r/4$. Thus $d_{C(\theta^n(t))}(c) \leq d_H (C(t), C(\theta^n(t))) \leq K_L h < \ell(U) $ by Assumption \ref{ass:C} and (\ref{hypo:h0}). Indeed, we recall that the quantity $\ell(U) $ is the same for all the sets $C(t), t \in [0,T]$ thanks to Remark \ref{rem:constantell} and Assumption \ref{ass:C}.
Consequently Theorem \ref{thm:projection} implies that there is $ c^n \in \mathrm P_{C(\theta^n(t))}(c)$. Since
\be{majccn}
d(c,c^n) = d_{C(\theta^n(t))}(c) \leq K_L h < r/4,
\end{equation}
we deduce from (\ref{hypo:J}) that
\begin{align*}
d(c^n , x(t_0)) \leq d(c^n , c) + d(c, x(t)) + d(x(t), x(t_0)) \leq 3r/4 < r.
\end{align*}
In other words, $c^n$ belongs to $U$.
Moreover with (\ref{hypo:h0}) we obtain
\begin{align}
d(c^n , x^n(\theta^n(t))) & \leq d(c^n , c) + d(c, x(t))+ d(x(t), x^n(t)) + d(x^n(t),x^n(\theta^n(t) ) \nonumber \\
& < r/2 + r/4 + Vh <r < \rho(U)
\label{maj1}
\end{align}
and
\begin{align}
d(c , x^n(\theta^n(t))) & \leq d(c, x(t))+ d(x(t), x^n(t)) + d(x^n(t),x^n(\theta^n(t) ) \nonumber\\
& < 3r/4 <\rho(U).
\label{maj2}
\end{align}
Firstly, the inequality \eqref{maj} is satisfied thanks to \eqref{maj1}. Secondly,
\begin{align*}
& \left| \langle L_{x^n(\theta^n(t)) \to x^n(t)} \left[ f(\theta^n(t),x^n(\theta^n(t)))\right]+ \mathrm R_2(h) \right. , \vspace{6pt}\\
& \hspace{1cm} \left. - L_{x^n(\theta^n(t)) \to x^n(t)} [\Gamma_{x^n(\theta^n(t)),c^n} - \Gamma_{x^n(\theta^n(t)),c}] \rangle \right| \vspace{6pt}\\
& \hspace{2cm} \leq ( \|f\|_\infty +\mathcal{O}(h)) | L_{x^n(\theta^n(t)) \to x^n(t)} [\Gamma_{x^n(\theta^n(t)),c^n} - \Gamma_{x^n(\theta^n(t)),c}] | \vspace{6pt} \\
& \hspace{2cm} \leq ( \|f\|_\infty +\mathcal{O}(h)) | \exp^{-1}_{x^n(\theta^n(t))} (c^n)-\exp^{-1}_{x^n(\theta^n(t))} (c) | \vspace{6pt}\\
& \hspace{2cm} \leq ( \|f\|_\infty +\mathcal{O}(h)) C_e(U) d(c^n,c) \textmd{ according to } \eqref{maj1} , \eqref{maj2} \textmd{ and Assumption } \ref{ass:ic} \vspace{6pt} \\
& \hspace{2cm} \leq ( \|f\|_\infty +\mathcal{O}(h)) C_e(U) d_H(C(t),C(\theta^n(t))) = \mathcal{O}(h) \textmd{ thanks to Assumption \ref{ass:C}}.
\end{align*}
Thirdly, using \eqref{majccn}, we have $$d(x^n(\theta^n(t)),c) \leq d(x^n(\theta^n(t)),c^n) + d(c, c^n) \leq d(x^n(\theta^n(t)),c^n) + K_L h .$$
Hence with \eqref{maj1},
\begin{align*}
d(x^n(\theta^n(t)),c)^2 \leq d(x^n(\theta^n(t)),c^n)^2 +2 K_L h \rho(U) + (K_L h)^2.
\end{align*}
The previous inequality always holds if $c^n$ and $c$ are switched (using \eqref{maj2}) and as a consequence
$$ d(x^n(\theta^n(t)),c^n)^2 -d(x^n(\theta^n(t)),c)^2 =\mathcal{O}(h).$$
Hence \eqref{maj} becomes
\begin{align*}
& \langle -\dot x^n(t) + L_{x^n(\theta^n(t)) \to x^n(t)}\left[ f(\theta^n(t),x^n(\theta^n(t)))\right]+ \mathrm R_2(h), L_{x^n(\theta^n(t)) \to x^n(t)} \Gamma_{x^n(\theta^n(t)),c} \rangle_{T_{x^n(t)}M} \vspace{6pt} \\
& \hspace{5cm} \leq \beta {E_U } d(x^n(\theta^n(t)),c)^2 +\mathcal{O}(h).
\end{align*}
For $n \to \infty$, since $f$ is Lipschitz continuous on $TM$ and with (\ref{eq:lim}), we have as previously
\begin{align*}
& \langle L_{x^n(\theta^n(t)) \to x^n(t)}\left[ f(\theta^n(t),x^n(\theta^n(t)))\right]+ \mathrm R_2 (h), - L_{x^n(\theta^n(t)) \to x^n(t)}\Gamma_{x^n(\theta^n(t)),c} \rangle_{T_{x^n(t)}M} \\
& \hspace{5cm} + \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M} = \mathcal{O}(h^{\frac{1}{2}}).
\end{align*}
Consequently, we have for large enough $k$
\begin{align*}
& \langle -\dot x^k(t) , L_{x^k(\theta^k(t)) \to x^k(t)} \Gamma_{x^k(\theta^k(t)),c} \rangle_{T_{x^k(t)}M} \\
& \hspace{2cm} \leq \beta {E_U } d(x^k(\theta^k(t)),c)^2 - \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M}+\mathcal{O}(h^{\frac{1}{2}}),
\end{align*}
which means
\begin{align*}
& \left \langle -D\Phi(x^k(t))^{-1} [\dot y^k(t)] , L_{x^k(\theta^k(t)) \to x^k(t)} \Gamma_{x^k(\theta^k(t)),c} \right \rangle_{T_{x^k(t)}M} \\
& \hspace{2cm} \leq \beta {E_U } d(x(t),c)^2 - \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M}+\mathcal{O}(h^{\frac{1}{2}})
\end{align*}
because $ d(x^k(\theta^k(t)),c) \leq d(x(t),c) + \mathcal{O}(h^{\frac{1}{2}})$
and so
\begin{align*}
& \left \langle - \dot y^k(t) , \left[D\Phi(x^k(t))^{-1}\right]^* L_{x^k(\theta^k(t)) \to x^k(t)} \Gamma_{x^k(\theta^k(t)),c} \right \rangle_{H} \\
& \hspace{2cm} \leq \beta {E_U } d(x(t),c)^2 - \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M}+\mathcal{O}(h^{\frac{1}{2}}).
\end{align*}
In addition
$$\left[D\Phi(x^k(t))^{-1}\right]^* \left( L_{x^k(\theta^k(t)) \to x^k(t)} \Gamma_{x^k(\theta^k(t)),c}\right ) = \left[D\Phi(x(t))^{-1}\right]^*( \Gamma_{x(t),c})+\mathcal{O}(h^{\frac{1}{2}}).$$
Indeed setting $ w_k = L_{x^k(\theta^k(t)) \to x^k(t)} \Gamma_{x^k(\theta^k(t)),c}$, $\tilde{w_k} = \Gamma_{x^k(t),c}$ and $w= \Gamma_{x(t),c} $, we have
\begin{align*}
& \left[D\Phi(x^k(t))^{-1}\right]^*(w_k) - \left[D\Phi(x(t))^{-1}\right]^* (w) \\
& \hspace{1cm}= \left[D\Phi(x^k(t))^{-1}\right]^*(w_k - \tilde{w_k} + \tilde{w_k} - L_{x(t) \to x^k(t)} w) \\
& \hspace{2cm} + \left[D\Phi(x^k(t))^{-1}\right]^* ( L_{x(t) \to x^k(t)} w) - \left[D\Phi(x(t))^{-1}\right]^* (w) \\
& \hspace{1cm} = \mathcal{O}(h^{\frac{1}{2}}),
\end{align*}
by smoothness of the chart $\Phi$ and because $w_k - \tilde{w_k} = \mathcal{O}(h) $, $ \tilde{w_k} - L_{x(t) \to x^k(t)} w = \mathcal{O}(h^{\frac{1}{2}}) $.
From this equality we deduce that
$$ \langle -\dot y^k(t) , \left[D\Phi(x(t))^{-1}\right]^* \Gamma_{x(t),c} \rangle_{H} \leq \beta {E_U } d(x(t),c)^2 - \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M}+\mathcal{O}(h^{\frac{1}{2}}).$$
So applying (\ref{eq:zn}) and then (\ref{eq:zn2}) give for almost every time $t\in \tilde{J}$ and all $c\in U \cap C(t)$ satisfying $d(x(t),c)< r/4$,
$$ \langle -\dot y(t) , \left[D\Phi(x(t))^{-1}\right]^* \Gamma_{x(t),c} \rangle_{H} \leq \beta {E_U} d(x(t),c)^2 - \langle f(t,x(t)), \Gamma_{x(t),c} \rangle_{T_{x(t)}M},$$
which is equivalent to
$$ \langle- \dot x(t) + f(t,x(t)) , \Gamma_{x(t),c} \rangle_{T_{x(t)}M} \leq \beta {E_U } d(x(t),c)^2.$$
As a consequence, Proposition \ref{prop:hypomo} yields that for such time $t\in \tilde{J}$,
\begin{align}
-\dot x(t) +f(t,x(t)) \in \mathrm{N}( C(t), x(t)).
\label{eq:incl}
\end{align}
We have shown that for every $t_0 \in J$, there is an open interval $\tilde{J}$ where inclusion \eqref{eq:incl} is satisfied almost everywhere. By a compactness argument, $J$ can be covered by a finite number of these intervals so finally \eqref{eq:incl} holds for $a. e. \, t \in J $. Furthermore by dividing $I= [0, T] $ into small intervals of length $|J|= \bar T$, we can again follow these five steps and obtain the same result for $a. e. \, t \in I $. In other words, the function $x$ (so built in the whole interval $I$) is a solution of (\ref{eq:sys}).
\medskip\noindent {\bf Sixth step: } Uniqueness \\
Let $x_1$, $x_2$ be two solutions of (\ref{eq:sys}). Since $x_1 $ and $x_2$ belong to $W^{1,\infty}(I,M)$, for all $t \in I, \, x_1(t), x_2(t) \in \mathcal{B}$,
where $\mathcal{B}$ is a bounded set of $M$. Furthermore setting $$ \displaystyle \alpha := \frac{\rho(\mathcal{B})}{ 2\max(\|\dot{x_1}\|_\infty, \|\dot{x}_2\|_\infty) } >0$$ we have for $i=1,2$ and for all $t \leq \alpha $, $d(x_i(t), x_0) \leq \rho(\mathcal{B}) /2$ and so $d(x_1(t),x_2(t)) \leq \rho(\mathcal{B}) $.\\
It comes from Lemma \ref{lem:hypo} that for a.e. $t \leq \alpha $,
\be{ineg1} \langle- \dot{x_1}(t) + f(t, x_1(t)), \Gamma_{x_1(t), x_2(t)} \rangle _{T_{x_1(t)}M} \leq E_\mathcal{B} |\dot{x_1}(t) - f(t, x_1(t)) | d(x_1(t),x_2(t))^2 , \end{equation}
and
\be{ineg2} \langle -\dot{x_2}(t) + f(t, x_2(t)), \Gamma_{x_2(t), x_1(t)} \rangle _{T_{x_2(t)}M} \leq {E_\mathcal{B} |\dot{x_2}(t) - f(t, x_2(t)) | } d(x_1(t),x_2(t))^2 . \end{equation}
Moreover with (\ref{eq:symmetry}), $L_{x_1(t) \to x_2(t)}\left(\Gamma_{x_1(t),x_2(t)}\right) = -\Gamma_{x_2(t),x_1(t)}$. Thus (\ref{ineg1}) becomes
\begin{align}
& \langle L_{x_1(t) \to x_2(t)} (\dot{x_1}(t) - f(t, x_1(t))), \Gamma_{x_2(t), x_1(t)} \rangle _{T_{x_2(t)}M} \nonumber \\
& \hspace{3cm} \leq {E_\mathcal{B} |\dot{x_1}(t) - f(t, x_1(t)) | } d(x_1(t),x_2(t))^2 . \label{ineg3}
\end{align}
By summing (\ref{ineg2}) and (\ref{ineg3}), we have
\begin{align}
& \langle L_{x_1(t) \to x_2(t)} (\dot{x_1}(t) - f(t, x_1(t))) - (\dot{x_2}(t) - f(t, x_2(t))), \Gamma_{x_2(t), x_1(t)} \rangle _{T_{x_2(t)}M} \nonumber \\
& \hspace{3cm} \leq E_\mathcal{B} F d(x_1(t),x_2(t))^2 , \label{ineg4}
\end{align}
where $F= 2 \|f\|_\infty + \|\dot{x_1}\|_\infty +\|\dot{x}_2\|_\infty $.
Assumption \ref{ass:f} implies that $$ | f(t, x_2(t))- L_{x_1(t) \to x_2(t)} ( f(t, x_1(t))) |\leq L_f d(x_1(t),x_2(t)). $$ As $|\Gamma_{x_2(t), x_1(t)} | = d(x_1(t),x_2(t))$,
we have
\be{ineg5} \langle f(t, x_2(t))- L_{x_1(t) \to x_2(t)} ( f(t, x_1(t))) , \Gamma_{x_2(t), x_1(t)} \rangle \leq L_f d(x_1(t),x_2(t))^2. \end{equation}
It follows from (\ref{ineg4}) and (\ref{ineg5}) that
\be{ineg6} \langle L_{x_1(t) \to x_2(t)} (\dot{x_1}(t)) - \dot{x_2}(t), \Gamma_{x_2(t), x_1(t)} \rangle _{T_{x_2(t)}M} \leq \left({E_\mathcal{B} F }+L_f \right) d(x_1(t),x_2(t))^2 , \end{equation}
Furthermore
$$\begin{array}{lll}
\frac{d}{dt} (d(x_1(t),x_2(t))^2 &= &\langle \dot{x_1}(t), \nabla_x d^2(x_1(t),x_2(t)) \rangle_{T_{x_1(t)}M} \vspace{6pt}\\
& & \hspace{2cm}+ \langle \dot{x_2}(t), \nabla_y d^2(x_1(t),x_2(t)) \rangle_{T_{x_2(t)}M} \vspace{6pt}\\
&=&\langle \dot{x_1}(t), -2 \Gamma_{x_1(t), x_2(t)} \rangle_{T_{x_1(t)}M} \vspace{6pt} \\
& & \hspace{2cm} + \langle \dot{x_2}(t), - 2 \Gamma_{x_2(t), x_1(t)} \rangle_{T_{x_2(t)}M} \vspace{6pt} \\
&=& 2\langle L_{x_1(t) \to x_2(t)} (\dot{x_1}(t)) - \dot{x_2}(t), \Gamma_{x_2(t), x_1(t)} \rangle _{T_{x_2(t)}M}.
\end{array}$$
We deduce that $$ \frac{1}{2}\frac{d}{dt} (d(x_1(t),x_2(t))^2 \leq \left( {E_\mathcal{B} F }+L_f \right) d(x_1(t),x_2(t))^2. $$ which is equivalent to
$$ \frac{d}{dt} (d(x_1(t),x_2(t))^2 \leq 2 \left( {E_\mathcal{B} F }+L_f \right) d(x_1(t),x_2(t))^2. $$ With the help of Gronwall Lemma, we conclude that $x_1=x_2$ in $[0, \alpha]$ since $x_1(0)=x_0=x_2(0)$ and obviously this equality holds in $[0,T]$ with the same arguments.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
Genome rearrangements are evolutionary events that shuffle genomic architectures.
Most frequent genome rearrangements are \emph{reversals} (that flip segments of a chromosome), \emph{translocations} (that exchange segments of two chromosomes),
\emph{fusions} (that merge two chromosomes into one), and \emph{fissions} (that split a single chromosome into two).
The minimal number of such events between two genomes is often used in phylogenomic studies to measure the evolutionary distance between the genomes.
These four types of rearrangements can be modeled by $2$-breaks~\cite{alekseyev2008} (also called DCJs~\cite{yancopoulos2005}),
which break a genome at two positions and glue the resulting fragments in a new order.
They simplify the analysis of genome rearrangements and allow one to efficiently compute the corresponding evolutionary distance between two genomes.
\emph{Transpositions} represent yet another type of genome rearrangements that cuts off continuous segments of a genome and moves them to different positions.
In contrast to reversal-like rearrangements, transpositions are rarely observed and believed to appear in a small proportion in the course of evolution
(e.g., in Drosophila evolution transpositions are estimated to constitute less than $10\%$ of genome rearrangements~\cite{ranz2003rearrangement}).
Furthermore, transpositions are hard to analyze; in particular, computing the transposition distance is known to be NP-complete~\cite{bulteau2012}.
To simplify analysis of transpositions, they can be modeled by $3$-breaks~\cite{alekseyev2008} that break the genome
at \textit{three} positions and glue the resulting fragments in a new order.
In the current work we propose a computational method for determining the proportion of transpositions (modeled as 3-breaks)
among the genome rearrangements (2-breaks and 3-breaks) between two genomes.
To the best of our knowledge, previously the proportion of transpositions was studied only from the perspective
of its bounding with the weighted distance model~ \cite{bader2007,fertin2009}, where reversal-like and transposition-like rearrangements are assigned different weights.
However, it was empirically observed~\cite{Blanchette1996} and then proved that the weighted distance model does not, in fact, achieve its design goal~\cite{jiang2011}.
We further remark that any approach to the analysis of genome rearrangements that controls the proportion of transpositions
would need to rely on a biologically realistic value, which can be estimated with our method.
We applied our method for different pairs among the rat, macaque, and human genomes and estimated the transpositions rate in all pairs to be around 0.26.
\section{Background}
For the sake of simplicity, we restrict our attention to circular genomes.
We represent a genome with $n$ blocks as a graph which contains $n$ directed
edges encoding blocks and $n$ undirected edges encoding block adjacencies. We denote the
tail and head of a block $i$ by $i^t$ and $i^h$, respectively.
A \emph{2-break} replaces any pair of adjacency edges $\{x,y\}$, $\{u,v\}$ in the genome graph
with either a pair of edges $\{x,u\}$, $\{y,v\}$ or a pair of edges
$\{u,y\}$, $\{v,x\}$.
Similarly, a \emph{3-break} replaces any triple of adjacency edges with another triple of edges forming a matching on the same six vertices (Fig.~\ref{fig:bpg}).
Let $P$ and $Q$ be genomes on the same set $S$ of blocks (e.g., synteny blocks or orthologous genes). We assume
that in their genome graphs the adjacency edges of $P$ are colored black and the adjacency edges of $Q$
are colored red.
The \emph{breakpoint graph} $G(P, Q)$ is defined on the set of vertices $\{i^t,i^h| i \in S\}$
with black and red edges inherited from genome graphs of $P$ and $Q$.
The black and red edges in $G(P,Q)$ form a collection of alternating black-red cycles (Fig.~\ref{fig:bpg}).
We say that a black-red cycle is an \emph{$\ell$-cycle} if it contains $\ell$ black edges (and $\ell$ red
edges), and we denote the number of $\ell$-cycles in $G(P,Q)$ by $c_{\ell}(P,Q)$.
We call $1$-cycles \textit{trivial cycles}\footnote{In the breakpoint graph constructed on synteny blocks, there are no trivial cycles since no adjacency
is shared by both genomes. However, in our simulations below this condition may not hold, which would result in the appearance of trivial cycles.}
and we call \textit{breakpoints} the vertices belonging to non-trivial cycles.
\begin{figure}%
\begin{center}
\includegraphics[width=\textwidth]{bpgnew_exp.pdf}
\caption{\textbf{a)} The breakpoint graph $G(P,Q_0)$ of ``black'' genome $P$ and ``red'' genome $Q_0=P$ , each consisting of a single circular chromosome $(1,2,3,4,5,6,7,8)$.
Here, $n = 8$, $b = 0$, and all cycles in $G(P,Q_0)$ are trivial.
\textbf{b)} The breakpoint graph of ``black'' genome $P$ and ``red'' genome $Q_1=(1,2,3,-7,-6,-5,-4,8)$ obtained from $Q_0$ with a reversal of a segment $4,5,6,7$
(represented as 2-break on the dotted edges shown in a). Here we use $-i$ to
denote opposite orientation of the block $i$.
The graph consists of $c_1 = 6$ trivial cycles and $c_2 = 1$ 2-cycle, and thus $b =2c_2 = 2$.
\textbf{c)} The breakpoint graph of ``black'' genome $P$ and ``red'' genome $Q_2=(1,2,-6,-5,3,-7,-4,8)$ obtained from $Q_1$ with a transposition of a segment $3,-7$ (represented as a single 3-break on the dotted edges shown in b).
The graph consists of $c_1 = 3$ trivial cycles, $c_2 = 1$ 2-cycle, and $c_3 = 1$ 3-cycle; thus $b = 2c_2+3c_3 = 5$.}
\label{fig:bpg}
\end{center}
\end{figure}
The $2$-break distance between genomes $P$ and $Q$ is the minimum number of $2$-breaks required to transform $P$ into $Q$.
\begin{theorem}[\cite{yancopoulos2005}]
The 2-break distance between circular genomes $P$ and $Q$ is
$$d(P, Q) = n(P, Q) - c(P, Q) \; ,$$
where $n(P, Q )$ and $c(P, Q)$ are, respectively, the number of blocks and cycles in $G(P, Q)$.
\end{theorem}
While 2-breaks can be viewed as particular cases of 3-breaks (that keep one of the affected edges intact),
from now on we will assume that 3-breaks change all three edges on which they operate.
\section{Estimation for the Transposition Rate}
In our model, we assume that the evolution represents a discrete Markov process, where
different types of genome rearrangements (2-breaks and 3-breaks) occur independently with fixed probabilities.
Let $p$ and $1-p$ be the rate (probability) of 3-breaks and 2-breaks, respectively.
For any two given genomes resulted from this process, our method estimates the value of $p$ as explained below.
In the next section we evaluate the accuracy of the proposed method on simulated genomes and further apply it to real mammalian genomes
to recover the proportion of transpositions in mammalian evolution.
Let the evolution process start from a ``black'' genome $P$ and result in a ``red'' genome $Q$.
It can be viewed as a transformation of the breakpoint graph $G(P,P)$, where red edges are parallel to black edges and form trivial cycles,
into the breakpoint graph $G(P,Q)$ with 2-breaks and 3-breaks operating on red edges.
There are observable and hidden parameters of this process. Namely, we can observe the following parameters:
\begin{itemize}
\item $c_{\ell}=c_{\ell}(P,Q)$, the number of $\ell$-cycles (for any $\ell \geq 2$) in $G(P,Q)$;
\item $b=b(P,Q) = \sum_{\ell \geq 2} \ell c_{\ell}$, the number of active (broken) fragile regions between $P$ and $Q$, also equal the number of synteny blocks between $P$ and $Q$ and the halved total length of all non-trivial cycles in $G(P,Q)$;
\item $d=d(P,Q)$, the $2$-break distance between $P$ and $Q$;
\end{itemize}
while the hidden parameters are:
\begin{itemize}
\item $n=n(P,Q)$, the number of (active and inactive) fragile regions in $P$ (or $Q$),
also equal the number of solid regions (blocks) and the halved total length of all cycles in $G(P,Q)$;
\item $k_2$, the number of $2$-breaks between $P$ and $Q$,
\item $k_3$, the number of $3$-breaks between $P$ and $Q$.
\end{itemize}
We estimate the rearrangement distance between genomes $P$ and $Q$ as $k_2+k_3$ and the rate $p$ of transpositions as
$$p = \frac{k_3}{k_2+k_3} \; .$$
We remark that in contrast to other probabilistic methods for estimation of evolutionary parameters (such as the evolutionary distance in~\cite{Lin08}),
in our method we assume that the number of trivial cycles $c_1$ is not observable.
While trivial cycles can be observed in the breakpoint graph constructed on homologous gene families (rather than synteny blocks),
their interpretation as conserved gene adjacencies (which happen to survive just by chance) implicitly adopts
the \emph{random breakage model} (RBM)~\cite{Ohno70,Nadeau84} postulating that every adjacency has equal probability to be broken
by rearrangements. The RBM however was recently refuted with the more accurate \emph{fragile breakage model} (FBM)~\cite{PevznerTeslerPNAS03}
and then the \emph{turnover fragile breakable model} (TFBM)~\cite{Alekseyev10b},
which postulate that only certain (``fragile'') genomic regions are prone to genome rearrangements.
The FBM is now supported by many studies (see \cite{Alekseyev10b} for further references and discussion).
\section{Estimation for the Hidden Parameters}
In this section, we estimate hidden parameters $n$, $k_2$, and $k_3$ using observable parameters, particularly $c_2$ and $c_3$.
Firstly, we find the probability that a red edge was never broken in the course of evolution between $P$ and $Q$.
An edge is not broken by a single 2-break with the probability $\left(1-\frac{2}{n}\right)$ and by a single 3-break with the probability
$\left(1-\frac{3}{n}\right)$. So, the probability for an edge to remain intact during the whole process of $k_2$ 2-breaks and $k_3$ 3-breaks is
$$ \left(1-\frac{2}{n}\right)^{k_2}\left(1-\frac{3}{n}\right)^{k_3} \approx e^{-\gamma} \; ,$$
where $\gamma = \frac{2k_2+3k_3}{n}$.
Secondly, we remark that for any fixed $\ell$, the number of $\ell$-cycles resulting from occasional splitting of longer cycles is negligible,\footnote{We remark
that under the parsimony condition long cycles are never split into smaller ones. Our method does not rely on the parsimony condition
and can cope with such splits when their number is significantly smaller than the number of blocks.}
since the probability of such splitting has order $\frac{b}{n^2}$.
In particular, this implies that the number of trivial cycles (i.e., 1-cycles) in $G(P,Q)$ is approximately equal to the number of
red edges that were never broken in the course of evolution between $P$ and $Q$.
Since the probability of each red edge to remain intact is approximately $e^{-\gamma}$, the number of such edges is approximated by $n\cdot e^{-\gamma}$.
On the other hand, the number of trivial cycles in $G(P,Q)$ is simply equal to $n-b$, the number of shared block adjacencies between $P$ and $Q$. That is,
\begin{equation}\label{eq:nb}
n-b \approx n e^{-\gamma} \; .
\end{equation}
Thirdly, we estimate the number of 2-cycles in $G(P,Q)$. By the same reasoning as above, such cycles mostly result from 2-breaks that merge pairs of trivial cycles.
The probability for a red edge to be involved in exactly one 2-break is
$\frac{2k_2}{n} \left(1-\frac{1}{n}\right)^{2k_2+3k_3-1}$. The probability that another red edge was involved in the same 2-break is
$\frac{1}{n} \left(1-\frac{1}{n}\right)^{2k_2+3k_3-1}$. Since the total number of edge pairs is $n(n-1)/2$, we have
the following approximate equality for the number of 2-cycles:
\begin{equation}\label{eq:c2}
c_2 \approx k_2 e^{-2\gamma} \; .
\end{equation}
And lastly, we estimate the number of 3-cycles in $G(P,Q)$. As above, they mostly result from either 3-breaks that merge three 1-cycles,
or 2-breaks that merge a 1-cycle and a 2-cycle.
The number of 3-cycles of the former type approximately equals $k_3 e^{-3\gamma}$ analogously to the reasoning above.
The number of 3-cycles of the latter type is estimated as follows.
Clearly, one of the red edges in such a 3-cycle results from two 2-breaks, say $\rho_1$ followed by $\rho_2$, which happens with the probability about
$$\frac{2k_2(2k_2-2)}{2n^2}\left(1-\frac{1}{n}\right)^{2k_2+3k_3-2} \approx 2\frac{k^2_2}{n^2} e^{-\gamma} \; .$$
One of the other two edges results solely from $\rho_1$, while the remaining one results solely from $\rho_2$,
which happens with the probability about $\left(\frac{1}{n}e^{-\gamma}\right)^2$.
Since there are about $n^3$ ordered triples of edges, we get the following approximate equality for the number of 3-cycles:
\begin{equation}\label{eq:c3}
c_3\approx k_3 e^{-3\gamma}+\frac{2k_2^2}{n}e^{-3\gamma} \; .
\end{equation}
Fig.~\ref{fig:c2c3} provides an empirical evaluation of the estimates (\ref{eq:c2}) and (\ref{eq:c3}) for the number of 2-cycles and 3-cycles in $G(P,Q)$,
which demonstrates that these estimates are quite accurate.
\begin{figure}%
\begin{center}
\includegraphics[width=\textwidth]{c2c3-eps-converted-to.pdf}
\caption{Empirical and analytical curves for the number of 2-cycles and 3-cycles averaged over 100 simulations on $n=400$ blocks
with proportion of 3-breaks $p=0.3$.}
\label{fig:c2c3}
\end{center}
\end{figure}
Below we show how one can estimate the probability $p$ from the (approximate) equations \eqref{eq:nb}, \eqref{eq:c2}, and \eqref{eq:c3}.
We eliminate $k_2$ from \eqref{eq:c3}, using \eqref{eq:c2}:
$$ c_3 \approx k_3 e^{-3\gamma}+\frac{2c_2^2}{n} e^{\gamma} \; .$$
Now we consider the following linear combination of the last equation and \eqref{eq:c2}:
$$2e^{-\gamma}(c_2 - k_2 e^{-2\gamma})+3\left(c_3 -(k_3 e^{-3\gamma}+\frac{2c_2^2}{n} e^{\gamma})\right)\approx 0 \; .$$
It gives us the following equation for $\gamma$ and $n$:
$$\gamma e^{-3\gamma} \approx \frac{1}{n}\left(2c_2 e^{-\gamma}+3c_3-\frac{6c_2^2e^\gamma}{n}\right) \; .$$
Using \eqref{eq:nb}, we eliminate $n$ from the last equation and obtain the following equation with respect to a single indeterminate $\gamma$:
\begin{equation}\label{eq:gamma}
\gamma e^{-3\gamma} \approx \frac{1-e^{-\gamma}}{b}\left(2c_2 e^{-\gamma}+3c_3-\frac{6c_2^2e^\gamma(1-e^{-\gamma})}{b}\right) \; .
\end{equation}
Solving this equation numerically (see Example \ref{example}, Section \ref{sec:simulated}), we obtain the numerical values
for $\gamma^{est}$, $n^{est}$, $k^{est}_2$ and $k^{est}_3$, and, finally,
$$p_{est} = \frac{k_3^{est}}{k^{est}_2+k^{est}_3} \; .$$
\section{Experiments and Evaluation}
\subsection{Simulated Genomes} \label{sec:simulated}
We performed a simulation with a fixed number of blocks $n = 1800$
and variable parameters $p$ and $\gamma$.
In each simulation, we started with a genome $P$ and applied a number of 2-breaks and 3-breaks with probability $1-p$ and $p$, respectively, until
we reached the chosen value of $\gamma$. We denote the resulting genome by $Q$ and estimate $p$ with our method as $p_{est}$.
We observed that the robustness of our method mostly depends on $p$ and $\gamma$, and it becomes unstable for $p_{est}<0.15$ (Fig.~\ref{fig:estp}).
So in our experiments we let $p$ range between $0.05$ and $1$ with step $0.05$ and $\gamma$ range between $0.2$ and $1.2$ with step $0.1$.
\begin{figure}%
\begin{center}
\includegraphics[width=0.8\textwidth]{p_vs_pest-eps-converted-to.pdf}
\caption{Boxplots for the value of $p$ as a function of $p_{est}$}
\label{fig:estp}
\end{center}
\end{figure}
In Fig.~\ref{fig:estp}, we present boxplots for the value of $p$ as a function of $p_{est}$ cumulative over the values of $\gamma$.
These evaluations demonstrate that $p_{est}$ estimates $p$ quite accurately with the absolute error below $0.1$ in $90\%$ of observations.
\begin{example}
\label{example}
Let us consider the example from our simulated dataset. In this example, the number of active blocks
$b = 716$, the number of 2-cycles $c_2 = 107$, the number of 3-cycles $c_3 = 48$, and the hidden parameters are:
the total number of blocks is $n = 1800$, the number of $2$-breaks $k_2 = 279$ and the number of $3$-breaks is $k_3 = 114$.
So, the value of $p$ in this example is $0.29$ and the value of $\gamma$ is 0.5.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{figure_eqn3.pdf}
\caption{Typical behavior of
$f(\gamma) = \gamma e^{-3\gamma}-\frac{1-e^{-\gamma}}{b}\left(2c_2 e^{-\gamma}+3c_3-\frac{6c_2^2e^\gamma(1-e^{-\gamma})}{b}\right)$,
the difference between right and left hand sides of \eqref{eq:gamma}, where $b = 716$, $c_2 = 107$, $c_3 = 48$.}
\label{fig:gamma}
\end{center}
\end{figure}
At first, using the bisection method, one can find roots of \eqref{eq:gamma}. In this case there are two roots:
$\gamma = 0.466$ and $\gamma = 1.007$ (See Fig.\ref{fig:gamma}). Let us check the root $0.466$ first.
Then, using \eqref{eq:nb}, one finds the estimated value of $n$: $716/(1-e^{0.466})\approx 1922$. Equation \eqref{eq:c2} gives us the estimated value of $k_2$: $107 e^{2\cdot0.466} \approx 272$.
One can estimate $k_3$ as $(\gamma n - 2k_2)/3 \approx 117$.
And finally we obtain the estimated value of $p$: $117/(117+272) \approx 0.3$.
In this example, using the second root of \eqref{eq:gamma} yields a negative value for $k_3$, so we do not consider it.
So, our method quite accurately estimates the value of $p$, and also values of $\gamma$ and $n$.
\end{example}
\subsection{Mammalian Genomes}
We analyzed a set of three mammalian genomes: rat, macaque, and human, represented
as sequences of $1,360$ synteny blocks~\cite{ma2006,alekseyev2009}.
For each pair of genomes, we circularized\footnote{While chromosome circularization introduces artificial edges to the breakpoint graph,
the number of such edges (equal to the number of chromosomes) is negligible as compared to the number of edges representing block adjacencies in the genomes. For subtle differences in analysis of circular and linear genomes see \cite{alekseyev2008multi}}
their chromosomes, constructed the breakpoint graph, obtained parameters $b$, $c_2$, $c_3$,
and independently estimated the value of $p$.
The results in Table~\ref{tab:mam} demonstrate consistency and robustness with respect to the evolutionary distance between the genomes
(e.g., the 2-break distance between rat and human genomes is 714, while the 2-break distance between macaque and human genomes is 106).
The rate of transpositions for all genome pairs is estimated to be around 0.26.
Numerical experiments suggest that the 95\% confidence interval for such values is [0.1, 0,4] (Fig.~\ref{fig:estp}).
\begin{table}
\caption{Observable parameters $b$, $c_2$, $c_3$ and estimation $p_{est}$ for the rate of evolutionary transpositions
between circularized rat, macaque, and human genomes.}
\label{tab:mam}
\centering
\begin{tabular}{| c || c | c | c | c | }
\hline
Genome pair & $b$ & $c_2$ & $c_3$ & $p_{est}$ \\
\hline\hline
rat-macaque & 1014 & 201 & 85 & 0.27 \\
\hline
rat-human & 1009 & 194 & 79 & 0.26 \\
\hline
macaque-human & 175 & 45 & 17 & 0.25 \\
\hline
\end{tabular}
\end{table}
\section{Discussion}
In the present work we describe a first computational method for estimation of the transposition rate
between two genomes from the distribution of cycle lengths in their breakpoint graph.
Our method is based on modeling the evolution as a Markov process under the assumption that the transposition rate remains constant.
The method does rely on the random breakage model \cite{Ohno70,Nadeau84} and thus is consistent with more prominent fragile breakage model \cite{PevznerTeslerPNAS03,Alekseyev10b} of chromosome evolution.
As a by-product, the method can also estimate the true rearrangement distance (as $k_2+k_3$) in the evolutionary model that includes
both reversal-like and transposition-like operations.
Application of our method on different pairs of mammalian genomes reveals that the transposition rate is almost the same
for distant genomes (such as rat and human genomes)
and close genomes (such as macaque and human genomes), suggesting that the transposition rate remains the same across different lineages in mammalian evolution.
In further development of our method, we plan to employ the technique of stochastic differential equations,
which may lead to a more comprehensive description of the $c_{\ell}$ behavior.
It appears to be possible to obtain equations, analogous to \eqref{eq:c2} and \eqref{eq:c3},
for $c_\ell$ with $\ell>3$. This could allow one to verify the model and estimate the transposition rate more accurately.
\subsection*{Acknowledgements}
The authors thank Jian Ma for providing the synteny blocks for mammalian genomes.
The work is supported by the National Science Foundation under the grant No. IIS-1462107. The work of
NA is also supported by grant 6.38.672.2013 of SPbSU and RFFB grant 13-01-12422-ofi-m.
\bibliographystyle{splncs}
|
\section{Introduction}
The theory of error-correcting codes generally study the codes over the finite field. In recent time, the codes over the finite rings have been studied extensively because of their important role in algebraic coding theory. Negacyclic codes, an important class of constacyclic codes, over finite rings also have been well studied these day.
In 1960's, Berlekamp {\rm\cite{Ber68, Ber84}} introduced negacyclic codes over the field $\F_p$, $p$ odd prime, and designed a decoding algorithm that corrects up to $t\left(<\frac{p-1}{2}\right)$ Lee errors. Wolfmann {\rm\cite{Wolf99}}, in 1999, studied negacyclic codes of odd length over $\Z_4$. In 2003, Blackford {\rm\cite{Blfd03}} extended these study to negacyclic codes of even length over $\Z_4$.
The structure of negacyclic codes of length $n$ over a finite chain ring such that the length is not divisible by the characteristic of the residue field is obtained by Dinh and L$\acute{\text{o}}$pez-Permouth {\rm\cite{Dinh-Lopez04}} in a more general setting in the year 2004. When the length $n$ of the code is divisible by the characteristic of the residue field then the code is called a repeated-root codes. Repeated-root negacyclic codes over finite rings have also been investigated by many authors. The structure of negacyclic codes of length $2^t$ over $\Z_{2^m}$ was obtained in {\rm\cite{Dinh-Lopez04}}. In 2005, Dinh {\rm\cite{Dinh05}} investigated negacyclic codes of length $2^s$ over the Galois ring $\text{GR}(2^a,m)$. S$\breve{\text{a}}$l$\breve{\text{a}}$gean {\rm\cite{Salag06}}, in 2006, has studied the repeated-root negacyclic codes over a finite chain ring and has shown that these codes are principally generated over the Galois ring $\text{GR}(2^a,m)$. Various kinds of distances of negacyclic
codes of length $2^s$ over $\Z_{2^a}$ are determined in {\rm\cite{Dinhh07}}. The structure of the negacyclic codes of length $2p^s$ over the ring $\F_{p^m}+u\F_{p^m}$ have been discussed in {\rm\cite{Dinh15}}.
Let $p$ be a odd prime. In this paper we study the structure of negacyclic codes of odd length over the non chain ring $\F_p[u,v]/\langle u^2, v^2, uv-vu\rangle$. We find a unique set of generators, rank and a minimal spanning set for these codes. We also find the Hamming distance of these codes for length $p^l$.
The structures of cyclic codes over the ring $R_{u^2,v^2,p} = \F_p[u,v]/\langle u^2, v^2, uv-vu\rangle$ have been discussed in {\rm\cite{Pkk-Bg15}}. We can view the cyclic and negacyclic codes over the ring $R_{u^2,v^2,p}$ as an ideal in the rings $R_{u^2,v^2,p}[x]/\langle x^n-1 \rangle$ and $R_{u^2,v^2,p}[x]/$ $\langle x^n+1 \rangle$ respectively. We define the ring isomorphism from the ring $R_{u^2,v^2,p}[x]/$ $\langle x^n-1 \rangle$ to the ring $R_{u^2,v^2,p}[x]/ \langle x^n+1 \rangle$ to get the structure of negacyclic code over the ring $R_{u^2,v^2,p}$.
\section{Preliminaries}\label{pre}
A linear code $C$ of length $n$ over a ring $R$ is negacyclic if $(-c_{n-1}, c_0, \cdots , c_{n-2}) $ $\in C$ whenever $(c_0, c_1, \cdots , c_{n-1})\in C$. We can consider a negacyclic code $C$ of length $n$ over a ring $R$ as an ideal in the ring $R[x]/\langle x^n+1\rangle$ via the correspondence $R^n\rightarrow R[x]/\langle x^n+1\rangle$, $(c_0, c_1, \cdots , c_{n-1})\rightarrow c_0+c_1x+\cdots+c_{n-1}x^{n-1}$. Let $R_{u^2,v^2,p}=\F_p + u\F_p+v\F_p + uv\F_p,
u^2=0$, $v^2=0$ and $uv=vu$. This ring is isomorphic to the ring $\F_p[u,v]/\langle u^2, v^2, uv-vu\rangle$. The ring $R_{u^2,v^2,p}$ is a finite commutative local ring with the unique maximal ideal $\langle u,v\rangle$. The set $\{ \{0\}, \langle u\rangle, \langle v\rangle, \langle uv\rangle, \langle u + \alpha v\rangle$, $\langle u,v\rangle, \langle 1\rangle\}$ gives list of all ideals of the ring $R_{u^2, v^2, p}$, where $\alpha$ is a non zero element of $\F_p$. Since the maximal ideal $\langle u,v\rangle$ is not principal, the ring $R_{u^2, v^2, p}$ is not a chain ring. The residue field $\overline{R}$ of a ring $R$ is define as $\overline{R} = R/M$, where $M$ is a maximal ideal. For the ring $R_{u^2,v^2,p}$ the residue field is $\F_p$. Let $\mu : R[x] \rightarrow \overline{R}[x]$ denote the natural ring homomorphism that maps $r \mapsto r + M$ and the variable $x$ to $x$.
We define the degree of the polynomial $f(x) \in R[x]$ as the degree of the polynomial $\mu(f(x))$ in $\overline{R}[x]$, i.e., $deg(f(x)) = deg(\mu(f(x))$ (see, for example, \cite{McDonald74}). A polynomial $f(x) \in R[x]$ is called regular if it is not a zero divisor. The following conditions are equivalent for a finite commutative local ring $R$.
\begin{prop} {\rm (cf. \cite[Exercise XIII.2(c)]{McDonald74})} \label{regular-poly}
Let $R$ be a finite commutative local ring. Let $f(x) = a_0+a_1x+ \cdots +a_nx^n$ be in $R[x]$, then the following are equivalent.
\begin{enumerate}[{\rm (1)}]
\item $f(x)$ is regular; \label{1}
\item $\langle a_0, a_1, \cdots , a_n \rangle = R$; \label{4}
\item $a_i$ is an unit for some $i$, $0 \leq i \leq n$; \label{3}
\item $\mu(f(x)) \neq 0$; \label{2}
\end{enumerate}
\end{prop}
Let $g(x)$ be a non zero polynomial in $\F_p[x]$. By above proposition, it is easy to see that the polynomial $g(x)+up_1(x)+vp_2(x)+uvp_3(x) \in R_{u^2,v^2,p}[x]$ is regular. Note that $\text{deg}(g(x)+up_1(x)+vp_2(x)+uvp_3(x)) = \text{deg}(g(x))$.
\section{Structures for negacyclic codes over the ring $R_{u^2, v^2, p}$}\label{generator}
In this section we assume that $n$ is an odd integer. Let $R_{u^2,v^2,p}=\F_p + u\F_p+v\F_p + uv\F_p,
u^2=0$, $v^2=0$ and $uv=vu$. The following theorem gives the ring isomorphism from the ring $R_{u^2,v^2,p}[x]/\langle x^n-1 \rangle$ to the ring $R_{u^2,v^2,p}[x]/\langle x^n+1 \rangle$
\begin{prop}\label{isomorphism}
Let $\phi :\frac{R_{u^2,v^2,p}[x]}{\langle x^n-1\rangle}\rightarrow\frac{R_{u^2,v^2,p}[x]}{\langle x^n+1\rangle}$ be a map defined as $\phi(f(x))=f(-x)$, for all $f(x)\in \frac{R_{u^2,v^2,p}[x]}{\langle x^n-1\rangle}$. The map $\phi$ is a ring isomorphism.
\end{prop}
\begin{proof}
For polynomials $f(x), g(x)\in R_{u^2,v^2,p}[x]$,
\[f(x)\equiv g(x) ~\text{mod}~ (x^n-1);\]
if and only if there exists a polynomial $h(x)\in R_{u^2,v^2,p}[x]$ such that
\[f(x)-g(x)=h(x)(x^n-1);\]
if and only if
\[f(-x)-g(-x)=h(-x)((-x)^n-1)=-h(-x)(x^n+1);\]
if and only if
\[f(-x)\equiv g(-x) ~\text{mod}~ (x^n+1);\]
This implies that for $f(x), g(x)\in \frac{R_{u^2,v^2,p}[x]}{\langle x^n-1\rangle}$, $\phi(f(x))=\phi(g(x))$ if and only if $f(x)=g(x)$. Hence, $\phi$ is well-defined and one-to-one. Since the rings $\frac{R_{u^2,v^2,p}[x]}{\langle x^n-1\rangle}$ and $\frac{R_{u^2,v^2,p}[x]}{\langle x^n+1\rangle}$ are finite and of same order, $\phi$ is an onto map. It is easy to see that $\phi$ is a ring homomorphism. So $\phi$ is a ring isomorphism.
\end{proof}
\begin{remark} \label{restr-isomorphism}
We restrict the isomorphism $\phi$ to an isomorphism $\phi:\frac{\F_p[x]}{\langle x^n-1\rangle}\rightarrow\frac{\F_p[x]}{\langle x^n+1\rangle}$.
\end{remark}
Throughout this paper we use the isomorphism $\phi$ and restriction of $\phi$ defined in Proposition \ref{isomorphism} and Remark \ref{restr-isomorphism}.
\begin{prop} \label{cyclic-negacyclic}
Let $R$ be a ring. Let $A\subseteq \frac{R[x]}{\langle x^n-1\rangle}$, $B\subseteq \frac{R[x]}{\langle x^n+1\rangle}$ be two sets such that $\phi(A)=B$. Then $A$ is an ideal of $\frac{R[x]}{\langle x^n-1\rangle}$ if and only if $B$ is an ideal of $\frac{R[x]}{\langle x^n+1\rangle}$. Equivalently, $A$ is a cyclic code of length $n$ over the ring $R$ if and only if $B$ is a negacyclic code of length $n$ over the ring $R$.
\end{prop}
\begin{proof}
The proof is obvious since the map $\phi$ is a ring isomorphism.
\end{proof}
\begin{theo} \label{negacyclic}
Let $C$ be a negacyclic code of length $n$ over the ring $R_{u^2,v^2,p}$. Then $C$ will be of the form $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x),ug_2(x)+vg_{22}(x)+vug_{23}(x),vg_3(x)+vug_{33}(x),vug_4(x)\rangle$, where $g_4(x)|g_2(x)|g_1(x)|(x^n+1)$ and $g_4(x)|g_3(x)|g_1(x)|(x^n+1)$.
\end{theo}
\begin{proof}
The code $C$ is a negacyclic codes of length $n$ over the ring $R_{u^2,v^2,p}$. From Proposition \ref{cyclic-negacyclic}, we know that for the negacyclic code $C$ there exist a cyclic code, say $A$ over the same ring and of same length. We know the structure of a cyclic code over the ring $R_{u^2,v^2,p}$ from {\rm\cite{Pkk-Bg15}}. Let the cyclic code over the ring $R_{u^2,v^2,p}$ be $A=\langle g(x)+up_1(x)+vq_1(x)+vur_1(x),ua_1(x)+vq_2(x)+vur_2(x),va_2(x)+vur_3(x),vua_3(x)\rangle$, where $a_3(x)|a_1(x)|g(x)|(x^n-1)$ and $a_3(x)|a_2(x)|g(x)|(x^n-1)$. Now the polynomials $g(x)+up_1(x)+vq_1(x)+vur_1(x)$, $ua_1(x)+vq_2(x)+vur_2(x)$, $va_2(x)+vur_3(x)$, $vua_3(x)\in \frac{R[x]}{\langle x^n-1\rangle}$. Therefore from the definition of $\phi$ from Proposition \ref{isomorphism}, we get $\phi(g(x)+up_1(x)+vq_1(x)+vur_1(x))=g(-x)+up_1(-x)+vq_1(-x)+vur_1(-x)=g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)$, $\phi(ua_1(x)+vq_2(x)+vur_2(x))=ua_1(-x)+vq_2(-x)+vur_2(-x)=ug_2(x)+vg_{22}(x)+vug_{23}(x)$, $\phi(va_2(x)+vur_3(x))
=va_2(-x)+vur_3(-x)=vg_3(x)+vug_{33}(x)$, $\
\phi(vua_3(x))=vua_3(-x)=vug_4(x)$ and $\phi(x^n-1)=-(x^n+1)$, where $g(-x)=g_1(x)$, $a_1(-x)=g_2(x)$, $a_2(-x)=g_3(x)$, $a_3(-x)=g_4(x)$, $p_1(-x)=g_{11}(x)$, $q_1(-x)=g_{12}(x)$, $r_1(-x)=g_{13}(x)$, $q_2(-x)=g_{22}(x)$, $r_2(-x)=g_{23}(x)$, $r_3(-x)=g_{33}(x)$. Again $\phi(A)=C$. Therefore the code $C$ can be written as $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x),ug_2(x)+vg_{22}(x)+vug_{23}(x),vg_3(x)+vug_{33}(x), vug_4(x)\rangle$, where $g_4(x)|g_2(x)|g_1(x)|(x^n+1)$ and $g_4(x)|g_3(x)|g_1(x)|(x^n+1)$.
\end{proof}
\begin{theo}\label{unique}
Any negacyclic code $C$ of length $n$ over the ring $R_{u^2,v^2,p}$ is uniquely generated by the polynomials $A_1=g_1(x)+ug_{11}(x)+vg_{12}(x)+uvg_{13}(x), \linebreak A_2=ug_2(x)+vg_{22}(x)+uvg_{23}(x), A_3=vg_3(x)+uvg_{33}(x), A_4=uvg_4(x)$, where, $g_{ij}(x)$ are zero polynomial or $\text{deg}(g_{ij}(x))<\text{deg}(g_{j+1}(x))$ for $1\leq i\leq 3$, $i\leq j\leq 3$.
\end{theo}
\begin{proof}
The code $C$ is generated by the polynomial $A_1, A_2, A_3$ and $A_4$. Let $A=\langle g(x)+up_1(x)+vq_1(x)+vur_1(x),ua_1(x)+vq_2(x)+vur_2(x),va_2(x)+vur_3(x),vua_3(x)\rangle$ be the cyclic code over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$. Now, for any polynomial $f(x) \in \F_p[x]$ we get $\text{deg}(f(-x))=\text{deg}(f(x))$. Therefore, $\text{deg}(\phi(f(x)))=\text{deg}(f(x))$. From above theorem, we have $\phi(g(x))=g_1(x)$, $\phi(a_1(x))=g_2(x)$, $\phi(a_2(x))=g_3(x)$, $\phi(a_3(x))=g_4(x)$, $\phi(p_1(x))=g_{11}(x)$, $\phi(q_1(x))=g_{12}(x)$, $\phi(r_1(x))=g_{13}(x)$, $\phi(q_2(x))=g_{22}(x)$, $\phi(r_2(x))=g_{23}(x)$, $\phi(r_3(x))=g_{33}(x)$. Also from Theorem $3.1$ of {\rm\cite{Pkk-Bg15}}, we have $\text{deg}(p_1(x))<\text{deg}(a_1(x))$, $\text{deg}(q_1(x))<\text{deg}(a_2(x))$, $\text{deg}(r_1(x))<\text{deg}(a_3(x))$, $\text{deg}(q_2(x))<\text{deg}(a_2(x))$, $\text{deg}(r_2(x))<\text{deg}(a_3(x))$, $\text{deg}(r_3(x))<\text{deg}(a_3(x))$. Therefore, from the relation $\text{deg}(\phi(f(x)))=\
text{deg}(f(
x))$ we can write
$\text{deg}(\phi(p_1(x)))<\text{deg}(\phi(a_1(x)))$, $\text{deg}(\phi(q_1(x)))<\text{deg}(\phi(a_2(x)))$, $\text{deg}(\phi(r_1(x)))<\text{deg}(\phi(a_3(x)))$, $\text{deg}(\phi(q_2(x)))<\text{deg}(\phi(a_2(x)))$, $\text{deg}(\phi(r_2(x)))<\text{deg}(\phi(a_3(x)))$, $\text{deg}(\phi(r_3(x)))<\text{deg}(\phi(a_3(x)))$. This implies that $\text{deg}(g_{ij}(x))<\text{deg}(g_{j+1}(x))$ for $1\leq i\leq 3$, $i\leq j\leq 3$.
To prove uniqueness we assume that the polynomial $A_1'=g_1(x)+ug'_{11}(x)+vg'_{12}(x)+uvg'_{13}(x) \in C$ satisfies degree result $\text{deg}(g'_{1j}(x))<\text{deg}(g_{j+1}(x))$ for $1\leq j\leq 3$. Since, $\phi(A)=C$, therefore, there exists a polynomial $g(x)+up'_1(x)+vq'_1(x)+vur'_1(x)\in A$, such that $\phi(g(x)+up'_1(x)+vq'_1(x)+vur'_1(x))=g_1(x)+ug'_{11}(x)+vg'_{12}(x)+uvg'_{13}(x)$, where $\phi(p'_1(x))=p'_1(-x)=g'_{11}(x)$, $\phi(q'_1(x))=q'_1(-x)=g'_{12}(x)$, $\phi(r_1(x))=r'(-x)=g'_{13}(x)$. Since, $\text{deg}(\phi(f(x)))=\text{deg}(f(x))$, for all $f(x) \in A$, thus, $\text{deg}(\phi(p'_1(x)))=\text{deg}(p'_1(x))=\text{deg}(g'_{11}(x))$. Similarly, $\text{deg}(q'_1(x))=\text{deg}(g'_{12}(x))$, $\text{deg}(r'_1(x))=\text{deg}(g'_{13}(x))$. Therefore, $\text{deg}(p'_1(x))=\text{deg}(g'_{11}(x))<\text{deg}(g_2(x))=\text{deg}(a_1(x))$. This implies that $\text{deg}(p'_1(x))<\text{deg}(a_1(x))$ Similarly we get, $\text{deg}(q'_1(x))<\text{deg}(a_2(x))$, $\text{deg}(r'_1(x))<\text{deg}(a_3(x))$. But
from Theorem $3.1$ of {\rm\cite{Pkk-Bg15}}, we know that the polynomial $g(x)+up_1(x)+vq_1(x)+vur_1(x) \in A$ is unique which satisfying the degree result. Therefore, $p_1(x)=p'_1(x)$, $q_1(x)=q'_1(x)$ and $r_1(x)=r'_1(x)$. This implies that $\phi(p_1(x))=\phi(p'_1(x))$, thus $g_{11}(x)=g'_{11}(x)$. Similarly, $g_{12}(x)=g'_{12}(x)$ and $g_{13}(x)=g'_{13}(x)$. Hence, $A_1$ is unique. Similarly we can prove that $A_2, A_3$ and $A_4$ are also unique.\\
\end{proof}
\begin{theo}
Let $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x),ug_2(x)+vg_{22}(x)+vug_{23}(x),vg_3(x)+vug_{33}(x),vug_4(x)\rangle$ be a negacyclic code of length $n$ over the ring $R_{u^2,v^2,p}$. Then we must have the following properties
\begin{align}
&g_4(x)|g_3(x)|g_1(x), g_4(x)|g_2(x)|g_1(x)|(x^n+1),\\
&g_{i+1}(x)|\frac{x^n+1}{g_i(x)}g_{ii}(x), ~ {\rm{for}} ~ 1 \leq i \leq 3,\\
&g_3(x)|\frac{g_1(x)}{g_2(x)}g_{22}(x)\\
&g_4(x)|g_{22}(x)\\
&g_4(x)|\left(g_{11}(x)-\frac{g_1(x)}{g_3(x)}g_{33}(x)\right)\\
&g_4(x)|\left(g_{12}(x)-\frac{g_1(x)}{g_2(x)}g_{23}(x)+\frac{g_1(x)}{g_2(x)g_3(x)}g_{22}(x)g_{33}(x)\right)\\
&g_{i+j+1}(x)|\frac{x^n+1}{g_i(x)}s_{i(i+j)} ~{\rm{for}}~ 1 \leq i \leq 2 ~{\rm{and ~ for ~ a ~ fix}}~ i ~{\rm{for}}~ 1 \leq j \leq 3-i,\notag \\&{\rm{where}},~ s_{ii}=g_{ii} ~ {\rm{and}} ~ s_{i(i+j)}=g_{i(i+j)}-\sum_{l=1}^j\frac{s_{i(i+l-1)}}{g_{i+l}(x)}g_{(i+l)(i+j)}(x).
\end{align}
\end{theo}
\begin{proof}
Let $A=\langle g(x)+up_1(x)+vq_1(x)+vur_1(x),ua_1(x)+vq_2(x)+vur_2(x),\linebreak va_2(x)+vur_3(x),vua_3(x)\rangle$ be the cyclic code over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$. Also, from Theorem \ref{negacyclic}, we have $\phi(g(x))=g_1(x)$, $\phi(a_1(x))=g_2(x)$, $\phi(a_2(x))=g_3(x)$, $\phi(a_3(x))=g_4(x)$, $\phi(p_1(x))=g_{11}(x)$, $\phi(q_1(x))=g_{12}(x)$, $\phi(r_1(x))=g_{13}(x)$, $\phi(q_2(x))=g_{22}(x)$, $\phi(r_2(x))=g_{23}(x)$, $\phi(r_3(x))=g_{33}(x)$. Now, from Remark \ref{restr-isomorphism}, we get that the map $\phi$, $\phi:\frac{\F_p[x]}{\langle x^n-1\rangle}\rightarrow\frac{\F_p[x]}{\langle x^n+1\rangle}$ such that $\phi(f(x))=f(-x)$, $\forall ~ f(x)\in \frac{\F_p[x]}{\langle x^n-1\rangle}$ is an isomorphism. Now from Proposition $3.2$ of {\rm\cite{Pkk-Bg15}}, we know that the properties are true for the ring $\frac{\F_p[x]}{\langle x^n-1\rangle}$. Therefore all of these properties are true for the ring $\frac{\F_p[x]}{\langle x^n+1\rangle}$.
\end{proof}
The following theorem characterizes the free negacyclic codes over the ring $R_{u^2,v^2,p}$.
\begin{theo} \label{freecode}
If $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x),ug_2(x)+vg_{22}(x)+vug_{23}(x),vg_3(x)+vug_{33}(x),vug_4(x)\rangle$ be a negacyclic code of length $n$ over the ring $R_{u^2,v^2,p}$, then $C$ is a free negacyclic code if and only if $g_1(x)=g_4(x)$. In this case, we have $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)\rangle$ and $g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)|(x^n+1)$ in $R_{u^2,v^2,p}$.
\end{theo}
\begin{proof}
We are given that $C$ is a negacyclic code over the ring $R_{u^2,v^2,p}$. Hence from Proposition \ref{cyclic-negacyclic}, there exist one and only one cyclic code $A$ over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$. Let the cyclic code be $A=\langle g(x)+up_1(x)+vq_1(x)+vur_1(x),ua_1(x)+vq_2(x)+vur_2(x),va_2(x)+vur_3(x),vua_3(x)\rangle$, where\linebreak $a_3(x)|a_1(x)|g(x)|(x^n-1)$ and $a_3(x)|a_2(x)|g(x)|(x^n-1)$. Therefore we have, $\phi(g(x)+up_1(x)+vq_1(x)+vur_1(x))=g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)$, $\phi(ua_1(x)+vq_2(x)+vur_2(x))=ug_2(x)+vg_{22}(x)+vug_{23}(x)$, $\phi(va_2(x)+vur_3(x))=vg_3(x)+vug_{33}(x)$, $\phi(vua_3(x))=vug_4(x)$ and $\phi(x^n-1)=-(x^n+1)$. Now, it is given $g_1(x)=g_4(x)$. Since $\phi$ is an isomorphism therefore $g(x)=a_3(x)$. We know from Proposition $3.3$ of {\rm\cite{Pkk-Bg15}} that $A=\langle g(x)+up_1(x)+vq_1(x)+vur_1(x)\rangle$ if and only if $g(x)=a_3(x)$. Now $\phi(g(x)+up_1(x)+vq_1(x)+vur_1(x))=g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)$, Hence $C=\langle g_1(x)
+ug_{11}(x)+vg_{12}(x)+vug_{13}(
x)\rangle$. Again we have $\phi(x^n-1)=-(x^n+1)$ and we know from Proposition $3.3$ of {\rm\cite{Pkk-Bg15}} that $g(x)+up_1(x)+vq_1(x)+vur_1(x)|(x^n-1)$. Hence $g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)|(x^n+1)$.
\end{proof}
Note that we get the simpler form for the generators of the negacyclic code over $R_{u^2,v^2,p}$, like in the above theorem, if we have $g_1(x) = g_2(x), g_3(x)$ or $g_4(x) = g_2(x), g_3(x)$.\\
In the following theorem we write the structure of $C$ when $n$ be relatively prime to $p$.
\begin{theo}
Let $C$ be a negacyclic code over the ring $R_{u^2,v^2,p}$ of length $n$. If $n$ is relatively prime to $p$, then we have $C=\langle g_1(x)+ug_2(x)+uvg_{13}(x), vg_3(x)+uvg_4(x)\rangle$ with $g_2(x)|g_1(x)|(x^n+1)$ and $ g_4(x)|g_3(x)|g_1(x)|(x^n+1)$.
\end{theo}
\begin{proof}
Let $C$ be a negacyclic code over the ring $R_{u^2,v^2,p}$. Hence from Proposition \ref{cyclic-negacyclic}, there exists one and only one cyclic code $A$ over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$. If $n$ is relatively prime to $p$, then from Theorem $3.4$ of {\rm\cite{Pkk-Bg15}}, we can write the cyclic code $A=\langle g(x)+ua_1(x)+uvr_1(x), va_2(x)+uva_3(x)\rangle$ with $a_1(x)|g(x)|(x^n-1)$ and $ a_3(x)|a_2(x)|g(x)|(x^n-1)$.
Let $\phi(g(x)+ua_1(x)+uvr_1(x))=g_1(x)+ug_2(x)+uvg_{13}(x)$ and $\phi(va_2(x)+uva_3(x))=vg_3(x)+uvg_4(x)$, where, $g(-x)
=g_1(x)$, $r_1(-x)=g_{13}(x)$, $a_1(-x)=g_2(x)$, $a_2(-x)=g_3(x)$, $a_3(-x)=g_4(x)$. Therefore $C$ can be written as $C=\langle g_1(x)+ug_2(x)+uvg_{13}(x), vg_3(x)+uvg_4(x)\rangle$ with $g_2(x)|g_1(x)|(x^n+1)$ and $ g_4(x)|g_3(x)|g_1(x)|(x^n+1)$.
\end{proof}
\section{The Ranks and the minimum distance}\label{rank}
In this section, we find the rank and minimal spanning set of negacyclic codes over the ring $R_{u^2,v^2,p}$. Following Dougherty and Shiromoto \cite[page 401]{Dou-Shiro01}, we define the rank of the code $C$ by the minimum number of generators of $C$ and define the free rank of $C$ by the maximum of the ranks of $R_{u^2,v^2,p}$-free submodules of $C$.\\
\begin{theo}
Let $n$ be not relatively prime to $p$. Let $C$ be a negacyclic code over the ring $R_{u^2,v^2,p}$ of length $n$. If $C=\langle g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x),ug_2(x)+vg_{22}(x)+vug_{23}(x),vg_3(x)+vug_{33}(x),vug_4(x)\rangle$ with $deg(g_1(x))\linebreak =r_1$, $deg(g_2(x))=r_2$, $deg(g_3(x))=r_3$, $deg(g_4(x))=r_4$, then the minimal spanning set of $C$ is $B=\{A_1, xA_1, \cdots , x^{n-r_1-1}A_1, A_2, xA_2, \cdots , x^{r_1-r_2-1}A_2,\linebreak A_3, xA_3, \cdots , x^{r_1-r_3-1}A_3, A_4, xA_4, \cdots, x^{r'-r_4-1}A_4\}$, where, $r'=min\{r_2, r_3\}$ and $A_1=g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)$, $A_2= ug_2(x)+vg_{22}(x)+vug_{23}(x)$, $A_3=vg_3(x)+vug_{33}(x)$, $A_4=vug_4(x)$ also $C$ has free rank $n-r_1$ and rank $n+r_1+r'-r_2-r_3-r_4$.
\end{theo}
\begin{proof}
Let $C$ be a negacyclic code over the ring $R_{u^2,v^2,p}$ of length $n$, where $n$ is not relatively prime to $p$. From the Proposition \ref{cyclic-negacyclic}, we get that there exists a cyclic code $A$ over the same ring such that $\phi(A)=C$. Now, we know from Theorem $4.1$ of {\rm\cite{Pkk-Bg15}}, the minimal spanning set of the cyclic code $A$ over the ring $R_{u^2,v^2,p}$ is $\{g(x)+up_1(x)+vq_1(x)+vur_1(x), x(g(x)+up_1(x)+vq_1(x)+vur_1(x)), \cdots, x^{n-r_1-1}(g(x)+up_1(x)+vq_1(x)+vur_1(x)), ua_1(x)+vq_2(x)+vur_2(x), x(ua_1(x)+vq_2(x)+vur_2(x)), \cdots, x^{r_1-r_2-1}(ua_1(x)+vq_2(x)+vur_2(x)), va_2(x)+vur_3(x), x(va_2(x)+vur_3(x)), \cdots, x^{r_1-r_3-1}(va_2(x)+vur_3(x)),\linebreak vua_3(x), x(vua_3(x)), \cdots, x^{r'-r_4-1}(vua_3(x))\}$, where, $r'=min\{r_2, r_3\}$. From Theorem \ref{negacyclic} we have $\phi(g(x)+up_1(x)+vq_1(x)+vur_1(x))=A_1, \phi(ua_1(x)+vq_2(x)+vur_2(x))=A_2, \phi(va_2(x)+vur_3(x))=A_3$ and $\phi(vua_3(x))=A_4$. Therefore the spanning set of negacyclic code $C$ over the ring $R_{
u^2,v^2,p}$ is $B=\{A_1, xA_1, \cdots , x^{n-r_
1-1}A_1, A_2, xA_2, \cdots , x^{r_1-r_2-1}A_2, A_3, xA_3, \cdots , x^{r_1-r_3-1}A_3, A_4,\linebreak xA_4, \cdots, x^{r'-r_4-1}A_4\}$, where, $r'=min\{r_2, r_3\}$ and $A_1=g_1(x)+ug_{11}(x)+vg_{12}(x)+vug_{13}(x)$, $A_2= ug_2(x)+vg_{22}(x)+vug_{23}(x)$, $A_3=vg_3(x)+vug_{33}(x)$, $A_4=vug_4(x)$.
\end{proof}
Let $n$ be a positive integer not relatively prime to $p$. Let $C$ be a negacyclic code of length $n$ over the ring $R_{u^2,v^2,p}$. We know that there exists a cyclic code $A$ of length $n$ over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$, where, $\phi$ is defined as $\phi(f(x))=f(-x)$, for $f(x) \in A$. The following lemma shows that the isomorphism $\phi$ is a distance preserving map.\\
\begin{lemma}\label{lm}
Let $\phi(A)=C$, where, $A$ and $C$ are the cyclic and negacyclic codes of length $n$ over the ring $R_{u^2,v^2,p}$ and $\phi$ is defined as $\phi(f(x))=f(-x)$, for $f(x) \in A$, then, $w_{H}(f(x)) = w_{H}(\phi(f(x)))$.
\end{lemma}
\begin{proof}
Let $f(x)=\sum_{i=0}^{n-1}c_ix^i$, where, $c_i \in R_{u^2,v^2,p}$. Now $\phi(f(x))=f(-x)=\sum_{i=0}^{n-1}(-1)^ic_ix^i$. Therefore the coefficient of $x^i$ of $f(x)$ and $f(-x)$ are $c_i$ and $(-1)^ic_i$. That is both coefficient are simultaneously $0$ or non $0$. Hence, $w_{H}(f(x))=w_{H}(\phi(f(x)))$.
\end{proof}
\begin{theorem} \label{md1}
Let $n$ be not relatively prime to $p$. If $ C = \langle A_1, A_2, A_3, A_4\rangle$ is a negacyclic code of length $n$ over the ring $R_{u^2,v^2,p}$. Then $w_{H}(C)=w_{H}(A)$, where, $A$ is the cyclic codes over the ring $R_{u^2,v^2,p}$ such that $\phi(A)=C$.
\end{theorem}
\begin{proof}
Let $h(x)$ be the minimum weighted polynomial in $A$ and the weight is $w_{H}(h(x))=m$. There exists a polynomial $f(x) \in C$ such that $\phi(h(x))=f(x)$. From Lemma \ref{lm}, the weight of $w_{H}(f(x))=m$. Now we prove that $f(x)$ is the minimum weighted polynomial in $C$. If possible, let $f_1(x)$ be the minimum weighted polynomial of $C$ and $w_{H}(f_1(x)) < m$. There exists a polynomial $h_1(x) \in A$ such that $\phi(h_1(x))=f_1(x)$. Again, from Lemma \ref{lm}, $w_{H}(h_1(x))=w_{H}(f_1(x))<m$. Hence, a contradiction that $h(x)$ be the minimum weighted polynomial in $A$. Therefore, $w_{H}(C)=w_{H}(A)$.
\end{proof}
\begin{definition}
Let $ m = b_{l-1}p^{l-1} + b_{l-2}p^{l-2} + \cdots + b_1p + b_0$, $b_i \in \F_p, 0
\leq i \leq l-1$, be the $p$-adic expansion of $m$.
\begin{enumerate} [{\rm (1)}]
\item If $ b_{l-i} \neq 0$ for all $1 \leq i \leq q, q < l, $ and $ b_{l-i} = 0 $ for all $i, q+1 \leq i \leq l$, then $m$ is said to have a $p$-adic length $q$ zero expansion.
\item If $ b_{l-i} \neq 0$ for all $1 \leq i \leq q, q < l, $ $b_{l-q-1} = 0$ and $ b_{l-i} \neq 0 $ for some $i, q+2 \leq i \leq l$, then $m$ is said to have $p$-adic length $q$ non-zero expansion.
\item If $ b_{l-i} \neq 0$ for $1 \leq i \leq l, $ then $m$ is said to have a $p$-adic length $l$ expansion or $p$-adic full expansion.
\end{enumerate}
\end{definition}
The following theorem follows from the above theorem and Theorem 5.4 of \cite{Pkk-Bg15}.
\begin{theorem} \label{md-thm}
Let $C$ be a negacyclic code over the ring $R_{u^2,v^2,p}$ of length $p^l$ where $l$ is a positive integer. Then, $C=\langle A_1, A_2, A_3, A_4 \rangle$, where, $g_1(x) = (x+1)^{t_1}, g_2(x) = (x+1)^{t_2}, g_3(x) = (x+1)^{t_3}, g_4(x) = (x+1)^{t_4}$, for some $t_1 > t_2 > t_4 > 0$, $t_1 > t_3 > t_4 > 0$ $(where A_i$'s and $g_i$'s are defined in Theorem \ref{unique}$)$
\begin{enumerate}[{\rm (1)}]
\item If $t_4 \leq p^{l-1},$ then $d(C) = 2$.
\item If $t_4 > p^{l-1}$, let $t_4 = b_{l-1}p^{l-1} + b_{l-2}p^{l-2} + \cdots + b_1p + b_0$ be the $p$-adic expansion of $t_4$ and $g_4(x)=(x+1)^{t_4}=(x^{p^{l-1}}+1)^{b_{l-1}}(x^{p^{l-2}}+1)^{b_{l-2}} \cdots (x^{p^{1}}+1)^{b_1}(x^{p^0}+1)^{b_0}$.
\begin{enumerate}[{\rm ($a$)}]
\item If $t_4$ has a $p$-adic length $q$ zero expansion or full expansion $(l=q)$, then $d(C) = (b_{l-1}+1)(b_{l-2}+1)\cdots(b_{l-q}+1)$.
\item If $t_4$ has a $p$-adic length $q$ non-zero expansion, then $d(C)=2(b_{l-1}+1)(b_{l-2}+1)\cdots(b_{l-q}+1)$.
\end{enumerate}
\end{enumerate}
\end{theorem}
\section{Examples} \label{exm}
\begin{example}
Negacyclic codes of length $5$ over $R_{u^2,v^2,5} = \F_5 + u \F_5 + v \F_5 + uv \F_5, u^2 = 0, v^2 = 0, uv = vu$: We have
$$ x^5+1=(x+1)^5 ~\text{over}~ R_{u^2,v^2,5}.$$
Let $g=x+1$ and $c_0, c_1, c_2, c_3, c_4, c_5 \in \F_5$. The non-zero negacyclic codes of length $5$ over $R_{u^2,v^2,5}$ with generator polynomial, rank and minimum distance are given in tables below.\\
\end{example}
\begin{center}
{\bf Table 1.} All non zero free negacyclic codes of length 5 over $R_{u^2,v^2,5}$.
\begin{tabular}{| l | c| c |}
\hline
Non-zero generator polynomials & Rank & d(C)\\
\hline
$<g^4+uc_0g^3+vc_1g^3+uvc_2g^3>$, $c_0c_1=0$&1&5\\
\hline
$<g^3+uc_0g^2+vc_1g^2+uv(c_2+c_3x)g>$&2&4\\
\hline
$<g^2+u(c_0+c_1x)+v(c_2+c_3x)+uv(c_4+c_5x)>$, &3&3\\ $c_0=c_1$ or $c_2=c_3$&&\\
\hline
$<g+uc_0+vc_1+uvc_2>$&4&2\\
\hline
$<1>$ &5&1\\
\hline
\end{tabular}\\
\end{center}
\newpage
{\bf Table 2.} All non zero non free single generated negacyclic codes of length 5 over $R_{u^2,v^2,5}$.
\begin{center}
\begin{tabular}{| l | c| c |}
\hline
Non-zero generator polynomials & Rank & d(C)\\
\hline
$<ug^4+vc_0g^4+uvc_1g^3>$&1&5 \\
\hline
$<vg^4+uvc_0g^3>$&1&5 \\
\hline
$<uvg^4>$&1&5 \\
\hline
$<ug^3+v(c_0+c_1x)g^3+uv(c_2+c_3x)g>$&2&4 \\
\hline
$<vg^3+uv(c_0+c_1x)g>$&2&4\\
\hline
$<uvg^3>$&2&4\\
\hline
$<ug^2+v(c_0+c_1x+c_2x^2)g^2+uv(c_3+c_4x)>$&3&3\\
\hline
$<vg^2+uv(c_0+c_1x)>$&3&3\\
\hline
$<uvg^2>$&3&3\\
\hline
$<ug+v(c_0+c_1x+c_2x^2+c_3x^3)g+uvc_4>$&4&2\\
\hline
$<vg+uvc_0>$&4&2\\
\hline
$<uvg>$&4&2\\
\hline
$<u+v(c_0+c_1x+c_2x^2+c_3x^3+c_4x^4)>$&5&1 \\
\hline
$<v>$&5&1 \\
\hline
$<uv>$&5&1 \\
\hline
\end{tabular}\\
\end{center}
\begin{center}
{\bf Table 3.} Some non zero non free negacyclic codes of length 5 over $R_{u^2,v^2,5}$.
\begin{tabular}{| l | c| c |}
\hline
Non-zero generator polynomials & Rank & d(C)\\
\hline
$<g^4+uc_0g^3+vc_1g^3+uvc_2g^2, uvg^3>$&2&4\\
\hline
$<ug^4+uvc_0g^3, vg^4+uvc_1g^3>$&2&5\\
\hline
$<ug^4+v(c_0+c_1x)g^3+uvg^2, uvg^3>$&2&4\\
\hline
$<ug^4+vc_0g^3+uvc_1g^2, vg^4>$&2&5\\
\hline
$<ug^4+uvc_0g^2, uvg^3>$&2&4\\
\hline
$<vg^4+uvc_0g^2, uvg^3>$&2&4\\
\hline
$<vg^4+uvc_0g, uvg^2>$ &3&3\\
\hline
$<g^3+uc_0g+vc_1g+uvc_2, ug^2+vc_3g+uvc_4$,&5&2\\ $vg^2+uvc_5, uvg>$, $c_0c_2=0$&&\\
\hline
$<ug^3+v(c_0+c_1x)g^3+uv(c_2+c_3x), uvg^2>$&3&3\\
\hline
$<vg^3+uvc_0, uvg>$ &4&2\\
\hline
$<g^2+uc_0+vc_1, ug+vc_2, vg, uv>$&6&1\\
\hline
$<ug^2+vc_0+uvc_1, vg^2+uvc_2, uvg>$&7&2\\
\hline
$<vg^2+uvc_2, uvg>$&4&2\\
\hline
$<g+uc_0+vc_1, uv>$&5&1\\
\hline
$<g+uc_0, v>$&5&1\\
\hline
$<g+vc_0, u+vc_1 >$&5&1\\
\hline
$<g, u, v>$&6&1\\
\hline
$<ug+vc_0, vg, uv>$&9&1 \\
\hline
$<vg, uv>$&5&1 \\
\hline
$< u, v >$&10&1 \\
\hline
\end{tabular}
\end{center}
\bibliographystyle{plain}
|
\section{Introduction}
Big data explosion due to advancement in social networking and IoT (Internet of Things) produces immense data, which urge the data scientists to strive for the development of better data storage medium. Optical, digital and cloud data storage \cite{dimakis2010network} medium have their own limitations and need to be maintained regularly. While the computer scientists are endeavoring to develop dense data storage medium, researchers at the other end thought of exploring natural medium to safeguard this data
Properties of DNA like scalability, density and long term stability makes it ideal for long term archival of data. A good deal of work has been done to store the data on DNA \cite{davis1996microvenus} \cite{cox2001long} \cite{wong2003organic} \cite{church2012next}.
By employing error correction for DNA data storage, the breakthrough was made by N. Goldman and his team \cite{goldman2013towards} in the year $2013.$ They developed very efficient and novel approach for storing the data of size $739$ KB on DNA and retrieving it back. They used four folds redundancy to tackle errors by retrieving correct data from one of the copies of DNA strands. But due to redundancy there was increase in the length of DNA which makes the technique expensive to use DNA for storage at the commercial level. In the year $2014,$ group at The Chinese University of Hong Kong developed method for larger DNA reads used to embed the data \cite{yim2014essential}. Low Density Parity-Check (LDPC) codes based method was introduced by Aldrin Kay-Yuen Yim et al., to encode data in DNA for improved length of DNA reads. Most of the methods used so far have limitations in size of the data inserted in DNA and cost associated with DNA synthesis and sequencing technology. Work of Goldman \cite{goldman2013towards} and Church \cite{church2012next} has laid a cornerstone for DNA based data storage systems. Recently model of DNA channel for data storage is proposed by Han Mao Kiah et al., \cite{KiahPM14}. Also a rewritable random access DNA based storage system has been very recently reported \cite{rewritableDNAoligica}. In \cite{grass2015robust}, Reed Solomon codes based DNA storage system has been reported. In this work, we have developed error correction scheme (a family of non linear ternary codes) by modifying Goldman's scheme \cite{goldman2013towards} for the error detection and correction along with improvement in storage capacity to store data on DNA.
The paper is organized as follows. Section 2 describes briefly related work and the Golddman's approach. Section 3 gives our work on non-linear ternary codes. Section 4 discuss the analysis of class DNA Golay code sub code. Final section 5 concludes the paper with general remarks.
\section{Error correction for DNA storage}
There are various error correction schemes used for DNA based data storage systems \cite{smith2003some} \cite{yachie2007alignment} \cite{haughton2011repetition} but the effective schemes was first introduce by Church \cite{church2012next} and Goldman \cite{goldman2013towards}.
Using next generation synthesis and sequencing technology, Church came up with efficient one bit per base algorithm of encoding information bits into fix length of DNA chunks (99 bases). Flanking primers at the beginning and end of information data was inserted to identify the specific DNA segment in which the particular data was encoded.
This method includes homo polymer repeated sequences that cause errors while writing and reading the data. This difficulty was taken care by Goldman's approach for DNA based data storage in $2013.$ Goldman used one bit per base system introduced by Church and modifying it by employing the improved base $3$ Huffman code (trits 0, 1 and 2) encoding scheme.
In this original file type to binary code which is then converted to a ternary code which is in turn converted to the triplet DNA code. It involved four steps shown in Fig \ref{AdvancedDNASchemetic}. Binary digits holding the ASCII codes was converted to base-3 Huffman code that replaces each byte with five or six base-3 digits (trits). Each of trit was encoded with one of the three nucleotides different from the previous one used (to avoid homopolymers that cause error in the synthesis of DNA). DNA strand was divided into chunks each of length $117$ base pair (bp). To include redundancy for error detection and correction, $75$ bases for each DNA information chunks were overlapped for four fold redundancy to get the data loss that occurred during synthesis and sequencing DNA. For the data security each redundant chunk was converted to reverse complement of the strand in every alternate chunks. Each DNA chunk was appended with an indexing information bit to determine the location of the data in each chunk and the file from which it is generated. At the end parity check bit was added for error detection. For details the reader is referred to \cite{goldman2013towards}. We modified the Goldman's scheme further as described in next section.
\begin{figure}[h]
\begin{center}
\scalebox{0.32}[0.35]{\includegraphics{Goldmanalgo.png}}
\caption{Stepwise encoding of data into DNA using Goldman's approach is explained in detail. In this original file type to binary code (0, 1) which is then converted to a ternary code (0, 1, 2), which is in turn converted to the DNA code. Four fold redundancy is added to each DNA chunk by introducing reverse complement to every alternate DNA chunk. DNA chunk index is addedto each DNA chunk to locate the file.}
\label{AdvancedDNASchemetic}
\end{center}
\end{figure}
\section{Family of Non-Linear Ternary Codes}
Goldman used four fold redundancy due to which was increase in the length of the DNA used to encode data which makes the DNA based storage medium more expensive. In this work, non-linear ternary error correction codes is used instead of Huffman codes as referred in Fig \ref{golaycode}. In DNA we have only three nucleotides different from the last one. So it is inescapable to have codewords in base-3 in order to avoid homopolymers runs \cite{goldman2013towards}. The non-linear ternary codes described here improve the storage capacity of DNA by decreasing the length of the DNA required for storing a file. Any arbitrary computer file can be converted into list of ASCII values ranging from $0$ to $255$. So we need set of $256$ ternary codewords, each corresponding to one value in $\{0,..,255\},$ to encode any such file into DNA string. From exhaustive search and enumeration of codewords we identified and constructed seven families of non linear ternary codes with the parameters $(9, 256, 3)_{3}$, $(11, 256, 5)_{3}$, $(15, 256, 7)_{3}$, $(18, 256, 9)_{3}$, $(21, 256, 11)_{3}$, $(24, 256, 13)_{3}$ and $(26, 256, 15)_{3}$ respectiveley \cite{codetables}. Among these, $(11, 256, 5)_{3}$ is a sub code of ternary Golay code (listed in the Table \ref{DNA golay code}) \cite{2401740}. We discuss encoding using the subcode $(11,256,5)_{3}$ in detail. The ternary Golay code consists of $729$ codewords of length $11,$ with minimum hamming distance $5.$ So it allows receiver to identify 4 trits of errors and correct 2 trits of errors that occur in codeword.
The steps b and c of the algorithm shown in Fig \ref{AdvancedDNASchemetic} were modified by using ternary Golay codes instead of Huffman codes (see Fig \ref{DNAgolay}). Each byte of a computer file is encoded to Golay code Table \ref{DNA golay code}. The table consists of $243$ codewords from $729$ Golay codewords such that minimum hamming distance between any two codewords is $6.$ These $243$ codewords were assigned to $243$ ASCII values having highest probability of occurrences according to frequency table used by Goldman as described in supplementary file. Remaining $13$ ASCII values were assigned with codewords chosen randomly from remaining set of ternary Golay codewords i.e. these $13$ codewords will have minimum hamming distance $5$ with other $243$ codewords. This ensures that by maximum likelihood decoding we can correct up to $2$ trit flip error. Since there exists only $243$ codewords in set of $3^{11}$ codewords with minimum hamming distance $6,$ it is not possible to construct code consisting of minimum $256$ codewords such that length of codewords is $11$ and minimum hamming distance more than $5.$ The file information (i.e. file size and file extension) was encoded using same table and appended at the end of this string with Golay codes of comma and semicolon as separators. Next, this string of trits was converted to DNA such that only one of three nucleotides, different from the last one, was used to encode current bit into nucleotide. The resulting DNA sequence was split into segments (or chunks) of $99$ bases each as shown in Fig \ref{chunk} where $(i= 99$) is number of bases used for storing file content. Here, $\lambda$ is number of bases required to store file index number (no of file index trits = 2, thus allowing maximum of $9$ files to be distinguished), $\mu$ is number of bases required for storing chunk index (no of segment index trits $\mu = \left \lceil \log_{3}{\mbox{( total no of segments)}} \right \rceil )$ and one base for odd parity-check trit \cite{goldman2013towards} were appended at the end of each segment of $99$ bases. This parity is obtained by summing odd bits for file identifier and chunk index.
\begin{figure}[h]
\begin{center}
\scalebox{0.4}[0.4]{\includegraphics{chunk.jpg}}
\caption{Chunk architecture for DNA Golay code approach. There is two main parts of chunk, one is chunk info bits and other is chunk header. Chunk info bits contains original data information and chunk header has index. Chunk header includes file index for identification of file and chunk index to identify chunk. Last is appended odd parity check information.}
\label{chunk}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\scalebox{0.32}[0.32]{\includegraphics{golaycode_flowchart.png}}
\caption{Schematic flow diagram for DNA Golay code approach. Computer file is compressed using standard compression methods. Compressed binary file is converted to non linear family of codes by assigning each character to codeword. This ternary code is converted to DNA and DNA sequence is fragmented into chunk of variable length. No redundancy is added to chunk. Error correction can be done by nearest-neighbour likelihood decoding approach.}
\label{golaycode}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\scalebox{0.22}[0.25]{\includegraphics{AdvancedDNA_Schemetic.png}}
\caption{Improved error correction approach for DNA data storage using DNA Golay code is developed here. It depicts the steps of conversion of file into DNA sequence by using DNA Golay code table. In first step, binary code (blue bar) was mapped to base 3 non linear ternary code (orange bar) where each byte was replaced with base 3 numbers called trits (for ternary Golay code with 11 trits). These trits were then converted to DNA code (green bar) by replacing each trit with one of the three nucleotides different from the previous one used to avoid homopolymers. Long DNA was divided into DNA chunks of length 99 base pairs without redundancy. Each DNA segment was appended with index, file identifier and parity trits.}
\label{DNAgolay}
\end{center}
\end{figure}
\begin{table}[ht]
\caption{Goldman base table designed to covert Huffman code to DNA nucleotides avoiding homo polymer}
\centering
\begin{tabular}{|l|l|l|l|}
\hline
&0&1&2 \\ \hline
A&C&G&T \\ \hline
C&G&T&A \\ \hline
G&T&A&C \\ \hline
T&A&C&G \\ \hline
\end{tabular}
\label{basetable}
\end{table}
\begin{example}
To understand the encoding and decoding procedure, here is demonstrated simple example. Message data is "DA". Each letter of message is converted to ASCII characters. ASCII characters for each letter are 68 and 65 respectively. ASCII characters are converted to sub code of ternary Golay code of length 11. Codes corresponding to letter D is 02221221120 and for A is 10111000101 referred from table \ref{DNA golay code}.
DNA sequence corresponding to message DNA is CATGATGCTGAGTCTCGTAGTC. This DNA sequence is divided to chunks $C_1$ is CATGATGCTGA and $C_2$ is GTCTCGTAGTC (here of length 11). Index of chunk i and file identifier (ID) is appended at the end of the message data. Let the chunk index for each DNA chunk be 0 and 1. File identifier for the file be 00. Also for checksum, parity bit P is added which can be calculated by summation of odd position trits in ID and i. Now each DNA chunk is appended by with chunk identifier bases (concatenating ID. i .P). Chunk identifier for each DNA chunk here is CGTA and CGAG respectively. So final DNA chunks are CATGATGCTGACGTA and GTCTCGTAGTCCGAG.
\end{example}
\subsection{Error Correction Analysis}
\begin{definition}
Let $\Sigma_{DNA} =\{A,T,G,C\}$ be the alphabet of DNA nucleotides. A DNA code $\mathscr{C}_{DNA}(n,M)$ is a sub set of $\Sigma_{DNA}^n$ of size $M$ with each codewords of length $n$.
\end{definition}
\begin{definition}
The Hamming distance $d_{H}(\bar{\textbf{x}}_{\textbf{DNA}}$,$\bar{\textbf{y}}_{\textbf{DNA}})$ between two DNA codewords $\bar{\textbf{x}}_{\textbf{DNA}}$ and $\bar{\textbf{y}}_{\textbf{DNA}}$ is the number of positions in which $\bar{\textbf{x}}_{\textbf{DNA}}$ and $\bar{\textbf{y}}_{\textbf{DNA}}$ differs. For instance, let $\bar{\textbf{x}}_{\textbf{DNA}}$ = ATGACT $\bar{\textbf{y}}_{\textbf{DNA}}$ = ATTAGC, then $d_{H}(\bar{\textbf{x}}_{\textbf{DNA}}$,$\bar{\textbf{y}}_{\textbf{DNA}})$ = 3.
\end{definition}
\begin{Lemma}\label{uniq}
Let $\phi$ be the map used for the conversion (using Table \ref{basetable}) of $\mathscr{C}_{Z_3}$ ternary code to $\mathscr{C}_{DNA}$ DNA code i.e, $\phi(\mathscr{C}_{Z_3})=\mathscr{C}_{DNA}$.
Observe that the map $\phi$ is not unique.
\label{map}
\end{Lemma}
\begin{proof} Easy to observe.
\end{proof}
\begin{Lemma}
Let $\bar{\textbf{x}}_{\textbf{DNA}}$ $\in \mathscr{C}_{DNA}$ send through a noisy channel and $\bar{\textbf{y}}_{\textbf{DNA}}$ is received. If
$d_{H}(\bar{\textbf{x}}_{\textbf{DNA}}$,$\bar{\textbf{y}}_{\textbf{DNA}})$ $=$ $t$ and corresponding ternary code obtained using Goldman's base Table \ref{basetable} for $\bar{\textbf{x}}_{\textbf{DNA}}$ and $\bar{\textbf{y}}_{\textbf{DNA}}$ are $\bar{\textbf{x}}$ and $\bar{\textbf{y}}$ respectively then $t <$ $d_{H}(\bar{\textbf{x}}$,$\bar{\textbf{y}})$ $\leq$ $2t$.
\label{distance}
\end{Lemma}
\begin{proof} Using Lemma \ref{uniq}.
\end{proof}
\begin{example}
let $\bar{\textbf{x}}_{\textbf{DNA}} = GTCTCGTAGTC$ and $\bar{\textbf{y}}_{\textbf{DNA}} = GAGTCGTAGTC$ then $\bar{\textbf{x}} = 10111000101 $ and $\bar{\textbf{y}} = 11101000101 $. The minimum hamming distance $d_{H}(\bar{\textbf{x}}_{\textbf{DNA}}$,$\bar{\textbf{y}}_{\textbf{DNA}}) = 2$ and $d_{H}(\bar{\textbf{x}}$,$\bar{\textbf{y}})= 2$.
\end{example}
\begin{Lemma}
Using double layer error correcting scheme, we can correct any 2 bit flips in DNA.
\end{Lemma}
\begin{proof}
Correction of 1 bit flip in DNA is straightforward using Lemma \ref{uniq}. In case of 2 bit flips we can prove it by contradiction and Lemma 1 and 2.
\end{proof}
\begin{remark}
If error occurs in chunk header, error cannot be corrected and data cannot be retrieved.
\end{remark}
\section{Analysis}
Performance of error correcting codes used for DNA data storage can be measured by Shannon information capacity, information ratio of base to bits.
Family of non linear codes described here has many advantages over codes used for DNA data storage. It has properties like higher data density, higher DNA Shannon information, lower DNA storage cost and code rate almost $0.5$. We describe them now briefly.
\subsection{DNA Information Density}
Amount of data that can be encoded in DNA can be quantified by DNA information density. \begin{definition} For a given DNA based information storage system, DNA information density is total amount of data that can be stored in unit gram of DNA.\end{definition} At theoretical maximum, one gram of single stranded genetic code can store $455$ EB (exabytes) of information \cite{church2012next}. Goldman achieved information density $2.2$ PB (petabytes) per gram of DNA. Using DNA Golay codes, computationally we achieved information density for DNA based storage medium as $1.15 \times 10^{20} = 115$ EB (Exabytes) per gram DNA.
\begin{Proposition}
DNA storage capacity using our chunk architecture for one gram of DNA is calculated by solving the following non-linear equation
$$
\begin{array}{c}
[(182 \times 10^{19} \times l) \div \left( \left( l+3\right) + log_3 \left( \frac{N\left( x+22\right)}{l} \right) \times N \right)] -22\\
= x,
\end{array}
$$ where x= number of bytes per one gram of DNA, l = Length of chunk without chunk index, N = Length of the error correcting code.
\end{Proposition}
In case of DNA Golay codes, DNA storage capacity obtained is $1.15 \times 10^{20}$ (115 Exabytes) bytes per gram of DNA by using $l = 99$ and $N =11$.
\subsection{DNA Storage Cost}
Cost of DNA synthesis and sequencing is a major limitation of DNA storage medium. Considering synthesis cost to be $\char36 0.05$ per base \cite{goldman2013towards}, total cost per MB for different file size (bytes) was plotted by us. Using DNA Golay code, required amount of the DNA decreased which resulted in decreased cost. We also plotted the cost to store data using various non-linear ternary codes of our family. We observe that with the increase in the amount of data, there is negligible increase in cost per unit data (1 MB) associated with DNA based storage using non linear family of codes developed by us. In contrary, cost for the DNA increases with the increase in data using Goldman method.
\subsection{Tradeoff for Code Rate }
The information rate of our encoding scheme is 0.73 bits per base by encoding 8 bits into 11 DNA bases per byte (i.e. 8/11 = 0.73 bits per base). It can be improved further to 0.89 by using 9 base per byte (i.e. 8/9 = 0.89 bits per base) by encoding each byte using code $(9, 256, 3)_{3}$, at the cost of reduction in error detection capacity to 2 trits and correction capacity to 1 trits per codeword.
It can be observed that error correction capacity t for each code increases linearly with increase in length of the code n. But with increase in these two parameters, there is declination in code rate. For reliable data storage, code rate is optimal with $(11, 256, 5)_{3}$ code, if one choose code with $(9, 256, 3)_{3}$, there is decrease in code rate. Hence there is open challenge for the construction of codes with optimal code rate for DNA based archival storage medium. Thus one can conclude that code of family $(11, 256, 5)_{3}$ is optimal code at present.
\begin{table}[h]
\caption{Codewords from subcode of Ternary Golay Code i.e. $[11,6,5]_{3}$ assigned to 256 ASCII values is given in the table.}
\begin{center}
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|}
\hline
ASCII &Golay codes & Wei- &ASCII & Golay codes& Wei-& ASCII & Golay codes&Wei-&ASCII &Golay codes & Wei-\\
Values& & ght &values& & ght & values& & ght &values& &ght\\ \hline\hline
86 & 00002111202 & 6&170 & 00001222101 &6&127& 00020220222 &6& 253&00022001121&6 \\ \hline
52 &00021112020 & 6&138 & 00010110111&6&41& 00012221010 &6&86&00011002212&6\\ \hline
42&00201010122 &6& 100 & 00200121021 &6&44&00202202220 &6& 250&00221200011&6\\ \hline
132 &00220011210 &6& 161 & 00222122112&9&98& 00211120200&6 & 8&00210201102&6\\ \hline
34 &00212012001&6& 10 & 00102020211 &6&149&00101101110&6 &87&00100212012&6 \\ \hline
21 &00122210100 &6& 74 & 00121021002 &6&36& 00120102201 &6& 69&00112100022&6\\ \hline
177 &00111211221 &9&20 & 00110022120&6&213& 02012212122&9&163&02011020021&6\\ \hline
229 &02010101220 &6& 255 & 02002102011 &6&197& 02001210210 &6& 133&02000021112&6\\ \hline
252 &02022022200 &6& 26 & 02021100102 &6&173&02020211001 &6&151&02210222211&9\\ \hline
82 &02212000110 &6& 75 & 02211111012&9 &37& 02200112100 &6&166&02202220002&6\\ \hline
191 &02201001201 &6&88& 02220002022 &6& 63& 02222110221&9& 68&02221221120 &9\\ \hline
150 &02111202000 &6&76 & 02110010202&5&4& 02112121101 &9& 154&02101122222 &9\\ \hline
234&02100200121 &6&22 & 02102011020&6 &162&02121012111&9& 105&02120120010&6\\ \hline
102 &02122201212&9 &171& 01021121211 &9&104& 01020202110&6 &169&01022010012&6\\ \hline
196& 01011011100 &6&208&01010122002&6& 84&01012200201&6&130&01001201022&6\\ \hline
146 &01000012221 &6& 72&01002120120&6 &16& 01222101000&6 & 66&01221212202&9\\ \hline
24 &01220020101 &6&106& 01212021222 &9& 223& 01211102121&9 & 58&01210210020&6\\ \hline
137 &01202211111 &9&73& 01201022010&6 & 101& 01200100212&6 & 168&01120111122&9\\ \hline
181&01122222021 &9&175&01121000220 &6& 251& 01110001011&6& 40&01112112210&9\\ \hline
140 &01111220112 &9& 17& 01100221200&6 & 83& 01102002102 &6&254 &01101110001&6\\ \hline
240&20121202122 &9&214& 20120010021&6 & 53& 20122121220 &9&202&20111122011 &9\\ \hline
25 &20110200210 &6&18&20112011112&9 & 247& 20101012200 &6& 174&20100120102&6\\ \hline
112 &20102201001 &6&89&20022212211&9 &210& 20021020110 &6& 217&20020101012&6\\ \hline
248 &20012102100 &6& 194&20011210002 &6&182& 20010021201&6&80&20002022022&6\\ \hline
79 &20001100221 &6& 195 &20000211120&6 &12& 20220222000 &6& 209&20222000202&6\\ \hline
165 &20221111101 &9&245&20210112222&9 &2& 20212220121 &9& 81&20211001020&6\\ \hline
38 &20200002111 &6& 141&20202110010 &6&211& 20201221212&9 & 239&22100111211&9\\ \hline
95 &22102222110 &9&43& 22101000012 &6&224&22120001100&6&203&22122112002&9\\ \hline
145 &22121220201&9 & 147&22110221022 &9&19& 22112002221 &9& 50&22111110120&9\\ \hline
136 &22001121000 &6& 107& 22000202202&6& 134& 22002010101&6 &109 &22021011222&9\\ \hline
153 &22020122121& 9&148& 22022200020 &6& 205& 22011201111 &9&212 &22010012010&6\\ \hline
54 &22012120212 &9& 241&22202101122&9& 156&22201212021&9 &115 &22200020220&6\\ \hline
116& 22222021011& 9& 78&22221102210 &9 & 67 & 22220210112 & 9 & 70 & 22212211200 &9 \\ \hline
178 &22211022102 &9& 159& 22210100001&6 & 142& 21112020000 &6&92 &21111101202&9\\ \hline
48&21110212101 &9& 90& 21102210222 &9& 218& 21101021121 &9&126 &21100102020&6\\ \hline
39 &21122100111 &9&219& 21121211010&9& 167&21120022212 &9&114 &21010000122&6\\ \hline
172&21012111021 &9& 14& 21011222220 &9& 120& 21000220011&6 &139 &21002001210&6\\ \hline
160&21001112112 &9& 33&21020110200 &6& 179& 21022221102&9&117&21021002001&6\\ \hline
225 &21211010211 &9&129&21210121110 &9& 183&21212202012 &9& 230&21201200100&6\\ \hline
35 &21200011002 &6&93&21202122201 &9& 6& 21221120022&9 &32&21220201221&9\\ \hline
56&21222012120 &9&158&10212101211 &9& 185& 10211212110&9 &47&10210020012&6\\ \hline
143 &10202021100 &6&123&10201102002 &6& 204& 10200210201 &6& 242&10222211022&9\\ \hline
111 &10221022221 &9&103&10220100120&6& 108& 10110111000&6& 9&10112222202&9\\ \hline
65 &10111000101 &6&249&10100001222 &6& 13&10102112121 &9& 180&10101220020&6\\ \hline
226&10120221111 &9&144&10122002010&6 & 15&10121110212 &9& 57&10011121122&9\\ \hline
128 &10010202021 &6&135&10012010220 &6& 243&10001011011&6 & 190&10000122210&6\\ \hline
207 &10002200112 &6&77&10021201200 &6& 45&10020012102 &6& 91&10022120001&6\\ \hline
192 &12221010000 &6&186&12220121202 &9& 216&12222202101&9 & 97&12211200222&9\\ \hline
118 &12210011121 &9&246&12212122020&9& 215&12201120111 &9& 51&12200201010&6\\ \hline
206 &12202012212 &9&184&12122020122 &9& 227&12121101021 &9& 233&12120212220&9\\ \hline
237 &12112210011&9 &188&12111021210 &9& 113&12110102112&9 & 49&12102100200&6\\ \hline
201 &12101211102&9& 155& 12100022001&6 & 222&12020000211&6&231 &12022111110&9\\ \hline
5 &12021222012 &9& 27& 12010220100 &6& 131& 12012001002 &6&164 &12011112201&9\\ \hline
3 &12000110022&6& 46& 12002221221&9 & 119& 12001002120&6 &28 &11200222122&9\\ \hline
176 &11202000021 &6 &23& 11201111220&9 &64&11220112011&9 &157 &11222220210&9\\ \hline
187&11221001112& 9&244& 11210002200&6 & 238& 11212110102 &9&96 &11211221001&9\\ \hline
235&11101202211 &9& 60&11100010110&6& 1&11102121012 &9&110 &11121122100&9\\ \hline
200 &11120200002&6& 221& 11122011201&9& 99&11111012022 &9&31 &11110120221&9\\ \hline
198 &11112201120 &9& 193&11002212000& 6&125&11001020202&6 &124 &11000101101&6\\ \hline
152 &11022102222& 9&122&11021210121&9& 71&11020021020&6 &94 &11012022111&9\\ \hline
220 &11011100010 &6& 29&11010211212&9& 199&00000201211 &5&61 &00000102122&5\\ \hline
11 &00002012110 &5&228&00002210021&5 & 62&00001021220 &5& 55&00001120012&5\\ \hline
121 &00020121100&5&7&00020022011 &5&30&00022100210 &5& 232&00022202002&5\\ \hline
189 &00021010201 &5&59&00010212200 &5& 236&00010011022&5 & 0&00000000000&0\\ \hline
\end{tabular}
\end{center}
\label{DNA golay code}
\end{table}
\section{Conclusions}
In this paper, we introduced non linear family of DNA codes with different parameters to develop efficient DNA based archival information storage system. Using non linear ternary Golay code (DNA Golay Code), we are able to correct 2 flip errors per block of 11 nucleotides along with significant improvement in storage capacity. Using this approach, we obtained significant difference in data storage density and length of DNA required when compared to Goldman's et al approach. We also developed a software DNA cloud 2.0 which implements our encoding schemes available at http://www.guptalab.org/dnacloud/. We hope our codes generate interest in coding theory community to construct better coding schemes for this upcoming area.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\begin{figure*}[!tb]
\begin{center}
\includegraphics[width=0.56\textwidth]{error_tejasOnly_pruned}
\caption {Comparison : {\em SPEC2006} suite \label{fig:serial_validation} }
\end{center}
\end{figure*}
\begin{figure*}[!tb]
\begin{center}
\includegraphics[width=0.47\textwidth]{tejas_valid_parallel}
\caption {Comparison : {\em SPLASH-2} suite \label{fig:parallel_validation} }
\end{center}
\end{figure*}
This article serves to establish Tejas ~\cite{partejas} as a validated micro-architectural simulator.
To do so, a range of serial and parallel benchmarks were run on a PowerEdge R620 server
(for details see Table~\ref{tab:hardware_details}). The linux ``perf'' command was
used to measure the number of cycles taken to execute the benchmarks.
The same set of benchmarks were then run
on the Tejas simulator, configured to mimic the DELL server as closely as possible
(see Table~\ref{tab:tejas_configuration} for details).
The comparison of the simulated cycle counts with their hardware counterparts is given in the next section.
\begin{table}[!htb]
\centering
\scriptsize
\begin{tabular}{|c|c|c|c|}
\hline
{\bf Parameter} & {\bf Value} & {\bf Parameter} & {\bf Value} \\
\hline
Microarchitecture & Intel Sandybridge & Number of cores & 12 \\
Main Memory & 32 GB & Memory Type & ECC DDR3 \\
L1 i-cache and d-cache & 32 KB & L2 cache & 256 KB \\
L3 Cache & 15 MB & Frequency & 2GHz \\
Hyper-threading & No & DVFS & Disabled \\
Load Buffer Size & 64 & Store Buffer Size & 64 \\
Reorder Buffer & 168 micro-ops &&\\
\cline{2-4}
Operating System & \multicolumn{3}{c|}{ Ubuntu 12.10 Linux 3.5.0-36-generic, 64-bit} \\
\hline
\end{tabular}
\caption{Details of the Reference Hardware \label{tab:hardware_details}}
\end{table}
\section{Results}
Figure~\ref{fig:serial_validation} shows the results for a set of 17 benchmarks from the
SPEC2006 suite. We compute the time it takes for a benchmark to complete on native hardware
(averaged across 10 runs). Then, we simulate the benchmark on Tejas, and compute the absolute error.
The average absolute error is 11.45\%. 10 out of 17 benchmarks have an error less than 10\%.
Only 4 benchmarks have errors in the 20-30\% range ({\em sjeng}, {\em astar}, {\em mcf},
and {\em gcc}).
Figure~\ref{fig:parallel_validation} shows the results for a set of 11 benchmarks from the
{\em SPLASH-2} suite. The average absolute error was observed to be 18.77\%. It has been computed
the same way as was done for the case of sequential benchmarks. In this case, the average
error is more primarily because the jitter introduced by the operating system is not
predictable, there are hardware events that induce jitter, and lastly we are not privy
to all details of the operation of
the cache coherence protocols in Intel systems. Only 3 benchmarks had errors more
than 25\% namely {\em radiosity}, {\em radix} and {\em water-spatial}. For most
of the benchmarks, the error ranges from 10 to 17\%.
\begin{table}
\scriptsize
\centering
\begin{tabular}{|c|c||c|c|}
\hline
{\bf Parameter} & {\bf Value} & {\bf Parameter} & {\bf Value} \\
\hline
\multicolumn{4}{|c|}{{\bf Pipeline}} \\
\hline
Retire Width & 4 & Integer RF (phy) & 160 \\
Issue Width & 6 & Float RF (phy) & 144 \\
ROB size & 168 & Predictor & TAGE~\cite{tage} \\
IW size & 54 & Bmispred penalty & 8 cycles \\
LSQ size & 64 & & \\
iTLB & 128 entries & dTLB & 128 entries \\
Integer ALU & 3 units & lat = 1 cycle & RoT = 1 \\
Integer Mul & 1 unit & lat = 3 cycles & RoT = 1 \\
Integer Div & 1 unit & lat = 21 cycles & RoT = 12\\
Float ALU & 1 units & lat = 3 cycles & RoT = 1\\
Float Mul & 1 unit & lat = 5 cycles & RoT = 1 \\
Float Div & 1 unit & lat = 24 cycles & RoT = 12 \\
\hline
\multicolumn{4}{|c|}{{\em RoT : reciprocal of throughput}} \\
\hline
\hline
\multicolumn{4}{|c|}{{\bf Private L1 i-cache, d-cache}} \\
\hline
Write-mode & Write-Through & Block size & 64 \\
Associativity & 8 & Size & 32 kB \\
Latency & 3 cycles & & \\
\hline
\hline
\multicolumn{4}{|c|}{{\bf Private Unified L2 cache}} \\
\hline
Write-mode & Write-Back & Block size & 64 \\
Associativity & 8 & Size & 256 kB \\
Latency & 6 cycles & & \\
\hline
\multicolumn{4}{|c|}{{\bf Shared L3 cache}} \\
\hline
Write-mode & Write-back & Block size & 64 \\
Associativity & 8 & Size & 15 MB \\
Latency & 29 cycles & & \\
\hline
\hline
\multicolumn{2}{|c||}{{\bf Main Memory Latency}} & \multicolumn{2}{c|}{{200 cycles}} \\
\hline
\hline
\multicolumn{4}{|c|}{{\bf NOC and Traffic}} \\
\hline
Topology & Bus & Latency & 1 cycle \\
Flit size & 32 bytes & & \\
\hline
\end{tabular}
\caption{Simulation parameters \label{tab:tejas_configuration}}
\end{table}
\section{Comparison with other Simulators}
Let us now put our numbers in the right perspective by comparing similar numbers
obtained on other simulators. We shall observe that Tejas is more accurate on
both serial and parallel benchmarks as compared to most of the other widely
used architecture simulators (for which published results are available).
MARSS~\cite{MARSS} is a cycle-accurate simulator based on PTLSim.
It is a tool built on QEMU and provides fast and full system simulation.
MARSS has been validated against a x86 target machine
with the Intel Xeon E5620 processor. For the {\em SPEC CPU 2006} benchmark suite,
it has errors ranging from -59.2\% to 50\%, with an average absolute error of 23.46\%.
Sniper~\cite{sniper} is an approximate simulator.
It tries to find a middle ground between the simulators that are fast but inaccurate
and the simulators that are accurate but slow.
It has been validated against a 4-socket Intel
Xeon X7460 Dunnington shared-memory machine.
An average absolute error of 25\% has been reported for the {\em SPLASH-2} benchmark suite (we report
18.77\%).
FastMP~\cite{FastMP}, is also an approximate simulator. It simulates a subset of cores in
detail. Subsequently, it performs a real-time analysis of the behavior of the cores that have been
simulated in detail and uses this data to approximate
the behavior of the other cores. FastMP has been validated against a real x86 machine using the {\em SPEC 2006}
benchmark suite. It suffers from an average error of 9.56\% but for some of the benchmarks the error is
as high as 40\% (our maximum error is roughly 27\%).
To the best of our knowledge other popular simulators such as SESC~\cite{SESC} and MacSim ~\cite{MacSim} have not been validated.
Multi2sim, which is a heterogeneous simulator, has been validated for the GPU framework but
its CPU simulation framework has not been validated and published (to the best of our knowledge).
\section{Conclusion}
Tejas has been demonstrated to be a reliable simulator that provides a highly accurate
reflection of the simulated hardware. It must be noted that many details of the underlying
hardware are not known -- the branch predictor, the coherence protocol, the
NOC parameters, the select/data forwarding logic, to name a few. Better knowledge of these
will allow further reduction of the error in the accuracy.
The incomplete knowledge aside, Tejas shows an average absolute error of 11.45\% in serial benchmarks,
and 18.77\% in parallel benchmarks, which is 5-10 percentage points better than some of the most
popular architecture simulators currently in use (as of 2014).
\bibliographystyle{plain}
|
\section{Introduction}\label{s1}
\subsection{}\label{s1s1}
In the present work we introduce and analyze some numerical results
obtained by experimenting. These experiments are connected with the well known
problem about the distribution of the zeros of Hermite--Pad\'{e} polynomials
for a collection of three functions $[f_0\equiv1,f_1,f_2]$, defined and holomorphic
at the infinity point $z=\infty$, $f_1,f_2\in\HH(\infty)$.
Our numerical experiments are restricted only to the Hermite--Pad\'{e} polynomials
of first kind. It is well known that Hermite--Pad\'{e} polynomials\footnote{Sometimes
these polynomials are called ``multiple orthogonal polynomials'',
see \cite{AsFi13}, \cite{Ass11}.}
of first and second kind
are closely related to each other, see \cite[\S 2, equation (2.1.9)]{Nut84}, \cite{Ass06},
\cite{AsFi13}, \cite{Ass11}. The numerical results obtained here
may permit interpretations about the Hermite--Pad\'{e} polynomials of the second kind.
Let
$\PP_n:=\CC_n[z]$, $n\in\NN$,
be the class of all polynomials with complex coefficients of degree $\leq{n}$.
For an arbitrary $n\in\NN$, define the Hermite--Pad\'{e} polynomials of first kind
$Q_{n,j}\in\PP_n$, $j=0,1,2$, $Q_{n,1},Q_{n,2}\not\equiv0$,
by the relation (see \cite{Nut84},
also \cite{GoRa81}, \cite{Sta88}, \cite{Gon03}, \cite{Apt08}, \cite{ApKu11},
\cite{Apt12}, \cite{AsFi13})
\begin{equation}
(Q_{n,0}\cdot1+Q_{n,1}f_1+Q_{n,2}f_2)(z)=O\(\frac1{z^{2n+2}}\),
\qquad z\to\infty.
\label{1}
\end{equation}
In the present work, we will use the terminology and notation
of the work by J. Nuttall \cite{Nut84} (see also H. Stahl \cite{Sta88}).
According to their paper, the classical Pad\'{e} polynomials $P_{n,0},P_{n,1}$ of the function
$f\in\HH(\infty)$ are, in fact, the Hermite--Pad\'{e} polynomials for the collection of two functions
$[f_0\equiv1,f_1=f]$, such that $P_{n,0},P_{n,1}\in\PP_n$, $P_{n,1}\not\equiv0$ and:
\begin{equation}
(P_{n,0}\cdot1+P_{n,1}f)(z)=O\(\frac1{z^{n+1}}\),
\qquad z\to\infty.
\label{2}
\end{equation}
We present the results from our numerical experiments about two cases.
In the first case the functions $f_1$ and $f_2$ have the following form:
\begin{equation}
f_1(z)=\frac{z}{\sqrt{(z-a_1)(z-b_1)}},\qquad
f_2(z)=\frac{z}{\sqrt{(z-a_2)(z-b_2)}},
\label{3}
\end{equation}
where $a_1,b_1,a_2,b_2\in\CC$, $a_1\neq b_1$, $a_2\neq b_2$, and
$\{a_1,b_1\}\cap\{a_2,b_2\}=\varnothing$. Therefore, the pair of functions
$f_1,f_2$ forms an {\it Angelesco system}
(see \cite{Kal79}, \cite{GoRa81}, \cite{Sta88}, \cite{GoRaSo97}, \cite{Gon03},
\cite{Apt08}).
In the second case, $f_1=f$, $f_2=f^2$, where
\begin{equation}
f(z)=(z^2-1)^{1/4}(z-a)^{-1/2},\quad f(\infty)=1,\quad a\not\in\RR,
\label{4}
\end{equation}
and the pair of functions $f_1,f_2$ forms a (generalized) {\it Nikishin system}
(see \cite{Nik80}, \cite{GoRaSo97},
\cite{Gon03}, \cite{Apt08}, \cite{ApLy10}, \cite{Rak11},
\cite{FiLo11}, \cite{DeLoLo12}).
\begin{remark}\label{rem1}
In the theory of Hermite--Pad\'{e} polynomials for a collection of three
functions $[1,f_1,f_2]$
two opposite situations are usually selected, which are connected
with the distribution of the branch points of the functions $f_1$ and $f_2$.
In the first case the sets of branch points $\mathcal{A}_1=\mathcal{A}(f_1)$ and
$\mathcal{A}_2=\mathcal{A}(f_2)$ {\it do not intersect each other}.
We say that the pair of functions forms an {\it Angelesco system},
see \cite{Gon03}, \cite{Apt08}, \cite{ApLy10}.
In the second case the sets of branch points $\mathcal{A}_1$ and $\mathcal{A}_2$
of the functions $f_1$ and $f_2$ {\it are equivalent}. We say that two functions $f_1,f_2$
forms a {\it Nikishin system}, see \cite{Nik80}, \cite{Gon03}, \cite{Apt08}, \cite{ApLy10}.
In the first case the interaction matrix $M$ for the theory-potential-equilibrium vector problem
is $M=\begin{pmatrix} 2&1\\1&2\end{pmatrix}$. In the second case it is
$M=\begin{pmatrix} 2&-1\\-1&2\end{pmatrix}$,
see \cite{GoRa85}, \cite{Gon03}, \cite{Apt08}, \cite{Rak11}.
The natural expectation is that
such differences between the distribution of the branch points
and the structure of the interaction matrices leads to essentially different
limiting distributions of the zeros of the corresponding Hermite--Pad\'{e} polynomials.
However, this is not always the case. Figures
\ref{Fig_ang(2_2)_4000_120_full}, \ref{Fig_nik(2_1)_4000_120_full},
\ref{Fig_ang(2_2)_4000_120_blbk} and \ref{Fig_nik(2_1)_4000_120_blrd}
show the distribution of the zeros of the polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points), $Q_{120,2}$ (black points) for two pair of functions:
one with different branch points
\begin{equation}
\label{ang1}
f_1(z)=\sqrt{(z-1)/(z+1)},\qquad f_2(z)=\sqrt{(z-2)/(z+2)}
\end{equation}
and the other with coincident branch points
\begin{equation}
\label{nik1}
f_1(z)=\(\frac{z-1}{z+1}\)^{1/3}\(\frac{z-2}{z+2}\)^{1/3},\quad
f_2(z)=\(\frac{z-1}{z+1}\)^{2/3}\(\frac{z-2}{z+2}\)^{1/3}.
\end{equation}
From the results on figures
\ref{Fig_ang(2_2)_4000_120_full}--\ref{Fig_nik(2_1)_4000_120_blrd}
it follows that the distribution of the zeros of Hermite--Pad\'{e} polynomials
for these two different systems of functions
will be equivalent (with regard to the fact that supports of the limit measures
for the polynomials $Q_{n,j}$ change over to each other). Note that
the system \eqref{nik1} is a special case of a two Markov functions system,
which was considered by E.~A.~Rakhmanov in \cite{Rak11}. The numerical results
for the system \eqref{nik1} obtained here are in a good agreement with
the results of E.~A.~Rakhmanov~\cite{Rak11}.
\end{remark}
\begin{remark}\label{rem2}
Instead of the pair of functions \eqref{ang1} with second order branch points
we can consider a pair of functions with arbitrary order of the branch points.
For example, the pair of functions
\begin{equation}
\label{ang2}
f_1(z)=\sqrt[3]{(z-1)/(z+1)},\qquad f_2(z)=\sqrt[3]{(z-2)/(z+2)}
\end{equation}
have the same distribution of the zeros of the corresponding Hermite--Pad\'{e} polynomials
as \eqref{ang1};
see fig. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}
and comp. fig. \ref{Fig_ang_im(_4)_3_3_4000_120_full}--\ref{Fig_ang_im(_4)_3_3_4000_120_rd}.
However, the pair of functions
\begin{equation}
\label{ang3}
f_1(z)=\sqrt{(z-1)/(z+1)},\qquad f_2(z)=\sqrt[3]{(z-2)/(z+2)}
\end{equation}
have another distribution of the zeros of the corresponding Hermite--Pad\'{e} polynomials,
see fig. \ref{Fig_ang(2_3)_4000_120_full}--\ref{Fig_ang(2_3)_4000_120_rd}
and comp. fig. \ref{Fig_ang_im(_4)_2_3_4000_120_full}--\ref{Fig_ang_im(_4)_2_3_4000_120_rd}.
Thus, the distribution depends non only of the type but also on
the degree of the branch points
(in \eqref{ang3} for the first function $f_1$ instead of degree $1/3$,
as in \eqref{ang2}, we took degree $1/2$).
The distribution of the zeros remains stable while moving the branch points
of the function $f_1$ along the imaginary axis, for example,
when we change the points $z=\pm1$ into the points $z=\pm1+i\cdot 0.4$:
\begin{align}
\label{ang4}
f_1(z)=\sqrt[3]{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))},\\
f_2(z)=\sqrt[3]{(z-2)/(z+2)},\nonumber
\end{align}
see fig. \ref{Fig_ang_im(_4)_3_3_4000_120_full}--\ref{Fig_ang_im(_4)_3_3_4000_120_rd}
and comp. \eqref{ang2}, and:
\begin{align}
\label{ang5}
f_1(z)=\sqrt{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))},\\
f_2(z)=\sqrt[3]{(z-2)/(z+2)},\nonumber
\end{align}
see fig. \ref{Fig_ang_im(_4)_2_3_4000_120_full}--\ref{Fig_ang_im(_4)_2_3_4000_120_rd}
and comp. \eqref{ang3}.
This confirms that the distribution of the zeros of the Hermite--Pad\'{e} polynomials
depends not only on the geometrical position of the branch points,
but also on the type of the branch points.
\end{remark}
\begin{remark}\label{rem3}
If instead of \eqref{nik1} we consider the pair of functions
\begin{equation}
\label{nik2}
f_1(z)=\(\frac{z-1}{z+1}\)^{1/3}\(\frac{z-2}{z+2}\)^{1/3},\quad
f_2(z)=\(\frac{z-1}{z+1}\)^{2/3}\(\frac{z-2}{z+2}\)^{-1/3},
\end{equation}
then the distribution of the zeros of the corresponding Hermite--Pad\'{e} polynomials
will be different than before,
see fig. \ref{Fig_nik(2_-1)_4000_120_full}--\ref{Fig_nik(2_-1)_4000_120_blrd}.
\end{remark}
\subsection{}\label{s1s2}
A numerous conjectures about the asymptotic behavior of the zeros of
the Hermite--Pad\'{e} polynomials of both first and second kind was, in
part, formulated in the fundamental work of J. Nuttall \cite{Nut84}
(see also \cite{Nut81}, \cite{Nut82}). These conjectures served as a
background for some of the numerous investigations. Among them the
prominent results of H.~Stahl~\cite{Sta85a}--\cite{Sta86b} and
A.~A.~Gonchar--E.~A.~Rakhmanov~\cite{GoRa81},~\cite{GoRa87}, first of
all should be noted, and also the results of A.~I.~Aptekarev with
co-authors \cite{ApKa86}, \cite{ApKuVa08}, \cite{Apt08}, \cite{ApLy10},
\cite{ApLyTu11}, \cite{ApYa11}. J. Nuttall's exact results
\cite{Nut81}, \cite{Nut82}, \cite{Nut84} about the asymptotics of the
Hermite--Pad\'{e} polynomials for a collection of $m$ functions
$[f_0\equiv1,f_1,\dots,f_{m-1}]$, $m\ge3$, are based mainly on the a
priori assumption of the existence of an associated $m$-sheeted Riemann
surface $\RS_m$ with a canonical decomposition into $m$ sheets
$\RS^{(j)}$, $j=1,\dots,m$. The sheets defined by an Abel integral of
third kind with purely imaginary periods and logarithmic singularities
of the form $(m-1)\log{z}$, $z=z^{(1)}\sim\infty^{(1)}$ and $-\log{z}$,
$z=z^{(j)}\sim\infty^{(j)}$, $j=2,\dots,m$, and having also property
that the first ``physical'' sheet $\RS^{(1)}$ is always connected (see
\cite{Nut81}, \cite{Nut84}, \cite{NuTr87}).\footnote{ In some cases,
considered by Nuttall, it appears that the multi-valued analytic
functions $f_1,\dots,f_{m-1}$ continue from the first sheet onto
$\RS_m\setminus\RS^{(m)}$ as single-valued meromorphic functions and
$[1,f_1,\dots,f_{m-1}]$ are in a sense independent.} In such case it is
possible to describe the asymptotical behavior of the Hermite--Pad\'{e}
polynomials in terms connected to this RS (see \cite{Nut84},
\cite{Nut86}, \cite{ApKuVa08}, \cite{Apt08}). However, the question how
to ``construct'' such a RS in the general case when $m\geq3$ remains
open.\footnote{ For $m=2$ the existence of a corresponding
hyperelliptic Riemann surface with two sheets follows directly from the
theorems by Stahl; see \cite{Nut84}, \cite{Sta97b}, \cite{ApYa11},
\cite{KovSu14}.} As for the present, results are achieved only for the
case $m=3$. There are two major methods to obtain the results of such
type. The first one mainly based on the {\it cubic equation} (see
\cite{ApKa86}, \cite{ApKuVa08}, \cite{ApTu12b}, also \cite{Nut86},
\cite{MaRaSu12}, \cite{MaRaSu13}). Using this method one can find an
explicit representation of the so-called Nuttall's
$\phi$-function\footnote{In the case of two sheeted hyperelliptic RS
such function is known as Deift's $g$-function; see~\cite{RaSu13}.}
(see \cite{ApKuVa08}, \cite{ApTu12b}, \cite{ApTu14}), i.e., an Abel
integral of third kind, which has the level curves that define the
needed canonical decomposition of RS $\RS_3$ into three sheets . The
other method consists of first solving the theoretical-potential
extremal problem, connected with the existence of the so-called {\it
Nuttall condenser} on the Riemann sphere $\myo\CC$ (see \cite{RaSu12},
\cite{RaSu13}, \cite{KovSu14}, \cite{Sue13b}). If such a condenser can
be found, then the RS $\RS_3$ with the needed decomposition into three
sheets is ``constructed'' on the base of this condenser by a specific
scheme (see \cite{RaSu13}, \cite{KovSu14}). Note that for now the
second method can be applied only when the pair $f_1,f_2$ forms a
Nikishin system \cite{RaSu13}, \cite{Sue13b}. Namely, for some general
enough functions $f_1,f_2$, which forms a {\it complex} Nikishin
system,\footnote{Under (general) Nikishin system, which contains two
functions $f_1,f_2$, we understand such a system for which the
corresponding vector theoretical-potential equilibrium problem is
defined by a Nikishin matrix; see \cite{Apt08}, \cite{ApLy10},
\cite{Lap12}.} in \cite{RaSu13} (see also \cite{RaSu12},
\cite{KovSu14}) an {\it existence} of the Nuttall condenser, consisting
of two non-intersecting plates and having certain symmetrical
properties, is proved. Such a condenser is an analogue of Stahl
compact, but for the case of Hermite--Pad\'{e} polynomials when $m=3$.
In \cite{KovSu14} a scheme for constructing a RS with three sheets and
with canonical decomposition made on the base of an already existing
Nuttall's condenser is proposed.
As mentioned before, the numerical results are obtained here for two opposite
cases:
for a pair of functions $f_1,f_2$ creating an Angelesco system (see \eqref{3})
and for a pair of functions $f_1=f,f_2=f^2$ creating a (generalized) Nikishin system (see \eqref{4}).
These new numerical results give rise to some new conjectures about
the asymptotical properties of the Hermite--Pad\'{e} polynomials of first kind.
It is well known that the Hermite--Pad\'{e} polynomials of first and second kind
are closely related, in particular, they are bi-orthogonal;
see \cite[\S 2, formula (2.1.9)]{Nut84},
also the recent works \cite{Ass01}, \cite{AsFi13}, \cite{Ass11}.
It seems that the conjectures presented here can be applied also to
the Hermite--Pad\'{e} polynomials of second kind.
For a better exposition of the new numerical results, we state
at the beginning some well known facts about the asymptotical
behavior of the zeros and poles of the classical Pad\'{e} approximants,
i.e. zeros of the Pad\'{e} polynomials (see \eqref{2}), and also for
two-point Pad\'{e} approximants \cite{Bus13}.
These results and their pictures (see
fig. \ref{Fig_pade10_2500_131}--\ref{Fig_pade103_5000_266_full} and
fig. \ref{Fig_bus210a_3000_90_full}--\ref{Fig_bus205c_4000_120_full})
are needed for the analysis of the numerical results,
connected with the behavior of the zeros of the Hermite--Pad\'{e} polynomials.
\section{Main results}\label{s2}
The main empirical results obtained in the paper are presented herein.
\subsection{Angelesco system}\label{s2s1}
The first part of the numerical results of the behavior of the
Hermite--Pad\'{e} polynomials for a collection of three functions
$[1,f_1,f_2]$ refers to the case when each of the functions $f_1$ and
$f_2$ has a pair of branch points of second order and the sets of the
branch points of the functions $f_1$ and $f_2$ do not intersect. Thus,
the pair of functions $f_1,f_2$ forms an Angelesco system (see
\eqref{3}). About the main properties of the Angelesco system, see
\cite{Nut84}, also \cite{GoRa81}. In \cite{GoRa81}, the first results
of general character about the convergence of the Hermite--Pad\'{e}
approximants of second kind for the collection of functions
$[1,f_1,\dots,f_{m-1}]$, $m\ge3$ were obtained, with the functions
$f_1,\dots,f_{m-1}$ creating an Angelesco system and
$f_j=\myh{\sigma}_j$ being Markov functions with non-intersecting
supports $\Delta_j=\supp{\sigma_j}$, $j=1,2,\dots,m-1$, along the real
axis, and containing a finite number of intervals. In \cite{GoRa81}, A.
A. Gonchar and E.~A.~Rakhmanov found a new property, which they called
{\it pushing} of the support of the equilibrium measures (see also
\cite{Nut84}, \cite{ApKa86}, \cite{Apt08}). For $m=3$ this method looks
as follows. Let the support sets $\Delta_1,\Delta_2$ of the measures
$\sigma_1,\sigma_2$ are the non-intersecting closed intervals
$\Delta_1=[a_1,b_1]$, $\Delta_2=[a_2,b_2]$ of different lengths and
contained in the real axis; for definiteness we suppose that
$|\Delta_1|<|\Delta_2|$ and that $\Delta_2$ be on the right of
$\Delta_1$. It turns out that in the case when the intervals are close
enough, the support $F_2\subset\Delta_2$ of the equilibrium measure
$\lambda_2$ is the interval $[a^*_2,b_2]$, where $a^*_2\in(a_2,b_2)$
(see \eqref{eq.kal84} and fig. \ref{Fig_mar_1_2000_90_rdbk}). Thus, the
equilibrium measures $\lambda_1,\lambda_2$ are absolutely continuous
with respect to the normalized linear Lebesgue measure, the density
$\lambda_1'(x)$ behaves, in the neighborhood of the points $a_1$ and
$b_1$, like $(x-a_1)^{-1/2}$ and $(b_1-x)^{-1/2}$, respectively, and
the density $\lambda_2'(x)$ behaves, in the neighborhood of the point
$a_2$, like $(x-a^*_1)^{1/2}$ and like $(b_2-x)^{-1/2}$ around $b_2$.
Thus, under a specific mutual positions the smaller interval pushes the
support of the equilibrium measure inside the larger interval; this
does not happen with the support of the equilibrium measure for a
smaller interval (see fig. \ref{Fig_mar_1_2000_90_rdbk}). When
$[a_1,b_1]=[-a,0]$, $[a_2,b_2]=[0,1]$, where $a\in(0,1)$, the point
$a_2^*$ is calculated by the following formula, found by V.~A.~Kalyagin
\cite{Kal84} (see also \cite[p. 5.3, formula (5.3.18)]{Nut84},
\cite{ApKa86}):
\begin{equation}
\label{eq.kal84}
a_2^*=\frac{(1-a)^3}{9(a^2-a+1)}.
\end{equation}
The authors have found, by experiments, a new property called {\it mutual
pushing} of the supports of the equilibrium measures, in the case when the functions $f_1$ and $f_2$
have a pair of branch points $a_1,b_1\in\CC$ and $a_2,b_2\in\CC$,
$\{a_1,b_1\}\cap\{a_2,b_2\}=\varnothing$, located not on one line,
but on two parallel lines, and the intervals
$\Delta_1=[a_1,b_1]$ and $\Delta_2=[a_2,b_2]$ have different lengths, i.e.
$|\Delta_1|>|\Delta_2|$ (see fig.
\ref{Fig_ang3(14)_5000_180_rdbk}).
\subsection*{Fig. \ref{Fig_ang3(1)_5000_120_rdbk}--\ref{Fig_ang3(1)_5000_120_full}}
First, when these lines are far enough from each other, there is no
collision of the support of the equilibrium measures and the supports
of $\lambda_1$ and $\lambda_2$ are two non-intersecting arcs,
respectively\footnote{It is well known that in the case of an Angelesco
system these two arcs in a sense are ``{\it attracted}'' to each other,
and in the case of a Nikishin system are they ``{\it repelled}'' from
each other (see \cite{Apt08}, also \cite{Nut84}, \cite{NuTr87},
\cite{GoRa85}, \cite{Gon03}).}, see \cite{Nut84}, \cite{NuTr87},
\cite{Apt08}. The measures $\lambda_1$ and $\lambda_2$ are absolutely
continuous with respect to the length of the arc $|dz|$, and their
densities $\lambda_j'$, $j=1,2$, in the neighborhoods of the branch
points $a_j,b_j$, behave like Chebyshev measures, i.e.
$\sim|z-a_j|^{-1/2}$ and $\sim|z-b_j|^{-1/2}$, respectively. This can
be seen very well on figure \ref{Fig_ang3(1)_5000_120_rdbk}, where the
red points are the zeros of the Hermite--Pad\'{e} polynomial
$Q_{120,1}$ and the black points are the zeros of $Q_{120,2}$. It is
obvious that the extremal compact sets $F_1$ and $F_2$ are attracted to
each other, and the zeros of the polynomials $Q_{120,j}$, which are
onto $F_j$, are repelled from each other and from the branch points
$a_j,b_j$, $j=1,2$. The zeros of the polynomial $Q_{120,0}$ (blue
points, see fig. \ref{Fig_ang3(1)_5000_120_full}) form a third extremal
compact $F_0$, which separates the compact sets $F_1$ and $F_2$. The
distribution of the zeros of the polynomial $Q_{n,0}$, when
$n\to\infty$, is described by the third extremal measure $\lambda_0$,
$\supp\lambda_0=F_0$. Thus, the following fact is true (see
\cite{Sta88}). Let $U_n(z):=\max\{\log|Q_{n,1}(z)|,\log|Q_{n,2}(z)|\}$,
$z\in\CC$. Then $U_n$ is subharmonic in $\CC$. Therefore,
$U_n(z)=-U^{\mu_n}(z)$, where $\mu_n$ is measure, $|\mu_n|\leq n$,
$U^{\mu_n}(z)=-\int\log|z-\zeta|\,d\mu_n(\zeta)$ is the logarithmic
potential with respect to $\mu_n$. For an arbitrary polynomial
$Q\in\CC[z]$ define the measure
$$
\chi(Q)=\sum_{\zeta:Q(\zeta)=0}\delta_\zeta ,
$$
which counts the number of zeros of the polynomial $Q$. Then
\begin{equation}
\label{eq_Sta88}
\lim_{n\to\infty}\frac1n\mu_n=
\lim_{n\to\infty}\frac1n\chi(Q_{n,0})=\lambda_0
\end{equation}
(the convergence in \eqref{eq_Sta88} is understood as weak convergence of measures).
\subsection*{Fig. \ref{Fig_ang3(2)_5000_180_rdbk}--\ref{Fig_ang3(2)_5000_180_full}}
Further convergence of the parallel lines along the imaginary axis leads to the following.
While the branch points are far enough from each other,
collision of the equilibrium measures does not occur.
However, on fig. \ref{Fig_ang3(2)_5000_180_rdbk} it is clearly seen,
that the upper extremal compact set $F_1$ has strongly curved towards
the lower extremal compact set $F_2$.
The third extremal compact set $F_0$, as before, separates $F_1$ and $F_2$.
\subsection*{Fig. \ref{Fig_ang3(8)_5000_180_rdbk}--\ref{Fig_ang3(8)_5000_180_full}}
Then, under further convergence of the branch points,
the upper extremal compact set $F_1$ has even strongly curved towards
the lower extremal compact set $F_2$, the support of the equilibrium measure
of the upper compact set $F_1$ starts to break down,
the second lower compact set almost does not change;
see fig. \ref{Fig_ang3(8)_5000_180_rdbk}.
The third extremal compact set $F_0$, as before,
separates the other two compact sets from each other,
but now it touches the second compact set $F_2$.
\subsection*{Fig. \ref{Fig_ang3(7)_5000_180_rdbk}--\ref{Fig_ang3(7)_5000_180_full}}
Finally, under certain relative positions of the pair of branch points,
the support of the equilibrium measure of the upper extremal compact
$F_1$ breaks down, while the second compact set $F_2$ has hardly changed,
see fig. \ref{Fig_ang3(7)_5000_180_rdbk}.
The third extremal compact set $F_0$, as before,
separates the other two compact sets from each other,
and touches the second compact set $F_2$,
see fig. \ref{Fig_ang3(7)_5000_180_full}.
\subsection*{Fig. \ref{Fig_ang3(9)_5000_180_rdbk}--\ref{Fig_ang3(9)_5000_180_full}}
Under further convergence of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
have reached the second (lower) compact set $F_2$. The second compact set $F_2$ has started to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
several points stand out,
which started to form another component.
Thus, a second component of the support of the equilibrium measure $\lambda_2$
started to form, i.e. the support of the equilibrium measure $\lambda_2$
started breaking down on two arcs; see fig. \ref{Fig_ang3(9)_5000_180_rdbk}.
The third extremal compact set $F_0$, as before, ``seeks'' to separate
the other two compact sets from each other, but now each of the compact sets $F_1$ and $F_2$
has two components. It is clearly seen, that the compact set $F_0$
now crosses the compact set $F_2$; see fig. \ref{Fig_ang3(9)_5000_180_full}.
\subsection*{Fig. \ref{Fig_ang3(14)_5000_180_rdbk}--\ref{Fig_ang3(14)_5000_180_full}}
Further, the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
cross the second compact set $F_2$. The second compact set $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues; see fig. \ref{Fig_ang3(14)_5000_180_rdbk}.
The third extremal compact set $F_0$ crosses the compact set $F_2$. As before,
it ``seeks'' to separate the other two compact sets $F_1$ and $F_2$ from each other,
but now each of these compact sets has by two components.
It is clearly seen, that at the junction of the red, black and blue points
appear two equilateral triangles with multicolored vertexes,
see fig. \ref{Fig_ang3(14)_5000_180_full}.
\subsection*{Fig. \ref{Fig_ang3(11)_5000_180_rdbk}--\ref{Fig_ang3(11)_5000_180_full}}
The two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
even further cross the second compact set $F_2$. The second compact set $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (even than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues. It is clearly seen, that at the junction of the red, black and blue points
appear two equilateral triangles with multicolored vertexes.
By analogy with classical Pad\'{e} approximants and two-point Pad\'{e} approximants
(see fig. \ref{Fig_pade10_2500_130_blu} and \ref{Fig_bus210b_4000_120_full}),
it is natural to assume, that the center of each triangle
has a Chebotarev point $v_1$, $v_2$ with zero density.
At the branch points $a_j,b_j$ the density of the measures $\lambda_1$
and $\lambda_2$ are proportional to $|z-a_j|^{-1/2},|z-b_j|^{-1/2}$, $j=1,2$,
respectively. There is a Froissart singlet (blue) on the imaginary axis;
see fig. \ref{Fig_ang3(11)_5000_180_full}, \ref{Fig_ang3(11)_5000_180_rdbk}.
\subsection*{Fig. \ref{Fig_ang3(13)_5000_180_rdbk}--\ref{Fig_ang3(13)_5000_180_full}}
Finally, under certain relative positions of the pairs of branch points,
the support $F_2$ of the equilibrium measure $\lambda_2$ is separated
on two practically equivalent arcs.
However, according to the distribution of the zeros of the polynomial $Q_{180,2}$,
the density $\lambda_2'$ of the equilibrium measures of each arc must be different.
On the upper arc it behaves like a Chebyshev measure, that is at the end points $a_2,b_2$
the density is proportional to $|z-a_2|^{-1/2}$ and $|z-b_2|^{-1/2}$, respectively.
The end points of the lower arc $v_1,v_2$ are the Chebotarev points
and their density is proportional to $|z-v_1|^{1/2}$ and $|z-v_2|^{1/2}$;
see fig. \ref{Fig_ang3(13)_5000_180_rdbk}.
At the junction of the red, black and blue points
appeared two equilateral triangles with multicolored vertexes,
and the center of each has a Chebotarev point $v_1,v_2$;
see fig. \ref{Fig_ang3(13)_5000_180_full}.
\subsection{Nikishin system}\label{s2s2}
In the theory of Pad\'{e} approximants it is well known
the property called ``{\it Froissart doublets}'',
which was experimentally found (see \cite{BaGr81}, \cite{Gil78}, \cite{GiKr03})
and means that in the maximal\footnote{This notion has been introduced by
H. Stahl \cite{Sta85a}--\cite{Sta85c}. The existence of such a domain
$D=D(f)=\myo\CC\setminus{S}$ follows from the classical theorem of Stahl
\cite{Sta97b}. In this regard, the domain $D$ is usually
called {\it Stahl domain} and the symmetrical compact set $S$ -- {\it Stahl
compact} for the multi-valued analytic function $f$.}
domain of holomorphy $D=D(f)$ of the multi-valued function $f$
for some $n\in\Lambda$, $\Lambda\subset\NN$ is an infinite sequence,
are positioned pairs of zero-pole of a diagonal Pad\'{e} approximant
(that is, different from each other zeros of the Pad\'{e} polynomials $P_{n,0},P_{n,1}$,
see fig. \ref{Fig_pade10_2500_130}--\ref{Fig_pade103_5000_266_full}).
For every fixed $n\in\Lambda$, these points are different from each other,
the pole is not dependent of any singularity of the original function,
the zero and the pole are infinitely close to each other when $n\to\infty$,
that is they are asymptotically ``canceled out''.
In other words, the residue of the Pad\'{e} approximant at such a pole
converges to zero when $n\to\infty$.
Because such poles do not correspond to the singularities of the original function,
sometimes they are called ``spurious'' poles and zeros or ``defects''
of the diagonal Pad\'{e} approximant.
In a typical case these poles and zeros are dense on the Riemann surface
$\myo\CC$ when $n\to\infty$, $n\in\Lambda$
(see \cite{Sta96}, \cite{Sta98}, \cite{Sue02b}, \cite{Sue04}, \cite{Sue10}).
Thus they are sometimes called wondering poles and zeros or
floating poles and zeros. For an arbitrary
algebraic function $f$ the number of these pairs
depends mainly on the genus of the corresponding Riemann surface,
and also of the number of zeros of the functions $\Delta{f}(\zeta)=(f^{+}-f^{-})(\zeta)$,
$\zeta\in{S}$, on the Stahl compact set $S=S(f)$.
It is shown in \cite{Sue04}, that the appearance of the Froissart doublets
is due to points of an ``incorrect'' interpolation of the diagonal Pad\'{e} approximants
$[n/n]_f:=-P_{n,0}/P_{n,1}$ in the Stahl domain $D(f)=\myo\CC\setminus{S}$ with another branch
$\myt{f}$ of the original function $f$ when $n\in\Lambda$. Namely, the existence
of Froissart doublets does not allow uniform convergence of the Pad\'{e} approximants
in the Stahl domain (for details see below).
The second part of the numerical experiments are in the case,
when the sets of singularity points for the two functions
$f_1$ and $f_2$ {\it intersect} each other.
Specifically, we select a collection of three functions $[1,f,f^2]$, where
the function $f$ is of the type \eqref{4} and thus, the pair of functions $f,f^2$
forms a Nikishin system (see \cite{ApLy10}, \cite{Rak11}).
In this case another new property has been found. Namely,
the appearance of triple zeros ({\it Froissart triplets}, see below), i.e.
zeros of the Hermite--Pad\'{e} polynomials $Q_{n,0},Q_{n,1},Q_{n,2}$,
which are very close to each other, but still have different values.
These zeros are in the domain of holomorphy
of the functions $f,f^2$, do not correspond to either zeros, nor singularities
of these function, for each $n$ they are practically identical to each other
and with the transfer from $n$ to $n+1$ they shifted in the complex plane as one
unit.
It is appropriate to compare these triplets with the very well known
Froissart doublets for classical Pad\'{e} polynomials
(see \cite{Fro69}, also \cite{BaGr81}, \cite{Gil78}, \cite{GiTr87},
\cite{GiKr03}, \cite{GoGu13}, \cite{AdIb13}, \cite{Bel09}, \cite{Sta97b}, \cite{Tre13}).
Froissart doublets are sometimes called ``defects'' \cite[Chapter 2, \S\,2.2]{BaGr81},
and also spurious or wondering (floating) zeros and poles of the Pad\'{e} approximant
\cite{BaGr81}, \cite{Chu80} (see also \cite{Sue04}).
It is considered that the existence of such zeros and poles of
the Pad\'{e} approximant does not allow uniform convergence of the Pad\'{e} approximant.
Because such zeros and poles are infinitely close to each other asymptotically,
then when the limit is taken, they are practically ``canceled out''.
For some classes of hyperelliptic functions $f$,
which allow the representation $f=\myh{\sigma}$, where the support
$S_\sigma=\bigsqcup\limits_{j=1}^{2g+2}[e_{2j-1},e_{2j}]\subset\RR$ of the measure
$\sigma$ consists of finite number of non-intersecting intervals, it
was shown \cite{Sue00}, \cite{Sue06}, \cite{ApBuMaSu11} that the movement
of such poles is subject to certain regularity.
Namely, the corresponding divisor of the Nuttall $\Psi$ function
(see \cite{Sue00}, \cite{Sue06}, \cite{ApBuMaSu11})
moves along a Riemann surface $\RS_2$ with two sheets of genus $g\ge1$
and is subject to the general {\it Dubrovin system}:
\begin{equation}
\label{eqdub}
\dot z_k=-\frac{2w(\zz_k)}{\prod\limits_{j\neq k}(z_k-z_j)}
\int_{e_{2g+2}}^{\infty}\frac{\prod\limits_{j\neq k}(x-z_j)}{w(x)}\,dx,
\qquad k=1,\dots,g,
\end{equation}
where $\dot z_k=dz_k/dt$, $e_{2g+2}$ is the rightmost point of the support
$S_\sigma=\bigsqcup\limits_{j=1}^{2g+2}[e_{2j-1},e_{2j}]\subset\RR$ of the measure
$\sigma$ (it is supposed, that the support of the measure $\sigma$ has $g\geq1$ gaps,
the endpoints of the intervals of the support are numbered according to the
ascending values, and the path of integration in \eqref{eqdub} is part
of the real axis $[e_{2g+2},+\infty)$; for details about the notations
see \cite{Sue02b}). The points
$\zz_1(t),\dots,\zz_{g}(t)$, $t\in\RR_{+}$, are on the corresponding
hyperelliptic Riemann surface $\RS_2$ of genus $g$.
If $r$ is a real-valued rational function, which has poles
only outside $\myh{S}_\sigma$, then for ${f}=\myh\sigma+r$ we have: {\it
$[n/n]_{{f}}\to {f}$ in the spherical metric locally uniformly in
$\myo{\CC}\setminus\myh{S}_\sigma$} ($[n/n]_f:=-P_{n,1}/P_{n,0}$);
thus, each pole $r$ attracts exactly the number of poles of $[n/n]_{{f}}$,
as its multiplicity (see \cite{Gon03}).
In this case the movement of the poles and the points of interpolation
of the Pad\'{e} approximants $[n/n]_{{f}}$, which are in the
gaps between intervals, are subject to the system \eqref{eqdub}.
The fact, that such poles of the Pad\'{e} approximants form a pair
with the zeros that are near them and in this pair they are asymptotically close to each other
and when taking the limit are canceled out,
has led some authors to believe that their appearance
is random and is not connected with the nature of the original function $f$.
Such an approach to this property reflected on the respective terminology:
such zeros and poles of the approximant began to be called ``spurious''.
Their appearance, when using Pad\'{e} approximants, was considered especially negatively
in \cite{BaGr81}, \cite{Bel09}, \cite{Sta97}, therefore when
$n\in\Lambda$ the Pad\'{e} approximants $[n/n]_f$ themselves
sometimes were thought to be defective or entirely excluded
from the research \cite[Chapter 2, \S\,2.3]{BaGr81} or was proposed of using the
so-called ``purification'' of the spurious poles \cite{Sta96}.
In \cite{Dum08} Dumas researched the problem of the asymptotical
behavior of the sequence $\{[n/n]_d\}_{n\in\NN}$ for
elliptic\footnote{Speaking about elliptic functions,
we follow to the terminology of the monograph \cite[Chapter 10, par. 10.10]{Spr60},
where under elliptic function is understood a single-valued function
defined on an elliptical Riemann surface.}
functions of the special form
\begin{equation}
d(z)=\sqrt{(z-e_1)\dotsb(z-e_4)}-z^2+\(e_1+\dotsb+e_4\)z/2,
\label{1.1}
\end{equation}
where the points $e_1,\dots,e_4\in\mathbb C$ are pairwise different
and such root branch is selected,
that the main member of which, in a neighborhood of the point $z=\infty$,
is equal to $z^2$; thus $d\in\HH(\infty)$.
Particularly, Dumas has shown that in
{\it ``general position'' the set of poles of the Pad\'{e} approximants $[n/n]_d$
is dense in $\overline{\mathbb C}$}.
In \cite{Sue03} the result of Dumas has been extended to some
{\it classes\/} of elliptical functions.
It was found \cite[\S\,1, par .2]{Sue04} the following property:
for each $n$, from some subsequence $\Lambda$,
always exists a point $\mb=\mb_n\in{D}=\myo{\CC}\setminus{S}$,
for which
$[n/n]_d(\mb_n)=\myt{d}(\mb_n)$, where
$\myt{d}(z)=-\sqrt{(z-e_1)\dotsb(z-e_4)}-z^2+(e_1+\dotsb+e_4)z/2$ is
{\it another branch\/} of the elliptical function \eqref{1.1} in $D$.
Thus, for each $n$, from some subsequence $\Lambda$,
the diagonal Pad\'{e} approximants $[n/n]_d$ {\it interpolate\/} at
some point from the domain $D$ {\it another branch} $\myt{d}$ of the function $d$.
In the case of general position, such points of the ``incorrect'' interpolation
$\{\mb_n\}_{n\in\Lambda}$ are dense in $\myo{\CC}$.
In \cite{Sue04} were obtained much more general results in this direction.
In the work of E.~A.~Rakhmanov \cite{Rak12} was obtained an electrostatic interpretation
of the poles of the diagonal Pad\'{e} approximant $[n/n]_f$
for some algebraic functions $f$ {\it for each fixed} $n$.
In this framework, the spurious poles play a special role:
they should be considered a part of the ``external field''.
\subsection{}\label{s2s3}
In the present work, experimentally was obtained a new property
for the Nikishin system $f_1=f,f_2=f^2$, where
\begin{equation}
f(z)=(z^2-1)^{1/4}(z-a)^{-1/2},\quad\text{where}\quad
f(\infty)=1,\quad a\notin\RR,
\label{1.2}
\end{equation}
which is connected with the behavior of the zeros of the Hermite--Pad\'{e}
polynomials of first kind. Namely, there appear, for some $n\in\Lambda=\Lambda(f)$,
single spurious zeros of the Hermite--Pad\'{e} polynomials,
and also triple spurious zeros of the Hermite--Pad\'{e} polynomials.
Under ``spurious'' zeros of the Hermite--Pad\'{e} polynomials we understand
such zeros that, first, are different from each other (they cannot be canceled out),
second, do not correspond to either zeros, nor singularities of the original function,
and third, significantly change their location, when we transfer from $n$ to $n+1$
(sometimes they may disappear). By analogue with Pad\'{e} polynomials, it is
naturally to call such zeros {\it Froissart singlets} and {\it Froissart triplets}.
It is natural to assume the same as for Pad\'{e} approximants,
in the ``typical'' case (i.e. for the branch points $a$, which are in
``general position'', that is $a\notin i\RR$, see \eqref{1.2})
these spurious zeros (singlets and triplets)
of the Hermite--Pad\'{e} polynomials are dense only in ``their'' domain,
which is the difference from Pad\'{e} approximants.
From the numerical experiments it follows that
(see fig. \ref{Fig_nik_(1_6)_5000_165_bl}, \ref{Fig_nik_(1_6)_5000_165_bk},
\ref{Fig_nik_(_9)_5000_180_blbk}) the zeros of the polynomials $Q_{n,0}$ and $Q_{n,2}$ of
the functions of type \eqref{1.2} when $a=i\cdot 1.6$ have the same limiting distribution,
the corresponding boundary compact set $F_2$ separates the Riemann surface into
three domains, two of which have an internal boundary arc.
Thus, they have the same structure (see fig. \ref{Fig_bus205c_4000_120_full})
as in the theorem of Buslaev \cite{Bus13}, \cite{BuMaSu12}
for two-point Pad\'{e} approximants. Remark that,
the single zeros of the polynomial $Q_{n,2}$ at the point $z\approx a$
corresponds to a simple pole of the function $f^2$ at the point $z=a$.
The corresponding compact set $F_2$ has four Chebotarev's points (three of
which have zero density and one has infinite density) for the equilibrium measure,
which corresponds to the limiting distribution of the zeros of
the Hermite--Pad\'{e} polynomials $Q_{n,0},Q_{n,2}$.
It is clear, that the distribution of the zeros of the Hermite--Pad\'{e} polynomials
$Q_{n,0}$ and $Q_{n,2}$ must be equivalent, since with the mapping
$f\mapsto 1/f$ the type of the singularity of the original function
stays the same (see \eqref{1.2}).
The limiting distribution of the zeros of the polynomial $Q_{n,1}$ must
be different. That is, it must correspond to the second compact set
$F_1$ (see fig. \ref{Fig_nik_(1_6)_5000_165_full},
\ref{Fig_nik_(1_6)_5000_165_rd} and fig.
\ref{Fig_nik_(_9)_5000_187_full}--\ref{Fig_nik_(_9)_5000_187_bk}). If
now we substitute $F:=F_1\cup F_2$, $F_1\cap F_2\neq\varnothing$, then
the complement $\myo\CC\setminus{F}$, as before, consists if three
domains, but now each of these domains has an internal boundary arc
(see fig. \ref{Fig_bus205c_4000_120_full}). Thus, $\myo\CC=F \sqcup
D_1\sqcup D_2\sqcup D_3$. Since the pair $f,f^2$ forms a Nikishin
system, then by the analogy with \cite{RaSu12}, \cite{RaSu13},
\cite{KovSu14}, \cite{Sue13b}, we can associate with the collection of
functions $[1,f,f^2]$ a Nuttall condenser. After then by the analogue
with \cite{RaSu13}, we can to describe in terms of the condenser the
limiting distribution of the zeros of the Hermite--Pad\'{e} polynomials
$Q_{n,0},Q_{n,1},Q_{n,2}$ and the corresponding Riemann surface with
the canonical (in Nuttall's sense) partition into three sheets. After
that, with the help of this Riemann surface, it might be possible to
find strong asymptotics for the Hermite--Pad\'{e} polynomials. The
question about how exactly to do this stays open; a heuristical method
for solving this problem for functions of the type
$f(z)=(z-a_1)^{\alpha_1}(z-a_2)^{\alpha_2}(z-a_3)^{\alpha_3}$,
$\alpha_1+\alpha_2+\alpha_3=0$, $2\alpha_j\in\CC\setminus\ZZ$, is
proposed in \cite{Sue13b}.
Figures \ref{Fig_nik_(1_6eps)_3000_121-130_full}--\ref{Fig_nik_(1_6eps)_3000_121-130_bk}
show the distribution of the zeros of the Hermite--Pad\'{e} polynomials
$Q_{n,0}$ (blue points), $Q_{n,1}$ (red points), $Q_{n,2}$ (black points),
$n=121,\dots,130$, for the collection of three functions $[1,f,f^2]$, where $f$ is
``perturbed'' with respect to the function \eqref{1.2}:
\begin{equation}
\label{nik_eps}
f(z)=(z^2-1)^{1/4}(z-(0.1+i\sqrt{3}\cdot1.6))^{-1/2}.
\end{equation}
It is clear, that by comparison with the unperturbed case \eqref{1.2},
the picture of the distribution of the zeros is entirely the same.
\section{Concluding remarks}\label{s3}
Thus, the main empirical results of the paper are as follows.
\subsection{}\label{s3s1}
In the case when the pair of functions $f_1,f_2$ forms
an Angelesco system, where the functions $f_1,f_2$ have the type \eqref{3},
there has been found numerically the property of the {\it mutual pushing}
of the supports of the measures, which are equilibrium for
the extremal compact sets.
\subsection{}\label{s3s2}
In the case when the pair of functions $f_1=f,f_2=f^2$ forms a (generalized)
Nikishin
system, where the function $f$ has the type \eqref{4}, there have been
found numerically the properties of {\it Froissart singlets and triplets}
presence.
Thus, in the present paper there have been found numerically new properties
related to the behavior of the zeroes of the Hermite--Pad\'{e} polynomials
of the first kind.
These numerical phenomena should be researched and strictly justified in
future.
\clearpage
\markboth{\bf Pade approximant}{\bf Pade approximant}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_1}}
\vskip-6mm
\caption{Zeros and poles of the diagonal Pad\'{e} approximant $[131/131]_f$ of the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))(z-(0.5-i\cdot1.2))\}^{1/3}$.
The zeros and poles of the diagonal Pad\'{e} approximant $[n/n]_f$ are distributed,
when $n\to\infty$, according to Stahl's theorem \cite{Sta97b}.
When $n=131$ the zeros and poles of the diagonal Pad\'{e} approximant $[131/131]_f$
of the function $f$ are distributed accordingly to the electrostatical model,
by E. A. Rakhmanov \cite{Rak12}, and they model
the Chebotarev-Stahl compact $S_{131}$, which depends on $n$.
Froissart doublets are not present when $n=131$.
}
\label{Fig_pade10_2500_131}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_2}}
\vskip-6mm
\caption{Zeros and poles of the diagonal Pad\'{e} approximant $[130/130]_f$ of the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8i))(z-(0.9+i\cdot1.5))
(z-(0.5-i\cdot1.2))\}^{1/3}$, distributed accordingly
to the electrostatical model by E. A. Rakhmanov \cite{Rak12}.
There is a Froissart doublet when $n=130$ (see also fig.
\ref{Fig_pade10_2500_130_red} and fig. \ref{Fig_pade10_2500_130_blu}).
Since the genus of the Riemann surface is $1$,
there might be at most one Froissart doublet.
In full compliance with the Rakhmanov model \cite{Rak12},
the Froissart doublet ``attracts'' the Stahl $S$-compact $S_{130}$,
comp. fig. \ref{Fig_pade10_2500_131}.
}
\label{Fig_pade10_2500_130}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_3}}
\vskip-6mm
\caption{The poles of the Pad\'{e} approximant $[130/130]_f$
approximate a Chebotarev point $v_{130}$
for the $S$-compact $S_{130}$. When $n\to\infty$ we have that $v_n\to v$ is a classical
Chebotarev point. There is one spurious pole of the Pad\'{e} approximant $[130/130]_f$,
it is accompanied by a spurious zero of the Pad\'{e} approximant $[130/130]_f$
(see fig. \ref{Fig_pade10_2500_130_blu}).
}
\label{Fig_pade10_2500_130_red}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_4}}
\vskip-6mm
\caption{The Chebotarev point should not be approximated by
zeros of the Pad\'{e} approximant $[130/130]_f$.
There is one spurious zero of the Pad\'{e} approximant $[130/130]_f$,
it is accompanied by a spurious pole of the Pad\'{e} approximant $[130/130]_f$
(see fig. \ref{Fig_pade10_2500_130_red}).
}
\label{Fig_pade10_2500_130_blu}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_5}}
\vskip-6mm
\caption{Zeros and poles of the diagonal Pad\'{e} approximants $[n/n]_f$, $n=121,\dots,130$,
of the function $f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))
(z-(0.5-i\cdot1.2))\}^{1/3}$ for each $n=121,\dots,130$
are distributed in the complex plane accordingly to the electrostatical model
by E. A. Rakhmanov \cite{Rak12}.
Since the genus of the Riemann surface is $1$,
there might be at most one Froissart doublet for each $n$.
Here are observed 5 Froissart doublets.
In full compliance with the Rakhmanov model \cite{Rak12}
the $n$-th Froissart doublet ``attracts''
the Stahl $S$-compact $S_{n}$, $n\in\{121,\dots,130\}$,
comp. fig. \ref{Fig_pade10_2500_131}.
}
\label{Fig_pade10_2500_121_130}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_6}}
\vskip-6mm
\caption{For the function $f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))
(z-(0.5-i\cdot1.2))\}^{1/3}$
poles of the Pad\'{e} approximants $[n/n]_f$, $n=121,\dots,130$,
approximate a Chebotarev point.
However, the presence of Froissart doublets
changes slightly the position of the existing points $v_n$
depending on $n\in\{121,\dots,130\}$.
This is in full accordance with the Rakhmanov model \cite{Rak12}.
}
\label{Fig_pade10_2500_121_130_red}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_7}}
\vskip-6mm
\caption{Zeros of the Pad\'{e} approximants $[n/n]_f$, $n=121,\dots,130$, for the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))(z-(0.5-i\cdot1.2))\}^{1/3}$.
The Chebotarev point should not be approximated by
zeros of the Pad\'{e} approximant $[n/n]_f$.
}
\label{Fig_pade10_2500_121_130_blu}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_8}}
\vskip-6mm
\caption{Zeros and poles of the diagonal Pad\'{e} approximants $[n/n]_f$
when $n=131$ and $n=132$ of the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))(z-(0.5-i\cdot1.2))\}^{1/3}$.
The Froissart doublets are not present when $n=131$ and $n=132$.
Therefore the $S$-compacts $S_{131}$ and $S_{132}$ are practically the same.
}
\label{Fig_pade10_2500_131_132}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_9}}
\vskip-6mm
\caption{Poles of the diagonal Pad\'{e} approximants $[n/n]_f$
when $n=131$ and $n=132$ of the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))(z-(0.5-i\cdot1.2))\}^{1/3}$.
The Froissart doublets are not present when $n=131$ and $n=132$.
Therefore, the Chebotarev points $v_{131}$ and $v_{132}$, approximated by
the poles of the diagonal Pad\'{e} approximants $[n/n]_f$ when
$n=131$ and $n=132$, are very close to each other and are practically the same.
}
\label{Fig_pade10_2500_131_132_red}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_10}}
\vskip-6mm
\caption{Zeros of the Pad\'{e} approximants $[n/n]_f$ when $n=131$ and $n=132$ of the function
$f(z)=1/\{(z-(-1.2+i\cdot0.8))(z-(0.9+i\cdot1.5))(z-(0.5-i\cdot1.2))\}^{1/3}$
do not approximate the Chebotarev points $v_{131}$ and $v_{132}$.
The Froissart doublets are not present when $n=131$ and $n=132$.
Therefore the zeros the diagonal Pad\'{e} approximants $[n/n]_f$ when $n=131$ and $n=132$
are practically the same.
}
\label{Fig_pade10_2500_131_132_blu}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_1_11}}
\vskip-6mm
\caption{Zeros and poles of the diagonal Pad\'{e} approximants $[103/103]_f$ of the function
$f(z)=1/\bigl\{(z+(4.3+i\cdot1.))
(z-(2.+i\cdot.5))\*(z+(2.+i\cdot2.))\*(z+(1.-i\cdot3.))
(z-(4.+i\cdot2.))(z-(3.+i\cdot5.))\bigr\}^{1/6}$.
In the limit when $n\to\infty$
the zeros and poles of the diagonal Pad\'{e} approximants $[n/n]_f$
are distributed accordingly with Stahl's theorem \cite{Sta97b}.
Under fixed $n=103$ these zeros and poles are distributed in a plane,
accordingly to the electrostatical model by Rakhmanov \cite{Rak12}.
Since the genus of the Riemann surface is $1$,
for each $n$ there might be no more than 4 Froissart doublets.
Here are observed 3 Froissart doublets.
When $n=103$, in full compliance with the Rakhmanov model,
the Froissart doublets ``attract'' the Stahl $S$-compact $S_{103}$.
}
\label{Fig_pade103_5000_266_full}
\end{figure}
\clearpage
\markboth{\bf Case One, Angelesco system}{\bf Case One, Angelesco system}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_2_1}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{90,1}$ (red points)
and $Q_{90,2}$ (black points) for the set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z+1)/z}$, $f_2=\sqrt{(z-3)/z}$ (we select such a root branch,
that $\sqrt{1\,}=1$). The functions $f_1,f_2$ are of Markov type:
$f_1(z)=\myh\sigma_1(z)$, $f_2(z)=\myh\sigma_2(z)$, where
$\sigma'_1(x)=\sqrt{(1+x)/(-x)}/\pi$, $x\in(-1,0)$,
$\Delta_1=\supp{\sigma_1}=[-1,0]$,
$\sigma'_2(x)=\sqrt{(3-x)/x}/\pi$, $x\in(0,3)$,
$\Delta_2=\supp{\sigma_2}=[0,3]$. It is known \cite{GoRa81}, \cite{Apt08}, that
the zeros of the polynomials $Q_{n,1}$ and $Q_{n,2}$, when taken the limit,
are distributed on the support of the equilibrium measures
$S_1=\lambda_1$ and $S_2=\lambda_2$ and in accordance with their densities
$\lambda_1'(x)$ and $\lambda_2'(x)$. The figure clearly shows, that
$S_1=\Delta_1$, but $S_2=[a_2^*,3]\subsetneq\Delta_2$, because
$|\Delta_1|<|\Delta_2|$. The point $a_2^*\in(0,3)$ is calculated by the formula \eqref{eq.kal84},
obtained by V. A. Kalyagin \cite{Kal84} (see also \cite{Nut84}, \cite{ApKa86}).
The density $\lambda_1'(x)$ behaves, in the neighborhoods of the points $a_1$ and $b_1$,
like $(x-a_1)^{-1/2}$ and $(b_1-x)^{-1/2}$, respectively, and the density
$\lambda_2'(x)$ behaves, in the neighborhood of the point $a_2$, like $(x-a^*_1)^{1/2}$
and in the neighborhood of the point $b_2$ like $(b_2-x)^{-1/2}$.
}
\label{Fig_mar_1_2000_90_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_2_2}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{90,0}$ (blue points)
and $Q_{90,2}$ (black points) for the set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z+1)/z}$, $f_2=\sqrt{(z-3)/z}$.
The zeros of the polynomial $Q_{90,0}$ (blue points, comp fig.
\ref{Fig_ang3(1)_5000_120_full}) create a third extremal compact $F_0$,
which ``separates'' the compacts $F_1$ and $F_2$.
}
\label{Fig_mar_1_3000_90_blbk}
\end{figure}
\clearpage
\markboth{\bf Angelesco system and Nikishin system}{\bf Angelesco system and Nikishin system}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_1}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt{(z-2)/(z+2)}$.
The sets of branch points of the functions $f_1$ and $f_2$ do not intersect each other,
and thus the pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Nikishin system
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$;
see fig. \ref{Fig_nik(2_1)_4000_120_full}--\ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_ang(2_2)_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_2}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$.
The pair $f_1,f_2$ create a Nikishin system,
because the branch points of the functions are equivalent.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution of the zeros for the Angelesco system
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt{(z-2)/(z+2)}$;
see fig. \ref{Fig_ang(2_2)_4000_120_full},
\ref{Fig_ang(2_2)_4000_120_blbk},
\ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_nik(2_1)_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_3}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt{(z-2)/(z+2)}$.
The pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Nikishin system
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$;
comp. fig. \ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_ang(2_2)_4000_120_blbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_4}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
and $Q_{120,1}$ (red points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$.
The pair $f_1,f_2$ create a Nikishin system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Angelesco system
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt{(z-2)/(z+2)}$;
comp. fig. \ref{Fig_ang(2_2)_4000_120_blbk}.
}
\label{Fig_nik(2_1)_4000_120_blrd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_5}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The sets of branch points of the functions $f_1$ and $f_2$ do not intersect each other,
and thus the pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Nikishin system
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$;
see fig. \ref{Fig_nik(2_1)_4000_120_full}--\ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_ang(3_3)_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_6}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Nikishin system
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$;
comp. fig. \ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_ang(3_3)_4000_120_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_7}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,1}$ (red points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is the same as the distribution for the Nikishin system
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{1/3}$;
comp. fig. \ref{Fig_nik(2_1)_4000_120_blrd}.
}
\label{Fig_ang(3_3)_4000_120_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_8}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The sets of branch points of the functions $f_1$ and $f_2$ do not intersect each other,
and thus the pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is different than the distribution of the pair
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
comp. fig. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
This confirms that the distribution of the zeros of the Hermite-Pad\'{e} polynomials
depends not only on the geometrical position of the branch points,
but also of the type of the branch points.
}
\label{Fig_ang(2_3)_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_9}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is different than the distribution of the pair
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
comp. fig. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
This confirms that the distribution of the zeros of the Hermite-Pad\'{e} polynomials
depends not only on the geometrical position of the branch points,
but also of the type of the branch points.
}
\label{Fig_ang(2_3)_4000_120_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_10}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,1}$ (red points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The pair $f_1,f_2$ create an Angelesco system.
However, the distribution of the zeros of the Hermite-Pad\'{e} polynomials
for this system is different than the distribution of the pair
$f_1(z)=\sqrt[3]{(z-1)/(z+1)}$, $f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
comp. fig. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
This confirms that the distribution of the zeros of the Hermite-Pad\'{e} polynomials
depends not only on the geometrical position of the branch points,
but also of the type of the branch points.
}
\label{Fig_ang(2_3)_4000_120_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_11}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_3_3_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_12}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_3_3_4000_120_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_13}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,1}$ (red points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt[3]{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(3_3)_4000_120_full}--\ref{Fig_ang(3_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_3_3_4000_120_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_14}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(2_3)_4000_120_full}--\ref{Fig_ang(2_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_2_3_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_15}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(2_3)_4000_120_full}--\ref{Fig_ang(2_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_2_3_4000_120_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_16}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,1}$ (red points)
for a set of three functions $[1,f_1,f_2]$, where
$f_1(z)=\sqrt{(z-(1+i\cdot 0.4))/(z-(-1+i\cdot 0.4))}$,
$f_2(z)=\sqrt[3]{(z-2)/(z+2)}$.
The set of the branch points of the functions $f_1$ and $f_2$ do not intersect each other,
thus the pair $f_1,f_2$ create an Angelesco system;
comp. \ref{Fig_ang(2_3)_4000_120_full}--\ref{Fig_ang(2_3)_4000_120_rd}.
}
\label{Fig_ang_im(_4)_2_3_4000_120_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_17}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points) and $Q_{120,2}$ (black points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{-1/3}$.
Here the pair $f_1,f_2$ create a Nikishin system,
because the branch points of the functions are equivalent.
}
\label{Fig_nik(2_-1)_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_3_18}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points)
and $Q_{120,1}$ (red points) for a set of three functions
$[1,f_1,f_2]$, where
$f_1(z)=\((z-1)/(z+1)\)^{1/3}\((z-2)/(z+2)\)^{1/3}$,
$f_2(z)=\((z-1)/(z+1)\)^{2/3}\((z-2)/(z+2)\)^{-1/3}$.
The pair $f_1,f_2$ create a Nikishin system.
}
\label{Fig_nik(2_-1)_4000_120_blrd}
\end{figure}
\clearpage
\markboth{\bf Case One, Angelesco system}{\bf Case One, Angelesco system}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_1}}
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{120,1}$ (red points),
$Q_{120,2}$ (black points) in the plane $\CC_z$ for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.7))(z-(1.0+i\cdot0.7))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.7))(z-(.65-i\cdot0.7))}$.
Since the branch points $a_1=-1.0+i\cdot0.7,b_1=1.0+i\cdot0.7$ and
$a_2=-0.65+i\cdot0.7,b_2=0.65+i\cdot0.7$ of the functions are
far enough from each other, there is no collision of the supports
of the equilibrium measures and the supports $\lambda_1$ and $\lambda_2$
are two non-intersecting arcs.
The measures $\lambda_1$ and $\lambda_2$ are absolutely continuous with respect
to the length of the arc $|dz|$, and their densities $\lambda_j'$, $j=1,2$,
behave, in the neighborhoods of the branch points $a_j,b_j$, like Chebyshev measures,
i.e. $\sim|z-a_j|^{-1/2}$ and $\sim|z-b_j|^{-1/2}$, respectively.
It is obvious that the extremal compacts $F_1$ and $F_2$ are converging to each other,
and the zeros of the polynomials $Q_{120,j}$, which are onto $F_j$,
are diverging from each other and from the branch points $a_j,b_j$, $j=1,2$.
The zeros of the polynomial $Q_{120,0}$ (blue points, see
fig. \ref{Fig_ang3(1)_5000_120_full}) create a third extremal compact
$F_0$, which separates the compacts $F_1$ and $F_2$.
}
\label{Fig_ang3(1)_5000_120_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_2}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{120,0}$ (blue points),
$Q_{120,1}$ (red points), $Q_{120,2}$ (black points) in the plane $\CC_z$
for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.7))(z-(1.0+i\cdot0.7))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.7))(z-(.65-i\cdot0.7))}$.
The red points are distributed onto the support of the equilibrium measure
$\lambda_1$, when taking the limit, accordingly to its density $\lambda'_1$.
The black points are distributed onto the support of the equilibrium measure
$\lambda_2$, when taking the limit, accordingly to its density $\lambda'_2$.
Same applies for the blue points.
Both measures $\lambda_1$ and $\lambda_2$ are absolutely continuous
on the corresponding extremal curves $F_1=\supp{\lambda_1}$ and $F_2=\supp{\lambda_2}$
with respect to the length of the arc $ds$, and their densities
behave like the Chebyshev measures $|z-a_j|^{-1/2},|z-b_j|^{-1/2}$, $j=1,2$
around the endpoints of the curves.
The blue points (zeros of the polynomial $Q_{120,0}$) separate the red points
(zeros of the polynomial $Q_{120,1}$) from the black points (zeros of the polynomial $Q_{120,2}$).
}
\label{Fig_ang3(1)_5000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_3}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.6))(z-(1.0+i\cdot0.6))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.6))(z-(.65-i\cdot0.6))}$.
The branch points have come closer to each other,
however they are still far enough from each other
and there is no collision of the supports of the equilibrium measures.
It is clearly seen, that the upper extremal compact $F_1$
has strongly bent towards the lower extremal compact $F_2$.
}
\label{Fig_ang3(2)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_4}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points), $Q_{180,2}$ (black points) for two functions
$f_1(z)=\sqrt{(1.0-(-1.0+i\cdot0.6)\zeta)(1.0-(1.0+i\cdot0.6)\zeta)}$,
$f_2(z)=\sqrt{(1.0-(-.65-i\cdot0.6)\zeta)(1.0-(.65-i\cdot0.6)\zeta)}$.
The branch points have come closer to each other,
however they are still far enough from each other
and there is no collision of the supports of the equilibrium measures.
It is clearly seen, that the upper extremal compact $F_1$
has strongly bent towards the lower extremal compact $F_2$.
The third extremal compact $F_0$, as before, separates $F_1$ and $F_2$.
}
\label{Fig_ang3(2)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_5}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.525))(z-(1.0+i\cdot0.525))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.525))(z-(.65-i\cdot0.525))}$.
It is clearly seen, that the upper extremal compact $F_1$
has even strongly bent towards the lower extremal compact $F_2$.
The third extremal compact $F_0$, as before, separates $F_1$ and $F_2$.
The support of the equilibrium measure of the upper extremal compact
$F_1$ starts to break down, while the second compact $F_2$ has hardly changed.
}
\label{Fig_ang3(8)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_6}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials
$Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.525))(z-(1.0+i\cdot0.525))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.525))(z-(.65-i\cdot0.525))}$.
It is clearly seen, that the upper extremal compact $F_1$
has even strongly bent towards the lower extremal compact $F_2$.
The third extremal compact $F_0$, as before, separates $F_1$ and $F_2$.
The support of the equilibrium measure of the upper extremal compact
$F_1$ starts to break down, while the second compact $F_2$ has hardly changed.
The third extremal compact $F_0$, as before,
separates the other two compacts from each other,
but now it touches the second compact $F_2$.
}
\label{Fig_ang3(8)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_7}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.5))(z-(1.0+i\cdot0.5))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.5))(z-(.65-i\cdot0.5))}$.
It is clearly seen, that under this position of the pair of branch points,
the support of the equilibrium measure of the upper extremal compact
$F_1$ breaks down, while the second compact $F_2$ has hardly changed.
The red points break down the support of the equilibrium measure $\lambda_1$
of the compact $F_1$ on two arcs.
}
\label{Fig_ang3(7)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_8}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials
$Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.5))(z-(1.0+i\cdot0.5))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.5))(z-(.65-i\cdot0.5))}$.
It is clearly seen, that under this position of the pair of branch points,
the support of the equilibrium measure of the upper extremal compact
$F_1$ breaks down, while the second compact $F_2$ has hardly changed.
The third extremal compact $F_0$, as before,
separates the other two compacts from each other,
and touches the second compact $F_2$
}
\label{Fig_ang3(7)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_9}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.45))(z-(1.0+i\cdot0.45))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.45))(z-(.65-i\cdot0.45))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
have reached the second (lower) compact $F_2$. The second compact $F_2$ has started to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$) several points stand out,
which started to form another component.
Thus, a second component of the support of the equilibrium measure $\lambda_2$
started to form, i.e. the support of the equilibrium measure $\lambda_2$
started breaking down on two arcs.
}
\label{Fig_ang3(9)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_10}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points), $Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.45))(z-(1.0+i\cdot0.45))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.45))(z-(.65-i\cdot0.45))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
have reached the second (lower) compact $F_2$. The second compact $F_2$ has started to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$) several points stand out,
which started to form another component.
Thus, a second component of the support of the equilibrium measure $\lambda_2$
started to form, i.e. the support of the equilibrium measure $\lambda_2$
started breaking down on two arcs.
The third extremal compact $F_0$, as before, ``seeks'' to separate
the other two compacts from each other, but now each of the compacts $F_1$ and $F_2$
has by two components. It is clearly seen, that the compact $F_0$
now crosses the compact $F_2$.
}
\label{Fig_ang3(9)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_11}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.4))(z-(1.0+i\cdot0.4))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.4))(z-(.65-i\cdot0.4))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
cross the second compact $F_2$. The second compact $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues.
}
\label{Fig_ang3(14)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_12}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials
$Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.4))(z-(1.0+i\cdot0.4))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.4))(z-(.65-i\cdot0.4))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
cross the second compact $F_2$. The second compact $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues.
The third extremal compact $F_0$ crosses the compact $F_2$. As before,
it ``seeks'' to separate the other two compacts $F_1$ and $F_2$ from each other,
but now it ``fights'' with two components.
It is clearly seen, that at the junction of the red, black and blue points
appear two ``equilateral'' triangles with multicolored vertexes.
}
\label{Fig_ang3(14)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_13}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.3))(z-(1.0+i\cdot0.3))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.3))(z-(.65-i\cdot0.3))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
even further cross the second compact $F_2$. The second compact $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (even than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues.
}
\label{Fig_ang3(11)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_14}}
\vskip-10mm
\caption{The distribution of the Hermite-Pad\'{e} polynomials
$Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points) and $Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.3))(z-(1.0+i\cdot0.3))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.3))(z-(.65-i\cdot0.3))}$.
It is clearly seen, that under this position of the pair of branch points,
the two arcs, which are the result of the breaking of the support of the measure $\lambda_1$,
even further cross the second compact $F_2$. The second compact $F_2$ continues to change:
from the total set of black points (zeros of the polynomial $Q_{180,2}$)
even more points stand out (even than before), which form the second component of $F_2$.
Thus, the forming of the second component of the support of the equilibrium measure
$\lambda_2$ continues. It is clearly seen, that at the junction of the red, black and blue points
appear two ``equilateral'' triangles with multicolored vertexes.
By analogy with classical Pad\'{e} approximants and two-point Pad\'{e} approximants
(see fig. \ref{Fig_pade10_2500_130_blu} and \ref{Fig_bus210b_4000_120_full})
it is natural to assume, that the center of each triangle
has a Chebotarev point $v_1$, $v_2$ with zero density.
At the branch points $a_j,b_j$ the density of the measures $\lambda_1$
and $\lambda_2$ are proportional to $|z-a_j|^{-1/2},|z-b_j|^{-1/2}$, $j=1,2$,
respectively. There is a Froissart singlet (blue) on the imaginary axis.
}
\label{Fig_ang3(11)_5000_180_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_15}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials $Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.15))(z-(1.0+i\cdot0.15))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.15))(z-(.65-i\cdot0.15))}$.
It is clearly seen, that under this position of the pair of branch points,
the support $F_2$ of the equilibrium measure $\lambda_2$ is separated
on two practically equivalent arcs.
However, according to the distribution of the zeros of the polynomial $Q_{180,2}$,
the density $\lambda_2'$ of the equilibrium measures of each arc must be different.
On the upper arc it behaves like a Chebyshev measure, that is at the end points $a_2,b_2$
the density is proportional to $|z-a_2|^{-1/2}$ and $|z-b_2|^{-1/2}$, respectively.
The end points of the lower arc $v_1,v_2$ are the Chebotarev points
and their density is proportional to $|z-v_1|^{1/2}$ and $|z-v_2|^{1/2}$.
}
\label{Fig_ang3(13)_5000_180_rdbk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_4_16}}
\caption{The distribution of the Hermite-Pad\'{e} polynomials
$Q_{180,0}$ (blue points),
$Q_{180,1}$ (red points),
$Q_{180,2}$ (black points) for two functions
$f_1(z)=1/\sqrt{(z-(-1.0+i\cdot0.15))(z-(1.0+i\cdot0.15))}$,
$f_2(z)=1/\sqrt{(z-(-.65-i\cdot0.15))(z-(.65-i\cdot0.15))}$.
It is clearly seen, that under this position of the pair of branch points,
the support $F_2$ of the equilibrium measure $\lambda_2$ is separated
on two practically equivalent arcs.
The branch points $a_2,b_2$ are on the upper arc,
the end points of the lower arc $v_1,v_2$ are the Chebotarev points
and their density is proportional to $|z-v_1|^{1/2}$ and $|z-v_2|^{1/2}$.
At the junction of the red, black and blue points
appeared two ``equilateral'' triangles with multicolored vertexes,
and the center of each has a Chebotarev point $v_1,v_2$.
}
\label{Fig_ang3(13)_5000_180_full}
\end{figure}
\clearpage
\markboth{\bf Case Two, Nikishin system}{\bf Case Two, Nikishin system}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_1}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$Q_{n,1}$ (red points), $Q_{n,2}$ (black points), $n=61,\dots,80$,
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis. There is a pair of Froissart triplets
inside these two domans for some $n\in\{61,\dots,80\}$,
there is a pair of Froissart singlets
in the complementary domains for some $n\in\{61,\dots,80\}$,
there is one Froissart doublet on the negative part of the imaginary axis.
This follows from the analysis of the next figures \ref{Fig_nik_(1_6)_2000_61-80_rd},
\ref{Fig_nik_(1_6)_2000_61-80_bl}, and \ref{Fig_nik_(1_6)_2000_61-80_bk}.
}
\label{Fig_nik_(1_6)_2000_61-80_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_2}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points),
$n=61,\dots,80$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_2000_61-80_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_3}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=61,\dots,80$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_2000_61-80_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_4}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=61,\dots,80$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{n,2}$, $n=61,\dots,80$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 1.6$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(1_6)_2000_61-80_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_5}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$Q_{n,1}$ (red points), $Q_{n,2}$ (black points), $n=121,\dots,130$,
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis. There is a pair of Froissart triplets
inside these two domans for some $n\in\{121,\dots,130\}$,
there is a pair of Froissart singlets
in the complementary domains for some $n\in\{121,\dots,130\}$,
there is a Froissart doublet on the negative part of the imaginary axis.
This follows from the analysis of the next figures \ref{Fig_nik_(1_6)_3000_121-130_rd},
\ref{Fig_nik_(1_6)_3000_121-130_bl} and \ref{Fig_nik_(1_6)_3000_121-130_bk}.
}
\label{Fig_nik_(1_6)_3000_121-130_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_6}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_3000_121-130_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_7}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_3000_121-130_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_8}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{n,2}$, $n=121,\dots,130$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 1.6$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(1_6)_3000_121-130_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_9}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{165,0}$ (blue points),
$Q_{165,1}$ (red points), $Q_{165,2}$ (black points) when $n=165$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis. There are no Froissart zeros when $n=165$.
}
\label{Fig_nik_(1_6)_5000_165_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_10}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points),
$n=121,\dots,130$, in the plane $\CC_z$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There are no Froissart zeros when $n=165$.
}
\label{Fig_nik_(1_6)_5000_165_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_11}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=121,\dots,130$, in the plane $\CC_z$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There are no Froissart zeros when $n=165$.
}
\label{Fig_nik_(1_6)_5000_165_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_12}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=121,\dots,130$, in the plane $\CC_z$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{n,2}$, $n=121,\dots,130$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 1.6$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
There are no Froissart zeros when $n=165$.
}
\label{Fig_nik_(1_6)_5000_165_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_13}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$Q_{n,1}$ (red points), $Q_{n,2}$ (black points), $n=127,128$,
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis. There is a pair of Froissart triplets
inside these two domans for some $n\in\{127,128\}$,
there is a pair of Froissart singlets
in the complementary domains for some $n\in\{127,128\}$,
there is one Froissart doublet on the negative part of the imaginary axis.
This follows from the analysis of the next figures \ref{Fig_nik_(1_6)_3000_127-128_rd},
\ref{Fig_nik_(1_6)_3000_127-128_bl} and \ref{Fig_nik_(1_6)_3000_127-128_bk}.
}
\label{Fig_nik_(1_6)_3000_127-128_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_14}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points),
$n=127,128$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_3000_127-128_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_15}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=127,128$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
}
\label{Fig_nik_(1_6)_3000_127-128_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_5_16}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=127,128$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 1.6z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{n,2}$, $n=127,128$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 1.6$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(1_6)_3000_127-128_bk}
\end{figure}
\clearpage
\markboth{\bf $Q_{n,0}\cdot 1+Q_{n,1}f+Q_{n,2}f^2$}{\bf $Q_{n,0}\cdot 1+Q_{n,1}f+Q_{n,2}f^2$}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_1}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$Q_{n,1}$ (red points), $Q_{n,2}$ (black points), $n=121,\dots,130$,
for a set of three functions $[1,f,f^2]$, where $f$ is a ``perturbed'' function:
$f(z)=(1-z^2)^{1/4}(1-(0.1+i\sqrt{3}\cdot 1.6)z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis. There is a pair of Froissart triplets
inside these two domans for some $n\in\{121,\dots,130\}$,
there is a pair of Froissart singlets
in the complementary domains for some $n\in\{121,\dots,130\}$,
there is one Froissart doublet on the negative part of the imaginary axis.
This follows from the analysis of the next figures \ref{Fig_nik_(1_6eps)_3000_121-130_rd},
\ref{Fig_nik_(1_6eps)_3000_121-130_bl} and \ref{Fig_nik_(1_6eps)_3000_121-130_bk}.
It is clearly seen, that the picture of the distribution of the zeros
is entirely the same compared to the unperturbed case.
}
\label{Fig_nik_(1_6eps)_3000_121-130_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_2}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f$ is a ``perturbed'' function:
$f(z)=(1-z^2)^{1/4}(1-(0.1+i\sqrt{3}\cdot 1.6)z)^{-1/2}$.
}
\label{Fig_nik_(1_6eps)_3000_121-130_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_3}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f$ is a ``perturbed'' function:
$f(z)=(1-z^2)^{1/4}(1-(0.1+i\sqrt{3}\cdot1.6)z)^{-1/2}$.
}
\label{Fig_nik_(1_6eps)_3000_121-130_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_4}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=121,\dots,130$, for a set of three functions $[1,f,f^2]$, where
$f$ is a ``perturbed'' function:
$f(z)=(1-z^2)^{1/4}(1-(0.1+i\sqrt{3}\cdot1.6)z)^{-1/2}$.
There are a simple zeros of the polynomials $Q_{n,2}$, $n=121,\dots,130$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=0.1+i\sqrt{3}\cdot1.6$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(1_6eps)_3000_121-130_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_5}}
\vskip-6mm
\caption{
The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$Q_{n,1}$ (red points), $Q_{n,2}$ (black points) when $n=166,\dots,170$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis.
There is a pair of Froissart triplets
inside these two domans for two $n$ of $n\in\{166,\dots,170\}$.
This follows from the analysis of the next figures \ref{Fig_nik_(_9)_4000_166-170_rd},
\ref{Fig_nik_(_9)_4000_166-170_bl}, \ref{Fig_nik_(_9)_4000_166-170_bk}.
There are no Froissart points in the complementary domains for $n=166,\dots,170$.
}
\label{Fig_nik_(_9)_4000_166-170_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_6}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,1}$ (red points)
$n=166,\dots,170$ for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
}
\label{Fig_nik_(_9)_4000_166-170_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_7}}
\vskip-6mm
\caption{
The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,0}$ (blue points),
$n=166,\dots,170$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
}
\label{Fig_nik_(_9)_4000_166-170_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_8}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{n,2}$ (black points),
$n=166,\dots,170$, for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i\sqrt{3}\cdot 0.9z)^{-1/2}$.
There are simple zeros of the polynomials $Q_{n,2}$, $n=166,\dots,170$,
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 0.9$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(_9)_4000_166-170_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_9}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{187,0}$ (blue points),
$Q_{187,1}$ (red points), $Q_{187,2}$ (black points) when $n=187$
for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
The Riemann sphere is decomposed into 3 domains by the zeros of the Hermite-Pad\'{e} polynomials,
one of them contains the infinity point, while the other two are symmetrical with respect
to the imaginary axis.
There are no Froissart points when $n=187$.
}
\label{Fig_nik_(_9)_5000_187_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_10}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{187,1}$ (red points),
when $n=187$ for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
}
\label{Fig_nik_(_9)_5000_187_rd}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_11}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{187,0}$ (blue points),
when $n=187$ for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i \sqrt{3}\cdot 0.9z)^{-1/2}$.
}
\label{Fig_nik_(_9)_5000_187_bl}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_12}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomial $Q_{187,2}$ (black points),
when $n=187$ for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i\sqrt{3}\cdot 0.9z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{187,2}$
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 0.9$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(_9)_5000_187_bk}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_6_13}}
\vskip-6mm
\caption{The distribution of the zeros of the Hermite-Pad\'{e} polynomials $Q_{180,0},Q_{180,2}$
(blue and black points),
when $n=180$ for a set of three functions $[1,f,f^2]$, where
$f(z)=(1-z^2)^{1/4}(1-i\sqrt{3}\cdot 0.9z)^{-1/2}$.
There is a simple zero of the polynomial $Q_{180,2}$
on the positive part of the imaginary axis at the point $z\approx a$, $a=i\sqrt{3}\cdot 0.9$,
corresponding to a simple pole of the function $f^2$ at the point $z=ia$.
}
\label{Fig_nik_(_9)_5000_180_blbk}
\end{figure}
\clearpage
\markboth{\bf Buslaev compact set for $f(z)=\sqrt{(z-a)/(z-b)}$}{\bf Buslaev
compact set for $f(z)=\sqrt{(z-a)/(z-b)}$}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_7_1}}
\vskip-6mm
\caption{The distribution of zeros and poles of the two-point Pad\'{e} approximant
(at points $z=0$ and $z=\infty$) of the function
$f(z)=\sqrt{(z-a)/(z-b)}$, $a=.5$, $b=2$.
Here are selected two ``same'' branches of the function $f$:
$f_0=\sqrt{(z-.5)/(z-2)}$, $f_\infty=\sqrt{(z-.5)/(z-2)}$.
The interval $[1/2,2]$ is the Buslaev compact.
}
\label{Fig_bus210a_3000_90_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_7_2}}
\vskip-6mm
\caption{The distribution of zeros and poles of the two-point Pad\'{e} approximant
(at points $z=0$ and $z=\infty$) of the function
$f(z)=\sqrt{(z-a)/(z-b)}$, $a=.5$, $b=2$.
Here are selected two ``different'' branches of the function $f$:
$f_0(z)=\sqrt{(z-.5)/(z-2)}$, $f_\infty(z)=-\sqrt{(z-.5)/(z-2)}$.
There is one Chebotarev point $v=\sqrt{ab}=1$ of zero density
on the Buslaev compact.
From this can be seen, that this point cannot be calculated
using of the two-point Pad\'{e} approximants, comp. \ref{Fig_pade10_2500_130_blu};
contrary to the case with classical Pad\'{e} approximants
(see fig. \ref{Fig_pade10_2500_130_red}, \ref{Fig_pade10_2500_130})
}
\label{Fig_bus210b_4000_120_full}
\end{figure}
\newpage
\begin{figure}[!ht]
\centerline{
\includegraphics[width=15cm,height=15cm]{FIG_7_3}}
\vskip-6mm
\caption{The distribution of zeros and poles of the two-point Pad\'{e} approximant
of order $[120/120]$ of the function $f(z)=(z-a_1)^\alpha(z-a_2)^{-\alpha}$, $\alpha=1/4$,
$a_1=0.9-i\cdot1.1$ ,$a_2=0.1+i\cdot0.2$,
when two ``quite different'' branches are selected,
$f_0=\sqrt[4]{(z-a_1)/(z-a_2)}$, $f_\infty=-\sqrt[4]{(z-a_1)/(z-a_2)}$.
The zeros (blue points) and the poles (red points) create a Buslaev compact.
}
\label{Fig_bus205c_4000_120_full}
\end{figure}
\clearpage
\markboth{\bf References}{\bf References}
\newpage
|
\section*{About this document}\addcontentsline{toc}{section}{About this document}
This document is a package vignette for the \pkg{ggRandomForests} package for ``Visually Exploring Random Forests'' (\url{http://CRAN.R-project.org/package=ggRandomForests}). The \pkg{ggRandomForests} package is designed for use with the \pkg{randomForestSRC} package~\citep[\url{http://CRAN.R-project.org/package=randomForestSRC}]{Ishwaran:RFSRC:2014} for growing random forests for survival (time to event response), regression (continuous response) and classification (categorical response) settings and uses the \pkg{ggplot2} package~\citep[\url{http://CRAN.R-project.org/package=ggplot2}]{Wickham:2009} for plotting diagnostic and variable association results. \pkg{ggRandomForests} is structured to extract data objects from \pkg{randomForestSRC} objects and provides functions for printing and plotting these objects.
The vignette is a tutorial for using the \pkg{ggRandomForests} package with the \pkg{randomForestSRC} package for building and post-processing random forests for regression settings. In this tutorial, we explore a random forest for regression model constructed for the Boston housing data set~\citep{Harrison:1978,Belsley:1980}, available in the \pkg{MASS} package~\citep{mass:2002}. We grow a random forest and demonstrate how \pkg{ggRandomForests} can be used when determining how the response depends on predictive variables within the model. The tutorial demonstrates the design and usage of many of \pkg{ggRandomForests} functions and features and also how to modify and customize the resulting \code{ggplot} graphic objects along the way.
The vignette is written in \LaTeX using the \pkg{knitr} package~\citep[\url{http://CRAN.R-project.org/package=knitr}]{Xie:2015, Xie:2014,Xie:2013}, which facilitates weaving \proglang{R}~\citep{rcore} code, results and figures into document text. Throughout this document, \proglang{R} code will be displayed in \emph{code blocks} as shown below. This code block loads the \proglang{R} packages required to run the source code listed in code blocks throughout the remainder of this document.
\begin{Schunk}
\begin{Sinput}
R> ################## Load packages ##################
R> library("ggplot2") # Graphics engine
R> library("RColorBrewer") # Nice color palettes
R> library("plot3D") # for 3d surfaces.
R> library("dplyr") # Better data manipulations
R> library("parallel") # mclapply for multicore processing
R>
R> # Analysis packages.
R> library("randomForestSRC") # random forests for survival, regression and
R> # classification
R> library("ggRandomForests") # ggplot2 random forest figures (This!)
R>
R> ################ Default Settings ##################
R> theme_set(theme_bw()) # A ggplot2 theme with white background
R>
R> ## Set open circle for censored, and x for events
R> event.marks <- c(1, 4)
R> event.labels <- c(FALSE, TRUE)
R>
R> ## We want red for death events, so reorder this set.
R> strCol <- brewer.pal(3, "Set1")[c(2,1,3)]
\end{Sinput}
\end{Schunk}
This vignette is available within the \pkg{ggRandomForests} package on the Comprehensive R Archive Network (CRAN)~\citep[\url{http://cran.r-project.org}]{rcore}. Once the package has been installed, the vignette can be viewed directly from within \proglang{R} with the following command:
\begin{Schunk}
\begin{Sinput}
R> vignette("randomForestSRC-Regression", package = "ggRandomForests")
\end{Sinput}
\end{Schunk}
A development version of the \pkg{ggRandomForests} package is also available on GitHub (\url{https://github.com}). We invite comments, feature requests and bug reports for this package at \url{https://github.com/ehrlinger/ggRandomForests}.
\section{Introduction} \label{S:introduction}
Random Forests~\citep{Breiman:2001} (RF) are a fully non-parametric statistical method which requires no distributional or functional assumptions on covariate relation to the response. RF is a robust, nonlinear technique that optimizes predictive accuracy by fitting an ensemble of trees to stabilize model estimates. Random Survival Forests (RSF)~\citep{Ishwaran:2007a,Ishwaran:2008} are an extension of Breiman's RF techniques to survival settings, allowing efficient non-parametric analysis of time to event data. The \pkg{randomForestSRC} package~\citep{Ishwaran:RFSRC:2014} is a unified treatment of Breiman's random forests for survival (time to event response), regression (continuous response) and classification (categorical response) problems.
Predictive accuracy make RF an attractive alternative to parametric models, though complexity and interpretability of the forest hinder wider application of the method. We introduce the \pkg{ggRandomForests} package for visually exploring random forest models. The \pkg{ggRandomForests} package is structured to extract intermediate data objects from \pkg{randomForestSRC} objects and generate figures using the \pkg{ggplot2} graphics package~\citep{Wickham:2009}.
Many of the figures created by the \pkg{ggRandomForests} package are also available directly from within the \pkg{randomForestSRC} package. However \pkg{ggRandomForests} offers the following advantages:
\begin{itemize}
\item Separation of data and figures: \pkg{ggRandomForests} contains functions that operate on either the \code{randomForestSRC::rfsrc} forest object directly, or on the output from \pkg{randomForestSRC} post processing functions (i.e. \code{plot.variable}, \code{var.select}, \code{find.interaction}) to generate intermediate \pkg{ggRandomForests} data objects. functions are provide to further process these objects and plot results using the \pkg{ggplot2} graphics package. Alternatively, users can use these data objects for their own custom plotting or analysis operations.
\item Each data object/figure is a single, self contained object. This allows simple modification and manipulation of the data or \pkg{ggplot2} objects to meet users specific needs and requirements.
\item The use of \pkg{ggplot2} for plotting. We chose to use the \pkg{ggplot2} package for our figures to allow users flexibility in modifying the figures to their liking. Each plot function returns either a single \code{ggplot} object, or a \code{list} of \code{ggplot} objects, allowing users to use additional \pkg{ggplot2} functions or themes to modify and customize the figures to their liking.
\end{itemize}
This document is formatted as a tutorial for using the \pkg{randomForestSRC} package for building and post-processing random forest models with the \pkg{ggRandomForests} package for investigating how the forest is constructed. In this tutorial, we use the Boston Housing Data (Section~\ref{S:data}), available in the \pkg{MASS} package~\citep{mass:2002}, to build a random forest for regression (Section~\ref{S:rfr}) and demonstrate the tools in the \pkg{ggRandomForests} package for examining the forest construction.
Random forests are not parsimonious, but use all variables available in the construction of a response predictor. We demonstrate a random forest variable selection (Section~\ref{S:varselection}) process using the Variable Importance (Section~\ref{S:vimp}) measure (VIMP)~\citep{Breiman:2001} as well as Minimal Depth (Section~\ref{S:minimaldepth})~\citep{Ishwaran:2010}, a property derived from the construction of each tree within the forest, to assess the impact of variables on forest prediction.
Once we have an idea of which variables we are want to investigate further, we will use variable dependence plots~\citep{Friedman:2000} to understand how a variable is related to the response (Section~\ref{S:dependence}). Marginal dependence plots (Section~\ref{S:variabledependence}) give us an idea of the overall trend of a variable/response relation, while partial dependence plots (Section~\ref{S:partialdependence}) show us a risk adjusted relation. These figures may show strongly non-linear variable/response relations that are not easily obtained through a parametric approach. We are also interested in examining variable interactions within the forest model (Section~\ref{S:interactions}). Using a minimal depth approach, we can quantify how closely variables are related within the forest, and generate marginal dependence (Section~\ref{S:coplots}) and partial dependence (Section!\ref{S:partialcoplots}) (risk adjusted) conditioning plots (coplots)~\citep{chambers:1992,cleveland:1993} to examine these interactions graphically.
\section{Data: Boston Housing Values} \label{S:data}
The Boston Housing data is a standard benchmark data set for regression models. It contains data for 506 census tracts of Boston from the 1970 census~\citep{Harrison:1978,Belsley:1980}. The data is available in multiple \proglang{R} packages, but to keep the installation dependencies for the \pkg{ggRandomForests} package down, we will use the data contained in the \pkg{MASS} package~\citep[\url{http://CRAN.R-project.org/package=MASS}]{mass:2002}, available with the base install of \proglang{R}. The following code block loads the data into the environment. We include a table of the Boston data set variable names, types and descriptions for reference when we interpret the model results.
\begin{Schunk}
\begin{Sinput}
R> # Load the Boston Housing data
R> data(Boston, package="MASS")
R>
R> # Set modes correctly. For binary variables: transform to logical
R> Boston$chas <- as.logical(Boston$chas)
\end{Sinput}
\end{Schunk}
\begin{table}
\caption{\code{Boston} housing data dictionary.}
\begin{tabular}{lll}
\toprule
Variable & Description & type\\
\midrule
crim & Crime rate by town. & numeric\\
zn & Proportion of residential land zoned for lots over 25,000 sq.ft. & numeric\\
indus & Proportion of non-retail business acres per town. & numeric\\
chas & Charles River (tract bounds river). & logical\\
nox & Nitrogen oxides concentration (10 ppm). & numeric\\
\addlinespace
rm & Number of rooms per dwelling. & numeric\\
age & Proportion of units built prior to 1940. & numeric\\
dis & Distances to Boston employment center. & numeric\\
rad & Accessibility to highways. & integer\\
tax & Property-tax rate per \$10,000. & numeric\\
\addlinespace
ptratio & Pupil-teacher ratio by town. & numeric\\
black & Proportion of blacks by town. & numeric\\
lstat & Lower status of the population (percent). & numeric\\
medv & Median value of homes (\$1000s). & numeric\\
\bottomrule
\end{tabular}
\end{table}
The main objective of the Boston Housing data is to investigate variables associated with predicting the median value of homes (continuous \code{medv} response) within 506 suburban areas of Boston.
\subsection{Exploratory Data Analysis} \label{S:eda}
It is good practice to view your data before beginning an analysis, what~\cite{Tukey:1977} refers to as Exploratory Data Analysis (EDA). To facilitate this, we use \pkg{ggplot2} figures with the \code{ggplot2::facet_wrap} command to create two sets of panel plots, one for categorical variables with boxplots at each level, and one of scatter plots for continuous variables. Each variable is plotted along a selected continuous variable on the X-axis. These figures help to find outliers, missing values and other data anomalies in each variable before getting deep into the analysis. We have also created a separate \pkg{shiny} app~\citep[\url{http://shiny.rstudio.com}]{shiny:2015}, available at \url{https://ehrlinger.shinyapps.io/xportEDA}, for creating similar figures with an arbitrary data set, to make the EDA process easier for users.
The Boston housing data consists almost entirely of continuous variables, with the exception of the ``Charles river'' logical variable. A simple EDA visualization to use for this data is a single panel plot of the continuous variables, with observation points colored by the logical variable. Missing values in our continuous variable plots are indicated by the rug marks along the x-axis, of which there are none in this data. We used the Boston housing response variable, the median value of homes (\code{medv}), for X variable.
\begin{Schunk}
\begin{Sinput}
R> # Use reshape2::melt to transform the data into long format.
R> dta <- melt(Boston, id.vars=c("medv","chas"))
R>
R> # plot panels for each covariate colored by the logical chas variable.
R> ggplot(dta, aes(x=medv, y=value, color=chas))+
+ geom_point(alpha=.4)+
+ geom_rug(data=dta
+ labs(y="", x=st.labs["medv"]) +
+ scale_color_brewer(palette="Set2")+
+ facet_wrap(~variable, scales="free_y", ncol=3)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-eda-1}
}
\caption[EDA variable plots]{EDA variable plots. Points indicate variable value against the median home value variable. Points are colored according to the chas variable.}\label{fig:eda}
\end{figure}
\end{Schunk}
This figure is loosely related to a pairs scatter plot~\citep{Becker:1988}, but in this case we only examine the relation between the response variable against the remainder. Plotting the data against the response also gives us a "sanity check" when viewing our model results. It's pretty obvious from this figure that we should find a strong relation between median home values and the \code{lstat} and \code{rm} variables.
\section{Random Forest - Regression} \label{S:rfr}
A Random Forest is grown by \emph{bagging}~\citep{Breiman:1996} a collection of \emph{classification and regression trees} (CART)~\citep{cart:1984}. The method uses a set of $B$ bootstrap~\citep{bootstrap:1994} samples, growing an independent tree model on each sub-sample of the population. Each tree is grown by recursively partitioning the population based on optimization of a \emph{split rule} over the $p$-dimensional covariate space. At each split, a subset of $m \le p$ candidate variables are tested for the split rule optimization, dividing each node into two daughter nodes. Each daughter node is then split again until the process reaches the \emph{stopping criteria} of either \emph{node purity} or \emph{node member size}, which defines the set of \emph{terminal (unsplit) nodes} for the tree. In regression trees, the split rule is based on minimizing the mean squared error, whereas in classification problems, the Gini index is used~\citep{Friedman:2000}.
Random Forests sort each training set observation into one unique terminal node per tree. Tree estimates for each observation are constructed at each terminal node, among the terminal node members. The Random Forest estimate for each observation is then calculated by aggregating, averaging (regression) or votes (classification), the terminal node results across the collection of $B$ trees.
For this tutorial, we grow the random forest for regression using the \code{rfsrc} command to predict the median home value (\code{medv} variable) using the remaining 13 independent predictor variables. For this example we will use the default set of $B=1000$ trees (\code{ntree} argument), $m=5$ candidate variables (\code{mtry}) for each split with a stopping criteria of at most \code{nodesize=5} observations within each terminal node.
Because growing random forests are computationally expensive, and the \pkg{ggRandomForests} package is targeted at the visualization of random forest objects, we will use cached copies of the \pkg{randomForestSRC} objects throughout this document. We include the cached objects as data sets in the \pkg{ggRandomForests} package. The actual \code{rfsrc} calls are included in comments within code blocks.
\begin{Schunk}
\begin{Sinput}
R> # Load the data, from the call:
R> # rfsrc_Boston <- rfsrc(medv~., data=Boston)
R> data(rfsrc_Boston)
R>
R> # print the forest summary
R> rfsrc_Boston
\end{Sinput}
\begin{Soutput}
Sample size: 506
Number of trees: 1000
Minimum terminal node size: 5
Average no. of terminal nodes: 79.911
No. of variables tried at each split: 5
Total no. of variables: 13
Analysis: RF-R
Family: regr
Splitting rule: regr
Error rate: 11.94
\end{Soutput}
\end{Schunk}
The \code{randomForestSRC::print.rfsrc} summary details the parameters used for the \code{rfsrc} call described above, and returns variance and generalization error estimate from the forest training set. The forest is built from 506 observations and 13 independent variables. It was constructed for the continuous \code{medv} variable using \code{ntree=1000} regression (\code{regr}) trees, randomly selecting 5 candidate variables at each node split, and terminating nodes with no fewer than 5 observations.
\subsection{Generalization error estimates} \label{S:error}
One advantage of Random Forests is a built in generalization error estimate. Each bootstrap sample selects approximately 63.2\% of the population on average. The remaining 36.8\% of observations, the Out-of-Bag (OOB)~\citep{BreimanOOB:1996e} sample, can be used as a hold out test set for each of the trees in the forest. An OOB prediction error estimate can be calculated for each observation by predicting the response over the set of trees which were NOT trained with that particular observation. The Out-of-Bag prediction error estimates have been shown to be nearly identical to n--fold cross validation estimates~\citep{StatisticalLearning:2009}. This feature of Random Forests allows us to obtain both model fit and validation in one pass of the algorithm.
The \code{gg_error} function operates on the \code{randomForestSRC::rfsrc} object to extract the error estimates as the forest is grown. The code block demonstrates part the \pkg{ggRandomForests} design philosophy, to create separate data objects and provide functions to operate on the data objects. The following code block first creates a \code{gg_error} object, then uses the \code{plot.gg_error} function to create a \code{ggplot} object for display.
\begin{Schunk}
\begin{Sinput}
R> # Plot the OOB errors against the growth of the forest.
R> gg_e <- gg_error(rfsrc_Boston)
R> plot(gg_e)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-error-1}
}
\caption[Random forest generalization error]{Random forest generalization error. OOB error convergence along the number of trees in the forest.}\label{fig:error}
\end{figure}
\end{Schunk}
This figure demonstrates that it does not take a large number of trees to stabilize the forest prediction error estimate. However, to ensure that each variable has enough of a chance to be included in the forest prediction process, we do want to create a rather large random forest of trees.
\subsection{Random Forest Prediction} \label{S:rf-predict}
The \code{gg_rfsrc} function extracts the OOB prediction estimates from the random forest. This code block executes the the data extraction and plotting in one line, since we are not interested in holding the prediction estimates for later reuse. Also note that we add in the additional \pkg{ggplot2} command (\code{coord_cartesian}) to modify the plot object. Each of the \pkg{ggRandomForests} plot commands return \code{ggplot} objects, which we can also store for modification or reuse later in the analysis.
\begin{Schunk}
\begin{Sinput}
R> # Plot predicted median home values.
R> plot(gg_rfsrc(rfsrc_Boston), alpha=.5)+
+ coord_cartesian(ylim=c(5,49))
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-rfsrc-1}
}
\caption[OOB predicted median home values]{OOB predicted median home values. Points are jittered to help visualize predictions for each observation. Boxplot indicates the distribution of the predicted values.}\label{fig:rfsrc}
\end{figure}
\end{Schunk}
The \code{gg_rfsrc} plot shows the predicted median home value, one point for each observation in the training set. The points are jittered around a single point on the x-axis, since we are only looking at predicted values from the forest. These estimates are Out of Bag, which are analogous to test set estimates. The boxplot is shown to give an indication of the distribution of the prediction estimates. For this analysis the figure is another model sanity check, as we are more interested in exploring the ``why'' questions for these predictions.
\section{Variable Selection} \label{S:varselection}
Random forests are not parsimonious, but use all variables available in the construction of a response predictor. Also, unlike parametric models, Random Forests do not require the explicit specification of the functional form of covariates to the response. Therefore there is no explicit p-value/significance test for variable selection with a random forest model. Instead, RF ascertain which variables contribute to the prediction through the split rule optimization, optimally choosing variables which separate observations. We use two separate approaches to explore the RF selection process, Variable Importance (Section~\ref{S:vimp}) and Minimal Depth (Section~\ref{S:minimaldepth}).
\subsection{Variable Importance.} \label{S:vimp}
\emph{Variable importance} (VIMP) was originally defined in CART using a measure involving surrogate variables (see Chapter 5 of~\cite{cart:1984}). The most popular VIMP method uses a prediction error approach involving ``noising-up'' each variable in turn. VIMP for a variable $x_v$ is the difference between prediction error when $x_v$ is noised up by randomly permuting its values, compared to prediction error under the observed values~\citep{Breiman:2001, Liaw:2002, Ishwaran:2007, Ishwaran:2008}.
Since VIMP is the difference between OOB prediction error before and after permutation, a large VIMP value indicates that misspecification detracts from the variable predictive accuracy in the forest. VIMP close to zero indicates the variable contributes nothing to predictive accuracy, and negative values indicate the predictive accuracy \emph{improves} when the variable is mispecified. In the later case, we assume noise is more informative than the true variable. As such, we ignore variables with negative and near zero values of VIMP, relying on large positive values to indicate that the predictive power of the forest is dependent on those variables.
The \code{gg_vimp} function extracts VIMP measures for each of the variables used to grow the forest. The \code{plot.gg_vimp} function shows the variables, in VIMP rank order, from the largest (Lower Status) at the top, to smallest (Charles River) at the bottom. VIMP measures are shown using bars to compare the scale of the error increase under permutation.
\begin{Schunk}
\begin{Sinput}
R> # Plot the VIMP rankings of independent variables.
R> plot(gg_vimp(rfsrc_Boston), lbls=st.labs)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-vimp-1}
}
\caption[Random forest VIMP plot]{Random forest VIMP plot. Bars are colored by sign of VIMP, longer blue bars indicate more important variables.}\label{fig:vimp}
\end{figure}
\end{Schunk}
For our random forest, the top two variables (\code{lstat} and \code{rm}) have the largest VIMP, with a sizable difference to the remaining variables, which mostly have similar VIMP measure. This indicates we should focus attention on these two variables, at least, over the others.
In this example, all VIMP measures are positive, though some are small. When there are both negative and positive VIMP values, the \code{plot.gg_vimp} function will color VIMP by the sign of the measure. We use the \code{lbls} argument to pass a named \code{vector} of meaningful text descriptions to the \code{plot.gg_vimp} function, replacing the often terse variable names used by default.
\subsection{Minimal Depth.} \label{S:minimaldepth}
In VIMP, prognostic risk factors are determined by testing the forest prediction under alternative data settings, ranking the most important variables according to their impact on predictive ability of the forest. An alternative method uses inspection of the forest construction to rank variables. \emph{Minimal depth} assumes that variables with high impact on the prediction are those that most frequently split nodes nearest to the trunks of the trees (i.e. at the root node) where they partition large samples of the population.
Within a tree, node levels are numbered based on their relative distance to the trunk of the tree (with the root at 0). Minimal depth measures the important risk factors by averaging the depth of the first split for each variable over all trees within the forest. Lower values of this measure indicate variables important in splitting large groups of patients.
The \emph{maximal subtree} for a variable $x$ is the largest subtree whose root node splits on $x$. All parent nodes of $x$'s maximal subtree have nodes that split on variables other than $x$. The largest maximal subtree possible is at the root node. If a variable does not split the root node, it can have more than one maximal subtree, or a maximal subtree may also not exist if there are no splits on the variable. The minimal depth of a variables is a surrogate measure of predictiveness of the variable. The smaller the minimal depth, the more impact the variable has sorting observations, and therefore on the forest prediction.
The \code{gg_minimal_depth} function is analogous to the \code{gg_vimp} function for minimal depth. Variables are ranked from most important at the top (minimal depth measure), to least at the bottom (maximal minimal depth). The vertical dashed line indicates the minimal depth threshold where smaller minimal depth values indicate higher importance and larger indicate lower importance.
The \code{randomForestSRC::var.select} call is again a computationally intensive function, as it traverses the forest finding the maximal subtree within each tree for each variable before averaging the results we use in the \code{gg_minimal_depth} call. We again use the cached object strategy here to save computational time. The \code{var.select} call is included in the comment of this code block.
\begin{Schunk}
\begin{Sinput}
R> # Load the data, from the call:
R> # varsel_Boston <- var.select(rfsrc_Boston)
R> data(varsel_Boston)
R>
R> # Save the gg_minimal_depth object for later use.
R> gg_md <- gg_minimal_depth(varsel_Boston)
R>
R> # plot the object
R> plot(gg_md, lbls=st.labs)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-minimaldepth-1}
}
\caption[Minimal Depth variables in rank order, most important at the top]{Minimal Depth variables in rank order, most important at the top. Vertical dashed line indicates the maximal minimal depth for important variables.}\label{fig:minimaldepth}
\end{figure}
\end{Schunk}
In general, the selection of variables according to VIMP is to rather arbitrarily examine the values, looking for some point along the ranking where there is a large difference in VIMP measures. The minimal depth threshold method has a more quantitative approach to determine a selection threshold. Given minimal depth is a quantitative property of the forest construction, \cite{Ishwaran:2010} also construct an analytic threshold for evidence of variable impact. A simple optimistic threshold rule uses the mean of the minimal depth distribution, classifying variables with minimal depth lower than this threshold as important in forest prediction. The minimal depth plot for our model indicates there are ten variables which have a higher impact (minimal depth below the mean value threshold) than the remaining three.
Since the VIMP and Minimal Depth measures use different criteria, we expect the variable ranking to be somewhat different. We use \code{gg_minimal_vimp} function to compare rankings between minimal depth and VIMP. In this call, we plot the stored \code{gg_minimal_depth} object (\code{gg_md}), which would be equivalent to calling \code{plot.gg_minimal_vimp(varsel_Boston)} or \code{plot(gg_minimal_vimp(varsel_Boston))}.
\begin{Schunk}
\begin{Sinput}
R> # gg_minimal_depth objects contain information about
R> # both minimal depth and VIMP.
R> plot.gg_minimal_vimp(gg_md)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-minimalvimp-1}
}
\caption[Comparing Minimal Depth and Vimp rankings]{Comparing Minimal Depth and Vimp rankings. Points on the red dashed line are ranked equivalently, points below have higher VIMP, those above have higher minimal depth ranking. Variables are colored by the sign of the VIMP measure.}\label{fig:minimalvimp}
\end{figure}
\end{Schunk}
The points along the red dashed line indicates where the measures are in agreement. Points above the red dashed line are ranked higher by VIMP than by minimal depth, indicating the variables are sensitive to misspecification. Those below the line have a higher minimal depth ranking, indicating they are better at dividing large portions of the population. The further the points are from the line, the more the discrepancy between measures. The construction of this figure is skewed towards a minimal depth approach, by ranking variables along the y-axis, though points are colored by the sign of VIMP.
In our example, both minimal depth and VIMP indicate the strong relation of \code{lstat} and \code{rm} variables to the forest prediction, which agrees with our expectation from the EDA (Section~\ref{S:eda}) done at the beginning of this document. We now turn to investigating how these, and other variables, are related to the predicted response.
\section{Response/Variable Dependence.} \label{S:dependence}
As random forests are not a parsimonious methodology, we can use the minimal depth and VIMP measures to reduce the number of variables we need to examine to a manageable subset. We would like to know how the forest response depends on some specific variables of interest. We often choose to examine variables of interest based on the study question, or other previous knowledge. In the absence of this, we will look at variables that contribute most to the predictive accuracy of the forest.
Although often characterized as a ``black box'' method, it is possible to express a random forest in functional form. In the end the forest predictor is some function, although complex, of the predictor variables $$\hat{f}_{rf} = f(x).$$ We use graphical methods to examine the forest predicted response dependency on covariates. We again have two options, variable dependence (Section~\ref{S:variabledependence}) plots are quick and easy to generate, and partial dependence (Section~\ref{S:partialdependence}) plots are computationally intensive but give us a risk adjusted look at the dependence.
\subsection{Variable Dependence} \label{S:variabledependence}
Variable dependence plots show the predicted response as a function of a covariate of interest, where each observation is represented by a point on the plot. Each predicted point is an individual observations, dependent on the full combination of all other covariates, not only on the covariate of interest. Interpretation of variable dependence plots can only be in general terms, as point predictions are a function of all covariates in that particular observation. However, variable dependence is straight forward to calculate, only requiring the predicted response for each observation.
We use the \code{gg_variable} function call to extract the training set variables and the predicted OOB response from \code{randomForestSRC::rfsrc} and \code{randomForestSRC::predict} objects. In the following code block, we will store the \code{gg_variable} data object for later use, as all remaining variable dependence plots can be constructed from this (\code{gg_v}) object. We will also use the minimal depth selected variables (minimal depth lower than the threshold value) from the previously stored \code{gg_minimal_depth} object (\code{gg_md$topvars}) to filter the variables of interest.
The \code{plot.gg_variable} function call operates in the \code{gg_variable} object. We pass it the list of variables of interest (\code{xvar}) and request a single panel (\code{panel=TRUE}) to display the figures. By default, the \code{plot.gg_variable} function returns a list of \code{ggplot} objects, one figure for each variable named in \code{xvar} argument. The next three arguments are passed to internal \code{ggplot} plotting routines. The \code{se} and \code{span} arguments are used to modify the internal call to \code{ggplot2::geom_smooth} for fitting smooth lines to the data. The \code{alpha} argument lightens the coloring points in the \code{ggplot2::geom_point} call, making it easier to see point over plotting. We also demonstrate modification of the plot labels using the \code{ggplot2::labs} function.
\begin{Schunk}
\begin{Sinput}
R> # Create the variable dependence object from the random forest
R> gg_v <- gg_variable(rfsrc_Boston)
R>
R> # We want the top ranked minimal depth variables only,
R> # plotted in minimal depth rank order.
R> xvar <- gg_md$topvars
R>
R> # plot the variable list in a single panel plot
R> plot(gg_v, xvar=xvar, panel=TRUE,
+ se=.95, span=1.2, alpha=.4)+
+ labs(y=st.labs["medv"], x="")
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-variable-1}
}
\caption[Variable dependence plot]{Variable dependence plot. Individual case predictions are marked with points. Loess smooth curve indicates the trend as the variables increase with shaded 95\% confidence band.}\label{fig:variable}
\end{figure}
\end{Schunk}
This figure looks very similar to the EDA (Section~\ref{S:eda}) figure, although with transposed axis as we plot the response variable on the y-axis. The closer the panels match, the better the RF prediction. The panels are sorted to match the order of variables in the \code{xvar} argument and include a smooth loess line~\citep{cleveland:1981, cleveland:1988}, with 95\% shaded confidence band, to indicates the trend of the prediction dependence over the covariate values.
There is not a convenient method to panel scatter plots and boxplots together, so we recommend creating panel plots for each variable type separately. The Boston housing data does contain a single categorical variable, the Charles river logical variable. Variable dependence plots for categorical variables are constructed using boxplots to show the distribution of the predictions within each category. Although the Charles river variable has the lowest importance scores in both VIMP and minimal depth measures, we include the variable dependence plot as an example of categorical variable dependence.
\begin{Schunk}
\begin{Sinput}
R> plot(gg_v, xvar="chas", points=FALSE,
+ se=FALSE, notch=TRUE, alpha=.4)+
+ labs(y=st.labs["medv"])
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-chas-1}
}
\caption[Variable dependence for Charles River logical variable]{Variable dependence for Charles River logical variable.}\label{fig:chas}
\end{figure}
\end{Schunk}
The figure shows that most housing tracts do not border the Charles river (\code{chas=FALSE}), and comparing the distributions of the predicted median housing values indicates no significant difference in home values. This reinforces the findings in both VIMP and Minimal depth, the Charles river variable has very little impact on the forest prediction of median home values.
\subsection{Partial Dependence.} \label{S:partialdependence}
Partial variable dependence plots are a risk adjusted alternative to variable dependence. Partial plots are generated by integrating out the effects of all variables beside the covariate of interest. Partial dependence data are constructed by selecting points evenly spaced along the distribution of the $X$ variable of interest. For each value ($X = x$), we calculate the average RF prediction over all other covariates in $X$ by
$$ \tilde{f}(x) = \frac{1}{n} \sum_{i = 1}^n \hat{f}(x, x_{i, o}), $$
where $\hat{f}$ is the predicted response from the random forest and $x_{i, o}$ is the value for all other covariates other than $X = x$ for the observation $i$~\citep{Friedman:2000}. Essentially, we average a set of predictions for each observation in the training set at the value of $X=x$. We repeating the process for a sequence of $X=x$ values to generate the estimated points to create a partial dependence plot.
Partial plots are another computationally intensive analysis, especially when there are a large number of observations. We again turn to our data caching strategy here. The default parameters for the \code{randomForestSRC::plot.variable} function generate partial dependence estimates at \code{npts=25} points (the default value) along the variable of interest. For each point of interest, the \code{plot.variable} function averages \code{n} response predictions. This is repeated for each of the variables of interest and the results are returned for later analysis.
\begin{Schunk}
\begin{Sinput}
R> # Load the data, from the call:
R> # partial_Boston <- plot.variable(rfsrc_Boston,
R> # xvar=gg_md$topvars,
R> # partial=TRUE, sorted=FALSE,
R> # show.plots = FALSE )
R> data(partial_Boston)
R>
R> # generate a list of gg_partial objects, one per xvar.
R> gg_p <- gg_partial(partial_Boston)
R>
R> # plot the variable list in a single panel plot
R> plot(gg_p, xvar=xvar, panel=TRUE, se=FALSE) +
+ labs(y=st.labs["medv"], x="")
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-partial-1}
}
\caption[Partial dependence panels]{Partial dependence panels. Risk adjusted variable dependence for variables in minimal depth rank order.}\label{fig:partial}
\end{figure}
\end{Schunk}
We again order the panels by minimal depth ranking. We see again how the \code{lstat} and \code{rm} variables are strongly related to the median value response, making the partial dependence of the remaining variables look flat. We also see strong nonlinearity of these two variables. The \code{lstat} variable looks rather quadratic, while the \code{rm} shape is more complex.
We could stop here, indicating that the RF analysis has found these ten variables to be important in predicting the median home values. That increasing \code{lstat} (percentage population of lower status) values are associated with decreasing median home values (\code{medv}) and increasing `rm > 6` (number of rooms $> 6$) are associated with increasing median home values. However, we may also be interested in investigating how these variables work together to help improve the random forest prediction of median home values.
\section{Variable Interactions} \label{S:interactions}
Using the different variable dependence measures, it is also possible to calculate measures of pairwise interactions among variables. Recall that minimal depth measure is defined by averaging the tree depth of variable $i$ relative to the root node. To detect interactions, this calculation can be modified to measure the minimal depth of a variable $j$ with respect to the maximal subtree for variable $i$~\citep{Ishwaran:2010, Ishwaran:2011}.
The \code{randomForestSRC::find.interaction} function traverses the forest, calculating all pairwise minimal depth interactions, and returns a $p \times p$ matrix of interaction measures. For each row, the diagonal terms are are related to the minimal depth relative to the root node, though normalized to minimize scaling issues. Each off diagonal minimal depth term is relative to the diagonal term on that row. Small values indicate that an off diagonal term typically splits close to the diagonal term, indicating an forest split proximity of the two variables.
The \code{gg_interaction} function wraps the \code{find.interaction} matrix for use with the provided plot and print functions. The \code{xvar} argument indicates which variables we're interested in looking at. We again use the cache strategy, and collect the figures together using the \code{panel=TRUE} option.
\begin{Schunk}
\begin{Sinput}
R> # Load the data, from the call:
R> # interaction_Boston <- find.interactions(rfsrc_Boston)
R> data(interaction_Boston)
R>
R> # Plot the results in a single panel.
R> plot(gg_interaction(interaction_Boston),
+ xvar=gg_md$topvars, panel=TRUE)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-interactions-1}
}
\caption[Minimal depth variable interactions]{Minimal depth variable interactions. Reference variables are marked with red cross in each panel. Higher values indicate lower interactivity with reference variable.}\label{fig:interactions}
\end{figure}
\end{Schunk}
The \code{gg_interaction} figure plots the interactions for the target variable (shown in the red cross) with interaction scores for all remaining variables. We expect the covariate with lowest minimal depth (\code{lstat}) to be associated with almost all other variables, as it typically splits close to the root node, so viewed alone it may not be as informative as looking at a collection of interactive depth plots. Scanning across the panels, we see each successive target depth increasing, as expected. We also see the interactive variables increasing with increasing target depth. Of interest here is the interaction of \code{lstat} with the \code{rm} variable shown in the \code{rm} panel. Aside from these being the strongest variables by both measures, this interactive measure indicates the strongest connection between variables.
\section{Coplots} \label{S:coplots}
Conditioning plots (coplots)~\citep{chambers:1992, cleveland:1993} are a powerful visualization tool to efficiently study how a response depends on two or more variables~\citep{cleveland:1993}. The method allows us to view data by grouping observations on some conditional membership. The simplest example involves a categorical variable, where we plot our data conditional on class membership, for instance on the Charles river logical variable. We can view a coplot as a stratified variable dependence plot, indicating trends in the RF prediction results within panels of group membership.
Conditional membership with a continuous variable requires stratification at some level. Often we can make these stratification along some feature of the variable, for instance a variable with integer values, or 5 or 10 year age group cohorts. However in the variables of interest in our Boston housing example, we have no "logical" stratification indications. Therefore we will arbitrarily stratify our variables into 6 groups of roughly equal population size using the \code{quantile_pts} function. We pass the break points located by \code{quantile_pts} to the \code{cut} function to create grouping intervals, which we can then add to the \code{gg_variable} object before plotting with the \code{plot.gg_variable} function. The simple modification to convert variable dependence plots into condition variable dependence plots is to use the \code{ggplot2::facet_wrap} command to generate a panel for each grouping interval.
We start by examining the predicted median home value as a function of \code{lstat} conditional on membership within 6 groups of \code{rm} ``intervals''.
\begin{Schunk}
\begin{Sinput}
R> # Find the rm variable points to create 6 intervals of roughly
R> # equal size population
R> rm_pts <- quantile_pts(rfsrc_Boston$xvar$rm, groups=6, intervals=TRUE)
R>
R> # Pass these variable points to create the 6 (factor) intervals
R> rm_grp <- cut(rfsrc_Boston$xvar$rm, breaks=rm_pts)
R>
R> # Append the group factor to the gg_variable object
R> gg_v$rm_grp <- rm_grp
R>
R> # Modify the labels for descriptive panel titles
R> levels(gg_v$rm_grp) <- paste("rm in ", levels(gg_v$rm_grp), sep="")
R>
R> # Create a variable dependence (co)plot, faceted on group membership.
R> plot(gg_v, xvar = "lstat", smooth = TRUE,
+ method = "loess", span=1.5, alpha = .5, se = FALSE) +
+ labs(y = st.labs["medv"], x=st.labs["lstat"]) +
+ theme(legend.position = "none") +
+ scale_color_brewer(palette = "Set3") +
+ facet_wrap(~rm_grp)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-coplots-1}
}
\caption[Variable Coplots]{Variable Coplots. Predicted median home values as a function of percentage of lower status population, stratified by average number of rooms groups.}\label{fig:coplots}
\end{figure}
\end{Schunk}
Each point in this figure is the predicted median home value response plotted against \code{lstat} value conditional on \code{rm} being on the interval specified. We again use the smooth loess curve to get an idea of the trend within each group. Overall, median values continue to decrease with increasing \code{lstat}, and increases with increasing \code{rm}. In addition to trends, we can also examine the conditional distribution of variables. Note that smaller homes (\code{rm}) in high status (lower \code{lstat}) neighborhoods still have high predicted median values, and that there are more large homes in the higher status neighborhoods (bottom right panel).
A single coplot gives us a grouped view of a variable (\code{rm}), along the primary variable dimension (\code{lstat}). To get a better feel for how the response depends on both variables, it is instructive to look at the complement coplot. We repeat the previous coplot process, predicted median home value as a function of the \code{rm} variable, conditional on membership within 6 groups \code{lstat} intervals.
\begin{Schunk}
\begin{Sinput}
R> # Find the lstat variable points to create 6 intervals of roughly
R> # equal size population
R> lstat_pts <- quantile_pts(rfsrc_Boston$xvar$lstat, groups=6, intervals=TRUE)
R>
R> # Pass these variable points to create the 6 (factor) intervals
R> lstat_grp <- cut(rfsrc_Boston$xvar$lstat, breaks=lstat_pts)
R>
R> # Append the group factor to the gg_variable object
R> gg_v$lstat_grp <- lstat_grp
R>
R> # Modify the labels for descriptive panel titles
R> levels(gg_v$lstat_grp) <- paste("lstat in ", levels(gg_v$lstat_grp), "
R>
R> # Create a variable dependence (co)plot, faceted on group membership.
R> plot(gg_v, xvar = "rm", smooth = TRUE,
+ method = "loess", span=1.5, alpha = .5, se = FALSE) +
+ labs(y = st.labs["medv"], x=st.labs["rm"]) +
+ theme(legend.position = "none") +
+ scale_color_brewer(palette = "Set3") +
+ #scale_shape_manual(values = event.marks, labels = event.labels)+
+ facet_wrap(~lstat_grp)
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-coplots2-1}
}
\caption[Variable Coplots]{Variable Coplots. Predicted median home value as a function of average number of rooms, stratified by percentage of lower status groups.}\label{fig:coplots2}
\end{figure}
\end{Schunk}
We get similar information from this view, predicted median home values decrease with increasing \code{lstat} percentage and decreasing \code{rm}. However viewed together we get a better sense of how the \code{lstat} and \code{rm} variables work together (interact) in the median value prediction.
Note that typically \cite{cleveland:1993} conditional plots for continuous variables included overlapping intervals along the grouped variable. We chose to use mutually exclusive continuous variable intervals for multiple reasons:
\begin{itemize}
\item Simplicity - We can create the coplot figures directly from the \code{gg_variable} object by adding a conditional group column directly to the object.
\item Interpretability - We find it easier to interpret and compare the panels if each observation is only in a single panel.
\item Clarity - We prefer using more space for the data portion of the figures than typically displayed in the \code{coplot} function available in base R, which require the bar plot to present the overlapping segments.
\end{itemize}
It is still possible to augment the \code{gg_variable} to include overlapping conditional membership with continuous variables by duplicating rows of the object, and setting the correct conditional group membership. The \code{plot.gg_variable} function recipe above could then be used to generate the panel plot, with panels ordered according to the factor levels of the grouping variable. We leave this as an exercise for the reader.
\subsection{Partial dependence coplots} \label{S:partialcoplots}
By characterizing conditional plots as stratified variable dependence plots, the next logical step would be to generate an analogous conditional partial dependence plot. The process is similar to variable dependence coplots, first determine conditional group membership, then calculate the partial dependence estimates on each subgroup using the \code{randomForestSRC::plot.variable} function with a the \code{subset} argument for each grouped interval. The \code{ggRandomForests::gg_partial_coplot} function is a wrapper for generating a conditional partial dependence data object. Given a random forest (\code{randomForestSRC::rfsrc} object) and a \code{groups} vector for conditioning the training data set observations, \code{gg_partial_coplot} calls the \code{randomForestSRC::plot.variable} function for a set of training set observations conditional on \code{groups} membership. The function returns a \code{gg_partial_coplot} object, a sub class of the \code{gg_partial} object, which can be plotted with the \code{plot.gg_partial} function.
The following code block will generate the data object for creating partial dependence coplot of the predicted median home value as a function of \code{lstat} conditional on membership within the 6 groups of \code{rm} "intervals" that we examined in the previous section.
\begin{Schunk}
\begin{Sinput}
R> partial_coplot_Boston <- gg_partial_coplot(rfsrc_Boston, xvar="lstat",
+ groups=rm_grp,
+ show.plots=FALSE)
\end{Sinput}
\end{Schunk}
Since the \code{gg_partial_coplot} makes a call to \code{randomForestSRC::plot.variable} for each group (6) in the conditioning set, we again resort to the data caching strategy, and load the stored result data from the \code{ggRandomForests} package. We modify the legend label to indicate we're working with groups of the ``Room'' variable, and use the \code{palette="Set1"} from the \pkg{RColorBrewer} package~\citep{rcolorbrewer:2014} to choose a nice color theme for displaying the six curves.
\begin{Schunk}
\begin{Sinput}
R> # Load the stored partial coplot data.
R> data(partial_coplot_Boston)
R>
R> # Partial coplot
R> plot(partial_coplot_Boston, se=FALSE)+
+ labs(x=st.labs["lstat"], y=st.labs["medv"],
+ color="Room", shape="Room")+
+ scale_color_brewer(palette="Set1")
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-prtl-coplots-1}
}
\caption[Partial Coplots]{Partial Coplots. Risk adjusted predicted median value as a function of Lower Status, conditional on groups of average number of rooms.}\label{fig:prtl-coplots}
\end{figure}
\end{Schunk}
Unlike variable dependence coplots, we do not need to use a panel format for partial dependence coplots because we are looking risk adjusted estimates (points) instead of population estimates. The figure has a loess curve through the point estimates conditional on the \code{rm} interval groupings. The figure again indicates that larger homes (\code{rm} from 6.87 and up, shown in yellow) have a higher median value then the others. In neighborhoods with higher \code{lstat} percentage, the Median values decrease with \code{rm} until it stabilizes from the intervals between 5.73 and 6.47, then decreases again for values smaller than 5.73. In lower \code{lstat} neighborhoods, the effect of smaller \code{rm} is not as noticeable.
We can view the partial coplot curves as slices along a surface viewed into the page, either along increasing or decreasing \code{rm} values. This is made more difficult by our choice to select groups of similar population size, as the curves are not evenly spaced along the \code{rm} variable. We return to this problem in the next section.
We also construct the complement view, for partial dependence coplot of the predicted median home value as a function of \code{rm} conditional on membership within the 6 groups of \code{lstat} ``intervals'', and cache the following \code{gg_partial_coplot} data call, and plot the results with the \code{plot.gg_variable} call:
\begin{Schunk}
\begin{Sinput}
R> partial_coplot_Boston2 <- gg_partial_coplot(rfsrc_Boston, xvar="rm",
+ groups=lstat_grp,
+ show.plots=FALSE)
\end{Sinput}
\end{Schunk}
\begin{Schunk}
\begin{Sinput}
R> # Load the stored partial coplot data.
R> data(partial_coplot_Boston2)
R>
R> # Partial coplot
R> plot(partial_coplot_Boston2, se=FALSE)+
+ labs(x=st.labs["rm"], y=st.labs["medv"],
+ color="Lower Status", shape="Lower Status")+
+ scale_color_brewer(palette="Set1")
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-prtl-coplots2-1}
}
\caption[Partial Coplots]{Partial Coplots. Risk adjusted predicted median value as a function of average number of rooms, conditional on groups of percentage of lower status population.}\label{fig:prtl-coplots2}
\end{figure}
\end{Schunk}
This figure indicates that the median home value does not change much until the \code{rm} increases above 6.5, then flattens again above 8, regardless of the \code{lstat} value. This agrees well with the \code{rm} partial plot(Section~\ref{S:partialdependence}) shown earlier. Again, care must be taken in interpreting the even spacing of these curves along the percentage of \code{lstat} groupings, as again, we chose these groups to have similar sized populations, not to be evenly spaced along the \code{lstat} variable.
\section{Partial plot surfaces} \label{S:plotsurface}
Visualizing two dimensional projections of three dimensional data is difficult, though there are tools available to make the data more understandable. To make the interplay of lower status and average room size a bit more understandable, we will generate a contour partial plot of the median home values. We could generate this figure with the coplot data we already have, but the resolution would be a bit strange. To generate the plot of \code{lstat} conditional on \code{rm} groupings, we would end up with contours over a grid of \code{lstat}= $25 \times$ \code{rm}= $6$, for the alternative \code{rm} conditional on \code{lstat} groups, we'd have the transpose grid of \code{lstat}= $6 \times$ \code{rm}= $25$.
Since we are already using the data caching strategy, we will generate another set of \code{gg_partial} objects with increased resolution in both the \code{lstat} and \code{rm} dimensions. For this exercise, we will find 50 points evenly spaced along the \code{rm} variable values, and generate a partial plot curve for each point. For these partial plots, we will evaluate the risk adjusted median home value over `npts=50` points along the \code{lstat} variable. This code block finds 50 \code{rm} values evenly spaced along the distribution of \code{rm}.
\begin{Schunk}
\begin{Sinput}
R> # Find the quantile points to create 50 cut points
R> rm_pts <- quantile_pts(rfsrc_Boston$xvar$rm, groups=50)
\end{Sinput}
\end{Schunk}
We use the following data call to generate the partial plots with the \code{randomForestSRC::plot.variable} call. Within the lapply call, we use scope to modify the value of the \code{rm} variable within the \code{rfsrc_Boston} training set. Since all values in the training set are the same, the averaged value of \code{rm} places each partial plot curve at a specific value of \code{rm}. This code block took about 20 minutes to run on a quad core Mac Air using a single processor.
The cached data is stored in the \code{partial_Boston_surf} data set in the \code{ggRandomForests} package. The data set is a \code{list} of 50 \code{plot.variable} objects. This code block loads the data, converts the \code{plot.variable} objects to \code{gg_partial} objects, attaches numeric values for the \code{rm} variable, and generates the contour plot.
\begin{Schunk}
\begin{Sinput}
R> # Generate the gg_partial_coplot data object
R> system.time(partial_Boston_surf <- lapply(rm_pts, function(ct){
+ rfsrc_Boston$xvar$rm <- ct
+ plot.variable(rfsrc_Boston, xvar = "lstat", time = 1,
+ npts = 50, show.plots = FALSE,
+ partial = TRUE)
+ }))
R> # user system elapsed
R> # 1109.641 76.516 1199.732
\end{Sinput}
\end{Schunk}
\begin{Schunk}
\begin{Sinput}
R> # Load the stored partial coplot data.
R> data(partial_Boston_surf)
R>
R> # Instead of groups, we want the raw rm point values,
R> # To make the dimensions match, we need to repeat the values
R> # for each of the 50 points in the lstat direction
R> rm.tmp <- do.call(c,lapply(rm_pts,
+ function(grp){rep(grp, 50)}))
R>
R> # Convert the list of plot.variable output to
R> partial_surf <- do.call(rbind,lapply(partial_Boston_surf, gg_partial))
R>
R> # attach the data to the gg_partial_coplot
R> partial_surf$rm <- rm.tmp
R>
R> # ggplot2 contour plot of x, y and z data.
R> ggplot(partial_surf, aes(x=lstat, y=rm, z=yhat))+
+ stat_contour(aes(colour = ..level..), binwidth = .5)+
+ labs(x=st.labs["lstat"], y=st.labs["rm"],
+ color="Median Home Values")+
+ scale_colour_gradientn(colours=topo.colors(10))
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-contour3d-1}
}
\caption[Partial coplot contour plot]{Partial coplot contour plot. Contours of median home value along the lstat/rm plane.}\label{fig:contour3d}
\end{figure}
\end{Schunk}
The contours are generated over the raw \code{gg_partial} estimation points, not smooth curves as shown in the partial plot (Section~\ref{S:partialdependence}) and partial coplot (Section~\ref{S:partialcoplots}) figures previously. Contour lines, like topographic maps, are concentrated where the slope of the surface is large. We use color to indicate the direction of the contour lines, so that lower median home values are concentrated in the lower right hand corner, and the values increase along the diagonal toward the upper right. The close contour lines indicate some thing like a step in values at 7 and 7.5 rooms, and at 5, 10 and 15\% lstat.
Contour plots are still a little difficult to interpret. However, we can also generate a surface with this data using the \pkg{plot3D} package~\citep{plot3D:2014} and the \code{plot3D::surf3D} function. Viewed in 3D, a surface can help to better understand what the contour lines are showing us.
\begin{Schunk}
\begin{Sinput}
R> # Modify the figure margins to make the figure larger
R> par(mai = c(0,0,0,0))
R>
R> # Transform the gg_partial_coplot object into a list of three named matrices
R> # for surface plotting with plot3D::surf3D
R> srf <- surface_matrix(partial_surf, c("lstat", "rm", "yhat"))
R>
R> # Generate the figure.
R> surf3D(x=srf$x, y=srf$y, z=srf$z, col=topo.colors(10),
+ colkey=FALSE, border = "black", bty="b2",
+ shade = 0.5, expand = 0.5,
+ lighting = TRUE, lphi = -50,
+ xlab="Lower Status", ylab="Average Rooms", zlab="Median Value"
+ )
\end{Sinput}
\begin{figure}[!htb]
{\centering \includegraphics[width=\maxwidth]{rf-surface3d-1}
}
\caption[Partial plot surface]{Partial plot surface.}\label{fig:surface3d}
\end{figure}
\end{Schunk}
These figures reinforce the previous findings, where lower home values are associated with higher \code{lstat} percentage, and higher values are associated with larger \code{rm}. The difference in this figure is we can see how the predicted values change as we move around the map of \code{lstat} and \code{rm} combinations. We do still need to be careful though, as partial plots average over values on the surface that are note supported by actual observations.
\section{Conclusion} \label{S:conclusion}
In this vignette, we have demonstrated the use of the \pkg{ggRandomForests} package to explore a regression random forest built with the \pkg{randomForestSRC} package. We have shown how to create a random forest model (Section~\ref{S:rfsrc}) and determine which variables contribute to the forest prediction accuracy using both VIMP (Section~\ref{S:vimp}) and Minimal Depth (Section~\ref{S:minimaldepth}) measures. We outlined how to investigate variable associations with the response variable using variable dependence (Section~\ref{S:variabledependence}) and the risk adjusted partial dependence (Section~\ref{S:partialdependence}) plots. We've also explored variable interactions by using pairwise minimal depth interactions (Section~\ref{S:interactions}) and directly viewed these interactions using variable dependence coplots (Section~\ref{S:coplots}) and partial dependence coplots (Section~\ref{S:partialcoplots}). Along the way, we've demonstrated the use of additional commands from the \pkg{ggplot2} package for modifying and customizing results from \pkg{ggRandomForests}.
|
\section{Introduction}
We consider a scenario in which we must encode $s$ message symbols using a length $n$ error-correcting code subject to a set of encoding constraints. Specifically, each coded symbol is a function of only a subset of the message symbols. This setup arises in various situations such as in the case of a sensor network in which each sensor can measure a certain subset of a set of parameters. The sensors would like to collectively encode the readings to allow for the possibility of measurement errors. Another scenario is one in which a client wishes to download data files from a set of servers, each of which stores information about a subset of the data files. The user should be able to recover all of the data even in the case when some of the file servers fail. Ideally, the user should also be able to download the files faster in the absence of server failures. To protect against errors, we would like the coded symbols to form an error-correcting code with reasonably high minimum distance. On the other hand, efficient download of data is permitted when the error-correcting code is of systematic form. Therefore, in this paper, we present an upper bound on the minimum distance of an error-correcting code when subjected to encoding constraints, reminiscent of the cut-set bounds presented in~\cite{Dikaliotis2011}. In certain cases, we provide a code construction that achieves this bound. Furthermore, we refine our bound in the case that we demand a systematic linear error-correcting code, and present a construction that achieves the bound. In both cases, the codes can be decoded efficiently due to the fact that our construction utilizes Reed-Solomon codes.
\subsection{Prior Work}
The problem of constructing error-correcting codes with constrained encoding has been addressed by a variety of authors. Dau et al.~\cite{Dau2013,Dau2014ISIT,Dau2014JSAC} considered the problem of finding linear MDS codes with constrained generator matrices. They have shown that, under certain assumptions, such codes exist over large enough finite fields, as well as over small fields in a special case.
A similar problem known as the weakly secure data exchange problem was studied in \cite{Yan2011},\cite{Yan2014}. The problem deals with a set of users, each with a subset of messages, who are interested in broadcasting their information securely when an eavesdropper is present. In particular, the authors of~\cite{Yan2014} conjecture the existence of secure codes based on Reed-Solomon codes and present a randomized algorithm to produce them.
The problem was also considered in the context of multisource multicast network coding in~\cite{Dikaliotis2011,Halbawi2014DRS,Halbawi2014DGC}. In~\cite{Halbawi2014DRS}, the capacity region of a simple multiple access network with three sources is achieved using Reed-Solomon codes. An analogous result is derived in~\cite{Halbawi2014DGC} for general multicast networks with 3 sources using Gabidulin codes.
There has been a recent line of work involving codes with local repairability properties, in which every parity symbol is a function of a predetermined set of data symbols \cite{Han2007, Huang2007, Gopalan2012, Pap2012, Tamo2013, Prakash2012, Kamath2013, Rawat2012,Tamo2014}. Another recent paper \cite{Mazumdar2014} represents code symbols as vertices of a partially connected graph. Each symbol is a function of its neighbors and, if erased, can be recovered from them. Our code also utilizes a graph structure, though only to describe the encoding procedure. There is not necessarily a notion of an individual code symbol being repairable from a designated local subset of the other code symbols.
\section{Problem Setup}
Consider a bipartite graph $G = (\mathcal{M},\mathcal{V},\mathcal{E})$ with $s =\card{\mathcal{M}} \leq \card{\mathcal{V}}=n$. The set $\mathcal{E}$ is the set of edges of the graph, with $(m_i, c_j) \in \mathcal{E}$ if and only if $m_i \in \mathcal{M}$ is connected to $c_j \in \mathcal{V}$. This graph defines a code where the vertices $\mathcal{M}$ correspond to message symbols and the vertices $\mathcal{V}$ correspond to codeword symbols. A bipartite graph with $s=3$ and $n=7$ is depicted in figure \ref{fig:eg}. Thus, if each $m_i$ and $c_j$ are assigned values in the finite field $\mathbb{F}_q$ with $q$ elements, then our messages are the vectors $\mathbf{m} = (m_1, \ldots, m_s) \in \mathbb{F}_q^s$ and our codewords are the vectors $\mathbf{c} = (c_1, \ldots, c_n) \in \mathbb{F}_q^n$. Each codeword symbol $c_j$ will be a function of the message symbols to which it is connected, as we will now formalize.
Henceforth, $[\mathbf{c}]_\mathcal{I}$ is the subvector of $\mathbf{c}$ with elements indexed by $\mathcal{I} \subseteq \{1, ..., n\}$, and $[\mathbf{A}]_{i,j}$ is the $(i,j)^\text{th}$ element of a matrix $\mathbf{A}$. Let $\mathcal{N}(c_j)$ denote the neighborhood of $c_j \in \mathcal{V}$, i.e. $\mathcal{N}(c_j) = \{m_i \in \mathcal{M} : (m_i, c_j) \in \mathcal{E}\}$. Similarly, define $\mathcal{N}(m_i) = \{c_j : (m_i, c_j) \in \mathcal{E}\}$. We will also consider neighborhoods of subsets of the vertex sets, i.e. for $\mathcal{V}' \subseteq \mathcal{V}$, $\mathcal{N}(\mathcal{V}') = \cup_{c_j \in \mathcal{V}'} \mathcal{N}(c_j)$. The neighborhood of a subset of $\mathcal{M}$ is defined in a similar manner. Let $m_i$ take values in $\mathbb{F}_q$ and associate with each $c_j \in \mathcal{V}$ a function $f_j: \mathbb{F}_q^s \longrightarrow \mathbb{F}_q$. We restrict each $f_j$ to be a function of $\mathcal{N}(c_j)$ only. Now consider the set $\mathcal{C} = \{(c_1,\ldots,c_n):c_j = f_j(\mathbf{m}),\mathbf{m} \in \mathbb{F}_q^s\}$. The set $\mathcal{C}$ is an error-correcting code of length $n$ and size at most $q^s$. We will denote the minimum distance of $\mathcal{C}$ as $d(\mathcal{C})$. If we restrict $f_j$ to be \emph{linear}, then we obtain a linear code with dimension at most $s$.
The structure of the code's generator matrix can be deduced from the graph $G$.
Let $\mathbf{g}_j \in \mathbb{F}_q^{s \times 1}$ be a column vector such that the $i^\text{th}$ entry is zero if $m_i \notin \mathcal{N}(c_j)$. Defining $f_j(\mathcal{N}(c_j)) = \mathbf{m}\mathbf{g}_j$ yields a linear function in which $c_j$ is a function of $\mathcal{N}(c_j)$ only, as required. A concatenation of the vectors $\mathbf{g}_j$ forms the following matrix:
\begin{equation}
\mathbf{G} = \left[\begin{array}{ccc}
| & & |\\
\mathbf{g}_1& \cdots & \mathbf{g}_n\\
| & & |
\end{array}\right]
\end{equation}
where $\mathbf{G} \in \mathbb{F}_q^{s \times n}$ is the generator matrix of the code $\mathcal{C}$.
We associate with the bipartite graph $G = (\mathcal{M},\mathcal{V},\mathcal{E})$ an adjacency matrix $\mathbf{A} \in \{0,1\}^{s \times n}$, where $[\mathbf{A}]_{i,j} = 1$ if and only if $(m_i, c_j) \in \mathcal{E}$. For the example in figure \ref{fig:eg}, this matrix is equal to
\begin{equation}
\mathbf{A} = \begin{bmatrix}
1 & 0 & 0 & 1 & 1 & 1 & 1\\
1 & 1 & 1 & 0 & 1 & 1 & 1\\
0 & 0 & 1 & 1 & 1 & 1 & 1\\
\end{bmatrix}
\label{eqn:adj}
\end{equation}
A \emph{valid} generator matrix $\mathbf{G}$ (in generic form) is built from $\mathbf{A}$ by replacing non-zero entries with indeterminates. The choice of indeterminates (from a suitably-sized finite field $\mathbb{F}_q$) determines the dimension of the code and its minimum distance. For general linear codes, the Singleton bound (on minimum distance) is tight over large alphabets.
In the presence of encoding constraints, the Singleton bound can be rather loose. In the next section, we derive an upper bound on the minimum distance of any code (linear or non-linear) associated with a bipartite graph. This bound is reminiscent of the cut-set bounds of Dikaliotis et al. in~\cite{Dikaliotis2011}.
\subsection{Subcodes of Reed-Solomon Codes}
\label{sec:RS}
Throughout this paper, we use the original definition of an $[n,k]_q$ Reed-Solomon code as in~\cite{Reed1960}, the $k$-dimensional subspace of $\mathbb{F}_q^n$ given by $\mathcal{C}_\text{RS} =\left\lbrace\left(m(\alpha_1), \ldots, m(\alpha_n)\right):\deg{m(x)} < k \right\rbrace$, where the $m(x)$ are polynomials over $\mathbb{F}_q$ of degree $\deg{m(x)}$, and the $\alpha_i \in \mathbb{F}_q$ are distinct (fixed) field elements. Each message vector $\mathbf{m} = \left(m_0, \ldots, m_{k-1}\right)$ is mapped to a message polynomial $m(x)=\sum_{i=0}^{k-1}m_i x^i$, which is then evaluated at the $n$ elements $\{\alpha_1,\alpha_2,\ldots,\alpha_n\}$ of $\mathbb{F}_q$, known as the defining set of the code. Reed-Solomon codes are MDS codes; their minimum distance attains the Singleton bound, i.e. $d(\mathcal{C}_\text{RS}) = n - k + 1$.
We can extract a subcode of a Reed-Solomon code that is valid for the bipartite graph $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$ as follows: First, let $\mathbb{F}_q$ be a finite field with cardinality $q \geq n$. Associate to each $c_j \in \mathcal{V}$ a distinct element $\alpha_j \in \mathbb{F}_q$. Consider the $i^\text{th}$ row of the adjacency matrix $\mathbf{A}$ of $G$, and let $t_i(x) = \prod_{j: [\mathbf{A}]_{i,j} = 0}(x-\alpha_j)$. For example, ${t_3(x) = (x-\alpha_1)(x-\alpha_2)}$ corresponds to the the third row of $\mathbf{A}$ in \eqref{eqn:adj}. Choose $k$ such that $k > \deg{t_i(x)}$, $\forall i$. If $\mathbf{t}_i \in \mathbb{F}_q^k$ is the (row) vector of coefficients of $t_i(x)$ and $\mathbf{G}_{\text{RS}}$ is the generator matrix of a Reed-Solomon code with defining set $\{\alpha_1,\ldots,\alpha_n\}$ and dimension $k $, then $\mathbf{t}_i\mathbf{G}_\text{RS} = (t_i(\alpha_1), \ldots, t_i(\alpha_n))$ is a vector that is valid for the $i^\text{th}$ row of $\mathbf{G}$, i.e. if $[\mathbf{A}]_{i,j} = 0$ then $[\mathbf{t}_i\mathbf{G}_\text{RS}]_j = 0$. A horizontal stacking of the vectors $\mathbf{t}_i$ results in a transformation matrix $\mathbf{T}$ that will produce a valid generator matrix $\mathbf{G}$ from $\mathbf{G}_\text{RS}$:
\begin{equation}
\mathbf{G} = \mathbf{T}\mathbf{G}_\text{RS} =
\begin{bmatrix}
&\mathbf{t}_1&\\
&\vdots&\\
&\mathbf{t}_s&\\
\end{bmatrix}
\begin{bmatrix}
1 & \cdots & 1\\
\alpha_1 & \cdots & \alpha_{n}\\
\vdots & \ddots & \vdots\\
\alpha_1^{(k-1)} & \cdots & \alpha_n^{(k-1)}
\end{bmatrix}
\end{equation}
The rank of $\mathbf{G}$ will be equal to the rank of $\mathbf{T}$, and the resulting code $\mathcal{C}$ will have a minimum distance $d(\mathcal{C})$ that is determined by $\mathcal{C}_\text{RS}$. Indeed, $d(\mathcal{C}) \geq d(\mathcal{C}_\text{RS})$.
\begin{figure}
\centering
\includegraphics[scale=0.9]{eg2.eps}
\caption{A bipartite graph representing with 3 message symbols and 7 code symbols}
\label{fig:eg}
\end{figure}
\section{Minimum Distance}
In this section, an upper bound on the minimum distance of a code defined by a bipartite graph $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$ is derived. The bound closely resembles the cut-set bounds of~\cite{Dikaliotis2011}. In most cases, this bound is tighter than the Singleton bound for a code of length $n$ and dimension $s$. For each $\mathcal{M}' \subseteq \mathcal{M}$ define $n_{\mathcal{M}'} :=\card{\mathcal{N}(\mathcal{M}')}$. This is the number of code symbols $c_j$ in $\mathcal{V}$ that are a function of the information symbols $\mathcal{M}'$.
The following proposition characterizes the minimum distance of any code defined by $G$.
\begin{prop}
\label{prop:dmin}
Fix a field $\mathbb{F}_q$. For any code $\mathcal{C}$ with $\card{\mathcal{C}} = q^s$ defined by a fixed graph $G = (\mathcal{M},\mathcal{V},\mathcal{E})$, the minimum distance $d(\mathcal{C})$ obeys
\begin{equation}
d(\mathcal{C}) \leq n_{\mathcal{M}'} - \card{\mathcal{M}'} + 1, \quad \forall \mathcal{M}' \subseteq \mathcal{M}. \label{eqn:dminbnd}
\end{equation}
\end{prop}
\begin{IEEEproof}
Working toward a contradiction, suppose $d(\mathcal{C}) > n_\mathcal{I} - \card{\mathcal{I}} + 1$ for some $\mathcal{I} \subseteq \mathcal{M}$. Let $\mathcal{C}'$ be the encoding of all message vectors $\mathbf{m}$ where $[\mathbf{m}]_{\mathcal{I}^\text{c}} \in \mathbb{F}_q^\card{\mathcal{I}^\text{c}}$ has some arbitrary but fixed value. Note that $[\mathbf{c}]_{\mathcal{N}(\mathcal{I})^\mathsf{c}}$ is the same for all $\mathbf{c} \in \mathcal{C}'$, since the symbols $\mathcal{N}(\mathcal{I})^\mathsf{c}$ are a function of $\mathcal{I}^\mathsf{c}$ only. Since $\card{\mathcal{I}} > n_\mathcal{I} - d(\mathcal{C}) + 1$, then by the pigeonhole principle there exist $\mathbf{c}_1, \mathbf{c}_2 \in \mathcal{C}'$ such that, without loss of generality, the first $n_\mathcal{I} - d(\mathcal{C}) + 1$ symbols of $[\mathbf{c}_1]_{\mathcal{N}(\mathcal{I})}$ and $[\mathbf{c}_2]_{\mathcal{N}(\mathcal{I})}$ are identical. Furthermore, $[\mathbf{c}_1]_{\mathcal{N}(\mathcal{I})^\mathsf{c}} = [\mathbf{c}_2]_{\mathcal{N}(\mathcal{I})^\mathsf{c}}$. Finally, since $\mathcal{N}(\mathcal{I})$ and $\mathcal{N}(\mathcal{I})^\text{c}$ partition $\mathcal{V}$, we obtain $d_\text{H}(\mathbf{c}_1, \mathbf{c}_2) \leq {n - (n_\mathcal{I} - d(\mathcal{C}) + 1 + (n - n_\mathcal{I}))} = d(\mathcal{C})-1$, a contradiction. Figure \ref{fig:dmin} illustrates the relation between $\mathcal{I}$ and the corresponding partition of $\mathcal{V}$.
\end{IEEEproof}
As a direct corollary, we obtain the following upper bound on $d(\mathcal{C})$:
\begin{clry}
\begin{equation}
d(\mathcal{C}) \leq \min_{\mathcal{M}'\subseteq \mathcal{M}}\{n_{\mathcal{M}'} - \card{\mathcal{M}'}\} + 1 \label{eqn:dmin}
\end{equation}
\end{clry}
Our next task is to provide constructions of codes that achieve this bound.
\begin{figure}
\centering
\includegraphics[scale=1]{dmin_proof.eps}
\caption{Partitions of $\mathcal{M}$ and of $\mathcal{V}$ used in the proof of proposition \protect{\ref{prop:dmin}}. The set $\mathcal{N}(\mathcal{I})$ is a function of both $\mathcal{I}$ and $\mathcal{I}^\mathsf{c}$, while the set $\mathcal{N}(\mathcal{I})^\mathsf{c}$ is a function of $\mathcal{I}^\mathsf{c}$ only.}
\label{fig:dmin}
\end{figure}
\section{Systematic Construction}
In this section, we provide a code construction that achieves the minimum distance bound stated in corollary \ref{clry:dsys}. We appeal to Hall's Theorem, a well-known result in graph theory that establishes a necessary and sufficient condition for finding a matching in a bipartite graph. Some terminology needed from graph theory is defined in the following subsection.
\subsection{Graph Theory Preliminaries}
Let $G = (\mathcal{S},\mathcal{T},\mathcal{E})$ be a bipartite graph. A \emph{matching} is a subset $\tilde{\mathcal{E}} \subseteq \mathcal{E}$ such that no two edges in $\tilde{\mathcal{E}}$ share a common vertex. A vertex is said to be \emph{covered} by $\tilde{\mathcal{E}}$ if it is incident to an edge in $\tilde{\mathcal{E}}$. An $\mathcal{S}$-\emph{covering} matching is one by which each vertex in $\mathcal{S}$ is covered. We will abuse terminology and say that an edge $e \in \tilde{\mathcal{E}}$ is \emph{unmatched} if $e \notin \tilde{\mathcal{E}}$. We can now state Hall's Theorem.
\begin{thm}
\label{thm:hall}
Let $G = (\mathcal{S},\mathcal{T},\mathcal{E})$ be a bipartite graph. There exists an $\mathcal{S}$-covering matching if and only if $\card{\mathcal{S}'} \leq \mathcal{N}(\mathcal{S}')$ for all $\mathcal{S}' \subseteq \mathcal{S}$.
\end{thm}
For a proof of the theorem, see e.g. \cite[p.53]{LintWilsonBook}.
Set $d_\mathsf{min} = \min_{\mathcal{M}'\subseteq \mathcal{M}}\{n_{\mathcal{M}'} - \card{\mathcal{M}'}\} + 1$. In order to construct a generator matrix $\mathbf{G} \in \mathbb{F}_q^{s\times n}$ for a code $\mathcal{C}$ with minimum distance $d_\mathsf{min}$, we will use an $[n,n-d_\mathsf{min} +1]$ Reed-Solomon code with generator matrix $\mathbf{G}_\text{RS}$. We will then extract $\mathcal{C}$ as a subcode using an appropriately built transformation matrix $\mathbf{T}$ to form $\mathbf{G} = \mathbf{T}\mathbf{G}_\text{RS}$ such that $\mathbf{G}$ is in systematic form, which implies that the dimension of $\mathcal{C}$ is $s$. Since $\mathcal{C}$ is a subcode of a code with minimum distance $d_\mathsf{min}$, we have $d(\mathcal{C}) \geq d_\mathsf{min}$. $\eqref{eqn:dmin}$ further implies that $d(\mathcal{C}) = d_\mathsf{min}$.
Our construction is as follows: consider a graph $G = (\mathcal{M},\mathcal{V},\mathcal{E})$ defining $\mathcal{C}$, and define the set $\mathcal{A} = \{c_j : \mathcal{N}(c_j) = \mathcal{M}\}$, i.e. $\mathcal{A}$ is the set of code symbols that are a function of \textit{every} message symbol. Note that $\mathcal{A} \subseteq \mathcal{N}(\mathcal{M}')$ for every $\mathcal{M}' \subseteq \mathcal{M}$. Therefore, if $a = \card{\mathcal{A}}$ then the size of the neighborhood of $\mathcal{N}(\mathcal{M}')$ can be expressed as $n_{\mathcal{M}'} = r_{\mathcal{M}'} + a$, where $r_{\mathcal{M}'}$ is the cardinality of the set $\mathcal{R}(\mathcal{M}') = \mathcal{N}(\mathcal{M}') \setminus \mathcal{A}$.
\begin{thm}
Let $G = (\mathcal{M},\mathcal{V},\mathcal{E})$. Set $d_\mathsf{min} = \min_{\mathcal{M}'\subseteq \mathcal{M}}\{n_{\mathcal{M}'} - \card{\mathcal{M}'}\} + 1$ and $k_\mathsf{min} = n - d_\mathsf{min} + 1$. A linear code $\mathcal{C}$ with parameters $[n,s,d_\mathsf{min}]$ valid for $G$ can be constructed with a systematic-form generator matrix provided that $k_\mathsf{min} \geq r_\mathcal{M}$.
\end{thm}
\begin{IEEEproof}
First, we establish a bound on $a$. Note that since $n = n_\mathcal{M} = r_\mathcal{M} + a$ and $k_\mathsf{min} \geq r_\mathcal{M}$, then we have $a \geq d_\mathsf{min} - 1$. Fix an arbitrary subset $\mathcal{A}^* \subseteq \mathcal{A}$ of size $a^* = a - (d_\mathsf{min} - 1)$, which is guaranteed to exist by virtue of the bound on $a$, and let $\mathcal{B} = \mathcal{A}\setminus\mathcal{A}^*$. Now, we focus on a particular subgraph of $G$ defined by $G^* = (\mathcal{M},\mathcal{V}^*,\mathcal{E}^*)$ where $\mathcal{V}^* = \mathcal{V} \setminus \mathcal{B}$, and $\mathcal{E}^*= \{(m_i,c_j) \in \mathcal{E}: c_j \in \mathcal{V}^*\}$ is the edge set corresponding to this subgraph.
Since $n_{\mathcal{M}'} = r_{\mathcal{M}'} + a$, then from the definition of $d_\mathsf{min}$ we have
\begin{equation}
\card{\mathcal{M}'} \leq r_{\mathcal{M}'} + a - (d_\mathsf{min} - 1), \quad \forall \mathcal{M}' \subseteq \mathcal{M} \label{eqn:Hall1}
\end{equation}
The neighborhood of every subset $\mathcal{M}'$ when restricted to $\mathcal{V}^*$ is exactly $\mathcal{N}^*(\mathcal{M}') = \mathcal{R}(\mathcal{M}') \cup \mathcal{A}^*$, with cardinality $n^*_{\mathcal{M}'} = r_{\mathcal{M}'} + a^*$. The bounds \eqref{eqn:Hall1} can now be expressed in a way suitable for the condition of Hall's theorem:
\begin{equation}
\card{\mathcal{M}'} \leq n^*_{\mathcal{M}'}, \quad \forall \mathcal{M}' \subseteq \mathcal{M} \label{eqn:Hall2}
\end{equation}
An $\mathcal{M}$-covering matching in $G^*$ can be found by letting $\mathcal{S} = \mathcal{M}$ and $\mathcal{T} = \mathcal{V}^*$ in theorem~\ref{thm:hall}. Let $\tilde{\mathcal{E}} = \{(m_i,c_{{j}(i)})\}_{i=1}^s\subseteq \mathcal{E}^*$ be such a matching, and $\tilde{\mathcal{V}}$ the subset of $\mathcal{V}^*$ that is covered by $\tilde{\mathcal{E}}$. Let $\mathbf{A}_{\tilde{\mathcal{E}}}$ be the adjacency matrix of $G$ when the edge set $\{(m_i,c_j) \in \mathcal{E}: c_j \in \tilde{\mathcal{V}}, j \neq j(i)\}$ is removed. The number of zeros in any row of $\mathbf{A}_{\tilde{\mathcal{E}}}$ is at most $n - d_\mathsf{min}$. To see this, note that the edges in $\mathcal{E}$ incident to $\mathcal{B}$ are not removed by the matching, and every $m_i \in \mathcal{M}$ is connected to at least one vertex in $\mathcal{V}^*$.
Next, we build a valid $\mathbf{G}$ for $G$ using $\mathbf{A}_{\tilde{\mathcal{E}}}$, utilizing the method described in section \ref{sec:RS}. Fix a $[n,n-d_\mathsf{min}+1]$ Reed-Solomon code with generator matrix $\mathbf{G}_\text{RS}$ and defining set $\{\alpha_1, \ldots, \alpha_n\}$. The $i^\text{th}$ transformation polynomial is $t_i(x)=\prod_{j:[\mathbf{A}_{\tilde{\mathcal{E}}}]_{i,j}=0}(x-\alpha_i)$. Since the number of zeros in any row of $\mathbf{A}_{\tilde{\mathcal{E}}}$ is at most $n - d_\mathsf{min}$, we have $\deg {t_i(x)} \leq n - d_\mathsf{min} = k - 1$ for all $i$. We use the $t_i(x)$, after normalizing by $t_i(\alpha_{j(i)})$, to construct a transformation matrix $\mathbf{T}$ and then $\mathbf{G} = \mathbf{T}\mathbf{G}_\text{RS}$ is valid for $G$. Note that $\mathbf{G}$ is in systematic form due the fact that the columns of $\mathbf{A}_{\tilde{\mathcal{E}}}$ indexed by $\{j(i)\}_{i=1}^s$ form a permutation of the identity matrix of size $s$. Lastly, $d(\mathcal{C}) = d_\text{min}$ since $d(\mathcal{C}) \leq d_\mathsf{min}$ by corollary \eqref{eqn:dmin}, and $d(\mathcal{C}) \geq d_\mathsf{min}$ since $\mathcal{C}$ is a subcode of a code with minimum distance $d_\mathsf{min}$.
\end{IEEEproof}
\section{Minimum Distance for Systematic Linear Codes}
In this section, we will restrict our attention to the case where a code valid for $G$ is linear, so that each $c_j \in \mathcal{V}$ is a linear function of the message symbols $m_i \in \mathcal{N}(c_j)$. We seek to answer the following: What is the greatest minimum distance attainable by a \textit{systematic} linear code valid for $G$?
Any systematic code must correspond to a matching $\tilde{\mathcal{E}} \subseteq \mathcal{E}$ which identifies each message symbol $m_i \in \mathcal{M}$ with a unique codeword symbol $c_{j(i)} \in \mathcal{V}$, where $j(i) \in \{1, \ldots, n\}$. Explicitly, $\tilde{\mathcal{E}}$ consists of $s$ edges of the form $\{(m_i, c_{j(i)})\}$ for $i = 1, \ldots, s$ such that $c_{j(i_1)} \ne c_{j(i_2)}$ for $i_1 \ne i_2$. As before, $\tilde{\mathcal{V}}$ is the subset of vertices in $\mathcal{V}$ which are involved in the matching: $\tilde{\mathcal{V}} = \{c_{j(i)}\}_{i = 1}^s$. Our code becomes systematic by setting $c_{j(i)} = m_i$ for $i = 1, \ldots, s$, and choosing each remaining codeword symbol $c_j \notin \tilde{\mathcal{V}}$ to be some linear function of its neighboring message symbols $m_i \in \mathcal{N}(c_j)$.
\begin{defn}
For $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$, let $\tilde{\mathcal{E}} \subseteq \mathcal{E}$ be an $\mathcal{M}$-covering matching so that $\tilde{\mathcal{E}} = \{(m_i, c_{j(i)})\}_{i = 1}^s$. Let $\tilde{\mathcal{V}} = \{c_{j(i)}\}_{i = 1}^s$ be the vertices in $\mathcal{V}$ which are covered by $\tilde{\mathcal{E}}$. Define the \textit{matched adjacency matrix} $\mathbf{A}_{\tilde{\mathcal{E}}} \in \{0, 1\}^{s \times n}$ so that $[\mathbf{A}_{\tilde{\mathcal{E}}}]_{i, j} = 1$ if and only if either $(m_i, c_j) \in \tilde{\mathcal{E}}$, or $c_j \notin \tilde{\mathcal{V}}$ and $(m_i, c_j) \in \mathcal{E}$. In other words, $\mathbf{A}_{\tilde{\mathcal{E}}}$ is the adjacency matrix of the bipartite graph formed by starting with $G$ and deleting the edges $\{(m_i, c_j) \in \mathcal{E} ~:~ c_j \in \tilde{\mathcal{V}} \text{ and } j \ne j(i)\}$.
\end{defn}
\begin{defn}
Let $\tilde{\mathcal{E}} \subseteq \mathcal{E}$ be a matching for the $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$ which covers $\mathcal{M}$. Let $z_{\tilde{\mathcal{E}}}$ be the maximum number of zeros in any row of the corresponding matched adjacency matrix $\mathbf{A}_{\tilde{\mathcal{E}}}$, and define $k_{\tilde{\mathcal{E}}} := z_{\tilde{\mathcal{E}}} + 1$. Furthermore, define $k_\mathsf{sys} = \text{min}_{\tilde{\mathcal{E}}} k_{\tilde{\mathcal{E}}}$ where $\tilde{\mathcal{E}}$ ranges over all matchings for $G$ which cover $\mathcal{M}$, and $d_\mathsf{sys} = n - k_\mathsf{sys} + 1$.
\end{defn}
\begin{lem}
For a given bipartite graph $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$ which merits a matching that covers $\mathcal{M}$, we have
\begin{equation}
s \le k_\mathsf{min} \le k_\mathsf{sys} \le n
\end{equation}
and
\begin{equation}
d_\mathsf{sys} \le d_\mathsf{min}.
\end{equation}
\end{lem}
\begin{IEEEproof}
Let $\mathbf{A}$ be the adjacency matrix of $G$.
For any subset $\mathcal{M}' \subseteq \mathcal{M}$ we have $d_\mathsf{min} \le n_{\mathcal{M}'} - |\mathcal{M}'| + 1$, and likewise $k_\mathsf{min} = n - d_\mathsf{min} + 1 \ge |\mathcal{M}'| + (n-n_{\mathcal{M}'})$. Taking $\mathcal{M}' = \mathcal{M}$ (and noting that in our framework, every $c_j \in \mathcal{V}$ is connected to at least one vertex in $\mathcal{M}$, hence $n_{\mathcal{M}} = n$) we obtain $k_\mathsf{min} \ge s$.
Now choose a set $\mathcal{M}'$ for which the above relation holds with equality, that is, $k_\mathsf{min} = |\mathcal{M}'| + (n-n_{\mathcal{M}'})$. Since $\mathcal{N}(\mathcal{M}')$ is simply the union of the support sets of the rows of $\mathbf{A}$ corresponding to $\mathcal{M}'$, then each of these rows must have at least $n - n_{\mathcal{M}'} = |\mathcal{N}(\mathcal{M}')^\mathsf{c}|$ zeros. Furthermore, any matching $\tilde{\mathcal{E}}$ which covers $\mathcal{M}$ must identify the rows of $\mathcal{M}'$ with columns of $\mathcal{N}(\mathcal{M}')$. Thus, in the matched adjacency matrix $\mathbf{A}_{\tilde{\mathcal{E}}}$, the row corresponding to $j \in \mathcal{M}'$ must have $|\mathcal{M}'| - 1$ zeros in the columns of $\mathcal{N}(\mathcal{M})$ which are matched to $\mathcal{M}' \setminus \{j\}$, in addition to the $n - n_{\mathcal{M}'}$ zeros in the columns corresponding to $\mathcal{N}(\mathcal{M}')$. This gives us $k_{\tilde{\mathcal{E}}} \ge |\mathcal{M}'| + (n-n_{\mathcal{M}'})$ for each matching $\tilde{\mathcal{E}}$, hence $k_\mathsf{sys} \ge k_\mathsf{min}$. It follows directly that $d_\mathsf{sys} \le d_\mathsf{min}$. Finally, it is clear from definition that for any $\mathcal{M}$-covering matching $\tilde{\mathcal{E}}$ we must have that $k_{\tilde{\mathcal{E}}}$ is less than the length of the adjacency matrix $\mathbf{A}$, which is $n$, hence $k_\mathsf{sys} \le n$.
\end{IEEEproof}
\begin{cor}
\label{clry:dsys}
Let $G = (\mathcal{M}, \mathcal{V}, \mathcal{E})$ be a bipartite graph which merits a systematic linear code. The largest minimum distance obtainable by a systematic linear code is $d_\mathsf{sys}$.
\end{cor}
\begin{IEEEproof}
Let $\mathcal{C}$ be a systematic linear code which is valid for $G$. Then $\mathcal{C}$ must have a codeword containing at least $k_\mathsf{sys} - 1$ zeros, i.e. a codeword of Hamming weight at most $n - k_\mathsf{sys} + 1 = d_\mathsf{sys}$. Since the code is linear, this Hamming weight is an upper bound for its minimum distance, so $d(\mathcal{C}) \le d_\mathsf{sys}$.
It remains to see that there are systematic linear codes which are valid for $G$ and achieve a minimum distance of $d_\mathsf{sys}$. Let $\tilde{\mathcal{E}}$ be an $\mathcal{M}$-covering matching for $G$ such that $k_{\tilde{\mathcal{E}}} = k_\mathsf{sys}$. Then for any $k \ge k_\mathsf{sys}$, we claim that an $[n, k]$ Reed-Solomon code contains a systematic linear subcode that is valid for $G$. Indeed, choose a set of $n$ distinct elements $\{\alpha_i\}_{i = 1}^n \subseteq \mathbb{F}_q$ as the defining set of our Reed-Solomon code. Then to form our subcode's generator matrix $\mathbf{G}$, note that (as mentioned before) $\mathbf{G}$ must have zero entries in the same positions as the zero entries of $\mathbf{A}_{\tilde{\mathcal{E}}}$, and indeterminate elements in the remaining positions. There are at most $k_\mathsf{sys} - 1$ zeros in any row of $\mathbf{A}_{\tilde{\mathcal{E}}}$ (and at \textit{least} $s-1$ zeros in each row, since there must be $s$ columns which have nonzero entries in exactly one row). For each row $i \in \{1,\ldots,s\}$ of $\mathbf{A}_{\tilde{\mathcal{E}}}$, let $\mathcal{I}_i \subseteq \{1,\ldots,n\}$ be the set of column indices $j$ such that $[\mathbf{A}_{\tilde{\mathcal{E}}}]_{i, j} = 0$. Then form the polynomial $t_i(x) = \prod_{j \in \mathcal{I}_i} (x - \alpha_j)$ and normalize by $t_i(\alpha_{j(i)})$, which accordingly has degree at most $k_\mathsf{sys}$ (and at least $s-1$). We now set the $i^{th}$ row of $\mathbf{G}$ to be $(t_i(\alpha_1),\ldots, t_i(\alpha_n))$, and we see that by construction this row has zeros precisely at the indices $j \in \mathcal{I}_i$ as desired.
The rows of $\mathbf{G}$ generate a code with minimum distance at least that of the original Reed-Solomon code, which is $n - k + 1$. Furthermore, by setting $k = k_\mathsf{sys}$ for our Reed-Solomon code, we see this new code $\mathcal{C}$ has minimum distance at least $n - k_\mathsf{sys} + 1 = d_\mathsf{sys}$. Since by our previous argument, $d(\mathcal{C}) \le d_\mathsf{sys}$, the minimum distance of $\mathcal{C}$ must achieve $d_\mathsf{sys}$ with equality.
\end{IEEEproof}
\section{Achievability Using MDS Codes}
Throughout this paper, we have utilized Reed-Solomon codes to construct systematic linear codes valid for a particular $G=(\mathcal{M},\mathcal{V},\mathcal{E})$ that attain the highest possible distance. It is worth mentioning that this choice is not necessary and in fact, the Reed-Solomon code utilized can be replaced with any linear MDS code with the same parameters.
\begin{lem}
Fix an arbitrary $[n,k]$ linear MDS code $\mathcal{C}$. For any $\mathcal{I} \subseteq [n]$ where $\card{\mathcal{I}} \leq k-1$ , there exists $\mathbf{c} \in \mathcal{C}$ such that $[\mathbf{c}]_\mathcal{I} = \mathbf{0}$.
\begin{IEEEproof}
Let $\mathbf{G} = [\mathbf{g}_i]_{i=1}^n$ be the generator matrix of $\mathcal{C}$ and let $\mathbf{G}_\mathcal{I}=[\mathbf{g}_i]_{i \in \mathcal{I}}$. Since $\card{\mathcal{I}} \leq k-1$, $\mathbf{G}_\mathcal{I}$ has full column rank and so it has a non-trivial left nullspace of dimension $k - \card{\mathcal{I}}$. If $\mathbf{h}$ is any vector in that nullspace then $\mathbf{c} = \mathbf{h}\mathbf{G}$ is such that $[\mathbf{c}]_\mathcal{I} = \mathbf{0}$.
\end{IEEEproof}
\end{lem}
Therefore, to produce a valid linear code $\mathcal{C}$ for $G=(\mathcal{M},\mathcal{V},\mathcal{E})$ with $d(\mathcal{C}) = d^*$, where $d^* \leq n_{m_i}$ for all $m_i \in \mathcal{M}$, we fix an arbitrary $[n,n-d^*+1]$ MDS code and then select vectors ${\mathbf{h}_1,\ldots,\mathbf{h}_s}$ such that $\mathbf{h}_i$ is in the left nullspace of $\mathbf{G}_{\mathcal{I}_i}$, where $\mathcal{I}_i =\{j:\mathbf{A}_{i,j} = 0\}$. Note that the specific selection of the $\mathbf{h}_i$ determines the dimension of $\mathcal{C}$.
For a systematic construction, in which the dimension of the code is guaranteed to be $s$, some extra care has to be taken when choosing the $\mathbf{h}_i$. We must choose each $\mathbf{h}_i$ such that its not in the nullspace of $\mathbf{g}_{j(i)}$, which the column corresponding to the systematic coordinate $c_{j(i)}$.
\section{Example}
In this section, we construct a systematic linear code that is valid for the graph in figure \ref{fig:eg}. The bound of corollary \ref{eqn:dmin} asserts that $d(\mathcal{C}) \leq 5$ for any $\mathcal{C}$ valid for $G$. However, lemma \ref{clry:dsys} shows that $d(\mathcal{C}_\mathsf{sys}) \leq4$ for any valid systematic linear code $\mathcal{C}_\mathsf{sys}$. A matching achieving this bound is given by the edges $\tilde{\mathcal{E}} = \{(m_1,v_1),(m_2,v_2),(m_3,v_3)\}$ and so the edges removed from the graph are $\{(m_2,v_1),(m_2,v_3)\}$. The new adjacency matrix $\mathbf{A}_{\tilde{\mathcal{E}}}$ is given by,
\begin{equation}
\mathbf{A}_{\tilde{\mathcal{E}}} = \begin{bmatrix}
1 & 0 & 0 & 1 & 1 & 1 & 1\\
\mathbf{0} & 1 & 0 & \mathbf{0} & 1 & 1 & 1\\
0 & 0 & 1 & 1 & 1 & 1 & 1\\
\end{bmatrix}
\end{equation}
where boldface zeros refer to those edges removed from $G$ because of the matching $\tilde{\mathcal{E}}$.
A generator matrix which is valid for $\mathbf{A}_{\tilde{\mathcal{E}}}$ can be constructed from that of a $[7,4]$ Reed-Solomon code over $\mathbb{F}_7$ with defining set $\{0,1,\alpha,\ldots,\alpha^5\}$ where $\alpha$ is a primitive element in $\mathbb{F}_7$, using the method described in \ref{sec:RS}.
The polynomials corresponding to the transformation matrix are given by,
\begin{IEEEeqnarray}{rCl}
t_1(x) & = & \alpha^{5}(x-1)(x-\alpha) \\
t_2(x) & = & \alpha^{4}x(x-\alpha)(x-\alpha^2) \\
t_3(x) & = &\alpha^{3}x(x-1)
\end{IEEEeqnarray}
Finally, the systematic generator matrix for $\mathcal{C}_\mathsf{sys}$ is,
\begin{equation}
\mathbf{G}_\mathsf{sys}= \begin{bmatrix}
1 & 0 & 0 & \alpha^2 & \alpha^5 & 1 & \alpha^5 \\
0 & 1 & 0 & 0 & 1 & \alpha^4 & 1 \\
0 & 0 & 1 & \alpha^5 & \alpha^5 & \alpha^2 & 1
\end{bmatrix}
\end{equation}
\section{Conclusion}
In this paper, we have studied the problem of analyzing and designing error-correcting codes when the encoding of every coded symbol is restricted to a subset of the message symbols. We obtain an upper bound on the minimum distance of any such code, similar to the cut-set bounds of~\cite{Dikaliotis2011}. By providing an explicit construction, we show that under certain assumptions this bound is achievable. Furthermore, the field size required for the construction scales linearly with the code length. The second bound is on the minimum distance of linear codes with encoding constraints when the generator matrix is required to be in systematic form. We provide a construction that always achieves this bound. Since all of our constructions are built as subcodes of Reed-Solomon codes, they can be decoded efficiently using standard Reed-Solomon decoders. For future work, it remains to show that the first upper bound is achievable in general over small fields.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\noindent
\emph{Network coding}, introduced 2000 in \cite{ahlswede2000network}, is a promising method for transmitting information over a network.
\cite{li2003linear} proved in 2003 that \emph{linear network coding} (LNC) is a \emph{max-flow} achieving approach for general multisource multicast networks.
In LNC, the packets within the networks are vectors over a finite field $\ensuremath{\mathbb F_{q}}$ and intermediate nodes transmit $\ensuremath{\mathbb F_{q}}$-linear combinations of the incoming packets.
In 2006, \cite{ho2006} showed that choosing the linear functions at the nodes of a network in a random fashion achieves the \emph{max-flow bound} with probability exponentially approaching 1 with the code length.
This method is called \emph{random linear network coding} (RLNC).
In this paper, we consider a unicast scenario, where one source node with $n$ outgoing edges wants to transmit information to one destination node with $\tilde{n} \ge n$ incoming edges over a network whose topology is neither known by the source nor by the destination.
We assume that the transmission takes place in \emph{generations}, which happen infinitely fast, or equivalenty, every node of the network waits until all incoming edges have sent their packets before sending the outgoing packets.
The source emits $n$ vectors from $\ensuremath{\mathbb F_{q}}^M$ (length-$M$ packets) into the network in parallel and the destination collects $\tilde{n} \geq n$ such packets from the network during one generation.
Within the network, nodes send---possibly different---$\ensuremath{\mathbb F_{q}}$-linear combinations of the packets of their incoming edges to the outgoing edges.
The vectors sent by the source during one generation can be seen as a matrix $\ve{X} \in \ensuremath{\mathbb F_{q}}^{n \times M}$ and the collected packets by the receiver analogously as $\ve{Y} \in \ensuremath{\mathbb F_{q}}^{\tilde{n} \times M}$, where the rows of the matrices correspond to the packets.
Errors are considered to be vectors added to nodes as additional incoming edges.
This channel model is equivalent to the operator channel described in \cite{koetter2008coding} under the assumption that the network stays constant during one generation.
The input-output relation of the considered scenario is referred to as the \emph{multiplicative additive matrix channel} (MAMC) \cite{silva2010}
\begin{align} \label{Eq_MAMC}
\ve{Y} = \ve{A} \cdot \ve{X} + \ve{B},
\end{align}
where $\ve{A} \in \ensuremath{\mathbb F_{q}}^{\tilde{n} \times n}$ is called the \emph{network channel matrix}, and $\ve{B} \in \ensuremath{\mathbb F_{q}}^{\tilde{n} \times M}$ is the \emph{additive error matrix}. It is commonly assumed (e.g., \cite{silva2010}) that $\tilde{n}=n$ and $\ve{A}$ is uniformly distributed at random among all regular matrices from $\ensuremath{\mathbb F_{q}}^{n \times n}$.
The latter is a good assumption for sufficiently large networks and field sizes.
We also assume that errors occur additively at arbitrary intermediate nodes.
In this case, it can be shown (cf.\ \cite{koetter2008coding}) that the rank of $\ve{B}$ is upper bounded by the number of additive errors within the network during one generation.
Error correction for MAMCs can be done using \emph{lifted rank-metric codes}.
Alternatively, \cite{seidl2013differential} showed that higher rates can be achieved by using differential precoding instead of lifting.
This noncoherent transmission scheme for MAMCs is called \emph{differential linear network coding} (DLNC).
However, the method requires the channel matrix to remain constant between generations.
This is a strong assumption which does not hold for all real-world communication networks.
Hence, it has to be determined which influence a ``slow'' variation of the network (change of the multiplicative matrix in the MAMC) has on the error structure.
We show that slow (i.e., the probability of a leaving/joining node during one generation is small) changes of the network topology result in impulsive error peaks.
For this case, we propose a suitable channel coding scheme based on \emph{rank-metric partial-unit-memory (PUM) codes} and analyze it in terms of transmission rates and error probabilities.
This paper is organized as follows: In Sec.~\ref{sec:DLNC} and \ref{sec:RankMetricCodes} we briefly explain DLNC and rank-metric codes.
An error model which includes slow changes of the network topology is introduced in Sec.~\ref{sec:Error_Model} and Sec.~\ref{sec:CodingScheme} shows a proper coding scheme for it and analyzes its performance.
We conclude the paper with a short summary in Sec.~\ref{sec:Conclusion}.
We make use of the following notations.
Let $q$ be a prime power, $m$ a positive integer.
We write $\ensuremath{\mathbb F_{q}}^{k \times \ell}$ for the set of all $k \times \ell$ matrices over $\ensuremath{\mathbb F_{q}}$.
Vectors are considered to be row vectors.
$\ve{I}_n$ denotes the identity matrix of size $n \times n$.
$\mathrm{GL}(n, \ensuremath{\mathbb F_{q}})$ is the set of all invertible $(n \times n)$-matrices over $\ensuremath{\mathbb F_{q}}$.
The entry of a matrix $\ve{A}$ in the $i^\mathrm{th}$ row and the $j^\mathrm{th}$ column is called $A_{ij}$.
Let $\ensuremath{\boldsymbol{\beta}} := \left( \beta_1, \dots, \beta_m \right)$ be a basis\footnote{E.g., $\ve{\beta}=\left(1,\alpha, \alpha^2, \dots, \alpha^{m-1}\right)$, where $\alpha$ is a primitive element of $\ensuremath{\mathbb F_{q^m}}$.} of $\ensuremath{\mathbb F_{q^m}}$ over $\ensuremath{\mathbb F_{q}}$. Then there is a bijective linear map
\begin{align*}
\ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}} : \, \ensuremath{\mathbb F_{q^m}}^n \rightarrow \ensuremath{\mathbb F_{q}}^{m \times n}, \, \ve{a} = \left(a_1, \dots, a_n \right) \mapsto \ve{A} = \left[ A_{ij} \right], \label{eq:bijective_linear_map}
\end{align*}
where $a_j = \sum_{i=1}^{m} A_{ij} \beta_i$ for all $j \in \left\{ 1, \dots, n \right\}$.
$f^{(*)\ell}(x)$ denotes the $\ell$-fold convolution of $f(x)$, i.e.,
$f^{(*)0}(x) = \delta(x)$, $f^{(*)1}(x) = f(x)$,
$f^{(*)2}(x) = f(x) * f(x)$,
$f^{(*)3}(x) = f(x) * f(x) * f(x)$, etc.
\section{Differential Linear Network Coding}\label{sec:DLNC}
\noindent
In this section, we briefly describe the DLNC precoding method introduced in \cite{seidl2013differential}.
The concept of DLNC can be compared with \emph{differential phase-shift keying}---information is not transmitted absolutely, but by the transition between two successive symbols.
We first restrict ourselves to square matrices, i.e., $n = M$.
For differential modulation, we assume to have a sequence of source words $\S_i \in \mathrm{GL}(n, \ensuremath{\mathbb F_{q}})$, $i=1,\dots,N$, for some $N \in \ensuremath{\mathbb{N}}$. Then we generate the DLNC transmit symbol $\ve{X}_i$ of the $i$th generation as follows
\begin{equation} \label{Eq_DLNC_X}
\ve{X}_i = \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}} \cdot \ve{S}_i,
\end{equation}
where $\ve{X}_0 := \ve{I}_n$ is the initialization word.
Differential demodulation at the destination node starts by calculating the pseudoinverse $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+$ of the previously received matrix $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{}$, followed by calculating the product $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \cdot \ve{Y}_i^{}$. The pseudoinverse $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+$ has the following properties
\begin{eqnarray}
\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \cdot \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} &=&
\ve{I}_n^{} + \ve{L} \ve{I}_\mathcal{U}^\mathrm{T}, \label{Eq_pinv1} \\
\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \cdot \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \cdot \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} &=&
\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{}.
\end{eqnarray}
Thereby, $\ve{I}_\mathcal{U}$ consists of a subset $\mathcal{U}$ of the rows of $\ve{I}_n$. $\ve{L}$ is a full-rank matrix of appropriate dimensions and $\ve{I}_\mathcal{U}^\mathrm{T} \ve{L} = -\ve{I}_{|\mathcal{U}|}$.
Thus, with $\ve{Y}_i = \ve{A}_i \ve{X}_i + \ve{B}_i$, and if we assume that the network channel matrix $\ve{A}_i$ stays constant between the two generations $\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}$ and $i$, the demodulation process results in the demodulated symbol $\tilde{\ve{S}}_i$
\begin{eqnarray}
\tilde{\ve{S}_i} \hspace*{-2mm} &=& \hspace*{-2mm} \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{Y}_i^{}
= \ve{S}_i +
\underbrace{%
\ve{L} \ve{I}_\mathcal{U}^\mathrm{T} \ve{S}_i^{}
- \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{S}_i^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{B}_i^{}
}_{=: \ve{E}_i} \label{Eq_E_i} \nonumber \\[-0.3cm]
&=& \hspace*{-2mm} \ve{S}_i + \ve{E}_i, \label{Eq_DLNC_AMC}
\end{eqnarray}
where $\ve{E}_i$ denotes the \emph{effective error matrix} in generation $i$. It can be directly seen from \eqref{Eq_DLNC_AMC} that differential modulation and demodulation transforms the MAMC with network channel matrix $\ve{A}_i$ and additive error matrix $\ve{B}_i$ into an \emph{additive matrix channel} (AMC) \cite{silva2010} with additive error matrix $\ve{E}_i$.
It is shown in \cite{seidl2013differential} that $\rk(\ve{E}_i) \leq \rk(\ve{B}_i) + \rk(\ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}})$.
In case of non-square matrices $\ve{X}_i$ ($n \neq M$), the differential encoding reads $\ve{X}_i = \left[\ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}\right]_{[n]} \cdot \ve{S}_i$, where $\left[\ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}\right]_{[n]} \in \mathrm{GL}(n,\ensuremath{\mathbb F_{q}})$ denotes the square matrix obtained from the first $n$ columns of $\ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}$.
The restriction to source matrices $\S_i$ such that $\left[\ve{X}_{i}\right]_{[n]}$ is invertible results in a rate loss of $L_{\mathrm{dlnc}} = \frac{n}{q M}$ (cf.\ \cite{seidl2013differential}), which is negligible for sufficiently large field sizes $q$.
Accordingly, differential demodulation is done via $\left[\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}\right]_{[n]}^+ \cdot \ve{Y}_i^{}$
\section{Rank-Metric Partial-Unit-Memory Codes}\label{sec:RankMetricCodes}
\noindent
In this section, we give an overview of \emph{rank-metric} \emph{partial-unit-memory (PUM) codes}.
We start by defining rank-metric block codes in general and Gabidulin codes in particular.
Next, we review how rank-metric PUM codes are constructed in \cite{wachter2011partial} and give an idea how to decode these codes and state a sufficient decoding condition (cf.\ \cite{wachter2014convolutional,wachter2012rank}).
\subsection{Rank-Metric Codes}
\noindent
Error-correcting codes are often assessed by the Hamming distance of their codewords. However, in 1978, Delsarte \cite{delsarte1978bilinear} introduced the so-called rank metric which turns out to be practical for certain channels, like the MAMC.
The \emph{rank weight} of an element $\mathbf{a} \in \ensuremath{\mathbb F_{q^m}}^n$ is defined as $\wtR\left(\a\right) := \rk\left( \ve{A} \right) = \rk\left( \ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}}(\a) \right)$.
Using this notation, we can define the \emph{rank metric} of two elements $\mathbf{a}, \mathbf{b} \in \ensuremath{\mathbb F_{q^m}}^n$ as $\dR\left( \a, \b \right) := \wtR\left( \a - \b\right) = \rk\left( \ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}}(\a) - \ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}}(\b) \right)$.
The \emph{minimum rank distance} of a code $\code{}$ is defined as $d := \min\left\{\dR\left( \a - \b \right) : \a, \b \in \code{} \land \a \neq \b\right\}$.
As for the Hamming distance, a Singleton-like upper bound on the minimum rank distance can be given, i.e., $d \leq n - k + 1$ (cf.\ \cite{delsarte1978bilinear, gabidulin1985theory, roth1991maximum}). Codes fulfilling this bound with equality are called \emph{maximum rank distance (MRD) codes}.
A special class of MRD codes was introduced by Delsarte \cite{delsarte1978bilinear} and later independently reintroduced by Gabidulin \cite{gabidulin1985theory} and Roth \cite{roth1991maximum} and is usually called \emph{Gabidulin codes}. Their structure and known decoding algorithms have a lot in common with \emph{Reed--Solomon codes} in Hamming metric.
\begin{definition}[Gabidulin Code \cite{gabidulin1985theory}]
Let $g_0, g_1, \dots, g_{n-1} \in \ensuremath{\mathbb F_{q^m}}$ be linearly independent over $\ensuremath{\mathbb F_{q}}$. Then a $\mathcal{G}[n,k] \subset \ensuremath{\mathbb F_{q^m}}^n$ Gabidulin code is a linear code given by the following generator matrix:\vspace{-0.3cm}
\begin{align}
\ve{G}_{\mathcal{G}} =
\begin{bmatrix}
g_0 & g_1 & \dots & g_{n-1} \\
g_0^{q^1} & g_1^{q^1} & \dots & g_{n-1}^{q^1} \\
\vdots & \vdots & \ddots & \vdots \\
g_0^{q^{k-1}} & g_1^{q^{k-1}} & \dots & g_{n-1}^{q^{k-1}}
\end{bmatrix}
\end{align}
\end{definition}
A proof that these codes are MRD can be found in \cite{gabidulin1985theory}.
\subsection{Partial-Unit-Memory Codes in Rank Metric}
\noindent
We are considering convolutional codes in rank metric. In general, a convolutional code can be described by a semi-infinite block-Toeplitz generator matrix \cite{johannesson1999fundamentals}.
We only consider terminated generator matrices, which is not a restriction of generality in our case because we are dealing with finite information sequences.
Such a generator matrix of a rate $R = \frac{k}{n}$ convolutional code of memory $\mu$ is given by
\begin{align}
\ve{G} =
\arraycolsep0.1cm
\begin{bmatrix}
\ve{G}^{(0)} & \ve{G}^{(1)} & \dots & \ve{G}^{(\mu)} &\ & & \ve{0} \\[0.1cm]
& \ve{G}^{(0)} & \ve{G}^{(1)} & \dots & \ve{G}^{(\mu)} & & \\[0.1cm]
& & \ddots & \ddots & \ddots & \ddots & \\[0.1cm]
\ve{0} & & & \ve{G}^{(0)} & \ve{G}^{(1)} & \dots & \ve{G}^{(\mu)}
\end{bmatrix}, \label{eq:generator_matrix}
\end{align}
where $\ve{G}^{(i)} \in \ensuremath{\mathbb F_{q}}^{k \times n}$ for all $i$ (cf.\ \cite{johannesson1999fundamentals}).
PUM codes are convolutional codes of memory $\mu = 1$, introduced by Lee \cite{lee1976short} and Lauer \cite{lauer1979some}. It can be shown (e.g.\ \cite[Thm. 8.28]{Bossert1999}) that any convolutional code can be represented as a PUM code. However, PUM codes are usually constructed using block codes to obtain a good algebraic understanding of the convolutional code. For the following definition, let $\ensuremath{k^{(1)}}, k \in \ensuremath{\mathbb{N}}$, such that $\ensuremath{k^{(1)}} \leq k \leq n-\ensuremath{k^{(1)}}$.
\begin{definition}
A $\ensuremath{\mathcal{PUM}}(n,k,\ensuremath{k^{(1)}})$ code over $\ensuremath{\mathbb F_{q}}$ is a rate $\frac{k}{n}$ convolutional code with memory $\mu=1$ and generator matrix submatrices $\Gsub{0}$ and $\Gsub{1}$ having $\rk\big(\Gsub{0}\big) = k$ and $\rk\big(\Gsub{1}\big) = \ensuremath{k^{(1)}}$.
\end{definition}
W.l.o.g. we can assume that only the first $\ensuremath{k^{(1)}}$ rows of $\Gsub{1}$ are nonzero (otherwise we can transform the information sequence such that it has this form) and therefore we can subdivide the matrices as follows:
\begin{align}
\Gsub{0} =
\begin{bmatrix}
\Gsub{00} \\
\Gsub{01}
\end{bmatrix},
\quad
\Gsub{1} =
\begin{bmatrix}
\Gsub{10} \\
\mathbf{0}
\end{bmatrix},
\end{align}
where $\Gsub{00}, \Gsub{10} \in \ensuremath{\mathbb F_{q}}^{\ensuremath{k^{(1)}} \times n}$ and $\Gsub{01} \in \ensuremath{\mathbb F_{q}}^{(k-\ensuremath{k^{(1)}}) \times n}$.
It is shown in \cite{wachter2011partial} and \cite{wachter2014convolutional} how PUM codes can be constructed based on block rank-metric codes.
One can construct PUM codes based on Gabidulin codes by choosing $\Gsub{00}$, $\Gsub{10}$ and $\Gsub{01}$ as submatrices of generator matrices of Gabidulin code~\eqref{eq:generator_matrix}.
In particular, the submatrices are chosen such that the codes in Tab.~\ref{table:codes} are Gabidulin codes with the properties given in the table.
We also choose a \emph{bounded minimum distance (BMD)} error-erasure decoder for each of the defined codes, e.g., from \cite{koetter2008coding} or \cite{gabidulin2008error}.
\begin{table}[h]
\centering
\caption{Code definitions using sub-matrices of the PUM code generator matrix \cite{wachter2014convolutional}}
\label{table:codes}
\scalebox{0.8}{
\begin{tabular}{c:c:c:c:c}
\hline
Code
& Generator Matrix
& Type
& Minimum Rank Distance
& Decoder
\vphantom{\scalebox{1.5}{M}}\\
\hline
$\C{0}$
& $\Gsub{0} = \begin{bmatrix} \Gsub{00} \\ \Gsub{01} \end{bmatrix}$
& $\mathcal{G}[n,k]$
& $d_0 = n - k + 1$
& $\BMD{\C{0}}$
\vphantom{\scalebox{1.5}{$\begin{bmatrix} \Gsub{00} \\ \Gsub{01} \end{bmatrix}$}} \\
\hdashline
$\C{1}$
& $\begin{bmatrix} \Gsub{01} \\ \Gsub{10} \end{bmatrix}$
& $\mathcal{G}[n,k]$
& $d_1 = n - k + 1$
& $\BMD{\C{1}}$
\vphantom{\scalebox{1.5}{$\begin{bmatrix} \Gsub{01} \\ \Gsub{10} \end{bmatrix}$}} \\
\hdashline
$\C{01}$
& $\Gsub{01}$
& $\mathcal{G}[n,k - \ensuremath{k^{(1)}}]$
& $d_{01} = n - k + \ensuremath{k^{(1)}} + 1$
& $\BMD{\C{01}}$
\vphantom{\scalebox{1.5}{$\begin{bmatrix} \Gsub{01} \\ \Gsub{10} \end{bmatrix}$}} \\
\hdashline
$\C{\sigma}$
& $\Gsub{\sigma} := \begin{bmatrix} \Gsub{00} \\ \Gsub{01} \\ \Gsub{10} \end{bmatrix}$
& $\mathcal{G}[n,k + \ensuremath{k^{(1)}}]$
& $d_{\sigma} = n - k -\ensuremath{k^{(1)}} + 1$
& $\BMD{\C{\sigma}}$
\vphantom{\scalebox{1.5}{$\begin{bmatrix} \Gsub{00} \\ \Gsub{01} \\ \Gsub{10} \end{bmatrix}$}} \\
\hline
\end{tabular}
}
\end{table}
\subsection{BMD Decoding of Partial-Unit-Memory Codes}\label{subsec:BMDDecPUM}
\noindent
We consider codewords $\c$ of PUM codes of length $nN$, obtained from an information sequence $\i$ of length $k(N-1)$ by multiplication with the (terminated) PUM generator matrix~\eqref{eq:generator_matrix}. We can divide the codeword $\c \in \ensuremath{\mathbb F_{q^m}}^{nN}$ into $N$ blocks of length $n$, $\c = \big( \csub{1} \dots \csub{N}\big)$, as well as the information word $\i \in \ensuremath{\mathbb F_{q^m}}^{k(N-1)}$ into $(N-1)$ blocks of length $k$, $\i = \big( \isub{1} \dots \isub{N-1}\big)$. With that notation, we can derive the simple encoding rule:
\begin{align}
\csub{j} = \Gsub{1} \, \isub{j-1} + \Gsub{0} \, \isub{j} \quad \forall j \in \{1,\dots,N\}, \label{eq:PUMencoding}
\end{align}
with $\isub{0} = \isub{N} = \mathbf{0}$. We use the decoder to correct an additive rank error $\ve e = \left( \esub{1} \dots \esub{N}\right) \in \ensuremath{\mathbb F_{q^m}}^{nN}$, so the received word is $\r = \c + \ve e = \left( \rsub{1} \dots \rsub{N}\right) \in \ensuremath{\mathbb F_{q^m}}^{nN}$.
It is known that convolutional codes can be ML decoded using the \emph{Viterbi} algorithm \cite{viterbi1967error}. However, the complexity of this procedure depends strongly on the size of the underlying field, which defines the number of states of the respective trellis.
The necessary field size over which a Gabidulin code must be defined grows exponentially with the codelength $n$.
Therefore, Viterbi's algorithm is not a good choice for decoding PUM codes based on Gabidulin codes.
An alternative is an algorithm introduced by Dettmar and Sorger \cite{dettmar1995bounded} for decoding PUM codes over Hamming metric.
It uses the block decoders of the underlying algebraic codes to find a much smaller subgraph of the trellis which contains the most likely code sequence under a certain condition.
Afterwards, the Viterbi algorithm finds this sequence in the reduced trellis in less time than without the reduction step.
In \cite{wachter2014convolutional}, a generalization of this algorithm to rank metric was proposed. It is shown that the following bound provides a sufficient condition for successful decoding, using the rank weight distribution of the additive error $\ve e$.
\begin{align}
&\sum\limits_{h=i}^{i+j-1} t^{(h)} < \frac{\delta_j}{2}, \quad
\begin{cases}
\forall \,i \in \{0, \dots, N\} \text{ and}\\
\forall j \in \{1, \dots, N-i+1\},
\end{cases} \label{eq:BMDcondition}
\end{align}
where $t^{(h)} := \wtR \left( \ve{E}^{(h)} \right) = \wtR \left(\ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}}\left(\esub{h}\right) \right)$ and
\begin{align}
\delta_j :=
\begin{cases}
d_{01}, & \text{if } j=1,\\
d_0 + (j-2) d_\sigma + d_1, & \text{else}.
\end{cases}
\end{align}
If the sequence is sent block-by-block using a Gabidulin code with dimension $k$ instead, all error patterns containing at least one block $h$ with $t^{(h)} \geq \frac{d_0}{2}$ will not be decoded correctly. PUM codes can handle such \emph{error peaks} if $t^{(h)} < \frac{d_{01}}{2}$ and if the errors in the surrounding blocks are not too large (cf.~\eqref{eq:BMDcondition}). This property makes them suitable for DLNC in slowly-varying networks because they can handle seldom occurring network changes resulting in impulsive error peaks.
It can be shown that in terms of the \emph{sequence error probability} $\PROB_\mathrm{fail}$, i.e., the probability that the decoder fails for at least one block of the sequence, the PUM decoder is strictly better than the block-by-block decoder.
\section{DLNC Channel Model for Slowly-Varying Networks} \label{sec:Error_Model}
\noindent
We analyze the statistical behavior of the rank of the effective error matrix $\ve{E}$, cf., (\ref{Eq_E_i}).
For that purpose, we need to discuss the statistical behavior of the additive error matrix $\ve{B}$, and the effects of a slowly-varying network topology.
In order to derive an analytic expression for the probability distributions, we make the following assumptions:
\begin{enumerate}
\item \label{itm:p_n} The probability $p_\mathrm{n}$ that an error occurs at a certain node during one generation is constant.
\item \label{itm:p_d} The probability $p_{\Delta \mathrm{n}}$ that a certain node leaves or joins the network during one generation is constant.
\item \label{itm:p_e} The probability $p_\mathrm{e}$ that there is a directed edge from node $i$ to $j$ ($i \neq j$) is constant.
\item \label{itm:N} The number of nodes $|\mathcal{N}|$ is sufficiently large and can be assumed to be constant due to very slow network changes.
\item \label{itm:q} The field size $q$ is sufficiently large such that the probability that two independent errors cancel is close to zero.
\end{enumerate}
\subsection{Additive Error Matrix} \label{sec:AddErrMatrix}
\noindent
It was shown in \cite{seidl2013differential} that the rank of the additive error matrix $\ve{B}$ is approximately distributed binomially with parameters $|\mathcal{N}|$ and $p_\mathrm{n}$ if Assumption \ref{itm:q} holds.
Hence, its \emph{probability mass function} (pmf) with mean value $\mu_\ve{B} = |\mathcal{N}| \cdot p_\mathrm{n}$ is given by
\begin{align} \label{Eq_rankBDistribution0}
f_{\rk(\ve{B})}(\tau) \approx {|\mathcal{N}| \choose \tau} \cdot p_\mathrm{n}^{\tau} \cdot (1-p_\mathrm{n})^{|\mathcal{N}|-\tau}.
\end{align}
\subsection{Slow Changes of the Network Topology} \label{sec:NetChanges}
\noindent
If nodes leave or join the network between generations, the network behavior changes and can be expressed as a difference in the channel matrices
\begin{equation} \label{Eq_Delta_A_i}
\ve{A}_i = \ve{A}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}} + \Delta \ve{A}_i,
\end{equation}
where we call $\Delta \ve{A}_i$ the \emph{channel deviation}.
Since the rank of $\Delta \ve{A}_i$ turns out to be important for the effective error matrix using DLNC, we are interested in its distribution.
We start by proving an important theorem.
\begin{theorem}\label{thm:leaving_node}
If exactly one node $\nu$ with $\In\{\nu\}$ incoming and $\Out\{\nu\}$ outgoing edges leaves the network between generation $i-1$ and $i$, the rank of $\Delta \ve{A}_i$ is upper bounded by $\rk \left( \Delta \ve{A}_i \right) \leq \min\left\{n, \Out\{\nu\}, \In\{\nu\}\right\}$.
\end{theorem}
\begin{proof}
Due to the dimensions of $\Delta \ve{A}_i$, its rank is upper bounded by $n$.
Let the outgoing packets be independent random linear combinations $\ve{x}_{\mathrm{o},j}$ of the incoming edges, sent to nodes $\nu_{\mathrm{o},j}$ with $j \in \{1,\dots,\Out\{\nu\}\}$.
The channel deviation can then be interpreted as errors of value $\ve{e}_{\mathrm{o},j} := -\ve{x}_{\mathrm{o},j}$ at each node $\nu_{\mathrm{o},j}$.
By the same argument as for the additive error matrix (cf.\ \cite{koetter2008coding}), the rank of $\Delta \ve{A}_i$ is then upper bounded by the number of these additive errors, namely $\Out\{\nu\}$.
Alternatively, the channel deviation can be seen as additive errors at the origin nodes of the $\In\{\nu\}$ incoming edges. Hence, $\rk \left( \Delta \ve{A}_i \right) \leq \In\{\nu\}$.
The proof is illustrated in Fig.~\ref{fig:channel_deviation_illustration}.
\end{proof}
\begin{figure}[h]
{
\tikzstyle{N} = [draw,circle, minimum width=1.5cm, inner sep=0pt]
\resizebox{0.5\textwidth}{!}{
\begin{tikzpicture}[scale=1.8, every path/.style={>=latex}]
\def1{1}
\def1.5{1.5}
\def1.5{1.5}
\def0.7{0.7}
\def4{4}
\node[font=\Large] (title1) at (0,1) {$\rk \left( \Delta \ve{A}_i \right) \leq \Out\{\nu\}$};
\node[font=\Large] (title1) at (4,1) {$\rk \left( \Delta \ve{A}_i \right) \leq \In\{\nu\}$};
\draw[thick] (4/2,1.5) -- (4/2,-1.5-1.5-0.5);
\node[N] (i1) at (-1,0) { $\nu_{\mathrm{i},1}$ };
\node (dots1) at (0,0) { $\dots$ };
\node[N] (i2) at (1,0) { $\nu_{\mathrm{i},\In\{\nu\}}$ };
\node[N] (v) at (0,-1.5) { $\nu$ };
\node[N] (o1) at (-1,-1.5-1.5) { $\nu_{\mathrm{o},1}$ };
\node (dots2) at (0,-1.5-1.5) { $\dots$ };
\node[N] (o2) at (1,-1.5-1.5) { $\nu_{\mathrm{o},\Out\{\nu\}}$ };
\draw[->,>=latex,thick] (i1) -- (v);
\draw[->,>=latex,thick] (i2) -- (v);
\draw[->,>=latex,thick] (v) -- (o1) node[pos=0.1,left] {$\ve{x}_{\mathrm{o},1}$};
\draw[->,>=latex,thick] (v) -- (o2) node[pos=0.1,right] {$\ve{x}_{\mathrm{o},\Out\{\nu\}}$};
\node (e1) at (-0.7*1/1.5-1,0.7-1.5-1.5) {};
\draw[->, >=latex,red,thick] (e1) -- (o1) node[pos=0,above] {$-\ve{x}_{\mathrm{o},1}$};
\node (e2) at (0.7*1/1.5+1,0.7-1.5-1.5) {};
\draw[->, >=latex,red,thick] (e2) -- (o2) node[pos=0,above] {$-\ve{x}_{\mathrm{o},\Out\{\nu\}}$};
\node[N] (i1) at (4-1,0) { $\nu_{\mathrm{i},1}$ };
\node (dots1) at (4+0,0) { $\dots$ };
\node[N] (i2) at (4+1,0) { $\nu_{\mathrm{i},\In\{\nu\}}$ };
\node[N] (v) at (4+0,-1.5) { $\nu$ };
\node[N] (o1) at (4-1,-1.5-1.5) { $\nu_{\mathrm{o},1}$ };
\node (dots2) at (4+0,-1.5-1.5) { $\dots$ };
\node[N] (o2) at (4+1,-1.5-1.5) { $\nu_{\mathrm{o},\Out\{\nu\}}$ };
\draw[->,>=latex,thick] (i1) -- (v) node[pos=0.8,left] {$\ve{x}_{\mathrm{i},1}$};
\draw[->,>=latex,thick] (i2) -- (v) node[pos=0.8,right] {$\ve{x}_{\mathrm{i},\In\{\nu\}}$};
\draw[->,>=latex,thick] (v) -- (o1);
\draw[->,>=latex,thick] (v) -- (o2);
\node (e1) at (4-0.7*1/1.5-1,-0.7) {};
\draw[->, >=latex,red,thick] (e1) -- (i1) node[pos=0,below] {$-\ve{x}_{\mathrm{i},1}$};
\node (e2) at (4+0.7*1/1.5+1,-0.7) {};
\draw[->, >=latex,red,thick] (e2) -- (i2) node[pos=0,below] {$-\ve{x}_{\mathrm{i},\In\{\nu\}}$};
\end{tikzpicture}
}
}
\caption{Illustration of the proof of Theorem \ref{thm:leaving_node}.}
\label{fig:channel_deviation_illustration}
\end{figure}
The same argument holds for nodes that join the network between generations.
For a leaving or joining node $\nu$, we define $w(\nu) := \min\left\{n, \Out\{\nu\}, \In\{\nu\}\right\}$, and $\ell_i$ to be the number of leaving nodes in generation $i$.
Its pmf $f_W(w)$ can be derived using Assumption \ref{itm:p_e}.
Due to the subadditivity of the rank, we obtain the following upper bound on $\rk \left( \Delta \ve{A}_i \right)$.
\begin{align*}
\rk \left( \Delta \ve{A}_i \right) \leq \sum\limits_{\substack{\nu \text{ leaving}\\\text{in gen. } i}} w(\nu) = \sum\limits_{j=1}^{\ell_i} w(\nu_j)
\end{align*}
This means that the rank of the deviation matrix $\Delta \ve{A}_i$ is composed by two random processes. The first one determines the number $\ell_i$ of leaving/joining nodes before the current generation according to the pmf $f_L(\ell)$, which is binomially distributed with $|\mathcal{N}|$ and $p_{\Delta \mathrm{n}}$.
The second process determines the $\ell_i$ corresponding node weights $w(\nu_j)$, which are distributed according to $f_W(w)$.
Hence, the pmf $f_{\rk(\Delta \ve{A}_i)}$ is approximately given by
\begin{equation} \label{Eq_tauDeltaA}
f_{\rk(\Delta \ve{A}_i)}(\tau) = \sum\nolimits_{\ell=0}^{|\mathcal{N}|} f_L(\ell) \cdot f_W^{(*)\ell}(\tau).
\end{equation}
\subsection{Effective Error Matrix}
\noindent
Based on the insights of the last two paragraphs, we are able to bound the rank of the effective error matrix $\ve{E}_i$.
\begin{theorem}
The rank of the effective error matrix $\ve{E}_i$, which is present in a
DLNC system applied to a slowly-varying network is upper bounded by
\begin{equation} \label{Eq_theoremEi}
\rk(\ve{E}_i) \le \rk(\ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}) + \rk(\ve{B}_i) + \rk(\Delta \ve{A}).
\end{equation}
Given Assumption \ref{itm:q}, the bound is tight with high probability.
\end{theorem}
\begin{proof}
Inserting (\ref{Eq_MAMC}) and (\ref{Eq_Delta_A_i}) into (\ref{Eq_E_i})
results in
\begin{eqnarray}
\tilde{\ve{S}_i}^{} \hspace*{-2mm} &=& \hspace*{-2mm}
\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \left[(\ve{A}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} + \Delta \ve{A}_i^{}) \ve{X}_i^{}
+ \ve{B}_i^{} \right] \nonumber \\
&=& \hspace*{-2mm} \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{A}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{X}_i^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \Delta \ve{A}_i^{} \ve{X}_i^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{B}_i^{}. \label{Eq_proof_Ei_1}
\end{eqnarray}
With the aid of (\ref{Eq_pinv1}) $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} = \ve{I}_n + \ve{L} \ve{I}_\mathcal{U}^{\mathrm{T}}$, we can rearrange
$\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}} = \ve{A}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}} + \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}$ in the following way
\begin{equation} \label{Eq_proof_Ei_2}
\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^+ \ve{A}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} = \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1}
+ \ve{L} \ve{I}_{\mathcal{U}}^{\mathrm{T}} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1}
- \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1}.
\end{equation}
Combining (\ref{Eq_proof_Ei_1}) and (\ref{Eq_proof_Ei_2}) we obtain
\begin{small}
\begin{eqnarray}
\tilde{\ve{S}}_i^{} \hspace*{-2mm} &=& \hspace*{-2mm} \left( \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1}
+ \ve{L} \ve{I}_{\mathcal{U}}^{\mathrm{T}} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1}
- \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1} \right)
\ve{X}_{i}^{} \nonumber \\
&& \hspace*{-2mm} + \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \Delta \ve{A}_{i}^{} \ve{X}_{i}^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{i}^{} \nonumber \\
&=&\hspace*{-2mm} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1} \ve{X}_{i}^{}
+ \ve{L}_{}^{} \ve{I}_{n}^{\mathrm{T}} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1} \ve{X}_{i}^{}
- \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{X}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{-1} \ve{X}_{i}^{} \nonumber \\
&&\hspace*{-2mm} + \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \Delta \ve{A}_{i}^{} \ve{X}_{i}^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{i}^{} \nonumber \\
&=&\hspace*{-2mm} \ve{S}_{i}^{}
+ \underbrace{\ve{L} \ve{I}_{\mathcal{U}}^{\mathrm{T}} \ve{S}_{i}^{}
- \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{} \ve{S}_{i}^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \ve{B}_{i}^{}
+ \ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{+} \Delta \ve{A}_{i}^{} \ve{X}_{i}^{}}_{=: \ve{E}_i}.\nonumber%
\end{eqnarray}%
\end{small}%
Thus, the effective error matrix consists of four parts.
The first part corresponds to the rank deficiency of $\ve{Y}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}^{}$, which is assumed to be zero.
This assumption is justifiable as long as $q$ is sufficiently large (cf. \cite{seidl2013differential}).
The second and the third part have the same rank as the preceding and the current additive error matrices, $\ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}$ and $\ve{B}_i$, respectively.
The rank of the last part equals to $\rk(\Delta \ve{A}_i)$.
Finally, taking the subadditivity of the rank into account, we obtain \eqref{Eq_theoremEi} as an upper bound on the rank of $\ve{E}_i$.
Due to Assumption \ref{itm:q}, with high probability, the sum of the matrices has the same rank as the sum of the ranks.
\end{proof}
As a consequence, the rank $\ve{E}_i$ can approximately be described by the sum of three random variables
\begin{equation}
\rk(\ve{E}_i) = \rk(\ve{B}_{\ensuremath{i \hspace*{0.5mm} \text{--} \hspace*{0.5mm} 1}}) + \rk(\ve{B}_i) + \rk(\Delta \ve{A}),
\end{equation}
where the first two summands, which describe the effect of the additive error in the MAMC are distributed according to~\eqref{Eq_rankBDistribution0}.
The third summand describes the effect of the slowly-varying network and its pmf is given by \eqref{Eq_tauDeltaA}.
The resulting approximate pmf can be described as
\begin{equation}
f_{\rk(\ve{E}_i)}(\tau) = f_{\rk(\ve{B})}^{(*)2}(\tau) * f_{\rk(\Delta \ve{A})}(\tau) \label{eq:pdf_E}.
\end{equation}
\begin{myEx}
Fig.~\ref{fig:statistics_id10013} illustrates the influence of the additive rank error and the rank error caused by slow network changes.
The plot shows the pmfs of the ranks of the additive error $\ve{B}_i$, the channel deviation $\Delta \ve{A}_i$, the additive error without the influence of network changes $\rk(\tilde{\ve{E}}_i) := \rk(\ve{B}_i)+\rk(\ve{B}_{i-1})$ and the effective error $\ve{E}_i$.
We consider a network with $|\mathcal{N}| = 100$ nodes and probability parameters $p_\mathrm{n} = 0.03$, $p_\mathrm{e} = 0.05$ and $p_{\Delta \mathrm{n}} = 0.01$.
We have chosen $p_{\Delta \mathrm{n}}$ relatively large such that the changes are more visible in the illustration.
Plot (iii) in Fig.~\ref{fig:statistics_id10013} depicts the effect caused by the differential demodulation in case of no network changes (cf. \cite{seidl2013differential}), i.e., the expected value of the rank of the effective error is approximately doubled with respect to the additive error given by the MAMC (plot (i)).
The effect of additional slow network changes can be seen in plot (iv).
In contrast to (iii), both variance and mean are increased, making errors with high rank more likely.
These error peaks can be better handled using PUM codes than using block codes.
\vspace{-0.3cm}
\begin{figure}[h]
\definecolor{mycolor1}{rgb}{0.00000,0.00000,0.56250}%
\begin{tikzpicture}
\draw (6.4,3.85) node {(i)};
\draw (6.4,1.575) node {(ii)};
\draw (6.4,-0.7) node {(iii)};
\draw (6.4,-2.95) node {(iv)};
\small
\begin{axis}[%
width=0.39\textwidth,
height=0.1\textwidth,
area legend,
scale only axis,
xmin=-1,
xmax=31,
xmajorgrids,
ymin=0,
ymax=0.4,
xticklabels={},
ytick={0, 0.2, 0.4},
ylabel={$\fprob{\rk(\Delta \ve{A}_i)}(\tau)$},
ymajorgrids,
name=plot2,
]
\addplot[ybar,bar width=0.01\textwidth,draw=black,fill=mycolor1] plot table[row sep=crcr] {%
0 0.36889018600201\\
1 0.0231180459860032\\
2 0.0586944355852956\\
3 0.0911433918438329\\
4 0.0981777700218436\\
5 0.0822248030761639\\
6 0.0616778091358218\\
7 0.0479092074304523\\
8 0.0395346038355769\\
9 0.032288769474248\\
10 0.0253376739819835\\
11 0.0190948423706681\\
12 0.0142877946335622\\
13 0.0104795373719017\\
14 0.0077693599147917\\
15 0.00567790541638334\\
16 0.0041304360907685\\
17 0.00289948911127024\\
18 0.00205113701325663\\
19 0.00142483196033241\\
20 0.00100829818089914\\
21 0.000696589862860375\\
22 0.000485615082349968\\
23 0.000312791474356274\\
24 0.000227342475974991\\
25 0.000147670311287091\\
26 0.000104103357182831\\
27 7.05255113400449e-05\\
28 4.76588779704058e-05\\
29 2.93655712746945e-05\\
30 1.88950602054386e-05\\
31 1.25164730023288e-05\\
32 7.58209422256455e-06\\
33 7.34139281867361e-06\\
34 4.57332667392783e-06\\
35 3.00876754863673e-06\\
36 9.62805615563753e-07\\
37 7.22104211672815e-07\\
38 7.22104211672815e-07\\
39 4.81402807781876e-07\\
40 7.22104211672815e-07\\
41 2.40701403890938e-07\\
42 1.20350701945469e-07\\
43 1.20350701945469e-07\\
44 0\\
45 0\\
46 0\\
47 0\\
48 0\\
49 0\\
50 0\\
};
\addplot [color=black,solid,forget plot]
table[row sep=crcr]{%
-1 0\\
51 0\\
};
\end{axis}
\begin{axis}[%
width=0.39\textwidth,
height=0.1\textwidth,
area legend,
scale only axis,
xmin=-1,
xmax=31,
xmajorgrids,
ymin=0,
ymax=0.4,
xticklabels={},
ytick={0, 0.2, 0.4},
ylabel={$\fprob{\rk(\ve{B}_i)}(\tau)$},
ymajorgrids,
at=(plot2.above north west),
anchor=below south west,
]
\addplot[ybar,bar width=0.01\textwidth,draw=black,fill=mycolor1] plot table[row sep=crcr] {%
0 0.0460447343559131\\
1 0.146953863558409\\
2 0.227493275404529\\
3 0.229012341964485\\
4 0.169828440074377\\
5 0.100105306864202\\
6 0.0488663565630247\\
7 0.0204779126374255\\
8 0.00758040931273732\\
9 0.00253518753648131\\
10 0.000789861656868114\\
11 0.000226500021061373\\
12 6.57114832622261e-05\\
13 1.51641884451291e-05\\
14 3.36981965447313e-06\\
15 1.20350701945469e-06\\
16 1.20350701945469e-07\\
17 2.40701403890938e-07\\
18 0\\
19 0\\
20 0\\
21 0\\
22 0\\
23 0\\
24 0\\
25 0\\
26 0\\
27 0\\
28 0\\
29 0\\
30 0\\
31 0\\
32 0\\
33 0\\
34 0\\
35 0\\
36 0\\
37 0\\
38 0\\
39 0\\
40 0\\
41 0\\
42 0\\
43 0\\
44 0\\
45 0\\
46 0\\
47 0\\
48 0\\
49 0\\
50 0\\
};
\addplot [color=black,solid,forget plot]
table[row sep=crcr]{%
-1 0\\
51 0\\
};
\end{axis}
\begin{axis}[%
width=0.39\textwidth,
height=0.1\textwidth,
area legend,
scale only axis,
xmin=-1,
xmax=31,
xmajorgrids,
ymin=0,
ymax=0.2,
ytick={0, 0.1, 0.2},
xticklabels={},
ylabel={$\fprob{\rk(\tilde{\ve{E}}_i)}(\tau)$},
ymajorgrids,
name=plot3,
at=(plot2.below south west),
anchor=above north west,
]
\addplot[ybar,bar width=0.01\textwidth,draw=black,fill=mycolor1] plot table[row sep=crcr] {%
0 0.00204669877410529\\
1 0.0135444890492528\\
2 0.042498534607397\\
3 0.0879037842804545\\
4 0.13497466310707\\
5 0.163434411139366\\
6 0.1635750249697\\
7 0.139743621075078\\
8 0.103406305566257\\
9 0.0680831289315842\\
10 0.0401536608288798\\
11 0.0215898106674687\\
12 0.0106607388624329\\
13 0.00489937883072907\\
14 0.00212124252520825\\
15 0.000851542619683554\\
16 0.000331455657704919\\
17 0.000120350701945469\\
18 4.31052003906731e-05\\
19 1.32631385817456e-05\\
20 2.70175045183706e-06\\
21 1.842102580798e-06\\
22 2.45613677439733e-07\\
23 0\\
24 0\\
25 0\\
26 0\\
27 0\\
28 0\\
29 0\\
30 0\\
31 0\\
32 0\\
33 0\\
34 0\\
35 0\\
36 0\\
37 0\\
38 0\\
39 0\\
40 0\\
41 0\\
42 0\\
43 0\\
44 0\\
45 0\\
46 0\\
47 0\\
48 0\\
49 0\\
50 0\\
};
\addplot [color=black,solid,forget plot]
table[row sep=crcr]{%
-1 0\\
51 0\\
};
\end{axis}
\begin{axis}[%
width=0.39\textwidth,
height=0.1\textwidth,
area legend,
scale only axis,
xmin=-1,
xmax=31,
xmajorgrids,
ymin=0,
ymax=0.2,
ytick={0, 0.1, 0.2},
ylabel={$\fprob{\rk(\ve{E}_i)}(\tau)$},
ymajorgrids,
at=(plot3.below south west),
anchor=above north west,
xlabel={$\tau$},
]
\addplot[ybar,bar width=0.01\textwidth,draw=black,fill=mycolor1] plot table[row sep=crcr] {%
0 0.00075378837606254\\
1 0.00500990498557695\\
2 0.0161725553978579\\
3 0.0343824762500784\\
4 0.0557657258148203\\
5 0.0738581205223859\\
6 0.0853304897819184\\
7 0.090394921003887\\
8 0.0898161323730002\\
9 0.0857213839495637\\
10 0.0794393229216877\\
11 0.0711941945768746\\
12 0.0620976464192166\\
13 0.0527427126728921\\
14 0.0438309888075075\\
15 0.0355262991459153\\
16 0.0283831165649355\\
17 0.0223183008347549\\
18 0.0173140449637591\\
19 0.0131960860478045\\
20 0.00997326617927907\\
21 0.00744602424526294\\
22 0.00550346567039209\\
23 0.00402167835439819\\
24 0.00289787297327269\\
25 0.00205197946817025\\
26 0.00147576978089663\\
27 0.00104741952744174\\
28 0.000732174372447844\\
29 0.000502525584041693\\
30 0.00034938545615802\\
31 0.000244754029568694\\
32 0.000167262914336458\\
33 0.000109052472783241\\
34 7.79823425871152e-05\\
35 4.82630876169075e-05\\
36 3.61052105836407e-05\\
37 2.07543557436574e-05\\
38 1.48596274851038e-05\\
39 1.14210360009476e-05\\
40 6.26314877471319e-06\\
41 3.80701200031586e-06\\
42 3.68420516159599e-06\\
43 2.33332993567746e-06\\
44 1.22806838719866e-06\\
45 7.36841032319199e-07\\
46 4.91227354879466e-07\\
47 4.91227354879466e-07\\
48 2.45613677439733e-07\\
49 0\\
50 2.45613677439733e-07\\
};
\addplot [color=black,solid,forget plot]
table[row sep=crcr]{%
-1 0\\
51 0\\
};
\end{axis}
\end{tikzpicture}%
\caption{Influence of slow network changes on the pmf of $\rk(\ve{E}_i)$.}
\label{fig:statistics_id10013}
\end{figure}
\end{myEx}
\section{Suitable Coding Schemes}\label{sec:CodingScheme}
\noindent
In this section, we describe how PUM codes can be used to make DLNC resilient against the additive error peaks caused by network changes.
We show how PUM codes can be combined with DLNC, give an idea how to choose code parameters and present numerical results which demonstrate that better results can be obtained using PUM codes instead of ordinary block codes.
\subsection{Combining DLNC and PUM Codes}
\noindent
An information sequence $\i^{(j)} \in \ensuremath{\mathbb F_{q^m}}^{k}$ for $j=1,\dots,N-1$ has to be transmitted using PUM codes in combination with DLNC.
We first determine the corresponding codeword sequence $\c^{(j)} \in \ensuremath{\mathbb F_{q^m}}^{n}$ for $j=1,\dots,N$ using the encoding rule \eqref{eq:PUMencoding} and calculate the corresponding matrix representation sequence $\S_j := \ensuremath{\mathbf{\Phi}_{\boldsymbol{\beta}}}(\c^{(j)})$ for $j=1,\dots,N$, which we use as source symbols for the differential encoding described in Sec.~\ref{sec:DLNC}.
After the transmission, the sequence is demodulated and the result $\tilde{\S}_j = \S_j + \ve{E}_j$ ($j=1,\dots,N$) can be decoded using the PUM BMD decoder described in Sec.~\ref{subsec:BMDDecPUM}.
The error sequence $\ve{E}_j$ is distributed according to the error model derived in Sec.~\ref{sec:Error_Model}.
\subsection{PUM Code Parameter Choices}
\noindent
When designing a PUM code for DLNC in slowly-varying networks, one has several possibilities to choose the code parameters $N$ and $\ensuremath{k^{(1)}}$.
Here, we assume that $n$ and $k$ are fixed, e.g., because the desired code rate and the packet size is given.
In general, $N$ should be chosen sufficiently large, such that the rate loss of the termination of the PUM code does not play a role, i.e., $\frac{N}{N+1} \cdot \frac{k}{n} \approx \frac{k}{n}$.
It is not easy to analytically derive a good range for $\ensuremath{k^{(1)}}$ for general DLNC channels in slowly-varying networks due to the involved analytic description of the pmf of the effective rank error \eqref{eq:pdf_E}.
However, for a given network, one can use the pmf of $\rk(\ve{E}_i)$ to get an idea how to choose $\ensuremath{k^{(1)}}$.
As already mentioned, the PUM decoder described in \cite{wachter2014convolutional} is able to decode up to $\lfloor\frac{d_{01}-1}{2}\rfloor$ errors under certain conditions.
Hence, a necessary condition for good decoding results is that $\PROB\big\{\rk(\ve{E}_i) \geq \tfrac{d_{01}}{2} = \tfrac{n-k+\ensuremath{k^{(1)}}+1}{2}\big\}$ is small, or equivalently, that $\ensuremath{k^{(1)}}$ is as large as possible.
On the other hand, $\ensuremath{k^{(1)}}$ should not be chosen too large, because otherwise $d_\sigma = n-k-\ensuremath{k^{(1)}}+1$ gets too small and the decoding capabilities decrease again (cf.\ \eqref{eq:BMDcondition}).
The following figure illustrates this behavior by showing how $\PROB_\mathrm{fail}$ changes as a function of $\ensuremath{k^{(1)}}$ for given $p_\mathrm{n}=0.03$, $p_\mathrm{e}=0.05$, $p_{\Delta \mathrm{n}}=0.005$, $|\mathcal{N}|=100$, $N=50$ and $d_0 = n-k+1 =41$.
It can be seen from Fig.~\ref{fig:k_1choice} that $\ensuremath{k^{(1)}}=24$ is an optimal choice for this given parameter set.
\begin{figure}[h]
\begin{tikzpicture}
\small
\begin{axis}[%
width=0.4\textwidth,
height=1.5in,
scale only axis,
xmin=0,
xmax=30,
xlabel={$\ensuremath{k^{(1)}}$},
xmajorgrids,
ymode=log,
ymin=0.001,
ymax=1,
yminorticks=true,
ylabel={$\PROB_\mathrm{fail}$},
ymajorgrids,
yminorgrids
]
\addplot [color=blue,solid,mark=asterisk,mark options={solid},forget plot]
table[row sep=crcr]{%
0 0.224676069870161\\
2 0.151481080480831\\
4 0.101884041718067\\
6 0.0677145485609147\\
8 0.0445697472755982\\
10 0.0298264088420353\\
12 0.0196344226412196\\
14 0.0135380534713882\\
16 0.00957667118693956\\
18 0.00745719631897923\\
20 0.00607214909755127\\
22 0.00538053223852067\\
24 0.00531540179378102\\
26 0.00551832162311634\\
28 0.00615683444263274\\
30 0.0075111632272912\\
};
\end{axis}
\end{tikzpicture}%
\caption{$\PROB_\mathrm{fail}$ as a function of $\ensuremath{k^{(1)}}$.}
\label{fig:k_1choice}
\end{figure}
\subsection{Numerical Results}
\noindent
Example \ref{ex:10012} shows numerical results obtained by simulating random linear networks with given parameters $\ensuremath{|\mathcal{N}|}$, $p_\mathrm{n}$ and $p_\mathrm{e}$ (cf.\ Sec.~\ref{sec:Error_Model}).
In order to run the simulations in sufficiently short time\footnote{Instead of simulating RLNC using real networks over finite fields, we evaluated the statistical behavior of the nodes and counted the number of additive errors and leaving/joining nodes and their number of incoming and outgoing edges.}, we assume that the field size $q$ is large enough, such that independent errors cancel only with negligibly small probability, and therefore the ranks of the error matrices are very likely to be equal to the number of errors happened.
If this assumption does not hold (e.g., if $q$ is relatively small), our results are still upper bounds on the sequence error probability.
We used rank-metric PUM codes with parameters $N$, $n$, $k$ and $\ensuremath{k^{(1)}}$ and checked if the PUM decoder, described in Sec.~\ref{subsec:BMDDecPUM}, is able to correct the error pattern.
$\PROB_\mathrm{fail}$ denotes the probability that the PUM decoder fails, i.e., at least one generation of the sequence is not contained in the subgraph of the trellis.
For comparison, we also checked the cases (block-by-block) when the source symbols for every generation are encoded using Gabidulin block codes with the same code dimension $k$, both differentially and via lifting.
\begin{myEx}\label{ex:10012}
Fig.~\ref{fig:num_res_10012} shows the sequence failure probability $\PROB_\mathrm{fail}$ as a function of the network change probability $p_{\Delta \mathrm{n}}$ of a PUM decoder of a code with $N = 50$, $n = 35$, $k=15$ and $\ensuremath{k^{(1)}} = 10$ and compares it to block-by-block decoders in combination with both DLNC and lifting.
The network is assumed to have $|\mathcal{N}| = 100$ nodes and probability parameters $p_\mathrm{n} = 0.01$ and $p_\mathrm{e} = 0.02$.
It can be seen that the PUM decoder is better than the block-by-block decoder for any $p_{\Delta \mathrm{n}}$ and also improves upon the lifting approach for $p_{\Delta \mathrm{n}} \le 10^{-2}$.
\begin{figure}[h]
\begin{tikzpicture}
\small
\begin{axis}[%
width=0.4\textwidth,
height=3in,
scale only axis,
xmode=log,
xmin=1e-07,
xmax=0.1,
xminorticks=true,
xlabel={$p_{\Delta n}$},
xmajorgrids,
xminorgrids,
ymode=log,
ymin=1e-05,
ymax=1,
yminorticks=true,
ylabel={$\PROB_\mathrm{fail}$},
ymajorgrids,
yminorgrids,
legend style={at={(0.03,0.97)},anchor=north west,draw=black,fill=white,legend cell align=left}
]
\addplot [color=blue,solid,mark=asterisk,mark options={solid}]
table[row sep=crcr]{%
0.0316 0.553097345132743\\
0.01 0.0055110386103365\\
0.00316 0.000170892866236799\\
0.001 4.02795129669822e-05\\
0.000316 2.64149898093611e-05\\
0.0001 2.38461613546375e-05\\
3.16e-05 2.34445820954008e-05\\
1e-05 2.25066826842226e-05\\
3.16e-06 2.2570587028166e-05\\
1e-06 2.24247821230594e-05\\
3.16e-07 2.47386093569773e-05\\
};
\addlegendentry{PUM decoder (DLNC)};
\addplot [color=red,solid,mark=asterisk,mark options={solid}]
table[row sep=crcr]{%
0.0316 0.981563421828909\\
0.01 0.17261307732722\\
0.00316 0.0153671422463229\\
0.001 0.00282055946900861\\
0.000316 0.00107524857517985\\
0.0001 0.000690918679089266\\
3.16e-05 0.000583879502225258\\
1e-05 0.000555404910826123\\
3.16e-06 0.00054556118258681\\
1e-06 0.000541977084204848\\
3.16e-07 0.000541026577845527\\
};
\addlegendentry{block-by-block decoding (DLNC)};
\addplot [color=red,dashed]
table[row sep=crcr]{%
0.0316 0.0317388354334053\\
0.01 0.0317388354334053\\
0.00316 0.0317388354334053\\
0.001 0.0317388354334053\\
0.000316 0.0317388354334053\\
0.0001 0.0317388354334053\\
3.16e-05 0.0317388354334053\\
1e-05 0.0317388354334053\\
3.16e-06 0.0317388354334053\\
1e-06 0.0317388354334053\\
3.16e-07 0.0317388354334053\\
};
\addlegendentry{block-by-block decoding (lifting)};
\end{axis}
\end{tikzpicture}%
\vspace{-0.2cm}
\caption{$\PROB_\mathrm{fail}$ of different codes/decoders as a function of $p_{\Delta \mathrm{n}}$.}
\label{fig:num_res_10012}
\end{figure}
Fig.~\ref{fig:num_res_10012_gain} shows the gain obtained by the new coding scheme, i.e., the fraction of $\PROB_\mathrm{fail}$ of the block-by-block decoder for DLNC and the PUM decoder.
It makes clear that not only the PUM decoder is better at any $p_{\Delta \mathrm{n}}$, but especially good in a region where $p_{\Delta \mathrm{n}}$ is relatively large ($3 \cdot 10^{-4}$ to $10^{-2}$).
\begin{figure}[h]
\begin{tikzpicture}
\small
\begin{axis}[%
width=0.4\textwidth,
height=1.5in,
scale only axis,
xmode=log,
xmin=1e-07,
xmax=0.1,
xminorticks=true,
xlabel={$p_{\Delta n}$},
xmajorgrids,
xminorgrids,
ymode=log,
ymin=1,
ymax=100,
yminorticks=true,
ylabel={$\PROB_\mathrm{fail}$ gain},
ymajorgrids,
yminorgrids
]
\addplot [color=blue,solid,mark=asterisk,mark options={solid},forget plot]
table[row sep=crcr]{%
0.0316 1.77466666666667\\
0.01 31.3213333333333\\
0.00316 89.9226666666667\\
0.001 70.0246666666667\\
0.000316 40.706\\
0.0001 28.974\\
3.16e-05 24.9046666666667\\
1e-05 24.6773333333333\\
3.16e-06 24.1713333333333\\
1e-06 24.1686666666667\\
3.16e-07 21.8697247706422\\
};
\end{axis}
\end{tikzpicture}%
\vspace{-0.2cm}
\caption{Gain of $\PROB_\mathrm{fail}$ of PUM codes compared to block-by-block decoding.}
\label{fig:num_res_10012_gain}
\end{figure}
\end{myEx}
\newpage
\section{Conclusion}\label{sec:Conclusion}
\noindent
This paper extended the results of \cite{seidl2013differential} to the case of slowly-varying networks.
At first, we derived a probabilistic DLNC channel model for this case by analyzing the effects of joining/leaving nodes on the network channel matrix, and with that, on the effective error (\ref{Eq_E_i}) in a DLNC system.
Furthermore, we showed that PUM rank-metric codes are the proper error correction strategy for the situation at hand.
We confirmed our considerations by numerical simulations, and showed that in slowly-varying networks, DLNC in combination with rank-metric PUM codes outperforms the conventional, lifting-based RLNC approach.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Young Massive Clusters (YMCs, e.g., Arches, Quintuplet, Westerlund 1, RSCG, GLIMPSE-CO1:
\citealp{Figer99,Clark05,Figer06,Davies11}) are gravitationally bound stellar systems with masses
$>$10$^{4}$\,M$_{\odot}$\,and ages $<$ 100\,Myr \citep{Portegies-Zwart10}. Despite their small numbers
and short lifetimes, they appear to be the `missing link' between open clusters and globular clusters
(e.g.,\, \citealp{Elmegreen97,Bastian08,Kruijssen14c}). As such, understanding their formation may
provide the link to understanding globular cluster formation. If YMCs and even globular clusters form
in a similar way to open clusters, then our knowledge of nearby stellar systems may potentially be used as a
framework to understanding star formation across cosmic time, back to globular cluster formation
$\sim$ 10 Gyr ago.
An understanding of the formation of clusters requires observations of their natal dust and gas well
before the onset of star formation. In recent surveys of dense cluster-forming molecular clumps, one
object, G0.253+0.016, stands out as extreme \citep{Longmore12}. Located at a distance of $\sim$8.5\,kpc,
G0.253+0.016\, lies $\sim$100\,pc from the Galactic centre \citep{Molinari11}. With a dust temperature of $\sim$20~K,
effective radius of $\sim$2.9\,pc, average volume density of $>$10$^{4}$\,${\rm cm}^{-3}$, mass of $\sim$10$^{5}$\,M$_{\odot}$,
\citep{Lis94, Lis98, Lis01} and devoid of prevalent star-formation, G0.253+0.016\, has exactly the properties
expected for the precursor to a high-mass, Arches-like cluster. Despite being identified over a decade
ago, its significance as a precursor to a high-mass cluster has only recently been investigated in detail
\citep{Longmore12,Kauffmann13,Rathborne-brick-malt90,Johnston14}.
Recent work shows that the
volume density threshold predicted by theoretical models of turbulence,
which depend on the gas mean volume density and 1-dimensional turbulent Mach number, is $\sim$10$^{4}$${\rm cm}^{-3}$\,
for the solar neighbhourhood and is $\sim$10$^{8}$${\rm cm}^{-3}$\, for the high-pressure inner 200\,pc environment of the Central
Molecular Zone (CMZ; \citealp{Kruijssen14,Rathborne-pdf}). The lack of current
star formation within G0.253+0.016\, is consistent with its location in the CMZ and this environmentally dependent volume density threshold for star formation,
which is orders of magnitude higher than that derived for the solar neighbourhood \citep{Rathborne-pdf}.
Because G0.253+0.016\, is close to virial equilibrium, has sufficient mass to form a YMC through direct collapse,
and shows structure on $\sim$ 0.3\,pc scales \citep{Longmore12}, it provides
an ideal template to investigate whether or not YMCs form through hierarchical fragmentation, in a scaled up version
of open cluster formation. Detailed studies of clumps like G0.253+0.016\, have the potential to reveal the gas structure in protoclusters
before star formation has turned on and significantly disrupted its natal material. If YMCs like the Arches are the missing
link between globular clusters and more typical open clusters, the detailed study of protoclusters like
G0.253+0.016\, may reveal the initial conditions for star formation not only within high-mass clusters, but potentially
for all clusters.
Both theoretical studies \citep{elmegreen08,kruijssen12d} and simulations \citep{bonnell08,krumholz12b,dale13}
suggest that high-mass clusters form through the hierarchical merging of overdensities in the interstellar medium. While some
of this hierarchical structure may also be evident in the initial stellar distribution \citep{maschberger10,hopkins12b,Parker14},
it can be easily erased over long time-scales, either via the structure's dispersal (if unbound) or relaxation (if bound). In
either case, these processes are expected to give rise to spherically symmetric stellar clusters and a
background field star population \citep[e.g.][]{Lada03}. Notwithstanding this theoretical picture, numerical studies
of the early phase of cluster evolution often model the gas as a spherically-symmetric background potential
\citep[e.g.][]{goodwin06,baumgardt07}, although notable exceptions exist \citep[e.g.][]{moeckel10,Kruijssen12e,Parker13}.
This disparity between cluster formation theory, which suggests a highly structured initial state, and the early evolutionary modelling, which usually posit a smooth spherical initial state can be resolved by characterising the
observed gas structure within high-mass protoclusters before star formation has begun.
The outcome of the high-mass cluster formation process is well-known. Observations show that YMCs are
centrally condensed with $>10^{4}$\,M$_{\odot}$\, of stars in their central $\sim 1$~pc. Thus, the puzzle is then:
is this central condensation of stars set by the initial gas structure or do they relax into this small,
centrally concentrated region over time?
To answer this question and to distinguish between the models, requires observations of protocluster clumps
with a significant improvement in angular resolution, dynamic range, and sensitivity compared to those obtained
to date. Fortunately, the new Atacama Large Millimeter/Sub-millimeter Array (ALMA) provides the necessary
improvements and allows us, for the first time, to image protoclusters across the Galaxy with comparable
physical resolution (i.e., $\lesssim$ 0.1\,pc) to clouds in the nearby solar neighbourhood.
Our goal is to determine the location, mass, and kinematics of the small-scale fragments within G0.253+0.016. The
new ALMA data reveal the internal structure and kinematics in this high-mass protocluster down to spatial
scales of 0.07\,pc (Fig.~\ref{three-colour-image}).
As such, these data can be used to determine how the material is distributed within this protocluster.
\section{ALMA observations}
We imaged the 3\,mm (90\,GHz) dust continuum and molecular line emission across G0.253+0.016\, using
25 antennas as part of ALMA's Early Science Cycle 0. Because ALMA's field of
view at this wavelength is $\sim$ 1.2' and the cloud is large on the sky (3' $\times$ 1'),
a 13-pointing mosaic was obtained to cover its full extent.
Using ALMA's Band 3, the correlator was configured to use 4 spectral windows in
dual polarization mode centred at 87.2, 89.1, 99.1 and 101.1\,GHz each with 1875\,MHz
bandwidth and 488\,kHz (1.4$-$1.7\,km~s$^{-1}$) channel spacing. With this setup we obtained the
3\,mm dust continuum emission and also covered a number of excellent tracers of dense
molecular and shocked gas, in particular: \hcnnt, \hcopnt, \cchnt, \siont, \sont, H$^{13}$CO$^{+}$, H$^{13}$CN,
\hntcnt, \hcctcnnt, \hncont, \htcsnt\, \chtshnt, \nhtchont, \chtchont, \nhtcnnt, and \hcfnnt\, (for brevity
in the text, we refer to the molecular transitions simply by the name of the molecule, e.g., \hcn\,
will be called `\hcnnt'; see Table~\ref{lines} for the quantum numbers of the transitions, frequency,
critical densities, and excitation energies).
To build up the signal to noise and improve the $uv$-coverage, G0.253+0.016\, was imaged on six
different occasions over the period 29 July -- 1 August 2012. We used the ALMA early science
`extended configuration'; the projected baselines range from 13--455 m.
All data reduction was performed using the CASA and Miriad software packages. Each individual
dataset was independently calibrated (by our ALMA support scientist) before being merged together.
We performed the cleaning and imaging to produce the final data products. Because
the emission from adjacent 1.7\,km~s$^{-1}$\, spectral channels are correlated, the final data cubes were
resampled to have a velocity width of 3.4\,km~s$^{-1}$\, per channel. The ALMA data have an
angular resolution of $\sim$1.7\arcsec\, ($\sim$0.07 pc), with a 1$\sigma$ continuum sensitivity of $\sim$
25$\mu$Jy beam$^{-1}$ and line brightness sensitivity of 1.0 mJy beam$^{-1}$ per 3.4\,km~s$^{-1}$\, channel.
\section{Combination of the ALMA interferometric data with single dish data}
Because the interferometric ALMA data are insensitive to emission on the largest angular scales ($>$ 1.2\,\arcmin),
where possible we used single-dish data to fill in the missing zero-spacing information. Given that the
emission from the clump is extended (3\arcmin$\times$1\arcmin) and the emission
from the dust continuum and many of the molecular transitions on spatial scales $>$ 1.2\arcmin\,
significant, it was critical that the zero-spacing information was included to properly sample the $uv$-plane
(e.g., \citealp{Stanimirovic02}).
The combination of the ALMA data with the zero-spacing information was performed in the Fourier
domain and was achieved using the corresponding single-dish data as a model during the CLEANing
process. The inclusion of the single-dish data in this manner is superior to using alternatives task such
as `feathering' (see \citealp{Ott-CASAGuide} for details). For this technique to be applicable there must
be an overlapping region of spatial frequencies between the two datasets. Fortunately, for both the
continuum and line datasets, existing single-dish data provided the necessary coverage for inclusion
with the ALMA data.
Ideally, the single dish data should have the same signal-to-noise as the interferometric data that they
are being combined with. Although the signal-to-noise of the single dish data is lower than that of the
ALMA data, it is high enough that the gain from recovering the zero-spacing information outweighs the
increased noise that they may introduce.
To recover the zero-spacing information for the ALMA line cubes (\hcnnt, \hcopnt, \cchnt, \siont, \hncont),
we used data from the MALT90 survey which were obtained using the Mopra 22\,m telescope
\citep{Foster-malt90,Jackson-malt90,Rathborne-brick-malt90}. To convert these data from T$^{*}_{A}$\, to the flux $F_{\nu}$ in
Jy\,beam$^{-1}$, we used a scaling factor of 23 K/Jy\,beam$^{-1}$, calculated via the equation
T$^{*}_{A}$ = 1.36\,$\lambda^{2}$\, $F_{\nu}$ /($\eta$\,$\theta^{2}$), using a FWHM beam size, $\theta$, of
38\arcsec, wavelength, $\lambda$, of 3\,mm, and a main beam efficiency, $\eta$, of 0.49 \citep{Ladd05}.
For the continuum image, since we do not have a direct measurement of the emission at 3\,mm on
the larger angular scales, it was necessary to estimate the flux at 3\,mm. To achieve this we scaled
an image of the dust continuum emission at 500 \,$\mu$m\, ({\it{Herschel}}, 33\arcsec\, angular resolution)
to what is expected at 3\,mm assuming that the flux scales like $\nu ^{(2+\beta)}$, where $\beta$
was assumed to be 1.2. {\it{Herschel}} data were used for this purpose as these data provide the most reliable recovery of the large
scale emission. We choose to scale the 500 \,$\mu$m\, {\it{Herschel}} image as it is the closest in
wavelength to 3 mm, minimising the effect of the assumption of $\beta$ (see \citealp{Rathborne-pdf} for more details).
Figure~\ref{alma-sd} compares the single dish, ALMA-only, and combined (ALMA and single dish)
images for both the continuum and the \hncont\, integrated intensity. For the 3\,mm dust continuum emission,
the ALMA-only image recovers $\sim$18\% of the total flux. The images clearly show the
significant improvement in angular resolution provided by ALMA and the recovery of the emission on
the large spatial scales provided by the inclusion of the zero-spacing information.
\section{The small-scale structure within a cluster forming clump revealed}
\subsection{Continuum emission}
Figure~\ref{three-colour-image} shows a two colour image of G0.253+0.016\, combining infrared images from the {\it Spitzer Space Telescope}\,
(24\,$\mu$m, $\sim$6\arcsec\, angular resolution, shown in cyan) with the new millimetre image from ALMA (3\,mm,
$\sim$1.7\arcsec\, angular resolution, shown in red). Contours show single dish dust continuum emission
(JCMT 450\,$\mu$m\,, $\sim$ 7.5\arcsec\, angular resolution). Identified as an infrared dark cloud (IRDC), G0.253+0.016\, is clearly seen as a prominent
extinction feature in mid-IR images. Because it is so cold, its thermal dust emission peaks at millimeter/sub-millimeter
wavelengths. Moreover, since the 3\,mm dust continuum emission is optically thin, internal structure is easily revealed
by the ALMA images. While the overall morphology of the dust continuum emission revealed by ALMA matches well the
mid-IR extinction and the large-scale dust continuum emission provided by the single dish image, the ALMA image shows
significant structure on the small-scale. The emission is neither smooth nor evenly distributed across the clump; it is highly
fragmented and there is a clear concentration of emission toward its centre. These small-scale fragments are presumably
the cores from which stars may form.
Because the 3\,mm continuum emission is optically thin , given standard assumptions about the dust emissivity, it
can be used to determine the column density and, hence, mass of its small-scale fragments. Recent observations from
{\it{Herschel}} of the 170--500 \,$\mu$m\, dust continuum
emission, show that the dust temperature (T$_{dust}$) within G0.253+0.016\, decreases smoothly from 27 K
on its outer edges to 19 K in its interior \citep{Longmore12}. External to G0.253+0.016\,, the dust temperatures
are considerably higher, $>$35 K. Within the cloud and on the spatial scales probed by the
{\it{Herschel}} data ($\sim$33\arcsec, $\sim$1 pc), no small volumes of heated dust are apparent.
All of the millimetre emission detected by ALMA falls within the
region for which T$_{dust}$\, was measured to be $<$ 22\,K. Thus, we assume that the
temperature of the dust detected by ALMA ranges from 19--22\,K and use a value of 20
$\pm$ 1\,K for all calculations.
Assuming this dust temperature, a standard value of the
gas to dust mass ratio of 100, a dust absorption coefficient ($\kappa_{3mm}$) of 0.27 cm$^{2}$ g$^{-1}$ (using
$\kappa_{1.2mm}$ = 0.8 cm$^{2}$ g$^{-1}$, and $\kappa \propto \nu^{\beta}$ \citealp{Chen08,Ossenkopf94},) and a
$\beta$ of 1.2 $\pm$ 0.1 , the intensity of the 3\,mm emission can be converted into the H$_{2}$\, column density, N(H$_{2}$), by multiplying
the measured fluxes by a single conversion factor of 1.9 $\times$10$^{23}$\,${\rm cm}^{-2}$\,/mJy. We assume that
T$_{dust}$\, is constant across the cloud. For the small range of measured dust temperatures in G0.253+0.016, this
approximation is well justified. Given these assumptions, the uncertainties for T$_{dust}$\, and $\beta$ introduce an
uncertainty of 10\% for the column density, volume density, mass, and virial ratio.
The column densities across G0.253+0.016\, measured from the ALMA-only image range from 0.5 - 82 $\times$ 10$^{22}$\,${\rm cm}^{-2}$, with a
mean value of 1.0 $\times$ 10$^{22}$\,${\rm cm}^{-2}$. Since this image detects only the highest contrast peaks within the cloud and does not
include the large-scale emission, we also determine the column density distribution using the combined (ALMA+SD) image which is
more representative of the total column of material associated with G0.253+0.016. When including the emission on all spatial scales, we
find that the mean column density is $\sim$ 8.6 $\times$ 10$^{22}$\,${\rm cm}^{-2}$\, and an order of magnitude higher compared to molecular clouds in the
solar neighbourhood (see \citealp{Rathborne-pdf} for a detailed discussion of the column density distribution across G0.253+0.016\, and its
comparison to solar neighbourhood clouds).
\subsection{Molecular line emission: spectra, moment maps, channel maps, position velocity diagrams}
The 3\,mm (90\,GHz) spectrum is rich in molecular lines that span a range in their excitation energies
and critical densities (see Table~\ref{lines}). As such, these observations also provide data cubes
from many molecular species and reveal a wealth of information about the gas kinematics and chemistry
within G0.253+0.016. Figure~\ref{integrated-intensity-images} shows, along with the dust continuum image, the
integrated intensity images for each of the detected molecular transitions. Clearly seen
are the differences in the morphology of the emission from the various tracers, indicating that the chemistry and
excitation conditions vary considerably within the clump.
To quantify the differences in the morphology of the integrated intensity for each molecular transition compared to the others
and the dust continuum emission, we have determined the 2-D cross-correlation function for all combinations of
image pairs. Figure~\ref{corr} shows the derived correlation coefficients for each transition with respect to all
others (the lower half of the figure shows the correlation coefficients via a colour scale, the upper half of the figure shows the
derived values). For each particular tracer (labelled along the diagonal) the molecular transition for which it has the highest
correlation coefficient can be found by searching for the darkest square either in the row to its left or in the column below
(the corresponding values are mirrored in the upper portion of the figure).
We find that the morphology of the dust continuum emission is not highly correlated with any gas tracer,
but is most correlated with the morphology of the emission from the molecules with high excitation energies (\nhtchont\,
in particular). In addition, we find that the morphology of the emission from \hcnnt\, and \hcopnt\, are highly correlated
with each other and little else, that the shock tracers (\siont\, and \sont) are correlated with each other, and that the dense
gas tracers with high excitation energies are well correlated with each other.
Figure~\ref{spectra} shows example spectra from the detected molecular
transitions at several locations across the clump (spectra from a single pixel were extracted at the positions
marked on Fig.~\ref{integrated-intensity-images} and listed in Table~\ref{spectra-table}). The spectra clearly
show the complexity of the emission and the differences in intensity and velocity detected from the various tracers.
Molecular species with lower excitation temperatures (E$_{u}$/$k$ $\sim$4\,K, e.g.,\, \hcnnt, \hcopnt, \cchnt)
show emission over a wider range in velocity ($\sim$ 10--15\,km~s$^{-1}$) compared to the isotolopogues
(e.g.,\, H$^{13}$CN, H$^{13}$CO$^{+}$, \hntcnt, and \hcctcnnt; $\sim$3 --10\,km~s$^{-1}$) )
and those with higher excitation temperatures (E$_{u}$/$k$ $>$10\,K, e.g.,\, \chtshnt, \nhtchont, \chtchont, \htcsnt, \nhtcnnt, and \hcfnnt; $\sim$ 3\,km~s$^{-1}$ ).
The shocked gas (traced by \siont\, and \sont) also shows emission over a wide range in velocity towards most locations in the clump ($\sim$ 3--20\,km~s$^{-1}$) .
For each molecular transition we characterise the emission via a moment analysis using the definitions of
\cite{Sault95}. All moments were calculated over the velocity range $-$10\,km~s$^{-1}$\,$<$V$_{LSR}$\,$<$\,70\,km~s$^{-1}$\,
(this range is delineated by dotted vertical lines on the spectra). The moment analysis allows us to characterise the emission at every point across
the map and to easily compare the emission from the various molecular transitions to each other and to the
dust continuum emission.
We compute the zeroth moment (M$_{0}$, or integrated intensity), the first moment (M$_{1}$,
the intensity weighted velocity field), and the second moment (M$_{2}$, the intensity weighted velocity
dispersion, $\sigma$). These are defined as
\[M_{0} = \int T_{A}^{*}(v)\,dv, \hspace{8mm}
M_{1} = \frac{\int T_{A}^{*}(v)v\, dv}{\int T_{A}^{*}(v)\,dv}, \hspace{8mm}
M_{2} = \sqrt{\frac{\int T_{A}^{*}(v) (v-M_{1})^{2}\,dv}{\int T_{A}^{*}(v)\,dv}},\]
\noindent where T$^{*}_{A}$$(v)$ is the measured brightness at a given velocity $v$. The peak intensity of the
emission was also determined via a three-point quadratic fit to each spectrum
(defined as moment $-$2; \citealp{Sault95} ). Note that \cchnt\, has well separated hyperfine
components which will affect the first and second moment determinations (HCN also has hyperfine
components, but their separation is less than the observed linewidth for HCN toward G0.253+0.016).
To include only the highest signal-to-noise emission in the moment analysis, the peak intensity, intensity
weighted velocity field, and the intensity weighted velocity dispersion were only computed at positions
across the clump where the emission was above an intensity threshold. For the emission from the
molecular species that are the brightest and include the large-scale emission from the single dish data
(i.e., \hcnnt, \hcopnt, \cchnt, \siont, and \hncont) we used a threshold in the
integrated intensity of $>$ 20\% of the peak. Since the noise level is different and the extended emission
is included in these cases, this threshold was chosen so that the moments were calculated only when the emission
was significantly detected above the noise. The choice of this threshold was visually checked (the region over which the moments were
calculated are clearly outlined by the edges of the moment maps) to ensure that it was appropriate for the moment calculation.
For the fainter lines with a lower signal to noise that do not include the
large-scale emission, we used a threshold in the peak intensity of $>$ 3.8 mJy\,beam$^{-1}$\, ($\sim$5$\sigma$).
At locations within the clump where the emission arises from well separated velocity features within the range $-$10\,km~s$^{-1}$\,$<$V$_{LSR}$\,$<$\,70\,km~s$^{-1}$,
the first moment (intensity weighted velocity field) will result in an intensity weighted average of their individual
velocities. In addition, the second moment (intensity weighted velocity dispersion) will clearly be larger than the
velocity dispersion of the individual velocity components. Thus, while this analysis may not describe
well every individual velocity component, it does provide a method by which to easily
characterise and compare the emission from the various tracers and the dust continuum across the clump.
Figure~\ref{HCN-moments} shows the integrated intensity image (zeroth moment, M$_{0}$), peak intensity
image, intensity weighted velocity field (first moment, M$_{1}$), and also the intensity weighted velocity
dispersion (second moment, M$_{2}$) derived from the \hcnnt\, emission toward G0.253+0.016\, (all other lines
are included on-line in Figs.~\ref{HCO+-moments}--\ref{HC5N-moments}). Included on the latter three images
are contours of the peak intensity: these highlight the small regions within the clump that are brightest and are
included to aid in the comparison to the velocity field and dispersion images. All images showing the velocity
field and dispersions use the same range of values for the colour coding (i.e., from $-$20 to 60\,km~s$^{-1}$\, and from
0 to 20\,km~s$^{-1}$\, respectively). This common colour coding for the different molecular transitions allows easy
comparison and readily highlights differences between them.
The images of the velocity field clearly show the well known large-scale velocity gradient across the clump; most
tracers of the dense gas show blue-shifted emission towards more positive declinations and red-shifted emission
towards more negative declinations (i.e., \siont, \sont, \hncont, and \htcsnt). Some of the more complex molecules
(\nhtchont, \chtchont, and \htcsnt) are also consistent with this trend. In contrast to the molecular
transitions with lower excitation energies, the transitions from the more complex molecules also reveal emission
toward the clump's centre at $\sim$20\,km~s$^{-1}$\, (coloured green in these images).
The obvious exception to these trends is seen in the emission from both \hcnnt\, and \hcopnt, which are clearly
red-shifted across the whole clump and with respect to all other tracers. The trend that these optically thick gas tracers
are systematically red-shifted with respect to the optically thin and hot gas tracers, indicates that G0.253+0.016\, exhibits
radial motions (either expanding or contracting, depending on the relative excitation temperatures of the inner and
outer portions of the cloud, see \citealp{Rathborne-brick-malt90} for an in-depth discussion).
The trend of red-shifted \hcnnt\, and \hcopnt\, emission seen in the single-dish data remains in the interferometric data
even when the single dish data are not included. Thus, radial motions are evident on all spatial scales.
Figure~\ref{HCN-channels} shows the channel maps for \hcnnt\, (all other channel maps are included on-line in Figs.~\ref{HCO+-channels}--\ref{HC5N-channels}).
The panels show the emission from each spectral channel (the corresponding velocity is labelled in the top left corner of each panel) toward positions
within the image where the emission was above the previously-defined intensity threshold. The channel maps also
show well the clump's large-scale velocity gradient. On the smaller scales, they reveal a complex network of emission
features with a complicated velocity structure: there is emission on all spatial scales, the morphology of which ranges
from small, compact regions to extended, filamentary structures. Toward many of the filamentary features we find
homogeneous velocity fields, indicating that they are in fact physically coherent structures.
Surprisingly, we find filamentary features in both emission and in absorption. The absorption filaments are evident
in the \hcnnt\, and \hcopnt\, integrated intensity images and channel maps (i.e., Figs.~\ref{HCN-moments}, \ref{HCO+-moments}
and Figs.~\ref{HCN-channels}, \ref{HCO+-channels} respectively). The most intriguing absorption filaments are those seen in \hcopnt\ that
extend for more than 20\,km~s$^{-1}$\, in velocity. A detailed analysis of these features is discussed in \cite{Bally-filaments}.
Position-velocity diagrams were also generated for each of the detected molecular transitions (Fig.~\ref{pv-diagrams}).
The major axis of the clump is defined along the 0-offset position in Right Ascension. The position-velocity diagrams
were made by averaging the emission over the clump's minor axis (i.e., in Right Ascension) and show the emission
along the major axis (Declination) as a function of velocity. These diagrams also show the complexity in the emission
and the systematic red-shift of the \hcopnt\, and \hcnnt\, emission.
\section{Statistical characterisation of the emission}
Given the complexity of the clump's structure and velocity field, we have employed several statistical methods in
order to characterise the observed emission from the dust continuum and the various molecular transitions. Our goal is to
answer the questions: How correlated is the emission from the various tracers with respect to one another? Which
molecular line, if any, traces well the dust column density? Does turbulence set the observed structure? What is the
structure of the gas; is it filamentary and hierarchically distributed or smooth and centrally condensed? Answers to
these questions will reveal both the physical conditions within G0.253+0.016\, and the mechanisms that are shaping its
internal structure.
\subsection{Molecular `families' tell the same story}
The combination of molecular transitions covered by these observations trace distinct chemical and physical
conditions. For comparison, we group the molecules into four `families': (1) those that have `low' excitation energies
E$_{u}$/$k$ $\sim$ 4\,K (i.e., \hcnnt, \hcopnt, \cchnt), (2) those that are tracers of shocked gas
(i.e., \siont\, and \sont), (3) the isotopologues (i.e., H$^{13}$CN, H$^{13}$CO$^{+}$, \hntcnt, and \hcctcnnt), and (4) those that have `high'
excitation energies E$_{u}$/$k$ $\gtrsim$ 10\,K (i.e., \hncont, \chtshnt, \nhtchont, \chtchont, \htcsnt, \nhtcnnt, and \hcfnnt).
Figures~\ref{compare_dust_lines0}--\ref{compare_dust_lines2} compare the derived parameters from the moment
analysis to the dust column density. These plots were generated by extracting the values from the moment maps and
column density image at every pixel where there were non-zero values in both images . Because the number of plotted
points within each panel is high, the data were binned so that a contour map of the distribution could be made.
Contours show the overall distribution and location of the most common values (the contour
levels are shown at 10 to 90\% in steps of 10\% of the peak number; the dotted contour outlines the 1\% level).
From top to bottom the rows show the normalised integrated intensities, normalised peak intensities,
intensity weighted velocity field, and intensity weighted velocity dispersion as a function of dust column density. For
ease of comparison, the molecular transitions are grouped according to the molecular `families'. Within each group
the molecular transitions are ordered (from left to right) by increasing excitation energies (see Table~\ref{lines} for the
specific values).
Figure~\ref{compare_dust_lines0} shows the plots for \hcnnt\, and \hcopnt. For comparison, we show the same
plots made including and excluding the large-scale emission (ALMA+SD and ALMA only; left and right respectively) to
show the effect of its inclusion when calculating the intensity weighted moments. As expected, when including the
large-scale emission, we find a clear correspondence between \hcnnt\, and \hcopnt\, and the lower dust column density material
(i.e., at column densities $<$ 5 $\times$ 10$^{22}$\,${\rm cm}^{-2}$). These molecules
clearly show an absence of blue-shifted emission, compared to most others (compare to Figs.~\ref{compare_dust_lines1}
and \ref{compare_dust_lines2}).
When the large-scale emission is included, the measured intensity weighted velocity dispersion is more centrally peaked
(because the emission is widespread, with a broad line profile, the `intensity weighting' will tend to result in a similar calculated
value for the intensity weighted velocity despite the emission being detected over a broad range). Moreover, a clear trend is
evident: the measured velocity dispersions decrease as the dust column density increases. This trend is not evident in the
plots made with the ALMA-only data because, in this case, there is no emission from the molecular transitions at the same
location as the low column density emission, which is required to see the trend.
The shock tracers (SiO and SO) show bright and widespread emission in both position and velocity across G0.253+0.016\, (Fig.~\ref{compare_dust_lines1}).
Moreover, their measured velocity dispersions extend to $\sim$ 20\,km~s$^{-1}$, which is higher than most other tracers and expected since SiO and SO
trace gas that is stirred up in regions of shocks. When including the large-scale emission, \siont\, shows the opposite
trend in its velocity dispersion compared to
\hcnnt\, and \hcopnt: its velocity dispersion increases as the dust column density increases, consistent with
increased shocks in the higher density material. For SO, while its typical measured dispersion is low ($\sim$
2\,km~s$^{-1}$), there is a significant tail in this distribution up to dispersions of $\sim$ 20\,km~s$^{-1}$.
In contrast to SiO, there is no clear trend for an increasing SO velocity dispersion with increased
dust column density: SO shows low velocity dispersions regardless of the column density.
The differences in the velocity dispersions measured for SiO and SO could be understood in
terms of the differences in their excitation energies and critical densities : while SO has a higher excitation energy
compared to SiO, it has a
critical density which is an order of magnitude lower. Thus these species may in fact be tracing different material within the clump.
There are many filamentary features found within G0.253+0.016\, that are detected in SiO and SO, that have high velocity dispersions, and
large steps in velocity over small spatial regions (i.e., linear features seen in the moment maps in Figs.~\ref{SiO-moments}
and \ref{SO-moments}). Several of these are also evident in the channel maps of H$^{13}$CN\, (at LSR velocities of
$\sim$ 10--17\,km~s$^{-1}$; see Figs.~\ref{SiO-channels}, \ref{SO-channels}, \ref{H13CN-channels}). They likely correspond to
shock fronts within the clump: their details will be presented in a future paper.
The isotopologues and the molecular transitions with higher excitation energies (e.g.,\,H$^{13}$CN, H$^{13}$CO$^{+}$, \hntcnt,
\hcctcnnt, \hncont, \chtshnt, \nhtchont, \chtchont, \htcsnt, \nhtcnnt, and \hcfnnt; Figs.~\ref{compare_dust_lines1} and \ref{compare_dust_lines2}),
are detected over a much smaller range in dust
column densities and are more sharply peaked at, or slightly above, the mean value of the dust column density ($\sim$ 8.6
$\times$10$^{22}$ \,${\rm cm}^{-3}$, shown as the vertical dotted line) compared to the molecules with lower excitation
energies\footnote{With the exception of \hncont, these plots were made using the ALMA only data and, thus, will not
include the lower column densities traced on the large scale by the dust emission. Given the conditions needed to excite
these transitions, large-scale emission is not expected.
Indeed, their data do not show obvious
effects that would indicate they suffer from the lack of the zero spacing information in a similar way to the data from the
molecular transitions with lower excitation energies and the dust continuum very clearly do (i.e.\, large negative `bowls' around bright emission).}.
Their integrated intensity distribution appears bi-modal,
indicating the presence of two distinct integrated intensity values. Their measured velocities span from 0 -- 50\,km~s$^{-1}$;
this velocity distribution is broader compared to that derived from \hcnnt\, and \hcopnt\footnote{The distribution in measured
velocities are narrower for \hcnnt\, and \hcopnt\, compared to many of the other transitions because they include the
large-scale emission from the single-dish data which, due to the `intensity weighting' of the moment calculation and
in the case of widespread emission with a broad line profile, will tend to result in a similar calculated value for the
intensity weighted velocity despite the emission being detected over a broad range. See Figure~\ref{compare_dust_lines0}
for a comparison of these plots for both \hcnnt\, and \hcopnt\, including and excluding the single dish data to see this effect.}.
While the peak intensities for the isotopologues and the molecules with the higher excitation energies are smoothly distributed
over a similar range in column densities to their integrated intensities, their velocity dispersion distributions are typically concentrated at low
values of the velocity dispersion. With the exception of \hncont\, which is far more abundant than these other molecules,
we find that the emission predominantly arises from very narrow velocity features ($\sigma \lesssim$
2\,km~s$^{-1}$). Such narrow velocity dispersions have also been observed toward G0.253+0.016\, on small spatial scales from other molecular
tracers (i.e. N$_{2}$H$^{+}$ (3--2); \citealp{Kauffmann13}). The tail in these distributions to considerably higher velocity dispersions
(dotted contours), indicates positions within the clump where the moment analysis has overestimated the velocity dispersion due to
multiple velocity components within the spectrum. This occurs in fewer than 1\% of the spectra.
The narrow range in dust column density over which the higher excitation energy molecular transitions are detected can be understood in terms
of their critical densities. These molecules are typically detected
when the gas has either a high temperature (T$_{gas}$$>$100\,K) or a high density (n $>$ n$_{crit}$), or both. Because
these conditions are needed for both their formation and excitation, the presence of these molecules can be used to
pinpoint gas with these physical conditions. These molecular transitions all have critical densities of $\sim 10^{5} -
10^{7}$\,${\rm cm}^{-3}$\, (see Table~\ref{lines}). Thus, while the gas temperature may be low toward the clump's centre,
the volume density may be very high. If the volume densities are sufficiently higher than their critical densities, the
molecules may be excited regardless of the gas temperature.
We see a clear trend that molecules with higher excitation energies and critical densities peak toward
the centre of the clump (compared to those with lower excitation energies and critical densities).
The position-velocity diagrams also reveal that the molecular transitions with
higher excitation energies and critical densities peak toward the centre of the cloud in both position and
velocity (declination offset of $\sim$ 0, and velocity of $\sim$ 20\,km~s$^{-1}$; e.g.\, \chtchont\, and \nhtcnnt\, in Fig.~\ref{pv-diagrams}). These
trends are exactly what is expected for a clump with a denser interior.
\subsection{Which molecules can be used to assess the dynamical state of the cores?}
While the molecular transitions detected by these observations all probe dense gas, they span a range in their
critical densities and excitation energies, trace distinct chemical conditions, and often suffer from high optical
depths. As such, they may often trace different material within G0.253+0.016\, and may not necessarily correlate well
with the optically thin dust continuum emission.
One of our key goals is to identify star-forming dust cores within the clump, to measure their sizes, masses, and virial state.
Determining these properties will allow us to ascertain the potential of G0.253+0.016\, to form a cluster and the
distribution of masses for its cores. In order to identify bound star-forming cores, we need to reliably associate
molecular line emission with each dust continuum core in order derive their velocity dispersions. Ideally, we need
to identify a molecular transition that is optically thin and is bright enough to be detected in most cores.
Given the limited velocity resolution of our data, a detailed analysis of the cores' virial states are
challenging with the current data.\footnote{The identification of discrete cores within G0.253+0.016\, from the 3\,mm
dust continuum emission, an analysis of their virial state, and derivation of the core mass function will be presented
in a future paper.} However, we can gauge, overall, which molecular transitions correlate
with the dust emission. We find that emission from the molecular transitions with higher excitation energies
typically provide a better morphological match to the dust column density compared to the
transitions with lower excitation energies, the isotopologues, or the
shock tracers (compare the morphology of the dust emission to \nhtchont, \chtchont, \hncont\,
and \htcsnt\, in Fig.~\ref{integrated-intensity-images} and also their 2-D correlation coefficients shown in
Figure~\ref{corr}). Moreover, since their velocity dispersions are typically
narrow ($\sim$ 2\,km~s$^{-1}$), these molecules may in fact be tracing individual high column density cores within the
clump with very narrow velocity dispersions.
\subsection{Is the gas structure set by turbulence?}
Supersonic turbulence is an important mechanism by which overdensities within molecular clouds
may be generated \citep{Vazquez-Semadeni94,Padoan97}. As such, understanding the degree to
which turbulence sets the initial structure within molecular clouds is an important part of understanding
core and, ultimately, star formation. Recent work using the ALMA dust continuum emission presented
here shows that the log-normal shape, dispersion, and mean of its column density probability distribution
function (N-PDF) very closely matches the predictions of theoretical models of supersonic turbulence in
gas given the high-pressure environment in which G0.253+0.016\, is immersed \citep{Rathborne-pdf}.
These results also show that while, overall, the N- PDF shows a log-normal distribution,
there is a small deviation from log-normal at the highest column densities, which
indicates self-gravitating gas. This small deviation is coincident with the location of a known water maser \citep{Lis94} and, thus,
pinpoints where star formation is beginning in the cloud. Assuming a dust temperature
of 20\,K, this small core (radius of 0.04\,pc) has a mass of $\sim$72\,M$_{\odot}$\, and virial ratio,
$\alpha_{\rm vir}$, of $\lesssim$ 1 (this is an upper limit since the associated molecular line emission
is unresolved in velocity, $\sigma <$ 1.4\,km~s$^{-1}$; see \citealp{Rathborne-pdf} for details). Since
$\alpha_{\rm vir}$ is close to unity, this core is gravitationally bound and unstable to collapse.
The fact that the N-PDF for G0.253+0.016\, matches
well predictions from turbulent theory suggests that our current understanding
of molecular cloud structure derived from the solar neighbourhood also holds in
high-pressure environments \citep{Rathborne-pdf}. This has important implications for understanding the
initial gas structure for star formation in a range of different environments across the Universe \citep{Kruijssen13}.
To further determine the extent to which turbulence shapes the observed gas structure within
G0.253+0.016, we have calculated the fractal dimension and the turbulent
power spectrum. These statistical methods are commonly used to describe the structure
of molecular clouds in both simulations and observations.
\subsubsection{Fractal structure on small spatial scales}
An important characteristic of a protocluster clump, and one that is easily revealed by these data, is the
distribution of its material on small spatial scales. If G0.253+0.016\, represents a clump that will give
rise to a high-mass cluster, then the distribution of the material within it before star formation has commenced
can provide important constraints for theories that describe cluster formation. As discussed previously, there
have been a wide range of theoretical and numerical studies showing that the formation of high-mass stellar
clusters should proceed hierarchically, but thus far the degree of hierarchal structure within high-mass protoclusters
has not been estimated.
The fractal dimension (D$_{f}$) can be used to characterise the structure within a molecular cloud. It is a
simple tool that determines the degree to which the emission is smooth and centrally condensed or structured
and hierarchical. Since observations of dust or molecular line emission reveal only the 2 dimensional projection
of a 3 dimensional structure, care needs to be taken to accurately convert the measured projected fractal
dimension (D$_{p}$) to the real three-dimensional fractal dimension (D$_{f}$). This conversion is often
assumed to be D$_{f}$ = D$_{p}$ + 1 \citep{Beech92}; however, using simulations of cloud structure,
\cite{Sanchez05} derive an empirical relation to convert between D$_{f}$ and D$_{p}$ (see their figure~8).
Rather than a single conversion factor of unity for all values of D$_{p}$, their simulations suggest that D$_{f}$
may actually increase linearly as D$_{p}$ decreases. Previous observations derive projected fractal dimensions
D$_{p}\sim$ 1.35 (see \citealp{Federrath09_Df} and references therein) leading to a quoted D$_{f}$ for molecular
clouds of $\sim$ 2.3 \citep{Elmegreen97-Fractal}.
To calculate the projected fractal dimension (D$_{p}$) for G0.253+0.016, we employ the perimeter-area method
\citep{Mandelbrot77} which is often used to measure the fractal dimension of interstellar gas clouds
(e.g., \citealp{Falgarone91,Federrath09_Df,Sanchez05}).
Structures were defined as
connected regions within a contour at a given intensity level. The area (A) was calculated as the sum of the
number of pixels within the contour, the perimeter (P) was calculated as the number of pixels externally
adjacent to the contour level. By varying the contour levels from the image minimum to its maximum,
values for P and A were determined on many spatial scales. With a power-law fit of the form
P $\propto$ A $^{D_{p}/2}$, we can determine the perimeter-area fractal dimension D$_{p}$. A smooth
structure will have D$_{p}$ = 1, while a very convoluted structure will have D$_{p}$ = 2 \citep{Federrath09_Df}.
Figure~\ref{fractal-dimension} shows the perimeter-area plots measured from contours of the dust and line
integrated intensity emission. For this analysis we use, where possible, the data that includes the large-scale
emission. Since this analysis is based on contours within an image, the effect of including the large-scale
emission simply increases the number of contours to lower levels of the intensity which translates to an increase
in the number of points plotted at large perimeters and areas. Their inclusion has little effect on the derived slope.
In all panels the solid lines show two extreme values for D$_{p}$ (1 for smooth, centrally
condensed emission, 2 for a highly fractal cloud). The dotted lines show linear fits to the data points: the
derived values for D$_{p}$ range from 1.4--1.7, depending on the molecular tracer. For the dust continuum
that includes the large-scale emission, we find a D$_{p}$ of 1.6 (plotted in the top left panel, marked
as ALMA+SD). For comparison we also calculate the fractal dimension derived from only the large-scale dust continuum
emission (in the same panel marked as SD), and find D$_{p}$ to be $\sim$1.2, consistent with
its smooth distribution.
Since the fractal dimension is best indicated by emission that is
optically thin, i.e., from the dust emission and the molecular transitions with higher excitation energies, we select
the derived D$_{p}$ from these molecules to characterise the emission from within G0.253+0.016\, (i.e., D$_{p}$ $\sim$ 1.6).
Using the empirical conversion from the two-dimensional fractal dimension D$_{p}$ to the three-dimensional fractal
dimension D$_{f}$ from \cite{Sanchez05}, we find that D$_{f}$ $\sim$ 1.8 for G0.253+0.016. Simulations of compressively
driven turbulence in the ISM \citep{Federrath09_Df} with Mach numbers, $\mathcal{M}$, of $\sim$ 5.5, reveal typical values for
D$_{f}$ of 2.3--2.6 (which correspond to D$_{p}$ $\sim$ 1.3--1.5 using the empirical relation derived by \citealp{Sanchez05}) . Since
$\mathcal{M}$\, is $\sim$ 30 for G0.253+0.016, it has a much higher level of turbulence compared to most other molecular clouds.
Given this higher level of turbulence, it may be expected that the sonic length may be lower and the
degree of fragmentation might be higher in G0.253+0.016\, (i.e.
a higher D$_{p}$, and thus, lower D$_{f}$ for higher $\mathcal{M}$). While detailed simulations are necessary to determine the effect of $\mathcal{M}$\,
on the fractal dimension, our results are consistent with a marginally higher value for D$_{p}$ for G0.253+0.016\, when
compared to typical molecular clouds. This analysis suggests that the clump is highly fragmented on small spatial scales
and its structure consistent with that produced by compressively driven turbulence.
\subsubsection{Turbulent power spectrum}
The power spectrum provides a measure of how turbulence is propagated within a molecular cloud
and it identifies the preferential spatial scale, if any, at which the structure changes. On large scales
a break in the power spectrum indicates the spatial scale at which turbulent energy is injected, on the small
scales it can indicate the spatial scales at which either turbulence is dissipated or injected . Measuring the small-scale
break point of the power spectrum can therefore reveal the spatial scale at which gravity overcomes the turbulent
pressure or the spatial scale at which energy may be injected (i.e. from protostellar outflows) .
We use the dust continuum image and several of the integrated intensity images to compute the power spectra.
Because the turbulent power spectrum is a statistical measure of the energy on various spatial scales,
it is important that the emission is well detected and well sampled across the clump. As such, we perform
this analysis using \hcnnt, \hcopnt, \hncont, and \siont\, because their data include the zero-spacing
information and the emission is detected with high signal-to-noise ratios at all spatial scales across G0.253+0.016.
Figure~\ref{power-spectrum} shows the annular averages of each of the power spectra, P(k), as a
function of the spatial frequency, k. The power spectra were generated using a Fast Fourier Transform
of the images. In all panels the dotted vertical line marks the spatial frequency corresponding to the angular
resolution of the data: the power spectra at spatial frequencies greater than this are not well sampled by these data.
In cases where the power spectrum is close to a power law (i.e., for \hcnnt, \hcopnt, \hncont, and \siont\, emission)
we fit it with a single power law to derive $\alpha$ where, P(k) $\propto$ k$^{-\alpha}$. In these cases, we
find that $\alpha \sim$ 3. A power-law slope of 3 is predicted for a line intensity power spectrum of an
optically thick medium \citep{Lazarian04, Burkhart13}. Thus, the derived power spectrum slope of $\sim$ 3
for these molecular transitions are consistent with them being optically thick toward G0.253+0.016.
In contrast, for the dust continuum emission which is optically thin, the power spectrum shows a clear break.
While this break occurs at large spatial frequencies (i.e. small spatial scales), it is lies above the spatial
frequencies corresponding to the angular resolution of these data (marked with the vertical dotted line).
For this power spectrum we fit two distinct power laws and derive their indicies and the break point. For
small values of k (i.e., large spatial scales) the derived power-law index is $\alpha \sim$ 2.9 compared
to $\alpha \sim$ 7.5 for large values of k (i.e., small spatial scales). Values for $\alpha$ of $\sim$ 2.8 are
observed toward other star-forming regions such as Perseus, Taurus and Rosetta \citep{Padoan04}.
The trend of a steeper spectrum on smaller spatial scales implies that the higher column density material is
concentrated on smaller spatial scales, while the lower column density material is distributed on larger spatial
scales. This is consistent with the distribution of the emission in the dust continuum image.
Since the effect of gravity on small scales will cause a steepening
of the power-law slope at large spatial frequencies , the spatial scale at which gravity may dominate can be determined
from the break point in the power spectrum. From the dust continuum emission power spectrum, we find that the
break occurs at spatial frequencies corresponding to size scales of $\sim$0.09\,pc. This size scale is comparable to
the size of the small region of self-gravitating gas that was identified as the deviation at high column densities via the N-PDF
\citep{Rathborne-pdf}. As such, this lends
support to the idea that gravity has overcome the internal pressure on these small spatial scales.
For cores within molecular clouds to collapse into stars, they need to be sufficiently dense that their
self-gravity overcomes the thermal pressure . The Jeans length, defined as the minimum length scale for gravitational
fragmentation to occur, was determined for G0.253+0.016\, using the relation
\begin{equation}
\lambda_{J} = 0.4\,pc \left ( \frac{c_{s}}{0.2\, km s^{-1}} \right) \left ( \frac{n(H_{2})}{10^{3}\, cm^{-3}} \right) ^{-0.5}
\end{equation}
\noindent where $c_{s}$ is the sound speed and n(H$_{2}$) the volume density.
While the measured dust temperature toward G0.253+0.016\, is $\sim$ 20\,K \citep{Longmore12}, recent observations
indicate that the gas temperature is considerably higher, $\sim$ 65\,K \citep{Ao13}, leading to the speculation that the
gas may be heated by the dissipation of turbulent energy and/or cosmic rays rather
than by photon heating. These observations are supported by SPH modelling of the dust and gas temperature
distribution in G0.253+0.016\, which suggest that the gas and dust temperatures are not coupled and that the high
interstellar radiation field (ISRF) and cosmic ray ionisation rate (CRIR) in the CMZ are responsible for the
discrepant dust and gas temperatures \citep{Clark13}.
Assuming a global gas temperature for G0.253+0.016\, of 65\,K \citep{Ao13} the corresponding gas sound speed, $c_{s}$, is 0.5\,km~s$^{-1}$.
Using an average volume density for this material of $\sim$ 10$^{5}$\,${\rm cm}^{-3}$,
we derive $\lambda_{J}$\,to be 0.10\,pc. This value is comparable to the size scale
derived from the break point in the power spectrum of the dust continuum emission
and of the size of the self-gravitating core identified via the N-PDF. Extrapolating the global
size-linewidth relation in Figure~\ref{profiles} (top right panel) to smaller scales, suggests that the line-widths
are subsonic on spatial scales of $\sim$0.1\,pc (the sonic length is $\sim$ 0.2\,pc). As such, since the sonic length exceeds
the derived break point in the power spectrum, the thermal Jeans length is indeed the relevant quantity for describing the minimum size scale for
gravitational fragmentation . An absence of
a similar break in the power spectra derived from the molecular line emission is consistent with
the transitions being optically thick and the dust continuum optically thin. Thus, the small, high-column density
features detected in the optically thin dust continuum emission may have formed because gravity was
able to overcome the thermal pressure on these small scales.
\section{Implications for the formation of young massive clusters}
\label{clusters}
Observations show that YMCs like the Arches are very centrally condensed, with stellar masses of
$>$10$^{4}$\,M$_{\odot}$\, within their central 0.5\,pc \citep{Portegies-Zwart10}. If G0.253+0.016\, is the molecular
cloud progenitor to a cluster like the Arches, then its internal structure can reveal the initial conditions
for the formation of such high-mass clusters and whether or not the central concentration of stellar
densities arises directly from the initial gas structure.
To determine its internal structure, we have calculated the mass and volume density profiles
of G0.253+0.016\, as a function of radius (left panels of Fig.~\ref{profiles}). We define each region
over which to calculate the mass and volume density using contours of column density; they
were calculated on many spatial scales by varying the contour levels from the
lowest to the highest values. An `effective radius' was determined for each region by summing the
number of pixels contained within that contour level and assuming that this area, A, forms a circle such that
$R_{eff} = \sqrt{A/\pi}$. The mass within each region was then calculated by summing the
column density, the volume density is calculated by dividing the mass by the volume of a sphere
with a radius of $R_{eff}$ (left panels of Fig.~\ref{profiles}).
Using the detected emission from the various molecular species, we have also calculated
the average velocity dispersion and virial ratio as a function of radius (right panels of Fig.~\ref{profiles}).
The virial ratio is $\alpha_{\rm vir}$, defined as $\alpha_{\rm vir}=5\sigma^2R/GM$, where $\sigma$
is the measured velocity dispersion, $R$ the radius, $G$ the gravitational constant, and $M$ the
mass. For these calculations the `effective radius' was determined from the total number of
pixels within the integrated intensity image for each of the molecular transitions.
Within these regions the total mass, $M$, was determined from the dust column density summed
over the same region, while $\sigma$ was determined by averaging the intensity weighted velocity dispersions.
The ALMA observations of the gas structure within G0.253+0.016\, show that the material is distributed in
a complex, filamentary network of structures that are highly fragmented. Its volume density profile
(lower left panel of Fig.~\ref{profiles}) shows that the average volume density increases with smaller radii, suggesting
that the material is more concentrated toward the clump's centre. We find that the volume density
profile follows very well a power law for radii $<$ 2\,pc with an index of $\sim -$1.2. In contrast to
what is expected for a Bonner-Ebert sphere and Plummer models that are often used to describe
the volume density profiles within starless cores and globular clusters, the volume density
profile for G0.253+0.016\, shows no flattening at small radii.
While the total mass of G0.253+0.016\, is $\sim$ 10$^{5}$\,M$_{\odot}$, its mass profile shows that the majority of the mass
originates from the low column density material in its outer edges, at large radii (upper left panel of Fig.~\ref{profiles}).
We find that the average velocity dispersion decreases on small spatial scales (upper panel of Fig.~\ref{profiles}),
similar to the results of \cite{Kauffmann13}. Moreover, we find that
the derived virial ratios are $>$ 1 for the molecular emission that is widespread (i.e.\,have large
effective radii) but $<$ 1 for the molecular species that are detected only over smaller
regions (lower panel of Fig.~\ref{profiles}). This trend of decreasing virial ratios as the radius decreases, implies that G0.253+0.016\, is unbound
at large radii, but bound at smaller radii. Thus, its outer envelope of lower density material may be unbound
and expanding, while the inner denser region is bound and, in the absence of other supporting
mechanisms, may be collapsing.
The notion that the G0.253+0.016\, may have a dense and collapsing interior with a low-density surface that
is expanding was posed recently by \cite{Bally-filaments} when discussing the nature and origin of the
\hcopnt\, absorption filaments seen toward G0.253+0.016. This idea is supported by the observations of the
large scale gas motions which show systematically red-shifted optically thick emission compared to the
optically thin and hot gas tracers \citep{Rathborne-brick-malt90}. In this scenario, the exterior layers of the cloud have
a lower excitation temperature compared to the inner layers of the cloud, because the emission is sub-thermally excited (n$ << $n$_{crit}$).
The expanding outer regions, coupled with a collapsing inner region, is consistent with
the scenario in which G0.253+0.016\, was formed recently as a result of
a pericentre passage close to Sgr A$^{*}$ \citep{Longmore13b,Kruijssen14d}
While a gravitationally bound and collapsing clump is consistent with what is expected in the very
early stage of mass assembly within a protocluster, we find that there is no large concentration
of mass in the centre of G0.253+0.016. Indeed, its total mass within a radius of 0.5\,pc
is only $\sim$ 6 $\times$ 10$^{3}$\,M$_{\odot}$: this is insufficient to form an Arches-like cluster by more than
2 orders of magnitude. Indeed, recent work characterising the structure within other high-mass, dense
clumps in the CMZ show that the lack of a strongly peaked central mass distribution
may be a general trend within these YMC-progenitor clouds \citep{Walker14}.
Thus, if G0.253+0.016\, is destined to form a high-mass ($>$10$^{4}$\,M$_{\odot}$) cluster,
then the structure of the resulting cluster must evolve from this initial dispersed, hierarchical structure
to a smooth, centrally condensed configuration via a dynamical process (e.g. \citealp{Allison10,moeckel10,Kruijssen12e,Parker13,Parker14}) .
As such, any model for cluster formation
that posits that stars form directly from an initially smooth, central concentration
of gas with a mass similar to that of the final stellar cluster (e.g. \citealp{Lada84a,Geyer01,Boily03,goodwin06,baumgardt07}) ,
is ruled out by these observations.
\section{Conclusions}
Using observations from the new ALMA telescope we have conducted a statistical analysis of the small-scale
structure and velocity within the high-mass protocluster clump G0.253+0.016. Its 3\,mm dust continuum and molecular line
emission reveal a complex network of emission features that have a complicated velocity structure. Filamentary
structures, that are coherent in velocity, are seen in both emission and absorption.
Widespread emission is detected from 17 different molecular transitions that probe a range in distinct physical and
chemical conditions. Comparing their intensity, velocity dispersion, and velocity fields, we find that the molecular `families'
trace the same type of material with the clump: transitions that are optically thick have broad velocity dispersions and trace
material on the large scales, while transitions that are optically thin have narrow velocity dispersions and trace the densest
material in the clump's centre. Indeed, the dust column density is best traced by molecules with higher excitation
energies and critical densities, consistent with a clump that has a denser interior.
The emission is found to be highly fragmented (D$_{f} \sim$ 1.8), consistent with structure that is produced by
compressively driven turbulence. We find a clear break in the turbulent power spectra derived from the optically
thin dust continuum emission at a spatial scale of $\sim$ 0.1\,pc. This size
scale is comparable to the Jeans length corresponding to the observed gas density and temperature, and also to
the size of the small core
of self-gravitating gas that was identified from the N-PDF. Thus, these structures
might correspond to the spatial scale at which gravity has overcome the turbulent pressure.
We speculate that G0.253+0.016\, is on the verge of forming a cluster from hierarchical, filamentary structures that arise
from a highly turbulent medium. Because the initial gas structure is fractal and hierarchical, its structure is quite
different from the smooth, compact, very centrally condensed structure observed in high-mass clusters. Consequently,
these observations rule out models of cluster formation that begin with such gas distributions.
The implication is that if stars within YMCs form from hierarchical gas structures such as these observed in G0.253+0.016,
then their distribution must change to a more centrally condensed configuration over a dynamical time scale.
Thus, any theoretical study that aims to
accurately characterise and understand the early phases of cluster formation ought to begin with an initial gas structure
that is hierarchical rather than one that is smooth and centrally concentrated.
\acknowledgements
The authors acknowledge their ALMA Contact Scientist, Dr Crystal Brogan, for assistance in preparing the observations and performing the initial
data calibration. We thank Phil Myers, Gary Fuller, Mark Krumholz, and Eli Bressert for valuable discussions.
J.M.R acknowledges funding support via CSIRO's Julius Career Award that was used to host a week-long team meeting
which laid the foundation for this work. J.M.R, S.N.L, J.M.D.K, J.B. and N.B. acknowledge the hospitality of the Aspen Center for Physics,
which is supported by the National Science Foundation Grant No.~PHY-1066293. JJ acknowledges funding support from NSF grant AS 1211844.
G.G. acknowledges support for CONICYT project PFB-06. This paper makes use of the following ALMA data:
ADS/JAO.ALMA\#2011.0.00217.S. ALMA is a partnership of ESO (representing its member states), NSF (USA) and NINS (Japan),
together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is
operated by ESO, AUI/NRAO and NAOJ.
{\it Facilities:} \facility{Mopra, ALMA, Herschel}.
\begin{sidewaystable}[h]
\caption{\label{lines} Summary of the molecular transitions detected toward G0.253+0.016$^{a}$.}
{\scriptsize{
\begin{tabular}{llrrcl}
\tableline
\tableline
& Molecular transition$^{b}$ & Frequency & E$_{u}$/$k$ & n$_{crit}$\, & Name \\
& & & & & \\
& & (GHz) & (K) & (${\rm cm}^{-3}$) & \\
\tableline
\multicolumn{6}{l}{Dense, `low' E$_{u}$/$k$:}\\
& HCN v=0 J=1--0 & 88.632 & 4.25 & 1$\times 10^{6}$ & Hydrogen Cyanide \\
& HCO$^{+}$ v=0 J=1--0 & 89.189 & 4.28 & 1$\times 10^{5}$ & Formylium \\
& HCO 1(0,1)--0(0,0), J=3/2-1/2, F= 2- 1 & 86.671 & 4.18 & $\sim$ 3$\times 10^{5}$ & Formyl Radical \\
& C$_{2}$H v = 0 N=1--0, J=3/2-1/2, F= 2- 1 & 87.317 & 4.19 & 4$\times 10^{5}$ & Ethynyl \\
\multicolumn{6}{l}{Shocks:}\\
& SiO v = 0 J=2--1 & 86.847 & 6.25 & 2$\times 10^{6}$ & Silicon Monoxide \\
& SO v = 0 3(2)--2(1) & 99.300 & 9.23 & 3$\times 10^{5}$ & Sulfur Monoxide \\
\multicolumn{6}{l}{Isotolopologues:}\\
& H$^{13}$CN v = 0 J=1--0 & 86.340 & 4.14 & 9$\times 10^{5}$ & Hydrogen Cyanide, isotopologue \\
& H$^{13}$CO$^{+}$ J=1--0 & 86.754 & 4.16 & 1$\times 10^{5}$ & Formylium, isotopologue \\
& HN$^{13}$C J=1--0 & 87.091 & 4.18 & 1$\times 10^{5}$ & Hydrogen Isocyanide, isotopologue \\
& HCC$^{13}$CN v = 0 J=11--10 & 99.661 & 28.70 & $\sim$ 8$\times 10^{6}$ & Cyanoacetylene, isotopologue \\
\multicolumn{6}{l}{Dense, `high' E$_{u}$/$k$:}\\
& HNCO v=0 4(0,4)--3(0,3) & 87.925 & 10.55 & 4$\times 10^{6}$ & Isocyanic Acid \\
& CH$_{3}$SH v = 0 4(0)+-3(0)+A & 101.139 & 12.14 & $\sim$ 9$\times 10^{5}$ & Methyl Mercaptan \\
& NH$_{2}$CHO 4(1,3)--3(1,2) & 87.849 & 13.52 & $\sim$ 4$\times 10^{6}$ & Formamide \\
& CH$_{3}$CHO v=0 5(1,4)--4(1,3) E & 98.863 & 16.59 & $\sim$ 3$\times 10^{6}$ & Acetaldehyde \\
& H$_{2}$CS 3(1,3)--2(1,2) & 101.478 & 22.91 & 2$\times 10^{5}$ & Thioformaldehyde \\
& NH$_{2}$CN 5(1,4)--4(1,3), v=0 & 100.629 & 28.99 & $\sim$ 9$\times 10^{6}$ & Cyanamide \\
& HC$_{5}$N v = 0 J=33--32 & 87.864 & 71.69 & $\sim$ 9$\times 10^{6}$ & Cyanobutadiyne \\
\tableline
\end{tabular}}}
\footnotetext[1]{{\scriptsize{Excitation energies (E$_{u}$/$k$), Einstein A coefficients, and collisional rates were obtained from the Leiden Atomic and Molecular
Database (LAMDA; \citealp{Schoier05}) and Cologne Database for Molecular Spectroscopy (CDMS; \citealp{Muller01,Muller05}) assuming a gas temperature of 20\,K.
To derive critical densities, we used calculated collisional rates ($\gamma$) where possible and the equation n$_{crit}$\, = A$_{u}$ / $\gamma$, where A$_{u}$ is the Einstein A coefficient.
For many of the more complex species, however, collisional rates have not been calculated. In these cases (marked as approximate in the table), we calculate an approximate critical
density via n$_{crit}$ = A$_{u}$ / ($v \sigma$), where $v$ is the velocity of the molecule and $\sigma$ is the collisional cross section which was assumed to be 10$^{-15}$\,${\rm cm}^{-2}$.}}}
\footnotetext[2]{{\scriptsize{The molecules are grouped according to the molecular `families' and are ordered by increasing excitation energies.}}}
\end{sidewaystable}
\begin{sidewaystable}[h]
\caption{Summary of the positions across G0.253+0.016\, from which spectra are displayed in Figure~\ref{spectra}. \label{spectra-table}}
{\scriptsize{
\begin{tabular}{lllrrl}
\tableline
\tableline
\multicolumn{5}{c}{Position} & Comment \\
& R.A. & Dec. & \multicolumn{1}{c}{Offset R.A.} & \multicolumn{1}{c}{Offset Dec.} & \\
& (deg) & (deg) & \multicolumn{1}{c}{(deg)} & \multicolumn{1}{c}{(deg)} & \\
\tableline
1 & 17:46:10.628 & --28.:42:17.75 & 0.004 & 0.004 & Peak in the dust column density \\
2 & 17:46:10.308 & --28.:42:28.60 & 0.003 & 0.001 & Centre of the clump \\
3 & 17:46:09.297 & --28.:42:12.50 & --0.001 & 0.006 & Position showing multiple velocity components \\
4 & 17:46:11.000 & --28.:43:34.05 & 0.005 & --0.017 & Lower part of the clump, where most molecules are detected \\
5 & 17:46:09.883 & --28.:43:15.50 & 0.001 & --0.012 & Lower part of the clump, where the linewidths from the molecules are very different \\
6 & 17:46:09.590 & --28.:41:20.70 & 0.000 & 0.020 & Upper part of the clump, dust continuum peak \\
\tableline
\end{tabular}}}
\end{sidewaystable}
|
\section{Introduction}
\subsection{Summary}
\label{ss:intro}
Let $\FM$ be an algebraically closed field, and let ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ be a connected reductive group over $\FM$ which is a product of simply-connected quasi-simple groups and of general linear groups. We assume that the characteristic $p$ of $\FM$ is very good for each quasi-simple factor of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. Let also $\check{G}$ be the complex Langlands dual group. The main result of this paper is an equivalence of categories relating tilting objects in the heart of Bezrukavnikov's \emph{exotic t-structure} on the derived category of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$-equivariant coherent sheaves on the Springer resolution $\widetilde{\NC}$ of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$, and Iwahori-constructible parity sheaves on the affine Grassmannian ${\EuScript Gr}$ of $\check{G}$, with coefficients in $\FM$ (in the sense of~\cite{jmw}).
We provide several applications of this result; in particular
\begin{enumerate}
\item
\label{it:intro-perverse}
a proof that spherical parity sheaves on ${\EuScript Gr}$, with coefficients in a field of good characteristic, are perverse, which provides the last missing step in the proof of the Mirkovi{\'c}--Vilonen conjecture~\cite{mv} on stalks of standard spherical perverse sheaves on ${\EuScript Gr}$ in the expected generality;
\item
\label{it:intro-equiv}
a construction of an equivalence of categories relating $D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ to the ``modular mixed derived category'' of Iwahori-constructible sheaves on ${\EuScript Gr}$ (in the sense of~\cite{modrap2}), which is a modular generalization of an equivalence due to Arkhipov--Bezrukavnikov--Ginzburg~\cite{abg};
\item
an ``extension'' of~\eqref{it:intro-equiv} to an equivalence relating
$D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ (where $\widetilde{\gg}$ is the Grothendieck resolution of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$) to the modular mixed derived category of Iwahori-\emph{equivariant} sheaves on ${\EuScript Gr}$.
\end{enumerate}
A weaker version of~\eqref{it:intro-perverse} was obtained earlier by Juteau--Mautner--Williamson \cite{jmw2} using a case-by-case argument. The application to the Mirkovi{\'c}--Vilonen conjecture is due to Achar--Rider~\cite{arider}. In the case $p=0$, the equivalence in~\eqref{it:intro-equiv} plays an important role in the representation theory of Lusztig's quantum groups at a root of unity, see~\cite{abg, bezru-tilting, bl}. We expect our equivalence to play a comparable role in the modular representation theory of connected reductive groups.\footnote{One year after this paper was written, this expectation was indeed confirmed in the article~\cite{prinblock} by P.~Achar and the second author.} A similar result has been obtained independently by Achar--Rider~\cite{arider2}, under the assumption that spherical parity sheaves are perverse. Our methods are different from theirs; see~\S\ref{ss:comparison} below for a detailed comparison.
\subsection{Main result}
To state our results more precisely, let us choose a Borel subgroup ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \subset {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$, and a maximal torus ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \subset {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$. Let $\bg$ be the Lie algebra of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$. Then the Springer resolution $\widetilde{\NC}$ is defined as
\[
\widetilde{\NC}:=\{(\xi, g{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}) \in \gg^* \times {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}/{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \mid \xi_{|g \cdot \bg}=0\}.
\]
This variety is endowed with a natural action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$, defined by
\[
(g,x) \cdot (\xi, h{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}) := (x^{-2} g \cdot \xi, gh{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}),
\]
so that we can consider the derived category $D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$-equiva\-riant coherent sheaves on $\widetilde{\NC}$. This category possesses a remarkable t-structure, called the \emph{exotic t-structure}, defined by Bezrukavnikov~\cite{bezru-tilting} in the case $p=0$, and studied in our generality in~\cite{mr}. The heart ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ of this t-structure has a natural structure of a graded highest weight category, with weights the lattice ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} = X^*({\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}})$ of characters of ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$, and ``normalized'' standard, resp.~costandard, objects denoted $\Delta^\lambda_{\widetilde{\NC}}$, resp.~$\nabla^\lambda_{\widetilde{\NC}}$. In particular, we will be interested in the category $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))$ of \emph{tilting objects} in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ (i.e.~those objects which possess both a standard filtration and a costandard filtration). This category is Krull--Schmidt, and its indecomposable objects are parametrized in a natural way by ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \times \ZM$. For $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, we denote by $\TC^\lambda$ the indecomposable object attached to $(\lambda,0)$. Then for any $n \in \ZM$ the object associated with $(\lambda,n)$ is $\TC^\lambda \langle n \rangle$, where $\langle n \rangle$ is the $n$-th power of the functor $\langle 1 \rangle$ of tensoring with the tautological $1$-dimensional $\Gm$-module (see~\S\ref{ss:notation-KW-tilting} for details). For any $\TC$ in $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))$, $\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$, we denote by $(\TC : \Delta^\lambda_{\widetilde{\NC}} \langle m \rangle)$, resp.~$(\TC : \nabla^\lambda_{\widetilde{\NC}} \langle m \rangle)$, the multiplicity of the standard object $\Delta^\lambda_{\widetilde{\NC}} \langle m \rangle$, resp.~of the costandard object $\nabla^\lambda_{\widetilde{\NC}} \langle m \rangle$, in a standard (resp. costandard) filtration of $\TC$.
Let now $\check{G}$ be a complex connected group, with a maximal torus $\check{T} \subset \check{G}$, and assume that $(\check{G}, \check{T})$ is Langlands dual to $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}, {\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}})$, in the sense that the root datum of $(\check{G}, \check{T})$ is dual to that of $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}, {\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}})$. In particular, we have an identification ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}=X_*(\check{T})$. We let $\check{B} \subset \check{G}$ be the Borel subgroup containing $\check{T}$ whose roots are the coroots of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$ (which we will consider as the \emph{negative} coroots). Let $\mathscr{O}:=\CM[ \hspace{-1pt} [z] \hspace{-1pt} ]$ and $\mathscr{K}:=\CM( \hspace{-1pt} (z) \hspace{-1pt} )$. Then the affine Grassmannian ${\EuScript Gr}$ of $\check{G}$ is defined as
\[
{\EuScript Gr} := \check{G}(\mathscr{K}) / \check{G} (\mathscr{O}),
\]
with its natural ind-variety structure. We denote by $\check{I}$ the Iwahori subgroup of $\check{G}(\mathscr{O})$ determined by $\check{B}$, i.e.~the inverse image of $\check{B}$ under the morphism $\check{G}(\mathscr{O}) \rightarrow \check{G}$ defined by the evaluation at $z=0$. Then $\check{I}$ acts naturally on ${\EuScript Gr}$ via left multiplication on $\check{G}(\mathscr{K})$, and the orbits of this action are parametrized in a natural way by ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$; we denote by ${\EuScript Gr}_\lambda$ the orbit associated with $\lambda$ and by $i_\lambda : {\EuScript Gr}_\lambda \hookrightarrow {\EuScript Gr}$ the inclusion. We let
\[
\Parity_{(\check{I})}({\EuScript Gr},\FM)
\]
be the category of parity sheaves on ${\EuScript Gr}$, with coefficients in $\FM$, with respect to the stratification by $\check{I}$-orbits (in the sense of~\cite{jmw}). This category is defined as an additive subcategory of the derived category $D^{\mathrm{b}}_{(\check{I})}({\EuScript Gr}, \FM)$ of $\check{I}$-constructible $\FM$-sheaves on ${\EuScript Gr}$. It is Krull--Schmidt, and its indecomposable objects are parametrized in a natural way by ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \times \ZM$. We denote by $\EC_\lambda$ the indecomposable object attached to $(\lambda,0)$; then for any $n \in \ZM$ the object associated with $(\lambda,n)$ is $\EC_\lambda[n]$.
The main result of this paper (whose proof is given in~\S\ref{ss:proof-main}) is the following.
\begin{thm}
\label{thm:main}
There exists an equivalence of additive categories
\begin{equation}
\label{eqn:equiv-main}
\Theta \colon
\Parity_{(\check{I})}({\EuScript Gr},\FM) \xrightarrow{\sim} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))
\end{equation}
which satisfies the following properties:
\begin{enumerate}
\item
\label{it:main-thm-shifts}
$\Theta \circ [1] \cong \langle -1 \rangle \circ \Theta$;
\item
\label{it:main-thm-ch}
for all $m \in \ZM$, $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $\EC \in \Parity_{(\check{I})}({\EuScript Gr},\FM)$, we have
\begin{align*}
(\Theta(\EC) : \Delta^\lambda_{\widetilde{\NC}} \langle m \rangle) &= \dim_\FM \bigl( \HM^{m-\dim({\EuScript Gr}_{-\lambda})}({\EuScript Gr}_{-\lambda}, i_{-\lambda}^* \EC) \bigr); \\
(\Theta(\EC) : \nabla^\lambda_{\widetilde{\NC}} \langle m \rangle) &= \dim_\FM \bigl( \HM^{m-\dim({\EuScript Gr}_{-\lambda})}({\EuScript Gr}_{-\lambda}, i_{-\lambda}^! \EC) \bigr);
\end{align*}
\item
\label{it:main-thm-indec}
$\Theta(\EC_\lambda) \cong \TC^{-\lambda}$
for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
\end{enumerate}
\end{thm}
\subsection{Outline of the proof}
\label{ss:outline}
Our strategy of proof of Theorem~\ref{thm:main} is based on the description of both sides in~\eqref{eqn:equiv-main} in terms of an appropriate category of ``Soergel bimodules.''\footnote{Our ``Soergel bimodules'' are in fact not bimodules over any ring, but rather modules over a certain algebra built out of two copies of a polynomial algebra. We use this terminology since these objects play the role which is usually played by actual Soergel bimodules.} This idea is very classical, see e.g.~\cite{soergel-kategorie, by, dodd} for examples in characteristic zero, and~\cite{soergel, modrap1} for examples in positive characteristic.
On the ``constructible side'' (i.e.~the left-hand side in~\eqref{eqn:equiv-main}), this description is obtained using the total cohomology functor. The arguments in this section follow well-established techniques; see~\S\ref{ss:key-lemma} below for a discussion of the only new idea that is needed.
On the ``coherent side'' (i.e.~the right-hand side in~\eqref{eqn:equiv-main}), we adapt a construction due to Dodd~\cite{dodd} in characteristic zero, which uses a ``Kostant--Whittaker reduction'' functor. This construction uses modular (and integral) versions of some classical results of Kostant, which are treated in the companion paper~\cite{riche3}. Our constructions are slightly different from Dodd's, however, in that we do not use a deformation to (asymptotic) $\mathcal{D}$-modules, but only to coherent sheaves on the Grothendieck resolution
\[
\widetilde{\gg}:=\{(\xi, g{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}) \in \gg^* \times {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}/{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \mid \xi_{|g \cdot \ng}=0\},
\]
where $\ng$ is the Lie algebra of the unipotent radical of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$. (On the constructible side, this deformation amounts to replacing the category $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ by the category $\Parity_{\check{I}}({\EuScript Gr}, \FM)$ of $\check{I}$-equivariant $\FM$-parity sheaves on ${\EuScript Gr}$.)
Our proof of fully-faithfulness of the Kostant--Whittaker reduction functor is also different from the proof in~\cite{dodd}. One of the crucial ideas in our proof, which we learnt in papers of Soergel~\cite{soergel-kategorie, soergel} and was used also in~\cite{modrap1}, is to compute $\Hom$-spaces between Soergel bimodules from the analogue for parity sheaves. We use this computation to prove the ``coherent side''; see the proof of Theorem~\ref{thm:main-geom} for more details.
We also use some ideas from categorification, related to the fact that both categories appearing in~\eqref{eqn:equiv-main} provide categorifications of the spherical module $\MM_{\mathrm{sph}}$ of the affine Hecke algebra $\HM_{\mathrm{aff}}$ attached to ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ (in the sense that their split Grothen\-dieck groups are equipped with natural actions of $\HM_{\mathrm{aff}}$, and are naturally isomorphic to $\MM_{\mathrm{sph}}$). On the left-hand side this uses the realization of $\HM_{\mathrm{aff}}$ in terms of constructible sheaves on the affine flag variety of $\check{G}$, and on the right-hand side this uses the Kazhdan--Lusztig--Ginzburg description of $\HM_{\mathrm{aff}}$ in terms of equivariant coherent sheaves on the Steinberg variety of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$, a ``categorical'' counterpart of which is provided by the ``geometric braid group action'' studied in~\cite{riche, br}.
Note that, surprisingly, our proof does \emph{not} use the geometric Satake equivalence from~\cite{mv}.
\subsection{Equivalences of triangulated categories}
One of our motivations for the study of the equivalence in Theorem~\ref{thm:main} was the desire to obtain a ``modular version'' of an equivalence of categories due to Arkhipov--Bezrukavnikov--Ginzburg~\cite{abg} in the case $p=0$. Our version involves the ``modular mixed derived category''
\[
D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM) := K^{\mathrm{b}} \Parity_{(\check{I})}({\EuScript Gr}, \FM),
\]
introduced and studied (in a more general setting) in~\cite{modrap2}. In particular, it can be endowed with a ``perverse t-structure'' and a ``Tate twist'' autoequivalence $\langle 1 \rangle$, and possesses ``standard objects'' $\Delta^\mix_{\lambda}$ ($\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$) and ``costandard objects'' $\nabla^\mix_\lambda$ ($\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$) which have the same properties as ordinary standard and costandard perverse sheaves. For $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, we denote by $\EC^\mathrm{mix}_\lambda$ the object $\EC_\lambda$, considered as an object of $D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM)$. In case $\FM=\overline{\QM}_\ell$, the category $D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM)$ is equivalent to the bounded derived category of the category $\tilde{\PC}$ of~\cite[\S 4.4]{bgs}, see~\cite[Remark~2.2]{modrap2}.
As an immediate application of Theorem~\ref{thm:main} we obtain the following theorem, whose proof is given in~\S\ref{ss:proof-equivalence-tNC}, and which provides a ``modular analogue'' of the equivalence of~\cite[Theorem~9.4.1]{abg}.
\begin{thm}
\label{thm:equivalence-tNC}
There exists an equivalence of triangulated categories
\[
\Phi \colon D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM) \xrightarrow{\sim} D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})
\]
which satisfies
\[
\Phi \circ \langle 1 \rangle \cong \langle 1 \rangle [1] \circ \Phi, \quad \Phi(\Delta^\mix_{\lambda}) \cong \Delta_{\widetilde{\NC}}^{-\lambda}, \quad \Phi(\nabla^\mix_{\lambda}) \cong \nabla_{\widetilde{\NC}}^{-\lambda}, \quad \Phi(\EC^\mathrm{mix}_\lambda) \cong \TC^{-\lambda}
\]
for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
\end{thm}
\begin{remark}
One should think of $\Phi$ as some kind of ``Ringel--Koszul duality,'' i.e.~a composition of a Koszul duality (as in~\cite[Theorem~2.12.6]{bgs}, which sends simple objects to projective objects) and a Ringel duality (which sends projective objects to tilting objects). The relevance of such equivalences in Lie-theoretic contexts was pointed out in~\cite{bg}; see also~\cite{by}. The idea that, in a modular context, simple perverse sheaves should be replaced by parity sheaves in ``Koszul-type'' statements is implicit in~\cite{soergel}, and was developed more explicitly in~\cite{rsw}.
\end{remark}
As explained in~\S\ref{ss:outline}, our proof of Theorem~\ref{thm:main} is based on the consideration of a ``deformation'' of the picture, replacing $\check{I}$-constructible sheaves by $\check{I}$-equivariant sheaves, and equivariant coherent sheaves on $\widetilde{\NC}$ by equivariant coherent sheaves on $\widetilde{\gg}$. Using these considerations, we define objects $\Delta^\lambda_{\widetilde{\gg}}$ and $\nabla^\lambda_{\widetilde{\gg}}$ in $D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ which satisfy
\[
\mathsf{L} i^*(\Delta^\lambda_{\widetilde{\gg}}) \cong \Delta^\lambda_{\widetilde{\NC}}, \qquad \mathsf{L} i^*(\nabla^\lambda_{\widetilde{\gg}}) \cong \nabla^\lambda_{\widetilde{\NC}}
\]
for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, where $i \colon \widetilde{\NC} \hookrightarrow \widetilde{\gg}$ is the inclusion (see~\S\ref{ss:standard-KW}). We also define an additive and Karoubian subcategory ${\EuScript Tilt}$ of $D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$, stable under the shift $\langle n \rangle$ for any $n \in \ZM$, and indecomposable objects $\widetilde{\TC}^\lambda$ in ${\EuScript Tilt}$ satisfying $\mathsf{L} i^*(\widetilde{\TC}^\lambda) \cong \TC^\lambda$ and such that ${\EuScript Tilt}$ is Krull--Schmidt with indecomposable objects $\widetilde{\TC}^\lambda \langle n \rangle$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $n \in \ZM$, see~\S\ref{ss:proof-equivalence-tgg}. On the other hand, we consider the equivariant modular mixed derived category
\[
D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr}, \FM) := K^{\mathrm{b}} \Parity_{\check{I}}({\EuScript Gr}, \FM),
\]
and denote by $\Delta^\mix_{\check{I}, \lambda}$, resp.~$\nabla^\mix_{\check{I}, \lambda}$, the standard, resp.~costandard, object associated with $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, the parity complex $\EC_\lambda$ can be naturally ``lifted'' to the category $\Parity_{\check{I}}({\EuScript Gr}, \FM)$; we again denote by $\EC_\lambda^\mathrm{mix}$ this object viewed as an object in $D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr}, \FM)$.
The following ``deformation'' of Theorem~\ref{thm:equivalence-tNC} is proved in~\S\ref{ss:proof-equivalence-tgg}. (In case $p=0$, a similar result can be deduced from the main result of~\cite{dodd}, though this equivalence is not explicitly stated there.)
\begin{thm}
\label{thm:equivalence-tgg}
There exists an equivalence of triangulated categories
\[
\Psi \colon D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr},\FM) \xrightarrow{\sim} D^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})
\]
which satisfies
\[
\Psi \circ \langle 1 \rangle \cong \langle 1 \rangle [1] \circ \Psi, \quad \Psi(\Delta^\mix_{\check{I}, \lambda}) \cong \Delta_{\widetilde{\gg}}^{-\lambda}, \quad \Psi(\nabla^\mix_{\check{I}, \lambda}) \cong \nabla_{\widetilde{\gg}}^{-\lambda}, \quad \Psi(\EC^\mathrm{mix}_{\lambda}) \cong \widetilde{\TC}^{-\lambda}
\]
for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
\end{thm}
In~\S\ref{ss:compatibilities} we also prove that the equivalences $\Psi$ and $\Phi$ are compatible in the natural way.
\subsection{Parity, tilting and the Mirkovi{\'c}--Vilonen conjecture}
\label{ss:mv-conjecture}
Our other main motivation for the study of Theorem~\ref{thm:main} comes from the Mirkovi{\'c}--Vilonen conjecture~\cite[Conjecture~13.3]{mv}. In this subsection, we let ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ be more generally any split connected reductive group scheme over an arbitrary commutative ring $k$. As before, we fix a (split) maximal torus ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$ and let ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ denote the weight lattice. We again let $\check{G}$ and $\check{T} \subset \check{G}$ denote the unique (up to isomorphism) complex connected reductive group and maximal torus with Langlands dual root datum, and consider its affine Grassmannian ${\EuScript Gr}:=\check{G}(\mathscr{K})/\check{G}(\mathscr{O})$.
The \emph{geometric Satake equivalence}, proven by Mirkovi{\'c}--Vilonen~\cite{mv} for any $k$ which is Noetherian and of finite global dimension, is an equivalence of abelian categories
\[
\mathsf{S}_k \colon \Perv_{(\check{G}(\mathscr{O}))}({\EuScript Gr}, k) \xrightarrow{\sim} \Rep({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}),
\]
where $\Perv_{(\check{G}(\mathscr{O}))}({\EuScript Gr}, k)$ is the category of $\check{G}(\mathscr{O})$-constructible $k$-perverse sheaves on ${\EuScript Gr}$, and $\Rep({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}})$ is the category of algebraic ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-modules which are of finite type over $k$. (This equivalence is compatible with the natural monoidal structures on these categories.)
After a choice of positive roots, the $\check{G}(\mathscr{O})$-orbits on ${\EuScript Gr}$ are parametrized in a natural way by the set of dominant weights ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+ \subset {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$; we denote by ${\EuScript Gr}^\lambda$ the orbit associated with $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, and by $i^\lambda \colon {\EuScript Gr}^\lambda \hookrightarrow {\EuScript Gr}$ the inclusion. For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, we consider the perverse sheaf
\[
\IC_!(\lambda, k) := {}^p (i^\lambda)_! \underline{k}_{{\EuScript Gr}^\lambda} [\dim({\EuScript Gr}^\lambda)].
\]
By~\cite[Proposition~13.1]{mv}, this object corresponds, under $\mathsf{S}_k$, to the Weyl ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-module associated with $\lambda$.
Towards the end of their paper, Mirkovi{\'c}--Vilonen state the following conjecture.
\begin{conj}[Mirkovi{\'c}--Vilonen~\cite{mv}]
\label{conj:mv}
The cohomology modules of the stalks of $\IC_!(\lambda, \ZM)$ are free.
\end{conj}
We understand this conjecture was formulated based on evidence in type $A$ and as part of an attempt to produce a ``modular analogue'' of results of Bezukavnikov and collaborators (in particular, the equivalence of~\cite[Theorem~9.4.1]{abg}), like the one we prove as Theorem~\ref{thm:equivalence-tNC}.
Using~\cite[Proposition~8.1(a)--(b)]{mv}, one can also state this conjecture equivalently as the property that, for any field $\Bbbk$, the dimensions
\[
\dim_\Bbbk \bigl( \HM^m({\EuScript Gr}_\mu, i_\mu^* \IC_!(\lambda, \Bbbk)) \bigr)
\]
are independent of $\Bbbk$.
Note that, by~\cite[Lemma~7.1]{mv}, if $\Bbbk$ has characteristic zero, we have $\IC_!(\lambda, \Bbbk) = \IC \CC({\EuScript Gr}^\lambda, \underline{\Bbbk})$, and the graded dimensions of stalks of these objects can be computed in terms of affine Kazhdan--Lusztig polynomials~\cite{kl} and vanish in either all odd or all even degrees.
It was realized by Juteau~\cite{juteau} that Conjecture~\ref{conj:mv} does not hold as stated, namely that the cohomology modules of the stalks of $\IC_!(\lambda, \ZM)$ can have $p$-torsion if $p$ is a prime number which is bad for $\check{G}$.
On the other hand, if $\IC_!(\lambda, \ZM)$ does not have $p$-torsion for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, then it follows that for any field $\Bbbk$ of characteristic $p$, the tilting objects in $\Perv_{(\check{G}(\mathscr{O}))}({\EuScript Gr}, \Bbbk)$ are parity complexes, or equivalently that all the parity sheaves $\EC_\lambda$ for $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ (i.e.~the parity sheaves which are $\check{G}(\mathscr{O})$-constructible) are perverse. (The equivalence between these properties follows from~\cite[Proposition~3.3]{jmw2}.)
The statement that the parity sheaves $\EC_\lambda$ for $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ with coefficients in a field $\Bbbk$ of characteristic $p$ are perverse, was proved using a case-by-case analysis by Juteau--Mautner--Williamson~\cite[Theorem 1.8]{jmw2} provided that $p$ is bigger than explicit bounds. The bounds in~\cite{jmw2} are a byproduct of the method of proof and are in general stronger than the condition that $p$ is good (most notably when ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ is quasi-simple of type $C_n$, in which case the bound is $p>n$).
Achar--Rider~\cite{arider} then proved that if all the parity sheaves $\EC_\lambda$ for $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ with coefficients in a field $\Bbbk$ of characteristic $p$ are perverse, then the cohomology modules of the stalks of the perverse sheaves $\IC_!(\lambda, \ZM)$ have no $p$-torsion.
In~\S\ref{ss:proof-parity-perverse} we give a uniform proof of the following.
\begin{cor}
\label{cor:parity-perverse}
If $p$ is good for $\check{G}$, then for any $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ the parity sheaf $\EC_\lambda$ is perverse.
\end{cor}
By the main result of~\cite{arider}, this implies:
\begin{thm}[Mirkovi{\'c}--Vilonen conjecture]
\label{thm:mv}
If $p$ is good for $\check{G}$, then
the cohomology modules of the stalks of the perverse sheaves $\IC_!(\lambda, \ZM)$ have no $p$-torsion.
\end{thm}
Our proof of Corollary~\ref{cor:parity-perverse} is based on the explicit description of the tilting objects $\TC^\lambda$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ obtained in~\cite{mr}.
\subsection{Other applications}
\label{ss:intro-other}
Now, let us come back to our assumptions on ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ and $p$ from~\S\ref{ss:intro}.
In this subsection we state some other applications of Theorem~\ref{thm:main}. Most of these results are already known (at least when $p=0$), but their proof usually requires the geometric Satake equivalence of~\S\ref{ss:mv-conjecture}. We find it useful to explain how these results can be derived directly using our methods.
We
denote by $\Parity_{(\check{G}(\mathscr{O}))}({\EuScript Gr},\FM)$ the subcategory of $\Parity_{(\check{I})}({\EuScript Gr},\FM)$ consisting of objects which are $\check{G}(\mathscr{O})$-constructible, i.e.~which are direct sums of objects $\EC_\lambda[i]$ where $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ and $i \in \ZM$.
We also denote by $\PParity_{(\check{G}(\mathscr{O}))}({\EuScript Gr},\FM)$ the subcategory of $\Parity_{(\check{G}(\mathscr{O}))}({\EuScript Gr},\FM)$ consisting of objects which are perverse sheaves. By Corollary~\ref{cor:parity-perverse}, this category consists of objects which are direct sums of parity sheaves $\EC_\lambda$ where $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$.
To conform to the notation used most of the time in this context, for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ we set $\EC^\lambda:=\EC_{w_0 \lambda}$ (where $w_0$ is the longest element in the Weyl group $W$ of $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}},{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}})$), and (as in~\S\ref{ss:mv-conjecture}) ${\EuScript Gr}^\lambda = \sqcup_{\mu \in W\lambda} {\EuScript Gr}_{\mu}$. We also denote by $L_\lambda$ the unique $\check{T}$-fixed point in ${\EuScript Gr}_\lambda$, and by $\imath_\lambda \colon \{L_\lambda\} \hookrightarrow {\EuScript Gr}$ the inclusion.
We denote by $\Tilt({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}})$ the additive category of tilting ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-modules, and by $\mathsf{T}(\lambda)$ the indecomposable tilting ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-module with highest weight $\lambda$ (for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$). We also denote by $\mathsf{M}(\lambda)$ and $\mathsf{N}(\lambda)$ the standard and costandard ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-modules with highest weight $\lambda$ respectively (see~\cite[\S 4.3]{mr}). We denote by $(\mathsf{T}(\lambda) : \mathsf{M}(\mu))$ and $(\mathsf{T}(\lambda) : \mathsf{N}(\mu))$ the corresponding multiplicities.
From Theorem~\ref{thm:main} one can deduce the following result (see~\S\ref{ss:proof-Satake}).
\begin{prop}
\label{prop:equiv-parity-tilt}
There exists an equivalence of additive categories
\[
\overline{\mathsf{S}}_\FM \colon \PParity_{(\check{G}(\mathscr{O}))}({\EuScript Gr},\FM) \xrightarrow{\sim} \Tilt({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}})
\]
which satisfies $\overline{\mathsf{S}}_\FM(\EC^{\lambda}) \cong \mathsf{T}(\lambda)$ for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$. Moreover, for $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ we have
\begin{equation}
\label{eqn:formula-stalks-parity}
\sum_{k \in \ZM} \dim \bigl( \HM^{k-\dim({\EuScript Gr}^{\mu})}(\imath_\mu^* \EC^\lambda) \bigr) \cdot \mathsf{v}^k = \sum_{\nu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+} \bigl( \mathsf{T}(\lambda) : \mathsf{N}(-w_0\nu) \bigr) \cdot \MC^{-w_0\mu}_\nu(\mathsf{v}^{-2}),
\end{equation}
where $\MC_\eta^\chi(\mathsf{v})$ is Lusztig's $q$-analogue~\cite{lusztig}.
\end{prop}
In case $p=0$, $\mathsf{M}(\lambda)$ and $\mathsf{T}(\lambda)$ both coincide with the simple ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$-module with highest weight $\lambda$, so that $\bigl( \mathsf{T}(\lambda) : \mathsf{M}(\nu) \bigr)=\delta_{\lambda, \nu}$. On the other hand we have $\EC^\lambda=\IC \CC({\EuScript Gr}^\lambda, \underline{\FM})$. By~\cite{kl}, the dimensions of the stalks of this perverse sheaf can be expressed in terms of Kazhdan--Lusztig polynomials for the affine Weyl group $W_{\mathrm{aff}}$ of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. Hence in this case, \eqref{eqn:formula-stalks-parity} provides a geometric proof of the relation between affine Kazhdan--Lusztig polynomials and Lusztig's $q$-analogue, conjectured in~\cite{lusztig} and proved by different methods in~\cite{ka}.
Once it is known that the parity sheaves $\EC^\lambda$ are perverse, as remarked in~\cite[Proposition~3.3]{jmw2}, it follows that $\mathsf{S}_\FM(\EC^\lambda) \cong \mathsf{T}(\lambda)$ (where $\mathsf{S}_\FM$ is the geometric Satake equivalence, see~\S\ref{ss:mv-conjecture}). Hence one can obtain a different construction of an equivalence $\overline{\mathsf{S}}_\FM$ by simply restricting $\mathsf{S}_\FM$.
In this setting, a sketch of a different proof of~\eqref{eqn:formula-stalks-parity} is given in~\cite[Remark~4.2]{jmw2}.
One can also apply our results to describe the (equivariant) cohomology of spherical parity sheaves and their costalks (see \S \ref{ss:proof-cohomology} for the proof).
\begin{prop}
\label{prop:cohomology}
\begin{enumerate}
\item
\label{it:cohom}
For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, there exist isomorphisms of graded vector spaces, resp.~of graded $\HM^\bullet_{\check{I}}(\mathrm{pt}; \FM)$-modules,
\[
\HM^\bullet({\EuScript Gr}, \EC^\lambda) \cong \mathsf{T}(\lambda), \qquad \HM_{\check{I}}^\bullet({\EuScript Gr}, \EC^\lambda) \cong \mathsf{T}(\lambda) \otimes \HM^\bullet_{\check{I}}(\mathrm{pt}; \FM),
\]
where the grading on $\mathsf{T}(\lambda)$ is obtained from the $\Gm$-action through the cocharacter given by the sum of the positive coroots.
\item
\label{it:cohom-stalks}
For any $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, there exist isomorphisms of graded vector spaces, resp.~of graded $\HM^\bullet_{\check{I}}(\mathrm{pt}; \FM)$-modules,
\begin{align*}
\HM^{\bullet-\dim({\EuScript Gr}^{\mu})}(\imath_\mu^! \EC^\lambda) &\cong \bigl( \mathsf{T}(\lambda) \otimes \Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}(-w_0 \mu)) \bigr)^{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}, \\
\HM_{\check{I}}^{\bullet-\dim({\EuScript Gr}^{\mu})}(\imath_\mu^! \EC^\lambda) &\cong \bigl( \mathsf{T}(\lambda) \otimes \Gamma(\widetilde{\gg}, \OC_{\widetilde{\gg}}(-w_0 \mu)) \bigr)^{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}},
\end{align*}
where in both cases $\mathsf{T}(\lambda)$ is in degree $0$, the global sections are equipped with the grading induced by the $\Gm$-actions on $\widetilde{\NC}$ and $\widetilde{\gg}$, and the $\HM^\bullet_{\check{I}}(\mathrm{pt}; \FM)$-action on $( \mathsf{T}(\lambda) \otimes \Gamma(\widetilde{\gg}, \OC_{\widetilde{\gg}}(-w_0 \mu)))^{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ is induced by the natural morphism $\widetilde{\gg} \rightarrow \mathrm{Lie}({\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}) = \Spec(\HM^\bullet_{\check{I}}(\mathrm{pt}; \FM))$.
\end{enumerate}
\end{prop}
The first isomorphism in~\eqref{it:cohom} can be alternatively deduced from the fact that $\mathsf{S}_\FM(\EC^\lambda) \cong \mathsf{T}(\lambda)$, see the comments after Proposition~\ref{prop:equiv-parity-tilt}. Then one can deduce the second isomorphism using~\cite[Lemma~2.2]{yz}. In case $p=0$, the isomorphisms in~\eqref{it:cohom-stalks} are proved in~\cite[Corollary~2.4.5 and Proposition~8.7.1]{gr}.\footnote{The conventions for the choice of positive roots and the normalization of line bundles are different in~\cite{gr}, which explains the formal difference between the two formulas. A more important difference is that in~\cite{gr} we construct an \emph{explicit and canonical} isomorphism, while here we only claim the \emph{existence} of such an isomorphism.}
\subsection{Comparison with~\cite{arider2}}
\label{ss:comparison}
As mentioned already, in~\cite{arider2} Achar and Rider have obtained a different proof of Theorem~\ref{thm:equivalence-tNC}. Their methods are quite different from ours, and closer to the methods used in~\cite{abg}. In fact, while for us most of the work is required on the ``coherent side,'' in their approach the most difficult constructions appear on the ``topological side.'' Moreover, the exotic t-structure does not play any role in the construction of their equivalence.\footnote{The relation between their equivalence and the exotic t-structure is studied in~\cite[Section~8]{arider2}, but only after the equivalence is constructed.} Another important difference is that their arguments rely on the geometric Satake equivalence, while ours do not.
The assumptions in~\cite{arider2} are also different from ours: in fact they assume that the field $\FM$ is such that
any spherical parity sheaf with coefficients in $\FM$ on ${\EuScript Gr}$ is perverse. Hence Corollary~\ref{cor:parity-perverse} allows to extend the validity of their results to all good characteristics.
\subsection{A key lemma}
\label{ss:key-lemma}
A very important role in our arguments is played by the following easy lemma (see~\cite[Lemma~3.3.3]{by}).
\begin{lem}
\label{lem:key}
Let $k$ be an integral domain, and let $\KM$ be its field of fractions. Let $A$ and $B$ be $k$-algebras, and let $\varphi \colon A \rightarrow B$ be an algebra morphism. If the morphism $\KM \otimes_k \varphi \colon \KM \otimes_k A \rightarrow \KM \otimes_k B$ is an isomorphism, then the ``restriction of scalars'' functor
\[
\mathrm{Mod}(B) \rightarrow \mathrm{Mod}(A)
\]
is fully-faithful on modules which are $k$-free.
\end{lem}
We learnt how useful this observation can be in~\cite{by}. It also plays an important role in~\cite{modrap1}. In practice, this lemma can be used
when we are interested in modules over a $k$-algebra $B$ which is ``complicated'' or not well understood, but for which we have a ``simplified model'' $A$ which is ``isomorphic to $B$ up to torsion'', i.e.~such that we have a $k$-algebra morphism $A \rightarrow B$ inducing an isomorphism $\KM \otimes_k A \xrightarrow{\sim} \KM \otimes_k B$.
For instance, on the ``constructible side'' of our proof of Theorem~\ref{thm:main}, working over $\FM$ directly we would have to consider the equivariant cohomology algebra $\HM^\bullet_{\check{I}}({\EuScript Gr};\FM)$. Using the results in~\cite{yz} (which rely on the geometric Satake equivalence) one can obtain a description of this algebra in terms of the distribution algebra of the universal centralizer associated with ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. This algebra is a rather complicated object; in particular it is not finitely generated over $\FM$. On the other hand, if $\KM$ is a field of characteristic zero, the algebra $\HM^\bullet_{\check{I}}({\EuScript Gr};\KM)$ has a nice description (in terms similar to the description of the cohomology of a finite flag variety as a coinvariant algebra) which follows from~\cite[Theorem~1]{bf}.\footnote{This simpler description is related to the preceding one using the fact that, over a field of characteristic zero, the distribution algebra of a smooth group scheme is isomorphic to the enveloping algebra of its Lie algebra.} Hence, instead of working over $\FM$, our main constructions are done with coefficients in a finite localization ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ of $\ZM$. In this case, we do not have a very explicit description of the algebra $\HM^\bullet_{\check{I}}({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. But the same construction as the one used by Bezrukavnikov--Finkelberg in the case of $\KM$ provides a ``simplified model'' for this algebra, which is isomorphic to $\HM^\bullet_{\check{I}}({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ ``up to ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-torsion,'' see~\S\ref{ss:cohomology-Gr}. Since the modules over this algebra that we want to consider are all ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, using Lemma~\ref{lem:key} we can replace the ``complicated algebra'' $\HM^\bullet_{\check{I}}({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ by its ``simplified model'' without loosing any information.
Given this strategy, we also have to work over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ on the ``coherent side.'' Most of the complications appearing in this setting are treated in~\cite{br} and in~\cite{riche3}.
\subsection{Contents}
In Section~\ref{sec:preliminaries} we prove some preliminary technical results, and introduce some objects which will play an important role in the later sections. In Section~\ref{sec:constructible-side} we describe the left-hand side in~\eqref{eqn:equiv-main} in terms of our Soergel bimodules. In Section~\ref{sec:KW-reduction} we define the Kostant--Whittaker reduction functor, building on the main results of~\cite{riche3}. In Section~\ref{sec:tilting-KW} we use this functor to describe the right-hand side in~\eqref{eqn:equiv-main} in terms of Soergel bimodules. Finally, in Section~\ref{sec:proofs} we prove the results stated in the introduction.
\subsection{Some notation and conventions}
All rings in this paper are tacitly assumed to be commutative and unital.
If $A$ is an algebra, we denote by $\mathrm{Mod}(A)$ the category of left $A$-modules.
If $A$ is a $\ZM$-graded algebra, we denote by $\Modgr(A)$ the category of $\ZM$-graded left $A$-modules. We denote by $\langle 1 \rangle$ the shift of the grading defined by $(M \langle 1 \rangle)_n=M_{n-1}$. If $M$ is a free graded $A$-module of finite rank, we denote by $\grk_A(M) \in \ZM[\mathsf{v},\mathsf{v}^{-1}]$ its graded rank, with the convention that $\grk_A(A\langle n \rangle)=\mathsf{v}^{n}$.
If $X$ is a scheme, we denote by $\OC_X$ its structure sheaf, and by $\OC(X)$ the global sections of $\OC_X$.
If $Y$ is a scheme and $X$ is a $Y$-scheme, we denote by $\Omega_{X/Y}$, or simply $\Omega_X$, the sheaf of relative differentials, and by $\Omega(X/Y)$, or simply $\Omega(X)$, its global sections.
If $k$ is a ring and $V$ is a free $k$-module of finite rank, by abuse we still denote by $V$ the affine $k$-scheme $\Spec \bigl( \mathrm{S}_k(\Hom_k(V,k)) \bigr)$, where $\mathrm{S}$ denotes the symmetric algebra.
If ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is a finite localization of $\ZM$, we define a \emph{geometric point} of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ to be an algebraically closed field whose characteristic $p \geq 0$ is not invertible in ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. If $\FM$ is a such a geometric point, then there exists a unique algebra morphism ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \FM$, so that tensor products on the form $\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (-)$ make sense.
If $X$ is a Noetherian scheme and ${\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}}$ is an affine $X$-group scheme, we denote by $\Rep({\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}})$ the category of representations of ${\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}}$ which are coherent as $\OC_X$-modules. If ${\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}}$ is flat over $X$, then this category is abelian.
At various points in the paper we consider certain schemes and affine group schemes that could be defined over various base rings. When it is not clear from context, we will use a subscript to specify the base ring. For example, we write $X_\ZM$, resp.~$X_\FM$, to denote the $\ZM$-scheme $X$, resp.~its base change to $\FM$. In order to avoid notational clutter, we will affix a single subscript $k$ to some constructions like fiber products, for example writing $(X \times_Z Y)_k$ and $\OC(X \times_Z Y)_k$ rather than $X_k \times_{Z_k} Y_k$ and $\OC(X_k \times_{Z_k} Y_k)$, or categories of equivariant coherent sheaves, writing $\Coh^G(X)_k$ instead of $\Coh^{G_k}(X_k)$. We will also use the abbreviation
$D^G(X)_k := D^{\mathrm{b}} \Coh^{G_k} (X_k)$.
\subsection{Acknowledgements}
Soon after we started working on this paper, we were informed by
Pramod Achar and Laura Rider that they were also working on a similar project, which was completed in~\cite{arider2}. We thank them for keeping us informed of their progress, and for useful discussions. We also thank the referee for helpful comments.
\section{Preliminary results}
\label{sec:preliminaries}
\subsection{Reminder on parity sheaves}
\label{ss:reminder-parity}
Let $X$ be a complex algebraic variety, and
\[
X=\bigsqcup_{s \in \mathscr{S}} X_s
\]
be a finite (algebraic) stratification of $X$ into affine spaces. For any $s \in \mathscr{S}$, we denote by $i_s \colon X_s \hookrightarrow X$ the embedding. If $k$ is a Noetherian ring of finite global dimension, we denote by $D^{\mathrm{b}}_\mathscr{S}(X,k)$ the $\mathscr{S}$-constructible derived category of sheaves of $k$-modules on $X$. Note that any $k$-local system on any stratum $X_s$ is constant. If $k'$ is a $k$-algebra which is also Noetherian and of finite global dimension, we will denote by
\[
k'(-) \colon D^{\mathrm{b}}_\mathscr{S}(X,k) \rightarrow D^{\mathrm{b}}_\mathscr{S}(X,k')
\]
the derived functor of extension of scalars.
The following is a slight generalization of~\cite[Definition~2.4]{jmw} (where it is assumed that $k$ is a complete local principal ideal domain).
\begin{defi}
\label{def:parity}
An object $\FC$ in $D^{\mathrm{b}}_\mathscr{S}(X,k)$ is called $*$-\emph{even}, resp.~$!$-\emph{even}, if $\HC^n(i_s^*\FC)$, resp.~$\HC^n(i_s^! \FC)$, vanishes if $n$ is odd, and is a projective $k$-local system if $n$ is even. It is called \emph{even} if it is both $*$-even and $!$-even.
An object $\FC$ is called $*$-\emph{odd}, resp.~$!$-\emph{odd}, resp.~\emph{odd} if $\FC[1]$ is $*$-even, resp.~$!$-even, resp.~even.
Finally, an object is called a \emph{parity complex} if it is isomorphic to the direct sum of an even and an odd object.
\end{defi}
We will denote by $\Parity_\mathscr{S}(X,k)$ the full subcategory of $D^{\mathrm{b}}_\mathscr{S}(X,k)$ consisting of parity complexes; it is stable under direct sums and direct summands. It is clear also that if $k'$ is a $k$-algebra satisfying the assumptions above, then the functor $k'(-)$ restricts to a functor from $\Parity_\mathscr{S}(X,k)$ to $\Parity_\mathscr{S}(X,k')$ (which we denote similarly).\footnote{This remark uses the property that the functor $k'(-)$ commutes with the functors $i_s^!$ and $i_s^*$. For $i_s^*$, this follows from~\cite[Proposition~2.6.5]{ks}. For $i_s^!$, we observe that~\cite[Proposition~3.1.11]{ks} provides a morphism of functors $k'(-) \circ i_s^! \rightarrow i_s^! \circ k'(-)$. To prove that this morphism is an isomorphism on $D^{\mathrm{b}}_{\mathscr{S}}(X,k)$, it suffices to remark that $D^{\mathrm{b}}_{\mathscr{S}}(X,k)$ is generated (as a triangulated category) by objects of the form $(i_{t})_*(\underline{M}_{X_t})$ where $t \in \mathscr{S}$ and $M$ is a finitely generated flat $k$-module, and that our morphism is clearly an isomorphism on such objects.}
We now assume that a connected complex algebraic group $A$ acts on $X$, stabilizing each stratum $X_s$. We will assume that $\HM^n_A(\mathrm{pt}; k)$ vanishes if $n$ is odd, and is projective over $k$ otherwise.
One can consider the $\mathscr{S}$-constructible $A$-equivariant derived category $D^{\mathrm{b}}_{A,\mathscr{S}}(X,k)$ in the sense of Bernstein--Lunts. We let $\For \colon D^{\mathrm{b}}_{A,\mathscr{S}}(X,k) \rightarrow D^{\mathrm{b}}_{\mathscr{S}}(X,k)$ be the forgetful functor. More generally, if $A' \subset A$ is a closed subgroup satisfying the same assumption as $A$, we have a forgetful functor $\For \colon D^{\mathrm{b}}_{A,\mathscr{S}}(X,k) \rightarrow D^{\mathrm{b}}_{A',\mathscr{S}}(X,k)$. We say that an object $\FC$ in $D^{\mathrm{b}}_{A,\mathscr{S}}(X,k)$ is a \emph{parity complex} if $\For(\FC)$ is a parity complex in the sense of Definition~\ref{def:parity}. We denote by $\Parity_{A,\mathscr{S}}(X,k)$ the subcategory of $D^{\mathrm{b}}_{A,\mathscr{S}}(X,k)$ consisting of parity complexes. For $k'$ a $k$-algebra satisfying the assumptions above, we also have an ``extension of scalars'' functor $k'(-)$ in this setting. If $\mathscr{S}$ is the stratification by $A$-orbits, we will omit it from the notation.
We refer to~\cite{jmw} for the main properties of parity sheaves, in the case $k$ is a complete local principal ideal domain. Here we will only need the properties below, which follow from~\cite[Proposition~2.6]{jmw}. (Note that the proof in \emph{loc}.~\emph{cit}.~does not use the running assumptions on the ring of coefficients.)
\begin{lem}
\label{lem:properties-parity}
\begin{enumerate}
\item
\label{it:parity-cohomology}
If $\FC$ is in $\Parity_{A, \mathscr{S}}(X,k)$, then $\HM^\bullet_{A}(X, \FC)$
is a finitely generated projective module over $\HM^\bullet_A(\mathrm{pt}; k)$. If $A' \subset A$ is a closed subgroup as above, then the natural morphism
\[
\HM^\bullet_{A'}(\mathrm{pt}; k) \otimes_{\HM^\bullet_A(\mathrm{pt}; k)} \HM^\bullet_{A}(X, \FC) \rightarrow \HM^\bullet_{A'}(X, \For(\FC))
\]
is an isomorphism. If $k'$ is a $k$-algebra as above, then the natural morphism
\[
k' \otimes_k \HM^\bullet_{A}(X, \FC) \rightarrow \HM^\bullet_{A}(X, k'(\FC))
\]
is also an isomorphism.
\item
\label{it:morph-parity-For}
If $\FC$, $\GC$ are in $\Parity_{A, \mathscr{S}}(X,k)$, then $\Hom^\bullet_{D^{\mathrm{b}}_{A,\mathscr{S}}(X, k)}(\FC, \GC)$ is a finitely generated projective $\HM^\bullet_A(\mathrm{pt}; k)$-module.
If $A' \subset A$ is a closed subgroup as above, then the natural morphism
\[
\HM^\bullet_{A'}(\mathrm{pt}; k) \otimes_{\HM^\bullet_A(\mathrm{pt}; k)} \Hom^\bullet_{D^{\mathrm{b}}_{A,\mathscr{S}}(X, k)}(\FC, \GC) \rightarrow \Hom^\bullet_{D^{\mathrm{b}}_{A',\mathscr{S}}(X, k)}(\For(\FC), \For(\GC))
\]
is an isomorphism.
If $k'$ is a $k$-algebra as above, then the natural morphism
\[
k' \otimes_k \Hom^\bullet_{D^{\mathrm{b}}_{A,\mathscr{S}}(X, k)}(\FC, \GC) \rightarrow \Hom^\bullet_{D^{\mathrm{b}}_{A,\mathscr{S}}(X, k')}(k'(\FC), k'(\GC))
\]
is also an isomorphism.\qed
\end{enumerate}
\end{lem}
\begin{remark}
In~\cite{jmw}, the ring $k$ is assumed to be a principal ideal domain, so every finitely generated projective $k$-module is free.
\end{remark}
We will also use the following observation.
\begin{lem}
\label{lem:parity-indec-For}
Assume that $k$ is complete local.
If $\FC$ is an object of $\Parity_{A,\mathscr{S}}(X,k)$ which is indecomposable, then its image $\For(\FC)$ in $\Parity_{\mathscr{S}}(X,k)$ is also indecomposable.
\end{lem}
\begin{proof}
Under our assumptions on $k$, $\Parity_{A,\mathscr{S}}(X,k)$ and $\Parity_{\mathscr{S}}(X,k)$ are Krull--Schmidt categories (see~\cite[Remark~2.1]{jmw}). Hence to prove the indecomposability of $\For(\FC)$ it suffices to prove that $\End_{D^{\mathrm{b}}_{\mathscr{S}}(X, k)}(\For(\FC))$ is a local ring. By Lemma~\ref{lem:properties-parity} the ring
$\Hom^\bullet_{D^{\mathrm{b}}_{\mathscr{S}}(X, k)}(\For(\FC), \For(\FC))$
is a quotient of $\Hom^\bullet_{D^{\mathrm{b}}_{A, \mathscr{S}}(X, k)}(\FC, \FC)$. Hence $\End_{D^{\mathrm{b}}_{\mathscr{S}}(X, k)}(\For(\FC))$ is a quotient of $\End_{D^{\mathrm{b}}_{A, \mathscr{S}}(X, k)}(\FC)$. Since the latter is local, the former is also local, which finishes the proof.
\end{proof}
Assume now that $k$ is an integral domain. Recall that, in this setting, the \emph{rank} of a projective $k$-module $M$, denoted $\rk_k(M)$, is the dimension of the $\mathrm{Frac}(k)$-vector space $\mathrm{Frac}(k) \otimes_k M$.
To $X$ one can associate the free $\ZM[\mathsf{v},\mathsf{v}^{-1}]$-module $\MM_X$ with basis $(\mathbf{e}_s : s \in \mathscr{S})$ indexed by $\mathscr{S}$, and to any $\FC$ in $\Parity_{A,\mathscr{S}}(X,k)$ the elements
\begin{align*}
\ch_X^*(\FC)=\sum_{\substack{s \in \mathscr{S} \\ j \in \ZM}} \rk_k \bigl( \HM^{j-\dim(X_s)}(X_s, i_s^*\FC) \bigr) \cdot \mathsf{v}^j \cdot \mathbf{e}_s, \\
\ch_X^!(\FC)=\sum_{\substack{s \in \mathscr{S}\\ j \in \ZM}} \rk_k \bigl( \HM^{j-\dim(X_s)}(X_s, i_s^!\FC) \bigr) \cdot \mathsf{v}^{-j} \cdot \mathbf{e}_s.
\end{align*}
We also define a bilinear form $\langle- , - \rangle$ on $\MM_X$, with values in $\ZM[\mathsf{v},\mathsf{v}^{-1}]$, by setting
\[
\langle \mathsf{v}^n \mathbf{e}_s, \mathsf{v}^m \mathbf{e}_t \rangle = \mathsf{v}^{-n-m} \delta_{s,t}.
\]
The following result is also an easy application of~\cite[Proposition~2.6]{jmw}, whose proof is left to the reader.
\begin{lem}
\label{lem:grk-Hom-parity}
Assume that $k$ is an integral domain.
\begin{enumerate}
\item
\label{it:ch-parity-D}
If $\FC$ is in $\Parity_{A, \mathscr{S}}(X,k)$, then its Grothendieck--Verdier dual $\DM_X(\FC)$ is also in $\Parity_{A, \mathscr{S}}(X,k)$, and moreover we have
\[
\ch_X^!(\DM_X(\FC)) = \ch_X^*(\FC).
\]
\item
\label{it:dim-Hom-parity-ch}
If $\FC$, $\GC$ are in $\Parity_{A, \mathscr{S}}(X,k)$, then we have
\[
\pushQED{\qed}
\grk_{\HM^\bullet_A(\mathrm{pt}; k)} \bigl( \Hom^\bullet_{D^{\mathrm{b}}_{A,\mathscr{S}}(X, k)}(\FC, \GC) \bigr) = \langle \ch_X^*(\FC), \ch_X^!(\GC) \rangle.
\qedhere
\popQED
\]
\end{enumerate}
\end{lem}
\begin{remark}
In the paper we will also use the straightforward generalization of the notions and results of this subsection to the case of ind-varieties, in the setting of~\cite[\S 2.7]{jmw}.
\end{remark}
\subsection{Equivariant coherent sheaves}
\label{ss:equiv-coh}
Let $X$ be a Noetherian scheme, and let ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$ be an affine group scheme over $X$. Consider a Noetherian $X$-scheme $Y$ endowed with an action of ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$, i.e.~we are given a morphism $a\colon {\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y \rightarrow Y$ of $X$-schemes which satisfies the natural compatibility property with the group structure on ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$. Let also $p \colon {\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y \rightarrow Y$ be the projection. Then
one can define the category $\Coh^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}(Y)$ of ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$-equivariant coherent sheaves on $Y$ as the category of pairs $(\FC, \phi)$ where $\FC$ is a coherent sheaf on $Y$ and $\phi \colon p^*(\FC) \xrightarrow{\sim} a^*(\FC)$ is an isomorphism which satisfies the usual cocyle condition. If ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$ is flat over $X$, then this category is abelian.
We define the ``universal stabilizer'' as the fiber product
\begin{equation}
\label{eqn:stabilizer}
{\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}:= Y \times_{Y \times_X Y} ({\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y)
\end{equation}
where the morphism $Y \rightarrow Y \times_X Y$ is the diagonal embedding (which we will denote by $\Delta$), and the morphism ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y \rightarrow Y \times_X Y$ is $a \times p$. Then ${\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}$ is an affine group scheme over $Y$, and the natural morphism ${\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}} \rightarrow {\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y$ is a closed embedding of $Y$-group schemes.
The goal of this subsection is to recall the construction of a faithful functor
\begin{equation}
\label{eqn:stabilizer-functor}
\Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}}(Y) \rightarrow \Rep({\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}})
\end{equation}
whose composition with the forgetful functor $\Rep({\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}) \rightarrow \Coh(Y)$ coincides with the natural forgetful functor $\For^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}_Y \colon \Coh^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}(Y) \rightarrow \Coh(Y)$.
We begin with a remark. Let $X'$ be a Noetherian $X$-scheme. Then one can consider the $X'$-group scheme ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}':={\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X X'$, which acts naturally on $Y':=Y \times_X X'$. If $f \colon Y' \rightarrow Y$ is the natural morphism, then we remark that the usual pullback functor $f^*\colon \Coh(Y) \rightarrow \Coh(Y')$ induces a functor from $\Coh^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}(Y)$ to $\Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}'}(Y')$, which we also denote $f^*$. In particular, using this construction for the morphism $Y \rightarrow X$, we obtain a functor
\[
p_1^* \colon \Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}}(Y) \rightarrow \Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y}(Y \times_X Y),
\]
where $Y \times_X Y$ is considered as a $Y$-scheme through the second projection.
Using this functor we can define~\eqref{eqn:stabilizer-functor} as the composition
\[
\Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}}(Y) \xrightarrow{p_1^*} \Coh^{{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}} \times_X Y}(Y \times_X Y) \rightarrow \Coh^{{\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}}(Y \times_X Y) \xrightarrow{\Delta^*} \Coh^{{\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}}(Y) = \Rep({\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}),
\]
where the second arrow is the restriction functor, and ${\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}$ acts trivially on $Y$. (Note that $\Delta$ is an ${\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}$-equivariant morphism of $Y$-schemes.)
More generally,
if $Z \subset Y$ is a (not necessarily ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$-stable) closed subscheme, one can consider the restriction ${\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}_Z:= Z \times_Y {\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}$, and the composition
\[
\Coh^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}(Y) \rightarrow \Rep({\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}_Z)
\]
of~\eqref{eqn:stabilizer-functor} with restriction to $Z$. The composition of this functor with the forgetful functor $\Rep({\mathbf S}} \def\sb{{\mathbf s}} \def\SC{{\mathcal{S}}_Z) \rightarrow \Coh(Z)$ is the composition $\Coh^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}(Y) \xrightarrow{\For^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}_Y} \Coh(Y) \rightarrow \Coh(Z)$, where the second functor is restriction.
\subsection{Deformation to the normal cone}
\label{ss:def-normal-cone}
Let
$A$ be a ring, and let
$I \subset A$ be an ideal. We define the associated \emph{deformation to the normal cone}\footnote{Our terminology comes from geometry: if $X$ is the spectrum of $A$ and $Y \subset X$ the closed subscheme associated with $I$, then the spectrum of $\mathrm{DNC}_I(A)$ is (a slight variant of) the usual deformation to the normal cone of $X$ along $Y$ considered e.g.~in~\cite[Chap.~5]{fulton}.} $\mathrm{DNC}_I(A)$ as the subalgebra of $A[\hbar^{\pm 1}]$ generated by $A[\hbar]$ together with the elements of the form $\hbar^{-1} f$ for $f \in I$.
\begin{lem}
\label{lem:def-normal-cone-BC}
Let $A'$ be a ring, and let $A \rightarrow A'$ be a flat ring morphism. We denote by $I'$ the ideal of $A'$ generated by the image of $I$. Then there exists a canonical isomorphism of $k$-algebras
\[
A' \otimes_A \mathrm{DNC}_I(A) \xrightarrow{\sim} \mathrm{DNC}_{I'}(A').
\]
\end{lem}
\begin{proof}
The
natural morphism
\[
A' \otimes_{A} \mathrm{DNC}_I(A) \rightarrow A' \otimes_{A} \bigl( A[\hbar^{\pm 1}] \bigr) = A'[\hbar^{\pm 1}]
\]
is injective. The image of this morphism is clearly $\mathrm{DNC}_{I'}(A')$, which provides the desired isomorphism.
\end{proof}
\subsection{Two lemmas on graded modules over polynomial rings}
\label{ss:nakayama}
Let $k$ be a ring, and let $V$ be a free $k$-module of finite rank. We denote by $A$ be the symmetric algebra of $V$ over $k$, which we consider as a $\ZM$-graded $k$-algebra with the generators $V \subset A$ in degree $2$. We consider the trivial $A$-module $k$ as a graded module concentrated in degree $0$.
\begin{lem}
\label{lem:nakayama}
\begin{enumerate}
\item
\label{it:nakayama-Tor1}
Let $M$ be a graded $A$-module which is bounded below for the internal grading, and assume that
\[
k \otimes_A M \text{ is graded free over $k$} \quad \text{and} \quad \Tor_{1}^{A}(k,M)=0.
\]
Then $M$ is graded free over $A$.
\item
\label{it:nakayama-complex}
Let $M$ be an object of the bounded derived category of graded $A$-modules, and assume that the cohomology modules of $M$ are all bounded below for the internal grading. Assume that the complex $k\@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A} M$ is concentrated in degree $0$, and graded free over $k$. Then $M$ is concentrated in degree $0$, and is graded free over $A$.
\end{enumerate}
\end{lem}
\begin{proof}
\eqref{it:nakayama-Tor1} Let $M'$ be a free graded $A$-module and let $f \colon M' \rightarrow M$ be a morphism such that the induced morphism $k \otimes_A M' \rightarrow k \otimes_{A} M$ is an isomorphism. (Such an $M'$ exists by our first assumption.) By the graded Nakayama lemma, $f$ is surjective. Let $M''$ be its kernel, and consider the exact sequence
\[
\Tor_{1}^{A}(k,M) \rightarrow k \otimes_{A} M'' \rightarrow k \otimes_{A} M' \rightarrow k \otimes_{A} M \rightarrow 0.
\]
This exact sequence shows, using our second assumption, that $k \otimes_{A} M''=0$. By the graded Nakayama lemma again, we deduce that $M''=0$, which finishes the proof.
\eqref{it:nakayama-complex} An easy argument using the graded Nakayama lemma shows that $M$ is concentrated in non-positive degrees. Now, consider the truncation triangle
\[
\tau_{<0} M \rightarrow M \rightarrow \mathsf{H}^0(M) \xrightarrow{+1}.
\]
Applying the functor $k \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A}(-)$ we obtain a distinguished triangle
\[
k \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A} \tau_{<0} M \rightarrow k \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A} M \rightarrow k \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A}\mathsf{H}^0(M) \xrightarrow{+1}.
\]
The associated long exact sequence of cohomology implies that $\Tor_{1}^{A}(k,\mathsf{H}^0(M))=0$, hence $\mathsf{H}^0(M)$ is free by~\eqref{it:nakayama-Tor1}. Considering again the long exact sequence we obtain that $k \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{A} \tau_{<0} M=0$, hence that $\tau_{<0} M=0$ by the graded Nakayama lemma, which finishes the proof.
\end{proof}
Now we assume
that we are given an open $k$-subscheme $V' \subset V$.
Then, for any algebraically closed field $\FM$ and any ring morphism $k \rightarrow \FM$, we set
\[
V_\FM := \Spec(\FM) \times_{\Spec(k)} V, \qquad
V'_\FM := \Spec(\FM) \times_{\Spec(k)} V'.
\]
\begin{lem}
\label{lem:morphism-grk}
Let $M$, $N$ be free graded $A$-modules of finite rank, and let $\varphi \colon M \rightarrow N$ be a morphism of graded $A$-modules. Assume that $\grk_{A}(M)=\grk_{A}(N)$, and that the morphism
\[
\varphi'_\FM \colon \OC(V'_\FM) \otimes_{A} M \rightarrow \OC(V'_\FM) \otimes_{A} N
\]
induced by $\varphi$ is injective for any algebraically closed field $\FM$ and any ring morphism $k \rightarrow \FM$. Then $\varphi$ is an isomorphism.
\end{lem}
\begin{proof}
Let $\varphi_\FM \colon \FM \otimes_k M \rightarrow \FM \otimes_k N$ be the morphism induced by $\varphi$. Then,
considering the commutative diagram
\[
\xymatrix@R=0.6cm{
\FM \otimes_k M \ar[d] \ar[r]^-{\varphi_\FM} & \FM \otimes_k N \ar[d] \\
\OC(V'_\FM) \otimes_{A} M \ar[r]^-{\varphi'_\FM} & \OC(V'_\FM) \otimes_{A} N
}
\]
(where the verticall arrows are injective since the morphism $V'_\FM \rightarrow V_\FM$ is an open embbeding) we see that $\varphi_\FM$ is injective, hence an isomorphism under our assumption on graded ranks. Now if $i \in \ZM$, the $k$-modules $M_i$ and $N_i$ are free of (the same) finite rank, and the determinant of the restriction $\varphi_i \colon M_i \rightarrow N_i$ of $\varphi$ (in any fixed choice of bases) does not belong to any prime ideal of $k$. Hence this determinant is invertible, proving that $\varphi$ is an isomorphism.
\end{proof}
\subsection{Affine braid groups and associated Hecke algebras}
\label{ss:Haff}
Let $({\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}},\Phi,{\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}},{\check \Phi})$ be a root datum, and let $\Phi^+ \subset \Phi$ be a system of positive roots. We will assume that ${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}/ \ZM {\check \Phi}$ has no torsion (or in other words that the connected reductive groups with root datum $({\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}},\Phi,{\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}},{\check \Phi})$ have a simply-connected derived subgroup). For $\alpha \in \Phi$, we denote by $\alpha^\vee$ the corresponding coroot, and by $s_\alpha$ the associated reflection.
Let $W$ be the corresponding Weyl group, and $W_{\mathrm{aff}} := W \ltimes {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ be the associated affine Weyl group. To avoid confusions, for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ we denote by $t_\lambda$ the corresponding element of $W_{\mathrm{aff}}$. We let $\ZM \Phi \subset {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ be the root lattice; then the subgroup $W_{\mathrm{aff}}^\mathrm{Cox}:=W \ltimes (\ZM \Phi) \subset W_{\mathrm{aff}}$ is a Coxeter group with generators given by reflections along walls of the fundamental dominant alcove. (In other words, we use the same conventions as in~\cite[\S 1.4]{LuAff}.)
The simple reflections which belong to $W$ will be called \emph{finite}; the ones which do not belong to $W$ will be called \emph{affine}.
Let us consider the length function $\ell \colon W_{\mathrm{aff}} \rightarrow \ZM_{\geq 0}$ defined as follows: for $w \in W$ and $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ we set
\begin{equation}
\label{eqn:length}
\ell(w \cdot t_\lambda)=\sum_{\alpha \in \Phi^+ \cap w^{-1} (\Phi^+)} |\langle \lambda, \alpha^\vee \rangle | + \sum_{\alpha \in \Phi^+ \cap w^{-1}(-\Phi^+)}
|1 + \langle \lambda, \alpha^\vee \rangle |.
\end{equation}
Then the restriction of $\ell$ to $W_{\mathrm{aff}}^\mathrm{Cox}$ is the length function associated with the Coxeter structure considered above. We denote by $\Omega$ the subgroup of $W_{\mathrm{aff}}$ consisting of elements of length $0$; it is a commutative group isomorphic to ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}/ \ZM\Phi$ via the composition of natural maps $\Omega \hookrightarrow W \ltimes {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \twoheadrightarrow {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \twoheadrightarrow {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}/\ZM\Phi$. Moreover,
any element of $W_{\mathrm{aff}}$ can be written in the form $\omega v$ for unique $\omega \in \Omega$ and $v \in W_{\mathrm{aff}}^\mathrm{Cox}$. (For all of this, see~\cite[\S 1.5]{LuAff}.)
We will also consider the braid group $\BM_{\mathrm{aff}}$ associated with $W_{\mathrm{aff}}$. It is defined as the group generated by elements $T_w$ for $w \in W_{\mathrm{aff}}$, with relations $T_{vw} = T_v T_w$ for all $v,w \in W_{\mathrm{aff}}$ such that $\ell(vw)=\ell(v)+\ell(w)$. There exists a canonical surjection $\BM_{\mathrm{aff}} \twoheadrightarrow W_{\mathrm{aff}}$ sending $T_w$ to $w$. One can define (following Bernstein and Lusztig), for each $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, an element $\theta_\lambda \in \BM_{\mathrm{aff}}$, see e.g.~\cite[\S 1.1]{riche} for details. (This element is denoted $\overline{T}_\lambda$ in~\cite[\S 2.6]{LuAff}.) Then $\BM_{\mathrm{aff}}$ admits a second useful presentation (usually called the \emph{Bernstein presentation}), with generators $\{T_w, \, w \in W\}$ and $\{\theta_\lambda, \, \lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}\}$, and the following relations (where $v,w \in W$, $\lambda, \mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, and $\alpha$ runs over simple roots):
\begin{enumerate}
\item
$T_v T_w = T_{vw}$ \quad if $\ell(vw)=\ell(v)+\ell(w)$;
\item
$\theta_\lambda \theta_\mu = \theta_{\lambda+\mu}$;
\item
$T_{s_\alpha} \theta_\lambda = \theta_\lambda T_{s_\alpha}$ \quad if $\langle \lambda, \alpha^\vee \rangle = 0$;
\item
$\theta_\lambda = T_{s_\alpha} \theta_{\lambda-\alpha} T_{s_\alpha}$ \quad if $\langle \lambda, \alpha^\vee \rangle = 1$.
\end{enumerate}
(See~\cite{br-appendix} for a proof of this fact in the case ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}/\ZM\Phi$ is finite; the general case is similar.)
The following lemma is proved in~\cite[Lemma~6.1.2]{riche2}.
\begin{lem}
\label{lem:affine-simple-conjugate}
For any affine simple reflection $s_0$, there exist a finite simple reflection $t$ and an element $b \in \BM_{\mathrm{aff}}$ such that $T_{s_0} = b \cdot T_t \cdot b^{-1}$. \qed
\end{lem}
We define the affine Hecke algebra $\HM_{\mathrm{aff}}$ as the quotient of the group algebra of $\BM_{\mathrm{aff}}$ over $\ZM[\mathsf{v},\mathsf{v}^{-1}]$ by the relations
\[
(T_s+\mathsf{v}^{-1})(T_s-\mathsf{v})=0
\]
for all finite simple reflections $s$. (Note that the same formula for affine simple roots automatically follows by Lemma~\ref{lem:affine-simple-conjugate}.) We denote by $\HM_{W}$ the subalgebra of $\HM_{\mathrm{aff}}$ generated by $\ZM[\mathsf{v},\mathsf{v}^{-1}]$ and the elements $T_w$ for $w \in W$, and by $\MM_{\mathrm{triv}}$ the $\HM_{W}$-module which is free of rank one over $\ZM[\mathsf{v},\mathsf{v}^{-1}]$, and where $T_s$ acts by multiplication by $\mathsf{v}$ for each finite simple reflection $s$.
We define the ``spherical'' right $\HM_{\mathrm{aff}}$-module
\[
\MM_{\mathrm{sph}} := \MM_{\mathrm{triv}} \otimes_{\HM_{W}} \HM_{\mathrm{aff}}.
\]
We denote by $\mathbf{m}_0 \in \MM_{\mathrm{sph}}$ the element $1 \otimes 1$. For $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, we denote by $w_\lambda$ the shortest representative in $W t_\lambda \subset W_{\mathrm{aff}}$, and set $\mathbf{m}_\lambda:=\mathbf{m}_0 \cdot T_{w_\lambda}$. Then the elements $\mathbf{m}_\lambda$, $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, form a $\ZM[\mathsf{v},\mathsf{v}^{-1}]$-basis of $\MM_{\mathrm{sph}}$. We define a bilinear form $\langle -, - \rangle$ on $\MM_{\mathrm{sph}}$, with values in $\ZM[\mathsf{v},\mathsf{v}^{-1}]$, by setting
\[
\langle \mathsf{v}^i \mathbf{m}_\lambda, \mathsf{v}^j \mathbf{m}_\mu \rangle = \mathsf{v}^{-j-i} \delta_{\lambda,\mu}.
\]
For any sequence of simple reflections $\underline{s}=(s_1, \cdots, s_r)$ and any $\omega \in \Omega$ we will consider the element
\[
\mathbf{m}(\omega, \underline{s}) := \mathbf{m}_0 \cdot T_{\omega} \cdot (T_{s_1} + \mathsf{v}^{-1}) \cdots (T_{s_r}+\mathsf{v}^{-1}) \qquad \in \MM_{\mathrm{sph}}.
\]
\begin{remark}
Our element $T_s \in \HM_{\mathrm{aff}}$ corresponds to the element denoted $H_s$ in~\cite{soergel-kl}, while our $\mathsf{v}$ corresponds to $v^{-1}$ in~\cite{soergel-kl}.
\end{remark}
\section{Constructible side}
\label{sec:constructible-side}
\subsection{Overview}
\label{ss:overview}
In this section we describe the category $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ in terms of an appropriate category of Soergel bimodules using a ``total cohomology'' functor. Similar constructions appear in~\cite{soergel-kategorie, soergel, modrap1} for flag varieties of reductive groups, and in~\cite{by} for flag varieties of Kac--Moody groups (with coefficients in characteristic zero).
The main difference with these works is that in our case the cohomology algebra is much more complicated. To overcome this difficulty we work over a certain ring ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ of integers, which allows to replace this cohomology algebra by a ``simplified model,'' see~\S\ref{ss:key-lemma} for a discussion of this idea.
After setting the notation in~\S\ref{ss:notation-constructible}, we introduce our ``Soergel bimodules'' in~\S\S\ref{ss:deformations}--\ref{ss:algebraic-BS}. In~\S\ref{ss:cohomology-Gr} we study the equivariant cohomology of ${\EuScript Gr}$. In~\S\ref{ss:topological-BS} we explain how the category $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ can be recovered from a certain category of (equivariant) ``Bott--Samelson parity sheaves'' over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Then in~\S\ref{ss:cohomology-functors} we introduce our ``total cohomology functor,'' and in~\S\S\ref{ss:equivalence}--\ref{ss:proof-H-ff} we prove that this functor induces an equivalence between ``Bott--Samelson'' parity sheaves and Soergel bimodules.
Finally, in~\S\ref{ss:graded-ranks-parity} we derive a formula for the graded rank of the space of morphisms between certain Soergel bimodules, which will play an important role in a proof on the ``coherent side'' (see~\S\ref{ss:equivalence-tilting}).
\subsection{Notation}
\label{ss:notation-constructible}
In this section we let $\check{G}$ be a connected reductive algebraic group over $\CM$, with a chosen Borel subgroup $\check{B} \subset \check{G}$ and maximal torus $\check{T} \subset \check{B}$. We let ${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}:=X^*(\check{T})$ be the lattice of characters of $\check{T}$, and ${\check \Phi} \subset {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$ be the roots of $\check{G}$. We also fix a finite localization ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ of $\ZM$. In the whole section we will make the following assumptions:
\begin{enumerate}
\item
${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}} / \ZM {\check \Phi}$ has no torsion (or in other words the connected reductive groups which are Langlands dual to $\check{G}$ have a simply-connected derived subgroup);
\item
all the torsion primes of the ``refined root system'' ${\check \Phi} \subset {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$ (in the sense of~\cite[Section~5]{demazure}) are invertible in ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\end{enumerate}
In the later sections we will apply our results in the case $\check{G}$ is a product of simple groups (of adjoint type) and general linear groups; in this case the first condition is automatic, and the second condition means that the prime numbers which are not very good for some simple factor of $\check{G}$ are invertible in ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
Let $\mathscr{O}:=\CM[ \hspace{-1pt} [z] \hspace{-1pt} ]$ and $\mathscr{K}:=\CM( \hspace{-1pt} (z) \hspace{-1pt} )$. We consider the affine Grassmannian
\[
{\EuScript Gr} := \check{G}(\mathscr{K}) / \check{G} (\mathscr{O}),
\]
with its natural ind-variety structure. We denote by $\check{I}$ the Iwahori subgroup of $\check{G}(\mathscr{O})$ determined by $\check{B}$, i.e.~the inverse image of $\check{B}$ under the morphism $\check{G}(\mathscr{O}) \rightarrow \check{G}$ defined by the evaluation at $z=0$. Then $\check{I}$ acts naturally on ${\EuScript Gr}$ via left multiplication on $\check{G}(\mathscr{K})$. We also let the multiplicative group $\Gm$ act on ${\EuScript Gr}$ by loop rotation (i.e.~via $x \cdot g(z) = g(x^{-1} z)$), so that we obtain an action of the semi-direct product $\check{I} \rtimes \Gm$.
The main players of this section are the categories
\[
\Parity_{\check{I} \rtimes \Gm} ({\EuScript Gr},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}), \quad \Parity_{\check{I}} ({\EuScript Gr},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \quad \text{and} \quad \Parity_{(\check{I})}({\EuScript Gr},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
Here, in the right-hand side, we use the notation $\Parity_{(\check{I})}({\EuScript Gr},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ for the category $\Parity_{\mathscr{S}}({\EuScript Gr},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ where $\mathscr{S}$ is the stratification of ${\EuScript Gr}$ by orbits of $\check{I}$. If $\FM$ is a field (not necessarily algebraically closed) whose characteristic is invertible in ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, we can consider the unique algebra morphism ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \FM$, and the categories
\begin{equation}
\label{eqn:Parity-FM}
\Parity_{\check{I} \rtimes \Gm} ({\EuScript Gr},\FM), \quad \Parity_{\check{I}} ({\EuScript Gr},\FM) \quad \text{and} \quad \Parity_{(\check{I})}({\EuScript Gr},\FM).
\end{equation}
We let ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}:=X_*(\check{T})$ be the lattice of cocharacters of $\check{T}$, and $\Phi \subset {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ be the coroots of $\check{G}$ (with respect to $\check{T}$). The choice of $\check{B}$ determines a system of positive roots: more precisely we denote by ${\check \Phi}^+ \subset {\check \Phi}$ the roots which are \emph{opposite} to the $\check{T}$-weights in the Lie algebra of $\check{B}$. We denote by $\Phi^+ \subset \Phi$ the corresponding system of positive coroots.
To these data one can associate the affine Weyl group $W_{\mathrm{aff}}$ and its length function $\ell$ as in~\S\ref{ss:Haff}.
Recall that the $\check{I}$-orbits on ${\EuScript Gr}$ are parametrized in a natural way by $W_{\mathrm{aff}}/W \cong {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, and that each $\check{I}$-orbit is stable under the action of $\check{I} \rtimes \Gm$. More precisely, any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ defines a point $z^\lambda \in \check{T}(\mathscr{K}) \subset \check{G}(\mathscr{K})$. We set $L_\lambda := z^\lambda \check{G}(\mathscr{O}) / \check{G}(\mathscr{O}) \in {\EuScript Gr}$, and ${\EuScript Gr}_\lambda:=\check{I} \cdot L_\lambda$. Then we have
\[
{\EuScript Gr} = \bigsqcup_{\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}} {\EuScript Gr}_\lambda.
\]
Moreover, the dimension of ${\EuScript Gr}_\lambda$ is the length of the shortest representative in $t_\lambda W \subset W_{\mathrm{aff}}$, i.e.~$\ell(w_{-\lambda})$.
For any $w \in t_\lambda W$ we also set ${\EuScript Gr}_w := {\EuScript Gr}_\lambda$.
Let $\FM$ be as above, and consider the categories
in~\eqref{eqn:Parity-FM}. By~\cite{jmw} these categories
are all Krull--Schmidt, and their indecomposable objects can be described as follows. For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ there exists a unique indecomposable object $\EC_\lambda$ in $\Parity_{\check{I} \rtimes \Gm} ({\EuScript Gr},\FM)$
which is supported on $\overline{{\EuScript Gr}_\lambda}$
and whose restriction to ${\EuScript Gr}_\lambda$
is
$\underline{\FM}_{{\EuScript Gr}_\lambda}[\dim({\EuScript Gr}_\lambda)]$.
Moreover any indecomposable object in $\Parity_{\check{I} \rtimes \Gm} ({\EuScript Gr},\FM)$
is isomorphic to $\EC_\lambda[n]$
for some unique $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $n \in \ZM$. By Lemma~\ref{lem:parity-indec-For}, the images of $\EC_\lambda$
under the appropriate forgetful functors to $\Parity_{\check{I}} ({\EuScript Gr},\FM)$ and $\Parity_{(\check{I})} ({\EuScript Gr},\FM)$
remain indecomposable; for simplicity these images will still be denoted by $\EC_\lambda$.
The same description of indecomposable objects as above applies in the categories $\Parity_{\check{I}} ({\EuScript Gr},\FM)$
and $\Parity_{(\check{I})} ({\EuScript Gr},\FM)$.
The connected components of ${\EuScript Gr}$ are parametrized by $\Omega$; for $\omega \in \Omega$ we denote by ${\EuScript Gr}_{(\omega)}$ the corresponding component. (In fact, if we identify $\Omega$ with ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}/\ZM\Phi$ as in~\S\ref{ss:Haff}, then ${\EuScript Gr}_{(\omega)}$ is the union of the orbits ${\EuScript Gr}_\lambda$ where $\lambda$ has image $\omega$ in ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}/\ZM\Phi$.)
Below we will also use the affine flag variety
\[
\FC l := \check{G}(\mathscr{K}) / \check{I},
\]
with its natural ind-variety structure, and the natural $\check{I}$-action. The $\check{I}$-orbits on $\FC l$ are parametrized in a natural way by $W_{\mathrm{aff}}$, and are stable under the loop rotation action. If $\FC l_w$ is the orbit associated with $w \in W_{\mathrm{aff}}$, then we have
$\dim(\FC l_w)=\ell(w)$, and the image of $\FC l_w$ under the natural projection $\FC l \twoheadrightarrow {\EuScript Gr}$ is ${\EuScript Gr}_w$.
For each simple reflection $s$, the orbit $\FC l_s$ is isomorphic to $\AM^1_{\CM}$, and its closure $\overline{\FC l_s}$ is isomorphic to $\PM^1_{\CM}$.
We define
\[
\tg := \check{\XB} \otimes_\ZM {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}, \qquad \tg^* := \Hom_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) = {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \otimes_\ZM {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
(In fact, $\tg$ is the Lie algebra of the ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-torus which is Langlands dual to $\check{T}$.) Then there exists a canonical isomorphism of graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-algebras
\begin{equation}
\label{eqn:cohomology-TD}
\OC(\tg^*) = \mathrm{S}(\tg) \xrightarrow{\sim} \HM^\bullet_{\check{T}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}),
\end{equation}
where $\tg$ is in degree $2$, and $\mathrm{S}(\tg)$ is the symmetric algebra of $\tg$. Using~\cite[Theorem~1.3(2)]{totaro}, we deduce, under our assumptions on ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$,\footnote{Recall that the torsion primes of $\check{G}$ (in the sense of~\cite{totaro}) are the same as the torsion primes of the ``refined root system'' ${\check \Phi} \subset {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$ (in the sense of~\cite{demazure}).} a canonical isomorphism
\begin{equation}
\label{eqn:cohomology-GD}
\OC(\tg^*/ W) \xrightarrow{\sim} \HM^\bullet_{\check{G}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\end{equation}
We will also identify $\HM^\bullet_{\Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ with ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$ (where $\hbar$ is an indeterminate, in degree $2$) in the natural way.
\begin{lem}
\label{lem:t/W-affine-space}
The ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-scheme $\tg^*/W$ is isomorphic to an affine space. Moreover, $\OC(\tg^*)$ is free over $\OC(\tg^*/W)$.
\end{lem}
\begin{proof}
The claims follow from~\cite[Th{\'e}or{\`e}me~2(c) \& Th{\'e}or{\`e}me~3]{demazure} and our assumption on ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, since $\OC(\tg^*/W)=\mathrm{S}_\ZM({\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}})^W \otimes_\ZM {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\end{proof}
\begin{remark}
The main result of this section will be proved in the $\check{I}$-equivariant setting; the $\check{I} \rtimes \Gm$-equivariant setting will be used only for technical purposes. However, similar results hold in the $\check{I} \rtimes \Gm$-equivariant case. On the ``coherent side'', replacing $\check{I}$ by $\check{I} \rtimes \Gm$ amounts to deforming coherent sheaves on $\widetilde{\gg}$ to asymptotic $\DC$-modules on $\mathscr{B}$; see~\cite{dodd} for details in the characteristic zero case.
\end{remark}
\subsection{Some algebras}
\label{ss:deformations}
We denote by $\Delta \subset \tg^* / W \times \tg^* / W$ the diagonal copy of $\tg^* / W$, and by $I_1 \subset \OC(\tg^* / W \times \tg^* / W)$ the associated ideal
(i.e.~the ideal generated by elements of the form $f \otimes 1 - 1 \otimes f$ for $f \in \OC(\tg^* / W)$). We also denote by $I_2 \subset \OC(\tg^* / W \times \tg^*)$, resp.~$I_3 \subset \OC(\tg^* \times \tg^*)$, the ideal generated by the image of $I_1$ under the ring morphism associated with the quotient morphism $\tg^* / W \times \tg^* \rightarrow \tg^* / W \times \tg^* /W$, resp.~$\tg^* \times \tg^* \rightarrow \tg^* / W \times \tg^* /W$.
We will consider the $\ZM$-graded algebras\footnote{Here the letter ``$C$'' stands for ``coinvariants,'' since these algebras will play the role played by the coinvariant algebra in~\cite{soergel-kategorie, soergel, modrap1}.}
\begin{align*}
C_\hbar := \mathrm{DNC}_{I_2}(\OC(\tg^* / W \times \tg^*)),& \qquad \widetilde{C}_\hbar := \mathrm{DNC}_{I_3}(\OC(\tg^* \times \tg^*)), \\
C := C_\hbar / \hbar \cdot C_\hbar,& \qquad \widetilde{C} := \widetilde{C}_\hbar / \hbar \cdot \widetilde{C}_\hbar.
\end{align*}
Here the grading is induced by the grading on $\OC(\tg^*)$ and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$ from~\S\ref{ss:notation-constructible}. We also set $C'_\hbar := \mathrm{DNC}_{I_1}(\OC(\tg^* / W \times \tg^*/W))$.
By Lemmas~\ref{lem:def-normal-cone-BC} and~\ref{lem:t/W-affine-space},
we have canonical isomorphisms
\begin{equation}
\label{eqn:isom-Chbar}
C_\hbar \cong \OC(\tg^* / W \times \tg^*) \otimes_{\OC(\tg^* / W \times \tg^*/W)} C'_\hbar, \quad
\widetilde{C}_\hbar \cong \OC(\tg^* \times \tg^*) \otimes_{\OC(\tg^* / W \times \tg^*)} C_\hbar.
\end{equation}
\begin{lem}
\label{lem:deformation-flat}
\begin{enumerate}
\item
The two natural
ring morphisms
$\OC(\tg^*/W)[\hbar] \rightarrow C'_\hbar$
are flat.
\label{it:def-flat-1}
\item
The natural
ring morphisms $\OC(\tg^*/W)[\hbar] \rightarrow C_\hbar$ and $\OC(\tg^*)[\hbar] \rightarrow C_\hbar$
are flat.
\label{it:def-flat-2}
\item
The two natural ring morphisms
$\OC(\tg^*)[\hbar] \rightarrow \widetilde{C}_\hbar$
are flat.
\label{it:def-flat-3}
\end{enumerate}
\end{lem}
\begin{proof}
First we treat~\eqref{it:def-flat-1}.
By symmetry, it is enough to prove the claim in the case of the morphism induced by the first projection $\tg^* / W \times \tg^*/W \rightarrow \tg^* / W$.
By Lemma~\ref{lem:t/W-affine-space}, we can fix an isomorphism of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-schemes
$\tg^* / W \cong \mathbb{A}^n_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ for some $n \in \ZM_{\geq 0}$. Then $\Delta$ identifies with the diagonal copy of $\mathbb{A}^n_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ in $\mathbb{A}^{2n}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Writing $\mathbb{A}^{2n}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ as the direct sum of the diagonal and antidiagonal copies of $\mathbb{A}^n_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, we obtain ring isomorphisms
\[
C'_\hbar \cong {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[x_1, \cdots, x_n] \otimes \mathrm{DNC}_{I_+}({\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[y_1, \cdots, y_n]) \\
\cong {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[x_1, \cdots, x_n, z_1, \cdots, z_n, \hbar],
\]
where $I_+ \subset {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[y_1, \cdots, y_n]$ is the ideal of the subscheme $\{0\} \subset \mathbb{A}^n_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}$, and $z_i := \hbar^{-1} y_i$.
With these identifications, the morphism under consideration is defined by $\hbar \mapsto \hbar$, $x_i \mapsto x_i + \hbar z_i$. This morphism is clearly flat.
The second claim in~\eqref{it:def-flat-2} is an immediate consequence of~\eqref{it:def-flat-1} and the first isomorphism in~\eqref{eqn:isom-Chbar}. To prove the first claim we decompose the morphism as the composition $\OC(\tg^*/W)[\hbar] \rightarrow C_\hbar' \rightarrow C_\hbar$. Now the first morphism is flat by~\eqref{it:def-flat-1}, and the second one is flat by the first isomorphism in~\eqref{eqn:isom-Chbar} since the projection $\tg^* \rightarrow \tg^*/W$ is flat (see Lemma~\ref{lem:t/W-affine-space}). This implies the desired claim.
Finally, in~\eqref{it:def-flat-3} the flatness of the morphism induced by the first projection $\tg^* \times \tg^* \rightarrow \tg^*$ follows from~\eqref{it:def-flat-2} and the second isomorphism in~\eqref{eqn:isom-Chbar}. Then the flatness of the other morphism follows by symmetry.
\end{proof}
Let us denote by
\[
f_1, f_2 \colon \OC(\tg^*)[\hbar] \rightarrow \widetilde{C}_\hbar
\]
the morphisms considered in Lemma~\ref{lem:deformation-flat}\eqref{it:def-flat-3}.
These morphisms
endow $\widetilde{C}_\hbar$ with the structure of a graded $\OC(\tg^*)[\hbar]$-bimodule.
In fact this algebra has a natural structure of bialgebra in the category of graded $\OC(\tg^*)[\hbar]$-bimodules, constructed as follows.
Consider the ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-algebra morphism
\begin{equation}
\label{eqn:comultiplication}
\OC(\tg^* \times \tg^*)[\hbar] \rightarrow \widetilde{C}_\hbar \otimes_{\OC(\tg^*)[\hbar]} \widetilde{C}_{\hbar}
\end{equation}
sending any $x$ in the first copy of $\OC(\tg^*)$ to $f_1(x) \otimes 1$, and any $y$ in the second copy of $\OC(\tg^*)$ to $1 \otimes f_2(y)$. One can easily check that the image under this morphism of any element of the form $g \otimes 1 - 1 \otimes g$ with $g \in \OC(\tg^*/W)$ belongs to $\hbar \cdot (\widetilde{C}_\hbar \otimes_{\OC(\tg^*)[\hbar]} \widetilde{C}_{\hbar})$. It follows from Lemma~\ref{lem:deformation-flat}\eqref{it:def-flat-3} that $\widetilde{C}_\hbar \otimes_{\OC(\tg^*)[\hbar]} \widetilde{C}_{\hbar}$ is flat over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$; in particular it has no $\hbar$-torsion. Hence~\eqref{eqn:comultiplication} factors in a unique way through a graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-algebra morphism
\[
\widetilde{C}_\hbar \rightarrow \widetilde{C}_\hbar \otimes_{\OC(\tg^*)[\hbar]} \widetilde{C}_{\hbar},
\]
which provides our comultiplication morphism.
Using this structure, if $M$ and $N$ are graded $\widetilde{C}_\hbar$-modules, then we obtain that the tensor product $M \otimes_{\OC(\tg^*)[\hbar]} N$ has a natural structure of graded $\widetilde{C}_\hbar$-module.
Similar constructions provide a structure of graded $(\OC(\tg^*/W)[\hbar], \OC(\tg^*)[\hbar])$-bimo\-dule on $C_\hbar$, and a graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-algebra morphism
\[
C_\hbar \rightarrow C_\hbar \otimes_{\OC(\tg^*)[\hbar]} \widetilde{C}_{\hbar}.
\]
Hence, if $M$ is a graded $C_\hbar$-module and $N$ is a graded $\widetilde{C}_\hbar$-module, we obtain that the tensor product $M \otimes_{\OC(\tg^*)[\hbar]} N$ has a natural structure of graded $C_\hbar$-module.
Applying the functor ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} (-)$, we also obtain graded algebra morphisms
\[
\widetilde{C} \rightarrow \widetilde{C} \otimes_{\OC(\tg^*)} \widetilde{C}, \qquad C \rightarrow C \otimes_{\OC(\tg^*)} \widetilde{C},
\]
and the corresponding structures for tensor products of graded modules.
\subsection{``Algebraic'' Bott--Samelson category}
\label{ss:algebraic-BS}
The group $W_{\mathrm{aff}}$ acts naturally on $\tg^* \times \AM^1_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}$, via the formulas
\[
v \cdot (\xi,x)=(v \cdot \xi, x), \qquad t_\lambda \cdot (\xi,x) = (\xi + x \lambda, x)
\]
for $\xi \in \tg^*$, $x \in \AM^1_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, $v \in W$ and $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. For this action, the subspace $\tg^* = \tg^* \times \{0\} \subset \tg^* \times \AM^1_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is stable, and the action of $W_{\mathrm{aff}}$ on this subspace factors through the natural action of $W=W_{\mathrm{aff}} / {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
The $W_{\mathrm{aff}}$-action on $\tg^* \times \AM^1_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ induces an action on the graded algebra $\OC(\tg^*)[\hbar] = \OC(\tg^* \times \AM^1_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. If $w \in W_{\mathrm{aff}}$, we denote by $(\OC(\tg^*)[\hbar])^w$ the subalgebra of $w$-invariants.
Below we will need the following easy lemma.
\begin{lem}
\label{lem:invariants-hbar}
For any simple reflection $s$, the morphism
\[
{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} (\OC(\tg^*)[\hbar])^s \rightarrow \OC(\tg^*)^s
\]
induced by the restriction morphism $\OC(\tg^*)[\hbar] \rightarrow \OC(\tg^*)$
is an isomorphism.
\end{lem}
\begin{proof}
If $s$ is finite, then the claim is obvious. The general case follows, using Lemma~\ref{lem:affine-simple-conjugate}.
\end{proof}
For any $w \in W_{\mathrm{aff}}$, we define the graded $\OC(\tg^* \times \tg^*)[\hbar]$-module $E^{\hbar}_w$ as follows. As a graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-module, we have $E^\hbar_w=\OC(\tg^*)[\hbar]$. The \emph{right} copy of $\OC(\tg^*)[\hbar]$ in $\OC(\tg^* \times \tg^*)[\hbar] = \OC(\tg^*)[\hbar] \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} \OC(\tg^*)[\hbar]$ acts in the natural way, by multiplication. And any $f$ in the \emph{left} copy of $\OC(\tg^*)[\hbar]$ acts by multiplication by $w^{-1} \cdot f$.
Since the induced action of $\OC(\tg^* \times \tg^*)$ on $E^\hbar_w / \hbar \cdot E^\hbar_w$ factors through an action of $\OC(\tg^* \times_{\tg^*/W} \tg^*)$, there exists a unique extension of the action of $\OC(\tg^* \times \tg^*)[\hbar]$ to an action of $\widetilde{C}_\hbar$ on $E^\hbar_w$. By restriction, one can also consider $E^\hbar_w$ as a graded $C_\hbar$-module.
If $s$ is a simple reflection, we also consider the graded $\OC(\tg^* \times \tg^*)[\hbar]$-module
\[
D^\hbar_s:= \OC(\tg^*)[\hbar] \otimes_{(\OC(\tg^*)[\hbar])^s} \OC(\tg^*)[\hbar] \langle -1 \rangle.
\]
Using the same arguments as above and Lemma~\ref{lem:invariants-hbar}, one can check that the action of $\OC(\tg^* \times \tg^*)[\hbar]$ on $D^\hbar_s$ extends in a canonical way to an action of $\widetilde{C}_\hbar$, so that $D^\hbar_s$ can be considered as a graded $\widetilde{C}_\hbar$-module.
Now, using the constructions of~\S\ref{ss:deformations} one can define, for any $\omega \in \Omega$ and
any sequence $\underline{s}=(s_1, \cdots, s_n)$ of simple reflections, the graded $C_\hbar$-module
\[
D_{\hbar}(\omega, \underline{s}) := E^\hbar_\omega \otimes_{\OC(\tg^*)[\hbar]} D^\hbar_{s_1} \otimes_{\OC(\tg^*)[\hbar]} \cdots \otimes_{\OC(\tg^*)[\hbar]} D^\hbar_{s_n}.
\]
We will also consider the corresponding constructions for $C$-modules: we set
\[
E_w:=E^\hbar_w / \hbar \cdot E^\hbar_w, \qquad D_s := D_s^\hbar / \hbar \cdot D_s^\hbar,
\]
and then
\[
D(\omega, \underline{s}) := D_{\hbar}(\omega, \underline{s}) / \hbar \cdot D_{\hbar}(\omega, \underline{s}) \cong E_\omega \otimes_{\OC(\tg^*)} D_{s_1} \otimes_{\OC(\tg^*)} \cdots \otimes_{\OC(\tg^*)} D_{s_n}.
\]
Note for later use that, by Lemma~\ref{lem:invariants-hbar}, we have a canonical isomorphism
\begin{equation}
\label{eqn:Ds-tensor-product}
D_s \xrightarrow{\sim} \OC(\tg^*) \otimes_{\OC(\tg^*)^s} \OC(\tg^*).
\end{equation}
With these definitions one can define the category $\mathsf{BS}^{\mathrm{alg}}$ with
\begin{itemize}
\item
\emph{objects}: triples $(\omega, \underline{s}, i)$ with $\omega \in \Omega$, $\underline{s}$ a sequence of simple reflections indexed by $(1, \cdots, n)$ for some $n \in \ZM$, and $i \in \ZM$;
\item
\emph{morphisms}:
for $\omega, \omega' \in \Omega$, $\underline{s}$ and $\underline{t}$ sequences of simple reflections, and $i,j \in \ZM$,
\begin{multline*}
\Hom_{\mathsf{BS}^{\mathrm{alg}}} \bigl( (\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr) =
\\
\begin{cases}
\Hom_{\Modgr(C)} \bigl( D(\omega, \underline{s}) \langle -i \rangle, D(\omega, \underline{t}) \langle -j \rangle \bigr) & \text{if $\omega=\omega'$;} \\
0 & \text{if $\omega \neq \omega'$.}
\end{cases}
\end{multline*}
\end{itemize}
\subsection{Equivariant cohomology of ${\EuScript Gr}$}
\label{ss:cohomology-Gr}
The graded algebras $C_\hbar$ and $C$ defined in~\S\ref{ss:deformations} can be used to describe the algebras $\HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and $\HM^\bullet_{\check{I}} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ ``up to torsion,'' as follows.
There exists a natural graded algebra morphism
\[
\OC(\tg^*)[\hbar] \xrightarrow[\sim]{\eqref{eqn:cohomology-TD}} \HM^\bullet_{\check{I} \rtimes \Gm}(\mathrm{pt};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
On the other hand, we have a canonical isomorphism
\[
\HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} \bigl( (\check{I} \rtimes \Gm) \backslash (\check{G}(\mathscr{K}) \rtimes \Gm);{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigr),
\]
so that there also exists a natural graded algebra morphism
\[
\OC(\tg^*/W)[\hbar] \xrightarrow[\sim]{\eqref{eqn:cohomology-GD}} \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm}(\mathrm{pt};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
induced by the multiplication of $\check{G}(\mathscr{O}) \rtimes \Gm$ on $\check{G}(\mathscr{K}) \rtimes \Gm$ on the \emph{right}.
Using the fact that the projection $\check{G}(\mathscr{K}) \rtimes \Gm \rightarrow \mathrm{pt}$ factors through $\Gm$, it is not difficult to check that the images of $\hbar$ under these two morphisms coincide (see the proof of Lemma~\ref{lem:morphism-cohomology-factorization} below for similar considerations). Hence
combining them we obtain an algebra morphism
\begin{equation}
\label{eqn:morp-cohomology-Gr}
\OC(\tg^*/W \times \tg^*)[\hbar] \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\end{equation}
\begin{remark}
Note that we have switched the order of the factors here: the left-hand factor of $\tg^*/W \times \tg^*$ is related to the multiplication of $\check{G}(\mathscr{O})$ on $\check{G}(\mathscr{K})$ on the right, while the right-hand factor is related to the multiplication of $\check{I}$ on $\check{G}(\mathscr{K})$ on the left. This choice of convention complicates some formulas in this section, but it will make the comparison with the constructions on the ``coherent'' side easier. Another option would have been to work with the (less customary) variety ${\EuScript Gr}':=\check{G}(\mathscr{O}) \backslash \check{G}(\mathscr{K})$ instead of ${\EuScript Gr}$. Here ${\EuScript Gr}'$ is isomorphic to ${\EuScript Gr}$ through $\check{G}(\mathscr{O}) g \mapsto g^{-1} \check{G}(\mathscr{O})$, but this isomorphism switches the role of left and right multiplications.
\end{remark}
\begin{lem}
\label{lem:morphism-cohomology-factorization}
Morphism~\eqref{eqn:morp-cohomology-Gr} factors in a unique way through a graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-algebra morphism
\[
\gamma_\hbar \colon C_\hbar \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
\end{lem}
\begin{proof}
There exist natural converging spectral sequences
\begin{align*}
E_2^{pq} = \HM^{p}_{\check{G}(\mathscr{O}) \rtimes \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \HM^{q}({\EuScript Gr}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) &\Rightarrow \HM^{p+q}_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}), \\
E_2^{pq} = \HM^{p}_{\check{I} \rtimes \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \HM^{q}({\EuScript Gr}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) &\Rightarrow \HM^{p+q}_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\end{align*}
In both cases, the spectral sequence degenerates since the left-hand side vanishes unless $p$ and $q$ are even. This implies in particular that $\HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-free, proving the unicity of the factorization. It also follows that the natural morphism
\[
\HM^\bullet_{\check{I} \rtimes \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \otimes_{\HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
is an isomorphism. Using this, we see that to prove the existence of the factorization it suffices to prove that the natural algebra morphism
\begin{equation}
\label{eqn:morp-cohomology-Gr-2}
\OC(\tg^* / W \times \tg^*/W)[\hbar] \rightarrow \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\end{equation}
defined in a way similar to~\eqref{eqn:morp-cohomology-Gr} factors through $C'_\hbar$.
The latter property can be proved as follows.\footnote{A similar claim is asserted without details in~\cite{bf}. We thank V.~Ginzburg for explaining this proof to one of us.} Using the same spectral sequence argument as above, one can check that $\HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]$-free, and that the natural morphism
\[
{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{\check{G}(\mathscr{O})} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
is an isomorphism. From these facts we see that it suffices to prove that the morphism
\[
\OC(\tg^*/W \times \tg^*/W) \rightarrow \HM^\bullet_{\check{G}(\mathscr{O})} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
defined as for~\eqref{eqn:morp-cohomology-Gr-2} (but with the $\Gm$-equivariance omitted) factors through $\OC(\Delta)$.
Now we make the following observation. Let $H$ be a topological group, acting on a topological space $X$, and let $Y:=H \times X$. We endow $Y$ with an action of $H \times H$ via $(h_1,h_2) \cdot (k,x) = (h_1 k h_2^{-1}, h_2 \cdot x)$. Then there exists a natural morphism
\begin{equation}
\label{eqn:morphism-equiv-coh}
\HM^\bullet_{H \times H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{H \times H}(Y; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\end{equation}
We claim that~\eqref{eqn:morphism-equiv-coh} factors through the morphism $\HM^\bullet_{H \times H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ induced by restriction to the diagonal copy of $H$. Indeed one can consider the composition
\begin{equation}
\label{eqn:morphism-equiv-coh-2}
\HM^\bullet_{H \times H}(Y; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{H}(Y; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{H}(X; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}),
\end{equation}
where the first morphism is induced by restriction to the diagonal copy, and the second morphism by restriction to the $H$-stable subspace $X= \{1\} \times X \subset Y$. Since $Y$ identifies with the induced variety $(H \times H) \times^H X$ (via the morphism $[(h_1,h_2) : x] \mapsto (h_1 h_2^{-1}, h_2 \cdot x)$), \eqref{eqn:morphism-equiv-coh-2} is an isomorphism. Since the composition of~\eqref{eqn:morphism-equiv-coh} and~\eqref{eqn:morphism-equiv-coh-2} clearly factors through $\HM^\bullet_H(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$, the same holds for~\eqref{eqn:morphism-equiv-coh}.
We take for $H$ a maximal compact subgroup of $\check{G}$, so that we have isomorphisms $\HM^\bullet_{\check{G}(\mathscr{O})}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{\check{G}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. Similarly, using the K{\"u}nneth formula (which is applicable here since our cohomology spaces are free over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$) we obtain isomorphisms
\[
\OC(\tg^*/W \times \tg^*/W) \cong \HM^\bullet_{\check{G}(\mathscr{O}) \times \check{G}(\mathscr{O})}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{H \times H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}),
\]
and one can identify the morphism $\OC(\tg^*/W \times \tg^*/W) \rightarrow \OC(\Delta)$ with the morphism $\HM^\bullet_{H \times H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{H}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ considered above.
If $\Omega H$ denotes the group of polynomial loops from the unit circle to $H$, then as in~\cite[\S 1.2]{ginzburg-2} we have a natural homeomorphism $\Omega H / H \xrightarrow{\sim} {\EuScript Gr}$. Writing $\Omega H = H \times \Omega^0 H$ (where $\Omega^0 H$ is the space of \emph{based} loops, i.e.~those sending the base point of the circle to the identity) we obtain isomorphisms
\[
\HM^\bullet_{\check{G}(\mathscr{O})} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_H(\Omega H / H; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{H \times H}(H \times \Omega^0 H; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
Hence we are in the setting considered above, with $X=\Omega^0 H$, and the desired claim follows from our general observation.
\end{proof}
Using a spectral sequence argument as in the proof of Lemma~\ref{lem:morphism-cohomology-factorization}, one can check that
the natural morphism
\[
\HM^\bullet_{\check{I}}(\mathrm{pt};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \otimes_{\HM^\bullet_{\check{I} \rtimes \Gm} (\mathrm{pt};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^\bullet_{\check{I}} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
is an isomorphism.
We denote by
\[
\gamma \colon C \rightarrow \HM^\bullet_{\check{I}} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
the composition of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} \gamma_\hbar$ with this isomorphism. Then $\gamma$ is a graded algebra morphism.
Since ${\EuScript Gr}$ is the disjoint union of its connected components ${\EuScript Gr}_{(\omega)}$ ($\omega \in \Omega$) which are $\check{I} \rtimes \Gm$-stable, there exist natural isomorphisms of graded algebras
\[
\HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \prod_{\omega \in \Omega} \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}), \qquad \HM^\bullet_{\check{I}} ({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \prod_{\omega \in \Omega} \HM^\bullet_{\check{I}} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
For any $\omega \in \Omega$, we denote by
\[
\gamma^{(\omega)}_\hbar \colon C_\hbar \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}), \qquad \text{resp.} \qquad \gamma^{(\omega)} \colon C \rightarrow \HM^\bullet_{\check{I}} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
the composition of $\gamma_\hbar$, resp.~$\gamma$, with the projection on the factor parametrized by $\omega$.
\begin{prop}
\label{prop:cohomology-Gr}
For all $\omega \in \Omega$, the morphisms $\QM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \gamma^{(\omega)}_\hbar$ and $\QM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \gamma^{(\omega)}$ are isomorphisms.
\end{prop}
\begin{proof}
It is sufficient to prove the claim for $\gamma^{(\omega)}_\hbar$. Then, by construction of this morphism, it is sufficient to prove that the similar morphism
\[
C'_\hbar \rightarrow \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
becomes an isomorphism after applying $\QM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (-)$. However,
since $\QM$ is flat over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, by Lemma~\ref{lem:def-normal-cone-BC} we have a natural isomorphism
\[
\QM \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} C'_\hbar \cong \mathrm{DNC}_{I_1^{\QM}}(\OC( \tg^*_\QM / W \times \tg^*_\QM / W )),
\]
where $\tg_\QM^*:=\QM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \tg^*$ and $I_1^{\QM}$ is the ideal of the diagonal copy of $\tg_\QM^*/W$. Similarly, we have
\[
\QM \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr}_{(\omega)};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm} ({\EuScript Gr}_{(\omega)};\QM).
\]
Hence our claim follows from~\cite[Theorem 1]{bf}.
\end{proof}
\begin{remark}
Unlike for the case of $\QM$, the morphisms $\gamma^{(\omega)}_\hbar$ and $\gamma^{(\omega)}$ are \emph{not} isomorphisms. In fact, $C_{\hbar}$ is a finitely generated ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-algebra, whereas $\HM^\bullet_{\check{I} \rtimes \Gm}({\EuScript Gr}_{(\omega)}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is not finitely generated in general, see~\cite{yz}.
\end{remark}
\subsection{``Topological'' Bott--Samelson category}
\label{ss:topological-BS}
Let $\EM$ be either ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ or $\FM$. Recall the standard convolution product on the category $D^{\mathrm{b}}_{\check{I} \rtimes \Gm}(\FC l, \EM)$, defined by
\[
\FC \star \GC := \mu_*(\FC \, \widetilde{\boxtimes} \, \GC),
\]
where $\FC l \, \widetilde{\times} \, \FC l$ is the quotient of $(\check{G}(\mathscr{K}) \rtimes \Gm) \times \FC l$ by the natural diagonal action of $\check{I} \rtimes \Gm$, $\mu \colon \FC l \, \widetilde{\times} \, \FC l \rightarrow \FC l$ is defined by $\mu([g : h \check{I}]) = gh \check{I}$, and $\FC \, \widetilde{\boxtimes} \, \GC$ is the ``twisted external product'' of $\FC$ and $\GC$, i.e.~the unique object whose pullback to $(\check{G}(\mathscr{K}) \rtimes \Gm) \times \FC l$ is the external product of the pullback of $\FC$ with $\GC$.
A similar construction provides a bifunctor
\[
(-) \star (-) \colon D^{\mathrm{b}}_{\check{I} \rtimes \Gm}(\FC l, \EM) \times D^{\mathrm{b}}_{\check{I} \rtimes \Gm}({\EuScript Gr}, \EM) \rightarrow D^{\mathrm{b}}_{\check{I} \rtimes \Gm}({\EuScript Gr}, \EM).
\]
We now introduce some ``Bott--Samel\-son objects'' in $\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr},\EM)$, as follows. Each connected component ${\EuScript Gr}_{(\omega)}$ contains a unique $0$-dimensional $\check{I}$-orbit; we denote by $\delta_\omega^{\EM}$ the constant (skyscraper) sheaf on this orbit (with coefficients $\EM$). On the other hand, for any simple reflection $s$, we have the $\check{I} \rtimes \Gm$-equivariant parity complex $\EC_{s,\EM}:=\underline{\EM}_{\overline{\FC l_s}}[1]$ on $\FC l$. Then, if $\omega \in \Omega$ and if $\underline{s}=(s_1, \cdots, s_r)$ is a sequence of simple reflections, we can consider the object
\[
\EC_\EM(\omega, \underline{s}) := \EC_{s_r,\EM} \star \cdots \star \EC_{s_1,\EM} \star \delta^\EM_{\omega^{-1}}
\]
in $D^{\mathrm{b}}_{\check{I} \rtimes \Gm}({\EuScript Gr},\EM)$. The arguments in~\cite[\S 4.1]{jmw} or in~\cite[\S 5.5]{FieWil} show that $\EC_\EM(\omega, \underline{s})$ belongs to the subcategory $\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr},\EM)$.
We will denote similarly the images of this object in $\Parity_{\check{I}}({\EuScript Gr},\EM)$ and in $\Parity_{(\check{I})}({\EuScript Gr},\EM)$.
We define the category $\mathsf{BS}^{\mathrm{top}}$ with
\begin{itemize}
\item \emph{objects}: triples $(\omega, \underline{s}, i)$ with $\omega \in \Omega$, $\underline{s}$ a sequence of simple reflections indexed by $(1, \cdots, n)$ for some $n \in \ZM$, and $i \in \ZM$;
\item \emph{morphisms}:
\[
\Hom_{\mathsf{BS}^{\mathrm{top}}} \bigl( (\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr) :=
\Hom_{\Parity_{\check{I}}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} (\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s}) [i], \EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega', \underline{t})[j]).
\]
\end{itemize}
\begin{prop}
\label{prop:parity-Karoubi}
The category $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ can be recovered from the category $\mathsf{BS}^{\mathrm{top}}$, in the sense that it is equivalent to the Karoubian closure of the additive envelope of the category
which has the same objects as $\mathsf{BS}^{\mathrm{top}}$, and morphisms from $(\omega, \underline{s}, i)$ to $(\omega', \underline{t}, j)$ which are given by the $(j-i)$-th piece of the graded vector space
\[
\FM \otimes_{\HM^\bullet_{\check{I}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \left( \bigoplus_{n \in \ZM} \Hom_{\mathsf{BS}^{\mathrm{top}}} \bigl( (\omega, \underline{s}, 0), (\omega', \underline{t}, n) \bigr) \right).
\]
\end{prop}
\begin{proof}
Denote (for the duration of the proof) by $\mathsf{A}$ the category which has the same objects as $\mathsf{BS}^{\mathrm{top}}$, and whose morphisms are defined as in the statement of the proposition. We observe that, if $\FM(-)$ is the ``modular reduction functor'' defined as in~\S\ref{ss:reminder-parity}, then we have we have a canonical isomorphism
\[
\FM \bigl( \EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s}) \bigr) \cong \EC_\FM(\omega, \underline{s})
\]
for any $(\omega, \underline{s})$ as above. (In fact, this follows from the commutation of the functor $\FM(-)$ with $*$-pullback, see~\cite[Proposition~2.6.5]{ks}, and with $!$-pushforward, see~\cite[Proposition~2.6.6]{ks}.) Using these isomorphisms and
Lemma~\ref{lem:properties-parity}\eqref{it:morph-parity-For},
we see that the assignment $(\omega, \underline{s}, i) \mapsto \EC_\FM(\omega,\underline{s})[i]$ defines an equivalence of categories from $\mathsf{A}$ to the full subcategory $\mathsf{A}'$ of $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ whose objects are of the form $\EC_\FM(\omega,\underline{s})[i]$.
Now we observe that, by the results of~\cite{jmw}, the category $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ is a Krull--Schmidt, Karoubian, additive category, and moreover that any indecomposable object in this category is isomorphic to a direct summand of an object of the form $\EC_\FM(\omega,\underline{s})[i]$. It follows that $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$ is equivalent to the Karoubian closure of the additive envelope of $\mathsf{A}'$, which finishes the proof.
\end{proof}
\begin{remark}
\label{rk:indec-parity}
More precisely, the results of~\cite[\S 4.1]{jmw}
imply that if $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and if $w_{-\lambda} = \omega s_1 \cdots s_r$ is a reduced expression for $w_\lambda$, then $\EC_\lambda$ can be characterized (up to isomorphism) as the unique direct summand of $\EC_\FM(\omega, (s_1, \cdots, s_r))$ which is not a direct summand of any object of the form $\EC_\FM(\omega', \underline{t})[i]$ where $\omega' \in \Omega$, $i \in \ZM$, and $\underline{t}$ is a sequence of simple reflections of length at most $r-1$.
\end{remark}
\subsection{Equivariant cohomology functors}
\label{ss:cohomology-functors}
For any $\omega \in \Omega$, we define the functor
\[
\mathsf{H}_{\check{I} \rtimes \Gm} \colon \Parity_{\check{I} \rtimes \Gm}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \Modgr(C_\hbar)
\]
as the composition
\begin{equation}
\label{eqn:functor-H}
\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\HM^\bullet_{\check{I} \rtimes \Gm}({\EuScript Gr},-)}
\Modgr(\HM^\bullet_{\check{I} \rtimes \Gm}({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})) \rightarrow \Modgr(C_\hbar),
\end{equation}
where the second functor is the ``restriction of scalars'' functor associated with the morphism $\gamma_\hbar$. We define similarly a functor
\[
\mathsf{H}_{\check{I}} \colon \Parity_{\check{I}}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \Modgr(C).
\]
The goal of this subsection is to prove the following.
\begin{prop}
\label{prop:H-BS}
For any $\omega \in \Omega$ and any sequence $\underline{s}=(s_1, \cdots, s_n)$ of simple reflections, there exists a canonical isomorphism of graded $C$-modules
\[
\mathsf{H}_{\check{I}} (\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s})) \cong D(\omega,\underline{s}).
\]
\end{prop}
Before proving the proposition, we remark that the same constructions as in~\S\ref{ss:cohomology-Gr} allow to define graded algebra morphisms
\[
\widetilde{C}_\hbar \rightarrow \HM^\bullet_{\check{I} \rtimes \Gm}(\FC l; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \qquad \text{and} \qquad \widetilde{C} \rightarrow \HM^\bullet_{\check{I}}(\FC l; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}),
\]
and then functors
\[
\widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm} \colon \Parity_{\check{I} \rtimes \Gm}(\FC l, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \Modgr(\widetilde{C}_\hbar), \qquad \widetilde{\mathsf{H}}_{\check{I}} \colon \Parity_{\check{I}}(\FC l, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \Modgr(\widetilde{C}).
\]
\begin{lem}
\label{lem:cohomology-basic}
\begin{enumerate}
\item
\label{it:cohomology-delta}
For any $\omega \in \Omega$, there exists a canonical isomorphism of graded $C_\hbar$-modules
\[
\mathsf{H}_{\check{I} \rtimes \Gm}(\delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong E^\hbar_{\omega^{-1}}.
\]
\item
\label{it:cohomology-Es}
For any simple reflection $s$, there exists a canonical isomorphism of $\widetilde{C}_\hbar$-modules
\[
\widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm}(\EC_{s,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong D^\hbar_s.
\]
\end{enumerate}
\end{lem}
\begin{proof}
\eqref{it:cohomology-delta}
By definition, if $\omega = t_\lambda v$ (with $v \in W$ and $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$) we have
\[
\mathsf{H}_{\check{I} \rtimes \Gm}(\delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) = \HM^\bullet_{\check{I} \rtimes \Gm}({\EuScript Gr}, \delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) = \HM^\bullet_{\check{I} \rtimes \Gm} \bigl( z^\lambda \cdot (\check{G}(\mathscr{O}) \rtimes \Gm) / \check{G}(\mathscr{O}) \rtimes \Gm; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigr).
\]
We deduce a canonical isomorphism
\[
\mathsf{H}_{\check{I} \rtimes \Gm}(\delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{(\check{I} \rtimes \Gm) \times (\check{G}(\mathscr{O}) \rtimes \Gm)} \bigl( z^\lambda \cdot (\check{G}(\mathscr{O}) \rtimes \Gm) ; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigr).
\]
Now if $\dot{v}$ is a lift of $v$ in $\check{G}$, the assignment $(a,b) \mapsto a \cdot z^\lambda \dot{v} \cdot b^{-1}$ induces an isomorphism
\[
(\check{I} \rtimes \Gm) \times (\check{G}(\mathscr{O}) \rtimes \Gm) / K \xrightarrow{\sim} z^\lambda \cdot (\check{G}(\mathscr{O}) \rtimes \Gm),
\]
where $K=\{(a,b) \in (\check{I} \rtimes \Gm) \times (\check{G}(\mathscr{O}) \rtimes \Gm) \mid b=(z^\lambda \dot{v})^{-1} a (z^\lambda \dot{v})\}$. This group is isomorphic to $(\check{I} \rtimes \Gm)$ (through $a \mapsto \bigl(a, (z^\lambda \dot{v})^{-1} a (z^\lambda \dot{v}) \bigr)$); hence we obtain canonical isomorphisms
\[
\mathsf{H}_{\check{I} \rtimes \Gm}(\delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_K(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \HM^\bullet_{\check{I} \rtimes \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\]
The right-hand side is isomorphic to $\OC(\tg^*)[\hbar]$ (see~\eqref{eqn:cohomology-TD}), with the natural action of the subalgebra $\OC(\tg^*)[\hbar] \subset C_\hbar$. And, via this isomorphism, the subalgebra $\OC(\tg^*/W)[\hbar] \subset C_\hbar$ acts via the composition of multiplication and the morphism
\[
\OC(\tg^*/W)[\hbar] \hookrightarrow \OC(\tg^*)[\hbar] \xrightarrow{f \mapsto \omega \cdot f} \OC(\tg^*)[\hbar].
\]
Hence we have constructed a canonical isomorphism of $\OC(\tg^*/W \times \tg^*)[\hbar]$-modules
\[
\mathsf{H}_{\check{I} \rtimes \Gm}(\delta_\omega^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong E^\hbar_{\omega^{-1}}.
\]
Since these spaces have no $\hbar$-torsion, this isomorphism is automatically an isomorphism of $C_\hbar$-modules.
\eqref{it:cohomology-Es}
Using Lemma~\ref{lem:affine-simple-conjugate}, one can assume that $s$ is finite, with associated simple root $\alpha^\vee \in {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$ and associated simple coroot $\alpha \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
By definition we have
\[
\widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm}(\EC_{s,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong \HM^{\bullet}_{\check{T} \times \Gm}(\overline{\FC l_s}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \langle -1 \rangle.
\]
Since $\overline{\FC l_s}$ is isomorphic to $\PM^1_\CM$,
it is well known (see e.g.~\cite{FieWil}) that the morphism
\[
\HM^{\bullet}_{\check{T} \times \Gm}(\overline{\FC l_s}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^{\bullet}_{\check{T} \times \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \oplus \HM^{\bullet}_{\check{T} \times \Gm}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
induced by restriction to the fixed points $\check{I} / \check{I}$ and $s \cdot \check{I} / \check{I}$ induces an isomorphism between $\HM^{\bullet}_{\check{T} \times \Gm}(\overline{\FC l_s}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and
\[
{}' \hspace{-1pt} D^{\hbar}_s:=\{(a,b) \in \OC(\tg^*)[\hbar] \oplus \OC(\tg^*)[\hbar] \mid a=b \ \mathrm{mod} \ \alpha^\vee\}.
\]
(Here, the right copy of $\OC(\tg^*)[\hbar]$ in $\widetilde{C}_\hbar$ acts diagonally, while any $f$ in the left copy of $\OC(\tg^*)[\hbar]$ in $\widetilde{C}_\hbar$ acts by multiplication by $(f,s(f))$.)
Now we consider the morphism $D^\hbar_s \rightarrow {}' \hspace{-1pt} D^\hbar_s$ defined by $f \otimes g \mapsto (fg, s(f)g)$. By our assumptions on ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, there exists ${\check \mu} \in {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}} \otimes_\ZM {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ such that $\langle {\check \mu}, \alpha \rangle = 1$.
Then our morphism sends the basis of $D^\hbar_s$ as an $\OC(\tg^*)[\hbar]$-module (via the action of the right copy of $\OC(\tg^*)[\hbar]$ in $\widetilde{C}_\hbar$) consisting of $1 \otimes 1$ and ${\check \mu} \otimes 1$ to the basis of ${}' \hspace{-1pt} D^{\hbar}_s$ consisting of $(1,1)$ and $({\check \mu}, {\check \mu}-\alpha^\vee)$; hence it is an isomorphism.
We have constructed an isomorphism of $\OC(\tg^* \times \tg^*)[\hbar]$-modules
\[
\widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm}(\EC_{s,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong D_s^\hbar.
\]
Since both sides have no $\hbar$-torsion, this isomorphism is automatically an isomorphism of $\widetilde{C}_\hbar$-modules.
\end{proof}
\begin{proof}[Proof of Proposition~{\rm \ref{prop:H-BS}}]
The arguments in~\cite[\S 4.1]{jmw} imply that for any $\FC$ in $\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and any sequence $\underline{t}=(t_1, \cdots, t_r)$ of simple reflections, the convolution
\[
\EC_{t_1, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \star \cdots \star \EC_{t_r, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \star \FC
\]
belongs to $\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. Then the same arguments as in the proof of~\cite[Proposition~3.2.1]{by} (using Lemma~\ref{lem:properties-parity}\eqref{it:parity-cohomology} instead of~\cite[Corollary~B.4.2]{by}) imply that there exists a canonical isomorphism of graded modules over $\HM^\bullet_{\check{I} \rtimes \Gm}({\mathrm{pt}}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}[\hbar]} \HM^\bullet_{\check{G}(\mathscr{O}) \rtimes \Gm}({\mathrm{pt}}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$:
\begin{multline*}
\mathsf{H}_{\check{I} \rtimes \Gm}(\EC_{t_1, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \star \cdots \star \EC_{t_r, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \star \FC) \cong \\
\widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm}(\EC_{t_1,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \otimes_{\HM^\bullet_{\check{I} \rtimes \Gm}({\mathrm{pt}}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \cdots \otimes_{\HM^\bullet_{\check{I} \rtimes \Gm}({\mathrm{pt}}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \widetilde{\mathsf{H}}_{\check{I} \rtimes \Gm}(\EC_{t_r,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \otimes_{\HM^\bullet_{\check{I} \rtimes \Gm}({\mathrm{pt}}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \mathsf{H}_{\check{I} \rtimes \Gm}(\FC).
\end{multline*}
Applying this remark to $\FC=\delta_{\omega^{-1}}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and $\underline{t}=(s_n, \cdots, s_1)$, and using Lemma~\ref{lem:cohomology-basic}, we obtain a canonical isomorphism of $\OC(\tg^*/W \times \tg^*)[\hbar]$-modules
\[
\mathsf{H}_{\check{I} \rtimes \Gm} (\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s})) \cong D_\hbar(\omega,\underline{s}).
\]
Since both sides have no $\hbar$-torsion, this isomorphism is automatically an isomorphism of $C_\hbar$-modules. Specializing $\hbar$ to $0$, we deduce the isomorphism of the proposition.
\end{proof}
\subsection{Equivalence}
\label{ss:equivalence}
The proof of the following proposition (which is independent of the rest of the section) is postponed to \S\ref{ss:proof-H-ff}.
\begin{prop}
\label{prop:global-coh-ff}
For any $\omega, \omega' \in \Omega$, any sequences $\underline{s},\underline{t}$ of simple reflections, and any $n \in \ZM$, the morphism
\begin{multline*}
\Hom_{\Parity_{\check{I}}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s}), \EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})[n]) \\
\rightarrow \Hom_{\Modgr ( \HM^\bullet_{\check{I}}({\EuScript Gr};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}))}(\HM^\bullet_{\check{I}}({\EuScript Gr}, \EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s})), \HM^\bullet_{\check{I}}({\EuScript Gr}, \EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})[n]))
\end{multline*}
induced by the functor $\HM^\bullet_{\check{I}}({\EuScript Gr},-)$ is an isomorphism.
\end{prop}
Now we define a functor
\[
\mathsf{H}_{\mathsf{BS}} \colon \mathsf{BS}^{\mathrm{top}} \rightarrow \mathsf{BS}^{\mathrm{alg}}
\]
as follows. This functor sends an object $(\omega, \underline{s}, i)$ of $\mathsf{BS}^{\mathrm{top}}$ to the corresponding object $(\omega, \underline{s}, i)$ of $\mathsf{BS}^{\mathrm{alg}}$. Then if $(\omega, \underline{s}, i)$ and $(\omega', \underline{t}, j)$ are objects of $\mathsf{BS}^{\mathrm{top}}$, the morphism
\[
\Hom_{\mathsf{BS}^{\mathrm{top}}} \bigl((\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr) \rightarrow \Hom_{\mathsf{BS}^{\mathrm{alg}}} \bigl((\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr)
\]
is defined as the trivial morphism if $\omega \neq \omega'$ (in which case both $\Hom$-spaces are $0$), and as the morphism induced by $\mathsf{H}_{\check{I}}$ if $\omega=\omega'$, using the canonical isomorphisms
\[
\mathsf{H}_{\check{I}}(\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s})[i]) \cong D(\omega,\underline{s})\langle -i \rangle, \qquad \mathsf{H}_{\check{I}}(\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})[j]) \cong D(\omega',\underline{t})\langle -j \rangle
\]
deduced from Proposition~\ref{prop:H-BS}.
The main result of this section is the following.
\begin{thm}
\label{thm:main-top}
The functor $\mathsf{H}_{\mathsf{BS}}$ is an equivalence of categories.
\end{thm}
\begin{proof}
The functor $\mathsf{H}_{\mathsf{BS}}$ clearly induces a bijection on objects. So, what we have to prove is that if $(\omega, \underline{s}, i)$ and $(\omega', \underline{t}, j)$ are objects of $\mathsf{BS}^{\mathrm{top}}$, then the corresponding morphism
\[
\Hom_{\mathsf{BS}^{\mathrm{top}}} \bigl((\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr) \rightarrow \Hom_{\mathsf{BS}^{\mathrm{alg}}} \bigl((\omega, \underline{s}, i), (\omega', \underline{t}, j) \bigr)
\]
is an isomorphism. This is obvious if $\omega \neq \omega'$.
Now, assume that $\omega=\omega'$. Then both parity complexes are supported on ${\EuScript Gr}_{(\omega)}$, so that our morphism is induced by the composition
\[
\Parity_{\check{I}}({\EuScript Gr}_{(\omega)}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\HM^\bullet_{\check{I}}({\EuScript Gr}_{(\omega)},-)} \Modgr \bigl( \HM^\bullet_{\check{I}}({\EuScript Gr}_{(\omega)}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \bigr) \rightarrow \Modgr(C),
\]
where the second arrow is the ``restriction of scalars'' functor associated with $\gamma^{(\omega)}$. Now the first functor is fully faithful on objects of the form $\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s})$ by Proposition~\ref{prop:global-coh-ff}, and the second functor is fully faithful on objects which are ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free by Lemma~\ref{lem:key} and Proposition~\ref{prop:cohomology-Gr}. Since the ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules $\HM^\bullet_{\check{I}}({\EuScript Gr}_{(\omega)},\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s}))$ are free by Lemma~\ref{lem:properties-parity}\eqref{it:parity-cohomology}, this finishes the proof of the theorem.
\end{proof}
\subsection{Proof of Proposition~\ref{prop:global-coh-ff}}
\label{ss:proof-H-ff}
The proof below is a simple variant of the main result of~\cite{ginzburg}. The key observation is from~\cite[Theorem~4.1]{arider}, where it is shown that the arguments from~\cite{ginzburg} involving weights can be replaced in a parity sheaf setting by parity arguments. Another exposition of Ginzburg's proof in an equivariant setting appears as~\cite[Lemma~3.3.1]{by} (for coefficients in characteristic zero).
To avoid unnecessary notational complications, in this proof we will say that an object $\EC$ in $\Parity_{\check{I}}({\EuScript Gr}, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is a \emph{Bott--Samelson parity complex} if it is isomorphic to $\EC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega, \underline{s})[i]$ for some $\omega \in \Omega$, $\underline{s}$ a sequence of simple reflections, and $i \in \ZM$.
Let us fix an extension of the partial order on ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ corresponding to the closure relations among the orbits ${\EuScript Gr}_\lambda$ to a total ordering $\sqsubseteq$ on ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, such that $({\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}, \sqsubseteq)$ is isomorphic (as an ordered set) to $\ZM_{\geq 0}$ with its standard order. Let ${\EuScript Gr}_{\sqsubseteq \lambda}$ denote the union of all ${\EuScript Gr}_\mu$ where $\mu \sqsubseteq \lambda$, and $i_{\sqsubseteq \lambda}:{\EuScript Gr}_{\sqsubseteq \lambda} \rightarrow {\EuScript Gr}$ the closed embedding. Let also $A_{\sqsubseteq \lambda} := \HM_{\check{I}}^\bullet({\EuScript Gr}_{\sqsubseteq \lambda};{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. We define analogously ${\EuScript Gr}_{\sqsubset \lambda}$, $i_{\sqsubset \lambda}$, and $A_{\sqsubset \lambda}$.
We begin with a number of preliminary lemmas.
\begin{lem}
\label{lem:cohom-lambda}
For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, there exists a canonical exact sequence of $\ZM$-graded ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules
\[
0 \rightarrow \HM^\bullet_{c,\check{I}}({\EuScript Gr}_\lambda; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow A_{\sqsubseteq \lambda} \rightarrow A_{\sqsubset \lambda} \rightarrow 0.
\]
\end{lem}
\begin{proof}
The lemma can be proved by induction, using the fact that the existence of the exact sequences for smaller $\lambda$'s implies that $A_{\sqsubset \lambda}$ is concentrated in even degrees, and the adjunction triangle
\[
(i_{\lambda})_! (i_\lambda)^* \underline{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_{{\EuScript Gr}_{\sqsubseteq \lambda}} \rightarrow \underline{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_{{\EuScript Gr}_{\sqsubseteq \lambda}} \rightarrow i_{\sqsubset \lambda *} i_{\sqsubset \lambda}^* \underline{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_{{\EuScript Gr}_{\sqsubseteq \lambda}} \xrightarrow{[1]},
\]
where $i_\lambda \colon {\EuScript Gr}_\lambda \hookrightarrow {\EuScript Gr}$ is the inclusion.
\end{proof}
Let again $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. To simplify notation, we set $Z= {\EuScript Gr}_{\sqsubset \lambda}$, $X = {\EuScript Gr}_{\sqsubseteq \lambda}$, $U= {\EuScript Gr}_\lambda$, and denote by $i:Z\rightarrow X$, resp.~$j:U \rightarrow X$, the closed, resp.~open, embedding.
A key ingredient we will need is the following lemma.
\begin{lem}
\label{lem:exact-sequences-adj}
Let $\EC$ be a Bott--Samelson parity complex, and let $\FC$ be either $i^*_{\sqsubseteq \lambda} \EC$ or $i^!_{\sqsubseteq \lambda} \EC$. Then the adjunction triangles induce short exact sequence of $A_{\sqsubseteq \lambda}$-modules
\begin{gather}
0 \rightarrow \HM_{\check{I}}^\bullet(X, j_!j^* \FC) \rightarrow \HM_{\check{I}}^\bullet(X,\FC) \rightarrow \HM_{\check{I}}^\bullet(X, i_*i^*\FC) \rightarrow 0, \label{eqn:adjses1}\\
0 \rightarrow \HM_{\check{I}}^\bullet(X, i_*i^! \FC) \rightarrow \HM_{\check{I}}^\bullet(X,\FC) \rightarrow \HM_{\check{I}}^\bullet(X, j_*j^*\FC) \rightarrow 0 \label{eqn:adjses2}.
\end{gather}
\end{lem}
\begin{proof}
For $\FC = i^*_{\sqsubseteq \lambda} \EC$, resp.~$i^!_{\sqsubseteq \lambda} \EC$, $\FC$ is either $*$-even or $*$-odd, resp.~either $!$-even or $!$-odd, and~\eqref{eqn:adjses1}, resp.~\eqref{eqn:adjses2}, is obtained using the long exact sequence of equivariant cohomology associated with the adjunction triangle $j_! j^* \FC \rightarrow \FC \rightarrow i_*i^* \FC \xrightarrow{[1]}$, resp.~$i_* i^! \FC \rightarrow \FC \rightarrow j_*j^* \FC \xrightarrow{[1]}$, because by induction the various terms are concentrated either all in even degrees, or all in odd degrees.
Consider now the sequence~\eqref{eqn:adjses2} for $\FC = i^*_{\sqsubseteq \lambda} \EC$. In this case it is shown in~\cite[Proposition~5.9]{FieWil}\footnote{In~\cite{FieWil}, this result (as well as the other results used in this proof) is stated for a coefficient ring that is a complete local principal ideal domain, but the same proof applies verbatim for general coefficients.} that the composition of restriction morphisms
\[
\HM^\bullet_{\check{I}}({\EuScript Gr}, \EC) \rightarrow \HM^\bullet_{\check{I}}({\EuScript Gr}_\lambda, i_\lambda^* \EC) \rightarrow \HM^\bullet_{\check{I}}(\{L_\lambda\}, \imath_\lambda^* \EC)
\]
is surjective, where (as in~\S\ref{ss:intro-other}) $\imath_\lambda \colon \{L_\lambda\} \hookrightarrow {\EuScript Gr}$ is the inclusion. Since the second morphism is an isomorphism by~\cite[Proposition~2.3]{FieWil}, the first one is also surjective. And since this first morphism factors through the morphism
\[
\HM^\bullet_{\check{I}}({\EuScript Gr}_{\sqsubseteq \lambda}, i^*_{\sqsubseteq \lambda} \EC) \rightarrow \HM^\bullet_{\check{I}}({\EuScript Gr}_\lambda, i_\lambda^* \EC)
\]
considered in~\eqref{eqn:adjses2}, the latter morphism is also surjective, which finishes the proof in this case.
Finally, let us consider the sequence~\eqref{eqn:adjses1} for $\FC=i^!_{\sqsubseteq \lambda} \EC$. In this case, the arguments are similar to the ones used in the preceding case, using the fact that the natural morphism
\[
\HM^\bullet_{\check{I}}(\{L_\lambda\}, \imath_\lambda^! \EC) \rightarrow \HM^\bullet_{\check{I}}({\EuScript Gr}, \EC)
\]
is injective. (This fact is shown in the proof of~\cite[Proposition 5.9]{FieWil}: in fact this morphism identifies with the morphism considered in~\cite[Proposition 5.8(2)]{FieWil}.)
\end{proof}
Since $U={\EuScript Gr}_\lambda$ is isomorphic to an affine space, the graded $\HM^\bullet_{\check{I}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-module $\HM^\bullet_{c,\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is free of rank one. Moreover, the natural morphism
\[
\HM^{2 \dim({\EuScript Gr}_\lambda)}_{c,\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \HM^{2 \dim({\EuScript Gr}_\lambda)}_{c}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
is an isomorphism, and if $x_\lambda \in \HM^{2 \dim({\EuScript Gr}_\lambda)}_{c,\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ is the inverse image of a generator of $\HM^{2 \dim({\EuScript Gr}_\lambda)}_{c}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (as an ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-module), then $x_\lambda$ gives a basis of $\HM^\bullet_{c,\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ over $\HM^\bullet_{\check{I}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. We still denote by $x_\lambda$ the image of this element in $A_{\sqsubseteq \lambda}$, under the injection of Lemma~\ref{lem:cohom-lambda}.
\begin{lem}
\label{lem:action-xlambda}
Let $\EC$ be a Bott--Samelson parity complex, and let $\FC$ be either $i^*_{\sqsubseteq \lambda} \EC$ or $i^!_{\sqsubseteq \lambda} \EC$. Then the morphism
\[
\HM^\bullet_{\check{I}}(X, \FC) \rightarrow \HM^{\bullet+2 \dim({\EuScript Gr}_\lambda)}_{\check{I}}(X,\FC)
\]
given by the action of $x_\lambda \in A_{\sqsubseteq \lambda}$ factors as a composition
\[
\HM_{\check{I}}^\bullet(X, \FC) \twoheadrightarrow \HM^\bullet_{\check{I}}(X, j_*j^* \FC) \xrightarrow{\sim} \HM_{\check{I}}^{\bullet + 2\dim({\EuScript Gr}_\lambda)}(X, j_! j^* \FC) \hookrightarrow \HM^{\bullet + 2\dim({\EuScript Gr}_\lambda)}_{\check{I}}(X, \FC),
\]
where the first, resp.~third, morphism is the morphism appearing in~\eqref{eqn:adjses2}, resp. \eqref{eqn:adjses1}, and the second one is an isomorphism.
\end{lem}
\begin{proof}
The element
$x_\lambda \in A_{\sqsubseteq \lambda}$ acts trivially on $\HM_{\check{I}}^\bullet(X, i_*i^*\FC)$ and $\HM_{\check{I}}^\bullet(X, i_*i^!\FC)$. Hence its action on the module $\HM^\bullet_{\check{I}}(X, \FC)$ factors through a morphism
\[
\HM^\bullet_{\check{I}}(X, j_*j^* \FC) \rightarrow \HM_{\check{I}}^{\bullet + 2\dim({\EuScript Gr}_\lambda)}(X, j_! j^* \FC).
\]
To show that the latter morphism is an isomorphism, we observe that $j^* \FC$ is a direct sum of shifts of constant sheaves $\underline{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_U$ (since $\EC$ is a parity complex), and that our morphism is the corresponding direct sum of shifts of the isomorphism
\[
\HM^\bullet_{\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\sim} \HM^\bullet_{\check{I}}(\mathrm{pt}; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\sim} \HM^{\bullet+2\dim({\EuScript Gr}_\lambda)}_{c,\check{I}}(U;{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
determined by our choice of $x_\lambda \in \HM^{2 \dim({\EuScript Gr}_\lambda)}_{c,\check{I}}(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$.
\end{proof}
\begin{lem}
\label{lem:morph-exact-sequence}
Let $\EC_1, \EC_2$ be Bott--Samelson parity complexes. If $\phi \colon \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \rightarrow \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)$ is a morphism of $A_{\sqsubseteq \lambda}$-modules, then the composition
\begin{equation}
\label{eqn:composition-phi}
\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \xrightarrow{\phi} \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2) \twoheadrightarrow \HM^\bullet_{\check{I}}(X, j_* j^* i_{\sqsubseteq \lambda}^! \EC_2)
\end{equation}
(where the second morphism is the surjection appearing in~\eqref{eqn:adjses2}) factors uniquely through a morphism of $A_{\sqsubseteq \lambda}$-modules
\[
\phi' \colon \HM^\bullet_{\check{I}}(X, j_* j^* i_{\sqsubseteq \lambda}^* \EC_1) \rightarrow \HM^\bullet_{\check{I}}(X, j_* j^* i_{\sqsubseteq \lambda}^! \EC_2).
\]
Moreover, we have $\phi'=0$ iff $\phi$ factors as a composition
\[
\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \twoheadrightarrow \HM^\bullet_{\check{I}}(Z, i^*_{\sqsubset \lambda} \EC_1) \xrightarrow{\phi''} \HM^\bullet_{\check{I}}(Z, i^!_{\sqsubset\lambda}\EC_2) \hookrightarrow \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)
\]
where $\phi''$ is a morphism of $A_{\sqsubset \lambda}$-modules and the other morphisms are the ones appearing in~\eqref{eqn:adjses1} and~\eqref{eqn:adjses2}.
\end{lem}
\begin{proof}
Using Lemma~\ref{lem:exact-sequences-adj}, we see that the first claim is equivalent to the statement that the composition
\[
\HM^\bullet_{\check{I}}(X, i_* i^! i_{\sqsubseteq \lambda}^* \EC_1) \hookrightarrow
\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \xrightarrow{\phi} \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2) \twoheadrightarrow \HM^\bullet_{\check{I}}(X, j_* j^* i_{\sqsubseteq \lambda}^! \EC_2)
\]
vanishes. And using Lemma~\ref{lem:action-xlambda}, to prove this it suffices to prove that the composition
\[
\HM^\bullet_{\check{I}}(X, i_* i^! i_{\sqsubseteq \lambda}^* \EC_1) \hookrightarrow
\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \xrightarrow{\phi} \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2) \xrightarrow{x_\lambda} \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)
\]
vanishes. This follows from the fact that $\phi$ commutes with the action of $x_\lambda$, and that $x_\lambda$ acts trivially on $\HM^\bullet_{\check{I}}(X, i_* i^! i_{\sqsubseteq \lambda}^* \EC_1)$.
Now we consider the second claim. The ``if'' direction is easy. Conversely, assume that $\phi'=0$. Then the composition~\eqref{eqn:composition-phi} vanishes, hence the image of $\phi$ is included in the image of the embedding $\HM^\bullet_{\check{I}}(Z, i^!_{\sqsubset\lambda}\EC_2) \hookrightarrow \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)$ of Lemma~\ref{lem:exact-sequences-adj}. On the other hand, using Lemma~\ref{lem:action-xlambda}, this also implies that the composition of $\phi$ with the action of $x_\lambda$ on $\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)$ vanishes hence, since $\phi$ is a morphism of $A_{\sqsubseteq \lambda}$-modules, that $\phi$ vanishes on the image of the action of $x_\lambda$ on $\HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1)$. By Lemma~\ref{lem:action-xlambda}, we deduce that the composition
\[
\HM^\bullet(X, j_! j^* i_{\sqsubseteq \lambda}^* \EC_1) \hookrightarrow \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^* \EC_1) \xrightarrow{\phi} \HM^\bullet_{\check{I}}(X, i_{\sqsubseteq \lambda}^! \EC_2)
\]
vanishes which, in view of Lemma~\ref{lem:exact-sequences-adj}, proves the existence of a morphism of $A_{\sqsubseteq \lambda}$-modules $\phi'' \colon \HM^\bullet_{\check{I}}(X, i_* i^*_{\sqsubset \lambda} \EC_1) \rightarrow \HM^\bullet_{\check{I}}(X, i_* i^!_{\sqsubset\lambda}\EC_2)$ as in the lemma. Finally, the fact that $\phi''$ is a morphism of $A_{\sqsubset \lambda}$-modules follows from Lemma~\ref{lem:cohom-lambda}.
\end{proof}
Now we are ready to prove Proposition~\ref{prop:global-coh-ff}. We will proceed by induction
on $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, showing that for any Bott--Samelson parity complexes $\EC_1,\EC_2$, the cohomology functor induces an isomorphism of graded vector spaces
\begin{equation}
\Hom^\bullet_{{\EuScript Gr}_{\sqsubseteq \lambda}} \bigl( i^*_{\sqsubseteq \lambda}\EC_1,i^!_{\sqsubseteq \lambda}\EC_2 \bigr) \xrightarrow{\sim} \Hom_{A_{\sqsubseteq \lambda}} \bigl( \HM^\bullet_{\check{I}}({\EuScript Gr}_{\sqsubseteq \lambda}, i^*_{\sqsubseteq \lambda}\EC_1),\HM_{\check{I}}^\bullet({\EuScript Gr}_{\sqsubseteq \lambda}, i^!_{\sqsubseteq \lambda}\EC_2) \bigr). \label{eqn:ff}
\end{equation}
If $\lambda$ is minimal, then ${\EuScript Gr}_\lambda$ is a point and $i^*_{\sqsubseteq \lambda}\EC_1$ and $i^!_{\sqsubseteq \lambda}\EC_2$ are also parity. On the other hand, the cohomology functor induces an equivalence of categories between $\Parity_{\check{I}}({\EuScript Gr}_\lambda, {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and the full subcategory of the category of finitely generated graded $A_{\sqsubseteq \lambda}$-modules consisting of free modules. Thus \eqref{eqn:ff} is indeed an isomorphism in this case.
Now fix $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, and suppose that~\eqref{eqn:ff} is an isomorphism for all $\lambda' \sqsubset \lambda$.
We use the same notation as above for $X$, $Z$, $U$, $i$ and $j$.
Note that Lemma~\ref{lem:morph-exact-sequence} is equivalent to the existence of a natural sequence
\begin{equation}
\label{eqn:exact-sequence-Hom}
\begin{split}
& \Hom_{A_{\sqsubset \lambda}}(\HM^\bullet_{\check{I}}(Z, i^*_{\sqsubset\lambda} \EC_1),\HM^\bullet_{\check{I}}(Z, i^!_{\sqsubset\lambda}\EC_2)) \rightarrow \\
& \qquad \qquad \Hom_{A_{\sqsubseteq \lambda}}(\HM_{\check{I}}^\bullet(X, i^*_{\sqsubseteq\lambda} \EC_1),\HM^\bullet_{\check{I}}(X, i^!_{\sqsubseteq\lambda}\EC_2)) \rightarrow \\
& \qquad \qquad \qquad \qquad \Hom_{\HM_{\check{I}}^\bullet(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\HM^\bullet_{\check{I}}(U, j^*i^*_{\sqsubseteq\lambda} \EC_1),\HM^\bullet_{\check{I}}(U,j^*i^!_{\sqsubseteq\lambda}\EC_2))
\end{split}
\end{equation}
which is exact at the middle term. It is also easy to check (using Lemma~\ref{lem:exact-sequences-adj}) that the first morphism is injective.
Consider the adjunction triangle
\[
i_{\sqsubset \lambda *} i_{\sqsubset \lambda}^! \EC_2 \rightarrow i_{\sqsubseteq \lambda}^! \EC_2 \rightarrow j_* j^* i_{\sqsubseteq \lambda}^! \EC_2 \xrightarrow{[1]}
\]
and the long exact sequence
\begin{multline*}
\cdots \rightarrow \Hom^n_Z(i^*_{\sqsubset\lambda} \EC_1,i^!_{\sqsubset\lambda}\EC_2) \rightarrow \Hom^n_X(i^*_{\sqsubseteq\lambda} \EC_1,i^!_{\sqsubseteq\lambda}\EC_2) \\
\rightarrow \Hom^n_U(j^*i^*_{\sqsubseteq\lambda} \EC_1,j^*i^!_{\sqsubseteq\lambda}\EC_2) \rightarrow \cdots
\end{multline*}
obtained by appying the functor $\Hom_X(i^*_{\sqsubseteq\lambda} \EC_1,-)$. Parity considerations and~\cite[Corollary~2.8]{jmw} imply that the connecting morphisms in this long exact sequence are trivial, so that
the maps form short exact sequences in each degree.
Then one can consider the commutative diagram
\begin{equation*}
\xymatrix@R=0.5cm@C=1.2cm{
\Hom_Z^\bullet(i^*_{\sqsubset\lambda} \EC_1,i^!_{\sqsubset\lambda}\EC_2) \ar[r]^-{\HM^\bullet_{\check{I}}(Z,-)} \ar@{^{(}->}[d] & \Hom_{A_{\sqsubset \lambda}}(\HM^\bullet_{\check{I}}(Z, i^*_{\sqsubset\lambda} \EC_1),\HM^\bullet_{\check{I}}(Z, i^!_{\sqsubset\lambda}\EC_2)) \ar@{^{(}->}[d] \\
\Ext_X^\bullet(i^*_{\sqsubseteq\lambda} \EC_1,i^!_{\sqsubseteq\lambda}\EC_2) \ar[r]^-{\HM^\bullet_{\check{I}}(X,-)} \ar@{->>}[d] & \Hom_{A_{\sqsubseteq \lambda}}(\HM_{\check{I}}^\bullet(X, i^*_{\sqsubseteq\lambda} \EC_1),\HM^\bullet_{\check{I}}(X, i^!_{\sqsubseteq\lambda}\EC_2)) \ar[d]\\
\Ext_U^\bullet(j^*i^*_{\sqsubseteq\lambda} \EC_1,j^*i^!_{\sqsubseteq\lambda}\EC_2) \ar[r]^-{\HM^\bullet_{\check{I}}(U,-)} & \Hom_{\HM_{\check{I}}^\bullet(U; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\HM^\bullet_{\check{I}}(U, j^*i^*_{\sqsubseteq\lambda} \EC_1),\HM^\bullet_{\check{I}}(U,j^*i^!_{\sqsubseteq\lambda}\EC_2)),
}
\end{equation*}
where the right-hand column is the sequence~\eqref{eqn:exact-sequence-Hom}, and both column are exact at the middle term. By induction, the upper horizontal arrow is an isomorphism.
The lower one is an isomorphism because the cohomology functor induces an equivalence of categories between $\Parity_{\check{I}}({\EuScript Gr}_\lambda,{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and the full subcategory of free modules in the category of finitely generated graded $\HM^\bullet_{\check{I}}({\EuScript Gr}_\lambda; {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-modules. By the five-lemma, this implies that the middle horizontal arrow is also an isomorphism, which completes the induction step.
\begin{remark}
It can be easily checked that the proof of Proposition~\ref{prop:global-coh-ff} applies to any ring $k$ of coefficients which is Noetherian and of finite global dimension, and also in the $\check{I} \rtimes \Gm$-equivariant setting, or for other partial flag varieties of Kac--Moody groups. If $k$ is a complete local principal ideal domain, one can also work directly with the categories $\Parity_{\check{I} \rtimes \Gm}({\EuScript Gr}, k)$ and $\Parity_{\check{I}}({\EuScript Gr}, k)$ instead of restricting to ``Bott--Samelson objects.''
\end{remark}
\subsection{Graded ranks of $\mathrm{Hom}$ spaces}
\label{ss:graded-ranks-parity}
We identify the $\ZM[\mathsf{v},\mathsf{v}^{-1}]$-module $\MM_{\FC l}$ associated with $\FC l$ and its stratification by $\check{I}$-orbits as in~\S\ref{ss:reminder-parity} with $\HM_{\mathrm{aff}}$, where $\mathbf{e}_w$ corresponds to $T_{w^{-1}}$ (for $w \in W_{\mathrm{aff}}$). We also identify $\MM_{{\EuScript Gr}}$ (where ${\EuScript Gr}$ is stratified by $\check{I}$-orbits) with $\MM_{\mathrm{sph}}$, where $\mathbf{e}_\lambda$
corresponds to $\mathbf{m}_{-\lambda}$ (for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$).
\begin{lem}
\label{lem:ch-parity-Gr}
Let $\EM$ be either $\FM$ or ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\begin{enumerate}
\item
\label{it:ch-parity-Fl}
For any $\FC$, $\GC$ in $\Parity_{\check{I}}(\FC l,\EM)$, we have
\[
\ch_{\FC l}^!(\FC \star \GC) = \ch_{\FC l}^!(\GC) \cdot \ch_{\FC l}^!(\FC) \quad \text{and} \quad \ch_{\FC l}^*(\FC \star \GC) = \ch_{\FC l}^*(\GC) \cdot \ch_{\FC l}^*(\FC)
\]
in $\HM_{\mathrm{aff}}$.
\item
\label{it:ch-parity-Gr}
For any
$\GC$ in $\Parity_{\check{I}}(\FC l,\EM)$, we have
\[
\ch_{{\EuScript Gr}}^!(\GC \star \delta^\EM_1) = \mathbf{m}_0 \cdot \ch_{\FC l}^!(\GC) \quad \text{and} \quad \ch_{{\EuScript Gr}}^*(\GC \star \delta^\EM_1) = \mathbf{m}_0 \cdot \ch_{\FC l}^*(\GC)
\]
in $\MM_{\mathrm{sph}}$.
\end{enumerate}
\end{lem}
\begin{proof}[Sketch of proof]
The case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ follows from the case $\EM=\FM$, so we concentrate on the latter case. Also, in each case, the formula for $\ch^*$ follows from the formula for $\ch^!$ using Lemma~\ref{lem:grk-Hom-parity}\eqref{it:ch-parity-D}, so we only consider the latter case.
\eqref{it:ch-parity-Fl}
First we consider the case $\GC=\EC_{s,\FM}$ for some simple reflection $s$. We let $\check{J}_s$ be the minimal standard parahoric subgroup of $\check{G}(\mathscr{K})$ associated with $s$, and define $\FC l^s:= \check{G}(\mathscr{K}) / \check{J}_s$. We let $p_s \colon \FC l \rightarrow \FC l^s$ be the projection, and set $\FC l^s_w:=p_s(\FC l_w)$ for $w \in W_{\mathrm{aff}}$. Then by base change we have
\[
\FC \star \GC = \FC \star \EC_{s,\FM} \cong (p_s)^* (p_{s})_* \FC[1],
\]
and $\ch_{\FC l}^!(\EC_{s,\FM})=T_s + \mathsf{v}^{-1}$. The formula in this case can be checked by a direct computation, using the fact that
\[
\HM^\bullet(\FC l^s_w, (i_w^{\FC l^s})^! (p_{s})_* \FC) = \HM^\bullet(\FC l_w, (i_w^{\FC l})^! \FC) \oplus \HM^\bullet(\FC l_{ws}, (i_{ws}^{\FC l})^! \FC)
\]
for $w \in W_{\mathrm{aff}}$, which can be derived from the base change theorem and~\cite[Proposition~2.6]{jmw}. (Here $i_w^{\FC l^s}$, $i_w^{\FC l}$ and $i_{ws}^{\FC l}$ are the obvious inclusions.)
The case where $\GC$ is the sky-scraper sheaf $\EC_{\omega,\FM}$ at the $\check{I}$-fixed point given by the unique point in $\FC l_\omega$ (for $\omega \in \Omega$) is easy.
Using these special cases one deduces that the formula holds when $\GC$ is of the form
\begin{equation}
\label{eqn:parity-Fl}
\EC_{s_1,\FM} \star \cdots \star \EC_{s_r,\FM} \star \EC_{\omega,\FM}[n]
\end{equation}
where $\omega \in \Omega$, $n \in \ZM$, and $s_1, \cdots, s_r$ are simple reflections. Then one can prove the formula when $\GC$ is indecomposable by induction on the dimension of its support, using the fact that any indecomposable parity complex on $\FC l$ appears as a direct summand of an object of the form~\eqref{eqn:parity-Fl} with $r$ the dimension of the support, see~\cite[\S 4.1]{jmw}. The general case follows.
\eqref{it:ch-parity-Gr}
We have $\GC \star \delta^\EM_1=p_* \GC$, where $p \colon \FC l \rightarrow {\EuScript Gr}$ is the natural projection. Then the lemma can be checked by a direct computation, using the formula
\[
\HM^\bullet({\EuScript Gr}_\lambda, (i_\lambda)^! p_* \GC) \cong \bigoplus_{w \in t_\lambda W} \HM^\bullet(\FC l_w, (i_w^{\FC l})^! \GC)
\]
for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, which can be derived from the base change theorem and~\cite[Proposition~2.6]{jmw}.
\end{proof}
\begin{remark}
Using similar arguments one can show that $\ch_{{\EuScript Gr}}^?(\FC \star \GC)=\ch_{{\EuScript Gr}}^?(\GC) \cdot \ch_{\FC l}^?(\FC)$ for any $\GC$ in $\Parity_{\check{I}}({\EuScript Gr},\EM)$,
$\FC$ in $\Parity_{\check{I}}(\FC l,\EM)$ and $?=!$ or $*$.
\end{remark}
\begin{prop}
\label{prop:ch-BS-parity}
For $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ or $\FM$, for any sequence $\underline{s}$ of simple reflections, and for any $\omega \in \Omega$ we have
\[
\ch_{{\EuScript Gr}}^!(\EC_{\EM}(\omega,\underline{s})) =\ch_{{\EuScript Gr}}^*(\EC_{\EM}(\omega,\underline{s})) = \mathbf{m}(\omega,\underline{s}).
\]
\end{prop}
\begin{proof}
This follows immediately from Lemma~\ref{lem:ch-parity-Gr}, using the facts that $\ch_{\FC l}^!(\EC_{s,\EM})=\ch_{\FC l}^*(\EC_{s,\EM})=(T_s + \mathsf{v}^{-1})$ and $\ch_{\FC l}^*(\EC_{\omega,\EM}) = \ch_{\FC l}^*(\EC_{\omega,\EM}) = T_{\omega^{-1}}$ (where $\EC_{\omega,\EM}$ is defined as in the proof of Lemma~\ref{lem:ch-parity-Gr}\eqref{it:ch-parity-Fl}).
\end{proof}
The following consequence of Theorem~\ref{thm:main-top} will play a crucial role in Section~\ref{sec:tilting-KW}.
\begin{cor}
\label{cor:grk-Hom-parity}
For any any $\omega, \omega' \in \Omega$ and any sequences $\underline{s}$, $\underline{t}$ of simple reflections, the graded $\OC(\tg^*)$-module
\[
\bigoplus_{n \in \ZM} \Hom_{\mathsf{BS}^{\mathrm{alg}}} \bigl( (\omega, \underline{s}, 0), (\omega', \underline{t}, n) \bigr)
\]
is free. Its graded rank is equal to
$\langle \mathbf{m}(\omega, \underline{s}), \mathbf{m}(\omega', \underline{t}) \rangle$.
\end{cor}
\begin{proof}
The first assertion follows from Theorem~\ref{thm:main-top} and Lemma~\ref{lem:properties-parity}\eqref{it:morph-parity-For}.
The second one follows from Lemma~\ref{lem:grk-Hom-parity}\eqref{it:dim-Hom-parity-ch} and Proposition~\ref{prop:ch-BS-parity}.
\end{proof}
\section{Kostant--Whittaker reduction}
\label{sec:KW-reduction}
In this section and the next one we work with derived categories of equivariant coherent sheaves, and usual derived functors between them. See~\cite[Appendix~A]{mr} for a brief reminder of the main definitions and properties of these objects.
\subsection{Overview}
\label{ss:overview-KW}
The goal of this section is to introduce and study the ``Kostant--Whittaker reduction'' functor for equivariant coherent sheaves on the Grothendieck resolution of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. This construction is a mild adaptation of a construction in~\cite{bf}; it relies in a crucial way on geometric results proved in~\cite{riche3}. A related construction also appears in~\cite{dodd}. This functor is used in Section~\ref{sec:tilting-KW} to obtain a description of the category $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))$ in terms of ``Soergel bimodules.''
After introducing our notation and assumptions in~\S\ref{ss:notation-KW}, we recall the definition of the ``geometric braid group action'' of~\cite{riche, br} in~\S\ref{ss:braid-groth}, and the main results of~\cite{riche3} in~\S\ref{ss:reminder}. In~\S\ref{ss:KW-functors} we define our functors. Then, after some preparation in \S\ref{ss:KW-line-bundle}, in~\S\ref{ss:KW-braid-group} we prove the main result of the section, a certain compatibility property between Kostant--Whittaker reduction and the geometric braid group action.
\subsection{Notation}
\label{ss:notation-KW}
Let ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$ be a group scheme over $\ZM$ which is a product of split simply connected quasi-simple groups, and general linear groups $\mathrm{GL}_{n,\ZM}$. In particular, ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$ is a split connected reductive group over $\ZM$.
We let ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\ZM \subset {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$ be a Borel subgroup, ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_\ZM \subset {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\ZM$ be a (split) maximal torus, and ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_\ZM \subset {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$ be the Borel subgroup which is opposite to ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\ZM$ (with respect to ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_\ZM$).
We denote by $r$ the rank of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$.
We let $N$ be the product of all the prime numbers which are not very good for some quasi-simple factor of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$, and set ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}:=\ZM[1/N]$. Throughout the section, we use the letter $\FM$ to denote an arbitrary geometric point of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. We use the letter $\EM$ to denote either ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ or $\FM$.
We let ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ be the groups obtained from ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\ZM$, ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_\ZM$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_\ZM$ by base change to ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, $\bg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, $\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, $\bg^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ be their respective Lie algebras.
We also denote by ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, resp.~${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, the unipotent radical of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, resp.~${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, by $\ng_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, resp.~$\ng^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, its Lie algebra, by $W$ the Weyl group of $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}, {\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$, and by ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}:=X^*({\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ the weight lattice.
We also set
\[
{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM := \Spec(\FM) \times_{\Spec({\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}},
\]
denote by $\gg_\FM$ the Lie algebra of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM$, by ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$, ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_\FM$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_\FM$, ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\FM$, ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\FM^+$ the base change of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+$, and by $\bg_\FM$, $\tg_\FM$, $\bg^+_\FM$, $\ng_\FM$, $\ng_\FM^+$ their respective Lie algebras. Note that we also have ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}=X^*({\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_\FM)$.
If $V$ is an $\EM$-module, we set $V^*:=\Hom_\EM(V,\EM)$.
In this section we will consider the Grothendieck resolution
\[
\widetilde{\gg}_\EM:= ({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times^{{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}} (\gg/\ng)^*)_\EM \hookrightarrow ({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}/{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \times \gg^*)_\EM.
\]
The scheme $\widetilde{\gg}_\EM$ is a vector bundle over the flag variety $\mathscr{B}_\EM:=({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}/{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}})_\EM$. It is endowed with an action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM \times \GmE$, where ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ acts naturally, and the action of $\GmE$ is induced by the action on $(\gg/\ng)_\EM^*$ where $x \in \GmE$ acts by multiplication by $x^{-2}$.
We will consider the derived categories of equivariant coherent sheaves
$D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM$
and $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM$.
We denote by $\langle 1 \rangle$ the functor of tensoring with the free rank one tautological $\GmE$-module, and by
\[
\FM(-) \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\FM, \quad \FM(-) \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\FM
\]
the ``modular reduction functors'' induced by the functor $\FM \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (-)$.
For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ we denote by $\OC_{\mathscr{B}_\EM}(\lambda)$ the line bundle on $\mathscr{B}_\EM$ associated naturally with $\lambda$, and by $\OC_{\widetilde{\gg}_\EM}(\lambda)$ the pullback of $\OC_{\mathscr{B}_\EM}(\lambda)$ to $\widetilde{\gg}_\EM$.
We will also consider the morphism
\[
\nu \colon \widetilde{\gg}_\EM \rightarrow \tg^*_\EM
\]
which is defined as follows. Consider the restriction morphism $(\gg / \ng)_\EM^* \rightarrow \tg^*_\EM$. It is easily checked that this morphism is ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\EM$-equivariant, where ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\EM$ acts trivially on $\tg^*_\EM$. Therefore, our morphism defines a morphism $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times^{{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}} (\gg/\ng)^*)_\EM \rightarrow \tg^*_\EM$, which is our morphism $\nu$. There is also a natural morphism
\[
\pi \colon \widetilde{\gg}_\EM \rightarrow \gg^*_\EM
\]
induced by the coadjoint action.
By~\cite[Lemma~4.2.3]{riche3} there exists a ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-invariant symmetric bilinear form on $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ which is a perfect pairing. We fix once and for all such a bilinear form, and we denote by $\varkappa \colon \gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \xrightarrow{\sim} \gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*$ the induced (${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-equivariant) isomorphism. We denote similarly the induced isomorphism $\gg_\FM \xrightarrow{\sim} \gg^*_\FM$.
In the case $\EM=\FM$, we
let $\gg^{{\mathrm{rs}}}_\FM \subset \gg^{{\mathrm{reg}}}_\FM \subset \gg_\FM$ be the open subsets consisting of regular semisimple elements and regular elements respectively (see~\cite[\S 2.3]{riche3}), let $\gg_\FM^{*,{\mathrm{rs}}} \subset \gg_\FM^{*,{\mathrm{reg}}} \subset \gg^*_\FM$ be their image under $\varkappa$, and let $\widetilde{\gg}_\FM^{{\mathrm{rs}}} \subset \widetilde{\gg}_\FM^{{\mathrm{reg}}}$ be the inverse images in $\widetilde{\gg}_\FM$. Then there exists a natural action of $W$ on $\widetilde{\gg}^{{\mathrm{reg}}}_\FM$ stabilizing $\widetilde{\gg}_\FM^{{\mathrm{rs}}}$ and commuting with the ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM \times \GmF$-action, see~\cite[\S 1.9]{br}. Moreover, the restriction $\nu_{\mathrm{reg}} \colon \widetilde{\gg}_\FM^{\mathrm{reg}} \rightarrow \tg^*_\FM$ is $W$-equivariant, where $W$ acts naturally on $\tg_\FM^*$. (Indeed, this property is easily checked for the restriction of $\nu_{\mathrm{reg}}$ to $\widetilde{\gg}^{\mathrm{rs}}$; then the claim follows by a density argument.)
\subsection{Geometric braid group action}
\label{ss:braid-groth}
We let ${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$ be the lattice of cocharacters of ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and denote by $\Phi \subset {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, resp.~${\check \Phi} \subset {\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}}$, the roots, resp.~coroots, of $({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}, {\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. We denote by $\Phi^+$ the positive roots, i.e.~the roots appearing in $\bg^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. To these data one can associate the affine Weyl group $W_{\mathrm{aff}}$ and the affine braid group $\BM_{\mathrm{aff}}$ as in~\S\ref{ss:Haff}. We will also denote
by $\rho$ the half-sum of positive roots.
Let $s$ be a finite simple reflection, associated with a simple root $\alpha$.
Recall the associated subscheme $Z_s^\EM \subset (\widetilde{\gg} \times_{\gg^*} \widetilde{\gg})_\EM$ defined in~\cite[\S\S 1.3--1.5]{br}. (If $\EM=\FM$, then $Z_s^\EM$ is the closure of the graph of the action of $s$ on $\widetilde{\gg}^{{\mathrm{reg}}}_\EM$.) We denote by
\[
\TM^s_{\widetilde{\gg}}, \, \SM^s_{\widetilde{\gg}} \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}\times \Gm}(\widetilde{\gg})_\EM \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM
\]
the Fourier--Mukai transforms with kernels $\OC_{Z_s^\EM} \langle -1 \rangle$ and $\OC_{Z_s^\EM}(-\rho, \rho-\alpha) \langle -1 \rangle$ respectively. (Here, $\OC_{Z_s^\EM}(-\rho, \rho-\alpha)$ denotes the tensor product of $\OC_{Z_s^\EM}$ with the line bundle on $\widetilde{\gg} \times \widetilde{\gg}_\EM$ which is the pullback of the line bundle $\OC_{\mathscr{B}_\EM}(-\rho) \boxtimes \OC_{\mathscr{B}_\EM}(\rho-\alpha)$ on $\mathscr{B} \times \mathscr{B}_\EM$.) By \cite[Lemma~1.5.1 \& Proposition~1.10.3]{br}, these functors are quasi-inverse equivalences of categories. We will use the same symbols to denote the similar autoequivalences of the category $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM$.
By~\cite[Section~1]{br}, there exists a right action\footnote{Here by a (left) action of a group on a category we mean a group morphism from the given group to the group of \emph{isomorphism classes} of autoequivalences of the category. As usual, a right action of a group is a left action of the opposite group.} of the group $\BM_{\mathrm{aff}}$ on the category $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM$, resp.~$D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM$, where $T_s$ acts by the functor $\TM^s_{\widetilde{\gg}}$ (for any finite simple reflection $s$), and where $\theta_\lambda$ acts by tensoring with the line bundle $\OC_{\widetilde{\gg}_\EM}(\lambda)$ (for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$). (The same remarks as in~\cite[\S 3.3]{mr} on the difference with the conventions of~\cite{riche, br} apply here: the (right) action considered in the present paper differs from the (left) action of~\cite{riche,br} by the composition with the anti-automorphism of $\BM_{\mathrm{aff}}$ fixing all generators $T_s$ for $s$ a simple reflection and $\theta_\lambda$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.) For $b \in \BM_{\mathrm{aff}}$, we denote by
\[
{\EuScript I}^\EM_b \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM \xrightarrow{\sim} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM
\]
the action of $b$. (This functor is defined only up to isomorphism.) It is easily checked that we have an isomorphism of functors
\begin{equation}
\label{eqn:F-IS}
\FM \circ {\EuScript I}_b^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong {\EuScript I}_b^\FM \circ \FM
\end{equation}
for any $b \in \BM_{\mathrm{aff}}$.
For $s$ a finite simple reflection, associated with a simple root $\alpha$, we also set $\widetilde{\gg}_s^\EM := {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times^{{\mathbf P}} \def\pb{{\mathbf p}} \def\PC{{\mathcal{P}}_s} (\gg / (\pg_s)^{\mathrm{nil}})^*_\EM$, where ${\mathbf P}} \def\pb{{\mathbf p}} \def\PC{{\mathcal{P}}^\EM_s$ is the minimal standard parabolic subgroup of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ associated with $s$, and $(\pg_s^\EM)^{\mathrm{nil}}$ is the Lie algebra of the unipotent radical of ${\mathbf P}} \def\pb{{\mathbf p}} \def\PC{{\mathcal{P}}^\EM_s$. There exists a natural morphism $\widetilde{\pi}_s \colon \widetilde{\gg}_\EM \rightarrow \widetilde{\gg}^\EM_s$.
By~\cite[Corollary~5.3.2]{riche}
there exists natural exact sequences
\begin{equation}
\label{eqn:ses-tgg}
\OC_{\Delta \widetilde{\gg}_\EM} \langle 2 \rangle \hookrightarrow \OC_{\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}_\EM} \twoheadrightarrow \OC_{Z_s^\EM}, \quad
\OC_{Z_s^\EM}(-\rho, \rho-\alpha) \hookrightarrow \OC_{\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}_\EM} \twoheadrightarrow \OC_{\Delta \widetilde{\gg}_\EM}
\end{equation}
in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg} \times \widetilde{\gg})_\EM$, where in each sequence the surjection is induced by restriction of functions, and where $\Delta \widetilde{\gg}_\EM \subset \widetilde{\gg} \times \widetilde{\gg}_\EM$ is the diagonal copy. (In fact, in~\cite{riche} only the case $\EM=\FM$ is treated, but one can easily check that the same arguments apply in the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.)
\subsection{Reminder on~\cite{riche3}: Kostant section and universal centralizer}
\label{ss:reminder}
We denote by $\Phi^{\mathrm{s}} \subset \Phi$ the subset of simple roots. As in~\cite[\S 4.3]{riche3},\footnote{Note that, compared to~\cite{riche3}, we have switched the roles of positive and negative roots. Another difference which appears below is that we work with $\gg^*$ instead of $\gg$, using the identification $\varkappa$.} for each $\alpha \in -\Phi^{\mathrm{s}}$ we choose an element $e_\alpha \in \gg_\ZM$ which forms a $\ZM$-basis of the $\alpha$-weight space in $\gg_\ZM$, and set
\[
e:= \sum_{\alpha \in -\Phi^{\mathrm{s}}} e_\alpha \quad \in \gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
We will consider the cocharacter ${\check \lambda}_\circ:=\sum_{\alpha \in -\Phi^+} \alpha^{\vee}$, and the $\GmR$-actions on $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*$ defined by
\[
x \cdot y = x^{-2} {\check \lambda}_\circ(x) \cdot y
\]
for $x \in \GmR$ and $y$ either in $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ or in $\gg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*$. With these definitions, $\varkappa$ is $\GmR$-equivariant, $e$ is fixed by the action, and $\bg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+$ and $\ng_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+$ are $\GmR$-stable.
By~\cite[Lemma~4.3.1]{riche3}, the quotient $\bg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+/[e,\ng^+_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}]$ is free of rank $r$; therefore one can choose a $\GmR$-stable ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-submodule $\sg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \subset \bg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+$ such that $\bg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+=\sg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \oplus [e, \ng_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^+]$. We set
\[
\SC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} := \varkappa(e + \sg)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}, \qquad \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}:=\SC \times_{\gg^*} \widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
The $\GmR$-action on $\gg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ defined above stabilizes $\SC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and contracts it to $\varkappa(e)$ as $t \rightarrow \infty$. Similarly, the action on $\widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ defined by
\[
x \cdot [g:\xi] = [{\check \lambda}_\circ(x) g : x^{-2} \xi]
\]
(for $x \in \GmR$, $g \in {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and $\xi \in (\gg/\ng)^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$) stabilizes $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and is contracting
(see~\cite[\S 3.5]{riche3} for details). We will denote by $\SC_\FM$, resp.~$\widetilde{\SC}_\FM$, the scheme obtained from $\SC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, resp.~$\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, by base change to $\FM$.
The following result is proved in~\cite[Propositions~3.5.5 \&~4.5.2]{riche3}. Here we consider the $\GmE$-action on $\tg^*_\EM$ where $x$ acts by multiplication by $x^{-2}$.
\begin{prop}
\label{prop:Kostant-section}
The morphism $\nu \colon \widetilde{\gg}_\EM \rightarrow \tg^*_\EM$ induces a $\GmE$-equivariant isomorphism of $\EM$-schemes
\[
\nu_\SC \colon \widetilde{\SC}_\EM \xrightarrow{\sim} \tg^*_\EM.
\]
\end{prop}
The following result is proved in~\cite[Lemmas 3.5.1 \& 4.5.1]{riche3}.
\begin{lem}
\label{lem:a-smooth}
The morphism
\[
a \colon {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \widetilde{\SC}_\EM \rightarrow \widetilde{\gg}_\EM
\]
induced by the ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$-action on $\widetilde{\gg}_\EM$ is smooth. When $\EM=\FM$, it factorizes through a surjective morphism ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \widetilde{\SC}_\FM \rightarrow \widetilde{\gg}^{\mathrm{reg}}_\FM$.
\end{lem}
We denote by $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\EM$ the universal centralizer associated with the action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ on $\widetilde{\gg}_\EM$, see~\eqref{eqn:stabilizer}. We will also denote by $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM$ the restriction of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\EM$ to $\widetilde{\SC}_\EM$. Then by~\cite[Corollary 3.5.8 \& Proposition 4.5.3]{riche3}, $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM$ is a commutative smooth group scheme over $\widetilde{\SC}_\EM$. We denote by $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^\EM$ its Lie algebra; it is a locally free sheaf of commutative $\OC_{\widetilde{\SC}_\EM}$-Lie algebras (see~\cite[\S 2.1]{riche3}).
Recall that the quotient $\tg^*_\EM/W$ is a smooth scheme, isomorphic to an affine space (see~\cite[Th{\'e}or{\`e}me~3 \& Corollaire on p.~296]{demazure}).
We will denote by $\varrho \colon \tg^*_\EM \rightarrow \tg^*_\EM/W$ the quotient morphism; then we denote by
$\eta_\SC \colon \widetilde{\SC}_\EM \rightarrow \tg^*_\EM/W$ the composition
\[
\widetilde{\SC}_\EM \xrightarrow{\nu_\SC} \tg^*_\EM \xrightarrow{\varrho} \tg^*_\EM/W.
\]
The following result follows from~\cite[Theorems~3.5.12 \& 4.5.5]{riche3}, using the fact that the natural morphism $\SC_\EM \rightarrow \tg^*_\EM / W$ is an isomorphism (see~\cite[Theorems 3.2.2 \& 4.3.3]{riche3}, combined with~\cite[Propositions 2.3.2 \& 4.2.1]{riche3}).
\begin{thm}
\label{thm:Lie-centralizer}
There exists a canonical isomorphism of sheaves of commutative $\OC_{\widetilde{\SC}_\EM}$-Lie algebras
\[
\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_{\SC}^\EM \cong (\eta_\SC)^* \Omega_{\tg^*_\EM/W}.
\]
\end{thm}
For the remainder of this subsection, let us
consider the case $\EM=\FM$. First we recall that there exists a unique smooth commutative group scheme $\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM$ over $\tg^*_\FM$ whose pull-back under $\nu_{\mathrm{reg}}$ is the restriction $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM_{\mathrm{reg}}$ of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM$ to $\widetilde{\gg}_\FM^{\mathrm{reg}}$, see~\cite[Proposition~3.3.9]{riche3}.
Now by~\cite[Remark~3.5.9]{riche3}, the composition
\[
\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg}^{\mathrm{reg}})_\FM \rightarrow \Rep(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM_{\mathrm{reg}}) \rightarrow \Rep(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM_\SC) \xrightarrow{\sim} \Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM)
\]
is an equivalence of categories, where the first functor is the functor~\eqref{eqn:stabilizer-functor} in our particular situation, the second functor is induced by restriction to $\widetilde{\SC}$, and the last functor is induced by the isomorphism of Proposition~\ref{prop:Kostant-section}.
Similarly, let $s$ be a finite simple reflection, let $W_s:=\{1,s\} \subset W$, and consider the natural morphism $\widetilde{\gg}_s^\FM \rightarrow \tg_\FM^*/W_s$. Let also $\widetilde{\gg}_s^{\FM,{\mathrm{reg}}} \subset \widetilde{\gg}_s^\FM$ be the inverse image of $\gg^{*,{\mathrm{reg}}}_\FM$ under the natural morphism $\widetilde{\gg}_s^\FM \rightarrow \gg^*_\FM$, and let $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM_{s}$ be the universal stabilizer associated with the action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM$ on $\widetilde{\gg}^\FM_s$. Then the same arguments as in the case of $\widetilde{\gg}_\FM$ (see in particular~\cite[Remark~3.5.7]{riche3}) show that there exists a unique commutative group scheme $\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}_s^\FM$ on $\tg^*_\FM / W_s$ whose pull-back under the morphism $\widetilde{\gg}^{\FM,{\mathrm{reg}}}_s \rightarrow \tg_\FM^*/W_s$ is the restriction of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\FM_{s}$ to $\widetilde{\gg}_s^{\FM,{\mathrm{reg}}}$; moreover, as above we have a natural equivalence of categories
$\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg}^{{\mathrm{reg}}}_s)_\FM \xrightarrow{\sim} \Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}_s^\FM)$.
There exists a canonical Cartesian diagram
\[
\xymatrix@R=0.4cm{
\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM \ar[r] \ar[d] & \widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}_s^\FM \ar[d] \\
\tg^*_\FM \ar[r] & \tg^*_\FM / W_s,
}
\]
so that the direct and inverse image functors under the quotient morphism $\tg^*_\FM \rightarrow \tg^*_\FM / W_s$ induce functors $\Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM) \rightarrow \Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM_s)$ and $\Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM_s) \rightarrow \Rep(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^\FM)$ respectively. Under the equivalences considered above, these functors correspond to the direct and inverse image functors under the restriction of $\widetilde{\pi}_s$ to $\widetilde{\gg}_\FM^{\mathrm{reg}}$, respectively.
\subsection{Definition of the functors}
\label{ss:KW-functors}
Let us denote by $\mathsf{T}(\tg^*_\EM/W)$ the tangent bundle of the smooth $\EM$-scheme $\tg^*_{\EM} / W$. We consider the $\GmE$-action on $\tg^*_{\EM} / W$ such that the corresponding grading on $\OC(\tg^*_{\EM} / W)=\OC(\tg^*_{\EM})^W$ is obtained by restriction of the grading on $\OC(\tg^*_{\EM})$ where the generators $\tg_\EM \subset \OC(\tg^*_{\EM})$ are in degree $2$, so that the morphism $\varrho$ of~\S\ref{ss:reminder} is $\GmE$-equivariant. This action induces a $\GmE$-equivariant structure on $\Omega_{\tg^*_\EM/W}$. We consider the $\GmE$-action on $\mathsf{T}(\tg^*_\EM/W)$ such that $\OC(\mathsf{T}(\tg^*_\EM/W))$ is the symmetric algebra of $\Omega_{\tg^*_\EM/W} \langle -2 \rangle$ as a graded algebra. In other words, the $\GmE$-action on $\mathsf{T}(\tg^*_\EM/W)$ is the combination of the action induced by the action on $\tg^*_\EM/W$, with multiplication by $x^2$ in the fibers of the projection $\mathsf{T}(\tg^*_\EM/W) \rightarrow \tg^*_\EM/W$.
We will consider the bounded derived categories of (equivariant) coherent sheaves
$D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_\EM$
and $D(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$.
In this context also we have ``modular reduction functors''
\begin{align*}
\FM(-) \colon D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} &\rightarrow D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\FM}, \\
\FM(-) \colon D(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} &\rightarrow D(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\FM}.
\end{align*}
Now we can explain the construction of the ``Kostant--Whittaker reduction'' functors
\begin{align}
\kappa_\EM \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM &\rightarrow D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}, \label{eqn:KW1} \\
\overline{\kappa}_\EM \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM &\rightarrow D(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}. \label{eqn:KW2}
\end{align}
First,
consider the functor
\[
\overline{\kappa}'_\EM \colon \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM \rightarrow \Coh(\tg^*)_{\EM}
\]
defined as the pullback functor associated with the embedding $\widetilde{\SC}_\EM \hookrightarrow \widetilde{\gg}_\EM$, where we identify $\widetilde{\SC}_\EM$ with $\tg^*_{\EM}$ using the isomorphism $\nu_\SC$ from Proposition~\ref{prop:Kostant-section}.
\begin{remark}
Consider the closed subvariety $\widetilde{\Upsilon}_\FM \subset \widetilde{\gg}_\FM$ defined in~\cite[\S 3.5]{riche3}, and the analogously defined closed subscheme $\widetilde{\Upsilon}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \subset \widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Then ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\EM$ acts freely on $\widetilde{\Upsilon}_\EM$, and $\nu$ induces an isomorphism of $\EM$-schemes
$\widetilde{\Upsilon} / {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\EM \xrightarrow{\sim} \tg^*_\EM$. (In the case $\EM=\FM$, this fact follows from~\cite[Proposition~3.2.1, Theorem~3.2.2 \& Proposition~3.5.5]{riche3}; the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is similar.) Using this isomorphism, the functor $\overline{\kappa}_\EM'$ can be described more canonically as the composition of restriction to $\widetilde{\Upsilon}_\EM$ with the natural equivalences $\Coh^{{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}}(\widetilde{\Upsilon})_\EM \xrightarrow{\sim} \Coh(\widetilde{\Upsilon} / {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\EM) \cong \Coh(\tg^*_{\EM})$.
\end{remark}
\begin{lem}
\label{lem:KW-exact}
The functor $\overline{\kappa}'_\EM$
is exact.
\end{lem}
\begin{proof}
Our functor can be written as the composition
\[
\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM \xrightarrow{a^*} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \widetilde{\SC})_\EM \xrightarrow{\sim} \Coh(\widetilde{\SC})_\EM \xrightarrow[\sim]{(\nu_\SC)_*} \Coh(\tg^*)_\EM,
\]
where $a$ is the morphism considered in Lemma~\ref{lem:a-smooth}, and the middle arrow is the obvious equivalence. Now $a$ is a smooth (in particular, flat) morphism by Lemma~\ref{lem:a-smooth}, which implies exactness.
\end{proof}
Now we can explain the definition of $\overline{\kappa}_\EM$. This functor will be induced by an exact functor $\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM \rightarrow \Coh(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$, which we denote similarly. In fact, starting from an equivariant coherent sheaf $\FC$ on $\widetilde{\gg}_\EM$, its restriction to $\widetilde{\SC}_\EM$ is naturally endowed with an action of the universal centralizer $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM$, see~\S\ref{ss:equiv-coh}. Differentiating this action we obtain an action of the Lie algebra $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^\EM$ of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM$. By Theorem~\ref{thm:Lie-centralizer} one can identify $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^\EM$ with $(\eta_\SC)^* \Omega_{\tg^*_\EM /W}$, considered as a sheaf of commutative Lie algebras on $\widetilde{\SC}_\EM$. Identifying $\widetilde{\SC}_\EM$ with $\tg^*_\EM$ via $\nu_\SC$ (see Proposition~\ref{prop:Kostant-section}), we obtain an action of the commutative Lie algebra $\varrho^* \Omega_{\tg_\EM^*/W}$ on $\overline{\kappa}'_\EM(\FC)$, hence also an action of its symmetric algebra $\varrho^*\mathrm{S}_{\OC_{\tg_\EM^*/W}}(\Omega_{\tg^*_\EM/W})$. This action defines a coherent sheaf $\overline{\kappa}_\EM(\FC)$ on $\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W)_{\EM}$ whose direct image under the affine morphism $\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W)_{\EM} \rightarrow \tg^*_\EM$ is $\overline{\kappa}'_\EM(\FC)$.
It follows from Lemma~\ref{lem:KW-exact} that the functor
\[
\overline{\kappa}_\EM \colon \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM \rightarrow \Coh(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}
\]
that we have just defined is exact. Then the functor~\eqref{eqn:KW2} is defined as the induced functor between bounded derived categories.
The construction of the functor $\kappa_\EM$ is similar, simply keeping track of the appropriate $\GmE$-actions. More precisely, recall the action of $\GmE$ on $\widetilde{\SC}_\EM$ defined in \S\ref{ss:reminder}. One can ``extend'' this action to the group scheme ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM \times \widetilde{\SC}_\EM$ via
\[
x \cdot (g,y)=({\check \lambda}_\circ(x) g {\check \lambda}_\circ(x)^{-1}, x \cdot y)
\]
for $x \in \GmE$, $g \in {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ and $y \in \widetilde{\SC}_\EM$. Then the subgroup $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_{\SC}^\EM$ is $\GmE$-stable, and the projection $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM \rightarrow \widetilde{\SC}_\EM$ is $\GmE$-equivariant. This action induces a $\GmE$-equivariant structure of the coherent sheaf $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^\EM$ (on $\widetilde{\SC}_\EM$), and it is easily seen that the isomorphism of Theorem~\ref{thm:Lie-centralizer} is $\GmE$-equivariant, where the action on the right-hand side is induced by the action on $\mathsf{T}(\tg^*_\EM/W)$ considered at the beginning of this subsection (see~\cite[Proof of Theorem~3.4.2]{riche3}). Now if $\FC$ is in $\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM$, then the restriction of $\FC$ to $\widetilde{\SC}_\EM$ is a $\GmE$-equivariant coherent sheaf, and the action of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^\EM$ is compatible with this structure in the obvious sense, which allows to define the exact functor
\[
\kappa_\EM \colon \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM \rightarrow \Coh^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}
\]
by the same recipe as for $\overline{\kappa}_\EM$. Then the functor~\eqref{eqn:KW1} is defined as the induced functor between bounded derived categories.
The proof of the following lemma is easy, and left to the reader.
\begin{lem}
\label{lem:F-kappa}
There exist canonical isomorphisms of functors
\[
\FM \circ \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong \kappa_\FM \circ \FM, \qquad \FM \circ \overline{\kappa}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong \overline{\kappa}_\FM \circ \FM,
\]
where, in both equations, the functor $\FM$ on the left-hand side is the modular reduction functor on the ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-scheme $(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}$, and the functor $\FM$ on the right-hand side is the similar functor on $\widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.\qed
\end{lem}
\subsection{Another geometric braid group action}
\label{ss:weyl-action}
The goal of this subsection is to define a ``geometric'' (right) action of the group $\BM_{\mathrm{aff}}$ on the category $D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$. We begin by defining an action of $W_{\mathrm{aff}}$.
First, there exists a natural action of $W$ on the $\EM$-scheme $(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$, induced by the action on $\tg^*_\EM$. For $w \in W$, we denote by
\[
{}' {\EuScript K}_{w}^\EM \colon D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM} \xrightarrow{\sim} D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}
\]
the pullback functor associated with the morphism given by the action of $w$.
Then, let $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. To define the functor associated with $t_\lambda$ we will identify $D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$ with the derived category of finitely generated graded modules over the commutative $\OC(\tg^*_\EM)$-Lie algebra $\bigl( \OC(\tg^*) \otimes_{\OC(\tg^*/W)} \Omega(\tg^*/W) \bigr)_\EM$. We define a module $F_\lambda^\EM$ for this Lie algebra as follows. As an $\OC(\tg^*_\EM)$-module, $F_\lambda^\EM$ is free of rank one. Then to define the action of the commutative Lie algebra $\OC(\tg^*) \otimes_{\OC(\tg^*/W)} \Omega(\tg^*/W)_\EM$ it is enough to define a morphism of $\OC(\tg^*_\EM)$-modules
\begin{equation}
\label{eqn:definition-F-lambda}
\bigl( \OC(\tg^*) \otimes_{\OC(\tg^*/W)} \Omega(\tg^*/W) \bigr)_\EM \rightarrow \OC(\tg^*_\EM).
\end{equation}
In order to do so, we interpret the left-hand side as the module of sections of the projection
\begin{equation}
\label{eqn:proj-T*}
(\tg^* \times_{\tg^* / W} \mathsf{T}^*(\tg^* / W))_\EM \rightarrow \tg^*_\EM,
\end{equation}
where $\mathsf{T}^*(\tg^*_{\EM} / W)$ is the cotangent bundle of the smooth $\EM$-scheme $\tg^*_{\EM} / W$.
The morphism $\varrho \colon \tg^*_\EM \rightarrow \tg^*_\EM/W$ defines a natural morphism
\[
d^*\varrho \colon (\tg^* \times_{\tg^* / W} \mathsf{T}^*(\tg^* / W))_\EM \rightarrow \mathsf{T}^*(\tg^*_\EM).
\]
Now since $\tg^*_\EM$ is an affine space, the right-hand side is canonically isomorphic to $\tg^*_\EM \times \tg_\EM$. Hence, starting with a section $\sigma$ of~\eqref{eqn:proj-T*} the composition of $\sigma$ with
\[
(\tg^* \times_{\tg^* / W} \mathsf{T}^*(\tg^* / W))_\EM \xrightarrow{d^*\varrho} \mathsf{T}^*(\tg^*_\EM) \cong \tg^* \times \tg_\EM \twoheadrightarrow \tg_\EM \xrightarrow{d\lambda} \EM
\]
defines an element in $\OC(\tg^*_\EM)$. This construction provides the definition of~\eqref{eqn:definition-F-lambda}, hence also of the module $F_\lambda^\EM$. It is clear that if $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $w \in W$ we have canonical isomorphisms of $\bigl( \OC(\tg^*) \otimes_{\OC(\tg^*/W)} \Omega(\tg^*/W) \bigr)_\EM$-modules
\begin{equation}
\label{eqn:compositions-F}
F_\lambda^\EM \otimes_{\OC(\tg^*_\EM)} F_\mu^\EM \cong F_{\lambda+\mu}^\EM \quad \text{and} \quad {}' {\EuScript K}_{w}^\EM(F_\lambda^\EM) = F^\EM_{w^{-1}\lambda}.
\end{equation}
(In the first isomorphism, the left-hand side is endowed with the natural action on the tensor product).
With this definition at hand, we define the functor
\[
{}' {\EuScript K}_{t_\lambda}^\EM \colon D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_\EM \xrightarrow{\sim} D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}
\]
as the functor of tensoring (over $\OC(\tg^*_\EM)$) with the module $F_\lambda^\EM$.
Using~\eqref{eqn:compositions-F} one can easily check that we have canonical isomorphisms
\[
{}' {\EuScript K}_{v}^\EM \circ {}' {\EuScript K}_{w}^\EM \cong {}' {\EuScript K}_{w v}^\EM, \quad {}' {\EuScript K}_{t_\lambda}^\EM \circ {}' {\EuScript K}_{t_\mu}^\EM \cong {}' {\EuScript K}_{t_{\lambda+\mu}}^\EM, \quad {}' {\EuScript K}_{w}^\EM \circ {}' {\EuScript K}_{t_\lambda}^\EM \cong {}' {\EuScript K}_{t_{w^{-1} \lambda}}^\EM \circ {}' {\EuScript K}_{w}^\EM
\]
for all $v,w \in W$ and $\lambda, \mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. In other words, these functors define a right action of the group $W_{\mathrm{aff}}$ on the category $D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$. For $w \in W_{\mathrm{aff}}$, we denote by $'{\EuScript K}^\EM_w$ the functor giving the action of $w$.
Now we ``renormalize'' this action as follows.
For $w \in W$ and $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, we set
\[
{\EuScript K}_{T_w}^\EM := {}' {\EuScript K}_w^\EM \langle -\ell(w) \rangle, \qquad {\EuScript K}_{\theta_\lambda}^\EM := {}' {\EuScript K}^\EM_{t_\lambda} \langle \lambda({\check \lambda}_\circ)\rangle.
\]
Then these functors extend to a right action of $\BM_{\mathrm{aff}}$ on $D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\EM}$. For any $b \in \BM_{\mathrm{aff}}$, we denote by ${\EuScript K}_b^\EM$ the action of $b$. Note that if $\overline{b}$ is the image of $b$ under the canonical surjection $\BM_{\mathrm{aff}} \twoheadrightarrow W_{\mathrm{aff}}$, then there exists $n(b) \in \ZM$ such that
\begin{equation}
\label{eqn:K-'K}
{\EuScript K}_b^\EM \cong {}' {\EuScript K}^\EM_{\overline{b}} \langle n(b) \rangle.
\end{equation}
Note also that if $\omega \in \Omega$ then we have necessarily
\begin{equation}
\label{eqn:n-Tomega}
n(T_\omega)=0.
\end{equation}
(Indeed,
if the claim is known for a power of $\omega$ then it follows for $\omega$. Now this claim is obvious if $\omega=t_\lambda$ for some $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ such that $\lambda( \alpha^\vee )=0$ for all $\alpha \in \Phi$. Since the quotient of $\Omega$ by the subgroup consisting of such elements is finite, this suffices to imply the claim for all $\omega \in \Omega$.)
\subsection{Kostant--Whittaker reduction of line bundles}
\label{ss:KW-line-bundle}
The goal of this subsection is to prove the following result (whose proof is quite technical, even though the statement is very natural).
\begin{prop}
\label{prop:KW-line-bundle}
For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, there exists an isomorphism
\[
\kappa_\EM(\OC_{\widetilde{\gg}_\EM}(\lambda)) \cong F_\lambda^\EM \langle \lambda({\check \lambda}_\circ) \rangle.
\]
\end{prop}
We start with two preliminary results.
The first lemma is a generalization of a lemma in~\cite{gk}.
\begin{lem}
\label{lem:S-position}
If $g \in {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM$ and $\xi \in (\gg/\ng)_\FM^*$ are such that $g \cdot \xi \in \SC_\FM$, then $g \in {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+_\FM \cdot {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$.
\end{lem}
\begin{proof}
Our assumption ensures that $[g:\xi] \in \widetilde{\SC}_\FM$.
Recall the contracting $\GmF$-actions on $\SC_\FM$ and $\widetilde{\SC}_\FM$ defined in~\S\ref{ss:reminder}. We have $\lim_{x \rightarrow \infty} x \cdot [g:\xi] = [1:\varkappa(e)]$, hence $\lim_{x \rightarrow \infty} {\check \lambda}_\circ(x) g {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM = 1 {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$.
Since $1 {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$ belongs to the ${\check \lambda}_\circ$-stable open subset ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\FM^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM / {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM \subset \mathscr{B}_\FM$, we deduce that
$g{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$ also belongs to ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}_\FM^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM / {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}_\FM$, which finishes the proof.
\end{proof}
\begin{lem}
\label{lem:Flambda-R-F}
Let $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$. Let $M$ be an object of $\Coh^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ which is flat over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and whose direct image to $\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is coherent. Assume that for any geometric point $\FM$ of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ there exists an isomorphism $\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} M \cong F_\lambda^\FM$ in $\Coh^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_\FM$. Then there exists an isomorphism $M \cong F_\lambda^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\end{lem}
\begin{proof}
In this proof we identify the category $\Coh^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ with the category of finitely generated graded $\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules, and similarly for $\FM$.
First we construct an isomorphism of graded $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-modules $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\sim} M$. In fact, if $M_0$ denotes the degree $0$-part of $M$ (a free ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-module of finite rank), our assumption ensures that $\dim_\CM(\CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} M_0)=1$; we deduce that $M_0$ has rank $1$.
Choosing any basis for this module, we obtain a morphism of graded $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-modules
\[
\phi \colon \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow M.
\]
Our assumption implies that the induced morphism $\OC(\tg^*_\FM) \rightarrow \FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} M$ is an isomorphism for all $\FM$. Since all graded pieces in $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and $M$ are free ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules of finite rank, we deduce that $\phi$ is an isomorphism.
We claim that $\phi$ induces an isomorphism of graded $\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules $F_\lambda^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \xrightarrow{\sim} M$. In fact, it is enough to prove that for all $\omega \in \Omega(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}/W)$ we have $\omega \cdot \phi(1) = (d\lambda)(\omega') \cdot \phi(1)$, where $\omega'$ is the image of $\omega$ in $\Omega(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) = \tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ as in~\S\ref{ss:weyl-action} (and we still write $d\lambda$ for the morphism $(d\lambda) \otimes 1 \colon \tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$). However,
the embedding
\[
\End_{\mathrm{Mod}^{\mathrm{gr}}(\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_{\CM})}(F_\lambda^\CM) \hookrightarrow \End_{\mathrm{Mod}^{\mathrm{gr}}(\OC(\tg^*_{\CM}))}(F_\lambda^\CM)
\]
is an isomorphism, since the right-hand side is reduced to scalars. Hence $\CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \phi$ induces an isomorphism of $\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W))_\CM$-modules $F_\lambda^\CM \xrightarrow{\sim} \CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} M$. We deduce that the image of $\omega \cdot \phi(1) - (d\lambda)(\omega') \cdot \phi(1)$ in $\CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} M$ is $0$; it follows that this element is zero, which finishes the proof.
\end{proof}
\begin{proof}[Proof of Proposition~{\rm \ref{prop:KW-line-bundle}}]
Using Lemmas~\ref{lem:F-kappa} and~\ref{lem:Flambda-R-F}, it is enough to prove the isomorphism in the case $\EM=\FM$.
For simplicity, in the proof we omit the subscripts ``$\FM$.''
We denote by $\widetilde{\gg}^\diamond$ the inverse image in $\widetilde{\gg}$ of the open subset ${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} / {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \subset \mathscr{B}$; more concretely we have $\widetilde{\gg}^\diamond = {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \times^{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} (\gg/\ng)^*$. This open subset is stable under the action of $\Gm$ obtained by restricting the action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$ along the embedding $\Gm \rightarrow {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$ sending $x$ to $({\check \lambda}_\circ(x), x)$. Moreover,
by Lemma~\ref{lem:S-position}, we have $\widetilde{\SC} \subset \widetilde{\gg}^\diamond$.
First, we claim that there exists a canonical isomorphism of $\Gm$-equivariant line bundles (in other words a canonical trivialization)
\begin{equation}
\label{eqn:rest-line-bundle}
\OC_{\widetilde{\gg}}(\lambda)_{| \widetilde{\gg}^\diamond} \xrightarrow{\sim} \OC_{\widetilde{\gg}^\diamond} \langle \lambda({\check \lambda}_\circ) \rangle.
\end{equation}
In fact, consider the variety $\XC:={\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times^{{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}} (\gg/\ng)^*$. This variety is endowed with a (free) action of ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$ defined by $t \cdot [g:\xi] = [g t^{-1}: t\cdot \xi]$, and $\widetilde{\gg}$ is the quotient of $\XC$ by this action. It is also endowed with a natural action of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$, such that the quotient morphism $q \colon \XC \rightarrow \widetilde{\gg}$ is ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$-equivariant. Moreover, it follows from the definition that we have
\[
\OC_{\widetilde{\gg}}(\lambda) = (q_* \OC_{\XC})^{{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}},-\lambda}
\]
as ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$-equivariant line bundles,
where on the right-hand side we mean sections on which ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$ acts by the character $-\lambda$. Consider now the following commutative diagram:
\[
\xymatrix@R=0.5cm{
{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^* \ar[r]^-{\sim} \ar[d]^{q^\diamond} & {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \times^{{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}} (\gg/\ng)^* \ar[d] \ar@{^{(}->}[r] & \XC \ar[d]^-{q} \\
{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^* \ar[r]^-{\sim} & {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \times^{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} (\gg/\ng)^* \ar@{^{(}->}[r] & \widetilde{\gg}.
}
\]
Here $q^\diamond$ is the projection, the right horizontal arrows are the natural (open) embeddings, and the left arrows are defined by $(t,u,\xi) \mapsto [ut^{-1}: t \cdot \xi]$ and $(u,\xi) \mapsto [u:\xi]$ respectively. All the maps in this diagram are $\Gm$-equivariant, if $\Gm$ acts on $\XC$ and $\widetilde{\gg}$ via the morphism $\Gm \rightarrow {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm$ considered above, and on ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*$, resp.~${\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*$, via
\begin{multline*}
x \cdot (t,u,\xi)=(t \cdot {\check \lambda}_\circ^{-1}(x), {\check \lambda}_\circ(x) u {\check \lambda}_\circ^{-1}(x), x^{-2}{\check \lambda}_\circ(x) \cdot \xi), \\
\text{resp.} \quad x \cdot (u,\xi)=({\check \lambda}_\circ(x) u {\check \lambda}_\circ^{-1}(x), x^{-2}{\check \lambda}_\circ(x) \cdot \xi).
\end{multline*}
We deduce an isomorphism of $\Gm$-equivariant line bundles
\[
\OC_{\widetilde{\gg}}(\lambda)_{| {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*} \xrightarrow{\sim} \bigl( (q^\diamond)_* \OC_{{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*} \bigr)^{{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}},-\lambda}.
\]
Now we have $(q^\diamond)_* \OC_{{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times {\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*} \cong \FM[{\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}] \otimes_\FM \OC_{{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*}$, and the subsheaf where ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$ acts by $-\lambda$ is $(\FM \lambda) \otimes_\FM \OC_{{\mathbf U}} \def\ub{{\mathbf u}} \def\UC{{\mathcal{U}}^+ \times (\gg/\ng)^*}$. We deduce~\eqref{eqn:rest-line-bundle}.
Using~\eqref{eqn:rest-line-bundle} we obtain a canonical isomorphism of $\Gm$-equivariant $\OC_{\tg^*}$-modules
\begin{equation}
\label{eqn:kappa-line-bundle}
\kappa(\OC_{\widetilde{\gg}}(\lambda)) \cong \OC_{\tg^*} \langle \lambda({\check \lambda}_\circ) \rangle.
\end{equation}
To conclude we have to identify the action of $\Omega(\tg^*/W)$ on $\kappa(\OC_{\widetilde{\gg}}(\lambda))$.
In order to do so we can restrict to the open subset $\tg^{*,{\mathrm{rs}}} \subset \tg^*$ (the complement of the coroot hyperplanes). Note that we have $\nu^{-1}(\tg^{*,{\mathrm{rs}}})=\widetilde{\gg}^{\mathrm{rs}}$. To compute the action we will use another, more elementary section of the restriction of $\nu$ to $\widetilde{\gg}^{\mathrm{rs}}$. Namely, we set
\[
\Sigma^{\mathrm{rs}} := \{1 {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}\} \times \varkappa(\tg^{\mathrm{rs}}) \subset \widetilde{\gg}^{\mathrm{rs}} \subset {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}/{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}} \times \gg^*,
\]
where $\tg^{\mathrm{rs}} := \gg^{\mathrm{rs}} \cap \tg$.
Clearly, $\nu$ restricts to an isomorphism $\Sigma^{\mathrm{rs}} \xrightarrow{\sim} \tg^{*,{\mathrm{rs}}}$. Hence, if $\widetilde{\SC}^{\mathrm{rs}} := \widetilde{\gg}^{\mathrm{rs}} \cap \widetilde{\SC}$, we have canonical identifications
$\widetilde{\SC}^{\mathrm{rs}} \xrightarrow{\sim} \tg^{*,{\mathrm{rs}}} \xleftarrow{\sim} \Sigma^{\mathrm{rs}}$.
In fact, since the restriction of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}$ to $\widetilde{\gg}^{\mathrm{reg}}$ is the pullback of the group scheme $\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}$ on $\tg^*$, we also obtain a canonical identification of the corresponding restrictions of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}$, and then of their Lie algebras:
\begin{equation}
\label{eqn:Lie-tIB-sections}
\vcenter{
\xymatrix@R=0.5cm{
\LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathrm{rs}}) \ar[r]^-{\sim} \ar[d] & \LM \mathrm{ie}(\widetilde{{\mathbf J}} \def\jb{{\mathbf j}} \def\JC{{\mathcal{J}}}^{\mathrm{rs}}) \ar[d] & \LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_{\Sigma}^{\mathrm{rs}}) \ar[l]_-{\sim} \ar[d] \\
\widetilde{\SC}^{\mathrm{rs}} \ar[r]^-{\sim} & \tg^{*,{\mathrm{rs}}} & \Sigma^{\mathrm{rs}}. \ar[l]_-{\sim}
}
}
\end{equation}
(Here, for ${\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}}$ a smooth group scheme over a scheme $X$, $\LM \mathrm{ie}({\mathbf A}} \def\ab{{\mathbf a}} \def\AC{{\mathcal{A}})$ denotes the vector bundle whose sections are the Lie algebra of $A$, considered as an $\OC_X$-module; the superscript ``${\mathrm{rs}}$'' means restriction to the regular semisimple locus, and $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^{\mathrm{rs}}_\Sigma$ is the restriction of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}$ to $\Sigma^{\mathrm{rs}}$.)
Using the notation we have just introduced,
Theorem~\ref{thm:Lie-centralizer} defines an isomorphism $\LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathrm{rs}}) \xrightarrow{\sim} \widetilde{\SC}^{\mathrm{rs}} \times_{\SC^{\mathrm{rs}} }\mathsf{T}^*(\SC^{\mathrm{rs}})$. In fact, since the morphism $\widetilde{\SC}^{\mathrm{rs}} \rightarrow \SC^{\mathrm{rs}}$ is {\'e}tale (because the restriction of $\varrho \colon \tg^* \rightarrow \tg^*/W$ to $\tg^{*,{\mathrm{rs}}}$ is {\'e}tale, see~\cite[Proof of Lemma~3.5.3]{riche3}), this can be rewritten as an isomorphism $\LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathrm{rs}}) \xrightarrow{\sim} \mathsf{T}^*(\widetilde{\SC}^{\mathrm{rs}})$. Now we clearly have $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\Sigma^{\mathrm{rs}} = {\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times \Sigma^{\mathrm{rs}} \subset {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Sigma^{\mathrm{rs}}$, hence $\LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\Sigma^{\mathrm{rs}}) = \tg \times \Sigma^{\mathrm{rs}} \xrightarrow{\sim} \mathsf{T}^*(\Sigma^{\mathrm{rs}})$, where we identify $\tg$ with $\varkappa(\tg)^*=\gg / (\ng^+ \oplus \ng)$ in the natural way. We claim that the following diagram commutes, where the horizontal isomorphisms are induced by the identifications in~\eqref{eqn:Lie-tIB-sections} and the vertical isomorphisms are the ones we have just defined:
\begin{equation}
\label{eqn:Lie-tIB-sections-2}
\vcenter{
\xymatrix@R=0.5cm{
\LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathrm{rs}}) \ar[r]^-{\sim} \ar[d]_-{\wr} & \LM \mathrm{ie}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_{\Sigma}^{\mathrm{rs}}) \ar[d]^-{\wr} \\
\mathsf{T}^*(\widetilde{\SC}^{\mathrm{rs}}) \ar[r]^-{\sim} & \mathsf{T}^*(\Sigma^{\mathrm{rs}}).
}
}
\end{equation}
In fact, let $y \in \varkappa(\tg^{\mathrm{rs}})$, and let $\tilde{z}$ the point of $\widetilde{\SC}^{\mathrm{rs}}$ corresponding to $\tilde{y}=(1 {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}},y) \in \Sigma^{\mathrm{rs}}$ under the identification in~\eqref{eqn:Lie-tIB-sections}. Let us also fix $g \in {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ such that $\tilde{z}=g \cdot \tilde{y}$. Then, unravelling the various definitions (see in particular~\cite[Remark~3.3.4]{riche3}), we see that the fiber over $\tilde{z} \leftrightarrow \tilde{y}$ of~\eqref{eqn:Lie-tIB-sections-2} can be described as
\[
\xymatrix@R=0.5cm{
\gg_{g \cdot y} \ar[rrrrr]_-{\mathrm{ad}_{g^{-1}}}^-{\sim} \ar@{^{(}->}[d] &&&&& \gg_y \ar@{=}[d] \\
\gg \ar@{->>}[d] &&&&& \tg \ar[d]^-{\wr} \\
\varkappa(\sg)^* & \gg \ar@{->>}[l] & [\gg,g \cdot y]^\bot \ar@{_{(}->}[l] \ar[r]^-{\mathrm{ad}_{g^{-1}}}_-{\sim} \ar@/_15pt/@{-->}[ll]_-{\sim} & [\gg, y]^\bot \ar@{^{(}->}[r] \ar@/^15pt/@{-->}[rr]^-{\sim} & \gg \ar@{->>}[r] & \varkappa(\tg)^*.
}
\]
Now we have $\gg_{g \cdot y} = [\gg, g \cdot y]^\bot$ and $\gg_y=[\gg, y]^\bot$, so the commutativity is obvious.
With this comparison at hand, we can now identify the restriction of $\kappa(\OC_{\widetilde{\gg}}(\lambda))$ to $\tg^{*,{\mathrm{rs}}}$ with $\OC_{\widetilde{\gg}}(\lambda)_{| \Sigma^{\mathrm{rs}}}$, endowed with the action of ${\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}_\Sigma^{\mathrm{rs}}$, identified with $\tg \otimes \OC_{\Sigma^{\mathrm{rs}}}$ as above. But the restriction of $\OC_{\widetilde{\gg}}(\lambda)$ to $\{1 {\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}\} \times (\gg/\ng)^*$ is $\OC_{(\gg/\ng)^*} \otimes_\FM \FM_\lambda$ (where $\FM_\lambda$ is the one-dimensional ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$-module defined by $\lambda$), and we finally deduce that~\eqref{eqn:kappa-line-bundle} defines an isomorphism $\kappa(\OC_{\widetilde{\gg}}(\lambda)) \cong F_\lambda \langle \lambda({\check \lambda}_\circ) \rangle$, as desired.
\end{proof}
\subsection{Kostant--Whittaker reduction and geometric actions}
\label{ss:KW-braid-group}
The main result of this subsection is the following.
\begin{prop}
\label{prop:KW-actions}
\begin{enumerate}
\item
\label{it:KW-actions-F}
For any $b \in \BM_{\mathrm{aff}}$, there exists an isomorphism of functors
\[
{\EuScript K}^\FM_b \circ \kappa_\FM \cong \kappa_\FM \circ {\EuScript I}^\FM_b.
\]
\item
\label{it:KW-actions-R}
Let $b \in \BM_{\mathrm{aff}}$, and let $\FC$ be an object of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ such that $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC)$ is concentrated in degree $0$, and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free. Then $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ {\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_b(\FC)$ is concentrated in degree $0$ and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, and moreover there exists an isomorphism
\[
{\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_b \circ \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC) \cong \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ {\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_b(\FC)
\]
in $D^{\Gm}(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^*/W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\end{enumerate}
\end{prop}
\begin{remark}
It is probably true that there exists an isomorphism of functors ${\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_b \circ \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ {\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_b$. The weaker statement in Proposition~\ref{prop:KW-actions}\eqref{it:KW-actions-R} will be sufficient for our purposes.
\end{remark}
Before proving this result we need a preliminary lemma. In this lemma we fix a finite simple reflection $s$, and consider the morphism $Z_s^\EM \rightarrow \widetilde{\gg}_\EM$ induced by the second projection.
\begin{lem}
\label{lem:fiber-product-Zs}
\begin{enumerate}
\item
The Cartesian square
\[
\xymatrix@R=0.5cm{
(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_\EM \ar@{^{(}->}[r] \ar[d] & Z_s^\EM \ar[d] \\
\widetilde{\SC}_\EM \ar@{^{(}->}[r] & \widetilde{\gg}_\EM
}
\]
is tor-independent in the sense of~\cite[Definition~3.10.2]{lipman}.
\label{it:square-tor-indep}
\item
The embedding $Z_s^\EM \hookrightarrow (\widetilde{\gg} \times \widetilde{\gg})_\EM$ identifies $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_\EM$ with a closed subscheme of $(\widetilde{\SC} \times \widetilde{\SC})_\EM$. Then,
identifying $\widetilde{\SC}_\EM$ with $\tg^*_\EM$ via the isomorphism of Proposition~{\rm \ref{prop:Kostant-section}}, $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_\EM$ identifies with the graph of the action of $s$ on $\tg^*_\EM$.
\label{it:Zs-graph}
\end{enumerate}
\end{lem}
\begin{proof}
\eqref{it:square-tor-indep}
We decompose our Cartesian square into two squares:
\[
\xymatrix@R=0.5cm{
\bigl(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s\bigr)_\EM \ar@{^{(}->}[r] \ar[d] & {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \bigl( \widetilde{\SC} \times_{\widetilde{\gg}} Z_s \bigr)_\EM \ar[r] \ar[d] & Z_s^\EM \ar[d] \\
\widetilde{\SC}_\EM \ar@{^{(}->}[r] & ({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \widetilde{\SC})_\EM \ar[r]^-{a} & \widetilde{\gg}_\EM.
}
\]
Here the left arrows are induced by the embedding $\Spec(\EM) \hookrightarrow {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ given by the identity. The left-hand hand square is obviously Cartesian and tor-independent (since ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$ is flat over $\Spec(\EM)$), and the right-hand square is Cartesian by ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\EM$-equivariance, and tor-independent since $a$ is flat (see Lemma~\ref{lem:a-smooth}). The claim follows.
\eqref{it:Zs-graph}
The first claim is a consequence of the definition of $\widetilde{\SC}_\EM$ and of the fact that $Z_s^\EM$ is included in $(\widetilde{\gg} \times_{\gg^*} \widetilde{\gg})_\EM$. Now we consider the second claim.
In the case $\EM=\FM$, the claim follows from the definition of $Z_s^\FM$, using the facts that $\widetilde{\SC}_\FM$ is included in $\widetilde{\gg}_\FM^{\mathrm{reg}}$ (see~\cite[Equation~(3.1.1)]{riche3}), and that the morphism $\nu_{\mathrm{reg}}$ is $W$-equivariant (see~\S\ref{ss:notation-KW}).
Now, let us deduce the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. It follows in particular from the first claim that $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is an affine scheme. Since ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \times (\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is flat over $Z_s^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (see the proof of~\eqref{it:square-tor-indep}), which is itself flat over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and since ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is a direct summand in $\OC({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$, the scheme $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is also flat over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
Let us denote by $f$ the morphism $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ induced by the first projection. We claim that $f$ is an isomorphism. In fact, consider the morphism
\[
f^* \colon \OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \OC(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
Since $f$ is projective, the right-hand side is finite over $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. Moreover, the morphism $\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (f^*)$ is an isomorphism for any geometric point $\FM$ of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, by the case of fields treated first. Using~\cite[Lemma~1.4.1]{br} we deduce that $f^*$ is an isomorphism, which finishes the proof of the claim.
Now, consider the morphism $\tau_s \colon \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ defined as the composition of the inverse of $f$ with the natural projection $(\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
The induced morphism $\widetilde{\SC}_\CM \rightarrow \widetilde{\SC}_\CM$ coincides with the action of $s$ (via the identification $\widetilde{\SC}_\CM \xrightarrow{\sim} \tg^*_\CM$). It follows that the morphism $\tau_s$ itself coincides with the action of $s$, finishing the proof.
\end{proof}
\begin{proof}[Proof of Proposition~{\rm \ref{prop:KW-actions}}]
In each case, it is sufficient to prove the claim when $b=T_s$ for $s$ a finite simple reflection, or when $b=\theta_\lambda$ for some $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
First, assume that $b=\theta_\lambda$. Since $\kappa_\EM$ is compatible with tensoring with a line bundle, Proposition~\ref{prop:KW-line-bundle} implies that there exists an isomorphism of functors ${\EuScript K}^\EM_{\theta_\lambda} \circ \kappa_\EM \cong \kappa_\EM \circ {\EuScript I}^\EM_{\theta_\lambda}$, which proves the claims in~\eqref{it:KW-actions-F} and~\eqref{it:KW-actions-R} in this case.
Now we consider the case $b=T_s$. In this case we have ${\EuScript I}^\EM_{T_s} \cong \mathsf{R} (p_s)_* \circ \mathsf{L} (q_s)^* \langle -1 \rangle$, where $p_s, q_s \colon Z_s^\EM \rightarrow \widetilde{\gg}_\EM$ are induced by the first and second projections, respectively. Let $\kappa'_\EM$ be the composition of $\kappa_\EM$ with the direct image under the (affine) morphism $\varsigma \colon (\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^*/W))_\EM \rightarrow \tg^*_\EM$. (In other words, $\kappa'_\EM$ is the composition of restriction to $\widetilde{\SC}_\EM$ with the functor $(\nu_\SC)_*$.)
By Lemma~\ref{lem:fiber-product-Zs}\eqref{it:square-tor-indep} and the base-change theorem (see~\cite[Theorem~3.10.3]{lipman}), we have
$\kappa'_\EM \circ {\EuScript I}^\EM_{T_s} \cong \mathsf{R} (p'_s)_* \circ \mathsf{L} (q'_s)^* \circ \kappa'_\EM \langle -1 \rangle$, where
$p_s', q_s' \colon (\widetilde{\SC} \times_{\widetilde{\gg}} Z_s)_\EM \rightarrow \tg_\EM^*$
are the compositions of $\nu_\SC$ with the morphisms obtained by restriction from $p_s, q_s$. Using Lemma~\ref{lem:fiber-product-Zs}\eqref{it:Zs-graph}, we deduce a canonical isomorphism
\begin{equation}
\label{eqn:isom-KW-action-Ts}
\kappa'_\EM \circ {\EuScript I}^\EM_{T_s} \cong (\tau'_s)^* \circ \kappa'_\EM \langle -1 \rangle,
\end{equation}
where $\tau'_s \colon \tg^*_\FM \rightarrow \tg^*_\FM$ is the action of $s$.
In the case $\EM=\FM$, since the automorphism $\tau_s \colon \widetilde{\SC}_\FM \xrightarrow{\sim} \widetilde{\SC}_\FM$ is the restriction of a ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM \times \GmF$-equivariant automorphism of $\widetilde{\gg}_\FM^{\mathrm{reg}}$, isomorphism~\eqref{eqn:isom-KW-action-Ts} is induced by an isomorphism of functors
\[
{\EuScript K}^\FM_{T_s} \circ \kappa_\FM \cong \kappa_\FM \circ {\EuScript I}^\FM_{T_s},
\]
which finishes the proof of~\eqref{it:KW-actions-F}. Then the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ follows from the case $\EM=\CM$: indeed by~\eqref{eqn:isom-KW-action-Ts} we have an isomorphism
\begin{equation}
\label{eqn:isom-KW-action-Ts-R}
\varsigma_* \bigl( {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{T_s} \circ \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC) \bigr) \cong \varsigma_* \bigl( \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ {\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{T_s}(\FC) \bigr).
\end{equation}
Hence if $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (\FC)$ is concentrated in degree $0$ and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, the same is true for $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ {\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{T_s}(\FC)$. And in this case, since the image of~\eqref{eqn:isom-KW-action-Ts-R} under $\CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} (-)$ is $\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^*/W))_\CM$-linear, we deduce that~\eqref{eqn:isom-KW-action-Ts-R} is $\OC(\tg^* \times_{\tg^* / W} \mathsf{T}(\tg^*/W))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-linear, which finishes the proof of~\eqref{it:KW-actions-R}.
\end{proof}
We finish this section with the following variant of Lemma~\ref{lem:fiber-product-Zs}, to be used later. In this lemma we consider the morphism $(\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg})_\EM \rightarrow \widetilde{\gg}_\EM$ induced by the second projection.
\begin{lem}
\label{lem:fiber-product-xis}
\begin{enumerate}
\item
\label{it:square-tor-indep-xi}
The Cartesian square
\[
\xymatrix@R=0.5cm{
\bigl( \widetilde{\SC} \times_{\widetilde{\gg}} (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}) \bigr)_\EM \ar@{^{(}->}[r] \ar[d] & (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg})_\EM \ar[d] \\
\widetilde{\SC}_\EM \ar@{^{(}->}[r] & \widetilde{\gg}_\EM
}
\]
is tor-independent in the sense of~\cite[Definition~3.10.2]{lipman}.
\item
\label{it:xis-graph}
Identifying $\widetilde{\SC}_\EM$ with $\tg^*_\EM$ via the isomorphism of Proposition~{\rm \ref{prop:Kostant-section}}, the fiber product $\bigl( \widetilde{\SC} \times_{\widetilde{\gg}} (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}) \bigr)_\EM$ identifies with the closed subscheme of $(\tg^* \times \tg^*)_\EM$ given by $(\tg^* \times_{\tg^* / W_s} \tg^*)_\EM$.
\end{enumerate}
\end{lem}
\begin{proof}
The proof of~\eqref{it:square-tor-indep-xi} is identical to the proof of Lemma~\ref{lem:fiber-product-Zs}\eqref{it:square-tor-indep}. The case $\EM=\FM$ of~\eqref{it:xis-graph} follows from the observation that the intersection of $(\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg})_\FM$ with $\widetilde{\gg}_\FM^{\mathrm{reg}} \times \widetilde{\gg}_\FM^{\mathrm{reg}}$ is the union of the diagonal copy of $\widetilde{\gg}_\FM^{\mathrm{reg}}$ and the graph of the action of $s$. Now, let us deduce the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
By the same arguments as in the proof of Lemma~\ref{lem:fiber-product-Zs}\eqref{it:Zs-graph}, the fiber product $\bigl( \widetilde{\SC} \times_{\widetilde{\gg}} (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is a flat ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-scheme, and a closed subscheme of the affine scheme $(\widetilde{\SC} \times \widetilde{\SC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Consider the algebra
\[
A:= \OC \bigl( \widetilde{\SC} \times_{\widetilde{\gg}} (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
We claim that $A$ is ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free. Indeed, consider the contracting $\GmR$-action on $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ considered in~\S\ref{ss:reminder}. The diagonal action on $(\widetilde{\SC} \times \widetilde{\SC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ stabilizes the subscheme $(\widetilde{\SC} \times_{\widetilde{\gg}} (\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg}))_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, so that $A$ is endowed with a $\ZM$-grading. Each graded piece of $A$ is finite over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and flat, hence free, which proves our claim.
Now, consider the surjection
\[
\OC(\tg^* \times \tg^*)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \twoheadrightarrow A.
\]
For any $x \in \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})^s$, the image of $x \otimes 1 - 1 \otimes x$ in $A$ becomes zero in $\CM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} A$,
by the case $\EM=\CM$. Since $A$ is ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, this implies that this image is $0$, hence that our morphism factors through a surjection
\[
\bigl( \OC(\tg^*) \otimes_{\OC(\tg^*)^s} \OC(\tg^*) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \twoheadrightarrow A.
\]
This surjection is compatible with the $\ZM$-gradings, and on both sides the graded pieces are free of finite rank over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. (In fact, this property has been proved above for $A$. For the left-hand side, we have $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})=\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})^s \oplus \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})^s \cdot \delta$, where $\delta \in \tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is any element such that $\langle \delta, \alpha \rangle=1$ -- see e.g.~\cite[Claim~3.9]{ew} --
which implies our claim.) Hence to conclude we only have to prove that the induced morphism
\[
\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigl( \OC(\tg^*) \otimes_{\OC(\tg^*)^s} \OC(\tg^*) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \twoheadrightarrow \FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} A
\]
is an isomorphism for all geometric points $\FM$ of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
Using again the case $\EM=\FM$ treated above, we only have to prove that the natural morphism
\[
\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigl( \OC(\tg^*) \otimes_{\OC(\tg^*)^s} \OC(\tg^*) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \bigl( \OC(\tg^*) \otimes_{\OC(\tg^*)^s} \OC(\tg^*) \bigr)_\FM
\]
is an isomorphism. In turn, this follows easily from the decompositions $\OC(\tg^*_\EM)=\OC(\tg^*_\EM)^s \oplus \OC(\tg^*_\EM)^s \cdot \delta$ for $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and $\FM$ (where $\delta$ is as above, and we denote similarly its image in $\tg_\FM$).
\end{proof}
\section{Tilting exotic sheaves}
\label{sec:tilting-KW}
In this section we use the same notation as in Section~\ref{sec:KW-reduction}.
\subsection{Overview}
\label{ss:overview-exotic}
In this section we give a description of the category of tilting objects in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ in terms of ``Soergel bimodules.'' This description is based on a ``Bott--Samelson type'' description of these tilting objects, due to Dodd~\cite{dodd} in the case $p=0$ and generalized to the modular setting in~\cite{mr}, and on the use of the ``Kostant--Whittaker reduction'' of Section~\ref{sec:KW-reduction}.
In~\S\ref{ss:notation-KW-tilting} we review the basic definitions and results on Bezrukavnikov's exotic t-structure, and in~\S\ref{ss:standard-KW} we introduce and study some variants of the associated ``standard'' and ``costandard'' objects. In~\S\ref{ss:BS-objects-geom} we recall the ``Bott--Samelson'' description of tilting objects in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$. In~\S\ref{ss:grk-Hom-tilting} we compute the graded ranks of $\Hom$-spaces between our ``Bott--Samelson objects.'' Finally, in~\S\S\ref{ss:BS-category-tilting}--\ref{ss:equivalence-tilting} we obtain the desired description in terms of Soergel bimodules.
\subsection{Reminder on the exotic t-structure}
\label{ss:notation-KW-tilting}
In this section we will
consider the Springer resolution
\[
\widetilde{\NC}_\EM:=({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times^{{\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}} (\gg/\bg)^*)_\EM \hookrightarrow \mathscr{B}_\EM \times \gg^*_\EM,
\]
a sub-vector bundle of the Grothendieck resolution $\widetilde{\gg}_\EM$ studied in Section~\ref{sec:KW-reduction}. We denote by $i \colon \widetilde{\NC}_\EM \hookrightarrow \widetilde{\gg}_\EM$ the inclusion. For $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ we denote by $\OC_{\widetilde{\NC}_\EM}(\lambda)$ the restriction of $\OC_{\widetilde{\gg}_\EM}(\lambda)$ to $\widetilde{\NC}_\EM$.
We will consider the derived categories of equivariant coherent sheaves
$D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM$ and $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM$.
By~\cite[Section~1]{br}, the geometric braid group actions considered in~\S\ref{ss:braid-groth} ``restrict'' to the categories $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM$ and $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM$ in the following sense. For $s$ a finite simple reflection, associated with a simple root $\alpha$, we define $Z_s^{\prime \EM}:= Z_s^\EM \cap (\widetilde{\NC}_\EM \times \widetilde{\NC}_\EM)$, and denote by
\[
\TM^s_{\widetilde{\NC}}, \, \SM^s_{\widetilde{\NC}} \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM
\]
the Fourier--Mukai transforms with kernels $\OC_{Z_s^{\prime \EM}} \langle -1 \rangle$ and $\OC_{Z_s^{\prime \EM}}(-\rho, \rho-\alpha) \langle -1 \rangle$ respectively. We use the same symbols for the analogous endofunctors of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM$. (Here $\OC_{Z_s^{\prime \EM}}(-\rho, \rho-\alpha)$ is defined by the same recipe as for $\OC_{Z_s^{\EM}}(-\rho, \rho-\alpha)$ in~\S\ref{ss:braid-groth}.) Then $\TM^s_{\widetilde{\NC}}$ and $\SM^s_{\widetilde{\NC}}$ are quasi-inverse equivalences of categories, and there exists a right action of the group $\BM_{\mathrm{aff}}$ on the categories $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM$ and $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM$ such that $T_s$ acts by $\TM^s_{\widetilde{\NC}}$ for any finite simple reflection $s$, and $\theta_\lambda$ acts by tensoring with $\OC_{\widetilde{\NC}_\EM}(\lambda)$ for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
For $b \in \BM_{\mathrm{aff}}$, we denote by
\[
{\EuScript J}^\EM_b \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM \xrightarrow{\sim} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\EM
\]
the action of $b$. Then there exist isomorphisms of functors
\begin{equation}
\label{eqn:action-i*}
{\EuScript I}_b^\EM \circ \mathsf{R} i_* \cong \mathsf{R} i_* \circ {\EuScript J}_b^\EM, \qquad \mathsf{L} i^* \circ {\EuScript I}_b^\EM \cong {\EuScript J}_b^\EM \circ \mathsf{L} i^*
\end{equation}
and
\begin{equation}
\label{eqn:action-modular-reduction}
\FM \circ {\EuScript J}_b^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong {\EuScript J}_b^\FM \circ \FM.
\end{equation}
Recall the elements $w_\lambda \in W_{\mathrm{aff}}$ defined in~\S\ref{ss:Haff}. We set
\[
\nabla^\lambda_{\widetilde{\NC},\EM} := {\EuScript J}^\EM_{T_{w_\lambda}}(\OC_{\widetilde{\NC}_\EM}), \qquad
\Delta^\lambda_{\widetilde{\NC},\EM} := {\EuScript J}^\EM_{(T_{w_\lambda^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\EM}).
\]
By~\eqref{eqn:action-modular-reduction}, these objects satisfy
\begin{equation}
\label{eqn:F-D-N-tNC}
\FM(\nabla^\lambda_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong \nabla^\lambda_{\widetilde{\NC},\FM}, \quad \FM(\Delta^\lambda_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong \Delta^\lambda_{\widetilde{\NC},\FM}.
\end{equation}
\begin{remark}
\label{rk:nabla-delta-Baff}
For any $w \in W$ we have ${\EuScript J}^\EM_{T_w}(\OC_{\widetilde{\NC}_\EM}) \cong \OC_{\widetilde{\NC}_\EM} \langle -\ell(w) \rangle$. (See~\cite[Equation~(3.10)]{mr} for the case $\EM=\FM$; the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ can be deduced using the arguments in the proof of~\cite[Proposition~1.4.3]{br}, or proved along the same lines.) Therefore, as in~\cite{mr}, for any $w \in Wt_\lambda$ we have
\[
\nabla^\lambda_{\widetilde{\NC},\EM} \cong {\EuScript J}^\EM_{T_w}(\OC_{\widetilde{\NC}_\EM}) \langle -\ell(w_\lambda) + \ell(w) \rangle, \quad
\Delta^\lambda_{\widetilde{\NC},\EM} \cong {\EuScript J}^\EM_{(T_{w^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\EM}) \langle - \ell(w) + \ell(w_\lambda) \rangle.
\]
\end{remark}
In the case $\EM=\FM$,
the objects $\nabla^\lambda_{\widetilde{\NC},\FM}$ and $\Delta^\lambda_{\widetilde{\NC},\FM}$ were studied in~\cite{mr} (see in particular~\cite[Proposition~3.7]{mr}). In fact, if we denote by $D^{\leq 0}$, resp.~$D^{\geq 0}$, the subcategory of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ generated under extensions by the objects $\Delta^\lambda_{\widetilde{\NC},\FM} \langle n \rangle [m]$ with $n \in \ZM$ and $m \in \ZM_{\geq 0}$, resp.~by the objects $\nabla^\lambda_{\widetilde{\NC},\FM} \langle n \rangle [m]$ with $n \in \ZM$ and $m \in \ZM_{\leq 0}$, then the pair $(D^{\leq 0}, D^{\geq 0})$ constitutes a bounded t-structure on $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$, called the \emph{exotic t-structure}. We denote by ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ the heart of this t-structure.
By~\cite[Corollary~3.10]{mr}, the objects $\Delta^\lambda_{\widetilde{\NC},\FM}$ and $\nabla^\lambda_{\widetilde{\NC},\FM}$ belong to ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$. Let us fix an order $\leq'$ on ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ as in~\cite[\S 2.5]{mr}. Then by~\cite[\S 3.5]{mr}, the category ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ is a graded highest weight category with weight poset $({\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}, \leq')$, standard objects $\{\Delta^\lambda_{\widetilde{\NC},\FM}, \, \lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}\}$ and costandard objects $\{\nabla^\lambda_{\widetilde{\NC}, \FM}, \, \lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}\}$. In particular, it makes sense to consider the \emph{tilting objects} in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$, i.e.~the objects which admit both a standard filtration (i.e.~a filtration with subquotients of the form $\Delta^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle$ with $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$),
and a costandard filtration (i.e.~a filtration with subquotients of the form $\nabla^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle$ with $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$).
The subcategory consisting of such objects will be denoted by $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM)$.
If $X$ admits a standard filtration, resp.~a costandard filtration, we denote by $(X : \Delta^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle)$, resp.~$(X : \nabla^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle)$, the number of times $\Delta^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle$, resp.~$\nabla^\lambda_{\widetilde{\NC}, \FM} \langle m \rangle$, appears in a standard, resp.~costandard, filtration of $X$. (This number does not depend on the filtration.) The general theory of graded highest weight categories implies that for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ there exists a unique (up to isomorphism) indecomposable object $\TC^\lambda$ in $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM)$ which satisfies
\[
(\TC^\lambda : \Delta_{\widetilde{\NC}, \FM}^\lambda)=1 \quad \text{and} \qquad (\TC^\lambda : \Delta^\mu_{\widetilde{\NC}, \FM} \langle m \rangle) \neq 0 \Rightarrow \mu \leq' \lambda.
\]
Moreover, every object in $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM)$ is a direct sum of objects of the form $\TC^\lambda \langle m \rangle$ for some $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$. (See e.g.~\cite[Appendix~A]{modrap2} for references on this subject.)
Recall also from~\cite[Equation~(2.3)]{mr} that we have
\begin{equation}
\label{eqn:Hom-Delta-nabla}
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM}(\Delta^\lambda_{\widetilde{\NC},\FM}, \nabla^\mu_{\widetilde{\NC},\FM} \langle n \rangle [m]) =
\begin{cases}
\FM & \text{if $\lambda=\mu$ and $n=m=0$;} \\
0 & \text{otherwise.}
\end{cases}
\end{equation}
\begin{remark}
One can easily check that none of our constructions depends on the choice of the order $\leq'$.
\end{remark}
\subsection{More ``standard'' and ``costandard'' objects}
\label{ss:standard-KW}
If ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$ is an affine group scheme over a ring $k$, we denote by $\mathsf{Inv}^{\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}_k$ the functor of derived ${\mathbf H}} \def\hb{{\mathbf h}} \def\HC{{\mathcal{H}}$-invariants; see~\cite[\S A.3]{mr}.
\begin{lem}
\label{lem:morphisms-fg}
For any $\FC,\GC$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$,
the complex of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules
\[
R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\FC,\GC)
\]
is bounded, and has finitely generated cohomology modules.
\end{lem}
\begin{proof}
By~\cite[Proposition~A.6]{mr},
we have a natural isomorphism
\[
R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\FC,\GC) \cong \mathsf{Inv}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ \mathsf{Inv}^{\GmR}_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \bigl( R\Hom_{D^{\mathrm{b}} \Coh(\widetilde{\NC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\FC,\GC) \bigr).
\]
Since the natural morphism $\widetilde{\NC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \rightarrow \gg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (the restriction of the morphism $\pi$ from~\S\ref{ss:notation-KW}) is projective, by~\cite[Theorem~III.8.8]{hartshorne} $R\Hom_{D^{\mathrm{b}} \Coh(\widetilde{\NC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\FC,\GC)$ is a bounded complex of finitely generated $\OC(\gg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-modules, which implies that the complex of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules $\mathsf{Inv}^{\GmR}_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} ( R\Hom_{D^{\mathrm{b}} \Coh(\widetilde{\NC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\FC,\GC) )$ is bounded, and has finitely generated cohomology modules. Hence the claim follows from the fact that if $M$ is a ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-module which is finitely generated over ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, then $\mathsf{Inv}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(M)$ is a bounded complex of finitely generated ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-modules, see~\cite[Lemma~II.B.5 and its proof]{jantzen}.
\end{proof}
\begin{prop}
\label{prop:morphism-D-N-tNC}
For $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $n,m \in \ZM$
we have
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\Delta^\lambda_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}, \nabla^{\mu}_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \langle n \rangle [m]) = \begin{cases}
{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} & \text{if $\lambda=\mu$ and $n=m=0$;} \\
0 & \text{otherwise.}
\end{cases}
\]
Moreover, the natural morphism
\[
\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\Delta^\lambda_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}, \nabla^{\mu}_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \langle n \rangle [m])
\rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM}(\Delta^\lambda_{\widetilde{\NC},\FM}, \nabla^{\mu}_{\widetilde{\NC},\FM} \langle n \rangle [m])
\]
is an isomorphism.
\end{prop}
\begin{proof}
By the same arguments as in~\cite[Proof of Lemma~4.11]{mr}, there exists a canonical isomorphism of complexes of $\FM$-vector spaces
\[
\FM \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\Delta^\lambda_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}, \nabla^{\mu}_{\widetilde{\NC},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \langle n \rangle)
\xrightarrow{\sim} R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM}(\Delta^\lambda_{\widetilde{\NC},\FM}, \nabla^{\mu}_{\widetilde{\NC},\FM} \langle n \rangle).
\]
By~\eqref{eqn:Hom-Delta-nabla}, the right-hand side is isomorphic to $\FM$ if $\lambda=\mu$ and $n=0$, and is $0$ otherwise. Using Lemma~\ref{lem:morphisms-fg}, we deduce the claim.
\end{proof}
As in~\cite[\S 4.1]{mr}, for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, we set
\[
\nabla^\lambda_{\widetilde{\gg},\EM} := {\EuScript I}^\EM_{T_{w_\lambda}}(\OC_{\widetilde{\gg}_\EM}), \qquad
\Delta^\lambda_{\widetilde{\gg},\EM} := {\EuScript I}^\EM_{(T_{w_\lambda^{-1}})^{-1}}(\OC_{\widetilde{\gg}_\EM}).
\]
Then by~\eqref{eqn:F-IS} we have
\[
\FM(\nabla^\lambda_{\widetilde{\gg},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong \nabla^\lambda_{\widetilde{\gg},\FM}, \qquad \FM(\Delta^\lambda_{\widetilde{\gg},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}) \cong \Delta^\lambda_{\widetilde{\gg},\FM},
\]
and by~\eqref{eqn:action-i*} we have
\begin{equation}
\label{eqn:delta-nabla-i^*}
\mathsf{L} i^*(\nabla^\lambda_{\widetilde{\gg},\EM}) \cong \nabla^\lambda_{\widetilde{\NC},\EM}, \qquad \mathsf{L} i^*(\Delta^\lambda_{\widetilde{\gg},\EM}) \cong \Delta^\lambda_{\widetilde{\NC},\EM}.
\end{equation}
Recall that for $\FC,\GC$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM$ the morphism
\[
\bigoplus_{n \in \ZM} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM}(\FC, \GC \langle -n \rangle) \rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\FC,\GC)
\]
induced by the forgetful functor
is an isomorphism. In particular, this isomorphism endows the right-hand side with a natural $\ZM$-grading.
\begin{prop}
\label{prop:morphism-D-N-tgg}
For $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$ there exists an isomorphism of graded $\OC(\tg^*_\EM)$-modules
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\Delta^\lambda_{\widetilde{\gg},\EM}, \nabla^{\mu}_{\widetilde{\gg},\EM} [m]) = \begin{cases}
\OC(\tg^*_\EM) & \text{if $\lambda=\mu$ and $m=0$;} \\
0 & \text{otherwise.}
\end{cases}
\]
Moreover, the natural morphisms
\[
\EM \otimes_{\OC(\tg^*_\EM)} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\Delta^\lambda_{\widetilde{\gg},\EM}, \nabla^{\mu}_{\widetilde{\gg},\EM} [m])
\rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM}(\Delta^\lambda_{\widetilde{\NC},\EM}, \nabla^{\mu}_{\widetilde{\NC},\EM} [m])
\]
(induced by $\mathsf{L} i^*$ via isomorphisms~\eqref{eqn:delta-nabla-i^*})
and
\[
\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\Delta^\lambda_{\widetilde{\gg},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}, \nabla^{\mu}_{\widetilde{\gg},{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} [m])
\rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\FM}(\Delta^\lambda_{\widetilde{\gg},\FM}, \nabla^{\mu}_{\widetilde{\gg},\FM} [m])
\]
are isomorphisms.
\end{prop}
\begin{proof}
As in~\cite[Lemma~4.11]{mr},
for $\FC,\GC$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM$
we have a canonical isomorphism
\[
\EM \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{\OC(\tg_\EM^*)} R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\FC,\GC) \cong R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM}(\mathsf{L} i^* \FC, \mathsf{L} i^* \GC).
\]
Choosing $\FC=\Delta^\lambda_{\widetilde{\gg},\EM}$ and $\GC=\nabla^{\mu}_{\widetilde{\gg},\EM}$ and using~\eqref{eqn:delta-nabla-i^*}, we deduce an isomorphism
\[
\EM \@ifnextchar_{\@lotimessub}{\@lotimesnosub}_{\OC(\tg_\EM^*)} R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\Delta^\lambda_{\widetilde{\gg},\EM}, \nabla^{\mu}_{\widetilde{\gg},\EM}) \cong R\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM}(\Delta^\lambda_{\widetilde{\NC},\EM}, \nabla^{\mu}_{\widetilde{\NC},\EM}).
\]
Using Proposition~\ref{prop:morphism-D-N-tNC}, the right-hand side is concentrated in degree $0$, and isomorphic either to $\EM$ (when $\lambda=\mu$) or to $0$ (when $\lambda \neq \mu$). By Lemma~\ref{lem:nakayama}\eqref{it:nakayama-complex}, this implies the first claim, and the first isomorphism.
The proof of the second isomorphism is similar to the proof of the corresponding claim in Proposition~\ref{prop:morphism-D-N-tNC}.
\end{proof}
\subsection{``Coherent'' Bott--Samelson objects}
\label{ss:BS-objects-geom}
As in~\cite[\S 4.2]{mr},
if $s$ is a finite simple reflection,
we denote by
\[
\Xi_s^\EM \colon D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM
\]
the Fourier--Mukai transform associated with the kernel $\OC_{(\widetilde{\gg} \times_{\widetilde{\gg}_s} \widetilde{\gg})_\EM} \langle -1 \rangle$ (see~\S\ref{ss:braid-groth} for the notation). We also make a similar definition for $s_0$ an affine simple reflection: by Lemma~\ref{lem:affine-simple-conjugate} we can choose (once and for all) a finite simple reflection $t$ and an element $b \in \BM_{\mathrm{aff}}$ such that $T_{s_0} = b T_t b^{-1}$, and set $\Xi^\EM_{s_0}:={\EuScript I}^\EM_{b^{-1}} \circ \Xi^\EM_t \circ {\EuScript I}^\EM_b$.
One can easily check that for any simple reflection $s \in W_{\mathrm{aff}}$ there exists an isomorphism
\begin{equation}
\label{eqn:F-Xi}
\FM \circ \Xi_s^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong \Xi_s^\FM \circ \FM.
\end{equation}
For $\underline{s}=(s_1, \cdots, s_n)$ a sequence of simple reflections and $\omega \in \Omega$, we set
\[
\MC_{\EM}(\omega,\underline{s}):= \Xi^\EM_{s_n} \circ \cdots \circ \Xi^\EM_{s_1} \circ {\EuScript I}^\EM_{T_\omega} (\OC_{\widetilde{\gg}_\EM}).
\]
By~\eqref{eqn:F-IS} and~\eqref{eqn:F-Xi}, these objects satisfy
\[
\FM(\MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega,\underline{s})) \cong \MC_{\FM}(\omega,\underline{s}).
\]
Our interest in the objects $\MC_{\EM}(\omega,\underline{s})$ comes from the following result, which
is proved in~\cite[Corollary~4.2]{mr}.
\begin{prop}
\label{prop:tilting-Bott-Samelson}
For any sequence $\underline{s}$ of simple reflections and $\omega \in \Omega$, the object $\mathsf{L} i^* \bigl( \MC_{\FM}(\omega,\underline{s}) \bigr)$ is in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$, and is tilting. Moreover, any indecomposable object of $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM)$ is a direct summand in an object of the form $\mathsf{L} i^* \bigl( \MC_{\FM}(\omega,\underline{s}) \bigr) \langle n \rangle$ with $(\omega, \underline{s})$ as before and $n \in \ZM$. \qed
\end{prop}
The properties of $\Hom$-spaces between objects of the form $\MC_{\EM}(\underline{s},\omega)$ are summarized in the following proposition.
\begin{prop}
\label{prop:morphisms-BS-geom}
For any sequences $\underline{s}$ and $\underline{t}$ of simple reflections, for $\omega, \omega' \in \Omega$, and for $k \in \ZM$, we have
\begin{multline*}
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\MC_{\EM}(\omega,\underline{s}), \MC_{\EM}(\omega',\underline{t}) [k]) = 0
\\
\text{and} \qquad \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM}( \mathsf{L} i^*(\MC_{\EM}(\omega,\underline{s})), \mathsf{L} i^*(\MC_{\EM}(\omega',\underline{t})) [k]) = 0
\end{multline*}
unless $k=0$. Moreover, the graded $\OC(\tg_\EM^*)$-module
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\MC_{\EM}(\omega,\underline{s}), \MC_{\EM}(\omega',\underline{t}) )
\]
is graded free, and the functor $\mathsf{L} i^*$ induces an isomorphism
\begin{multline*}
\EM \otimes_{\OC(\tg_\EM^*)} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\EM}(\MC_{\EM}(\omega,\underline{s}), \MC_{\EM}(\omega',\underline{t})) \\
\xrightarrow{\sim} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\EM}(\mathsf{L} i^*(\MC_{\EM}(\omega,\underline{s})), \mathsf{L} i^*(\MC_{\EM}(\omega',\underline{t})) ).
\end{multline*}
Finally, the functor $\FM(-)$ induces isomorphisms of graded $\FM$-vector spaces
\[
\FM \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}( \MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega,\underline{s}), \MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega',\underline{t}) )
\xrightarrow{\sim} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\FM}(\MC_{\FM}(\omega,\underline{s}), \MC_{\FM}(\omega',\underline{t}) )
\]
and
\begin{multline*}
\FM \otimes_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}( \mathsf{L} i^*(\MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega,\underline{s})), \mathsf{L} i^*(\MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega',\underline{t})) ) \\
\xrightarrow{\sim} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\FM}( \mathsf{L} i^*(\MC_{\FM}(\omega,\underline{s})), \mathsf{L} i^* (\MC_{\FM}(\omega',\underline{t})) ).
\end{multline*}
\end{prop}
\begin{proof}
By~\cite[Lemma~4.1]{mr}, the objects of the form $\MC_{\EM}(\omega,\underline{s})$ belong to the subcategory of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\EM$ generated under extensions by the objects $\nabla^\lambda_{\widetilde{\gg},\EM}\langle m \rangle$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$, and also to the subcategory generated under extensions by the objects $\Delta^\lambda_{\widetilde{\gg},\EM}\langle m \rangle$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$. (In~\cite{mr} only the case $\EM=\FM$ is considered, but the same proof applies in the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$; in fact this proof relies on the existence of the exact sequences~\eqref{eqn:ses-tgg}.) Then the claims follow from Propositions~\ref{prop:morphism-D-N-tNC} and~\ref{prop:morphism-D-N-tgg}.
\end{proof}
\subsection{Graded ranks of $\mathrm{Hom}$-spaces}
\label{ss:grk-Hom-tilting}
Recall the $\HM_{\mathrm{aff}}$-module $\MM_{\mathrm{sph}}$ of \S\ref{ss:Haff}. If $\FC$ is an object of ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ which admits a standard filtration, we define
\[
\ch_\Delta(\FC) := \sum_{\substack{\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \\ n \in \ZM}} (\FC : \Delta^\lambda_{\widetilde{\NC}} \langle n \rangle) \cdot \mathsf{v}^n \cdot \mathbf{m}_\lambda \quad \in \MM_{\mathrm{sph}}.
\]
Similarly, if $\GC$ is an object of ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ which admits a costandard filtration, we define
\[
\ch_\nabla(\GC) := \sum_{\substack{\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}} \\ n \in \ZM}} (\GC : \nabla^\lambda_{\widetilde{\NC}} \langle n \rangle) \cdot \mathsf{v}^{-n} \cdot \mathbf{m}_\lambda \quad \in \MM_{\mathrm{sph}}.
\]
One can easily check that for $\FC,\GC$ tilting objects in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ we have
\begin{equation}
\label{eqn:grk-tilting}
\grk_\FM \bigl( \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\FM}(\FC,\GC) \bigr) = \langle \ch_\Delta(\FC), \ch_\nabla(\GC) \rangle.
\end{equation}
\begin{prop}
\label{prop:ch-tilting}
For any sequence $\underline{s}$ of simple reflections and any $\omega \in \Omega$ we have
\[
\ch_\Delta(\mathsf{L} i^* \MC_\FM(\omega,\underline{s})) = \ch_\nabla(\mathsf{L} i^* \MC_\FM(\omega,\underline{s})) = \mathbf{m}(\omega,\underline{s}).
\]
\end{prop}
\begin{proof}
We only consider the case of $\ch_\Delta$; the case of $\ch_\nabla$ is similar. For any $w \in W_{\mathrm{aff}}$ we set
\[
\Delta^w_{\widetilde{\NC},\FM} := {\EuScript J}_{(T_{w^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}), \qquad \Delta^w_{\widetilde{\gg},\FM} := {\EuScript I}_{(T_{w^{-1}})^{-1}}(\OC_{\widetilde{\gg}_\FM})
\]
and
$\mathbf{m}_w:=\mathbf{m}_0 \cdot T_w \in \MM_{\mathrm{sph}}$. Then $\Delta^w_{\widetilde{\NC},\FM}$ is a standard object in ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$, and we have
\[
\ch_\Delta(\Delta^w_{\widetilde{\NC},\FM}) = \mathbf{m}_w.
\]
Indeed, the formula holds by definition if $w$ is minimal in $Ww$. For a general $w \in W_{\mathrm{aff}}$, write $w=uv$ with $u \in W$ and $v$ minimal in $Ww$. Then $T_{w^{-1}}=T_{v^{-1}} T_{u^{-1}}$, so that we have
\begin{multline*}
\Delta^w_{\widetilde{\NC},\FM} = {\EuScript J}_{(T_{w^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}) = {\EuScript J}_{(T_{u^{-1}})^{-1} \cdot (T_{v^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}) \\
= {\EuScript J}_{(T_{v^{-1}})^{-1}} \circ {\EuScript J}_{(T_{u^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}) = {\EuScript J}_{(T_{v^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}) \langle \ell(u) \rangle
\end{multline*}
(see Remark~\ref{rk:nabla-delta-Baff}),
so that
\[
\ch_\Delta(\Delta^w_{\widetilde{\NC},\FM}) = \mathsf{v}^{\ell(u)} \ch_{\Delta}(\Delta^v_{\widetilde{\NC},\FM}) = \mathsf{v}^{\ell(u)} \mathbf{m}_v=\mathbf{m}_w.
\]
Using this fact and the proof of~\cite[Lemma~4.1]{mr}, one can check that for all $w \in W_{\mathrm{aff}}$ and all simple reflections $s$, the object $\mathsf{L} i^*(\Xi_s^\FM (\Delta^w_{\widetilde{\gg},\FM}))$ belongs to ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ and admits a standard filtration, and that moreover we have
\[
\ch_\Delta \bigl( \mathsf{L} i^*(\Xi_s^\FM (\Delta^w_{\widetilde{\gg},\FM})) \bigr) = \mathbf{m}_w \cdot (T_s + \mathsf{v}^{-1}) = \ch_\Delta \bigl( \mathsf{L} i^* (\Delta^w_{\widetilde{\gg},\FM}) \bigr) \cdot (T_s + \mathsf{v}^{-1}).
\]
Then one deduces that for any object $\FC$ in the subcategory of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_\FM$ generated under extensions by the objects $\Delta^\lambda_{\widetilde{\gg},\FM} \langle n \rangle$ ($\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, $n \in \ZM$), the object $\mathsf{L} i^*(\Xi_s^\FM ( \FC))$ belongs to ${\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ and admits a standard filtration, and that moreover we have
\[
\ch_\Delta \bigl( \mathsf{L} i^*(\Xi_s^\FM ( \FC)) \bigr) = \ch_\Delta(\mathsf{L} i^*(\FC)) \cdot (T_s + \mathsf{v}^{-1}).
\]
Since, for $\omega \in \Omega$, we have
\[
\ch_\Delta \bigl( \mathsf{L} i^*({\EuScript I}^\FM_{T_\omega}(\OC_{\widetilde{\gg}_\FM})) \bigr) = \ch_\Delta \bigl( {\EuScript J}^\FM_{(T_{\omega^{-1}})^{-1}}(\OC_{\widetilde{\NC}_\FM}) \bigr) = \ch_\Delta(\Delta^\omega_{\widetilde{\NC},\FM}) = \mathbf{m}_\omega = \mathbf{m}_0 \cdot T_\omega,
\]
the formula follows.
\end{proof}
\begin{prop}
\label{prop:grk-Hom-tilting}
Let $\underline{s},\underline{t}$ be sequences of simple reflections, and let $\omega,\omega' \in \Omega$. Then the graded $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-module
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}( \MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega,\underline{s}), \MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega',\underline{t}))
\]
is graded free, of graded rank $\langle \mathbf{m}(\omega,\underline{s}),\mathbf{m}(\omega',\underline{t}) \rangle$.
\end{prop}
\begin{proof}
The first assertion follows from Proposition~\ref{prop:morphisms-BS-geom}. For the second assertion, we observe that, by Proposition~\ref{prop:morphisms-BS-geom} again, the graded rank under consideration is equal to the graded dimension of the graded $\FM$-vector space
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})_\FM} \bigl( \mathsf{L} i^*(\MC_{\FM}(\omega,\underline{s})), \mathsf{L} i^*(\MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega',\underline{t})) \bigr)
\]
(where $\FM$ is any geometric point of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$).
Then the formula follows from~\eqref{eqn:grk-tilting} and Proposition~\ref{prop:ch-tilting}.
\end{proof}
\subsection{``Coherent'' Bott--Samelson category}
\label{ss:BS-category-tilting}
We define a ``coherent'' category of Bott--Samelson objects $\mathsf{BS}^{\mathrm{coh}}$ as follows. The objects in this category are the triples $(\omega, \underline{s}, n)$ as in \S\ref{ss:topological-BS}. The morphisms are given by
\[
\Hom_{\mathsf{BS}^{\mathrm{coh}}} \bigl( (\omega, \underline{s}, n), (\omega', \underline{t}, m) \bigr) =
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega,\underline{s}) \langle -n \rangle, \MC_{{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\omega',\underline{t}) \langle -m \rangle ).
\]
\begin{prop}
\label{prop:tilt-Karoubi}
The category $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \GM}(\widetilde{\NC})_\FM)$ can be recovered from the category $\mathsf{BS}^{\mathrm{coh}}$, in the sense that it is equivalent to the Karoubian closure of the additive envelope of the category
which has the same objects as $\mathsf{BS}^{\mathrm{coh}}$, and morphisms from $(\omega, \underline{s}, n)$ to $(\omega', \underline{t}, m)$ which are given by the $(m-n)$-th piece of the graded vector space
\[
\FM \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \left( \bigoplus_{k \in \ZM} \Hom_{\mathsf{BS}^{\mathrm{coh}}} \bigl( (\omega, \underline{s}, 0), (\omega', \underline{t}, k) \bigr) \right).
\]
\end{prop}
\begin{proof}
Consider the category which has the same objects as $\mathsf{BS}^{\mathrm{coh}}$, and whose morphisms are defined as in the statement of the proposition. By Proposition~\ref{prop:morphisms-BS-geom}, this category is equivalent to the full subcategory $\mathsf{A}$ of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM$ whose objects are of the form $\mathsf{L} i^*(\MC_\FM(\omega,\underline{s}))$.
It follows from Proposition~\ref{prop:tilting-Bott-Samelson} that the category $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})_\FM)$ is equivalent to the Karoubian closure of the additive envelope of $\mathsf{A}$; the claim follows.
\end{proof}
\subsection{Kostant--Whittaker reduction and Bott--Samelson objects}
Let us consider the constructions of~\S\S\ref{ss:deformations}--\ref{ss:algebraic-BS}, with ${\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ now defined as $X^*({\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$. Then the space $\tg^*$ of Section~\ref{sec:constructible-side} coincides with the space denoted $\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ in the present section, and we can define the algebra $C$ and the category $\mathsf{BS}^{\mathrm{alg}}$ of ``Bott--Samelson'' $C$-modules with the present data. We use the same notation as in~\S\ref{ss:algebraic-BS} for the modules $E_w$ and $D_s$.
\begin{lem}
\label{lem:isom-C-tangent}
There exists a natural $W$-equivariant isomorphism of graded algebras
\[
C \xrightarrow{\sim} \OC \bigl( \tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
\end{lem}
\begin{proof}
By definition, the algebra $C$ is generated by $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and the images of the elements of the form $\hbar^{-1}(f \otimes 1 - 1 \otimes f) \in C_\hbar$ for $f \in \OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})^W$. On the other hand, the algebra $\OC \bigl( \tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is the symmetric algebra (over $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$) of the module
\[
\bigl( \OC(\tg^*) \otimes_{\OC(\tg^*)^W} \Omega(\tg^*/W) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
(Note that this module is free of finite rank since $\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}/W$ is an affine space, see Lemma~\ref{lem:t/W-affine-space}.) One can easily check that the assignment
\[
\hbar^{-1}(f \otimes 1 - 1 \otimes f) \mapsto d(f)
\]
induces a morphism of graded $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-algebras $C \rightarrow \OC \bigl( \tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, and that this morphism is a $W$-equivariant isomorphism.
\end{proof}
From now on we identify the two algebras in Lemma~\ref{lem:isom-C-tangent} via the isomorphism constructed in its proof. Then one can consider the functor $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ of~\S\ref{ss:KW-functors} as a functor taking values in $D^{\mathrm{b}} \Modgr(C)$.
\begin{prop}
\label{prop:kappa-BS}
For any sequence $\underline{s}$ of simple reflections, any $\omega \in \Omega$, and any $n \in \ZM$, there exists an isomorphism
\[
\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \bigl( \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}) \langle n \rangle \bigr) \cong D(\omega,\underline{s}) \langle n \rangle \qquad \text{in $D^{\mathrm{b}} \Modgr(C)$.}
\]
\end{prop}
Before proving the proposition we begin with a lemma.
\begin{lem}
\label{lem:kappa-Xi}
Let $s$ be a finite simple reflection. Let $\FC$ be an object of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ such that $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC)$ is concentrated in degree $0$, and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free. Then $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ \Xi^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_s(\FC)$ is concentrated in degree $0$ and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, and moreover there exists an isomorphism of graded $C$-modules
\[
\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ \Xi^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_s(\FC) \cong \kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} D_s.
\]
\end{lem}
\begin{proof}
The proof is very similar to the proof of the case $b=T_s$ of Proposition~\ref{prop:KW-actions}\eqref{it:KW-actions-R}. In fact, using the same notation as in this proof, by Lemma~\ref{lem:fiber-product-xis} and the base change theorem (and using also~\eqref{eqn:Ds-tensor-product}) we have a canonical isomorphism of functors
\begin{equation}
\label{eqn:isom-kappa-Xi}
\kappa'_\EM \circ \Xi^\EM_s(-) \cong \kappa'_\EM(-) \otimes_{\OC(\tg^*_\EM)} (\EM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} D_s)
\end{equation}
for $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ or $\FM$. (Note that $D_s$ is free over $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ -- see the proof of Lemma~\ref{lem:fiber-product-xis}\eqref{it:xis-graph} -- so that the tensor product on the right-hand side makes sense in the derived category.)
When $\EM=\FM$, using the fact that $\Xi_s^\FM \cong \mathsf{L} (\widetilde{\pi}_s)^* \circ \mathsf{R} (\widetilde{\pi}_s)_*$
(see~\cite[Proposition~5.2.2]{riche}) and the remarks at the end of~\S\ref{ss:reminder}, one can check that this isomorphism is induced by an isomorphism of functors
\[
\kappa_\FM \circ \Xi^\FM_s(-) \cong \kappa_\FM(-) \otimes_{\OC(\tg^*_\FM)} (\FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} D_s)
\]
(where the tensor product on the right-hand side is defined a similar way as in the case of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$).
Then we can come back to the case $\EM={\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$: if $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\FC)$ is concentrated in degree $0$ and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$-free, \eqref{eqn:isom-kappa-Xi} shows that the same holds for $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \circ \Xi^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_s(\FC)$. And the final claim follows from the case $\EM=\CM$ treated above.
\end{proof}
\begin{proof}[Proof of Proposition~{\rm \ref{prop:kappa-BS}}]
We remark that the statement of Lemma~\ref{lem:kappa-Xi} also holds when $s$ is an \emph{affine} simple reflection, by definition of $\Xi_s$ in this case and Proposition~\ref{prop:KW-actions}\eqref{it:KW-actions-R} (see also~\eqref{eqn:K-'K}). This observation reduces the proof of Proposition~\ref{prop:kappa-BS} to the proof that
\[
\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}({\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{T_\omega}(\OC_{\widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}})) \cong E_\omega
\]
when $\omega \in \Omega$. However by Proposition~\ref{prop:KW-actions} (see also~\eqref{eqn:K-'K} and~\eqref{eqn:n-Tomega}) we have
\[
\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}({\EuScript I}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{T_\omega}(\OC_{\widetilde{\gg}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}})) \cong {}' {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_\omega(\OC_{\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*}),
\]
where $\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*$ is identified with the zero section in $\tg^* \times_{\tg^*/W} \mathsf{T}(\tg^*/W)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Now writing $\omega=v t_\lambda$ (with $v \in W$ and $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$) we have
\[
{}' {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_\omega(\OC_{\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*}) \cong {}' {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{t_\lambda} \circ {}' {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_v (\OC_{\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*}) \cong {}' {\EuScript K}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}_{t_\lambda} (\OC_{\tg_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}^*}) \cong F_\lambda^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}.
\]
Similarly we have $E_\omega \cong E_{t_\lambda}$, and one can check that $F_\lambda^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and $E_{t_\lambda}$ are isomorphic under our identification $C \cong \OC \bigl( \tg^* \times_{\tg^* / W} \mathsf{T}(\tg^* / W) \bigr)_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\end{proof}
\subsection{Equivalence}
\label{ss:equivalence-tilting}
Let us fix isomorphisms as in Proposition~\ref{prop:kappa-BS}, for all $\underline{s}$ and $\omega$. Then the functor $\kappa_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ induces a functor
\[
\kappa_{\mathsf{BS}} \colon \mathsf{BS}^{\mathrm{coh}} \rightarrow \mathsf{BS}^{\mathrm{alg}}.
\]
The main result of this section is the following.
\begin{thm}
\label{thm:main-geom}
The functor $\kappa_{\mathsf{BS}}$ is an equivalence of categories.
\end{thm}
Before proving the theorem we first establish a lemma. In this statement, we identify quasi-coherent sheaves on $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ with $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-modules, and denote by $\Rep(\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ the abelian category of representations of the commutative $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-Lie algebra $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$.
\begin{lem}
\label{lem:differentiation-fully-faithful}
The natural functor
\[
\Rep(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_{\SC}^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \Rep(\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})
\]
defined by differentiating the action is fully faithful on representations which are free over $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$.
\end{lem}
\begin{proof}
The group scheme $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is smooth over $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (see~\S\ref{ss:reminder}), hence it is infinitesimally flat as an $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$-group scheme (in the sense of~\cite[\S I.7.4]{jantzen}) by~\cite[Expos{\'e}~VII, Proposition~1.10]{sga6}. This group scheme is also integral, since it is flat over $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ and its pullback to $\widetilde{\SC}_\CM \cap \widetilde{\gg}_\CM^{\mathrm{rs}}$ is irreducible. (In fact, it can be deduced from~\cite[Lemma~2.3.3]{riche3} that the restriction of $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}^\CM$ to $\widetilde{\gg}_\CM^{\mathrm{rs}}$ identifies with ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}} \times \widetilde{\gg}_\CM^{\mathrm{rs}}$.) Using~\cite[Lemma~I.7.16]{jantzen} we deduce that, if we denote by $\mathrm{Dist}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ the distribution algebra of this group scheme, the natural functor
\[
\Rep(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \mathrm{Mod} \bigl( \mathrm{Dist}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \bigr)
\]
is fully faithful on representations which are projective over $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$.
Now, consider the natural algebra morphism
\[
\UC_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \rightarrow \mathrm{Dist}(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}),
\]
where the left-hand side is the universal enveloping algebra of $\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, see~\cite[\S I.7.10]{jantzen}. If $\KM$ denotes the fraction field of $\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$, we claim that this morphism induces an isomorphism
\begin{equation}
\label{eqn:morph-Lie-Dist}
\KM \otimes_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \UC_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \xrightarrow{\sim} \KM \otimes_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Dist(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}).
\end{equation}
Indeed, the left-hand side is isomorphic to $\UC_{\KM}(\KM \otimes_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$ and, if we set $\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\KM:=\Spec(\KM) \times_{\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}} \widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$, by~\cite[Lemma~2.1.1]{riche3} and~\cite[\S I.7.4, Equation (1)]{jantzen} respectively, there are natural isomorphisms
\[
\KM \otimes_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \widetilde{{\mathfrak I}} \def\ig{{\mathfrak i}} \def\IM{{\mathbb{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \cong \Lie(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\KM), \qquad \KM \otimes_{\OC(\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Dist(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \cong \Dist(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\KM),
\]
so that~\eqref{eqn:morph-Lie-Dist} gets identified with the natural morphism
\[
\UC_\KM(\Lie(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\KM)) \rightarrow \Dist(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\KM).
\]
Since $\KM$ is a field of characteristic zero, the latter morphism is an isomorphism by~\cite[\S I.7.10, Equation (1)]{jantzen}, which finishes the proof of our claim.
Now that the claim is established, the lemma follows from Lemma~\ref{lem:key}.
\end{proof}
\begin{proof}[Proof of Theorem~{\rm \ref{thm:main-geom}}]
By definition, $\kappa_{\mathsf{BS}}$ is essentially surjective. Hence what we have to prove is that for any sequences $\underline{s}, \underline{t}$ of simple reflections and any $\omega, \omega' \in \Omega$, the morphism
\[
\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}), \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})) \rightarrow \Hom_{C}(D(\omega,\underline{s}), D(\omega',\underline{t}))
\]
induced by $\overline{\kappa}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (via the isomorphisms of Proposition~\ref{prop:kappa-BS}) is an isomorphism. By Corollary~\ref{cor:grk-Hom-parity} and Proposition~\ref{prop:grk-Hom-tilting}, both sides are graded free over $\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})$, of the same graded rank. Hence, using Lemma~\ref{lem:morphism-grk}, to finish the proof it suffices to prove that the image of our morphism under the functor $\OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} (-)$ is injective for any geometric point $\FM$ of ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$. Then, using the definition of $\overline{\kappa}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ (see~\S\ref{ss:KW-functors}) and Lemma~\ref{lem:differentiation-fully-faithful}, it suffices to prove that the morphism
\begin{multline*}
\OC(\tg^*_{{\mathrm{rs}},\FM}) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}), \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})) \\
\rightarrow \OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Hom_{\Rep(\widetilde{{\mathbf I}} \def\ib{{\mathbf i}} \def\IC{{\mathcal{I}}}_\SC^{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}){}_{| \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}, \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t}){}_{| \widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}})
\end{multline*}
induced by restriction to $\widetilde{\SC}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is injective. And finally, to prove this injectivity it suffices to prove that the natural morphism
\begin{multline*}
\OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}), \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})) \\
\rightarrow \Hom_{\Coh(\widetilde{\SC}_\FM^{\mathrm{rs}})}(\MC_\FM(\omega,\underline{s}){}_{| \widetilde{\SC}_\FM^{\mathrm{rs}}}, \MC_\FM(\omega',\underline{t}){}_{| \widetilde{\SC}_\FM^{\mathrm{rs}}})
\end{multline*}
is injective.
Consider the following chain of isomorphisms:
\begin{multline*}
\OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}})} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}), \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega', \underline{t})) \\
\cong \OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_\FM)} \bigl( \FM \otimes_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}}(\MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega,\underline{s}), \MC_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}(\omega',\underline{t})) \bigr) \\
\cong \OC(\tg^{*,{\mathrm{rs}}}_{\FM}) \otimes_{\OC(\tg^*_\FM)} \bigl( \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})_\FM}(\MC_\FM(\omega,\underline{s}), \MC_\FM(\omega',\underline{t})) \bigr) \\
\cong \Hom_{\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg}^{\mathrm{rs}})_\FM}(\MC_\FM(\omega,\underline{s}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}}, \MC_\FM(\omega',\underline{t}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}}).
\end{multline*}
Here the second isomorphism follows from Proposition~\ref{prop:morphisms-BS-geom}, and the other ones are easy. (Observe that, on the last line, the objects $\MC_\FM(\omega,\underline{s}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}}$ and $\MC_\FM(\omega',\underline{t}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}}$ are concentrated in degree $0$, i.e.~equivariant coherent sheaves.)
Using these identifications, the morphism under consideration is the morphism
\begin{multline*}
\Hom_{\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM}(\widetilde{\gg}^{\mathrm{rs}}_\FM)} \bigl( \MC_\FM(\omega,\underline{s}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}}, \MC_\FM(\omega',\underline{t}){}_{|_{\widetilde{\gg}^{\mathrm{rs}}_\FM}} \bigr) \\
\rightarrow \Hom_{\Coh(\widetilde{\SC}^{\mathrm{rs}})_\FM} \bigl( \MC_\FM(\omega,\underline{s}){}_{| \widetilde{\SC}_\FM^{\mathrm{rs}}}, \MC_\FM(\omega',\underline{t}){}_{| \widetilde{\SC}_\FM^{\mathrm{rs}}} \bigr)
\end{multline*}
induced by restriction to $\widetilde{\SC}_\FM^{\mathrm{rs}}$. Then injectivity follows from the observation that the functor $\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg}^{\mathrm{rs}})_\FM \rightarrow \Coh(\widetilde{\SC}^{\mathrm{rs}})_\FM$ is faithful, since it can be written as the composition
\[
\Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg}^{\mathrm{rs}})_\FM \rightarrow \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \widetilde{\SC}^{\mathrm{rs}})_\FM \xrightarrow{\sim} \Coh(\widetilde{\SC}^{\mathrm{rs}})_\FM
\]
where the first functor is the inverse image under the natural morphism $a^{{\mathrm{rs}}} \colon {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\FM \times \widetilde{\SC}_\FM^{\mathrm{rs}} \rightarrow \widetilde{\gg}_\FM^{\mathrm{rs}}$ (which is faithful since $a^{\mathrm{rs}}$ is smooth and surjective, see Lemma~\ref{lem:a-smooth}), and the second functor is the natural equivalence.
\end{proof}
\section{Proofs of the main results}
\label{sec:proofs}
In this section (except in~\S\ref{ss:proof-parity-perverse}) we come back to the assumptions of~\S\ref{ss:intro}: ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ is a product of simply-connected quasi-simple groups and general linear groups over $\FM$, and the characteristic $p$ of $\FM$ is very good for each quasi-simple factor of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. Such a group can be obtained by base change to $\FM$ from a split connected reductive group scheme ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$ over $\ZM$, and we set ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_{\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}=\Spec({\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}) \times_{\Spec(\ZM)} {\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}_\ZM$, where ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ is the localization of $\ZM$ at all the prime numbers which are not very good for a quasi-simple factor of ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$. These data satisfy the assumptions of Sections~\ref{sec:KW-reduction}--\ref{sec:tilting-KW}, and we use the same notation as in these sections. (The only exception is that we drop the subscripts ``$\FM$,'' since the geometric point will not vary anymore.)
We let also $\check{G}$ be the complex Langlands dual group. This group and ${\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ satisfy the assumptions of Section~\ref{sec:constructible-side}, and we also use the notation of this section. (Note a slight conflict of notation: $\tg$ denotes ${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}} \otimes_\ZM {\mathfrak R}} \def\rg{{\mathfrak r}} \def\RM{{\mathbb{R}}$ in Section~\ref{sec:constructible-side}, while in the present section it will denote ${\check {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}} \otimes_\ZM \FM$.)
We assume that the roots of $\check{B}$ with respect to $\check{T}$ are the coroots of ${\mathbf B}} \def\bb{{\mathbf b}} \def\BC{{\mathcal{B}}$ with respect to ${\mathbf T}} \def\tb{{\mathbf t}} \def\TC{{\mathcal{T}}$.
\subsection{Proof of Theorem~\ref{thm:main}}
\label{ss:proof-main}
Combining Theorem~\ref{thm:main-top} and Theorem~\ref{thm:main-geom} we obtain an equivalence of categories
\begin{equation}
\label{eqn:equiv-BS}
\mathsf{BS}^{\mathrm{top}} \xrightarrow{\sim} \mathsf{BS}^{\mathrm{coh}}
\end{equation}
which is the identity morphism on objects. Then using Proposition~\ref{prop:parity-Karoubi} and Proposition~\ref{prop:tilt-Karoubi} we deduce the desired equivalence of additive categories
\[
\Theta \colon
\Parity_{(\check{I})}({\EuScript Gr}, \FM) \xrightarrow{\sim} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})).
\]
By construction, this equivalence
satisfies
\begin{equation}
\label{eqn:Theta'-BS}
\Theta \bigl( \EC_\FM(\omega,\underline{s}) \bigr) \cong \mathsf{L} i^* \MC_\FM(\omega,\underline{s})
\end{equation}
for any sequence $\underline{s}$ of simple reflections and any $\omega \in \Omega$, and $\Theta \circ [1] \cong \langle -1 \rangle \circ \Theta$.
It also satisfies property~\eqref{it:main-thm-indec},
by~\eqref{eqn:Theta'-BS} and a standard argument based on induction on $\ell(w_\lambda)$ (see in particular Remark~\ref{rk:indec-parity} and \cite[Remark~4.3]{mr}).
Finally we prove property~\eqref{it:main-thm-ch}. In fact, this property can be expressed as the following equalities for any $\EC$ in $\Parity_{(\check{I})}({\EuScript Gr},\FM)$:
\begin{equation}
\label{eqn:equality-characters}
\ch_{{\EuScript Gr}}^*(\EC) = \ch_\Delta(\Theta(\EC)); \qquad \ch_{{\EuScript Gr}}^!(\EC) = \ch_{\nabla}(\Theta(\EC)).
\end{equation}
However these formulas hold when $\EC=\EC_\FM(\omega,\underline{s})$ by Proposition~\ref{prop:ch-BS-parity}, Proposition~\ref{prop:ch-tilting}, and~\eqref{eqn:Theta'-BS}. And one can easily check that if they hold for $\EC$ then they hold for $\EC[1]$. Using these observations one can prove, by standard arguments, that the formulas~\eqref{eqn:equality-characters} hold when $\EC=\EC_\lambda[i]$ for some $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $i \in \ZM$, by induction on $\ell(w_\lambda)$. The general case follows.
\subsection{Proof of Theorem~\ref{thm:equivalence-tNC}}
\label{ss:proof-equivalence-tNC}
We construct the equivalence $\Phi$ as the composition
\[
D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM) := K^{\mathrm{b}} \Parity_{(\check{I})}({\EuScript Gr}, \FM) \xrightarrow[\sim]{K^{\mathrm{b}}(\Theta)} K^{\mathrm{b}} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))
\xrightarrow{\sim} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}),
\]
where the last equivalence is provided by~\cite[Proposition~3.11]{mr}.
Then it follows from this construction and the properties of $\Theta$ proved in Theorem~\ref{thm:main} that we have
\[
\Phi \circ \langle 1 \rangle \cong \langle 1 \rangle [1] \circ \Phi \qquad \text{and} \qquad \Phi(\EC^\mathrm{mix}_\lambda) \cong \TC^{-\lambda}.
\]
It remains to prove that
\[
\Phi(\Delta^\mix_{-\lambda}) \cong \Delta_{\widetilde{\NC}}^{\lambda} \qquad \text{and} \qquad \Phi(\nabla^\mix_{-\lambda}) \cong \nabla_{\widetilde{\NC}}^{\lambda}
\]
for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
We explain the proof of the first isomorphism; the proof of the second one is similar.
Our proof is similar to the proof of a similar claim in~\cite[Lemma~5.2]{modrap2}. Namely, we prove the isomorphism by induction on $\ell(w_\lambda)$. If $\ell(w_\lambda)=0$, then we have $\Delta^{\mathrm{mix}}_{-\lambda} \cong \EC^{\mathrm{mix}}_{-\lambda}$ and $\Delta^\lambda_{\widetilde{\NC}} \cong \TC^\lambda$, hence our claim is clear. For a general $\lambda$, we consider a non-zero morphism $\Delta^\lambda_{\widetilde{\NC}} \rightarrow \TC^\lambda$ (which is unique up to an invertible scalar), and the associated triangle
\begin{equation}
\label{eqn:triangle-Delta-Phi}
\Delta^\lambda_{\widetilde{\NC}} \rightarrow \TC^\lambda \rightarrow N \xrightarrow{[1]}.
\end{equation}
Then $N$ belongs to the triangulated subcategory of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ generated by the objects $\Delta^\mu_{\widetilde{\NC}} \langle m \rangle$ with $m \in \ZM$ and $\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ which satisfies ${\EuScript Gr}_{-\mu} \subset \overline{{\EuScript Gr}_{-\lambda}}$ and $\mu \neq \lambda$ (see Theorem~\ref{thm:main}\eqref{it:main-thm-ch}--\eqref{it:main-thm-indec}). In particular, using induction we deduce that $\Phi^{-1}(N)$ is supported on $\overline{{\EuScript Gr}_{-\lambda}} \smallsetminus {\EuScript Gr}_{-\lambda}$.
Using again induction we also observe that
\[
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM)}(\Phi^{-1}(\Delta^\lambda_{\widetilde{\NC}}), \Delta^{\mathrm{mix}}_{-\mu} \langle m \rangle [n]) = 0
\]
for all $n,m \in \ZM$ and $\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ such that ${\EuScript Gr}_{-\mu} \subset \overline{{\EuScript Gr}_{-\lambda}}$ and $\mu \neq \lambda$. Hence the triangle
\[
\Phi^{-1}(\Delta^\lambda_{\widetilde{\NC}}) \rightarrow \EC^\mathrm{mix}_{-\lambda} \rightarrow \Phi^{-1}(N) \xrightarrow{[1]}
\]
obtained from~\eqref{eqn:triangle-Delta-Phi}
satisfies the two properties which characterize uniquely the triangle whose first arrow is the unique (up to scalar) non-zero morphism $\Delta^{\mathrm{mix}}_{-\lambda} \rightarrow \EC^{\mathrm{mix}}_{-\lambda}$. In particular, we deduce the wished-for isomorphism $\Phi^{-1}(\Delta^\lambda_{\widetilde{\NC}}) \cong \Delta^{\mathrm{mix}}_{-\lambda}$.
\begin{remark}
\label{rk:real-tNC}
The equivalence $K^{\mathrm{b}} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))
\xrightarrow{\sim} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$ used in the proof of Theorem~\ref{thm:equivalence-tNC} is the composition
\begin{equation}
\label{eqn:real-1}
K^{\mathrm{b}} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})) \hookrightarrow K^{\mathrm{b}} {\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}) \xrightarrow{\mathsf{real}} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})
\end{equation}
where $\mathsf{real}$ is the functor constructed in~\cite[Lemme~3.1.11]{bbd}. On the other hand it follows from~\cite[Corollary~4.16]{mr} that the objects in $\Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))$ are concentrated in degree $0$, so that one can also construct such an equivalence as the composition
\begin{equation}
\label{eqn:real-2}
K^{\mathrm{b}} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})) \hookrightarrow K^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}) \xrightarrow{\mathsf{can}} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}),
\end{equation}
where $\mathsf{can}$ is the canonical functor. We claim that~\eqref{eqn:real-1} and~\eqref{eqn:real-2} are isomorphic. Indeed, it is clear that $\mathsf{can}$ coincides with the realization functor from~\cite[Lemme~3.3.11]{bbd} associated with the tautological t-structure on $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$. Then the claim follows from the observation that if $\mathscr{C}$ and $\mathscr{C}'$ are the hearts of two bounded t-structures on $D^{\mathrm{b}} \mathscr{A}$ for some abelian category $\mathscr{A}$, then the following diagram commutes up to isomorphism:
\[
\xymatrix@R=0.6cm{
K^{\mathrm{b}} (\mathscr{C} \cap \mathscr{C}') \ar@{^{(}->}[r] \ar@{^{(}->}[d] & K^{\mathrm{b}} \mathscr{C} \ar[d]^-{\mathsf{real}} \\
K^{\mathrm{b}}\mathscr{C}' \ar[r]^{\mathsf{real'}} & D^{\mathrm{b}} \mathscr{A}.
}
\]
\end{remark}
\subsection{Proof of Theorem~\ref{thm:equivalence-tgg}}
\label{ss:proof-equivalence-tgg}
Before considering the proof of Theorem~\ref{thm:equivalence-tgg}, we need some preliminary results.
First, recall that
we use the same notation for the indecomposable objects in $\Parity_{\check{I}}({\EuScript Gr}, \FM)$
and in $\Parity_{(\check{I})}({\EuScript Gr}, \FM)$, see~\S\ref{ss:notation-constructible}.
We define the additive subcategory ${\EuScript Tilt}$ of $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ as the Karoubian closure of the additive envelope of the subcategory generated by the objects $\MC_\FM(\omega,\underline{s})$ for $\underline{s}$ a sequence of simple reflections and $\omega \in \Omega$. Using~\eqref{eqn:equiv-BS} it is not difficult to construct an equivalence of categories
\begin{equation}
\label{eqn:equiv-Parity-Tilt-equiv}
\Parity_{\check{I}}({\EuScript Gr}, \FM) \xrightarrow{\sim} {\EuScript Tilt}
\end{equation}
sending $\EC_\FM(\omega,\underline{s}) [n]$ to $\MC_\FM(\omega,\underline{s}) \langle -n \rangle$.
This equivalence shows that the category ${\EuScript Tilt}$ is Krull--Schmidt, and that one can transfer the known classification of indecomposable objects in $\Parity_{\check{I}}({\EuScript Gr}, \FM)$ proved in~\cite{jmw} to deduce a classification of the indecomposable objects in ${\EuScript Tilt}$. Namely, for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$, if $w_{\lambda}=\omega s_1 \cdots s_r$ is a reduced decomposition, then there exists a unique indecomposable factor $\widetilde{\TC}^\lambda$ of $\MC_\FM(\omega,(s_1, \cdots, s_r))$ which is not isomorphic to any object of the form $\MC_\FM(\omega',\underline{t}) \langle m \rangle$ where $m \in \ZM$, $\omega' \in \Omega$, and $\underline{t}$ is a sequence of simple reflections of length at most $r-1$. This object does not depend on the choice of the reduced decomposition up to isomorphism. Moreover, any indecomposable object in ${\EuScript Tilt}$ is isomorphic to some $\widetilde{\TC}^\lambda \langle m \rangle$ for some $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$. Finally, the image of $\EC_{-\lambda}$ under~\eqref{eqn:equiv-Parity-Tilt-equiv} is $\widetilde{\TC}^\lambda$.
The following lemma implies that the objects $\widetilde{\TC}^\lambda$ are ``deformations'' of the tilting exotic sheaves $\TC^\lambda$.
\begin{lem}
\label{lem:tTC-i*}
For any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ we have $\mathsf{L} i^*(\widetilde{\TC}^\lambda) \cong \TC^\lambda$.
\end{lem}
\begin{proof}
The result easily follows if we can prove that $\mathsf{L} i^*(\widetilde{\TC}^\lambda)$ is indecomposable. However, since $\widetilde{\TC}^\lambda$ is a direct summand in an object of the form $\MC_\FM(\omega,\underline{s})$, it follows from Proposition~\ref{prop:morphisms-BS-geom} that the functor $\mathsf{L} i^*$ induces an isomorphism
\[
\FM \otimes_{\OC(\tg^*)} \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\widetilde{\TC}^\lambda) \xrightarrow{\sim} \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})} \bigl( \mathsf{L} i^* (\widetilde{\TC}^\lambda) \bigr).
\]
Then the proof proceeds as for Lemma~\ref{lem:parity-indec-For}.
\end{proof}
\begin{lem}
\label{lem:Tilt-generates}
The subcategory ${\EuScript Tilt}$ generates $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ as a triangulated category.
\end{lem}
\begin{proof}
Looking at the proof of~\cite[Corollary~4.2]{mr}, one can check that the triangulated subcategory generated by ${\EuScript Tilt}$ coincides with the triangulated subcategory generated by the objects $\Delta^\lambda_{\widetilde{\gg}} \langle m \rangle$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$. By~\cite[Lemma~1.11.3(2)]{br}, for $w \in W$ we have ${\EuScript I}_{T_w}(\OC_{\widetilde{\gg}}) \cong \OC_{\widetilde{\gg}} \langle - \ell(w) \rangle$. We deduce (as in Remark~\ref{rk:nabla-delta-Baff}) that we have
\[
\Delta^\lambda_{\widetilde{\gg}} \cong {\EuScript I}_{(T_{t_{-\lambda}})^{-1}}(\OC_{\widetilde{\gg}}) \langle - \ell(t_\lambda) + \ell(w_\lambda) \rangle.
\]
Then, using~\cite[Lemma~2.3]{mr} and~\cite[Lemma~1.11.3]{br}, one can check that the triangulated subcategory generated by the object $\Delta^\lambda_{\widetilde{\gg}}\langle m \rangle$ coincides with the triangulated subcategory generated by the objects $\OC_{\widetilde{\gg}}(\lambda) \langle m\rangle$ for $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$ and $m \in \ZM$. Hence the claim of the lemma is equivalent to the claim that $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ is generated by this collection of objects, which can be proved by the same arguments as the corresponding claim for $\widetilde{\NC}$ in~\cite[Corollary~5.8]{achar}.
\end{proof}
\begin{lem}
\label{lem:KbTilt-tgg}
There exists an equivalence of triangulated categories
\[
K^{\mathrm{b}} {\EuScript Tilt} \xrightarrow{\sim} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})
\]
commuting with $\langle 1 \rangle$ and
sending $\widetilde{\TC}^\lambda$ to $\widetilde{\TC}^\lambda$ for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}$.
\end{lem}
\begin{proof}
It follows from~\cite[Corollary~4.16]{mr} that the complexes $\TC^\lambda$ are concentrated in degree $0$. Using Lemma~\ref{lem:tTC-i*} and Lemma~\ref{lem:nakayama}\eqref{it:nakayama-complex} (over an affine open covering of $\widetilde{\gg}$) we deduce that the complexes $\widetilde{\TC}^\lambda$ are also concentrated in degree $0$. We construct the functor of the lemma as the composition
\[
K^{\mathrm{b}} {\EuScript Tilt} \rightarrow K^{\mathrm{b}} \Coh^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg}) \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg}).
\]
It follows from standard arguments, using Proposition~\ref{prop:morphisms-BS-geom} and Lemma~\ref{lem:Tilt-generates}, that this functor is an equivalence of categories.
\end{proof}
\begin{proof}[Proof of Theorem~{\rm \ref{thm:equivalence-tgg}}]
We define $\Psi$ as the composition
\[
D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr}, \FM) := K^{\mathrm{b}} \Parity_{\check{I}}({\EuScript Gr}, \FM) \xrightarrow[\sim]{\eqref{eqn:equiv-Parity-Tilt-equiv}} K^{\mathrm{b}} {\EuScript Tilt}
\xrightarrow[\sim]{{\rm Lemma~\ref{lem:KbTilt-tgg}}} D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg}).
\]
By construction, this equivalence satisfies $\Psi \circ \langle 1 \rangle \cong \langle 1 \rangle [1] \circ \Psi$ and $\Psi(\EC^{\mathrm{mix}}_{-\lambda}) \cong \widetilde{\TC}^{\lambda}$. The isomorphisms involving standard and costandard objects can be proved using the same arguments as for their counterparts in Theorem~\ref{thm:equivalence-tNC}.
\end{proof}
\subsection{Compatibility}
\label{ss:compatibilities}
In this subsection we prove
that the functors $\Phi$ and $\Psi$ as constructed in~\S\S\ref{ss:proof-equivalence-tNC}--\ref{ss:proof-equivalence-tgg}
are compatible in the natural way.
We will denote by
$\For \colon D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr}, \FM) \rightarrow D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM)$
the ``forgetful functor'' induced by the forgetful functor $\Parity_{\check{I}}({\EuScript Gr}, \FM) \rightarrow \Parity_{(\check{I})}({\EuScript Gr}, \FM)$.
\begin{prop}
\label{prop:Phi-Psi-For}
The following diagram commutes up to isomorphisms of functors:
\[
\xymatrix@R=0.6cm@C=1.5cm{
D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr}, \FM) \ar[r]^-{\Psi}_-{\sim} \ar[d]_-{\For} & D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg}) \ar[d]^-{\mathsf{L} i^*} \\
D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}, \FM) \ar[r]^-{\Phi}_-{\sim} & D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}).
}
\]
\end{prop}
\begin{proof}
It is clear from construction that the diagram
\[
\xymatrix@R=0.6cm@C=1.5cm{
\Parity_{\check{I}}({\EuScript Gr}, \FM) \ar[r]_-{\sim}^-{\eqref{eqn:equiv-Parity-Tilt-equiv}} \ar[d]_-{\For} & {\EuScript Tilt} \ar[d]^-{\mathsf{L} i^*} \\
\Parity_{(\check{I})}({\EuScript Gr}, \FM) \ar[r]_-{\sim}^-{\eqref{eqn:equiv-main}} & \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}))
}
\]
commutes. Hence to conclude we only have to prove that the diagram
\[
\xymatrix@R=0.6cm@C=1.5cm{
K^{\mathrm{b}} {\EuScript Tilt} \ar[r]^-{\sim} \ar[d]_-{K^{\mathrm{b}}(\mathsf{L} i^*)} & D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg}) \ar[d]^-{\mathsf{L} i^*} \\
K^{\mathrm{b}} \Tilt({\EuScript E}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})) \ar[r]^-{\sim} & D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})
}
\]
commutes, where the horizontal arrows are the functors considered in~\S\ref{ss:proof-equivalence-tgg} and~\S\ref{ss:proof-equivalence-tNC} respectively. The latter fact follows from Remark~\ref{rk:real-tNC} and the fact that the objects in ${\EuScript Tilt}$, considered as equivariant coherent sheaves, are acyclic for the functor $i^*$ (see Lemma~\ref{lem:nakayama}\eqref{it:nakayama-complex}).
\end{proof}
\subsection{Proof of Corollary~\ref{cor:parity-perverse}}
\label{ss:proof-parity-perverse}
We begin with an easy lemma. Let $\Bbbk$ be a field, and $X=\bigsqcup_{s \in \mathscr{S}} X_s$ be an algebraic variety endowed with an algebraic stratification, where $X_s$ is simply connected for any $s \in \mathscr{S}$. We denote by $\DM_X$ the Grothendieck--Verdier duality functor.
\begin{lem}
\label{lem:perverse-criterion}
Let $\FC$ be an object in $D^{\mathrm{b}}_{\mathscr{S}}(X,\Bbbk)$ which satisfies $\DM_X(\FC) \cong \FC$ and $\Hom(\FC, \FC[n])=0$ for any $n \in \ZM_{<-1}$. Then $\FC$ is perverse.
\end{lem}
\begin{proof}
Assume that $\FC$ is not perverse, and
let $N=\max\{n \in \ZM_{> 0} \mid {}^p \hspace{-1pt} \HC^n(\FC) \neq 0\}$. Then since $\DM_X(\FC) \cong \FC$, $N$ is also the largest integer such that ${}^p \hspace{-1pt} \HC^{-N}(\FC) \neq 0$, and we have $\HC^{-N}(\FC) \cong \DM_X(\HC^{N}(\FC))$. In particular, we deduce that $\mathrm{top}(\HC^{N}(\FC)) \cong \DM_X(\mathrm{soc}(\HC^{-N}(\FC)))$. Under our assumptions each simple $\mathscr{S}$-constructible perverse sheaf on $X$ is stable under $\DM_X$, hence we deduce that there exists an isomorphism $\mathrm{top}(\HC^{N}(\FC)) \cong \mathrm{soc}(\HC^{-N}(\FC))$; in particular there exists a non-zero morphism $\phi \colon \HC^{N}(\FC) \rightarrow \HC^{-N}(\FC)$. Now consider the following morphism (where the first and third morphisms come from the appropriate perverse truncation triangles):
\[
\psi \colon \FC \rightarrow \HC^{N}(\FC)[-N] \xrightarrow{\phi[-N]} \HC^{-N}(\FC)[-N] \rightarrow \FC[-2N].
\]
Then ${}^p \hspace{-1pt} \HC^N(\psi) \neq 0$, hence $\psi$ is a non zero element in $\Hom(\FC, \FC[-2N])$, contradicting our assumption.
\end{proof}
\begin{proof}[Proof of Corollary~{\rm \ref{cor:parity-perverse}}]
By standard reductions (see e.g.~\cite[Lemma~3.6]{jmw2}) one can assume that ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ is a product of simply connected quasi-simple groups not of type $A$ and general linear groups. Then ${\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}$ and $\FM$ satisfy the assumptions of Theorem~\ref{thm:main}.
Using Lemma~\ref{lem:perverse-criterion}, to prove our claim it suffices to prove that for $\lambda \in -{\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ and $n<-1$ we have
\[
\Hom_{\Parity_{(\check{I})}({\EuScript Gr}, \FM)}(\EC_\lambda, \EC_\lambda[n])=0.
\]
However we have $\Theta(\EC_\lambda) \cong \TC^{-\lambda}$ by Theorem~\ref{thm:main}\eqref{it:main-thm-indec}, hence $\Theta(\EC_\lambda) \cong \mathsf{T}(-\lambda) \otimes \OC_{\widetilde{\NC}}$ by~\cite[Corollary~4.8]{mr}. We deduce that
\[
\Hom_{\Parity_{(\check{I})}({\EuScript Gr}, \FM)}(\EC_\lambda, \EC_\lambda[n]) \cong \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})}(\mathsf{T}(-\lambda) \otimes \OC_{\widetilde{\NC}}, \mathsf{T}(-\lambda) \otimes \OC_{\widetilde{\NC}} \langle -n \rangle).
\]
Now by~\cite[Proposition~A.6]{mr} the right-hand side is isomorphic to
\begin{multline*}
\mathsf{H}^0 \bigl( \mathsf{Inv}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}} \circ \mathsf{Inv}^{\Gm}(\mathsf{T}(-\lambda)^* \otimes \mathsf{T}(-\lambda) \otimes R\Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}) \langle -n \rangle) \bigr) \\
\cong \mathsf{H}^0 \Bigl( \mathsf{Inv}^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}} \bigl(\mathsf{T}(-\lambda)^* \otimes \mathsf{T}(-\lambda) \otimes \mathsf{Inv}^{\Gm} ( R\Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}) \langle -n \rangle) \bigr) \Bigr).
\end{multline*}
Now, using the same arguments as in~\cite[Lemma~1.4.2]{br} one can check that
\[
\mathsf{Inv}^{\Gm}(R\Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}) \langle -n \rangle)=0
\]
unless $n \in 2\ZM_{\geq 0}$. The claim follows.
\end{proof}
\subsection{Proof of Proposition~\ref{prop:equiv-parity-tilt}}
\label{ss:proof-Satake}
By~\cite[Corollary~4.8]{mr}, for any $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$ we have $\TC^\lambda \cong \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}$, hence $\Theta(\EC_{-\lambda}) \cong \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}$.
Now we observe that, by the same arguments as in the proof of Corollary~\ref{cor:parity-perverse} and since $\mathsf{Inv}^{\Gm} \bigl( R\Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}) \bigr) = \FM$, the functor
\[
\Tilt({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}) \rightarrow D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC}) \colon V \mapsto V \otimes \OC_{\widetilde{\NC}}
\]
is fully faithful.
Hence $\Theta$ induces an equivalence
\begin{equation}
\label{eqn:Theta'-perv}
\PParity_{(\check{G}(\mathscr{O}))}({\EuScript Gr}) \xrightarrow{\sim} \Tilt({\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}})
\end{equation}
sending $\EC^\lambda = \EC_{w_0 \lambda}$ to $\mathsf{T}(-w_0 \lambda)$ for all $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$.
Now by~\cite[Proposition~2.1]{mv} the categories of $\check{G}(\mathscr{O})$-constructible and $\check{G}(\mathscr{O})$-equivariant perverse sheaves on ${\EuScript Gr}$ are canonically equivalent. Using this property and the antiautomorphism of $\check{G}(\mathscr{K})$ defined by $g \mapsto g^{-1}$, one can construct an autoequivalence of $\PParity_{(\check{G}(\mathscr{O}))}({\EuScript Gr})$ sending $\EC^\lambda$ to $\EC^{-w_0\lambda}$.
Composing this equivalence with~\eqref{eqn:Theta'-perv} provides the equivalence $\overline{\mathsf{S}}_\FM$.
Formula~\eqref{eqn:formula-stalks-parity} follows from the following chain of equalities for $\lambda,\mu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$:
\begin{align*}
\sum_{k \in \ZM} \dim \bigl( \HM^{k-\dim({\EuScript Gr}^{\mu})}(\imath_\mu^* \EC^\lambda) \bigr) \cdot \mathsf{v}^k &= \sum_{k \in \ZM} \dim \bigl( \HM^{k-\dim({\EuScript Gr}_{w_0 \mu})}({\EuScript Gr}_{w_0 \mu}, i_{w_0 \mu}^* \EC_{w_0\lambda}) \bigr) \cdot \mathsf{v}^k \\
&= \sum_{k \in \ZM} (\TC^{-w_0\lambda} : \Delta_{\widetilde{\NC}}^{-w_0 \mu} \langle k \rangle) \cdot \mathsf{v}^k \\
&= \sum_{\nu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+} (\mathsf{T}(-w_0\lambda) : \mathsf{M}(\nu)) \cdot \MC_\nu^{-w_0 \mu}(\mathsf{v}^{-2})
\\
& = \sum_{\nu \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+} \bigl( \mathsf{T}(\lambda) : \mathsf{N}(-w_0\nu) \bigr) \cdot \MC^{-w_0\mu}_\nu(\mathsf{v}^{-2}).
\end{align*}
Here the second equality follows from
Theorem~\ref{thm:main}\eqref{it:main-thm-ch}, the third one follows from the formula for $\ch_\Delta(\mathsf{T}(-w_0 \lambda) \otimes \OC_{\widetilde{\NC}})$ provided by~\cite[Proposition~4.6]{mr}, and the last one from the fact that $(\mathsf{T}(-w_0 \lambda) : \mathsf{N}(\nu)) = (\mathsf{T}(\lambda) : \mathsf{M}(-w_0 \nu))$ (since $\mathsf{T}(\lambda)^* \cong \mathsf{T}(-w_0 \lambda)$).
\subsection{Proof of Proposition~\ref{prop:cohomology}}
\label{ss:proof-cohomology}
Before giving the proof of the proposition, we start with a lemma describing the objects $\widetilde{\TC}^\lambda$ introduced in~\S\ref{ss:proof-equivalence-tgg} in the case $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$.
\begin{lem}
\label{lem:ttilting-dominant}
For $\lambda \in {\mathbf X}} \def\xb{{\mathbf x}} \def\XC{{\mathcal{X}}^+$, we have $\widetilde{\TC}^\lambda \cong T(\lambda) \otimes \OC_{\widetilde{\gg}}$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$.
\end{lem}
\begin{proof}
Recall that, by~\cite[Corollary~4.8]{mr}, we have $\TC^\lambda \cong \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}$.
Using this fact,
the isomorphisms
\[
\mathsf{L} i^*(\widetilde{\TC}^\lambda) \cong \TC^\lambda, \qquad \mathsf{L} i^*(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}) \cong \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}
\]
(see Lemma~\ref{lem:tTC-i*}),
and the same arguments as in the proof of Proposition~\ref{prop:morphism-D-N-tgg},
one can check that the graded $\OC(\tg^*)$-modules $\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\widetilde{\TC}^\lambda, \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}})$, $\Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}, \widetilde{\TC}^\lambda)$, $\End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\widetilde{\TC}^\lambda)$ and $\End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}})$ are free, and that the morphisms
\begin{align*}
\FM \otimes_{\OC(\tg^*)} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\widetilde{\TC}^\lambda, \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}) & \rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\TC^\lambda, \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}), \\
\FM \otimes_{\OC(\tg^*)} \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}, \widetilde{\TC}^\lambda) & \rightarrow \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}, \TC^\lambda), \\
\FM \otimes_{\OC(\tg^*)} \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\widetilde{\TC}^\lambda) & \rightarrow \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\TC^\lambda), \\
\FM \otimes_{\OC(\tg^*)} \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\gg})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}) & \rightarrow \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}})
\end{align*}
induced by $\mathsf{L} i^*$
are isomorphisms. Since $\End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}) \cong \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}}}(\widetilde{\NC})}(\TC^\lambda)$ is concentrated in non-negative degrees (see the proof of Corollary~\ref{cor:parity-perverse}), we deduce that the morphisms
\begin{align*}
\End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})}(\widetilde{\TC}^\lambda) &\rightarrow \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})}(\TC^\lambda), \\ \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}) &\rightarrow \End_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})}(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}})
\end{align*}
induced by $\mathsf{L} i^*$ are isomorphisms.
Choose a pair of inverse isomorphisms $f \colon \TC^\lambda \xrightarrow{\sim} \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}$ and $g \colon \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}} \xrightarrow{\sim} \TC^\lambda$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})$. Then by the preceding observations there exist morphisms $f' \colon \widetilde{\TC}^\lambda \rightarrow \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}$ and $g' \colon \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}} \rightarrow \widetilde{\TC}^\lambda$ in $D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\gg})$ such that $\mathsf{L} i^*(f')=f$, $\mathsf{L} i^*(g')=g$. And these observations also show that the fact that $g \circ f=\mathrm{id}$, resp.~$f \circ g = \mathrm{id}$, implies that $g' \circ f'=\mathrm{id}$, resp.~$f' \circ g' = \mathrm{id}$.
\end{proof}
\begin{proof}[Proof of Proposition~{\rm \ref{prop:cohomology}}]
First we prove~\eqref{it:cohom}.
It follows from the construction of our equivalence~\eqref{eqn:equiv-Parity-Tilt-equiv} that the following diagram commutes:
\[
\xymatrix@R=0.5cm{
\Parity_{\check{I}}({\EuScript Gr}, \FM) \ar[rr]_-{\sim}^-{\eqref{eqn:equiv-Parity-Tilt-equiv}} \ar[rd]_-{\HM^\bullet_{\check{I}}({\EuScript Gr},-)} & & {\EuScript Tilt} \ar[ld]^-{\kappa} \\
& \mathrm{Mod}^\mathrm{gr}(\OC(\tg^*)). &
}
\]
(In both downward arrows, we omit the forgetful functor from graded $C$-modules to graded $\OC(\tg^*)$-modules.) It is clear that $\kappa(\mathsf{T}(\lambda) \otimes \OC_{\widetilde{\gg}}) \cong \mathsf{T}(\lambda) \otimes \OC(\tg^*)$, with the grading indicated in the statement of the proposition. On the other hand, we have
\begin{multline*}
\HM^\bullet_{\check{I}}({\EuScript Gr}, \EC^\lambda) \cong \HM^\bullet_{\check{I}}(\mathrm{pt}; \FM) \otimes_{\HM^\bullet_{\check{G}(\mathscr{O})}(\mathrm{pt}; \FM)} \HM^\bullet_{\check{G}(\mathscr{O})}({\EuScript Gr}, \EC^\lambda) \\
\cong \HM^\bullet_{\check{I}}(\mathrm{pt}; \FM) \otimes_{\HM^\bullet_{\check{G}(\mathscr{O})}(\mathrm{pt}; \FM)} \HM^\bullet_{\check{G}(\mathscr{O})}({\EuScript Gr}, \EC^{-w_0 \lambda}) \cong \HM^\bullet_{\check{I}}({\EuScript Gr}, \EC_{-\lambda}).
\end{multline*}
(Here we use the autoequivalence considered in~\S\ref{ss:proof-Satake} and the fact that the two natural morphisms $\HM^\bullet_{\check{G}(\mathscr{O})}(\mathrm{pt}; \FM) \rightarrow \HM^\bullet_{\check{G}(\mathscr{O})}({\EuScript Gr}; \FM)$ coincide, see the proof of
Lemma~\ref{lem:morphism-cohomology-factorization}.)
These observations prove the second isomorphism in~\eqref{it:cohom}. The first one follows, using Lemma~\ref{lem:properties-parity}\eqref{it:parity-cohomology}.
Now we prove the first isomorphism in~\eqref{it:cohom-stalks}. By adjunction, for $m \in \ZM$ we have
\begin{multline*}
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM)}^{m}(\Delta^{\mathrm{mix}}_{-\mu}, \EC^{\mathrm{mix}}_{-\lambda} \langle - m \rangle) \\
\cong \Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}_{-\mu},\FM)}(\underline{\FM}_{{\EuScript Gr}_{-\mu}}\{\dim ({\EuScript Gr}_{-\mu})\}, (i_{-\mu})^! \EC^{\mathrm{mix}}_{-\lambda} \{ m \}).
\end{multline*}
(Here, as in~\cite{modrap2}, $\{1\} = \langle - 1 \rangle [1]$ is the autoequivalence of the triangulated category $D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}_{-\mu},\FM)=K^{\mathrm{b}} \Parity_{(\check{I})}({\EuScript Gr}_{-\mu}, \FM)$ induced by the cohomological shift in $\Parity_{(\check{I})}({\EuScript Gr}_{-\mu},\FM)$.)
By~\cite[Remark~2.7]{modrap2}, $(i_{-\mu})^! \EC^{\mathrm{mix}}_{-\lambda}$ is the complex whose $0$-th term is the parity complex $(i_{-\mu})^! \EC_{-\lambda}$, and whose other terms vanish. We deduce that
\begin{multline*}
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr}_{-\mu},\FM)}(\underline{\FM}_{{\EuScript Gr}_{-\mu}}\{\dim({\EuScript Gr}_{-\mu})\}, (i_{-\mu})^! \EC^{ \mathrm{mix}}_{-\lambda} \{ m \}) \\
\cong \HM^{m-\dim({\EuScript Gr}_{-\mu})}({\EuScript Gr}_{-\mu}, (i_{-\mu})^! \EC_{-\lambda}).
\end{multline*}
Finally, using the same considerations as in~\S\ref{ss:proof-Satake} (or as in the proof of~\eqref{it:cohom}), we deduce a canonical isomorphism
\[
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM)}^{m}(\Delta^{\mathrm{mix}}_{-\mu}, \EC^{\mathrm{mix}}_{-\lambda} \langle - m \rangle) \cong \HM^{m-\dim({\EuScript Gr}^\mu)}(\imath_\mu^! \EC^\lambda).
\]
On the other hand, using the equivalence $\Phi$
we obtain that
\[
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM)}^{m}(\Delta^{\mathrm{mix}}_{-\mu}, \EC^{\mathrm{mix}}_{-\lambda} \langle - m \rangle) \cong \Hom_{D^{{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}} \times \Gm}(\widetilde{\NC})}(\Delta^\mu_{\widetilde{\NC}}, \TC^\lambda \langle -m \rangle).
\]
Using~\cite[Equation~(4.10)]{mr} and the fact that $\TC^\lambda \cong \mathsf{T}(\lambda) \otimes \OC_{\widetilde{\NC}}$, we deduce that
\[
\Hom_{D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM)}^{m}(\Delta^{\mathrm{mix}}_{-\mu}, \EC^{\mathrm{mix}}_{-\lambda} \langle - m \rangle) \cong \bigl( \mathsf{T}(\lambda) \otimes \Gamma(\widetilde{\NC}, \OC_{\widetilde{\NC}}(-w_0 \mu))_m \bigr)^{\mathbf G}} \def\gb{{\mathbf g}} \def\GC{{\mathcal{G}},
\]
where the subscript ``$m$'' denotes the $m$-th graded part.
This finishes the proof.
The proof of the second isomorphism in~\eqref{it:cohom-stalks} is similar, using Lemma~\ref{lem:ttilting-dominant} and replacing $D^{\mathrm{mix}}_{(\check{I})}({\EuScript Gr},\FM)$ by $D^{\mathrm{mix}}_{\check{I}}({\EuScript Gr},\FM)$, ordinary cohomology by equivariant cohomology, $\Phi$ by $\Psi$, and $\widetilde{\NC}$ by $\widetilde{\gg}$.
\end{proof}
|
\section{Introduction}\quad
Distributed control of networked cooperative multiagent systems has received much attention in recent years due to the rapid advances in computing and communication technologies. Examples include agreement or consensus problems\cite{RenAtkins_IJRNC_2007,JiEgerstedt_TR_2007,Olfati-SaberMurray_TAC_2004}, in which a group of agents seek to agree upon certain quantities of interest, formation control of robots and vehicles\cite{ConsoliniMorbidiPrattichizzoTosques_Automatica_2008,CaoAndersonMorseYu_CDC_2008}, and distributed estimation\cite{SperanzonFischioneJohansson_CDC_2006,Olfati-SaberShamma_CDC_2005}, etc.
In these applications, an important aspect is to determine the controller actuation schemes. Although continuous feedback of states have always been used in early implementations of distributed control, it is not suitable for agents equipped with embedded microprocessors that have very limited resources to transmit and gather information. To overcome this difficulty, event-triggered control scheme\cite{FanFengWangSong_Automatica_2013,SeybothDimarogonasJohansson_Automatica_2013,LiuChenYuan_JSSC_2012,ManuelTabuada_TAC_2011,WangLemmon_TAC_2011,WangLemmon_CDC_2008,HeemelsSandeeBosch_IJC_2007,Dimarogonas_TAC_2012,Tabuada_TAC_2007} was proposed to reduce the controller updates. In fact, event-driven strategies
for multi-agent systems can be viewed as a linearization and discretization
process, which was considered and investigated in early papers
\cite{LC2004,LC2007}. For example, in the paper \cite{LC2004}, following
algorithm was investigated
\begin{align}
x^{i}(t+1)=f(x^{i}(t))+c_{i}\sum_{j=1}^{m}a_{ij}(f(x^{j}(t)))\label{cml}
\end{align}
which can be considered as nonlinear consensus algorithm. In particular, let $f(x(t))=x(t)$ and $c_{i}=(t_{k+1}^{i}-t_{k}^{i})$.
Then (\ref{cml}) becomes
\begin{align}
x^{i}(t_{k+1}^{i})=x^{i}(t_{k}^{i})+(t_{k+1}^{i}-t_{k}^{i})\sum_{j=1}^{m}a_{ij}x^{j}(t_{k}^{i}),
\end{align}
which is some variant of the event triggering (distributed, self triggered) model for
consensus problem. In centralized control, the bound for
$(t_{k+1}^{i}-t_{k}^{i})=(t_{k+1}-t_{k})$ to reach synchronization was
given in the paper \cite{LC2004} when the coupling graph is indirected and
in \cite{LC2007} for the directed coupling graph.
The key point in event-triggered algorithm is the design of a decision maker that determines when the next actuator update should occur. The existing event-triggered algorithms are all based on the observation of states. For example, in a typical event-triggered algorithm proposed in \cite{Tabuada_TAC_2007}, an update is triggered when a certain error of the states becomes large enough with respect to the norm of the state, which requires a continuous observation of the states. In addition, self-triggered control\cite{AntaTabuada_TAC_2010,ManuelAntaTabuada_Automatica_2010,WangLemmon_TAC_2009,MazoTabuada_CDC_2008,AntaTabuada_PACC_2008} has been proposed as a natural extension of event-triggered control, in which each agent predicts its next update time based on discontinuous state observation to further reduce resource usage for the control systems.
Both the event-triggered and self-triggered control algorithms that have been proposed till now have some drawbacks in common. First, the resulting system in event-triggered control is one that with system delays, especially with self delays, which generally is difficult to handle. And the existing analysis for such algorithms are always based on some quadratic Lyapunov function, which has very strict restrictions on the network structure. For example, to the best of our knowledge, the latest analysis is still restricted to static networks. Second, since these algorithms are based on the observation of states, which are generally much complicated and untraceable, making it difficult to predict and exclude some unexpected possibilities such as the occurrence of zero inter-execution times.
To overcome such difficulties, we proposed a new kind of self-triggered consensus algorithms that are structure-based. That is, each agent predicts its next update time based on its coupling structure with its neighbours instead of the observation of their states. This can bring several advantages. First, since the coupling structure is relatively simpler and more traceable than the states, it is relatively easier to handle. Actually, in our algorithm, the occurrence of zero inter-execution time has been excluded by directly providing a positive lower bound on the update intervals.
Second, although the resulting system of the self-triggered algorithm is one with self delays, by some proper transformation, we showed that the system can corresponding to a discrete-time system without self-delays. We have proposed the algorithm in both centralized and distributed approaches. In the centralized approach, the system can be directly related to a discrete-time system without delays, which can be seen as a discrete version of the nominal system. However, in the distributed approach, the situation is much more complicated and the system can not be directly transformed into its discretized version. However, by some proper indirect transformation, we showed that the convergence of the nominal system can be reduced to that of a discrete-time consensus algorithm with off-diagonal delays but without self delays. Thus, other than using a quadratic Lyapunov function, the convergence of the original algorithm can be solved by the analysis of a discrete-time consensus algorithm. This brings another possibility that is also considered as a big advantage and a major contribution of this work. That is, we can analyze networks with switching topologies, especially stochastically switching topologies.
The rest of the paper is organized as follows: Section \ref{secPreliminaries} provides some preliminaries on matrix and graph theory that will be used in the main text. Section \ref{secSelfTriggered} provides the self-triggered consensus algorithm in both centralized and distributed approach with convergence analysis. Section \ref{secSimulation} provide an example with numerical simulation to illustrate the theoretical results obtained in the previous section. The paper is concluded in Section \ref{secConclusion}.
\section{Preliminaries}\label{secPreliminaries}
In this section, we present some definitions and results in matrix and graph theories and probability theories that will be used later.
For a matrix $A=[a_{ij}]$, $a_{ij}$ represents the entry of $A$ on the $i$th row and $j$th column, which is sometimes also denoted as $[A]_{ij}$. A matrix $A=[a_{ij}]$ is called a {\em nonnegative matrix} if $a_{ij}\ge 0$ for all $i$, $j$. And $A$ is called a {\em stochastic matrix} if $A$ is square and $\sum_{j}a_{ij}=1$ for each $i$. Given a nonnegative matrix $A$ and $\delta>0$, the {\em $\delta$-matrix} of $A$ is a matrix that has nonzero entries only in the place that $A$ has entries equal to or greater than $\delta$. For two matrix $A$, $B$ of the same dimension, we write $A\ge B$ if $A-B$ is a nonnegative matrix. Throughout this paper, we use $\prod_{i=1}^{k}A_{i}=A_{k}A_{k-1}\cdots A_{1}$ to denote the left product of matrices.
A {\em directed graph} $\mathcal{G}$ is defined by its vertex set $\mathcal{V}(\mathcal{G})=\{v_{1},\cdots,v_{n}\}$ and edge set $\mathcal{E}(\mathcal{G})\subseteq \{(v_{i},v_{j}): v_{i}, v_{j}\in \mathcal{V}(\mathcal{G})\}$, where an edge is an ordered pair of vertices in $\mathcal{V}(\mathcal{G})$. If $(v_{i},v_{j})\in \mathcal{E}(\mathcal{G})$, then $v_{i}$ is called the {\em (in-)neighbour} of $v_{j}$. $\mathcal{N}_{i}=\{j:~(v_{j},v_{i})\in \mathcal{E}(\mathcal{G})\}$ is the set of neighbours of agent $i$. A {\em directed path} in $\mathcal{G}$ is an ordered sequence of vertices $v_{1}$, $\cdots$, $v_{s}$ such that $(v_{i},v_{i+1})\in \mathcal{E}(\mathcal{G})$ for $i=1,\cdots,s-1$. A {\em directed tree} is a directed graph where every vertex has exactly one in-neighbour except one ,called the {\em root}, without any in-neighbour. And there exists a directed path from the root to any other vertex of the graph. A subgraph of $\mathcal{G}$ is a directed graph $\mathcal{G}_{S}$ satisfying $\mathcal{V}(\mathcal{G}_{S})\subseteq \mathcal{V}(\mathcal{G})$, and $\mathcal{E}(\mathcal{G})\subseteq \mathcal{E}(\mathcal{G})$. A {\em spanning subgraph} of $\mathcal{G}$ is a subgraph of $\mathcal{G}$ that has the same vertex set with $\mathcal{G}$. We say $\mathcal{G}$ has a {\em spanning tree} if $\mathcal{G}$ has a spanning subgraph that is a tree.
A {\em weighted directed graph} is a directed graph where each edge is equipped with a weight. Thus, a graph $\mathcal{G}$ of $n$ vertices corresponds to an $n\times n$ nonnegative matrix $A=[a_{ij}]$, called the {\em weight matrix} in such way that $(v_{j},v_{i})\in \mathcal{E}(\mathcal{G})$ if and only if $a_{ij}>0$. On the other hand, given an $n\times n$ nonnegative matrix $A$, it corresponds to a weighted directed graph $\mathcal{G}(A)$ such that $\mathcal{G}(A)$ has $A$ as its weight matrix. Using this correspondence, we can introduce the concept of $\delta$-graph. The {\em $\delta$-graph} of $\mathcal{G}$ is a weighted directed graph that has the $\delta$-matrix of $\mathcal{G}$'s weight matrix as its weight matrix. A graph $\mathcal{G}$ has a {\em $\delta$ spanning tree} if its $\delta$-graph has a spanning tree.
From the weight matrix $A=[a_{ij}]$ of a graph $\mathcal{G}$, we define its graph Laplacian $L=[l_{ij}]$ as follows:
\begin{eqnarray*}
l_{ij}=\left\{
\begin{array}{cc}
\sum_{k\ne i}a_{ik}, & i=j;\\
-a_{ij}, & i\ne j.
\end{array}
\right.
\end{eqnarray*}
Thus $L$ has zero row sum with $l_{ii}>0$ while $l_{ij}\le 0$ for $i,j=1,2,\cdots,n$ and $i\ne j$. And the nonnegative linear combination of several graph Laplacians is also a graph Laplacian of some graph. There is a one-to-one correspondence between a graph and its weight matrix or its Laplacian matrix. In the following, for the sake of simplicity in presentation, sometimes we don't explicitly distinguish a graph from its weight matrix or Laplacian matrix, i.e., when we say a matrix has some property that is usually associated with a graph, what we mean is that property is held by the graph corresponding to this matrix. For example, when we say a nonnegative matrix $A$ has a spanning tree, what we mean is that the graph of $A$ has a spanning tree. And it is of similar meaning when we say that a graph has some property that is usually associated with a matrix.
Let $\{\Omega, \mathcal{F},\Prb\}$ be a probability space, where $\Omega$ is the sample space, $\mathcal{F}$ is $\sigma$-algebra on $\Omega$, and $\Prb$ is the probability on $\mathcal{F}$. We use $\E\{\cdot\}$ to denote the mathematical expectation and $\E\{\cdot|\mathcal{F}\}$ the conditional expectation with respect to $\mathcal{F}$, i.e., $\E\{\cdot|\mathcal{F}\}$ is a random variable that is measurable with respect to $\mathcal{F}$.
\begin{definition}[adapted process/sequence]
Let $\{A_{k}\}$ be a stochastic process defined on the basic
probability space $\{\Omega,\mathcal{F},\Prb\}$, and let
$\{\mathcal{F}_{k}\}$ be a filtration, i.e., a sequence of nondecreasing
sub-$\sigma$-algebras of $\mathcal{F}$. If $A_{k}$ is measurable with respect to
$\mathcal{F}_{k}$, then the sequence
$\{A_{k},\mathcal{F}_{k}\}$ is called an adapted process/sequence.
\end{definition}
\begin{remark}
For example, let $X(\omega)$ be a random variable defined on $\Omega$, then $\{\E\{X(\omega)|\mathcal{F}_{k}\},\mathcal{F}_{k}\}$ is an adapted sequence.
\end{remark}
\section{Self-Triggered Consensus Algorithm}\label{secSelfTriggered}
In a network of $n$ agents, with $x_{i}\in \mathbb{R}$ being the state of agent $i$,
we assume that each agent's dynamics obey a single integrator model
\begin{eqnarray}
\dot{x}_{i}(t)=u_{i}(t), ~~i=1,2,\cdots,n,
\end{eqnarray}
where $u_{i}(t)$ is the control law for agent $i$ at time $t$. In consensus algorithm, the control law is usually given by(\cite{Olfati-SaberMurray_TAC_2004,FaxMurray_IFAC_2002,DimarogonasKyriakopoulos_Automatica_2008})
\begin{eqnarray}
u_{i}(t)=-\sum_{j\in \mathcal{N}_{i}}l_{ij}[x_{j}(t)-x_{i}(t)]
=-\sum_{j=1}^{n}l_{ij}[x_{j}(t)-x_{i}(t)].
\end{eqnarray}
Thus the consensus algorithm can be written as
\begin{eqnarray}
\dot{x}_{i}(t)=-\sum_{j=1}^{n}l_{ij}[x_{j}(t)-x_{i}(t)],
\end{eqnarray}
or in matrix form as
\begin{eqnarray}
\dot{x}(t)=-Lx(t),
\end{eqnarray}
where $x(t)=[x_{1}(t),\cdots, x_{n}(t)]^{\top}\in \mathbb{R}^{n}$ is the stack vector of the agents' states. In event/self-triggered control, the control law $u(t)$ is piecewise constant, i.e.,
\begin{eqnarray}
u(t)=u(t_{k}), ~t\in [t_{k},t_{k+1}),
\end{eqnarray}
where $\{t_{k}\}$ is the time sequence that an update of $u(\cdot)$ occurs. The central point of such algorithm is the choice of appropriate $\{t_{k}\}$ such that some desired properties of the algorithm such as stability and convergence can be preserved. This can be done in a centralized approach or a distributed approach, both of which will be discussed in the following.
\subsection{Centralized Approach}
In this subsection, we first consider centralized self-triggered consensus algorithm. The sequence of the update time is denoted by: $t_{0}, t_{1}, t_{2}, \cdots$, then the self-triggered algorithm has the form:
\begin{eqnarray}\label{eqnCentralizedFix}
\dot{x}(t)=-Lx(t_{k}),~t\in [t_{k},t_{k+1}).
\end{eqnarray}
Denote $\Delta t_{k}=t_{k+1}-t_{k}$ for $k=0,1,\cdots$. We have the following Theorem.
\begin{theorem}\label{thmCentralFixed}
If the graph $\mathcal{G}(L)$ has a spanning tree and there exists $\delta\in (0,1/2)$ such that $\delta/l_{\max}\le\Delta t_{k}\le (1-\delta)/l_{\max}$ for each $k$, where $l_{\max}=\max_{i}\{l_{ii}\}$, then the algorithm will achieve consensus asymptotically.
\end{theorem}
\begin{proof}
From the algorithm, for each $t\in (t_{k},t_{k+1}]$, since
$$\dot{x}(t)=-Lx(t_{k}),$$ we have
$$x(t)=x(t_{k})-(t-t_{k})Lx(t_{k})=[I-(t-t_{k})L]x(t_{k}).$$
Particularly, let $t=t_{k+1}$, we have
$$x(t_{k+1})=(I-\Delta t_{k}L)x(t_{k})= A_{k}x(t_{k}),$$
where $A_{k}=I-\Delta t_{k}L$. It is easy to verify that $A_{k}$ is a stochastic matrix for each $k$. In fact, since $\delta/l_{\max}\le\Delta t_{k}\le (1-\delta)/l_{\max}$,
$$[A_{k}]_{ii}=1-\Delta t_{k}l_{ii}\ge 1-\Delta t_{k}l_{\max}\ge 1-\frac{1-\delta}{l_{\max}}l_{\max}=\delta.$$ For $i\ne j$,
$$[A_{k}]_{ij}=-\Delta t_{k} l_{ij}\ge 0.$$ It implies that $[A_{k}]_{ij}$ is positive if and only if $-l_{ij}$ is positive, and the positive elements of $[A_{k}]$ is uniformly lower bounded by a positive scalar since both $\Delta t_{k}$ and $-l_{ij}$ is uniformly lower bounded by a positive scalar. Furthermore,
$\sum_{j=1}^{n}[A_{k}]_{ij}=1-\Delta t_{k}\sum_{j=1}^{n}l_{ij}=1.$
By a standard argument from the theory of products of stochastic matrices, there exists $x^{*}\in \mathbb{R}$ such that
$$\lim_{k\to+\infty}x_{i}(t_{k})=x^{*}.$$
On the other hand, it is not difficult to verify that the function $\max_{i}\{x_{i}(t)\}$ is nonincreasing and $\min_{i}\{x_{i}(t)\}$ is nondecreasing. Thus,
$$\lim_{t\to+\infty}x_{i}(t)=x^{*}.$$
\end{proof}
Using a similar argument as that in Theorem \ref{thmCentralFixed}, it is not difficult to propose a centralized event-triggered algorithm on networks with stochastic switching topologies. For example, consider the following consensus algorithm:
\begin{eqnarray}\label{eqnCentralSwitching}
\dot{x}(t)=-L^{k}x(t_{k}),~t\in [t_{k},t_{k+1}),
\end{eqnarray}
where $L^{k}=[l_{ij}^{k}]$ is uniformly bounded in $k$ and $l_{\max}^{k}=\max_{i}\{l_{ii}^{k}\}\in [l_{\min},l_{\max}]$ for some $0<l_{\min}<l_{\max}$. Before provide the next theorem, we first summarize the following assumption.
\begin{assumption}\label{assumCentralSwitching}
\begin{enumerate}
\item[(i)] There exists $\delta\in (0,1/2)$ such that $\Delta t_{k}\in [\delta/l_{\max}^{k},(1-\delta)/l_{\max}^{k}]$;
\item[(ii)]$\{L^{k},\mathcal{F}_{k}\}$ is an adapted random sequence, and there exists $h>0$, $\delta'>0$ such that the graph corresponding to the conditional expectation $\E\{\sum_{m=k+1}^{k+h}L^{m}|\mathcal{F}_{k}\}$ has a $\delta'$-spanning tree;
\end{enumerate}
\end{assumption}
Now we have
\begin{theorem}
Under assumption\ref{assumCentralSwitching}, the self-triggered consensus algorithm \eqref{eqnCentralSwitching} will reach consensus almost surely.
\end{theorem}
\begin{proof}
Similar as in the proof of Theorem \ref{thmCentralFixed}, we have $x(t_{k+1})=A_{k}x(t_{k})$ with $A_{k}=I-\Delta t_{k}L^{k}$. Since $\{\Delta t_{k}\}$ is a deterministic sequence, $\{A_{k},\mathcal{F}_{k}\}$ is also an adapted sequence, and
\begin{eqnarray*}
\E\{\sum_{m=k+1}^{k+h}A_{m}|\mathcal{F}_{k}\}&=&hI-\E\{\sum_{m=k+1}^{k+h}\Delta t_{m}L^{m}|\mathcal{F}_{k}\}\\
&\ge&h\delta I-\frac{\delta}{l_{\max}}\E\{\sum_{m=k+1}^{k+h}L_{0}^{m}|\mathcal{F}_{k}\},
\end{eqnarray*}
where $[L^{m}_{0}]_{ij}=[L^{m}]_{ij}$ for $i\ne j$ and $[L_{0}^{m}]_{ii}=0$ for all $i$. Thus, $\E\{\sum_{m=k+1}^{k+h}A_{m}|\mathcal{F}_{k}\}$ has a $\delta''$-spanning tree with $\delta''=\delta'\delta/l_{\max}>0$. From Theorem 3.1 of \cite{LiuB_SICON_2011}, the sequence $\{x(t_{k})\}$ will reach consensus almost surely, thus $x(t)$ will also reach consensus almost surely.
\end{proof}
As corollaries in special cases, it can be shown that almost sure consensus still holds if we replace item (ii) in Assumption \ref{assumCentralSwitching} with one of the following more special ones.
\begin{enumerate}
\item[(ii')]$\{L^{k}\}$ is an independent and identically distributed sequence, and there exists $\delta'>0$ such that the graph corresponding to the expectation $\E L^{k}$ has a $\delta'$-spanning tree;
\item[(ii'')]$\{L^{k}\}$ is a homogenous Markov chain with a stationary distribution $\pi$, and there exists $\delta'>0$ such that the graph corresponding to the expectation with respect to $\pi$, $\E_{\pi}L^{k}$, has a $\delta'$-spanning tree;
\end{enumerate}
\subsection{Distributed Approach}
In this section, we discuss the distributed event-triggered consensus algorithm.
Let $\{t_{i}^{k}\}_{k=0}^{+\infty}$ be the time sequence such that $t_{i}^{k}$ is the $k$th time that agent $i$ updates its control law. Then the distributed self-triggered algorithm has the following form:
\begin{eqnarray}
x_{i}(t)=-\sum_{j=1}^{n}l_{ij}x_{j}(t_{j}^{k_{j}(t)}),~~i=1,\cdots,n,
\end{eqnarray}
where $k_{j}(t)=\max\{k:~t_{j}^{k}\le t\}$.
Denote $\Delta t_{i}^{k}=t_{i}^{k+1}-t_{i}^{k}$. Then we have the following Theorem.
\begin{theorem}\label{thmDistributedFixTopology}
If the graph $\mathcal{G}(L)$ has a spanning tree and there exists $\delta_{i}\in (0,1/2)$ for $i=1,2,\cdots,n$ such that $\Delta t_{i}^{k}\in [\delta_{i}/l_{ii},(1-\delta_{i})/l_{ii}]$, then the distributed event triggered algorithm can reach a consensus as $t\to \infty$.
\end{theorem}
The proof will be divided into two steps corresponding to two lemmas. Now we give and prove the first lemma.
\begin{lemma}
If there exists $x^{*}\in \mathbb{R}$ such that
$$
\lim_{k\to \infty}x(t_{i}^{k})=x^{*}, ~~i=1,2,\cdots,n,
$$
then the distributed event-triggered algorithm will reach a consensus, i.e.,
$$
\lim_{t\to\infty}x_{i}(t)=x^{*}, ~~i=1,2,\cdots,n.
$$
\end{lemma}
\begin{proof}
By the definition of the algorithm, for agent $i$,
$$
\dot{x}_{i}(t)=-\sum_{j=1}^{n}l_{ij}x(t_{j}^{k_{j}(t)})=-\sum_{j=1,j\ne i}^{n}l_{ij}[x_{j}(t_{j}^{k_{j}(t)})-x_{i}(t_{i}^{k_{i}(t)})].
$$
Since $\lim_{t\to\infty}k_{j}(t)=+\infty,$ $$
\lim_{t\to\infty}\dot{x}_{i}(t)=-\sum_{j=1,j\ne i}^{n}l_{ij}(x^{*}-x^{*})=0.
$$
Thus,
\begin{eqnarray*}
\lim_{t\to\infty}|x_{i}(t)-x^{*}|&\le & \lim_{t\to\infty}|x_{i}(t)-x_{i}(t_{i}^{k_{i}(t)})|
+\lim_{t\to\infty}|x_{i}(t_{i}^{k_{i}(t)})-x^{*}|\\
&\le&\lim_{t\to\infty}\Delta {t_{i}^{k_{i}(t)}}\max_{s\in [t_{i}^{k_{i}(t)},t]}|\dot{x}_{i}(s)|+\lim_{t\to\infty}|x_{i}(t_{i}^{k_{i}(t)})-x^{*}|\\
&\le& \frac{1-\delta_{i}}{l_{ii}}\lim_{t\to\infty}\max_{s\in [t_{i}^{k_{i}(t)},t]}|\dot{x}_{i}(s)|+\lim_{t\to\infty}|x_{i}(t_{i}^{k_{i}(t)})-x^{*}|\\
&=&0.
\end{eqnarray*}
\end{proof}
Now we are to provide and prove the second lemma.
\begin{lemma}
Under the assumption in Theorem \ref{thmDistributedFixTopology}, there exists $x^{*}$ such that
$$
\lim_{k\to\infty}x_{i}(t_{i}^{k})=x^{*}.
$$
\end{lemma}
\begin{proof}
Let $\{t_{k}\}$ be the time sequence that the $k$th update occurred in the network, i.e., $\{t_{k}\}=\cup_{i=1}^{n}\{t_{i}^{k'}\}$, and $\Delta t_{k}=t_{k+1}-t_{k}$. In case for that $t_{k}=t_{i}^{k'}=t_{j}^{k''}$ for some some $k$, $k'$, $k''$ and $i\ne j$, we consider the update of agent $i$ and $j$ at time $t_{k}$ as one update of the network. Now, construct an auxiliary sequence $\{y(k)\}$, where $y(k)=[y_{1}(k),y_{2}(k),\cdots,y_{n}(k)]^{\top}\in \mathbb{R}^{n}$ with $y_{i}(k)=x_{i}(t_{i}^{k_{i}(t_{k})})$, i.e.,
$y_{i}(k)$ is the latest state that agent $i$ uses in its control law at time $t_{k}$. Now consider the evolution of $\{y(k)\}$.
\begin{enumerate}
\item[Case 1.] Agent $i$ does not update at time $t_{k+1}$:
In this case, by definition,
$$y_{i}(k+1)=y_{i}(k);$$
\item[Case 2.] Agent $i$ does update at time $t_{k+1}$:
In this case, $t_{i}^{k_{i}(t_{k+1})}=t_{k+1}$, assume $t_{i}^{k_{i}(t_{k})}=t_{k-d_{ik}}$ be the last update of agent $i$ before $t_{k+1}$, with $d_{ik}\ge 0$ being the number of updates occur at all other agents $x_{j}$, $j\ne i$, between the two successive updates of agent $i$, i.e., in the interval $(t_{i}^{k_{i}(t_{k})},t_{i}^{k_{i}(t_{k+1})})$. By definition $y_{i}(k)=\cdots=y_{i}(k-d_{ik})$, and
$\dot{x}_{i}(t)=-\sum_{j=1}^{n}l_{ij}y_{j}(k')$ for $t\in (t_{k'},t_{k'+1})$ for all $k'$. Since $[t_{i}^{k_{i}(t_{k})},t_{i}^{k_{i}(t_{k+1})})=\cup_{m=k-d_{ik}}^{k}[t_{m},t_{m+1})$,
\begin{eqnarray*}
y_{i}(k+1)&=&x_{i}(t_{k+1})=x_{i}(t_{i}^{k_{i}(t_{k})})+\int_{t_{i}^{k_{i}(t_{k})}}^{t_{i}^{k_{i}(t_{k+1})}}\dot{x}_{i}(t)dt\\
&=&x_{i}(t_{k-d_{ik}})+\sum_{m=0}^{d_{ik}}\int_{t_{k-d_{ik}+m}}^{k-d_{ik}+m+1}\dot{x}_{i}(t)dt\\
&=&x_{i}(t_{k-d_{ik}})-\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m}\sum_{j=1}^{n}l_{ij}y_{j}(k-d_{ik}+m)\\
&=&y_{i}(k-d_{ik})-\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m}l_{ii}y_{i}(k-d_{ik}+m)\\
&&-\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m}\sum_{j=1,j\ne i}^{n}l_{ij}y_{j}(k-d_{ik}+m)\\
&=&y_{i}(k)-\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m}l_{ii}y_{i}(k)-\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m}\sum_{j=1,j\ne i}^{n}l_{ij}y_{j}(k-d_{ik}+m)\\
&=&(1-l_{ii}\sum_{m=0}^{d_{ik}}\Delta t_{k-d_{ik}+m})y_{i}(k)-\sum_{m=0}^{d_{ik}}\sum_{j=1,j\ne i}^{n}\Delta t_{k-d_{ik}+m}l_{ij}y_{j}(k-d_{ik}+m)\\
&=&(1-\Delta t_{i}^{k_{i}(t_{k})}l_{ii})y_{i}(k)-\sum_{m=0}^{d_{ik}}\sum_{j=1,j\ne i}^{n}\Delta t_{k-d_{ik}+m}l_{ij}y_{j}(k-d_{ik}+m)\\
&=&\sum_{m=0}^{d_{ik}}\sum_{j=1}^{n}a_{ij}^{m}(k)y_{j}(k-m),
\end{eqnarray*}
where $a_{ii}^{0}(k)=1-\Delta t_{i}^{k_{i}(t_{k})}l_{ii}$, $a_{ii}^{m}(k)=0$ for $m=1,2,\cdots,d_{ik}$ and $a_{ij}^{m}(k)=-\Delta t_{k-m}l_{ij}$.
It is easy to verify that
$a_{ii}^{0}(k)=1-\Delta t_{i}^{k_{i}(t_{k})}l_{ii}\ge 1-(1-\delta_{i})=\delta_{i}$, $a_{ij}^{m}=-\Delta t_{k-m}l_{ij}\ge 0$, and
$$
\sum_{m=0}^{d_{ik}}\sum_{j=1}^{n}a_{ij}^{m}(k)=1.
$$
Besides, by the assumption $\delta_{i}/l_{ii}\le\Delta t_{i}^{k}\le (1-\delta_{i})/l_{ii}$ for each $i$, $k$, it is clear that in the interval $[t_{i}^{k},t_{i}^{k+1}]$, only finite updates occur for agents $x_{j}$ with $j\ne i$, and the number of updates is uniformly upper bounded, i.e., there exists $\tau>0$ independent of $i$, $k$, such that $d_{ik}\le \tau$. For example, let $\delta_{\min}=\min_{i}\{\delta_{i}/l_{ii}\}$, and $\delta_{\max}=\max_{i}\{\delta_{i}/l_{ii}\}$, then $\delta_{\min}\le\Delta t_{i}^{k}\le \delta_{\max}$ for all $i$, $k$, and on each time interval of length $m\delta_{\min}$, at most $m+1$ updates occur for each agent. Let $h'$ be the smallest positive integer satisfying $h'\delta_{\min}\ge \delta_{\max}$, then on each time interval of length $\delta_{\max}$, at most $h'+1$ updates occur for each agent. Thus on the interval $(t_{i}^{k},t_{i}^{k+1})$, at most $(n-1)(h'+1)$ updates occur for all agents $j$ with $i\ne j$. Pick $\tau=(n-1)(h'+1)$, then $d_{ik}\le \tau$ for all $i$, $k$.
\end{enumerate}
If agent $i$ does not update at time $t_{k+1}$, i.e., case 1, we define $a_{ij}^{m}(k)=0$ for all $m=0,1,\cdots,\tau$ and $i,j=1,2,\cdots,n$ except for $a_{ii}^{0}(k)=1$,
If agent $i$ updates at time $t_{k+1}$, i.e., case 2, we define $a_{ij}^{m}(k)=0$ for all $i,j=1,2,\cdots,n$ and $m=d_{ik}+1,\cdots,\tau$ when $d_{ik}<\tau$.
For both cases, we can give a uniform iterative formula for $y(k)$ as:
\begin{eqnarray}
y(k+1)=\sum_{l=0}^{\tau}\sum_{j=1}^{n}a_{ij}^{l}(k)y_{j}(k-l),
\end{eqnarray}
where
$a_{ij}^{l}(k)\ge 0$ for all $i,j,l,k$, and
$$
\sum_{l=0}^{\tau}\sum_{j=1}^{n}a_{ij}^{l}(k)=1,
$$
for all $i$, $k$.
Denote $\delta=\min_{i}\{\delta_{i}\}>0$, we have $a_{ii}^{0}(k)\ge \delta$ for all $i$, $k$.
Construct a new matrix $B(k)=[b_{ij}(k)]$, where
\begin{align}
b_{ii}(k)=a_{ii}^{0}(k)
\end{align}
and
\begin{align}
b_{ij}(k)=\sum_{l=0}^{\tau}a_{ij}^{l}(k),~~ if~ i\ne j
\end{align}
Obviously, $B(k)\ge \delta I$, and $b_{ij}(k)=\Delta t_{i}^{k_{i}(t_{k})} l_{ij}\ge \delta_{\min}l_{ij}$ if agent $i$ updates at time $t_{k+1}$. If there exists $h>0$ such that on each interval $[t_{k},t_{k+h}]$ all agents update at least once, then $\sum_{m=k}^{k+h}B(m)\ge -\delta_{\min}L$ for each $k$. Since $\mathcal{G}(L)$ has a spanning tree, there exists $\delta'>0$ such that $\sum_{m=k}^{k+h}B(m)$ has a $\delta'$-spanning tree for each $k$. From Corollary \ref{corollaryAppendixDeterministic} in the appendix, the sequence $\{y(k)\}$ will reach a consensus, i.e., there exists $x^{*}\in \mathbb{R}$ such that
$$
\lim_{k\to\infty}y_{i}(k)=x^{*}.
$$
This is equivalent to
$$
\lim_{k\to\infty}x_{i}(t_{i}^{k})=x^{*}.
$$
At last, we give some hint on how to calculate the quantity $h$ mentioned above. It is easy to see that each agent updates at least once on each time interval of length $\delta_{\max}$, since $\Delta t_{i}^{k}\le \delta_{\max}$ for all $i$, $k$. And in each time interval of length $\delta_{\max}$, there are at most $n(h'+1)$ updates of all the agents, where $h'$ is defined as before. Pick $h=n(h'+1)$, then for each $k$, each agent updates at least once in the time interval $[t_{k},t_{k+h}]$.
\end{proof}
\begin{remark}
It is not difficult to see that this algorithm is applicable to leader-follower networks, in which the leader receives no information from others. For example, if agent $i$ is the leader, then $l_{ii}=0$ and the update time interval $\Delta t_{i}^{0}=+\infty$ by definition, which mean the leader $i$ does not need to update its state. Generally, if agent $i$ wants a long update time interval, all it needs to do is to choose small coupling weights so that $l_{ii}$ is small, and this can be done independent of all other agents.
\end{remark}
From the above analysis, it is not difficult to see that the distributed event-triggered algorithm can also be applied to networks with switching topologies. For example, at each update time $t_{i}^{k}$, agent $i$ may reset its coupling weights $l_{ij}$ such that the algorithm can be written in the form:
\begin{eqnarray}\label{eqnDistributedSwitching}
\dot{x}_{i}(t)=-\sum_{j=1}^{n}l_{ij}^{k}x_{j}(t_{j}^{k_{j}(t)}),~~t\in [t_{i}^{k},t_{i}^{k+1}),~i=1,2,\cdots,n,
\end{eqnarray}
where $\delta_{i}/l_{ii}^{k}\le\Delta t_{i}^{k}\le (1-\delta_{i})/l_{ii}^{k}$. One of the great advantage of this algorithm is that each agent can adjust its update time interval independently based on its own need. If agent $i$ wants a longer update time interval, then it may decrease the coupling weight, otherwise, increase the coupling weight. For example, each agent $i$ may store a scaling factor $\epsilon_{i}^{k}$ such that
\begin{eqnarray}\label{eqnWeightScaling}
l_{ij}^{k}=\epsilon_{i}^{k}l_{ij}.
\end{eqnarray}
By a similar analysis as in the case of fixed topology, we can prove
\begin{theorem}
Suppose that $\epsilon_{\min}$ and $\epsilon_{\max}$ are two positive number satisfying $0<\epsilon_{\min}<\epsilon_{\max}$ and for all $i$, $k$, $\epsilon_{\min}\le\epsilon_{i}^{k}\le \epsilon_{\max}$. If the graph $\mathcal{G}(L)$ has a spanning tree,
then the distributed event-triggered consensus algorithm with switching topology \eqref{eqnDistributedSwitching} and weight updating rule \eqref{eqnWeightScaling} can reach consensus as $t\to \infty$.
\end{theorem}
Finally, we investigate event-triggered consensus algorithm in networks with stochastically switching topologies in which each agent $i$ updates its coupling weights independently at the same time it updates its state.
Given $0<a<b$, and let $S_{i}(a,b)=\{s=[s_{1},\cdots,s_{n}]:~s_{i}\in [a,b], s_{j}\le 0, j\ne i, \sum_{j=1}^{n}s_{j}=0\}$. The event-triggered algorithm can be formulated as follows:
\begin{eqnarray}\label{eqnDistributedIIdSwitching}
\dot{x}_{i}(t)=-\sum_{j=1}^{n}\tilde{l}_{ij}^{k}x_{j}(t_{j}^{k_{j}(t)}),~t\in [t_{i}^{k},t_{i}^{k+1}),
\end{eqnarray}
where $\tilde{L}_{i}^{k}=[\tilde{l}_{i1}^{k},\cdots,\tilde{l}_{in}^{k}]\in S_{i}(a_{i},b_{i})$ for some given $0<a_{i}<b_{i}$, and $\{\tilde{L}_{i}^{k}\}$ is an independent and identically distributed sequence.
Before stating the convergence result, we summarize the following assumption.
\begin{assumption}\label{assumDistributedIIdSwitching}
\begin{enumerate}
\item[(i)] There exist $\delta_{i}\in (0,1/2)$ such that $\Delta t_{i}^{k}\in [\delta_{i}/\tilde{l}_{ii}^{k},(1-\delta_{i})/\tilde{l}_{ii}^{k}]$;
\item[(ii)] There exists $\delta>0$ such that the graph with Laplacian being $\E \tilde{L}^{k}$ has a $\delta$-spanning tree, where $\tilde{L}^{k}=[\tilde{L}_{1}^{k\top},\cdots,\tilde{L}_{n}^{k\top}]^{\top}$ is a matrix with its $i$th row being $\tilde{L}_{i}^{k}$;
\end{enumerate}
\end{assumption}
\begin{theorem}\label{thmDistributedIIdSwitching}
Under Assumption \ref{assumDistributedIIdSwitching}, the event-triggered consensus algorithm \eqref{eqnDistributedIIdSwitching} will reach a consensus almost surely.
\end{theorem}
\begin{proof}
The algorithm \eqref{eqnDistributedIIdSwitching} can be reformulated as:
\begin{eqnarray}\label{eqnDistributedDependentSwitching}
\dot{x}_{i}(t)=-\sum_{j=1}^{n}l_{ij}^{k}x_{j}(t_{j}^{k_{j}(t)}),~t\in [t_{k},t_{k+1}),
\end{eqnarray}
where $l_{ij}^{k}=\tilde{l}_{ij}^{k_{i}(t_{k})}$.
Let $L^{k}=[l_{ij}^{k}]$, then generally the sequence $\{L^{k}\}$ is not an independent sequence since for each $k$, $l_{ij}^{k}=l_{ij}^{k+1}$ if no update of agent $i$ occurs at time $t_{k+1}$. However, similar to the analysis given above, it can be seen that $L^{k}$ is independent of $L^{k'}$ for $k'\ge k+N$, where $N=n(h'+1)$ and
$h'$ is the smallest positive integer satisfying $h'\min_{i}\{\delta_{i}/b_{i}\}\ge \max_{i}\{(1-\delta_{i})/a_{i}\}$.
Similarly, we have
\begin{eqnarray}
y(k+1)=\sum_{l=0}^{N}\sum_{j=1}^{n}a_{ij}^{l}(k)y_{j}(k-l),
\end{eqnarray}
where
$a_{ij}^{l}(k)\ge 0$ for all $i,j,l,k$, and
$$
\sum_{l=0}^{N}\sum_{j=1}^{n}a_{ij}^{l}(k)=1,
$$
for all $i$, $k$.
Since $a_{ij}^{l}(k)=-\Delta t_{k-l}l_{ij}^{k-l}$, we can see that $A^{0}(k),\cdots,A^{N}(k)$ are independent of $A^{0}(k'),\cdots,A^{N}(k')$, when
$k'\ge k+2N$. On the other hand, in each time interval $(t_{k},t_{k+N})$, each agent there at least updates once. Thus we have
$$
\sum_{m=k+1}^{k+N}\E B(k)\ge \delta'\tilde{L}_{0},
$$
where $B(k)$ is defined as before and $\delta'=\min_{i,k}\{\Delta t_{i}^{k}\}=\min_{i}\{\delta_{i}/b_{i}\}>0$, and $[\tilde{L}_{0}]_{ij}=-[\E\tilde{L}^{k}]_{ij}$ for $i\ne j$ and $[\tilde{L}_{0}]_{ii}=0$ for all $i$. This implies that $\sum_{m=k+1}^{k+N}\E B(k)$ has a $\delta''$-spanning tree with $\delta''=\delta\delta'>0$. The conclusion follows from Corollary \ref{corollaryAppendixIndependent} in the Appendix.
\end{proof}
\section{Numerical Simulation}\label{secSimulation}
In this section, we provide an example to illustrate the theoretical results in the previous section. We use the distributed event-triggered algorithm in networks with stochastically switching topologies as considered in Theorem \ref{thmDistributedIIdSwitching}.
Consider a network of four agents with
\begin{align*}
&\tilde{L}_{1}^{k}\in \{[1, -1 ,0 ,0], [1,0, 0,-1]\}, \\
&\tilde{L}_{2}^{k}=\{[-1,1,0,0], [0,1,-1,0]\},\\
&\tilde{L}_{3}^{k}\in \{[0,-1,1,0], [0,0,1,-1]\},\\
&\tilde{L}_{4}^{k}\in \{[0,-1,0,1], [0,0,-1,1]\},
\end{align*}
and each agent selects its coupling weights using a uniform distribution. We choose $\delta_{i}=0.1$ for all the agents. The next update time is randomly chosen from the permissible range. It is not difficult to verify that the conditions in Theorem \ref{thmDistributedIIdSwitching} are satisfied. The simulation results are provided in Fig. \ref{figIIdSwitching} with the initial value of the four agents being randomly chosen. It can be seen that the agents actually reached consensus.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{figIIdSwitching.eps}
\caption{Self-triggered consensus in networks of four agents with stochastically switching topologies.}\label{figIIdSwitching}
\end{figure}
\section{Conclusion}\label{secConclusion}
In this paper, we proposed a new structure-based self-triggered consensus algorithm in both centralized approach and distributed approach. Different from existing works that used a quadratic Lyapunov function as the analysis tool, we reduce the self-triggered consensus algorithm to a discrete-time consensus algorithm by some proper transformation, which enables the application of the product theory of stochastic matrices to the convergence analysis. Compared to existing work, our method provides several advantages. First, each agent does not need to calculate the system error to determine its update time. Second, it provides explicit positive lower and upper bounds for the update interval of each agent based on its coupling weights, which can be adjusted by each agent independent of other agents. We also used our method to investigate networks with switching topologies, especially networks with stochastically switching topologies. Our work reveals that the event/self-triggered algorithms are essentially discrete and thus more suitable to a discrete analysis framework. And it is predictable that more results can be obtained for self-triggered algorithms under our new analysis framework in the future.
\section*{Appendix: Discrete-time consensus analysis}
The appendix is devoted to the discussion of discrete-time consensus algorithms in networks of multi-agents with time delays based on the product theory of stochastic matrices and provide some results that is used in the main text.
\subsection{Preliminaries}
This subsection will provide some preliminaries on stochastic matrices that will be used in the proof of Theorem \ref{thmAppendixMain} in the next subsection.
A nonnegative matrix $A$ is called a {\em stochastic indecomposable and aperiodic (SIA)} if it is a stochastic matrix and there exists a column vector $v$ such that $\lim_{k\to\infty}A^{k}={\bf e}_{n}v^{\top}$, where ${\bf e}_{n}$ is an $n$-dimensional column vector with all entries $1$. A sequence of stochastic matrices $\{A_{k}\}$ is {\em ergodic} if there exists a column vector $v$ such that $\lim_{k\to\infty}\prod_{k=1}^{\infty}A_{k}={\bf e}_{n}v^{\top}$. $A$ is called {\em $\delta$-SIA} if its $\delta$-matrix is SIA. A stochastic matrix $A$ is called {\em scrambling} if for any pair ($i$, $j$), there exists an index $k$ such that both $a_{ik}>0$ and $a_{jk}>0$. Similarly, $A$ is called {\em $\delta$-scrambling} if its $\delta$-matrix is scrambling. For a matrix $A$, $\diag(A)$ refers to the diagonal matrix formed by the diagonal elements of $A$, i.e., $[\diag(A)]_{ii}=[A]_{ii}$ for all $i$ and $[\diag(A)]_{ij}=0$ for $i\ne j$. And we define $\overline{\diag}(A)=A-\diag(A)$.
We use $\Prb\{\cdot|\mathcal{F}\}$ to denote the conditional probability with respect to $\mathcal{F}$ on a probability space $\{\Omega,\Prb,\mathcal{F}\}$, which can also be expressed into the conditional expectation of an indicator function, i.e.,
$\Prb\{\cdot|\mathcal{F}\}=\E\{{\bf 1}_{\{\cdot\}}|\mathcal{F}\}$, where ${\bf 1}_{\{\cdot\}}$ is an indicator function which is defined as
\begin{eqnarray*}
{\bf 1}_{S}(x)=\left\{
\begin{array}{cc}
1, & x\in S,\\
0, & x\not\in S.
\end{array}
\right.
\end{eqnarray*}
Thus, $\Prb\{\cdot|\mathcal{F}\}$ is also a random variable that is measurable with respect to $\mathcal{F}$.
In the following, we will introduce some lemmas that will be used in the proof of the main results.
Let $A=[a_{ij}]$ be an $n\times n$ stochastic matrix. Define
\begin{eqnarray*}
\Delta(A)=\max_{j}\max_{i_{1},i_{2}}|a_{i_{1}j}-a_{i_{2}j}|.
\end{eqnarray*}
It can be seen that $\Delta(A)$ measures how different the rows of $A$ are. $\Delta(A)=0$ if and only if the rows of $A$ are identical.
Define
\begin{eqnarray*}
\lambda(A)=1-\min_{i_{1},i_{2}}\sum_{j}\min\{a_{i_{1}j},a_{i_{2}j}\}.
\end{eqnarray*}
\begin{remark}
It can be seen that if $A$ is scrambling, then $\lambda<1$, if $A$ is $\delta$-scrambling for some $\delta>0$, then $\lambda(A)\le 1-\delta$.
\end{remark}
\begin{lemma}\cite{Wolfowitz}\label{lemContractionOfProduct}
For any stochastic matrices $A_{1}$, $A_{2}$, $\cdots$, $A_{k}$, $k>0$,
\begin{eqnarray*}
\Delta(A_{1}A_{2}\cdots A_{k})\le \prod_{i=1}^{k}\lambda(A_{i}).
\end{eqnarray*}
\end{lemma}
\begin{remark}\label{remErgodicityOfProducts}
From Lemma \ref{lemContractionOfProduct}, it can be seen that for a sequence of stochastic matrices $\{A_{k}\}$, if there exists a sequence $\{k_{i}\}$ such that $k_{i}<k_{i+1}$, $\lim_{i\to\infty}k_{i}=\infty$, and $\prod_{i=1}^{\infty}\lambda(\prod_{m=k_{i}+1}^{k_{i+1}}A_{m})=0$, then $\lim_{k\to\infty}\Delta(\prod_{m=1}^{k} A_{m})=0$, i.e., $\{A_{k}\}$ is ergodic. Particularly, from the property of scrablingness, if there exists $\delta>0$ such that there are infinite matrices in the sequence $\{\prod_{m=k_{i}+1}^{k_{i+1}}A_{m}\}$ that are $\delta$-scrambling, then $\{A_{k}\}$ is ergodic.
\end{remark}
\begin{lemma}\cite{Wolfowitz,XiaoWang_TAC_2008}\label{lemSIAtoScrambling}
Let $A_{1}$, $A_{2}$, $\cdots$, $A_{k}$ (repetitions permitted) be $r\times r$ SIA matrices with the property that for any $1<k_{1}<k_{2}\le k$, $\prod_{i=k_{1}}^{k_{2}}A_{i}$ is SIA. If $k>{\bf n}(r)$, then $\prod_{i=1}^{k}A_{i}$ is a scrambling matrix. Here ${\bf n}(r)$ is the number of different types of all $r\times r$ stochastic matrices.
\end{lemma}
\begin{lemma}\cite{XiaoWang_TAC_2008}\label{lemHighDimensionSpanningTree}
Let $A_{1}$, $\cdots$, $A_{m}$ be $n\times n$ nonnegative matrices, let
\begin{eqnarray*}
D=\left[
\begin{array}{cccc}
A_{1}& A_{2}& \cdots & A_{m}\\
0& 0& \cdots & 0\\
\cdots & \cdots & \ddots & \cdots\\
0 & 0& \cdots &0
\end{array}
\right]_{mn\times mn},
\end{eqnarray*}
let
\begin{eqnarray*}
M_{0}=\left[
\begin{array}{ccccc}
I & 0& \cdots& 0&0\\
I& 0&\cdots& 0&0\\
0&I&\cdots& 0&0\\
\vdots& \vdots&\ddots& 0&0\\
0&0&\cdots& I&0
\end{array}
\right]_{mn\times mn},
\end{eqnarray*}
and let $M_{k}=D+M_{0}^{k}$ for any $k\in \{1,2,\cdots,m-1\}$. Then if $\mathcal{G}(\sum_{i=1}^{m}A_{i})$ contains a spanning tree, then $\mathcal{G}(M_{k})$ contains a spanning tree with the property that the root vertex of the spanning tree has a self-loop in $\mathcal{G}(M_{k})$.
\end{lemma}
\begin{remark}
Actually, since $M_{0}^{k}=M_{0}^{m-1}=\left[\begin{array}{cccc}I & 0 & \cdots & 0 \\
I & 0 & \cdots & 0 \\ \vdots & \vdots & \ddots & \vdots \\
I & 0 & \cdots & 0\end{array}\right]$ for all $k\ge m-1$, thus Lemma \ref{lemHighDimensionSpanningTree} holds for all $k\in \mathbb{N}$. This is important for our further results based on this lemma.
\end{remark}
\begin{remark}
If the condition is $\mathcal{G}(\sum_{i=1}^{m}A_{i})$ has a $\delta$-spanning tree for some $\delta>0$, then the conclusion becomes that $\mathcal{G}(M_{k})$ has a $\delta'$-spanning tree whose root vertex has a self-loop, where $\delta'\ge\delta/m$.
\end{remark}
\begin{lemma}\cite{WangL_JCtrl_2006}\label{lemSpanningTreeToSIA}
Let $A$ be a stochastic matrix. If $\mathcal{G}(A)$ contains a spanning tree with the property that the root vertex of the spanning tree has a self-loop in $\mathcal{G}(A)$, then A is SIA.
\end{lemma}
From Lemma \ref{lemHighDimensionSpanningTree} and \ref{lemSpanningTreeToSIA}, we can have the following Lemma.
\begin{lemma}\label{lemProdHighDimensionSpanningTree}
Let $A_{1}^{i}$, $\cdots$, $A_{m}^{i}$, $i=1,\cdots,k$ be $n\times n$ nonnegative matrices. Let
\begin{eqnarray*}
D_{i}=\left[
\begin{array}{cccc}
A_{1}^{i}& A_{2}^{i}& \cdots & A_{m}^{i}\\
0& 0& \cdots & 0\\
\cdots & \cdots & \ddots & \cdots\\
0 & 0& \cdots &0
\end{array}
\right]_{mn\times mn},
\end{eqnarray*}
and $M_{0}$ being defined as in Lemma \ref{lemHighDimensionSpanningTree}. Then if $\mathcal{G}(\sum_{i=1}^{k}\sum_{j=1}^{m}A_{j}^{i})$ has a $\delta$-spanning tree for some $\delta>0$, then the product
$\prod_{i=1}^{k}(D_{i}+M_{0})$ has a $\delta'$-spanning tree for some $0<\delta'<\delta$. And the root of the spanning tree has a self-loop, thus $\prod_{i=1}^{k}(D_{i}+M_{0})$ is $\delta'$-SIA.
\end{lemma}
\begin{remark}\label{remProdHighDimensionSpanningTree}
It is obvious from Lemma \ref{lemHighDimensionSpanningTree} that if there exists $\epsilon>0$ such that $D'_{i}\ge \epsilon (M_{0}+D_{i})$, then $\prod_{i=1}^{k}D_{i}'$ is $\delta'$-SIA for some $0<\delta'<\delta$. This is the case that will appear in the proof of Theorem \ref{thmAppendixMain}.
\end{remark}
\begin{proof}First, we have
\begin{eqnarray}\label{eqnProductSIA}
\prod_{i=1}^{k}(M_{0}+D_{i})\ge M_{0}^{k}+\sum_{i=1}^{k}M_{0}^{k-i}D_{i}M_{0}^{i-1}\ge M_{0}^{k}+\sum_{i=1}^{k}D_{i}M_{0}^{i-1},
\end{eqnarray}
where the second inequality is due to the fact that $M_{0}^{j}D_{i}\ge D_{i}$ for any $j\ge 0$. And it is not difficult to verify that the first block row sum of $D_{i}$ is preserved in $D_{i}M_{0}^{i-1}$ for $i=1,\cdots,k$. Thus the first block row sum of $\sum_{i=1}^{k}D_{i}M_{0}^{i-1}$ equals that of $\sum_{i=1}^{k}D_{i}$, i.e., $\sum_{i=1}^{k}\sum_{j=1}^{m}A_{j}^{i}$. This means $\mathcal{G}(\sum_{i=1}^{k}D_{i}M_{0}^{i-1})$ has a $\delta$-spanning tree. From Lemma \ref{lemHighDimensionSpanningTree}, $\mathcal{G}(M_{0}^{k}+\sum_{i=1}^{k}D_{i}M_{0}^{i-1})$ has a $\delta'$-spanning tree for some $0<\delta'<\delta$, and the root has a self-loop. This is also true for the product $\prod_{i=1}^{k}(M_{0}+D_{i})$ due to \eqref{eqnProductSIA}. Thus, $\prod_{i=1}^{k}(M_{0}+D_{i})$ is $\delta'$-SIA from Lemma \ref{lemSpanningTreeToSIA}. The proof is completed.
\end{proof}
From Lemma \ref{lemProdHighDimensionSpanningTree}, we can easily obtain the following corollary.
\begin{corollary}\label{corollaryProdHighDimensionSpanningTree}
Let $A_{1}^{i}$, $\cdots$, $A_{m}^{i}$, $i=1,2,\cdots,$ be uniformly bounded $n\times n$ nonnegative matrices. Let $D_{i}$ and $M_{0}$ be defined as in Lemma \ref{lemProdHighDimensionSpanningTree}. If there exists $k$ such that $\mathcal{G}(\sum_{i=1}^{k}\sum_{j=1}^{m}A_{j}^{i})$ has a $\delta$-spanning tree for some $\delta>0$, then for each $k'>k$, the product
$\prod_{i=1}^{k'}(D_{i}+M_{0})$ is $\delta'(k')$-SIA, where $0<\delta'(k')<\delta$ depends on $k'$.
\end{corollary}
\begin{proof}
Since $A_{j}^{i}\ge 0$, if $\mathcal{G}(\sum_{i=1}^{k}\sum_{j=1}^{m}A_{j}^{i})$ has a $\delta$-spanning tree for some $\delta>0$, then for each $k'>k$, $\mathcal{G}(\sum_{i=1}^{k'}\sum_{j=1}^{m}A_{j}^{i})$ also has a $\delta$-spanning tree. The conclusion is obvious from Lemma \ref{lemProdHighDimensionSpanningTree}.
\end{proof}
\begin{remark}\label{remProductSIA}
If $A_{1}^{i}$, $\cdots$, $A_{m}^{i}$, $i=1,2,\cdots,$ is a random sequence, then it is obvious from Corollary \ref{corollaryProdHighDimensionSpanningTree} that the event $\mathcal{G}(\sum_{i=1}^{k}\sum_{j=1}^{m}A_{j}^{i})$ has a $\delta$-spanning tree is contained in the event $\prod_{i=1}^{k'}(D_{i}+M_{0})$ is $\delta'(k')$-SIA.
\end{remark}
\begin{lemma}[second Borel--Cantelli lemma \cite{Durrett}]\label{lemSecondBorelCantelli}
Let $\{\mathcal{F}_{n}, n\ge 0\}$ be a filtration with
$\mathcal{F}_{0}=\{\emptyset,\Omega\}$ and $\{X_{n}, n\ge 1\}$ a
sequence of events with $X_{n}\in\mathcal{F}_{n}$. Then
\begin{eqnarray*}
\{X_{n}~\textnormal{ \it occurs~infinitely~often
}\}=\Big\{\sum_{n=1}^{+\infty}\Prb\{X_{n}|\mathcal{F}_{n-1}\}=+\infty\Big\},
\end{eqnarray*}
with a probability $1$, where ``infinitely often" means that an
infinite number of events from $\{X_{n}\}_{n=1}^{\infty}$ occur.
\end{lemma}
\begin{lemma}\label{lemProbabilityToLongerSeqence}
Let $\{A_{k},\mathcal{F}_{k}\}$ be an adapted random sequence. If for a sequence $\{X_{k}\}$ defined as $X_{k}=\{A_{k}\in S_{k}\}$ for some given set $S_{k}$ there exists $\delta>0$ such that the conditional probability
$\Prb\{X_{m+1}|\mathcal{F}_{m}\}>\delta$, then for each $m, h>0$,
\begin{eqnarray*}
\Prb\{X_{m+1},\cdots,X_{m+h}|\mathcal{F}_{m}\}>\delta^{h}.
\end{eqnarray*}
\end{lemma}
\begin{proof}
By the definition of conditional probability, we have
\begin{eqnarray*}
&&\Prb\{X_{m+1},\cdots,X_{m+h}|\mathcal{F}_{m}\}\\
&=&\E\{{\bf 1}_{X_{m+1}}\cdots{\bf 1}_{X_{m+h}}|\mathcal{F}_{m}\}\\
&=&\E\{{\bf 1}_{X_{m+1}}\cdots\E\{{\bf 1}_{X_{m+h}}|\mathcal{F}_{m+h-1}\}|\mathcal{F}_{m}\}\\
&\ge&\delta\E\{{\bf 1}_{X_{m+1}}\cdots{\bf 1}_{X_{m+h-1}}|\mathcal{F}_{m}\}\\
&\ge&\delta^{2}\E\{{\bf 1}_{X_{m+1}}\cdots{\bf 1}_{X_{m+h-2}}|\mathcal{F}_{m}\}\\
&\ge&\cdots\\
&\ge&\delta^{h}.
\end{eqnarray*}
\end{proof}
\begin{lemma}(Lemma 5.4 in \cite{LiuB_SICON_2011})\label{lemExpectationToProbability1}
Let $\{\Omega, \mathcal{F},\Prb\}$ be a probability space, and $f$ be a
random variable with $0\le f\le 1$. If for a $\sigma$-algebra
$\mathcal{F}'\subseteq \mathcal{F}$, a set $S\in \mathcal{F}'$ with
$\Prb\{S\}>0$, $\E\{f|\mathcal{F}'\}\ge\delta$ holds on $S$ for some
$\delta>0$, then we have
\begin{eqnarray*}
\Prb\{f\ge\frac{\delta}{2}|\mathcal{F}'\}\ge\frac{\delta}{2},
\end{eqnarray*}
on $S$ and particularly,
\begin{eqnarray*}
\E\{f{\bf
1}_{\{f\ge\frac{\delta}{2}\}}|\mathcal{F}'\}\ge\frac{\delta^{2}}{4}
\end{eqnarray*}
on $S$.
\end{lemma}
\subsection{Discrete-time delayed consensus analysis}
In this section, we will discuss discrete-time consensus algorithms in networks with switching topologies and time delays. We first consider general stochastic processes described by an adapted sequence and establish a most general result. Then we use it to obtain two corollaries under two special cases that will appear in the main text.
Consider the following discrete time dynamical systems with switching topologies:
\begin{eqnarray}\label{eqnDiscrete}
x_{i}(k+1)=\sum_{l=0}^{\tau}\sum_{j=1}^{n}a_{ij}^{l}(k)x_{j}(k-l),~~i=1,\cdots,n,
\end{eqnarray}
where $k$ is the time index, $x(k)=[x_{1}(k),\cdots,x_{n}(k)]^{\top}\in \mathbb{R}^{n}$ is the state variable of the system at time $k$, $\tau\ge 0$ is the bound of the time delay, where $\tau=0$ corresponds to the case of no time delay.
In the following, we always assume that $a_{ii}^{0}(k)>0$, $a_{ii}^{l}(k)=0$ for $l\ne 0$, $a_{ij}^{l}(k)\ge 0$ for each $i$, $j$, $l$, $k$, and
\begin{eqnarray*}
\sum_{l=0}^{\tau}\sum_{j=1}^{n}a_{ij}^{l}(k)=1
\end{eqnarray*}
holds for each $k$ and $i$.
Define $A^{l}(k)=[a_{ij}^{l}(k)]$, and $B(k)=[b_{ij}(k)]=\sum_{l=0}^{\tau}A^{l}(k)$.
We have the following Theorem.
\begin{theorem}\label{thmAppendixMain}
Let $\{A^{0}(k),\cdots,A^{\tau}(k),\mathcal{F}_{k}\}$ be an adapted process. If there exists $\delta>0$, $h>0$ such that $B(k)\ge \delta I$ and the conditional expectation $\E\{\sum_{k=m+1}^{m+h}B(k)|\mathcal{F}_{m}\}$ has a $\delta$-spanning tree, then the system (\ref{eqnDiscrete}) will reach consensus almost surely.
\end{theorem}
\begin{remark}
When $\tau=0$, the system has no time delays,and Theorem \ref{thmAppendixMain} reduces to Theorem 3.1 of \cite{LiuB_SICON_2011}. From this point of view, this result can be seen as an extension of that in \cite{LiuB_SICON_2011}.
\end{remark}
The proof of this theorem will be divided into several steps.
First, we will prove the following lemma, which shows that the conditional expectation of a spanning tree implies a positive conditional probability of existence of a spanning tree for a longer length of the summation.
\begin{lemma}\label{lemExpectationToProbability}
Let $\{A_{k},\mathcal{F}_{k}\}$ be an adapted random sequence of $n\times n$ stochastic matrices, if there exists $\delta>0$ such that for each $m$, $\E\{A_{m+1}|\mathcal{F}_{m}\}$ has a $\delta$ -spanning tree, then there exist $h>0$, $0<\delta'<\delta$ such that
\begin{eqnarray}\label{eqnProbabilitySpanningTree}
\Prb\{\sum_{k=m+1}^{m+h}A_{k} \textnormal{ has a $\delta'$-spanning tree }|\mathcal{F}_{m}\}>\delta'.
\end{eqnarray}
\end{lemma}
\begin{proof}
Denote $\Sp(n)$ the number of different types of spanning trees composed of $n$ vertices. For some fixed $m$, we decompose the probability space $\Omega$ with respect to $\mathcal{F}_{m}$, i.e., $S_{m}^{(i_{1})}\in \mathcal{F}_{m}$, $\Omega=\cup S_{m}^{(i_{1})}$, $i_{1}=1,2,\cdots,s_{1}$ where $s_{1}\le \Sp(n)$, and $S_{m}^{(i)}\cap S_{m}^{(j)}=\emptyset$ for $i\ne j$ such that on each $S_{m}^{(i)}$, $\E\{A_{m+1}~|~\mathcal{F}_{m}\}$ has a (fixed) specified spanning tree. Denote $\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}}$ the restriction of $\E\{A_{m}~|~\mathcal{F}_{m}\}$ on ${S_{m}^{(i_{1})}}$, i.e.,
\begin{eqnarray*} \E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}}=\E\{A_{m+1}~|~\mathcal{F}_{m}\}\times{\bf 1}_{S_{m}^{(i_{1})}}.
\end{eqnarray*}
Let $\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}})$ be a $\delta$-spanning tree contained in $\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}}$, which is arbitrarily selected when more than one choices are available. For each $S_{m}^{(i_{1})}$, we pick an edge $(j,i)\in \mathcal{E}(\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}}))$,
and let $S_{m+1}^{(i_{1})}=\{[A_{m+1}]_{ij}\ge \delta/2\}\cap S_{m}^{(i_{1})}$. Then $S_{m+1}^{(i_{1})}\in \mathcal{F}_{m+1}$, and from Lemma \ref{lemExpectationToProbability1},
\begin{eqnarray*}
\Prb\{S_{m+1}^{(i_{1})}~|~\mathcal{F}_{m}\}\ge \frac{\delta}{2}
\end{eqnarray*}
holds on $S_{m}^{(i_{1})}$.
Similarly, we decompose each $S_{m+1}^{(i_{1})}$ with respect to $\mathcal{F}_{m+1}$, i.e., $S_{m+1}^{(i_{1})}=\cup_{i_{2}}S_{m+1}^{(i_{1}i_{2})}$ with $S_{m+1}^{(i_{1}i_{2})}\in\mathcal{F}_{m+1}$ and $S_{m+1}^{(i_{1}i_{2})}\cap S_{m+1}^{(i_{1}i_{2}')}$ for $i_{2}, i_{2}'=1,2,\cdots,s_{2}$, $i_{2}\ne i_{2}'$, where $s_{2}\le \Sp(n)$ depends on the index $i_{1}$ such that $\E\{A_{m+2}~|~\mathcal{F}_{m+1}\}_{S_{m+1}^{(i_{1}i_{2})}}$ has a specified $\delta$-spanning tree on each $S_{m+1}^{(i_{1}i_{2})}$.
For each $S_{m+1}^{(i_{1}i_{2})}$, we pick an edge $(j,i)\in \mathcal{E}(\ST_{\delta}(\E\{A_{m+2}~|~\mathcal{F}_{m+1}\}_{S_{m+1}^{(i_{1}i_{2})}}))$ and let $S_{m+2}^{(i_{1}i_{2})}=\{[A_{m+2}]_{ij}\ge \delta/2\}\cap S_{m+1}^{(i_{1}i_{2})}$, then $S_{m+2}^{(i_{1}i_{2})}\in \mathcal{F}_{m+2}$ and from Lemma \ref{lemExpectationToProbability1},
\begin{eqnarray*}
\Prb\{S_{m+2}^{(i_{1}i_{2})}~|~\mathcal{F}_{m+1}\}\ge \frac{\delta}{2}
\end{eqnarray*}
holds on $S_{m+1}^{(i_{1}i_{2})}$.
Continuing this process, we can get a sequence:
\begin{eqnarray*}
S_{m}^{(i_{1})},S_{m+1}^{(i_{1})}, S_{m+1}^{(i_{1}i_{2})},S_{m+2}^{(i_{1}i_{2})},\cdots,S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}, S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}
\end{eqnarray*}
such that for $l=1,2,\cdots,k$, $S_{m+l}^{(i_{1}i_{2}\cdots i_{l})}\in \mathcal{F}_{m+l}$, $S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l-1})}=\cup_{i_{l}} S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}$, $\E\{A_{m+l}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}}$ has a (fixed) spanning tree, and
\begin{eqnarray}\label{eqnChooseEdge}
S_{m+l}^{(i_{1}i_{2}\cdots i_{l})}=\{[A_{m+l}]_{ij}>\delta/2\}\cap S_{m+l-1}^{i_{1}i_{2}\cdots i_{l}},
\end{eqnarray}
where the edge $(j,i)\in \mathcal{E}(\ST_{\delta}(\E\{A_{m+l}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}}))$.
For any fixed sequence $i_{1},i_{2},\cdots, i_{k}$, $S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}\subseteq S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}\subseteq S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1})}\subseteq \cdots \subseteq S_{m+1}^{(i_{1})}$. We choose the edge $(j,i)$ in (\ref{eqnChooseEdge}) for each $S_{m+l}^{(i_{1}i_{2}\cdots i_{l})}$ in such way that if $\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}})$ hasn't appeared in the sequence, $\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1} )}})$, $\cdots$, $\ST_{\delta}(\E\{A_{m+1-1}~|~\mathcal{F}_{m+l-2}\}_{S_{m+l-2}^{(i_{1}i_{2}\cdots i_{l-1})}})$, then we choose $(j,i)\in \mathcal{E}(\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}}))$ arbitrarily. Otherwise, we choose $(j,i)\in \mathcal{E}(\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}}))$ which hasn't been chosen before when \\ $\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m+l-1}\}_{S_{m+l-1}^{(i_{1}i_{2}\cdots i_{l})}})$ appears in the sequence \\ $\ST_{\delta}(\E\{A_{m+1}~|~\mathcal{F}_{m}\}_{S_{m}^{(i_{1})}})$, $\cdots$, $\ST_{\delta}(\E\{A_{m+1-1}~|~\mathcal{F}_{m+l-2}\}_{S_{m+l-2}^{(i_{1}i_{2}\cdots i_{l-1})}})$ if there still exists such an edge.
Since there at most $\Sp(n)$ different types of spanning trees, and each spanning tree has $n-1$ edges, it is obvious when $k=(n-1)\Sp(n)$, the sets
\begin{eqnarray*}
S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}\subseteq \{\sum_{l=1}^{k}A_{m+l} \textnormal{ has $\delta/2$-spanning tree} \}.
\end{eqnarray*}
This is because for each fixed sequence $i_{1}$, $\cdots$, $i_{k}$, at least one type of spanning trees has appeared at least $n-1$ times, thus by the edge chosen strategy, each of its edges has been chosen at least once.
So we have
\begin{eqnarray}
&&\Prb\{\sum_{l=1}^{k}A_{m+l} \textnormal{ has $\delta/2$-spanning tree}~|~\mathcal{F}_{m}\}\\
&\ge & \Prb\{\cup_{i_{1},i_{2},\cdots,i_{k}}S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}~|~\mathcal{F}_{m}\}\nonumber\\
&=&\E\{\sum_{i_{1},i_{2},\cdots,i_{k}}{\bf 1}_{S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}}~|~\mathcal{F}_{m}\}\nonumber\\
&=&\E\{\E\{\sum_{i_{1},i_{2},\cdots,i_{k}}{\bf 1}_{S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}}~|~\mathcal{F}_{m+k-1}\}~|~\mathcal{F}_{m}\}\label{stepIneq1}\\
&\ge&\frac{\delta}{2}\E\{\sum_{i_{1},i_{2},\cdots,i_{k}}{\bf 1}_{S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}}~|~\mathcal{F}_{m}\}\label{stepEq1}\\
&=&\frac{\delta}{2}\E\{\sum_{i_{1},i_{2},\cdots,i_{k-1}}{\bf 1}_{S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1})}}~|~\mathcal{F}_{m}\}\label{stepEq2}\\
&\ge&\cdots\nonumber\\
&\ge&(\frac{\delta}{2})^{k-1}\E\{\sum_{i_{1}}{\bf 1}_{S_{m+1}^{(i_{1})}}~|~\mathcal{F}_{m}\}\nonumber\\
&\ge&(\frac{\delta}{2})^{k}\E\{\sum_{i_{1}}{\bf 1}_{S_{m}^{(i_{1})}}~|~\mathcal{F}_{m}\}\nonumber\\
&=&(\frac{\delta}{2})^{k}\nonumber,
\end{eqnarray}
The inequality from (\ref{stepIneq1}) to (\ref{stepEq1}) is due to the fact
\begin{eqnarray*}
\Prb\{S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}~|~\mathcal{F}_{m+k-1}\}\ge \delta/2
\end{eqnarray*}
on $S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}\in \mathcal{F}_{m+k-1}$, which implies
\begin{eqnarray*}
\E\{{\bf 1}_{S_{m+k}^{(i_{1}i_{2}\cdots i_{k})}}~|~\mathcal{F}_{m+k-1}\}\ge \frac{\delta}{2} {\bf 1}_{S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}}.
\end{eqnarray*}
The equality from (\ref{stepEq1}) to (\ref{stepEq2}) is due to the fact that $S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1})}=\cup_{i_{k}}S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k})}$,
and $S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1}j)}\cap S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1}j')}=\emptyset$ for $j\ne j'$, which implies
\begin{eqnarray*}
{\bf 1}_{S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1})}}=\sum_{i_{k}}{\bf 1}_{S_{m+k-1}^{(i_{1}i_{2}\cdots i_{k-1}i_{k})}}
\end{eqnarray*}
for each fixed sequence $i_{1}$, $i_{2}$, $\cdots$, $i_{k-1}$.
Let $h=(n-1)\Sp(n)$, and $\delta'=(\delta/2)^{h}$, then (\ref{eqnProbabilitySpanningTree}) holds. The proof is completed.
\end{proof}
Now we come to the proof of Theorem \ref{thmAppendixMain}.
{\em Proof of Theorem \ref{thmAppendixMain}}
Denote $y_{k}=[x(k), x(k-1),\cdots, x(k-\tau)]^{\top}$, then we have:
\begin{eqnarray}\label{eqnHighDimension}
y_{k+1}=C_{k}y_{k},
\end{eqnarray}
where
\begin{eqnarray*}
C_{k}=\left[\begin{array}{ccccc}
A^{0}(k)& A^{1}(k)& \cdots & A^{\tau-1}(k)& A^{\tau}(k)\\
I&0&\cdots&0& 0\\
0& I & \cdots &0&0\\
\vdots& \vdots & \ddots & \vdots&\vdots\\
0& 0& \cdots & I &0
\end{array}
\right].
\end{eqnarray*}
Let $B_{k'}'=\sum_{k=(k'-1)h+1}^{k'h}B_{k}$, and $\mathcal{F}_{k'}'=\mathcal{F}_{k'h}$. Then $\{B_{k}',\mathcal{F}_{k}'\}$ also forms an adapted sequence, and
$\E\{B_{m+1}'~|~\mathcal{F}_{m}'\}$ has a $\delta$-spanning tree.
From Lemma \ref{lemExpectationToProbability}, there exists $\delta'>0$ such that
\begin{eqnarray*}
\Prb\Big\{\sum_{k=m+1}^{m+(n-1)\Sp(n)}B_{k}'\textnormal{ has a $\delta'$-spanning tree}|~\mathcal{F}_{m}'\Big\}>\delta'.
\end{eqnarray*}
Since
\begin{eqnarray*}
C_{k}&=&\left[\begin{array}{ccccc}
\diag(A^{0}(k))& 0& \cdots & 0& 0\\
I&0&\cdots&0& 0\\
0& I & \cdots &0&0\\
\vdots& \vdots & \ddots & \vdots&\vdots\\
0& 0& \cdots & I &0
\end{array}
\right]
+\left[\begin{array}{ccccc}
\overline{\diag}(A^{0}(k))& A^{1}(k)& \cdots & A^{\tau-1}(k)& A^{\tau}(k)\\
0&0&\cdots&0& 0\\
0& 0 & \cdots &0&0\\
\vdots& \vdots & \ddots & \vdots&\vdots\\
0& 0& \cdots & 0 &0
\end{array}
\right]\\
&\ge&\delta M_{0}+D_{k}\ge \delta(M_{0}+D_{k}),
\end{eqnarray*}
where $D_{k}$ refers to the second matrix on the righthand side of the first line.
From Lemma \ref{lemProdHighDimensionSpanningTree}, Corollary \ref{corollaryProdHighDimensionSpanningTree}, and Remark \ref{remProdHighDimensionSpanningTree}, Remark \ref{remProductSIA}, there exist $\delta''(p)\in (0,\delta')$, $p\ge 1$ such that
\begin{eqnarray*}
\Prb\bigg\{\prod_{k=m+1}^{m+p(n-1)\Sp(n)}C_{k}\textnormal{ is $\delta''(p)$-SIA},p=1,2,\cdots~|~\mathcal{F}_{m}'\bigg\}>\delta'.
\end{eqnarray*}
Let $C'_{k'}=\prod_{k=(k'-1)(n-1)\Sp(n)+1}^{k'(n-1)\Sp(n)}C_{k}$, $\mathcal{F}_{k'}''=\mathcal{F}_{k'(n-1)\Sp(n)}'$, then $\{C_{k}',\mathcal{F}_{k}''\}$ forms an adapted sequence and
\begin{eqnarray*}
\Prb\Big\{\prod_{k=1}^{p}C_{m+k}'\textnormal{ is $\delta''(p)$-SIA },p=1,2,\cdots~|~\mathcal{F}_{m}''\Big\}>\delta'.
\end{eqnarray*}
Let $k_{n}={\bf n}(n)+1$, then from Lemma \ref{lemSIAtoScrambling} and Lemma \ref{lemProbabilityToLongerSeqence},
\begin{eqnarray*}
&&\Prb\Big\{\prod_{i=1}^{k_{n}}C_{m+i}' \textnormal{ is $\delta''(1)^{k_{n}}$-scrambling} |~\mathcal{F}_{m}''\Big\}\\
&\ge&\Prb\Big\{\prod_{k=1}^{p}C_{m'+k}' \textnormal{ is $\delta''(p)$-SIA}, ~m'=m,\cdots,m+k_{n}-1,p=1,2,\cdots.|~\mathcal{F}_{m}''\Big\} \\
&>&\delta'^{k_{n}}.
\end{eqnarray*}
Let $C_{m}''=\prod_{k=(m-1)k_{n}+1}^{mk_{n}}C_{k}'$, and $\mathcal{F}_{m}'''=\mathcal{F}_{mk_{n}}''$.
Then $\{C_{k}'',\mathcal{F}_{k}'''\}$ forms an adapted sequence, and
\begin{eqnarray*}
\Prb\{C_{m+1}'' \textnormal{ is $\delta''(1)^{k_{n}}$-scrambling}~|~\mathcal{F}_{m}'''\}>\delta'^{k_{n}}.
\end{eqnarray*}
From the Second Borel-Cantelli Lemma (Lemma \ref{lemSecondBorelCantelli}), this implies
\begin{eqnarray*}
\Prb\{C_{k}'' \textnormal{ is $\delta''(1)^{k_{n}}$-scrambling infinitely often}\}=1.
\end{eqnarray*}
From Lemma \ref{lemContractionOfProduct} and Remark \ref{remErgodicityOfProducts}, we have
\begin{eqnarray*}
\Prb\Big\{\lim_{k\to+\infty}\Delta\Big(\prod_{m=1}^{k}C_{m}\Big)=0\Big\}=1.
\end{eqnarray*}
This implies the system (\ref{eqnHighDimension}) reaches consensus almost surely. On the other hand it is not difficult to show that consensus of (\ref{eqnHighDimension}) can imply consensus of (\ref{eqnDiscrete}). And the proof is completed.
From Theorem \ref{thmAppendixMain}, we can have the following two corollaries. The first one is for deterministic case.
\begin{corollary}\label{corollaryAppendixDeterministic}
Let $\{A^{0}(k),\cdots,A^{\tau}(k)\}$ be a deterministic sequence with $A^{0}(k)\ge \delta I$ for some $\delta>0$, and $[A^{i}(k)]_{jj}=0$ for $i=1,2,\cdots,\tau$, $j=1,\cdots,n$. If there exists $h>0$ such that $\sum_{k=m+1}^{m+h}B(k)$ has a $\delta$-spanning tree, then the system \eqref{eqnDiscrete} will reach consensus.
\end{corollary}
The second one is for a random sequence that will become independent after a fixed length of time.
\begin{corollary}\label{corollaryAppendixIndependent}
Let $\{A^{0}(k),\cdots,A^{\tau}(k)\}$ be a random sequence with $A^{0}(k)\ge \delta I$ for some $\delta>0$, and $[A^{i}(k)]_{jj}=0$ for $i=1,2,\cdots,\tau$, $j=1,\cdots,n$. And there exists $N>0$ such that $A^{0}(k),\cdots,A^{\tau}(k)$ is independent of $A^{0}(k'),\cdots,A^{\tau}(k')$ whenever $k'\ge k+N$. If there exists $h>0$ such that $\sum_{k=m+1}^{m+h}\E B(k)$ has a $\delta$-spanning tree, then the system \eqref{eqnDiscrete} will reach consensus almost surely.
\end{corollary}
\begin{proof}
Let $\mathcal{F}_{k}=\sigma\{A^{0}(m),\cdots,A^{\tau}(m),~m\le k\}$ be the $\sigma$-algebra formed by $A^{0}(m),\cdots,A^{\tau}(m)$, $m\le k$. Then $\{A^{0}(k),\cdots,A^{\tau}(k),\mathcal{F}_{k}\}$ forms an adapted process. From the assumption on $\{A^{0}(k),\cdots,A^{\tau}(k)\}$, $A^{0}(k'),\cdots,A^{\tau}(k')$ is independent of $\mathcal{F}_{k}$ whenever $k'\ge k+N$. Since $B(k)\ge 0$, we have
$$
\E\{\sum_{m=k+1}^{k+N+h}B(m)|\mathcal{F}_{k}\}\ge \E\{\sum_{m=k+N+1}^{k+N+h}B(m)|\mathcal{F}_{k}\}=\E\{\sum_{m=k+N+1}^{k+N+h}B(m)\}
=\sum_{m=k+N+1}^{k+N+h}\E B(m).
$$
Thus, $\E\{\sum_{m=k+1}^{k+N+h}B(m)|\mathcal{F}_{k}\}$ has a $\delta$-spanning tree and the conclusion follows from Theorem \ref{thmAppendixMain}.
\end{proof}
|
\section{Introduction}
\label{Introduction}
The tangent lines to a space curve form a ruled surface, which is called the {\it tangent surface}
or the
{\it tangent developable} or the {\it tangential variety} to the space curve
(see for instance \cite{Kreyszig}\cite{Cayley}\cite{Ishikawa4}\cite{Lawrence}).
Tangent surfaces appear in various geometric problems and applications naturally,
providing several important examples of non-isolated singularities in applications of geometry
(see for instance \cite{AG}\cite{CI}\cite{FP}\cite{IMT0}\cite{IKY}\cite{INS}\cite{Nuno Ballesteros}\cite{NS}).
It is known, in the three dimensional Euclidean space ${E}^3$,
that the tangent surface to a generic space curve $\gamma : I \to {E}^3$
is locally diffeomorphic to the {\it cuspidal edge}
or to the {\it folded umbrella} (also called, {\it cuspidal cross cap}), as is found by Cayley and Cleave \cite{Cleave}.
Cuspidal edge singularities appear along ordinary points where
$\gamma', \gamma'', \gamma'''$ are linearly independent, while
the folded umbrella appears at an isolated point of zero torsion where
$\gamma', \gamma'', \gamma'''$ are linearly dependent but $\gamma', \gamma'', \gamma''''$
are linearly independent (see \cite{BG}\cite{Porteous}).
The classification is generalized to more degenerate cases
by Mond \cite{Mond1}\cite{Mond3} and Shcherbak \cite{Shcherbak1}\cite{Shcherbak2}\cite{Arnol'd3}.
See also \cite{Ishikawa3}\cite{Ishikawa4}.
The classifications were performed mainly in locally projectively flat cases so far.
However more general cases, namely, not necessarily projectively flat cases have never been treated as far as the authors recognize.
\
In this paper we give the complete solution to
the local diffeomorphism classification problem of generic singularities which appear in
tangent surfaces, in as wider situations as possible.
We interpret {\it geodesics} as \lq\lq lines" whenever a (semi-)Riemannian metric,
or, more generally, an affine connection $\nabla$ is given in an ambient space of arbitrary dimension. Then, given
an immersed curve, or, more generally a {\it directed} curve or a {\it frontal} curve which has well-defined tangent directions along the curve, we define {\it $\nabla$-tangent surface}
by the ruled surface by tangent geodesics to the curve.
Then the main theorems in this paper are as follows:
\begin{Th}
\label{genericity-theorem} {\rm (Genericity 1: Singularities of tangent surfaces to generic curves)}
Let $\nabla$ be any affine connection on a manifold $M$ of dimension $m \geq 3$.
The singularities of the $\nabla$-tangent surface
to a generic curve in $M$ on a neighborhood of the curve
are only the cuspidal edges and the folded umbrellas if $m = 3$,
and the embedded cuspidal edges if $m \geq 4$.
\stepcounter{Cor
\begin{Th}
\label{genericity-theorem2} {\rm (Genericity 2: Singularities of tangent surfaces to generic directed curves) }
Let $\nabla$ be any affine connection on a manifold $M$ of dimension $m \geq 3$.
The singularities of the $\nabla$-tangent surface to a generic directed
curve in $M$ on a neighborhood of the curve
are only the cuspidal edges, the folded umbrellas and the swallowtails if $m = 3$,
and the embedded cuspidal edges and open swallowtails if $m \geq 4$.
\stepcounter{Cor
The genericity is exactly given in each case of genericity 1 and 2 (see Propositions \ref{genericity1}, \ref{genericity2}) using
Whitney $C^\infty$ topology on appropriate space of curves.
\begin{Th}
\label{characterization-theorem}
{\rm (Characterization)}
Let $\nabla$ be a torsion free affine connection on a manifold $M$.
Let $\gamma : I \to M$ be a $C^\infty$ curve from an open interval $I$.
{\rm (1)} Let $\dim(M) = 3$.
If $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0)$ are linearly independent,
then the $\nabla$-tangent surface
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is locally diffeomorphic to the cuspidal edge at $(t_0, 0) \in I \times \mathbf{R}$.
If $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0)$ are linearly dependent,
and $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^4\gamma)(t_0)$ are linearly independent, then
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is
locally diffeomorphic to the folded umbrella at $(t_0, 0) \in I \times \mathbf{R}$.
{\rm (2)} Let $\dim(M) = 3$.
If $(\nabla\gamma)(t_0) = 0$ and $(\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0), (\nabla^4\gamma)(t_0)$
are linearly independent, then
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is
locally diffeomorphic to the swallowtail at $(t_0, 0) \in I \times \mathbf{R}$.
{\rm (3)} Let $\dim(M) \geq 4$.
If $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0)$ are linearly independent,
then the $\nabla$-tangent surface $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is locally diffeomorphic to the embedded
cuspidal edge at $(t_0, 0) \in I \times \mathbf{R}$.
{\rm (4)} Let $\dim(M) \geq 4$.
If $(\nabla\gamma)(t_0) = 0$ and
$(\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0), (\nabla^4\gamma)(t_0), (\nabla^5\gamma)(t_0)$
are linearly independent, then
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is
locally diffeomorphic to the open swallowtail at $(t_0, 0) \in I \times \mathbf{R}$.
\stepcounter{Cor
\
A map-germ $f : (\mathbf{R}^2, p) \to M$ is locally {\it diffeomorphic} at $p$ to another map-germ
$g : (\mathbf{R}^2, p') \to M'$
if there exist diffeomorphism-germs
$\sigma : (\mathbf{R}^2, p) \to (\mathbf{R}^2, p')$ and $\tau : (M, f(p)) \to (M', g(p'))$ such that
$\tau\circ f = g\circ \sigma : (\mathbf{R}^2, p) \to (M', g(p'))$.
The {\it cuspidal edge} is defined by the map-germ $(\mathbf{R}^2, 0) \to (\mathbf{R}^m, 0)$, $m \geq 3$,
$$
(t, s) \mapsto (t + s, \ t^2 + 2st, \ t^3 + 3st^2, \ 0, \ \dots, \ 0),
$$
which is diffeomorphic to $(u, w) \mapsto (u, w^2, w^3, 0, \dots, 0)$.
The cuspidal edge singularities are originally defined only in the three dimensional space.
Here we are generalizing the notion of the cuspidal edge in higher dimensional space. In Theorem \ref{characterization-theorem} (2), we emphasize it by writing \lq\lq embedded" cuspidal edge. In what follows, we call it just
cuspidal edge for simplicity even in the case $m \geq 4$.
The {\it folded umbrella} (or the {\it cuspidal cross cap}) is defined by the map-germ
$(\mathbf{R}^2, 0) \to (\mathbf{R}^3, 0)$,
$$
(t, s) \mapsto (t + s, \ t^2 + 2st, \ t^4 + 4st^3),
$$
which is diffeomorphic to $(u, t) \mapsto (u, t^2 + ut, t^4 + {\textstyle \frac{2}{3}}ut)$.
The {\it swallowtail} is defined by the map-germ $(\mathbf{R}^2, 0) \to (\mathbf{R}^3, 0)$
$$
(t, s) \mapsto (t^2 + s, \ t^3 + {\textstyle \frac{3}{2}}st, \ t^4 + 2st^2),
$$
which is diffeomorphic to $(u, t) \mapsto (u, t^3 + ut, t^4 + \frac{2}{3}ut^2)$.
The {\it open swallowtail} is defined by the map-germ $(\mathbf{R}^2, 0) \to (\mathbf{R}^m, 0)$, $m \geq 4$,
$$
(t, s) \mapsto (t^2 + s, \ t^3 + {\textstyle \frac{3}{2}}st, \ t^4 + 2st^2, \ t^5 + {\textstyle \frac{5}{2}}st^3, \ 0, \ \dots, \ 0),
$$
which is diffeomorphic to $(u, t) \mapsto (u, t^3 + ut, t^4 + \frac{2}{3}ut^2, t^5 + \frac{5}{9}ut^3, 0, \dots, 0)$.
The open swallowtail singularity was introduced by Arnol'd (see \cite{Arnold}) as a singularity of Lagrangian varieties in symplectic geometry. Here we abstract its diffeomorphism class as the singularity of tangent surfaces (see \cite{Givental}\cite{Ishikawa4}).
\
In \cite{INS}, Izumiya, Nagai, Saji introduced and studied the class
\lq\lq E-flat" great circular surfaces in the standard three sphere $S^3$
in detail, which contains the class of tangent surfaces to curves in $S^3$.
The generic classifications given there (Theorems 1.2, 1.3 of \cite{INS}) in the sphere geometry
become different from ours,
because of the differences of topology and mappings spaces defining the genericity.
Because we treat singularities in a general ambient space, we need the intrinsic
characterizations of singularities found in \cite{KRSUY}\cite{FSUY}.
Note that the characterization of swallowtails was applied to hyperbolic geometry in \cite{KRSUY}
and to Euclidean and affine geometries in \cite{IM}.
The characterization of folded umbrellas is applied to Lorenz-Minkowski geometry in \cite{FSUY}.
In this paper we apply to non-flat projective geometry the characterizations and their some generalization
via the notion of openings introduced by the first author (\cite{Ishikawa4}, see also \cite{Ishikawa2}).
\
In \S \ref{Affine connection and geodesic}, we recall on affine connections and related notions.
In \S \ref{Tangent surface and frontal}, we define the tangent surface to an immersed curve and
show that the tangent surface is a frontal under certain general conditions (Lemma \ref{non-degenerate-frontal}).
In \S \ref{Tangent surface to directed curve} we introduce the notion of directed curves and
define their tangent surfaces.
We recall the criteria of singularities in \S \ref{Cuspidal edge and folded umbrella} and
\S \ref{Swallowtail and open swallowtail}. After a preliminary calculations in \S \ref{Characteristic vector field},
we show Theorem \ref{characterization-theorem}
in \S \ref{Euclidean case}, using the criteria of singularities for the Euclidean case and
in \S \ref{Proof of the characterization theorem} in general.
In \S \ref{Proof of the genericity theorem 1} we show Theorem \ref{genericity-theorem},
and in \S \ref{Proof of the genericity theorem 2}
we prove Theorem \ref{genericity-theorem2}.
Apart from main theorems, but related to them, we
give an observation on the singularities of tangent surfaces to torsionless curves in
\S \ref{Tangent surfaces to torsionless curves and fold singularities}. In flat case the
tangent surface to a torsionless curve necessarily has the fold singularity. However, in non-flat case
we have an example where $(2, 5)$-cuspidal edge singularity appears
on the tangent surface of some torsionless curve.
\
In this paper all manifolds and mappings are assumed to be of class $C^\infty$ unless otherwise stated.
\section{Affine connection and geodesic}
\label{Affine connection and geodesic}
Let $M$ be an $m$-dimensional manifold with an affine connection
$\nabla$ (see \cite{Helgason}\cite{KN}). For any vector fields $X, Y$ on $M$, a
vector field $\nabla_XY$ on $M$, which is called the {\it covariant derivative} of $Y$ by $X$, is assigned such that
$$
\nabla_{hX + kY}Z = h\nabla_XZ + k\nabla_YZ, \quad
\nabla_X(hY + kZ) = h\nabla_XY + (Xh)Y + k\nabla_XZ + (Xk)Z,
$$
for any vector fields $X, Y, Z$ and functions $h, k$ on $M$.
For a system of local coordinates $x^\lambda\, (\lambda = 1, 2, \dots, m)$, we write
$$
\nabla_{\frac{\partial}{\partial x^\mu}}{\textstyle\frac{\partial}{\partial x^\nu}} \ = \ \Gamma^\lambda_{\mu \nu}\ {\textstyle\frac{\partial}{\partial x^\lambda}},
$$
using the Christoffel symbols (coefficients of the connection)
$\Gamma^\lambda_{\mu \nu}$ and the Einstein convention.
In general, for a given mapping $g : N \to (M, \nabla)$, we define the notion of
{\it covariant derivative} $\nabla_\eta^g v : N \to TM$
of a vector field $v : N \to TM$ along $g$
by a vector field $\eta : N \to TN$ over a manifold $N$ (see \cite{FSUY}).
Using a local presentation
$\eta(t) = \eta^i(t)\frac{\partial}{\partial t^i}(t)$,
$v(t) = v^\lambda(t)\frac{\partial}{\partial x^\lambda}(g(t))$,
for local coordinates $t = (t_1, \dots, t_n)$ of $N$ and $x = (x_1, \dots, x_m)$ of $M$,
and using the Einstein convention, we define
$$
(\nabla_\eta^g v)(t) := \left\{ \eta^i(t) \frac{\partial v^\lambda }{\partial t^i}(t)
+ \Gamma^\lambda_{\mu\nu}(g(t))\eta^i(t) \frac{\partial g^\mu}{\partial t^i}(t)v^\nu(t)\right\}\frac{\partial}{\partial x^\lambda}(g(t)).
$$
Here $g^\mu = x^\mu\circ g$. The definition is naturally derived from the parallelisms on $M$ induced by the connection
$\nabla$.
Then we have, as in the usual case,
$$
\nabla_{h\eta + k\xi}^f v = h\nabla_\eta^f v + k \nabla_\xi^f v,
\quad
\nabla_\eta^f(hv + kw) = h\nabla_\eta^f v + k\nabla_\eta^f w
+ (\eta h)v + (\eta k)w,
$$
for any vector fields $v, w$ along $f$, $\eta, \xi$ over $N$, and functions $h, k$ on $N$.
If $g : M \to M$ is the identity mapping, then $\nabla_\eta^g v$ is just
the ordinary covariant derivative $\nabla_\eta v$ for vector fields $\eta, v$ over $M$.
The covariant derivative along a mapping is well-defined also for any tensor field over the mapping,
that is compatible with any contractions.
Then we get the notion of geodesics:
A curve $\varphi : I \to M, \varphi = \varphi(s)$ is called a $\nabla$-{\it geodesic}
if $\nabla_{\partial/\partial s}^\varphi \frac{d \varphi}{ds} = 0$.
For $x \in M, v \in T_xM$, let $\varphi(x, v, s)$ denote the $\nabla$-geodesic
determined by the differential equation
$$
\dfrac{\partial^2\varphi}{\partial s^2}^\lambda(x, v, s) \ + \ \Gamma^\lambda_{\mu \nu}(\varphi(x, v, s))
\dfrac{\partial \varphi}{\partial s}^\mu(x, v, s)\dfrac{\partial \varphi}{\partial s}^\nu(x, v, s) = 0,
$$
with the initial conditions $\varphi(x, v, 0) = x$ and $\frac{\partial \varphi}{\partial s}(x, v, 0) = v$.
Here $\varphi^\lambda$ denotes the $\lambda$-th component $x^\lambda\circ \varphi$ of $\varphi$.
Note that $\varphi : U (\subset TM\times\mathbf{R}) \to M$
is defined on an open neighborhood $U$ of $TM\times\{ 0\}$.
\begin{Lem}
\label{geodesic-expression}
At each point of M, there exists an open neighborhood $U$ of the point such that
the $\nabla$-geodesics $\varphi(x, v, s)$ are written
$$
\varphi(x, v, s) = x + s\, v + \frac{1}{2}s^2\, h(x, v, s)
$$
for some $C^\infty$ mapping $h(x, v, s)$ on an open neighborhood of $TU\times \{ 0\}$ in
$TU\times\mathbf{R}$.
\stepcounter{Th
\noindent{\it Proof\,}:\
Since $\varphi(x, v, 0) = x$, we have locally
$\varphi(x, v, s) = x + s\tilde{\varphi}(x, v, s)$ for some $C^\infty$ mapping $\tilde{\varphi}$.
Then $\frac{\partial \varphi}{\partial s}(x, v, s) = \tilde{\varphi}(x, v, s) + s\frac{\partial \tilde{\varphi}}{\partial s}(x, v, s)$.
Since $\tilde{\varphi}(x, v, 0) = \frac{\partial \varphi}{\partial s}(x, v, 0) = v$, we have locally
$\tilde{\varphi}(x, v, s) = v + \frac{1}{2}s\, h(x, v, s)$ for some $C^\infty$ mapping $h$.
Thus we have the result.
\hfill $\Box$
\begin{Lem}
\label{reminder}
Let $\varphi(x, v, s) = x + s\, v + \frac{1}{2}s^2\, h(x, v, s)$ be the expression of the $\nabla$-geodesics
as in Lemma \ref{geodesic-expression}. Then we have
\begin{align*}
h^\lambda(x, v, 0) & = \frac{\partial^2 \varphi^\lambda}{\partial s^2}(x, v, 0) = - \Gamma^\lambda_{\mu\nu}(x)v^\mu v^\nu,
\\
\frac{\partial h^\lambda}{\partial x^\kappa}(x, v, 0) & = - \Gamma^\lambda_{\mu\nu,\kappa}(x)v^\mu v^\nu,
\\
\frac{\partial h^\lambda}{\partial v^\rho}(x, v, 0) & = - \Gamma^\lambda_{\rho\nu}(x)v^\nu - \Gamma^\lambda_{\mu\rho}(x)v^\mu,
\\
\frac{\partial h^\lambda}{\partial s}(x, v, 0) & = \frac{1}{3}\frac{\partial^3\varphi}{\partial s^3}(x, v, 0)
= \frac{1}{3}(-\Gamma_{\mu\nu,\kappa}^\lambda + \Gamma_{\rho\kappa}^\lambda\Gamma_{\mu\nu}^\rho
+ \Gamma_{\kappa\rho}^\lambda\Gamma_{\mu\nu}^\rho)v^\mu v^\nu v^\kappa
\end{align*}
where $\Gamma^\lambda_{\mu\nu,\kappa} = \partial \Gamma^\lambda_{\mu\nu}/ \partial x^\kappa$.
\stepcounter{Th
\noindent{\it Proof\,}:\
For a fixed $(x, v)$, $\varphi(x, v, s)$ is a $\nabla$-geodesic, therefore we have
$$
\dfrac{\partial^2 \varphi^\lambda}{\partial s^2}(x, v, s)
= - \Gamma^\lambda_{\mu\nu}(\varphi(x, v, s))\frac{\partial \varphi}{\partial s}^\mu(x, v, s)\frac{\partial \varphi}{\partial s}^\nu(x, v, s).
$$
Since $\frac{\partial \varphi}{\partial s}(x, v, 0) = v$, we have that $\frac{\partial^2 \varphi}{\partial s^2}(x, v, 0) = h(x, v, 0)$,
and that
$h^\lambda(x, v, 0) = - \Gamma^\lambda_{\mu\nu}(x)v^\mu v^\nu$.
Moreover we have
\begin{align*}
\dfrac{\partial^3 \varphi^\lambda}{\partial s^3}(x, v, s)
= & - \Gamma^\lambda_{\mu\nu,\kappa}(\varphi(x, v, s))\frac{\partial \varphi}{\partial s}^\kappa(x, v, s)\frac{\partial \varphi}{\partial s}^\mu(x, v, s)\frac{\partial \varphi}{\partial s}^\nu(x, v, s)
\\
&
- \Gamma^\lambda_{\mu\nu}\frac{\partial^2 \varphi}{\partial s^2}^\mu(x, v, s)\frac{\partial \varphi}{\partial s}^\nu(x, v, s)
- \Gamma^\lambda_{\mu\nu}\frac{\partial \varphi}{\partial s}^\mu(x, v, s)\frac{\partial^2 \varphi}{\partial s^2}^\nu(x, v, s).
\end{align*}
By setting $s = 0$ and by the first equality, we have the fourth equality.
By differentiating both sides of
$h^\lambda(x, v, 0) = - \Gamma^\lambda_{\mu\nu}(x)v^\mu v^\nu$ (the first equality),
by $x^\kappa$ and $v^\rho$ we have the second and the third equalities.
\hfill $\Box$
\
A connection $\nabla$ on $M$ is called {\it torsion free} if
$T(X, Y) := \nabla_XY - \nabla_YX - [X, Y] = 0, $ for any vector fields $X, Y$ on $M$. A connection
$\nabla$ is torsion free if and only if, for any system of local coordinates,
$\Gamma^\lambda_{\mu\nu} = \Gamma^\lambda_{\nu\mu}, 1 \leq \lambda, \mu, \nu \leq m$.
\begin{Rem}
\label{torsion-free}
{\rm
We observe that the equation on geodesics
$$
\dfrac{\partial^2\varphi}{\partial s^2}^\lambda(x, v, s) \ + \ \Gamma^\lambda_{\mu \nu}(\varphi(x, v, s))
\dfrac{\partial \varphi}{\partial s}^\mu(x, v, s)\dfrac{\partial \varphi}{\partial s}^\nu(x, v, s) = 0,
$$
is symmetric on the indices $\mu, \nu$. Therefore
The geodesics $\varphi(x, v, s)$ and the tangent surfaces ${\mbox {\rm Tan}}(\gamma)$
remain same if the connection $\Gamma^\lambda_{\mu \nu}$ is replaced by the torsion free
connection $\frac{1}{2}(\Gamma^\lambda_{\mu \nu} + \Gamma^\lambda_{\nu \mu})$,
in other word, if $\nabla$ is replaced by the torsion free connection $\widetilde{\nabla}$,
defined by $\widetilde{\nabla}_XY = \nabla_XY - \frac{1}{2}T(X, Y)$.
}
\enr
\
Let $\gamma : I \to M$ be a curve which is not necessarily a geodesic nor
an immersed curve. Then the first derivative $(\nabla\gamma)(t)$ means just the velocity
vector field $\gamma'(t)$. The second derivative $(\nabla^2\gamma)(t)$ is defined, in terms of
covariant derivative along the curve $\gamma$, by
$$
(\nabla^2\gamma)(t) := \nabla_{\partial/\partial t}^\gamma (\nabla\gamma)(t).
$$
Note that $\gamma$ is a $\nabla$-geodesic if and only if $\nabla^2 \gamma = 0$.
In general, we define $k$-th covariant derivative of $\gamma$ inductively by
$$
(\nabla^k\gamma)(t) := \nabla_{\partial/\partial t}^\gamma (\nabla^{k-1}\gamma)(t), \ (k \geq 2).
$$
Then we have by direct calculations:
\begin{Lem}
\label{iteration}
\begin{align*}
(\nabla\gamma)^\lambda & = (\gamma')^\lambda,
\\
(\nabla^2\gamma)^\lambda & = (\gamma'')^\lambda + \Gamma^\lambda_{\mu\nu}(\gamma')^\mu(\gamma')^\nu,
\\
(\nabla^3\gamma)^\lambda &
= (\gamma''')^\lambda +
(\Gamma^\lambda_{\mu\nu, \kappa} + \Gamma^\lambda_{\kappa\rho}\Gamma^\rho_{\mu\nu})
(\gamma')^\mu(\gamma')^\nu(\gamma')^\kappa
+ (2\Gamma^\lambda_{\mu\nu} + \Gamma^\lambda_{\nu\mu})(\gamma')^\mu(\gamma'')^\nu.
\end{align*}
\stepcounter{Th
\section{Tangent surface and frontal}
\label{Tangent surface and frontal}
Let $\gamma : I \to M$ be a $C^\infty$ immersion from an open interval $I$.
Then the tangent surface to $\gamma$ is the ruled surface by tangent $\nabla$-geodesics to $\gamma$.
More precisely the $\nabla$-tangent surface $f = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma) : V (\subset I \times \mathbf{R}) \to M$ to the curve
$\gamma$ is defined by
$$
f (t, s) := \varphi(\gamma(t), \gamma'(t), s),
$$
on an open neighborhood $V$ of $I \times \{0\}$.
The mapping $f$ has singularity at least along $\{ s = 0\}$. In fact
$f_*(\frac{\partial}{\partial t})(t, 0) = \gamma'(t) = f_*(\frac{\partial}{\partial s})(t, 0)$ and the kernel of the differential $f_*$
is generated by the vector field $\eta = \frac{\partial}{\partial t} - \frac{\partial}{\partial s}$ along $\{ s = 0 \}$ on the $t$-$s$-plane.
A map-germ $f : (\mathbf{R}^n, p) \to M$, $n \leq m = \dim(M)$ is called a {\it frontal} if
there exists a $C^\infty$ {\it integral} lifting $\widetilde{f} : (\mathbf{R}^n, p) \to {\mbox {\rm Gr}}(n, TM)$ of $f$.
Here ${\mbox {\rm Gr}}(n, TM)$ means the Grassmannian bundle over $M$ consisting of
tangential $n$-planes in $TM$,
$$
{\mbox {\rm Gr}}(n, TM) := \{ \Pi \mid \Pi \subseteq T_xM, \Pi {\mbox{\rm \ is a linear subspace,\ }}\dim(\Pi) = n, \, x \in M\},
$$
with the canonical projection $\pi : {\mbox {\rm Gr}}(n, TM) \to M$ to the base manifold $M$,
and we call $\tilde{f}$ is {\it integral} if $f_*(T_q\mathbf{R}^n) \subseteq \widetilde{f}(q)$ for any $q$
in a neighborhood of $p$ in
$\mathbf{R}^n$, after taking a representative of $f$.
The definition generalizes the preceding definition of \lq\lq frontal" in the case $m = n+1$ (see \cite{FSUY}).
The definitions in the case $m = n+1$ are equivalent to each other as is easily seen.
Note that, in \cite{Ishikawa4},
we have introduced the same notion of frontal mapping under the restriction that the locus of immersive points
of $f$ is dense, where the integral lifting $\widetilde{f}$ is uniquely determined.
Let $f : (\mathbf{R}^n, p) \to M$ is a frontal and $\widetilde{f}$ is an integral lifting of $f$.
Then there exists a frame $V_1, V_2, \dots, V_n : (\mathbf{R}^n, p) \to TM$
along $f$ associated with $\widetilde{f}$ such that
$$
\widetilde{f}(q) = \langle V_1(q), V_2(q), \dots, V_n(q)\rangle_{\mathbf{R}},
$$
for any $q$ in a neighborhood of $p$ in
$\mathbf{R}^n$.
Then there is a $C^\infty$ function-germ $\sigma : (\mathbf{R}^n, p) \to \mathbf{R}$ such that
$$
(\dfrac{\partial f}{\partial t_1} \wedge \dfrac{\partial f}{\partial t_2} \wedge \cdots \wedge \dfrac{\partial f}{\partial t_n})(t)
= \sigma(t)(V_1 \wedge V_2 \wedge \cdots \wedge V_n)(t),
$$
as germs of $n$-vector fields $(\mathbf{R}^n, p) \to \wedge^n TM$ over $f$.
Then the singular locus (non-immersive locus) $S(f)$ of $f$ coincides with the zero locus $\{ \sigma = 0\}$
of $\sigma$.
We call $\sigma$ a {\it signed area density function} or briefly an {\it $s$-function} of the frontal $f$.
Note that the function $\sigma$ is essentially the same thing with the function $\lambda$ introduced in
\cite{KRSUY}\cite{FSUY} in the case $\dim(M) = 3$. However we avoid the notation $\lambda$ here because
we use it for index of Christoffel symbols.
We say that a frontal $f : (\mathbf{R}^n, p) \to M$ has a {\it non-degenerate} singular point at $p$ if
the $s$-function $\sigma$ of $f$ satisfies $\sigma(p) = 0$ and $d\sigma(p) \not= 0$.
The condition is independent of the choice of the integral lifting $\widetilde{f}$
and the associated frame $V_1, V_2, \dots, V_n$.
If $f$ has a non-degenerate singular point at $p$, then $f$ is of corank $1$ such that the singular locus
$S(f) \subset (\mathbf{R}^n, p)$ is a regular hypersurface.
In this paper we concern with only the cases $n = 1$ and $n = 2$.
\
Returning to our situation, we have
\begin{Lem}
\label{non-degenerate-frontal}
Suppose $(\nabla\gamma)(t_0)$ and $(\nabla^2\gamma)(t_0)$ are linearly independent.
Then the germ of tangent surface $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$
is a frontal with the non-degenerate singular point at $(t_0, 0)$ and
with the singular locus $S(\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)) = \{ s = 0\}$.
\stepcounter{Th
\noindent{\it Proof\,}:\
We set $f(t, s) = {\mbox {\rm Tan}}(\gamma)(t, s) = \varphi(\gamma(t), \gamma'(t), s)$.
By Lemma \ref{geodesic-expression}, write $\varphi = x + s\, v + \frac{1}{2}s^2\, h$. Then we have
$$
f(t, s) = \gamma(t) + s\, \gamma'(t) + \frac{1}{2} s^2 \, h(\gamma(t), \gamma'(t), s),
$$
and
\begin{align*}
\frac{\partial f}{\partial t} & = \gamma' + s\gamma'' +
\frac{1}{2}s^2\, (\gamma')^\mu\frac{\partial h}{\partial x^\mu}(\gamma, \gamma', s) +
\frac{1}{2}s^2\, (\gamma'')^\nu\frac{\partial h}{\partial v^\nu}(\gamma, \gamma', s),
\\
\frac{\partial f}{\partial s} & = \gamma' + s\, h(\gamma, \gamma', s) + \frac{1}{2}s^2\, \frac{\partial h}{\partial s}(\gamma, \gamma', s).
\end{align*}
Then we see that $S(f) \supseteq \{ s = 0 \}$.
Let $s \not= 0$. Then
\begin{equation*}
\begin{split}
\frac{1}{s}(\frac{\partial f}{\partial t} - \frac{\partial f}{\partial s}) =
& \
\gamma'' + \frac{1}{2}s\, (\gamma')^\mu\frac{\partial h}{\partial x^\mu}(\gamma, \gamma', s)
+ \frac{1}{2}s\, (\gamma'')^\nu\frac{\partial h}{\partial v^\nu}(\gamma, \gamma', s)
\\
&
\hspace{0.5truecm}
- h(\gamma, \gamma', s) - \frac{1}{2}s\, \frac{\partial h}{\partial s}(\gamma, \gamma', s).
\end{split}
\end{equation*}
We define $F(t, s)$ by the right hand side. Then $F(t, s) = \frac{1}{s}(\frac{\partial f}{\partial t} - \frac{\partial f}{\partial s})$ if
$s \not= 0$. Moreover $F$ is $C^\infty$ also on $s = 0$ and
$$
F(t, 0) = \gamma''(t) - h(\gamma(t), \gamma'(t), 0).
$$
By Lemma \ref{reminder}, we have
$$
h^\lambda(\gamma(t), \gamma'(t), 0) = - \Gamma^\lambda_{\mu\nu}(\gamma(t))\, (\gamma'(t))^\mu(\gamma'(t))^\nu.
$$
Hence we have
$$
F^\lambda(t, 0) = (\gamma''(t))^\lambda + \Gamma^\lambda_{\mu\nu}(\gamma(t))\, (\gamma'(t))^\mu(\gamma'(t))^\nu
= (\nabla^2\gamma)^\lambda(t).
$$
Therefore if
$(\nabla \gamma)(t), (\nabla^2\gamma)(t)$ are linearly independent at $t = t_0$, then
$\frac{\partial f}{\partial t}(t, s)$ and $F(t, s)$ are linearly independent around $(t_0, 0)$ and satisfies that
$$
(\frac{\partial f}{\partial t} \wedge \frac{\partial f}{\partial s})(t, s) = -s (\frac{\partial f}{\partial t} \wedge F)(t, s).
$$
Therefore we see that $\frac{\partial f}{\partial t}(t, s)$ and $F(t, s)$ define the integral lifting of $f$,
$f$ is frontal with non-degenerate singular point at $(t_0, 0)$, and that
$S(f) = \{ s = 0\}$.
\hfill $\Box$
\begin{Rem}
{\rm
In the Euclidean case,
Lemma \ref{non-degenerate-frontal} holds globally on $I\times \mathbf{R}$. However, even in a locally projectively flat case,
Lemma \ref{non-degenerate-frontal} holds just locally near $I \times \{ 0\}$.
For example, let $M$ be the standard three dimensional sphere $S^3 \subset \mathbf{R}^4$
with the standard (Levi-Civita) connection. Then geodesics in $S^3$ are given by great circles (with periodic parametrizations) and
we observe, via the natural double covering $S^3 \to \mathbf{R} P^3$,
that the tangent surface to any curve in $S^3$
has singularities not only along the original curve, but also along the antipodal of the curve (cf. \cite{INS}).
}
\enr
\section{Tangent surface to directed curve}
\label{Tangent surface to directed curve}
Let $PTM = {\mbox {\rm Gr}}(1, TM)$ denote the projective tangent bundle over the manifold $M$,
and $\pi : PTM \to M$ the natural projection. The fiber of $\pi$ over $x \in M$ is the
projective space $P(T_xM)$ of dimension $m-1$.
A curve $\gamma : I \to M$, which is not necessarily an immersion, is called {\it directed}
if there assigned a $C^\infty$ lifting $\widetilde{\gamma} : I \to PTM$ of $\gamma$ for $\pi$ which
satisfies $\gamma'(t) \in \widetilde{\gamma}(t) \subset T_{\gamma(t)}M$ for any $t \in I$.
Here $\widetilde{\gamma}(t) \in P(T_{\gamma(t)}M)$ is regarded as a one-dimensional
linear subspace of $T_{\gamma(t)}M$. The notion of directed curves is nothing but the
notion of frontal maps introduced in \S \ref{Tangent surface and frontal} in the case $n = 1$
with assignment of an integral lifting.
Then we regard the direction $\widetilde{\gamma}(t_0)$ is assigned to
each point $\gamma(t_0)$ on $\gamma$. Note that
if $\gamma'(t_0) \not= 0$, then $\widetilde{\gamma}(t_0)$ is uniquely determined by
the tangent line $\langle \gamma'(t_0)\rangle_{\mathbf{R}} \subset T_{\gamma(t_0)}M$.
Let $\gamma : I \to M$ be a directed curve and $\widetilde{\gamma}$ its integral lifting.
Then there exists a frame $u : I \to TM$ of $\widetilde{\gamma}$ which satisfies
$\widetilde{\gamma}(t) = \langle u(t)\rangle_{\mathbf{R}}, u(t) \not= 0$ for any $t \in I$.
Note that there exists a unique function $a(t)$ such that
$\gamma'(t) = a(t)u(t)$.
Then define the $\nabla$-tangent surface $f = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma) : V (\subset I \times \mathbf{R}) \to M$ by
$$
f(t, s) := \varphi(\gamma(t), u(t), s),
$$
using the family of $\nabla$-geodesics $\varphi = \varphi(x, v, s)$ and a frame $u(t)$.
The $\nabla$-tangent surface for an immersed curve $\gamma$ in \S\ref{Tangent surface and frontal}
was defined by the frame $u(t) = \gamma'(t)$.
\begin{Lem}
If the immersion locus of a directed curve $\gamma : I \to M$ is dense in $I$,
then the integral lifting $\widetilde{\gamma}$ is uniquely determined.
The right equivalence class of the germ of $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma) : (I \times \mathbf{R}, I \times\{ 0\}) \to M$
for a directed curve $\gamma$
is independent of the choice of the frame $u$.
\stepcounter{Th
\noindent{\it Proof\,}:\
The first half is clear because $\widetilde{\gamma}$ is $C^\infty$, so is continuous.
The second half is achieved by the diffeomorphism $(t, s) \to (t, c(t)s)$ for
another choice $c(t)u(t), c(t) \not= 0$.
\hfill $\Box$
\begin{Lem}
\label{Lemma-}
Let $k \geq 2$.
Suppose $(\nabla\gamma)^i(t_0) = 0, 1 \leq i < k$ and $(\nabla\gamma)^k(t_0) \not= 0$.
Then we have:
{\rm (1)}
For any coordinates of $M$ around $\gamma(t_0)$,
$\gamma^{(i)}(t_0) = 0, 1 \leq i < k$ and $\gamma^{(k)}(t_0) = (\nabla^k\gamma)(t_0) \not= 0$.
Moreover we have $\gamma^{(k+1)}(t_0) = (\nabla^{k+1}\gamma)(t_0)$.
{\rm (2)}
Set
$$
u(t) = \frac{1}{k(t - t_0)^{k-1}}\gamma'(t).
$$
Then $u$ is a $C^\infty$ vector field along $\gamma$ on a neighborhood of $t_0$.
The curve $\gamma$ is directed on a neighborhood of $t_0$ by the frame $u$.
{\rm (3)}
For any frame $u(t)$ of the directed curve $\gamma$ around $t_0$, and for any $\ell \geq 0$,
$$
(\nabla^k\gamma)(t_0), \ (\nabla^{k+1}\gamma)(t_0), \ \dots, \ (\nabla^{k+\ell}\gamma)(t_0)
$$
are linearly independent if and only if
$$
u(t_0), \ (\nabla^\gamma_{\partial/\partial t}u)(t_0), \ \dots, \ ((\nabla^\gamma_{\partial/\partial t})^\ell u)(t_0)
$$
are linearly independent.
In particular, for the frame in {\rm (2)}, we have
$$
u(t_0) = \frac{1}{k!}(\nabla^k\gamma)(t_0), \
(\nabla u)(t_0)
= \frac{1}{k\cdot k!}(\nabla^{k+1}\gamma)(t_0), \ \dots, \
(\nabla^{\ell} u)(t_0)
= \frac{\ell !}{k\cdot (k+\ell - 1)!}(\nabla^{k+\ell}\gamma)(t_0).
$$
where $\nabla^i u = (\nabla^\gamma_{\partial/\partial t})^i u$.
\stepcounter{Th
\noindent{\it Proof\,}:\
(1)
Let $k = 2$. Then $\gamma'(t_0) = (\nabla \gamma)(t_0) = 0$. By Lemma \ref{iteration},
we have $\gamma''(t_0) = (\nabla^2\gamma)(t_0) \not= 0, \gamma'''(t_0) = (\nabla^3\gamma)(t_0)$.
Let $k \geq 3$. Then
$(\nabla^k\gamma)^\lambda$ is a sum of $(\gamma^{(k)})^\lambda$
and a polynomial of $\Gamma^\lambda_{\mu\nu}$, their partial derivatives
and $\gamma^{(i)}, i < k$, each monomial of which
contains a $\gamma^{(i)}$ with $i \leq k-2$ (cf. Lemma \ref{iteration}).
Thus we have $\gamma^{(i)}(t_0) = (\nabla\gamma)^i(t_0) = 0, 1 \leq i < k$. Moreover we have
$0 \not= (\nabla^k\gamma)(t_0) = \gamma^{(k)}(t_0)$ and $(\nabla^{k+1}\gamma)(t_0) = \gamma^{(k+1)}(t_0)$.
\\
(2) is clear.
\\
(3)
We have that $c(t)u(t) = \gamma'(t)$
for some function $c(t)$. If $k \geq 2$, then $c(t_0) = 0$.
By operating $\nabla^\gamma_{\partial/\partial t}$ to both sides of $c(t)u(t) = \gamma'(t)$,
we have
$$
c'(t)u(t) + c(t)(\nabla^\gamma_{\partial/\partial t})u(t) = (\nabla^2\gamma)(t).
$$
If $k \geq 3$, then $c(t_0) = 0, c'(t_0) = 0$.
In general we have
$$
c(t_0) = c'(t_0) = \dots = c^{(k-2)}(t_0) = 0, c^{(k-1)}(t_0) \not= 0,
$$
and
$$
\begin{array}{ccc}
c^{(k-1)}(t)u(t) + (k-1)c^{(k-2)}(t)(\nabla u)(t) + { }_{k-1}C_2 c^{(k-3)}(t)(\nabla^2 u)(t) + \cdots & = & (\nabla^{k}\gamma)(t)
\\
c^{(k)}(t)u(t) + kc^{(k-1)}(t)(\nabla u)(t) + { }_{k}C_2 c^{(k-2)}(t)(\nabla^2 u)(t) + \cdots & = & (\nabla^{k+1}\gamma)(t)
\\
\vdots & & \vdots
\\
c^{(k+\ell)}(t)u(t) + \cdots + { }_{k+\ell - 1}C_{k-1} c^{(k-1)}(t)(\nabla^\ell u)(t) + \cdots & = &
(\nabla^{k+\ell}\gamma)(t).
\end{array}
$$
Evaluating at $t_0$, we have the result.
\hfill $\Box$
\
Lemma \ref{non-degenerate-frontal} is generalized as follows:
\begin{Lem}
\label{non-degenerate-frontal2}
Let $\gamma : I \to M$ be a $C^\infty$ curve, $t_0 \in I$, and $k \geq 1$.
Suppose that $(\nabla^i\gamma)(t_0) = 0, 1 \leq i < k$ and that
$(\nabla^k\gamma)(t_0), (\nabla^{k+1}\gamma)(t_0)$ are linearly independent.
Then the germ of $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is a frontal with non-degenerate singular point at $(t_0, 0)$
and with the singular locus $S(\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)) = \{ s = 0\}$.
\stepcounter{Th
\noindent{\it Proof\,}:\
Suppose $k \geq 2$. Let $u(t)$ be a frame around $t_0$ of the directed curve $\gamma$
and $c(t)u(t) = \gamma'(t)$, $u(t_0) \not= 0$.
(For instance $c(t) = k(t - t_0)^{k-1}$).
Since $f(t, s) = \gamma(t) + s u(t) + \frac{1}{2}s^2 h(\gamma(t), u(t), s)$, we have
\begin{align*}
\frac{\partial f}{\partial t} & = \gamma' + su' +
\frac{1}{2}s^2\, (\gamma')^\mu\frac{\partial h}{\partial x^\mu}(\gamma, u, s) +
\frac{1}{2}s^2\, (u')^\nu\frac{\partial h}{\partial v^\nu}(\gamma, u, s),
\\
\frac{\partial f}{\partial s} & = u + s\, h(\gamma, u, s) + \frac{1}{2}s^2\, \frac{\partial h}{\partial s}(\gamma, u, s).
\end{align*}
Then we see that $S(f) \supseteq \{ s = 0 \}$ and the kernel field of $f_*$ along $\{ s = 0\}$ is given by
$\eta = \frac{\partial}{\partial t} - c(t)\frac{\partial}{\partial s}$.
Let $s \not= 0$. Then
\begin{equation*}
\begin{split}
\frac{1}{s}(\frac{\partial f}{\partial t} - c(t)\frac{\partial f}{\partial s}) =
& \
u' + \frac{1}{2}s\, (\gamma')^\mu\frac{\partial h}{\partial x^\mu}(\gamma, u, s)
+ \frac{1}{2}s\, (u')^\nu\frac{\partial h}{\partial v^\nu}(\gamma, u, s)
\\
&
\hspace{0.5truecm}
- c(t) h(\gamma, u, s) - \frac{1}{2}s c(t) \frac{\partial h}{\partial s}(\gamma, u, s).
\end{split}
\end{equation*}
We define $F(t, s)$ by the right hand side. Then
$F(t, s) = \frac{1}{s}(\frac{\partial f}{\partial t} - c(t)\frac{\partial f}{\partial s})$ if
$s \not= 0$. Moreover $F$ is $C^\infty$ also on $s = 0$ and
$$
F(t, 0) = u'(t) - c(t) h(\gamma(t), u(t), 0).
$$
By Lemma \ref{reminder},
$$
F(t, 0) = u'(t) + c(t)\Gamma^\lambda_{\mu\nu}(\gamma(t))\, (u(t))^\mu(u(t))^\nu
= (\nabla^\gamma_{\partial/\partial t}u)(t).
$$
By Lemma \ref{Lemma-} (3),
if
$(\nabla^k \gamma)(t_0), (\nabla^{k+1} \gamma)(t_0)$ are linearly independent, then
$\frac{\partial f}{\partial s}(t, s)$ and $F(t, s)$ are linearly independent around $(t_0, 0)$.
Moreover they satisfies
$$
(\frac{\partial f}{\partial t} \wedge \frac{\partial f}{\partial s})(t, s) = -s (\frac{\partial f}{\partial s} \wedge F)(t, s).
$$
Therefore we see that $\frac{\partial f}{\partial s}(t, s)$ and $F(t, s)$ define an integral lifting of $f$, $f$ is frontal with
non-degenerate singular point at $(t_0, 0)$, and that
$S(f) = \{ s = 0\}$.
\hfill $\Box$
\section{Cuspidal edge and folded umbrella}
\label{Cuspidal edge and folded umbrella}
Let $f : (\mathbf{R}^2, p) \to M^3$ be a frontal with a non-degenerate singular point at $p$
and $\widetilde{f} : (\mathbf{R}^2, p) \to {\mbox {\rm Gr}}(2, TM)$ the integral lifting of $f$.
Let $V_1, V_2 : (\mathbf{R}^2, p) \to TM$ be an associated frame with $\widetilde{f}$.
Let $L : (\mathbf{R}^2, p) \to T^*M \setminus \zeta$ be an annihilator of
$\widetilde{f}$. The condition is that $\langle L, V_1\rangle = 0, \langle L, V_2\rangle = 0$.
Here $\zeta$ means the zero section.
Let $c : (\mathbf{R}, t_0) \to (\mathbf{R}^2, p)$ be a parametrization of the singular locus $S(f)$, $p = c(t_0)$, and
$\eta : (\mathbf{R}^2, p) \to T\mathbf{R}^2$ be a vector field which restricts to the kernel field of $f_*$ on $S(f)$.
Suppose that $V_2(p) \not\in f_*(T_p\mathbf{R}^2)$.
Then, for any affine connection $\nabla$ on $M$, we define
$$
\psi(t) := \langle L(c(t)), (\nabla_\eta^f V_2)(c(t))\rangle.
$$
Note that the vector field $(\nabla_\eta^f V_2)(c(t))$ is independent of the extension $\eta$ and the choice of affine connection $\nabla$, since $\eta\vert_{S(f)}$ is a kernel field of $f_*$.
We call the function $\psi(t)$
the {\it characteristic function} of $f$.
Then the following characterizations of cuspidal edges and folded umbrellas are given in \cite{KRSUY}\cite{FSUY}:
\begin{Th}
\label{characterization}
{\rm (Theorem 1.4 of \cite{FSUY})}.
Let $f : (\mathbf{R}^2, p) \to M^3$ be a germ of frontal with a non-degenerate singular point at $p$.
Let $c : (\mathbf{R}, t_0) \to (\mathbf{R}^2, p)$ be a parametrization of the singular locus of $f$.
Suppose $f_*c'(t_0) \not= 0$. Then, for the characteristic function $\psi$,
\\
{\rm (1)}
$f$ is diffeomorphic to the cuspidal edge if and only if $\psi(t_0) \not= 0$.
\\
{\rm (2)} $f$ is diffeomorphic to the folded umbrella if and only if
$\psi(t_0) = 0, \psi'(t_0) \not= 0$.
\stepcounter{Cor
Note that the conditions appeared in Theorem \ref{characterization} are invariant under diffeomorphism equivalence.
\begin{Rem}
{\rm
In the situation of Theorem \ref{characterization}, we set $\gamma(t) = f(c(t))$. Then we have
$\psi(t_0) \not= 0$ if and only if $V_1(c(t_0)), V_2(c(t_0)), (\nabla_\eta^f V_2)(c(t_0))$ are linearly independent.
}
\enr
\
The above construction is generalized to the case $m = \dim(M) \geq 4$.
In general, let $f : (\mathbf{R}^2, p) \to M^m, m \geq 4,$ be a frontal with a non-degenerate singular point at $p$
and $\widetilde{f} : (\mathbf{R}^2, p) \to {\mbox {\rm Gr}}(2, TM)$ the integral lifting of $f$.
Let $V_1, V_2 : (\mathbf{R}^2, p) \to TM$ be an associated frame with $\widetilde{f}$.
We take a coframe $L_1, \dots, L_{m-2} : (\mathbf{R}^2, p) \to T^*M$ satisfying that
$$
\langle L_i(t, s), V_1(t, s)\rangle = 0, \langle L_i(t, s), V_2(t, s)\rangle = 0, \quad (1 \leq i \leq m-2),
$$
and that $L_1(t, s), \dots, L_{m-2}(t, s)$ are linearly independent for $(t, s) \in (\mathbf{R}^2, p)$.
We define the {\it characteristic (vector valued) function} $\psi : (\mathbf{R}, t_0) \to \mathbf{R}^{m-2}$ by $\psi = (\psi_1, \dots, \psi_{m-2})$,
$$
\psi_i(t) := \langle L_i(c(t)), (\nabla_\eta^f V_2)(c(t))\rangle.
$$
Let $g : (\mathbf{R}^n, p) \to (\mathbf{R}^\ell, q)$ be a map-germ.
A map germ $f : (\mathbf{R}^n, p) \to \mathbf{R}^{\ell+r}$ is called an {\it opening} of $g$ if there exist
functions $h_1, \dots, h_r : (\mathbf{R}^n, p) \to \mathbf{R}$ and functions
$a_{ij} : (\mathbf{R}^n, p) \to \mathbf{R}, (1 \leq i \leq r, 1 \leq j \leq \ell)$ such that
$$
dh_i = \sum_{j=1}^\ell a_{ij}dg_j, \quad f = (g, h_1, \dots, h_r),
$$
(see for example \cite{Ishikawa4}). If $\ell = n$, then the condition on $h$ is equivalent to that
$f$ is frontal associated with an integral lifting $\widetilde{f} : (\mathbf{R}^n, p)
\to {\mbox {\rm Gr}}(n, T\mathbf{R}^{n+r})$ having Grassmannian coordinates $(a_{ij})$
such that $\widetilde{f}(p)$ projects isomorphically to
$T_{g(p)}\mathbf{R}^n$ by the projection $\mathbf{R}^{n+r} = \mathbf{R}^n\times \mathbf{R}^r \to \mathbf{R}^n$.
\begin{Rem}
{\rm
A map-germ $f : (\mathbf{R}^n, p) \to M$ is a frontal with a non-degenerate singular point at $p$
if and only if $f$ is diffeomorphic to an opening of a map-germ $g : (\mathbf{R}^n, p) \to (\mathbf{R}^n, q)$
of Thom-Boardman singularity type $\Sigma^1$ at $p$, i.e.
$g$ is of corank one and $j^1g : (\mathbf{R}^n, p) \to J^1(\mathbf{R}^n, \mathbf{R}^n)$ is transversal to
the variety of singular $1$-jets (see for example \cite{GG}).
}
\enr
We can summarize several known results as those on openings of the fold:
\begin{Th}
\label{characterization2}
Let $f : (\mathbf{R}^2, p) \to M^m, m \geq 2$ be a germ of frontal with a non-degenerate singular point at $p$,
$\widetilde{f} : (\mathbf{R}^2, p) \to {\mbox {\rm Gr}}(2, TM)$ the integral lifting of $f$ and
$V_1, V_2 : (\mathbf{R}^2, p) \to TM$ an associated frame with $\widetilde{f}$.
Let $c : (\mathbf{R}, t_0) \to (\mathbf{R}^2, p)$ be a parametrization of the singular locus of $f$.
Suppose $f_*c'(t_0) \not= 0$.
Then $f$ is diffeomorphic to an opening of the fold, namely to the germ $(u, w) \mapsto (u, \frac{1}{2}w^2)$.
Moreover we have:
{\rm (0)} Let $m = 2$. Then $f$ is diffeomorphic to the fold.
{\rm (1)} Let $m \geq 3$. Then $f$ is diffeomorphic to the cuspidal edge if and only if
$\psi(t_0) \not= 0$.
{\rm (2)} Let $m = 3$. Then $f$ is diffeomorphic to the folded umbrella if and only if
$\psi(t_0) = 0, \psi'(t_0) \not= 0$.
\stepcounter{Cor
\
\noindent
{\it Proof of Theorem \ref{characterization2}:}
The assertion (0) follows from Whitney's theorem (see \cite{Whitney}\cite{SUY}\cite{Saji}).
(1) The condition $\psi(t_0) \not= 0$ is equivalent to that $f$ is a front, namely, that $\widetilde{f}$ is an immersion.
Suppose $\dim(M) = 3$. Then by Proposition 1.3 of \cite{KRSUY},
we see that $f$ is diffeomorphic to the cuspidal edge.
In general cases $m \geq 2$, we see that
there exists a submersion $\pi : (M, f(p)) \to (\mathbf{R}^2, 0)$ such that
$$
T_{f(p)}(\pi^{-1}(0)) + \widetilde{f}(p) = T_{f(p)}M,
$$
and that $\pi\circ f$ satisfies the same condition with $f$,
i.e., $\pi\circ f$ is a frontal with the non-degenerate singular point at $p$ with the same singular locus with
$f$ and $\eta(c(t_0))$ and $c'(t_0)$ are linearly independent, but $m = 2$. Thus by the assertion (0),
the map-germ $\pi\circ f$ is diffeomorphic to a fold.
Moreover we see $f$ is an opening of $\pi\circ f$ because
$f$ is frontal.
The condition that $\widetilde{f}$ is an immersion is equivalent, in this case,
to that $f$ is a versal opening of $\pi\circ f$ (\S 6 of \cite{Ishikawa4}).
In fact, up to diffeomorphism equivalence, let
$$
f(u, w) = (u, \frac{1}{2}w^2, h_1(u, w), \dots, h_r(u, w)),
$$
$(m = 2+r)$ and
$$
dh_i(u, w) = a_i(u, w)du + b_i(u, w)d(\frac{1}{2}w^2) = a_i(u, w)du + wb_i(u, w)dw,
$$
for some functions $a_i, b_i, 1 \leq i \leq r$.
Then we have
$$
\frac{\partial f}{\partial u} = (1, 0, a_1, \dots, a_r), \quad \frac{\partial f}{\partial w} = (0, w, w b_1, \dots, w b_r)
$$
and the pair $V_1 = \frac{\partial f}{\partial u}, V_2 = \frac{1}{w}\frac{\partial f}{\partial w} = (0, 1, b_1, \dots, b_r)$
gives a frame of the frontal $f$. Moreover $\eta = \frac{\partial}{\partial w}$ gives the kernel field of $f_*$
along $\{w = 0\}$.
For any connection $\nabla$, we have the characteristic vector field
$\nabla^f_\eta V_2(u, 0) = (0, 0, \frac{\partial b_1}{\partial w}, \dots, \frac{\partial b_r}{\partial w})$.
Then $\psi(0) \not= 0$ if and only if $V_1(0, 0), V_2(0, 0), (\nabla^f_\eta V_2)(0, 0)$
are linearly independent. The condition is equivalent to that $h_1(0, w), \dots, h_r(0, w)$
generate $m_1^3/m_1^4$ over $\mathbf{R}$ and it is equivalent to that $f$ is a versal opening of
the Whitney's cusp. Here $m_1$ means the ideal consisting of function-germs $h(w)$ with $h(0) = 0$.
Then we see that $f$ is diffeomorphic to cuspidal edge (see Proposition 6.8 (3) $\ell = 2$ of \cite{Ishikawa4}).
The assertion (2) follows from Theorem 1.4 of \cite{FSUY}.
\hfill $\Box$
\
\section{Swallowtail and open swallowtail}
\label{Swallowtail and open swallowtail}
Based on results in \cite{KRSUY} and \cite{Ishikawa4},
we summarize the characterization results on openings of the Whitney's cusp map-germ:
\begin{Th}
\label{characterization3}
Let $f : (\mathbf{R}^2, p) \to M^m, m \geq 2$ be a germ of frontal with a non-degenerate singular point at $p$,
$V_1, V_2 : (\mathbf{R}^2, p) \to TM$ an associated frame with $\widetilde{f}$ with $V_2(p) \not\in f_*(T_p\mathbf{R}^2)$,
and $\eta : (\mathbf{R}^2, p) \to T\mathbf{R}^2$ an extension of a kernel field along of $f_*$.
Let $c : (\mathbf{R}, t_0) \to (\mathbf{R}^2, p)$ be a parametrization of the singular locus of $f$.
Set $\gamma = f\circ c : (\mathbf{R}, t_0) \to M$.
Suppose $(\nabla\gamma)(t_0) = 0$ and $(\nabla^2\gamma)(t_0) \not= 0$.
Then $f$
is diffeomorphic to an opening of Whitney's cusp,
namely to the germ $(u, t) \mapsto (u, t^3 + ut)$. Moreover we have
{\rm (0)} Let $m = 2$. Then $f$ is diffeomorphic to Whitney's cusp.
{\rm (1)} Let $m = 3$. Then $f$ is diffeomorphic to the swallowtail if and only if
$$
V_1(c(t_0)), \ V_2(c(t_0)), \ (\nabla^f_\eta V_2)(c(t_0))
$$
are linearly independent in $T_{f(p)}M$.
{\rm (2)} Let $m \geq 4$. Then $f$ is diffeomorphic to the open swallowtail if and only if
$$
(V_1\circ c)(t_0), \ (V_2\circ c)(t_0), \ ((\nabla^f_\eta V_2)\circ c)(t_0), \
(\nabla^\gamma_{\partial/\partial t}((\nabla^f_\eta V_2)\circ c))(t_0)
$$
are linearly independent in $T_{f(p)}M$.
\stepcounter{Cor
\noindent{\it Proof\,}:\
The assertion (0) follows from Whitney's theorem (also see \cite{Whitney}\cite{SUY}\cite{Saji}).
(1) follows from Proposition 1.3 of \cite{KRSUY}.
In general cases $m \geq 2$, we see that
there exists a submersion $\pi : (M, f(p)) \to (\mathbf{R}^2, 0)$ such that $\pi^{-1}(0)$ is
transverse to $\widetilde{f}(0) \subset T_{f(p)}M$, $\pi\circ f$ satisfies the same condition with $f$,
namely, that $\pi\circ f$ is a frontal with the non-degenerate singular point at $p$ and with the same singular locus with
$f$ and $\eta(c(t_0))$ and $c'(t_0)$ are linearly independent, but $m = 2$. Thus by the assertion (0),
the map-germ $\pi\circ f$ is diffeomorphic to the Whitney's cusp.
Moreover we see $f$ is an opening of Whitney's cusp because $f$ is frontal.
Let $f(u, t) = (u, t^3 + ut, h_1(u, t), \dots, h_r(u, t)), m = 2 + r$ and
$dh_i = a_idu + b_id(t^3 + ut) = (a_i + tb_i)du + (3t^2 + u)b_i dt, $
for some functions $a_i = a_i(u, t), b_i = b_i(u, t), 1 \leq i \leq r$.
Then we have
$$
\frac{\partial f}{\partial u} = (1, t, a_1 + t b_1, \dots, a_r + t b_r),
\quad \frac{\partial f}{\partial t} = (0, 3t^2 + u, (3t^2 + u)b_1, \dots, (3t^2 + u)b_r),
$$
a frame $V_1 = \frac{\partial f}{\partial u}, V_2 = \frac{1}{3t^2 + u}\frac{\partial f}{\partial t} = (0, 1, b_1, \dots, b_r)$
of the frontal $f$, and a kernel field $\eta = \frac{\partial}{\partial t}$ of $f_*$.
We have
\begin{align*}
V_1(0, 0) = (1, 0, a_1(0, 0), \dots, a_r(0, 0)), \quad V_2(0, 0) = (0, 1, b_1(0, 0), \dots, b_r(0, 0)),
\\
(\nabla^f_\eta V_2)(0, 0) = (0, 0, \frac{\partial b_1}{\partial t}(0, 0), \dots, \frac{\partial b_r}{\partial t}(0, 0)).
\end{align*}
Let $c(t) = (- 3t^2, t)$. Then $\gamma(t) = f(c(t)) = (-3t^2, -2t^3, h_1(-3t^2, t), \dots, h_r(-3t^2, t))$ and
$$
\nabla^f_\eta V_2(c(t)) = (0, 0, \frac{\partial b_1}{\partial t}(c(t)), \dots, \frac{\partial b_r}{\partial t}(c(t))).
$$
Then we have
$$
\nabla^\gamma_{\partial/\partial t}((\nabla^f_\eta V_2)\circ c)\vert_{t=0} =
(0, 0, \frac{\partial^2 b_1}{\partial t^2}(0, 0), \dots, \frac{\partial^2 b_r}{\partial t^2}(0, 0)).
$$
Thus the condition of (2) is equivalent, in our case,
to that $f$ is a versal opening of $\pi\circ f$ and
then we see $f$ is diffeomorphic to the open swallowtail
(see Proposition 6.8 (3) $\ell = 3$ of \cite{Ishikawa4}).
Thus we have the characterization (2).
\hfill $\Box$
\section{Characteristic vector field}
\label{Characteristic vector field}
Let $\gamma : I \to M$ be an immersion. We set $f = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ and
suppose $\nabla\gamma, \nabla^2\gamma$ are linearly independent at $t = t_0$.
Note that $\eta = \frac{\partial}{\partial t} - \frac{\partial}{\partial s}$ generates the field of kernels ${\mbox {\rm Ker}}(f_*)$ of the differential
$f_*$ along $s = 0$.
Let $F(t, s) = \frac{1}{s}(\frac{\partial f}{\partial t} - \frac{\partial f}{\partial s})$ as in the proof of Lemma \ref{non-degenerate-frontal}.
Then we have
$$
(\nabla_\eta^f F)(t, s) = \{(\frac{\partial}{\partial t} - \frac{\partial}{\partial s})F^\lambda + (\Gamma^\lambda_{\mu\nu}(f(t, s))
((\frac{\partial}{\partial t} - \frac{\partial}{\partial s}) f^\mu) F^\nu \}\frac{\partial}{\partial x^\lambda}(f(t, s)).
$$
Therefore we have
$$
(\nabla_\eta^f F)(t, 0) = (\frac{\partial F}{\partial t} - \frac{\partial F}{\partial s})(t, 0).
$$
We call the vector field $(\nabla_\eta^f F)(t, 0)$ along $\gamma$ the {\it characteristic vector field} of
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$.
By straightforward calculations we have
\begin{Lem}
\label{characteristic vector}
The characteristic vector field of $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is given by
\begin{align*}
(\nabla_\eta^f F)^\lambda(t, 0)
& =
(\gamma''')^\lambda + (\Gamma^\lambda_{\mu\nu,\kappa} + \frac{1}{2}\Gamma^\lambda_{\rho\mu}\Gamma^\rho_{\nu\kappa}
+ \frac{1}{2}\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\nu\kappa})(\gamma')^\mu(\gamma')^\nu(\gamma')^\kappa
+ \frac{3}{2}(\Gamma^\lambda_{\mu\nu} + \Gamma^\lambda_{\nu\mu})(\gamma')^\mu(\gamma'')^\nu.
\end{align*}
\stepcounter{Th
\begin{Lem}
\label{third derivative}
$(\nabla_\eta^f F)(t, 0) = (\nabla^3 \gamma)(t)$ if the affine connection $\nabla$ is torsion free.
\stepcounter{Th
\noindent{\it Proof\,}:\ We compare Lemma \ref{characteristic vector} and Lemma \ref{iteration}.
The equality $\frac{3}{2}(\Gamma^\lambda_{\mu\nu} + \Gamma^\lambda_{\nu\mu}) =
2\Gamma^\lambda_{\mu\nu} + \Gamma^\lambda_{\nu\mu}$ holds if and only if
$\Gamma^\lambda_{\mu\nu} = \Gamma^\lambda_{\nu\mu}$. Then
the equality
$\frac{1}{2}\Gamma^\lambda_{\rho\mu}\Gamma^\rho_{\nu\kappa}
+ \frac{1}{2}\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\nu\kappa}
= \Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\nu\kappa}$ holds.
\hfill $\Box$
\
Now let $\gamma : I \to M$ be a directed curve and $\gamma'(t_0) = 0$ for a $t_0 \in I$. Suppose
$c(t)u(t) = \gamma'(t)$ in a neighborhood of $t_0$ for some frame $u(t), u(t_0) \not= 0$ and a function $c(t)$.
(In Lemma \ref{Lemma-}, we can take $c(t) = k(t - t_0)^{k-1}$).
We have defined in \S \ref{Tangent surface to directed curve}
$$
f(t, s) = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)(t, s) := \varphi(\gamma(t), u(t), s),
$$
using $\nabla$-geodesics $\varphi(x, v, s)$.
Then $V_1 = \frac{\partial f}{\partial s}$ and
$$
V_2 = F(t, s) = \frac{1}{s}\left(\frac{\partial f}{\partial t}(t, s) - c(t)\frac{\partial f}{\partial s}(t, s)\right)
$$
form a frame of the integral lifting of $f$.
Let $\eta = \frac{\partial}{\partial t} - c(t)\frac{\partial}{\partial s}$,
the kernel field. Then we have
\begin{Lem}
\label{characteristic-vector-field}
If $\nabla$ is torsion free, then
$$
(\nabla^f_\eta F)(t, 0) = ((\nabla^\gamma_{\partial/\partial t})^2 u)(t).
$$
\stepcounter{Th
\noindent{\it Proof\,}:\
By straightforward calculations, we have
\begin{align*}
(\nabla^f_\eta F)^\lambda(t, 0)
& =
(u'')^\lambda + c'\Gamma_{\mu\nu}^\lambda u^\mu u^\nu
+ \frac{3}{2}c(\Gamma^\lambda_{\mu\nu} + \Gamma^\lambda_{\nu\mu})u^\nu (u')^\mu
\\
& \qquad \qquad
+ \frac{1}{2}c^2(2\Gamma_{\mu\nu, \kappa}^\lambda + \Gamma^\lambda_{\rho\mu}\Gamma^\rho_{\kappa\nu} +
\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\kappa\nu})u^\mu u^\nu u^\kappa
\\
& =
(u'')^\lambda + \Gamma_{\mu\nu}^\lambda (\gamma'')^\mu u^\nu +
(\frac{1}{2}\Gamma_{\nu\mu}^\lambda + \frac{3}{2}\Gamma_{\mu\nu}^\lambda)(\gamma')^\mu(u')^\nu
\\
& \qquad \qquad
+ \frac{1}{2}(2\Gamma_{\mu\nu, \kappa}^\lambda + \Gamma^\lambda_{\rho\mu}\Gamma^\rho_{\kappa\nu} +
\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\kappa\nu})(\gamma')^\mu(\gamma')^\kappa u^\nu.
\end{align*}
On the other hand, we have
$$
(\nabla^\gamma_{\partial/\partial t} u)^\lambda = (u')^\lambda + \Gamma^\lambda_{\mu\nu}(\gamma')^\mu u^\nu,
$$
and then
$$
((\nabla^\gamma_{\partial/\partial t})^2 u)^\lambda = (u'')^\lambda
+ \Gamma_{\mu\nu}^\lambda(\gamma'')^\mu u^\nu
+ 2\Gamma^\lambda_{\mu\nu}(\gamma')^\mu(u')^\nu +
(\Gamma_{\mu\nu,\kappa}^\lambda +
\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\kappa\nu})(\gamma')^\mu(\gamma')^\kappa u^\nu.
$$
Since $\nabla$ is torsion free, we have
$\frac{1}{2}\Gamma_{\nu\mu}^\lambda + \frac{3}{2}\Gamma_{\mu\nu}^\lambda = 2\Gamma^\lambda_{\mu\nu}$ and
$\frac{1}{2}(2\Gamma_{\mu\nu, \kappa}^\lambda + \Gamma^\lambda_{\rho\mu}\Gamma^\rho_{\kappa\nu} +
\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\kappa\nu}) =
\Gamma_{\mu\nu,\kappa}^\lambda +
\Gamma^\lambda_{\mu\rho}\Gamma^\rho_{\kappa\nu}$. Thus we have
$(\nabla^f_\eta F)(t, 0) = ((\nabla^\gamma_{\partial/\partial t})^2 u)(t)$.
\hfill $\Box$
\section{Euclidean case}
\label{Euclidean case}
Now we show Theorem \ref{characterization-theorem} in Euclidean case, using the characterization results
of singularities prepared in previous sections.
Let $M = {E}^m$ be the Euclidean $m$-space with the connection $\Gamma^\lambda_{\mu\nu} = 0$.
Then $\varphi(x, v, s) = x + s\, v$, that is $h(x, v, s) = 0$.
If $\gamma'(t_0) \not= 0$, then the tangent surface is given by
$f(t, s) = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)(t, s) = \gamma(t) + s\gamma'(t)$ near $t_0 \times \mathbf{R}$.
Then $\frac{\partial f}{\partial t} = \gamma'(t) + s\gamma''(t),
\frac{\partial f}{\partial s} = \gamma'(t)$ and we have $\frac{\partial f}{\partial t} - \frac{\partial f}{\partial s} = s\gamma''(t)$.
$F(t, s) = \frac{1}{s}(\frac{\partial f}{\partial t} - \frac{\partial f}{\partial s}) = \gamma''(t)$. Moreover we have
$(\nabla_{\partial/\partial t - \partial/\partial s}^f F)(t, 0) = \gamma'''(t)$.
Suppose $\gamma'(t_0)$ and $\gamma''(t_0)$ are linearly independent.
By Lemma \ref{non-degenerate-frontal}, $f$ is a frontal with non-degenerate singular point at $(t_0, 0)$.
We apply Theorem \ref{characterization} to this situation. We take $\eta = \frac{\partial}{\partial t} - \frac{\partial}{\partial s}$ and
$c(t) = (t, 0)$. Then $c'(t_0)$ and $\eta(c(t_0))$ are linearly independent, that is, $f_*(c'(t_0)) \not= 0$.
Take a coframe, namely, a system of
germs of $1$-forms $L_i = L_i(t, s) : (I\times \mathbf{R}, (t_0, 0)) \to T^*{E}^m \setminus \zeta$ along $f$, $i = 1, \dots, m-2$, which satisfy that
$$
\langle L_i(t, s), \gamma'(t)\rangle = 0, \quad \langle L_i(t, s), \gamma''(t)\rangle = 0, (i = 1, \dots, m-2),
$$
and $L_1(t, s), \dots L_{m-2}(t, s)$ are linearly independent for $(t, s) \in (I\times\mathbf{R}, (t_0, 0))$.
Here $\zeta$ means the zero section. Actually we can take $L$ to be independent of $s$ in this case.
Set $\ell_i(t) = L_i(t, 0)$ and
set $\psi_i(t) = \langle \ell_i(t), \gamma'''(t)\rangle$ and
$$
\psi(t) = (\psi_1(t), \dots, \psi_{n-2}(t)) = (\langle \ell_1(t), \gamma'''(t)\rangle, \dots, \langle \ell_{m-2}(t), \gamma'''(t)\rangle),
$$
the characteristic function. Then we have that
$\psi(t_0) = 0$ if and only if $\gamma'(t_0), \gamma''(t_0), \gamma'''(t_0)$ are linearly dependent.
We have
$$
\langle \ell_i(t), \gamma'(t)\rangle = 0, \quad \langle \ell_i(t), \gamma''(t)\rangle = 0, (i = 1, \dots, m-2),
$$
and
\begin{align*}
0 & = \langle \ell_i'(t), \gamma'(t)\rangle + \langle \ell_i(t), \gamma''(t)\rangle = \langle \ell_i'(t), \gamma'(t)\rangle,
\\
0 & = \langle \ell_i'(t), \gamma''(t)\rangle + \psi_i(t).
\end{align*}
Suppose $\psi(t_0) = 0$. Then $\langle \ell_i'(t_0), \gamma''(t_0)\rangle = 0$ for any $i, (1 \leq i \leq m-2)$.
Since we have also $\langle \ell_i'(t_0), \gamma'(t_0)\rangle = 0$, we obtain $\langle \ell_i'(t_0), \gamma'''(t_0)\rangle = 0$,
because $\gamma'(t_0)$ and $\gamma''(t_0)$ are linearly independent and
$\gamma'(t_0), \gamma''(t_0), \gamma'''(t_0)$ are linearly dependent.
Now we have
$$
\psi_i'(t) = \langle \ell_i'(t), \gamma'''(t)\rangle + \langle \ell_i(t), \gamma^{(4)}(t)\rangle.
$$
Therefore we have $\psi_i'(t_0) = \langle \ell_i(t_0), \gamma^{(4)}(t_0)\rangle$.
Thus under the condition $\psi(t_0) = 0$, we have
$\psi'(t_0) \not= 0$ if and only if $\gamma'(t_0), \gamma''(t_0), \gamma^{(4)}(t_0)$ are linearly independent.
Thus Theorem \ref{characterization} and Theorem \ref{characterization2} (1) imply
Theorem \ref{characterization-theorem} (1)(3) in the Euclidean case.
\
Suppose now $\gamma'(t_0) = 0, \gamma''(t_0) \not= 0$. Let $c(t)u(t) = \gamma'(t)$ with $u(t_0) \not= 0$.
Then $c(t_0) = 0, c'(t_0) \not= 0$.
The tangent surface is defined by $f(t, s) = \gamma(t) + s u(t)$. Then $f$ is a frontal with a frame
$V_1(t, s) = \frac{\partial f}{\partial s}(t, s) = u(t)$ and $V_2(t, s) =
F(t, s) = \frac{1}{s}(\frac{\partial f}{\partial t} - c\frac{\partial f}{\partial s}) = u'(t)$.
Moreover $\nabla^f_{\eta}F(t, 0) = u''(t)$.
Now $u(t_0), u'(t_0), u''(t_0)$ are linearly independent if and only if
$\gamma''(t_0), \gamma'''(t_0), \gamma^{(4)}(t_0)$ are linearly independent.
Then in the case $m = 3$, by Theorem \ref{characterization3} (1),
we have Theorem \ref{characterization-theorem} (2).
Let $m \geq 4$. Then we have $\nabla^\gamma_{\partial/\partial t}\nabla^f_{\eta}F(t, 0) = u'''(t)$.
Now $u(t_0), u'(t_0), u''(t_0), u'''(t_0)$ are linearly independent if and only if
$\gamma''(t_0), \gamma'''(t_0), \gamma^{(4)}(t_0), \gamma^{(5)}(t_0)$ are linearly independent.
By Theorem \ref{characterization3} (2), we have Theorem \ref{characterization-theorem} (4).
\section{Proof of the characterization theorem}
\label{Proof of the characterization theorem}
\noindent
{\it Proof of Theorem \ref{characterization-theorem} (1)(3) in the general torsion free case:}
\\
Suppose $(\nabla \gamma)(t_0)$ and $(\nabla^2\gamma)(t_0)$ are linearly independent.
Then $f = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is a frontal with the frame
$V_1(t, s) = \frac{\partial f}{\partial t}(t, s)$ and $V_2(t, s) = F(t, s)$ for the integral lifting $\widetilde{f}$.
We take coframe $L = (L_1, \dots, L_{m-2})$, $L_i : (I\times\mathbf{R}, (t_0, 0)) \to T^*M \setminus \zeta$ along $f$ satisfying
$$
\langle L_i(t, s), V_1(t, s) \rangle = 0, \quad \langle L_i(t, s), V_2(t, s)\rangle = 0, \ (1 \leq i \leq m-2).
$$
Set $\ell_i(t) = L_i(t, 0)$.
Then we have
$$
\langle \ell_i(t), (\nabla \gamma)(t) \rangle = 0, \quad \langle \ell_i(t), (\nabla^2\gamma)(t) \rangle = 0.
$$
Set
$$
\psi_i(t) := \langle \ell_i(t), (\nabla_\eta^f F)(t, 0) \rangle, (1 \leq i \leq m-2).
$$
Since $\nabla$ is torsion free, by Lemma \ref{third derivative},
we have $(\nabla_\eta^f F)(t, 0) = (\nabla^3\gamma)(t)$
and so
$$
\psi_i(t) = \langle \ell_i(t), (\nabla^3\gamma)(t)\rangle.
$$
Define the vector valued function
$$
\psi : I \to \mathbf{R}^{m-2}, \quad \psi(t) = (\psi_1(t), \dots, \psi_{m-2}(t)).
$$
Then $\psi(t_0) = 0$ if and only if $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), \nabla^3\gamma (t_0)$
are linearly dependent.
Note that the covariant derivative $(\nabla_{\frac{\partial}{\partial t}}^\gamma \psi_i)(t)$ of the function $\psi_i(t)$ is equal to the
ordinary derivative $\psi'_i(t)$. So we have
$$
\psi_i'(t) = (\nabla_{\frac{\partial}{\partial t}}^\gamma \psi_i)(t) = \langle (\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t), (\nabla^3\gamma)(t) \rangle
+ \langle \ell_i(t), (\nabla^4 \gamma) (t) \rangle.
$$
Since $\langle \ell_i(t), (\nabla\gamma)(t)\rangle = \langle \ell_i(t), (\nabla^2\gamma)(t)\rangle = 0$, we have
$$
0 = \langle (\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t), (\nabla\gamma)(t) \rangle + \langle \ell_i(t), (\nabla^2\gamma)(t) \rangle =
\langle (\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t), (\nabla\gamma)(t) \rangle.
$$
If $\psi(t_0) = 0$ namely if $(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0)$
are linearly dependent, then we have that
$
\langle \ell_i(t_0), (\nabla^3\gamma)(t_0) \rangle = 0, (1 \leq i \leq m-2),
$
since $(\nabla^3\gamma)(t_0)$ is a linear combination of $(\nabla\gamma)(t_0)$ and $(\nabla^2\gamma)(t_0)$.
Thus we have
$$
0 = \langle(\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t_0), (\nabla^2\gamma)(t_0) \rangle + \langle \ell_i(t_0),
(\nabla^3\gamma)(t_0) \rangle =
\langle(\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t_0), (\nabla^2\gamma)(t_0) \rangle.
$$
Moreover we have $\langle(\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t_0), (\nabla^3\gamma)(t_0) \rangle = 0$.
Therefore we have
$$
\psi_i'(t_0) = \langle (\nabla_{\frac{\partial}{\partial t}}^\gamma \ell_i)(t_0), (\nabla^3\gamma)(t_0)\rangle +
\langle \ell_i(t_0), (\nabla^4\gamma)(t_0) \rangle = \langle \ell_i(t_0), (\nabla^4\gamma)(t_0)\rangle.
$$
Therefore, if $\psi(t_0) = 0$ and
$(\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), (\nabla^4\gamma)(t_0)$ are linearly independent,
then $\psi'(t_0) \not= 0$.
Now, by Theorem \ref{characterization},
we see that, if $\dim(M) = m = 3$, and $\psi(t_0) = 0, \psi'(t_0) \not= 0$,
then $f$ is diffeomorphic to the folded umbrella (cuspidal cross cap)
at $(t_0, 0)$.
Moreover, if $\dim(M) \geq 3$ and $\psi(t_0) \not= 0$, then $f$ is diffeomorphic to
the cuspidal edge at $(t_0, 0)$.
\hfill $\Box$
\
\noindent
{\it Proof of Theorem \ref{characterization-theorem} (2)(4) in the general torsion free case:}
\\
Let $\gamma : I \to M$ is not an immersion at $t_0$, $\gamma'(t_0) = 0$, but $\gamma''(t_0) \not= 0$.
Let $c(t)u(t) = \gamma'(t), u(t_0) \not= 0$.
Then the $\nabla$-tangent surface is defined by $f(t, s) = \varphi(\gamma(t), u(t), s)$ using the
geodesics $\varphi(x, v, s)$ on $TM$.
Then we have the frame
$$
V_1(t, s) = \frac{\partial f}{\partial s}(t, s), \quad
V_2(t, s) = F(t,s) = \frac{1}{s}(\frac{\partial f}{\partial t} - c(t)\frac{\partial f}{\partial s}).
$$
We set $\eta = \frac{\partial}{\partial t} - c(t)\frac{\partial}{\partial s}$.
Then, by Lemma \ref{characteristic-vector-field}, we have
$(\nabla^f_\eta F)(t, 0) = ({\nabla^\gamma_{\partial/\partial t}}^2 \, u)(t)$.
Therefore we have
$\nabla^\gamma_{\partial/\partial t}((\nabla^f_\eta F)(t, 0)) = ({\nabla^\gamma_{\partial/\partial t}}^3 \, u)(t)$.
Now, by Lemma \ref{Lemma-},
$V_1(t_0, 0), V_2(t_0, 0), (\nabla^f_\eta F)(t_0, 0)$
are linearly independent if and only if
$u(t_0), (\nabla^\gamma_{\partial/\partial t} u)(t_0), ({\nabla^\gamma_{\partial/\partial t}}^2 \, u)(t_0)$
are linearly independent, and the condition is equivalent to that
$(\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0), (\nabla^4\gamma)(t_0)$ are linearly independent.
Then in the case $m = 3$, by Theorem \ref{characterization3} (1),
we have Theorem \ref{characterization-theorem} (2).
Let $m \geq 4$. Then
$V_1(t_0, 0), V_2(t_0, 0), (\nabla^f_\eta F)(t_0, 0), \nabla^\gamma_{\partial/\partial t}((\nabla^f_\eta F)(t, 0))\vert_{t = t_0}$
are linearly independent if and only if
$u(t_0), (\nabla^\gamma_{\partial/\partial t} u)(t_0), ({\nabla^\gamma_{\partial/\partial t}}^2 \, u)(t_0),
({\nabla^\gamma_{\partial/\partial t}}^3 \, u)(t_0)$ are linearly independent, and the condition is equivalent to that
$(\nabla^2\gamma)(t_0), (\nabla^3\gamma)(t_0), (\nabla^4\gamma)(t_0), (\nabla^5\gamma)(t_0)$
are linearly independent.
By Theorem \ref{characterization3} (2), we have Theorem \ref{characterization-theorem} (4).
\section{Proof of genericity theorem 1}
\label{Proof of the genericity theorem 1}
In general, let $\gamma : I \to M$ be a $C^\infty$ curve and $t_0 \in I$. Define
$$
a_1 := \inf\left\{ k \ \left\vert \ k \geq 1, (\nabla^k\gamma)(t_0) \not= 0 \right.\right\}.
$$
Note that $\gamma$ is an immersion at $t_0$ if and only if $a_1 = 1$.
If $a_1 < \infty$, then define
$$
a_2 := \inf\left\{ k \left\vert \ {\mbox {\rm rank}}\left( (\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), \dots, (\nabla^k\gamma)(t_0) \right)
= 2 \right.\right\}.
$$
We have $1 \leq a_1 < a_2$. If $a_i < \infty, 1 \leq i < \ell \leq m$, then define $a_\ell$ inductively by
$$
a_\ell := \inf\left\{ k \ \left\vert \
{\mbox {\rm rank}}\left( (\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), \dots, (\nabla^k\gamma)(t_0) \right)
= \ell \right.\right\}.
$$
If $a_m < \infty$, then we call the strictly increasing sequence $(a_1, a_2, \dots, a_m)$ of
natural numbers the {\it $\nabla$-type} of $\gamma$ at $t_0$.
To obtain our Theorem \ref{genericity-theorem}, we show
\begin{Prop}
\label{genericity1}
Let $(M, \nabla)$ be a manifold of dimension $m$ with an affine connection $\nabla$.
Then there exists an open dense set ${\mathcal U}$ in the set of $C^\infty$ curves from $I$ to $M$
in Whitney $C^\infty$ topology, such that for any $\gamma$ belonging to ${\mathcal U}$ and for any $t_0 \in I$,
$\gamma$ is of $\nabla$-type
$(1, 2, 3)$ or $(1, 2, 4)$ if $m = 3$, and
$(1, 2, 3, \dots, m-1, m)$ or $(1, 2, 3, \dots, m-1, m+1)$ if $m \geq 4$, at $t_0$.
\stepcounter{Th
\noindent
{\it Proof of Propositions \ref{genericity1}:}
First we remark that, for any local coordinates on $M$, and
for each $k = 1, 2, \dots$, the iterated covariant derivative
$(\nabla^{k}\gamma)(t)$ is expressed as
$(\nabla^{k}\gamma)(t) = \gamma^{(k)}(t) + P$, by a polynomial $P$ of $\gamma^{(i)}(t), 0 \leq i < k$ and
$(\partial^\alpha \Gamma^\lambda_{\mu\nu} / \partial x^\alpha)(\gamma(t)), \vert\alpha\vert \leq k - 2$
(cf. Lemma \ref{iteration}). Therefore, for positive integer $r$,
there exists an algebraic diffeomorphism of $\Phi : J^r(I, M)_{t_0, q} \to J^r(I, M)_{t_0, q}$
of the $r$-jet space $J^r(I, M)_{t_0, q} = \{ j^r\gamma(t_0) \mid \gamma : (I, t_0) \to (M, q)\}$
satisfying the following conditions:
if $\Phi(j^r\gamma(t_0)) = j^r\beta(t_0), \beta : (I, t_0) \to (M, q)$, then
$(\nabla^k\gamma)(t_0) = \beta^{(k)}(t_0), 1 \leq k \leq r$. In particular,
for any $1 \leq \ell \leq r$, and for any $1 \leq i_1 < i_2 < \cdots < i_\ell \leq r$ and for any $j, 0 \leq j \leq m$,
${\mbox {\rm rank}}((\nabla^{i_1}\gamma)(t_0), \dots, (\nabla^{i_\ell}\gamma)(t_0)) = j$ if and only if
${\mbox {\rm rank}}(\beta^{(i_1)}(t_0), \dots, \beta^{i_\ell}(t_0)) = j$.
Let $r \geq m+1$.
We set
\begin{align*}
S_{\nabla} := \{ j^{r}\gamma(t_0) \ \mid \ & (\nabla \gamma)(t_0), (\nabla^2 \gamma)(t_0),
\dots, (\nabla^{m}\gamma)(t_0)
{\mbox{\rm \ are linearly dependent and }}
\\
&
(\nabla \gamma)(t_0), (\nabla^2 \gamma)(t_0), \dots, (\nabla^{m-1}\gamma)(t_0), (\nabla^{m+1}\gamma)(t_0)
{\mbox{\rm \ are linearly dependent.}}\}
\end{align*}
is an algebraic set of codimension $2$.
The above diffeomorphism $\Phi$ maps $S_{\nabla}$ to $S_{\nabla_0}$. Here $\nabla_0$ is the trivial connection on $\mathbf{R}^m$.
Thus the calculation is reduced to the trivial case which is well known (see \cite{GG}\cite{Ishikawa4}).
Note that $S_{\nabla}$ is intrinsically defined by the given connection $\nabla$.
Then we have the associated closed stratified subbundle
$S_{\nabla}(I, M)$ of the $r$-jet bundle $J^r(I, M) \to I\times M$ of codimension $2$.
By the transversality theorem
$$
{\mathcal U} := \{ \gamma : I \to M \mid j^r\gamma : I \to J^r(I, M) {\mbox{\rm \ is transverse to }} S_{\nabla}(I, M)
\}
$$
is open dense in Whitney $C^\infty$ topology (see \cite{GG}).
Let $\gamma \in {\mathcal U}$ and $t_0 \in I$. Since $S_{\nabla}(I, M)$ is codimension $2$,
$j^r\gamma(t_0) \not\in S_{\nabla}(I, M)$. This means that
$$
(\nabla \gamma)(t_0), (\nabla^2 \gamma)(t_0), \dots, (\nabla^{m}\gamma)(t_0)
$$
are linearly independent, or, they are linearly dependent but
$$
(\nabla \gamma)(t_0), (\nabla^2 \gamma)(t_0), \dots, (\nabla^{m-1}\gamma)(t_0), (\nabla^{m+1}\gamma)(t_0)
$$
are linearly independent. In the first case, $\gamma$ is of $\nabla$-type $(1, 2, \dots, m-1, m)$.
In the second case, $\gamma$ is of type $\nabla$-type $(1, 2, \dots, m-1, m+1)$.
Thus we have Proposition \ref{genericity1}.
\hfill $\Box$
\begin{Rem}
{\rm
Let $\boldsymbol{a} = (a_1, a_2, \dots, a_m)$ be any strictly increasing sequence of positive integers.
For an integer $r \geq a_m$, we define, in the $r$-jet bundle $J^r(I, M)$,
$$
\Sigma_{\boldsymbol{a}}(I, M) : = \{ j^r\gamma(t_0) \in J^r(I, M) \mid
\gamma {\mbox{\rm \ is of \ }} \nabla{\mbox{\rm -type\ }} \boldsymbol{a} \}.
$$
Then $\Sigma_{\boldsymbol{a}}(I, M)$ is a stratified subbundle of $J^r(I, M)$ over $I\times M$ with
an algebraic typical fiber. It can be shown, as in the proof of Propositions \ref{genericity1},
that the codimension of
$\Sigma_{\boldsymbol{a}}(I, M)$ in $J^r(I, M)$ is independent of $\nabla$ and
is given by $\sum_{i = 1}^m (a_i - i)$ (see \cite{Ishikawa4} for the flat case).
}
\enr
\noindent
{\it Proof of Theorem \ref{genericity-theorem}:}
We may suppose $\nabla$ is torsion free (Remark \ref{torsion-free}). Then Proposition \ref{genericity1},
and Theorem \ref{characterization-theorem} imply Theorem \ref{genericity-theorem}.
\section{Perturbations of integral curves}
\label{Proof of the genericity theorem 2}
To treat directed curves (see \S \ref{Tangent surface to directed curve}),
we consider $PTM = {\mbox {\rm Gr}}(1, TM)$ with the natural projection
$\pi : PTM \to M$ and
the tautological subbundle $D \subset TPTM$
on the tangent bundle of $PTM$:
For any $(x, \ell) \in PTM$ and for any $v \in T_{(x, \ell)}PTM$,
$v \in D_{(x, \ell)}$ if and only if $\pi_*(v) \in \ell \subset T_xM$.
A curve $\widetilde{\gamma} : I \to PTM$ is called {\it integral}
if $\widetilde{\gamma}_*(\partial/\partial t) \in D_{\widetilde{\gamma}(t)}$,
for any $t \in I$. Recall that $\gamma = \pi\circ \widetilde{\gamma}$
with the lifting $\widetilde{\gamma}$ is called a {\it directed curve}.
Then we have
\begin{Prop}
\label{genericity2}
Let $M$ be a manifold of dimension $m$ with an affine connection $\nabla$.
Then there exists an open dense subset ${\mathcal O}$ in the space of $C^\infty$ integral curves
$I \to PTM$ with Whitney $C^\infty$ topology
such that for any $\widetilde{\gamma} \in {\mathcal O}$ and for any $t_0 \in I$,
$\gamma = \pi\circ \widetilde{\gamma} : I \to M$ is of $\nabla$-type
$$
(1, 2, 3), (1, 2, 4), {\mbox{\rm \ or \ }} (2, 3, 4),
$$
if $m = \dim(M) = 3$, and
$$
(1, 2, 3, 4, \dots, m-1, m), (1, 2, 3, 4, \dots, m-1, m+1), {\mbox{\rm \ or \ }} (2, 3, 4, 5, \dots, m, m+1).
$$
if $m \geq 4$, at $t_0$.
\stepcounter{Th
To treat directed curves, we introduce the following notion:
Let $u : I \to TM$ be a vector field along a curve $\gamma : I \to M$.
For $t_0 \in I$, we set
$$
b_i :=
\inf\left\{ k \left\vert \
{\mbox {\rm rank}}\left( u(t_0), (\nabla^\gamma_{\partial/\partial t} u)(t_0), \dots, ((\nabla^\gamma_{\partial/\partial t})^{k-1} u)(t_0) \right)
= i \right.\right\}.
$$
We have $1 \leq b_1 < b_2 < \cdots < b_m$, if each $b_i < \infty$.
Then we call the strictly increasing sequence $(b_1, b_2, \dots, b_m)$ of
natural numbers the {\it $\nabla$-type} of $u$ at $t_0$.
Note that the $\nabla$-type of a curve $\gamma$ which we have defined in above is the
$\nabla$-type of the velocity vector field $\gamma' : I \to TM$ along $\gamma$.
Let $\gamma : (\mathbf{R}, t_0) \to M$ be a germ of directed curve
with an integral lifting $\widetilde{\gamma} : (\mathbf{R}, t_0)
\to PTM$ generated by a frame $u : (\mathbf{R}, t_0) \to TM$, $u(t_0) \not= 0$.
Then $b_1 = 1$ for $u$, since $u(t_0) \not= 0$.
Generalizing Lemma \ref{Lemma-} (3), we have:
\begin{Lem}
\label{c, u}
If $\nabla$-type of $u$ is $(1, b_2, \dots, b_m)$ and the order of $c$ at $t_0$ is $\ell$,
that is, $c(t_0) = \cdots = c^{(\ell - 1)}(t_0) = 0, c^{(\ell)}(t_0) \not= 0$, then
$\gamma$ is of $\nabla$-type $(a_1, a_2, \dots, a_m) = (1+\ell, b_2+\ell, \dots, b_m+\ell)$.
\stepcounter{Th
\noindent{\it Proof\,}:\
By taking covariant derivative $\nabla$ $\ell$-times of the both sides of $c(t)u(t) = \gamma'(t)$, we have
$(\nabla\gamma)(t_0) = \cdots = (\nabla^\ell\gamma)(t_0) = 0, (\nabla^{\ell+1}\gamma)(t_0) \not= 0$.
Therefore $a_1 = 1+\ell$. In general, we have inductively
$$
{\mbox {\rm rank}}\left( (\nabla\gamma)(t_0), (\nabla^2\gamma)(t_0), \dots, (\nabla^k\gamma)(t_0) \right)
=
{\mbox {\rm rank}}\left( u(t_0), (\nabla u)(t_0), \dots, (\nabla^{k-\ell-1} u)(t_0) \right),
$$
for any $k \geq 1 + \ell$.
Therefore we have $a_i = b_i + \ell, 1 \leq i \leq m$.
\hfill $\Box$
\
We need also the following lemma on local perturbations of integral curves.
\begin{Lem}
\label{perturbation}
Let $a < t_1 < t_2 < b$ and $\widetilde{\gamma}, \widetilde{\alpha} : (a, b) \to PT\mathbf{R}^m$ be
integral curves. Then there exists an integral curve $\widetilde{\beta} : (a, b) \to
PT\mathbf{R}^m$ such that $\widetilde{\beta}(t) = \widetilde{\alpha}(t), a < t \leq t_1$ and
$\widetilde{\beta}(t) = \widetilde{\gamma}(t), t_2 \leq t < b$. If $\widetilde{\alpha}$
is sufficiently close to $\widetilde{\gamma}$ on $[t_1, t_2]$ in Whitney $C^\infty$ topology,
then $\widetilde{\beta}$ can be taken to be close to $\widetilde{\gamma}$
on $(a, b)$ in Whitney $C^\infty$ topology.
\stepcounter{Th
\noindent{\it Proof\,}:\
Let $x = (x^\lambda)$ be a system of coordinates of $\mathbf{R}^m$ and $(x, \xi) = (x^\lambda, \xi_\lambda)$ be
the associated system of coordinates of $T\mathbf{R}^m$.
Let $(x\circ \widetilde{\gamma})'(t) = c(t)u(t)$, $(x\circ \widetilde{\alpha})'(t) = e(t)v(t)$, for
some $c, e : (a, b) \to \mathbf{R}$ and $u, v : (a, b) \to \mathbf{R}^m \setminus \{ 0\}$.
Then we take a function $f(t)$ and $w : (a, b) \to \mathbf{R}^m$ such that
$f(t) = e(t)$ on $(a, t_1]$, $f(t) = c(t)$ on $[t_2, b)$,
$w(t) = v(t)$ on $(a, t_1]$, $w(t) = u(t)$ on $[t_2, b)$ and
$$
\dint{t_1}{t_2} f(t) w(t) dt = (x\circ\widetilde{\gamma})(t_1) - (x\circ\widetilde{\alpha})(t_1) + \dint{t_1}{t_2} c(t) u(t) dt.
$$
Then we have the required $\widetilde{\beta}$ by $(\xi\circ \widetilde{\beta})(t) = w(t)$ and
$$
(x\circ \widetilde{\beta}) (t) = (x\circ\widetilde{\alpha})(t_1) + \dint{t_1}{t} f(t) w(t) dt, \quad (a < t < b).
$$
\hfill $\Box$
\
\noindent
{\it Proof of Proposition \ref{genericity2}:}
Let $\widetilde{\gamma} : (\mathbf{R}, t_0) \to PTM$ be a germ of integral curve with
$\gamma = \pi\circ\widetilde{\gamma}$.
Let $c(t)u(t) = \gamma'(t)$ for some frame $u : (\mathbf{R}, t_0) \to TM$
along $\gamma$, $u(t_0) \not= 0$, and for some function $c : (\mathbf{R}, t_0) \to \mathbf{R}$.
Note that $\widetilde\gamma$ is determined by the frame $u$.
The frame $u$ is determined up to the multiplication of functions $b(t)$ with $b(t_0) \not= 0$.
Given the initial point $q = \gamma(t_0)$, the pair $(u, c)$ determines the directed curve $\gamma$ uniquely.
Moreover $(\nabla^k u)(t_0) = u^{(k)}(t_0) + Q$, by a polynomial $Q$
of $u^{(i)}(t_0), c^{(i)}(t_0), 0 \leq i < k$ and
$(\partial^\alpha \Gamma^\lambda_{\mu\nu} / \partial x^\alpha)(q), \vert\alpha\vert \leq k - 2$.
In particular $(\nabla^k u)(t_0)$ depends only on $k$-jet of $(c, u)$ and just on
the position $q = \gamma(t_0)$.
Let us consider the $r$-jet bundle $J^r(I, \mathbf{R}\times(TM \setminus \zeta))$ over
$I\times \mathbf{R}\times(TM \setminus \zeta)$, where $\zeta$ is the zero-section.
For the projection $I\times \mathbf{R}\times(TM \setminus \zeta) \to I \times M$, take the fiber
$J^r(I, \mathbf{R}\times(TM \setminus \zeta))_{t_0, q}$
over a $(t_0, q) \in I\times M$, and consider the set
\begin{align*}
S_{\nabla} := \ &\{ j^r(c, u)(t_0) \mid \ u(t_0), (\nabla u)(t_0),
\dots, (\nabla^{m-1} u)(t_0)
{\mbox{\rm \ are linearly dependent }}
\\
&
\quad {\mbox{\rm and }}
u(t_0), (\nabla u)(t_0), \dots, (\nabla^{m-2} u)(t_0), (\nabla^{m} u)(t_0)
{\mbox{\rm \ are linearly dependent}}\}.
\\
S'_{\nabla} := \ &\{ j^r(c, u)(t_0) \mid \
u(t_0), (\nabla u)(t_0),
\dots, (\nabla^{m-1} u)(t_0)
{\mbox{\rm \ are linearly dependent}}
\\
&
\quad {\mbox{\rm and }} c(t_0) = 0
\}
\\
S''_{\nabla} := \ &\{ j^r(c, u)(t_0) \mid c(t_0) = c'(t_0) = 0 \}.
\end{align*}
Then, for any but fixed system of local coordinates around $q$ of $M$,
$S_{\nabla}, S'_{\nabla}, S''_{\nabla}$ are
algebraic sets of codimension $\geq 2$.
Let $S_{\nabla}(I, M), S'_{\nabla}(I, M), S''_{\nabla}(I, M)$ be the corresponding
subbundle of $J^r(I, \mathbf{R}\times(TM \setminus \zeta))$ over $I\times M$.
For any subinterval $J \subset I$,
we set
\begin{align*}
{\widetilde{\mathcal O}}_J :=
\{ (c, u) : I \to \mathbf{R} \times (TM \setminus \zeta) \mid & \
j^r(c, u) : I \to J^r(I, \mathbf{R}\times(TM\setminus\zeta))
\\
&
{\mbox{\rm \ is transverse to }}
S_{\nabla}(I, M),
S'_{\nabla}(I, M), S''_{\nabla}(I, M)
{\mbox{\rm \ over \ }} J
\}.
\end{align*}
Then $\widetilde{{\mathcal O}} = \widetilde{{\mathcal O}}_I$
is open dense in Whitney $C^\infty$ topology.
Let $(c, u) \in \widetilde{{\mathcal O}}$ and $t_0 \in I$.
Then $j^r(c, u)(t_0) \not \in S_\nabla \cup S'_{\nabla} \cup S''_{\nabla}$.
Since $j^r(c, u)(t_0) \not\in S_{\nabla}$,
we have that the $\nabla$-type of $u$ is $(1, 2, \dots, m-1, m)$ or $(1, 2, \dots, m-1, m+1)$.
Since $j^r(c, u)(t_0) \not\in S'_{\nabla}$, if $c(t_0) = 0$ then
$\nabla$-type of $u$ must be $(1, 2, \dots, m-1, m)$.
On the other hand, since $j^r(c, u)(t_0) \not\in S''_{\nabla}$,
we have that $c(t_0) \not= 0$ or $c(t_0) = 0, c'(t_0) \not= 0$, i.e.
the order of $c$ at $t_0$ is $0$ or $1$.
We set
$$
{\mathcal O}_J := \{ \widetilde{\gamma} : I \to PTM {\mbox{\rm \ integral\ }} \mid \exists (c, u) \in {\widetilde{\mathcal O}}_J, \ \widetilde{\gamma}(t) = \langle u(t)\rangle_{\mathbf{R}}, (\pi\circ\widetilde{\gamma})'(t) = c(t)u(t) \}.
$$
We will show, for any compact subinterval $J \subset I$, that ${\mathcal O}_J$ is open dense
and ${\mathcal O} = {\mathcal O}_I$ is open dense in the space of
integral curves with Whitney $C^\infty$ topology.
That ${\mathcal O}_J$ and ${\mathcal O}$ are open is clear, since ${\widetilde{\mathcal O}}$ is open.
We will show ${\mathcal O}_J$ is dense.
Let $\widetilde{\gamma} : I \to PTM$ be any integral curve and
${\mathcal I}$ be any open neighborhood of $\widetilde{\gamma}$.
Set $\gamma = \pi\circ{\widetilde{\gamma}} : I \to M$.
Take any frame $u$ associated to $\widetilde{\gamma}$.
Then there exists uniquely $c : I \to \mathbf{R}$ which satisfies $c(t)u(t) = \gamma'(t), t \in I$.
Take a compact subinterval $J' \subset I$ such that
$J \subsetneq J'$.
We approximate $(c, u)$ by some $(e, v) \in {\widetilde{\mathcal O}}_J$ and
that $(e, v) = (c, u)$ outside of $J'$.
Then $v$ generates a curve $\rho : I \to PTM, \rho(t) = \langle v(t)\rangle_{\mathbf{R}}$ ,
which approximates $\widetilde{\gamma}$, however $\rho$ may not be an integral curve.
Consider the vector field $(\frac{\partial}{\partial t}, e v)$ along the graph of $\pi\circ\rho$ in $I\times M$.
Extend $(\frac{\partial}{\partial t}, v(t))$ to a vector field $(\frac{\partial}{\partial t}, V(t, x))$ over $I\times M$ with
a support contained in $I\times K$ for some compact $K \subset M$. Take $t_0 \in J$.
Take the integral curve $\alpha : I \to M$
of the vector field $(\frac{\partial}{\partial t}, e(t)V(t, x))$ through $(t_0, \alpha(t_0))$.
Then $\alpha'(t) = e(t) V(t, \alpha(t))$. Define the vector field $w : I \to TM$ over $\alpha$ by
$w(t) = V(t, \alpha(t))$. Then we have $\alpha'(t) = e(t)w(t)$.
If we choose $(e, v)$ sufficiently close to $(c, u)$,
then $w(t) \not= 0$ and $(e, w) \in {\widetilde{\mathcal O}}_J$.
However the integral curve $\widetilde{\alpha}$ defined by $w$ may not belong to ${\mathcal I}$,
an open set for Whitney $C^\infty$ topology.
Further we modify the perturbation $(e, v)$ over $J' \setminus J$
and the extension $V$ over $(J' \setminus J)\times M$ to obtain
an integral curve $\widetilde{\beta}$ such that $\widetilde{\beta} = \widetilde{\alpha}$ on $J$ and
$\widetilde{\beta} = \widetilde{\gamma}$ outside of $J'$, using the method of Lemma \ref{perturbation}.
Then the integral curve $\widetilde{\beta}$
approximates $\widetilde{\gamma}$ and belongs to ${\mathcal I}$, while
$(e, w) \in \widetilde{\mathcal O}_J$.
Thus we have seen that ${\mathcal O}_J$ is dense, for any compact subinterval $J \subset I$.
Since ${\mathcal O} = \cap_{J \subset I} {\mathcal O}_J$, the intersection over
compact subintervals $J \subset I$, we have that
${\mathcal O}$ is residual, and therefore that ${\mathcal O}$ is dense in Whitney $C^\infty$ topology
\cite{GG}.
Thus we have that ${\mathcal O}$ is open dense in Whitney $C^\infty$ topology.
Then, using Lemma \ref{c, u},
we have the required result.
\hfill $\Box$
\begin{Rem}
{\rm
By the same method as above, we have that
the codimension of jets of integral curves such that the projections are of $\nabla$-type
$(a_1, a_2, \dots, a_m)$ is given by
$$
\ell + \sum_{i=1}^m (b_i - i) = a_1 - 1 + \sum_{i=2}^m (a_i - a_1 - i +1),
$$
for any affine connection $\nabla$. Note that
the codimension is calculated in Theorem 5.6 of \cite{Ishikawa4} in
the flat case
(cf. Theorem 5.8, Theorem 3.3 of \cite{Ishikawa4}).
}
\enr
\
\noindent
{\it Proof of Theorem \ref{genericity-theorem2}:}
We may suppose $\nabla$ is torsion free (Remark \ref{torsion-free}). Then Proposition \ref{genericity2}
and Theorem \ref{characterization-theorem} imply Theorem \ref{genericity-theorem2}.
\hfill $\Box$
\
\section{Tangent surfaces to torsionless curves and fold singularities}
\label{Tangent surfaces to torsionless curves and fold singularities}
Let $(M, \nabla)$ be a manifold of dimension $m \geq 3$
with a torsion free affine connection $\nabla$.
Consider a curve $\gamma : I \to M$ such that $(\nabla\gamma)(t), (\nabla^2\gamma)(t)$ are linearly
independent for any $t \in I$. Though the torsion is not defined in general, we can define:
\begin{Def}
{\rm
A curve $\gamma : I \to M$ is called {\it torsionless} if
$(\nabla\gamma)(t), (\nabla^2\gamma)(t)$ are linearly
independent but
$(\nabla\gamma)(t), (\nabla^2\gamma)(t), (\nabla^3\gamma)(t)$
are linearly dependent everywhere.
}
\stepcounter{Th
The situation is, by any means, non-generic. However
to study torsionless curves is an interesting geometric problem (cf. \cite{Cartan}).
\begin{Rem}
{\rm
Let $M$ be a Riemannian manifold with the Levi-Civita connection $\nabla$. The notion of torsion is well-defined for the curve $\gamma$ parametrized by the arc-length, if $\nabla\gamma, \nabla^2\gamma$ are linearly independent.
Then the condition that the torsion of $\gamma$ is zero if and only if $\gamma$ satisfies the third order non-linear ordinary differential equation,
$$
\nabla^3\gamma - \frac{\Vert\nabla^2\gamma\Vert'}{\Vert\nabla^2\gamma\Vert} \, \nabla^2\gamma
+ \Vert\nabla^2\gamma\Vert^2 \, \nabla\gamma = 0.
$$
Then in particular $\gamma$ is torsionless in our sense.
}
\enr
If $(M, \nabla)$ is projectively flat, namely, if it is
projectively equivalent to the Euclidean space $(E^m, \nabla_0)$
with the standard connection $\nabla_0$, then it is well-known that any torsionless curve is a \lq\lq plane curve",
therefore its $\nabla$-tangent surface is \lq\lq folded" into a totally geodesic surface.
Note that a Riemannian manifold is locally projectively flat if and only if it has a constant curvature
(Beltrami's Theorem, see p.352 of \cite{Sharpe} or p.97 of \cite{Eisenhart}).
A map-germ $f : (\mathbf{R}^2, p) \to (M, f(p))$ is called a {\it fold}, or an {\it embedded fold}, if it is diffeomorphic to
$$
(t, s) \mapsto (t + s, t^2 + 2st, 0, \dots, 0),
$$
which is diffeomorphic also to $(u, w) \mapsto (u, \frac{1}{2}w^2, 0, \dots, 0)$.
The fold singularities appear in other geometric problems also (see \cite{FRUYY} for instance).
For our problem, we have
\begin{Prop}
\label{tangent-to-torsionless-curve}
Let $(M, \nabla)$ be locally projectively flat around $q \in M$ and
$\gamma : (\mathbf{R}, t_0) \to (M, q)$ a germ of torsionless curve.
Then
the germ of $\nabla$-tangent surface $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma) : (\mathbf{R}^2, (t_0, 0) \to
(M, q)$ to $\gamma$ is a fold.
In particular it is a generically two-to-one mapping.
\stepcounter{Th
\Proof
There exits a germ of projective equivalence $\varphi : (M, q) \to (E^m, 0)$ such that
$\varphi\circ \gamma : (\mathbf{R}, 0) \to (E^m, 0)$ is a strictly convex plane curve in $E^2\times \{ 0\} \subset E^m$.
Then, by Theorem \ref{characterization2},
the tangent surface
$\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\varphi\circ\gamma) = \varphi\circ(\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma))$
is diffeomorphic to a fold regarded as a map-germ $(\mathbf{R}^2, (0, 0)) \to (\mathbf{R}^2, (0, 0))$.
Therefore $\nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)$ is diffeomorphic to a fold,
regarded as a surface-germ $(\mathbf{R}^2, (t_0, 0)) \to (M, q)$.
\hfill $\Box$
\
A map-germ $f : (\mathbf{R}^2, p) \to (M^3, f(p))$ to a three dimensional space
is called a {\it $(2, 5)$-cuspidal edge}, if it is diffeomorphic to
$$
(u, w) \mapsto (u, w^2, w^5).
$$
Then $f$ is a frontal with a non-degenerate singular point at $p$,
and the characteristic function $\psi$ vanishes identically (see \S \ref{Cuspidal edge and folded umbrella}).
In fact, the model germ
$(u, w) \mapsto (u, w^2, w^5)$
(the \lq\lq suspension" of $(2, 5)$-cusp) is a frontal
with a non-degenerate singular point at $0$, the kernel field $\partial/\partial u$ for $f_*$ is transverse to the singular curve $w = 0$, and
the characteristic function $\psi(u) \equiv 0$.
However $f$ is injective and it is never diffeomorphic to a tangent surface to any
torsionless curve in a projectively flat space, by Proposition \ref{tangent-to-torsionless-curve}.
Nevertheless it can be a $\nabla$-tangent surface of a torsionless curve
for a torsion free affine connection on $M^3$.
\begin{Ex}
{\rm
Let $\nabla$ be the torsion free affine connection on $\mathbf{R}^3$ with coordinates $x_1, x_2, x_3$
defined by
$\Gamma^\lambda_{\mu\nu} = x_1 + x_2^2$ if $(\lambda, \mu, \nu) = (3, 1, 2), (3, 2, 1)$ and
otherwise $\Gamma^\lambda_{\mu\nu} = 0$. Let $\gamma : \mathbf{R} \to \mathbf{R}^3$
be the immersion defined by $\gamma(t) = (-t^2, t, 0)$. Then $\Gamma^\lambda_{\mu\nu} = 0$ along
$\gamma$. We have $(\nabla\gamma)(t) = (-2t, 1, 0), (\nabla^2\gamma)(t) = (-2, 0, 0),
(\nabla^3\gamma)(t) = (0, 0, 0)$. Therefore $\gamma$ is torsionless.
For any $t_0 \in \mathbf{R}$, the $\nabla$-geodesic with the initial condition $(\gamma(t_0), \gamma'(t_0))$
is given by $(-2t_0s - t_0^2, s + t_0, \frac{1}{3}t_0 s^4)$. Therefore we have
$$
f(t, s) = \nabla{\mbox{\rm -}}{\mbox {\rm Tan}}(\gamma)(t, s) = (-2ts - t^2, \ s+t, \ \frac{1}{3}ts^4).
$$
Set $u = s+t, w = s$. Then we see that $f$ is diffeomorphic to
$(u, w) \mapsto (-u^2+w^2, u, \frac{1}{3}uw^4 - \frac{1}{3}w^5)$, which is diffeomorphic to
$(u, w) \mapsto (u, w^2, w^5)$. Therefore $f$ is a $(2, 5)$-cuspidal edge.
}
\ene
We conclude this section by posing the problem on singularities of $\nabla$-tangent surfaces
to torsionless curves in the case of a general affine connection $\nabla$.
{\small
|
\section{Introduction}
All maps and manifolds in the paper are assumed to be smooth and by compact manifold we also mean closed (without boundary). Let $f\!:M\to X$ be an immersion of a real compact $2n$ dimensional manifold $M$ into a complex $n\!+\!1$ dimensional manifold $(X,J)$. A point $p\in M$ is called \textit{CR regular} for the immersion $f$ if the complex dimension of $f_*T_pM\cap Jf_*T_pM\subset T_{f(p)}X$ equals the expected dimension $n\!-\!1$. Points on $M$ that are not CR regular are called \textit{complex} or \textit{CR singular} points, and the complex tangent space at those points has the full dimension $n$.
\begin{definition} An immersion (embedding) $f\!:M\to X$ is \textit{CR regular}, if all points on $M$ are CR regular for the immersion (embedding).
\end{definition}
Although it will also be obvious from the next section, we give a simple argument that gives a necessary condition for an oriented $4$-manifold $M$ to have a CR regular immersion into $\mathbb{C}^3$. Let $f\!:M\to \mathbb{C}^3$ be a CR regular immersion. Then $L=f^*(f_*TM\cap if_*TM)\to M$ is a complex line bundle on $M$. Let $E\to M$ be its orthogonal complement in $TM$ ($E$ is an $SO(2)=U(1)$ bundle). Since $(E\otimes\mathbb{C})\oplus L=f^*T\mathbb{C}^3$ is a trivial complex bundle, it follows from the Whitney sum formula for Chern classes that
\[(1+c_1(E\otimes\mathbb{C})+c_2(E\otimes\mathbb{C}))\cup (1+c_1(L))=1.\]
Since $c_1(E\otimes\mathbb{C})=0$, we have $c_1(L)=0$, and thus also $c_2(E\otimes\mathbb{C})=-p_1(E)=0$. So $L$ is trivial and, since $TM=L\oplus E$, $\chi(M)=0$ and $p_1(M)=p_1(E)=0$. We have shown that if $M$ can be CR regularly immersed in $\mathbb{C}^3$, is must satisfy $\chi(M)=0$ and $p_1(M)=0$. If, in addition, $f\!:M\to \mathbb{C}^3$ is an embedding, then by a result of Thom, the $\mathrm{mod}\ 2$ Euler class of the normal bundle of $M$ vanishes, and so by the Whitney sum formula, $w_2(M)$ also vanishes (we thank Peter Landweber for pointing this out).
The main theorem of this paper shows that the above conditions are also sufficient for the existence of CR regular immersions and embeddings of compact oriented $4$-manifolds into $\mathbb{C}^3$.
\begin{theorem}\label{thm} Let $M$ be a real compact orientable $4$-manifold. Then
\begin{itemize}
\item[(i)] $M$ has a CR regular immersion into $\mathbb{C}^3$ if and only if $\chi(M)=0$ and $p_1(M)=0$,
\item[(ii)] $M$ has a CR regular embedding into $\mathbb{C}^3$ if and only if $\chi(M)=0$, $p_1(M)=0$ and $w_2(M)=0$.
\end{itemize}
\end{theorem}
The vanishing of the second Stiefel-Whitney class $w_2(M)$ (and $M$ being orientable) is equivalent to $M$ being spin. By Hirzebruch's signature theorem we have $\sigma(M)=\frac{p_1(M)[M]}{3}$, where $\sigma(M)=b_2^+-b_2^-$ is the signature of (the intersection form of) $M$, so the vanishing of the first Pontryagin class $p_1(M)$ is equivalent to $\sigma(M)=0$. Since $SO(4)$-bundles are determined by their Euler, first Pontryagin and second Stiefel-Whitney class, we get the following corollary of the main theorem:
\begin{corollary} A compact orientable $4$-manifold has a CR regular embedding into $\mathbb{C}^3$ if and only if it is parallelizable.
\end{corollary}
\begin{examples} We give some applications of the above theorem:
\begin{itemize}[leftmargin=*]
\item Since every compact orientable $3$-manifold $Y$ is parallelizable, $Y\times S^1$ can be CR regularly embedded into $\mathbb{C}^3$.
\item No simply connected compact orientable $4$-manifold can be CR regularly immersed into $\mathbb{C}^3$, since it always has the Euler characteristic at least $2$.
\item $\mathbb{C}\mathrm{P}^2\#\overline\mathbb{C}\mathrm{P}^2\#2(S^1\times S^3)$ can be regularly immersed into $\mathbb{C}^3$. In fact, any compact oriented $4$-manifold can be immersed into $\mathbb{C}^3$ after perhaps a connected sum with some copies of $\mathbb{C}\mathrm{P}^2$ or $\overline\mathbb{C}\mathrm{P}^2$, and $S^1\times S^3$ or $S^2\times S^2$.
\end{itemize}
\end{examples}
\begin{remark} If $M$ is a compact oriented $4$-manifolds, then by a recent result of Jacobowitz and Landweber \cite{JL}, the vanishing of the first Pontryagin class $p_1(M)$ is equivalent to $M$ having both an independent map into $\mathbb{C}^3$ and a totally real immersion into $\mathbb{C}^4$.
\end{remark}
\section{Complex points}
We assume from now on that $M$ is a compact oriented $4$-manifold, and $X$ is a complex $3$-manifold, and let $f\!:M\to X$ be an immersion. Let $Gf\!:M\to\widetilde{\mathrm{Gr}}_{4} f^*TX$ be the Gauss map of $M$ into the oriented real Grassmann bundle of the pull-back of $TX$, and let $(\P_\c f^*T^*X)^\pm\subset \widetilde{\mathrm{Gr}}_{4} f^*T^*X$ be the subbundle of complex hyperplanes, with their orientation either agreeing ($+$) or disagreeing ($-$) with the induced orientation from $TX$. By Thom's transversality theorem, for generic $f$ (which we also assume from now on), $\mathrm{G} f(M)$ and $(\P_\c f^*T^*X)^\pm$ intersect transversally, and so $M$ has only finitely many complex points by a simple dimension count.
A complex point $p\in M$ is called \textit{positive}, if at $p$, the manifold $M$ intersect $(\P_\c f^*T^*X)^+$, and \textit{negative}, if it intersect $(\P_\c f^*T^*X)^-$; it is called \textit{elliptic}, if that intersection is positive, and \textit{hyperbolic}, if it is negative. Let $e_\pm(M)$ and $h_\pm(M)$ be the numbers of positive or negative elliptic or hyperbolic complex points on $M$. For \textit{Lai indices}, $I_\pm(M)=e_\pm(M)-h_\pm(M)$, we have topological formulas (\cite{Lai}) in terms of Chern classes of $X$ and Euler classes of the tangent and normal bundle of $M$:
\[2I_{\pm}(M)=(e(TM)\pm c_2(X)|_M+e(NM)\cup c_1(X)|_M\pm e(NM)^2)[M].\]
Since $e(NM)^2=p_1(NM)=p_1(TX)|_M-p_1(TM)$, in the case of $X=\mathbb{C}^3$, Lai formulas have a simple form
\[\label{lai2}2I_{\pm}(M)=\chi(M)\mp p_1(M).\]
Since Lai indices are invariant under regular homotopy of immersions, the vanishing of the indices is a necessary condition for the immersion to be regularly homotopic to a CR regular immersion. As a direct consequence of the Cancellation theorem (\cite{S1}) we have a converse statement:
\begin{proposition}\label{cancel} Let $f\!:M\to X$ be a smooth generic immersion (embedding) of a real compact oriented $4$-manifold $M$ into a complex $3$-manifold $X$. Let $d$ be some metric on $X$ and let $\varepsilon>0$. If $I_\pm(M)=0$, there exists a regular homotopy (isotopy) $f_t:M\times [0,1]\to X$, so that
\begin{itemize}
\item[(i)] $d(f_t(p),f(p))<\varepsilon$ for every $t\in [0,1]$ and every $p\in M$.
\item[(ii)] $f_1\!:M\to X$ is a CR regular immersion (embedding).
\end{itemize}
\end{proposition}
\begin{remark} The Cancellation theorem in \cite{S1} is proven for embeddings, but it is trivially adapted to immersions by passing to the normal bundle. The theorem holds in all dimensions and is essentially a consequence of Gromov's convex integration (\cite{Gr}). For real surfaces in complex surfaces, this is a classical result of Eliashberg and Harlamov (\cite{EH}).
\end{remark}
\section{Proof of Theorem \ref{thm}}
The proof of the main theorem is a combination of classical results about immersions and embedding of orientable $4$-manifolds into $\mathbb{R}^6$ and Proposition \ref{cancel}. While some of the arguments below are probably well known to the experts, we include them both for completeness and because we were not able to find them in print.
Let us first look at CR regular immersions into $\mathbb{C}^3$. We have already seen that the vanishing of the Euler characteristic and the first Pontryagin class is a necessary condition for the existence of CR regular immersions. Let us look at the converse statement.
By Hirsh's immersion principle, an oriented compact $4$-manifold $M$ can be immersed into $\mathbb{R}^6$ if and only if there exists an $SO(2)$-bundle $L$ on $M$, so that $TM\oplus L$ is trivial ($L$ is the normal bundle of such an immersion). By a result of Cappell and Shaneson (\cite{CS}), the existence of such an $L$ is equivalent to the existence of an integral lift $\alpha\in H^2(M,\mathbb{Z})$ of the second Stiefel-Whitney class $w_2(M)$, so that $\alpha\cup \alpha=-p_1(M)$. The cohomology conditions are obviously necessary. We sketch the proof of sufficiency (details can be found in \cite{K}). Let $L$ be an $SO(2)=U(1)$-bundle with $c_1(L)=\alpha$, $w_2(M)=\alpha\ (\mathrm{mod}\ 2)$ and $\alpha\cup \alpha=-p_1(M)$. Since $TX\oplus L$ is orientable, it can be trivialized over the $1$-skeleton of $M$. The obstruction to trivialization over the $2$-skeleton of $M$ is precisely $w_2(TX\oplus L)\in H^2(M,\pi_1(SO(6))=H^2(M,\mathbb{Z}_2)$, which by construction vanishes. Since $\pi_2(SO(6))$ is trivial, $TX\oplus L$ can be further trivialized over the $3$-skeleton of $M$, and the obstruction for the trivialization over the whole $M$ turns out to be $p_1(TX\oplus L)\in H^4(M,\pi_3(SO(6))=H^4(M,\mathbb{Z})$. Since $p_1(TX\oplus L)=p_1(M)+c_1(L)^2=0$, $TX\oplus L$ is trivial.
We now show that if $p_1(M)=0$, $M$ can be immersed into $\mathbb{R}^6$. By Hirzebruch's signature theorem, $\sigma(M)=0$, and by classification of signature $0$ unimodular forms (see for example \cite{GS}), the intersection form
\[Q:H^2(M,\mathbb{Z})/\mathrm{Tor}\times H^2(M,\mathbb{Z})/\mathrm{Tor}\to\mathbb{Z}\] is equivalent to either $k\langle 1\rangle\oplus k\langle -1\rangle$ or $kH$, where $H=\left[\begin{smallmatrix} 0&1\\ 1&0\end{smallmatrix}\right]$. Since every oriented $4$-manifold is spin$^\mathrm{c}$, there exist an integral lift $\alpha\in H^2(M,\mathbb{Z})$ of $w_2(M)$. Let $\alpha=\alpha_F+\alpha_T\in H^2(M,\mathbb{Z})/\mathrm{Tor}\oplus \mathrm{Tor}$ be the splitting of $\alpha$ into the free and the torsion part. Let us first assume that $Q$ is equivalent to $k\langle 1\rangle\oplus k\langle-1\rangle$, and let $a_1,a_2,\ldots,a_{2k}$ be the corresponding basis of $H^2(M,\mathbb{Z})/\mathrm{Tor}$. Let us write $\alpha_F=n_1a_1+n_2a_2+\cdots+n_{2k}a_{2k}$ in this basis. By Wu's formula, $n_j=Q(\alpha_F,a_j)$ and $Q(a_j,a_j)$ have the same parity for every $1\le j\le 2k$, so $n_j$ must all be odd. Define $\widetilde\alpha=a_1+a_2+\cdots+a_{2k}+\alpha_T\in H^2(X,\mathbb{Z}).$ Since the $\mathrm{mod}\ 2$ reduction of $\widetilde \alpha-\alpha=(n_1-1)a_1+(n_2-1)a_2+\cdots+(n_{2k}-1)a_{2k}$ is trivial, $w_2(M)=\widetilde\alpha\ (\mathrm{mod}\ 2)$. We also have
$(\widetilde\alpha\cup\widetilde\alpha)[M]=Q(a_1+a_2+\cdots+a_{2k},a_1+a_2+\cdots+a_{2k})=0=-p_1(M)[M],$ so $M$ can be immersed in $\mathbb{R}^6$. For $Q$ equivalent to $kH$, the proof is essentially the same, except that in this case, all $n_j$ must be even, and we can just take $\widetilde\alpha=\alpha_T$.
Let now $M$ be an orientable manifold with $\chi(M)=0$ and $p_1(M)=0$. The above argument gives us an immersion of $M$ into $\mathbb{C}^3$ and since $2I_\pm(M)=\chi(M)\mp p_1(M)[M]=0$, we get that $M$ can be regularly homotoped to a CR regular immersion by Proposition \ref{cancel}.
We now prove the statement about embeddings. By a theorem of Cappell and Shaneson (\cite{CS}), an oriented $4$-manifold can be embedded into $\mathbb{R}^6$ if and only if $M$ is spin, $\chi(M)=0$ and $p_1(M)=0$ (for a complete proof, see \cite{R}). Again by Proposition \ref{cancel}, such an embedding can be isotoped to a CR regular embedding.
\begin{remark} We can see from the proof above that if $\chi(M)=0$ and $p_1(M)=0$, one can construct many regularly non-homotopic CR immersions into $\mathbb{C}^3$ (unless $H^2(M,\mathbb{Z})$ is completely torsion). If the intersection form of $M$ is equivalent to $k\langle 1\rangle\oplus k\langle -1\rangle$, we could, in the above proof, take any $\tilde\alpha=m_1a_1+m_2a_2+\cdots+m_{2k}a_{2k}+\alpha_T$, as long as all $m_j$ are odd and $m_1^2+\cdots+m_k^2=m_{k+1}^2+\cdots+m_{2k}^2$. If the intersection form of $M$ is equivalent to $kH$, we could take any $\tilde\alpha=m_1a_1+m_2a_2+\cdots+m_{2k}a_{2k}+\alpha_T$, as long as all $m_j$ are even and $m_1^2+m_3^2+\cdots+m_{2k-1}^2=m_2^2+m_4^2+\cdots+m_{2k}^2$. All these different choices provide us with regularly non-homotopic CR immersions, since they all have different normal bundles.
\end{remark}
\begin{remark} For a final remark, notice that if a compact oriented $4$-manifold $M$ has a CR regular immersion into any complex $3$-manifold $(X,J)$, $f\!:M\to X$, then $TM$ splits as a direct sum of two $SO(2)=U(1)$-bundles, $TM=L\oplus E$, where $L=f^*f_*TM\cap f^*Jf_*TX$, and $E$ its orthogonal complement in $TM$ (in the case of $X=\mathbb{C}^3$, $L$ is trivial, as we have seen in the first section). So $M$ has an almost complex structure. A simple calculation using Lai's formulas shows, that all CR regular embeddings into $\mathbb{C}\mathrm{P}^3$ are homologically trivial.
\end{remark}
\bibliographystyle{amsplain}
|
\section{Introduction}
Circumnuclear star formation rings are present in approximately 20\% of nearby disk galaxies \citep{Comeron2010}. Early theories suggested that the radius at which the rings form was the inner Lindblad resonance (ILR) radius, which is set by an asymmetric component in the stellar disk such as a large scale bar. However, further modeling of galaxy gravitational potentials has shown that circumnuclear ring formation is linked more particularly to the presence of a sufficient number of $x_2$ orbits \citep{Regan2003}. Such orbits are (near-)circular, in contrast to the elongated $x_1$ orbits at larger radii, which make up the large scale bar \citep{1980A&A....92...33C}. The radii of $x_2$ orbits are close to the ILR radius.
As gas and dust lose angular momentum, they move inward. At the overlap between $x_1$ to $x_2$ orbits, gas and dust will settle on the lower energy $x_2$ orbits as a result of collisions, forming a circumnuclear ring. The discrete locations where this transition happens are generally called `inflow points'. Such inflow points are most obvious in ring systems with large scale stellar bars, where the connection between the spiral shocks of the bar and the circumnuclear ring can often be clearly seen \citep[for example NGC\,6951,][]{Tessel}. However, circumnuclear rings also exist in galaxies classified as non-barred \citep[see, e.g.,][for a discussion on the difference, or lack thereof, between barred and non-barred systems in this context]{2010AJ....139.2465G}. Any gravitational asymmetry in the disk may suffice, as long as enough $x_2$ orbits are present in the gravitational potential.
The gas and dust can experience shocks as they transition from $x_1$ to $x_2$ orbits at the inflow points. However, after the gas and dust settles on the $x_2$ orbits of the ring, the accreted ISM experiences very little large-scale friction. With that in mind, it is not too surprising that it has been found in multiple galaxies that most (if not all) star formation in circumnuclear rings takes place at the inflow points \citep{2001MNRAS.323..663R,2005ApJ...633L..25A,Boker2008,2008ApJS..174..337M,Tessel2,Tessel3,2014MNRAS.438..329F}. Consequently, it is thought that, owing to the absence of large scale friction or temporal variations in gravitational potential, stars and gas primarily move through the circumnuclear ring together. The stellar age gradient regularly seen stretching out from each inflow point is often referred to as `pearls-on-a-string', first coined by \citet{Boker2008}. An alternative theory \citep{Elmegreen1994} suggests that the inflow points play no special role in star formation, and that a build up of gas will trigger a burst of star formation after a critical limit has been reached anywhere in a ring. This is often referred to as the `popcorn' theory of star formation in rings.
Recent star formation is often tracked by observing hydrogen recombination lines such as H$\alpha$ or Br$\gamma$. The equivalent width of the Balmer lines is, for example, dependent on stellar age from $\sim$4\,Myr \citep[e.g.][]{1995ApJS...96....9L,1997AJ....113..975B,1999ApJS..125..489G}. Alternatively, emission lines from other atoms and molecules stimulated by stellar radiation, such as warm H$_2$, HeI, and [FeII], in the optical and NIR can be used. With a combination of these lines it is also possible to observe potential age gradients, as demonstrated for the circumnuclear ring in NGC\,613 \citep{Boker2008,2014MNRAS.438..329F}. However, strong extinction can make observations in the optical and NIR range difficult, especially at the inflow points where dust and gas build up, potentially limiting the effectiveness of these methods.
When there are no large-scale mechanical stresses, stellar radiation is most likely to be the prevalent form of heating for the neutral gas in the ring (excluding contributions of CR or X-ray ionization). Thus, an alternative way of tracking recent star formation is to study the heating and cooling of the neutral interstellar medium (ISM) via strong emission lines of the neutral gas in the FIR. In the presence of star formation, the heating of the neutral ISM is dominated by photo-electric heating \citep{1994ApJ...427..822B,2001ApJS..134..263W,2001ApJ...548L..73H}. Dust grains and PAHs eject electrons as the result of the absorption of FUV (6eV $<$ h$\nu$ $<$ 13.6eV) photons from young O and B type stars. These electrons in turn interact with the gas, heating atoms and molecules. The cooling times for dust and gas are much shorter than the Myr astronomical timescales of interest here. Thus, gas heating and cooling, via photo-electric heating, may be considered in balance at all times, and the observed cooling flux can be used as a proxy for the heating.
To test whether the heating of the neutral ISM in the ring is indeed dominated by stellar radiative heating from stars formed at the inflow points, we investigate the case of the circumnuclear ring in NGC\,4736, a nearby (D=4.66\,Mpc, $i\sim35\degr$, PA$\sim300\degr$) SAab spiral galaxy. The ring in NGC\, 4736 has a radius of 40$\arcsec$ (0.9\,kpc). Based on a previously observed HI velocity field \citep{2008AJ....136.2563W} and assuming that the spiral arms are concave, the north side of the disk is the near side, and the south the far side, with a clockwise rotation. We have combined available data to form a large multiwavelength dataset, tracing the gas and the recent star formation, including {\it Herschel} PACS data from the KINGFISH (Key Insights on Nearby Galaxies: A Far-Infrared Survey with Herschel) Open Time Key Program \citep{2011PASP..123.1347K}. The PACS observations enable us to probe the main neutral gas cooling lines of $[$CII$]$ at 158$\mu$m and $[$OI$]$ at 63$\mu$m, as well as the FIR dust continuum. We also include data obtained with the {\it Spitzer} Space Telescope as part of the SINGS ({\it Spitzer} Infrared Nearby Galaxies Survey) Legacy program \citep{2003PASP..115..928K}, to probe the mid-infrared wavelength range, which includes emission from large dust grains, PAH's, stars, and molecular and atomic gas.
In \S2 we present the data. In \S3 the atomic and molecular gas mass surface densities, and subsequent dust mass surface density, are computed. In \S4 a census is made of the star formation rate density throughout the ring. In \S5 the contribution of the neutral ISM to the observed flux is determined. We then compare the dust and gas cooling with the stellar heating in \S6. The implications are discussed in \S7, with a summary in \S8.
\section{Data sets}
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics[width=17cm]{N4736_overlay.eps}}
\caption{Spatial coverage of the spectral data sets presented in this study, overlaid on the {\it GALEX} FUV image of NGC\,4736. The circumnuclear ring at a radius of r$\sim$40$\arcsec$ (0.9\,kpc) stands out clearly in the {\it GALEX} image. The different regions mapped spectroscopically, and indicated here in color, are: {\it Spitzer/IRS} long-low (14-38 $\mu$m) spectra (dark blue contours), {\it Herschel/PACS} [CII] and [OI] emission line maps (green contours), {\it Herschel/PACS} [NII]122$\mu$m, [NII]205$\mu$m emission line maps (yellow contours), and {\it Spitzer/IRS} short-low (5-14$\mu$m) spectra (purple contours). In this image, north is up and east is to the left, and the total size is approximately 7.3 arc minutes on a side.}
\label{fig:overlay}
\end{figure}
\subsection{Herschel PACS}\label{sect:data}
Both spectra and images were obtained for NGC\,4736 with the {\it Herschel} PACS instruments as part of the KINGFISH Open Time Key Program. The PACS spectroscopic instrument is an IFU (integral field unit) with 5$\times$5, 9\farcs4 pixels \citep{2010A&A...518L...2P}. This gives an instantaneous Field-of-View (FoV) of 47\arcsec\,$\times$\,47\arcsec. The PACS beam FWHM is 5\farcs2 in the wavelength range 60-85\,$\mu$m, and 12\arcsec\, at 130-210\,$\mu$m. For the spectra several targeted pointings were made for the various lines. The [NII]122$\mu$m and [NII]205$\mu$m lines were observed with 2 adjacent pointings, one focussed on the nucleus and one on the circumnuclear ring (see yellow contours in Fig. \ref{fig:overlay}). The [CII]158$\mu$m and [OI]63$\mu$m lines were observed in a strip 200\arcsec\, long (green contours), with an orientation equal to previous {\it Spitzer} IRS data (blue contours), and one adjacent pointing on the circumnuclear ring (equal to the [NII] off-nucleus pointing). The FIR continuum was probed at 70\,$\mu$m, 100\,$\mu$m, and 160\,$\mu$m with the PACS photometer over the full extent of the galaxy.
The PACS observations were reduced consistently for the complete KINGFISH sample using the Scanamorphus package \citep{2013PASP..125.1126R}. The KINGFISH observations are described in \citet{2012ApJ...745...95D} and \citet{2012ApJ...747...81C}. In all cases, the images were rescaled to a pixel scale of 4\arcsec. The spectra zeroth-moment maps were convolved to the PACS\,160$\mu$m resolution of 12\arcsec\,, using a pixel scale of 4\arcsec\,, since the PACS\,160 beam is essentially the same as the [CII] beam.
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics[width=17cm]{FIR_paperfigure5.eps}}
\caption{Distribution of the Total InfraRed (TIR) continuum emission ({\it left}), [CII]\,158\,$\mu$m line emission ({\it middle}), and [OI]\,63\,$\mu$m line emission ({\it right}) in the circumnuclear ring of NGC\ 4736. The 12\arcsec\, apertures used to extract fluxes are indicated in the TIR and [CII] panels. The spatial scale is given in the bottom right corner, and the PACS\,160$\mu$m beam of $\sim$12\arcsec\, is indicated in the top left corner of each panel.}
\label{fig:FIRgas}
\end{figure*}
The [CII] and [OI] intensity maps, as well as the total-IR (TIR) continuum map are shown in Fig. \ref{fig:FIRgas}. The TIR continuum is constructed from a combination of four individual Spitzer and Herschel images, at 8\,$\mu$m, 24\,$\mu$m, 70\,$\mu$m, and 160\,$\mu$m, following the equation in \citet{2007ApJ...657..810D},
\begin{equation}\label{eq1}
TIR = \left(0.95\langle\nu\,S_{\nu}\rangle_{8} + 1.15\langle\nu\,S_{\nu}\rangle_{24} + \langle\nu\,S_{\nu}\rangle_{70} + \langle\nu\,S_{\nu}\rangle_{160}\right).
\end{equation}
\noindent Even though the FoV of the spectra clips the northern and southern part of the circumnuclear ring, the [CII] and [OI] emission is clearly concentrated there. The TIR continuum also has elevated emission at those locations, but is mostly dominated by the nucleus.
Integrated fluxes are extracted by covering the ring with 3x3 pixel (12\arcsec\,x12\arcsec\,) apertures. In this manner we measured the [CII] and [OI] emission line and continuum fluxes at 13 individual points, seven on the east side, and six on the west side of the ring. For the TIR continuum an additional six apertures are positioned on the ring outside the spectroscopic FoV (see Fig. \ref{fig:FIRgas}). This east (west) side corresponds to the blue (red) shifted side of the galaxy. The positions of, and fluxes in, these 12\arcsec\, apertures are given in Table \ref{tab:FIRfluxes}.
\begin{table*}
\begin{minipage}{2\columnwidth}
\centering
\caption{[CII], [OI], and TIR surface brightnesses in defined apertures}\label{tab:FIRfluxes}
\begin{tabular}{r c c c c c c c c c}
\hline\hline
\# & pos x & pos y & Angle & [CII] & [CII]$_{unc}$ & [OI] & [OI]$_{unc}$ & TIR & TIR$_{unc}$ \\
& [pix] & [pix] & [deg] & W/m$^2$/sr & rel. & W/m$^2$/sr & rel. & W/m$^2$/sr & rel. \\
\hline
1 &149 &144 & 15 & -- & -- & -- & -- & 1.29e-4 & 0.09 \\
2 &146 &143 & 29 & 1.12e-6 & 0.01 & 2.43e-7 & 0.12 & 2.64e-4 & 0.07 \\
3 &143 &145 & 49 & 1.13e-6 & 0.01 & 3.33e-7 & 0.06 & 2.84e-4 & 0.04 \\
4 &141 &148 & 68 & 1.62e-6 & 0.01 & 4.91e-7 & 0.04 & 4.54e-4 & 0.04 \\
5 &141 &151 & 85 & 1.75e-6 & 0.01 & 5.40e-7 & 0.03 & 4.88e-4 & 0.03 \\
6 &142 &154 & 102 & 1.60e-6 & 0.01 & 4.24e-7 & 0.05 & 3.98e-4 & 0.03 \\
7 &143 &156 & 117 & 1.11e-6 & 0.01 & 2.72e-7 & 0.09 & 2.94e-4 & 0.03 \\
8 &145 &158 & 135 & 8.87e-7 & 0.02 & 2.00e-7 & 0.23 & 2.56e-4 & 0.04 \\
9 &148 &159 & 157 & -- & -- & -- & -- & 2.70e-4 & 0.03 \\
10 &151 &159 & 180 & -- & -- & -- & -- & 2.56e-4 & 0.03 \\
11 &155 &161 & 205 & 8.50e-7 & 0.02 & 1.31e-7 & 0.19 & 2.20e-4 & 0.03 \\
12 &157 &161 & 213 & 7.83e-7 & 0.02 & 1.69e-7 & 0.11 & 2.12e-4 & 0.03 \\
13 &160 &160 & 229 & 1.05e-6 & 0.01 & 2.94e-7 & 0.06 & 2.66e-4 & 0.03 \\
14 &161 &157 & 243 & 1.34e-6 & 0.01 & 3.70e-7 & 0.05 & 3.28e-4 & 0.03 \\
15 &161 &154 & 259 & 1.44e-6 & 0.01 & 3.82e-7 & 0.05 & 3.54e-4 & 0.03 \\
16 &161 &151 & 275 & 1.40e-6 & 0.01 & 4.34e-7 & 0.08 & 3.38e-4 & 0.03 \\
17 &159 &148 & 297 & -- & -- & -- & -- & 2.88e-4 & 0.05 \\
18 &156 &146 & 320 & -- & -- & -- & -- & 3.42e-4 & 0.10 \\
19 &154 &145 & 337 & -- & -- & -- & -- & 3.14e-4 & 0.12 \\
20 &152 &144 & 353 & -- & -- & -- & -- & 2.56e-4 & 0.12\\
\hline
\end{tabular}\\
\begin{flushleft}
{\bf Notes:} [CII] and [OI] emission line fluxes, TIR continuum flux, and their relative uncertainties in 3x3 pixel (12\arcsec\,x12\arcsec\,) apertures. Central positions (in TIR map pixels) of the apertures are given in the 2nd and 3rd columns. The order of the apertures (angles) here is clockwise (subsequently following the direction of gas motion), starting from due south. For reference, pixel (151,152) corresponds to the galaxy center (RA 12h50m53.18s, Dec 41d07m12.0s).
\end{flushleft}
\end{minipage}
\end{table*}
\subsection{Spitzer}
As part of the SINGS survey NGC\,4736 was observed with {\it Spitzer}-IRS using the `short-low' and `long-low' spectral modules in spectral mapping mode. The circumnuclear ring is incompletely covered by the `long-low' footprint (see again Fig. \ref{fig:overlay}, blue contour), and partially by the three off-nuclear `short-low' pointings (purple contours). The `short-low' (SL) module has a pixel scale of 1\farcs85 and a resolving power of 60-120, decreasing with increasing wavelength. The observations have a FoV of 3\farcs6 by 27\arcsec\, and cover the wavelength range 5-14$\mu$m. The `long-low' (LL) module has a pixel scale of 5\farcs1, a resolving power of approximately 60-120, again decreasing with wavelength, a FoV of 10\farcs6 by 84\arcsec\,, and covers the wavelength range 14-28$\mu$m. The reduced and fully calibrated data cubes are available on the SINGS Spitzer and Ancillary Data website\footnote{http://irsa.ipac.caltech.edu/data/SPITZER/SINGS/}.
For this work we extracted the [SIII] emission line flux at 18.7\,$\mu$m and 33.5\,$\mu$m by fitting the lines and the underlying warm dust continuum with PAHFIT \citep{2007ApJ...656..770S} on the LL data cube alone. The [SIII] emission lines are not blended, and the dust emission is a smoothly varying composition of black-body curves of different temperatures. Consequently, the separation of line and continuum emission at these wavelengths with PAHFIT is straight-forward. A check at the positions with both SL and LL data showed that the exclusion of the SL data cubes had no detrimental effect on the fitting of the long-wavelength dust continuum. Since the IRS data are not combined with the other data sets - they will only be used in Sect. \ref{sect:gn} - and have a limited FoV, no effort was made to convolve the cubes to the PACS160 resolution.
Spitzer-IRAC imaging at 8\,$\mu$m and 24\,$\mu$m are also presented, and were convolved to the PACS160 resolution using kernels from \citet{2011PASP..123.1218A}.
\subsection{Ancillary data}
The THINGS survey\footnote{http://www.mpia-hd.mpg.de/THINGS/Overview.html} \citep{2008AJ....136.2563W} observed the HI distribution of NGC\,4736 with the VLA between March 2003 and March 2004 in the BCD configurations. The HI intensity maps (natural and robust weighting) are publicly available from the THINGS website. The beam size of the observations is, depending on the chosen weighting, 10\farcs22\,$\times$\,9\farcs07\, (natural), or 5\farcs96\,$\times$\,5\farcs55\, (robust)
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics[width=17cm]{Gas_paperfigure3.eps}}
\caption{{\it Left and middle:} Mass surface density maps of atomic (HI) and molecular (CO(1-0)) gas, at the PACS\,160$\mu$m ($\sim$12\arcsec) resolution. {\it Right:} Total gas mass surface density map, based on HI and CO maps, with a 26\% correction for helium. The radius of the circumnuclear ring and the locations of the inflow points are indicated with a dashed circle and stars, respectively. Clockwise gas flow is indicated with the white arrows.}
\label{fig:gas}
\end{figure*}
The BIMA SONG survey \citep{2003ApJS..145..259H} observed the CO(1-0) distribution in NGC\,4736 in October 1997. The total FoV of these data are $\sim$190\arcsec\, (4.3\,kpc). The data have a resolution of 6\arcsec\, and are short spacing corrected. The resulting data cube and moment maps are publicly available from NED.
NGC\,4736 was observed with GALEX in both the NUV and FUV bands between April and June, 2004. The observations are publicly available through the MAST GALEX public access site\footnote{http://galex.stsci.edu/GR6/}. The FoV is 1.2\degr\,, with a FWHM angular resolution of 4\farcs0 (NUV) and 5\farcs6 (FUV).
An H$\alpha$ map was obtained from the NASA/IPAC Extraglactic Database (NED), but was originally observed, reduced, and continuum and [NII] subtracted by \citet{2004A&A...426.1135K}. It was obtained with the Isaac Newton 2.5m telescope. The angular resolution is 1\farcs4$\times$1\farcs4 and the pixel scale is 0\farcs3.
\section{ISM distribution in the ring}
\subsection{Atomic and molecular gas}
Both the HI and CO(1-0) intensity distributions, convolved to the PACS\,160$\mu$m resolution, are shown in the first two panels of Fig. \ref{fig:gas}. The HI distribution is dominated by two arcs which are co-spatial with the circumnuclear star forming ring (whose radius is approximated with the dashed ellipse, representing an $i\sim35\degr$ inclined ring). The nucleus is devoid of atomic gas. By comparison, the CO(1-0) distribution shows emission in a more spiral arm-like pattern. (This is even more apparent at native resolution.) The spiral arms only partly overlap with the circumnuclear ring. Unlike the HI, the CO(1-0) intensity distribution further shows strong emission from the nucleus.
The `total' gas mass density is computed via these two tracers, by assuming a CO-conversion factor, X$_{CO}$, of 7.4\,$\times$\,$10^{19}$ [cm$^{-2}$\,(K km/s)$^{-1}$] \citep{2013ApJ...777....5S}. This value is lower than the Galactic X$_{CO}$ value of 2.0\,$\times$\,$10^{20}$ [cm$^{-2}$\,(K km/s)$^{-1}$] most often assumed for nearby galaxies, but \citet{2013ApJ...777....5S} have shown that X$_{CO}$ drops to this value towards the galactic nuclear region of this galaxy (averaged at a resolution of 37\farcs5). Due to the lower X$_{CO}$ factor, the molecular gas composes a smaller fraction of the total gas mass. The total gas mass map is still dominated by molecular gas, as can be expected in the molecule rich centers of galaxies, but atomic gas provides about 1/3 of the gas mass in the ring. A 26\% correction for helium and metals has been applied to the total gas map.
Ring inflow points are defined as the position where the gas spiral arm(s) cross onto the ring. In a bi-symmetric potential, e.g. with a stellar bar, the two points should be more or less symmetric about the nucleus. An indication of their position in this galaxy is given in Fig. \ref{fig:gas}. The total gas mass map shows two, possibly three, regions of high mass surface density at the radius of the ring, in the east, the northwest, and at lower significance to the southwest. The first two are near the inflow points.
\subsection{Dust}
By comparing our gas mass map with a map of the dust mass, we can determine whether or not the dust and gas in this region are well mixed, which is a necessary condition for our upcoming analysis and comparison to models of the star formation in the ring. The dust mass distribution in NGC\,4736 is derived using via the \citet{2007ApJ...657..810D} model, including Spitzer, Herschel PACS\footnote{the same observations as used in this work} and SPIRE observations of the source, to constrain the IR dust emission \citep{2012ApJ...756..138A}. We find that our mass distribution of gas shows a one-on-one morphology correspondence, confirming that dust and gas are qualitatively well mixed. Equally, when we compare our TIR continuum map (Fig. \ref{fig:FIRgas}) with the dust luminosity found by the model a one-on-one morphology correspondence can be observed. Additionally we find that our total gas mass map also matches quantitatively, i.e., with reasonable conversion values for the dust-to-gas ratio (DGR) we recover the values in the dust map.
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics[width=17cm]{SFR_paperfigure4.eps}}
\caption{Star formation rate density maps, based on H$\alpha$+24$\mu$m ({\it left}), NUV+24$\mu$m ({\it middle}), and FUV+24$\mu$m ({\it right}), at the Spitzer 24$\mu$m ($\sim$5\arcsec) resolution. The H$\alpha$+24$\mu$m derived SFRD values in this image have been multiplied by a factor two for visualization purposes. The galactic nucleus and larger radii have been masked. The radius of the circumnuclear ring and the locations of the inflow points are indicated with a dashed circle and stars, respectively. The Spitzer 24$\mu$m beam of $\sim$5\arcsec\, is indicated in the top left corner of each subpanel.}
\label{fig:SFR}
\end{figure*}
\section{Star formation distribution in the ring}\label{sect:SFR}
The star formation rate in the circumnuclear ring can be derived from either the H$\alpha$, NUV, or FUV data sets. To compensate for the stellar light absorbed by dust, all three star formation rate tracers are combined with the Spitzer 24$\mu$m image to obtain dust-corrected values.
The star formation rate is computed based on the SFR calibrations given in \citet{2011ApJ...737...67M}, \citet{2011ApJ...741..124H}, and \citet{2012ARA&A..50..531K}. The resulting SFR surface density (SFRD) maps are shown in Fig. \ref{fig:SFR}. Differences between the three maps reflect the uncertainties and physical/temporal variations among the SFR tracers. The region for which the SFRD is computed is, in all cases, limited at small radii by a circular mask at 20\arcsec\, to exclude emission from the nucleus. At the outer edge, a limit has been placed at the minimum of either 70\arcsec\, radius, or the position where the flux in the tracers falls to a 2$\sigma$ level. The maps shown in Fig. \ref{fig:SFR} are at the resolution of the 24$\mu$m observations, which is $\sim$5\arcsec\,, and have a pixel scale of 1\farcs5. SFRD maps at the PACS\,160\,$\mu$m resolution were also computed.
In all three panels, the regions of highest star formation rates are found in the southeast and northwest quadrants of the ring. These positions are near/at the ring inflow points, as expected. At the other two quadrants of the ring, the SFRD is lower (by factors of three to six). This could indicate a secondary star formation event of lower-level intensity, or the continued emission from an ageing burst. The SFRD contrast around the ring is strongest in the H$\alpha$+24$\mu$m map. The SFRD values found for H$\alpha$+24$\mu$m were checked against previous results by \citet{2004ApJ...605..183W}. \citet{2004ApJ...605..183W} found a SFRD for the circumnuclear ring, based on H$\alpha$ fluxes alone, of 0.04\,M$_{\sun}$/yr/kpc$^2$. We match this value with our H$\alpha$ observations.
Comparing with the gas maps shown in Fig. \ref{fig:gas}, the SFR distribution is most similar to that of the atomic gas. This is surprising, given the fact that stars form from molecular not atomic gas. We speculate here that the young stars in the ring of NGC\ 4736 are energetic enough to dissociate large quantities of molecular gas in their immediate vicinity (without ionizing that gas as well), and will investigate this matter in an upcoming paper (Van der Laan, in prep.)
The star formation regions near the inflow points each cover about a fifth of the ring. The orbital time for a star at this radius (r=0.9\,kpc) is about 25\,Myr, given a 220\,km/s rotational velocity (value derived from the BIMA SONG first moment map). Therefore the duration of the starburst cannot be more than 5\,Myr. This is also a reasonable time span to explain the difference between the H$\alpha$+24$\mu$m, and NUV+24$\mu$m and FUV+24$\mu$m maps. When we `subtract' the H$\alpha$ distribution from either the NUV or FUV distribution (not shown), what remains is a more uniform distribution of flux, expected if the UV emission drops more slowly with time ($\sim$5.5\,Myr for H$\alpha$, $\sim$11\,Myr for NUV/FUV, as based on Starburst99 stellar burst synthesis models). The UV tracers are still tracing the previous `event' when they have moved through half the ring.
With these observations it is not possible to directly distinguish an age-gradient of star formation in the hot spots seen at the inflow points. Inspection of the observations shows that the hot spots are strongest in the 24\,$\mu$m observations and nearly unseen in the FUV and H$\alpha$, implying they are (strongly) extincted regions. However, an equivalent width map of the H$\alpha$ emission, given that a suitable continuum image is found, could potentially show age-related variations between the hot spots.
Averaged over the ring, the SFRD values correspond to a star formation rate density of $\sim$0.065\,M$_{\sun}$/yr/kpc$^2$. Within individual hotspots, the SFRD reaches more than 0.25\,M$_{\sun}$/yr/kpc$^2$. High SFRD are a common characteristic of circumnuclear star forming rings \citep{2004ARA&A..42..603K,2012ARA&A..50..531K}. Integrating over the full area of the ring leads to SFR values of 0.16, 0.27, and 0.25\,M$_{\sun}$/yr, for H$\alpha$+24$\mu$m, NUV+24$\mu$m, and FUV+24$\mu$m, respectively. These values are lower than previously derived SFRs for the whole galaxy of 0.38\,M$_{\sun}$/yr \citep{2010ApJ...714.1256C}, 0.43\,M$_{\sun}$/yr \citep{2006PhDT........12L}, and 0.70\,M$_{\sun}$/yr \citep{2011ApJ...738...89S}, but show the major role the circumnuclear ring plays in the overall star formation in NGC\ 4736.
\section{Radiation field in the ring}\label{sect:gn}
In order to investigate the effect the young stars have on the local ISM, it is important to measure the emission from the dense neutral regions (the PDRs) close to the forming stars.
As a first step, it is important to estimate, and then remove, the contribution of [CII] from the diffuse, ionized ISM. Neutral Carbon has an ionization potential (11.3eV) which is slightly lower than that of Hydrogen (13.6eV). Consequently, it will be present in both the neutral and ionized ISM. The contribution of [CII] from the diffuse ionized ISM can be determined from an observed [CII]/[NII]205\,$\mu$m ratio, since N$^+$ emission arrises exclusively from the diffuse ionized medium \citep{1994ApJ...434..587B}, and the critical density of [CII] and [NII]\,205$\mu$m are nearly the same. The observed [CII]/[NII] ratio is nearly constant ($\sim$4) in the ionized ISM over a large range of electron density, n$_e$. The difference between the predicted and measured [CII]/[NII]205$\mu$m ratio can then be used to estimate the ionized and neutral fractions of [CII] emission.
\setlength{\unitlength}{1cm}
\begin{figure}
\begin{tabular}{c}
\put(0,0){\includegraphics[width=8.6cm]{Figure_gn_ring90.ps}}
\put(1.5,3.2){\includegraphics[width=2.6cm]{Figure_gn_ring90_sub.ps}}
\end{tabular}
\caption{Plot of [CII]/[OI] vs. ([CII]+[OI])/TIR to determine the ambient UV radiation field incident on the PDR regions. The [CII] emission has been corrected for a 10\% contribution from the diffuse ionized ISM. The underlying grid is the \citet{1999ApJ...527..795K} PDR model for different densities and stellar radiation field strengths. The inset sub-panel is a zoom-in of the region of interest. Data points in the east/west side of the ring are indicated in blue/red.}
\label{fig:gnplot}
\end{figure}
\setlength{\unitlength}{1cm}
\begin{figure}
\begin{tabular}{c}
\put(0,0){\includegraphics[width=8.6cm]{Figure_gn_ring90fpdr.ps}}
\end{tabular}
\caption{Same as Fig. \ref{fig:gnplot}, except with a further assumption of 20\% TIR, and 60\% [CII] contribution from PDRs illuminated by a high radiation field intensity, with the rest originating from more diffuse neutral ISM components.}
\label{fig:gnplot_fpdr}
\end{figure}
Since the [NII]205$\mu$m line is not significantly detected in the NGC\ 4736 ring, we follow an analysis similar to the one discussed in \citet{2012ApJ...747...81C} and use the [CII]/[NII]122$\mu$m ratio to determine the ionized [CII] emission. The [CII]/[NII]122$\mu$m ratio is dependent on electron density in the range 1-100 cm$^{-3}$. While the [SIII]18.7/33.5$\mu$m ratio is available from our IRS maps, the ratio [SIII]18.7/33.5 is sensitive to much higher electron densities in the range $10^{2-5}$\,cm$^{-3}$, and the [SIII] ratio (0.56) is consistent with gas in the low density limit.
The [NII]122/205 ratio is sensitive to electron densities of 0.1-100 cm$^{-3}$. Since the [NII] 205$\mu$m line is not detected in the ring, the observed [NII]122$\mu$m signal-to-noise ratio in the ring ($5-10$) can be used as a proxy for the upper limit of the [NII]122/205 ratio. From \citet{2012ApJ...751..144B} we know that [NII]122$\mu$m/[NII]205$\mu$m increases with increasing electron density, with the ratio reaching 10 at n$_{e}$ = 100\,cm$^{-3}$. Thus, in the ring we are dealing with an n$_{e}$ lower than 100\,cm$^{-3}$, consistent with our Spitzer/IRS measurements. At an n$_{e}$ of 100\,cm$^{-3}$ the [CII]/[NII]122$\mu$m ratio from the diffuse ionized component of the ISM would be 1.24 (this is an upper limit). All additional [CII] emission should be from the neutral component. In the region of the ring for which we have [NII]122$\mu$m emission maps, we find [CII]/[NII]122$\mu$m ratios of 8-16, with the value increasing away from the highest intensity regions. These values correspond to 85 to 92\% contributions of the neutral phase to the [CII] emission. At lower electron density the [CII] fraction in the neutral phase rises. We therefore assume that 90\% of the measured [CII] emission in the ring is from the neutral phase.
In Fig.\,\ref{fig:gnplot} the (diffuse ionized emission corrected) [CII]/[OI] and ([CII]+[OI])/TIR ratios are overlaid on a PDR model from \citet[][hereafter K99]{1999ApJ...527..795K}. As described in Sect. \ref{sect:data}, the [CII], [OI], and TIR fluxes were extracted in square apertures of 12\arcsec\, by 12\arcsec\, size, covering the circumnuclear ring. No difference is seen between the two (Fig. \ref{fig:gnplot}, red/blue points) sides of the ring. Based on these [CII]/[OI] and ([CII]+[OI])/TIR ratios the radiation field G$_0$, a dimensionless variable in multiples of the radiation field in the Solar neighborhood, in the PDR regions of the circumnuclear ring is found to have a uniform value of $\sim$200, and the gas density, n$_H$, in the PDR regions to be $\sim$\,10$^{3}$\,cm$^{-1}$.
We can compare our G$_0$ derived above from the FIR emission lines with the average $<$U$>$ obtained from modeling the dust emission. $<$U$>$ is a measure of the interstellar radiation field in units of the solar neighborhood \citep{2007ApJ...657..810D}. For NGC\, 4736, $<$U$>$ is roughly 10 (Aniano et al., priv. comm.), about a factor of 20 lower than the G$_0$ inferred above. This large mis-match between the radiation field intensity derived from the gas and dust has been noted in other studies as well \citep{2012ApJ...747...81C}, and can be understood by noting that the PDR fraction, the fraction of the dust illuminated by the high intensity radiation field, can be much less than unity. Aniano et al. find a f$_{PDR}$ (the fraction of dust mass illuminated by G$_{0}$ greater than $\sim$110) in the ring of order 20\%. If we apply this fraction to the TIR, and assume only 60\% of the neutral [CII] emission is from dense PDRs illuminated by the high radiation field intensity, we can reconcile the two estimates (see Fig. \ref{fig:gnplot_fpdr}).
\section{Heating-cooling balance}
In the `pearls-on-a-string' model, the heating in the ring as a function of azimuth is greatest close to the inflow points, followed by a slow decay as the stars age and the ring rotates. To see if the observed cooling in the ring follows this pattern, we compare the dust continuum flux and the far-IR line emission, against the expected available stellar heating.
\subsection{Energy input into the ISM}
By making several reasonable assumptions, the available energy for dust and gas heating from the newly formed stars in the ring can be derived from the measured SFRD. An instantaneous starburst was modeled using Starburst99 \citep{1999ApJS..123....3L,2010ApJS..189..309L}, assuming a standard Kroupa IMF and solar metallicity. The output synthetic line spectrum was integrated between 912\,\AA{} and 1120\,\AA{} (6-13.6eV), to obtain the photo-electric heating energy as a function of time. The resulting available heating power per M$_{\odot}$ is shown in Fig. \ref{fig:LW_curve}. As can be seen, the power in the 6\,eV to 13.6\,eV range of the integrated stellar population spectrum is strongly dependent on time. The flux plateaus in the first 3\,Myr, but within 11\,Myr, it has dropped by an order of magnitude.
\begin{figure}
\includegraphics[width=8.6cm]{Figure_LWflux_burst.ps}
\caption{Integrated FUV flux between 6-13.6\,eV in the case of an instantaneous starburst, rescaled to 1M$_{\odot}$. Within 4\,Myr the energy has halved, and within 11\,Myr the energy has dropped by an order of magnitude. (Based on Starburst99 v6.0.3. models)}
\label{fig:LW_curve}
\end{figure}
Instead of a single instantaneous starburst, physically the star formation will occur over a longer time interval. Like a Ferris wheel, the ring can be thought of as having discrete volume elements (`cabins'), scooping up gas over a fixed time interval as it enters onto the ring. In this model, each volume element will have a star formation event lasting a fixed fraction of the rotation time, with the mean age of the event depending on the distance from the inflow point. By tracing back the observed star formation under this premise, we can recreate the heating flux throughout the ring.
As a reminder: the average SFRD measured in the ring was 0.065\,M$_{\odot}$/yr/kpc$^2$. To be able to compare the available heating to the observed TIR emission later on, we multiply by an average 24$\mu$m/UV flux ratio of 0.33, based on our observations, as a proxy of how much FUV flux is heating the dust, rather than immediately escaping the galaxy. This ratio is close to the measured dust/stellar flux ratio of 0.25 by \citet{2011ApJ...738...89S}.
\begin{figure*}
\begin{tabular}{c c}
\includegraphics[width=8.6cm]{Figure_LWbalans_2ndburst_deg.ps} &
\includegraphics[width=8.6cm]{Figure_LWbalans_gas_2ndburst_deg.ps} \\
\end{tabular}
\caption{{\it Left:} Available FUV heating flux (black saw-tooth line) as a function of azimuth, derived from an proposed star formation history. The proposed star formation history is indicated at the bottom of the panel (black box function). The grey lines (underlying the black) are an alternative example of a star formation history and FUV heating flux. The observed FIR (dust) flux is shown in blue/red (east/west side of the ring). Details in the text. {\it Right:} Similar, except now for the [CII]+[OI] (gas) emission lines. The FoV of the Herschel spectra does not cover the full ring, thus for several apertures the gas information is lacking.}
\label{fig:heatcool}
\end{figure*}
\subsection{Comparison to dust}
A SFR history (for one full 25\,Myr orbit) is postulated where there are two starbursts of 5\,Myr duration, 12.5\,Myrs apart, with a burst SFR intensity of 0.01\,M$_{\odot}$/yr (thus a total of 5$\times$10$^4$M$_{\odot}$ stellar mass per starburst is created). The heating curve is modeled by convolving the instantaneous starburst, shown in Fig. \ref{fig:LW_curve}, with this proposed (box-)function of SFR history to obtain the instantaneous starburst heating as a function of time/position throughout the ring (left panel Fig. \ref{fig:heatcool}, grey saw-tooth curve). This heating curve is compared with the TIR continuum sampled throughout the ring (colored points). We may assume an instantaneous response from the dust to the heating, since the timescales here are of order $\sim$Myr, while dust responds to heating on order 10$^4$yr, regardless of dust geometry (B. Draine, priv. comm.). In Fig. \ref{fig:heatcool} the correspondence between the {\it measured} TIR continuum and the {\it modeled} heating is very good. However, it is clear that the heating power falls too steeply around the ring, and that the second peak is significantly lower than the first. A much better match (black curve) can be made by shortening both initial bursts, lowering of the second burst to 0.008\,M$_{\odot}$/yr, and adding two short bursts of star formation at $\sim$8.75-10Myr and 20-21.25Myr. The inclusion of the latter is a divergence from the `pearls-on-a-string' scenario.
\subsection{Comparison to PDR gas}\label{sect:gascool}
To compare the response of the PDR gas to the modeled heating we need an observational measure of the photoelectric heating efficiency. This is the ratio of photon energy going into gas heating to the total available energy. This ratio, typically taken as ([CII] + [OI])/TIR, is equal to the ratio of gas cooling line flux to TIR continuum flux. In Fig. \ref{fig:heatciioi} this gas to dust ratio is plotted against, the ratio $\langle\nu\,S_{\nu}\rangle_{70}/\langle\nu\,S_{\nu}\rangle_{100}$, which is used as a proxy for the dust temperature. From Fig. \ref{fig:heatciioi} it can be seen that the photoelectric heating efficiency shows very little spread over the circumnuclear ring, with values of 0.004-0.006, about 0.5\%.
\begin{figure}
\includegraphics[width=8.6cm]{Figure_heat_ciioi_wfneut.ps}
\caption{Gas heating efficiency proxy, ([CII] + [OI])/TIR, as a function of the dust temperature (blue (east) and red (west) dots). For comparison, the results from \citet{2012ApJ...751..144B} in the circumnuclear star forming ring of NGC\,1097 (`x'), and \citet{2012ApJ...747...81C} in sub-kpc regions in NGC\,1097 and NGC\,4559 (dashed line) are indicated. The [CII] has been corrected for its 10\% diffuse ionized component.}
\label{fig:heatciioi}
\end{figure}
In the right panel of Fig. \ref{fig:heatcool}, this fixed 0.5\% scaling has been applied to the [CII]+[OI] flux in each region. It is immediately clear that the spread in the emission line flux is due to the FUV heating variation. The two cooling lines again follow the modeled FUV heating quite closely, although we lack the data points away from the inflow points to confirm the secondary star forming events, due to limited Herschel spectral coverage.
The photoelectric heating efficiencies are comparable to the values seen for other kpc-sized star forming regions, as for example in \citet{2012ApJ...747...81C} or \citet{2013ApJ...776...65P}. These studies find values of $\sim$0.5\% at lower dust temperatures, with a significant drop of heating efficiency at higher dust temperatures. However, \citet{2012A&A...548A..91L} and \citet{2014ApJ...787...16P} do not find a reduction in the gas heating efficiency at higher dust temperatures in a star forming region of the LMC, and the disk of Centaurus A, respectively. The NGC\,4736 ring does, however, have a higher photoelectric heating efficiency, at only marginally higher dust temperature, than starburst ring in NGC\,1097 ring \citep{2012ApJ...751..144B}.
\section{Implications for ring star formation theory}
We started this investigation to study if, following the `pearls-on-a-string' paradigm for circumnuclear star forming rings, all star formation takes place at the inflow points of the ring. In this scenario, as the system rotates, the subsequent star forming events should create a clear azimuthal age gradient \citep{Boker2008}.
Our simple modeling of the PDR dust and gas heating as function of azimuth requires secondary star formation events to be consistent with the observed cooling flux. While the exact strength, timing, and duration of these secondary events in the NGC\,4736 ring is model dependent, it is clear that their absence leads to a too large contrast in heating/cooling flux around the ring. Since the FUV heating power available from young stars diminishes by an order of magnitude in 11\,Myr, it seems unlikely that the drop be compensated for by stellar 'after-glow'.
Are these `secondary' star forming events truly random, taking place from left-over gas from the `primary' event? Or, are both events part of the `popcorn' scenario \citep{Elmegreen1994}, where star formation occurs anywhere in the ring? Given the relative strength of the events at both inflow points we consider this unlikely. This seems consistent with the fact that a number of galaxies show strong star formation events at the inflow points of their circumnuclear rings \citep{2001MNRAS.323..663R,2005ApJ...633L..25A,Boker2008,2008ApJS..174..337M,Tessel2,Tessel3}.
A combination of `pearls-on-a-string' and `popcorn' seems like a reasonable scenario to explain the star formation in the ring of NGC\ 4736. We know star formation is not a 100\% efficient process. Star formation efficiencies (SFE) are usually estimated to be of order a few percent. Although there may be an indication that the SFE in the center of NGC\,4736 is actually higher (10\%, K. Sandstrom, priv. comm.), even then a significant portion of gas remains. Therefore, we can ask if the remaining gas has enough chance to cool and is dense enough to collapse again.
The total gas mass surface density in the ring remains $\sim$50-80\,M$_{\odot}$/pc$^2$ away from the inflow points (Fig. \ref{fig:gas}). This is higher than the $\sim$8\,M$_{\odot}$/pc$^2$ star formation threshold noticed in \citet{2008AJ....136.2846B}. Additionally, while the temporal distance between the inflow points and these modeled secondary bursts is short (a few Myr), it is comparable to free fall times of giant molecular clouds.
Can we expect this hybrid star formation scenario in all circumnuclear rings? The life times of circumnuclear rings are of order 1-2\,Gyr. Inflow rates of gas onto the ring of 2\,M$_{\odot}$/yr are a reasonable upper limit \citep{Tessel,Tessel2}, although values of 0.1-0.25\,M$_{\odot}$/yr are more likely. So, even at extreme SFEs and low gas inflow rates, the gas mass in a ring will see an increase of about 10$^8$\,M$_{\odot}$ (0.1\,M$_{\odot}$/yr$\times$1\,Gyr) over the life time of the ring. Given the observed range of radii and widths found for circumnuclear rings (40-1000\,pc radii, $\sim$r/2 widths, \citet{Comeron2010}), this translates easily to gas mass surface densities similar to those found in NGC\ 4736 (10-10$^2$\,M$_{\odot}$/pc$^2$). Thus, the prevalence of `popcorn' star formation events should increase with the lifetime of the circumnuclear ring, even if starbursts, `pearls-on-a-string', continue to be at the inflow points.
\section{Summary}
New Herschel FIR observations of the circumnuclear ring in NGC\,4736 have been combined with archival observations of the star formation, and atomic and molecular gas distributions. This combination of data sets is used to characterize the mode of star formation in the ring. Due to the circular orbits of the ring itself, gas and dust will only experience large-scale shocks and compression at the inflow points, the locations where material transitions onto the ring. Star formation events, therefore, should be uniquely found close to the inflow points. We have investigated the heating and cooling of the interstellar medium in the ring of NGC\,4736, under this assumption. The stellar FUV flux available for heating was derived from measurements of the star formation rate, geometry-based estimates of time scales, and standard assumptions on the initial mass function (IMF). The FIR continuum and gas emission line fluxes were measured from Herschel PACS and archival Spitzer observations, and used to estimate the heating and cooling of the dust and gas around the ring.
We find that, while heating from a significant burst of star formation at the inflow points can explain the gas and dust cooling there, an additional star formation component is required at other positions in the ring. This additional component is smaller than the star formation events at the inflow points, and most likely arises from the general increase of gas density in the ring over its lifetime. Therefore, in the NGC\ 4736 ring, it is a combination of the `popcorn' and `pearls-on-a-string` models that best explains the data.
\begin{acknowledgements}
This project was started during a '11/'12 IPAC visiting graduate fellowship awarded to TvdL. TvdL would like to thank IPAC, especially P. Appleton and L. Armus, for hosting her.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sec:introduction}
In this paper, we explore the relation between positive definite
functions on rigid C$^*$-tensor categories and their Drinfeld
centers. Our inspiration for seeking such a relation comes from recent
developments in the so called central approximation properties of
discrete quantum groups. In particular, we give a categorification of
the correspondence between the completely positive central
functions on discrete quantum groups and the positive linear functionals on the character
algebra constructed from the spherical representations of the Drinfeld
double, as observed in~\cite{MR3238527}.
Such a categorification has been already obtained in the recent work of Popa and
Vaes~\cite{MR3406647}, and similarly to their work our main motivation is to understand approximation properties of monoidal categories. In this respect the paper can be seen as a natural continuation of our previous
work~\cite{arXiv:1405.6572}, where we studied amenability of monoidal
categories.
In fact, a connection between approximation properties of monoidal
categories and their Drinfeld centers has already appeared in
subfactor theory, although in a disguised form. Ocneanu introduced the
notion of asymptotic inclusion based on iterated basic extensions and
studied the associated $3$-dimensional topological quantum field
theory for finite depth
subfactors~\citelist{\cite{MR996454}}. Subsequently its relation to
the Drinfeld center was clarified through the work of
Evans--Kawahigashi~\cite{MR1316301}, Longo--Rehren~\cite{MR1332979},
Izumi~\cite{MR1782145}, and
M\"{u}ger~\citelist{\cite{MR1966524}\cite{MR1966525}}, to name a
few. In a related direction, Popa introduced the notion of symmetric
enveloping algebra $M \boxtimes_{e_N} M^\mathrm{op}$ associated with a
subfactor $N \subset M$ as a byproduct of his celebrated
classification program~\cite{MR1302385}. This notion specializes to
asymptotic inclusion in the finite depth case but has a better
universality property for infinite depth and non-irreducible
subfactors.
What arises from Popa's work is the principle that
approximation properties of the combinatorial data encoding the
original subfactor correspond to approximation properties of the
SE-inclusion $M\bar \otimes M^\mathrm{op} \subset M \boxtimes_{e_N} M^\mathrm{op}$
formulated in the language of Hilbert bimodules (correspondences). In
particular, based on the general theory of correspondences and
rigidity developed in~\cite{popa-corr-preprint}, Popa introduced the
notion of property (T) for subfactors~\cite{MR1729488} as an
antithesis of amenability, which played a central role in the
classification. Since any finitely generated rigid C$^*$-tensor category can
be realized as a part of the standard invariant~\cite{MR1334479}, it
is natural to try to borrow from this theory to formulate various
notions for C$^*$-tensor categories. Popa and Vaes~\cite{MR3406647} achieved this by axiomatizing the notions
of completely positive multipliers and completely bounded multipliers on rigid
C$^*$-tensor categories, and by relating them to properties of
asymptotic inclusions. They defined C$^*$-completions of the fusion algebras by considering what they called admissible representations, which are characterized by the property that their matrix coefficients are completely positive multipliers, and using these completions defined analogues of various approximation properties for such categories.
Aiming for a more direct connection with the Drinfeld center, we start
our work by considering unitary half-braidings on ind-objects of a
rigid C$^*$-tensor category~$\mathcal{C}$. For every such half-braiding, we
construct a $*$-representation of the fusion algebra. This
construction is an analogue of the restriction of spherical
representations of the Drinfeld double of a quantum group to the
subspace of spherical vectors. As an example if we consider the category
of $\Gamma$-graded Hilbert spaces for a discrete group $\Gamma$, then
we recover all unitary representations of $\Gamma$. In this scheme
there is a distinguished half-braiding, which we call the regular
half-braiding, giving rise to the regular representation of the fusion
algebra. Morally, it corresponds to the algebra object representing
the forgetful functor on the Drinfeld center which appeared in the
work of
Brugui\'{e}res--Virelizier~\citelist{\cite{MR2355605}\cite{MR3079759}}
and Brugui\'{e}res--Natale~\cite{MR2863377} in the framework of fusion
categories. The representations defined by the half-braidings lead to
a completion $C^*(\mathcal{C})$ of the fusion algebra to a C$^*$-algebra. As in~\cite{MR3406647}, the algebra $C^*(\mathcal{C})$ can then be
used to formulate analogues of various approximation properties of
groups for C$^*$-tensor categories. In particular, by considering the
isolation property of the representation defined by the unit object,
we get a natural definition of property (T).
In order to illustrate the general theory we consider categories $\mathcal{C}$
of Hilbert bimodules over a II$_1$-factor~$M$. Expanding on Izumi's
work~\cite{MR1782145} on the Longo--Rehren inclusion, we show that in
this case the Drinfeld center is monoidally equivalent to a category
of Hilbert $ B$-bimodules, where $A\subset B$ is the Longo--Rehren
inclusion associated with $\mathcal{C}$. As a particular case, this gives an equivalence between the category of representations of $C^*(\mathcal{C})$ and the category of Hilbert $B$-bimodules generated by $A$-central vectors, which has been already established by Popa and Vaes in their approach for categories arising from extremal finite index subfactors.
This equivalence can be used to connect the categorical notion of property (T)
to Popa's property (T) of the Longo--Rehren inclusion. In the case when $\mathcal{C}$ is the category associated with a finite index
extremal subfactor $N\subset M$, this means that property (T) for~$\mathcal{C}$ is equivalent to Popa's property~(T) for $N\subset M$. As has been observed in~\cite{MR3406647}, combined
with the results of~\cite{arXiv:1410.6238} this, in turn, can be used to
construct subfactors with property~(T) that do not come from discrete
groups.
\smallskip
The results described above had been obtained when we received preprint~\cite{MR3406647} by Popa and Vaes. A natural task was then to compare the two approaches.
It turned out that they are equivalent, and even the classes of representations of the fusion algebras are the same, so that a representation is admissible in the sense of Popa and Vaes if and only if it is defined by a unitary half-braiding. In light of this, our definition of property (T)
is a reformulation of theirs. The approaches naturally complement each other and
have their own advantages. For example, in our setting it is almost
immediate that for representation categories of compact quantum groups
the completion of the fusion algebra coincides with the one obtained
by embedding it in the Drinfeld double~\cite{MR3238527}. On the other
hand, for general categories it is more difficult to see in our
approach that the regular representation of the fusion algebra extends
to its C$^*$-algebra completion. We also mention that soon after both papers were posted, yet another alternative approach to representation theory of monoidal categories was suggested by Ghosh and C.~Jones~\cite{MR3447719}.
\smallskip
In the last section we study weakly monoidally Morita equivalent categories. This notion was
introduced by M\"{u}ger~\cite{MR1966524}. A prototypical example is the categories of Hilbert bimodules
over a factor and its finite index subfactor. As was pointed out by M\"{u}ger, a result of Schauenburg~\cite{MR1822847} implies that weakly Morita equivalent categories have monoidally equivalent Drinfeld
centers. This does \emph{not} imply that the corresponding fusion algebras are Morita equivalent, and the precise relation between these algebras will be discussed elsewhere. What we prove in the present paper, is that the property of weak containment of the unit object is preserved under Schauenburg's equivalence, which allows us to compare approximation properties of the original categories. In particular, we show that property (T) is invariant under weak monoidal Morita equivalence.
\medskip\noindent
{\bf Acknowledgement.} We are grateful to Sorin Popa and Stefaan Vaes for fruitful correspondence and in particular for informing us about their work and for their interest in ours.
\bigskip
\section{Preliminaries}
\label{sec:preliminaries}
\subsection{C\texorpdfstring{$^*$}{*}-tensor categories}
\label{sec:c-tensor-categories}
The main object of our study is rigid C$^*$-tensor categories, and in general we keep the conventions of~\cite{arXiv:1405.6572}, see also~\cite{MR3204665} for the proofs. For the convenience of the reader let us summarize the basic terminology.
A \emph{C$^*$-category} $\mathcal{C}$ is a linear category over the complex number field $\mathbb{C}$, endowed with Banach space norms on the morphism sets $\mathcal{C}(X, Y)$ and a conjugate linear anti-multiplicative involution $\mathcal{C}(X, Y) \to \mathcal{C}(Y, X)$, $T \mapsto T^*$ satisfying the C$^*$-identity $\norm{T^* T} = \norm{T}^2 = \norm{T T^*}$. We always assume that a C$^*$-category is closed under taking subobjects, so that any projection in the C$^*$-algebra $\mathcal{C}(X) = \mathcal{C}(X, X)$ corresponds to a subobject of $X$. We also assume that $\mathcal{C}$ is closed under finite direct sums. A C$^*$-category is called \emph{semisimple} if its morphism sets are finite dimensional. In such categories one can always take a decomposition of an object~$X$ into a direct sum of simple objects using minimal projections in the finite-dimensional C$^*$-algebra $\mathcal{C}(X)$. A \emph{unitary functor}, or a \emph{C$^*$-functor}, between C$^*$-categories is a linear functor $F$ compatible with involutions: $F(T^*) = F(T)^*$.
A \emph{C$^*$-tensor category} is a C$^*$-category $\mathcal{C}$ endowed with a bifunctor $\otimes \colon \mathcal{C} \times \mathcal{C} \to \mathcal{C}$, a distinguished object~$\mathds{1}$, and natural \emph{unitary} isomorphisms
\begin{align*}
\mathds{1} \otimes X &\to X \leftarrow X \otimes \mathds{1}, &
\Phi\colon (X \otimes Y) \otimes Z &\to X \otimes (Y \otimes Z)
\end{align*}
satisfying the standard set of axioms for monoidal categories. In this paper the unit $\mathds{1}$ is always assumed to be simple, namely $\mathcal{C}(\mathds{1}) \cong \mathbb{C}$. A \emph{unitary tensor functor}, or a \emph{C$^*$-tensor functor}, between C$^*$-tensor categories is a C$^*$-functor $F$ together with a unitary isomorphism $F_0\colon \mathds{1} \to F(\mathds{1})$ and natural unitary isomorphisms $F_2\colon F(X)\otimes F(Y) \to F(X \otimes Y)$ satisfying the standard compatibility conditions. If there is no fear of confusion we use $F$ instead of $(F, F_0, F_2)$ to denote C$^*$-tensor functors. The natural transformation of C$^*$-tensor functors $F \to G$ is defined in the same way as in the case of monoidal functors, but with the additional requirement that the structure morphisms are all unitary. If two C$^*$-tensor categories $\mathcal{C}$, $\mathcal{C}'$ are related by C$^*$-tensor functors $F\colon \mathcal{C} \to \mathcal{C}'$ and $G\colon \mathcal{C}' \to \mathcal{C}$ such that there exist natural isomorphisms of C$^*$-tensor functors $\Id_\mathcal{C} \to G F$ and $\Id_{\mathcal{C}'} \to F G$, we say that $\mathcal{C}$ and $\mathcal{C}'$ are \emph{unitarily monoidally equivalent}.
A C$^*$-tensor category $\mathcal{C}$ is said to be \emph{rigid} if every object $X$ in $\mathcal{C}$ has a dual. Assuming for simplicity that~$\mathcal{C}$ is strict, this means that there is an object $\bar{X}$ in $\mathcal{C}$ and morphisms $R \in \mathcal{C}(\mathds{1}, \bar{X} \otimes X)$ and $\bar{R} \in \mathcal{C}(\mathds{1}, X \otimes \bar{X})$ satisfying the \emph{conjugate equations}
\begin{align*}
(\iota_{\bar{X}} \otimes \bar{R}^*) (R \otimes \iota_{\bar{X}}) &= \iota_{\bar{X}},&
(\iota_X \otimes R^*) (\bar{R} \otimes \iota_X) &= \iota_X.&
\end{align*}
Rigid C$^*$-tensor categories (with simple units) are always semisimple.
A rigid C$^*$-tensor category has a good notion of dimension, defined by
$$
d^\mathcal{C}(X) = \min_{(R, \bar{R})} \norm{R} \norm{\bar{R}},
$$
where $(R, \bar{R})$ runs over the solutions of the conjugate equations for $X$. If there is no fear of confusion we simply write $d(X)$ instead of $d^\mathcal{C}(X)$. A solution $(R, \bar{R})$ satisfying $\norm{R} = \norm{\bar{R}} = d(X)^{{1/2}}$ is called \emph{standard}, and such solutions are unique up to transformations of the form $(R,\bar R)\mapsto((T \otimes \iota)R, (\iota \otimes T)\bar{R})$ for unitary morphisms~$T$. We often denote a choice of standard solution for the conjugate equations for $X$ as $(R_X, \bar{R}_X)$. As a convenient shorthand, when $(X_i)_{i \in I}$ is a parametrized family of objects in $\mathcal{C}$, we write $(R_i, \bar{R}_i)$ instead of $(R_{X_i}, \bar{R}_{X_i})$. Similarly, for many other constructions we use index $i$ instead of $X_i$, so for example we write $d_i$ for $d(X_i)$. If the family is self-dual, we also write $\bar i$ for the index corresponding to the dual of $X_i$.
There are several constructions based on standard solutions. For example, if $X, Y \in \mathcal{C}$, then $(R_X + R_Y, \bar{R}_X + \bar{R}_Y)$ is a standard solution for $X \oplus Y$. Similarly, $((\iota_{\bar{Y}} \otimes R_X \otimes \iota_Y) R_Y, (\iota_{\bar{Y}} \otimes \bar{R}_Y \otimes \iota_Y) \bar{R}_X)$ is a standard solution for $X \otimes Y$. The \emph{categorical trace} is the trace on $\mathcal{C}(X)$ given by
$$
\Tr_X(T) = R^*_X(\iota \otimes T)R_X = \bar{R}^*_X(T \otimes \iota)\bar{R}_X,
$$
which is independent of the choice of standard solutions $(R_X, \bar{R}_X)$. The second equality above characterizes the standard solutions. The normalized categorical traces are defined by $\tr_X=d(X)^{-1}\Tr_X$.
More generally, we can define \emph{partial categorical traces}
$$
\Tr_X\otimes\iota\colon \mathcal{C}(X\otimes Y,X\otimes Z)\to\mathcal{C}(Y,Z)\ \ \text{by}\ \ (\Tr_X\otimes\iota)(T)=(R^*_X\otimes\iota_Z)(\iota_{\bar X} \otimes T)(R_X\otimes\iota_Y),
$$
and similarly define $\iota\otimes\Tr_X$.
For $X, Y \in \mathcal{C}$ and a choice of standard solutions $(R_X, \bar{R}_X)$ and $(R_Y, \bar{R}_Y)$, we can define a linear anti-multiplicative map $\mathcal{C}(X, Y) \to \mathcal{C}(\bar{Y}, \bar{X})$, denoted by $T \mapsto T^\vee$, which is characterized by $(T \otimes \iota) \bar{R}_X = (\iota \otimes T^\vee) \bar{R}_Y$. This map can be also characterized by $(\iota \otimes T) R_X = (T^\vee \otimes \iota) R_Y$ and satisfies $T^{\vee *} = T^{* \vee}$ for the choice of standard solutions $(\bar{R}_X, R_X)$, $(\bar{R}_Y, R_Y)$ for $\bar{X}$ and $\bar{Y}$.
\subsection{Ind-objects in \texorpdfstring{C$^*$}{C*}-categories}
\label{subsec:ind-obj-C*-cat}
Let $\mathcal{C}$ be a semisimple C$^*$-category. By an \emph{ind-object} of $\mathcal{C}$ we will mean an inductive system $\{u_{ji}\colon X_i\to X_j\}_{i\prec j}$ in~$\mathcal{C}$, where $u_{ji}$ are isometries. We define a morphism between two such objects $\{u_{ji}\colon X_i\to X_j\}_{i\prec j}$ and $\{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$ as a collection $T$ of morphisms $T_{ki}\colon X_i\to Y_k$ in~$\mathcal{C}$ such that
$$
v^*_{lk}T_{li}=T_{ki}\ \ \text{if}\ k\prec l,\ \
T_{kj}u_{ji}=T_{ki}\ \ \text{if}\ i\prec j,
\ \ \text{and}\ \ \|T\|:=\sup_{k,i}\|T_{ki}\|<\infty.
$$
For ind-objects $X_* = \{u_{ji}\colon X_i\to X_j\}_{i\prec j}$, $Y_* = \{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$, and $Z_* = \{w_{nm}\colon Z_m\to Z_n\}_{m\prec n}$, the composition of morphisms $T\colon X_* \to Y_*$ and $S\colon Y_* \to Z_*$ is defined by
$$
(ST)_{ni}=\lim_k S_{nk}T_{ki}.
$$
In order to see that this is well-defined we need the following.
\begin{lemma}
For any morphism $T\colon\{u_{ji}\colon X_i\to X_j\}_{i\prec j}\to\{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$, index $k$ and $\varepsilon>0$ there exists an index $i_0$ such that for all $j\succ i\succ i_0$ we have
$$
\|T_{kj}-T_{ki}u_{ji}^*\|<\varepsilon.
$$
\end{lemma}
\begin{proof} Since $T_{kj}u_{ji}=T_{ki}$, the net $\{T_{ki}T^*_{ki}\}_i$ in the finite dimensional C$^*$-algebra $\End_\mathcal{C}(Y_k)$ is increasing. Since it is also bounded, it converges in norm. Hence we can find $i_0$ such that for all $j\succ i\succ i_0$ we have
$$
\|T_{kj}T^*_{kj}-T_{ki}T^*_{ki}\|<\varepsilon^2.
$$
It remains to observe that
$$
\|T_{kj}-T_{ki}u_{ji}^*\|^2=\|T_{kj}T^*_{kj}-T_{kj}u_{ji}T_{ki}^*-T_{ki}u_{ji}^*T_{kj}^*+T_{ki}u_{ji}^*u_{ji}T_{ki}^*\|= \|T_{kj}T^*_{kj}-T_{ki}T^*_{ki}\|,
$$
which proves the assertion.
\end{proof}
\begin{lemma}
The composition of morphisms of ind-objects is well-defined and is associative.
\end{lemma}
\begin{proof}
With $X_*, Y_*, Z_*$ as above, consider morphisms $T\colon X_* \to Y_*$ and $S\colon Y_* \to Z_*$. By the previous lemma, for fixed $i$ and $n$ we can find $k_0$ such that for all $l\succ k\succ k_0$ the morphism $S_{nl}$ is close to $S_{nk}v_{lk}^*$. But then $S_{nl}T_{li}$
is close to
$$
S_{nk}v_{lk}^*T_{li}=S_{nk}T_{ki}.
$$
It follows that the net $\{S_{nk}T_{ki}\}_k$ is convergent. Therefore the composition $ST$ is well-defined.
Assume now we are given one more morphism $R\colon Z_* \to \{t_{qp}\colon W_p\to\ W_q\}_{p\prec q}$. By definition we have
$$
[R(ST)]_{pi}=\lim_n\lim_k R_{pn}S_{nk}T_{ki}.
$$
As above, by the previous lemma we can find $n_0$ such that $R_{pm}$ is close to $R_{pn}w_{mn}^*$ for $m\succ n\succ n_0$. Similarly, applying the lemma to the morphism $T^*=(T^*_{ki})_{i,k}$, we can find $k_0$ such that $T_{li}$ is close to $v_{lk}T_{ki}$ for $l\succ k\succ k_0$. Then $R_{pm}S_{ml}T_{li}$ is close to $R_{pn}S_{nk}T_{ki}$. It follows that
$$
[R(ST)]_{pi}=\lim_{n,k}R_{pn}S_{nk}T_{ki}.
$$
In a similar way we get the same expression for $[(RS)T]_{pi}$.
\end{proof}
We denote by $\indcat{\CC}$ the category of ind-objects of $\mathcal{C}$. It is easy to see that this is a C$^*$-category. Moreover, the simple objects of $\mathcal{C}$ remain simple in $\indcat{\CC}$. In particular, if $X \in \mathcal{C}$ is irreducible and $Y_*$ is any ind-object, the morphism set $\Mor_\indcat{\CC}(X, Y_*)$ is a Hilbert space, with the inner product such that $(S, T)\iota_X = T^* S$.
\begin{remark}
The morphisms between ind-objects $\{u_{ji}\colon X_i\to X_j\}_{i\prec j}$ and $\{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$ can be described similarly to the purely algebraic case as
$$
\lim_i\colim_k\mathcal{C}(X_i,Y_k),
$$
where limit and colimit are understood in the topological (Banach space theoretic) sense. One disadvantage of this picture is that one has to check not only that the compositions but also that the adjoints are well-defined.
\end{remark}
In the following we assume that $\mathcal{C}$ is essentially small. We will mainly be interested in ind-objects defined by inductive systems of objects of the form $\oplus_{i\in F}X_i$ for finite $F\subset I$ with obvious inclusion maps between them. We denote such ind-objects by $\oplus_{i\in I}X_i$. A morphisms between two such ind-objects $\oplus_{i\in I}X_i$ and $\oplus_{k\in K}Y_k$ is a collection of morphisms $T_{ki}\colon X_i\to Y_k$ such that the morphisms
$$
(T_{ki})_{k\in G, i\in F}\colon \oplus_{i\in F}X_i\to\oplus_{k\in G}Y_k
$$
are uniformly bounded when $F$ and $G$ run over all finite subsets of $I$ and $K$, respectively. In fact, there is no loss of generality in considering only such ind-objects.
\begin{proposition}
Any ind-object of $\mathcal{C}$ is isomorphic to an object of the form $\oplus_{i\in I}X_i$.
\end{proposition}
\begin{proof}
Consider an ind-object $Y_* = \{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$. Fix a simple object $X$ and assume first for simplicity that every $Y_k$ is isotypic to $X$. Consider $H=\Mor_\indcat{\CC}(X,Y_*)$. As remarked before, this is a Hilbert space with inner product such that $T^*S=(S,T)\iota_U$. Choose an orthonormal basis $\{\xi_i\}_{i\in I}$ in $H$. By definition, every basis vector $\xi_i$ is a collection of morphisms $\xi_{ki}\colon X\to Y_k$. For every finite subset $F\subset I$ these morphisms define a morphism $u_{k,F}=(\xi_{ki})_{i\in F}\colon\oplus_{i\in F}U\to Y_k$. The morphisms $u_{k,F}$ define, in turn, a morphism of ind-objects $u\colon\oplus_{i\in I}X\to Y_*$. Orthonormality of the vectors $\xi_i$ implies that $u$ is well-defined and isometric. We claim that $u$ is unitary. This means that for every $k$ the morphisms $u_{k,F}u_{k,F}^*$ converge to the identity morphism of $Y_k$ as $F\to I$. In order to show this, it suffices to check that for every $T\colon U\to Y_k$ we have
$$
\lim_F u_{k,F}u_{k,F}^*T=T.
$$
The morphisms $v_{lk}T\colon X\to Y_l$ define a morphism $\zeta\colon X\to Y_*$. Then, as long as $F$ is large enough, this morphism is close to $\sum_{i\in F}(\zeta,\xi_i)\xi_i=\sum_{i\in F}\xi_i\xi_i^*\zeta$, so that $T$ is close to
$$
\sum_{i\in F}(\xi_i\xi_i^*\zeta)_k=\lim_{l}\sum_{i\in F}\xi_{ki}\xi_{li}^*v_{lk}T=u_{k,F}u_{k,F}^*T,
$$
and our claim is proved. Thus $Y\cong\oplus_{i\in I}X$.
\smallskip
In the general case we can decompose the objects $Y_k$ into isotypic components and repeat the above arguments.
\end{proof}
Let us choose representatives $(U_s)_{s\in\Irr(\mathcal{C})}$ of the isomorphism classes of simple objects of $\mathcal{C}$. Then the proposition and its proof show that any ind-object can be represented by a formal direct sum $\bigoplus_s U_s\otimes H_s$, where~$H_s$ are Hilbert spaces, cf.~\citelist{\cite{MR654325}\cite{MR3121622}}. The morphism space between two such direct sums $\bigoplus_s U_s\otimes H_s$ and $\bigoplus_s U_s\otimes H'_s$ is defined as
$$
\ell^\infty\text{-}\bigoplus_s B(H_s,H'_s).
$$
While this gives a very clear picture of $\indcat{\CC}$, it is not always convenient, as we will see soon, to decompose ind-objects into direct sums of simple objects.
\bigskip
\section{Drinfeld center}
\label{sec:half-braidings}
From now on we assume that $\mathcal{C}$ is an essentially small strict rigid C$^*$-tensor category satisfying our standard assumptions: $\mathcal{C}$ is closed under finite direct sums and subobjects, and the unit of $\mathcal{C}$ is simple.
\subsection{Half-braidings in rigid \texorpdfstring{C$^*$}{C*}-tensor categories}
\label{subsec:half-br}
The category $\indcat{\CC}$ is itself a C$^*$-tensor category: the tensor product of ind-objects defined by inductive systems $\{u_{ji}\colon X_i\to X_j\}_{i\prec j}$ and $\{v_{lk}\colon Y_k\to Y_l\}_{k\prec l}$ is represented by the inductive system
$\{u_{ji}\otimes v_{lk}\colon X_i\otimes Y_k\to X_j\otimes Y_l\}_{i\prec j, k\prec l}$. The category $\indcat{\CC}$ is again closed under direct sums and subobjects, and the unit of $\indcat{\CC}$ is simple, but $\indcat{\CC}$ is no longer rigid. More precisely, the only ind-objects that have conjugates are the ones lying in $\mathcal{C}$.
Consider now the Drinfeld center, or the Drinfeld double, ${\Dcen(\indC)}$ of $\indcat{\CC}$ in the C$^*$-algebraic sense, meaning that it is constructed using unitary half-braidings. More precisely, recall that given an ind-object~$Z$, a half-braiding on $Z$ is a collection of natural in $X\in\indcat{\CC}$ isomorphisms $c_X\colon X\otimes Z\to Z\otimes X$ such that for all objects $X$ and $Y$ in $\indcat{\CC}$ we have
\begin{equation} \label{eq:halfbr}
c_{X\otimes Y}=(c_X\otimes\iota_Y)(\iota_X\otimes c_Y).
\end{equation}
We will only consider unitary half-braidings.
\begin{remark}\label{rhb}
A unitary half-braiding is completely determined by its values on objects of $\mathcal{C}$. In other words, having a unitary half-braiding on $Z$ is the same thing as having a collection of natural in $X\in\mathcal{C}$ unitary isomorphisms $c_X\colon X\otimes Z\to Z\otimes X$ such that for all objects $X$ and $Y$ in $\mathcal{C}$ identity~\eqref{eq:halfbr} holds.
\end{remark}
By definition, the objects of ${\Dcen(\indC)}$ are pairs $(Z,c)$, where $Z$ is an ind-object of $\mathcal{C}$ and $c$ is a unitary half-braiding on $Z$. The morphisms are defined as the morphisms of $\indcat{\CC}$ respecting the half-braidings. Then ${\Dcen(\indC)}$ is a C$^*$-tensor category with the tensor product
$$
(Z,c)\otimes(Z',c')=(Z\otimes Z', (\iota_Z\otimes c')(c\otimes\iota_{Z'})).
$$
Furthermore, ${\Dcen(\indC)}$ is braided, with the unitary braiding defined by
$$
\sigma_{(Z,c),(Z',c')}=c'_Z.
$$
The Drinfeld center ${\Dcen(\CC)}$ of the category $\mathcal{C}$ is a full C$^*$-tensor subcategory of ${\Dcen(\indC)}$. It consists exactly of the objects that have duals: it is not difficult to see that as a dual of $(Z,c)$, with $Z\in\mathcal{C}$, we can take $(\bar Z,\bar c)$, where $\bar c_X=(c_{\bar X})^\vee$.
\subsection{Regular half-braidings}
\label{sec:regul-half-braid}
Our goal now is to construct a particular element of ${\Dcen(\indC)}$ playing the role of the regular representation. Fix representatives $(U_s)_{s \in \Irr(\mathcal{C})}$ of isomorphism classes of simple objects in~$\mathcal{C}$. Denote the index corresponding to the class of $\mathds{1}$ by $e$ and assume for convenience that $U_e=\mathds{1}$. Consider the ind-object
$$
Z_{\mathrm{reg}} = Z_{\mathrm{reg}}(\mathcal{C}) =\bigoplus_{s \in \Irr(\mathcal{C})} U_s\otimes\bar U_s.
$$
Recall that once standard solutions are fixed, we have anti-multiplicative maps $\mathcal{C}(X,Y)\to \mathcal{C}(\bar Y,\bar X)$, $T\mapsto T^\vee$, defined by either of the following identities:
$$
(\iota\otimes T)R_X=(T^\vee\otimes\iota)R_Y,\ \ (T\otimes\iota)\bar R_X=(\iota\otimes T^\vee)\bar R_Y.
$$
Let us now fix an object $X$ and choose a standard solution $(R_X,\bar R_X)$ of the conjugate equations. Let us also fix once for all standard solutions $(R_s,\bar R_s)$ for $U_s$. For every $s$ and $t$ choose isometries $u^\alpha_{st}\colon U_t\to X\otimes U_s$ such that $\sum_\alpha u^\alpha_{st}u^{\alpha*}_{st}$ is the projection onto the isotypic component of $X\otimes U_s$ corresponding to $U_t$. We then define
$$
c_{X,ts}\colon X\otimes U_s\otimes\bar U_s\to U_t\otimes\bar U_t\otimes X
$$
by
$$
c_{X,ts}=\left(\frac{d_s}{d_t}\right)^{1/2}\sum_\alpha (u^{\alpha*}_{st}\otimes u^{\alpha\vee}_{st}\otimes\iota_X)(\iota_X\otimes\iota_s\otimes\iota_{\bar s}\otimes R_X).
$$
Here $d_s$ and $d_t$ denote the quantum dimensions of $U_s$ and $U_t$, while to define $u^{\alpha\vee}_{st}$ we take as the dual of $X\otimes U_s$ the tensor product $\bar U_s\otimes\bar X$, with the standard solutions defined in the usual way from our fixed standard solutions for $X$ and $U_s$:
$$
R_{X\otimes U_s}=(\iota\otimes R_X\otimes\iota)R_s,\ \ \bar R_{X\otimes U_s}=(\iota\otimes \bar R_s\otimes\iota)\bar R_X.
$$
\begin{lemma}
The morphisms $c_{X,ts}$ depend neither on the choice of isometries $u^\alpha_{st}$ nor on the choice of standard solutions for $X$ (assuming that $R_s$ are fixed). Furthermore, these morphisms are natural in $X$.
\end{lemma}
\begin{proof} The claim that $c_{X,ts}$ does not depend on the choice of $u^\alpha_{st}$ is standard and easy to check. As for dependence on the standard solutions, recall that any other standard solution $(R_X',\bar R_X')$ of the conjugate equations for $X$ has the form $R_X'=(u\otimes\iota)R_X$ and $\bar R_X'=(\iota\otimes u)\bar R_X$ for a unitary $u$. This changes $u^{\alpha\vee}_{st}$ into $u^{\alpha\vee}_{st}(\iota_{\bar s}\otimes u^*)$. But then we see that $(u^{\alpha\vee}_{st}\otimes\iota_X)(\iota_{\bar s}\otimes R_X)$ remains unchanged. More explicitly, a direct computation shows that
$$
(u^{\alpha\vee}_{st}\otimes\iota)(\iota\otimes R_X)=(\iota_{\bar t}\otimes\iota_X\otimes \bar R_s^*)(\iota_{\bar t}\otimes u^\alpha_{st}\otimes\iota_{\bar s}) (R_t\otimes\iota_{\bar s}).
$$
Finally, the last statement of the lemma follows easily from the first two, since in order to prove it, it suffices to check that the morphisms $c_{X,ts}$ respect the embeddings $X\to X\oplus Y$ and projections $X\oplus Y\to X$.
\end{proof}
Note for future reference that
\begin{equation} \label{eq:matrixte}
c_{s,te}=\delta_{st}d_s^{-1/2}(\iota_s\otimes R_s).
\end{equation}
Later, see identity~\eqref{eq:regbr}, we will also obtain the following expression for $c_{X,ts}$:
$$
c_{X,ts}=d_t^{1/2}d_s^{1/2}(\iota_t\otimes\iota_{\bar t}\otimes\iota_X\otimes \bar R^*_s)(\iota_t\otimes p^{\bar U_t\otimes X\otimes U_s}_e\otimes\iota_{\bar s})(\bar R_t\otimes\iota_X\otimes\iota_s\otimes\iota_{\bar s}),
$$
where $p^U_e$ is the projection onto the isotypic component of $U$ corresponding to the unit object.
Observe next that the matrix $(c_{X,ts})_{t,s}$ is row and column finite, so when taking compositions of such matrices we will not have to worry about convergence.
\begin{lemma}
\label{lem:reg-half-br-candid-unitary}
The morphisms $c_{X,ts}$ define a unitary $c_X\colon X\otimes Z_{\mathrm{reg}}\toZ_{\mathrm{reg}}\otimes X$.
\end{lemma}
\begin{proof} Let us first check that the morphisms $c_{X,ts}$ define an isometry $c_X\colon X\otimesZ_{\mathrm{reg}}\to Z_{\mathrm{reg}}\otimes X$. It suffices to check that for all $r$ and $s$, we have
$$
\sum_t c_{X,tr}^*c_{X,ts}=\delta_{rs}\iota_{X\otimes U_s\otimes\bar U_s}.
$$
By definition this means that we have to check that
$$
\frac{(d_rd_s)^{1/2}}{d_t}\sum_{t,\alpha,\beta}(\iota\otimes\Tr_{\bar X})(u^\alpha_{rt}u^{\beta*}_{st}\otimes u^{\alpha\vee*}_{rt}u^{\beta\vee}_{st})
=\delta_{rs}\iota.
$$
For this, in turn, it suffices to check that if $U_t\prec X\otimes U_s$, then
\begin{equation} \label{eorth1}
\frac{(d_rd_s)^{1/2}}{d_t}(\iota\otimes\Tr_{\bar X})(u^{\alpha\vee*}_{rt}u^{\beta\vee}_{st})=\delta_{rs}\delta_{\alpha\beta}\iota_{\bar s}.
\end{equation}
Since $\bar U_r$ and $\bar U_s$ are simple, the left hand side is zero if $r\ne s$. If $r=s$, the left hand side is a scalar multiple of the identity morphism. Therefore in this case in order to check the identity we can take categorical traces of both sides. Then the right hand side gives $\delta_{\alpha\beta}d_s$, while the left hand side gives
$$
\frac{d_s}{d_t}\Tr_{\bar U_s\otimes\bar X}( u^{\alpha\vee*}_{st}u^{\beta\vee}_{st} )
=\frac{d_s}{d_t}\Tr_{\bar t}(u^{\beta\vee}_{st}u^{\alpha\vee*}_{st})
=\frac{d_s}{d_t}\Tr_{\bar t}((u^{\alpha*}_{st}u^{\beta}_{st})^\vee)=\delta_{\alpha\beta}d_s,
$$
which is what we need.
\medskip
We next check that $c_X$ is unitary. We have to show that for all $t$ and $\tau$ we have
$$
\sum_s c_{X,ts}c_{X,\tau s}^*=\delta_{t\tau}\iota_{U_t\otimes\bar U_t\otimes X}.
$$
Since $u^{\alpha*}_{st}u^\beta_{s\tau}=\delta_{t\tau}\delta_{\alpha\beta}\iota_{U_t}$, the above identity is immediate for $t\ne\tau$, while for $t=\tau$ the left hand side equals
$$
\sum_{s,\alpha}\frac{d_s}{d_t}(\iota_{t}\otimes u^{\alpha\vee}_{st}\otimes\iota_X)(\iota_{t}\otimes\iota_{\bar s}\otimes R_X R_X^*)(\iota_{t}\otimes u^{\alpha\vee*}_{st}\otimes\iota_X).
$$
Therefore in order to finish the proof it suffices to show that for every $s$, the morphism
$$
\sum_\alpha\frac{d_s}{d_t}(u^{\alpha\vee}_{st}\otimes\iota_X)(\iota_{\bar s}\otimes R_XR_X^*)(u^{\alpha\vee*}_{st}\otimes\iota_X)
$$
is the projection onto the isotypic component of $\bar U_t\otimes X$ corresponding to $\bar U_s$. For this, observe that by Frobenius reciprocity the morphisms
$$
w^\alpha_{ts}=\left(\frac{d_s}{d_t}\right)^{1/2}(u^{\alpha\vee}_{st}\otimes\iota_X)(\iota_{\bar s}\otimes R_X)\colon \bar U_s\to \bar U_t\otimes X
$$
form a basis in $\mathcal{C}(\bar U_s,\bar U_t\otimes X)$. We claim that we also have the orthogonality $w^{\alpha*}_{ts}w^\beta_{ts}=\delta_{\alpha\beta}\iota$ with respect to this relation. By definition of the categorical trace this is equivalent to
$$
\frac{d_s}{d_t}(\iota\otimes\Tr_{\bar X})(u^{\alpha\vee*}_{st}u^{\beta\vee}_{st})=\delta_{\alpha\beta}\iota_{\bar U_s}.
$$
But this follows from \eqref{eorth1}, so our claim is proved. We conclude that $\sum_\alpha w^{\alpha}_{ts}w^{\alpha*}_{ts}$ is the projection onto the isotypic component of $\bar U_t\otimes X$ corresponding to $\bar U_s$.
\end{proof}
\begin{theorem} \label{them:reg-half-braiding}
The unitaries $c_X\colon X\otimesZ_{\mathrm{reg}}\toZ_{\mathrm{reg}}\otimes X$ form a half-braiding on $Z_{\mathrm{reg}}$.
\end{theorem}
\begin{proof}
It remains only to check identity~\eqref{eq:halfbr}. In order to compute $c_{X\otimes Y}$, choose isometries $u^\alpha_{rt}\colon U_t\to X\otimes U_r$ as before, and similarly choose isometries $v^\beta_{sr}\colon U_r\to Y\otimes U_s$. Then using the isometries
$$
(\iota_X\otimes v^\beta_{sr})u^\alpha_{rt}\colon U_t\to X\otimes Y\otimes U_s
$$
in the definition of $c_{X\otimes Y,ts}$, we get
\begin{align*}
c_{X\otimes Y,ts}&=\left(\frac{d_s}{d_t}\right)^{1/2}\sum_{r,\alpha,\beta}(u^{\alpha*}_{rt}(\iota_X\otimes v^{\beta*}_{sr})\otimes u^{\alpha\vee}_{rt}(v^{\beta\vee}_{sr}\otimes\iota_{\bar X})\otimes\iota_{X\otimes Y})\\
&\qquad\qquad\qquad\qquad\qquad\qquad (\iota_{X\otimes Y\otimes U_s\otimes\bar U_s \otimes\bar Y}\otimes R_X\otimes\iota_Y)(\iota_{X\otimes Y\otimes U_s\otimes\bar U_s}\otimes R_Y)\\
&=\sum_r(c_{X,tr}\otimes\iota_Y)(\iota_X\otimes c_{Y,rs}).
\end{align*}
This means that $c_{X\otimes Y}=(c_X\otimes\iota_Y)(\iota_X\otimes c_Y)$.
\end{proof}
We will often denote the object $(Z_{\mathrm{reg}}, c)$ by just one symbol $Z_{\mathrm{reg}}$ or $Z_{\mathrm{reg}}(\mathcal{C})$.
\subsection{Unitary half-braidings and amenability}
We have shown that ${\Dcen(\indC)}$ is always rich. Expanding on ideas of Longo and Roberts~\cite{MR1444286}*{Section~5}, we will now show that generally this is not the case for~${\Dcen(\CC)}$, so we do need to consider ind-objects in order to construct nontrivial unitary half-braidings. These considerations are not going to be used in the subsequent sections, so we will be somewhat brief.
\smallskip
For every object $X$ in $\mathcal{C}$ denote by $\Gamma_X=(a^X_{st})_{s,t}\in B(\ell^2(\Irr(\mathcal{C})))$ the matrix describing decompositions of $X\otimes Y$ into simple objects, so $a^X_{st}=\dim \mathcal{C}(U_s,X\otimes U_t)$. Then $\norm{\Gamma_X}\le d(X)$, and the category $\mathcal{C}$ is called amenable if $\|\Gamma_X\|=d(X)$ for all objects $X$ in $\mathcal{C}$. Let us say that an object $X$ is amenable, if the full rigid C$^*$-tensor subcategory of $\mathcal{C}$ generated by $X$ is amenable.
We remark that it is not difficult to show, see e.g.~the proof of~\cite{MR1644299}*{Proposition~4.8}, that the norm of the matrix $\Gamma_X$ remains the same if we replace $\mathcal{C}$ by any full rigid C$^*$-tensor subcategory of $\mathcal{C}$ containing $X$. Therefore $\mathcal{C}$ is amenable if and only if every object of $\mathcal{C}$ is amenable.
\begin{theorem}
Assume that for a rigid C$^*$-tensor category $\mathcal{C}$ there exists a unitary half-braiding on an object $X\in\mathcal{C}$. Then $X$ is amenable.
\end{theorem}
\begin{proof}
We may assume that $\mathcal{C}$ is generated by $X$ as a rigid C$^*$-tensor category. Replacing, if necessary, $X$ by~$X\oplus \bar X$, we may also assume that every simple object embeds into $X^{\otimes n}$ for some $n\ge1$. Consider the Poisson boundary $\mathcal{P}$ of $\mathcal{C}$ with respect to the probability measure on $\Irr(\mathcal{C})$ defined by the normalized categorical trace on $X$~\cite{arXiv:1405.6572}. We will prove that the Poisson boundary is trivial, which by~\cite{arXiv:1405.6572}*{Theorem~5.7} implies amenability of $\mathcal{C}$.
We view $\mathcal{C}$ as a C$^*$-tensor subcategory of $\mathcal{P}$. By definition, the elements of $\mathcal{P}(Z)$ are bounded collections~$\xi=(\xi_Y)_Y$ of natural in $Y$ morphisms $Y\otimes Z\to Y\otimes Z$ that are harmonic, meaning that
$$
(\tr_X\otimes\iota)(\xi_{X\otimes Y})=\xi_Y\ \ \text{for all objects}\ \ Y\in\mathcal{C}.
$$
They can be realized as follows~\cite{arXiv:1405.6572}*{Proposition~3.3}. The algebras $\mathcal{N}_Z^{(n)}=\mathcal{C}(X^{\otimes n}\otimes Z)$, equipped with the normalized categorical traces and the embeddings $T\mapsto\iota_X\otimes T$, form an inductive system. In the limit we get a finite von Neumann algebra $\mathcal{N}_Z$. For any $\xi\in\mathcal{P}(Z)$, the elements $\xi^{[n]}=\xi_{X^{\otimes n}}\in\mathcal{N}^{(n)}_Z$ converge in the strong$^*$ operator topology to an element $\xi^{[\infty]}\in\mathcal{N}_Z$, and the map $\xi\mapsto\xi^{[\infty]}$ gives an algebra embedding of $\mathcal{P}(Z)$ into~$\mathcal{N}_Z$.
Take $\xi\in\mathcal{P}(Z)$. Then $(\iota_X\otimes\xi)^{[n]}=\xi_{X^{\otimes(n+1)}}$. On the other hand, if $c$ is a unitary half-braiding on $X$, then
$$
(c_{Z}(\xi\otimes\iota_X)c_{Z}^*)^{[n]}=(\iota\otimes c_{Z})(\xi_{X^{\otimes n}}\otimes\iota)(\iota\otimes c_{Z}^*)
=(c^*_{X^{\otimes n}}\otimes\iota)
(\iota\otimes\xi_{X^{\otimes n}})(c_{X^{\otimes n}}\otimes\iota).
$$
Since $\iota_X\otimes\xi_{X^{\otimes n}}$ is the image of $\xi^{[n]}$ under the embedding $\mathcal{N}^{(n)}_Z\hookrightarrow \mathcal{N}^{(n+1)}_Z$, as $n$ grows, the last expression becomes close in the trace-norm to
$$
(c^*_{X^{\otimes n}}\otimes\iota)\xi^{[n+1]}(c_{X^{\otimes n}}\otimes\iota)
=(c^*_{X^{\otimes n}}\otimes\iota)\xi_{X^{\otimes(n+1)}}(c_{X^{\otimes n}}\otimes\iota)=\xi_{X^{\otimes(n+1)}}=(\iota_X\otimes\xi)^{[n]}.
$$
It follows that $c_{Z}(\xi\otimes\iota_X)c_{Z}^*=\iota_X\otimes\xi$, that is,
$$
(\iota_Y\otimes c_{Z})(\xi_Y\otimes\iota_X)(\iota_Y\otimes c_{Z}^*)=\xi_{Y\otimes X}\ \ \text{for all}\ \ Y.
$$
The left hand side equals $(c^*_{Y}\otimes\iota_Z)(\iota_X\otimes\xi_Y)(c_{Y}\otimes\iota_Z)$. Therefore by conjugating by $c_{Y}\otimes\iota_Z$ we get
$$
\iota_X\otimes\xi_Y=\xi_{X\otimes Y}\ \ \text{for all}\ \ Y.
$$
A simple induction shows then that the same identity holds with $X$ replaced by $X^{\otimes n}$, hence it holds for any simple object $U$ in place of $X$. Letting $Y=\mathds{1}$ we then get $\iota_U\otimes\xi_\mathds{1}=\xi_U$. Thus, under our embedding of~$\mathcal{C}(Z)$ into~$\mathcal{P}(Z)$, we have $\xi=\xi_\mathds{1}\in\mathcal{C}(Z)$.
\end{proof}
In particular, if $\mathcal{C}$ admits a unitary braiding, or even weaker, if $\mathcal{C}$ is generated as a rigid C$^*$-tensor category by objects admitting unitary half-braidings, then $\mathcal{C}$ is amenable. This is a categorical analogue of the fact that abelian groups are amenable.
\begin{example}
If $\mathcal{C}=\Rep G$ is the representation category of a compact quantum group $G$, then a necessary condition for amenability of $U\in\Rep G$ is the equality $\dim U=\dim_q U$. Therefore if $\dim U<\dim_q U$, there exists no unitary half-braiding on $U$.
\end{example}
\bigskip
\section{Representations of the character algebra}
\label{sec:character-algebra}
We continue to assume that $\mathcal{C}$ is a rigid C$^*$-tensor category as in the previous section.
\subsection{From half-braidings to representations}
Recall that there is a semiring structure on the semigroup $\mathbb{Z}_+[\Irr(\mathcal{C})]$, with the product defined by
$$
[U] \cdot [V] = \sum_{s\in\Irr(\mathcal{C})} \dim \mathcal{C}(U_s, U \otimes V) [U_s].
$$
The operation $[U] \mapsto [\bar{U}]$ extends to an anti-multiplicative involution of this semiring. We embed the involutive semiring $\mathbb{Z}_+[\Irr(\mathcal{C})]$ into the involutive $\mathbb{C}$-algebra $\mathbb{C}[\Irr(\mathcal{C})]$.
Suppose that $(c_X\colon X \otimes Z \to Z \otimes X)_{X \in \mathcal{C}}$ is a unitary half-braiding on an ind-object $Z$. We want to define a $*$-representation of $\mathbb{C}[\Irr(\mathcal{C})]$ on the Hilbert space $\Mor_{\indcat{\CC}}(\mathds{1}, Z)$ with scalar product defined by $(\xi,\zeta)\iota=\zeta^*\xi$. Let $X$ be an object in $\mathcal{C}$ and $(R_X, \bar{R}_X)$ be a standard solution of the conjugate equations for~$X$. If $\xi \in \Mor_\indcat{\CC}(\mathds{1}, Z)$, we obtain a new element in the same morphism set by
$$
\pi_{(Z,c)}([X])\xi = (\iota_Z \otimes \bar{R}_X^*)(c_X \otimes \iota_{\bar{X}})(\iota_X \otimes \xi \otimes \iota_{\bar{X}})\bar{R}_X\colon \mathds{1} \to X \otimes \bar{X} \to X \otimes Z \otimes \bar{X} \to Z \otimes X \otimes \bar{X} \to Z.
$$
Since any other choice of $(R_X, \bar{R}_X)$ is of the form $((T \otimes \iota) R_X, (\iota \otimes T) \bar{R}_X)$ for some unitary $T$, the above definition does not depend on the choice of a standard solution. In order to simplify the notation we write~$\pi_Z$ instead of $\pi_{(Z,c)}$ when there is no danger of confusion.
It is clear that $\|\pi_Z([X])\|\le \|\bar R_X\|^2=d(X)$. It is also easy to see that $\pi_Z([X])$ is additive in $X$. The half-braiding axiom~\eqref{eq:halfbr} implies that $\pi_Z([X])$ is multiplicative in $X$. Thus we obtain a representation $\pi_Z$ of the algebra $\mathbb{C}[\Irr(\mathcal{C})]$ on $\Mor_\indcat{\CC}(\mathds{1}, Z)$.
Next we want to check the compatibility with the involution. For this we need the following lemma, which we will also repeatedly use later.
\begin{lemma} \label{lem:braidinv}
We have $(\iota_{\bar X}\otimes c_X)(R_X\otimes\iota_Z)=(c^*_{\bar X}\otimes\iota_X)(\iota_Z\otimes R_X)$.
\end{lemma}
\begin{proof} Since $c_\mathds{1}=\iota$, we have $c_{\bar X\otimes X}(R_X\otimes\iota_Z)=\iota_Z\otimes R_X$. Using then that $c_{\bar X\otimes X}=(c_{\bar X}\otimes\iota_X)(\iota_{\bar X}\otimes c_X)$, we get the result.
\end{proof}
\begin{lemma}
\label{lem:char-alg-act-compat-adj}
For any $\xi, \eta \in \Mor_\indcat{\CC}(\mathds{1}, Z)$ and $X \in \mathcal{C}$, we have $(\pi_Z([X]) \xi, \eta) = (\xi, \pi_Z([\bar{X}]) \eta)$.
\end{lemma}
\begin{proof}
We have to show the equality
\begin{equation} \label{eq:starrep}
\eta^*(\iota_Z \otimes \bar{R}_X^*)(c_X \otimes \iota_{\bar{X}}) (\iota_X \otimes \xi \otimes \iota_{\bar{X}}) \bar{R}_X = R_X^* (\iota_{\bar{X}} \otimes \eta^* \otimes \iota_X) (c_{\bar{X}}^* \otimes \iota_X) (\iota_Z \otimes R_X) \xi.
\end{equation}
The left hand side can be written as
$$
\bar{R}_X^*(\eta^*\otimes\iota_X\otimes\iota_{\bar X})(c_X \otimes \iota_{\bar{X}}) (\iota_X \otimes \xi \otimes \iota_{\bar{X}}) \bar{R}_X.
$$
Consider the morphism $T=(\eta^*\otimes\iota_X)c_X (\iota_X \otimes \xi)\in\mathcal{C}(X)$. Then the left hand side of \eqref{eq:starrep} equals
$$
\bar{R}_X^*(T\otimes\iota_{\bar X})\bar R_X=\Tr_X(T).
$$
On the other hand, since $(c^*_{\bar X}\otimes\iota_X)(\iota_Z\otimes R_X)=(\iota_{\bar X}\otimes c_X)(R_X\otimes\iota_Z)$ by Lemma~\ref{lem:braidinv}, the right hand side of~\eqref{eq:starrep} equals
$$
R_X^*(\iota\otimes T)R_X=\Tr_X(T),
$$
so we get the desired equality.
\end{proof}
From now on by a representation of $\mathbb{C}[\Irr(\mathcal{C})]$ we mean a $*$-representation.
\begin{definition}
We define the \emph{C$^*$-character algebra} $C^*(\mathcal{C})$ of $\mathcal{C}$ to be the C$^*$-completion of the $*$-algebra $\mathbb{C}[\Irr(\mathcal{C})]$ with respect to the representations $\pi_Z$ for all objects $(Z, c)\in{\Dcen(\indC)}$.
\end{definition}
As we already observed, $\|[X]\|\le d(X)$ in $C^*(\mathcal{C})$. The next example shows that this is actually equality.
\begin{example}
Consider the trivial half-braiding $(X \otimes \mathds{1} \to \mathds{1} \otimes X)_X$ for $\mathds{1}$. Then we obtain a representation of $\mathbb{C}[\Irr(\mathcal{C})]$ on $\mathbb{C}$, that is, a character. Expanding the relevant definitions we see that $\pi_\mathds{1}([X]) = d(X)$. We call $\pi_\mathds{1}$ the \emph{trivial representation} of $\mathbb{C}[\Irr(\mathcal{C})]$.
\end{example}
\begin{example} \label{ex:regrep}
Consider the half-braiding $(X \otimes Z_{\mathrm{reg}} \to Z_{\mathrm{reg}} \otimes X)_X$ constructed in Section~\ref{subsec:half-br}. The Hilbert space $\Mor_\indcat{\CC}(\mathds{1}, Z_{\mathrm{reg}})$
has an orthonormal basis consisting of the vectors $\xi_s=d_s^{-1/2}\bar{R}_s$, $s\in\Irr(\mathcal{C})$. It follows from \eqref{eq:matrixte} that $\pi_{Z_{\mathrm{reg}}}([U_s])\xi_e=\xi_s$. Therefore $\pi_{Z_{\mathrm{reg}}}$ can be identified with the regular representation of $\mathbb{C}[\Irr(\mathcal{C})]$ on $\ell^2(\Irr(\mathcal{C}))$.
\end{example}
\begin{remark}\label{rmk:q-grp-char-corr}
If $\mathcal{C} = \Rep G$ for some compact quantum group $G$, the half-braidings correspond to the $*$-representations of the Drinfeld double $\mathcal{O}_c(\hat{D}(G)) = \mathcal{O}(G) \bowtie c_c(\hat{G})$ via the standard argument (cf.~\cite{MR1321145}*{Section~IX.5}). The C$^*$-algebra $C^*(\mathcal{C})$ coincides with the C$^*$-completion of the character algebra of $G$ with respect to the embedding $\chi_U \mapsto \sigma_{-i/2}(\chi_U) h$ and the norm on $\mathcal{O}_c(\hat{D}(G))$ induced by the ``spherical unitary representations'', where $h$ is the Haar state and $\sigma_z$ is its modular automorphism group, see~\cite{MR3238527}*{Remark~31}.
\end{remark}
\subsection{Positive definite functions}\label{sec:positivedefinite}
Given an object $(Z,c)\in{\Dcen(\indC)}$ and a vector $\xi\in\Mor_{\indcat{\CC}}(\mathds{1},Z)$, the cyclic representation of $\mathbb{C}[\Irr\mathcal{C}]$ on $\overline{\pi_Z(\mathbb{C}[\Irr\mathcal{C}])\xi}$ is completely determined by the function $\phi(s)=d_s^{-1}(\pi([U_s])\xi,\xi)$ on $\Irr(\mathcal{C})$. It is natural to call such functions positive definite. While this definition would be sufficient for the theory we develop in the subsequent sections, it is clearly unsatisfactory. A correct intrinsic definition has been given by Popa and Vaes~\cite{MR3406647}. We will present it in a way convenient for our applications.
For a function $\phi$ on $\Irr(\mathcal{C})$ denote by $M^\phi$ the endomorphism of the identity functor on $\mathcal{C}$ such that $M^\phi_s\colon U_s\to U_s$ is the scalar morphism $\phi(s)$ for every $s\in\Irr(\mathcal{C})$. For $s,t\in\Irr(\mathcal{C})$ define a morphism
$$
A^\phi_{st}=d_s^{1/2}d_t^{1/2}(\iota_s\otimes\iota_{\bar s}\otimes \bar R^*_t)(\iota_s\otimes M^\phi_{\bar U_s\otimes U_t}\otimes\iota_{\bar t})(\bar R_s\otimes\iota_t\otimes\iota_{\bar t})\colon U_t\otimes\bar U_t\to U_s\otimes\bar U_s.
$$
\begin{definition}
A function $\phi$ on $\Irr(\mathcal{C})$ is called \emph{positive definite}, or a \emph{cp-multiplier}, if for any $s_1,\dots,s_n\in\Irr(\mathcal{C})$ the morphism
$$
(A^\phi_{s_i,s_j})^n_{i,j=1}\colon\bigoplus^n_{k=1}U_{s_k}\otimes\bar U_{s_k}\to \bigoplus^n_{k=1}U_{s_k}\otimes\bar U_{s_k}
$$
is positive.
\end{definition}
In the original definition of Popa and Vaes a cp-multiplier is defined by requiring certain maps $\theta^\phi_{U,V}$ on $\mathcal{C}(U\otimes V)$ to be completely positive for all $U,V\in\mathcal{C}$. But it is shown in~\cite{MR3406647}*{Lemma~3.7} that it suffices to check positivity of $\theta^\phi_{U,\bar U}(\bar R_U\bar R_U^*)$ for all $U$. Expanding the definitions one can check that
$$
\theta^\phi_{U,\bar U}(\bar R_U\bar R^*_U)=( \iota_U\otimes\iota_{\bar U}\otimes \bar R_U^*)(\iota_U\otimes M^\phi_{\bar U\otimes U} \otimes \iota_{\bar U})(\bar R_U\otimes\iota_U\otimes\iota_{\bar U}).
$$
For $U=\oplus^n_{i=1}U_{s_i}$ positivity of the above expression means exactly positivity of $(A^\phi_{s_i,s_j})^n_{i,j=1}$. Thus the above definition is equivalent to
the one in~\cite{MR3406647}.
\begin{example}
Let $\Gamma$ be a discrete group and $\mathcal{C}=\Hilb_{\Gamma,f}$ be the category of $\Gamma$-graded finite dimensional Hilbert spaces, or in other words, the representation category of the dual compact quantum group $\hat\Gamma$. Thus $\Irr(\mathcal{C})=\Gamma$ and we can choose representatives $U_s$, $s\in\Gamma$, of isomorphism classes of simple objects such that $U_s\otimes U_t=U_{st}$. Then $A^\phi_{st}$ is the scalar endomorphism $\phi(s^{-1}t)$ of $U_e$. Therefore a function $\phi$ on $\Gamma$ is positive definite in the above sense if and only if the matrix $(\phi(s_i^{-1}s_j))^n_{i,j=1}$ is positive for any $s_1,\dots,s_n\in\Gamma$, which is the standard definition of positive definite functions on groups.
\end{example}
\begin{example}\label{ex:deltae}
Consider an arbitrary rigid C$^*$-tensor category $\mathcal{C}$ and the function $\phi=\delta_e$. In this case $A^\phi_{st}=\delta_{st}\iota$, since $M^\phi_{\bar U_s\otimes U_t}=\delta_{st}d_s^{-1}R_sR_s^*$. Therefore the function $\phi=\delta_e$ is positive definite.
\end{example}
\begin{theorem} \label{thm:positivedef}
For any function $\phi$ on $\Irr(\mathcal{C})$ the following conditions are equivalent:
\begin{itemize}
\item[(i)] $\phi$ is positive definite;
\item[(ii)] $\phi(s)=d_s^{-1}(\pi_Z([U_s])\xi,\xi)$ for some $(Z,c)\in{\Dcen(\indC)}$ and $\xi\in\Mor_{\indcat{\CC}}(\mathds{1},Z)$;
\item[(iii)] $\phi(s)=d_s^{-1}\omega([U_s])$ for a positive linear functional $\omega$ on $C^*(\mathcal{C})$.
\end{itemize}
\end{theorem}
Popa and Vaes defined a C$^*$-algebra $C_u(\mathcal{C})$ for a rigid C$^*$-tensor category $\mathcal{C}$ as the C$^*$-envelope of $\mathbb{C}[\Irr(\mathcal{C})]$ with respect to the representations $\pi\colon \mathbb{C}[\Irr(\mathcal{C})]\to B(H)$ such that $s \mapsto d_s^{-1}(\pi([U_s]) \xi, \xi)$ is a cp-multiplier for any $\xi \in H$. As an immediate consequence of the above theorem we get the following.
\begin{corollary} \label{cor:twodef}
The identity map on $\mathbb{C}[\Irr(\mathcal{C})]$ extends to an isomorphism of $C^*(\mathcal{C})$ onto $C_u(\mathcal{C})$.
\end{corollary}
Turning to the proof of Theorem~\ref{thm:positivedef}, (ii) obviously implies (iii). Let us prove that (iii) implies (i). Take a representation $\pi\colon C^*(\mathcal{C})\to B(H)$ and a vector~$\xi$. We want to show that the function $\phi(s)=d_s^{-1}(\pi([U_s])\xi,\xi)$ is positive definite. Since any representation of $C^*(\mathcal{C})$ is weakly contained in a direct sum of representations defined by objects of ${\Dcen(\indC)}$ and the set of positive definite functions is closed under convex combinations and pointwise limits, without loss of generality we may assume that $\pi$ is defined by an object $(Z,c)\in{\Dcen(\indC)}$. Then $\xi \in \Mor_\indcat{\CC}(\mathds{1}, Z)$. In other words, it suffices to show that (ii) implies (i).
\begin{lemma} \label{lem:Mphi}
For every object $U$ of $\mathcal{C}$ the endomorphism $M^\phi_U$ is defined by the composition
$$
U\xrightarrow{\iota\otimes\xi}U\otimes Z\xrightarrow{c_U}Z\otimes U\xrightarrow{\xi^*\otimes\iota}U.
$$
\end{lemma}
\begin{proof}
It is clear that the composition in the formulation is natural in $U$. Therefore it suffices to consider $U=U_s$. Then the above composition is a scalar endomorphism $\alpha_s$. It follows that
$$
d_s\phi(s)=(\pi([U_s]) \xi, \xi)= \xi^*(\iota_Z \otimes \bar{R}_s^*)(c_s \otimes \iota_{\bar{s}})(\iota_s \otimes \xi \otimes \iota_{\bar{s}})\bar{R}_s=\alpha_s\bar{R}_s^*\bar{R}_s,
$$
so $\phi(s)=\alpha_s$.
\end{proof}
\begin{proof}[Proof of the implication {\rm(ii)}$\Rightarrow${\rm(i)} in Theorem~\ref{thm:positivedef}]
Consider the function $\phi(s)=d_s^{-1}(\pi([U_s])\xi,\xi)$ as above. By the previous lemma we have
\begin{align*}
A^\phi_{st}&=d_s^{1/2}d_t^{1/2}(\iota_s\otimes\iota_{\bar s}\otimes \bar R^*_t)(\iota_s\otimes\xi^*\otimes\iota_{\bar s}\otimes \iota_t\otimes\iota_{\bar t})
(\iota_s\otimes c_{\bar U_s\otimes U_t}\otimes\iota_{\bar t})(\iota_s\otimes\iota_{\bar s}\otimes\iota_t\otimes\xi\otimes\iota_{\bar t})
(\bar R_s\otimes\iota_t\otimes\iota_{\bar t})\\
&=d_s^{1/2}d_t^{1/2}(\iota_s\otimes\xi^*\otimes\iota_{\bar s})(\iota_s\otimes\iota_Z\otimes\iota_{\bar s}\otimes \bar R_t^*)
(\iota_s\otimes c_{\bar U_s\otimes U_t}\otimes\iota_{\bar t})(\bar R_s\otimes\iota_t\otimes\iota_Z\otimes\iota_{\bar t})
(\iota_t\otimes\xi\otimes\iota_{\bar t}).
\end{align*}
Using that $c_{\bar U_s\otimes U_t}=(c_{\bar s}\otimes\iota_t)(\iota_{\bar s}\otimes c_t)$ we get
$$
A^\phi_{st}=d_s^{1/2}d_t^{1/2}(\iota_s\otimes\xi^*\otimes\iota_{\bar s})(\iota_s\otimes c_{\bar s})(\bar R_s\otimes\iota_Z)(\iota_Z\otimes\bar R_t^*)(c_t\otimes\iota_{\bar t})(\iota_t\otimes\xi\otimes\iota_{\bar t}).
$$
Since $(\iota_s\otimes c_{\bar s})(\bar R_s\otimes\iota_Z)=(c^*_s\otimes\iota_Z)(\iota_Z\otimes\bar R_s)$ by Lemma~\ref{lem:braidinv}, we therefore see that $A^\phi_{st}=T_s^*T_t$, where
$$
T_t=d_t^{1/2}(\iota_Z\otimes\bar R_t^*)(c_t\otimes\iota_{\bar t})(\iota_t\otimes\xi\otimes\iota_{\bar t})\colon U_t\otimes\bar U_t\to Z.
$$
This obviously implies positive definiteness of $\phi$.
\end{proof}
Next, starting from a positive definite function we want to construct a unitary half-braiding. The construction will be a modification of our construction of $Z_{\mathrm{reg}}$. Let us first describe the framework within which we will define such a modification.
Consider an ind-object $\{u_{ji}\colon X_i\to X_j\}_{i\prec j}$ and assume that for every $i$ we are given a positive morphism $A_i\colon X_i\to X_i$ such that
\begin{equation} \label{eq:poscompatibl}
A_i=u_{ji}^*A_ju_{ji}\ \ \text{for}\ \ i\prec j.
\end{equation}
From this data we can construct a new ind-object as follows. For every $i$ choose an object $Y_i$ and a surjective morphism $v_i\colon X_i\to Y_i$ such that $v_i^*v_i=A_i$. It is easy to see that such a pair $(Y_i,v_i)$ exists and is unique up to a unitary isomorphism. For example, we can take $Y_i$ to be the subobject of $X_i$ corresponding to the complement of the kernel of $A_i$ and then take $v_i=A_i^{1/2}$. But it is more instructive to think of $Y_i$ as a quotient of $X_i$ with a new inner product on morphisms into $Y_i$: given morphisms $S,T\colon X\to X_i$ we have
\begin{equation} \label{eq:newscalarpr}
(v_iT)^*v_iS=T^*A_iS.
\end{equation}
The following lemma is immediate by definition.
\begin{lemma} \label{lem:indiso}
Assume $T\colon X_i\to U$ is a morphism such that $T^*T=A_i$. Then there exists a unique isometry $\tilde T\colon Y_i\to U$ such that $T=\tilde T v_i$.
\end{lemma}
In particular, applying this to $T=v_ju_{ji}$ we conclude that for $i\prec j$ there exists a unique isometric morphism $w_{ji}\colon Y_i\to Y_j$ such that $v_ju_{ji}=w_{ji}v_i$. We thus get a new ind-object $\{w_{ji}\colon Y_i\to Y_j\}_{i\prec j}$.
We will use this construction for ind-objects of the form $\oplus_{i\in I}X_i$. In this case to be given positive endomorphisms $A_F$ of $X_F=\oplus_{i\in F}X_i$ for all finite sets $F\subset I$ satisfying \eqref{eq:poscompatibl} is the same thing as to have morphisms $A_{ji}\colon X_i\to X_j$ such that $(A_{ij})_{i,j\in F}$ is positive for any finite $F$. In this case, by slightly abusing the terminology, we simply say that $A=(A_{ij})_{i,j\in I}$ is positive. Therefore, starting from an ind-object $\oplus_{i\in I}X_i$ and a positive matrix of morphisms $A=(A_{ij})_{i,j\in I}$ we get a new ind-object, which we denote by
$$
A\text{-}\oplus_{i\in I}X_i.
$$
Note that by definition for any finite set $F\subset I$ we have a canonical morphism $\oplus_{i\in F}X_i\to A\text{-}\oplus_{i\in I}X_i$ obtained by composing $v_F\colon \oplus_{i\in F}X_i\to A\text{-}\oplus_{i\in F}X_i$ with the canonical isometry $A\text{-}\oplus_{i\in F}X_i\to A\text{-}\oplus_{i\in I}X_i$. But in general these morphisms do not define a bounded morphism $\oplus_{i\in I} X_i\to A\text{-}\oplus_{i\in I}X_i$.
In some cases an endomorphism of the original ind-object defines an endomorphism of the new one. The following will be sufficient for our purposes.
\begin{lemma} \label{lem:indunitary}
Let $\oplus_{i\in I}X_i$ and $\oplus_{k\in K}X'_k$ be ind-objects, $A=(A_{ij})_{i,j\in I}$ and $B=(B_{kl})_{k,l\in K}$ be positive matrices of morphisms $A_{ij}\colon X_j\to X_i$ and $B_{kl}\colon X'_l\to X'_k$, and $U=(U_{ki})_{k\in K,i\in I}\colon\oplus_iX_i\to\oplus_kX'_k$ be a unitary such that the matrix $(U_{ki})_{k,i}$ is row and column finite and $UA=BU$, that is,
$$
\sum_j U_{kj}A_{ji}=\sum_l B_{kl}U_{li}
$$
for all $i\in I$ and $k\in K$. Then $U$ defines a unitary $V\colon A\text{-}\oplus_i X_i\to B\text{-}\oplus_k X'_k$, meaning that for any finite set $F\subset I$ and all sufficiently large finite sets $G\subset K$ the diagram
$$
\begin{xymatrix}{
\oplus_{i\in F}X_i\ar[rr]^{(U_{ki})_{k\in G,i\in F}}\ar[d] & & \ar[d]\oplus_{k\in G}X'_k\\
A\text{-}\oplus_{i\in I} X_i\ar[rr]_V& & B\text{-}\oplus_{k\in K} X'_k
}\end{xymatrix}
$$
commutes.
\end{lemma}
\begin{proof} Take a finite set $F\subset I$ and let $G\subset K$ be any finite set such that $U_{ki}=0$ if $i\in F$ and $k\notin G$. Consider the morphisms $A_F=(A_{ij})_{i,j\in F}$, $B_G=(B_{kl})_{k,l\in G}$ and $U_{G,F}=(U_{ki})_{k\in G,i\in F}$. By the choice of $G$ and the assumptions of the lemma, for any $i,j\in F$ we have
$$
A_{ij}=\sum_{k,l\in K}U^*_{ki}B_{kl}U_{lj}=\sum_{k,l\in G}U^*_{ki}B_{kl}U_{lj},
$$
so $A_F=U^*_{G,F}B_GU_{G,F}$. By Lemma~\ref{lem:indiso} this implies that $U_{G,F}\colon\oplus_{i\in F}X_i\to \oplus_{k\in G}X'_k$ induces an isometry $V_{G,F}\colon A\text{-}\oplus_{i\in F}X_i\to B\text{-}\oplus_{k\in G}X'_k$. It is easy to see that the family of isometries $V_{G,F}$ is consistent and hence defines an isometry $V\colon A\text{-}\oplus_i X_i\to B\text{-}\oplus_k X'_k$ satisfying the statement of the lemma. Using $U^*$ instead of~$U$ we can similarly construct an isometry $V'\colon B\text{-}\oplus_k X'_k\to A\text{-}\oplus_i X_i$. It is straightforward to check that the isometries $V$ and $V'$ are inverse to each other.
\end{proof}
Note that, under the assumptions of the previous lemma, if we denote by $\pi_F\colon\oplus_{i\in F}X_i\to A\text{-}\oplus_{i\in I}X_i$ and $\pi'_G\colon\oplus_{k\in G}X'_k\to B\text{-}\oplus_{k\in K}X'_k$ the canonical morphisms, then this lemma together with identity~\eqref{eq:newscalarpr} imply that for any morphisms $S=(S_i)_{i\in F}\colon X\to \oplus_{i\in F}X_i$ and $T=(T_k)_{k\in G}\colon X\to \oplus_{k\in G}X'_k$ we have
\begin{equation} \label{eq:newscalarpr1}
(\pi'_G T)^*V\pi_FS=\sum_{k,l,i}T_k^*B_{kl}U_{li}S_i.
\end{equation}
\begin{proof}[Proof of the implication {\rm(i)}$\Rightarrow${\rm(ii)} in Theorem~\ref{thm:positivedef}]
Let $\phi$ be a positive definite function on $\Irr(\mathcal{C})$, so that we have a positive matrix $A^\phi$ of morphisms $A^\phi_{st}\colon U_t\otimes\bar U_t\to U_s\otimes\bar U_s$. We can then define an ind-object
$$
Z_\phi=A^\phi\text{-}\bigoplus_{s\in\Irr(\mathcal{C})} U_s\otimes\bar U_s.
$$
We claim that the half-braiding $c$ for $Z_{\mathrm{reg}}=\oplus_s U_s\otimes\bar U_s$ constructed in Section~\ref{subsec:half-br} defines a unitary half-braiding $c_\phi$ for $Z_\phi$. For objects $X\in\mathcal{C}$, we have natural unitary isomorphisms
$$
X\otimes Z_\phi\cong(\iota_X\otimes A^\phi)\text{-}\oplus_s X\otimes U_s\otimes\bar U_s,\ \
Z\otimes X_\phi\cong(A^\phi\otimes\iota_X)\text{-}\oplus_s U_s\otimes\bar U_s\otimes X.
$$
Therefore in order to show that $c_X$ defines a unitary $c_{\phi,X}\colon X\otimes Z_\phi\to Z_\phi\otimes X$, by Lemma~\ref{lem:indunitary} it suffices to show that
\begin{equation} \label{eq:intertwin}
\sum_s c_{X,ps}(\iota_X\otimes A^\phi_{st})=\sum_q (A^\phi_{pq}\otimes\iota_X)c_{X,qt}.
\end{equation}
As in Section~\ref{subsec:half-br}, for all $q$ and $t$ choose a maximal family of isometries $u^\alpha_{tq}\colon U_q\to X\otimes U_t$ with mutually orthogonal ranges. Then the right hand side of \eqref{eq:intertwin} equals
\begin{align*}
&\sum_{\alpha,q} d_p^{1/2}d_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes \bar R^*_q\otimes\iota_X)(\iota_p\otimes M^\phi_{\bar U_p\otimes U_q}\otimes\iota_{\bar q}\otimes\iota_X)(\bar R_p\otimes\iota_q\otimes\iota_{\bar q}\otimes\iota_X) \\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad (u^{\alpha*}_{tq}\otimes u^{\alpha\vee}_{tq}\otimes\iota_X)(\iota_X\otimes\iota_t\otimes\iota_{\bar t}\otimes R_X)\\
&=\sum_{\alpha,q} d_p^{1/2}d_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes \bar R^*_q\otimes\iota_X)
(\iota_p\otimes\iota_{\bar p}\otimes u^{\alpha*}_{tq}\otimes u^{\alpha\vee}_{tq}\otimes\iota_X)(\iota_p\otimes M^\phi_{\bar U_p\otimes X\otimes U_t}\otimes\iota_{\bar t}\otimes\iota_{\bar X}\otimes\iota_X)\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad (\bar R_p\otimes\iota_X\otimes\iota_t\otimes\iota_{\bar t}\otimes R_X).
\end{align*}
Now observe that by definition of $u^{\alpha\vee}_{tq}$ we have
$$
\sum_\alpha \bar R^*_q(u^{\alpha*}_{tq}\otimes u^{\alpha\vee}_{tq})=\sum_\alpha\bar R^*_{X\otimes U_t}(u^{\alpha}_{tq}u^{\alpha*}_{tq}\otimes\iota_{\bar t}\otimes\iota_{\bar X})
=\bar R^*_{X\otimes U_t}(p^{X\otimes U_t}_q\otimes\iota_{\bar t}\otimes\iota_{\bar X}),
$$
where $p^U_q$ denotes the projection onto the isotypic component of $U$ corresponding to $U_q$. Taking the summation over $q$ we conclude that the right hand side of \eqref{eq:intertwin} equals
$$
d_p^{1/2}d_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes \bar R^*_{X\otimes U_t}\otimes\iota_X)(\iota_p\otimes M^\phi_{\bar U_p\otimes X\otimes U_t}\otimes\iota_{\bar t}\otimes\iota_{\bar X}\otimes\iota_X)(\bar R_p\otimes\iota_X\otimes\iota_t\otimes\iota_{\bar t}\otimes R_X).
$$
Recalling that $\bar R_{X\otimes U_t}=(\iota_X\otimes \bar R_t\otimes\iota_{\bar X})\bar R_X$, we see that this expression equals
\begin{equation}\label{eq:intertwin1}
d_p^{1/2}d_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes \bar R^*_t)(\iota_p\otimes M^\phi_{\bar U_p\otimes X\otimes U_t}\otimes\iota_{\bar t})(\bar R_p\otimes\iota_X\otimes\iota_t\otimes\iota_{\bar t}).
\end{equation}
Note that in the particular case of $\phi=\delta_e$, when $A_{st}=\delta_{st}\iota$ by Example~\ref{ex:deltae}, the equality of \eqref{eq:intertwin1} to the right hand side of \eqref{eq:intertwin} gives the identity
\begin{equation}\label{eq:regbr}
c_{X,pt}=d_p^{1/2}d_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes \bar R^*_t)(\iota_p\otimes p^{\bar U_p\otimes X\otimes U_t}_e\otimes\iota_{\bar t})(\bar R_p\otimes\iota_X\otimes\iota_t\otimes\iota_{\bar t}).
\end{equation}
Now, using this identity we see that the left hand side of \eqref{eq:intertwin} equals
\begin{align*}
&\sum_{s} d_p^{1/2}d_sd_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes \bar R^*_s)(\iota_p\otimes p^{\bar U_p\otimes X\otimes U_s}_e\otimes\iota_{\bar s})(\bar R_p\otimes\iota_X\otimes\iota_s\otimes\iota_{\bar s})\\
&\qquad\qquad\qquad\qquad
(\iota_X\otimes\iota_s\otimes\iota_{\bar s}\otimes \bar R^*_t)(\iota_X\otimes\iota_s\otimes M^\phi_{\bar U_s\otimes U_t}\otimes\iota_{\bar t})(\iota_X\otimes\bar R_s\otimes\iota_t\otimes\iota_{\bar t})\\
&=\sum_s d_p^{1/2}d_sd_t^{1/2}(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes \bar R^*_t)
(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes\bar R^*_s\otimes\iota_t\otimes\iota_{\bar t})
(\iota_p\otimes p^{\bar U_p\otimes X\otimes U_s}_e\otimes\iota_{\bar s}\otimes\iota_t\otimes\iota_{\bar t})\\
&\qquad\qquad\qquad\qquad
(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes\iota_s\otimes M^\phi_{\bar U_s\otimes U_t}\otimes\iota_{\bar t})
(\iota_p\otimes\iota_{\bar p}\otimes\iota_X\otimes\bar R_s\otimes\iota_t\otimes\iota_{\bar t})
(\bar R_p\otimes\iota_X\otimes\iota_t\otimes\iota_{\bar t}).
\end{align*}
In order to show that this expression equals \eqref{eq:intertwin1} it suffices to show that
$$
M^\phi_{\bar U_p\otimes X\otimes U_t}=\sum_s d_s(\iota_{\bar p}\otimes\iota_X\otimes\bar R^*_s\otimes\iota_t)
(p^{\bar U_p\otimes X\otimes U_s}_e\otimes\iota_{\bar s}\otimes\iota_t)(\iota_{\bar p}\otimes\iota_X\otimes\iota_s\otimes M^\phi_{\bar U_s\otimes U_t})
(\iota_{\bar p}\otimes\iota_X\otimes\bar R_s\otimes\iota_t).
$$
By naturality of $M^\phi$ we can rewrite the right hand side as
\begin{multline*}
\sum_s d_s(\iota_{\bar p}\otimes\iota_X\otimes\bar R^*_s\otimes\iota_t)
(p^{\bar U_p\otimes X\otimes U_s}_e\otimes\iota_{\bar s}\otimes\iota_t)M^\phi_{\bar U_p\otimes X\otimes U_s\otimes\bar U_s\otimes U_t}
(\iota_{\bar p}\otimes\iota_X\otimes\bar R_s\otimes\iota_t)\\
=M^\phi_{\bar U_p\otimes X\otimes U_t}\sum_s d_s(\iota_{\bar p}\otimes\iota_X\otimes\bar R^*_s\otimes\iota_t)
(p^{\bar U_p\otimes X\otimes U_s}_e\otimes\iota_{\bar s}\otimes\iota_t)
(\iota_{\bar p}\otimes\iota_X\otimes\bar R_s\otimes\iota_t).
\end{multline*}
Therefore it remains to check that
$$
\sum_s d_s(\iota_{\bar p}\otimes\iota_X\otimes\Tr_s)(p^{\bar U_p\otimes X\otimes U_s}_e)=\iota_{\bar p}\otimes\iota_X.
$$
But this is clearly true, since for any $q$ we have
$$
\sum_s d_s(\iota_q\otimes\Tr_s)(p^{U_q\otimes U_s}_e)=d_q(\iota_q\otimes\Tr_{\bar q})(p^{U_q\otimes\bar U_q}_e)=\iota_q.
$$
Thus we have proved that both sides of \eqref{eq:intertwin} are equal to \eqref{eq:intertwin1}. This completes the construction of the unitary~$c_{\phi,X}$. Naturality of this construction and the half-braiding condition easily follow from the corresponding properties of $c$.
Consider the representation $\pi_\phi$ of $C^*(\mathcal{C})$ defined by $(Z_\phi,c_\phi)$. Denote by $\xi_\phi$ the canonical morphism $\mathds{1}=U_e\otimes\bar U_e\to Z_\phi=A^\phi\text{-}\oplus_s U_s\otimes\bar U_s$. Then
$$
\phi(s)=d_s^{-1}(\pi_\phi([U_s])\xi_\phi,\xi_\phi).
$$
Indeed, by Lemma~\ref{lem:Mphi} we have
$$
d_s^{-1}(\pi_\phi([U_s])\xi_\phi,\xi_\phi)\iota_s=(\xi_\phi^*\otimes\iota_s)c_{\phi,s}(\iota_s\otimes\xi_\phi).
$$
By \eqref{eq:newscalarpr1} the last expression equals
$
\sum_t(A^\phi_{et}\otimes\iota_s)c_{s,te}.
$
Recalling that $c_{s,te}=\delta_{st}d_s^{-1/2}(\iota_s\otimes R_s)$ by \eqref{eq:matrixte} and using that $A^\phi_{et}=d_t^{1/2}\phi(t)\bar R^*_t$, we get
$$
\sum_t(A^\phi_{et}\otimes\iota_s)c_{s,te}=\phi(s)\iota_s,
$$
and the proof of the theorem is complete.
\end{proof}
The triples $(Z_\phi,c_\phi,\xi_\phi)$ constructed in the proof of Theorem~\ref{thm:positivedef} have the following universal property.
\begin{proposition}\label{prop:Zphi-universal}
Let $(Z,c)$ be an object in ${\Dcen(\indC)}$ and $\xi\in\Mor_\indcat{\CC}(\mathds{1},Z)$. Consider the positive definite function $\phi(s)=d_s^{-1}(\pi_Z([U_s])\xi,\xi)$. Then there exists a unique isometric morphism $T\colon (Z_\phi,c_\phi)\to (Z,c)$ in ${\Dcen(\indC)}$ such that $T\xi_\phi=\xi$.
\end{proposition}
\begin{proof} Let us first prove the uniqueness. Assume $T\colon (Z_\phi,c_\phi)\to (Z,c)$ is a morphism. Denote by $T_s$ the composition of $T$ with the canonical morphism $U_s\otimes\bar U_s\to Z_\phi$, so in particular we have $T_e=T\xi_\phi$. Clearly, the morphism $T$ is completely determined by the morphisms $T_s$, so we just have to check that $T_s$ is determined by $T_e$. By Lemma~\ref{lem:indunitary} and formula~\eqref{eq:matrixte} for the half-braiding on $Z_{\mathrm{reg}}$, we have the commutative diagram
$$
\begin{xymatrix}{
U_s\ar[d]_{\iota_s\otimes\xi_\phi}\ar[rr]^{d_s^{-1/2}(\iota_s\otimes R_s)\qquad\ \ } & & (U_s\otimes\bar U_s)\otimes U_s\ar[d]\\
U_s\otimes Z_\phi\ar[rr]_{c_{\phi,s}}& & Z_\phi\otimes U_s.
}\end{xymatrix}
$$
Applying $T$ and using that $c_s(\iota_s\otimes T)=(T\otimes\iota_s)c_{\phi,s}$, we get $c_s(\iota_s\otimes T_e)=d_s^{-1/2}(T_s\otimes\iota_s)(\iota_s\otimes R_s)$, that is,
\begin{equation} \label{eq:intertbr}
T_s = d_s^{1/2} (\iota_Z \otimes \bar{R}_s^*) ( c_s \otimes \iota_{\bar s}) (\iota_s \otimes T_e \otimes \iota_{\bar s}).
\end{equation}
Thus $T_s$ is indeed determined by $T_e$.
\smallskip
For the existence, we let $T_e=\xi$ and define $T_s\colon U_s\otimes\bar U_s\to Z$ by \eqref{eq:intertbr}. In order to show that the morphisms~$T_s$ define an isometry $T\colon Z_\phi\to Z$, by Lemma~\ref{lem:indiso} it suffices to check that for any finite set $F\subset\Irr(\mathcal{C})$ for the morphism $T_F=(T_s)_{s\in F}\colon\oplus_{s\in F}U_s\otimes\bar U_s\to Z$ we have $T_F^*T_F=(A^\phi_{s,t})_{s,t\in F}$, that is, $T_s^*T_t=A^\phi_{st}$. But this is exactly the computation we made in the proof of the implication (ii)$\Rightarrow$(i) in Theorem~\ref{thm:positivedef}.
\smallskip
It remains to check that $T$ intertwines the half-braidings. By the construction of $T$ we already have
$$
c_X(\iota_X\otimes T)(\iota_X\otimes\xi_\phi)=(T\otimes\iota_X)c_{\phi,X}(\iota_X\otimes\xi_\phi)
$$
for $X=U_s$, hence for all $X\in\mathcal{C}$. Applying this to $X\otimes U_s$ in place of $X$ and using the multiplicativity property of half-braidings, we get
\begin{equation} \label{eq:intertwin2}
(c_X\otimes\iota_s)(\iota_X\otimes T\otimes\iota_s)(\iota_X\otimes B_s)
=(T\otimes\iota_X\otimes\iota_s)(c_{\phi,X}\otimes\iota_s)(\iota_X\otimes B_s),
\end{equation}
where $B_s=c_{\phi,s}(\iota_s\otimes\xi_\phi)\colon U_s\to Z_\phi\otimes U_s$. But by equation~\eqref{eq:intertbr} for the identity map on $Z_\phi$ we know that
$$
d_s^{1/2} (\iota_{Z_\phi} \otimes \bar{R}_s^*) ( B_s \otimes \iota_{\bar s})
$$
is the canonical morphism $U_s\otimes\bar U_s\to Z_\phi$. Since \eqref{eq:intertwin2} holds for all $s$, we can therefore conclude that $c_X(\iota_X\otimes T)=(T\otimes\iota_X)c_{\phi,X}$.
\end{proof}
From the proof we also get the following.
\begin{corollary} \label{cor:cyclic}
The vector $\xi_\phi$ is cyclic for the representation $\pi_\phi$ of $C^*(\mathcal{C})$ defined by $(Z_\phi,c_\phi)$.
\end{corollary}
\begin{proof} Let $\xi_{\phi,s}$ be the morphism $\mathds{1}\to Z_\phi$ obtained by composing $d_s^{-1/2}\bar R_s\colon \mathds{1}\to U_s\otimes\bar U_s$ with the canonical morphism $U_s\otimes\bar U_s\to Z_\phi$. From the construction of $Z_\phi$ one can see that the vectors $\xi_{\phi,s}$, $s\in\Irr(\mathcal{C})$, span a dense subspace of $\Mor_{\indcat{\CC}}(\mathds{1},Z_\phi)$. On the other hand, from equality~\eqref{eq:intertbr} for the identity morphism $T$ on~$Z_\phi$ we have
$$
\xi_{\phi,s}=d_s^{-1/2}T_s\bar R_s=(\iota_Z \otimes \bar{R}_s^*) ( c_s \otimes \iota_{\bar s}) (\iota_s \otimes \xi_\phi \otimes \iota_{\bar s})\bar R_s=\pi_\phi([U_s])\xi_\phi.
$$
Hence the vector $\xi_\phi$ is indeed cyclic.
\end{proof}
By decomposing representations of $C^*(\mathcal{C})$ into direct sums of cyclic representations, we now obtain the following result.
\begin{corollary} \label{cor:tworepclasses}
Any representation of $C^*(\mathcal{C})$ is unitarily equivalent to the representation $\pi_Z$ defined by an object $(Z,c)\in{\Dcen(\indC)}$.
\end{corollary}
This result can also be formulated as follows. Let us say that an object $(Z,c)\in{\Dcen(\indC)}$ is \emph{spherical} if for any $T\in\End_{\Dcen(\indC)}((Z,c))$ such that $T\xi=0$ for all $\xi\in\Mor_\indcat{\CC}(\mathds{1},Z)$ we have $T=0$. Such objects form a full C$^*$-subcategory ${\Dcen_s(\indC)}$ of ${\Dcen(\indC)}$ closed under direct sums and subobjects. Note that in general this is not a tensor category.
\begin{proposition}
The category ${\Dcen_s(\indC)}$ is unitarily equivalent to the representation category of $C^*(\mathcal{C})$.
\end{proposition}
\begin{proof}
Consider the unitary functor $F\colon{\Dcen_s(\indC)}\to\Rep C^*(\mathcal{C})$ mapping an object $(Z,c)$ into the corresponding representation $\pi_{(Z,c)}$ of $C^*(\mathcal{C})$. By the definition of spherical objects this functor is faithful. Since the objects $(Z_\phi,c_\phi)$ are spherical by Proposition~\ref{prop:Zphi-universal}, this functor is also essentially surjective. It remains to show that it is full. By Proposition~\ref{prop:Zphi-universal}, any spherical object decomposes into a direct sum of objects $(Z_\phi,c_\phi)$. Therefore it suffices to show that any bounded operator $T\colon\Mor_\indcat{\CC}(\mathds{1},Z_\phi)\to \Mor_\indcat{\CC}(\mathds{1},Z)$ intertwining the representations $\pi_\phi$ and $\pi_{(Z,c)}$ is defined by a morphism $(Z_\phi,c_\phi)\to (Z,c)$ in ${\Dcen(\indC)}$. If~$T$~is isometric, this is true, again by Proposition~\ref{prop:Zphi-universal}. The general case follows from this, since any contraction $T\colon H\to K$ between Hilbert spaces can be dilated to an isometry $H\to K\oplus H$, $\xi\mapsto (T\xi,(1-T^*T)^{1/2}\xi)$.
\end{proof}
\begin{example}
Consider the regular representation of $C^*(\mathcal{C})$ on $\ell^2(\Irr(\mathcal{C}))$. Denote by $W^*(\mathcal{C})$ the von Neumann algebra generated by $C^*(\mathcal{C})$ in this representation. Since the vector $\delta_e$ is a cyclic trace vector for $W^*(\mathcal{C})$, the commutant $W^*(\mathcal{C})'$ is anti-isomorphic to $W^*(\mathcal{C})$. On the other hand, as we essentially observed in Example~\ref{ex:regrep}, the regular representation corresponds to $Z_{\mathrm{reg}}(\mathcal{C})$ under the equivalence $\Rep C^*(\mathcal{C})\cong{\Dcen_s(\indC)}$. Hence $\End_{{\Dcen(\indC)}}(Z_{\mathrm{reg}}(\mathcal{C}))\cong W^*(\mathcal{C})^\mathrm{op}$.
\end{example}
\subsection{Property (T)}
\label{sec:property-t}
With the algebra $C^*(\mathcal{C})$ at our disposal, the following definition is very natural.
\begin{definition}
\label{defn:cat-prop-T}
We say that \emph{$\mathcal{C}$ has property (T)} if any representation of $C^*(\mathcal{C})$ which weakly contains the trivial representation $\pi_\mathds{1}$, contains $\pi_\mathds{1}$ as a subrepresentation.
\end{definition}
Given a representation $\pi$ of $C^*(\mathcal{C})$ on a Hilbert space $H$, let us say that a vector $\xi\in H$ is \emph{invariant} if $\pi([X])\xi=d(X)\xi$ for all $X\in\mathcal{C}$. Any nonzero invariant vector gives an embedding of $\pi_\mathds{1}$ into~$\pi$. More generally, let us say that unit vectors $\xi_i\in H$, indexed by a directed set $I$, are \emph{almost invariant} if $\lim_i(\pi([X])\xi_i,\xi_i)=d(X)$ for all $X\in\mathcal{C}$.
Since $\|\pi([X])\|\le d(X)$, in this terminology the above definition of property (T) means that if a representation $\pi\colon C^*(\mathcal{C})\to B(H)$ has almost invariant vectors, then it has nonzero invariant vectors.
Almost invariance can be phrased in different ways.
\begin{lemma}\label{lem:aa}
Let $\pi=\pi_{(Z,c)}$ be the representation defined by an object $(Z,c)\in{\Dcen(\indC)}$ and $\{\xi_i\}_i$ be a net of unit vectors in $\Mor_\indcat{\CC}(\mathds{1},Z)$. Then the following conditions are equivalent:
\begin{itemize}
\item[(i)] the vectors $\xi_i$ are almost invariant;
\item[(ii)] we have $\lim_i\|\pi([X])\xi_i-d(X)\xi_i\|=0$ for all objects $X\in\mathcal{C}$;
\item[(iii)] we have $\lim_i\norm{c_X (\iota_X \otimes {\xi}_i) - \xi_i \otimes \iota_X} = 0$ for all objects $X\in\mathcal{C}$.
\end{itemize}
\end{lemma}
\begin{proof} The equivalence of (i) and (ii) is immediate from $\norm{\pi([X])}\le d(X)$. By faithfulness of $\Tr_X$ on $\mathcal{C}(X)$, condition (iii) is equivalent~to
$$
\Tr_X((c_X (\iota_X \otimes {\xi}_i) - \xi_i \otimes \iota_X)^* (c_X (\iota_X \otimes {\xi}_i) - \xi_i \otimes \iota_X)) \to 0.
$$
Expanding the product inside $\Tr_X$ and using that $(\pi([X])\xi_i,\xi_i)=\Tr_X((\xi^*_i \otimes \iota_X) c_X (\iota_X \otimes \xi_i))$ by the proof of Lemma~\ref{lem:char-alg-act-compat-adj}, we have
$$
\Tr_X((c_X (\iota_X \otimes {\xi}_i) - \xi_i \otimes \iota_X)^* (c_X (\iota_X \otimes {\xi}_i) - \xi_i \otimes \iota_X)) = 2 d(X) - 2 \Re (\pi([X]) \xi_i, \xi_i).
$$
Using once again that $\norm{\pi([X])}\le d(X)$ we conclude that (iii) holds if and only if $(\pi([X]) \xi_i, \xi_i)\to d(X)$, so
(iii) is equivalent to (i).
\end{proof}
As in the group case, there are many equivalent ways of formulating property (T). Let us list some of them.
\begin{proposition} \label{prop:propertyT}
The following conditions are equivalent:
\begin{itemize}
\item[(i)] the category $\mathcal{C}$ has property (T);
\item[(ii)] there exist a finite set $F\subset \Irr(\mathcal{C})$ and $\varepsilon>0$ such that if $\pi\colon C^*(\mathcal{C})\to B(H)$ is a representation and $\xi\in H$ is a unit vector such that $|(\pi([U_s])\xi,\xi)- d_s|<\varepsilon$ for all $s\in F$, then there exists an invariant unit vector in $H$;
\item[(iii)] there exists a nonzero projection $p\in C^*(\mathcal{C})$ such that $[X]p=d(X)p$ for all $X\in\mathcal{C}$.
\end{itemize}
\end{proposition}
\begin{proof}
(i)$\Rightarrow$(iii) This can be proved in the same way as the existence of Kazhdan projections, see e.g.~\cite{MR2391387}*{Section~17.2}. Consider any family of representations $\{\pi_\lambda\}_\lambda$ of $C^*(\mathcal{C})$ without nonzero invariant vectors such that the representation $\pi_\mathds{1}\oplus(\oplus_\lambda\pi_\lambda)$ is faithful. For every $\lambda$ take countably many copies of $\pi_\lambda$ and denote by~$\sigma$ the direct sum of all these representations for all $\lambda$. We claim that $\sigma$ is not faithful. Assume this is not the case. Then for any separable C$^*$-subalgebra $A\subset C^*(\mathcal{C})$ the representation $\sigma|_A$ is faithful and essential, so by Voiculescu's theorem it weakly contains any other representation. Since this is true for any~$A$, we conclude that $\pi_\mathds{1}$ is weakly contained in $\sigma$. But then $\sigma$ must have nonzero invariant vectors, which is a contradiction. Therefore $J=\ker\sigma\ne0$. Since $\pi_\mathds{1}\oplus\sigma$ is faithful, it follows that $J=\mathbb{C} p$ for a projection $p$. Clearly, $p$ has the required property.
\smallskip
(iii)$\Rightarrow$(ii) Since $[X]p=d(X)p$ for all $X\in\mathcal{C}$, we have $ap=\pi_\mathds{1}(a)p$ for all $a\in C^*(\mathcal{C})$, and letting $a=p$ we get $\pi_\mathds{1}(p)=1$. Now choose an element $x=\sum_{s\in F}\alpha_s[U_s]$, with $F\subset\Irr(\mathcal{C})$ finite, such that $\|p-x\|<1/2$. Then $|1-\sum_{s\in F}\alpha_s d_s|<1/2$. We can find $\varepsilon>0$ such that whenever $|\beta_s-d_s|<\varepsilon$ for all $s\in F$, we still have $|1-\sum_{s\in F}\alpha_s \beta_s|<1/2$. Then, assuming that $|(\pi([U_s])\xi,\xi)- d_s|<\varepsilon$ for all $s\in F$, we get a nonzero invariant vector $\pi(p)\xi$. Thus the pair $(F,\varepsilon)$ has the required property.
\smallskip
The implication (ii)$\Rightarrow$(i) is obvious.
\end{proof}
Property (T) for rigid C$^*$-tensor categories has been also introduced by Popa and Vaes~\cite{MR3406647}.
\begin{corollary} \label{cor:twodefT}
Definition \ref{defn:cat-prop-T} of property (T) is equivalent to the definition of Popa and Vaes.
\end{corollary}
\begin{proof}
This follows from Corollary~\ref{cor:twodef}, Proposition~\ref{prop:propertyT} and~\cite{MR3406647}*{Proposition~5.5}.
\end{proof}
\begin{remarks}\mbox{\ }
\noindent
(i) If $G$ is a compact quantum group, then by Remark~\ref{rmk:q-grp-char-corr} the category $\Rep G$ has property (T) if and only if the dual discrete quantum group $\hat G$ has central property (T) in the sense of~\cite{arXiv:1410.6238}. Together with Corollary~\ref{cor:twodefT} this gives an alternative proof of~\cite{MR3406647}*{Proposition~6.3}.
\smallskip\noindent
(ii) Recall that by Corollary~\ref{cor:tworepclasses} any representation of $C^*(\mathcal{C})$ is equivalent to a representation of the form $\pi_Z$. If we did not know this, we would have the dilemma of defining property (T) using either all representations or only representations $\pi_Z$. It is, however, not difficult to see that these two approaches are equivalent independently of the results of Section~\ref{sec:positivedefinite}. The key point is that the implication (i)$\Rightarrow$(iii) in Proposition~\ref{prop:propertyT} remains true if we define property (T) using only representations~$\pi_Z$. In order to see this, we have to use representations $\pi_Z$ in the proof, and for this we have to be able to split any representation~$\pi_Z$ into a direct sum of copies of $\pi_\mathds{1}$ and a representation of the same form without invariant vectors. For this, in turn, we have to show that if $(Z, c)\in{\Dcen(\indC)}$ and $\xi$ is an invariant unit vector in $\Mor_\indcat{\CC}(\mathds{1}, Z)$, then $\xi$ is a morphism in ${\Dcen(\indC)}$, that is, $c_X(\iota\otimes\xi_X)=\xi\otimes\iota_X$ for all $X$. But this is true by Lemma~\ref{lem:aa}. This was our initial approach before the appearance of~\cite{MR3406647}.
\end{remarks}
\bigskip
\section{Categories of Hilbert bimodules}
In this section we give an interpretation of our results and constructions in terms of Hilbert bimodules. Throughout the whole section $M$ denotes a fixed II$_1$-factor.
\subsection{Duality for Hilbert bimodules} \label{sec:Hilbdual}
Let us briefly review a few basic facts from the theory of Hilbert modules, see e.g.~\citelist{\cite{MR1245354}\cite{MR1642584}} for more details.
Denote by $\tau$ the unique tracial state on $M$. Let $X$ be a Hilbert $M$-bimodule, that is, a Hilbert space together with two commuting normal unital representations of $M$ and $M^\mathrm{op}$ on $X$, where $M^\mathrm{op}$ is the factor~$M$ with the opposite product. Denote by $\dim(X_M)$ the Murray--von Neumann dimension of $X$ considered as a right $M$-module, so if $X_M\cong pL^2(M)^n$ for a projection $p\in\Mat_n(M)$, then $\dim(X_M)=(\Tr\otimes\tau)(p)$, and if no such $p$ and $n\in\mathbb N$ exist, then $\dim(X_M)=\infty$ . We can similarly define $\dim({}_MX)$.
For a Hilbert $M$-bimodule $X$, consider the subspace $X^0\subset X$ of left bounded vectors, that is, vectors $\xi \in X$ satisfying $(\xi x, \xi) \le c \tau(x)$ for some $c \ge 0$ and all $x\in M_+$. Then $X^0$ is a sub-bimodule, and it admits a unique $M$-valued inner product $\langle\xi, \eta\rangle_M$ (antilinear in $\xi$, linear in $\eta$) satisfying $\langle\xi, \eta x\rangle_M = \langle\xi, \eta\rangle_M x$ and $\tau(\langle\xi, \eta\rangle_M)=(\eta,\xi)$. Then, given another Hilbert $M$-bimodule $Y$, the tensor product $X\otimes_MY$ is defined as the tensor product $X^0\otimes_M Y$ in the sense of Hilbert C$^*$-modules. This way the category $\Hilb_M$ of Hilbert $M$-bimodules becomes a C$^*$-tensor category.
Let us now describe the duality in $\Hilb_M$. Assume $X$ is a Hilbert $M$-bimodule such that both $\dim(X_M)$ and $\dim({_M}X)$ are finite. In this case the spaces of left and right bounded vectors in $X$ coincide and therefore~$X^0$ carries also the structure of a left Hilbert C$^*$-module over $M$, so it is equipped with an $M$-valued inner product $_M\langle\xi, \eta\rangle$ which is linear in $\xi$ and antilinear in $\eta$. The complex conjugate space $\bar{X}$ is also an $M$-bimodule such that $x \bar{\xi} y = \overline{y^* \xi x^*}$, and the inner products are related by $\langle\bar{\xi}, \bar{\eta}\rangle_M = {}_M\langle\xi, \eta\rangle$. We can choose a basis $\{\rho_i\}^n_{i=1} \subset X^0$ of $X$ as a right $M$-module, in the sense that $\xi = \sum_i \rho_i \langle\rho_i, \xi\rangle_M$ for all $\xi \in X^0$. The formula
$$
\Tr^r_X(T)=\sum_i\tau(\langle \rho_i,T\rho_i\rangle_M)=\sum_i(T\rho_i,\rho_i)
$$
defines a trace on $\End(X_M^0)$, and we have $\Tr^r_X(1)=\dim(X_M)$. Similarly, we have a trace $\Tr^l_X$ on $\End({}_MX^0)$ such that
$$
\Tr^l_X(T)=\sum_j(T\lambda_j,\lambda_j)
$$
for any basis $\{\lambda_j\}^m_{j=1}$ of the left $M$-module $X$.
There exists a unique Hilbert $M$-bimodule map $\bar R_X\colon L^2(M) \to X \otimes_M \bar{X}$ such that $\bar R_X(1)=\sum_i \rho_i \otimes \bar{\rho}_i $. The map $\bar R_X$ does not depend on the choice of $\{\rho_i\}^n_{i=1}$ and the adjoint map is given by $\bar R_X^*(\xi\otimes\bar\eta)={}_M\langle \xi,\eta\rangle$. It follows that
$$
\|\bar R_X\|^2=\bar R^*_X\bar R_X(1)=\sum_i(\rho_i,\rho_i)=\dim(X_M).
$$
Consider now $R_X=\bar R_{\bar X}\colon L^2(M)\to \bar X\otimes_M X$, so $R_X(1)=\sum_j \bar\lambda_j\otimes\lambda_j$. We have
$$
R^*_X(\bar\xi\otimes\eta)=\langle\xi,\eta\rangle_M\ \ \text{and}\ \ \|R_X\|^2=\dim(\bar X_M)=\dim({}_MX).
$$
From this we see that $(\iota\otimes R^*_X)(\bar R_X\otimes\iota)=\iota$ and conclude that $(R_X,\bar R_X)$ is a solution of the conjugate equations for $X$. Thus $\bar X$ is dual to $X$ and
$$
d(X)\le \sqrt{\dim({}_M X) \dim(X_M)}.
$$
In general this is a strict inequality and the solution $(R_X,\bar R_X)$ is not standard. The general criterion of standardness $\bar R^*_X(T\otimes\iota)\bar R_X=R_X^*(\iota\otimes T)R_X$ becomes
\begin{equation} \label{eq:eqtrace}
\Tr^r_X=\Tr^l_X\ \ \text{on}\ \ \End_{M\mhyph M}(X),
\end{equation}
in which case we also have $d(X)=\dim({}_M X)=\dim(X_M)$.
\subsection{Drinfeld center and Longo--Rehren construction}\label{subsec:LR}
Let $\mathcal{C}$ be a rigid full C$^*$-tensor subcategory of~$\Hilb_M$ such that condition~\eqref{eq:eqtrace} holds for all $X\in\mathcal{C}$. As usual, we also assume that $\mathcal{C}$ is closed under finite direct sums and subobjects. We will use the solutions $(R_X,\bar R_X)$ of conjugate equations defined in Section~\ref{sec:Hilbdual}. Note that since these solutions, as well as the left and right bases of Hilbert modules~\cite{MR1642584}*{Proposition~9.62}, behave well with respect to direct sums and tensor products, it suffices to check~\eqref{eq:eqtrace} on a set of bimodules generating the C$^*$-tensor category $\mathcal{C}$ by taking direct sums, tensor products and subobjects.
\begin{example} \label{ex:subfactorcat}
Assume $N\subset M$ is an extremal finite index subfactor. Consider the corresponding Jones tower $N\subset M\subset M_1\subset\dots$. Let $\mathcal{C}=\mathcal{C}_{N\subset M}$ be the full C$^*$-tensor subcategory of $\Hilb_M$ generated by the module~$L^2(M_1)$. Then~$\mathcal{C}$ satisfies the above assumptions. Indeed, since the Hilbert $M$-bimodule $L^2(M_1)$ is self-dual, the category $\mathcal{C}$ is rigid. Next, we have a canonical isomorphism $\End({}_ML^2(M_1)_M)\cong M'\cap M_{2}$. Under this isomorphism the traces $\Tr^r$ and $\Tr^l$ on $\End({}_ML^2(M_1)_M)$ introduced in Section~\ref{sec:Hilbdual} coincide, up to the factor $[M:N]$, with the restriction of the tracial states on $M_{2}$ and $M'\subset B(L^2(M_1))$, respectively, to $M'\cap M_{2}$. The extremality assumption implies that these traces are equal, so the condition~\eqref{eq:eqtrace} holds for all $X\in\mathcal{C}$.
In fact, any finitely generated rigid C$^*$-tensor category $\mathcal{C}$ (satisfying our standard assumptions) is unitarily monoidally equivalent to a category of the form $\mathcal{C}_{N\subset M}$ for some $M$ and $N$. More precisely, taking an object $X\in\mathcal{C}$ such that $X\otimes\bar X$ is a generating object, we have a standard $\lambda$-lattice consisting of the algebras $\mathcal{C}(X \otimes \bar{X} \otimes \cdots)$ ($n$ factors) and $\mathcal{C}(\bar{X} \otimes X \otimes \cdots)$ ($n-1$ factors), with the Jones projections given by copies of $\frac{1}{d(X)} R_X R_X^*$ and $\frac{1}{d(X)} \bar{R}_X \bar{R}_X^*$. Then a result of Popa~\cite{MR1334479} gives an extremal finite index subfactor $N \subset M$ such that $\mathcal{C}_{N\subset M}$ is unitarily monoidally equivalent to $\mathcal{C}$, and such that under this equivalence the module $L^2(M_1)$ corresponds to $X \otimes \bar{X}$.
\end{example}
With every category $\mathcal{C}$ as above one can associate an inclusion $A\subset B$ of II$_1$-factors, called the Longo-Rehren inclusion~\citelist{\cite{MR1332979} \cite{MR1742858}}. Namely,
put $A=M\bar\otimes M^\mathrm{op}$. Recall that there is a functorial construction of a Hilbert $M^\mathrm{op}$-bimodule $X^\natural$ from a Hilbert $M$-bimodule $M$: as a linear space we put $X^\natural = \bar{X}$, and then define the bimodule structure by $x\bar\xi=\overline{x^*\xi}$ and $\bar\xi x=\overline{\xi x^*}$. With this definition we have a natural identification of $(X \otimes_M Y)^\natural$ with $ X^\natural \otimes_{M^\mathrm{op}} Y^\natural$. Thus, $X \mapsto X^\natural$ is a monoidal functor which is antilinear on morphisms.
Now, choose a complete system of representatives of isomorphism classes of simple modules $\{X_s\}_{s \in I}$ in $\mathcal{C}$. Then $ B$ is generated by the spaces $X_s^0 \otimes X_s^{0\natural}$, with the product
\begin{equation}\label{eq:prod-in-tB}
(\xi_1 \otimes \bar \xi_2) \cdot (\eta_1 \otimes \bar \eta_2) = \sum_{\alpha} \left(\frac{d_s d_t}{d_{r_\alpha}}\right)^{{1/2}} w^\alpha_{st}(\xi_1 \otimes \eta_1) \otimes \overline{w^{\alpha}_{st}(\xi_2 \otimes \eta_2)}
\end{equation}
for $\xi_1 \otimes \bar\xi_2 \in X_s^0 \otimes X_s^{0\natural}$ and $\eta_1 \otimes \bar\eta_2 \in X_t^0 \otimes X_t^{0\natural}$,
where $\{w^\alpha_{st}\}_\alpha$ is any family of coisometries $X_s\otimes_MX_t\to X_{r_\alpha}$ defining a decomposition of $X_s\otimes_M X_t$ into simple bimodules. The $*$-structure is defined by
$$
(\xi_1 \otimes \bar\xi_2)^*=\bar\xi_1\otimes \xi_2,
$$
where we identify $\bar X_s^0\otimes\bar X_s^{0\natural}$ with a subspace of $ B$ using the map $\bar\xi_1\otimes \xi_2\mapsto J_s\bar\xi_1\otimes \overline{J_s\bar \xi_2}$, where $J_s$ is any unitary isomorphism of Hilbert $M$-bimodules $\bar X_s\to X_{\bar s}$. If $e\in I$ corresponds to $L^2(M)$, then $A$ is identified with the subalgebra formed by the bounded vectors in $X_e\otimes X_e^\natural$. The projection onto this subalgebra composed with the trace on $A$ defines a tracial state on $ B$; it is worth noting that this is the point where condition~\eqref{eq:eqtrace} is used. By construction $L^2( B)$ decomposes into the direct sum of the simple Hilbert $A$-bimodules $X_s\otimes X_s^\natural$. In particular, $A'\cap B=\C1$ and so $ B$ is a II$_1$-factor.
\begin{remark}
For $\mathcal{C}=\mathcal{C}_{N\subset M}$ as in Example~\ref{ex:subfactorcat}, Masuda~\cite{MR1742858} proved that $ B$ is isomorphic to the symmetric enveloping algebra $M \boxtimes_{e_N} M^\mathrm{op}$ of Popa~\cite{MR1302385}.
\end{remark}
Our goal is to prove the following result.
\begin{theorem} \label{thm:mon-eqv-D-center-B-bimod}
Let $\mathcal{C}\subset\Hilb_M$ be a C$^*$-tensor category as above and $A\subset B$ be the associated Longo--Rehren inclusion. Then ${\Dcen(\indC)}$ is unitarily monoidally equivalent to the full subcategory $\ZHC$ of $\Hilb_ B$ consisting of the Hilbert $ B$-bimodules $X$ such that as a Hilbert $A$-bimodule $X$ decomposes into a direct sum of copies of~$X_s\otimes X_t^\natural$.
\end{theorem}
Note that it is not immediately obvious, but will become clear from the proof, that $\ZHC$ is a tensor category. Let us also remark that the objects of $\ZHC$ can equivalently be characterized as Hilbert bimodules that are generated, as $ B$-bimodules, by $A$-sub-bimodules isomorphic to $X_s\otimes X_t^\natural$. In order to see this, it suffices to show that given a Hilbert $ B$-bimodule $H$ and a copy of $X_s\otimes X_t^\natural$ in ${}_AH_A$, the $ B$-bimodule structure defines bounded maps $L^2( B)\otimes_A(X_s\otimes X_t^\natural)\to H$ and
$(X_s\otimes X_t^\natural)\otimes_AL^2( B)\to H$. This, in turn, follows from the following general result (compare with \cite{MR3406647}*{Lemma~2.8}).
\begin{lemma}\label{lem:modstructure}
Assume that $P\subset Q$ is an irreducible inclusion of II$_1$-factors, $H$ is a Hilbert $Q$-$P$-module, and $X\subset H$ is a Hilbert $P$-sub-bimodule such that $\dim(X_P)$ and $\dim({}_PX)$ are finite. Then the map $Q\otimes X\ni a\otimes\xi\mapsto a\xi$ extends to a bounded map $L^2(Q)\otimes_PX\to H$.
\end{lemma}
\begin{proof} Choose a basis $\{\rho_i\}^n_{i=1}$ of $X_P$ and a basis $\{\lambda_j\}^m_{j=1}$ of ${}_PX$. Define a normal positive linear functional~$\psi$ on~$Q$ by
$$
\psi(a)=\sum_i(a\rho_i,\rho_i).
$$
The standard argument shows that $\psi$ is independent of the choice of $\{\rho_i\}_i$: if $\{\rho'_k\}_k$ is another basis, then
$$
\sum_k(a\rho'_k,\rho'_k)=\sum_{i,k}(a\rho_i\langle\rho_i,\rho'_k\rangle_P,\rho'_k)
=\sum_{i,k}(a\rho_i,\rho'_k\langle\rho'_k,\rho_i\rangle_P)=\sum_i(a\rho_i,\rho_i).
$$
In particular, since for any unitary $u\in P$ we can take the basis $\{u\rho_i\}^n_{i=1}$, we see that $u$ is contained in the centralizer of $\psi$. Since $P'\cap Q=\C1$, we conclude that $\psi$ coincides, up to a scalar factor, with the tracial state~$\tau$ on~$Q$ (e.g.~ because Connes' Radon--Nikodym cocycle $[D\psi : D\tau]_t$ lies in $P'\cap Q$). Thus $\psi=\dim(X_P)\tau$. It follows that for every $i=1,\dots,n$ the map $Q\ni a\mapsto a\rho_i$ extends to a bounded map $T_i\colon L^2(Q)\to H$ satisfying $\norm{T_i} \le \sqrt{\dim(X_P)}$.
Next, for every $j=1,\dots,m$ consider the map $L_j\colon L^2(Q)\to L^2(Q)\otimes_P X$ defined by $L_j\zeta=\zeta\otimes\lambda_j$. This map is bounded, and the adjoint map is given by
$$
L_j^*(\zeta\otimes\xi)=\zeta{}_P\langle\xi,\lambda_j\rangle\ \ \text{for}\ \ \zeta\in L^2(Q),\ \xi\in X^0.
$$
For $a\in Q$ and $\xi\in X^0$ we then have
$$
a\xi=\sum_{i,j}T_iL_j^*(a\otimes\xi)\langle\rho_i,\lambda_j\rangle_P,
$$
which shows that the map $a\otimes\xi\mapsto a\xi$ is bounded on $L^2(Q)\otimes_PX$.
\end{proof}
We will often use the following identities, which are immediate from the definition of $ B$.
\begin{lemma} \label{lem:prodrelations}
If $\bar R_s(1)=\sum_\alpha\rho_{s\alpha}\otimes\bar\rho_{s\alpha}$ and $R_s(1)=\sum_\beta\bar\lambda_{s\beta}\otimes\lambda_{s\beta}$, then for any $\xi,\eta\in X_s^0$ we have the following identities in $ B$:
$$
d_s^{-1}\sum_\alpha (\rho_{s\alpha}\otimes\bar\xi)(\bar\rho_{s\alpha}\otimes\eta)=1\otimes{}_M\langle\eta,\xi\rangle,\ \
d_s^{-1}\sum_\beta (\bar\lambda_{s\beta}\otimes\xi)(\lambda_{s\beta}\otimes\bar\eta)=1\otimes\langle\eta,\xi\rangle_M,
$$
$$
d_s^{-1}\sum_\alpha (\xi\otimes\bar\rho_{s\alpha})(\bar\eta\otimes\rho_{s\alpha})={}_M\langle\xi,\eta\rangle\otimes1,\ \
d_s^{-1}\sum_\beta (\bar\xi\otimes\lambda_{s\beta})(\eta\otimes\bar\lambda_{s\beta})=\langle\xi,\eta\rangle_M\otimes1.
$$
\end{lemma}
Turning to the proof of the theorem, we start by constructing a half-braiding from a Hilbert $ B$-bimodule.
The following observation will play a crucial role.
\begin{proposition}\label{prop:move-Xi-thru-bimod}
For any Hilbert $ B$-bimodule $H$ and any $s \in I$, there are unitary isomorphisms of Hilbert $A$-bimodules
$$
(X_s \otimes L^2(M^\mathrm{op})) \otimes_A H \cong (L^2(M) \otimes \bar{X}_s^\natural) \otimes_A H
\ \ \text{and}\ \ H \otimes_A (X_s \otimes L^2(M^\mathrm{op})) \cong H \otimes_A (L^2(M) \otimes \bar{X}_s^\natural).
$$
\end{proposition}
\begin{proof}
Using the isomorphism $L^2(M^\mathrm{op})\cong L^2(M)^\natural$ and the isometry
$d_s^{-1/2}R_s^\natural\colon L^2(M)^\natural \to \bar{X}_s^\natural \otimes_{M^\mathrm{op}} X_s^\natural$ , we get a map
$$
(X_s \otimes L^2(M^\mathrm{op})) \otimes_A H \to (X_s \otimes (\bar{X}_s^\natural \otimes_{M^\mathrm{op}} X_s^\natural)) \otimes_A H
\cong (L^2(M)\otimes \bar X_s^\natural)\otimes_A(X_s\otimes X_s^\natural)\otimes_A H.
$$
Now, the $ B$-module structure gives us a map $(X^0_s \otimes X_s^{0\natural})\otimes H \to H$, so in combination with the above we get a map
\begin{equation}\label{eq:isomodule1}
(X_s \otimes L^2(M^\mathrm{op})) \otimes_A H\to (L^2(M) \otimes \bar{X}_s^\natural) \otimes_A H,\quad
(\xi\otimes1)\otimes\zeta\mapsto d_s^{-1/2}\sum_\beta(1\otimes\lambda_{s\beta})\otimes (\xi\otimes\bar\lambda_{s\beta})\zeta.
\end{equation}
Since generally the $ B$-module structure does not define a bounded map $L^2(B)\otimes_A H\to H$, we still have to check that the above map is well-defined and isometric. For $\xi,\xi'\in X_s^0$ and $\zeta,\zeta'\in H$ we compute:
\smallskip
$\displaystyle
d_s^{-1}\sum_{\beta,\beta'}\big((1\otimes\lambda_{s\beta})\otimes (\xi\otimes\bar\lambda_{s\beta})\zeta,(1\otimes\lambda_{s\beta'})\otimes (\xi'\otimes\bar\lambda_{s\beta'})\zeta'\big)
$
\begin{align*}
&=d_s^{-1}\sum_{\beta,\beta'}\big((1\otimes{}_M\langle\lambda_{s\beta},\lambda_{s\beta'}\rangle)(\xi\otimes\bar\lambda_{s\beta})\zeta,(\xi'\otimes\bar\lambda_{s\beta'})\zeta'\big)\\
&=d_s^{-1}\sum_{\beta}\big((\xi\otimes\bar\lambda_{s\beta'})\zeta,(\xi'\otimes\bar\lambda_{s\beta'})\zeta'\big)\\
&=d_s^{-1}\sum_{\beta}\big((\bar\xi'\otimes\lambda_{s\beta'})(\xi\otimes\bar\lambda_{s\beta'})\zeta,\zeta'\big)\\
&=\big((\langle\xi',\xi\rangle_M\otimes1)\zeta,\zeta'\big),
\end{align*}
where in the last step we used Lemma~\ref{lem:prodrelations}. Thus the map \eqref{eq:isomodule1} is indeed isometric.
Similarly we get a map
$$
(L^2(M) \otimes \bar{X}_s^\natural) \otimes_A H\to (X_s \otimes L^2(M^\mathrm{op})) \otimes_A H
$$
such that
\begin{equation}\label{eq:isomodule2}
(1\otimes\xi)\otimes\zeta\mapsto d_s^{-1/2}\sum_\alpha(\rho_{s\alpha}\otimes 1)\otimes (\bar\rho_{s\alpha}\otimes\xi)\zeta.
\end{equation}
Using again Lemma~\ref{lem:prodrelations} it is easy to check that the maps \eqref{eq:isomodule1} and \eqref{eq:isomodule2} are inverse to each other. This proves the first isomorphism in the formulation of the lemma.
The second isomorphism is proved similarly. Namely, the map
$$
H \otimes_A (X_s \otimes L^2(M^\mathrm{op})) \to H \otimes_A (L^2(M) \otimes \bar{X}_s^\natural)
$$
is defined by
\begin{equation*}\label{eq:isomodule3}
\zeta\otimes(\xi\otimes1)\mapsto d_s^{-1/2}\sum_\alpha\zeta(\xi\otimes\bar\rho_{s\alpha})\otimes (1\otimes\rho_{s\alpha}),
\end{equation*}
and the inverse map is given by
\begin{equation*}\label{eq:isomodule4}
\zeta\otimes(1\otimes\xi)\mapsto d_s^{-1/2}\sum_\beta\zeta(\bar\lambda_{s\beta}\otimes\xi)\otimes (\lambda_{s\beta}\otimes1).
\end{equation*}
This proves the assertion.
\end{proof}
Let $H$ be a Hilbert $ B$-bimodule. Denote by $X_{H,s}$ the space of $M^\mathrm{op}$-bimodule homomorphisms from $X_s^\natural$ to $H$. It has a natural inner product $(T, S) = S^* T \in \Hom_{M^\mathrm{op}\text{-}M^\mathrm{op}}(X_s^\natural, X_s^\natural) = \mathbb{C}$ and inherits the structure of a $M$-bimodule from $H$. We also put $X_H = X_{H,e}$ and identify $X_H$ with the space of $M^\mathrm{op}$-central vectors in $H$. Define maps
$$
l_s\colon X^0_s\otimes X_H\to X_{H,s}\ \ \text{and}\ \ r_s\colon X_H\otimes X_s^0\to X_{H,s}
$$
by
$$
l_s(\xi\otimes\zeta)\bar\eta=(\xi\otimes\bar\eta)\zeta\ \ \text{and}\ \ r_s(\zeta\otimes\xi)\bar\eta=\zeta(\xi\otimes\bar\eta)\ \ \text{for}\ \ \eta\in X^0_s.
$$
\begin{lemma} \label{lem:lr-iso}
The maps $l_s$ and $r_s$ define unitary isomorphisms of Hilbert $M$-bimodules
$$
X_s\otimes_M X_H\cong X_{H,s}\ \ \text{and}\ \ X_H\otimes_M X_s\cong X_{H,s},
$$
which we denote by the same symbols $l_s$ and $r_s$.
\end{lemma}
\begin{proof}
We have unitary isomorphisms
$$
X_{H,s}\cong\Hom_{M^\mathrm{op}\text{-}M^\mathrm{op}}(L^2(M)^\natural, \bar X_s^\natural\otimes_{M^\mathrm{op}}H)\cong \Hom_{M^\mathrm{op}\text{-}M^\mathrm{op}}(L^2(M)^\natural, X_s\otimes_MH),
$$
where the first isomorphism is the normalized Frobenius isomorphism $T\mapsto d_s^{-1/2}(\iota\otimes T)R^\natural_s$ and the second comes from Proposition~\ref{prop:move-Xi-thru-bimod}. Note that the $M^\mathrm{op}$-bimodule structure on $X_s\otimes_MH$ is defined by that on $H$. It follows that the space of $M^\mathrm{op}$-central vectors in $X_s\otimes_MH$ coincides with $X_s\otimes_M X_H$. We thus get a unitary isomorphism $X_{H,s}\cong X_s\otimes_M X_H$. Explicitly, using the formula for $R_s$ and \eqref{eq:isomodule2}, this isomorphism is given by
\begin{equation*}\label{eq:li-inverse}
T\mapsto d_s^{-1}\sum_{\alpha,\beta}\rho_{s\alpha}\otimes (\bar\rho_{s\alpha}\otimes\lambda_{s\beta})T\bar\lambda_{s\beta}.
\end{equation*}
Using Lemma~\ref{lem:prodrelations} it is straightforward to check that $l_s$ defines a right inverse of this map.
Similarly it is proved that $r_s$ defines a unitary isomorphism $X_H\otimes_M X_s\cong X_{H,s}$, with the inverse given by
\begin{equation}\label{eq:ri-inverse}
T\mapsto d_s^{-1}\sum_{\alpha,\beta}(T\bar\rho_{s\alpha})(\bar\lambda_{s\beta}\otimes\rho_{s\alpha})\otimes\lambda_{s\beta}.
\end{equation}
This proves the assertion.
\end{proof}
By the previous lemma we obtain a unitary isomorphism of Hilbert $M$-bimodules $c^H_s=r_s^*l_s\colon X_s \otimes_M X_H \to X_H \otimes_M X_s$. Explicitly, using formula~\eqref{eq:ri-inverse} for $r_s^*$, we have
\begin{equation} \label{eq:halfbr-bimod}
c^H_s(\xi\otimes\zeta)=d_s^{-1}\sum_{\alpha,\beta}(\xi\otimes\bar\rho_{s\alpha})\zeta(\bar\lambda_{s\beta} \otimes\rho_{s\alpha})\otimes\lambda_{s\beta}.
\end{equation}
In a more invariant form we can say that $c^H_s$ is characterized by the identity
\begin{equation*} \label{eq:halfbr-bimod1}
(\xi\otimes\bar\eta)\zeta=c^H_s(\xi\otimes\zeta)_1(c^H_s(\xi\otimes\zeta)_2\otimes\bar\eta)\ \ \text{for}\ \ \xi,\eta\in X^0_s,\ \zeta\in X_H,
\end{equation*}
where we use Sweedler's sumless notation $c^H_s(\xi\otimes\zeta)_1\otimes c^H_s(\xi\otimes\zeta)_2$ for $c^H_s(\xi\otimes\zeta)$. Equivalently, $c^{H*}_s$ is characterized by
\begin{equation} \label{eq:halfbr-bimod2}
\zeta(\xi\otimes\bar\eta)=(c^{H*}_s(\zeta\otimes\xi)_1\otimes\bar\eta)c^{H*}_s(\zeta\otimes\xi)_2\ \ \text{for}\ \ \xi,\eta\in X^0_s,\ \zeta\in X_H.
\end{equation}
The family of isomorphisms $\{c^H_s\}_{s\in I}$ defines a natural family of unitary isomorphisms $(c^H_X\colon X \otimes X_H \to X_H \otimes X)_{X \in \mathcal{C}}$ (where we write $\otimes$ for the tensor product $\otimes_M$ in $\Hilb_M$).
\begin{lemma}\label{lem:XH-tensor-central}
The unitaries $c^H_X$ satisfy the half-braiding condition $c^H_{X\otimes Y}=(c^H_X\otimes\iota_Y)(\iota_X\otimes c^H_Y)$.
\end{lemma}
\begin{proof}
It suffices to consider $X=X_s$ and $Y=X_t$. But in this case the result follows immediately from the explicit formula~\eqref{eq:halfbr-bimod} using that left and right bases in $X_s\otimes_M X_t$ can be obtained either by taking tensor products of bases in $X_s$ and $X_t$, or by decomposing $X_s\otimes_MX_t$ into direct sums of the simple modules $X_k$ and choosing bases in $X_k$.
\end{proof}
So far we have used only that $H\in\Hilb_{ B}$, in which case we cannot say much about $X_H\in\Hilb_M$. But if we assume that $H\in\ZHC$, then $X_H$ decomposes into a direct sum of copies of $X_s$, $s\in I$. Consider the full subcategory of $\Hilb_M$ consisting of bimodules allowing such a decomposition. We have an obvious functor from $\indcat{\CC}$ into this category, which is a unitary monoidal equivalence. In order to not introduce yet another notation, in the remaining part of the proof we do not distinguish between these two equivalent categories. Thus, if $H\in\ZHC$, then $X_H\in\indcat{\CC}$ and therefore $(X_H,c^H)\in{\Dcen(\indC)}$.
Observe also that if $H\in\ZHC$, then we can reconstruct $H$ from $(X_H,c^H)$. Indeed, first of all we have a unitary isomorphism $\oplus_s X_{H,s}\otimes X_s^\natural\cong H$ of Hilbert $A$-bimodules, mapping $T\otimes\bar\eta$ into $T\bar\eta$. So by Lemma~\ref{lem:lr-iso} we get a unitary isomorphism of Hilbert $A$-bimodules
$$
\bigoplus_s(X_s\otimes X_s^\natural)\otimes_A(X_H\otimes L^2(M^\mathrm{op}))\cong H,\ \ (\xi\otimes\bar\eta)\otimes(\zeta\otimes1)\mapsto(\xi\otimes\bar\eta)\zeta.
$$
Identifying the $A$-bimodule on the left with $L^2( B)\otimes_A (X_H\otimes L^2(M^\mathrm{op}))$, we see that we actually get an isomorphism of $ B$-$A$-modules. Note in passing that this isomorphism easily implies that $\ZHC$ is closed under tensor products.
Next, the unitaries $c^H_s$ define a unitary isomorphism $U$ of Hilbert $A$-bimodules
$$
\bigoplus_s(X_s\otimes X_s^\natural)\otimes_A(X_H\otimes L^2(M^\mathrm{op}))\cong \bigoplus_s(X_H\otimes L^2(M^\mathrm{op}))\otimes_A(X_s\otimes X_s^\natural),
$$
\begin{equation} \label{eq:unitaryU}
U\big((\xi\otimes\bar\eta)\otimes(\zeta\otimes1)\big)=(c^H(\xi\otimes\zeta)_1\otimes1)\otimes(c^H(\xi\otimes\zeta)_2\otimes\bar\eta).
\end{equation}
Composing $U^*$ with the isomorphism $L^2( B)\otimes_A (X_H\otimes L^2(M^\mathrm{op}))\cong H$, we get an isomorphism
\begin{equation*}\label{eq:isoright}
(X_H\otimes L^2(M^\mathrm{op}))\otimes_A L^2( B)\cong H
\end{equation*}
of Hilbert $A$-bimodules. From \eqref{eq:halfbr-bimod2} we see that this isomorphism maps $(\zeta\otimes1)\otimes(\xi\otimes\bar\eta)$ into $\zeta(\xi\otimes\bar\eta)$ and hence respects the right actions of $ B$. Thus up to an isomorphism the Hilbert $ B$-bimodule $H$ can be reconstructed from the Hilbert $M$-module $X_H$ and the unitary $U$ defined by the half-braiding $c^H$.
\smallskip
Now take an object $(X,c)\in{\Dcen(\indC)}$. We want to construct a Hilbert $ B$-bimodule $H=H_{(X,c)}\in\ZHC$ out of~$(X,c)$. For this we basically have to repeat the procedure we used above to reconstruct a Hilbert $ B$-module from the corresponding object of ${\Dcen(\indC)}$. Thus, we put $H=L^2( B)\otimes_A (X\otimes L^2(M^\mathrm{op}))$ and consider $H$ as a Hilbert $ B$-$A$-module. Next, using the braiding $c$ instead of $c^H$ define a unitary $U$ by~\eqref{eq:unitaryU}. We can use the right $ B$-module structure on $(X\otimes L^2(M^\mathrm{op}))\otimes_A L^2( B)$ to define such a structure on~$H$, so for $\xi\in H$ and $b\in B$ we let $\xi b=U^*((U\xi)b)$.
\begin{lemma}\label{lem:half-br-and-prod-of-tB}
The left and right actions of $ B$ on $H$ commute, so $H$ is a Hilbert $ B$-module.
\end{lemma}
\begin{proof} For convenience write $\hat X$ for $X\otimes L^2(M^\mathrm{op})$. Let us also denote by $m$ the product on $ B$. We claim that
\begin{equation} \label{eq:Izu}
U (m \otimes \iota_{\hat{X}}) = (\iota_{\hat{X}} \otimes m) (U \otimes \iota_ B) (\iota_ B \otimes U)
\end{equation}
on a dense subspace of $L^2( B)\otimes_AL^2( B)\otimes_A\hat X$, which is an analogue of~\cite{MR1782145}*{(4.1)}. More precisely, we claim that the above identity holds on the $M\otimes_{\mathrm{alg}}M^\mathrm{op}$-sub-bimodule spanned by vectors of the form $(\xi_s\otimes\bar\eta_s)\otimes(\zeta\otimes 1)\otimes(\xi_t\otimes\bar\eta_t)$, with $\xi_s,\eta_s\in X^0_s$, $\xi_t,\eta_t\in X^0_t$ and $\zeta\in X$. On this sub-bimodule both sides of~\eqref{eq:Izu} make perfect sense, since the spaces $X_s\otimes_M X$ and $X\otimes_M X_s$ coincide with the algebraic tensor products $X_s^0\otimes_M X$ and $X\otimes_M X_s^0$ as both $(X_s)_M$ and ${}_MX_s$ have finite Murray--von Neumann dimensions.
Choose coisometries $w_\alpha\colon X_s\otimes_M X_t\to X_{k_\alpha}$ defining a decomposition of $X_s\otimes_M X_t$ into simple modules, and put $\xi_\alpha=w_\alpha(\xi_s\otimes \xi_t)$ and $\eta_\alpha=w_\alpha(\eta_s\otimes\eta_t)$, so that in $ B$ we have
$$
(\xi_s\otimes\bar\eta_s)(\xi_t\otimes\bar\eta_t)=\sum_\alpha\left(\frac{d_sd_t}{d_{r_\alpha}}\right)^{1/2}\xi_\alpha\otimes\bar\eta_\alpha.
$$
Applying both sides of \eqref{eq:Izu} to $(\xi_s\otimes\bar\eta_s)\otimes(\zeta\otimes 1)\otimes(\xi_t\otimes\bar\eta_t)$ we have to check that
\begin{multline} \label{eq:Izu2}
\sum_\alpha\left(\frac{d_sd_t}{d_{r_\alpha}}\right)^{1/2} c_{k_\alpha}(\xi_\alpha\otimes\zeta)\otimes\bar\eta_\alpha
\\=c_s\big(\xi_s\otimes c_t(\xi_t\otimes\zeta)_1\big)_1\otimes\big(c_s\big(\xi_s\otimes c_t(\xi_t\otimes\zeta)_1\big)_2\otimes\bar\eta_s\big)
\big(c_t(\xi_t\otimes\zeta)_2\otimes\bar\eta_t\big).
\end{multline}
By the half-braiding condition we have
$$
(c_s\otimes\iota)(\iota\otimes c_t)(\xi_s\otimes\xi_t\otimes\zeta)
=\sum_\alpha(\iota\otimes w_\alpha^*)c_{k_\alpha}(\xi_\alpha\otimes\zeta).
$$
This implies that the right hand side of \eqref{eq:Izu2} equals
$$
\sum_\alpha c_{k_\alpha}(\xi_\alpha\otimes\zeta)_1\otimes \big((w^*_\alpha c_{k_\alpha}(\xi_\alpha\otimes\zeta)_2)_1
\otimes\bar\eta_s\big)\big((w^*_\alpha c_{k_\alpha}(\xi_\alpha\otimes\zeta)_2)_2\otimes\bar\eta_t\big).
$$
We remark that the above expression is still meaningful, since the algebraic tensor product $X^0_s\otimes_M X^0_t$ coincides with $(X_s\otimes_M X_t)^0$ and hence the vector $(\iota\otimes w_\alpha^*)c_{k_\alpha}(\xi_\alpha\otimes\zeta)$ lies in the algebraic tensor product $X\otimes_M X^0_s\otimes_M X^0_t$. By definition of the product in $ B$ we then see that the above expression equals
$$
\sum_{\alpha,\beta}\left(\frac{d_sd_t}{d_{r_\beta}}\right)^{1/2} c_{k_\alpha}(\xi_\alpha\otimes\zeta)_1\otimes w_\beta w^*_\alpha c_{k_\alpha}(\xi_\alpha\otimes\zeta)_2\otimes \bar\eta_\beta,
$$
which is the left hand side \eqref{eq:Izu2} as the coisometries $w_\alpha$ have mutually orthogonal domains. Thus \eqref{eq:Izu} is proved.
Now, for $v=\xi\otimes1\in\hat X$ and $b,c\in B$ lying in the linear span of $X^0_s\otimes X^{0\natural}_s$, we get
\begin{equation}\label{eq:unitaryU1}
(b \otimes v) c = U^* (\iota \otimes m) (U \otimes \iota)(b \otimes v \otimes c) = (m \otimes \iota) (\iota \otimes U^*) (b \otimes v \otimes c)
=(b\otimes1)U^*(v\otimes c),
\end{equation}
from which it becomes obvious that the left and right actions of $ B$ on $L^2( B)\otimes_A\hat X$ commute.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:mon-eqv-D-center-B-bimod}]
The construction of the object $(X_H,c^H)$ out of $H\in\ZHC$ defines a unitary functor $G\colon\ZHC\to{\Dcen(\indC)}$, and the construction of the module $H_{(X,c)}$ out of $(X,c)\in{\Dcen(\indC)}$ defines a unitary functor $F\colon{\Dcen(\indC)}\to\ZHC$. The discussion following Lemma~\ref{lem:XH-tensor-central} implies that $FG$ is naturally unitarily isomorphic to the identity functor. In the opposite direction, we also have obvious natural isomorphisms $GF(X,c)\cong X$ in $\indcat{\CC}$. It is slightly less obvious that these isomorphisms respect the half-braidings. In fact, this is automatically the case. Indeed, since $FGF\cong F$, it suffices to show that if two half-braidings~$c$ and~$c'$ for some $X\in\indcat{\CC}$ define the same right $ B$-module structure on $L^2( B)\otimes_A(X\otimes L^2(M^\mathrm{op}))$, then $c=c'$. But this follows from formula~\eqref{eq:unitaryU1}, which shows that the unitary $U$ is completely determined by the $ B$-bimodule structure. Thus $GF$ is naturally unitarily isomorphic to the identity functor on ${\Dcen(\indC)}$.
Therefore $F$ and $G$ are unitary equivalences between the categories ${\Dcen(\indC)}$ and $\ZHC$. In order to get a unitary monoidal equivalence, it remains to turn either of these two functors into a unitary tensor functor. Let us do this for the functor $G\colon\ZHC\to{\Dcen(\indC)}$. Given $H,K\in{\Dcen(\indC)}$, define $G_2\colon X_H\otimes_M X_K\to X_{H\otimes_ B K}$ by $\xi\otimes\zeta\mapsto\xi\otimes\zeta$. This is easily seen to be a well-defined unitary isomorphism of Hilbert $M$-bimodules, since the embeddings $X_H\hookrightarrow H$ and $X_K\hookrightarrow K$ extend to unitary isomorphisms $(X_H\otimes L^2(M^\mathrm{op}))\otimes_AL^2( B)\cong H$ and $L^2( B)\otimes_A (X_K\otimes L^2(M^\mathrm{op}))\cong K$ of Hilbert $A$-$ B$- and $ B$-$A$-modules, respectively. It remains to check that $G_2$ respects the half-braidings. By \eqref{eq:halfbr-bimod}, for $\xi\in X_s^0$, $\zeta\in X_H$ and $\eta\in X_K$, we have
$$
(\iota\otimes c^H_s)(c^K_s\otimes\iota)(\xi\otimes\zeta\otimes\eta) = d_s^{-2}\sum_{\alpha,\beta,\alpha',\beta'}(\xi\otimes\bar\rho_{s\alpha})\zeta(\bar\lambda_{s\beta}\otimes\rho_{s\alpha}) \otimes
(\lambda_{s\beta}\otimes\bar\rho_{s\alpha'})\eta(\bar\lambda_{s\beta'}\otimes\rho_{s\alpha'})\otimes\lambda_{s\beta'}.
$$
Using that $d_s^{-1}\sum_\beta(\bar\lambda_{s\beta}\otimes\rho_{s\alpha}) (\lambda_{s\beta}\otimes\bar\rho_{s\alpha'}) =1\otimes\langle\rho_{s\alpha'},\rho_{s\alpha}\rangle_M$ by Lemma~\ref{lem:prodrelations}, then the $M^\mathrm{op}$-centrality of~$\zeta$, and finally that $\sum_\alpha\rho_{s\alpha}\langle\rho_{s\alpha},\rho_{s\alpha'}\rangle_M=\rho_{s\alpha'}$, we see that the above expression equals
$$
d_s^{-1}\sum_{\alpha',\beta'}(\xi\otimes\bar\rho_{s\alpha'})\zeta\otimes
\eta(\bar\lambda_{s\beta'}\otimes\rho_{s\alpha'})\otimes\lambda_{s\beta'}.
$$
This is $c^{H\otimes_ B K}_s(\xi\otimes\zeta\otimes\eta)$, as we need.
\end{proof}
We finish this section by noting that it is very plausible that the results we have obtained remain true without assumption~\eqref{eq:eqtrace} on $\mathcal{C}$. In this case, however, the factor $B$ is no longer of type II$_1$ (see \cite{MR3406647}*{Remark~2.7}) and more care is needed in dealing with Hilbert bimodules.
\subsection{Pointed bimodules and representations of the character algebra}
\label{sec:rep-char-alg-on-bimod}
We continue to consider a category $\mathcal{C}\subset\Hilb_M$ as in Section~\ref{subsec:LR}, with the associated Longo--Rehren inclusion $A\subset B$. For categories of the form $\mathcal{C}_{N\subset M}$ as in Example~\ref{ex:subfactorcat}, the results that follow have been obtained by Popa and Vaes~\cite{MR3406647} by different methods.
Given a Hilbert $ B$-module $H\in\ZHC$, we have the corresponding object $(X_H,c^H)\in{\Dcen(\indC)}$, and hence a representation $\pi_H$ of $C^*(\mathcal{C})$ on $\Mor_{\indcat{\CC}}(\mathds{1},X_H)=\Hom_{M\text{-}M}(L^2(M),X_H)$. Recall that $X_H\subset H$ is the subspace of $M^\mathrm{op}$-central vectors. Then $\Hom_{M\text{-}M}(L^2(M),X_H)$ can be identified with the subspace of $X_H$ of $M$-central vectors, that is, with the subspace $H_0\subset H$ of $A$-central vectors. Recalling formula~\eqref{eq:halfbr-bimod} for the half-braiding~$c^H$ and how the representation associated with a half-braiding is defined, we get the following formula for~$\pi_H$:
$$
\pi_H([X_s])\xi=d_s^{-1}\sum_{\alpha,\alpha'}(\rho_{s\alpha}\otimes\bar\rho_{s\alpha'})\xi(\bar\rho_{s\alpha}\otimes\rho_{s\alpha'})\ \ \text{for}\ \ \xi\in H_0,
$$
where $\{\rho_{s\alpha}\}_\alpha$ is a basis of the right Hilbert $M$-module $X_s$.
Every $A$-central unit vector $\xi\in H_0$ defines an $A$-bimodular normal ucp map $\Phi_\xi\colon B\to B$, see e.g.~\cite{MR2215135}*{Section~1.1}. Namely, since $A'\cap B=\C1$, we have $(x\xi,\xi)=(\xi x,\xi)=\tau(x)$ for all $x\in B$. Therefore the map $ B\ni x\mapsto \xi x$ extends to an isometry $T_\xi\colon L^2( B)\to H$. Viewing then elements of $ B$ as operators acting on the left on $L^2( B)$ and $H$, define $\Phi_\xi(x) = T_\xi^*x T_\xi$. In other words, $\Phi_\xi(x)$ is characterized by the identity
\begin{equation} \label{eq:Phi}
\tau(\Phi_\xi(x)b)=(x\xi b,\xi)\ \ \text{for}\ \ b\in B.
\end{equation}
Conversely, given an $A$-bimodular normal ucp map $\Phi\colon B\to B$, one can construct a Hilbert $ B$-bimodule $H$ together with a distinguished $A$-central unit vector $\xi\in H$: we obtain $H$ from $ B$ using the pre-inner product $(x,y)=\tau(\Phi(x)\Phi(y)^*)$ and then let $\xi$ be the image of $1\in B$ in $H$. This gives a one-to-one correspondence between $A$-bimodular normal ucp maps $\Phi\colon B\to B$ and isomorphism classes of pointed Hilbert bimodules over $A\subset B$, by which one means pairs $(H,\xi)$ consisting of a Hilbert $ B$-bimodule $H$ and an $A$-central unit vector $\xi\in H$ such that $ B\xi B$ is dense in $H$. Note that if $(H,\xi)$ is a pointed Hilbert bimodule, then $H\in\ZHC$, since $H$ is generated as a $ B$-bimodule by a copy of $L^2(A)$.
The ucp map $\Phi_\xi$ extends to a contraction on $L^2( B)$. Since it is $A$-bimodular, it follows that on $X_s\otimes X^\natural_s$ this extension acts as a scalar $\alpha_{\xi,s}$. Taking $x=\rho_{s\alpha}\otimes\bar\rho_{s\alpha'}$ and $b=\bar\rho_{s\alpha}\otimes\rho_{s\alpha'}=x^*$ in~\eqref{eq:Phi} and then summing up over $\alpha,\alpha'$, we get $\dim({}_A(X_s\otimes X^\natural_s))=d_s^2\alpha_{\xi,s}$ on the left and $d_s(\pi_H([X_s])\xi,\xi)$ on the right, that~is,
\begin{equation} \label{eq:bimodchar}
d_s\alpha_{\xi,s}=(\pi_H([X_s])\xi,\xi).
\end{equation}
Therefore $(\alpha_{\xi,s})_s$ is exactly the positive definite function on $I=\Irr(\mathcal{C})$ associated with $(X_H,c^H)\in{\Dcen(\indC)}$ and $\xi\in \Mor_{\indcat{\CC}}(\mathds{1},X_H)$. By Theorems~\ref{thm:positivedef} and~\ref{thm:mon-eqv-D-center-B-bimod} every positive definite function $\phi$ with $\phi(e)=1$ arises this way: the corresponding pointed Hilbert bimodule is $(F(Z_\phi),F(\xi_\phi))$. Note that the fact that~$F(Z_\phi)$ is generated by $F(\xi_\phi)$ as a $ B$-module follows from the universality property of $(Z_\phi,\xi_\phi)$ established in Proposition~\ref{prop:Zphi-universal}.
To summarize the above discussion, we have proved the following result.
\begin{theorem}
We have a one-to-one correspondence between $A$-bimodular normal ucp maps $\Phi\colon B\to B$ and positive definite functions $\phi$ on $I=\Irr(\mathcal{C})$ such that $\phi(e)=1$. Namely, the map $\Phi$ corresponding to $\phi$ is defined by $\Phi(x)=\phi(s)x$ for $x\in X^0_s\otimes X^{0\natural}_s\subset B$ and $s\in I$.
\end{theorem}
\begin{example}
The function $\phi\equiv1$ defines the identity map. The corresponding pointed bimodule is $(L^2( B),1)$. The function $\phi=\delta_e$ defines the conditional expectation $ B\to A$. The corresponding pointed bimodule is $(L^2( B)\otimes_A L^2( B),1\otimes1)$. In particular, the bimodule $L^2( B)\otimes_A L^2( B)$ corresponds to $Z_{\mathrm{reg}}(\mathcal{C})$ under the equivalence between $\ZHC$ and ${\Dcen(\indC)}$.
\end{example}
Consider now the following rigidity condition. One says~\citelist{\cite{popa-corr-preprint}*{Chapter~4}\cite{MR914742}} that $ B$ has property~(T) relative to $A$, or that $A$ is corigid in $ B$, if there are a finite set $F \subset B$ and $\varepsilon > 0$ such that if $(H, \xi)$ is a pointed Hilbert bimodule over $A \subset B$ with $\norm{x \xi - \xi x} < \varepsilon$ for $x \in F$, then there is a $ B$-central unit vector in~$H$. (This is not to be confused with rigidity of the inclusion $A \subset B$~\cite{MR2215135}.) For some other equivalent conditions see~\cite{popa-corr-preprint}*{4.1.5} and~\cite{MR1729488}*{9.1}.
\begin{proposition}
\label{prop:cat-T-vs-subfactor-T}
The C$^*$-tensor category $\mathcal{C}$ has property (T) if and only if $ B$ has property (T) relative to~$A$.
\end{proposition}
\begin{proof} Let $H\in\ZHC$ be a Hilbert $ B$-module and $\xi\in H$ be an $A$-central unit vector. Then by~\eqref{eq:Phi}, for any $x\in B$ we have
$$
\|x\xi-\xi x\|^2=2\tau(x^*x)-\tau(\Phi_\xi(x)x^*)-\tau(\Phi_\xi(x^*)x).
$$
Hence for any $x\in X^0_s\otimes X^{0\natural}_s$ we have
$$
\|x\xi-\xi x\|^2=2(1-\Re \alpha_{\xi,s})\|x\|^2_2.
$$
As $d_s\alpha_{\xi,s}=(\pi_H([X_s])\xi,\xi)$ by \eqref{eq:bimodchar}, we see that an $A$-central vector is $ B$-central if and only if it is invariant with respect to the representation $\pi_H\colon C^*(\mathcal{C})\to B(H_0)$. It also follows that $ B$ has property~(T) relative to~$A$ if and only if there are a finite set $F \subset I$ and $\varepsilon > 0$ such that if $H\in\ZHC$ and $\xi$ is an $A$-central unit vector satisfying $|(\pi_H([X_s])\xi,\xi)-d_s| < \varepsilon$ for all $s \in F$, then there is an invariant unit vector in $H_0$. But in view of the equivalence between the categories ${\Dcen(\indC)}$ and $\ZHC$, this is exactly condition (ii) in Proposition~\ref{prop:propertyT}.
\end{proof}
\bigskip
\section{Weakly Morita equivalent categories}
\label{sec:induct-thro-weak-Morita}
In this section we study Drinfeld centers and representation theory for weakly monoidally Morita equivalent C$^*$-tensor categories. As in Sections~\ref{sec:half-braidings} and~\ref{sec:character-algebra}, let $\mathcal{C}$ be an essentially small strict rigid C$^*$-tensor category satisfying our standard assumptions.
\subsection{\texorpdfstring{$Q$}{Q}-systems}
\label{sec:prelim-q-system}
A \emph{$Q$-system}~\cite{MR1257245} in $\mathcal{C}$ is a triple $(Q, v, w)$, where $Q$ is an object in $\mathcal{C}$, $v$ is an isometry in $\mathcal{C}(\mathds{1}, Q)$, and $w$ is a scalar multiple of an isometry in $\mathcal{C}(Q, Q \otimes Q)$, satisfying
\begin{itemize}
\item[]\emph{unit constraint}: $(v^* \otimes \iota) w = \iota = (\iota \otimes v^*) w$,
\item[]\emph{associativity}: $(w \otimes \iota) w = (\iota \otimes w) w$,
\item[]\emph{Frobenius condition}: $(w^* \otimes \iota) (\iota \otimes w) = w w^* = (\iota \otimes w^*) (w \otimes \iota)$.
\end{itemize}
(Of course, other normalizations than $v^*v=\iota$ are possible and used in the literature, and the last equality is redundant; in fact, it can be shown that the entire Frobenius condition is redundant~\cite{MR1444286}.)
In other words, a $Q$-system is a Frobenius object in $\mathcal{C}$ such that the algebra and coalgebra structures on it are obtained from each other by taking adjoints, and the coproduct is a scalar multiple of an isometry. Depending on the context, we will sometimes write $m_Q$ for the product $w^*$.
The object~$Q$ is self-dual, with $(w v, w v)$ being a solution of the conjugate equations.
In the following we also assume that
$$
\text{the}\ Q\text{-system} \ Q \ \text{is \emph{standard} and \emph{simple}},
$$
see~\cite{MR3308880}. The first assumption means that $w^*w=d(Q)\iota$, so that $(w v, w v)$ is a standard solution of the conjugate equations. The second assumption means that $Q$ is simple as a $Q$-bimodule. In the last section instead of simplicity we will require $Q$ to be \emph{irreducible}, meaning that $Q$ is simple as a left and/or right $Q$-module.
A left $Q$-module is an object $M \in \mathcal{C}$ together with a morphism $m_M = m^l_M \in \mathcal{C}(Q \otimes M, M)$ satisfying the associativity condition $m_M (m_Q \otimes \iota) = m_M (\iota \otimes m_M)$, the unit constraint $m_M (v \otimes \iota) = \iota$, and the $*$-compatibility $m_M^* = (\iota \otimes m_M) (w v \otimes \iota)$. By the Frobenius condition, $Q$ itself is a left $Q$-module in this sense. Furthermore, for general $M$ the map $m_M$ has the following properties similar to $m_Q=w^*$.
\begin{lemma}
\label{lem:comput-m-m-star}
For any left $Q$-module $M$, we have
$$
m_M m_M^{*} = d(Q) \iota\ \ \text{and}\ \ m^*_Mm_M=(\iota_Q\otimes m_M)(w\otimes\iota_M).
$$
\end{lemma}
\begin{proof}
We have
$$
m_M m_M^{*} = m_M (\iota \otimes m_M) (w v \otimes \iota)=m_M(w^* \otimes\iota)(w v \otimes \iota) = d(Q) \iota,
$$
since $w^*w=d(Q)\iota$ by our assumptions.
For the second identity, we get an equivalent one if we tensor it on the left by $\iota_Q$ and then multiply by $R_Q\otimes\iota_M=(wv)^*\otimes\iota_M$. Then the left hand side gives
$$
((wv)^*\otimes\iota_M)(\iota_Q\otimes\iota_Q \otimes m_M) (\iota_Q\otimes w v \otimes \iota_M)(\iota_Q\otimes m_M)=m_M(\iota_Q\otimes m_M),
$$
while the right hand side gives
$$
((wv)^*\otimes\iota_M)(\iota_Q\otimes\iota_Q\otimes m_M)(\iota_Q\otimes w\otimes\iota_M)=m_M(m_Q\otimes\iota_M),
$$
since $((wv)^*\otimes\iota)(\iota\otimes w)=w^*=m_Q$. Thus the lemma is proved.
\end{proof}
A $Q$-morphism between two left $Q$-modules $M$ and $N$ is a morphism $T \in \mathcal{C}(M, N)$ satisfying $T m_M = m_N (\iota \otimes T)$.
\begin{remark}
Let us call a left $Q$-module without the $*$-compatibility an algebraic left $Q$-module. Any algebraic left $Q$-module $M$ is isomorphic to a left $Q$-module as follows. Put $A = m_M m^*_M$. This is a positive morphism majorizing $\iota = m_M (vv^*\otimes\iota) m_M^*$, so in particular $A$ is invertible. Then it can be checked that $A^{-{1/2}} m_M (\iota \otimes A^{{1/2}})$ defines the structure of a left $Q$-module on $M$, and $A^{{1/2}}$ gives an isomorphism of this module with $(M, m_M)$.
\end{remark}
We denote the category of left $Q$-modules by $Q\mhyph\mathrm{mod}_\mathcal{C}$ or, if it is clear from the context, just by $Q\mhyph\mathrm{mod}$.
By the $*$-compatibility condition, $Q\mhyph\mathrm{mod}$ is closed under the involution $T \mapsto T^*$, so it is a C$^*$-category. Since $\mathcal{C}$ is closed under subobjects and has finite dimensional morphism spaces, it follows that $Q\mhyph\mathrm{mod}$ is semisimple. We also note that $Q\mhyph\mathrm{mod}$ is a right $\mathcal{C}$-module category.
For any $X\in\mathcal{C}$ and any left $Q$-module $M$ we have
an isomorphism
$$
\mathcal{C}(X,M)\cong\Mor_{Q\mhyph\mathrm{mod}}(Q\otimes X,M), \ \ T\mapsto m_M(\iota\otimes T),
$$
with the inverse $T'\mapsto T'(v\otimes\iota)$. In particular, $\mathcal{C}(\mathds{1},Q)\cong\End_{Q\mhyph\mathrm{mod}}(Q)$. Therefore irreducibility of $Q$ is equivalent to $\dim\mathcal{C}(\mathds{1},Q)=1$. We also remark that the above isomorphism implies that either $\Irr(\mathcal{C})$ and $\Irr(Q\mhyph\mathrm{mod})$ are finite or they are both infinite of the same cardinality.
The notions of right $Q$-modules and $Q$-bimodules are defined similarly, and the corresponding categories are denoted by $\mathrm{mod}\mhyph Q$ and $Q\mhyph\mathrm{mod}\mhyph Q$.
\begin{example}
Let $X$ be a simple object in $\mathcal{C}$. Then $Q_X = \bar{X} \otimes {X}$ has the canonical structure of an irreducible standard $Q$-system, with $v = d(X)^{-{1/2}} R_X$ and $w = d(X)^{1/2} \iota_{\bar{X}} \otimes \bar{R}_X \otimes \iota_X$. If $(M, m_M \colon Q_X \otimes M \to M)$ is a left $Q_X$-module, then $p = d(X)^{-3/2} (\iota_X \otimes m_M)(\bar{R}_X \otimes \iota_X\otimes\iota_M) \in \mathcal{C}(X \otimes M)$ is a projection. Let $N$ be the subobject of $X \otimes M$ specified by $p$. Then $\bar X\otimes N$ has a $Q_X$-module structure given by $d(X)^{1/2}\iota \otimes \bar{R}_X^* \otimes \iota$, and it is not difficult to show that the map $d(X)^{1/2}(\iota_{\bar X}\otimes p)(R_X\otimes\iota_M)=d(X)^{-1}m^*_M$ is an isometric $Q_X$-module isomorphism $M\cong \bar{X} \otimes N$. This way we obtain an equivalence $Q_X\mhyph\operatorname{mod} \cong \mathcal{C}$.
\end{example}
\begin{example}
Let $\mathcal{C}=\mathcal{C}_{N\subset M}$ be the category of Hilbert $M$-bimodules defined by an extremal finite index subfactor as in Example~\ref{ex:subfactorcat}. Then $Q = L^2(M_1)$ has the structure of a standard simple $Q$-system in~$\mathcal{C}$, where $v\colon L^2(M) \to L^2(M_1)$ is the inclusion map and $w\colon L^2(M_1) \to L^2(M_1)\otimes_M L^2(M_1)=L^2(M_2)$ is the inclusion map multiplied by $[M : N]^{{1/2}}$. This $Q$-system is irreducible if and only if $N\subset M$ is irreducible.
Up to monoidal equivalence, any standard simple $Q$-system $(Q, v, w)$ such that~$Q$ is a generating object is obtained this way.
Indeed, we get a $\lambda$-lattice by taking the algebras $\mathcal{C}(Q^{\otimes k})$, $\End_{Q\mhyph\mathrm{mod}}(Q^{\otimes k})$, $\End_{\mathrm{mod}\mhyph Q}(Q^{\otimes k})$, and $\End_{Q\mhyph\mathrm{mod}\mhyph Q}(Q^{\otimes k})$ with the natural inclusions maps and the normalized categorical traces. In order to verify the axioms of $\lambda$-lattices it suffices to show that for any $Q$-bimodule $M$ we get a commuting square
$$
\begin{matrix}
\End_{\mathrm{mod}\mhyph Q}(M) &\subset & \mathcal{C}(M)\\
\cup& &\cup\\
\End_{Q\mhyph\mathrm{mod}\mhyph Q}(M) &\subset & \End_{Q\mhyph\mathrm{mod}}(M).
\end{matrix}
$$
We claim that the trace-preserving conditional expectation $E^l_M\colon \mathcal{C}(M)\to\End_{Q\mhyph\mathrm{mod}}(M)$ is given by $$E^l_M(T)=d(Q)^{-1}m_M^l(\iota_Q\otimes T)m_M^{l*}.$$ It follows easily from Lemma~\ref{lem:comput-m-m-star} that this formula defines a conditional expectation onto $\End_{Q\mhyph\mathrm{mod}}(M)$, so we only need to show that the trace is preserved. This is equivalent to showing that $(\tr_Q\otimes\iota)(m^{l*}_Mm^l_M)=\iota_M.$
But this follows from the second identity in Lemma~\ref{lem:comput-m-m-star}, since
$$
(\tr_Q\otimes\iota)(w)=d(Q)^{-1}((wv)^*\otimes\iota)(\iota\otimes w)wv=d(Q)^{-1}w^*wv=v.
$$
The explicit formula shows that the conditional expectation $E^l_M$ maps $ \End_{\mathrm{mod}\mhyph Q}(M)$ into $ \End_{Q\mhyph\mathrm{mod}\mhyph Q}(M)$, so we indeed get a commuting square.
\end{example}
The semisimplicity implies that the finite colimits exist in $\mathcal{C}$, as can be seen by taking isotypic decompositions and reducing the statement to the category of finite dimensional Hilbert spaces.\footnote{However, infinite colimits do not make sense in general even in $\indcat{\CC}$, because we require uniform boundedness of morphisms.} In particular, we have a natural relative tensor product operation
$$
\mathrm{mod}\mhyph Q \times Q\mhyph\mathrm{mod} \to \mathcal{C}, \quad (M, N) \mapsto M \otimes_Q N = \text{coequalizer of } M \otimes Q \otimes N \rightrightarrows M \otimes N.
$$
We denote by $P_{M,N}$ the structure morphism $M\otimes N\to M\otimes_QN$ .
\begin{remark}
Although we will never need this, the tensor products over $Q$ can be explicitly constructed as follows. Consider the morphism
$$
p=d(Q)^{-1}(\iota\otimes v^*w^*\otimes\iota)(m^{r*}_M\otimes m^{l*}_N)\in\mathcal{C}(M\otimes N).
$$
Then $p$ is a projection, so there exists an object $X$ and an isometry $u\colon X\to M\otimes N$ such that $uu^*=p$, and as $P_{M,N}\colon M\otimes N\to M\otimes_Q N$ we can take $u^*\colon M\otimes N\to X$.
\end{remark}
By taking the polar decomposition of the adjoint of the structure morphism $P_{M,N}\colon M\otimes N\to M\otimes_Q N$ we may first assume that~$P_{M,N}$ is a coisometry, and then in view of Lemma~\ref{lem:comput-m-m-star} we rescale it so that $$P_{M,N}P^*_{M,N}=d(Q)\iota.$$Therefore for a left $Q$-module $M$ as a model of $Q\otimes_Q M$ we take $M$ and $P_{Q,M}=m_M$, and similarly for the right $Q$-modules. The common normalization of the structure maps ensures that the natural map $\Mor_{\mathrm{mod}\mhyph Q}(M, M') \times \Mor_{Q\mhyph\mathrm{mod}}(N, N') \to \mathcal{C}(M \otimes_Q N, M' \otimes_Q N')$ is compatible with the involution. The category $Q\mhyph\mathrm{mod}\mhyph Q$ then becomes a C$^*$-tensor category. This category is not strict on the nose: the associativity morphisms $\Phi^{Q}_{X,Y,Z}\colon (X\otimes_QY)\otimes_Q Z\to X\otimes_Q(Y\otimes_Q Z)$ are characterized by the identities $$\Phi^{Q}_{X,Y,Z}P_{X\otimes_Q Y,Z}(P_{X,Y}\otimes\iota_Z)=P_{X, Y\otimes_Q Z}(\iota_X\otimes P_{Y,Z}).$$
Note that our normalization of the structure maps $P$ implies that $\Phi^Q$ is unitary, as needed in C$^*$-tensor categories. As is common, from now on we will ignore the associativity morphisms and work with the category $Q\mhyph\mathrm{mod}\mhyph Q$ as if it was strict. Since by assumption $Q$ is simple as a $Q$-bimodule, the tensor unit in $Q\mhyph\mathrm{mod}\mhyph Q$ is simple.
Following M\"{u}ger~\cite{MR1966524}, we say that $\mathcal{C}$ and $Q\mhyph\mathrm{mod}\mhyph Q$ are \emph{weakly monoidally Morita equivalent}.
The \emph{dual $Q$-system} is given by the object $\hat{Q} = Q \otimes Q$ in $Q\mhyph\mathrm{mod}\mhyph Q$, together with the morphisms $$
\hat{v} = d(Q)^{-{1/2}} w\ \ \text{and}\ \ \hat{w} = d(Q)^{{1/2}} \iota \otimes v \otimes \iota
$$
under the identification of $\hat{Q} \otimes_Q \hat{Q}$ with $Q^{\otimes 3}$. Using that the unit of $\mathcal{C}$ is simple it is easy to check that this $Q$-system is simple (even without the simplicity assumption on $Q$; see also Proposition~\ref{prop:hat-Q-bimod-eqv-C} below). As we will see shortly, the dimension of $\hat Q=Q\otimes Q$ in $Q\mhyph\mathrm{mod}\mhyph Q$ equals the dimension of $Q$ in $\mathcal{C}$. Since $\hat w^*\hat w=d(Q)\iota$, it follows that the $Q$-system $\hat Q$ is standard.
If $Q$ is irreducible, then $\hat Q$ is also irreducible, as follows from the isomorphism
$$
\End_{Q\mhyph\mathrm{mod}}(Q)\cong\Mor_{Q\mhyph\mathrm{mod}\mhyph Q}(Q,Q\otimes Q), \ \ T\mapsto (T\otimes \iota)w.
$$
\subsection{Duality for \texorpdfstring{$Q$}{Q}-modules}
\label{sec:mod-cat-Frob-recip}
We continue to assume that $Q$ is a standard simple $Q$-system. Given any left $Q$-module $M$, its conjugate $\bar{M}$ has the natural structure of a right $Q$-module defined by
$m_{\bar M}=m_M^{\vee*}\in \mathcal{C}(\bar M\otimes Q,\bar M)$. Here, as usual, we fix a standard solution $(R_M,\bar R_M)$ of the conjugate equations for $M$ and then use it, together with the solution $(wv,wv)$ for $Q$, to define solutions of the conjugate equations for tensor products. A direct computation shows that $w^\vee=w^*$ and $v^\vee=v^*$, which then immediately implies by applying $\vee$ to the identities involving $m_M$ that $m_{\bar M}$ indeed defines the structure of a right $Q$-module on~$\bar M$.
Note that by definition we have
$$
(m^\vee_M\otimes\iota_{\bar M}) R_M=(\iota_{\bar M}\otimes\iota_Q\otimes m_M)(\iota_{\bar M}\otimes wv\otimes\iota_M)R_M=(\iota_{\bar M}\otimes m_M^*)R_M.
$$
Therefore $m_{\bar M}$ is characterized by the identity
\begin{equation} \label{eq:dualmod}
R_M^*(m_{\bar M}\otimes\iota)=R_M^*(\iota\otimes m_M)\ \ \text{on}\ \ \bar M\otimes Q\otimes M.
\end{equation}
If $M$ is a right $Q$-module, we can define the structure of a left $Q$-module on $\bar{M}$ in the same way by $m_{\bar{M}}^l = (m_M^r)^{\vee *}$. If $M$ is a $Q$-bimodule, then $\bar M$ becomes a $Q$-bimodule.
\smallskip
Now, assuming again that $M$ is a left $Q$-module, $M \otimes \bar{M}$ has the natural structure of a $Q$-bimodule. Since $m_M\colon Q\otimes M\to M$ is a left $Q$-module map, the morphism
$$
\bar{S}_M = (m_M \otimes \iota) (\iota \otimes \bar{R}_M)\in \mathcal{C}(Q, M \otimes \bar{M})
$$
is a left $Q$-module map. Note that by \eqref{eq:dualmod} we have $m_M=(\iota\otimes R^*_M)(\iota\otimes m_{\bar M}\otimes\iota)(\bar R_M\otimes\iota\otimes\iota)$, which implies that $\bar S_M=(\iota \otimes m_{\bar{M}}) (\bar{R}_M \otimes \iota)$. Hence $\bar S_M$ is also a morphism of right $Q$-modules.
Next, by \eqref{eq:dualmod} the morphism $R_M^*$ descends to $\bar{M} \otimes_Q M$. Let us denote the induced morphism $\bar{M} \otimes_Q M \to \mathds{1}$ by $[R_M^*]$.
We then have the following duality between left and right $Q$-modules.
\begin{lemma}
\label{lem:Q-mod-duality-1}
We have $(\iota \otimes [R_M^*]) (\bar{S}_M \otimes_Q \iota) = \iota_M$ and $([R_M^*] \otimes \iota) (\iota \otimes_Q \bar{S}_M) = \iota_{\bar{M}}$.
\end{lemma}
\begin{proof}
Since the situation is symmetric, we just give a proof of the first identity. The left hand side of this identity is the morphism $Q\otimes_Q M\to M$ defined by the morphism
$$
(\iota_M \otimes R_M^*) (\bar{S}_M \otimes\iota_M)\colon Q\otimes M\to M.
$$
As $\bar{S}_M = (m_M \otimes \iota) (\iota \otimes \bar{R}_M)$, the latter morphism is simply $m_M$, so it descends to the identity morphism $M=Q\otimes_Q M\to M$.
\end{proof}
This allows us to prove two versions of Frobenius reciprocity:
$$
\mathcal{C}(L \otimes_Q M, N) \cong \Mor_{\mathrm{mod}\mhyph Q}(L, N \otimes \bar{M})\ \ \text{and}\ \ \mathcal{C}(\bar{M} \otimes_Q L, N) \cong \Mor_{Q\mhyph\mathrm{mod}}(L, M \otimes N),
$$
where $L$ is a right $Q$-module in the first case and a left $Q$-module in the second. Namely, the first isomorphism is given by the map
$T \mapsto (T \otimes \iota) (\iota \otimes_Q \bar{S}_M)$, and its inverse is given by
$T' \mapsto (\iota \otimes R_M^*) (T' \otimes_Q \iota)$.
The second isomorphism is defined similarly.
\smallskip
Now, suppose that $M$ is a $Q$-bimodule in $\mathcal{C}$, so that $\bar{M}$ is also a $Q$-bimodule.
\begin{lemma} \label{lem:Sdescent}
The morphism $\bar S_M^*\colon M\otimes\bar M\to Q$ descends to a morphism $[\bar S^*_M]\colon M\otimes_Q\bar M\to Q$ of $Q$-bimodules.
\end{lemma}
\begin{proof} We have to check that $\bar S^*_M(m^r_M\otimes\iota_{\bar M})=\bar S^*_M(\iota_M\otimes m^l_{\bar M})$. By the characterization
$\bar R^*_M(m^r_M\otimes\iota)=\bar R^*_M(\iota\otimes m^l_{\bar M})$, we compute:
\begin{multline*}
\bar S^*_M(\iota_M\otimes m^l_{\bar M})= (\iota_Q \otimes \bar{R}_M^*)(m^{l*}_M \otimes \iota_{\bar M})(\iota_M\otimes m^l_{\bar M})\\
=(\iota_Q \otimes \bar{R}_M^*)(\iota_Q\otimes\iota_M\otimes m^l_{\bar M})(m^{l*}_M\otimes\iota_Q \otimes \iota_{\bar M})
=(\iota_Q \otimes \bar{R}_M^*)(\iota_Q\otimes m^r_M\otimes\iota_{\bar M})(m^{l*}_M\otimes\iota_Q \otimes \iota_{\bar M}).
\end{multline*}
Therefore, as $\bar S^*_M(m^r_M\otimes\iota_{\bar M})=(\iota_Q \otimes \bar{R}_M^*)(m^{l*}_M\otimes\iota_{\bar M})(m^r_M\otimes\iota_{\bar M})$, we have to show that
\begin{equation*}\label{eq:mrml}
m^{l*}_Mm^r_M=(\iota\otimes m^r_M)(m^{l*}_M\otimes\iota).
\end{equation*}
But this identity means simply that $m_M^{l*}$ is a morphism of right $Q$-modules, so the lemma is proved.
\end{proof}
We can now define duality morphisms for $Q$-bimodules. First, similarly to $\bar S_M$, define
$$
S_M=(m^l_{\bar M}\otimes\iota)(\iota\otimes R_{M})=(\iota\otimes m^r_M)(R_M\otimes\iota) \colon Q\to \bar M\otimes M.
$$
Then put
$$
R^Q_M=d(Q)^{-1}P_{\bar M,M}S_M\ \ \text{and}\ \ \bar R^Q_M=d(Q)^{-1}P_{M,\bar M}\bar S_M.
$$
By Lemma~\ref{lem:Sdescent} we can write $\bar S_M^*=[\bar S^*_M]P_{M,\bar M}$, and $(P_{M,\bar M}\bar S_M)^*=[\bar S^*_M]P_{M,\bar M}P^*_{M,\bar M}=d(Q)[\bar S^*_M]$. Similar identities hold for $S_M$, so we get
\begin{equation}\label{eq:RQstar}
\bar R^{Q*}_MP_{M,\bar M}=\bar S^*_M\ \ \text{and}\ \ R^{Q*}_MP_{\bar M,M}=S^*_M.
\end{equation}
\begin{lemma}
\label{lem:S-M-conjug-sol}
The pair $(R^Q_M,\bar R^Q_M)$ solves the conjugate equations for $M$ in $Q\mhyph\mathrm{mod}\mhyph Q$.
\end{lemma}
\begin{proof}
We will only show that $(\bar R^{Q*}_M\otimes_Q\iota)(\iota\otimes_Q R^Q_M)=\iota$. We have
\begin{align*}
(\bar R^{Q*}_M\otimes_Q\iota)(\iota\otimes_Q R^Q_M)P_{M,Q} &=d(Q)^{-1}(\bar R^{Q*}_M\otimes_Q\iota)
P_{M,\bar M\otimes_Q M}(\iota\otimes P_{\bar M,M}S_M)\\
&=d(Q)^{-1}P_{Q,M}(\bar R^{Q*}_MP_{M,\bar M}\otimes\iota)(\iota\otimes S_M)\\
&=d(Q)^{-1}P_{Q,M}(\bar S_M^*\otimes\iota)(\iota\otimes S_M),
\end{align*}
where we used the associativity $P_{M, \bar{M}\otimes_Q M} (\iota_M \otimes P_{\bar{M},M}) = P_{M \otimes_Q \bar{M}, M} (P_{M,\bar{M}} \otimes \iota)$ of the tensor product and then that $\bar R^{Q*}_MP_{M,\bar M}=\bar S_M^*$
by~\eqref{eq:RQstar}. We have
\begin{align*}
(\bar{S}_M^* \otimes \iota) (\iota \otimes S_M)&=
(\iota_Q\otimes \bar R_{M}^*\otimes\iota_M)(m^{l*}_M\otimes\iota_{\bar M}\otimes\iota_M)
(\iota_M\otimes\iota_{\bar M}\otimes m^r_M)(\iota_M\otimes R_M\otimes\iota_Q)\\
&=(\iota_Q\otimes m^r_M)(m^{l*}_M\otimes\iota_Q)=m^{l*}_Mm^r_M.
\end{align*}
Therefore, using that $P_{Q,M}=m^l_M$, $P_{M,Q}=m^r_M$ and $m^l_Mm^{l*}_M=d(Q)\iota$ by Lemma~\ref{lem:comput-m-m-star}, we get
$$
(\bar R^{Q*}_M\otimes_Q\iota)(\iota\otimes_Q R^Q_M)P_{M,Q}
=d(Q)^{-1}P_{Q,M}m^{l*}_Mm^r_M=m^r_M=P_{M,Q}.
$$
Hence $(\bar R^{Q*}_M\otimes_Q\iota)(\iota\otimes_Q R^Q_M)=\iota_M$.
\end{proof}
We thus see that $Q\mhyph\mathrm{mod}\mhyph Q$ is a rigid C$^*$-tensor category. Let us decorate the constructions related to $Q\mhyph\mathrm{mod}\mhyph Q$, such as the categorical trace, the dimension function, and so on, with superindex $Q$.
\begin{proposition}
\label{prop:S-M-resc-std-sol}
The solution $(R^Q_M,\bar R^Q_M)$ of the conjugate equations for $M$ in $Q\mhyph\mathrm{mod}\mhyph Q$ is standard and $d^Q(M)=d(Q)^{-1}d(M)$. Furthermore,
for any $T\in\End_{Q\mhyph\mathrm{mod}\mhyph Q}(M)$ we have $\Tr^Q_M(T)=d(Q)^{-1}\Tr_M(T)$.
\end{proposition}
\begin{proof}
For any $T\in\End_{Q\mhyph\mathrm{mod}\mhyph Q}(M)$ we have
$$
\bar R^{Q*}_M(T\otimes_Q\iota)\bar R^{Q}_M=d(Q)^{-1}\bar R^{Q*}_M(T\otimes_Q\iota)P_{M,\bar M}\bar S_M
=d(Q)^{-1}\bar R^{Q*}_MP_{M,\bar M}(T\otimes\iota)\bar S_M=d(Q)^{-1}\bar S^*_M(T\otimes\iota)\bar S_M,
$$
where in the last step we used \eqref{eq:RQstar}. This expression is a scalar endomorphism of $Q$, so in order to detect this scalar we can compose it with $v^*$ on the right and $v$ on the left. But $\bar S_M v=\bar R_M$, hence
$$
\bar R^{Q*}_M(T\otimes_Q\iota)\bar R^{Q}_M=d(Q)^{-1}\bar R^*_M(T\otimes\iota) R_M=d(Q)^{-1}\Tr_M(T).
$$
By a similar argument $R^{Q*}_M(\iota\otimes_Q T) R^{Q}_M$ gives the same result. Hence the solution $(R^Q_M,\bar R^Q_M)$ is standard and $\Tr^Q_M(T)=d(Q)^{-1}\Tr_M(T)$.
\end{proof}
For $Q$-bimodules, the contravariant functor $T \to T^\vee$ can be a priori defined in two ways, using standard solutions either in $\mathcal{C}$ or in $Q\mhyph\mathrm{mod}\mhyph Q$. However, these two definitions give the same result. In fact, the following slightly more general statement is true.
\begin{lemma}
\label{lem:vee-functor-same}
Let $T \colon M \to N$ be a morphism of left $Q$-modules. If $T^\vee \in \mathcal{C}(\bar{N}, \bar{M})$ satisfies $(T \otimes \iota) \bar{R}_M = (\iota \otimes T^\vee) \bar{R}_N$, then $T^\vee$ is a morphism of right $Q$-modules and $(T \otimes_Q \iota) P_{M,\bar{M}} \bar{S}_M = (\iota \otimes_Q T^\vee) P_{N,\bar{N}} \bar{S}_N$.
\end{lemma}
\begin{proof}
It is immediate that $T^\vee$ is a morphism of right $Q$-modules.
In order to show the second claim it is enough to establish the identity $(T \otimes \iota) \bar{S}_M = (\iota \otimes T^\vee) \bar{S}_N$. We have
\begin{multline*}
(\iota_N\otimes T^\vee)\bar S_N=(\iota_N\otimes T^\vee)(m_N\otimes\iota_{\bar N})(\iota_Q\otimes\bar R_N)
=(m_N\otimes\iota_{\bar M})(\iota_Q\otimes T\otimes\iota_{\bar M})(\iota_Q\otimes\bar R_M)\\
=(T\otimes\iota_{\bar M})(m_M\otimes\iota_{\bar M})(\iota_Q\otimes\bar R_M)
=(T\otimes\iota_{\bar M})\bar S_M,
\end{multline*}
and the lemma is proved.
\end{proof}
\subsection{Schauenburg's induction}
\label{sec:schauen-induc}
If $\Irr(\mathcal{C})$ is finite and $Q$ is generating, in which case $Q$ corresponds to a finite depth subfactor, Ocneanu observed that $\mathcal{Z}(\mathcal{C})$ and $\mathcal{Z}(Q\mhyph\mathrm{mod}\mhyph Q)$ have the same fusion rules through a characterization of the fusion rules in terms of the associated TQFT~\cite{MR1642584}*{p.~641 and Theorem~12.29}. In fact, by a result of Schauenburg~\cite{MR1822847} this can be strengthened to an equivalence of tensor categories $\mathcal{Z}(\mathcal{C}) \cong \mathcal{Z}(Q\mhyph\mathrm{mod}\mhyph Q)$ without the finiteness assumption on $\mathcal{C}$ and the generating assumption on $Q$. Although this result is stated in the algebraic context without involution, it can be further extended to our framework of ind-C$^*$-tensor categories in a straightforward way.
Before we explain this, let us remark that we can identify the categories $\indcat{Q\mhyph\mathrm{mod}_\mathcal{C}}$ and $Q\mhyph\mathrm{mod}_{\indcat{\CC}}$. Indeed, for any object $M$ in $Q\mhyph\mathrm{mod}_{\indcat{\CC}}$ the morphism $d(Q)^{-1/2}m_M^*$ gives an isometric $Q$-modular embedding of $M$ into $Q\otimes M$. Since $Q\otimes M$ lies in $\indcat{Q\mhyph\mathrm{mod}_\mathcal{C}}$ and this category is closed under subobjects, it follows that~$M$ is isomorphic to an object in $\indcat{Q\mhyph\mathrm{mod}_\mathcal{C}}$. The same remark applies to $Q$-bimodules.
Turning to the equivalence between ${\Dcen(\indC)}$ and $\mathcal{Z}(\indcatQ\mhyph\mathrm{mod}\mhyph Q)$, the starting point is that, as observed by Schauenburg, if $c$ is a unitary half-braiding on an ind-object $X$ of $\mathcal{C}$, then $X \otimes Q$ has the structure of a $Q$-bimodule in $\indcat{\CC}$, hence $X\otimes Q$ can be considered as an ind-object of $Q\mhyph\mathrm{mod}\mhyph Q$.
\begin{lemma}
The pair $(m^r, m^l) = (\iota \otimes m_Q, (\iota \otimes m_Q) (c_Q \otimes \iota))$ defines the structure of a $Q$-bimodule on $X \otimes Q$.
\end{lemma}
\begin{proof}
Let us concentrate on the $*$-compatibility of $m^l$, since this is the only new property compared to~\cite{MR1822847}. We have $m^{l*}=(c_Q^* \otimes \iota_Q) (\iota_X \otimes w)$ and we want to show that this is equal to $(\iota_Q \otimes m^l)(wv\otimes\iota_X\otimes\iota_Q)$. Since $\iota\otimes c_Q=(c_Q^*\otimes\iota)c_{Q\otimes Q}$, we have
$$
(\iota_Q \otimes m^l)(wv\otimes\iota_X\otimes\iota_Q) = (c_Q^* \otimes m_Q) (c_{Q \otimes Q} \otimes \iota_Q) (w v \otimes \iota_X\otimes\iota_Q).
$$
Using the naturality of the half-braiding, this is equal to $(c_Q^* \otimes m_Q) (\iota_X \otimes w v \otimes \iota_Q)$, which is indeed equal to $(c_Q^* \otimes \iota_Q) (\iota_X \otimes w)$ by the $*$-compatibility of $m_Q$.
\end{proof}
Next, we define a unitary half-braiding $\tilde c$ on $X\otimes Q\in\indcat{Q\mhyph\mathrm{mod}\mhyph Q}$. Let $Y$ be a $Q$-bimodule. Then $Y \otimes X$ and $X \otimes Y$ are models of $Y \otimes_Q (Q \otimes X)$ and $(X \otimes Q) \otimes_Q Y$, with the structure morphisms of the tensor product given by $m_Y^r \otimes \iota_X$ and $\iota_X \otimes m_Y^l$. Since $c_Q$ is by definition a morphism of left $Q$-modules, it induces a unitary morphism $Y \otimes_Q (Q \otimes X) \to Y \otimes_Q (X \otimes Q)$, so we get an isomorphism $Y \otimes_Q (X \otimes Q)\cong Y\otimes X$. Then, up to these isomorphisms, we define $\tilde{c}_Y\colon Y \otimes_Q (X \otimes Q) \to (X \otimes Q) \otimes_Q Y$ simply as the morphism $c_Y\colon Y\otimes X\to X\otimes Y$.
\begin{lemma} \label{lem:schaun-braiding}
The unitaries $\tilde c_Y$ form a half-braiding on $X\otimes Q\in\indcat{Q\mhyph\mathrm{mod}\mhyph Q}$.
\end{lemma}
\begin{proof}
Since the proof of the corresponding statement in~\cite{MR1822847} is omitted, let us briefly indicate the argument. The $Q$-bimodule structure on $Y\otimes X\cong Y\otimes_Q(X\otimes Q)$ is given by
$$
m^l_{Y\otimes X}=m^l_Y\otimes\iota_X, \ \ m^r_{Y\otimes X}=(m^r_Y\otimes\iota_X)(\iota_Y\otimes c_Q^*),
$$
and similarly the $Q$-bimodule structure on $X\otimes Y\cong (X\otimes Q)\otimes_Q Y$ is given by
$$
m^l_{X\otimes Y}=(\iota_X\otimes m^l_Y)(c_Q\otimes\iota_Y),\ \ m^r_{X\otimes Y}=\iota_X\otimes m^r_Y.
$$
Using this it is easy to check that the morphism $c_Y\colon Y\otimes X\to X\otimes Y$ is $Q$-bimodular.
Next, the morphism $\tilde c_Y\colon Y\otimes_Q(X\otimes Q)\to (X\otimes Q)\otimes_Q Y$ is defined by the morphism
$$
\sigma_Y=(\iota_X\otimes v\otimes\iota_Y)c_Y(m^r_Y\otimes\iota_X)(\iota_Y\otimes c^*_Q)\colon Y\otimes X\otimes Q\to X\otimes Q\otimes Y.
$$
As $c_Y(m^r_Y\otimes\iota)=(\iota\otimes m^r_Y)(c_Y\otimes\iota)(\iota\otimes c_Q)$, we have
$$
\sigma_Y=(\iota_X\otimes v\otimes\iota_Y)(\iota_X\otimes m^r_Y)(c_Y\otimes\iota_Q).
$$
From this we deduce that $(\sigma_Z\otimes\iota_Y)(\iota_Z\otimes\sigma_Y)=\sigma_{Z\otimes Y}$. By the naturality of $c$, we have $(\iota_{X\otimes Q}\otimes P_{Z,Y})\sigma_{Z\otimes Y}=\sigma_{Z\otimes_QY}(P_{Z,Y}\otimes\iota_{X\otimes Q})$. Hence $\tilde c$ satisfies $(\tilde c_Z\otimes_Q\iota_Y)(\iota_Z\otimes_Q\tilde c_Y)=\tilde c_{Z\otimes_Q Y}$.
\end{proof}
Thus, putting $F(X, c) = (X \otimes Q, \tilde{c})$, we obtain a C$^*$-tensor functor $F\colon{\Dcen(\indC)} \to \mathcal{Z}(\indcatQ\mhyph\mathrm{mod}\mhyph Q)$.
\smallskip
We will show that an inverse functor can be obtained by exactly the same construction using the dual $Q$-system $\hat Q$, modulo an equivalence of the categories $\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}$ and $\mathcal{C}$ that we are now going to explain.
Let $L \colon \mathcal{C} \to Q\mhyph\mathrm{mod}$ be the free module functor $U \mapsto Q \otimes U$, and $O \colon Q\mhyph\mathrm{mod} \to \mathcal{C}$ be the forgetful functor. As was already observed in Section~\ref{sec:prelim-q-system}, we have the adjunction
$$
\mathcal{C}(O (M), U) \cong \Mor_{Q\mhyph\mathrm{mod}}(M, L (U)),
$$
induced by the natural transformations
\begin{align*}
\eta_M &= d(Q)^{-{1/2}} m^{l*}_M\colon M \to L O (M) = Q \otimes M,&
\epsilon_U &= d(Q)^{{1/2}} v^* \otimes \iota_U\colon OL(U)\to U.
\end{align*}
The composition $T = L O$ has the structure of a \emph{monad}~\cite{MR1712872}*{Chapter~VI}. With the above normalization of~$\eta$ and~$\epsilon$, the multiplication $\mu\colon T^2\to T$ is given by $\hat w^*\otimes_Q\iota_M\colon \hat Q\otimes_Q\hat Q\otimes_Q M=Q\otimes Q\otimes M\to \hat Q\otimes_Q M=Q\otimes M$. Then a $T$-algebra structure $\hat{m}_M \colon T( M)=\hat Q\otimes_Q M = Q \otimes M\to M$ on $M \in Q\mhyph\mathrm{mod}$ is precisely an algebraic left $\hat{Q}$-module structure on $M$. We consider only $T$-algebras satisfying the $*$-compatibility condition, which can be written as $\hat{m} = \eta^* \mu T(\hat{m}^*)$. Then the category $(Q\mhyph\mathrm{mod})^T$ of $T$-algebras is the category $\hat{Q}\mhyph\mathrm{mod}_{Q\mhyph\mathrm{mod}}$ of left $\hat Q$-modules in $Q\mhyph\mathrm{mod}$.
As observed before, the categories $\mathcal{C}$ and $Q\mhyph\mathrm{mod}$ both have coequalizers, and the Frobenius reciprocity implies that a coequalizer in $Q\mhyph\mathrm{mod}$ is also a one in $\mathcal{C}$. This means that the functor $L$ creates coequalizers, and Beck's theorem~\cite{MR1712872}*{Section~VI.7} implies that $(Q\mhyph\mathrm{mod})^T$ is equivalent
to $\mathcal{C}$. More precisely, the comparison functor $P\colon (Q\mhyph\mathrm{mod})^T \to \mathcal{C}$ is characterized as the coequalizer diagram $Q \otimes W \rightrightarrows W \to P (W)$ for the morphisms $d(Q)^{{1/2}} v^* \otimes \iota_W$ and $\hat{m}_W$ for $W \in (Q\mhyph\mathrm{mod})^T$ and the inverse functor $\mathcal{C}\to (Q\mhyph\mathrm{mod})^T=\hat{Q}\mhyph\mathrm{mod}_{Q\mhyph\mathrm{mod}}$ is given by $U\mapsto Q\otimes U$.
The above considerations can be carried out for the right $\hat{Q}$-modules in $\mathrm{mod}\mhyph Q$ and the right and/or left $\hat{Q}$-modules in $Q\mhyph\mathrm{mod}\mhyph Q$, which leads to the equivalences of C$^*$-categories
\begin{align*}
\mathrm{mod}_{\mathrm{mod}\mhyph Q}\mhyph\hat{Q} &\cong \mathcal{C},&
{\hat{Q}\mhyph\mathrm{mod}_{Q\mhyph\mathrm{mod}\mhyph Q}} &\cong \mathrm{mod}\mhyph Q,&
\hat Q\mhyph\mathrm{mod}_{Q\mhyph\mathrm{mod}\mhyph Q}\mhyph\hat Q &\cong \mathcal{C}.
\end{align*}
We of course also have similar equivalences for the ind-categories.
\begin{proposition}\label{prop:hat-Q-bimod-eqv-C}
The equivalence $\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q} \cong \mathcal{C}$ of C$^*$-categories can be extended to an equivalence of C$^*$-tensor categories.
\end{proposition}
\begin{proof}
The equivalence $\mathcal{C} \to \hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}$ is given by $X \mapsto Q \otimes X \otimes Q$ at the level of objects. We have to show that $Q \otimes X \otimes Y \otimes Q$ becomes a model of $(Q \otimes X \otimes Q) \otimes_{\hat{Q}} (Q \otimes Y \otimes Q)$ in a natural way.
The right $\hat{Q}$-module structure on $Q \otimes X \otimes Q$ is given by $d(Q)^{1/2}\iota_{Q \otimes X} \otimes v^* \otimes \iota_Q$, where we as usual use $Q \otimes X \otimes Q \otimes Q$ as a model of $(Q \otimes X \otimes Q) \otimes_Q \hat{Q}$ with the structure morphism $P_{Q \otimes X \otimes Q, \hat{Q}} = \iota_{Q \otimes X} \otimes m_Q \otimes \iota_Q$. The left $\hat{Q}$-module structure on $Q \otimes Y \otimes Q$ can be described in a similar way, and we also have
$$
(Q \otimes X \otimes Q) \otimes_Q \hat{Q} \otimes_Q (Q \otimes Y \otimes Q) = Q \otimes X \otimes Q \otimes Q \otimes Y \otimes Q.
$$
Thus, $(Q \otimes X \otimes Q) \otimes_{\hat{Q}} (Q \otimes Y \otimes Q)$ is a coequalizer of $\iota_{Q \otimes X} \otimes v^* \otimes \iota_{Q \otimes Y \otimes Q}$ and $\iota_{Q \otimes X \otimes Q} \otimes v^* \otimes \iota_{Y \otimes Q}$. Since the endofunctors on $\mathcal{C}$ of the form $Z \mapsto V \otimes Z \otimes W$ for $V, W \in \mathcal{C}$ are exact, we thus see that it is enough to show that $v^*\colon Q\to\mathds{1}$ is a coequalizer of the morphisms $v^* \otimes \iota_Q$ and $\iota_Q \otimes v^*$. But this is obvious as $v^*v=\iota$.
\end{proof}
Now, the same construction as for $F$ using the dual $Q$-system $\hat{Q}$ provides a C$^*$-tensor functor $$\tilde{F}\colon \mathcal{Z}(\indcat{Q\mhyph\mathrm{mod}\mhyph Q}) \to \mathcal{Z}(\indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}}).$$
Take $(X,c)\in{\Dcen(\indC)}$. Then $F(X,c)=(X\otimes Q,\tilde c)$ and $\tilde F F(X,c)=((X\otimes Q)\otimes_Q\hat Q,\tilde{\tilde c})$. We can identify $(X\otimes Q)\otimes_Q\hat Q$ with $X\otimes\hat Q=X\otimes Q\otimes Q$. Under this identification the right $Q$- and $\hat Q$-module structures are obvious, but this is less so for the left module structures.
\begin{lemma}
The morphism $c_Q\otimes\iota_Q\colon Q\otimes X\otimes Q\to X\otimes Q\otimes Q=(X\otimes Q)\otimes_Q\hat Q$ is an isomorphism of $Q$- and $\hat Q$-bimodules.
\end{lemma}
\begin{proof} We only have to consider the left module structures. As was already stated in the proof of Lemma~\ref{lem:schaun-braiding}, the morphism $c_{\hat Q}\colon \hat Q\otimes X\to X\otimes\hat Q= (X\otimes Q)\otimes_Q\hat Q$ is a morphism of $Q$-bimodules. Furthermore, since the isomorphism
$\hat Q\otimes_Q(X\otimes Q)\cong \hat Q\otimes X$ defined by the isomorphisms $c^*_Q\colon X\otimes Q\to Q\otimes X$ and $\hat Q\otimes_Q Q\cong \hat Q$ does not affect the first factor of $\hat Q=Q\otimes Q$, by definition of the left $\hat Q$-module structure on $(X\otimes Q)\otimes_Q\hat Q$ we also see that $c_{\hat Q}\colon \hat Q\otimes X\to X\otimes\hat Q= (X\otimes Q)\otimes_Q\hat Q$ is a morphism of left $\hat Q$-modules. Therefore in order to prove the lemma it suffices to check that $c^*_{\hat Q}(c_Q\otimes\iota_Q)\colon Q\otimes X\otimes Q\to\hat Q\otimes X$ is a morphism of left $Q$- and $\hat Q$-modules. But this is obvious as $c^*_{\hat Q}(c_Q\otimes\iota_Q)=\iota_Q\otimes c_Q^*$.
\end{proof}
The equivalence $\indcat{\CC}\cong \indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}}$, $X\mapsto Q\otimes X\otimes Q$, defines an equivalence of the corresponding Drinfeld centers ${\Dcen(\indC)}\cong\mathcal{Z}(\indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}})$.
\begin{lemma}
The functor $\tilde F F$ is naturally unitarily isomorphic to the equivalence functor ${\Dcen(\indC)}\to\mathcal{Z}(\indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}})$.
\end{lemma}
\begin{proof} We have to find out what happens with the half-braiding $\tilde{\tilde c}$ under the isomorphism $(X\otimes Q)\otimes_Q\hat Q=X\otimes\hat Q\cong Q\otimes X\otimes Q$ from the previous lemma.
Take an object $Y\in\mathcal{C}$. Consider the morphism
$$
\tilde {\tilde c}_{Q\otimes Y\otimes Q}\colon (Q\otimes Y\otimes Q)\otimes_{\hat Q}((X\otimes Q)\otimes_Q\hat Q)\to
((X\otimes Q)\otimes_Q\hat Q)\otimes_{\hat Q}(Q\otimes Y\otimes Q).
$$
If we identify the right hand side with $(X\otimes Q)\otimes_Q(Q\otimes Y\otimes Q)$, then by the proof of Lemma~\ref{lem:schaun-braiding} this morphism is implemented by the morphism
$$
\tilde c_{Q\otimes Y\otimes Q}\colon (Q\otimes Y\otimes Q)\otimes_{Q}(X\otimes Q)\to
(X\otimes Q)\otimes_Q(Q\otimes Y\otimes Q)
$$
together with the multiplication map $(Q\otimes Y\otimes Q)\otimes_Q\hat Q\to Q\otimes Y\otimes Q$ defining the right $\hat Q$-module structure on $Q\otimes Y\otimes Q$. Similarly, if we identify $(X\otimes Q)\otimes_Q(Q\otimes Y\otimes Q)$ with $X\otimes Q\otimes Y\otimes Q$, then $\tilde c_{Q\otimes Y\otimes Q}$ is implemented by the morphism $c_{Q\otimes Y\otimes Q}$ together with the multiplication map $Q\otimes Y\otimes Q\otimes Q\to Q\otimes Y\otimes Q$. To summarize, if we identify $((X\otimes Q)\otimes_Q\hat Q)\otimes_{\hat Q}(Q\otimes Y\otimes Q)$ with $X\otimes Q\otimes Y\otimes Q$, then $\tilde {\tilde c}_{Q\otimes Y\otimes Q}$ is induced by the morphism
$$
d(Q)^{1/2}(\iota_{X\otimes Q\otimes Y}\otimes v^*w^{(2)*}\otimes\iota_Q)(c_{Q\otimes Y\otimes Q}\otimes\iota_{Q\otimes\hat Q})\colon Q\otimes Y\otimes Q\otimes X\otimes Q\otimes\hat Q\to X\otimes Q\otimes Y\otimes Q,
$$
where $w^{(2)}=(w\otimes\iota)w=(\iota\otimes w)w$. Identifying $(X\otimes Q)\otimes_Q\hat Q$ with $X\otimes\hat Q$, we can equivalently write that $\tilde {\tilde c}_{Q\otimes Y\otimes Q}$ is induced by
$$
d(Q)^{1/2}(\iota_{X\otimes Q\otimes Y}\otimes v^*w^*\otimes\iota_Q)(c_{Q\otimes Y\otimes Q}\otimes\iota_{\hat Q})\colon Q\otimes Y\otimes Q\otimes X\otimes\hat Q\to X\otimes Q\otimes Y\otimes Q,
$$
The half-braiding $\tilde{\tilde c}$ on $(X\otimes Q)\otimes_Q\hat Q=X\otimes\hat Q$ defines a half-braiding $c'$ on $Q\otimes X\otimes Q\cong X\otimes \hat Q$. Then we conclude that $c'_{Q\otimes Y\otimes Q}$ is implemented by the morphism $Q\otimes Y\otimes Q\otimes Q\otimes X\otimes Q\to Q\otimes X\otimes Y\otimes Q$ given by
\begin{multline*}
d(Q)^{1/2}(c_Q^*\otimes\iota_{Y\otimes Q})(\iota_{X\otimes Q\otimes Y}\otimes v^*w^*\otimes\iota_Q)(c_{Q\otimes Y\otimes Q}\otimes\iota_{\hat Q})(\iota_{Q\otimes Y\otimes Q}\otimes c_Q\otimes\iota_Q)\\
=d(Q)^{1/2}(\iota_Q\otimes c_Y\otimes\iota_Q)(\iota_{Q\otimes Y}\otimes v^*w^*\otimes\iota_{X\otimes Q}).
\end{multline*}
Since $d(Q)^{1/2}(\iota_{Q\otimes Y}\otimes v^*w^*\otimes\iota_{X\otimes Q})$ is precisely the structure morphism
$Q\otimes Y\otimes Q\otimes Q\otimes X\otimes Q\to (Q\otimes Y\otimes Q)\otimes_{\hat Q}(Q\otimes X\otimes Q)=Q\otimes Y\otimes X\otimes Q$, we thus see that $c'$ coincides with the half-braiding on $Q\otimes X\otimes Q$ defined by the half-braiding $c$ on $X$ using the equivalence functor $\indcat{\CC}\to \indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}}$.
\end{proof}
It follows that $F$ defines an equivalence between the category ${\Dcen(\indC)}$ and a subcategory of the category $\mathcal{Z}(\indcat{\QmodQ})$. For the same reason $\tilde F$ defines an equivalence between $\mathcal{Z}(\indcat{\QmodQ})$ and a subcategory of $\mathcal{Z}(\indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}})$. We then conclude that $F$ is an equivalence of the categories ${\Dcen(\indC)}$ and $\mathcal{Z}(\indcat{\QmodQ})$ and, modulo the equivalence ${\Dcen(\indC)} \cong \mathcal{Z}(\indcat{\hat{Q}\mhyph\mathrm{mod}\mhyph\hat{Q}})$, $\tilde F$ is an inverse functor. We thus have the following version of the result of Schauenburg~\cite{MR1822847}.
\begin{theorem}
The functor $F\colon{\Dcen(\indC)}\to\mathcal{Z}(\indcat{\QmodQ})$, $(X,c)\mapsto (X\otimes Q,\tilde c)$, is a unitary monoidal equivalence of categories.
\end{theorem}
We remark that Schauenburg's result does not give an equivalence of the categories ${\Dcen_s(\indC)}$ and $\mathcal{Z}_s(\indcat{\QmodQ})$ of spherical objects, or in other words, an equivalence of the representation categories of~$C^*(\mathcal{C})$ and $C^*(Q\mhyph\mathrm{mod}\mhyph Q)$. These categories are not equivalent in general. But the result does give rise to functors $\tilde F\colon\Rep C^*(\mathcal{C})\to\Rep C^*(Q\mhyph\mathrm{mod}\mhyph Q)$ and $\tilde G\colon\Rep C^*(Q\mhyph\mathrm{mod}\mhyph Q)\to \Rep C^*(\mathcal{C})$ such that $\tilde F\tilde G$ and $\tilde G\tilde F$ map every representation to a subrepresentation.
\subsection{Comparison of almost invariant vectors}
If $(X,c) \in {\Dcen(\indC)}$, the invariant vectors for the fusion algebra of $\mathcal{C}$ are the vectors in $\Mor_{\Dcen(\indC)}(\mathds{1}, X)$. Since Schauenburg's induction is a C$^*$-tensor functor, they correspond to the vectors in $\Mor_{\mathcal{Z}(\indcat{\QmodQ})}(Q, X \otimes Q)$, or the invariant vectors for the fusion algebra of $Q\mhyph\mathrm{mod}\mhyph Q$. It is less obvious what happens with almost invariant vectors, since this is a notion that does not make sense within the Drinfeld center itself.
\begin{theorem}
\label{thm:almost-inv-vec-mor-eqv}
For any $(X,c) \in {\Dcen(\indC)}$, the representation of $C^*(\mathcal{C})$ defined by $(X,c)$ weakly contains the trivial representation if and only if the representation of $C^*(Q\mhyph\mathrm{mod}\mhyph Q)$ defined by the image $F(X,c)=(X\otimes Q,\tilde c)$ of $(X,c)$ under Schauenburg's induction weakly contains the trivial representation.
\end{theorem}
For the proof we need to understand the morphisms $Q\to X\otimes Q$ in $\indcat{\QmodQ}$.
\begin{lemma} \label{lem:char-vec-sp-F-X}
For any $(X, c) \in {\Dcen(\indC)}$, the morphisms in $\Mor_{\indcat{\QmodQ}}(Q, X \otimes Q)$ are of the form
$$
T^S=(\iota_X\otimes v^*w^*\otimes\iota_Q)(c_Q\otimes\iota_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q\otimes\iota_Q)w^{(3)},
$$
where $w^{(n)}=(w\otimes\iota)w^{(n-1)}$ ($w^{(1)}=w$) and $S\in\Mor_\indcat{\CC}(Q,X)$.
\end{lemma}
\begin{proof}
The right $Q$-modular morphisms $Q\to X\otimes Q$ have the form $(T\otimes\iota)w$. By definition, such a morphism respects the left actions of $Q$ if and only if
\begin{equation}\label{eq:bimod-hom-char-in-orig}
(T\otimes\iota_Q)w w^*=(\iota_X \otimes w^*)(c_Q \otimes \iota_Q) (\iota_Q \otimes T \otimes \iota_Q) (\iota_Q \otimes w).
\end{equation}
Therefore in order to prove the lemma it suffices to show that a morphism $T$ satisfies \eqref{eq:bimod-hom-char-in-orig} if and only if it has the form
$$
T=(\iota_X\otimes v^*w^*)(c_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q)w^{(2)}
$$
for some $S$. If $T$ satisfies \eqref{eq:bimod-hom-char-in-orig}, then multiplying \eqref{eq:bimod-hom-char-in-orig} by $\iota\otimes v^*$ on the left and by $w$ on the right we see that~$T$ indeed has the above form, with $S=d(Q)^{-1}T$. Conversely, let $T$ be as above. Then
\begin{align*}
(T\otimes\iota_Q)w w^*&=(\iota_X\otimes v^*w^*\otimes\iota_Q)(c_Q\otimes\iota_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q\otimes\iota_Q)w^{(3)}w^*\\
&=(\iota_X\otimes v^*w^*\otimes\iota_Q)(\iota_X\otimes\iota_Q\otimes w)(c_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q)w^{(2)}w^*\\
&=(\iota_X\otimes w^*)(c_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q)w^{(2)}w^*.
\end{align*}
On the other hand,
\smallskip
$\displaystyle
(\iota_X \otimes w^*)(c_Q \otimes \iota_Q) (\iota_Q \otimes T \otimes \iota_Q) (\iota_Q \otimes w)$
\begin{align*}
&=(\iota_X \otimes w^*)(c_Q \otimes \iota_Q) (\iota_Q\otimes\iota_X\otimes v^*w^*\otimes\iota_Q)(\iota_Q\otimes c_Q\otimes\iota_Q\otimes\iota_Q)(\iota_Q\otimes\iota_Q\otimes S\otimes\iota_Q\otimes\iota_Q)(\iota_Q \otimes w^{(3)})\\
&=(\iota_X \otimes w^*)(c_Q \otimes \iota_Q) (\iota_Q\otimes\iota_X\otimes w^*)(\iota_Q\otimes c_Q\otimes\iota_Q)(\iota_Q\otimes\iota_Q\otimes S\otimes\iota_Q)(\iota_Q \otimes w^{(2)})\\
&=(\iota_X \otimes w^*)(\iota_X\otimes\iota_Q\otimes w^*)(c_{Q\otimes Q}\otimes\iota_Q)(\iota_Q\otimes\iota_Q\otimes S\otimes\iota_Q)(\iota_Q \otimes w^{(2)})\\
&=(\iota_X \otimes w^*)(\iota_X\otimes w^*\otimes\iota_Q)(c_{Q\otimes Q}\otimes\iota_Q)(\iota_Q\otimes\iota_Q\otimes S\otimes\iota_Q)(\iota_Q \otimes w^{(2)})\\
&=(\iota_X \otimes w^*)(c_Q\otimes\iota_Q)(\iota_Q\otimes S\otimes\iota_Q)(w^*\otimes\iota_Q\otimes\iota_Q)(\iota_Q \otimes w^{(2)}).
\end{align*}
Since by applying the Frobenius condition twice we obtain
$$
w^{(2)}w^*=(w\otimes\iota)(w^*\otimes\iota)(\iota\otimes w)=(w^*\otimes\iota\otimes\iota)(\iota\otimes w^{(2)}),
$$
this proves the lemma.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:almost-inv-vec-mor-eqv}]
Let $\{\xi_i\}_i$ be a net of almost invariant unit vectors for $C^*(\mathcal{C})$ in $\Mor_\indcat{\CC}(\mathds{1},X)$. Recall that by Lemma~\ref{lem:aa} this means that $c_Y(\iota_Y\otimes\xi_i)-\xi_i\otimes \iota_Y\to0$ for all $Y\in\mathcal{C}$. Consider the vectors
$$
\tilde\xi_i=d(Q)^{-1}T^{\xi_iv^*}\in\Mor_{\indcat{\QmodQ}}(Q,X\otimes Q).
$$
We claim that they are almost unit and almost invariant for $C^*(Q\mhyph\mathrm{mod}\mhyph Q)$.
By definition, as $i$ grows the vector $\tilde\xi_i$ becomes close (in $\Mor_\indcat{\CC}(Q,X\otimes Q)$) to
$$
d(Q)^{-1}(\iota_X\otimes v^*w^*\otimes\iota_Q)(\xi_iv^*\otimes\iota_Q\otimes\iota_Q\otimes\iota_Q)w^{(3)}
=d(Q)^{-1}(\xi_i\otimes\iota_Q)w^*w=\xi_i\otimes\iota_Q,
$$
hence $\|\tilde\xi_i\|\to1$.
Now, take a $Q$-bimodule $Y$. We want to check that $\tilde c_Y(\iota_Y\otimes_Q\tilde\xi_i)$ becomes close $\tilde\xi_i\otimes_Q\iota_Y$ for large $i$. Recall that as a model of $Y\otimes_Q(X\otimes Q)$ we take $Y\otimes X$ with the structure morphism $P_{Y,X\otimes Q}=(m^r_Y\otimes\iota_X)(\iota_Y\otimes c_Q^*)$, as a model of $(X\otimes Q)\otimes_Q Y$ we take $X\otimes Y$ with $P_{X\otimes Q,Y}=\iota_X\otimes m^l_Y$, and then $\tilde c_Y=c_Y$. Since $\iota_Y\otimes v$ is a right inverse of $P_{Y,Q}=m^r_Y$ and $v\otimes\iota_Y$ is a right inverse of $P_{Q,Y}=m^l_Y$, it follows that
\begin{align}
\tilde c_Y(\iota_Y\otimes_Q\tilde\xi_i)&=c_Y(m^r_Y\otimes\iota_X)(\iota_Y\otimes c_Q^*)(\iota_Y\otimes\tilde\xi_i)(\iota_Y\otimes v)=(\iota_X\otimes m^r_Y)(c_Y\otimes\iota_Q)(\iota_Y\otimes\tilde\xi_i)(\iota_Y\otimes v),\label{eq:aa-comparison1}\\
\tilde\xi_i\otimes_Q\iota_Y&=(\iota_X\otimes m^l_Y)(\tilde\xi_i\otimes\iota_Y)(v\otimes\iota_Y).\label{eq:aa-comparison2}
\end{align}
Since, as we have already shown, $\tilde\xi_i\approx\xi_i\otimes\iota_Q$, and $c_Y(\iota_Y\otimes\xi_i)\approx\xi_i\otimes\iota_Y$, both \eqref{eq:aa-comparison1} and \eqref{eq:aa-comparison2} become close to $\xi_i\otimes\iota_Y$ for large $i$. Thus, rescaling $\tilde{\xi}_i$ to a unit vector we obtain almost invariant unit vectors for $C^*(Q\mhyph\mathrm{mod}\mhyph Q)$.
Finally, the opposite implication also follows from the above, since an inverse of $F$ is defined in the same way as $F$ using the dual $Q$-system $\hat{Q}$.
\end{proof}
\begin{corollary}
Property (T) is invariant under weak monoidal Morita equivalence of C$^*$-tensor categories.
\end{corollary}
\begin{remark}
If $Q$ generates $\mathcal{C}$, so that $\mathcal{C}$, $Q\mhyph\mathrm{mod}$, $\mathrm{mod}\mhyph Q$ and $Q\mhyph\mathrm{mod}\mhyph Q$ can be regarded as bimodule categories defined by an extremal finite index subfactor, the above corollary has been already obtained by Popa~\cite{MR1729488}*{Proposition~9.8} using the universality of the symmetric enveloping algebra and permanence of relative property (T) for finite index intermediate algebras.
\end{remark}
\subsection{Comparison of regular half-braidings}
\label{sec:compar-reg-hbr}
In this last section we assume that $Q$ is a standard \emph{irreducible} $Q$-system.
Let $(M_k)_k$ be representatives of the isomorphism classes of simple objects in $\mathrm{mod}\mhyph Q$. Since $Q\in\mathrm{mod}\mhyph Q$ is simple by assumption, we may assume that $M_e = Q$ for some index $e$. We fix standard solutions $(R_k, \bar{R}_k)$ for $M_k$ once and for all, and denote by $(S_k, \bar{S}_k)$ the corresponding morphisms defined as in the Section~\ref{sec:mod-cat-Frob-recip}. Our goal is to show that the ind-object $Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q) = \oplus_k M_k \otimes_Q \bar{M}_k$ admits a half-braiding which corresponds to $Z_{\mathrm{reg}}(Q\mhyph\mathrm{mod}\mhyph Q)$ under Schauenburg's induction.
First, let us construct a half-braiding on $Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)$. This goes completely analogously to the construction of $Z_{\mathrm{reg}}(\mathcal{C})$. We fix $X \in \mathcal{C}$ and a standard solution $(R_X, \bar{R}_X)$. For each $k$ and $l$, choose an orthonormal basis $(u^\alpha_{k l})_\alpha$ in $\Mor_{\mathrm{mod}\mhyph Q}(M_l,X \otimes M_k)$, so that $\sum_\alpha u_{k l}^\alpha u_{k l}^{\alpha *}$ is the projection of $X \otimes M_k$ onto the isotypic component corresponding to $M_l$. We then define $c_{X, l k}\colon X \otimes M_k \otimes_Q \bar{M}_k \to M_l \otimes_Q \bar{M}_l \otimes X$ by
$$
c_{X, l k} = \left(\frac{d_k}{d_l}\right)^{{1/2}} \sum_\alpha (u^{\alpha *}_{k l} \otimes_Q u^{\alpha \vee}_{k l} \otimes \iota_X) (\iota_{ X \otimes M_k \otimes_Q \bar{M}_k } \otimes R_X),
$$
where $d_k=d(M_k)$ is the dimension of $M_k$ in $\mathcal{C}$ and $u^{\alpha \vee}_{k l}$ is defined using $\bar{R}_l$ and $\bar{R}_{X \otimes M_k} = (\iota \otimes \bar{R}_k \otimes \iota) \bar{R}_X$.
\begin{lemma}
The morphism $c_X \colon X \otimes Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q) \to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q) \otimes X$ defined by the morphisms $(c_{X, l k})_{l, k}$ is unitary.
\end{lemma}
\begin{proof}
The proof follows that of Lemma~\ref{lem:reg-half-br-candid-unitary}. First, in order to see that it is isometry, we need to show
$$
\frac{d_k}{d_l} u^{\beta}_{k l} u^{\alpha *}_{k l} \otimes_Q (\iota \otimes \Tr_{\bar X}) (u^{\beta * \vee}_{k l} u^{\alpha \vee}_{k l} ) = \delta_{\alpha, \beta} \iota.
$$
Since $\bar M_k$ is a simple left $Q$-module, the endomorphism $(\iota \otimes \Tr_{\bar X}) (u^{\beta * \vee}_{k l} u^{\alpha \vee}_{k l} ) $ of $\bar M_k$ is scalar, and we find this scalar as in the proof of Lemma~\ref{lem:reg-half-br-candid-unitary} by computing its trace.
Next, to verify that $c_X$ is unitary, we need to show that $\sum_k c_{X, l k} c_{X, l k}^* = \iota_{M_l \otimes_Q \bar{M}_l \otimes X}$, or
$$
\sum_{k, \alpha} \frac{d_k}{d_l} (\iota_l \otimes_Q u_{k l}^{\alpha \vee} \otimes \iota_X) (\iota_{X\otimes M_k\otimes_Q\bar M_k} \otimes R_X R_X^*) (\iota_l \otimes_Q u_{k l}^{\alpha \vee*} \otimes \iota_X) = \iota_{M_l \otimes_Q \bar{M}_l \otimes X}.
$$
This would follow if for each $k$ the morphism
$$
\sum_{\alpha} \frac{d_k}{d_l} (u_{k l}^{\alpha \vee} \otimes \iota_X) (\iota_{\bar k} \otimes R_X R_X^*) (u_{k l}^{\alpha \vee*} \otimes \iota_X)
$$
was the projection of $\bar{M}_l \otimes X$ onto the isotypic component corresponding to $\bar{M}_k$. This is again proved as in Lemma~\ref{lem:reg-half-br-candid-unitary} by observing that the Frobenius reciprocity isomorphism
$\mathcal{C}(\bar{M}_l, \bar{M}_k \otimes \bar{X})\cong\mathcal{C}(\bar{M}_l \otimes X, \bar{M}_k)$ defines an isomorphism
$\Mor_{Q\mhyph\mathrm{mod}}(\bar{M}_l, \bar{M}_k \otimes \bar{X}) \to \Mor_{Q\mhyph\mathrm{mod}}(\bar{M}_l \otimes X, \bar{M}_k)$.
\end{proof}
\begin{proposition}
The unitaries $c_X \colon X \otimes Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q) \to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q) \otimes X$ form a half-braiding on $\mathcal{C}$.
\end{proposition}
\begin{proof}
This is proved in the same way as Theorem~\ref{them:reg-half-braiding}, by considering the decomposition of $X\otimes Y\otimes M_k$ defined by decompositions of $Y\otimes M_k$ and $X\otimes M_l$.
\end{proof}
When we need to be specific, we will write $c^{\mathrm{mod}\mhyph Q}$ for the above half-braiding $(c_X)_X$. But if there is no danger of confusion, we will omit the half-braiding altogether and simply write $Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)$ for the object $(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q),c^{\mathrm{mod}\mhyph Q})\in{\Dcen(\indC)}$.
\begin{theorem}
\label{thm:compare-Zreg-Qmod-and-Zreg-QmodQ}
The image of $Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)$ under Schauenburg's induction ${\Dcen(\indC)}\to\mathcal{Z}(\indcat{\QmodQ})$ is isomorphic to
$Z_{\mathrm{reg}}(Q\mhyph\mathrm{mod}\mhyph Q)$.
\end{theorem}
We need some preparation to prove this theorem. Consider the canonical morphism $\zeta_e\colon Q=M_e\otimes_Q\bar M_e\toZ_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)$ in $\indcat{\CC}$.
\begin{lemma}
\label{lem:characterization-mor-from-ZregmodQ}
The morphism $\zeta_e$ satisfies identity~\eqref{eq:bimod-hom-char-in-orig}.
\end{lemma}
\begin{proof}
Consider the right hand side of~\eqref{eq:bimod-hom-char-in-orig} for $T=\zeta_e$, that is, the expression
$$
(\iota \otimes w^*) (c_Q^{\mathrm{mod}\mhyph Q} \otimes \iota)(\iota\otimes \zeta_e\otimes\iota) (\iota\otimes w).
$$
Composing it on the left with the projection onto $M_l\otimes_Q\bar M_l\otimes Q$ we get
$$
\left(\frac{d_e}{d_l}\right)^{{1/2}}\sum_\alpha(\iota_{M_l\otimes_Q\bar M_l}\otimes w^*)
(u_l^{\alpha*} \otimes_Q u_l^{\alpha\vee}\otimes\iota_Q\otimes\iota_Q)(\iota_Q\otimes\iota_Q\otimes wv\otimes\iota_Q)(\iota_Q\otimes w),
$$
where $(u^\alpha_l)_\alpha$ is an orthonormal basis in $\Mor_{\mathrm{mod}\mhyph Q}(M_l,Q \otimes Q)$ and we identify $(Q\otimes Q)\otimes_Q(Q\otimes Q)$ with~$Q^{\otimes 3}$. Using that $(\iota\otimes w^*)(wv\otimes\iota)=w$, we see that the above expression equals
$$
\left(\frac{d_e}{d_l}\right)^{{1/2}}\sum_\alpha(\iota_{M_l\otimes_Q\bar M_l}\otimes \iota_Q)
(u_l^{\alpha*} \otimes_Q u_l^{\alpha\vee}\otimes\iota_Q)(\iota_Q\otimes w^{(2)}).
$$
If $l\ne e$, then this expression is zero, since $u_l^{\alpha\vee}w\colon Q\to \bar M_l$ is a morphism of left $Q$-modules. For $l=e$ we may assume that $u_e^{\alpha_0} = d(Q)^{-{1/2}} w$ for some $\alpha_0$. Then the above formula picks up only the term corresponding to $\alpha_0$, and using that $u_e^{\alpha_0} w=d(Q)^{-1/2}w^*w=d(Q)^{1/2}\iota$ we see that it gives
$$
(w^*\otimes_Q\iota_Q\otimes\iota_Q)(\iota_Q\otimes w)=ww^*\colon Q\otimes Q\to Q\otimes Q=M_e\otimes_Q\bar M_e\otimes Q.
$$
Thus the right hand side of~\eqref{eq:bimod-hom-char-in-orig} equals $(\zeta_e\otimes\iota)ww^*$, so $\zeta_e$ satisfies~\eqref{eq:bimod-hom-char-in-orig}.
\end{proof}
\begin{lemma}
\label{lem:restr-from-ZregmodQ-inj}
For any $(X,c) \in {\Dcen(\indC)}$, the map $\Mor_{\Dcen(\indC)}(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q), X) \to \Mor_\indcat{\CC}(Q, X)$, $T\mapsto T\zeta_e$, is injective.
\end{lemma}
\begin{proof} Fix an index $k$. We claim that the canonical morphism $M_k\otimes_Q \bar M_k\to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)$ is given by
$$
\zeta_k=d_e^{-3/2}d_k^{1/2}(\iota \otimes \bar{R}_k^*) (c^{\mathrm{mod}\mhyph Q}_{k} \otimes \iota_{\bar k}) (\iota_{k} \otimes \zeta_e \otimes \iota_{\bar k}) (m_{k}^{r *} \otimes_Q m_{\bar k}^{l *}),
$$
where we identify $(M_k\otimes Q)\otimes_Q(Q\otimes\bar M_k)$ with $M_k\otimes Q\otimes\bar M_k$. Indeed, as in the previous lemma, composing this with the projection onto $M_l\otimes_Q\bar M_l$ we see that the above expression gives
$$
d_e^{-1}d_k^{1/2}d_l^{-1/2}\sum_\alpha
(u_l^{\alpha*} \otimes_Q u_l^{\alpha\vee})(m_{k}^{r *} \otimes_Q m_{\bar k}^{l *}),
$$
where $(u^\alpha_l)_\alpha$ is an orthonormal basis in $\Mor_{\mathrm{mod}\mhyph Q}(M_l,M_k \otimes Q)$. If $l\ne k$, this expression is zero. If $l=k$, then it picks up the term corresponding to the isometry $u_k^{\alpha_0}=d(Q)^{-1/2}m_{k}^{r*}$. Since $m_{\bar k}^{l *}=m^{r\vee}_{k}$ by definition, we then see that the above expression is the identity morphism. Thus our claim is proved.
Now, if $T\in\Mor_{\Dcen(\indC)}(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q), X)$, we get
$$
T\zeta_k=d_e^{-3/2}d_k^{1/2}(\iota_X \otimes \bar{R}_k^*) (c_{k} \otimes \iota_{\bar k}) (\iota_{k} \otimes T \zeta_e \otimes \iota_{\bar k}) (m_{k}^{r *} \otimes_Q m_{\bar k}^{l *}).
$$
Since $T$ is determined by the morphisms $T\zeta_k$, we conclude that it is determined by $T\zeta_e$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:compare-Zreg-Qmod-and-Zreg-QmodQ}]
Consider the image $(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q),\tilde c)$ of $(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q),c^{\mathrm{mod}\mhyph Q})$ under Schauenburg's induction.
Since by Lemma~\ref{lem:characterization-mor-from-ZregmodQ} the morphism $\zeta_e$ satisfies identity~\eqref{eq:bimod-hom-char-in-orig}, we have an isometric morphism $\xi=d(Q)^{-1/2}(\zeta_e\otimes\iota)w\colon Q\to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q$ in $\indcat{\QmodQ}$. Let $\phi$ be the positive definite function on $\Irr(Q\mhyph\mathrm{mod}\mhyph Q)$ defined by $\xi$. Then by Proposition~\ref{prop:Zphi-universal} there exists a unique isometric morphism $T\colon Z_\phi\to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q$ in $\mathcal{Z}(\indcat{\QmodQ})$ such that $T\xi_\phi=\xi$.
The morphism $T$ is unitary. Indeed, since Schauenburg's induction is an equivalence of categories, the endomorphism $p=\iota-TT^*$ of $Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q$ must be of the form $S\otimes\iota$ for some $S\in\End_{\Dcen(\indC)}(Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q))$. Since $pT=0$, we have $p\xi=0$, hence $S\zeta_e=0$. By Lemma~\ref{lem:restr-from-ZregmodQ-inj} it follows that $S=0$, so $TT^*=\iota$.
To finish the proof it remains to show that $\phi=\delta_e$, as then $Z_\phi=Z_{\mathrm{reg}}(Q\mhyph\mathrm{mod}\mhyph Q)$. Let us fix representatives $(X_a)_a$ of the isomorphism classes of simple objects in $Q\mhyph\mathrm{mod}\mhyph Q$, so that $Z_\phi=A^\phi\text{-}\oplus_a X_a\otimes_Q\bar X_a$. We continue to denote by $e$ the index corresponding to the unit object and assume that $X_e=Q$. Then by \eqref{eq:intertbr} the composition $T_a$ of $T$ with the canonical morphism $X_a\otimes_Q\bar X_a\to Z_\phi$ is given by
$$
T_a = (d^Q_a)^{{1/2}}(\iota \otimes_Q \bar R^{Q*}_a) (\tilde{c}_a \otimes_Q \iota_{\bar{a}}) (\iota_a \otimes_Q \xi \otimes_Q \iota_{\bar{a}})\colon X_a\otimes_Q \bar X_a\to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q.
$$
(Recall that the upper index $Q$ indicates that we are dealing with the category $Q\mhyph\mathrm{mod}\mhyph Q$.)
Since $\pi_\phi([X_a])\xi=(d^Q_a)^{-1/2}T_a\bar R^Q_a$, in order to check that $\phi=\delta_e$ it suffices to show that $\xi^*T_a=T_e^*T_a=0$ for $a\ne e$.
Let us compute the morphisms $T_a$ more explicitly. As we already used in the proof of Lemma~\ref{lem:schaun-braiding}, the morphism $\tilde c_a$ is implemented by the morphism $X_a\otimes Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q\to Z_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q\otimes X_a$ given by $$(\iota\otimes v\otimes\iota_a)(\iota\otimes m^r_a)(c^{\mathrm{mod}\mhyph Q}_a\otimes\iota_Q).$$
Since we also have $\bar R^{Q*}_a P_{\bar a,a}=\bar S^*_a$ by \eqref{eq:RQstar}, and $P_{Q,Q}=w^*$, we conclude that $T_a$ is implemented by the morphism $X_a\otimes Q\otimes \bar X_a\toZ_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q$ given by
$$
(d^Q_a)^{{1/2}}(\iota\otimes w^*)(\iota\otimes\iota_Q\otimes \bar S^*_a)(\iota\otimes v\otimes\iota_a\otimes\iota_{\bar a})(\iota\otimes m^r_a\otimes\iota_{\bar a})(c^{\mathrm{mod}\mhyph Q}_a\otimes\iota_Q\otimes\iota_{\bar a}) (\iota_a \otimes \xi \otimes \iota_{\bar{a}}).
$$
Since $w^*(v\otimes\iota)=\iota$ and $\bar S^*_a(m^r_a\otimes\iota)=\bar S^*_a(\iota\otimes m^l_{\bar a})$ by Lemma~\ref{lem:Sdescent}, the above expression equals
$$
(d^Q_a)^{{1/2}}(\iota\otimes \bar S^*_a)(\iota\otimes \iota_a\otimes m^l_{\bar a})(c^{\mathrm{mod}\mhyph Q}_a\otimes\iota_Q\otimes\iota_{\bar a}) (\iota_a \otimes \xi \otimes \iota_{\bar{a}}).
$$
Using that $\xi=d_e^{-1/2}(\zeta_e\otimes\iota)w$ and $(\iota\otimes m^l_{\bar a})(w\otimes\iota)=m^{l*}_{\bar a}m^l_{\bar a}$ by Lemma~\ref{lem:comput-m-m-star}, and then that $P_{Q,X_{\bar a}}=m^l_{\bar a}$, we finally get that the morphism $T_a$ is implemented by the morphism
$$
\tilde T_a= (d^Q_a)^{{1/2}}d_e^{-1/2}(\iota\otimes \bar S^*_a)(c^{\mathrm{mod}\mhyph Q}_a\otimes\iota_{\bar a}) (\iota_a \otimes \zeta_e\otimes \iota_{\bar{a}})(\iota_a\otimes m^{l*}_{\bar a})\colon X_a\otimes X_{\bar a}\toZ_{\mathrm{reg}}(\mathrm{mod}\mhyph Q)\otimes Q.
$$
Consider the component $\tilde T_{ka}$ of this morphism mapping $X_a\otimes X_{\bar a}$ into $M_k\otimes_Q\bar M_k\otimes Q$. By definition,
$$
\tilde T_{ka}= (d^Q_a)^{1/2}d_k^{-1/2}\sum_\alpha(\iota\otimes \bar S^*_a)
(u^{\alpha *}_{ak} \otimes_Q u^{\alpha \vee}_{ak} \otimes \iota_a\otimes\iota_{\bar a}) (\iota_a\otimes\iota_Q\otimes R_a\otimes\iota_{\bar a})(\iota_a\otimes m^{l*}_{\bar a}),
$$
where $(u^\alpha_{ak})_\alpha$ is an orthonormal basis in $\Mor_{\mathrm{mod}\mhyph Q}(M_k,X_a\otimes Q)$. Since by the definition of the maps $\bar S$ we have $(\iota \otimes \bar{S}_a^*) (R_a \otimes \iota) = m_{\bar{a}}^{r *}$, we get
$$
\tilde T_{ka}= (d^Q_a)^{1/2}d_k^{-1/2}\sum_\alpha
(u^{\alpha *}_{ak} \otimes_Q u^{\alpha \vee}_{ak} \otimes \iota_Q) (\iota_a\otimes\iota_Q\otimes m_{\bar{a}}^{r *})(\iota_a\otimes m^{l*}_{\bar a}).
$$
From this we see that $\tilde T^*_{kb}\tilde T_{ka}=0$ for $b\ne a$, since
$$
m^{l}_{\bar b}(\iota_Q\otimes m_{\bar{b}}^{r})(u^{\beta \vee*}_{bk} \otimes \iota_Q)(u^{\alpha \vee}_{ak} \otimes \iota_Q)(\iota_Q\otimes m_{\bar{a}}^{r *})m^{l*}_{\bar a}
$$
is a morphism $X_{\bar a}\to X_{\bar b}$ of $Q$-bimodules. Hence $T^*_bT_a=0$ for $b\ne a$, and the theorem is proved.
\end{proof}
\bigskip
\raggedright
\begin{bibdiv}
\begin{biblist}
\bib{MR914742}{article}{
author={Anantharaman-Delaroche, Claire},
title={On {C}onnes' property {$T$} for von {N}eumann algebras},
date={1987},
ISSN={0025-5513},
journal={Math. Japon.},
volume={32},
number={3},
pages={337\ndash 355},
review={\MR{914742 (89h:46086)}},
}
\bib{arXiv:1410.6238}{misc}{
author={Arano, Yuki},
title={Unitary spherical representations of {D}rinfeld doubles},
how={preprint},
date={2014},
eprint={\href{http://arxiv.org/abs/1410.6238}{{\tt arXiv:1410.6238
[math.QA]}}},
note={to appear in J. Reine Angew. Math.},
doi={10.1515/crelle-2015-0079},
}
\bib{MR3308880}{book}{
author={Bischoff, Marcel},
author={Kawahigashi, Yasuyuki},
author={Longo, Roberto},
author={Rehren, Karl-Henning},
title={Tensor categories and endomorphisms of von {N}eumann
algebras---with applications to quantum field theory},
series={Springer Briefs in Mathematical Physics},
publisher={Springer, Cham},
date={2015},
volume={3},
ISBN={978-3-319-14300-2; 978-3-319-14301-9},
url={http://dx.doi.org/10.1007/978-3-319-14301-9},
doi={10.1007/978-3-319-14301-9},
review={\MR{3308880}},
}
\bib{MR2391387}{book}{
author={Brown, Nathanial~P.},
author={Ozawa, Narutaka},
title={{$C^*$}-algebras and finite-dimensional approximations},
series={Graduate Studies in Mathematics},
publisher={American Mathematical Society},
address={Providence, RI},
date={2008},
volume={88},
ISBN={978-0-8218-4381-9; 0-8218-4381-8},
review={\MR{MR2391387 (2009h:46101)}},
}
\bib{MR2863377}{article}{
author={Brugui{{\`e}}res, Alain},
author={Natale, Sonia},
title={Exact sequences of tensor categories},
date={2011},
ISSN={1073-7928},
journal={Int. Math. Res. Not. IMRN},
number={24},
pages={5644\ndash 5705},
eprint={\href{http://arxiv.org/abs/1006.0569}{{\tt arXiv:1006.0569
[math.QA]}}},
url={http://dx.doi.org/10.1093/imrn/rnq294},
doi={10.1093/imrn/rnq294},
review={\MR{2863377}},
}
\bib{MR2355605}{article}{
author={Brugui{{\`e}}res, Alain},
author={Virelizier, Alexis},
title={Hopf monads},
date={2007},
ISSN={0001-8708},
journal={Adv. Math.},
volume={215},
number={2},
pages={679\ndash 733},
eprint={\href{http://arxiv.org/abs/math/0604180}{{\tt arXiv:math/0604180
[math.QA]}}},
url={http://dx.doi.org/10.1016/j.aim.2007.04.011},
doi={10.1016/j.aim.2007.04.011},
review={\MR{2355605 (2009b:18006)}},
}
\bib{MR3079759}{article}{
author={Brugui{{\`e}}res, Alain},
author={Virelizier, Alexis},
title={On the center of fusion categories},
date={2013},
ISSN={0030-8730},
journal={Pacific J. Math.},
volume={264},
number={1},
pages={1\ndash 30},
eprint={\href{http://arxiv.org/abs/1203.4180}{{\tt arXiv:1203.4180
[math.QA]}}},
url={http://dx.doi.org/10.2140/pjm.2013.264.1},
doi={10.2140/pjm.2013.264.1},
review={\MR{3079759}},
}
\bib{MR3238527}{article}{
author={De~Commer, Kenny},
author={Freslon, Amaury},
author={Yamashita, Makoto},
title={C{CAP} for {U}niversal {D}iscrete {Q}uantum {G}roups},
date={2014},
ISSN={0010-3616},
journal={Comm. Math. Phys.},
volume={331},
number={2},
pages={677\ndash 701},
eprint={\href{http://arxiv.org/abs/1306.6064}{{\tt arXiv:1306.6064
[math.OA]}}},
url={http://dx.doi.org/10.1007/s00220-014-2052-7},
doi={10.1007/s00220-014-2052-7},
review={\MR{3238527}},
}
\bib{MR3121622}{article}{
author={De~Commer, Kenny},
author={Yamashita, Makoto},
title={Tannaka-{K}re\u\i n duality for compact quantum homogeneous
spaces. {I}. {G}eneral theory},
date={2013},
ISSN={1201-561X},
journal={Theory Appl. Categ.},
volume={28},
pages={No. 31, 1099\ndash 1138},
eprint={\href{http://arxiv.org/abs/1211.6552}{{\tt arXiv:1211.6552
[math.OA]}}},
review={\MR{3121622}},
}
\bib{MR654325}{article}{
author={Deligne, Pierre},
author={Milne, James S.},
title={Tannakian categories},
conference={
title={Hodge cycles, motives, and Shimura varieties},},
book={
series={Lecture Notes in Mathematics},
volume={900},
publisher={Springer-Verlag, Berlin-New York},},
date={1982},
pages={101\ndash 228},
isbn={3-540-11174-3},
url={http://www.jmilne.org/math/xnotes/tc.html},
review={\MR{654325 (84m:14046)}},
}
\bib{MR1316301}{article}{
author={Evans, David~E.},
author={Kawahigashi, Yasuyuki},
title={On {O}cneanu's theory of asymptotic inclusions for subfactors,
topological quantum field theories and quantum doubles},
date={1995},
ISSN={0129-167X},
journal={Internat. J. Math.},
volume={6},
number={2},
pages={205\ndash 228},
url={http://dx.doi.org/10.1142/S0129167X95000468},
doi={10.1142/S0129167X95000468},
review={\MR{1316301 (96d:46080)}},
}
\bib{MR1642584}{book}{
author={Evans, David~E.},
author={Kawahigashi, Yasuyuki},
title={Quantum symmetries on operator algebras},
series={Oxford Mathematical Monographs},
publisher={The Clarendon Press Oxford University Press},
address={New York},
date={1998},
ISBN={0-19-851175-2},
note={Oxford Science Publications},
review={\MR{1642584 (99m:46148)}},
}
\bib{MR3447719}{article}{
author={Ghosh, Shamindra~Kumar},
author={Jones, Corey},
title={Annular representation theory for rigid {$C^*$}-tensor
categories},
date={2016},
ISSN={0022-1236},
journal={J. Funct. Anal.},
volume={270},
number={4},
pages={1537\ndash 1584},
eprint={\href{http://arxiv.org/abs/1502.06543}{{\tt arXiv:1502.06543
[math.OA]}}},
url={http://dx.doi.org/10.1016/j.jfa.2015.08.017},
doi={10.1016/j.jfa.2015.08.017},
review={\MR{3447719}},
}
\bib{MR1749868}{article}{
author={Hayashi, Tomohiro},
author={Yamagami, Shigeru},
title={Amenable tensor categories and their realizations as {AFD}
bimodules},
date={2000},
ISSN={0022-1236},
journal={J. Funct. Anal.},
volume={172},
number={1},
pages={19\ndash 75},
url={http://dx.doi.org/10.1006/jfan.1999.3521},
doi={10.1006/jfan.1999.3521},
review={\MR{1749868 (2001d:46092)}},
}
\bib{MR1644299}{article}{
author={Hiai, Fumio},
author={Izumi, Masaki},
title={Amenability and strong amenability for fusion algebras with
applications to subfactor theory},
journal={Internat. J. Math.},
volume={9},
date={1998},
number={6},
pages={669--722},
issn={0129-167X},
review={\MR{1644299 (99h:46116)}},
doi={10.1142/S0129167X98000300},
}
\bib{MR1782145}{article}{
author={Izumi, Masaki},
title={The structure of sectors associated with {L}ongo-{R}ehren
inclusions. {I}. {G}eneral theory},
date={2000},
ISSN={0010-3616},
journal={Comm. Math. Phys.},
volume={213},
number={1},
pages={127\ndash 179},
url={http://dx.doi.org/10.1007/s002200000234},
doi={10.1007/s002200000234},
review={\MR{1782145 (2002k:46160)}},
}
\bib{MR1321145}{book}{
author={Kassel, Christian},
title={Quantum groups},
series={Graduate Texts in Mathematics},
publisher={Springer-Verlag},
address={New York},
date={1995},
volume={155},
ISBN={0-387-94370-6},
review={\MR{1321145 (96e:17041)}},
}
\bib{MR1257245}{article}{
author={Longo, Roberto},
title={A duality for {H}opf algebras and for subfactors. {I}},
date={1994},
ISSN={0010-3616},
journal={Comm. Math. Phys.},
volume={159},
number={1},
pages={133\ndash 150},
url={http://projecteuclid.org/getRecord?id=euclid.cmp/1104254494},
review={\MR{1257245 (95h:46097)}},
}
\bib{MR1332979}{article}{
author={Longo, R.},
author={Rehren, K.-H.},
title={Nets of subfactors},
date={1995},
ISSN={0129-055X},
journal={Rev. Math. Phys.},
volume={7},
number={4},
pages={567\ndash 597},
eprint={\href{http://arxiv.org/abs/hep-th/9411077}{{\tt
arXiv:hep-th/9411077 [hep-th]}}},
url={http://dx.doi.org/10.1142/S0129055X95000232},
doi={10.1142/S0129055X95000232},
note={Workshop on Algebraic Quantum Field Theory and Jones Theory
(Berlin, 1994)},
review={\MR{1332979 (96g:81151)}},
}
\bib{MR1444286}{article}{
author={Longo, R.},
author={Roberts, J.~E.},
title={A theory of dimension},
date={1997},
ISSN={0920-3036},
journal={$K$-Theory},
volume={11},
number={2},
pages={103\ndash 159},
eprint={\href{http://arxiv.org/abs/funct-an/9604008}{{\tt
arXiv:funct-an/9604008 [math.FA]}}},
url={http://dx.doi.org/10.1023/A:1007714415067},
doi={10.1023/A:1007714415067},
review={\MR{1444286 (98i:46065)}},
}
\bib{MR1712872}{book}{
author={Mac~Lane, Saunders},
title={Categories for the working mathematician},
edition={Second},
series={Graduate Texts in Mathematics},
publisher={Springer-Verlag, New York},
date={1998},
volume={5},
ISBN={0-387-98403-8},
review={\MR{1712872 (2001j:18001)}},
}
\bib{MR1742858}{article}{
author={Masuda, Toshihiko},
title={Generalization of {L}ongo-{R}ehren construction to subfactors of
infinite depth and amenability of fusion algebras},
date={2000},
ISSN={0022-1236},
journal={J. Funct. Anal.},
volume={171},
number={1},
pages={53\ndash 77},
url={http://dx.doi.org/10.1006/jfan.1999.3523},
doi={10.1006/jfan.1999.3523},
review={\MR{1742858 (2001f:46093)}},
}
\bib{MR1966524}{article}{
author={M{\"u}ger, Michael},
title={From subfactors to categories and topology. {I}. {F}robenius
algebras in and {M}orita equivalence of tensor categories},
date={2003},
ISSN={0022-4049},
journal={J. Pure Appl. Algebra},
volume={180},
number={1-2},
pages={81\ndash 157},
eprint={\href{http://arxiv.org/abs/math/0111204}{{\tt
arXiv:math/0111204}}},
url={http://dx.doi.org/10.1016/S0022-4049(02)00247-5},
doi={10.1016/S0022-4049(02)00247-5},
review={\MR{1966524 (2004f:18013)}},
}
\bib{MR1966525}{article}{
author={M{\"u}ger, Michael},
title={From subfactors to categories and topology. {II}. {T}he quantum
double of tensor categories and subfactors},
date={2003},
ISSN={0022-4049},
journal={J. Pure Appl. Algebra},
volume={180},
number={1-2},
pages={159\ndash 219},
eprint={\href{http://arxiv.org/abs/math/0111205}{{\tt
arXiv:math/0111205}}},
url={http://dx.doi.org/10.1016/S0022-4049(02)00248-7},
doi={10.1016/S0022-4049(02)00248-7},
review={\MR{1966525 (2004f:18014)}},
}
\bib{MR3204665}{book}{
author={Neshveyev, Sergey},
author={Tuset, Lars},
title={Compact quantum groups and their representation categories},
series={Cours Sp{\'e}cialis{\'e}s [Specialized Courses]},
publisher={Soci{\'e}t{\'e} Math{\'e}matique de France, Paris},
date={2013},
volume={20},
ISBN={978-2-85629-777-3},
review={\MR{3204665}},
}
\bib{arXiv:1405.6572}{misc}{
author={Neshveyev, Sergey},
author={Yamashita, Makoto},
title={Poisson boundaries of monoidal categories},
how={preprint},
date={2014},
eprint={\href{http://arxiv.org/abs/1405.6572}{{\tt arXiv:1405.6572
[math.OA]}}},
}
\bib{MR996454}{incollection}{
author={Ocneanu, Adrian},
title={Quantized groups, string algebras and {G}alois theory for
algebras},
date={1988},
booktitle={Operator algebras and applications, {V}ol.\ 2},
series={London Math. Soc. Lecture Note Ser.},
volume={136},
publisher={Cambridge Univ. Press},
address={Cambridge},
pages={119\ndash 172},
review={\MR{MR996454 (91k:46068)}},
}
\bib{popa-corr-preprint}{misc}{
author={Popa, Sorin},
title={Correspondences},
how={preprint},
date={1986},
note={INCREST preprint},
url={http://www.math.ucla.edu/~popa/popa-correspondences.pdf}
}
\bib{MR1302385}{article}{
author={Popa, Sorin},
title={Symmetric enveloping algebras, amenability and {AFD} properties
for subfactors},
date={1994},
ISSN={1073-2780},
journal={Math. Res. Lett.},
volume={1},
number={4},
pages={409\ndash 425},
review={\MR{MR1302385 (95i:46095)}},
}
\bib{MR1334479}{article}{
author={Popa, Sorin},
title={An axiomatization of the lattice of higher relative commutants of
a subfactor},
date={1995},
ISSN={0020-9910},
journal={Invent. Math.},
volume={120},
number={3},
pages={427\ndash 445},
url={http://dx.doi.org/10.1007/BF01241137},
doi={10.1007/BF01241137},
review={\MR{1334479 (96g:46051)}},
}
\bib{MR1729488}{article}{
author={Popa, Sorin},
title={Some properties of the symmetric enveloping algebra of a
subfactor, with applications to amenability and property {T}},
date={1999},
ISSN={1431-0635},
journal={Doc. Math.},
volume={4},
pages={665\ndash 744 (electronic)},
review={\MR{MR1729488 (2001c:46116)}},
}
\bib{MR2215135}{article}{
author={Popa, Sorin},
title={On a class of type {${\rm II}_1$} factors with {B}etti numbers
invariants},
date={2006},
ISSN={0003-486X},
journal={Ann. of Math. (2)},
volume={163},
number={3},
pages={809\ndash 899},
eprint={\href{http://arxiv.org/abs/math/0209130}{{\tt
arXiv:math/0209130}}},
url={http://dx.doi.org/10.4007/annals.2006.163.809},
doi={10.4007/annals.2006.163.809},
review={\MR{2215135 (2006k:46097)}},
}
\bib{MR3406647}{article}{
author={Popa, Sorin},
author={Vaes, Stefaan},
title={Representation theory for subfactors, {$\lambda$}-lattices and
{$\rm C^*$}-tensor categories},
date={2015},
ISSN={0010-3616},
journal={Comm. Math. Phys.},
volume={340},
number={3},
pages={1239\ndash 1280},
eprint={\href{http://arxiv.org/abs/1412.2732}{{\tt arXiv:1412.2732
[math.OA]}}},
url={http://dx.doi.org/10.1007/s00220-015-2442-5},
doi={10.1007/s00220-015-2442-5},
review={\MR{3406647}},
}
\bib{MR1822847}{article}{
author={Schauenburg, Peter},
title={The monoidal center construction and bimodules},
date={2001},
ISSN={0022-4049},
journal={J. Pure Appl. Algebra},
volume={158},
number={2-3},
pages={325\ndash 346},
url={http://dx.doi.org/10.1016/S0022-4049(00)00040-2},
doi={10.1016/S0022-4049(00)00040-2},
review={\MR{1822847 (2002f:18013)}},
}
\bib{MR1245354}{article}{
author={Yamagami, Shigeru},
title={A note on {O}cneanu's approach to {J}ones' index theory},
date={1993},
ISSN={0129-167X},
journal={Internat. J. Math.},
volume={4},
number={5},
pages={859\ndash 871},
url={http://dx.doi.org/10.1142/S0129167X9300039X},
doi={10.1142/S0129167X9300039X},
review={\MR{1245354 (95f:46114)}},
}
\end{biblist}
\end{bibdiv}
\bigskip
\end{document}
|
\section{Introduction and motivations}
There are quantum mechanical problems in which the Hamiltonian depends explicitly on time,
for example, the interaction of the system with an external
time-dependent field where $H(x,t)\, =\, H_0(x)+H'(x,t)$. In such cases, the system does not
remain in any stationary state, and the
behavior of it is governed by the time-dependent Schr\"odinger equation
\begin{equation}
i\, \hbar\, \frac{\partial\, \Psi(X,t)}{\partial\, t}\, =\, H(X,t)\, \Psi(X,t). \nonumber
\end{equation}
The problem will now be an initial-value (Cauchy) problem of solving
this partial differential equation for a prescribed initial
condition at $t_0$. When problems of this short are discussed
formally, it is common to speak of the perturbation ($H'$) as
causing {\it transitions} between eigenstates of $H_0$. If this
statement is interpreted to mean that the state has changed from its
inital value to a final value, than it is incorrect. The effect of a
time-dependent perturbation is to produce a nonstationary state,
rather than to cause a jump from one stationary state to another
which were determined by considering boundary conditions solely.
Experiments open the possibility to investigate dynamical properties
of confined systems. In that field, of particular interest is the
response of the system to time-dependent variations of the confining
field. Then, even assuming complete isolation of the many-body
confined system from the environment, there is the question of how
the correlation properties change. Furthermore, one can consider an
atom in its ground state. If at times $t>t_0$ it becomes subject to
a time-varying potential energy caused by a charged heavy projectile
passing by, its electron cloud will be shaken up so that its energy
increases. The energy change, related to the stopping power, is a
measurable quantity. However, when we have shake-up processes, the
application of one-particle auxiliary pictures is not necessarily
useful {\it a priori}.
Motivated by such problems, in this work we consider a simplified
interacting model system introduced by Heisenberg
\cite{Heisenberg26,Moshinsky68,Ballentine98} in the early days of
quantum mechanics
\begin{equation}
\hat{H}_0(x_1,x_2)\, =\, -\, \frac{\hbar^2}{2m}\left(\frac{d^2}{dx_1^2}\, +
\frac{d^2}{dx_2^2}\right) +\frac{m}{2}\, \omega_0^2({x}_1^2+{x}_2^2)
-\frac{m}{2}\, \Lambda\, \omega_0^2({x}_1-{x}_2)^2,
\end{equation}
where $\Lambda\in{[0,0.5]}$ measures the strength of repulsive
interparticle interaction. We will use ideas and methods
\cite{Popov70,Kagan96,Campo13} for solving the quantum motion of a
particle in a harmonic oscillator with time-dependent frequency, by
adding a quadrupolar ($Q$) time-dependent perturbation, of
$\hat{H}'(t\rightarrow{-\infty})=\hat{H}'(t\rightarrow{\infty})=0$
character, to the above ground-state Hamiltonian
\begin{equation}
\hat{H}'(x_1,x_2,t)\, =\, \frac{m}{2}\, Q\, F(t)\, ({x}_1^2+{x}_2^2).
\end{equation}
In the field of heavy-particle interaction with atoms, such term could mimic the shake-up process
due to close encounters. There, as was stressed \cite{Fermi40} by Fermi,
quantum mechanics is needed since Bohr's classical treatment is
valid \cite{Bohr13,Dobson94} only for dipolar perturbation.
One might think that a two-particle model is a bit trivial to test
time-dependent many-body methods. But, this is not the case. Indeed,
it is the few-body correlated dynamics that serve \cite{Bauer13} as
benchmark for methods beyond, for instance, Time Dependent Density
Functional Theory, TD-DFT. Furthermore, to the theory of breathing
modes of many-body systems in harmonic confinement, it was utilized
\cite{Brabec13} to replace a given (say, with Coulombic
interparticle interaction) Hamiltonian with a quadratic Hamiltonian
for which an analytic solution exist. A mapping at the Hamiltonian
level could be a practical version of the isospectral deformation
discussed earlier \cite{Dreizler86} in the light of the,
unfortunately, formal character of pure existence theorems and
mapping lemmas behind TD-DFT.
We believe that experience about relevant details in interacting
systems is best governed by studying time-dependent quantities based
on exactly solvable instructive examples. In particular, the
simultaneous treatment of the notorious kinetic and interparticle
energy components could be more accurate than in TD-DFT where both
are approximated. Besides, our result allows a transparent
implementation of the mapping-formalism \cite{Ullrich12,Leeuwen15}
designed to construct independent-particle potentials (and to
discuss \cite{Schirmer07} an alarming paradox arising there) by
using exact probabilistic quantities of the correlated model as
inputs.
As a last motivation, we note that there have been efforts to
clarify the basic relations between interparticle correlation and
information-theoretic measures for entanglement. For instance, the
stationary ground-state of the correlated model applied in the
present work has already been investigated in this respect. A
surprising duality between R\'enyi's entropies characterizing
entangled systems, with attractive ($\Lambda<0$) and repulsive
interparticle interactions was found
\cite{Pipek09,Gross13,Glasser13} and explained \cite{Schilling14}
recently. The question of how such a remarkable duality will change
when the correlated system is perturbed by a time-dependent external
field is an exciting one of broad relevance. One of our goals here
is to provide an answer.
We stress that we restrict ourselves to the linear-response limit,
by taking our $QF(t)$ sufficiently small or sudden, since the main
goal is on the time-dependent spectral aspects of
correlation-dynamics and not on comparison with data. However, we
should note that similarly to the Born series of stationary
scattering theory, an order-by-order expansion could result in only
an asymptotic series. Indeed, higher-order response to external
field is an important topic in the energy change during shake-up
dynamics of a nucleus \cite{Robinson61} or an atom
\cite{Merzbacher74}. The idea that by renormalization of the kinetic
energy one could \cite{Prelovsek10} extend the validity of
linear-response formalism is also a challenging one. Such questions
need accurate solutions in linear and nonlinear responses. The
nonlinear version of the exact determination of energy changes in
our correlated model system will be published separately.
The rest of the paper is organized as follows. Section II contains
our theoretical results. Section III is devoted to a short summary
and few relevant comments. We will use natural, rather than atomic,
units in this work, except where the opposite is explicitly stated.
\section{Results and discussion}
By introducing the normal coordinates
$X_1\equiv{(x_1+x_2)/\sqrt{2}}$ and $X_2\equiv{(x_1-x_2)/\sqrt{2}}$,
one can easily rewrite \cite{Moshinsky68} the unperturbed
Hamiltonian in the form
\begin{equation}
\hat{H}_0(X_1,X_2)\, =\, -\, \frac{\hbar^2}{2m}\left(\frac{d^2}{dX_1^2}\, +
\frac{d^2}{dX_2^2}\right) +\frac{m}{2}\, \omega_1^2\, X_1^2 + \frac{m}{2}\, \omega_2^2\, X_2^2,
\end{equation}
where $\omega_1\equiv{\omega_0}$ and
$\omega_2\equiv{\omega_0\sqrt{1-2\Lambda}}$ denote the resulting
independent-mode frequencies. It is this separated form which shows
that a time-dependent perturbation of dipolar character,
$D(t)(x_1+x_2)=[\sqrt{2}D(t)]\, X_1$, will couple only to one normal
mode characterized by the unperturbed angular frequency
$\omega_1=\omega_0$. Therefore, such perturbation
\cite{Bohr13,Robinson61,Merzbacher74,Dobson94} produces an energy
change independent of the correlated aspect of our model.
In the case investigated here, we have $QF(t)\,
(x_1^2+x_2^2)=QF(t)(X_1^2+X_2^2)$, i.e., there is time-dependent
quadrupolar perturbation in both independent normal modes, which
will evolve in time independently. Thus, we proceed by one
oscillator [$\hat{h}(X,t)$] in a time-dependent harmonic
confinement, following the established theoretical path, where one
has to solve a time-dependent Schr\"odinger equation, of the form
given by Eq. (1), with
\begin{equation}
\hat{h}(X,t)\, =\, -\, \frac{\hbar^2}{2m}\, \frac{d^2}{dX^2}\, + \frac{m}{2}\, \Omega^2(t)\, X^2 .
\end{equation}
The solution rests on making proper changes of the time and distance
scales \cite{Popov70,Kagan96} in order to consider frequency
variations in $\Omega(t)$ as it changes from $\Omega_0$ during
time-evolution. The exact mode, a nonstationary evolving state
$\psi[X,\Omega_0,R(\Omega_0),t]$, contains these scales as
\begin{equation}
\psi(X,\Omega_0,R,t)=\left[\frac{m\, \Omega_0}{\hbar\, R^2(t)\, \pi}\right]^{1/4}\, \exp\left[-\frac{X^2}{2}\, \frac{m\, \Omega_0}{\hbar \, R^2(t)}
\left(1-i\, \frac{R(t)\dot{R}(t)}{\Omega_0}\right)\right]\, e^{-i\, \gamma(t)}.
\end{equation}
This exact quantum mechanical, time-dependent solution is obtained \cite{Popov70,Kagan96} by
considering a classical equation of motion
\begin{equation}
\ddot{\xi}(t) +\, \Omega(t)\, \xi(t)\, =\, 0,
\end{equation}
with complex $\xi(t)=R(t)\exp[i \gamma(t)]$ substitution, where the real $R(t)$ gives the length scale at the instant $t$.
The nonlinear differential equation, determining this scale becomes
\begin{equation}
\ddot{R} + R(t)\, \Omega^2(t)\, =\, \frac{\Omega_0^2}{[R(t)]^3},
\end{equation}
after taking $\dot{\gamma}(t)\equiv{\Omega_0/R^2(t)}$. The initial
conditions are $R(t_0=0)=1$ and $\dot{R}(t=0)=0$. Since we are
dealing with a weak external perturbation in this work on spectral
dynamics, we solve Eq. (7) via the substitutions $R(t)=1+r(t)$ and
$\dot{R}(t)=\dot{r}(t)$.
\noindent With $r\ll{1}$ to Eq. (7), we get a
forced-oscillator-like differential equation
\begin{equation}
\ddot{r}(t)\, +\, (2\Omega_0)^2\, r(t)\, =\, Q\, F(t),
\end{equation}
which is treated by going to the complex notation
$w(t)=[\dot{r}(t)+i(2\Omega_0)r(t)]$ in order to derive a
first-order differential equation for $w(t)$. Remarkably, Eq. (8)
shows transparently that the breathing frequency \cite{Brabec13} is
precisely $2\times{\Omega_0}$, i.e., the double of the confinement
frequency characterizing the ground-state mode of the unperturbed
$\hat{h}_0(X)$.
Once an explicit form for $F(t)$ is prescribed to be used in both
($i=1,2$) differential equations for $r_i(t)$ [considering the two
independent {\it modes}, with $\Omega_0\rightarrow{\omega_1}$ and
$\Omega_0\rightarrow{\omega_2}$] and the solutions for
$r_i(\omega_i,t)$ are found, the time-dependent wave function
becomes
\begin{equation}
\Psi(X_1,X_2,t)\, =\, \psi(X_1,\omega_1,R_1,t)\, \psi(X_2,\omega_2,R_2,t),
\end{equation}
which is valid for $|Q|<{\omega_i}$, and to which the form for
$\psi(X_i)$ is given by Eq. (5). By using this
normal-coordinate-based representation for the exact wave-function
we can easily calculate with it and the {\it unperturbed}
Hamiltonian, the expectation values of the kinetic and potential
energy components, $<K(t)>$ and $<V(t)>$, respectively. We get for
these quantities
\begin{equation}
<K(t)>\, \, =\, \, \sum_{i=1}^2\, \frac{\hbar\, \omega_i}{4}\,
\left[\frac{1}{R_i^2(\omega_i,t)}\, +
\frac{\dot{R}_i^2(\omega_i,t)}{\omega_i^2} \right]
\end{equation}
\begin{equation}
<V(t)>\, \, =\, \sum_{i=1}^2\, \frac{\hbar\, \omega_i}{4}\,
R^2_i(\omega_i,t).
\end{equation}
In the linear-response limit, the total energy {\it change} $\Delta
E(t)\equiv{E(t)-E(t_0)}$ becomes
\begin{equation}
\Delta E(t)\, =\, \sum_{i=1}^2\, \Delta E_i(t)\,\Rightarrow\,{
\sum_{i=1}^2\, \frac{\hbar\, \omega_i}{4}\,
\left[\frac{(2\omega_i)^2\, r_i^2(\omega_i,t)+
\dot{r}_i^2(\omega_i,t)}{\omega_i^2}\right]}.
\end{equation}
To arrive at the consistent r.h.s. above, which is valid for weak
external fields, we linearized Eqs. (10-11). In this case [with Eq.
(8)] one gets $\dot{E}_i(t)=(\hbar\omega_i/4)[2\,
\dot{r}_i(\omega_i,t)\, Q \, F(t)/\omega_i^2]$ for the rate. Next,
we derive for the overlap $O(t)=|\Psi(t_0)\Psi(t)|^2$ the following
expression
\begin{equation}
O(t)\, =\, \frac{1}{\sqrt{1+\Delta E_1(t)/\hbar \omega_1}\, \,
\sqrt{1+\Delta E_2(t)/\hbar \omega_2}},
\end{equation}
in terms of $\Delta E_i(t)$. This quantity is a measure of
correlated (when $\omega_1\neq{\omega_2})$ dynamics. Its product
form reflects the independence of two normal modes during
propagation. When we are close to the stability limit at
$\Lambda\rightarrow{0.5}$, even a weak external field could result
in an almost perfect orthogonality at certain, $F(t)$-dependent,
time $t>t_0$.
Now, we turn to quantities which show explicitly the entangled
nature of the correlated two-particle system in the time domain. By
rewriting the wave function in terms of original coordinates, we
determine the reduced single-particle density matrix
[$\Gamma_1(x_1,x_2,t)$] from
\begin{equation}
\Gamma_1(x_1,x_2,t)\, =\, \int_{-\infty}^{\infty}\, dx_3\,
\Psi^{*}(x_1,x_3,t)\, \Psi(x_2,x_3,t),
\end{equation}
After a long, but straightforward, calculation we obtain
\begin{equation}
\Gamma_1(x_1,x_2,t)=\phi_s(x_1,t)\, \phi_s^{*}(x_2,t)\,
e^{-\frac{m}{2\hbar}A(t)(x_1-x_2)^2},
\end{equation}
where, for further clarifications below, we introduced the following
abbreviations
\begin{equation}
\phi_s(x,t)=\left[\frac{m \omega_s(t)}{\hbar \pi}\right]^{1/4}\,
e^{-\frac{m}{2\hbar}\omega_s(t)\, x^2[1-i\, \alpha(t)/\omega_s(t)]},
\end{equation}
\begin{equation}
\frac{\alpha(t)}{\omega_s(t)}\, =\, \frac{1}{2}\,
\left[\frac{R_1(\omega_1,t)\dot{R}_1(\omega_1,t)}{\omega_1}+\frac{R_2(\omega_2,t)\dot{R}_2(\omega_2,t)}{\omega_2}
\right]
\end{equation}
\begin{equation}
A(t)\, =\, \frac{1}{4}\, \frac{[\omega_1(t)-\omega_2(t)]^2
+[\dot{R}_1(\omega_1,t)/R_1(\omega_1,t) -
\dot{R}_2(\omega_2,t)/R_2(\omega_2,t)]^2} {\omega_1(t)+\omega_2(t)}
\end{equation}
with
$\omega_s(t)=2\omega_1(t)\omega_2(t)/[\omega_1(t)+\omega_2(t)]$,
where $\omega_i(t)=\omega_i/[R_i(\omega_i,t)]^2$. The one-matrix is
Hermitian, as it must be, since
$\Gamma_1(x_1,x_2,t)=\Gamma_1^{*}(x_2,x_1,t)$. Later we will give,
as in the stationary case \cite{Glasser13} earlier, a direct
spectral decomposition of this important Hermitian matrix in
Eq.(15), without considering, as more usual, an eigenvalue problem.
\newpage
The diagonal of the reduced single-particle density matrix gives
the basic variable of TD-DFT, i.e., the one-particle probability
density
\begin{equation}
n(x,t)=\left[\frac{m \omega_s(t)}{\hbar \pi}\right]^{1/2}\,
e^{-\frac{m}{\hbar}\omega_s(t)\, x^2}.
\end{equation}
The single-particle probability current, $j(x,t)$, is calculated as
usual \cite{Ballentine98,Lein05,Nagy13} from
\begin{equation}
j(x,t)\, =\, \frac{2 \hbar}{m}\, Re\, \int\,
\Psi^{*}(x,x',t)\frac{\partial}{i\, \partial x}\, \Psi(x,x',t)\,
dx'.
\end{equation}
which, in terms of the functions introduced above, becomes
\begin{equation}
j(x,t)\, =\, \frac{2 \hbar}{m}\, x\, n(x,t)\, \alpha(t).
\end{equation}
These probability densities satisfy the continuity equation:
$\partial_t n(x,t)+\partial_x j(x,t)=0$.
By taking the $A(t)\rightarrow{0}$ substitution in Eq.(15), i.e.,
neglecting the important role of the interparticle coordinate, we
can define an auxiliary independent-particle model, where the
particles move in a {\it certain} external field. The wave function
of this modeling becomes
\begin{equation}
\Psi_s(x_1,x_2,t)\, =\, \phi_s(x_1,t)\, \phi_s(x_2,t).
\end{equation}
This approximate state will still result in the exact probability
density and exact probability current, by construction. So, one may
consider it as an {\it output} of TD-DFT. Indeed, using Eqs.
(6.50-6.52) of \cite{Ullrich12} or Eqs. (21-22) of \cite{Campo13},
an effective potential $V_{s}(x,t)$ is given by
\begin{equation}
V_{s}[x,t,\omega_s(t)]\, =\, \frac{1}{2}\omega_{s}^2(t)\, x^2\, -\,
\frac{1}{2}\left[\sqrt{\omega_{s}(t)}\, \frac{d^2}{d
t^2}\left(\frac{1} {\sqrt{\omega_{s}(t)}}\right)\right] x^2,
\end{equation}
at least upto a time-dependent constant \cite{Lein05,Nagy13}. In the
so-called adiabatic treatment in TD-DFT, one uses \cite{Ullrich12}
only the first term as a time-dependent external field.
But, with such a single-particle potential, which produces from a
time-dependent wave equation the $\Psi_s(x_1,x_2,t)$ state, we also
arrive at a dilemma. We know the exact Schr\"odinger Hamiltonian.
Especially, we know how the unperturbed Hamiltonian is modified by
adding a time-dependent quadrupolar perturbation to it. Furthermore,
since the perturbation acts only over a limited time in our problem,
in the Schr\"odinger picture we can follow the standard quantum
mechanical prescription to determine the energy change by
calculating expectation values with the exact wave function
$\Psi(X_1,X_2,t)$ and unperturbed Hamiltonian $\hat{H}_0(X_1,X_2)$.
In the light of $V_{s}[x,t,\omega_s(t)]$, the situation in the
orbital version of TD-DFT is not so clear. For instance, it will
contain a $\ddot{\omega}_s(t)$-proportional higher-order derivative
\cite{Ullrich12,Schirmer07,Campo13} . Quite unfortunately, in
TD-DFT, which is based \cite{Ullrich12,Leeuwen15,Schirmer07} on a
density-potential mapping lemma, we do not \cite{Dreizler86} know
universal functionals to determine an energy change via them.
\newpage
After the above remarks on open questions in applied DT-DFT, and
motivated partly by Dreizler's early suggestion \cite{Dreizler86} on
using an isospectral deformation ($d$) and a constrained search, we
turn to the decomposition of our reduced one-particle density
matrix, and, in addition, put forward an idea on the possibility of
using an {\it approximate} nonidempotent one-matrix,
$\Gamma_1^s(x_1,x_2,t)$, designed below. Following our experience
\cite{Glasser13} on the decomposition of the stationary
$\Gamma_1(x_1,x_2)$, we will use Mehler's formula
\cite{Erdelyi53,Koscik15} now to $\Gamma_1(x_1,x_2,t)$.
Thus, via the
$[\omega_s(t)+A(t)]=\bar{\omega}(t)[1+z^2(t)]/[1-z^2(t)]$ and
$A(t)=\bar{\omega}(t)2z(t)/[1-z^2(t)]$ correspondences, we introduce
two new variable $\bar{\omega}(t)$ and $z(t)$. The solutions are
\begin{equation}
\bar{\omega}(t)\, =\, \omega_s(t)\, \frac{1+z(t)}{1-z(t)},
\end{equation}
\begin{equation}
z(t)\, =\, \left[\frac{\omega_s(t)}{A(t)} + 1\right] -
\sqrt{\left[\frac{\omega_s(t)}{A(t)}+1\right]^2-1}.
\end{equation}
Now, in term of these we get our point-wise decomposition directly
\begin{equation}
\Gamma_1(x_1,x_2,t)\, =\, \sum_{l=0}^{\infty}\, [1-z(t)][z(t)]^l\,
\phi_l(x_1,t)\, \phi_l^{*}(x_2,t),
\end{equation}
where the $\phi_l(x,t)$ elements of the orthonormal basis set
(natural orbitals) are
\begin{equation}
\phi_l(x,t)\, =\,
\left[\frac{m\bar{\omega}(t)}{\hbar\pi}\right]^{1/4}\, e^{i\,
\frac{m}{2\hbar}\alpha(t)x^2}\, \frac{1}{\sqrt{2^l\, l!}}\,
e^{-\frac{m}{2\hbar}\bar{\omega}(t)x^2}\,
H_m(\sqrt{m\bar{\omega}(t)/\hbar}x).
\end{equation}
Of course, extracted quantities, like the exact one-matrix
$\Gamma_1(x_1,x_2,t)$ and the associated exact probability density
$n(x,t)$, contain less and less information then the wave function
$\Psi(x_1,x_2,t)$ and the pair-function
$n_2(x_1,x_2,t)=\Psi(x_1,x_2,t)\Psi^{*}(x_1,x_2,t)$. However, even
by these extracted quantities one can calculate the exact kinetic,
and the external-potential energies, respectively, while in TD-DFT
only the last energy is, in principle, exact. Thus, to formulate an
idea, which is based on inversion from {\it known} probability
density $n(x,t)$ and probability current $j(x,t)$ (i.e., from the
crucial \cite{Schirmer07,Rapp14} $\alpha(t)$ phase), we make a
rewriting
\begin{equation}
n(x,t)=\sum_{l=0}^{\infty}\, [1-z(t)][z(t)]^l
|\phi_l[x,t,\bar{\omega}(t)|]^{2}= \sum_{l=0}^{\infty}
[1-z_d(t)][z_d(t)]^l|\phi_l[x,t,\omega_d(t)]|^{2},
\end{equation}
which is valid if and only if $\omega_d(t)\, =\, \omega_s(t)\,
[1+z_d(t)]/[1-z_d(t)]$.
So, with one-particle probabilistic functions as input, we can
prescribe, from the r.h.s. of the above equation, a trace-conserving
$\Gamma_1^s(x_1,x_2,t)$ which has the form of $\Gamma_1(x_1,x_2,t)$.
Such nonidempotent form contains two (interconnected) parameters
$\omega_d(t)$ and $z_d(t)$, instead of the uniquely derived
$\bar{\omega}(t)$ and $z(t)$. Whether this deformed one-matrix could
result in, because of its {\it tunable} flexibility, an essentially
better approximation for the kinetic energy than the TD-DFT
[$z(t)=z_d(t)\equiv{0}$] where the one-matrix is idempotent,
requires future investigations within the framework of a constrained
search \cite{Dreizler86,Riveros14}.
The information-theoretic aspects of iterparticle interaction, i.e.,
of a crucial term in the Hamiltonian, may be considered via entropic
measures which do not use expectation values with physical
dimensions but use pure probabilities encoded in the occupation
numbers
\begin{equation}
P_l(t)\, \equiv{[1 - z(t)]\, [z(t)]^l},
\end{equation}
which are, in the present case, time-dependent quantities as well.
First, we calculate R\'enyi's entropy \cite{Renyi70}, with
$q\neq{1}$ and $q>0$, from
\begin{equation}
S_R(q,\Lambda,t)\, \equiv{\frac{1}{1-q}\,
\ln\frac{[1-z(t)]^q}{1-[z(t)]^q}}.
\end{equation}
The von Neumann entropy is obtained as a limiting case when
$q\rightarrow{1}$. It is given by
\begin{equation}
S_N(\Lambda,t)\, \equiv{- \ln[1-z(t)]\, -\frac{z(t)}{1-z(t)}\,
\ln[z(t)]}.
\end{equation}
In the knowledge of $z(t)$, these entropies characterize the
deviation from the stationary case described by $z(t_0)$, i.e.,
without external perturbation on the correlated two-body system. By
considering repulsive [$\Lambda\in(0,0.5)$] and attractive
[$\Lambda<0$] versions, we can investigate the duality
\cite{Glasser13,Schilling14} problem with respect to entropies, now
in the time domain. Notice, that the knowledge of R\'enyi entropies
for arbitrary values of $q$ provides, mainly from
information-theoretic point of view, more information than the von
Neumann entropy alone, since one could extract \cite{Calabrese15}
from them the full spectrum of the one-matrix. But, as Eq. (26)
shows transparently, to physical expectation values one also needs
the proper basis set.
Armed the above, strongly interrelated, theoretical details on the
dynamics of interparticle correlation we come to their
implementation. We use a simple model for illustration, with
$H'(t)\, =\, Q\, F(t) = Q\, \Theta(t)\exp(-\beta\, t)$. By tuning
the value of the effective rate of change (denoted by $\beta$) we
can easily investigate the sudden and adiabatic limits. We get
\begin{equation}
R_i(\Lambda,\omega_i,t)\, =\, 1 +\, \frac{Q}
{\beta^2+(2\omega_i)^2}\left[e^{-\beta t} -\cos(2\omega_i t) +
\frac{\beta}{2\omega_i}\sin(2\omega_i t)\right]
\end{equation}
\begin{equation}
\dot{R}_i(\Lambda,\omega_i,t)\, =\, \frac{Q}
{\beta^2+(2\omega_i)^2}\left[-\beta\, e^{-\beta t}
+(2\omega_i)\sin(2\omega_i t) + \beta\, \cos(2\omega_i t)\right]
\end{equation}
\begin{equation}
\Delta E_i(\Lambda,\omega_i,t)\, =\, \frac{\hbar\, \omega_i}{4}\,
\frac{Q^2}{\beta^2+(2\omega_i)^2}\,
\frac{1}{\omega_i^2}\, \left[1 + e^{-2\beta\, t} - 2\, e^{-\beta\,
t}\, \cos(2\omega_i t)\right]
\end{equation}
The sum ($i=1,2$) of $\Delta E_i(\Lambda,\omega_i,t)$ measures the
total deviation from the unperturbed ground-state energy
$E(t=0)=(\hbar/2)(\omega_1+\omega_2)$. In the kick-limit, where
$\beta t\rightarrow{0}$, we have $\Delta
E_i(\Lambda,\omega_i,t)\propto{\hbar \omega_i(Q\beta
t/\omega_i)^2/[\beta^2+(2\omega_i)^2]}$. As expected on physical
grounds, the change at long time, i.e., taking
$t\rightarrow{\infty}$ first, becomes a constant which tends to zero
in the $\beta\rightarrow{\infty}$ sudden limit. In the long-time
limit the overlap parameter of Eq. (13), which measures the
time-evolution of orthogonality between exact states, also tends to
a constant value.
Simple independent-particle modelings of the correlated model
system could be based on effective ($e$), prefixed external fields
$(m/2)\omega^2_e(x^2_1+x^2_2)$ before time-dependent perturbation.
In such cases one gets $\Delta
E_e(t=\infty,\omega_e)=2\times{{(\hbar
\omega_e/4)(Q/\omega_e)^2/[\beta^2+(2\omega_e)^2]}}$ for the total
energy change, following the standard quantum mechanical averaging
procedure discussed at Eq. (23). Two options are analyzed here, and
a remarkable result is found, as Figure 1 signals. Namely, with
$\omega_e\rightarrow{\omega_s(t=0)}$, i.e., with the density-optimal
modeling, the corresponding energy {\it change} tends to the exact
one at high enough values of $\beta$, since $2 \omega_s^{-1}(t=0) =
(\omega_1^{-1}+\omega_2^{-1})$. Somewhat surprisingly, by using the
so-called energy-optimal Hartree-Fock \cite{Moshinsky68,Pipek09}
modeling, where $\omega_e\rightarrow{\omega_0\sqrt{1-\Lambda}}$, one
cannot get agreement with the exact result at any value of $\beta$.
For illustration, a convenient ratio-function is defined as
\begin{equation}
\mathcal{R}(\Lambda,\beta,t)\, =\, \frac{\Delta
E_e(\Lambda,\beta,t)}{\Delta E(\Lambda,\beta,t)},
\end{equation}
and plotted in Figure 1 as a function of the $\beta$ with fixed,
density-optimal, $\omega_e=\omega_s(t=0)$. We have checked, using
both sides of Eq. (12), that with a $Q=0.2$ value we are in the
linear-response limit, where the energy changes are proportional to
$Q^2$. Notice, that we use standard Hartree atomic units, where
$m=\hbar=1$, to both Figures of this work.
\begin{figure}
\scalebox{0.3}[0.3] {\includegraphics{fig-TD2-energy-ratio.eps}}
\caption {Ratio of the energy changes $\mathcal{R}(\Lambda,t)$, as a
function of $t$ which is measured in atomic units. The solid and
dashed curves refer to interparticle repulsion ($\Lambda=0.4$) and
attraction ($\Lambda=-2.0$), respectively. We used $Q=0.2$ and
$\omega_0=1$. The top, middle, and bottom panels refer,
respectively, to the $\beta=0.25$, $\beta=0.5$, and $\beta=1.0$
values. \label{figure1}}
\end{figure}
As we can see from the Figure, far from the sudden limit, i.e., at a
small $\beta$ value, the density-optimal modeling underestimates the
true energy change in the correlated case. We speculate that such
sensitivity close to the adiabatic limit could be, at least
partially, behind yet unclarified controversies in the applied
TD-DFT \cite{Zeb13,Mao14} to energy transfer by slow ions in
wide-gap insulators, where $\Lambda$ is positive of course. In
Section III, we will return to the informations given in Figure 1,
considering there the energy changes for negative $\Lambda$.
At the end of this Section, we turn to illustrations of the spectral
aspect of dynamic interparticle correlation. We apply von Neumann
entropy to get a precise insight into the time-dependence of an
information-theoretic measure. In the light of facts on
Hamiltonian-based measures, the energy-change and the wave-function
overlap at $t\rightarrow{\infty}$, such analysis is desirable. Based
on it, one can discuss, and maybe understood, challenging
interrelations between measures which are based on different
emphasizes within quantum mechanics. We stress here that the exact
energy-changes depend, as expected, on the sign of $\Lambda$.
\begin{figure}
\scalebox{0.3}[0.3] {\includegraphics{fig-TD1-entropy.eps}} \caption
{Information-theoretic, von Neumann, entropies $S_N(\Lambda,t)$, as
a function of the time $t$, which is measured in atomic units. The
solid and dashed curves refer to a moderate interparticle repulsion
$\Lambda=0.3$, and the corresponding (see the text) interparticle
attraction $\Lambda=-0.75$, respectively. Three values for the rate
parameter $\beta$, measured in inverse time-units, are employed to
illustration. Top panel: $\beta=2$. Middle panel: $\beta=5$. Bottom
panel: $\beta=10$. \label{figure2}}
\end{figure}
In Figure 2 we exhibit $S_N(\Lambda,t)$, by taking $\omega_0=1$,
$Q=0.2$, $\Lambda=0.40$ and $\Lambda=-2.0$. These couplings for
repulsion and attraction, respectively, are fixed to yield equal
entropies at $t=0$, in order to shed further light on duality
\cite{Pipek09,Glasser13,Schilling14}, now in the time-domain as
well. Notice, that in the unperturbed case to {\it any} allowed
[$\Lambda\in(0,0.5)]$ repulsive coupling there exists a
corresponding attractive ($\Lambda<0$) one for which the calculated
entropies are equal. For instance, when we are close to the
stability limit in the repulsive ($r$) case, i.e., when
$\Lambda_r\rightarrow{0.5}$, the corresponding attractive ($a$)
coupling becomes $\Lambda_a\rightarrow{(1/2)(2\Lambda_r -1)^{-1}}$.
The solid and dashed curves refer to repulsion and attraction, as we
mentioned. They are plotted as a function of the time measured in
atomic units. The rate-parameter of the time-dependent external
perturbation is described by $\beta$ and measured in inverse units
of time. In order to discuss tendencies, we used three values for
this parameter. As expected, at very short times the entropies,
similarly to the energy changes, grow in time. After about 1-2
atomic units (i.e., about 25-50 attoseconds) they start to reflect
the oscillating character builded into the time-dependent occupation
numbers $P_l(t)$, the primary root of which is in the time-evolving
correlated wave function. The entropies will keep, in absence of an
environment-made dissipative \cite{Albrecht75,Tokatly13} coupling,
such oscillating behavior without limitation. From this point of
view, they signal the time-evolvement of the closed interacting
system. More importantly, one can see from the illustration that the
initial dual character of entropies, fixed at $t=0$, will {\it
disappear} due to a new physical variable, i.e., the time. Finally,
there are reductions in the oscillation-amplitudes by increasing
$\beta$. Clearly, at the so-called sudden limit
($\beta\rightarrow{\infty}$) for a time-dependent perturbation, they
will diminish as expected, since the system's reaction-rapidity is
limited by its normal mode frequencies.
In order to justify the above argument more directly, we take now a
somewhat less realistic modeling by using an abrupt quench of {\it
all} interactions at $t_0=0$ in the Hamiltonian. In that case,
instead of Eqs. (33-34), one gets \cite{Kagan96} the following exact
expressions
\begin{equation}
R_i(\Lambda,\omega_i,t)\, =\, \sqrt{1+t^2\, \omega_i^2}
\end{equation}
\begin{equation}
\dot{R}_i(\Lambda,\omega_i,t)\, =\, \frac{t\,
\omega_i^2}{\sqrt{1+t^2\, \omega_i^2}}.
\end{equation}
There is no change in the system's kinetic energy since
$[1/R_i(\omega_i,t)]^2+[\dot{R}_i(\omega_i,t)/\omega_i]^2=1$. By
using time-dependent $A(t)$ and $\omega_s(t)$, which are needed to
Eq. (25), we get $z(t)=z(t=0)$ for all $t\geq{0}$ since
$[\omega_s(t)/A(t)]\equiv{2[2\sqrt{\omega_1\omega_2}/(\omega_1-\omega_2)]^2}$.
Thus, there is no entropy-{\it change} in the total-quench case.
This $S(t)=constant$ character could be a useful constraint in
practical attempts, with a good starting $n(x,t)$, using the
interconnected $z_d(t)$ and $\omega_d(t)$ variables of Eq. (28). One
can take $\omega_d(t)\Rightarrow{\omega_d/R^2(\omega_d,t)}$, where
$R(\omega_d,t)=\sqrt{1+t^2\omega_d^2}$.
\section{summary and comments}
Based on the time-dependent Schr\"odinger equation, an exact
calculation is performed for the time-evolving energy change, and
orthogonality of initial and final states, in a correlated
two-particle model system driven by a weak external field of
quadrupolar character. Besides, considering the structure of the
exact time-dependent one-matrix, an independent-particle model is
defined from it which contains exact information on the
single-particle probability density and probability current. The
resulting noninteracting auxiliary state is used to construct an
effective potential and discuss its applicability. We analyzed the
energy changes in a comparative manner, and pointed out (see, below
as well) the limited capability of an optimized independent-particle
modeling for the underlying time-dependent process.
A point-wise decomposition of the reduced one-particle density
matrix is given in terms of time-dependent occupation numbers and
time-dependent orthonormal orbitals. Based on such exact spectral
decomposition on the time domain, an entropic measure of inseparable
correlation is also investigated. It is found that the duality
behavior characterizing the stationary ground-state may disappear
depending on the rapidity of the external time-dependent
perturbation. At realistic rapidity, the undamped time-dependent
oscillations, found in an entropic measure, reflect the fact that
the correlated system evolves in time. In this respect, we have a
signal on the excited state, similarly as one can see in the
time-evolving probability density and probability current.
Our first comment is based on results obtained for the energy
changes, and plotted in Figure 1, by considering the sign of
$\Lambda$. In the case of an attractive interparticle interaction,
the case of a nucleus \cite{Moshinsky68}, the exact result for
$\Delta E(\Lambda<0,\beta,t)$ becomes bigger at moderate $\beta$
values than the result obtained within the density-optimal framework
$\Delta E_e(\Lambda<0,\beta,t)$. So, we speculate that a recent
result \cite{Stetcu15} on the excitation of a deformed nucleus
within TD-DFT could have the same, underestimating character in its
quadrupole channel, as the atomistic case has. Thus, the true
enhancement in excitations over the result based on the modeling of
Teller \cite{Teller48}, might also need further theoretical
refinement for that channel.
In the second comment, we focus on the role of the switching rate
$\beta$ in the case of energy changes. As a preliminary step to
future realistic calculations, here we would like to estimate a
characteristic interaction-time ($T\propto{\beta^{-1}}$) in
classical atom-atom collision. To do that, we model the
three-dimensional screened interaction via a finite-range
($r\leq{R}$) potential
\begin{equation}
V(r)\, =\, \frac{Z_1\, Z_2\, e^{2}}{r}\left(1 - \frac{r}{R}\right).
\end{equation}
Solving the classical problem \cite{Nagy94} for the collision time
($T$) we get
\begin{equation}
T(b,v,R)=\frac{2}{v}\left[\frac{\sqrt{R^2-b^2}}{c}+
\frac{Z_1Z_2e^2}{2E_1}\frac{1}{c^{3/2}}\ln\frac{\sqrt{c}\sqrt{R^2-b^2}+R+(Z_1Z_2e^2/2E_1)}{\sqrt{(Z_1Z_2e^2/2E_1)^2+c\,
b^2}}\right]
\end{equation}
where $E_1=M_1v^{2}/2$ is the energy of a heavy charged ($Z_1e$)
projectile moving with velocity $v$ and colliding at impact
parameter $b$ with a fixed center of {\it effective} charge $Z_2e$.
The dimensionless parameter introduced is $c=[1+(Z_1Z_2e^2/R)/E_1]$,
i.e., it is related to the ratio of potential and kinetic energies.
A closer inspection of this compact (and reasonable at short-range
solid-state conditions) form shows that practically one may use
\begin{equation}
T(b,v,R)\, \simeq{\alpha(v)\, \frac{R}{v}\,
\frac{E_1}{E_1+Z_1Z_2e^2/R}}\, \sqrt{1-\frac{b^2}{R^2}}
\end{equation}
to get a very acceptable estimation as a function of the heavy
projectile velocity $v$, with which $\alpha(v)$ is {\it decreasing}
from 4 (at $v\rightarrow{0}$) to 2 (at $v\rightarrow{\infty}$). If
we take $\beta(v)\simeq{T^{-1}(v)}$ and consider the energy change
determined above, we get a $\Delta
E(v)\sim{[\beta(v)]^{-2}}\sim{[\alpha(v)v]^2}$ character at very low
velocities. For the opposite, high-velocity, limit $\Delta
E(v)\sim{[\beta(v)]^{-2}}\sim{[\alpha(v)v]^{-2}}$ is the scaling,
which also looks quite reasonable physically. However, the detailed
work on the exactly determined non-linear energy changes is left for
a dedicated publication.
There, and also in Bohr's pioneering modeling
\cite{Bohr13,Merzbacher74} with time-dependent dipole fields of a
passing charge, the time-scale derived above using collision theory,
could play a more quantitative role to understand fine details
behind experimental predictions \cite{Moller04,Markin09} on energy
losses of slow ions in insulators. Furthermore, in the renewed field
of time-dependent energy losses, a proper time-scale also could help
to analyze further a remarkable theoretical prediction
\cite{Correa12} on the strong interplay of nuclear and electronic
stopping components of the observable total energy transfer. Indeed,
few-electron shake-up processes can be dominant at close encounters
in condensed matter. In such dynamical cases, one-electron
approximations, with double-occupancy for an auxiliary spatial
orbital, could result in inaccuracies.
\begin{acknowledgments}
We thank Professor P. M. Echenique for the very warm hospitality at
the DIPC. One of us (IN) is grateful to Professor P. Bauer and
Professor D. S\'anchez-Portal for useful discussions on energy
transfer processes in insulators. This work was supported in part by
the Spanish Ministry of Economy and Competitiveness MINECO (Project
No. FIS2013-48286-C2-1-P).
\end{acknowledgments}
|
\section{Introduction}
\label{ellWF:ellWFintro}
Compact binary systems composed of neutron stars and/or black holes are one of
the most promising sources for the second generation of earth-bound
gravitational-wave (GW) detectors such as Advanced LIGO
\cite{2014arXiv1411.4547T} and Advanced Virgo \cite{TheVirgo:2014hva} as well
as for the proposed space-based detector eLISA \cite{2012arXiv1201.3621A}.
Detection of such systems in GW detectors relies on a data-analysis technique
known as {\it matched filtering} which in turn requires very accurate modeling
of GW signals from these sources \cite{Th300}. The compact binaries are known
to have significant eccentricities when they are formed. However, since the GW
radiation reaction effects tend to circularize the binary's
orbit~\cite{PM63,Pe64}, for most long-lived binary systems one can expect that
their orbits would have circularized by the time they enter the sensitivity
band of ground-based detectors. This has motivated the GW community to perform
searches of GW signals from coalescing compact binary systems (CCBs) using
circular orbit templates.\\
Many astrophysical scenarios have been proposed which suggest the possible
existence of close eccentric binary systems.\footnote{Eccentric binary systems,
with small orbital separations, formed through the capture process. Unlike the
long-lived CCBs such systems are expected to enter the detector with nonzero
eccentricities.} One such scenario may exist in the cores of dense globular
clusters due to a mechanism known as the ``Kozai mechanism'' \cite{KozaiOsc62}.
This mechanism can also come into effect in scenarios involving formation of
hierarchical triples of supermassive black holes due to subsequent mergers of
galaxies \cite{BLSKozai02}. Another scenario might involve formation of {\it
close} eccentric compact binary systems in dense stellar systems like globular
clusters \cite{BenacquistaLiving}. Compact binaries involving intermediate mass
black holes in globular clusters might be seen in the eLISA band with residual
eccentricities of $0.1 \lesssim e \lesssim 0.2$ \cite{GMHimbh04}. Other scenarios
involve formation of close eccentric compact binary systems at centers of
galaxies \cite{Kupi:2006mh} and NS-BH binary systems which can become eccentric
as a consequence of multi stage mass transfer from the NS to the BH
\cite{DavLevKing04}. In light of these possibilities it becomes necessary to
compute accurate waveforms accounting for the eccentricity of the binary's
orbit. \\
A number of investigations concerning the sensitivity of searches using
circular orbit templates to detect eccentric binary systems have been performed
in the past. The first such investigation was presented in
Ref.~\cite{PhysRevD.60.124008} where the authors studied the loss of
signal-to-noise ratio in detecting signals from binaries with residual
eccentricities using template banks constructed with quasi-circular waveform
models with leading order effects (both {\it conservative} and {\it secular}).
They argued that even if the system has a residual eccentricity ($e_0 \lesssim
0.13$ for binary system with two $1.4M_\odot$ neutron stars or $e_0 \lesssim 0.3$
for a binary with two $6M_\odot$ black holes), use of circular orbit templates
will be sufficient to detect signals from such systems.\footnote{They chose a
lower cutoff for the fitting factor (${\rm FF}\vert {\rm min}=90\%$)
corresponding to a loss in event rates of about less than $27\%$.} However,
this result has been subsequently weakened due to two independent
investigations \cite{PhysRevD.78.084029, PhysRevD.81.024007}. Both of the
investigations suggest that if the eccentricity of the binary when it enters
the sensitivity band of detector is greater than 0.1, then it will not be
possible to detect such systems using circular orbit templates. These
investigations only dealt with sources that will be seen in ground-based
detectors. However, the capabilities of circular waveforms to detect signals
from the coalescence of supermassive black holes (visible in the eLISA
frequency band) have been investigated in \cite{Porter:2010mb}. The results
presented in Ref.~\cite{Porter:2010mb} suggest that even to search signals from
sources with initial eccentricities of the order $10^{-4}$ one would need
waveforms which accurately account for the effects of eccentricities. In
addition, in a recent work, Huerta and Brown \cite{Huerta:2013qb} showed that
searches for CCBs with eccentricity $\geq 0.05$ would require eccentric
template banks to avoid significant loss in the sensitivity of the search.
Lastly, systematic errors due to the orbital eccentricity in measuring the
source parameters of double NS systems was investigated recently by
Favata~\cite{Favata2013} which again indicated the necessity to incorporate the
effects of eccentricity to measure the parameters of a double NS system if it
has non-negligible eccentricity when detected.\\
Evolution of a compact binary system can be divided into three stages: the
early inspiral, late inspiral and merger and the final ringdown. The early
inspiral phase can be very well modeled using the approximation schemes such
as multipolar post-Minkowskian (MPM) approximation matched to
post-Newtonian(PN)~\cite{Blanchet:2013haa} whereas the late inspiral, merger
and ringdown phases can be modeled using numerical relativity
(NR)~\cite{Pretorius07Review} or effective one body approach~\cite{BuonD98}. In
fact, it is now possible to perform numerical simulations to track the
evolution of the BH binary systems over many inspiral orbits and the subsequent
merger and ringdown phases. However, {\it high} computational cost of
generating numerical waveforms covering the entire parameter space of
coalescing binary black holes (BBHs) has led to the construction of hybrid
waveforms (by combining PN and NR waveforms), which further are used to
phenomenologically construct a waveform model which has sufficient overlap with
the hybrid waveform \cite{ Pan:2011gk, Ajith:2007qp, Ajith:2007kx,
Ajith:2009bn, Buonanno:2007pf, Ajith:2012tt}. In addition to this, one needs to
check the consistency between these two waveforms (PN $\&$ NR) in a regime
where both of them are valid. This would not only tell us about the
compatibility of the two waveforms but also would indicate the limits up to
which PN waveforms are reliable. There have been many such investigations
involving nonspinning and nonprecessing BBH in quasi-circular orbits
\cite{BCCKM06, BCP07NR, NRPNGoddard07, NRPNJena07, Gopakumar:2007vh,
Boyle:2008ge, Damour:2007vq, Boyle:2009dg, Hannam:2010ec, Berti:2007fi,
NRPNCaltech07} and quasi-eccentric orbits \cite{2010PhRvD..82b4033H}. The need
for such comparisons and matching of the two waveforms (PN and NR) has led to
the high accuracy computations of spherical harmonic modes of the PN waveforms
in case of CCBs moving in quasi-circular orbits \cite{BFIS08, K08,
Faye:2012we}. Evidently, in order to perform similar comparisons for eccentric
binaries, one would need high accuracy eccentric PN waveforms for such systems.
\\
The leading order (or Newtonian) expressions for the GW polarizations $(h_+
~{\rm and}~ h_\times)$ were obtained in the context of spacecraft Doppler
detection of GWs from an isolated compact binary in an eccentric orbit
\cite{1987GReGr..19.1101W}. This work was then extended to 1PN and the next
1.5PN order in \cite{WagW76, BS89, JunkS92, BS93, RS97}. At the 2PN order, the
transverse-traceless radiation field $(h^{\rm TT}_{ij})$ due to an isolated
binary composed of two compact stars moving in eccentric orbits was computed in
\cite{WWi96, GI97}. Although the two works, \cite{WWi96} and \cite{GI97},
followed two different approaches, their final findings were in perfect
agreement with each other. Under the {\it adiabatic} approximation, associated
2PN GW polarizations ($h_+, h_\times$) were obtained in \cite{GI02} for the
inspiral phase of binaries in quasi-eccentric orbits. Later in
Refs.~\cite{DGI04, KG06}, the method of variation of constants was used to
compute post-adiabatic corrections (varying on the orbital time scale and
$1/c^{5}$ times smaller) to the secular variation due to radiation reaction.
Using the 3PN generalized quasi-Keplerian representation of the conservative
dynamics of compact binary systems with arbitrary mass ratios moving in
eccentric orbits presented in \cite{MGS04}, Ref.~\cite{KG06} provides the
evolution of the orbital phase with relative 1PN accuracy (absolute 3.5PN).
The energy and angular momentum fluxes as well as evolution of orbital elements
up to 3PN order was calculated in Refs.~\cite{ABIQ07tail,ABIQ07,Arun:2009mc}.
Recently, computations of the frequency domain waveforms and the orbital
dynamics (both at the 2PN order) were presented for eccentric binaries in
harmonic coordinates \cite{Tessmer:2010ii}. On the NR front, the first
simulations involving nonspinning equal mass BBHs in bound eccentric orbits
were performed in \cite{JenaEccentric07, HinderEccentric07} and the effects of
eccentricity on the final mass and spin were studied. Another recent work
\cite{2010PhRvD..82b4033H} presents numerical simulations for a nonspinning
equal mass binary system with an initial eccentricity of $e\sim 0.1$ and
compares the NR waveforms with those of the PN models.\\
In this paper we present the computation of instantaneous part\footnote{The
part of the gravitational radiation which depends on the state of its source at
a given retarded time.} of various modes of the waveform ($h^{\ell m}$) for
general orbits using the basis of spherical harmonics of spin weight -2. In
addition we also specialize to the case of compact binaries in quasi-elliptical
orbits and provide 3PN instantaneous expressions for various modes using 3PN
quasi-Keplerian representation of the conserved dynamics of compact binaries in
eccentric orbits \cite{MGS04, ABIQ07}. Note again that investigations
presented here involve only the contributions from the instantaneous terms
which must be complemented by computations accounting for the hereditary
effects.\footnote{The part of the gravitational radiation which depends on the
entire dynamical history of the source and is complementary to the
instantaneous part of the radiation.} Computations of hereditary parts to
various modes of the waveform will form the basis for a companion paper
\cite{Mishra14}.\\
This paper is organized in the following manner. In
Sec.~\ref{ellWF:ellWFhlm-genformulas} we first introduce general formulas for
spherical harmonic modes of the gravitational waveform, $h^{\ell m}$, in terms
of the radiative mass and current multipole moments, $U^{\ell m}$ and $V^{\ell
m}$. Section~\ref{ellWF:ellWFmoments} recalls some of the important aspects of
the MPM-PN formalism and lists various inputs that are needed for computing 3PN
expressions for various modes. These inputs involve relations connecting
radiative moments to source moments, expressions for various source multipole
moments for an isolated compact binary system and equations of motion. In
Sec.~\ref{ellWF:ellWF3pnsphhar} we provide our results related to the
instantaneous part of the spherical harmonic modes of the waveform for a
nonspinning compact binary system in terms of variables that describe the
radiation from a generic compact binary. We find that these expressions are
quite large and run over several pages. Keeping this in mind we choose to list
only the dominant mode ($h_{22}$) in the main text of the paper and provide all
the relevant modes contributing to the 3PN waveform in a separate file readable
in MATHEMATICA (Hlm-GenOrb.m) that will be made available on the journal web
page as Supplemental Material \cite{SupplementalMaterial} along with the paper.
In Sec.~\ref{ellWF:ellWFQKR} we specialize to the case of CCBs moving in
quasi-elliptical orbits and provide the corresponding expression for the
dominant mode, $h^{22}$, in terms of the time-eccentricity $e_t$, a PN
parameter related to the orbital frequency $x$ and the eccentric anomaly $u$.
Similar to the general orbit case, in the case of CCBs in quasi-elliptical
orbits, expressions for all the relevant modes contributing to the 3PN waveform
will be listed in a separate file (Hlm-EllOrb.m). Finally in
Sec.~\ref{ellWF:ellWFconclusion} we conclude the paper by providing a brief
summary of our results and the future plans.
\section{Spherical harmonic modes of the gravitational
waveform} \label{ellWF:ellWFhlm-genformulas}
For an isolated source of GWs, the spherical harmonic modes of the waveform
($h^{\ell m}$), in terms of the radiative mass-type ($U^{\ell m}$) and
current-type multipole moments ($V^{\ell m}$) \cite{BFIS08, Th80, K08,
Faye:2012we}, are given as
\begin{equation} \label{ellWF:inv} h^{\ell m} =
-\frac{G}{\sqrt{2}\,R\,c^{\ell+2}}\left[U^{\ell m}-\frac{i}{c}V^{\ell
m}\right], \end{equation}
where $R$ is the distance of the source in radiative coordinates, $G$ is
Newton's gravitational constant and $c$ is the speed of the light. The
radiative multipole moments, $U^{\ell m}$ and $V^{\ell m}$, appearing above are
related to the symmetric trace-free (STF) radiative moments $U_L$ and $V_L$ as
\begin{subequations}\label{ellWF:UV}\begin{align} U^{\ell m} &=
\frac{4}{\ell!}\,\sqrt{\frac{(\ell+1)(\ell+2)}{2\ell(\ell-1)}} \,\alpha_L^{\ell
m}\,U_L,\\ V^{\ell m} &=
-\frac{8}{\ell!}\,\sqrt{\frac{\ell(\ell+2)}{2(\ell+1)(\ell-1)}}
\,\alpha_L^{\ell m}\,V_L. \end{align}\end{subequations}
Here $\alpha_L^{\ell m}$ denote STF tensors which connect the usual basis of
spherical harmonics $Y^{\ell m}(\Theta, \Phi)$ to the set of STF tensors
$N_{\langle L\rangle}=N_{\langle i_1}\cdots N_{i_\ell\rangle}$ as\footnote{Here
$L=i_1i_2\cdots i_l$ represents a multi-index composed of $l$ spatial indices
and the angular brackets ($\langle} \def\ra{\rangle \ra$) surrounding indices denote symmetric
trace-free projections.}
\begin{subequations}\label{ellWF:NY}\begin{align} N_{\langle
L\rangle}(\Theta,\Phi) &= \sum_{m=-\ell}^{\ell} \alpha_L^{\ell m}\,Y^{\ell
m}(\Theta,\Phi),\\Y^{\ell m}(\Theta,\Phi) &= \frac{(2\ell+1)!!}{4\pi \ell!}\,
\overline{\alpha}_L^{\ell m}\,N_{\langle L\rangle}(\Theta,\Phi).
\end{align}\end{subequations}
In the above, ${\bf N}={\bf X}/{\it R}$ is a unit vector pointing towards the
detector along the line joining the source to the detector. For instance, if
the binary's plane is the $x$-$y$ plane then ${\bf N}$, in terms of angles
($\Theta$, $\Phi$) giving the location of the binary, can be given as
\begin{align} {\bf N}&=\sin \Theta\,\cos \Phi\,\hat{\bf e}_x+\sin \Theta\,\sin
\Phi\,\hat{\bf e}_y +\cos \Theta \hat{\bf e}_z. \end{align}
The STF tensorial coefficients $\alpha_{L}^{\ell m}$ in terms of $N_{\langle
i_1}\cdots N_{i_\ell\rangle}$ and $Y^{\ell m}(\Theta, \Phi)$ can be written as
\footnote{The notation used in \cite{BFIS08, Faye:2012we} (which we follow
here) to the one in \cite{Th80,K08} is related by $\mathcal{Y}_L^{\ell
m}=\frac{(2\ell+1)!!}{4\pi \ell!} \, \overline{\alpha}_L^{\ell m}$.}
\begin{align}\label{ellWF:alpha} \alpha_L^{\ell m} &= \int d\Omega\,N_{\langle
L\rangle}\, \overline{Y}^{\,\ell m}\,,\end{align}
where the usual basis of spherical harmonics is given as
\begin{widetext} \begin{align} Y^{\ell m}(\Theta, \Phi)&=(-)^m\,{1 \over
2^\ell\,\ell!}\left[{2\,\ell+1 \over 4\,\pi} {(\ell-m)!
\over(\ell+m)!}\right]^{1/2}\mathrm{e}^{i\,m\,\Phi}(\sin \Theta)^m \,{{\rm
d}^{\ell+m} \over {\rm d}(\cos \Theta)^{\ell+m}}
\left(\cos^2\Theta-1\right)^\ell. \end{align} \end{widetext}
It is important to note that for nonspinning binaries, there exists a mode
separation as pointed out in Ref.~\cite{K08} and explicitly shown in
Ref.~\cite{Faye:2012we}. The mode $h^{\ell m}$ is completely determined by
mass-type radiative multipole moment ($U^{\ell m}$) when $\ell+m$ is even, and
by current-type radiative multipole moment ($V^{\ell m}$) when $\ell+m$ is odd.
This allows us to write for various modes
\bse \label{ellWF:hlm-modesep} \begin{align} h^{\ell m} &=
-\frac{G}{\sqrt{2}\,R\,c^{\ell+2}}\,U^{\ell m} \quad\quad\quad\quad\quad {\rm
if}\,\ell+m\,\,{\rm is\, even},\\ h^{\ell m} &=
\frac{i\,G}{\sqrt{2}\,R\,c^{\ell+3}}\,V^{\ell m}
\,\,\,\,\,\quad\quad\quad\quad\quad {\rm if}\,\ell+m\,\,{\rm is\, odd}.
\end{align} \end{subequations}
\section{Inputs for computing the 3PN waveform} \label{ellWF:ellWFmoments}
\subsection{Relations connecting the radiative moments to the source
moments}
\label{ellWF:ellWFmoments}
In the MPM-PN formalism \cite{Blanchet:2013haa, BFeom, BIJ02, BI04mult, BDE04,
BDEI05}, the radiative multipole moments ($U_{L}$, $V_{L}$) are first written
in terms of two sets of canonical moments ($M_{L}$, $S_{L}$), which in turn are
expressed in terms of six sets of source moments ($I_L$, $J_L$, $W_L$, $Y_L$,
$X_L$, $Z_L$). Relations connecting radiative moments to the canonical moments
and those connecting the canonical moments to the source moments, with the PN
accuracy desired for the waveform computations at the 3PN order, have been
established and have been listed in Ref.~\cite{BFIS08} (see Eqs.~(5.4)-(5.8) and
Eqs.~(5.9)-(5.11) there). Using these inputs we can parametrize the set of
radiative moments (and, hence, the modes) in terms of source multipole moments.
Below we list all the relevant radiative multipole moments in terms of the
source multipole moments with PN accuracy desired for the present work.
Further, these expressions can be decomposed into two parts namely the {\it
instantaneous} contribution and the {\it hereditary} contribution.\\
The only radiative moment required at the 3PN order is the one related to the
mass quadrupole ($U_{ij}$) and is given by
\be U_{ij}=U_{ij}^{\rm inst}+U_{ij}^{\rm hered}, \label{ellWF:U2decom} \ee
where the instantaneous and hereditary parts in terms of the source multipole
moments read
\begin{widetext} \bse\label{ellWF:U2} \begin{align}\label{ellWF:U2inst}
U_{ij}^{\rm inst}(U) &= I^{(2)}_{ij} (U) +{G \over c^5}\left\{{1
\over7}I^{(5)}_{a\langle i}I_{j\rangle a} - {5 \over7} I^{(4)}_{a\langle
i}I^{(1)}_{j\rangle a} -{2 \over7} I^{(3)}_{a\langle i}I^{(2)}_{j\rangle a}
+{1 \over3}\varepsilon_{ab\langle i}I^{(4)}_{j\rangle a}J_{b} +4\left[W^{(4)}
I_{ij}+W^{(3)} I_{ij}^{(1)}\right.\right.\nonumber\\&\left.\left.-W^{(2)} I_{ij}^{(2)}- W^{(1)}
I_{ij}^{(3)}\right]\right\} +\,\, \mathcal{O}\left(\frac{1}{c^7}\right),\\
U_{ij}^{\rm hered}(U) &= {2G M \over c^3} \int_{-\infty}^{U} d \tau \left[ \ln
\left({U-\tau \over 2\tau_0}\right)+{11 \over12} \right] I^{(4)}_{ij} (\tau)
+{G \over c^5}\left\{-{2 \over7}\int_{-\infty}^{U} d\tau I^{(3)}_{a\langle
i}(\tau)I^{(3)}_{j\rangle a}(\tau) \right\}\nonumber\\ & + 2\left({G M \over
c^3}\right)^2\int_{-\infty}^{U} d \tau \left[ \ln^2 \left({U-\tau
\over2\tau_0}\right)+{57 \over70} \ln\left({U-\tau \over
2\tau_0}\right)+{124627 \over44100} \right] I^{(5)}_{ij} (\tau) +\,\,
\mathcal{O}\left(\frac{1}{c^7}\right)\,. \label{ellWF:U2hered}\end{align} \end{subequations}
\end{widetext}
In the above, the quantity $M$ represents the mass monopole moment or the
Arnowitt, Deser and Misner (ADM) mass of the source. The constant $\tau_0$
appearing in the above integrals is related to an arbitrary length scale $r_0$
by $\tau_0=r_0/c$ and was originally introduced in the MPM formalism. Note
that, numbers in the parenthesis (appearing as superscripts of the source
moments) denote the $p^\mathrm{th}$ time derivatives. The Levi-Civita tensor is
denoted by $\varepsilon_{ijk}$, such that $\varepsilon_{123}=+1$ and ${\cal
O}(1/c^7)$ indicates that we ignore contributions of order 3.5PN and
higher.\\
As may be seen from the above, computing the instantaneous part requires source
multipole moments given at a retarded time $U$. On the other hand, the
hereditary part involves integrals over time and would require the knowledge of
the source multipole moments at any instant of time before $U$ in the past
dynamical history of the source. Further, the hereditary terms are of two
kinds: those with and without the logarithmic factors (see
Eq.~\eqref{ellWF:U2hered} above). The first integral appearing in
Eq.~\eqref{ellWF:U2hered} (the one with the logarithmic kernel inside) is
called the ``tail-integral'' and the one in the last line is called the
``tail-of-tail'' integral whereas the integral without the logarithmic factor
(in the first line) is known as the ``memory'' integral. This paper only
focuses on computing the {\it instantaneous} contribution to various modes of
gravitational waveforms and the computation of hereditary contributions shall
be discussed elsewhere \cite{Mishra14}.\\
Moments required with 2.5PN accuracy are the mass octupole moment $U_{ijk}$ and
the current quadrupole moment $V_{ij}$. The mass octupole moment $U_{ijk}$ is given
as
\begin{align} U_{ijk}&=U_{ijk}^{\rm inst}+U_{ijk}^{\rm hered},
\label{ellWF:U3decom} \end{align}
where $U_{ij}^{\rm inst}$ and $U_{ijk}^{\rm hered}$ in terms of source
multipole moments read
\begin{widetext} \begin{subequations}\label{ellWF:U3}\begin{align} U_{ijk}^{\rm
inst} (U) &= I^{(3)}_{ijk} (U) +{G \over c^5}\left\{ -{4
\over3}I^{(3)}_{a\langle i}I^{(3)}_{jk\rangle a}-{9 \over4}I^{(4)}_{a\langle
i}I^{(2)}_{jk\rangle a} + {1 \over4}I^{(2)}_{a\langle i}I^{(4)}_{jk\rangle a} -
{3 \over4}I^{(5)}_{a\langle i}I^{(1)}_{jk\rangle a} +{1
\over4}I^{(1)}_{a\langle i}I^{(5)}_{jk\rangle a}+ {1 \over12}I^{(6)}_{a\langle
i}I_{jk\rangle a}\right.\nonumber\\&\left.+{1 \over4}I_{a\langle i}I^{(6)}_{jk\rangle a} + {1
\over5}\varepsilon_{ab\langle i}\left[-12J^{(2)}_{ja}I^{(3)}_{k\rangle
b}-8I^{(2)}_{ja}J^{(3)}_{k\rangle b} -3J^{(1)}_{ja}I^{(4)}_{k\rangle
b}-27I^{(1)}_{ja}J^{(4)}_{k\rangle b}-J_{ja}I^{(5)}_{k\rangle
b}-9I_{ja}J^{(5)}_{k\rangle b} \right.\right.\nonumber\\&\left.\left.-{9 \over4}J_{a}I^{(5)}_{jk\rangle
b}\right]+{12 \over5}J_{\langle i}J^{(4)}_{jk\rangle}+4\left[W^{(2)}
I_{ijk}-W^{(1)} I_{ijk}^{(1)} +3\, I_{\langle
ij}Y_{k\rangle}^{(1)}\right]^{(3)}\right\}
+\,\mathcal{O}\left(\frac{1}{c^6}\right) \label{ellWF:U3inst},\\
U_{ijk}^{\rm hered} (U) &={2G M \over c^3} \int_{-\infty}^{U} d\tau\left[ \ln
\left({U-\tau \over 2\tau_0}\right)+{97 \over60} \right] I^{(5)}_{ijk} (\tau)
+{G \over c^5}\left\{ \int_{-\infty}^{U} d\tau \left[-{1
\over3}I^{(3)}_{a\langle i} (\tau)I^{(4)}_{jk\rangle a} (\tau)\right.\right.\nonumber\\&\left.\left.-{4
\over5}\varepsilon_{ab\langle i} I^{(3)}_{ja} (\tau)J^{(3)}_{k\rangle b}
(\tau)\right]\right\} +\,\mathcal{O}\left(\frac{1}{c^6}\right)
\label{ellWF:U3hered}. \end{align}\end{subequations}\end{widetext}
The current quadrupole moment $V_{ij}$ is given as
\begin{align} V_{ij}&=V_{ij}^{\rm inst}+V_{ij}^{\rm hered},
\label{ellWF:V2decom} \end{align}
where $V_{ij}^{\rm inst}$ and $V_{ij}^{\rm hered}$ in terms of source multipole
moments read
\begin{widetext}\begin{subequations}\label{ellWF:V2}\begin{align} V_{ij}^{\rm
inst} (U) &= J^{(2)}_{ij} (U) + {G \over7\,c^{5}}\left\{4J^{(2)}_{a\langle
i}I^{(3)}_{j\rangle a}+8I^{(2)}_{a\langle i}J^{(3)}_{j\rangle a}
+17J^{(1)}_{a\langle i}I^{(4)}_{j\rangle a}-3I^{(1)}_{a\langle
i}J^{(4)}_{j\rangle a}+9I_{a\langle i}I^{(5)}_{j\rangle a} -3I_{a\langle
i}J^{(5)}_{j\rangle a}\right.\nonumber\\&\left.-{1 \over4}J_{a}I^{(5)}_{ija} -7\varepsilon_{ab\langle
i}J_{a}J^{(4)}_{j\rangle b} +{1 \over2}\varepsilon_{ac\langle
i}\left[3I^{(3)}_{ab}I^{(3)}_{j\rangle bc} +{353 \over24}I^{(2)}_{j\rangle
bc}I^{(4)}_{ab} -{5 \over12}I^{(2)}_{ab}I^{(4)}_{j\rangle bc}+{113
\over8}I^{(1)}_{j\rangle bc}I^{(5)}_{ab} \right.\right.\nonumber\\&\left.\left.-{3
\over8}I^{(1)}_{ab}I^{(5)}_{j\rangle bc}+{15 \over4}I_{j\rangle bc}I^{(6)}_{ab}
+{3 \over8}I_{ab}I^{(6)}_{j\rangle bc}\right]+14\,\left[\varepsilon_{ab\langle
i}\left(-I_{j\rangle b}^{(3)}W_{a}-2I_{j\rangle b}Y_{a}^{(2)} +I_{j\rangle
b}^{(1)}Y_{a}^{(1)}\right)\right.\right.\nonumber\\&\left.\left.+3J_{\langle i}Y_{j\rangle
}^{(1)}-2J_{ij}^{(1)}W^{(1)}\right]^{(2)} \right\}
+\,\mathcal{O}\left(\frac{1}{c^6}\right)\,.\label{ellWF:V2inst}\\
V_{ij}^{\rm hered} (U) &= {2G M \over c^3} \int_{-\infty}^{U} d \tau \left[ \ln
\left({U-\tau \over 2\tau_0}\right)+{7 \over6} \right] J^{(4)}_{ij}(\tau)
+\,\mathcal{O}\left(\frac{1}{c^6}\right)\,.\label{ellWF:V2hered}
\end{align}\end{subequations}\end{widetext}
At the 2PN order the required moments are $U_{ijkl}$ and $V_{ijk}$. The moment
$U_{ijkl}$ is given by
\begin{align} U_{ijkl}&=U_{ijkl}^{\rm inst}+U_{ijkl}^{\rm hered},
\label{ellWF:U4decom} \end{align}
where $U_{ijkl}^{\rm inst}$ and $U_{ijkl}^{\rm hered}$ are related to the source
multipole moments by
\begin{widetext}\begin{subequations}\label{ellWF:U4}\begin{align} U_{ijkl}^{\rm
inst}(U) &= I^{(4)}_{ijkl} (U) + {G \over c^3} \left\{ -{21
\over5}I^{(5)}_{\langle ij}I_{kl\rangle }
- {63 \over5}I^{(4)}_{\langle ij}I^{(1)}_{kl\rangle }- {102
\over5}I^{(3)}_{\langle ij}I^{(2)}_{kl\rangle }\right\}
+\,\mathcal{O}\left(\frac{1}{c^5}\right)\label{ellWF:U4inst},\\
U_{ijkl}^{\rm hered} (U) &={G \over c^3} \left\{ 2 M \int_{-\infty}^{U} d \tau
\left[ \ln \left({U-\tau \over 2\tau_0}\right)+{59 \over30} \right]
I^{(6)}_{ijkl}(\tau) +{2 \over5}\int_{-\infty}^{U} d\tau I^{(3)}_{\langle
ij}(\tau)I^{(3)}_{kl\rangle }(\tau) \right\}
+\,\mathcal{O}\left(\frac{1}{c^5}\right).\label{ellWF:U4hered}
\end{align}\end{subequations}\end{widetext}
The moment $V_{ijk}$ is given by
\begin{align} V_{ijk}&=V_{ijk}^{\rm
inst}+V_{ijk}^{\rm hered}, \label{ellWF:V3decom} \end{align}
where
$V_{ijk}^{\rm inst}$ and $V_{ijk}^{\rm hered}$ are given in terms of the source multipole
moments as
\begin{widetext}\begin{subequations}\label{ellWF:V3}\begin{align}
V_{ijk}^{\rm inst} (U) &= J^{(3)}_{ijk} (U) + {G \over c^3} \left\{ {1
\over10}\varepsilon_{ab\langle i}I^{(5)}_{ja}I_{k\rangle b}- {1
\over2}\varepsilon_{ab\langle i}I^{(4)}_{ja}I^{(1)}_{k\rangle b} - 2 J_{\langle
i}I^{(4)}_{jk\rangle } \right\}
+\,\mathcal{O}\left(\frac{1}{c^5}\right),\label{ellWF:V3inst}\\ V_{ijk}^{\rm
hered} (U) &= {2 G M \over c^3}\int_{-\infty}^{U} d \tau \left[ \ln
\left({U-\tau \over 2\tau_0}\right)+{5 \over3} \right] J^{(5)}_{ijk} (\tau)
+\,\mathcal{O}\left(\frac{1}{c^5}\right).\label{ellWF:V3hered}
\end{align}\end{subequations}\end{widetext}
The moments required at the 1.5PN order are $U_{ijklm}$ and $V_{ijkl}$. The
mass-type moment $U_{ijklm}$ is given as
\begin{align} U_{ijklm}&=U_{ijklm}^{\rm inst}+U_{ijklm}^{\rm hered},
\label{ellWF:U5decom} \end{align} where in terms of the source multipole moments, $U_{ijklm}^{\rm inst}$ and
$U_{ijklm}^{\rm hered}$ read
\begin{widetext}\begin{subequations}\label{ellWF:U5}\begin{align}
U_{ijklm}^{\rm inst}(U) &= I^{(5)}_{ijklm}(U) + {G \over c^3}\left\{ -{710
\over21}I^{(3)}_{\langle \ij}I^{(3)}_{klm\rangle} -{265 \over7}I^{(2)}_{\langle
ijk}I^{(4)}_{lm\rangle} -{120 \over7}I^{(2)}_{\langle ij}I^{(4)}_{klm\rangle}
-{155 \over7}I^{(1)}_{\langle ijk}I^{(5)}_{lm\rangle} \right.\nonumber\\&\left.-{41
\over7}I^{(1)}_{\langle ij}I^{(5)}_{klm\rangle} -{34 \over7}I_{\langle
ijk}I^{(6)}_{lm\rangle} -{15 \over7}I_{\langle ij}I^{(6)}_{klm\rangle}\right\}
+\mathcal{O}\left(\frac{1}{c^4}\right)\label{ellWF:U5inst},\\
U_{ijklm}^{\rm hered}(U) &={G \over c^3}\left\{2 M \int_{-\infty}^{U} d \tau
\left[ \ln \left({U-\tau \over2\tau_0}\right)+ {232 \over105} \right]
I^{(7)}_{ijklm} (\tau)+{20 \over21}\int_{-\infty}^{U} d \tau I^{(3)}_{\langle
ij} (\tau) I^{(4)}_{klm\rangle} (\tau) \right\}\nonumber\\&
+\mathcal{O}\left(\frac{1}{c^4}\right)\label{ellWF:U5inst},
\end{align}\end{subequations}\end{widetext}
The current-type moment $V_{ijkl}$ is given by \begin{align}
V_{ijkl}&=V_{ijkl}^{\rm inst}+V_{ijkl}^{\rm hered}, \label{ellWF:V4decom}
\end{align}
where $V_{ijkl}^{\rm inst}$ and $V_{ijkl}^{\rm hered}$ in terms of the source
multipole moments read
\begin{widetext}\begin{subequations}\label{ellWF:V4}\begin{align} V_{ijkl}^{\rm
inst}(U) &=J^{(4)}_{ijkl}(U) + {G \over c^3}\left\{-{35 \over3}S^{(2)}_{\langle
ij}I^{(3)}_{kl\rangle} -{25 \over3}I^{(2)}_{\langle ij}J^{(3)}_{kl\rangle} -{65
\over6}J^{(1)}_{\langle ij}I^{(4)}_{kl\rangle} -{25 \over6}I^{(1)}_{\langle
ij}J^{(4)}_{kl\rangle} -{19 \over6}J_{\langle ij}I^{(5)}_{kl\rangle}
\right.\nonumber\\ & -{11 \over6}I_{\langle ij}J^{(5)}_{kl\rangle} -{11
\over12}J_{\langle i}I^{(5)}_{jkl\rangle} +{1 \over6}\varepsilon_{ab\langle
i}\left[ -5I^{(3)}_{ja}I^{(3)}_{kl\rangle b} -{11
\over2}I^{(4)}_{ja}I^{(2)}_{kl\rangle b} -{5
\over2}I^{(2)}_{ja}I^{(4)}_{kl\rangle b} -{1
\over2}I^{(5)}_{ja}I^{(1)}_{kl\rangle b} \right.\nonumber\\&\left.\left. +{37
\over10}I^{(1)}_{ja}I^{(5)}_{kl\rangle b} +{3 \over10}I^{(6)}_{ja}I_{kl\rangle
b} +{1 \over2}I_{ja}I^{(6)}_{kl\rangle b}\right] \right\}
+\mathcal{O}\left(\frac{1}{c^4}\right),\label{ellWF:V4inst}\\
V_{ijkl}^{\rm hered}(U) &= {2 G M \over c^3} \int_{-\infty}^{U} d \tau \left[
\ln \left({U-\tau \over2\tau_0}\right)+{119 \over60} \right]J^{(6)}_{ijkl}
(\tau)+\mathcal{O}\left(\frac{1}{c^4}\right)\,.\label{ellWF:V4hered}
\end{align}\end{subequations}\end{widetext}
Other mass-type moments $U_{L}$ contributing to 3PN waveform are given as
\be U_{L}=U_{L}^{\rm inst}+U_{L}^{\rm hered}, \label{ellWF:ULdecom} \ee where
$U_{L}^{\rm inst}$ and $U_{L}^{\rm hered}$ are related to the source multipole moments
as \begin{subequations}\label{ellWF:UL} \begin{align} U_{L}^{\rm inst} (U) &=
I^{(\ell)}_L(U) + \mathcal{O}\left(\frac{1}{c^3}\right)\,,\\ U_{L}^{\rm
hered}(U) &= \mathcal{O}\left(\frac{1}{c^3}\right)\,.
\end{align}\end{subequations}
Other current-type moments $V_{L}$ contributing to 3PN waveform are given as
\be V_{L}=V_{L}^{\rm inst}+V_{L}^{\rm hered}, \label{ellWF:VLdecom} \ee
where finally $V_{L}^{\rm inst}$ and $V_{L}^{\rm hered}$ in terms of source multipole
moments read
\begin{subequations}\label{ellWF:VL} \begin{align} V_{L}^{\rm inst} (U) &=
J^{(\ell)}_L(U) + \mathcal{O}\left(\frac{1}{c^3}\right)\,,\\ V_{L}^{\rm
hered}(U) &= \mathcal{O}\left(\frac{1}{c^3}\right)\,.
\end{align}\end{subequations}
\subsection{Source multipole moments in general dynamical variables}
\label{ellWF:sourcemoms}
What we need next are expressions for various source multipole moments with the
PN accuracy sufficient for the present computation. Expressions for various
multipole moments presented here are generalizations of related circular orbit
expressions presented in \cite{ABIQ04, BFIS08} to the case of general orbits
and have been computed using the methods presented in \cite{BIJ02, BI04mult}. We
skip all the details of the computation and list the final expressions for the
source multipole moments related to a source composed of two nonspinning
compact objects moving in general orbits.\\
The only moment required here with 3PN accuracy is the mass quadrupole,
$I_{ij}$, which for CCBs in general orbit was computed in Ref.~\cite{BI04mult} and
listed in Ref.~\cite{ABIQ07} in standard harmonic (SH)
coordinates.\footnote{Note that Ref.~\cite{ABIQ07} lists explicit expressions
for all the source multipole moments for binaries in general orbits needed for
computing 3PN energy flux.} As was argued in Ref.~\cite{ABIQ07}, though the use
of SH coordinate is useful in performing algebraic checks on PN computations,
quantities when expressed in these coordinates involve some {\it
gauge-dependent} logarithmic terms and are not suitable for numerical
calculations. It was suggested in Ref.~\cite{ABIQ07} that such logarithms can
be transformed away by using some coordinate transformations. They showed how
the use of a modified harmonic (MH) coordinate system (or alternatively an ADM
coordinate system) removes these logarithms. We skip the details related to
those transformations and directly write the expression for the mass quadrupole
moment in MH coordinates. In MH coordinates, to 3PN accuracy, $I_{ij}$ reads
\begin{widetext} \begin{align} \label{ellWF:I2} I_{ij}&=\,\nu\,m\,\left\{
\left[A_1-\frac{24}{7}\,\frac{\,\nu}{c^5}\,\frac{G^2\,m^2}{r^2}\,\dot{r}\right]\,
x_{\langle} \def\ra{\rangle i}x_{j \ra}+\left[A_2\,\frac{r\,\dot{r}}{c^2}+\frac{48}{7}
\,\frac{\,\nu}{c^5}\,\frac{G^2\,m^2}{r}\right]\,x_{\langle} \def\ra{\rangle i}v_{j \ra}
+A_3\,\frac{r^2}{c^2}\,v_{\langle} \def\ra{\rangle i}v_{j\ra} \right\}+
\mathcal{O}\left(\frac{1}{c^7}\right), \end{align} \end{widetext}
where, \bse \label{ellWF:A1-A3} \begin{widetext} \begin{align} A_1&=1 +{1\over
c^2}\left[v^2 \left(\frac{29}{42}-\frac{29 \,\nu }{14}\right) +{G\,m\over r}
\left(-\frac{5}{7}+\frac{8 \,\nu }{7}\right)\right] +{1\over
c^4}\left[{G\,m\over r} v^2 \left(\frac{2021}{756}-\frac{5947 \,\nu }{756}
-\frac{4883 \,\nu ^2}{756}\right)\right.\nonumber\\&\left.+{G\,m\over r} \dot{r}^2
\left(-\frac{131}{756} +\frac{907 \,\nu}{756}-\frac{1273 \,\nu ^2}{756}\right)
+{G^2\,m^2\over r^2} \left(-\frac{355}{252} -\frac{953 \,\nu }{126}+\frac{337
\,\nu ^2}{252}\right)+v^4\left(\frac{253}{504} -\frac{1835 \,\nu
}{504}\right.\right.\nonumber\\&\left.\left.+\frac{3545 \,\nu ^2}{504}\right)\right]
+{1\over c^6}\left[v^6 \left(\frac{4561}{11088}-\frac{7993 \,\nu }{1584}
+\frac{117067 \,\nu ^2}{5544}-\frac{328663 \,\nu ^3}{11088}\right)
+{G^2\,m^2\over r^2} \dot{r}^2\left(-\frac{8539}{20790}\right.\right.\nonumber\\&\left.\left.+\frac{52153 \,\nu
}{4158} -\frac{4652 \,\nu ^2}{231}-\frac{54121 \,\nu
^3}{5544}\right)+{G\,m\over r} \dot{r}^4 \left(\frac{2}{99}-\frac{1745 \,\nu
}{2772}+\frac{16319 \,\nu ^2}{5544} -\frac{311 \,\nu ^3}{99}\right)+{G\,m\over
r} v^4\,\times\right.\nonumber\\&\left.\times \left(\frac{307}{77} -\frac{94475
\,\nu}{4158}+\frac{218411 \,\nu ^2}{8316}+\frac{299857 \,\nu ^3}{8316}\right)
+v^2 \left({G^2\,m^2\over r^2} \left(\frac{187183}{83160}-\frac{605419
\,\nu}{16632} +\frac{434909 \,\nu ^2}{16632}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{37369 \,\nu
^3}{2772}\right) +{G\,m\over r} \dot{r}^2 \left(-\frac{757}{5544}+\frac{5545
\,\nu }{8316} -\frac{98311\,\nu ^2}{16632}+\frac{153407 \,\nu
^3}{8316}\right)\right) +{G^3\,m^3\over r^3}
\left(\frac{6285233}{207900}\right.\right.\nonumber\\&\left.\left.+\frac{15502 \,\nu }{385} -\frac{3632 \,\nu
^2}{693}+\frac{13289\,\nu ^3}{8316}-\frac{428}{105}
\log\left[\frac{r}{r_0}\right]\right)\right],\\
A_2&=-\frac{4}{7}+\frac{12 \,\nu }{7} +{1\over c^2}\left[v^2
\left(-\frac{26}{63}+\frac{202 \,\nu }{63} -\frac{418 \,\nu
^2}{63}\right)+{G\,m\over r} \left(-\frac{155}{54} +\frac{4057 \,\nu }{378}
+\frac{209 \,\nu ^2}{54}\right)\right] \nonumber\\&+{1\over
c^4}\left[{G\,m\over r} v^2 \left(-\frac{2839}{693} +\frac{237893 \,\nu
}{8316}-\frac{188063 \,\nu ^2}{4158} -\frac{58565 \,\nu ^3}{2079}\right)
+{G\,m\over r} \dot{r}^2 \left(\frac{305}{2772}+\frac{3233 \,\nu
}{2772}\right.\right.\nonumber\\&\left.\left.-\frac{8611 \,\nu ^2}{2772} -\frac{895 \,\nu
^3}{77}\right)+{G^2\,m^2\over r^2} \left(-\frac{12587}{20790} +\frac{406333
\,\nu}{8316}-\frac{2713 \,\nu ^2}{198}+\frac{4441 \,\nu ^3}{1386}\right) +v^4
\left(-\frac{457}{1386} \right.\right.\nonumber\\&\left.\left.+\frac{6103 \,\nu }{1386}-\frac{13693 \,\nu
^2}{693} +\frac{40687 \,\nu^3}{1386}\right)\right],\\
A_3&=\frac{11}{21}-\frac{11 \,\nu }{7} +{1\over c^2}\left[{G\,m\over r}
\left(\frac{106}{27}-\frac{335 \,\nu }{189} -\frac{985 \,\nu
^2}{189}\right)+\dot{r}^2 \left(\frac{5}{63}-\frac{25 \,\nu }{63} +\frac{25
\,\nu^2}{63}\right)+v^2 \left(\frac{41}{126}\right.\right.\nonumber\\&\left.\left.-\frac{337 \,\nu }{126}
+\frac{733 \,\nu ^2}{126}\right)\right]+{1\over c^4}\left[v^4
\left(\frac{1369}{5544} -\frac{19351 \,\nu }{5544}+\frac{45421 \,\nu
^2}{2772}-\frac{139999 \,\nu ^3}{5544}\right) +{G\,m\over r}
\dot{r}^2\left(\frac{79}{77} \right.\right.\nonumber\\&\left.\left.-\frac{5807 \,\nu }{1386} +\frac{515 \,\nu
^2}{1386}+\frac{8245 \,\nu ^3}{693}\right) +v^2 \left(\dot{r}^2
\left(\frac{115}{1386}-\frac{1135\,\nu }{1386} +\frac{1795 \,\nu
^2}{693}-\frac{3445 \,\nu ^3}{1386}\right) +{G\,m\over r} \left(\frac{587}{154}
\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{67933 \,\nu }{4158}+\frac{25660 \,\nu^2}{2079} +\frac{129781
\,\nu ^3}{4158}\right)\right)+{G^2\,m^2\over r^2} \left(-\frac{40716}{1925}
-\frac{10762 \,\nu }{2079}+\frac{62576 \,\nu ^2}{2079}-\frac{24314
\,\nu^3}{2079} \right.\right.\nonumber\\&\left.\left.+\frac{428}{105}
\log\left[\frac{r}{r_0}\right]\right)\right]. \end{align} \end{widetext} \end{subequations}
In the above, $x_i$ and $v_i$ denote the binary's relative separation and relative
velocity, respectively, whereas $\dot{r}$ denotes the radial velocity. As we
see, the above expression still has a dependence on some logarithms
$(\log[r_0])$, where the quantity $r_0$ is related to the arbitrary constant
$\tau_0$ appearing in tail integrals by $\tau_0=r_0/c$. It has been argued and
shown that it disappears from all the physical quantities like the radiation field
at infinity and the far-zone energy flux~\cite{B98tail, BFIS08, K08, ABIQ07}.\\
The expression for mass octupole, $I_{ijk}$, at 2.5PN order reads
\begin{widetext}\begin{align} \label{ellWF:I3}
I_{ijk}&=-\,\nu\,m\,\Delta\left\{\left[B_1-{56 \over9}{\,\nu \over
c^5}{G^2\,m^2 \over r^2} \,\dot{r}\right]\,x_{\langle} \def\ra{\rangle
ijk\ra}+\left[B_2\,{r\,\dot{r} \over c^2}+{\,\nu\,r \over c^5} \left({232
\over15}{G^2\,m^2 \over r^2}-{12 \over5}{G\,m \over r}\,v^2\right)\right]
\,x_{\langle} \def\ra{\rangle ij}v_{k\ra}\right.\nonumber\\&\left.+B_3\,{r^2 \over c^2}\,x_{\langle} \def\ra{\rangle i}v_{jk\ra}
+B_4\,{r^3\,\dot{r} \over c^4}\,v_{\langle} \def\ra{\rangle ijk\ra}\right\}+
\mathcal{O}\left(\frac{1}{c^6}\right), \end{align}\end{widetext}
where,
\bse \begin{widetext} \label{ellWF:B1-B4} \begin{align}
B_1&=1+{1 \over c^2} \left[v^2 \left(\frac{5}{6}-\frac{19 \,\nu
}{6}\right)+{G\,m \over r}\left(-\frac{5}{6} +\frac{13 \,\nu
}{6}\right)\right]+{1 \over c^4}\left[{G\,m \over r} v^2 \left(
\frac{3853}{1320}-\frac{14257 \,\nu }{1320}-\frac{17371 \,\nu ^2}{1320}\right)
\right.\nonumber\\&\left.+{G^2\,m^2 \over r^2} \left(-\frac{47}{33}-\frac{1591 \,\nu}{132} +\frac{235
\,\nu ^2}{66}\right)+v^4 \left(\frac{257}{440} -\frac{7319 \,\nu
}{1320}+\frac{5501 \,\nu ^2}{440}\right)+{G\,m \over r}\,\dot{r}^2
\left(-\frac{247}{1320}+\frac{531 \,\nu }{440}\right.\right.\nonumber\\&\left.\left.-\frac{1347 \,\nu
^2}{440}\right) \right],\\
B_2&=-\left(1-2 \,\nu\right)+{1 \over c^2}\left[v^2
\left(-\frac{13}{22}+\frac{107 \,\nu }{22} -\frac{102 \,\nu
^2}{11}\right)+{G\,m \over r} \left(-\frac{2461}{660} +\frac{8689 \,\nu
}{660}+\frac{1389 \,\nu^2}{220}\right)\right],\\
B_3&=1-2 \,\nu+{1 \over c^2}\left[{G\,m \over r}
\left(\frac{1949}{330}+\frac{62 \,\nu }{165} -\frac{483 \,\nu
^2}{55}\right)+v^2 \left(\frac{61}{110}-\frac{519 \,\nu }{110}+\frac{504 \,\nu
^2}{55}\right)\right.\nonumber\\&\left.+\dot{r}^2\left(-\frac{1}{11}+\frac{4 \,\nu }{11}-\frac{3
\,\nu ^2}{11}\right) \right],\\
B_4&=\left(\frac{13}{55}-\frac{52 \,\nu }{55}+\frac{39 \,\nu ^2}{55}\right).
\end{align} \end{widetext} \end{subequations}
The remaining mass-type source multipole moments with PN accuracy required in
the present work are \bse \label{ellWF:I4-I8} \begin{widetext} \begin{align}
I_{ijkl}&=\,\nu\,m\left\{
x_{\langle} \def\ra{\rangle ijkl \ra}\left[1-3\,\nu+{1 \over c^2}\left[{G\,m \over r}
\left(-\frac{10}{11} +\frac{61 \,\nu }{11}-\frac{105 \,\nu ^2}{11}\right)+v^2
\left(\frac{103}{110} -\frac{147 \,\nu }{22}+\frac{279 \,\nu
^2}{22}\right)\right] \right.\right.\nonumber\\&\left.\left.+{1 \over c^4}\left[v^4
\left(\frac{3649}{5720}-\frac{50191 \,\nu }{5720} +\frac{112357 \,\nu
^2}{2860}-\frac{325687 \,\nu ^3}{5720}\right)+{G^2\,m^2 \over r^2}
\left(-\frac{15549}{10010}-\frac{9457 \,\nu }{715}+\frac{7961 \,\nu ^2}{143}
\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.-\frac{5829 \,\nu ^3}{286}\right)+{G\,m \over r} v^2
\left(\frac{11049}{3575} -\frac{152489\,\nu }{7150}+\frac{15124 \,\nu
^2}{715}+\frac{46934 \,\nu ^3}{715}\right) +{G\,m \over r}\,\dot{r}^2
\left(-\frac{659}{3575}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+\frac{12619 \,\nu }{7150} -\frac{10557
\,\nu^2}{1430}+\frac{9617 \,\nu ^3}{715}\right) \right] \right]
+x_{\langle} \def\ra{\rangle ijk}v_{l \ra}{r\,\dot{r} \over c^2}\left[-\frac{72}{55}+\frac{72 \,\nu
}{11} -\frac{72 \,\nu ^2}{11}+{1 \over c^2}\left[{G\,m \over
r}\,\times\right.\right.\right.\nonumber\\&\left.\left.\left.\,\times \left(-\frac{15463}{3575}+\frac{98374 \,\nu
}{3575}-\frac{25606 \,\nu ^2}{715} -\frac{18839 \,\nu ^3}{715}\right) +v^2
\left(-\frac{476}{715}+\frac{1228 \,\nu}{143}-\frac{23512 \,\nu ^2}{715}
\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+\frac{25796 \,\nu ^3}{715}\right)\right] \right]
+x_{\langle} \def\ra{\rangle ij}v_{kl \ra}{r^2 \over c^2}\left[\frac{78}{55}-\frac{78 \,\nu }{11}
+\frac{78 \,\nu ^2}{11}+{1 \over c^2}\left[v^2 \left(\frac{553}{715}-\frac{6913
\,\nu }{715} +\frac{25994 \,\nu ^2}{715}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.-\frac{28207 \,\nu
^3}{715}\right) +{G\,m \over r} \left(\frac{27818}{3575}-\frac{72474
\,\nu}{3575} -\frac{17202 \,\nu ^2}{715}+\frac{27568 \,\nu
^3}{715}\right)+\dot{r}^2\left(-\frac{4}{13} +\frac{28 \,\nu
}{13}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.-\frac{56 \,\nu ^2}{13}+\frac{28 \,\nu ^3}{13}\right) \right]
\right]
+x_{\langle} \def\ra{\rangle i}v_{jkl \ra}{r^3\,\dot{r} \over c^4}\left[\frac{304}{715}-\frac{2128
\,\nu }{715} +\frac{4256 \,\nu ^2}{715}-\frac{2128 \,\nu ^3}{715} \right]
+v_{\langle} \def\ra{\rangle ijkl \ra}{r^4 \over
c^4}\,\times\right.\nonumber\\&\left.\,\,\times\left[\frac{71}{715}-\frac{497 \,\nu }{715}+\frac{994
\,\nu ^2} {715}-\frac{497 \,\nu ^3}{715} \right] \right\}+
\mathcal{O}\left(\frac{1}{c^5}\right),\label{ellWF:I4}\\
I_{ijklm}&=-\,\nu\,m\,\Delta\left\{
x_{\langle} \def\ra{\rangle ijklm\ra}\left[1-2\,\nu+{1 \over c^2}\left[{G\,m \over r}
\left(-\frac{25}{26} +\frac{139 \,\nu }{26}-\frac{109 \,\nu ^2}{13}\right)+v^2
\left(\frac{79}{78} -\frac{511 \,\nu }{78}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+\frac{137 \,\nu
^2}{13}\right)\right]\right]
+x_{\langle} \def\ra{\rangle ijkl}v_{m\ra}{r\,\dot{r} \over c^2}\left[-\frac{20}{13}+\frac{80 \,\nu
}{13} -\frac{60 \,\nu ^2}{13}\right]
+x_{\langle} \def\ra{\rangle ijk}v_{lm\ra}{r^2 \over c^2}\left[\frac{70}{39}-\frac{280 \,\nu }{39}
+\frac{70 \,\nu ^2}{13}\right]
\right\}\nonumber\\&+\mathcal{O}\left(\frac{1}{c^4}\right),\label{ellWF:I5}\\
I_{ijklmn}&=\,\nu\,m\left\{
x_{\langle} \def\ra{\rangle ijklmn\ra}\left[1-5 \,\nu +5 \,\nu ^2+{1 \over c^2}\left[v^2
\left(\frac{15}{14} -\frac{21 \,\nu }{2}+33 \,\nu ^2-\frac{63 \,\nu
^3}{2}\right) -{G\,m \over r} \left(1-9 \,\nu \right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+27 \,\nu ^2-26 \,\nu
^3\right)\right]\right]
-x_{\langle} \def\ra{\rangle ijklm}v_{n\ra}\frac{12}{7}{r\,\dot{r} \over c^2}\left(1-7 \,\nu +14
\,\nu ^2 -7 \,\nu ^3\right)
+x_{\langle} \def\ra{\rangle ijkl}v_{mn\ra}\frac{15}{7}{r^2 \over c^2}\left(1-7 \,\nu \right.\right.\nonumber\\&\left.\left.+14 \,\nu
^2 -7 \,\nu ^3\right) \right\}+
\mathcal{O}\left(\frac{1}{c^4}\right),\label{ellWF:I6}\\
\label{ellWF:I7} I_{ijklmno}&=-\,\nu\,m\,\Delta\left(1-4 \,\nu +3 \,\nu
^2\right)x_{\langle} \def\ra{\rangle ijklmno\ra}+ \mathcal{O}\left(\frac{1}{c^2}\right),\\
\label{ellWF:I8} I_{ijklmnop}&=\,\nu\,m\,\left(1-7 \,\nu +14 \,\nu ^2-7 \,\nu
^3\right)x_{\langle} \def\ra{\rangle ijklmnop\ra}+ \mathcal{O}\left(\frac{1}{c^2}\right).
\end{align} \end{widetext} \end{subequations}
The current quadrupole moment is needed at 2.5PN order and given as
\begin{widetext} \begin{align} \label{ellWF:J2}
J_{ij}&=-\,\nu\,m\,\Delta\left\{\left[C_1 -{62 \over7}{{\dot{r}}\,\nu \over
c^5}{G^2\,m^2 \over r^2} \right]\epsilon_{ab\langle} \def\ra{\rangle i}x_{j\ra a}v_{b}\right.\nonumber\\&\left.
+\left[C_2\,{r\,\dot{r} \over c^2} +{r\,\nu \over c^5}{G\,m \over r} \left({216
\over35}{G\,m \over r} -{4 \over5}\,v^2\right)\right]\epsilon_{ab\langle} \def\ra{\rangle i}v_{j\ra
b}x_{a}\right\}+ \mathcal{O}\left(\frac{1}{c^6}\right),
\end{align}\end{widetext}
where,
\bse \begin{widetext} \label{ellWF:C1-C2} \begin{align}
C_1&=1+{1 \over c^2}\left[v^2 \left(\frac{13}{28}-\frac{17 \,\nu }{7}\right)
+{G\,m \over r} \left(\frac{27}{14}+\frac{15 \,\nu }{7}\right)\right] +{1 \over
c^4}\left[{G\,m \over r} v^2 \left(\frac{671}{252}-\frac{1297 \,\nu }{126}
-\frac{121 \,\nu ^2}{12}\right)\right.\nonumber\\&\left.+{G\,m \over r} \dot{r}^2
\left(-\frac{5}{252} -\frac{241 \,\nu}{252}-\frac{335 \,\nu ^2}{84}\right)
+{G^2\,m^2 \over r^2} \left(-\frac{43}{252}-\frac{1543 \,\nu }{126} +\frac{293
\,\nu ^2}{84}\right)+v^4 \left(\frac{29}{84}-\frac{11\,\nu }{3} \right.\right.\nonumber\\&\left.\left.+\frac{505
\,\nu ^2}{56}\right)\right],\\
C_2&=\frac{5}{28} (1-2 \,\nu )+\frac{1}{504} {1 \over c^2}\left[{G\,m \over r}
\left(824 +1348 \,\nu -1038 \,\nu ^2\right)+75 v^2 \left(1-7 \,\nu +12 \,\nu
^2\right)\right]. \end{align} \end{widetext} \end{subequations}
Other current-type source multipole moments with PN accuracies sufficient for
present calculations read \bse \label{ellWF:J3-J7} \begin{widetext}
\begin{align} \label{ellWF:J3} J_{ijk}&=\,\nu\,m\,\epsilon_{ab\langle} \def\ra{\rangle i}\left\{
x_{jk\ra a}v_{b}\left[1-3 \,\nu+{1 \over c^2}\left[{G\,m \over r}
\left(\frac{14}{9} -\frac{16 \,\nu }{9}-\frac{86 \,\nu ^2}{9}\right)+v^2
\left(\frac{41}{90}-\frac{77 \,\nu }{18} +\frac{185 \,\nu
^2}{18}\right)\right]\right.\right.\nonumber\\&\left.\left.+{1 \over c^4}\left[v^4 \left(\frac{1349}{3960}
-\frac{4159 \,\nu }{792}+\frac{52409 \,\nu ^2}{1980}-\frac{171539 \,\nu
^3}{3960}\right) +{G^2\,m^2 \over r^2} \left(-\frac{45}{44}-\frac{988
\,\nu}{99}+\frac{9925 \,\nu ^2}{198} \right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.-\frac{8099 \,\nu
^3}{396}\right)+{G\,m \over r} \dot{r}^2 \left(-\frac{23}{396} -\frac{637 \,\nu
}{990}-\frac{1861 \,\nu^2}{990}+\frac{32221 \,\nu ^3}{1980}\right) +{G\,m \over
r} v^2 \left(\frac{1597}{660}-\frac{19381 \,\nu }{990} \right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+\frac{6307
\,\nu ^2}{198}+\frac{21127 \,\nu^3}{396}\right)\right]\right]
+x_{j}v_{k\ra b}x_{a}{r\,\dot{r} \over c^2}\left[\frac{2}{9}-\frac{10 \,\nu
}{9} +\frac{10 \,\nu ^2}{9}+{1 \over c^2}\left[v^2
\left(\frac{73}{495}-\frac{841 \,\nu }{495} +\frac{3002 \,\nu
^2}{495}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.-\frac{3151 \,\nu ^3}{495}\right)+{G\,m \over r}
\left(\frac{133}{66} -\frac{81 \,\nu}{55}-\frac{3914 \,\nu ^2}{165}+\frac{3089
\,\nu ^3}{330}\right)\right] \right]
+v_{jk\ra b}x_{a}{r^2 \over c^2}\left[\frac{7}{45} \left(1-5 \,\nu +5 \,\nu
^2\right) \right.\right.\nonumber\\&\left.\left.+{1 \over c^2}\left[v^2 \left(\frac{119}{990}-\frac{259 \,\nu
}{198}+\frac{2219 \,\nu ^2}{495} -\frac{4529 \,\nu ^3}{990}\right)+\dot{r}^2
\left(\frac{14}{165}-\frac{98 \,\nu}{165} +\frac{196 \,\nu ^2}{165}-\frac{98
\,\nu ^3}{165}\right) \right.\right.\right.\nonumber\\&\left.\left.\left.+{G\,m \over r}\left(\frac{751}{495}-\frac{1792
\,\nu }{495}-\frac{227 \,\nu ^2}{99} +\frac{427 \,\nu^3}{99}\right)\right]
\right] \right\}+ \mathcal{O}\left(\frac{1}{c^5}\right),\\
J_{ijkl}&=-\,\nu\,m\,\Delta\,\epsilon_{ab\langle} \def\ra{\rangle i}\left\{
x_{jkl\ra a}v_{b}\left[1-2 \,\nu+{1 \over c^2}\left[{G\,m \over r}
\left(\frac{15}{11} +\frac{35 \,\nu }{44}-\frac{185 \,\nu ^2}{22}\right)+v^2
\left(\frac{5}{11} -\frac{95 \,\nu }{22}\right.\right.\right.\right.\nonumber\\&\left.\left.\left.\left.+\frac{195 \,\nu
^2}{22}\right)\right]\right]
+\frac{5}{22} \,x_{jk}v_{l\ra b}x_{a}{r\,\dot{r} \over c^2} \left(1-4 \,\nu +3
\,\nu ^2\right)
+\frac{4}{11}\,x_{j}v_{kl\ra b}x_{a}{r^2 \over c^2} \left(1-4 \,\nu +3 \,\nu
^2\right) \right\}\nonumber\\&} \def\p{\partial+ \mathcal{O}\left(\frac{1}{c^4}\right),\\
J_{ijklm}&=\,\nu\,m\,\epsilon_{ab\langle} \def\ra{\rangle i}\left\{
x_{jklm\ra a}v_{b}\left[1-5 \,\nu +5 \,\nu ^2+{1 \over c^2}\left[v^2
\left(\frac{83}{182} -\frac{161 \,\nu }{26}+\frac{317 \,\nu ^2}{13}-\frac{707
\,\nu ^3}{26}\right)\right.\right.\right.\nonumber\\&\left.\left.\left.+{G\,m \over r} \left(\frac{81}{65}-\frac{138 \,\nu
}{65}-\frac{210\,\nu ^2}{13} +\frac{339 \,\nu ^3}{13}\right)\right]\right]
+\frac{20}{91}\,x_{jkl}v_{m\ra b}x_{a}{r\,\dot{r} \over c^2} \left(1-7 \,\nu
+14 \,\nu ^2-7 \,\nu ^3\right)
\right.\nonumber\\&\left.+\frac{54}{91}\,x_{jk}v_{lm\ra b}x_{a}{r^2 \over c^2} \left(1-7 \,\nu +14
\,\nu ^2-7 \,\nu ^3\right)\right\}+ \mathcal{O}\left(\frac{1}{c^4}\right),\\
J_{ijklmn}&=-\,\nu\,m\Delta\,\epsilon_{ab\langle} \def\ra{\rangle i}x_{jklmn\ra}x_{a}v_{b}\left(1-4
\,\nu +3 \,\nu ^2\right)+ \mathcal{O}\left(\frac{1}{c^2}\right),\\
J_{ijklmno}&=\,\nu\,m\,\epsilon_{ab\langle} \def\ra{\rangle i}x_{jklmno\ra}x_{a}v_{b} \left(1-7
\,\nu +14 \,\nu ^2-7 \,\nu ^3\right)+ \mathcal{O}\left(\frac{1}{c^2}\right).
\end{align} \end{widetext} \end{subequations}
The required gauge moments, the monopolar moment $W$, and two dipolar moments
$W_{i}$ and $Y_{i}$ are finally given by
\bse \label{ellWF:WY} \begin{align} W &=
\frac{1}{3}\,\nu\,m\,r\,\dot{r}+\mathcal{O}\left(\frac{1}{c^2}\right)
\label{ellWF:W0}\,,\\ W_i &= \frac{1}{10}\,\nu\,m\,\Delta\,r^2\left[v^i
-3\,\frac{\dot{r}}{r}\,x^i\right]+ \mathcal{O}\left(\frac{1}{c^2}\right)
\label{ellWF:W1}\,,\\ Y_i &= \frac{1}{5}\,\nu\,m\,\Delta\left[{1
\over2}\frac{G\,m}{r}\,x^i+{1
\over2}\,v^2\,x^i-\frac{3}{2}\,r\,\dot{r}\,v^i\right]+
\mathcal{O}\left(\frac{1}{c^2}\right)\,. \label{ellWF:Y1} \end{align} \end{subequations}
\subsection{The post-Newtonian compact binary dynamics}
\label{ellWF:ellWFbinarydynamics}
Since relations connecting the radiative multipole moment to the source
multipole moment involve time derivatives of the source multipole moments,
computations of various modes will require a knowledge of the equations of
motion (EOM) with the PN accuracy with which one wants to compute various
modes. Before we write expressions for the EOM, with the PN accuracy required
for the present work, let us recall the definitions of various dynamical
variables as well as some other related results which will be used in
calculations performed here. The binary's relative separation, $x^i$ is given
by
\begin{align} \label{ellWF:x} x^i&=y_1^i-y_2^i=r\,n^i \end{align}
with $r=|{\bf x}|$, where $\bf x$ is the relative separation vector. Here,
$y_1^i$ and $y_2^i$ are position vectors of the individual components of the
binary and ${n^i}$ are components of the unit vector, $\bf {\hat{n}}$, along
the relative separation vector.\\
For the relative velocity and relative acceleration we have
\begin{align} \label{ellWF:v-a} v^i&={{\rm d}x^i \over {\rm
d}t}\,\,\,\,\,\,{\rm and}\,\,\,\,\, \dot{r}={\bf \hat{n}}\cdot {\bf v},\\
a_i&={{\rm d}v^i \over {\rm d}t}={{\rm d^2}x^i \over {\rm d^2}t}. \end{align}
In addition we also would need expressions for $\dot{v}$ and $\ddot{r}$, which
can be given as
\begin{align} \label{ellWF:vdot-rdot} \dot{v}&={{\bf a}\cdot{\bf v} \over v},\\
\ddot{r}&={1 \over r}\left[\left(v^2-\dot{r}^2\right)+{\bf a}\cdot{\bf
x}\right]. \end{align} with $v=|{\bf v}|$.\\
While computing the time derivatives of the source multipole moments, whenever
quantities like $a^i$ or $\dot{v}$ or $\ddot{r}$ appear, they are consistently
replaced by their expressions in terms of variables related to the position and
velocities $(r, \dot{r}, v)$. Computations of various modes at the 3PN order
would require the knowledge of the 3PN EOM governing the compact binary
dynamics. EOM associated with SH coordinates at 3PN order for a system of two
compact objects moving in general orbits is available in the literature
\cite{BFeom, BDE04, DJSequiv, ABF01, DJSdim} and are given in terms of the
variables related to the position and the velocity of individual constituents
of the binary. Since we are using expressions for the source multipole moments
in the center-of-mass frame of the system, we need the EOM reduced to the CM
frame. The 3PN accurate expression for relative acceleration, reduced to the
CM frame, in SH coordinates, were obtained in \cite{BI03CM}. However, as
discussed in the previous section (about using MH or ADM coordinates instead of
SH coordinates), we wish to use EOM reduced to CM frame associated with MH
coordinate which is given in \cite{MW03} and takes the following form,
\begin{widetext} \begin{align} \label{ellWF:arel} a^i&=-{G\,m \over
r^2}\left\{\left[P_1-{\nu \over c^5}\left(\frac{136}{15} {G^2\,m^2 \over r^2}
\dot{r}+\frac{24}{5} {G\,m \over r} \dot{r} v^2\right)\right]\,n^i
+\left[P_2\,{\dot{r} \over c^2}+{\nu \over c^5}\left(\frac{24}{5}{G^2\,m^2
\over r^2} +\frac{8}{5} {G\,m \over r} v^2\right)\right]\,v^i\right\},
\end{align} \end{widetext} where, \begin{widetext} \bse \label{ellWF:P1-P2}
\begin{align}
P_1&=1+{1 \over c^2}\left[{G\,m \over r} (-4-2 \nu )-\frac{3 \dot{r}^2 \nu }{2}
+v^2 (1+3 \nu )\right]+{1 \over c^4}\left[{G^2\,m^2 \over r^2} \left(9
+\frac{87 \nu }{4}\right) +\dot{r}^4 \left(\frac{15 \nu }{8}-\frac{45 \nu
^2}{8}\right) \right.\nonumber\\&\left.+v^4 \left(3 \nu -4 \nu ^2\right)+{G\,m \over r}\dot{r}^2
\left(-2-25 \nu -2 \nu ^2\right) +v^2 \left({G\,m \over r} \left(-\frac{13 \nu
}{2}+2 \nu ^2\right) +\dot{r}^2 \left(-\frac{9 \nu }{2}+6
\nu^2\right)\right)\right] \nonumber\\&+{1 \over c^6}\left[{G^3\,m^3 \over r^3}
\left(-16-\frac{1399 \nu }{12} +\frac{41 \pi ^2 \nu }{16}-\frac{71 \nu
^2}{2}\right) +{G\,m \over r} \dot{r}^4 \left(79 \nu -\frac{69 \nu ^2}{2}
-30\nu ^3\right)+\dot{r}^6 \left(-\frac{35 \nu }{16}\right.\right.\nonumber\\&\left.\left.+\frac{175 \nu ^2}{16}
-\frac{175 \nu ^3}{16}\right)+{G^2\,m^2 \over r^2} \dot{r}^2
\left(1+\frac{22717 \nu}{168} +\frac{615 \pi ^2 \nu }{64}+\frac{11 \nu ^2}{8}-7
\nu ^3\right)+v^6 \left(\frac{11 \nu }{4} -\frac{49 \nu ^2}{4}\right.\right.\nonumber\\&\left.\left.+13 \nu
^3\right)+v^4 \left(\dot{r}^2\left(-\frac{15 \nu }{2} +\frac{237 \nu
^2}{8}-\frac{45 \nu ^3}{2}\right)+{G\,m \over r} \left(\frac{75 \nu }{4} +8 \nu
^2-10 \nu ^3\right)\right)+v^2\left({G^2\,m^2 \over r^2}
\times\right.\right.\nonumber\\&\left.\left.\times\left( -\frac{20827 \nu }{840} -\frac{123 \pi ^2 \nu }{64}+\nu
^3\right)+{G\,m \over r} \dot{r}^2 \left(-121 \nu +16 \nu ^2+20
\nu^3\right)+\dot{r}^4 \left(\frac{15 \nu }{2}-\frac{135 \nu ^2}{4}
\right.\right.\right.\nonumber\\&\left.\left.\left.+\frac{255 \nu ^3}{8}\right)\right)\right],\\
P_2&=-4+2 \nu+{1 \over c^2}\left[v^2 \left(-\frac{15 \nu }{2}-2 \nu
^2\right)+\dot{r}^2 \left(\frac{9 \nu }{2}+3 \nu ^2\right)+{G\,m \over r}
\left(2+\frac{41 \nu }{2} +4 \nu ^2\right)\right]\nonumber\\&+{1 \over
c^4}\left[\dot{r}^4 \left(-\frac{45 \nu }{8}+15 \nu ^2 +\frac{15 \nu
^3}{4}\right)+v^4 \left(-\frac{65 \nu }{8}+19 \nu ^2+6 \nu ^3\right) +{G^2\,m^2
\over r^2}\left(-4-\frac{5849 \nu }{840}-\frac{123 \pi ^2 \nu }{32}\right.\right.\nonumber\\&\left.\left.+25 \nu
^2 +8 \nu ^3\right)+{G\,m \over r} \dot{r}^2 \left(\frac{329 \nu }{6}+\frac{59
\nu ^2}{2} +18\nu ^3\right)+v^2 \left(\dot{r}^2 \left(12 \nu -\frac{111 \nu
^2}{4}-12 \nu ^3\right) +{G\,m \over r} \left(-15 \nu \right.\right.\right.\nonumber\\&\left.\left.\left.-27 \nu ^2-10 \nu
^3\right)\right)\right]. \end{align} \end{subequations} \end{widetext}
We now have all the inputs which are needed to compute the instantaneous
expressions for various spherical harmonic modes ($h^{\ell m}$) associated with
3PN gravitational waveforms of GW signals from CCBs moving in general orbits.
With this motivation, we shall proceed towards the next section where we shall
present our results.
\section{Instantaneous terms in the 3PN gravitational waveform for CCBs in
general orbits}
\label{ellWF:ellWF3pnsphhar}
Combing Eq.~\eqref{ellWF:UV} and Eq.~\eqref{ellWF:hlm-modesep} we can write the
instantaneous part of various modes as
\begin{align}\label{ellWF:hlm-modesep-inst} h^{\ell m}_{\rm inst}
&=-\frac{G}{{\sqrt 2}\,R\,c^{\ell+2}}\frac{4}{\ell!}
\,\sqrt{\frac{(\ell+1)(\ell+2)}{2\ell(\ell-1)}}\,\alpha_L^{\ell m}\,U_L^{\rm
inst} \nonumber\\&} \def\p{\partial \quad\quad\quad\quad {\rm if}\,\ell+m\,\,{\rm is\, even},\\ h^{\ell
m}_{\rm inst} &=-\frac{i\,G}{{\sqrt 2}\,R\,c^{\ell+3}}\frac{8}{\ell!}
\,\sqrt{\frac{\ell(\ell+2)}{2(\ell+1)(\ell-1)}}\,\alpha_L^{\ell m}\,V_L^{\rm
inst} \nonumber\\&} \def\p{\partial \quad\quad\quad\quad {\rm if}\,\ell+m\,\,{\rm is\, odd}. \end{align}
Relations connecting the instantaneous part of STF radiative moments ($U_L^{\rm
inst}$ and $V_L^{\rm inst}$) to the source multipole moments have been listed
in the previous section. This allows one to write the instantaneous part of
various modes ($h^{\ell m}_{\rm inst}$) in terms of the source multipole
moments. With expressions for the source multipole moments for CCBs moving in
general orbits and their relevant time derivatives, one can write expressions
for various modes in terms of dynamical variables related to the position and
velocity ($r$, $\dot{r}$, $\phi$, $v$).\footnote{Alternatively, one can also
compute various modes associated with the gravitational waveform using
polarization waveforms (see Sec. II and IX of Ref.~\cite{BFIS08} for the
details).}
Again, since $v^2=\dot{r}^2+r^2\dot{\phi}^2$, we can write various modes of
the waveform in terms of the dynamical variables, namely, the radial separation
$(r)$, radial velocity $(\dot{r})$, orbital phase $(\phi)$, and the angular
velocity $(\dot{\phi})$ . The structure of $h^{\ell m}$ reads
\be \label{ellWF:hlm} h^{\ell m}_{\rm inst}={4\,G\,m\,\nu \over c^4 R}{\sqrt
{\pi \over 5}}\,e^{-i\,m\,\phi}\,\hat{H}^{\ell m}_{\rm inst}. \ee
For the dominant mode ($\ell=2, m=2$), with 3PN accuracy, various PN pieces of
the coefficient $\hat{H}^{22}_{\rm inst}$ read
\begin{widetext} \bse \label{ellWF:H22} \begin{align} (\hat{H}^{22}_{\rm
inst})_{\rm Newt}&=\frac{G m}{r}+r^2 \dot{\phi }^2+2 i r \dot{r} \dot{\phi
}-\dot{r}^2,\\
(\hat{H}^{22}_{\rm inst})_{\rm 1PN}&={1 \over c^2}\left[\frac{G^2
m^2}{r^2}\left(-5+\frac{\,\nu }{2}\right) +\frac{G m
\dot{r}^2}{r}\left(-\frac{15}{14}-\frac{16 \,\nu
}{7}\right)+\left(-\frac{9}{14} +\frac{27 \,\nu}{14}\right)
\dot{r}^4+r\left(\frac{9 i}{7}\right.\right.\nonumber\\&\left.\left.-\frac{27 i \,\nu }{7}\right) \dot{r}^3
\dot{\phi } +G m r \left(\frac{11}{42}+\frac{26 \,\nu }{7}\right) \dot{\phi
}^2+r^4 \left(\frac{9}{14}-\frac{27 \,\nu }{14}\right) \dot{\phi }^4 +\dot{r}
\left(G m \left(\frac{25 i}{21}+\frac{45 i \,\nu }{7}\right) \dot{\phi }
\right.\right.\nonumber\\&\left.\left.+r^3\left(\frac{9 i}{7}-\frac{27 i \,\nu }{7}\right) \dot{\phi
}^3\right)\right],\\
(\hat{H}^{22}_{\rm inst})_{\rm 2PN}&={1 \over c^4}\left[\frac{G^3 m^3}{r^3}
\left(\frac{757}{63}+\frac{181 \,\nu }{36} +\frac{79 \,\nu
^2}{126}\right)+\left(-\frac{83}{168}+\frac{589 \,\nu }{168}-\frac{1111 \,\nu
^2}{168}\right) \dot{r}^6+r \left(\frac{83 i}{84} \right.\right.\nonumber\\&\left.\left.-\frac{589 i \,\nu
}{84}+\frac{1111 i \,\nu ^2}{84}\right) \dot{r}^5 \dot{\phi} +G^2 m^2
\left(-\frac{11891}{1512}-\frac{5225 \,\nu }{216}+\frac{13133 \,\nu
^2}{1512}\right) \dot{\phi }^2+G m r^3 \left(\frac{835}{252}\right.\right.\nonumber\\&\left.\left.+\frac{19
\,\nu }{252}-\frac{2995 \,\nu ^2}{252}\right) \dot{\phi }^4+r^6
\left(\frac{83}{168}-\frac{589 \,\nu }{168} +\frac{1111 \,\nu ^2}{168}\right)
\dot{\phi }^6 +\dot{r}^4 \left(\frac{G m}{r}\left(-\frac{557}{168}+\frac{83
\,\nu }{21} \right.\right.\right.\nonumber\\&\left.\left.\left.+\frac{214 \,\nu ^2}{21}\right)+r^2
\left(-\frac{83}{168}+\frac{589 \,\nu }{168} -\frac{1111 \,\nu
^2}{168}\right)\dot{\phi }^2\right)+\dot{r}^3 \left(G m \left( \frac{863
i}{126}-\frac{731 i \,\nu }{63} \right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{211 i \,\nu ^2}{9}\right)
\dot{\phi } +r^3 \left(\frac{83 i}{42}-\frac{589 i \,\nu }{42}+\frac{1111 i
\,\nu ^2}{42}\right) \dot{\phi }^3\right)+\dot{r}^2\left(\frac{G^2
m^2}{r^2}\left(\frac{619}{252} -\frac{2789 \,\nu}{252}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{467 \,\nu
^2}{126}\right) +G m r \left(\frac{11}{28} -\frac{169 \,\nu }{14}-\frac{58
\,\nu ^2}{21}\right) \dot{\phi }^2 +r^4\left(\frac{83}{168}-\frac{589 \,\nu
}{168} +\frac{1111 \,\nu ^2}{168}\right) \dot{\phi }^4\right) \right.\nonumber\\&\left.+\dot{r}
\left(\frac{G^2 m^2}{r}\left(-\frac{773 i}{189}-\frac{3767 i \,\nu
}{189}+\frac{2852 i \,\nu ^2}{189}\right)\dot{\phi }+G m r^2 \left(\frac{433
i}{84} +\frac{103 i \,\nu }{12} -\frac{1703 i \,\nu ^2}{84}\right) \dot{\phi
}^3\right.\right.\nonumber\\&\left.\left.+r^5 \left(\frac{83 i}{84}-\frac{589 i \,\nu }{84}+\frac{1111 i \,\nu
^2}{84} \right)\dot{\phi }^5\right)\right],\\
(\hat{H}^{22}_{\rm inst})_{\rm 2.5PN}&={1 \over c^5}\left[-\frac{122 G^2 m^2
\,\nu \dot{r}^3}{35 r^2} -\frac{468 i G^3 m^3 \,\nu \dot{\phi }}{35
r^2}+\frac{184 i G^2 m^2 \,\nu \dot{r}^2 \dot{\phi }}{35 r}-\frac{316}{35} i
G^2 m^2 r \,\nu \dot{\phi }^3 \right.\nonumber\\&\left.+\dot{r} \left(\frac{2 G^3 m^3 \,\nu }{105
r^3}-\frac{121}{5} G^2 m^2 \,\nu \dot{\phi }^2\right)\right],\\
(\hat{H}^{22}_{\rm inst})_{\rm 3PN}&={1 \over c^6}\left[\frac{G^4
m^4}{r^4}\left(-\frac{512714}{51975}+\left( -\frac{1375951}{13860}+\frac{41 \pi
^2}{16}\right) \,\nu +\frac{1615 \,\nu ^2}{616} +\frac{2963
\,\nu^3}{4158}-\frac{214}{105} \log \left[\frac{r}{r_0}\right]\right)
\right.\nonumber\\&\left.+\left(-\frac{507}{1232}+\frac{6101 \,\nu }{1232} -\frac{12525
\,\nu^2}{616}+\frac{34525 \,\nu ^3}{1232}\right) \dot{r}^8+r \left( \frac{507
i}{616}-\frac{6101 i \,\nu }{616}+\frac{12525 i \,\nu ^2}{308}
\right.\right.\nonumber\\&\left.\left.-\frac{34525 i \,\nu^3}{616}\right) \dot{r}^7 \dot{\phi }+\frac{G^3
m^3}{r} \left(\frac{42188851}{415800}+\left(\frac{190703}{3465} -\frac{123 \pi
^2}{64}\right) \,\nu -\frac{18415 \,\nu^2}{308}+\frac{281473 \,\nu^3}{16632}
\right.\right.\nonumber\\&\left.\left.-\frac{214}{15} \log\left[\frac{r}{r_0}\right]\right) \dot{\phi }^2+G^2
m^2 r^2 \left(\frac{328813}{55440}-\frac{374651 \,\nu }{33264}+\frac{249035
\,\nu ^2}{4158} -\frac{1340869 \,\nu ^3}{33264}\right) \dot{\phi }^4\right.\nonumber\\&\left.+G m r^5
\left(\frac{12203}{2772}-\frac{36427 \,\nu }{2772}-\frac{13667 \,\nu ^2}{1386}
+\frac{49729 \,\nu ^3}{924}\right) \dot{\phi }^6+r^8
\left(\frac{507}{1232}-\frac{6101 \,\nu }{1232}+\frac{12525 \,\nu ^2}{616}
\right.\right.\nonumber \\ & \left.\left. -\frac{34525 \,\nu ^3}{1232}\right)
\dot{\phi }^8+\dot{r}^4 \left(\frac{G^2 m^2 }{r^2}
\left(-\frac{92567}{13860}+\frac{7751 \,\nu}{396}+\frac{400943 \,\nu ^2}{11088}
+\frac{120695 \,\nu ^3}{3696}\right) \right.\right.\nonumber\\&\left.\left.+G m r
\left(-\frac{42811}{11088}+\frac{6749 \,\nu }{1386}+\frac{19321 \,\nu^2}{693}
-\frac{58855 \,\nu ^3}{1386}\right) \dot{\phi }^2\right)+\dot{r}^6
\left(\frac{G m}{r} \left(-\frac{5581}{1232}+\frac{4694 \,\nu
}{231}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{3365 \,\nu^2}{462}-\frac{1850 \,\nu ^3}{33}\right)+r^2
\left(-\frac{507}{616}+\frac{6101 \,\nu }{616} -\frac{12525 \,\nu ^2}{308}
+\frac{34525 \,\nu ^3}{616}\right)\dot{\phi}^2\right)\right.\nonumber\\&\left. +\dot{r}^5\left(G m
\left(\frac{17233 i}{1848} -\frac{31532 i \,\nu }{693}+\frac{65575 i \,\nu
^2}{2772}+\frac{85145 i \,\nu ^3}{693}\right) \dot{\phi} +r^3 \left(\frac{1521
i}{616}-\frac{18303 i \,\nu }{616}\right.\right.\right.\nonumber\\&\left.\left.\left.+\frac{37575 i \,\nu ^2}{308}
-\frac{103575 i \,\nu ^3}{616}\right) \dot{\phi }^3\right)+\dot{r}^3
\left(\frac{G^2 m^2}{r}\left(\frac{39052 i}{3465}-\frac{154114 i \,\nu }{2079}
-\frac{246065 i \,\nu ^2}{4158}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{365725 i \,\nu ^3}{4158}\right)
\dot{\phi }+G m r^2 \left(\frac{13867 i}{792}-\frac{191995 i \,\nu }{2772}
-\frac{8741 i \,\nu ^2}{5544}+\frac{52700 i \,\nu ^3}{231}\right) \dot{\phi }^3
+r^5 \left(\frac{1521 i}{616}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{18303 i \,\nu }{616}+\frac{37575 i
\,\nu ^2}{308} -\frac{103575 i \,\nu ^3}{616}\right) \dot{\phi
}^5\right)+\dot{r}^2 \left(\frac{G^3 m^3 }{r^3}\left(\frac{913799}{29700}
+\left(\frac{174679}{2310}+\frac{123 \pi ^2}{32}\right) \,\nu
\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{158215 \,\nu ^2}{2772} -\frac{12731 \,\nu
^3}{4158}-\frac{428}{105} \log\left[\frac{r}{r_0}\right]\right)+G^2 m^2
\left(\frac{20191}{18480} -\frac{3879065 \,\nu }{33264}-\frac{411899 \,\nu
^2}{8316} \right.\right.\right.\nonumber \\ & \left.\left.\left. -\frac{522547
\,\nu ^3}{33264}\right) \dot{\phi }^2+G m r^3 \left(\frac{381}{77}-\frac{101237
\,\nu }{2772} +\frac{247505 \,\nu ^2}{5544}+\frac{394771 \,\nu ^3}{5544}\right)
\dot{\phi }^4+r^6 \left(\frac{507}{616}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{6101 \,\nu
}{616}+\frac{12525 \,\nu ^2}{308} -\frac{34525 \,\nu ^3}{616}\right) \dot{\phi
}^6\right)+\dot{r} \left(\frac{G^3 m^3 }{r^2}\left(-\frac{68735
i}{378}+\left(-\frac{57788 i}{315} +\frac{123 i \pi ^2}{32}\right) \,\nu
\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{701 i \,\nu ^2}{27}+\frac{11365 i \,\nu ^3}{378} +\frac{428}{21}
i \log\left[\frac{r}{r_0}\right]\right) \dot{\phi } +G^2 m^2 r
\left(\frac{91229 i}{13860}+\frac{97861 i \,\nu }{4158}+\frac{919811 i \,\nu
^2}{8316}\right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{556601 i \,\nu ^3}{8316}\right) \dot{\phi }^3 +G m r^4
\left(\frac{6299 i}{792}-\frac{68279 i \,\nu }{5544}-\frac{147673 i
\,\nu^2}{2772} +\frac{541693 i \,\nu ^3}{5544}\right) \dot{\phi }^5+r^7
\left(\frac{507 i}{616} \right.\right.\right.\nonumber\\&\left.\left.\left.-\frac{6101 i \,\nu }{616}+\frac{12525 i \,\nu
^2}{308}-\frac{34525 i \,\nu ^3}{616}\right) \dot{\phi }^7\right)\right].
\end{align} \end{subequations} \end{widetext}
As mentioned in Sec.~\ref{ellWF:ellWFintro}, since expressions for various
modes are very large and would run over several pages, we have decided to
provide an additional file (Hlm-GenOrb.m) containing expressions for all the
modes which will be made available along with the paper. Finally,
circular-orbit limit of the instantaneous $h_{\ell m}$ can be obtained by
replacing related expressions for $\dot{\phi}(=\omega)$, $\dot{r}$ and $r$
given in Sec.~IV of \cite{BFIS08}. \\
Note that $h_{\ell m}$ can directly be used to write the polarization waveforms
($h_+$, $h_{\times}$) using the standard decomposition of $h_{+}$ and
$h_{\times}$ in terms of spherical harmonic modes of spin weight -2 given in
Ref.~\cite{BFIS08, K08},
\begin{equation} \label{ellWF:spinw} h_+ - i h_\times =
\sum^{+\infty}_{\ell=2}\sum^{\ell}_{m=-\ell} h^{\ell m}\,Y^{\ell
m}_{-2}(\Theta,\Phi)\,, \end{equation}
where $Y^{\ell m}_{-2}$'s (the spin-weighted spherical harmonics of weight
$-2$) are functions of the spherical angles $(\Theta,\Phi)$ defining the
binary's location and given as
\begin{widetext}\begin{subequations}\label{ellWF:harm}\begin{align} Y^{\ell
m}_{-2} &= \sqrt{\frac{2\ell+1}{4\pi}}\,d^{\,\ell m}_{\,2}(\Theta)\,e^{i \,m
\,\Phi},\\d^{\,\ell m}_{\,2} &= \sum_{k=k_1}^{k_2}\frac{(-)^k}{k!}
\frac{\sqrt{(\ell+m)!(\ell-m)!(\ell+2)!(\ell-2)!}}
{(k-m+2)!(\ell+m-k)!(\ell-k-2)!}\left(\cos\frac{\Theta}{2}\right)^{2\ell+m-2k-2}
\!\!\!\left(\sin\frac{\Theta}{2}\right)^{2k-m+2}\,,
\end{align}\end{subequations}\end{widetext}
with $k_1=\mathrm{max}(0,m-2)$ and $k_2=\mathrm{min}(\ell+m,\ell-2)$.\\
Further, the polarization waveform can be used to write the
transverse-traceless part of the radiation field ($h_{ij}$) by using the
following relation~\cite{BFIS08, K08}
\begin{equation} \label{ellWF:hij} h_{ij}^{\rm TT}=2
\left(h_+\,e_{ij}^{+}+h_{\times}\,e_{ij}^{\times}\right)\,, \end{equation}
where \bse\label{ellWF:epec} \begin{align} \label{ellWF:ep} e_{ij}^{+}&={1
\over2}\left(P_i P_j-Q_i Q_j\right)\,,\\ \label{ellWF:ep} e_{ij}^{\times}&={1
\over2}\left(P_i Q_j+P_j Q_i\right)\,. \end{align} \end{subequations}
Here {\bf P} and {\bf Q} are the two unit polarization vectors and they have
been chosen following the convention used in \cite{BFIS08}.
\section{Quasi-Keplerian representation}
\label{ellWF:ellWFQKR}
In the previous section, we presented general orbit expressions for the
dominant mode of the gravitational waveform from a compact binary system. In
this section we aim to specialize to the case of compact binary systems in
quasi-elliptical orbits. Expressions for various $h_{\ell m}$ describing the
radiation from binaries in quasi-elliptical orbits can simply be obtained by
using relations connecting generic dynamical variable $r$, $\dot{r}$, $\phi$
and $\dot{\phi}$ to a set of parameters associated with elliptical orbits in
general orbit expressions for various modes. Such relations can be established
by using the generalized quasi-Keplerian (QK) representation of the
conservative dynamics of the binary moving in eccentric orbits, which indeed is
available to us due to the work of Memmesheimer, Gopakumar and Sch{\"a}fer
(hereafter MGS) \cite{MGS04}.\\
The QK representation was first introduced by Damour and Deruelle \cite{DD85}
and dealt with the binary dynamics at 1PN order. The generalized QK
representation at 2PN order in ADM-type coordinates was given in
Refs.~\cite{DS88, SW93, Wex95}. MGS provides the 3PN generalized QK
representation in both the ADM and MH coordinates, which involve expressions of
the orbital elements associated with the orbit of the binary in terms of the
conserved energy and orbital angular momentum of the binary. Before we get into
the details of the parametrization, we first summarize equations describing the
radial and angular motion of the binary in terms of various orbital elements
associated with elliptical orbits (see Refs.~\cite{ABIQ07, KG06} for details).
In the parametric form, the radial separation, $r$, is given by
\be \label{ellWF:r} r=a_r\left(1-e_r\,\cos u\right), \ee
where $a_r$ is the semi-major axis of the orbit and $e_r$ is the eccentricity
of the orbit (both labeled after the radial coordinate, $r$). The quantity $u$
is called eccentric anomaly and at the 3PN order it is related to the mean
anomaly ($l$) by the relation
\be \label{ellWF:l-u} l=u-e_t\,\sin u+f_t\,\sin
V+g_{t}\left[V-u\right]+i_t\,\sin 2 V +h_t\,\sin 3 V. \ee
The orbital phase, $\phi$, at the 3PN order reads
\begin{align} \label{ellWF:Phi} \phi&=\phi_P+K\,\left[V+f_\phi\,\sin 2
V+g_\phi\,\sin 3 V +i_\phi\,\sin 4 V \right.\nonumber\\&\left. + h_\phi\,\sin 5 V\right], \end{align}
where $\phi_P$ is the initial phase at the first passage of the periastron and
$V$ is the true anomaly that takes the form
\be \label{ellWF:Vuephi} V\equiv V(u,e_\phi)=2\,\arctan\left[{\sqrt{1+e_{\phi}
\over 1-e_{\phi}}\,\tan \left({u \over2}\right)}\right]. \ee
Also, the mean anomaly, $l$, is related to the time as
\be \label{ellWF:l-n} l=n\left(t-t_P\right), \ee
where $t_P$ is the instant of the first passage at the periastron and
$n=2\,\pi/P$ is the mean motion with P being the orbital period.\\
In addition to this, expressions for the radial and angular velocity can be
given as
\bse \label{ellWF:rdot-phidot} \begin{align} \label{ellWF:rdot}
\dot{r}&=a_r\,e_r\,\sin u\,\left({\partial l \over\partial
u}\right)^{-1}\,{\partial l \over\partial t},\\
\label{ellWF:Phidot} \dot{\phi}&=K\left(1+2\,f_{\phi}\,\cos 2
V+3\,g_{\phi}\,\cos 3 V+4\,i_{\phi}\,\cos 4 V \right.\nonumber\\&\left.+5\,h_{\phi}\,\cos 5
V\right){\partial V \over\partial u} \left({\partial l \over\partial
u}\right)^{-1}{\partial l \over\partial t}. \end{align} \end{subequations} It may be seen
from Eqs.~\eqref{ellWF:l-u}, \eqref{ellWF:Vuephi} and \eqref{ellWF:l-n} that
\bse \label{ellWF:dldt-dldu} \begin{align} \label{ellWF:dldt} {\p l \over\p
t}&=n,\\ \label{ellWF:dldu} {\p l \over\p u}&=1-e_t\,\cos u+f_t\,\cos V\,{\p V
\over\p u}+g_t({\p V \over\p u}-1) \nonumber\\&} \def\p{\partial+2\,i_t\,{\cos 2 V}\,{\p V \over\p
u}+3\,h_t\,{\cos 3 V}\,{\p V \over\p u}.\\
\label{ellWF:dVdu} {\p V \over\p u}&={\left(1-{e_\phi}^2\right)^{1/2}
\over(1-e_\phi\,\cos u)}. \end{align} \end{subequations}
In the above, $e_\phi$ and $e_t$ denote eccentricities related to the
coordinates $\phi$ and $t$, respectively. $K$ is related to the advance of the
periastron per orbit and is given by $K=\Phi/(2\,\pi)$, where $\Phi$ is the
angle of return to the periastron. In this parametrization, $f_t$, $f_\phi$,
$g_t$ and $g_\phi$ contribute both at 2PN and 3PN order whereas $i_t$,
$i_\phi$, $h_t$ and $h_\phi$ contribute only at the 3PN order (see
Ref.~\cite{ABIQ07} for related details).\\
Once we have written equations connecting the generic dynamical variables ($r$,
$\dot{r}$, $\phi$, and $\dot{\phi}$) to the orbital elements of the elliptical
orbit, we can use inputs from MGS to express them in terms of a suitable set of
parameters of our choice. The main result of MGS is that, it provides 3PN
accurate expressions for various orbital elements ($a_r, e_t, e_r, e_\phi...$)
associated with the elliptical orbits in terms of the corresponding conserved
energy per unit reduced mass ($E$) and the parameter $h$, related to the
reduced angular momentum ($J$), by $h=J/G\,m$.\footnote{In Ref.~\cite{ABIQ07},
which uses results obtained in MGS, orbital elements are expressed in terms of
the parameters $\left\{\epsilon, j\right\}$ instead of $\left\{E, h\right\}$,
where $\epsilon$ and $j$ are defined as $\epsilon=-2\,E/c^2$ and
$j=-2\,E\,h^2$.} Using these relations one can express the dynamical variables
$r$, $\dot{r}$, $\phi$ and $\dot{\phi}$ in terms of $E$, $h$ and $u$. Here one
should note that this is not the only way in which the orbital dynamics can be
parametrized. In fact, one can re-express $E$ and $h$ in terms of any of the
two orbital elements to write equations describing the orbital motion of the
binary; however a parametrization involving gauge invariant parameters is
sometimes preferred as such parametrization is suitable for making comparisons
with related numerical results. This led MGS to use a parametrization involving
$n$ and $K=\Phi/(2\,\pi)$ (both are independent of the coordinate system used
when expressed in terms of $E$ and $h$ in order to describe the orbital motion
of the binary in elliptical orbits).\footnote{In fact, MGS uses $x_{\rm MGS}=
\left({G\,m\,n/c^3}\right)^{2/3}$ and the parameter $k'=(K-1)/3$.} Here, $n$ is
the mean motion and $K=\Phi/(2\,\pi)$ denotes the angle of the advance of the
periastron per orbital revolution. In a work related to the phasing of the GWs
from inspiralling compact binary in elliptical orbit due to Damour, Gopakumar
and Iyer \cite{DGI04}, the orbital dynamics has been described using $n$ and
$e_t$ as parametrizing variables. Following the conventions of \cite{DGI04},
K{\"o}nigsd{\"o}fer and Gopakumar \cite{KG06} provided the 3PN accurate
expressions for $r$, $\dot{r}$, $\phi$ and $\dot{\phi}$ in terms of $n$, $e_t$
and $u$.\footnote{In fact, \cite{KG06} is the extension of \cite{DGI04} and
discusses the 3PN conservative dynamics of the binaries in elliptical orbits.}
Reference~\cite{ABIQ07} makes an alternative choice of parametrization in terms
of variables $x$ and $e_t$, where $x$ is related to the orbital frequency
$\omega$ by, $x=(G\,m\,\omega/c^3)^{2/3}$, and is independent of the choice of
the coordinate system used.\footnote{Parameters $n$ and $\omega$ are related by
$n=K\omega$, and $K$ has been defined above.} The choice $\left\{x,
e_t\right\}$ as parametrizing variables leads to expressions which can be
reduced to those related to the circular orbit case ($e_t\rightarrow0$) which
uses $x$ as the expansion parameter. In addition to this, in another related
work Hinder {\it et al.} \cite{2010PhRvD..82b4033H} compared the two
parameterizations and concluded that the choice of $x$ as compared to $n$
provides better agreement with NR results. They separately discuss these two PN
models (based on the choice of parametrization as $x$ or $n$) and call them
x-model and n-model. Note that although \cite{2010PhRvD..82b4033H} uses 3PN QK
representation describing the conserved dynamics it used only 2PN accurate
dissipative dynamics and, hence, when used to obtain quasi-circular orbit limit
of the orbital phase it shows some deviations from related standard results as
was pointed out in Ref.~\cite{Huerta:2014eca}. As discussed below our results
must be coupled with 3PN evolution equations for QK variables given in
Ref.~\cite{Arun:2009mc} and, hence, would be more suitable for obtaining
quasi-circular limits of the amplitude and phase of various modes.\\
Following the arguments presented above we choose to parametrize the dynamical
variables $\{r, \dot{r}, \phi, \dot{\phi}\}$ in terms of the QK variables $\{x,
e_t, u\}$ which can further be used to write expressions for various $h_{\ell
m}$ in terms of the QK variables. Note that Ref.~\cite{2010PhRvD..82b4033H}
already lists the 3PN expressions for $r$ and $\dot{\phi}$ in terms of $\{x,
e_t, u\}$; however, they numerically differentiate $r$ to obtain $\dot{r}$ and
perform numerical integration of $\dot{\phi}$ to obtain $\phi$ in order to
avoid the use of long and complicated expressions for ${\dot r}$ and $\phi$ in
their numerical code. As mentioned above, Ref.~\cite{KG06} lists 3PN accurate
expressions for $\{r, \dot{r}, \phi, \dot{\phi}\}$ in terms of the QK variables
$\{n, e_t, u\}$ in MH coordinates (see Eqs.~(23)-(27) there). In fact, they use
a variable $\xi$ related to $n$ by $\xi=(G m n/c^3)$. Hence to obtain related
expressions parametrized in terms of $\{x, e_t, u\}$, all we need to know is how
$\xi$ is related to $x$ and $e_t$. The relation between $\xi$ and our QK
variables $x$ and $e_t$ with 3PN accuracy in MH coordinates is given as
\begin{widetext} \begin{align} \label{ellWF:xi2x} \xi &= {x^{3/2}
\over(1-e_t^2)^3}\Biggl\{1-3 e_t^2+3 e_t^4-e_t^6+x \left(-3+6 e_t^2-3
e_t^4\right)+x^2 \left[-\frac{9}{2}+7 \nu +\left(-\frac{33}{4}-\frac{\nu
}{2}\right) e_t^2+\left(\frac{51}{4}-\frac{13 \nu }{2}\right)
e_t^4\right]\nonumber\\&} \def\p{\partial+x^3 \left[\frac{3}{2}+\nu \left(\frac{457}{4}-\frac{123 \pi
^2}{32}\right)-7 \nu ^2+\left(-\frac{267}{4}+\nu \left(\frac{279}{2}-\frac{123
\pi ^2}{128}\right)-40 \nu ^2\right) e_t^2+\left(-\frac{39}{2}+\frac{55 \nu
}{4}-\frac{65 \nu ^2}{8}\right) e_t^4\right.\nonumber\\&\left.+\sqrt{1-e_t^2} \left(-15+6 \nu
+(-30+12 \nu ) e_t^2\right)\right] \Biggr\}\,. \end{align} \end{widetext}
Using the above in Eqs.~(23)-(27) of Ref.~\cite{KG06} one can write expressions
for $r$, $\dot{r}$, $\phi$ and $\dot{\phi}$ in terms of the variables $\{x,
e_t, u\}$ and then use them to obtain the expressions for the spherical
harmonic modes of the waveform for quasi-elliptical orbits.\\
We now are in a position to use the expressions for $r$, $\dot{r}$, $\phi$ and
$\dot{\phi}$ in terms of the variables $\{x, e_t, u\}$ to reexpress the
instantaneous part of spin-weighted spherical harmonic modes ($h^{\ell m}_{\rm
inst}$) in terms of parameters $x$, $e_t$ and $u$. Just like the expressions
for various $h_{\ell m}$ in general orbits we find that the expressions for
various modes in QK representation is too large to be given in the main text of
the paper. Hence we will just provide the expression for the dominant mode here
and list all the relevant modes in an additional file (Hlm-EllOrb.m). In
addition, since even the $h_{22}$ expression runs over many pages we provide
the PN structure of $h_{22}$ in the main text of the paper and list explicit
expressions for various PN pieces in Appendix A to maintain the flow of
discussion in the main text of the paper.\\
The structure of various modes, $h^{\ell m}$, remains the same as in
Eq.\eqref{ellWF:hlm}; however, now the coefficients $\hat{H}^{\ell m}_{\rm
inst}$ and $\phi$ are functions of the parameters $\{x, e_t, u\}$,
\be \label{ellWF:hlm-QKR} h^{\ell m}_{\rm inst}={4\,G\,m\,\nu\,x \over c^2
R}{\sqrt {\pi \over 5}}\,e^{-i\,m\,\phi}\,\hat{H}^{\ell m}_{\rm inst}. \ee
The instantaneous part of the dominant mode ($h^{22}_{\rm inst}$) reads
\begin{align} \label{ellWF:h22-QKR} h^{22}_{\rm inst}&={4\,G\,m\,\nu\,x \over
c^2 R}{\sqrt {\pi \over 5}}\,e^{-2\,i\,\phi}\,\hat{H}^{22}_{\rm inst}\,,
\end{align}
with
\begin{align} \label{ellWF:H22-QKR-decomp} H^{22}_{\rm inst}&= H^{22}_{\rm
Newt}+H^{22}_{\rm 1PN} + H^{22}_{\rm 2PN}+H^{22}_{\rm 2.5PN} + H^{22}_{\rm
3PN}\,, \end{align}
where various PN pieces appearing in Eq.~\eqref{ellWF:H22-QKR-decomp} are listed in
Appendix~A. Before we move to the concluding section, we would like to make a
few remarks about the results presented here.\\
We observe logarithmic dependences on the arbitrary length scale $r_0$ in
Eq.~\eqref{ellWF:H22-QKR-3PN} through the quantity $x_0$ which is related to
$r_0$ by $x_0=(G m/c^2 r_0)$. This dependence is due to the presence of terms
involving logarithms of $r_0$ in the expression for the mass quadrupole moment
($I_{ij}$) given by Eqs.~\eqref{ellWF:I2}-\eqref{ellWF:A1-A3}. As was discussed
in Sec.~\ref{ellWF:sourcemoms} we would expect such dependences to disappear
from final expression for various modes (for instance see Ref.~\cite{BFIS08}
which lists $h_{\ell m}$ for binaries in quasi-circular orbits). As has been
observed in Refs.~\cite{B98tail, BFIS08, K08, ABIQ07}, it turns out the
hereditary contribution has equal and opposite dependences on the arbitrary
length scale $r_0$ and cancels out from the final expression. Since we do not
provide hereditary contributions in this paper, such cancellation can not be
shown here. However, we explicitly show this cancellation in \cite{Mishra14}
which deals with computation of hereditary effects in various modes at 3PN
order.\\
In this paper we used the 3PN accurate QK representation which describes
the conserved dynamics of CCBs in eccentric orbits to obtain various modes in terms
of QK variables ($x$, $e_t$). However, it should be noted that the parameters
$x$, $e_t$ evolve with time over radiation reaction time scales and these
secular effects start to show at 2.5PN order \cite{KG06, Arun:2009mc}. Hence,
in order to correctly account for the reactive dynamics of the binary at 3PN
order our results should always be coupled with equations describing secular
time-evolution of $x$ and $e_t$. Equations describing the secular evolution of
orbital elements with relative 3PN accuracies were presented in
Ref.~\cite{Arun:2009mc} (see Sec. VI there). Evolution equations (due to
instantaneous terms in energy and angular momentum loss) for orbital frequency
($d\omega/dt$ (related to $x$ by $\omega=(c^3\,x^{3/2}/G m)$) and for
time-eccentricity parameter ($de_t/dt$) in terms of $x$ and $e_t$ have been
listed as Eqs.~(6.14)-(6.15) and Eqs.~(6.18)-(6.19), respectively, in
Ref.~\cite{Arun:2009mc}.\\
\section{Summary and concluding remarks}
\label{ellWF:ellWFconclusion}
In this paper we presented computations of the instantaneous contributions to
all the relevant modes of the 3PN accurate gravitational waveform of the GW
signal from nonspinning coalescing compact binaries in general orbits. The
expression for the instantaneous part of the dominant mode ($h^{\ell m}_{\rm
inst}$), in terms of the variables $r$, $\dot{r}$, $\phi$ and $\dot{\phi}$, has
been given by Eqs.~\eqref{ellWF:hlm}-\eqref{ellWF:H22} above, whereas
expressions for other subdominant modes (along with the dominant mode) have
been listed in a separate file (Hlm-GenOrb.m) that is being made available
along with the paper. Next, we specialized to the case of CCBs in
quasi-elliptical orbits using the 3PN quasi-Keplerian representation of the
conserved dynamics of compact binaries in eccentric orbits in
Sec.~\ref{ellWF:ellWFQKR}. Related 3PN accurate expressions for the
instantaneous part of the dominant mode ($h^{\ell m}_{\rm inst}$), in terms of
the variables, namely, the time-eccentricity $e_t$, a PN parameter $x$ and
eccentric anomaly $u$, is given by
Eqs.~\eqref{ellWF:h22-QKR}-\eqref{ellWF:H22-QKR-decomp} and
Eq.~\eqref{ellWF:H22-QKR}. The expressions for other sub-dominant modes (along
with the dominant mode) have been listed in a separate file (Hlm-EllOrb.m) that
is being made available along with the paper.\\
We once again remind the readers that the results presented here only account
for the contributions from the instantaneous terms in the waveform which must
be complemented by computations accounting for the hereditary effects. Our
investigations suggest that it is not possible to provide closed-form
analytical expressions for the hereditary terms for binaries moving in general
orbits. Moreover, even for the special case of CCBs in quasi-elliptical orbits
it may not be possible to have closed-form analytical expressions for the
hereditary terms valid for systems with arbitrary eccentricities. However, we
find that such computations can be performed assuming an expansion in the
eccentricity parameter ($e_t$) \cite{Mishra14}. Unlike the results presented
in this paper which can be applied to a binary with arbitrary eccentricity,
results of \cite{Mishra14} can only be applied to systems with small
eccentricities. However, the positive side of the work is that we shall have
complete 3PN analytical expression for the waveform for binaries in
quasi-elliptical orbits that can be used for comparison with related numerical
relativity results, which is one of the motivations for high PN order
computations of the gravitational waveforms. In addition, with complete
waveforms at hand one would be able to write complete polarization waveforms
(as discussed in Sec.~\ref{ellWF:ellWF3pnsphhar}) which would be useful for
data-analysis purposes.\\
\begin{acknowledgments}
We thank L Blanchet and G Faye for useful discussions. CKM thanks Chennai
Mathematical Institute for hospitality. KGA thanks Raman Research Institute for
hospitality. KGA's research was partly funded by a grant from Infosys
Foundation. All the calculations reported in this paper were carried out with
the aid of the algebraic computing software {\tt MATHEMATICA}.
\end{acknowledgments}
|
\section{introduction}
Let $(A,\mathfrak{m} )$ be a Noetherian local ring of dimension $d$ and let $M$ be a finitely generated $A$-module of dimension $r$. Let $I$
be an $\mathfrak{m} $-primary ideal. Let $\ell(N)$ denote the length of an $A$-module $N$. The function $H^{(1)}_I(M,n) = \ell(M/I^{n+1}M)$ is called the \emph{Hilbert-Samuel} function of $M$ with respect to $I$. It is well-known that there exists a polynomial $P_I(M,X) \in \mathbb{Q}[X]$ of degree $r$ such that $P_I(M,n) = H^{(1)}_I(M,n)$ for $n \gg 0$. The polynomial $P_I(M,X)$ is called the Hilbert-Samuel polynomial of $M$ with respect to $I$. We write
$$ P_I(M,X) = \sum_{i = 0}^{r}(-1)^ie_i^I(M)\binom{X+ r -i}{r-i}.$$
The integers $e_i^I(M)$ are called the $i^{th}$-Hilbert coefficent of $M$ with respect to $I$. The zeroth Hilbert coefficent $e_0^I(M)$ is called the \emph{multiplicity} of $M$ with respect to $I$.
For $j \geq 0$ let $\operatorname{Syz}^A_j(M)$ denote the $j^{th}$ syzygy of $M$.
In this paper we investigate the function $j \mapsto e_i^I(\operatorname{Syz}^A_j(M))$ for $i \geq 0$.
It becomes quickly apparent that for reasonable answers we need that the minimal resolution of $M$ should have some structure. Minimal resolutions of modules over complete intersection rings have a good structure. If $A = B/(f_1,\cdots,f_c)$ with $\mathbf{f} = f_1,\cdots,f_c$ a $B$-regular sequence and $\operatorname{projdim}_B M$ is finite then also the minimal resolution of $M$ has a nice structure. The definitive class of modules with a good structure theory of their minimal resolution is the class of modules with finite complete intersection dimension, see \cite{AGP}. We are able to prove our results for a more restrictive class of modules than modules of finite CI-dimension.
\begin{definition}\label{gci}
We say the $A$ module $M$ has finite GCI-dimension if there is a flat local extension $(B,\mathfrak{n} )$ of $A$ such that
\begin{enumerate}
\item
$\mathfrak{m} B = \mathfrak{n} $.
\item
$B = Q/(f_1,\cdots,f_c)$ , where $Q$ is local and $f_1,\ldots,f_c$ is a $Q$-regular sequence.
\item
$\operatorname{projdim}_Q M\otimes_A B$ is finite.
\end{enumerate}
\end{definition}
We note that every finitely generated module over an abstract complete intersection ring has finite GCI dimension. If $A = R/(f_1,\cdots,f_c)$ with $f_1,\cdots,f_c$ a $R$-regular sequence and $\operatorname{projdim}_R M$ is finite then also $M$ has finite GCI-dimension. We also note that if $M$ has finite GCI dimesnion then it has finite CI-dimension. If $M$ has finite CI-dimension then the function $i \mapsto \ell(\operatorname{Tor}^A_i(M,k))$ is of quasi-polynomial type with degree two. Set $\operatorname{cx}(M) = $ degree of this function $+ 1$. (See \ref{degree} for degree of a function of quasi-polynomial type).
Let $G_I(A) = \bigoplus_{n \geq 0} I^n/I^{n+1}$ be the associated graded ring of $A$ with respect to \ $I$. Let $G_I(M) = \bigoplus_{n \geq 0} I^nM/I^{n+1}M$ be the associated graded module of $M$ with respect to \ $I$.
Our main result is
\begin{theorem} \label{main}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d $ and let $M$ be a maximal Cohen-Macaulay \ $A$-module. Let $I$ be an $\mathfrak{m} $-primary ideal. Assume $M$ has finite GCI dimension. Then
for $i = 0,1,2$, the function $j \mapsto e_i^I(\operatorname{Syz}^A_j(M))$
is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{theorem}
Next we consider the asymptotic behavior of depth of associated graded modules of syzygy modules. We prove
\begin{theorem}\label{main-depth}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $\leq 2$ and let $M$ be a maximal Cohen-Macaulay \ $A$-module. Let $I$ be an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Assume $M$ has finite GCI dimension. Then
the functions $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j+1}(M))$ are constant for $j \gg 0$.
\end{theorem}
If $M$ is a Cohen-Macaulay \ $A$-module and $I$ is $\mathfrak{m} $-primary then it is known that $\operatorname{depth} G_{I^s}(M)$ is constant for all $s \gg 0$, see \cite[2.2]{Elias} (also see \cite[7.6]{Pu2}). Set $\xi_I(M) = \lim \operatorname{depth} G_{I^s}(M)$. We prove the following:
\begin{theorem}\label{main-depth-a}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $2$ and let $M$ be a maximal Cohen-Macaulay \ $A$-module. Let $I$ be an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Assume $M$ has finite GCI dimension. Then
the functions $j \mapsto \xi_I(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \xi_I(\operatorname{Syz}^A_{2j+1}(M))$ are constant for $j \gg 0$.
\end{theorem}
\s \emph{Dual Hilbert-Samuel function:} Assume $A$ has a canonical module $\omega$. The function $D^I(M,n) = \ell( \operatorname{Hom}_A(M, \omega/I^{n+1}\omega)$ is called the \emph{dual Hilbert-Samuel} function of $M$ with respect to \ $I$. In \cite{PuZ} it is shown that there exist a polynomial $t^I(M,z) \in \mathbb{Q}[z]$ of degree $d$ such that $t^I(M,n) = D^I(M,n)$ for all $n \gg 0$.
We write
$$ t^I(M,X) = \sum_{i = 0}^{d}(-1)^ic_i^I(M)\binom{X+ r -i}{r-i}.$$
The integers $c_i^I(M)$ are called the $i^{th}$- dual Hilbert coefficient of $M$ with respect to $I$. The zeroth dual Hilbert coefficient $c_0^I(M)$ is equal to $e_0^I(M)$, see \cite[2.5]{PuZ}.
We prove:
\begin{theorem}\label{main-dual}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d$, with a canonical module $\omega$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Assume $M$ has finite GCI dimension. Then
for $i = 0,1, \cdots, d$ the function $j \mapsto c_i^I(\operatorname{Syz}^A_j(M))$
is of quasi-polynomial type with period two. If $A$ is a complete intersection then the degree of each of the above functions $\leq \operatorname{cx}(M)-1$.
\end{theorem}
Although Theorem \ref{main-dual} looks more complicated than Theorem \ref{main}, its proof is
considerably simpler.
Let $H^i(-)$ denote the $i^{th}$ local cohomology functor of $G_I(A)$ with respect to \ $G_I(A)_+ = \bigoplus_{n > 0}I^n/I^{n+1}$. Set
\[
\operatorname{reg}(G_I(M)) = \max\{ i+j \mid H^i(G_I(M))_j \neq 0 \},
\]
the \emph{regularity} of $G_I(M)$. Set
\[
a_i(G_I(M)) = \max \{ j \mid H^i(G_I(M))_j \neq 0 \}.
\]
Assume that the residue field of $A$ is infinite. Let $J$ be a minimal reduction of $I$. Say $I^{r+1} = JI^r$. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Then it is well-known that $a_d(G_I(M)) \leq r - d $; see \cite[3.2]{T}(also see \cite[18.3.12]{BS}). We prove
\begin{theorem} \label{main-reg}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d \geq 2 $ and let $M$ be a maximal Cohen-Macaulay \ $A$-module. Let $I$ be an $\mathfrak{m} $-primary ideal. Assume $M$ has finite GCI dimension. Then the set
\[
\left\{ \frac{a_{d-1}(G_I(\operatorname{Syz}^A_i(M))}{i^{\operatorname{cx}(M)-1}} \right\}_{i \geq 1 }
\]
is bounded.
\end{theorem}
Here is an overview of the contents of this paper. In section two we discuss many preliminary facts that we need.
In section three we introduce a technique which is useful to prove our results.
In section four(five) we discuss the case when dimension of $A$ is one(two) respectively. The proof of Theorem \ref{main} is divided in these two sections. The proof of Theorem \ref{main-depth} is in section five. In section six we prove Theorem \ref{main-depth-a}. Theorem \ref{main-dual} is proved in section seven. Finally in section eight we prove Theorem \ref{main-reg}.
\section{Preliminaries}
In this section we collect a few preliminary results that we need.
\s
\label{hilbcoeff}
The \emph{Hilbert function} of $M$ with respect to $I$ is the function
\[
H_{I}(M,n) = \lambda(I^nM/I^{n+1}M)\quad \text{for all} \ n \geq 0.
\]
It is well known that the formal power series $\sum_{n \geq 0}H_{I}(M,n)z^n$
represents a rational function of a
special type:
\begin{equation*}
\sum_{n \geq 0}H_{I}(M,n) z^n = \frac{h_{I}(M ,z)}{(1-z)^{r}}\quad \text{where}
\ r = \dim M \ \text{and} \ h_{I}(M,z) \in\mathbb{Z}[z].
\end{equation*}
It can be shown that
$e_{i}^{I}(M) = (h_I(M,z))^{(i)}(1)/i! $ for all $ 0 \leq i \leq r$. It is convenient to set $e_{i}^{I}(M) = (h_I(M,z))^{(i)}(1)/i! $ even when $i \geq r$.
The number $e_{0}^{I}(M)$ is also called the \emph{multiplicity} of $M$ with respect
to $I$.
\textbf{I:} \textit{Superficial elements.}\\
For definition and basic properties of superficial sequences see \cite[p.\ 86–-87]{Pu1}.
The following result is well-known to experts. We give a proof due to lack of a reference.
\begin{lemma}\label{count-sup}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d \geq 1$ and let $I$ be an $\mathfrak{m} $-primary ideal. Assume the residue field $k = A/\mathfrak{m} $ is uncountable. Let $\{M_n \}_{n \geq 1}$ be a sequence of maximal Cohen-Macaulay \ $A$-modules. Then there exists $\mathbf{x} = x_1,\ldots,x_d \in I$ such that $\mathbf{x}$ is an $M_n$-superficial sequence (with respect to \ $I$) for all $n \geq 1$.
\end{lemma}
\begin{proof}
We recall the construction of a superficial element for a finitely generated module $M$ of dimension $r \geq 1$ (note $M$ need not be Cohen-Macaulay). Let $\mathcal{M} $ be the maximal ideal of $G_I(A)$. Let
\[
\operatorname{Ass}^*(G_I(M)) = \{ P \mid P \in \operatorname{Ass}(G_I(M)) \ \text{and} \ P \neq \mathcal{M} \}.
\]
Set $L = I/I^2$ and $V = L/\mathfrak{m} L = I/ \mathfrak{m} I$. For $P \in \operatorname{Spec} G_I(A)$ let $P_1 = P \cap L$ and
$\overline{P_1} = (P_1 + \mathfrak{m} I)/\mathfrak{m} I$.
If $P \neq \mathcal{M} $ then $P_1 \neq L$. By Nakayama's Lemma it follows easily that $\overline{P_1} \neq V$. If $k$ is infinite then
$$ S(M) = V \setminus \bigcup_{P \in \operatorname{Ass}^*(G_I(M))} \overline{P_1} \neq \emptyset. $$
Then $u \in I$ such that $\overline{u} \in S(M)$ is an $M$-superficial element with respect to $I$.
Now assume that $k$ is uncountable and let $\mathcal{F} = \{M_n \}_{n \geq 1}$ be a sequence of maximal Cohen-Macaulay \ $A$-modules. Set $M_0 = A$. Then
$$ S(\mathcal{F}) = V \setminus \bigcup_{n \geq 0} \bigcup_{P \in \operatorname{Ass}^*(G_I(M_n))} \overline{P_1} \neq \emptyset. $$
If $u \in I$ such that $\overline{u} \in S(\mathcal{F})$ then $\overline{u} \in S(M_n)$ for all $n \geq 1$. Thus $u$ is $M_n$-superficial element with respect to $I$ for all $n \geq 0$. Set $x_1 = u$. If $d \geq 2$ then note that $\{ M_n/x_1 M_n \}$ is a sequence of maximal Cohen-Macaulay \ $A/(x_1)$ modules. So by induction the result follows.
\end{proof}
\s\textbf{ Associated graded module and Hilbert function mod a superficial element: }\label{mod-sup-h} \\
Let $x \in I$
be $M$-superficial. Set $N = M/xM$. Let $r = \dim M$. The following is well-known cf., \cite{Pu1}.
\begin{enumerate}[\rm(1)]
\item Set $b_I(M,z) = \sum_{i \geq 0}\ell\big( (I^{n+1}M \colon_M x)/ I^nM \big)z^n$. Since
$x$ is $M$-superficial we have $b_I(M,z) \in \mathbb{Z} [z]$.
\item $h_I(M,z) = h_I(N,z) - (1-z)^rb_I(M,z)$; cf., \cite[Corollary 10 ]{Pu1}.
\item So we have
\begin{enumerate}[\rm(a)]
\item $e_i(M) = e_i(N)$ for $i = 0,\ldots, r-1$.
\item $e_r(M) = e_r(N) - (-1)^r b_I(M,1)$.
\end{enumerate}
\item
The following are equivalent
\begin{enumerate}[\rm(a)]
\item $x^*$ is $G_I(M)$-regular.
\item $G_I(N) = G_I(M)/x^*G_I(M)$
\item $b_I(M,z) = 0$
\item $e_r(M) = e_r(N)$.
\end{enumerate}
\item
\emph{(Sally descent)} If $\operatorname{depth} G_I(N) > 0$ then $x^*$ is $G_I(M)$-regular.
\end{enumerate}
\textbf{II:}\textit{Base change.}
\s\label{AtoA'}
Let $\phi \colon (A,\mathfrak{m} ) \rightarrow (A',\mathfrak{m} ')$ be a flat local ring homomorphism with $\mathfrak{m} A' = \mathfrak{m} '$. Set $I' = IA'$ and if
$N$ is an $A$-module set $N' = N\otimes A'$.
It can be easily seen that
\begin{enumerate}[\rm (1)]
\item
$\lambda_A(N) = \lambda_{A'}(N')$.
\item
$\operatorname{projdim}_A N = \operatorname{projdim}_{A'} N'$.
\item
$H^I(M,n) = H^{I'}(M',n)$ for all $n \geq 0$.
\item
$\dim M = \dim M'$ and $\operatorname{grade}(K,M) = \operatorname{grade}(KA',M')$ for any ideal $K$ of $A$.
\item
$\operatorname{depth} G_{I}(M) = \operatorname{depth} G_{I'}(M')$.
\end{enumerate}
\noindent The specific base changes we do are the following:
(i) If $M$ is a maximal Cohen-Macaulay \ $A$-module with finite GCI dimension then by definition there exists a flat homomorphism $(A,\mathfrak{m} ) \rightarrow (B,\mathfrak{n} )$ with $\mathfrak{m} B = \mathfrak{n} $ and $B = Q/(f_1,\ldots,f_c)$ for some local ring $Q$ and a $Q$-regular sequence $f_1,\ldots, f_c$ such that $\operatorname{projdim}_Q M\otimes_A B < \infty$. In this case we set $B = A^\prime$.
(ii) If $k \subseteq k'$ is an extension of fields then it is well-known that there exists a flat
local ring homomorphism $(A,\mathfrak{m} ) \rightarrow (A',\mathfrak{m} ')$ with $\mathfrak{m} A' = \mathfrak{m} '$ and $A'/\mathfrak{m} ' = k'$.
We use this construction when the residue field $k$ of $A$ is finite or countably infinite. We take $k'$ to be any uncountable field containing $k$.
(iii) We can also take $A'$ to be the completion of $A$.
\begin{remark}\label{base-change}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring and let $M$ be a maximal Cohen-Macaulay \ $A$-module with
finite GCI dimension. After doing the base-changes (i), (ii) and (iii) above we may assume that $A = Q/(f_1,\ldots,f_c)$ for some \emph{complete} Cohen-Macaulay \ local ring $Q$ and a $Q$-regular sequence $f_1,\ldots, f_c$ such that $\operatorname{projdim}_Q M < \infty$. Furthermore we may assume that
the residue field of $A$ is uncountable.
\end{remark}
\textbf{III:}\textit{Quasi-polynomial functions of period 2}. \\
Let us recall that a function $f \colon \mathbb{Z} _{\geq 0} \rightarrow \mathbb{Z} _{\geq 0}$ is said to be of \textit{ quasi-polynomial type} with period $g \geq 1$ if there exists polynomials $P_0, P_1,\ldots, P_{g-1} \in \mathbb{Q}[X]$ such that $f(mg + i) = P_i(m)$ for all $m \gg 0$ and $i = 0, \cdots, g-1$.
We need the following well-known result regarding quasi-polynomials of period two.
\begin{lemma}\label{g2}
Let $f \colon \mathbb{Z} _{\geq 0} \rightarrow \mathbb{Z} _{\geq 0}$. The following are equivalent:
\begin{enumerate}[\rm (i)]
\item
$f$ is of quasi-polynomial type with period $2$.
\item
$\sum_{n\geq 0} f(n)z^n = \frac{h(z)}{(1-z^2)^c}$ for some $h(z) \in \mathbb{Z} [z]$ and $c \geq 0$.
\end{enumerate}
Furthermore if $P_0, P_1 \in Q[X]$ are polynomials such that $f(2m + i) = P_i(m)$ for all $m \gg 0$ and $i = 0, 1$ then $\deg P_i \leq c-1$. \qed
\end{lemma}
\textit{Convention:} We set the degree of the zero-polynomial to be $-\infty$.
\begin{remark}\label{degree}
Let $f \colon \mathbb{Z} _{\geq 0} \rightarrow \mathbb{Z} _{\geq 0}$ be of quasi-polynomial type with period $2$. Let $P_0, P_1 \in Q[X]$ be polynomials such that $f(2m + i) = P_i(m)$ for all $m \gg 0$ and $i = 0, 1$. Then set $\deg f = \max\{\deg P_0, \deg P_1 \}$.
\end{remark}
As a consequence we get that
\begin{corollary}\label{g2-c}
Let $f \colon \mathbb{Z} _{\geq 0} \rightarrow \mathbb{Z} _{\geq 0}$. If the function $g(n) = f(n) + f(n-1)$ is of quasi-polynomial type with period $2$ then so is $f$. Furthermore
$\deg f = \deg g$.
\end{corollary}
\begin{proof}[Sketch of a proof]
Set $u(z) = \sum_{n \geq 0}f(n)z^n$. Then by our hypothesis and Lemma \ref{g2} we have
\[
(1+z)u(z) + f(0)z = \frac{h(z)}{(1-z^2)^c},
\]
for some $h(z) \in \mathbb{Z} [z]$ and $c \geq 0$. An easy computation now shows that
\[
u(z) = \frac{r(z)}{(1-z^2)^c},
\]
for some $r(z) \in \mathbb{Z} [z]$. Thus $f$ is of quasi-polynomial type with period two.
As $f(n) \leq g(n)$ we clearly get $\deg f \leq \deg g$. As $g(2n) = f(2n) + f(2n-1)$ and $g(2n + 1) = f(2n+1) + f(2n)$ we get that $\deg g \leq \deg f$.
\end{proof}
The following result is also well-known:
\begin{lemma}\label{g2-growth}
Let $f \colon \mathbb{Z} _{\geq 0} \times \mathbb{Z} _{\geq 0} \rightarrow \mathbb{Z} _{\geq 0}$ be a function such that
\[
\sum_{m,n\geq 0} f(m,n)z^mw^n = \frac{h(z,w)}{(1-z^2)^c(1-w)^{d}}
\]
Then for $m,n \gg 0$ we have
\[
f(m,n) = \sum_{i = 0}^{d-1}(-1)^ie_i(m)\binom{n+d-1-i}{d-1-i},
\]
where the functions $m \mapsto e_i(m)$ for $i = 0,\ldots, d-1$ are of quasi-polynomial type with period $2$ and degree $\leq c-1$. \qed
\end{lemma}
\textbf{IV:} \textit{Eisenbud operators}\\
Let $Q$ be a Noetherian ring and let $\mathbf{f}= f_1,\ldots f_c$ be a regular sequence in $Q$. Set $A = Q/(\mathbf{f})$. Let $M$ be a finitely generated $A$-module with $\operatorname{projdim}_Q M$ finite.
\s Let $\mathbb{F}: \cdots F_n \rightarrow \cdots F_1 \rightarrow F_0 \rightarrow 0$
be a free resolution of $M$ as a $A$-module.
Let $t_1,\ldots t_c \colon \mathbb{F}(+2) \rightarrow \mathbb{F} $ be the \emph{Eisenbud-operators}; see
\cite[section 1.]{Eisenbud-80}. Then
\begin{enumerate}
\item $t_i$ are uniquely determined up to homotopy.
\item $t_i, t_j$ commute up to homotopy.
\end{enumerate}
Let $T = A[t_1,\ldots,t_c]$ be a polynomial ring over $A$ with variables $t_1,\ldots,t_c$ of degree $2$.
Let $D$ be an $A$-module. The operators $t_j$ give well-defined maps
\[
t_j \colon \operatorname{Ext}^{i}_{A}(M,D) \rightarrow \operatorname{Ext}^{i+2}_{R}(M,D) \quad \ \text{for} \ 1 \leq j \leq c \ \text{and all} \ i,
\]
\[
t_j \colon \operatorname{Tor}^{i+2}_{A}(M,D) \rightarrow \operatorname{Tor}^{i}_{R}(M,D) \quad \ \text{for} \ 1 \leq j \leq c \ \text{and all} \ i.
\]
This turns $\operatorname{Ext}_A^*(M,D) = \bigoplus_{i \geq 0} \operatorname{Ext}^i_A(M,D)$ and $\operatorname{Tor}_A^*(M,D) = \bigoplus_{i \geq 0} \operatorname{Tor}^i_A(M,D)$ into modules over $T$ (here we give an element $t \in \operatorname{Tor}^i_A(M,D)$ degree $-i$). Furthermore these structure depend only on $ \mathbf{f}$, are natural in both module arguments and commute with the connecting maps induced by short exact sequences.
\s
Gulliksen, \cite[3.1]{Gulliksen}, proved that if $\operatorname{projdim}_Q M$ is finite then
$\operatorname{Ext}_A^*(M,D) $ is a finitely generated $T$-module.
If $A$ is local and $D = k$, the residue field of $A$, Avramov in \cite[3.10]{LLAV} proved a converse; i.e., if
$\operatorname{Ext}_A^*(M,k)$ is a finitely generated $T$-module then $\operatorname{projdim}_Q M$ is finite. For a more general result, see \cite[4.2]{AGP}.
\begin{definition}(with notation as above:) Assume $A$ is local with residue field $k$.
Set $\operatorname{cx} M = \dim_T \operatorname{Ext}_A^*(M,k)$, the \emph{complexity} of $M$.
\end{definition}
We need the following result regarding the growth of lengths of certain Tor's. Recall a graded module $X$ over $T = A[t_1,\ldots, t_c]$ is said to be *-Artinian $T$-module if every descending chain of
graded submodules of $X$ terminates.
\begin{proposition}\label{growth}
Let $(Q,\mathfrak{n} )$ be a complete Noetherian ring and let $\mathbf{f}= f_1,\ldots f_c$ be a regular sequence in $Q$. Set $A = Q/(\mathbf{f})$ and $\mathfrak{m} = \mathfrak{n} /( \mathbf{f})$. Let $M$ be a finitely generated $A$-module with $\operatorname{projdim}_Q M$ finite. Let $D$ be a non-zero $A$-module of finite length. Let $\operatorname{Ext}_A^*(M,D) $ be a finitely generated $T = A[t_1,\ldots, t_c]$-module as above. Then
\begin{enumerate}[\rm (1)]
\item
$\dim_T \operatorname{Ext}_A^*(M,D) \leq \operatorname{cx}(M)$.
\item
The function $n \mapsto \ell( \operatorname{Ext}^n_A(M, D))$ is of quasi-polynomial type with period $2$ and degree $\leq \operatorname{cx}(M)-1$.
\item
$\operatorname{Tor}_A^*(M,D)$ is a *-Artinian $T$-module. Here $t \in \operatorname{Tor}^A_n(M, D)$ has degree $-n$.
\item
The function $n \mapsto \ell( \operatorname{Tor}^n_A(M, D))$ is of quasi-polynomial type with period $2$ and degree $\leq \operatorname{cx}(M)-1$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) This follows from \cite[Theorem 5.3]{AGP}.
(2) This easily follows from (1).
For (3), (4) we use the following result. Let $E$ be the injective hull of $k$. Then
we have
\[
\operatorname{Hom}_A(\operatorname{Tor}^A_n(M, D), E) \cong \operatorname{Ext}^n_A(M, \operatorname{Hom}_A(D, E)). \tag{$\dagger$}
\]
(3) From $(\dagger)$ it follows that the Matlis dual of
$\operatorname{Tor}_A^*(M,D)$ is \\ $\operatorname{Ext}^*_A(M, \operatorname{Hom}_A(D, E))$. By Gulliksen's result we have that
$\operatorname{Ext}^*_A(M, \operatorname{Hom}_A(D, E))$ is a finitely generated graded $T$-module. So by Matlis-duality \cite[3.6.17]{BH} we get that \\ $\operatorname{Tor}_A^*(M,D)$ is a *-Artinian $T$-module.
(4) This follows from (2) and $(\dagger)$.
\end{proof}
As an immediate consequence we obtain:
\begin{corollary}\label{mult}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. If $M$ has finite GCI dimension over $A$ then the function
$i \mapsto e_0^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period $2$ and degree $= \operatorname{cx}(M) - 1$.
\end{corollary}
\begin{proof}
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with infinite residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$.
Then $ \operatorname{rank} F_i = \ell(\operatorname{Tor}^A_i(M,k))$. So by \ref{growth} we get that the function $i \mapsto \operatorname{rank} F_i$ is quasi-polynomial of period two.
Set $M_i = \operatorname{Syz}^A_i(M)$. The exact sequence
$0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ yields
$e_0^I(M_i) + e^I_0(M_{i+1}) = (\operatorname{rank} F_i) e_0^I(A)$. The result
now follows from \ref{g2-c}.
\end{proof}
We also prove:
\begin{proposition}\label{hilbSyz}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Assme $M$ has finite GCI dimension over $A$.
Fix $n \geq 0$. Then the function
\[
i \mapsto \ell\left(\frac{I^n\operatorname{Syz}^{i}_A(M)}{I^{n+1}\operatorname{Syz}^{i}_A(M)}\right)
\]
is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{proposition}
\begin {proof}
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with infinite residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$. Fix $n \geq 0$. The exact sequence
$0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ yields an exact sequence
\[
0 \rightarrow \operatorname{Tor}^A_{i+1}(M, A/I^{n+1}) \rightarrow M_{i+1}/I^{n+1}M_{i+1} \rightarrow F_i/I^{n+1}F_i \rightarrow M_i/I^{n+1}M_i \rightarrow 0.
\]
Similarly we have an exact sequence
\[
0 \rightarrow \operatorname{Tor}^A_{i+1}(M, A/I^{n}) \rightarrow M_{i+1}/I^{n}M_{i+1} \rightarrow F_i/I^{n}F_i \rightarrow M_i/I^{n}M_i \rightarrow 0.
\]
Computing lengths we have
\begin{align*}
\ell\left(\frac{I^nM_i}{I^{n+1}M_i}\right) + \ell\left(\frac{I^nM_{i+1}}{I^{n+1}M_{i+1}}\right) &= (\operatorname{rank}{F_i})\ell\left(\frac{I^n}{I^{n+1}}\right) + \ell\left(\operatorname{Tor}^A_{i+1}(M, A/I^{n+1}) \right) \\
&\ \ \ - \ell \left( \operatorname{Tor}^A_{i+1}(M, A/I^n) \right).
\end{align*}
By \ref{growth} the terms in the right hand side are of quasi-polynomial type with period $2$ and degree $\leq \operatorname{cx}(M)-1$. The result follows from
Corollary \ref{g2-c}.
\end{proof}
As a consequence of the above Proposition we get:
\begin{corollary}\label{zero-dim}
Let $(A,\mathfrak{m} )$ be an Artin local ring and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a finitely generated $A$-module of finite GCI-dimension. Then for all $j \geq 0$ the function
$i \mapsto e_j^I(\operatorname{Syz}^A_{i}(M))$ is of quasi-polynomial type with period $2$ and degree $\leq \operatorname{cx}(M)-1$.
\end{corollary}
\begin{proof}
We note that $I^r = 0$ for some $r \geq 1$. The result now follows from Corollary \ref{hilbSyz}.
\end{proof}
\s\label{Eis-reg} \emph{Eisenbud operators modulo a regular element.} \\
Let $x $ be $A \oplus M$-regular. Let $y \in Q$ be a pre-image of $x$ in $Q$. As permutation of regular sequences is also a regular sequence we also get that $y$ is $Q$-regular.
We now note that $\operatorname{projdim}_{Q/(y)}(M/xM)$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$ over $A$. Then $\mathbb{F}/x\mathbb{F}$ is a minimal resolution of $M/xM$ over $A/(x)$. Let $t_1,\ldots, t_c$ be the Eisenbud operators over $\mathbb{F}$. By the construction of Eisenbud operators it follows that $t_1,\ldots,t_c$ induce operators $t_1^*,\ldots, t_c^*$ over $\mathbb{F}/x\mathbb{F}$. Let $N$ be an $A/(x)$-module. Set $E = \operatorname{Tor}^A_*(M, N) = \operatorname{Tor}^{A/(x)}_*(M/xM, N)$. Then the action of $t_i$ on $E$ is same as that of $t_i^*$.
\s \label{bigraded} Now assume that $Q$ be a Noetherian ring and let $\mathbf{f}= f_1,\ldots f_c$ be a regular sequence in $Q$. Set $A = Q/(\mathbf{f})$. Let $M$ be a finitely generated $A$-module with $\operatorname{projdim}_Q M$ finite.
Let $\mathbb{F}$ be a minimal resolution of $M$. Let $t_1,\cdots, t_c \colon \mathbb{F}(+2) \rightarrow \mathbb{F}$ be the Eisenbud operators.
Let $I$ be an $\mathfrak{m} $-primary ideal of $A$ and let $\mathcal{R} = A[Iu]$ be the Rees algebra of $A$ with respect to \ $I$. Let $N = \bigoplus_{n \geq 0} N_n$ be an $\mathcal{R} $-module (not necessarily finitely generated). We claim that $\bigoplus_{i \geq 0} \operatorname{Tor}^A_i(M, N) = \bigoplus_{i, n \geq 0} \operatorname{Tor}^A_i(M, N_n)$ has a bigraded $\mathcal{R} [t_1,\ldots,t_c]$-module structure. Here $t_i $ have degree $(0,2)$ and if $a \in I^n$ then $\deg au^n = (n, 0)$. To see this let $v = au^s \in \mathcal{R} _s$.
The map $N_n \xrightarrow{v} N_{n+s}$
yields the following commutative diagram
\[
\xymatrix
{
\mathbb{F}\otimes N_n \ar@{->}[d]^{v}\ar@{->}[r]_{t_j}
&\mathbb{F}\otimes N_n(-2) \ar@{->}[d]^{v}
\\
\mathbb{F}\otimes N_{n+s} \ar@{->}[r]_{t_j}
&\mathbb{F}\otimes N_{n+s}(-2).
}
\]
Taking homology we get the required result.
An analogous argument yields that $\bigoplus_{i ,n\geq 0} \operatorname{Ext}^i_A(M, N_n)$ is a bigraded $\mathcal{R} [t_1,\ldots, t_c]$-module.
\section{$L^I_i(M)$}
In this section we extend and simplify a technique from \cite{PuMCM} and \cite{Pu2}.
Let $A$ be a Noetherian ring, $I$ an $\mathfrak{m} $-primary ideal and $M$ a finitely generated
$A$-module.
\s Set $L^I_0(M) = \bigoplus_{n \geq 0} M/I^{n+1}M$. Let $\mathcal{R} = A[Iu]$ be the \emph{Rees-algebra} of $I$. Let $\mathcal{S} = A[u]$. Then $\mathcal{R} $ is a subring of $\mathcal{S} $. Set
$M[u] = M\otimes_A \mathcal{S} $ an $\mathcal{S} $-module and so an $\mathcal{R} $-module. Let $\mathcal{R} (M) = \bigoplus_{n \geq 0}I^nM$ be the Rees-module of $M$ with respect to $I$. We have the following exact sequence of $\mathcal{R} $-modules
\[
0 \rightarrow \mathcal{R} (M) \rightarrow M[u] \rightarrow L^I_0(M)(-1) \rightarrow 0.
\]
Thus $L^I_0(M)(-1)$ (and so $L^I_0(M)$) is a $\mathcal{R} $-module. We note that $L^I_0(M)$ is \emph{not} a finitely generated $\mathcal{R} $-module. Also note that $L^I_0(M) = M \otimes_A L^I_0(A)$.
\s For $i \geq 1$ set
$$L^I_i(M) = \operatorname{Tor}^A_i(M, L^I_0(A)) = \bigoplus_{n \geq 0 } \operatorname{Tor}^A_i(M, A/I^{n+1}). $$
We assert that $L^I_i(M)$ is a finitely generated $\mathcal{R} $-module for $i \geq 1$. It is sufficient to prove it for $i = 1$. We tensor the exact sequence $0 \rightarrow \mathcal{R} \rightarrow \mathcal{S} \rightarrow L^I_0(A)(-1) \rightarrow 0$ with $M$ to obtain a sequence of $\mathcal{R} $-modules
\[
0 \rightarrow L^I_1(M)(-1) \rightarrow \mathcal{R} \otimes_A M \rightarrow M[u] \rightarrow L^I_0(M)(-1) \rightarrow 0.
\]
Thus $ L^I_1(M)(-1)$ is a $\mathcal{R} $-submodule of $\mathcal{R} \otimes_A M$. The latter module is a finitely generated $\mathcal{R} $-module. It follows that $L^I_1(M)$ is a finitely generated $\mathcal{R} $-module.
\s Now assume that $A$ is Cohen-Macaulay \ and $M$ is maximal Cohen-Macaulay. Set $N = \operatorname{Syz}^A_1(M)$ and $F = A^{\mu(M)}$ (here $\mu(M)$ is the cardinality of a minimal generator set of $M$). We tensor the exact sequence
\[
0 \rightarrow N \rightarrow F \rightarrow M \rightarrow 0,
\]
with $L^I_0(A)$ to obtain an exact sequence of $\mathcal{R} $-modules
\[
0 \rightarrow L^I_1(M) \rightarrow L^I_0(N) \rightarrow L^I_0(F) \rightarrow L^I_0(M) \rightarrow 0.
\]
As $e_0^I(F) = e_0^I(M) + e_0^I(N)$ we get that the function $n \rightarrow \ell(\operatorname{Tor}^A_1(M, A/I^{n+1}))$ is of polynomial type and degree $\leq d - 1$. Thus $\dim L^I_1(M) \leq d$.
We need the following result.
\begin{proposition}\label{H1}
Let $(A,\mathfrak{m} )$ be Cohen-Macaulay \ local ring of dimension $d \geq 1$ with infinite residue field and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a MCM \ $A$-module. Set $M_1 = \operatorname{Syz}^A_1(M)$. Let $x$ be $A \oplus M \oplus M_1$-superficial with respect to \ $I$. Let $\mathcal{R} = A[Iu]$ be the Rees algebra of $A$ with respect to \ $ I$. Set $X = xu \in \mathcal{R} _1$. If $N$ is a $\mathcal{R} $-module, let $H_1(X, N)$ denote the first Koszul homology of $N$ with respect to \ $X$. Then
\begin{enumerate}[\rm (1)]
\item
$H_1(X, L_0^I(M)) = \bigoplus_{n \geq 0}(I^{n+1}M \colon x)/I^nM$.
\item
$X$ is $L_0^I(M)$ regular if and only if $x^*$ is $G_I(M)$-regular.
\item
If $x^*$ is $G_I(A)$-regular then
$H_1(X,L_1^I(M)) \cong H_1(X, L_0^I(M_1))$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) The map $M/I^nM \xrightarrow{X} M/I^{n+1}M$ is given by $m + I^nM \mapsto xm + I^{n+1}M$. The result follows.
(2) This follows from (1).
(3) We have an exact sequence $0 \rightarrow L_1^I(M) \rightarrow L_0^I(M_1) \rightarrow L_0^I(F_0) \xrightarrow{\pi} L_0^I(M) \rightarrow 0$. Let $E = \ker \pi$. As $x^*$ is $G_I(A)$-regular we get that $H_1(X, L_0^I(A)) = 0$. So $H_1(X, E) = 0$. The result follows.
\end{proof}
\textbf{II}: \textit{Ratliff-Rush filtration.}
Let $(A,\mathfrak{m} )$ be a Noetherian local ring and let $I$ be an $\mathfrak{m} $-primary ideal in $A$. Let $M$ be a finitely generated $A$-module. The Ratliff-Rush submodule $\widetilde{IM}$ of $M$ with respect to \ $I$
is defined by
\[
\widetilde{IM} = \bigcup_{k \geq 0} (I^{k+1}M \colon I^k).
\]
If $\operatorname{depth} M > 0 $ then $\widetilde{I^kM} = I^kM $ for $k \gg 0$.
\s Now assume $d = \dim A > 0$ and $A$ is Cohen-Macaulay. Also assume $M$ is MCM $A$-module. Let $\mathcal{R} $ be the Rees algebra of $A$ with respect to \ $I$ and let $\mathcal{M}$ be its unique maximal graded ideal. Then by \cite[4.7]{Pu2} we have
\[
H^0_\mathcal{M}(L^I_0(M)) = \bigoplus_{n \geq 0} \frac{ \widetilde{I^{n+1}M}}{I^{n+1}M}.
\]
We need the following result:
\begin{proposition}\label{RR}[with assumptions as above]
If $\operatorname{depth} G_I(A) > 0$ then
$$H^0_\mathcal{M}(L^I_1(M)) \cong H^0_\mathcal{M}(L^I_0(\operatorname{Syz}^A_1(M))). $$
\end{proposition}
\begin{proof}
Set $M_1 = \operatorname{Syz}^A_1(M)$. The exact sequence $0 \rightarrow M_1 \rightarrow F \rightarrow M \rightarrow 0$ yields an exact sequence
\[
0 \rightarrow L_1^I(M) \rightarrow L^I_0(M_1) \rightarrow L_0^I(F) \rightarrow L^I_0(M) \rightarrow 0.
\]
Therefore we have an exact sequence
\[
0 \rightarrow H^0_\mathcal{M} (L^I_1(M)) \rightarrow H^0_\mathcal{M} (L^I_0(M_1)) \rightarrow H^0_\mathcal{M} (L_0^I(F))
\]
Now as $\operatorname{depth} G_I(A) > 0$ we get that $\widetilde{I^n} = I^n $ for all $n \geq 1$. So $H^0_\mathcal{M} (L^I_0(A)) = 0$. Thus $H^0_\mathcal{M} (L^I_0(F)) = 0$. The result follows.
\end{proof}
\s \label{bi} Now assume that $A = Q/(\mathbf{f})$ where $\mathbf{f} = f_1,\ldots, f_c$ is a $Q$-regular sequence and $M$ is a finitely generated $A$-module with $\operatorname{projdim}_Q M$ finite. Let
$I$ be an $\mathfrak{m} $-primary ideal. Set $L(M) = \bigoplus_{i \geq 0}L_i^I(M) = \bigoplus_{i \geq 0}\operatorname{Tor}^A_i(M, L^I_0(A))$. Give $L(M)$ a structure of a bigraded $S = \mathcal{R} [t_1,\cdots,t_c]$-module as discussed in \ref{bigraded}.
Then $H^0_\mathcal{M} (L(M)) = \bigoplus_{i\geq 0}H^0_\mathcal{M} (L^I_i(M))$ is a bigraded $S$-module. To see this note that $H^0_\mathcal{M} (L(M)) = H^0_{\mathcal{M} S}(L(M))$.
\section{dimension one}
In this section we assume $\dim A = 1$. We prove that if $M$ is MCM $A$-module with finite GCI dimension then the function $i \mapsto e_1^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period two. If $G_I(A)$ is Cohen-Macaulay \ then we show that the functions $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j+1}(M))$ is constant for $j \gg 0$.
The following Lemma is crucial.
\begin{lemma}\label{crucial}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension one and with an infinite residue field. Let $I$ be an $\mathfrak{m} $-primary ideal and let $x$ be $A$-superficial with respect to $I$. Assume $I^{r+1} = xI^r$. Let $M$ be a MCM $A$-module. Then
\begin{enumerate}[\rm (1)]
\item
$\deg h^I(M,z) \leq r$
\item
The postulation number of the Hilbert-Samuel function of $M$ with respect to \ $I$ is $\leq r - 2$.
\item
For $n \geq r$ and for all $i \geq 1$ we have
\[
\operatorname{Tor}^A_i(M, A/I^{n}) \cong \operatorname{Tor}^A_i(M, A/I^{n+1}) \quad \text{as $A$-modules}.
\]
\item
$\widetilde{I^nM} = I^nM$ for all $n \geq r$.
\end{enumerate}
\end{lemma}
\begin{proof}
We first note that as $x$ is $A$-regular it is also $M$-regular. Thus the equality $I^{r+1}M = xI^rM$ implies that $(I^{n+1}M \colon x) = I^nM$ for $n \geq r$. Therefore $x$ is $M$-superficial with respect to \ $I$.
(1) This follows from \cite[Proposition 13]{Pu1}.
(2) This follows from (1).
(3) Set $B = A/(x)$. For $n \geq r $ we have an exact sequence
\[
0 \rightarrow A/I^n \xrightarrow{\alpha_n} A/I^{n+1} \rightarrow B \rightarrow 0.
\]
Here $\alpha_n( a + I^n) = xa + I^{n+1}$. We have also used that for $n \geq r $, $I^{n+1} \subseteq (x)$. As $\operatorname{Tor}^A_i(M, B) = 0$ for $i \geq 1$ we get the required result.
(4) As $(I^{n+1}M \colon x) = I^n M$ for all $n \geq r$, by \cite[2.7]{Pu2} we get $\widetilde{I^nM} = I^nM$ for $n \geq r$.
\end{proof}
As an immediate corollary we obtain
\begin{theorem}\label{e1}
Let $(A,\mathfrak{m} )$ be Cohen-Macaulay \ local ring of dimension $d$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a MCM $A$-module. Assume $M$ has finite GCI-dimension over $A$.
Then the function $i \mapsto e_1^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{theorem}
\begin{proof}
By Corollary \ref{zero-dim} the result holds when $A$ is Artin.
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$. If $d \geq 2$ then by \ref{count-sup} we may choose $\mathbf{x} = x_1,\ldots,x_{d-1}$ which is $A \oplus M_i$-superficial for all $i \geq 0$. As $e^I_1(M_i) = e_1^I(M_i/ \mathbf{x}M_i)$ for all $i \geq 0$, we may assume $\dim A = 1$.
Let $x$ be $A$-superficial with respect to \ $I$. Say $I^{r+1} = xI^r$. Then by
Lemma \ref{crucial} $x$ is $M_i$-superficial for all $i \geq 0$ and the postulation number of the Hilbert-Samuel function of $M_i$ with respect to \ $I$ is $\leq r - 2$.
Fix $n \geq r$. The exact sequence
$0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ yields an exact sequence
\[
0 \rightarrow \operatorname{Tor}^A_{i+1}(M, A/I^{n+1}) \rightarrow M_{i+1}/I^{n+1}M_{i+1} \rightarrow F_i/I^{n+1}F_i \rightarrow M_i/I^{n+1}M_i \rightarrow 0.
\]
Using Lemma \ref{crucial},
it follows that
\[
\ell(\operatorname{Tor}^A_{i+1}(M, A/I^{n+1})) = (\operatorname{rank} F_i)e_1^I(A) - e_1^I(M_i) - e_1^I(M_{i+1}).
\]
The functions $i \mapsto \operatorname{rank} F_i$ and $i \mapsto \ell(\operatorname{Tor}^A_{i}(M, A/I^n))$ are of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$. The result now follows from \ref{g2-c}.
\end{proof}
We now prove:
\begin{theorem}\label{asymp-depth-1}
Let $(A,\mathfrak{m} )$ be Cohen-Macaulay \ local ring of dimension one and let $I$ be an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Let $M$ be a MCM $A$-module. Assume $M$ has finite GCI dimension over $A$. Then the functions $j \mapsto \operatorname{depth} G_{I}(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j+1}(M))$ are constant for $j \gg 0$
\end{theorem}
\begin{proof}
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with infinite residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$.
Let $x$ be $I$-superficial. Then $x^*$ is $G_I(A)$-regular. Say $I^{r+1} = xI^r$. Then by Lemma \ref{crucial} we get that $x$ is $M_i$-superficial for all $i \geq 0$. We have
$(I^{n+1}M_i \colon x) = I^nM_i$ for all $n \geq r$ and for all $i \geq 0$.
Let $t_1,\ldots,t_c$ be the Eisenbud operators on $\mathbb{F}$. Let $\mathcal{R} $ be the Rees algebra of $A$ with respect to \ $I$. By \ref{bigraded} we get that $E = \bigoplus_{i \geq 0} L_i^I(M)$ is a bigraded $S = \mathcal{R} [t_1,\ldots, t_c]$-module.
Set $X = xu$. Then $H_1(X,E)$ is a bigraded $S$-module. By \ref{H1} we get that $H_1(X,E) = \bigoplus_{i \geq 0} H_1(X, L_0^I(M_i))$. We now note that $H_1(X,E) \subseteq E_{n \leq r-1}$ and so is an Artinian $A[t_1,\ldots, t_c]$-module. It follows that the function $j \mapsto \ell(H_1(X, L_0^I(M_i)))$ is of quasi-polynomial type with period two. By \ref{H1} the result follows.
\end{proof}
We now prove
\begin{proposition}\label{e2-dim1}
Let $(A,\mathfrak{m} )$ be Cohen-Macaulay \ local ring of dimension one and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a MCM $A$-module of finite GCI dimension over $A$.
Then the function $i \mapsto e_2^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{proposition}
\begin{proof}
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with infinite residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$.
Let $x$ be $I$-superficial. Say $I^{r+1} = xI^r$. Then by Lemma \ref{crucial} we get that
\begin{equation}\label{eqn-1}
\operatorname{Tor}^A_i(M, A/I^n) \cong \operatorname{Tor}^A_i(M, A/I^{n+1}) \quad \text{ for all $n \geq r$.}
\end{equation}
Fix $n \geq 0$. Then by \ref{growth} the function $i \mapsto \ell(\operatorname{Tor}^A_i(M, A/I^{n+1}))$ is quasi-polynomial of period two and degree $\leq \operatorname{cx}(M)-1$. Thus by (\ref{eqn-1}) we get that there exists $p(z,w) \in \mathbb{Z}[z,w]$ such that
\[
\sum_{i \geq 1, n \geq 0} \ell(\operatorname{Tor}^A_i(M, A/I^{n+1}))z^iw^n = \frac{p(z,w)}{(1-z^2)^{\operatorname{cx}(M)}(1-w)}.
\]
Set
$$f(i,n) = \sum_{m = 0}^{n}\ell(\operatorname{Tor}^A_i(M, A/I^{m+1})).$$
Multiplying the previous equation
by $1/(1-w)$ we get that
\[
\sum_{i \geq 1, n \geq 0} f(i,n)z^iw^n = \frac{p(z,w)}{(1-z^2)^{\operatorname{cx}(M)}(1-w)^2}.
\]
By \ref{g2-growth} we get that
for $i,n \gg 0$
$$ f(i,n) = g_0(i)(n+1) - g_1(i) $$
where for $j = 0, 1$ the function $i \mapsto g_j(i)$ is of quasi-polynomial type with period $2$ and degree $ \leq \operatorname{cx}(M) - 1$.
It remains to note that
\begin{align*}
g_0(i) &= (\operatorname{rank} F_i)e_1^I(A) - e_1^I(M_i) - e_1^I(M_{i-1}), \text{and} \\
g_1(i) &= (\operatorname{rank} F_i)e_2^I(A) - e_2^I(M_i) - e_2^I(M_{i-1}).
\end{align*}
The result now follows from Corollary \ref{g2-c}.
\end{proof}
We need the following result in the next section.
\begin{proposition}\label{RR-artin-dim1}
Let $(A,\mathfrak{m} )$ be a one dimensional Cohen-Macaulay \ local ring with infinite residue field and let $I$ be an $\mathfrak{m} $-primary ideal of $A$. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Assume $A = Q/(\mathbf{f})$ where $\mathbf{f} = f_1,\ldots, f_c$ is a $Q$-regular sequence and $\operatorname{projdim}_Q M $ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$ over $A$. Let $t_1,\ldots,t_c$ be the Eisenbud operators over $\mathbb{F}$. Let $\mathcal{R} = A[Iu]$ be the Rees algebra of $A$ with respect to \ $I$ and let $\mathcal{M} $ be its unique maximal homogeneous ideal.
Then $\bigoplus_{i \geq 0} H^0_\mathcal{M} (L^I_i(M))$ is an Artinian $A[t_1,\ldots,t_c]$-module. Furthermore the function $i \mapsto \ell(H^0_\mathcal{M} (L^I_i(M)))$ is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{proposition}
\begin{proof}
Let $x$ be $A$-superficial with respect to \ $I$. Say $I^{r+1} = xI^r$. Set $M_i = \operatorname{Syz}^A_i(M)$ for $i \geq 0$. Then by Lemma \ref{crucial}.4 we get that $\widetilde{I^n M_i} = I^nM_i$ for $n \geq r$. So $H^0_\mathcal{M} (L^I_0(M_i))_n =0$ for $n \geq r-1$. The exact sequence $0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ yields an exact sequence
\[
0 \rightarrow L_{i+1}^I(M) \rightarrow L_0^I(M_{i+1}) \rightarrow L_0^I(F_i) \rightarrow L_0^I(M_i) \rightarrow 0.
\]
So we have an inclusion $0 \rightarrow H^0_\mathcal{M} (L^I_{i+1}(M)) \rightarrow H^0_\mathcal{M} (L_0^I(M_{i+1}))$. It follows that $H^0_\mathcal{M} (L^I_{i+1}(M))_n = 0$ for $n \geq r-1$.
By \ref{bigraded}, $L = \bigoplus_{i \geq 0}L^I_i(M)$ is a bigraded $S = \mathcal{R} [t_1,\ldots,t_c]$-module. Then \\ $H^0_\mathcal{M} (L) = \bigoplus_{i \geq 0} H^0_\mathcal{M} (L^I_i(M))$ is also a bigraded $S$-module. As $H^0_\mathcal{M} (L) \subseteq E = L_{n \leq r -1}$ it follows that $H^0_\mathcal{M} (L)$ is an Artinian $A[t_1,\ldots,t_c]$-module. It follows that the function $i \mapsto \ell(H^0_\mathcal{M} (L^I_i(M)))$ is of quasi-polynomial type with period two. To compute its degree note that the function $i \mapsto \ell(E_{i})$ has degree $\leq \operatorname{cx}(M)-1$.
\end{proof}
\section{dimension 2}
In this section we assume $(A,\mathfrak{m} )$ is Cohen-Macaulay \ with $\dim A = 2$ . Let $I$ be an $\mathfrak{m} $-primary ideal. If $M$ is a MCM $A$-module with finite GCI dimension over $A$ then we prove that the function $i \mapsto e_2^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period two. If $G_I(A)$ is Cohen-Macaulay \ then we show that the functions $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j+1}(M))$ are constant for $j \gg 0$.
\begin{theorem}\label{e2}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a MCM $A$-module. Assume that $M$ has finite GCI dimension over $A$.
Then the function $i \mapsto e_2^I(\operatorname{Syz}^A_i(M))$ is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{theorem}
\begin{proof}
By Corollary \ref{zero-dim} the result holds when $A$ is Artin. By Proposition \ref{e2-dim1} the result holds when $d = 1$.
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$. If $d \geq 3$ then by \ref{count-sup} we may choose $\mathbf{x} = x_1,\ldots,x_{d-2}$ which is $A \oplus M_i$-superficial for all $i \geq 0$. As $e_2^I(M_i) = e_2^I(M_i/ \mathbf{x}M_i)$ for all $i \geq 0$, we may assume $\dim A = 2$.
We note that for any $n \geq 1$ we have $e_2^{I^n}(M) = e_2^I(M)$. For $n \gg 1$ we have $\operatorname{depth} G_{I^n}(A) \geq 1$. Thus we can assume $\operatorname{depth} G_I(A) > 0$.
Let $x$ be $A\oplus M_i$-superficial for all $i \geq 0$. Then $x^*$ is $G_I(A)$-regular. Let $\mathcal{R} = A[Iu]$ be the Rees algebra of $A$ with respect to \ $I$. Set $X = xu \in \mathcal{R} _1$ and $\overline{A} =
A/(x)$. Also set $\overline{M_i} = M_i/xM_i$.
Let $t_1,\ldots, t_c$ be the Eisenbud operators over $\mathbb{F}$. Then $L(M) = \bigoplus_{i \geq 0} L^I_i(M)$ is a bigraded $S = \mathcal{R} [t_1,\ldots, t_c]$-module (see \ref{bi}).
Notice $X$ is $L^I_0(A)$-regular. So we have an exact sequence of $\mathcal{R} $-modules
\[
0 \rightarrow L^I_0(A)(-1) \xrightarrow{X} L^I_0(A) \rightarrow L^I_0(\overline{A}) \rightarrow 0.
\]
This induces an exact sequence of $S$-modules
\begin{equation}\label{two}
0 \rightarrow K \rightarrow L(M)(-1,0) \xrightarrow{X} L(M) \rightarrow L(\overline{M})
\end{equation}
By Proposition \ref{H1} we get that
\begin{equation}\label{2.5}
K = \bigoplus_{i \geq 0} H_1(X, L^I_0(M_i)).
\end{equation}
Let $\mathcal{M} $ be the unique maximal homogeneous maximal ideal of $\mathcal{R} $. We take local cohomology with respect to \ $\mathcal{M} $ (on (\ref{two})).
As $K$ is $\mathcal{M} $-torsion we get an exact sequence of $S$-modules
\begin{equation}\label{three}
0 \rightarrow K \rightarrow H^0_\mathcal{M} (L(M))(-1,0) \rightarrow H^0_\mathcal{M} (L(M)) \xrightarrow{\rho} H^0_\mathcal{M} (L(\overline{M})).
\end{equation}
By \ref{Eis-reg} and \ref{RR-artin-dim1} we get that $H^0_\mathcal{M} (L(\overline{M}))$ is an Artin
module over the subring
$T = A[t_1,\cdots, t_c]$ of $S$. So $E = \operatorname{image} \rho$ is also an Artin $T$-module.
By \ref{RR} we get that
\[
H^0_\mathcal{M} (L(M)) = \bigoplus_{i \geq 0}H^0_\mathcal{M} (L^I_0(M_i)).
\]
Thus for a fixed $i$ we get that $H^0_\mathcal{M} (L(M))_{(i,n)} = 0$ for $n \gg 0$.
It follows that $\ell(K_i) = \ell(E_i)$. As $E$ is an Artin $T$-module the function
$i \mapsto \ell(E_i)$ is of quasi-polynomial type with period two. By \ref{RR-artin-dim1} its degree is $\leq \operatorname{cx}(M) - 1$. Thus the function
$i \mapsto \ell(K_i)$ is of quasi-polynomial type with period two and degree $ \leq \operatorname{cx}(M)-1$. By \ref{mod-sup-h}(3b)
\ref{H1}(1) and (\ref{2.5}) we have
\[
e_2^I(M_i) = e_2^I(\overline{M_i}) - \ell(K_i),
\]
is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
\end{proof}
Next we prove:
\begin{theorem}\label{asymp-depth-2}
Let $(A,\mathfrak{m} )$ be Cohen-Macaulay \ local ring of dimension two and let $I$ be an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Let $M$ be a MCM $A$-module. Assume $M$ has finite GCI dimension over $A$. Then the functions $j \mapsto \operatorname{depth} G_{I}(\operatorname{Syz}^A_{2j}(M))$ and $j \mapsto \operatorname{depth} G_I(\operatorname{Syz}^A_{2j+1}(M))$ are constant for $j \gg 0$
\end{theorem}
To prove this theorem we need the following:
\begin{lemma}\label{depth1}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension two and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be an MCM $A$-module. Let $x,y$ be an $A\oplus M$-superficial sequence. Set $J = (x,y)$ and assume $I^{r+1} = JI^r$. Then the following are equivalent:
\begin{enumerate}[\rm (1)]
\item
$x^*$ is $G_I(M)$-regular.
\item
$(I^{n+1}M \colon x) = I^nM$ for all $n \geq 1$.
\item
$(I^{n+1}M \colon x) = I^nM$ for $n = 1,\cdots, r$.
\end{enumerate}
\end{lemma}
\begin{proof}
The equivalence (1) and (2) is well-known. Also clearly (2) implies (3).
Now assume (3). Set $N = M/xM$. We have the following exact sequence
\[
0 \rightarrow \frac{(I^{n}M \colon J)}{I^{n-1}M} \xrightarrow{\gamma_n} \frac{(I^{n}M \colon x)}{I^{n-1}M} \xrightarrow{\beta_n} \frac{(I^{n+1}M \colon x)}{I^nM} \xrightarrow{\alpha_n} \frac{I^{n+1}M}{JI^nM} \xrightarrow{\rho_n} \frac{I^{n+1}N}{yI^nN}\rightarrow 0,
\]
where $\rho_n$ is the natural surjection, $\gamma_n$ is the natural inclusion, $\alpha_n( a + I^nM) = xa + JI^nM$ and $\beta_n(a + I^{n-1}M) = ya + I^nM$.
Set
\[
U = \bigoplus_{n \geq 0}\frac{(I^{n}M \colon J)}{I^{n-1}M} \quad \text{and} \quad V = \bigoplus_{n \geq 0}\frac{(I^{n+1}M \colon x)}{I^nM}.
\]
If $(I^{n+1}M \colon x) = I^nM$ for $n = 1,\cdots, r$ then by the above exact sequence we have an exact sequence
\[
0 \rightarrow U \rightarrow V[-1] \rightarrow V \rightarrow 0
\]
As $\ell(V)$ is finite we get that $V[-1] \cong V$ and this implies $V = 0$, so (2) holds.
\end{proof}
We now give
\begin{proof}[Proof of Theorem \ref{asymp-depth-2}]
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$.
Let $x,y$ be $M_i\oplus A$-superficial sequence with respect to \ $I$ for all $i$. Such an element exists as the residue field of $A$ is uncountable, see \ref{count-sup}. Then $x^*$ is $G_I(A)$-regular.
Let $\mathcal{R} = A[Iu]$ be the Rees algebra of $A$ with respect to \ $I$. Set $X = xu \in \mathcal{R} _1$ and $\overline{A} =
A/(x)$. Also set $\overline{M_i} = M_i/xM_i$.
Let $t_1,\ldots, t_c$ be the Eisenbud operators over $\mathbb{F}$. Then $L(M) = \bigoplus_{i \geq 0} L^I_i(M)$ is a bigraded $S = \mathcal{R} [t_1,\ldots, t_c]$-module (see \ref{bi}).
Notice $X$ is $L^I_0(A)$-regular. So we have an exact sequence of $\mathcal{R} $-modules
\[
0 \rightarrow L^I_0(A)(-1) \xrightarrow{X} L^I_0(A) \rightarrow L^I_0(\overline{A}) \rightarrow 0.
\]
This induces an exact sequence of $S$-modules
\begin{equation}\label{four}
0 \rightarrow K \rightarrow L(M)(-1,0) \xrightarrow{X} L(M) \rightarrow L(\overline{M}).
\end{equation}
By Proposition \ref{H1} we get that
\begin{equation}\label{five}
K = \bigoplus_{i \geq 0} H_1(X, L^I_0(M_i)).
\end{equation}
Let $I^{r+1} = (x,y)I^r$. We note that $E = K_{n \leq r}$ is an Artin $A[t_1,\ldots, t_c]$-module. Thus the function $i \rightarrow \ell(E_i)$ is of quasi-polynomial type with period two. Therefore by Lemma \ref{depth1} we get that for each $j = 0, 1$ either $\operatorname{depth} G_I(M_{2i+j}) = 0 $ for $i \gg 0$ or $\operatorname{depth} G_I(M_{2i+j}) \geq 1 $ for $i \gg 0$. We now go mod $x$ and use Theorem \ref{asymp-depth-1} and \ref{mod-sup-h}(5) to conclude.
\end{proof}
\section{asymptotic depth}
In this section $(A,\mathfrak{m} )$ is a Cohen-Macaulay \ local ring of dimension two and $I$ is an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Let $M$ be a MCM $A$-module with finite GCI-dimension over $A$. Let $\xi_I(M) = \lim_{n \rightarrow \infty} \operatorname{depth} G_{I^n}(M)$. In this section we prove that the functions $i \mapsto \xi_I(\operatorname{Syz}^A_{2i}(M))$ and $i \mapsto \xi_I(\operatorname{Syz}^A_{2i+1}(M))$ are constant for $i \gg 0$.
Let $\mathcal{R} (I)$ denote the Rees algebra of $A$ with respect to \ $I$ and let $\mathcal{M} $ denote its unique maximal homogeneous ideal.
\s \label{reductions} We note that $\xi_I(M) \geq 1$. Furthermore $\xi_I(M) \geq 2$ if and only if \\ $H^1_\mathcal{M} (L^I_0(M))_{-1} = 0$, see \cite[9.2]{Pu2}. As $\dim A = 2$ the only possible values of $\xi_I(M)$ is $1$ or $2$. We now note that $L^I_0(M)(-1) = \bigoplus_{n \geq 0} M/I^nM$ behaves very well with respect to \ to the Veronese functor. Clearly for $m \geq 1$ we have
\[
L^I_0(M)(-1)^{<m>} = \left( \bigoplus_{n \geq 0} M/I^nM \right)^{<m>} = \bigoplus_{n \geq 0}M/I^{nm} = L^{I^m}_{0}(M)(-1).
\]
Also note that $\mathcal{M} ^{<m>}$ is the unique maximal homogeneous ideal of $\mathcal{R} (I^m)$. It follows that
\[
H^1_\mathcal{M} (L^I_0(M))_{-1} \cong H^1_{\mathcal{M} ^{<m>}}(L^{I^m}_0(M))_{-1} \quad \text{as $A$-modules}.
\]
We need the following:
\begin{lemma}\label{need}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension two and let $I$ be an $\mathfrak{m} $-primary ideal. Let $J = (x,y)$ be a minimal reduction of $I$ and assume that $I^{r+1} = JI^r$. Let
$M$ be a maximal Cohen-Macaulay \ $A$-module. Then $H^1_\mathcal{M} (L^I_0(M))_n = 0$ for $n \geq r - 2$. In particular if $m \geq r$ then
\[
H^1_{\mathcal{M} ^{<m>}}(L^{I^m}_0(M))_n = 0 \quad \text{for} \ n \geq 0.
\]
\end{lemma}
\begin{proof}
Let $a_2(G_I(M)) = \max \{ n \mid H^2_{\mathcal{M} }(G_I(M))_n \neq 0 \}$. Then by
\cite[3.2]{T} we get that $a_2(G_I(M)) \leq r - 2$. The exact sequence
$0 \rightarrow G_I(M) \rightarrow L^I_0(M) \rightarrow L^I_0(M)(-1) \rightarrow 0$ induces an exact sequence
\[
H^1_\mathcal{M} (L^I_0(M)) \rightarrow H^1_\mathcal{M} (L^I_0(M))(-1) \rightarrow H^2_\mathcal{M} (G_I(M)).
\]
By \cite[6.4]{Pu2} we get that $H^1_\mathcal{M} (L^I_0(M))_n = 0$ for $n \gg 0$. From the above exact sequence it follows that
$H^1_\mathcal{M} (L^I_0(M))_n = 0 $ for all $n \geq r - 2$. The rest of the assertion follows from \ref{reductions}.
\end{proof}
We now state and prove the main result of this section.
\begin{theorem}\label{asymp}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension two and let $I$ be an $\mathfrak{m} $-primary ideal with $G_I(A)$ Cohen-Macaulay. Let $M$ be a MCM $A$-module and assume $M$ has finite GCI dimension over $A$. Then
the functions $i \mapsto \xi_I(\operatorname{Syz}^A_{2i}(M))$ and $i \mapsto \xi_I(\operatorname{Syz}^A_{2i+1}(M))$ are constant for $i \gg 0$.
\end{theorem}
\begin{proof}
By Remark \ref{base-change} we may assume $A = Q/(\mathbf{f})$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. Let $\mathbb{F}$ be a minimal resolution of $M$. Set $M_i = \operatorname{Syz}^A_i(M)$. Let $\mathcal{R} (I) = A[Iu]$ denote the Rees algebra of $A$ with respect to \ $I$ and let $\mathcal{M} $ denote its unique maximal homogeneous ideal. By Lemma \ref{need} and \ref{reductions} we may assume that $H^1_\mathcal{M} (L^I_0(M_i))_n = 0 $ for $n \geq 0$ and for all $i \geq 0$.
We will show that the vanishing of $H^1_\mathcal{M} (L^I_0(M_i))_{-1}$ can be detected by a quasi-polynomial of period two.
\noindent\textit{Claim-1:} $H^1_\mathcal{M} (L^I_1(M_i))_{-1} \cong H^1_\mathcal{M} (L_0^I(M_{i+1}))_{-1}$ and
$H^1_\mathcal{M} (L^I_1(M_i))_{j} = 0$ for $j \geq 0$.
The exact sequence $0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ induces
an exact sequence $0 \rightarrow L^I_1(M_i) \rightarrow L^I_0(M_{i+1}) \rightarrow L^I_0(F_i) \xrightarrow{\pi_i} L^I_0(M_i) \rightarrow 0$. Let $C = \ker \pi_i$.
As $G_I(A)$ is Cohen-Macaulay \ we have $H^i_\mathcal{M} (L^I_0(A)) = 0$ for $i = 0, 1$, see \cite[5.2]{Pu2}. Thus
$H^0_\mathcal{M} (C) = 0$ and $H^1_\mathcal{M} (C) \cong H^0_\mathcal{M} (L^I_0(M_i))$. In particular we have $H^1_\mathcal{M} (C)_{-1} = 0$.
The exact sequence $0 \rightarrow L^I_1(M_i) \rightarrow L^I_0(M_{i+1}) \rightarrow C \rightarrow 0$ induces an exact sequence
\[
0 \rightarrow H^1_\mathcal{M} (L^I_1(M_i)) \rightarrow H^1_\mathcal{M} (L^I_0(M_{i+1})) \rightarrow H^1_\mathcal{M} (C)
\]
As $H^1_\mathcal{M} (C)_{-1} = 0$ we get $H^1_\mathcal{M} (L^I_1(M_i))_{-1} \cong H^1_\mathcal{M} (L_0^I(M_{i+1}))_{-1}$. As \\ $H^1_\mathcal{M} (L^I_0(M_{i+1}))_j = 0$ for $j \geq 0$ we also get
$H^1_\mathcal{M} (L^I_1(M_i))_{j} = 0$ for $j \geq 0$. Thus the Claim is proved.
Let $x$ be $M_i \oplus A$-superficial for each $i \geq 0$. Set $X = xu$, $\overline{A} = A/(x)$ and $\overline{M_i} = M_i/xM_i$. As $x^*$ is $G_I(A)$-regular, we have an exact sequence
\[
0 \rightarrow L^I_0(A)(-1) \xrightarrow{X} L^I_0(A) \rightarrow L^I_0(\overline{A}) \rightarrow 0.
\]
After tensoring with $M$ this induces for $i \geq 0$ an exact sequence
\[
L^I_{i+1}(M)(-1) \xrightarrow{X} L^I_{i+1}(M) \rightarrow L^I_{i+1}(\overline{M}) \rightarrow
L^I_{i}(M)(-1) \xrightarrow{X} L^I_{i}(M).
\]
We note that $L^I_{i+1}(M) \cong L^I_1(M_i)$. Also by \ref{H1} we get that
$H_1(X, L^I_1(M_i)) \cong H_1(X, L^I_0(M_{i+1}))$. Thus we have an exact sequence
\begin{align*}
0 &\rightarrow H_1(X, L^I_0(M_{i+1})) \rightarrow L^I_1(M_i)(-1) \xrightarrow{X} L^I_1(M_i) \rightarrow \\
&\xrightarrow{\alpha_i} L^I_1(\overline{M_i}) \rightarrow H_1(X, L^I_0(M_{i})) \rightarrow 0.
\end{align*}
Let $W_i = \ker \alpha_i$ and $D_i = \operatorname{image} \alpha_i$. As $H_1(X, L^I_0(M_{i+1}))$
has finite length we get a surjection $H^0_\mathcal{M} (L^I_1(M_i))(-1) \rightarrow H^0_\mathcal{M} (W_i)$ and a
isomorphism $H^1_\mathcal{M} (W_i) \cong H^1_\mathcal{M} (L^I_1(M_i))(-1)$. Thus
\begin{equation}\label{5.5}
H^0_\mathcal{M} (W_i)_0 = 0 \quad \text{and} \quad H^1_\mathcal{M} (W_i)_{0} \cong H^1_\mathcal{M} (L^I_1(M_i))_{-1}.
\end{equation}
The exact sequence $ 0 \rightarrow W_i \rightarrow L^I_1(M_i) \rightarrow D_i \rightarrow 0$ induces an exact sequence
\[
0 \rightarrow H^0_\mathcal{M} (W_i) \rightarrow H^0_\mathcal{M} (L^I_1(M_i)) \rightarrow H^0_\mathcal{M} (D_i) \rightarrow H^1_\mathcal{M} (W_i) \rightarrow
H^1_\mathcal{M} (L^I_1(M_i))
\]
Evaluating at $n = 0$ we get an exact sequence
\begin{equation}\label{six}
0 \rightarrow H^0_\mathcal{M} (L^I_1(M_i))_0 \rightarrow H^0_\mathcal{M} (D_i)_0 \rightarrow H^1_\mathcal{M} (W_i)_0 \rightarrow 0.
\end{equation}
By the exact sequence $0 \rightarrow D_i \rightarrow L^I_1(\overline{M_i}) \rightarrow H_1(X, L^I_0(M_{i})) \rightarrow 0$ we obtain an exact sequence
\[
0 \rightarrow H^0_\mathcal{M} (D_i) \rightarrow H^0_\mathcal{M} (L^I_1(\overline{M_i})) \rightarrow H_1(X, L^I_0(M_{i}))
\]
As $H_1(X, L^I_0(M_{i}))_0 = 0$ we get $H^0_\mathcal{M} (D_i)_0 \cong H^0_\mathcal{M} (L^I_1(\overline{M_i}))_0$.
By (\ref{six}) we get that
\[
\ell(H^0_\mathcal{M} (L^I_1(M_i))_0) \leq \ell(H^0_\mathcal{M} (L^I_1(\overline{M_i}))_0) \quad \text{with equality iff} \ H^1_\mathcal{M} (W_i)_0 = 0.
\]
The latter condition holds by (\ref{5.5}) if and only if $H^1_\mathcal{M} (L^I_1(M_i))_{-1} = 0$ and this by our claim holds if and only if $H^1_\mathcal{M} (L^I_0(M_{i+1}))_{-1} = 0$. This holds if and only if $\xi_I(M_{i+1}) \geq 2$.
We now claim that the functions $i \mapsto \ell(H^0_\mathcal{M} (L^I_1(M_i))_0)$ and
$i \mapsto \ell(H^0_\mathcal{M} (L^I_1(\overline{M_i}))_0)$ are of quasi-polynomial type with period two. This will prove our assertion. To see that these functions are of quasi-polynomial type, let $t_1,\ldots, t_c$ be the Eisenbud operators over $\mathbb{F}$. Then $L(M) = \bigoplus_{i \geq 0}L^I_i(M)$ is a bigraded module over $S = \mathcal{R} [t_1,\ldots,t_c]$, see \ref{bi}. Thus $H^0_\mathcal{M} (L(M)) = \bigoplus_{i \geq 0}H^0_\mathcal{M} (L^I_i(M))$ is a bigraded module over $S$. By \ref{RR} we get that $H^0_\mathcal{M} (L^I_i(M))_0 \cong \widetilde {IM_i}/IM_i$. As $\bigoplus_{i \geq 0}L^I_i(M)_0$ is an Artin $A[t_1,\ldots, t_c]$-module we get that $\bigoplus_{n\geq 0}H^0_\mathcal{M} (L^I_i(M))_0 $ is an Artin $A[t_1,\ldots, t_c]$-module. It follows that the function $i \mapsto \ell(H^0_\mathcal{M} (L^I_1(M_i))_0)$ is of quasi-polynomial type with period two. A similar argument yields that the function \\ $i \mapsto \ell(H^0_\mathcal{M} (L^I_1(\overline{M_i}))_0)$ is of quasi-polynomial type with period two.
\end{proof}
\section{Dual Hilbert Coefficients}
In this section we assume that $(A,\mathfrak{m} )$ is a Cohen-Macaulay \ local ring with a canonical module $\omega$. Let $I$ be an $\mathfrak{m} $-primary ideal and let $M$ be a maximal Cohen-Macaulay \ $A$-module. The function $D^I(M,n) = \ell( \operatorname{Hom}_A(M, \omega/I^{n+1}\omega)$ is called the \emph{dual Hilbert-Samuel} function of $M$ with respect to \ $I$. In \cite{PuZ} it is shown that there exist a polynomial $t^I(M,z) \in \mathbb{Q}[z]$ of degree $d$ such that $t^I(M,n) = D^I(M,n)$ for all $n \gg 0$.
We write
$$ t^I(M,X) = \sum_{i = 0}^{d}(-1)^ic_i^I(M)\binom{X+ r -i}{r-i}.$$
The integers $c_i^I(M)$ are called the $i^{th}$- dual Hilbert coefficient of $M$ with respect to $I$. The zeroth dual Hilbert coefficient $c_0^I(M)$ is equal to $e_0^I(M)$.
We prove:
\begin{theorem}\label{main-dual-2}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d$, with a canonical module $\omega$ and let $I$ be an $\mathfrak{m} $-primary ideal. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Assume $M$ has finite GCI dimension. Then
for $i = 0,1, \cdots, d$ the function $j \mapsto c_i^I(\operatorname{Syz}^A_j(M))$
is of quasi-polynomial type with period two. If $A$ is a complete intersection then the degree of each of the above functions $\leq \operatorname{cx}(M)-1$.
\end{theorem}
\begin{proof}
By Remark \ref{base-change} we may
assume $A = Q/(f_1,\ldots, f_c)$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite.
Let $\mathbb{F}$ be a minimal resolution of $M$. Let $t_1,\ldots,t_c$ be the Eisenbud operators over $\mathbb{F}$.
Let $\mathcal{R} = A[ut]$ be the Rees algebra of $A$ with respect to \ $I$ and let $N = \bigoplus_{n \geq 0} N_n$ is a finitely generated $\mathcal{R} $-module. Give
$E(N) = \bigoplus_{i,n \geq 0} \operatorname{Ext}^i_A(M,N_n)$ a bi-graded $S = \mathcal{R} [t_1,\ldots,t_c]$
structure as described in \ref{bigraded}. Then by \cite[1.1]{Pu3}, $E(N)$ is a finitely generated $S$-module. In particular $E_I(\omega, M) = \bigoplus_{i,n \geq 0} \operatorname{Ext}^i_A(M,I^n\omega)$ is a finitely generated $S$-module. We now note that \\
$\operatorname{Ext}^{i}_A(M, \omega/I^{n+1}\omega) \cong \operatorname{Ext}^{i+1}_A(M,I^{n+1}\omega)$ for all $i \geq 1$ and for all $n \geq 0$. Thus
$$D_I(\omega, M) = \bigoplus_{i \geq 1,n \geq 0}\operatorname{Ext}^{i}_A(M, \omega/I^{n+1}\omega)$$
is a submodule of $E_I(\omega, M)(0,1)$ and so is finitely generated. Set \\ $K = \operatorname{ann}_S D_I(\omega, M)$. Then $\mathfrak{q} = K_{0,0}$ is $\mathfrak{m} $-primary and so in particular $D_I(\omega, M)$ is a finitely generated $T = \mathcal{R} /\mathfrak{q} \mathcal{R} [t_1,\ldots,t_c]$-module. Therefore the Hilbert series of $D_I(\omega, M)$ is of the form
\[
\frac{h(z,w)}{(1-z)^d(1-w^2)^c}
\]
By \ref{g2-growth} we get that for $i, n \gg 0$
\[
\ell(\operatorname{Ext}^{i}_A(M, \omega/I^{n+1}\omega)) = \sum_{l = 0}^{d-1}(-1)^lv_l^I(i)\binom{n+d-1-l}{d-1-l}
\]
where the functions $i \mapsto v_l^I(i)$ for $i = 0,\ldots,d-1$ are of quasi-polynomial type with period two and degree $\leq c -1$. We note that it is possible that some of the $v_l^I(i)$ is identically zero.\\
Now set $M_i = \operatorname{Syz}^A_i(M)$. The exact sequence $0 \rightarrow M_{i+1} \rightarrow F_i \rightarrow M_i \rightarrow 0$ induces an exact sequence
\begin{align*}
0 \rightarrow &\operatorname{Hom}_A(M_i, \omega/I^{n+1}\omega) \rightarrow \operatorname{Hom}_A(F_i, \omega/I^{n+1}\omega) \rightarrow \operatorname{Hom}_A(M_{i+1}, \omega/I^{n+1}\omega)\\
\rightarrow &\operatorname{Ext}^{i+1}_A(M, \omega/I^{n+1}\omega) \rightarrow 0.
\end{align*}
Thus for $1 \leq l \leq d$ we have that
\[
\operatorname{rank}(F_i)e_l^I(\omega) - c_l^I(M_i) - c_l^I(M_{i+1}) = v_{l-1}^I(i+1).
\]
Using \ref{g2-c} we get that for $1 \leq l \leq d$ the function $i \mapsto c_l^I(M_i)$ is of quasi-polynomial type with period two and degree $\leq c-1$. Also note that $c_0^I(M_i) = e_0^I(M_i)$. So for $l = 0$ the result follows from \ref{mult}.
If $A$ is a complete intersection then we may take $Q$ to be a complete regular local ring. Then by \cite[3.9]{LLAV} there exists a complete local ring $R$ with $A = R/(g_1,\ldots,g_r)$ with $r = \operatorname{cx}(M)$, $g_1,\ldots,g_r$ a $R$-regular sequence and $\operatorname{projdim}_R M$ finite. It now follows that the degree of the functions $i \mapsto c_l^I(M_i)$ is $\leq \operatorname{cx}(M) -1$.
\end{proof}
\section{regularity}
Let $H^i(-)$ denote the $i^{th}$ local cohomology functor of $G_I(A)$ with respect to \ $G_I(A)_+ = \bigoplus_{n > 0}I^n/I^{n+1}$.
Set
\[
a_i(G_I(M)) = \max \{ j \mid H^i(G_I(M))_j \neq 0 \}.
\]
Assume that the residue field of $A$ is infinite. Let $J$ be a minimal reduction of $I$. Say $I^{r+1} = JI^r$. Let $M$ be a maximal Cohen-Macaulay \ $A$-module. Then it is well-known that $a_d(G_I(M)) \leq r - d $. We prove
\begin{theorem} \label{main-reg-2}
Let $(A,\mathfrak{m} )$ be a Cohen-Macaulay \ local ring of dimension $d \geq 2 $ and let $M$ be a maximal Cohen-Macaulay \ $A$-module. Let $I$ be an $\mathfrak{m} $-primary ideal. Assume $M$ has finite GCI dimension. Then the set
\[
\left\{ \frac{a_{d-1}(G_I(\operatorname{Syz}^A_i(M))}{i^{\operatorname{cx}(M)-1}} \right\}_{i \geq 1 }
\]
is bounded.
\end{theorem}
\begin{proof}
By Remark \ref{base-change} we may
assume $A = Q/(f_1,\ldots, f_c)$ where $Q$ is complete with uncountable residue field, $\mathbf{f} = f_1,\cdots, f_c$ is a $Q$-regular sequence and
$\operatorname{projdim}_Q M$ is finite. We prove the result by induction on $d = \dim A$.
We first consider the case when $d = 2$. Set $M_i = \operatorname{Syz}^A_i(M)$ for $i \geq 0$. As $k$ is uncountable we can choose
$x,y$ an $M_i \oplus A$-superficial sequence for all $i \geq 0$. Set $J = (x,y)$. Then $I^{r+1} = JI^r$ for some $r \geq 0$. Let $\mathcal{R} = A[ut]$ be the Rees algebra of $A$ with respect to \ $I$ and let $\mathcal{M} $ be its unique maximal homogeneous ideal. We note that for all $i \geq 0$ we have an isomorphism of $\mathcal{R} $-modules
$H^i_\mathcal{M} (G_I(E)) \cong H^i_{G_I(A)_+}(G_I(E))$ for any finitely generated $A$-module $E$. The exact sequence of $\mathcal{R} $-modules $0 \rightarrow G_I(M_i) \rightarrow L^I_0(M_i) \rightarrow L^I_0(M_i)(-1) \rightarrow 0$ induces an exact sequence for all $n \in \mathbb{Z} $
\begin{align*}
H^0_\mathcal{M} (L^I_0(M_i))_{n-1} &\rightarrow H^1(G_I(M_i))_n \rightarrow H^1_\mathcal{M} (L^I_0(M_i))_n \rightarrow H^1_\mathcal{M} (L^I_0(M_i))_{n-1} \rightarrow
\\
&\rightarrow H^2(G_I(M_i))_n.
\end{align*}
As $H^2(G_I(M_i))_n = 0$ for $n \geq r-1$ we get that $H^I_\mathcal{M} (L^I_0(M_i))_n = 0$ for
$n \geq r - 2$.
Set $\rho^I(M_i) = \min\{ j \mid \widetilde {I^nM_i} = I^nM_i \ \text{for all} \ n \geq j \}$.
As $H^0_\mathcal{M} (L^I_0(M_i)) = \bigoplus_{n\geq 0}\widetilde {I^{n+1}M}/I^{n+1}M$, to prove our result it suffices to show that
\[
\left\{\frac{\rho^I(M_i)}{i^{\operatorname{cx}(M)-1}} \right\}_{i \geq 1 } \quad \text{is bounded.}
\]
We have nothing to show if $\operatorname{cx}(M) \leq 1$ for then either $M$ is free (if $\operatorname{cx}(M) = 0$) or$M$ has a periodic resolution with period two (if $\operatorname{cx}(M) = 1$). So assume $\operatorname{cx}(M) \geq 2$.
Set
\[
B^I(M_i) = \bigoplus_{n \geq 0} \frac{I^{n+1}M_i \colon x}{I^nM_i} = H_1(xu, L^I_0(M_i)).
\]
Set $\overline{M_i} = M_i/xM_i$.
By \cite[2.9]{Pu2} we have an exact sequence
\[
0 \rightarrow \frac{I^{n+1}M_i \colon x}{I^nM_i} \rightarrow \frac{\widetilde {I^nM_i}}{I^nM_i} \rightarrow
\frac{\widetilde {I^{n+1}M_i}}{I^{n+1}M_i} \rightarrow \frac{\widetilde {I^{n+1}\overline{M_i}}}{I^{n+1}\overline{M_i}}.
\]
By Lemma \ref{crucial} we get that $\widetilde {I^n\overline{M_i}} = I^n\overline{M_i}$ for all $n \geq r$. By the above exact sequence we get that
\[
\ell\left( \frac{\widetilde {I^rM_i}}{I^rM_i} \right) = \sum_{n \geq r} \ell \left( \frac{I^{n+1}M_i \colon x}{I^nM_i} \right) \leq \ell(B^I(M_i)), \text{and}
\]
\[
\ell\left(\frac{\widetilde {I^nM_i}}{I^nM_i} \right) \geq \ell\left( \frac{\widetilde {I^{n+1}M_i}}{I^{n+1}M_i} \right) \quad \text{for} \ n \geq r.
\]
Claim: If $n \geq r$ and $\widetilde {I^nM_i} \neq I^nM_i$ then we have a strict inequality
\[
\ell\left(\frac{\widetilde {I^nM_i}}{I^nM_i} \right) > \ell\left( \frac{\widetilde {I^{n+1}M_i}}{I^{n+1}M_i} \right)
\]
Proof of Claim: If the result does not hold then we have $(I^{n+1}M_i \colon x) = I^nM_i$. For all $m \geq 1$ we have an exact sequence
\[
\frac{(I^{m}M_i \colon x)}{I^{m-1}M_i} \xrightarrow{\beta_m} \frac{(I^{m+1}M_i \colon x)}{I^mM_i} \xrightarrow{\alpha_m} \frac{I^{m+1}M_i}{JI^mM_i} \xrightarrow{\rho_m} \frac{I^{m+1}\overline{M_i}}{yI^m\overline{M_i}}\rightarrow 0,
\]
(see proof of Lemma \ref{depth1}).
So if $n \geq r$ and $(I^{n+1}M_i \colon x) = I^nM_i$ then $(I^{m+1}M_i \colon x) = I^mM_i$ for all $m \geq n$. So we get
\[
\ell\left( \frac{\widetilde {I^nM_i}}{I^nM_i} \right) = \sum_{m \geq n} \ell \left( \frac{I^{m+1}M_i \colon x}{I^mM_i} \right) = 0, \quad \text{a contradiction}.
\]
It follows that
\[
\rho^I(M_i) \leq r + 1 + \ell(B^I(M_i)).
\]
We consider the following two cases.
Case I: $\operatorname{depth} G_I(A) > 0$. So $x^*$ is $G_I(A)$-regular. By proof of Theorem \ref{e2} the function $i \mapsto \ell(B^I(M_i))$ is of quasi-polynomial type with period two and degree $\leq \operatorname{cx}(M)-1$.
It follows that
\[
\left\{\frac{\rho^I(M_i)}{i^{\operatorname{cx}(M)-1}} \right\}_{i \geq 1 } \quad \text{is bounded.}
\]
Case II. $\operatorname{depth} G_I(A) = 0$.
We note that $\operatorname{depth} G_{I^m}(A) \geq 1$ for all $m \gg 0$. Choose $s$ such that
$\operatorname{depth} G_{I^s}(A) \geq 1$. For $n \geq r$ we have
\[
\ell\left(\frac{\widetilde {I^nM_i}}{I^nM_i} \right) \geq \ell\left( \frac{\widetilde {I^{n+1}M_i}}{I^{n+1}M_i} \right).
\]
It follows that
\[
\rho^I(M_i) \leq \max \{ r, s\rho^{I^s}(M_i) \}.
\]
By our Case I the result follows.
Now assume $d \geq 3$ and that the result holds when dimension of the ring is $= d -1$.
Let $x$ be $M_i \oplus A$-superficial for all $i \geq 0$. Set $\overline{M_i} = M_i/xM_i$.
Then notice $\overline{M}$ has finite GCI-dimension over $\overline{A}$ and $\overline{M_i} \cong \operatorname{Syz}^{\overline{A}}_i(\overline{M})$ for $i \geq 0$. We have exact sequences
\begin{align*}
0 \rightarrow U_i &\rightarrow G_I(M_i)(-1)\xrightarrow{x^*} G_I(M_i) \rightarrow G_I(M_i)/x^*G_I(M_i) \rightarrow 0 \ \ \text{and} \\
0\rightarrow V_i &\rightarrow G_I(M_i)/x^*G_I(M_i) \rightarrow G_I(\overline{M_i}) \rightarrow 0;
\end{align*}
where $U_i, V_i$ are $G_I(A)$-modules of finite length. As $d \geq 3$ we have an exact sequence for all $n \in \mathbb{Z} $
\[
H^{d-2}(G_I(\overline{M_i}))_n \rightarrow H^{d-1}(G_I(M_i))_{n-1} \rightarrow H^{d-1}(G_I(M_i))_{n}
\]
It follows that
\[
a_{d-1}(G_I(M_i)) \leq a_{d-2}(G_I(\overline{M_i})) - 1.
\]
The result now holds by induction hypothesis.
\end{proof}
|
\section{Introduction}
\label{sec:sec1}
Consider a multivariate dataset, given as an $n \times p$ data matrix
$X$, and suppose we want to explore the existence of any clusters.
One way to detect clusters is by projecting the data onto a lower
dimensional subspace for which the data are maximally non-normal.
Hence, methods that are sensitive to non-normality can be used to
detect clusters.
One set of methods based on this principle is invariant coordinate
selection (ICS), introduced by \citet{Tyler}, together with a
one-dimensional variant called projection pursuit (PP), introduced by
\citet{friedman1974}. ICS involves the use of two scatter matrices,
$S_1 = S_1(X)$ and $S_2 = S_2(X)$ with $S_2$ chosen to be more robust than
$S_1$.
An eigen-decomposition of $S_2^{-1}S_1$ is carried out. If the data
can be partitioned into two clusters, then typically the eigenvector
corresponding to the smallest eigenvalue is a good estimate of the
clustering direction. The main choice for the user when carrying out
ICS is the choice of the two scatter matrices.
However, in numerical experiments based on a simple mixture of two
bivariate normal distributions, some strange behaviour was noticed.
In certain circumstances, ICS, and its variant PP, badly failed to
pick out the right clustering direction. Eventually, it was
discovered that the cause was the use of different location measures
in the two scatter matrices. The purpose of this paper is to explore the
reasons for this strange behaviour in detail and to demonstrate the
benefits of using common location measures.
Section \ref{sec:sec2} gives some examples of scatter matrices and reviews the
use of ICS and PP as clustering methods. Section \ref{sec:sec3} sets out
the multivariate normal mixture model with two useful standardizations of the
coordinate system. Section \ref{sec:sec4} demonstrates in the population setting an ideal situation where
ICS and PP work as expected and where an analytic solution is
available --- the two-group normal mixture
model where the two scatter matrices are given by the covariance
matrix and a kurtosis-based matrix. Some examples with other
robust estimators are given in Sections \ref{sec:sec5}--\ref{sec:sec6}, which show how ICS and PP can go
wrong when different location measures are used and how the problem is fixed by using a common location measure.
{\em Notation.} Univariate random variables, and their realizations,
are denoted by lowercase letters, $x$, say. Multivariate random
vectors, and their realizations, are denoted by lowercase bold
letters, $\boldsymbol{x}$, say. A capital letter, $X$, say is used for
$n \times p$ data matrix containing $p$ variables or measurements on
$n$ observations; $X$ can be written in terms of its rows as
$$
X=\begin{pmatrix} \boldsymbol{x}_1^T\\ \vdots \\ \boldsymbol{x}_n^T
\end{pmatrix},
$$
with $i$th row $\/{x}_i^T=(x_{i1}, \ldots ,x_{ip}),\ i=1,
\ldots, n$.
\section{Background}
\label{sec:sec2}
\subsection{Scatter matrices}
A scatter matrix
$S(X)$, as a function of an $n \times p$ data matrix $X$ is a $p
\times p$ affine equivariant positive definite matrix. Following
\citet{Tyler}, it is convenient to classify scatter matrices into
three classes depending on their robustness.
\begin{itemize}
\item [(1)]Class I: is the class of non-robust scatter matrices with
zero breakdown point and unbounded influence function. Examples
include the covariance matrix defined below in \eqref{eq:var} and the
kurtosis-based matrix in \eqref{eq:kmat}.
\item [(2)] Class II: is the class of scatter matrices that are
locally robust, in the sense that they have bounded influence
function and positive breakdown points not greater than
$\frac{1}{p+1}$. An example from this class is the class of
multivariate M-estimators, such as the M-estimate for the
$t$-distribution \citep[e.g., ][]{kc94a,arslan1995convergence}.
\item [(3)] Class III: is the class of scatter matrices with high breakdown
points such as the Stahel-Donoho estimate, the minimum volume
ellipsoid (mve) \citep{van2009minimum} and the constrained M-estimates,
\citep[e.g., ][]{kent1996constrained}.
\end{itemize}
Each scatter matrix has an implicit location measure. Let
us look at the main examples in more detail, and note what happens in
$p=1$ dimension. The labels in parentheses are used as part of the
notation later in the paper.
The sample covariance matrix (var) is defined by
\begin{equation}
S = \frac{1}{n} \sum_{i=1}^n (\boldsymbol{x}_i -
\bar{\boldsymbol{x}})(\boldsymbol{x}_i - \bar{\boldsymbol{x}})^T,
\label{eq:var}
\end{equation}
where for convenience here a divisor of $1/n$ is used, and where
$\bar{\boldsymbol{x}}$ is the sample mean vector. The implicit
measure of location is just the sample mean.
The kurtosis-based matrix (kmat) is defined by
\begin{equation}
K = \frac{1}{n}
\sum_{i=1}^n \{(\boldsymbol{x}_i - \bar{\boldsymbol{x}})^T S^{-1}
(\boldsymbol{x}_i - \bar{\boldsymbol{x}})\}
(\boldsymbol{x}_i - \bar{\boldsymbol{x}})(\boldsymbol{x}_i -
\bar{\boldsymbol{x}})^T. \label{eq:kmat}
\end{equation}
Note that outlying observations are given higher weight than for the
covariance matrix, so that $K$ is less robust than $S$. Again the
implicit measure of location is just the sample mean. When $p=1$,
the scatter matrix $S^{-1}K$ reduces to 3 plus the usual univariate kurtosis.
The $M$-estimator of scatter based on the multivariate
$t_\nu$-distribution for fixed $\nu$ is the maximum likelihood
estimate obtained by maximizing the likelihood jointly over scatter
matrix $\Sigma$ and location vector $\/\mu$. If both parameters are
unknown and $\nu \geq 1$, then under mild conditions on the data, the mle of
$(\boldsymbol{\mu}, \Sigma)$, is is the unique stationary point of the
likelihood. Similarly, if $\nu \geq 0$ and $\/\mu$ is known, the mle of
$\Sigma$, is is the unique stationary point of the likelihood
\citep{kc94a}. In either case, an iterative
numerical algorithm is needed. Note that when $\/\mu$ is to be
estimated as well as $\Sigma$, the mle of $\mu$ is the
implicit measure of location for this scatter matrix. For this paper
we limit attention to the choice $\nu=2$ (and label it below by t2).
The minimum volume ellipsoid (mve) estimate of scatter
$S_{\text{mve}}$, introduced by \citet{rousseeuw1985}, is the
ellipsoid that has the minimum volume among all ellipsoids containing
at least half of observations, and its implicit estimate of location,
$\bar {\boldsymbol{x}}_{\text{mve}}$, say, is the centre of that
ellipsoid. Calculating the exact mve requires extensive computation.
In practice, it is calculated approximately by considering only a
subset of all subsamples that contain $50\%$ of the observations,
\citep[e.g., ][]{van2009minimum, maronna2006}. If the location vector is
specified, the search is limited to ellipsoids centred at this
location measure.
When $p=1$, the mve reduces to the lshorth, defined as the length of
the shortest interval that contains at least half of observations.
The corresponding estimate of location,
$\bar{\boldsymbol{x}}_\text{lshorth}$, say, is the midpoint of this
interval. Calculating the lshorth around a known measure of location
is trivial; just find the length of the interval that contains half of
observations centered at this location measure. The lshorth was
introduced by \cite{grubel1988length}, building on earlier suggestion
of \citet{andrews1972robust} to use
$\bar{\boldsymbol{x}}_\text{lshorth}$, which they called the shorth,
as a location measure.
The minimum covariance determinant estimate of scatter (mcd),
$S_{\text{mcd}}$ is defined as the covariance matrix of half of
observations with the smallest determinant. The mcd location measure,
$\bar{\boldsymbol{x}}_{\text{mcd}}$, say, is the sample mean of those
observations. The mcd can be calculated approximately by considering
only a subset of all subsamples that contain at least half of
observations, \citep[e.g., ][]{rousseeuw1999fast}. The mcd estimate of
scatter with respect to a known location measure $\boldsymbol{\mu}$ is
defined as the covariance matrix about $\boldsymbol{\mu}$ of half of
observations with the smallest determinant. Recall that the
covariance matrix about $\boldsymbol{\mu}$ for a dataset is given by
$S +
(\boldsymbol{\mu}-\bar{\boldsymbol{x}})(\boldsymbol{\mu}-\bar{\boldsymbol{x}})^T$,
where $S$ and $\bar{\boldsymbol{x}}$ are the sample covariance matrix
and mean vector of the dataset.
When $p=1$, the mcd reduces to a truncated variance, $v_{\text{trunc}}$, say, defined as
the smallest variance of half the observations. Its implicit measure
of location, $\bar{x}_{\text{trunc}}$, say, is the sample mean of that interval. Also, a
modified definition of $v_{\text{trunc}}$ using a known location measure is
trivial and does not require any search; just find the interval that
contains half of observations centered at the given location measure
and calculate the variance.
Routines are available in \texttt{R} \citep{R} to compute (at least
approximately) these robust covariance matrices and their implicit
location measures, in particular, \texttt{tM} from the package
\texttt{ICS} \citep{nordhausen2008} for the multivariate
$t$-distribution, \texttt{cov.rob} from the package \texttt{MASS}
\citep{venables2010package} for mve, and \texttt{CovMcd} from the
package \texttt{rrcov} \citep{todorov2008package} for mcd. Modified
versions of these routines have been written by us to deal with the
case of known location measures.
\subsection{Invariant coordinate selection and projection pursuit}
\label{sec:sec22}
Given an $n \times p$ data matrix $X$, the ICS objective function is
given by the ratio of quadratic forms
\begin{equation}
\kappa_{\text{ICS}}(\boldsymbol{a})=
\frac{\boldsymbol{a}^TS_1\boldsymbol{a}}
{\boldsymbol{a}^TS_2 \boldsymbol{a}},
\quad \boldsymbol{a} \in \mathbb{R}^p,
\label{eq:kics}
\end{equation}
where $S_1=S_1(X)$ and $S_2=S_2(X)$ are two scatter matrices. By
convention, $S_2$ is chosen to be more robust than $S_1$. For
exploratory statistical analysis, attention is focused on the choices
for $\boldsymbol{a}$ maximizing or minimizing
$\kappa_{\text{ICS}}(\boldsymbol{a})$. These values can be calculated
analytically as the eigenvectors of $S_2^{-1}S_1$ corresponding to the
maximum/minimum eigenvalues.
The original ICS method did not make a strong distinction between the
largest and the smallest eigenvalues. However for clustering purposes between
two groups, when the mixing proportion is not too far from $1/2$, it is the
minimum eigenvalue which is of interest; see Section \ref{sec:sec4}.
The method of PP can be regarded as a one-dimensional version of ICS.
It looks for a linear projection $\boldsymbol{a}$ to maximize or
minimize the criterion,
\begin{equation}
\kappa_{\text{PP}}(\boldsymbol{a})=\frac{s_1(X\boldsymbol{a})}
{s_2(X\boldsymbol{a})}. \label{eq:kpp}
\end{equation}
where $s_1=s_1(X\boldsymbol{a})$ and $s_2=s_2(X\boldsymbol{a})$ are
two one-dimensional measures of spread. In general, optimizing
$\kappa_{\text{PP}}(\boldsymbol{a})$ must be carried out
numerically. Searching for a global optimum is computationally
expensive, and the complexity of the search increases as the dimension
$p$ increases. Alternatively, we can search for a local optimum
starting from a sensible initial solution, such as the ICS optimum
direction.
Both ICS and PP are equivariant under affine transformations. That
is, if $X$ is transformed to $U= \boldsymbol{1}_n \boldsymbol{h}^T +
XQ^T$, where $Q(p \times p)$ is nonsingular and $\boldsymbol{h}$ is a
translation vector in $\mathbb{R}^p$, then for either ICS or PP the
new optimal vector $\boldsymbol{b}$, say, for $U$ is related to
the corresponding optimal vector $\boldsymbol{a}$ for $X$ by
\begin{equation}
\boldsymbol{b} \propto Q^{-T}\boldsymbol{a}. \label{eq:scale}
\end{equation}
For numerical work it is convenient to have an explicit notation for the
different choices in ICS and PP. If Scat1 and Scat2 are the names of
two types of multivariate scatter matrix, each computed with its own implicit
location measure, then the corresponding versions of ICS and PP will be
denoted
\begin{equation}
\text{ICS}:\text{Scat1}:\text{Scat2}, \quad \text{and} \quad
\text{PP}:\text{Scat1}:\text{Scat2}. \nonumber
\end{equation}
Note that PP is based on the univariate versions of Scat1 and Scat2.
For example, ICS based on the covariance matrix and the minimum volume
ellipsoid will be denoted by ICS:var:mve. Other choices for scatter
matrices have been summarized in Section \ref{sec:sec2}.
When a common location measure is imposed on Scat1 and Scat2, then
this restriction will be indicated by the augmented notation
\begin{equation}
\text{ICS}:\text{Scat1}:\text{Scat2}:\text{Loc}, \nonumber
\end{equation}
and similarly for PP. In this paper the only choice used for the
location measure is the sample mean (mean). For example, ICS based on the
covariance matrix and the minimum volume ellipsoid, both computed
with respect to the mean vector, is denoted
$$
\text{ICS}:\text{var}:\text{mve}:\text{mean}.
$$
\section{The two-group multivariate normal mixture model}
\label{sec:sec3}
The simple model used to demonstrate the main points of this paper is the
two group multivariate normal mixture model, with density
\begin{equation}
f(\boldsymbol{x}) = q \phi_p(\boldsymbol{x}, \boldsymbol{\mu}_1, \Omega) +
(1-q) \phi_p(\boldsymbol{x}, \boldsymbol{\mu_2}, \Omega), \nonumber
\end{equation}
where $\phi_p$ is the multivariate normal density, $\boldsymbol{\mu}_1$ and $\boldsymbol{\mu}_2$
are two mean vectors, $\Omega$ is a common covariance matrix, and
$0<q<1$ is the mixing proportion. Even in this simple case, major
problems with ICS and PP can arise.
Since ICS and PP are affine equivariant, we may without loss of
generality choose the coordinate system so that
$$
\boldsymbol{\mu_1} = \alpha \boldsymbol{e_1}, \quad \boldsymbol{\mu_2} = -\alpha \boldsymbol{e_1}, \quad \Omega = I_p,
$$
where $\boldsymbol{e_1} = (1,0,\ldots, 0)^T$ is a unit vector along
the first coordinate axis, and $\alpha>0$. That is,
$\boldsymbol{\mu_1}$ and $\boldsymbol{\mu_2}$ lie equally spaced about
the origin along the first coordinate axis, and the covariance matrix
of each component equals the identity matrix.
A random vector $\/x$ from the mixture model can also be given a
stochastic representation,
\begin{equation}
\boldsymbol{x} = \alpha s \boldsymbol{e}_1 + \boldsymbol{\epsilon}, \nonumber
\end{equation}
where $\boldsymbol{\epsilon} \sim N_p(0,I_p)$ independently of an indicator variable $s$,
\begin{equation}
s=\left\{
\begin{array}{rl}
1& \text{ with probability } q\\
-1& \text{ with probability } (1-q)\\
\end{array} \right. \nonumber.
\end{equation}
Moments under the mixture model are calculated most simply in terms of this
stochastic representation. In particular,
\begin{equation}
\boldsymbol{\mu}_x= E(\boldsymbol{x}) = q \boldsymbol{\mu}_1+ (1-q) \boldsymbol{\mu}_2=(2q-1)\alpha \boldsymbol{e}_1, \quad
E(\boldsymbol{xx}^T) = \alpha^2 \boldsymbol{e}_1\boldsymbol{e}_1^T + I_p, \nonumber
\end{equation}
so that the covariance matrix is
\begin{equation}
\label{eq:sigx}
\Sigma_x=\text{var}(\boldsymbol{x}) = E(\boldsymbol{xx}^T) - \boldsymbol{\mu}_x \boldsymbol{\mu}_x^T =
4q(1-q)\alpha^2 \boldsymbol{e}_1 \boldsymbol{e}_1^T + I_p.
\end{equation}
For practical work it is also convenient to consider a
standardization for which the overall covariance matrix is the
identity matrix. That is, define a new random vector
\begin{equation}
\boldsymbol{y} = C^{-1}\boldsymbol{x}, \label{eq:y}
\end{equation}
where $C^{-1} = \text{diag}(1/c_1,\ldots, 1/c_p)$, where $c_1 = \{1+4
q(1-q)\alpha^2\}^{1/2}$, and $c_2 = \cdots = c_p = 1$. Then $\boldsymbol{y}$
has a stochastic representation
$$
\boldsymbol{y} = \delta s \boldsymbol{e}_1 + \boldsymbol{\eta},
$$
where
\begin{equation}
\label{eq:delta}
\delta = \alpha / \{1+4q(1-q)\alpha^2 \}^{1/2},
\end{equation}
and
\begin{equation}
\boldsymbol{\eta }\sim N_p(0, \text{diag}(\sigma^2_\eta, 1, \ldots, 1)) \nonumber
\end{equation}
where the first diagonal term $\sigma^2_\eta$ has two equivalent formulas,
$$
\sigma^2_\eta=\{1+4 \alpha^2 q(1-q)\}^{-1} \quad \text{or} \quad
\sigma^2_\eta=1-4q(1-q)\delta^2
$$
The first two moments of $\boldsymbol{y}$ are
$$
\boldsymbol{\mu}_y = (2q-1)\delta \boldsymbol{e}_1, \quad \Sigma_y = I_p.
$$
\section{A population example: PP based on the kurtosis and ICS based on the kurtosis-based matrix and the covariance matrix}
\label{sec:sec4}
In this section we look at ICS:kmat:var and PP:kmat:var in the
population case. In this setting it is possible
to derive analytic results. Note that since kmat is based on fourth
moments it is less robust than the variance matrix; hence kmat is
listed first.
Recall the kurtosis of a univariate random variable $u$, say, with mean
$\mu_u$, is defined by
\begin{equation}
\text{kurt}(u)=\frac{\text{E}\{(u-\mu_u)^4\}}{\left[\text{E}\{(u-\mu_u)^2\}\right]^2}-3. \nonumber
\end{equation}
The univariate kurtosis is zero when the random variable has normal
distribution. For non-normal distributions the kurtosis lies in the
interval $[-2, \infty]$ and is often nonzero.
\citet{pena2001} studied the population version of PP:kmat:var and
showed that when the mixing proportion is not too far from $1/2$ (more
precisely, if $q(1-q)>1/6$, i.e. $0.21 < q <0.79$), then minimizing the
PP objective function picks out the correct clustering direction.
Their result can be derived simply as follows. Let $\/a \in
\mathbb{R}^p$ be a unit vector. Write $\/a^T\/x = \alpha a_1 s + v$,
where $v = \/{a}^T \/{\epsilon} \sim N(0,1)$ is independent of $s$.
The moments of $s$ are E$(s)=$E$(s^3)=m$, say, where
\begin{equation}
\label{eq:c}
m= 2q-1,
\end{equation}
and E$(s^2)=$ E$(s^4) = 1$. Hence, var$(s)= \sigma^2$, say, where
\begin{equation}
\label{eq:sigma}
\sigma^2= 4q(1-q).
\end{equation}
Then
$$
\text{kurt}(s) = - 6 + 4/\sigma^2.
$$
It can be checked that $\text{kurt}(s)<0$ provided $q(1-q)>1/6$.
Next, we use the property that if $u_1, u_2$ are independent random
variables with the same variance, and if $\delta_1, \delta_2$ are
coefficients satisfying $\delta_1^2 + \delta_2^2 = 1$, then
$$
\text{kurt}(\delta_1 u_1 + \delta_2 u_2 ) =
\delta_1^4 \text{kurt}(u_1) + \delta_2^4 \text{kurt}(u_2).
$$
Applying this result to $\/a^T\/x$ yields
\begin{equation}
\label{eq:PP}
\text{kurt}(\/a^T\/x) =\frac{a_1^4 \alpha^4 \sigma^4}{(\alpha^2 a_1^2 \sigma^2+1)^2}
\text{kurt}(s).
\end{equation}
Provided $\text{kurt}(s)<0$, \eqref{eq:PP} is minimized when $a_1^2$ is
maximized, that is, if $a_1^2=1$, so that $\/a = \pm e_1$ picks out the
first coordinate axis.
The ICS calculations proceed similarly. First note that, the first diagonal term in
$\Sigma_x$, defined in \eqref{eq:sigx}, can be expressed in terms of $\sigma^2$, defined in \eqref{eq:sigma}, as
$\alpha^2 \sigma^2 + 1$.
The first factor in the population version of $K$ defined in \eqref{eq:kmat}, $K_x$, say, is given by
\begin{equation}
(\boldsymbol{x}-\boldsymbol{\mu}_x)^T \Sigma_{\boldsymbol{x}}^{-1}(\boldsymbol{x}-\boldsymbol{\mu}_x)=
\frac{(x_1-\alpha m)^2}{1+\alpha^2 \sigma^2}+x_2^2 + \cdots + x_p^2 = D^2, \text{ say,}
\nonumber
\end{equation}
where $m$ is defined in \eqref{eq:c}.
Note that $D^2$ is an even function in $x_2, \ldots, x_p$. Hence by symmetry
all the off-diagonal terms in $K_x$ vanish. The first diagonal term is
given by
\begin{equation}
E\{D^2 (x_1 - \alpha m)^2\} = (1+\alpha^2 \sigma^2)(p+2)+
\frac{\alpha^4 \sigma^4 \text{kurt}(s)}{(1+\alpha^2 \sigma^2)}. \nonumber
\end{equation}
The remaining diagonal terms, $j=2, \ldots, p$ are given by
\begin{equation}
E\{D^2 x_j^2\} = p+2. \nonumber
\end{equation}
Hence $\Sigma_x^{-1} K_x$ reduces to
$$
\text{diag}(p+2+\frac{\text{kurt}(s) \alpha^4 \sigma^4}{(1+\alpha^2 \sigma^2)}, p+2, \ldots, p+2).
$$
These diagonal values are the eigenvalues. Hence provided $\text{kurt}(s)<0$,
$\kappa_{\text{ICS}}$ is minimized when $\boldsymbol{a} = \boldsymbol{e}_1$, that is, when $\boldsymbol{a}$ picks out the clustering
direction.
If $p=2$, we can write a unit vector as $\boldsymbol{a} =
(\cos \theta, \sin \theta)^T$, and since $\/a$ and $-\boldsymbol{a}$
define the same axis, we can parameterize the ICS and PP objective
functions in terms of $\theta, \ -\pi/2 \leq \theta \leq \pi/2$.
Plots of $\kappa_\text{ICS}(\theta)$ and $\kappa_\text{PP}(\theta)$ for $\alpha=3$
and $q=1/2$ are shown in Figure \ref{fig:fig1}.
\begin{figure}[h!]
\centering
\includegraphics[width=6cm,height=6cm]{pop_k_kurt_a3.pdf}
\caption{Plot of the population criteria $\kappa_{\text{ICS}}(\theta)$ (red dotted line), and $\kappa_{\text{PP}}(\theta)$ (solid black line) versus $\theta$, for $q=1/2$, $\alpha=3$.}
\label{fig:fig1}
\end{figure}
For numerical work, especially when the underlying mixture model is
unknown, the only feasible standardization is to ensure the overall
variance matrix $\Sigma_y$ is the identity rather than the within
group variance matrix. In terms of the population model of this
section, it means working with $\/y$ from \eqref{eq:y} rather than
$\/x$. If $p=2$ and $\/b \propto (\cos \phi, \sin \phi)^T$, say, is
also written in polar coordinates, then from \eqref{eq:scale} and
\eqref{eq:y} $\/a$ and $\/b$ are related by
$$
\/b \propto C \/a,
$$
hence, $\phi$ and $\theta$ are related by
\begin{equation}
\begin{pmatrix} \cos \phi \\ \sin \phi \end{pmatrix} \propto \begin{pmatrix} c_1 & 0 \\ 0& c_2 \end{pmatrix}
\begin{pmatrix} \cos \theta \\ \sin \theta \end{pmatrix}. \nonumber
\end{equation}
Thus,
$$
\tan \phi = c \tan \theta,
$$
where $c=c_2/c_1$.
The plot of the ICS and PP objective functions in Figure
\ref{fig:fig2} shows that there is a sharper minimum in $\phi$ coordinates than in $\theta$ coordinates because
under our mixture model $c$ is less than 1. If $\/x$ is scaled as in \eqref{eq:y} with $c_1>c_2$, i.e $c>1$, then there
will be a wider minimum in $\phi$.
\begin{figure}[h!]
\centering
\includegraphics[width=6cm,height=6cm]{pops_k_kurt_d95.pdf}
\caption{Plot of the population criteria $\kappa_{\text{ICS}}(\phi)$ (red dotted line), and $\kappa_{\text{PP}}(\phi)$ (solid black line) versus $\phi$, for $q=1/2$, and $\delta=0.95$.}
\label{fig:fig2}
\end{figure}
\section{The effect of using a common location measure
on ICS and PP}
\label{sec:sec5}
As we mentioned earlier in Section \ref{sec:sec22}, the ICS and PP
criteria are expected to have
similar behaviour to the kurtosis-based criteria in Section
\ref{sec:sec4}. Namely, they are expected to be minimized in the
clustering direction when the mixing proportion is not too far from 1/2.
However, when applying ICS with at least one robust estimate of
scatter (mainly from Class III), some peculiar behaviour was observed.
In particular, the ICS criterion was often maximized in the clustering
direction rather than minimized.
Here is an explanation. Under the two-group mixture model with one
group slightly bigger than the other, a class III scatter matrix will
typically home in on the larger group, with its corresponding location
measure at the center of this group and its estimate of the scatter
matrix capturing the spread of this group. The other scatter matrix
(Class I or II) will measure the overall scatter of the data with its
corresponding location measure at the overall center of the data. The
result is erratic behaviour in $\kappa_\text{ICS}$ and $\kappa_\text{PP}$.
Imposing a common location measure on the two scatter matrices fixes
this problem. Here is an example in $p=2$ dimensions to
illustrate the issues in greater detail.
In this example we look at ICS:var:mve for the
population bivariate normal mixture model in Section \ref{sec:sec3},
with $q=1/2$ and any value of $\alpha>0$, i.e. $0 \leq \delta \leq 1$, where $\delta$ is given in \eqref{eq:delta}.
Standardize the coordinate
system so that the overall covariance matrix is the identity,
$\Sigma_y=I_2$. Let $\Sigma_{\text{mve}}$ denote the population
minimum volume ellipsoid scatter matrix.
Then it turns out that $\Sigma_{\text{mve}}$ is the within-group
covariance matrix for (either) one of the groups,
\begin{equation}
\label{eq:mvenc}
\Sigma_{\text{mve}}=\begin{pmatrix}1-\delta^2 &0 \\ 0&1 \end{pmatrix},
\end{equation}
where $0 \leq \delta \leq 1$ is given in (\ref{eq:delta}). The
implicit estimate of the center of the data will be given by the
center of either group, $\pm \delta \boldsymbol{e}_1$; both values fit equally
well.
\begin{figure}[h!]
\centering
\includegraphics[width=6cm,height=6cm]{mve_pop.pdf}
\caption{Plot of the population criterion of ICS:var:mve vs. $\phi$ for $\delta=0.9$.}
\label{fig:fig3}
\end{figure}
Figure \ref{fig:fig3} shows that the clustering direction estimated by the ICS:var:mve method is the direction
that minimizes $\kappa_{\text{ICS}}$ (the eigenvector of the smallest eigenvalue of $\Sigma_{\text{mve}}^{-1}$),
namely $(0,1)^T$, i.e. $\phi=\pm \pi/2$. However, the true direction of group separation
direction is $(1,0)^T$, i.e. $\phi=0$.
Next consider ICS:var:mve:mean, i.e. the common mean version of the previous example. The overall mean of the data
is at
the origin. When $\Sigma_{\text{mve}}$ is constrained to have its
location measure at the origin, then the ICS criterion now picks out the true
clustering direction. In order to give an analytic proof of this result,
we restrict attention to the the limiting case of the balanced mixture model, i.e
when $\delta=1,\ q=1/2$.
Hence, the group components will lie on two parallel vertical lines with means
\begin{equation}
\mu_1=(1,0)^T, \ \ \mu_2=(-1,0)^T, \nonumber
\end{equation}
and within-group covariance matrix
\begin{equation}
\begin{pmatrix}0 & 0 \\0 & 1 \end{pmatrix}. \nonumber
\end{equation}
In this setting, it can be shown that the population version of the MVE matrix,
$\Sigma_{\text{mve}}$, say, takes the form
\begin{equation}
\Sigma_{\text{mve}}=c_t \Sigma_t= \begin{pmatrix}2 &0 \\ 0& 2d^2 \end{pmatrix},
\nonumber
\end{equation}
where $d = \Phi^{-1}(.75) = 0.674$, the 75th quantile of the standard
normal distribution (see the Appendix). Hence the dominant
eigenvector is $\/e_1$. The ellipse of $\Sigma_{\text{mve}}$ is
plotted in Figure \ref{fig:mve}. Figure \ref{fig:fig4} shows that the
criterion of ICS:var:mve:mean, $\kappa_{\text{ICS}:\mu}(\phi)$ picks out the
correct clustering direction $\/e_1$.
\begin{figure}[h!]
\centering
\includegraphics[width=9cm,height=9cm]{pop_mve_m3.pdf}
\caption{Plot of the ellipse of the constrained $\Sigma_{\text{mve}}$.}
\label{fig:mve}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=6cm,height=6cm]{mve_pop_mean.pdf}
\caption{Plot of the population criteria of ICS:var:mve:mean, $\kappa_{\text{ICS}:\mu}(\phi)$.}
\label{fig:fig4}
\end{figure}
Like ICS, PP can fail to detect the clustering direction if applied
using different location measures. The reason for that is the
projection direction that separates the data into two groups with one
slightly bigger than the other, the more robust measure of spread will
be located at the larger group. In Section \ref{sec:sec6}, we give a
detailed numerical example of the problem arising from using two
different location measures in PP:var:mcd, and how the problem is
fixed by using a common location measure.
\section{Examples}
\label{sec:sec6}
\subsection*{Overview}
In this section, we give numerical examples that demonstrate different
ways in which ICS and/or PP can go wrong. We also show the effect of
using common location measures in these examples. We use one
simulated data set and apply different ICS and PP methods, with and
without imposing a common location measure (the mean).
A two-dimensional data set of size $n=500$ is generated from the
balanced mixture model, defined in Section \ref{sec:sec3}, with
$q=1/2$, and $\alpha=3$, so that $\delta=0.95$. Thus the two groups
are well-separated and no sensible statistical method should have any
problem finding the two clusters. All calculations are done after
standardization with respect to the ``total'' coordinates. That is,
the data matrix $Y (500 \times 2)$ is standardized to have sample mean
$\/0$ and sample covariance matrix $I_2$.
The ICS and PP methods used are:
\begin{itemize}
\item[(1)] (PP,ICS):var:t2 with corresponding criteria $\kappa_{\text{ICS}}^{1}$,
and $\kappa_{\text{ICS}}^{1}$.
\item[(2)](PP,ICS):var:mcd with corresponding criteria $\kappa_{\text{ICS}}^{2}$, and $\kappa_{\text{PP}}^{2}$.
\item[(3)](PP,ICS):var:mve with corresponding criteria $\kappa_{\text{ICS}}^{3}$, and $\kappa_{\text{PP}}^{3}$.
\item[(4)](PP,ICS):t2:mcd with corresponding criteria $\kappa_{\text{ICS}}^{4}$, and $\kappa_{\text{PP}}^{4}$.
\item[(5)](PP,ICS):t2:mve with corresponding criteria $\kappa_{\text{ICS}}^{5}$, and $\kappa_{\text{PP}}^{5}$.
\end{itemize}
When imposing the mean as the common location measure, the ICS and PP criteria will be denoted by $\kappa_{\text{ICS:mean}}^{j}$
and $\kappa_{\text{PP:mean}}^{j}$, where $j=1, \ldots, 5$.
To understand the behaviour of the ICS and PP, their criteria are plotted against $-\pi/2\leq\phi\leq\pi/2$.
The plots are shown in Figure \ref{fig:fig5}.
\newpage
\begin{figure}[t!]
\centering
\subfloat[(PP,ICS):var:t2]{%
\includegraphics[width=5cm,height=5cm]{criterion1.pdf}%
}%
\subfloat[(PP,ICS):var:t2:var]{%
\includegraphics[width=5cm,height=5cm]{criterion1_mean.pdf}}
\\
\centering
\subfloat[(PP,ICS):var:mcd]{\includegraphics[width=5cm,height=5cm]{criterion2.pdf}}%
\subfloat[(PP,ICS):var:mcd:mean]{\includegraphics[width=5cm,height=5cm]{criterion2_mean.pdf}}
\\
\centering
\subfloat[(PP,ICS):var:mve]{\includegraphics[width=5cm,height=5cm]{criterion3.pdf}}%
\subfloat[(PP,ICS):var:mve:mean]{\includegraphics[width=5cm,height=5cm]{criterion3_mean.pdf}}
\caption{}
\end{figure}
\newpage
\begin{figure}
\ContinuedFloat
\centering
\subfloat[(PP,ICS):t2:mcd]{\includegraphics[width=5cm,height=5cm]{criterion4.pdf}}%
\subfloat[(PP,ICS):t2:mcd:mean]{\includegraphics[width=5cm,height=5cm]{criterion4_mean_2.pdf}}
\\
\centering
\subfloat[(PP,ICS):t2:mve]{\includegraphics[width=5cm,height=5cm]{criterion5.pdf}}%
\subfloat[(PP,ICS):t2:mve:mean]{\includegraphics[width=5cm,height=5cm]{criterion5_mean_2.pdf}}
\caption{For $\delta=0.95$ and $q=1/2$, plots of different ICS (red dashed curve) and PP (black solid curve) criteria without (left) and with imposing a common location measure (right).}
\label{fig:fig5}
\end{figure}
From the panels in Figure \ref{fig:fig5}, we make the following remarks based on the simulated data set:
\begin{itemize}
\item[(1)] Panel (a) shows that ICS:var:t2 and PP:var:t2 work well
since $\bar{\/y}$ and $\bar{\/y_{t2}}$ are approximately equal. Hence,
imposing a common location measure has little effect, as shown in
(b).
\item[(2)] Panels (c), (e), (g), (i) show examples when ICS and/or PP
go wrong because of the difference in the location measures.
\item[(3)] Using a common location measure fixes the problem in panel
(d) for (PP, ICS):var:mcd, panel (f) for (PP, ICS):var:mve, and
panel(h) for (PP, ICS):t2:mcd.
\item[(4)] From panel (j), using a common location measure in
PP:t2:mve:mean does not seem to work well. The reason might
be due to the unstable behaviour of the mve and lshorth.
\item[(5)] The plots generally suggest that PP will be more accurate than ICS,
since the PP plots are narrower at the clustering direction than the
ICS plot. This property has been confirmed empirically in \citet{Fatimah}
for certain multivariate normal mixture models and choices of scatter matrix.
\item[(6)] Similar patterns are seen with most simulated data sets
from this model.
\end{itemize}
\subsection*{Behaviour of ICS:var:mcd}
To gain a deeper understanding of the behaviour of ICS:var:mcd in
panel \ref{fig:fig5} (c) and the effect of forcing a common location
measure on mcd in panel (d), we plot the ellipses of $S_{\text{mcd}}$
( with and without imposing a common location meaure) and superimpose
it on the data points of our example. The plots are shown in panels
\ref{fig:fig6} (a) and (b). The behaviour in this example agrees with
the interpretation given for the population example in Section
\ref{sec:sec5}.
\begin{figure}[H]
\centering
\subfloat[]
{\includegraphics[width=6cm,height=6cm]{mcd_plot2.pdf}}
\subfloat[]
{\includegraphics[width=6cm,height=6cm]{mcd_plot.pdf}}
\caption{Plots of the ellipses of mcd scatter matrix based on (a) mcd location measure, and (b) the sample mean, superimposed on data of size $n=500$, distributed as mixtures of two normal distributions.}
\label{fig:fig6}
\end{figure}
\subsection*{Behaviour of PP:var:mcd}
The objective function for PP:var:mcd, has a similar problem to ICS;
it is maximized rather than minimized near the
correct clustering direction.
To understand this behaviour in more detail, we plot in Figure \ref{fig:fig8}
one-dimensional histograms after projections by the following choices
for the angle $\phi$: $0^\circ$, $15^\circ$, $30^\circ$, and
$90^\circ$. For each histogram, we plot the $50\%$ of the data that
has the smallest variance, and the corresponding location measure
$\bar{x}_{\text{trunc}}$. The plots are repeated where the location measure is
constrained at the sample mean $\bar x=0$.
\newpage
\begin{figure}[h!]
\centering
\subfloat[ $v_{\text{trunc}}=0.08$, $\bar{x}_{\text{trunc}}= -0.93$]
{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_0.pdf}}
\subfloat[ $v_{\text{trunc}}=0.54$, $\bar x=0$ ]{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_m0.pdf}}
\\[-3ex]
\subfloat[ $v_{\text{trunc}}=0.14$, $\bar{x}_{\text{trunc}}=-0.92 $]{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_15.pdf}}
\subfloat[$v_{\text{trunc}}=0.42$, $\bar x= 0$ ]{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_m15.pdf}}
\\[-3ex]
\subfloat[ $v_{\text{trunc}}=0.21 $, $\bar{x}_{\text{trunc}}=0.63$]{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_30.pdf}}
\subfloat[ $v_{\text{trunc}}= 0.26$, $\bar{x}_{\text{trunc}}= 0.02$]{\includegraphics[width=0.4 \columnwidth, height=4cm ]{pp_mcd_m30.pdf}}
\\[-3ex]
\subfloat[ $v_{\text{trunc}}=0.12 $, $\bar{x}_{\text{trunc}}=0.18 $]{\includegraphics[width=0.4 \columnwidth, height=4cm]{pp_mcd_90.pdf}}
\subfloat[ $v_{\text{trunc}}=0.14 $, $\bar{x}_{\text{trunc}}= 0$]{\includegraphics[width=0.4 \columnwidth, height=4cm]{pp_mcd_m90.pdf}}
\caption{Histograms of $0^\circ, 15^\circ, 30^\circ$ and $90^\circ$ projections. Left panels show
the vectors of $50\%$ of data with the smallest variance (the blue lines), and its
location measure (the red lines), right panels show the $50\%$ of data with the smallest
variance computed around the mean 0.}
\label{fig:fig8}
\end{figure}
The shape of the histograms depends on of the projection directions. Note that
as $v_{\text{trunc}}$ gets smaller, the PP criterion $\kappa_{\text{PP}}$ gets larger.
\begin{itemize}
\item[(1)] The $0^\circ$ projection produces two widely separated
groups with one group is slightly bigger than the other. In this
case, $\bar{x}_{\text{trunc}}$ is at the larger group and $v_{\text{trunc}}$ is essentially
the variance of this group. Hence $v_{\text{trunc}}$ takes its smallest value
and $\kappa_{\text{PP}}$ is largest.
\item[(2)] The $15^\circ$ projection produces two slightly separated
groups with within-group variance is larger than in the $0^\circ$
projection. The value of $v_{\text{trunc}}$ is larger than for $0^\circ$.
\item[(3)] The $30^\circ$ projection produces one group, with a
pseudo-uniform distribution. The value of $v_{\text{trunc}}$ is larger than for
$15^\circ$.
\item[(4)] The $90^\circ$ projection produces one normally distributed
group. The value for $v_{\text{trunc}}$ becomes small again.
\end{itemize}
Constraining the mean to be at the origin fixes the problem. The value of
$v_{\text{trunc}}$ steadily decreases from $0^\circ$ to $90^\circ$.
\section*{Appendix}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
Consider the limiting balanced bivariate normal mixture model,
$$
\/y = s \/e_1 + z \/e_2,
$$
where $s = \pm 1$, each with probability 1/2, independent of $z \sim
N(0,1)$, and $\/e_1 = (1,0)^T$, $\/e_2 = (0,1)^T$. This model is
standardized with respect to the ``total'' coordinates; i.e. $E(\/y) =
\/0$ and $\text{var}(\/y) = I_2$. The model can also be described in terms
of a mixture of two normal distributions, concentrated on the vertical
lines $y_1=1$ and $y_1 = -1$.
In this appendix we shall show that the population version of the mve,
constrained to be centred at at the origin, is given by
$$
\Sigma_{\text{mve}} = \begin{bmatrix} 2 & 0 \\ 0 & d^2 \end{bmatrix},
$$
where $d = \Phi^{-1}(.75)$ in terms of the cumulative distribution function
of the $N(0,1)$ distribution.
First let $u_1 < u_2$ be two possible values
for $y_2$ and consider and ellipse based on a matrix $\Sigma$, with inverse
$\Sigma^{-1} = \Omega$,
\begin{equation}
\label{A:ellipse}
\/y^T \Omega \/y = 1
\end{equation}
which intersects the vertical lines at these points,
\begin{equation}
\label{A:ellipse-lines}
\begin{bmatrix} 1 & u_1 \end{bmatrix} \Omega
\begin{bmatrix} 1 \\ u_1\end{bmatrix} = 1, \quad
\begin{bmatrix} 1 & u_2 \end{bmatrix} \Omega
\begin{bmatrix} 1 \\ u_2\end{bmatrix} = 1.
\end{equation}
By symmetry the ellipse also intersects the points $(-1,-u_1)^T$ and
$(-1,-u_2)^T$. Note that $\Sigma$ will be a candidate for the mve
matrix if the interior of the ellipse covers 50\% of the probability
mass, that is,
\begin{equation}
\label{A:coverage}
\Phi(u_2) = \Phi(u_1)+1/2.
\end{equation}
If $u_1$ and $u_2$ are finite, then necessarily $u_1 < 0$ and $u_2 >0$.
The proof will proceed in two stages. First, for fixed $u_1, \ u_2$
satisfying \eqref{A:coverage}, we choose $\Sigma$ to minimize
$\det(\Sigma)$ (or equivalently maximize $\det(\Omega)$). Secondly,
we optimize over the choice of $u_1, \ u_2$.
Thus, start with a fixed pair of values $u_1, \ u_2$
satisfying \eqref{A:coverage}. If $\/y = (1,u)^T$ represents a point on one of the vertical lines, then the intersection with the ellipse \eqref{A:ellipse} can be written
$$
\omega_{11} + 2 \omega_{12} u + \omega_{22} u^2 = 1,
$$
or equivalently as the quadratic equation in u,
$$
A u^2 + B u + C = 0,
$$
where $A = \omega_{22}$, $B = 2 \omega_{12}$, $C = \omega_{11} -1$.
If this ellipse passes through $(1,u_1)^T$ and $(1,u_2)^T$, then then
$u_1, \ u_2$ are roots of the quadratic equation, so
\begin{equation}
\label{A:quad-roots}
u_1,u_2 = \frac{-B \pm \sqrt{B^2 - 4AC}}{2A}.
\end{equation}
In particular, setting $M = (u_1+u_2)/2$ to be the mean of the roots,
and $P = u_1u_2$ to be the product of the roots, we have
\begin{equation}
\label{A:roots}
M = -\frac{B}{2A} = -\frac{\omega_{12}}{\omega_{22}}, \quad
P = \frac{C}{A} = \frac{\omega_{11} - 1}{\omega_{22}}.
\end{equation}
Let us try to maximize $\det(\Omega)$ subject to the ellipse
satisfying \eqref{A:ellipse-lines}. Start with an arbitrary $\omega_{22}>0$.
Then \eqref{A:roots} determines the remaining elements of $\Omega$,
$$
\omega_{12} = -M \omega_{22}, \quad \omega_{11} = 1+P \omega_{22}.
$$
Hence
$$
\det(\Omega) = \omega_{11} \omega_{22} - \omega_{12}^2 =
\omega_{22} - Q \omega_{22}^2,
$$
where
\begin{equation}
\label{A:Q}
Q = M^2 - P = \frac{1}{4}(u_1 - u_2)^2> 0.
\end{equation}
Maximizing $\det(\Omega)$ with respect to the choice of $\omega_{22}$
leads to $\omega_{22} = 1/(2Q)$ and
$$
\det(\Omega) = 1/(4Q).
$$
The remaining task is to choose $u_1<0$ (which determines $u_2>0$ by
\eqref{A:coverage}) to maximize $\det(\Omega)$, or equivalently, to
minimize $Q$ in \eqref{A:Q}.
Recall a basic result from calculus. If $t=f(u)$ and $u = g(t)$ are
monotone functions which are inverse to one another, then $g(f(u)) =
u$. Differentiating two times yields the relation between the
derivatives,
$$
g' = 1/f', \quad g'' = -f''/\{f'\}^3.
$$
In particular, consider $f(u) = \Phi(u)$, with derivatives $f'(u) =
\phi(u)$ and $f''(u) = -u \phi(u)$, where $\phi(u)$ is the probability
density function of $N(0,1)$. Then $g(t) = \Phi^{-1}(t)$ with
derivatives $g'(t) = 1/\phi(u)$ and $g''(t) = u/\{\phi(u)\}^2$, where
$u = \Phi^{-1}(t)$.
With this notation, write $u_1 = g(t)$ for $0 < t < 1/2$. Then $u_2 =
g(t+1/2)$. Write $\phi_1 = \phi(u_1), \ \phi_2 = \phi(u_2)$. The
quantity $Q$ in \eqref{A:Q}, treated as a function of $t$, has
derivatives
\begin{align*}
Q' &= \frac{1}{2} \left\{ u_1 u_1' - u_1 u_2' - u_1'u_2 + u_2 u_2'\right\}\\
&= \frac{1}{2} \left\{u_1(1/\phi_1 -1/\phi_2) +
u_2(1/\phi_2 -1/\phi_1) \right\}\\
Q'' &= \frac{1}{2} \left\{ u_1 u_1''+ (u_1')^2 -u_1 u_2'' - 2 u_1' u_2' -
u_1''u_2 + u_2 u_2''+ (u_2')^2 \right\}\\
&= \frac{1}{2} \left\{ u_1^2/\phi_1^2 + 1/\phi_1^2 - u_1u_2/\phi_2^2 -
2/(\phi_1 \phi_2) - u_1u_2/\phi_1^2 + u_2^2/\phi_2^2 + 1/\phi_2^2 \right\}\\
&= \frac{1}{2} \left\{ (1/\phi_1 - 1/\phi_2)^2 +
u_1^2/\phi_1^2 + - u_1u_2/(\phi_1^2 + \phi_2^2) + u_2^2/\phi_2^2 \right\}.
\end{align*}
If $u_1=-d$, then $u_2=d$ and $\phi_1 = \phi_2$ so that the first
derivative vanishes. For all $(0 < t< 1/2)$, the second derivative is
positive, so the function is convex. Hence $Q$ is minimized for
$u_1=-d, \ u_2=d$. Then $M=0, Q=-P = d^2$ and the optimal $\Sigma$
becomes
$$
\Sigma = \Omega^{-1} = \begin{bmatrix} 2 & 0 \\ 0 & 2 d^2 \end{bmatrix},
$$
as required.
\clearpage
\bibliographystyle{apalike}
|
\section{Introduction}
Satisfiability of the Bernays-Sch\"onfinkel-Ramsey (BSR) fragment of first-order logic is decidable and NEXPTIME-complete~\cite{Lewis1980}.
The complexity remains if the fragment is restricted to a clause normal form. Only further restrictions on the number of literals per clause
enable better complexity results~\cite{Plaisted84}.
Its extension with linear arithmetic is undecidable. For example, Halpern~\cite{Halpern1991} showed that the combination of
Presburger Arithmetic with one unary predicate yields undecidability. His proof relies on $\forall\exists$ quantifier alternations
on the arithmetic part. The naturals can be defined on the basis of linear real arithmetic with the help of one (extra) unary predicate.
Fietzke and Weidenbach~\cite{Fietzke2012} showed undecidability for the combination of linear real arithmetic with several
binary or one ternary predicate. The proof is based on a reduction of the two-counter machine model~\cite{Minsky1967} to this fragment where a purely universal
quantifier prefix suffices. Two-counter machine instructions are translated into clauses of the form
\begin{center}
$\begin{array}{rr@{\;\parallel\;}l@{\;\rightarrow\;}l}
(i) & x' = x + 1 & P(i,x,y) & P(i+1,x',y)\\
(ii) & x = 0 & P(i,x,y) & P(j,x',y)\\
& x > 0, x' = x -1 & P(i,x,y) & P(i+1,x',y)\\
\end{array}$
\end{center}
where $P(i,x,y)$ models the program at instruction $i$ with counter values $x$, $y$.
Then, clause (i)~models the increment of counter $x$ (analogous for $y$) and a go to the next instruction;
clauses (ii)~model the conditional decrement of $x$ (analogous for $y$) and, otherwise, a jump to instruction $j$.
The start state is represented by a clause $\parallel \rightarrow P(1,n,m)$ for two positive integer values $n$, $m$, and
the halt instruction is represented by an atom $P(\text{halt},x,y)$ and reachability of the halting state
by a clause $\parallel P(\text{halt},x,y)\rightarrow \Box$. Then, a two-counter machine program halts if and only if
the BSR clause set of linear arithmetic with one ternary predicate constructed out of the program
is unsatisfiable. Note that for this reduction it
does not matter whether integer or real arithmetic is the underlying arithmetic semantics.
The first argument of $P$ is always a natural.
Our first contribution is refinements of the two-counter machine reduction where the arithmetic constraints
are further restricted to linear difference inequations $x-y \triangleleft c$,
simple additive inequations $x+y \triangleleft c$,
quotient inequations $x \triangleleft c\cdot y$ and
multiplicative inequations $x \cdot y \triangleleft c$ where $c\in\mathbb{R}$, and $\triangleleft\in\{<, \leq, =,\neq,\geq, >\}$.
Under all these restrictions, the combination
remains undecidable, respectively (see Section~\ref{section:UndecidableFragments}).
On the positive side, we prove decidability of the restriction to arithmetic constraints consisting of
simple bounds of the form $x\triangleleft c$, where $\triangleleft$ and $c$ are as above. Underlying
the result is the observation that similar to the finite model property of BSR, only finitely many test points are sufficient
for the arithmetic simple bounds constraints. Our construction is motivated by
results from quantifier elimination~\cite{Loos1993} and hierarchic superposition~\cite{Bachmair1994b,Kruglov2012,Fietzke2012}.
For example, consider the two clauses
\begin{center}
$\begin{array}{r@{\;\parallel\;}l@{\;\rightarrow\;}l}
x_2\neq 5 & R(x_1) & Q(u_1,x_2)\\
y_1<7, y_2\leq 2 & & Q(c,y_2), R(y_1)\\
\end{array}$
\end{center}
where $u_1$ is a free first-order variable, $x_i$, $y_i$ are variables over the reals, and
$c$ is a free first-order constant.
Our main result reveals that this clause set is satisfiable if and only if for every variable at the first argument of $Q$
the constant $c$ is substituted (Corollary~\ref{corollary:EquisatisfiabilityFreeInstantiation}), for the second argument of $Q$ the abstract real values
$5+\varepsilon$ and $-\infty$ and for $R$ the value $-\infty$ (Definitions~\ref{definition:BaseInstantiationPoints}, \ref{definition:BaseVariableInstantiationPoints}, Lemma~\ref{lemma:EquisatisfiabilityBaseInstantiation}).
The instantiation does not need to consider the simple bounds $y_1<7$, $y_2\leq 2$, because it is sufficient to explore the reals either from
$-\infty$ upwards or from $+\infty$ downwards, as is similarly done in linear quantifier elimination~\cite{Loos1993}. Also instantiation does not need
to consider the value $5+\varepsilon$ for $R$, motivated by the fact that hierarchic superposition will not derive the respective simple bound for the first
argument of $R$ in any generated clause~\cite{Bachmair1994b}. This idea can be extended to free argument positions of predicates and the respective
free constants. Every BSR clause set combined with simple bounds is always sufficiently complete because it does not contain a free non-constant
function symbol.
All abstract values are represented by Skolem constants over the reals, together with defining axioms.
For the example, we introduce the fresh Skolem constants $\alpha_{-\infty}$ to represent $-\infty$ and $\alpha_{5+\varepsilon}$ to represent $5+\varepsilon$
together with axioms expressing $\alpha_{-\infty} < 2 < 5 < \alpha_{5+\varepsilon}$.
Eventually, we obtain the ground clause set
\begin{center}
$\begin{array}{r@{\;\parallel\;}l@{\;\rightarrow\;}l}
\alpha_{-\infty} \geq 2 & & \Box \\
5 \geq \alpha_{5+\varepsilon} & &\Box \\
\alpha_{5+\varepsilon}\neq 5 & R(\alpha_{-\infty}) & Q(c,\alpha_{5+\varepsilon})\\
\alpha_{-\infty}\neq 5 & R(\alpha_{-\infty}) & Q(c,\alpha_{-\infty})\\
\alpha_{-\infty}<7, \alpha_{5+\varepsilon}\leq 2 & & Q(c,\alpha_{5+\varepsilon}), R(\alpha_{-\infty})\\
\alpha_{-\infty}<7, \alpha_{-\infty}\leq 2 & & Q(c,\alpha_{-\infty}), R(\alpha_{-\infty})\\
\end{array}$
\end{center}
which has the model
$\alpha_{-\infty}^{\mathcal{A}} = 1$, $\alpha_{5+\varepsilon}^{\mathcal{A}} = 6$, $R^{\mathcal{A}} = \{1\}$, $Q^{\mathcal{A}} = \{(c,6), (c,1)\}$, for instance.
Using the result that every BSR clause set with simple bounds has a satisfiability-preserving ground instantiation
with respect to finitely many constants, we prove NEXPTIME-completeness of the fragment in Section~\ref{section:Complexity}.
For this result, the fine grained instantiation introduced in Section~\ref{section:BaseModelTheory} and explained above by example
is not needed. However, our further goal is to develop useful reasoning procedures for the fragment for which a smaller set
of instances improves efficiency a lot. For the same reason, we do not restrict our attention to the BS fragment, but consider
equality as a first class citizen of the logic from the very beginning.
Once a BSR clause set with simple bounds is grounded, there are a variety of efficient decision procedures known, such as SMT solvers
employing the Nelson-Oppen~\cite{NelsonOpppen79} principle. However, for a large number of Skolem constants or large clause sets, an
a-priori grounding may, due to its exponential increase in size, not be affordable. We are not aware of any calculus that is actually
a decision procedure for the BSR fragment with simple bounds, although some work in this direction
has been done already~\cite{BaumgartnerFT08,Ruemmer08,BonacinaEtAl11,Kruglov2012}.
The decidability result on simple bounds can be lifted to constraints of the form $x \triangleleft s$ where $x$ is the only
variable and $s$ a ground expression. The lifting is done by the introduction of further Skolem constants for complicated ground terms,
following ideas of~\cite{Kruglov2012} and presented in Section~\ref{section:Basification}. The paper ends with a conclusion, Section~\ref{sec:conclusion}.
\section{Basic Definitions}
Hierarchic combinations of first-order logic with background theories \cite{Bachmair1994b} build upon sorted logic with equality. A \emph{(sorted) signature} $\Sigma$ consists of a set $\Xi$ of sorts, a set $\Omega$ of function symbols together with sort information and a set $\Pi$ of predicate symbols also equipped with sort information. For the sake of simplicity, we restrict ourselves to two sorts: the \emph{base sort $\mathcal{R}$} -- interpreted as the set $\mathbb{R}$ of reals by all the interpretations we shall consider -- and the \emph{free sort $\S$} -- interpreted by some freely selectable nonempty domain.
Throughout the paper we use convenient notation such as $P:\xi_1\times\ldots\times\xi_m\in\Pi$ to address an $m$-ary predicate symbol with full sort information; $P/m\in\Pi$ if there is some predicate symbol $P$ of arity $m$ in $\Pi$; $P\in\Pi$ if there is some $m$ such that $P/m\in\Pi$. To avoid confusion, we assume that $\Pi$ contains at most one pair $P : \xi_1\times\ldots\times\xi_m$ for every $P$, and do not allow for multiple occurrences of $P$ with different arity or sorting. Similar conventions shall hold for function symbols. In addition, we occasionally use the notation $\Omega\cap\mathcal{C}$ (where $\mathcal{C}$ is an arbitrary set of constant symbols without annotated sorting information) to denote the set $\{c \mid c:\xi \in \Omega \text{ for some sort $\xi$ and } c\in\mathcal{C}\}$.
We instantiate the framework of hierarchic specification from \cite{Bachmair1994b} by the hierarchic combination of the BSR fragment of first-order logic with a base theory allowing for the formulation of simple constraints on real-valued constant symbols and variables.
Formally, the used \emph{specification of the base theory} comprises the base signature $\Sigma_\text{\upshape{LA}} := \bigl\< \{\mathcal{R}\},\; \Omega_\text{\upshape{LA}}\cup\Omega_\text{\upshape{LA}}^\alpha, \Pi_\text{\upshape{LA}}\bigr\>$, where
\begin{align*}
\Omega_\text{\upshape{LA}} &:= \{r \!:\! \mathcal{R} \mid r\in\mathbb{R} \} \cup \{c_1\!:\!\mathcal{R}, \ldots, c_\kappa\!:\!\mathcal{R}\}, \\
\Omega_\text{\upshape{LA}}^\alpha & := \{\alpha_{-\infty}\!:\!\mathcal{R}\} \cup \{\alpha_{c+\varepsilon}\!:\!\mathcal{R} \,|\, c\in\Omega_\text{\upshape{LA}}\}, \\
\Pi_\text{\upshape{LA}} &:= \{ <\,:\mathcal{R}^2,\; \leq\,:\mathcal{R}^2,\; =\,:\mathcal{R}^2,\; \not=\,:\mathcal{R}^2,\; \geq\,:\mathcal{R}^2,\; >\,:\mathcal{R}^2\},
\end{align*}
together with the class $\mathfrak{M}$ of models containing one \emph{base model $\mathcal{M}$} for every possible allocation of the Skolem constants $c_1,\ldots, c_\kappa$ in $\Omega_\text{\upshape{LA}}$ and the ones in $\Omega_\text{\upshape{LA}}^\alpha$ to real numbers. Moreover, each of the base models $\mathcal{M} \in \mathfrak{M}$ shall extend the \emph{standard model $\mathcal{M}_\text{\upshape{LA}}$ of linear arithmetic over the reals}, i.e.\ $\mathcal{R}^\mathcal{M} = \mathbb{R}$ and $c^\mathcal{M} = c$ for every constant symbol $c\in\Omega_\text{\upshape{LA}}\cap \mathbb{R}$ and $\triangleleft^\mathcal{M} = \triangleleft$ for every $\triangleleft\in\{<,\leq,=,\not=,\geq,>\}$ -- in other words all real numbers serving as constant symbols represent their canonical value under each $\mathcal{M}\in\mathfrak{M}$ and every such $\mathcal{M}$ interprets all predicate symbols $<,\leq,=,\not=,\geq,>$ by their standard meaning on the reals.
The definition of the base theory leaves the exact number $\kappa$ of additional Skolem constants open; it may be $0$. While adding all reals as constant symbols to the signature serves the aim of defining $\mathfrak{M}$ to contain term-generated models only, the Skolem constants $c_1, \ldots, c_\kappa$ serve at least two purposes. Firstly, the modeling capabilities of our logical language are enhanced by real-valued constant symbols of which the exact value is \emph{not} predetermined. Secondly, in Section \ref{section:Basification} we outline a technique which allows us to soften our requirements towards the syntax a bit so that ground non-constant terms $s$ of the base sort become admissible in addition to constant symbols and variables. The method has already been described in \cite{Kruglov2012} under the name \emph{basification} and introduces defining unit clauses such as $c= s$ where $c$ is a fresh Skolem constant and $s$ is a variable-free term.
In fact, we introduce even more Skolem constants by adding the set $\Omega_\text{\upshape{LA}}^\alpha$ to the base signature. For notational convenience, we syntactically distinguish this special kind of base-sort constant symbols $\alpha_\t$, $\t$ being of the form $-\infty$ or $d+\varepsilon$ for arbitrary base-sort constant symbols $d\in\Omega_\text{\upshape{LA}}$. These will play a key role when we instantiate base-sort variables later on. We associate an inherent meaning with constant symbols $\alpha_\t$ that will be formalized by means of special of axioms.
We hierarchically extend the base specification $\<\Sigma_\text{\upshape{LA}}, \mathfrak{M}\>$ by the free sort $\S$ and finite sets $\Omega$ and $\Pi$ of \emph{free constant symbols} and \emph{free predicate symbols}, respectively, each equipped with appropriate sort information. We use the symbol $\approx$ to denote the built-in equality predicate on $\S$. To avoid confusion, we assume each constant or predicate symbol to occur in at most one of the sets $\Omega_\text{\upshape{LA}}$, $\Omega_\text{\upshape{LA}}^\alpha$, $\Omega$, $\Pi_\text{\upshape{LA}}$, $\Pi$ and that none of these sets contains the $\approx$ symbol.
In the light of sorted terms we consider two disjoint countably infinite sets of variables $V_\mathcal{R}$ -- the \emph{variables of the base sort}, usually denoted $x, y, z$ -- and $V_\S$ -- the \emph{free-sort variables}, usually denoted $u, w$.
\begin{definition}[BSR Clause Fragment with Simple Bounds]\label{definition:BSRwithConstrSyntax} ~\\
Let $\bigl\<\{\mathcal{R}, \S\}, \Omega, \Pi\bigr\>$ be a signature such that $\Omega$ exclusively contains constant symbols $c$ of the free sort and no function symbols of greater arity, and such that for every predicate symbol $P:\xi_1\times\ldots\times\xi_m \in\Pi$ and every argument position $i\leq m$ it holds $\xi_i\in\{\mathcal{R}, \S\}$.
An \emph{atomic constraint} is of the form $c\triangleleft d$ or $x\triangleleft d$ or $\alpha_\t \triangleleft d$ or $x= \alpha_\t$ or $\alpha_\t = \alpha_{\t'}$ with $c,d\in\Omega_\text{\upshape{LA}}$, $x\in V_\mathcal{R}$, $\alpha_\t, \alpha_{\t'}\in\Omega_\text{\upshape{LA}}^\alpha$ and $\triangleleft \in \{<,\leq,=,\not=,\geq,>\}$.
A \emph{free atom} $A$ is either of the form $s\approx s'$ with $s,s'$ being free-sort constant symbols in $\Omega$ or free-sort variables in $V_\S$, respectively, or $A$ is of the form $P(s_1, \ldots, s_m)$, where $P:\xi_1\times\ldots\times\xi_m \in\Pi$ is an $m$-ary predicate symbol. For each $i\leq m$ the term $s_i$ shall be of the sort $\xi_i$. If $\xi_i = \mathcal{R}$, then $s_i$ must be a variable $x\in V_\mathcal{R}$, and in case of $\xi_i = \S$, $s_i$ may be a variable $u\in V_\S$ or a constant symbol $c:\S\in\Omega$.
A \emph{clause} has the form $\Lambda \;\|\; \Gamma \to \Delta$, where $\Lambda$ is a multiset of atomic constraints, and $\Gamma$ and $\Delta$ are multisets of free atoms. We usually refer to $\Lambda$ as the \emph{constraint part} and to $\Gamma\to\Delta$ as the \emph{free part} of the clause.
We conveniently denote the union of two multisets $\Theta$ and $\Theta'$ by juxtaposition $\Theta, \Theta'$. Moreover, we often write $\Theta, A$ as an abbreviation for the multiset union $\Theta \cup \{A\}$. In our clause notation empty multisets are usually omitted left of ``$\to$'' and denoted by $\Box$ right of ``$\to$'' (where $\Box$ at the same time stands for \emph{falsity}).
\end{definition}
Intuitively, clauses $\Lambda \;\|\; \Gamma \to \Delta$ can be read as $\bigl(\bigwedge\Lambda \wedge \bigwedge\Gamma\bigr) \to \bigvee\Delta$.
Put into words, the multiset $\Lambda$ stands for a conjunction of all atomic constraints it contains, $\Gamma$ stands for a conjunction of the free atoms in it and $\Delta$ stands for a disjunction of the contained free atoms. All occurring variables are implicitly universally quantified.
Requiring the free part $\Gamma\to\Delta$ of clauses to not contain any base-sort constant symbols does not pose any restriction to expressiveness. Every base-sort constant symbol $c$ in the free part can safely be replaced by a fresh base-sort variable $x_c$ when an atomic constraint $x_c = c$ is added to the constraint part of the clause (a process known as purification, cf.\ \cite{Bachmair1994b, Kruglov2012}).
In the rest of the paper we omit the phrase ``over the BSR fragment with simple bounds'' when talking about clauses and clause sets, although we mainly restrict our considerations to this fragment.
A \emph{hierarchic interpretation} over the \emph{hierarchic specification} $\bigl\<\<\Sigma_\text{\upshape{LA}}, \mathfrak{M}\>, \<\{\S, \mathcal{R}\}, \Omega, \Pi\>\bigr\>$ is an algebra $\mathcal{A}$ which \emph{extends} a model $\mathcal{M}\in\mathfrak{M}$, i.e.\ $\mathcal{A}$ and $\mathcal{M}$ interpret the base sort and all constant and predicate symbols in $\Omega_\text{\upshape{LA}}\cup\Omega_\text{\upshape{LA}}^\alpha$ and $\Pi_\text{\upshape{LA}}$ in exactly the same way. Moreover, $\mathcal{A}$ comprises a nonempty domain $\S^\mathcal{A}$, assigns to each constant symbol $c\!:\!\S$ in $\Omega$ a domain element $c^\mathcal{A} \in \S^\mathcal{A}$ and interprets every predicate symbol $P\!:\!\xi_1\times\ldots\times\xi_m \in\Pi$ by a set $P^\mathcal{A}\subseteq \xi_1^\mathcal{A}\times\ldots\times\xi_m^\mathcal{A}$. Summing up, $\mathcal{A}$ extends the standard model of linear arithmetic and adopts the standard approach to semantics of (sorted) first-order logics when interpreting the free part of clauses.
Given a hierarchic interpretation $\mathcal{A}$, a \emph{variable assignment} is a sort-respecting total mapping $\beta: V_\mathcal{R}\cup V_\S \to \mathbb{R} \cup \S^\mathcal{A}$ so that $\beta(x)\in\mathbb{R}$ for every variable $x\in V_\mathcal{R}$ and $\beta(u)\in\S^\mathcal{A}$ for every $u\in V_\S$.
We write $\mathcal{A}(\beta)(s)$ to mean the \emph{value of the term $s$ under $\mathcal{A}$ with respect to the variable assignment $\beta$}. In accordance with the notation used so far, we thus define $\mathcal{A}(\beta)(v) := \beta(v)$ for every variable $v$ and $\mathcal{A}(\beta)(c) := c^\mathcal{A}$ for every constant symbol $c$.
As usual, we use the symbol $\models$ to denote truth under a hierarchic interpretation $\mathcal{A}$, possibly with respect to a variable assignment $\beta$.
In detail, we have the following for atomic constraints, equational and nonequational free atroms and clauses, respectively:
\begin{itemize}
\item $\mathcal{A},\beta\models s\triangleleft s'$ if and only if $\mathcal{A}(\beta)(s) \triangleleft \mathcal{A}(\beta)(s')$,
\item $\mathcal{A},\beta\models s\approx s'$ if and only if $\mathcal{A}(\beta)(s) = \mathcal{A}(\beta)(s')$,
\item $\mathcal{A},\beta\models P(s_1, \ldots, s_m)$ if and only if\\ $\bigl\<\mathcal{A}(\beta)(s_1), \ldots, \mathcal{A}(\beta)(s_m)\bigr\>\in P^\mathcal{A}$,
\item $\mathcal{A},\beta\models \Lambda\;\|\;\Gamma\to\Delta$ if and only if
\begin{itemize}
\item $\mathcal{A},\beta\not\models s\triangleleft s'$ for some atomic constraint $(s\triangleleft s')\in\Lambda$,
\item $\mathcal{A},\beta\not\models A$ for some free atom $A\in\Gamma$, or
\item $\mathcal{A},\beta\models B$ for some free atom $B\in\Delta$.
\end{itemize}
\end{itemize}
The variables occurring in clauses shall be universally quantified. Therefore, given a clause $C$, we call $\mathcal{A}$ a \emph{hierarchic model of $C$}, denoted $\mathcal{A}\models C$, if and only if $\mathcal{A},\beta\models C$ holds for every variable assignment $\beta$. For clause sets $N$, we say $\mathcal{A}$ is a \emph{hierarchic model of $N$}, also denoted $\mathcal{A}\models N$, if and only if $\mathcal{A}$ is a model of all clauses in $N$.
We call a clause $C$ (a clause set $N$) \emph{satisfiable w.r.t.\ $\mathfrak{M}$} if and only if there exists a hierarchic model $\mathcal{A}$ of $C$ (of $N$). Most of the time, we will omit the explicit reference to $\mathfrak{M}$, although we shall only consider models as satisfying that extend the standard model of linear real arithmetic.
From now on, we implicitly base all further considerations on a hierarchic specification\\ $\bigl\<\<\Sigma_\text{\upshape{LA}}, \mathfrak{M}\>, \<\{\S, \mathcal{R}\}, \Omega, \Pi\>\bigr\>$ whose extension part $\<\{\S, \mathcal{R}\}, \Omega, \Pi\>$ fulfills the requirements stipulated in Definition \ref{definition:BSRwithConstrSyntax} and which we simply take for granted. Often the sets $\Omega_\text{\upshape{LA}}\setminus\mathbb{R}$, $\Omega$ and $\Pi$ will coincide with the constant and predicate symbols that occur in a given clause set, and we thus assume a proper definition of the hierarchic specification.
\emph{Substitutions} $\sigma$ shall be defined in the standard way as sort-respecting mappings from variables to terms over our underlying signature. The \emph{restriction of the domain of a substitution $\sigma$ to a set $V$ of variables} is denoted by $\sigma|_V$ and shall be defined so that $v\sigma|_V := v\sigma$ for every $v\in V$ and $v\sigma|_V = v$ for every $v\not\in V$. While the application of a substitution $\sigma$ to terms, atoms and multisets thereof can be defined as usual, we need to be more specific for clauses. Consider a clause $C = \Lambda \;\|\; \Gamma \to \Delta$ and let $x_1, \ldots, x_k$ denote all base-sort variables occurring in $C$ for which $x_i\sigma \neq x_i$. We then set $C\sigma := \Lambda\sigma, x_1 = x_1\sigma,\, \ldots,\, x_k = x_k\sigma \;\|\; (\Gamma\to\Delta)\sigma|_{V_\S}$.
\emph{Simultaneous substitution} of $k \geq 1$ distinct variables $v_1, \ldots\, v_k$ with terms $s_1, \ldots, s_k$ shall be denoted in vector notation $\subst{v_1, \ldots, v_k}{s_1, \ldots, s_k}$ (or $\subst{\bar{v}}{\bar{s}}$ for short), where we require every $s_i$ to be of the same sort as $v_i$, of course.
Consider a clause $C$ and let $\sigma$ be an arbitrary substitution. The clause $C\sigma$ and variable-renamed variants thereof are called \emph{instances} of $C$.
A term $s$ is called \emph{ground}, if it does not contain any variables.
A clause $C$ shall be called \emph{essentially ground} if it does not contain free-sort variables and for every base-sort variable $x$ occurring in $C$, there is an atomic constraint $x= d$ in $C$ for some constant symbol $d\in\Omega_\text{\upshape{LA}}\cup\Omega_\text{\upshape{LA}}^\alpha$. A clause set $N$ is \emph{essentially ground} if all the clauses it contains are essentially ground.
We can restrict the syntactic form of clauses even further without limiting expressiveness.
\begin{definition}[Normal Form of Clauses and Clause Sets]\label{definition:BSRwithConstrNormalform}
A clause $\Lambda \;\|\; \Gamma \to \Delta$ is in \emph{normal form} if all base-sort variables which occur in $\Lambda$ do also occur in $\Gamma \to \Delta$.
A clause set $N$ over the BSR fragment with simple bounds is in \emph{normal form} if all clauses in $N$ are in normal form and pairwise variable disjoint.
\end{definition}
Let us briefly clarify why the above requirement on clauses does not limit expressiveness.
Any base-sort variable $x$ not fulfilling the stated requirement can be removed from the clause $\Lambda\;\|\;\Gamma\to\Delta$ by existential quantifier elimination methods that transform $\Lambda$ into an equivalent constraint $\Lambda'$ in which $x$ does not occur.\footnote{Methods for the elimination of existentially quantified real variables include Fourier-Motzkin variable elimination \cite{Dantzig1973}, the Loos-Weispfenning procedure \cite{Loos1993} and many others, see e.g.\ Chapter 5 in \cite{Kroening2008} for further details.} Moreover, $\Lambda'$ can be constructed in such a way that it contains only atomic constraints of the form admitted in Definition \ref{definition:BSRwithConstrSyntax} and so that no variables or constant symbols other than the ones in $\Lambda$ are necessary.
Given a clause set $N$, we will use the following notation: the set of all constant symbols occurring in $N$ shall be denoted by $\text{\upshape{consts}}(N)$. While the set $\text{\upshape{bconsts}}(N)$ exclusively contains all base-sort constant symbols from $\Omega_\text{\upshape{LA}}$ that occur in $N$, all base-sort constant symbols $\alpha_\t$ appearing in $N$ shall be collected in the set $\text{\upshape{$\alpha$consts}}(N) := \text{\upshape{consts}}(N)\cap\Omega_\text{\upshape{LA}}^\alpha$. The set of all free-sort constant symbols in $N$ is called $\text{\upshape{fconsts}}(N)$. Altogether, the sets $\text{\upshape{bconsts}}(N)$, $\text{\upshape{$\alpha$consts}}(N)$ and $\text{\upshape{fconsts}}(N)$ together form a partitioning of $\text{\upshape{consts}}(N)$ for every clause set $N$.
We moreover denote the set of all variables occurring in a clause $C$ (clause set $N$) by $\text{\upshape{vars}}(C)$ ($\text{\upshape{vars}}(N)$).
\section{Instantiation of Base-Sort Variables}\label{section:BaseModelTheory}
We first summarize the overall approach described in this section in an intuitive way. To keep the informal exposition simple, we pretend that all base-sort constant symbols are taken from $\mathbb{R}$ and thus are interpreted by their canonical value in all hierarchic interpretations. Consequently, we can speak of real values instead of constant symbols, and even refer to improper values such as $-\infty$ (a ``sufficiently small'' real value) and $r+\varepsilon$ (a real value ``slightly larger than $r$ but not too large''). A formal treatment with proper definitions will follow.
Given a finite clause set $N$, we aim at partitioning the reals $\mathbb{R}$ into finitely many partitions so that whenever $N$ is satisfiable, there exists a hierarchic model $\mathcal{A}$ of $N$
\begin{enumerate}[label=($\mathfrak{Pr}$), ref=($\mathfrak{Pr}$)]
\item\label{enum:IntervalModel} whose interpretation of predicate symbols does not distinguish between values within the same partition.
\end{enumerate}
As soon as we found such a finite partitioning $\P$, we pick one real value $r_p \in p$ as representative from every partition $p\in\P$. The following observation motivates why we can use those representatives instead of using universally quantified variables: given a clause $C$ that contains a base-sort variable $x$, and given a set $\{c_1, \ldots, c_{k}\}$ of constant symbols such that $\{c_1^\mathcal{A}, \ldots, c_k^\mathcal{A}\} = \{r_p \mid p\in\P\}$ it holds
\begin{equation}\label{eqn:InstantiationEquivalence} \mathcal{A}\models C \quad\Longleftrightarrow\quad \mathcal{A}\models \bigl\{C\subst{x}{c_i} \bigm| 1\leq i \leq k \bigr\}~.
\end{equation}
The equivalence claims that we can transform universal quantification over the base domain into finite conjunction over all representatives of partitions in $\P$. The formal version of this statement is given in Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation}, and we will see that hierarchic models complying with property \ref{enum:IntervalModel} play a key role in its proof.
Since $\P$ is supposed to be finite, the resulting set of instances $\bigl\{C\subst{x}{c_i} \mid 1\leq i\leq k\bigr\}$ is finite, too.
It turns out that the notion of \emph{elimination sets} described by Loos and Weispfenning in \cite{Loos1993} (in the context of quantifier elimination for linear arithmetic) can be adapted to yield reasonable sets of representatives from which we can construct a finite partitioning exhibiting the described characteristics. In this case the partitions are intervals on the real axis.
Intuitively speaking, we start with the partitioning $\P_0$ consisting of a single partition $\P_0 = \bigl\{(-\infty, +\infty)\bigr\}$. This initial partition shall be represented by $-\infty$ (the ``default representative''). We then successively extract from the given clause set $N$ larger and larger points $r$ on the real axis at which we have to introduce a new boundary, cut off the interval that is currently unbounded from above at that boundary and introduce the cut-off part as a new interval to the partitioning. For instance, an interval $[r', +\infty)$ might be cut into two parts $[r', r]$ and $(r, +\infty)$ for some point $r\geq r'$. Both parts become then new partitions in the partitioning and while the interval $[r', r]$ will be represented by $r'$, the other partition $(r,+\infty)$ will have $r+\varepsilon$ as its representative.
At the end of this process the overall result will be a finite partitioning of the real numbers with the desired properties.
In fact, we will operate on a more fine-grained level than described so far, as we define such partitionings independently for certain groups of base-sort variables. The possible benefit may be a significant decrease in the number of necessary instances. But there is even more potential for savings. The complete line of definitions and arguments will be laid out in detail for one direction of instantiation along the real axis, namely in the positive direction starting from $-\infty$ and going on to larger and larger instantiation points. However, one could as well proceed in the opposite way, starting from $+\infty$ and becoming smaller and smaller thereafter. While theoretically this duality is not worth much more than a side note, it appears to be quite interesting from a practical point of view. As it turns out, one can choose the direction of instantiation independently for each base-sort variable that needs to be instantiated. Hence, one could always pick the direction that results in less instantiation points. Such a strategy might again considerably cut down the number of instances in the resulting clause set and therefore might lead to shorter processing times when applying automated reasoning procedures.
Formally, the aforementioned representatives are induced by constant symbols when interpreted under a hierarchic interpretation. Independently of any interpretation these constant symbols will serve as symbolic \emph{instantiation points}, i.e.\ they will be used to instantiate base-sort variables similar to the $c_i$ in equivalence (\ref{eqn:InstantiationEquivalence}).
We start off by defining the set of instantiation points that need to be considered for every base-sort variable. This set depends on the constraints affecting such a variable. However, we first need to develop a proper notion of what it means for a base-sort variable to be affected by a constraint. The following example shall illustrate the involved issues.
\begin{example}
Consider the following clauses:
\[
\begin{array}{rclcll}
x_1\not= -5 &\| & &\to &T(x_1), &Q(x_1, x_2) ~, \\
y_1 < 2,\; y_2 < 0 &\| & &\to & &Q(y_1, y_2) ~, \\
z_1 \geq 6 &\| &T(z_1) &\to &\Box ~.
\end{array}
\]
Obviously, the variables $x_1$, $y_1$, $y_2$ and $z_1$ are affected by the constraints in which they occur explicitly. In the given clauses variables are just names addressing argument positions of predicate symbols. Thus, it is more suitable to speak of the \emph{argument position} $\<T,1\>$ instead of variables $x_1$ and $z_1$ that occur as the first argument of predicate symbol $T$ in the first and third clause. Speaking in such terms, argument position $\<T,1\>$ is directly affected by the constraints $x_1 \not= -5$ and $z_1 \geq 6$, argument position $\<Q,1\>$ is directly affected by $x_1\not= -5$ and $y_1 < 2$, and finally $\<Q,2\>$ is affected by $y_2 < 0$.
%
As soon as we take logical consequences into account, the notion of ``being affected'' needs to be extended. The above clause set, for instance, logically entails the clause $x \geq 6 \;\|\; \to Q(x,y)$. Hence, although not \emph{directly} affected by the constraint $z_1 \geq 6$ in the clause set, the argument position $\<Q,1\>$ is still indirectly subject to this constraint. The source of this effect lies in the first clause as it establishes a connection between argument positions $\<T,1\>$ and $\<Q,1\>$ via the simultaneous occurrence of variable $x_1$ in both argument positions.
\end{example}
One lesson learned from the example is that argument positions can be connected by variable occurrences. Such links in a clause set $N$ shall be expressed by the relation $\leftrightharpoons_N$.
\begin{definition}[Connections between Argument Positions]\label{definition:ConnectedArgumentPositions}
Let $N$ be a clause set in normal form.
We define the relation $\leftrightharpoons_N$ to be the smallest equivalence relation over pairs in $\Pi\times\mathbb{N}$ such that $\<Q,j\> \leftrightharpoons_N \<P,i\>$ whenever there is a clause in $N$ containing free atoms $Q(\ldots,v,\ldots)$ and $P(\ldots,v,\ldots)$ in which the variable $v$ occurs at the $j$-th and $i$-th argument position, respectively. (Note that $Q=P$ or $j=i$ is possible.)
The relation $\leftrightharpoons_N$ induces the set $\bigl\{[\<P,i\>]_{\leftrightharpoons_N} \mid P/m\in\Pi, 1\leq i\leq m \bigr\}$ of equivalence classes. To simplify notation a bit, we write $[\<P,i\>]$ instead of $[\<P,i\>]_{\leftrightharpoons_N}$ when the set $N$ is clear from the context.
\end{definition}
As we have argued earlier, it is more precise to speak of argument positions rather than variables, since their names are of no particular relevance. Nevertheless, variable names are a syntactical necessity. The following definition is supposed to provide a means to address the argument position class a variable stands for.
\begin{definition}[Argument Position Class]\label{definition:ArgumentPositionClass}
Let $N$ be a clause set in normal form. Consider a variable $v$ that occurs at the $i$-th argument position of a free atom $P(\ldots, v, \ldots)$ in $N$. We denote the \emph{equivalence class of argument positions $v$ is related to in $N$} by $\text{\upshape{ap}}_N(v)$, i.e.\ $\text{\upshape{ap}}_N(v) := [\<P,i\>]_{\leftrightharpoons_N}$.
If $v$ is a free-sort variable that exclusively occurs in equations $v\approx s$ in $N$, we set $\text{\upshape{ap}}_N(v) := \emptyset$.
\end{definition}
There is a crucial subtlety in this definition that guarantees well-definedness, namely that the clauses in $N$ are variable disjoint. Given a clause $C\in N$ and a variable $v$ which occurs in different argument positions $\<Q,j\>$ and $\<P,i\>$ in $C$, the definition of $\leftrightharpoons_N$ entails $\<Q,j\> \leftrightharpoons_N \<P,i\>$. Therefore, $[\<Q,j\>]$ and $[\<P,i\>]$ are identical. Since $v$ does not occur in any other clause in $N$, the class $\text{\upshape{ap}}_N(v)$ is well-defined to be $[\<P,i\>]$.
Next, we collect the instantiation points that are necessary to eliminate base-sort variables by means of finite instantiation. In order to do this economically, we rely on the same idea that also keeps elimination sets in \cite{Loos1993} comparatively small.
\begin{definition}[Instantiation Points for Base-Sort Argument Positions]\label{definition:BaseInstantiationPoints}
Let $N$ be a clause set in normal form and let $P: \xi_1\times\ldots\times \xi_m \in\Pi$ be a free predicate symbol occurring in $N$. Let $J := \{i \;|\; \xi_i = \mathcal{R}, 1\leq i\leq m \}$ be the indices of $P$'s base-sort arguments. For every $i\in J$, we define $\mathcal{I}_{P,i,N}$ to be the smallest set fulfilling
\begin{enumerate}[label=(\roman{*}), ref=(\roman{*})]
\item $d\in\mathcal{I}_{P,i,N}$ if there exists a clause $C$ in $N$ containing an atom $P(\ldots, x, \ldots)$ in which $x$ occurs as the $i$-th argument and a constraint $x= d$ or $x \geq d$ with $d\in\Omega_\text{\upshape{LA}}$ appears in $C$, and
\item $\alpha_{d+\varepsilon}\in\mathcal{I}_{P,i,N}$ if there exists a clause $C$ in $N$ containing an atom of the form $P(\ldots, x, \ldots)$, in which $x$ is the $i$-th argument and a constraint of the form $x\not= d$ or $x > d$ or $x= \alpha_{d+\varepsilon}$ with $d\in\Omega_\text{\upshape{LA}}$ appears in $C$.
\end{enumerate}
\end{definition}
The most apparent peculiarity about this definition is that atomic constraints of the form $x < d$ and $x \leq d$ are completely ignored when collecting instantiation points for $x$'s argument position. First of all, this is one of the aspects that makes this definition interesting from the efficiency point of view, because the number of instances that we have to consider might decrease considerably in this way. To develop an intuitive understanding why it is enough to consider constraints $x\triangleleft d$ with $\triangleleft \in \{=, \not=, \geq, >\}$ when collecting instantiation points, the following example may help.
\begin{example}
Consider two clauses $C = x > 2,\, x \leq 5 \;\| \to T(x)$ and $D = x < 0 \;\|\; T(x)\to\Box$. Recall that we are looking for a finite partitioning $\P$ of $\mathbb{R}$ so that we can construct a hierarchic model $\mathcal{A}$ of the clause set $\{C, D\}$ that complies with \ref{enum:IntervalModel}, i.e.\ for every partition $p\in \P$ and arbitrary real values $r_1, r_2 \in p$ it shall hold $r_1 \in T^\mathcal{A}$ if and only if $r_2 \in T^\mathcal{A}$. Of course, there exist infinitely many candidates for $\P$. A natural one is $\{(-\infty,0), [0,2], (2,5], (5,+\infty) \}$ which takes every atomic constraint in $C$ and $D$ into account. Correspondingly, we find a candidate predicate for $T^\mathcal{A}$, namely the interval $(2,5]$, so that $\mathcal{A}$ is a hierarchic model of $C$ and $D$ alike and it obeys \ref{enum:IntervalModel} with respect to the proposed partitioning $\P$.
But there are other interesting possibilities, too, for instance, the more coarse-grained partitioning $\{(-\infty, 2], (2, +\infty)\}$ together with the predicate $T^\mathcal{A} = (2, +\infty)$.
This latter candidate partitioning completely ignores the constraints $x < 0$ and $x \leq 5$ that constitute upper bounds on $x$. Dually, we could have concentrated on the upper bounds instead (completely ignoring the lower bounds). This would have lead to the partitioning $\{(-\infty, 0), [0, 5], (5, +\infty)\}$ and candidate predicates $T^\mathcal{A} = [0, 5]$ (or $T^\mathcal{A} = [0, +\infty)$). Both ways are possible, but the first one induces a simpler partitioning.
\end{example}
The example has revealed quite some freedom in choosing an appropriate partitioning of the reals. A larger number of partitions directly leads to a larger number of representatives (one for each interval). In fact, we use the collected instantiation points as representatives and taken together they induce the partitioning (as we will see in the next definition). It is due to this direct correspondence of partitions and instantiation points that a more fine-grained partitioning entails a larger number of instances that need to be considered. Hence, regarding efficiency, it is of high interest to keep the partitioning coarse.
\begin{definition}[Instantiation Points for Argument Position Classes and Induced Partitioning of $\mathbb{R}$]\label{definition:BaseVariableInstantiationPoints}
Let $N$ be a clause set in normal form and let $\mathcal{A}$ be a hierarchic interpretation (over the same signature).
For every equivalence class $[\<P,i\>]$ induced by $\leftrightharpoons_N$ we define the following:
The set $\mathcal{I}_{[\<P,i\>], N}$ of \emph{instantiation points for $[\<P,i\>]$} is defined by
\[\mathcal{I}_{[\<P,i\>], N} := \{\alpha_{-\infty}\} \cup\, \bigcup_{\<Q,j\>\in[\<P,i\>]} \mathcal{I}_{Q,j,N} ~.\]
The sequence $r_1, \ldots, r_k$ shall comprise all real values in the set $\bigl\{c^\mathcal{A} \bigm| c\in\mathcal{I}_{[\<P,i\>], N}\cap\Omega_\text{\upshape{LA}} \text{ or}$ $\alpha_{c+\varepsilon}\in \mathcal{I}_{[\<P,i\>], N} \bigr\}$ ordered so that $r_1 < \ldots < r_k$.
Given a real number $r$, we say \emph{$\mathcal{I}_{[\<P,i\>], N}$ $\mathcal{A}$-covers $r$} if there exists an instantiation point $c\in\mathcal{I}_{[\<P,i\>],N}\cap\Omega_\text{\upshape{LA}}$ with $c^\mathcal{A} = r$; analogously $(r+\varepsilon)$ is \emph{$\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$} if there is an instantiation point $\alpha_{c+\varepsilon}\in\mathcal{I}_{[\<P,i\>],N}$ with $c^\mathcal{A} = r$.
The partitioning $\P_{[\<P,i\>],N,\mathcal{A}}$ of the reals into finitely many intervals shall be the smallest partitioning (smallest w.r.t.\ the number of partitions) fulfilling the following requirements:
\begin{enumerate}[label=(\roman{*}), ref=\roman{*}]
\item\label{enum:BaseVariableInstantiationPoints:One} If $r_1$ is $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, then $(-\infty, r_1)\in\P_{[\<P,i\>],N,\mathcal{A}}$; otherwise, $(-\infty, r_1]\in\P_{[\<P,i\>],N,\mathcal{A}}$.
\item\label{enum:BaseVariableInstantiationPoints:Two} For every $j$, $1\leq j\leq k$, if $r_j$ and $r_{j}+\varepsilon$ are both $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, then $[r_j, r_j] = \{r_j\}\in\P_{[\<P,i\>],N,\mathcal{A}}$.
\item\label{enum:BaseVariableInstantiationPoints:Three} For every $\ell$, $1\leq \ell< k$,
\begin{enumerate}[label=(\ref{enum:BaseVariableInstantiationPoints:Three}.\roman{*}), ref=(\ref{enum:BaseVariableInstantiationPoints:Three}.\roman{*})]
\item\label{enum:BaseVariableInstantiationPoints:ThreeOne} if $r_\ell+\varepsilon$ and $r_{\ell+1}$ are both $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, then $(r_\ell, r_{\ell+1})\in\P_{[\<P,i\>],N,\mathcal{A}}$,
\item\label{enum:BaseVariableInstantiationPoints:ThreeTwo} if $r_\ell+\varepsilon$ is $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$ but $r_{\ell+1}$ is not, then $(r_\ell, r_{\ell+1}]\in\P_{[\<P,i\>],N,\mathcal{A}}$,
\item\label{enum:BaseVariableInstantiationPoints:ThreeThree} if $r_\ell+\varepsilon$ is not $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$ but $r_{\ell+1}$ is, then $[r_\ell, r_{\ell+1})\in\P_{[\<P,i\>],N,\mathcal{A}}$,
\item\label{enum:BaseVariableInstantiationPoints:ThreeFour} if neither $r_\ell+\varepsilon$ nor $r_{\ell+1}$ is $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, then $[r_\ell, r_{\ell+1}]\in\P_{[\<P,i\>],N,\mathcal{A}}$.
\end{enumerate}
\item\label{enum:BaseVariableInstantiationPoints:Four} If $r_k+\varepsilon$ is $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, then $(r_k, +\infty)\in\P_{[\<P,i\>],N,\mathcal{A}}$; otherwise $[r_k, +\infty)\in\P_{[\<P,i\>],N,\mathcal{A}}$.
\end{enumerate}
\end{definition}
Please note that partitionings as described in the definition do always exist, and do not contain empty partitions.
Moreover, it is worth to notice that for all instantiation points $c$ and $\alpha_{c+\varepsilon}$ in a set $\mathcal{I}_{Q,j,N}$ we have $c\in\Omega_\text{\upshape{LA}}$. Hence, the concrete values assigned to constant symbols $\alpha_\t$ do not contribute to the respective partitionings of $\mathbb{R}$.
We now fix the semantics of constant symbols $\alpha_\t$ by giving appropriate axioms that make precise what it means for $\alpha_{-\infty}$ to be ``small enough'' and for $\alpha_{c+\varepsilon}$ to be ``a little larger than $c$ but not too large'' under any hierarchic model $\mathcal{A}$ that satisfies the corresponding axioms.
\begin{definition}[Axioms for Instantiation Points $\alpha_\t$]\label{definition:BaseInstantiationAxioms}
Let $\mathcal{I}$ be a set of instantiation points and $\mathcal{C}\subseteq \Omega_\text{\upshape{LA}}$ be a set of base-sort constant symbols. We define the \emph{set of axioms}
\begin{align*}
\text{\upshape{Ax}}_{\mathcal{I}, \mathcal{C}}:=\; &\bigl\{ \alpha_{-\infty} < c \bigm| \alpha_{-\infty}\in\mathcal{I} \text{ and } c\in\mathcal{C}\bigr\} \;\cup\; \bigl\{ d < \alpha_{d+\varepsilon} \bigm| \alpha_{d+\varepsilon}\in\mathcal{I}\} \\
&\cup\; \bigl\{ d < c \;\to\; \alpha_{d+\varepsilon} < c \bigm| \alpha_{d+\varepsilon}\in\mathcal{I} \text{ and } c\in\mathcal{C}\setminus\{d\} \bigr\} \\
&\cup\; \bigl\{ d = c \;\to\; \alpha_{d+\varepsilon} = \alpha_{c+\varepsilon} \bigm| \alpha_{d+\varepsilon}\in\mathcal{I} \text{ and } c\in\mathcal{C}\setminus\{d\} \bigr\} ~.
\end{align*}
\end{definition}
The axioms we just introduced are clearly not in the admissible clause form. But they can easily be transformed into proper clauses with empty free parts. For convenience, we stick to the notation given here, but keep in mind that formally the axioms have a form in accordance with Definition \ref{definition:BSRwithConstrSyntax}.
The following proposition shows that every hierarchic interpretation can be turned into a model of a given set of axioms of the above shape by modifying the values assigned to constant symbols $\alpha_\t$. The proposition relies on three facts concerning the ordering $(\mathbb{R}, <)$: totality of $<$, the lack of minimal and maximal elements, and the density of $(\mathbb{R}, <)$, i.e.\ for arbitrary reals $r_1 < r_2$ there is an $r'$ between them.
\begin{proposition}\label{proposition:ImproperAxiomSatisfiability}
Let $\mathcal{I}$ be a nonempty set of instantiation points and let $\mathcal{C}$ be a nonempty set of base-sort constant symbols from $\Omega_\text{\upshape{LA}}$. Given an arbitrary hierarchic interpretation $\mathcal{A}$, we can construct a hierarchic model $\mathcal{B}$ so that (ii) $\mathcal{B}\models \text{\upshape{Ax}}_{\mathcal{I}, \mathcal{C}}$ and $\mathcal{B}$ and $\mathcal{A}$ differ only in the interpretation of the constant symbols in $\text{\upshape{$\alpha$consts}}(\text{\upshape{Ax}}_{\mathcal{I}, \mathcal{C}})$.
\end{proposition}
\begin{proof}
The model $\mathcal{B}$ can be derived from $\mathcal{A}$ by redefining the interpretation of all constant symbols $\alpha_\t \in \mathcal{I}$ as follows:
\begin{align*}
\alpha_{-\infty}^\mathcal{B} &:= \min\{c^\mathcal{A} \mid c\in\mathcal{C}\} - 1 ~,\\
\alpha_{d+\varepsilon}^\mathcal{B} &:= \tfrac{1}{2} \bigl( d^\mathcal{A} + \min(\{c^\mathcal{A} \mid c\in\mathcal{C},\; c^\mathcal{A} > d^\mathcal{A}\}\cup \{d^\mathcal{A}+1\}) \bigr) ~.
\end{align*}
Everything else is taken over from $\mathcal{A}$.
\end{proof}
Previously, we have been speaking of representatives of partitions $p$ in a partitioning $\P$ of the reals. As far as we know by now, these are just real values $r_p \in p$ from the specific interval they represent. The next lemma shows that we can find constant symbols $c_p$ in the set of instantiation points to address them. Furthermore, we will see where the values of these constant symbols lie within the respective interval.
In order to have a compact notation at hand when doing case distinctions on intervals, we use ``$\mbox{$(\!\cdot$}$'' to stand for ``$($'' as well as for ``$[$''. Analogously, ``$\mbox{$\cdot\!)$} $'' is used to address the two cases ``$)$'' and ``$]$'' at the same time.
\begin{lemma}\label{lemma:PartitionRepresentatives}
Let $N$ be a clause set in normal form and let $\mathcal{A}$ be a hierarchic model of\\ $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$ for an arbitrary argument position pair $\<P,i\>$.
For every partition $p\in \P_{[\<P,i\>], N, \mathcal{A}}$ we find some constant symbol $c_{[\<P,i\>], p} \in \mathcal{I}_{[\<P,i\>], N}$ so that $c_{[\<P,i\>], p}^\mathcal{A} \in p$ and
\begin{enumerate}[label=(\roman{*}), ref=(\roman{*})]
\item\label{enum:PartitionRepresentatives:One} if $p = (-\infty, r_u\mbox{$\cdot\!)$} $ or $p = (-\infty, +\infty)$, then $c_{[\<P,i\>], p}^\mathcal{A} < d^\mathcal{A}$ for every $d\in\text{\upshape{bconsts}}(N)$,
\item\label{enum:PartitionRepresentatives:Two} if $p = [r_\ell, r_u\mbox{$\cdot\!)$} $ or $p = [r_\ell, +\infty)$, then $c_{[\<P,i\>], p}^\mathcal{A} = r_\ell$.
\item\label{enum:PartitionRepresentatives:Three} if $p = (r_\ell, r_u\mbox{$\cdot\!)$} $ or $p = (r_\ell, +\infty)$, then $r_\ell < c_{[\<P,i\>], p}^\mathcal{A}$ and $c_{[\<P,i\>], p}^\mathcal{A} < d^\mathcal{A}$ for every $d\in\text{\upshape{bconsts}}(N)$ with $r_\ell < d^\mathcal{A}$.
\end{enumerate}
\end{lemma}
\begin{proof}
We proceed by case distinction on the form of the partition $p$.
\begin{description}
\item Case $p = (-\infty, r_u\mbox{$\cdot\!)$} $ for some real value $r_u$.
By construction of $\P_{[\<P,i\>], N, \mathcal{A}}$, there is a constant symbol $e\in\text{\upshape{bconsts}}(N)$ so that $e^\mathcal{A} = r_u$.
Moreover, we find the instantiation point $\alpha_{-\infty}$ in $\mathcal{I}_{[\<P,i\>], N}$ and thus $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$ contains the axiom $\alpha_{-\infty} < d$ for every constant symbol $d\in\text{\upshape{bconsts}}(N)$, in particular $\alpha_{-\infty} < e$. Hence, $\alpha_{-\infty}^\mathcal{A} \in (-\infty, r_u)$.
\item Case $p = (-\infty, +\infty)$.
We find the instantiation point $\alpha_{-\infty}$ in $\mathcal{I}_{[\<P,i\>], N}$, and obviously $\alpha_{-\infty}^\mathcal{A}$ lies in $p$. Moreover, $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$ contains the axiom $\alpha_{-\infty} < d$ for every constant symbol $d\in\text{\upshape{bconsts}}(N)$.
\item Case $p = [r_\ell, r_u\mbox{$\cdot\!)$} $ for real values $r_\ell, r_u$ with $r_\ell \leq r_u$.
\begin{description}
\item If $p$ is a point interval $[r_\ell, r_\ell]$, then we find a constant symbol $e\in\mathcal{I}_{[\<P,i\>], N}\cap\Omega_\text{\upshape{LA}}$ so that $p = \{e^\mathcal{A}\}$.
\item If $p$ is not a point interval, i.e.\ $r_\ell < r_u$, then the definition of $\P_{[\<P,i\>], N, \mathcal{A}}$ entails the existence of a constant symbol $e \in \text{\upshape{bconsts}}(N)$ such that $e^\mathcal{A} = r_\ell$ and either $e \in \mathcal{I}_{[\<P,i\>], N}$ or $\alpha_{e+\varepsilon}\in\mathcal{I}_{[\<P,i\>], N}$. But since the partition $p$ is not a point interval, $r_\ell + \varepsilon$ cannot be $\mathcal{A}$-covered by $\mathcal{I}_{[\<P,i\>],N}$, and thus $\alpha_{e+\varepsilon}$ cannot be in $\mathcal{I}_{[\<P,i\>],N}$. Consequently, $e$ must be in $\mathcal{I}_{[\<P,i\>],N}$ and we can choose $c_{[\<P,i\>], p} := e$.
\end{description}
\item The case of $p = [r_\ell, +\infty)$ can be argued analogously to the previous case.
\item Case $p = (r_\ell, r_u\mbox{$\cdot\!)$} $ for real values $r_\ell, r_u$ with $r_\ell < r_u$.
By construction of $\P_{[\<P,i\>], N, \mathcal{A}}$, we find some instantiation point $\alpha_{e+\varepsilon}\in \mathcal{I}_{[\<P,i\>], N}$ so that $e^\mathcal{A} = r_\ell$. Moreover, there exists a second constant symbol $e'\in\text{\upshape{bconsts}}(N)$ for which $e'^\mathcal{A} = r_u$.
As a consequence of $\alpha_{e+\varepsilon}\in \mathcal{I}_{[\<P,i\>], N}$ the set $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$ contains the axioms $e < \alpha_{e+\varepsilon}$ and $e<d \to \alpha_{e+\varepsilon} < d$ for every constant symbol $d\in\text{\upshape{bconsts}}(N)$, in particular for $d=e'$. And since $\mathcal{A}$ is a model of $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$, requirement \ref{enum:PartitionRepresentatives:Three} is satisfied and furthermore $\alpha_{e+\varepsilon}^\mathcal{A}\in p$ follows, since $e^\mathcal{A} < \alpha_{e+\varepsilon}^\mathcal{A} < e'^\mathcal{A}$.
\item Case $p = (r_\ell, +\infty)$. There must be an instantiation point $\alpha_{e+\varepsilon}$ in $\mathcal{I}_{[\<P,i\>],N}$ such that $e^\mathcal{A} = r_\ell$.
As we have already done in Definition \ref{definition:BaseVariableInstantiationPoints}, we denote by $r_1, \ldots, r_k$ all real values in ascending order which $\mathcal{A}$ assigns to constant symbols $d\in\text{\upshape{bconsts}}(N)$ that either themselves are instantiation points in $\mathcal{I}_{[\<P,i\>],N}$ or for which $\alpha_{d+\varepsilon}$ occurs in $\mathcal{I}_{[\<P,i\>],N}$.
By construction of $\P_{[\<P,i\>], N, \mathcal{A}}$, we know $r_\ell = r_k > r_{k-1} > \ldots > r_1$, and thus there is no base-sort constant symbol $e'$ in $\text{\upshape{bconsts}}(N)$ fulfilling $r_\ell < e'^\mathcal{A}$. Consequently, \ref{enum:PartitionRepresentatives:Three} is satisfied if we choose $c_{[\<P,i\>], p} := \alpha_{e+\varepsilon}$. Finally, we may conclude $\alpha_{e+\varepsilon}^\mathcal{A} \in p$, because $\text{\upshape{Ax}}_{\mathcal{I}_{[\<P,i\>],N}, \text{\upshape{bconsts}}(N)}$ contains the axiom $e < \alpha_{e+\varepsilon}$.
\end{description}
\end{proof}
We have arrived at one of the core results of the present paper. The next lemma shows that we can eliminate base-sort variables $x$ from clauses $C$ in a clause set $N$ by replacing $C$ with finitely many instances in which $x$ is substituted with the instantiation points that we computed for $x$. The resulting clause set shall be called $\widehat{N}_x$. In addition, the axioms that stipulate the meaning of newly introduced constant symbols $\alpha_\t$ need to be added to $\widehat{N}_x$. Iterating this step for every base-sort variable in $N$ eventually leads to a clause set that is essentially ground with respect to the constraint parts of the the clauses it contains (free-sort variables need to be treated separately, of course, see Section \ref{section:FreeModelTheory}).
The line of argument leading from a hierarchic model of the original clause set $N$ to a hierarchic model of the modified clause set $\widehat{N}_x$ (with instances and additional axioms) is almost trivial. The converse direction, however, rests on a model construction that yields a hierarchic model $\mathcal{B}$ which complies with property \ref{enum:IntervalModel} discussed in the beginning of the current section. Given a hierarchic model $\mathcal{A}$ of the instantiated clause set $\widehat{N}_x$, we construct the partitioning $\P_{\text{\upshape{ap}}_N(x), N, \mathcal{A}}$ based on the argument position class associated to $x$ in the original clause set $N$. By virtue of Lemma \ref{lemma:PartitionRepresentatives}, we know that each partition $p$ in the partitioning is represented by an instantiation point $c_p$ in $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ and for each of these instantiation points there is one clause in the modified clause set $\widehat{N}_x$ in which $x$ is instantiated by $c_p$.
Since $\mathcal{B}$ is supposed to comply with \ref{enum:IntervalModel}, i.e.\ it shall not distinguish between real values that stem from the same partition, the information how the model $\mathcal{A}$ treats the representative of a partition can be transferred to all values from this partition.
An example might be best suited to illustrate the key ideas.
\begin{example}
Consider the clause set $N = \bigl\{x > 2,\, z= 4 \;\|\; Q(x, z) \to T(x)\bigr\}$.
With respect to $N$ there are two instantiation points for variable $x$, namely $\alpha_{-\infty}$ and $\alpha_{2+\varepsilon}$. This will lead to the set of instances and axioms
\begin{align*}
\widehat{N}_x =\; &\bigl\{ \alpha_{-\infty} > 2,\, z= 4,\, x= \alpha_{-\infty} \;\|\; Q(x, z) \to T(x) \;,\\
&\quad \alpha_{2+\varepsilon} > 2,\, z= 4,\, x= \alpha_{2+\varepsilon} \;\|\; Q(x, z) \to T(x) \bigr\}\\
&\cup\; \text{\upshape{Ax}}_{\{\alpha_{-\infty}, \alpha_{2+\varepsilon}\}, \{2,4\}}
\end{align*}
where the axioms express the fact that the value of $\alpha_{-\infty}$ is strictly smaller than $2$ and the value of $\alpha_{2+\varepsilon}$ shall lie within the interval $(2, 4)$ in any hierarchic model of $\widehat{N}_x$. This already reveals redundancy of the first instance in the presence of the axioms, since the atomic constraint $\alpha_{-\infty} > 2$ is false under every hierarchic model of the axiom $\alpha_{-\infty} < 2$ that is contained in $\text{\upshape{Ax}}_{\{\alpha_{-\infty}, \alpha_{2+\varepsilon}\}, \{2,4\}}$.
We assume to have a hierarchic model $\mathcal{A}$ of $\widehat{N}_x$ at hand with $\alpha_{-\infty}^\mathcal{A} = 0$ and $\alpha_{2+\varepsilon}^\mathcal{A} = 3$. Then the two intervals $(-\infty, 2]$ and $(2,+\infty)$ constitute the partitioning $\P_{\text{\upshape{ap}}_N(x), N, \mathcal{A}}$, and the partitioning $\P_{\text{\upshape{ap}}_N(z), N, \mathcal{A}}$ is given by $\bigl\{(-\infty, 4), [4, +\infty)\bigr\}$. Moreover, assume that the set $Q^\mathcal{A}$ contains the pairs $\<0, 0\>$, $\<3,4\>$ and $\<7,4\>$ while the set $T^\mathcal{A}$ shall be the union of intervals $(0, 5] \cup [6,7)$. Clearly, $\mathcal{A}$ is a hierarchic model of $\widehat{N}_x$. (However, $\mathcal{A}$ is not a hierarchic model of $N$, since the pair $\<7,4\>$ is in $Q^\mathcal{A}$ but $7$ does not belong to $T^\mathcal{A}$.)
We now construct a hierarchic model $\mathcal{B}$ from $\mathcal{A}$ so that $\mathcal{B}$ complies with \ref{enum:IntervalModel}.
First of all, $\mathcal{B}$ takes over all values assigned to constant symbols, i.e.\ $\alpha_{-\infty}^\mathcal{B} := \alpha_{-\infty}^\mathcal{A}$ and $\alpha_{2+\varepsilon}^\mathcal{B} := \alpha_{2+\varepsilon}^\mathcal{A}$. The definition of $T$ under $\mathcal{B}$ will be piecewise with respect to the partitioning $\P_{[\<T,1\>], N, \mathcal{A}} = \P_{\text{\upshape{ap}}_N(x), N, \mathcal{A}} = \{(-\infty, 2], (2, +\infty)\}$. To do so, we use the idea of representatives: the interval $(-\infty, 2]$ is represented by $\alpha_{-\infty}^\mathcal{A} = 0$ and the interval $(2, +\infty)$ has $\alpha_{2+\varepsilon}^\mathcal{A} = 3$ as its representative. At the same time, $\alpha_{-\infty}^\mathcal{A}$ also represents the interval $(-\infty, 4)$ in partitioning $\P_{[\<Q,2\>], N, \mathcal{A}}$, and $[4, +\infty)$ from the same partitioning is represented by the constant $4$.
%
Putting the abstract idea of property \ref{enum:IntervalModel} into action, the set $T^\mathcal{B}$ will be defined so that $(-\infty, 2] \subseteq T^\mathcal{B}$ if and only if $\alpha_{-\infty}^\mathcal{A} \in T^\mathcal{A}$, and $(2, +\infty) \subseteq T^\mathcal{B}$ if and only if $\alpha_{2+\varepsilon}^\mathcal{A} \in T^\mathcal{A}$. For the set $Q^\mathcal{B}$ matters are technically more involved but follow the same scheme. For instance, all pairs $\<r_1, r_2\>$ with $r_1 \in (2, +\infty)$ and $r_2 \in [4, +\infty)$ shall be in $Q^\mathcal{B}$ if and only if the pair $\<\alpha_{2+\varepsilon}^\mathcal{A}, 4\>$ is in $Q^\mathcal{A}$.
Consequently, we end up with $T^\mathcal{B} = (2, +\infty]$ and $Q^\mathcal{B} = \{ \<r_1, r_2\> \mid r_1 \in (-\infty, 2] \text{ and } r_2 \in (-\infty, 4) \} \cup \{\<r_1, r_2\> \mid r_1 \in (2, +\infty) \text{ and } r_2 \in [4, +\infty)\}$.
The net results is a hierarchic interpretation $\mathcal{B}$ that is a model of $\widehat{N}_x$ just as $\mathcal{A}$ is. But beyond that $\mathcal{B}$ is also a hierarchic model of the original clause set $N$. This is by no means a coincidence as the proof of Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} shows.
\end{example}
\begin{lemma}\label{lemma:EquisatisfiabilityBaseInstantiation}
Let $N$ be a clause set in normal form, and assume that $N$ contains all the axioms in $\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(N), \text{\upshape{bconsts}}(N)}$ and for every $\alpha_{e+\varepsilon} \in \text{\upshape{$\alpha$consts}}(N)$ we have $e\in\text{\upshape{bconsts}}(N)$.
Suppose there is a clause $C$ in $N$ which contains a base-sort variable $x$. Let the clause set $\widehat{N}_x$ be constructed as follows: $\widehat{N}_x := \bigl(N\setminus\{ C \}\bigr) \;\cup\; \bigl\{C\subst{x}{c} \;\bigm|\; c\in\mathcal{I}_{\text{\upshape{ap}}_N(x), N} \bigr\} \;\cup\; \text{\upshape{Ax}}_{\mathcal{I}_{\text{\upshape{ap}}_N(x), N}, \text{\upshape{bconsts}}(N)}$.
The original clause set $N$ is satisfiable if and only if $\widehat{N}_x$ is satisfiable.
\end{lemma}
\begin{proof}
The ``only if''-part is almost trivial to show:
Let $\mathcal{A}$ be a hierarchic model of $N$.
Every axiom in $\text{\upshape{Ax}}_{\mathcal{I}_{\text{\upshape{ap}}_N(x), N}, \text{\upshape{bconsts}}(N)}$ which does not appear in $N$ concerns constant symbols $\alpha_\t$ that do not occur in $N$. By virtue of Proposition \ref{proposition:ImproperAxiomSatisfiability}, we can derive a hierarchic model $\mathcal{B}$ from $\mathcal{A}$ that differs from $\mathcal{A}$ only in how these $\alpha_\t$ are interpreted and that is a hierarchic model of $\text{\upshape{Ax}}_{\mathcal{I}_{\text{\upshape{ap}}_N(x), N}, \text{\upshape{bconsts}}(N)} \setminus N$. Thus, $\mathcal{B}\models N$ follows from $\mathcal{A}\models N$.
Since the variable $x$ is universally quantified in $C$, $\mathcal{B}\models C$ entails $\mathcal{B}\models C\subst{x}{c}$ for every instance $C\subst{x}{c}$.
Altogether, we obtain $\mathcal{B}\models \widehat{N}_x$.
The ``if''-part requires a more sophisticated argument:
Let $\mathcal{A}$ be a hierarchic model of $\widehat{N}_x$.
We use $\mathcal{A}$ to construct the hierarchic model $\mathcal{B}$ that extends a model in $\mathfrak{M}$ as follows.
For the domain $\S^\mathcal{B}$ we reuse $\mathcal{A}$'s free domain $\S^\mathcal{A}$.
For all base-sort and free-sort constant symbols $c\in \text{\upshape{consts}}(N)$, we set $c^{\mathcal{B}} := c^{\mathcal{A}}$.
For every predicate symbol $P:\xi_1\times\ldots\times\xi_m \in \Pi$ that occurs in $N$ and for every argument position $i$, $1\leq i\leq m$, Lemma \ref{lemma:PartitionRepresentatives} guarantees the existence of a base-sort constant symbol $c_{[\<P,i\>], p}\in \mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ for every partition $p\in \P_{[\<P,i\>], N, \mathcal{A}}$, such that $c_{[\<P,i\>], p}^\mathcal{A} \in p$ and the requirements \ref{enum:PartitionRepresentatives:One} to \ref{enum:PartitionRepresentatives:Three} of the lemma are met.
%
Based on this, we define the family of functions $\varphi_{[\<P,i\>]} : \mathbb{R}\cup\S^\mathcal{B} \to \mathbb{R}\cup\S^\mathcal{A}$ by
\[ \varphi_{[\<P,i\>]}(a) :=
\begin{cases}
c_{[\<P,i\>], p}^\mathcal{A} &\text{if $\xi_i = \mathcal{R}$ and $p\in \P_{[\<P,i\>], N, \mathcal{A}}$}\\
&\text{is the partition $a$ lies in,} \\
a &\text{if $\xi_i = \S$.}
\end{cases}
\]
Using the functions $\varphi_{[\<P,i\>]}$, we define the $P^\mathcal{B}$ so that for all domain elements $a_1,\ldots,a_m$ of appropriate sorts $\bigl\<a_1, \ldots, a_m\bigr\> \in P^{\mathcal{B}}$ if and only if $\bigl\<\varphi_{[\<P,1\>]}(a_1), \ldots,$ $\varphi_{[\<P,m\>]}(a_m)\bigr\> \in P^\mathcal{A}$.
We next show $\mathcal{B}\models N$. Consider any clause $C' = \Lambda' \;\|\; \Gamma' \to \Delta'$ in $N$ and let $\beta : V_\mathcal{R}\cup V_\S\to \mathbb{R}\cup \S^\mathcal{B}$ be an arbitrary variable assignment.
From $\beta$ we derive a special variable assignment $\beta_\varphi$ for which we will infer $\mathcal{A},\beta_\varphi \models C'$ as an intermediate step:
\[ \beta_\varphi(v) :=
\begin{cases}
c_{\text{\upshape{ap}}_N(v), p}^\mathcal{A} &\text{if $v\in V_\mathcal{R}$ and $\beta(v)\in p \in \P_{\text{\upshape{ap}}_N(v), N, \mathcal{A}}$,}\\
\beta(v) &\text{if $v\in V_\S$,}
\end{cases}
\]
for every variable $v$.
If $C' \neq C$, then $\widehat{N}_x$ already contains $C'$, and thus $\mathcal{A},\beta_\varphi\models C'$ must hold.
In case of $C' = C$, let $p_*$ be the partition in $\P_{\text{\upshape{ap}}_N(x), N, \mathcal{A}}$ containing the value $\beta(x)$, and let $c_*$ be an abbreviation for $c_{\text{\upshape{ap}}_N(x), p_*}$. Due to $\beta_\varphi(x) = c_*^\mathcal{A}$ and since $\mathcal{A}$ is a model of the clause $C\subst{x}{c_*}$ in $\widehat{N}_x$, we conclude $\mathcal{A},\beta_\varphi \models C$.
%
Hence, in any case we can deduce $\mathcal{A},\beta_\varphi\models C'$. By case distinction on why $\mathcal{A},\beta_\varphi \models C'$ holds, we may transfer this result to obtain $\mathcal{B},\beta\models C'$, too.
\begin{description}
\item Case $\mathcal{A}, \beta_\varphi \not\models c\triangleleft d$ for some atomic constraint $c\triangleleft d$ in $\Lambda'$ with base-sort constant symbols $c, d \in \Omega_\text{\upshape{LA}}\cup\Omega_\text{\upshape{LA}}^\alpha$ and $\triangleleft \in \{<, \leq, =, \not=, \geq, >\}$. Since $\mathcal{B}$ and $\mathcal{A}$ interpret constant symbols in the same way and independently of a variable assignment, we immediately get $\mathcal{B}, \beta\not\models c\triangleleft d$.
\item Case $\mathcal{A},\beta_\varphi \not\models (y\triangleleft d) \in \Lambda'$ for an arbitrary base-sort variable $y$ and a constant symbol $d\in\Omega_\text{\upshape{LA}}$. This translates to $\beta_\varphi(y) {\not\!\triangleleft}\, d^\mathcal{A}$. Let $p\in\P_{\text{\upshape{ap}}_N(y), N, \mathcal{A}}$ be the partition which contains $\beta(y)$ and therefore also $\beta_\varphi(y)$.
\begin{description}
\item If $d^\mathcal{A}$ lies outside of $p$, then $\beta_\varphi(y) \triangleleft d^\mathcal{A}$ if and only if $\beta(y) \triangleleft d^\mathcal{A}$, since $\beta_\varphi(y) \in p$. Thus, $d^\mathcal{B} = d^\mathcal{A}$ entails $\mathcal{B}, \beta \not\models y\triangleleft d$.
\item If $p$ is the point interval $p = \{d^\mathcal{A}\}$, then $\beta(y) = \beta_\varphi(y)$, and thus $\mathcal{B},\beta \not\models y\triangleleft d$.
\item If $p = \mbox{$(\!\cdot$} r_\ell, r_u\mbox{$\cdot\!)$} $ and $r_\ell < d^\mathcal{A} \leq r_u$, then $\triangleleft \not\in \{<, \leq, \not=\}$, since $\beta_\varphi(y) = c_{\text{\upshape{ap}}_N(y), p}^\mathcal{A} < d^\mathcal{A}$ (by \ref{enum:PartitionRepresentatives:Two} and \ref{enum:PartitionRepresentatives:Three} of Lemma \ref{lemma:PartitionRepresentatives}). Moreover, we conclude $d\not\in \mathcal{I}_{\text{\upshape{ap}}_N(y), N}$, since otherwise $p$ would be of the form $p = [d^\mathcal{A}, r_u\mbox{$\cdot\!)$} $ by the construction of $\P_{\text{\upshape{ap}}_N(y),N,\mathcal{A}}$ (requirements (\ref{enum:BaseVariableInstantiationPoints:Two}), \ref{enum:BaseVariableInstantiationPoints:ThreeThree} and \ref{enum:BaseVariableInstantiationPoints:ThreeFour}). Therefore, $\triangleleft \not\in \{=, \geq\}$, since otherwise the instantiation point $d$ would be in $\mathcal{I}_{\text{\upshape{ap}}_N(y), N}$. This only leaves $\triangleleft = {>}$. Hence, the constraint $y > d$ occurs in $N$ and thus we find the instantiation point $\alpha_{d+\varepsilon}$ in $\mathcal{I}_{\text{\upshape{ap}}_N(y), N}$. Consequently, requirements \ref{enum:BaseVariableInstantiationPoints:ThreeTwo} and \ref{enum:BaseVariableInstantiationPoints:ThreeFour} of the construction of $\P_{\text{\upshape{ap}}_N(y),N,\mathcal{A}}$ entail $p = \mbox{$(\!\cdot$} r_\ell, d^\mathcal{A}]$, which leads to $\beta(y) \leq d^\mathcal{A}$ because of $\beta(y)\in p$. Therefore, $\mathcal{B},\beta \not\models y > d$ must be true.
The cases $p = \mbox{$(\!\cdot$} r_\ell, +\infty)$, $p = (-\infty, r_u\mbox{$\cdot\!)$} $ and $(-\infty, +\infty)$ with $r_\ell < d^\mathcal{A} \leq r_u$ can be handled by similar arguments.
\item If $p = [d^\mathcal{A}, r_u\mbox{$\cdot\!)$} $ and $d^\mathcal{A} < r_u$, then $\beta_\varphi(y) = c_{\text{\upshape{ap}}_N(y), p}^\mathcal{A} = d^\mathcal{A}$ by \ref{enum:PartitionRepresentatives:Two} of Lemma \ref{lemma:PartitionRepresentatives}. Consequently, \mbox{$\triangleleft \not\in \{\leq, =, \geq\}$}. We conclude $\alpha_{d+\varepsilon} \not\in \mathcal{I}_{\text{\upshape{ap}}_N(y), N}$, because otherwise $[d^\mathcal{A}, r_u\mbox{$\cdot\!)$} $ would be a point interval, contradicting $d^\mathcal{A} < r_u$. Hence, the only remaining possibility is $\triangleleft = {<}$. But by $\beta(y)\in p$ we deduce $\beta(y) \geq d^\mathcal{A}$. Therefore, we clearly get $\mathcal{B},\beta \not\models y < d$.
The case $p = [d^\mathcal{A}, +\infty)$ is covered by analogous arguments.
\end{description}
\item Case $\mathcal{A}, \beta_\varphi \not\models (y= \alpha_\t) \in \Lambda'$ for an arbitrary variable $y$ and a constant symbol $\alpha_\t\in\Omega_\text{\upshape{LA}}^\alpha$.
Let $p\in\P_{\text{\upshape{ap}}_N(y), N, \mathcal{A}}$ be the partition which contains $\beta(y)$.
We immediately conclude $\alpha_\t^\mathcal{A} \neq c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A}$ since $c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A} = \beta_\varphi(y) \neq \alpha_\t^\mathcal{A}$.
\begin{description}
\item If $\alpha_\t^\mathcal{A}$ does not lie in $p$, then $\beta(y) \neq \alpha_\t^\mathcal{A}$, and thus $\mathcal{B},\beta \not\models y= \alpha_\t$.
\item Assume $\alpha_\t^\mathcal{A}$ lies in $p$.
By $c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A} \in p$ and the facts $\alpha_\t^\mathcal{A} \neq c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A}$ and $\alpha_\t^\mathcal{A} \in p$, $p$ cannot be a point interval.
Moreover, $p$ cannot be of the form $(-\infty, r_u\mbox{$\cdot\!)$} $, since then $c_{\text{\upshape{ap}}_N(y),p} = \alpha_{-\infty}$ and $r_u < \alpha_{c+\varepsilon}^\mathcal{A}$ for all $\alpha_{c+\varepsilon} \in \text{\upshape{$\alpha$consts}}(N)$ would follow by construction of $\P_{\text{\upshape{ap}}_N(y),N,\mathcal{A}}$ (since $\mathcal{A}\models \text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(N), \text{\upshape{bconsts}}(N)}$), contradicting either $\alpha_\t \neq c_{\text{\upshape{ap}}_N(y),p}$ or $\alpha_\t^\mathcal{A} \in p$.
Consequently, there exists a constant symbol $d\in\text{\upshape{bconsts}}(N)$ so that $\alpha_\t = \alpha_{d+\varepsilon}$. Hence, the atomic constraint $y= \alpha_{d+\varepsilon}$ contributes the instantiation point $\alpha_{d+\varepsilon}$ to $\mathcal{I}_{\text{\upshape{ap}}_N(y),N}$.
The constant symbol $c_{\text{\upshape{ap}}_N(y),p}$ cannot stem from $\Omega_\text{\upshape{LA}}$, since then $p$ would be of the form $[c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A}, +\infty)$ or $[c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A}, r_u\mbox{$\cdot\!)$} $ for some real value $r_u$ and we would have $c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A} = d^\mathcal{A}$, i.e.\ $c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A}$ and $c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A} + \varepsilon$ would be $\mathcal{A}$-covered by $\mathcal{I}_{\text{\upshape{ap}}_N(y), N}$ -- a contradiction, as we have already argued then $p$ cannot be a point interval.
Hence, there exists a constant symbol $e\in\text{\upshape{bconsts}}(N)$ so that $e \neq d$ and $c_{\text{\upshape{ap}}_N(y),p} = \alpha_{e+\varepsilon}$ and thus $p$ is either of the form $(e^\mathcal{A}, +\infty)$ or $(e^\mathcal{A}, r_u\mbox{$\cdot\!)$} $ for some real value $r_u$.
In case of $d^\mathcal{A} \neq e^\mathcal{A}$, the instantiation point $\alpha_{d+\varepsilon}$ results in a partition $p' \in \P_{\text{\upshape{ap}}_N(y),N,\mathcal{A}}$ such that $\alpha_{d+\varepsilon}^\mathcal{A} \in p'$ and $p' \neq p$. But this contradicts our assumption $\alpha_\t^\mathcal{A} = \alpha_{e+\varepsilon}^\mathcal{A} \in p$.
Consequently, $d^\mathcal{A} = e^\mathcal{A}$. As $\widehat{N}_x$ contains $\text{\upshape{Ax}}_{\{\alpha_{d+\varepsilon}\}, \{e\}}$ as a subset, $\mathcal{A}$ must be a hierarchic model of the axiom $d = e \to \alpha_{d+\varepsilon} = \alpha_{e+\varepsilon}$. This yields a contradiction, since above we concluded $\alpha_{d+\varepsilon}^\mathcal{A} = \alpha_\t^\mathcal{A} \neq c_{\text{\upshape{ap}}_N(y),p}^\mathcal{A} = \alpha_{e+\varepsilon}^\mathcal{A}$.
\end{description}
\item Case $\mathcal{A}, \beta_\varphi \not\models s\approx s'$ for some free atom $s\approx s' \in \Gamma'$. Hence, $s$ and $s'$ are either variables or constant symbols of the free sort, which means they do not contain subterms of the base sort. Since $\mathcal{B}$ and $\mathcal{A}$ behave identical on free-sort constant symbols and $\beta(u) = \beta_\varphi(u)$ for any variable $u\in V_\S$, it must hold $\mathcal{B}, \beta \not\models s\approx s'$.
\item Case $\mathcal{A}, \beta_\varphi \models s\approx s'$ for some $s\approx s' \in \Delta'$. Analogous to the above case, $\mathcal{B}, \beta \models s\approx s'$ holds.
\item Case $\mathcal{A}, \beta_\varphi \not\models P(s_1, \ldots, s_m)$ for some free atom $P(s_1, \ldots, s_m)\in\Gamma'$. This translates to\\ $\bigl\<\mathcal{A}(\beta_\varphi)(s_1), \ldots, \mathcal{A}(\beta_\varphi)(s_m)\bigr\> \not\in P^\mathcal{A}$.
\begin{description}
\item Every $s_i$ of the free sort is either a constant symbol or a variable. Thus, we have $\mathcal{A}(\beta_\varphi)(s_i) = \mathcal{B}(\beta)(s_i) = \varphi_{[\<P,i\>]}(\mathcal{B}(\beta)(s_i))$, since free-sort constant symbols are interpreted in the same way by $\mathcal{A}$ and $\mathcal{B}$, and because $\beta_\varphi(u) = \beta(u)$ for every free-sort variable $u$.
\item Every $s_i$ that is of the base sort must be a variable. Hence, $\mathcal{A}(\beta_\varphi)(s_i) = c_{[\<P,i\>],p}^\mathcal{A} = \varphi_{[\<P,i\>]}(\mathcal{B}(\beta)(s_i))$, where $p$ is the partition in $\P_{[\<P,i\>], N, \mathcal{A}}$ which contains $\beta(s_i)$ (and thus also $\beta_\varphi(s_i)$) and where we have $\text{\upshape{ap}}_N(s_i) = [\<P,i\>]$.
\end{description}
Put together, this yields $\bigl\<\varphi_{[\<P,1\>]}(\mathcal{B}(\beta)(s_1)), \ldots, \varphi_{[\<P,m\>]}(\mathcal{B}(\beta)(s_m))\bigr\> \not\in P^\mathcal{A}$. But then, by construction of $\mathcal{B}$, we have $\bigl\<\mathcal{B}(\beta)(s_1), \ldots, \mathcal{B}(\beta)(s_m)\bigr\> \not\in P^\mathcal{B}$, which entails $\mathcal{B}, \beta \not\models P(s_1, \ldots,$ $s_m)$.
\item Case $\mathcal{A}, \beta_\varphi \models P(s_1, \ldots, s_m)$ for some free atom $P(s_1, \ldots, s_m)\in\Delta'$. Analogous to the above case we conclude $\mathcal{B}, \beta \models P(s_1, \ldots, s_m)$.
\end{description}
%
Altogether, we have shown $\mathcal{B}\models N$.
\end{proof}
We have already pointed out that we intend to iteratively instantiate base-sort variables in an initial clause set $N$ by means of the construction described in Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation}. However, for efficiency reasons it is not desirable to recompute all necessary information such as argument position classes, sets of instantiation points and the like from scratch at each stage. The next lemma shows that the ingredients for the elimination of base-sort variables are invariant under instantiation. The main reason is that when $x$ is instantiated, constraints $x = c$ are introduced to the clause and free atoms $P(\ldots, x, \ldots)$ remain untouched (however, $x$ might be renamed afterwards). Consequently, it is indeed sufficient to compute, for instance, the relation $\leftrightharpoons_N$ and the set of instantiation points $\mathcal{I}_{[\<P,i\>],N}$ once and for all at the beginning for every argument position pair $\<P,i\>$ that is of interest and only use it at the appropriate stages of the overall instantiation process. How this can be done will be described in detail in Section \ref{section:Complexity}.
\begin{lemma}\label{lemma:IterativeInstantiation}
Let $N$ be a clause set in normal form that contains a clause $C$ in which a base-sort variable $x$ occurs. Further assume $\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(N), \text{\upshape{bconsts}}(N)} \subseteq N$ and for every $\alpha_{e+\varepsilon} \in \text{\upshape{$\alpha$consts}}(N)$ we have $e\in\text{\upshape{bconsts}}(N)$. Suppose $\widehat{N}_x$ is constructed from $N$ as described in Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} and variables have been renamed so that all clauses in $\widehat{N}_x$ are variable disjoint.
%
We observe the following facts:
\begin{enumerate}[label=(\roman{*}), ref=(\roman{*})]
\item\label{enum:IterativeInstantiation:One} ${\leftrightharpoons_N} = {\leftrightharpoons_{\widehat{N}_x}}$.
\item\label{enum:IterativeInstantiation:Two} $\text{\upshape{bconsts}}(N) = \text{\upshape{bconsts}}(\widehat{N}_x)$ and $\text{\upshape{fconsts}}(N) = \text{\upshape{fconsts}}(\widehat{N}_x)$.
\item\label{enum:IterativeInstantiation:Three} For every argument position class $[\<P,i\>]$ induced by $\leftrightharpoons_{N}$ we have $\mathcal{I}_{[\<P,i\>], N} = \mathcal{I}_{[\<P,i\>], \widehat{N}_x}$.
\item\label{enum:IterativeInstantiation:Four} $\widehat{N}_x$ is in normal form, $\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(\widehat{N}_x), \text{\upshape{bconsts}}(\widehat{N}_x)} \subseteq \widehat{N}_x$, and for every $\alpha_{e+\varepsilon} \in \text{\upshape{$\alpha$consts}}(\widehat{N}_x)$ we have $e\in \text{\upshape{bconsts}}(\widehat{N}_x)$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove the different parts separately.\medskip
\noindent\ref{enum:IterativeInstantiation:One}:
The relations $\leftrightharpoons_N$ and $\leftrightharpoons_{\widehat{N}_x}$ exclusively depend on the free parts of the clauses in $N$ and $\widehat{N}_x$, respectively, irrespective of variable names. Since $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ is nonempty, all free parts of clauses in $N$ do occur in $\widehat{N}_x$ (modulo variable renaming), and vice versa.\medskip
\noindent\ref{enum:IterativeInstantiation:Two}:
$\text{\upshape{bconsts}}(N) \subseteq \text{\upshape{bconsts}}(\widehat{N}_x)$ and $\text{\upshape{fconsts}}(N) \subseteq \text{\upshape{fconsts}}(\widehat{N}_x)$ hold due to $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ being nonempty. $\text{\upshape{bconsts}}(\widehat{N}_x) \subseteq \text{\upshape{bconsts}}(N)$ is a consequence of $\mathcal{I}_{\text{\upshape{ap}}_N(x), N} \cap \Omega_\text{\upshape{LA}} \subseteq \text{\upshape{bconsts}}(N)$. $\text{\upshape{fconsts}}(\widehat{N}_x) \subseteq \text{\upshape{fconsts}}(N)$ is true, because free parts of clauses are merely copied and variables renamed.\medskip
\noindent\ref{enum:IterativeInstantiation:Three}:
We start with two observations:
\begin{enumerate}[label=(\arabic{*}),ref=(\arabic{*})]
\item\label{enum:proofComplexity:Three:One} For every clause $D \neq C$ in $N$, we find a clause $\widehat{D}$ in $\widehat{N}_x$ such that $D$ and $\widehat{D}$ are the same modulo variable renaming. Conversely, every $\widehat{D}$ in $\widehat{N}_x$ is either identical to a clause $D$ in $N$ modulo renaming of variables or it is equal to $C\subst{x}{c}$ up to variable renaming for some constant symbol $c$.
\item\label{enum:proofComplexity:Three:Two} Regarding $C$ we find the instance $C\subst{x}{\alpha_{-\infty}}$ in $\widehat{N}_x$, possibly with renamed variables.
Consequently, for every base-sort variable $y\neq x$ for which $C$ contains an atomic constraint $y\triangleleft s$ and a free atom $Q(\ldots, y, \ldots)$ with $y$ in the $j$-th argument position, $C\subst{x}{\alpha_{-\infty}}$ contains the constraint $y\triangleleft s$ and the free atom $Q(\ldots, y, \ldots)$ with $y$ in the $j$-th argument position -- again, modulo variable renaming.
The converse also holds.
\end{enumerate}
That said, we distinguish two cases for every argument position pair $\<P,i\>$.
\begin{description}
\item If $\<P,i\>$ does not belong to the equivalence class $\text{\upshape{ap}}_N(x)$, i.e.\ $\text{\upshape{ap}}_N(x) \neq [\<P,i\>]$, then
we get $\mathcal{I}_{[\<P,i\>],N} = \mathcal{I}_{[\<P,i\>],\widehat{N}_x}$, since \ref{enum:proofComplexity:Three:One} and \ref{enum:proofComplexity:Three:Two} together entail $\mathcal{I}_{Q,j,N} = \mathcal{I}_{Q,j,\widehat{N}_x}$ for all $\<Q,j\> \in [\<P,i\>]$.
\item If $[\<P,i\>] = \text{\upshape{ap}}_N(x)$, then there is an argument position pair $\<Q,j\> \in [\<P,i\>]$ so that $C$ contains a free atom $Q(\ldots, x, \ldots)$ in which $x$ is the $j$-th argument. Moreover, $\widehat{N}_x$ is a superset of $\bigl\{ C\subst{x}{c} \bigm| c\in\mathcal{I}_{\text{\upshape{ap}}_N(x), N} \bigr\}$ modulo renaming of variables. This entails $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}\setminus\{\alpha_{-\infty}\} \subseteq \mathcal{I}_{Q,j,\widehat{N}_x}$, and therefore, $\mathcal{I}_{\text{\upshape{ap}}_N(x), N} \subseteq \mathcal{I}_{[\<P,i\>], \widehat{N}_x}$.
On the other hand, assume there is an instantiation point in $\mathcal{I}_{[\<P,i\>], \widehat{N}_x}$ that is not in $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$, i.e.\ there is a clause $\widehat{D}$ in $\widehat{N}_x$ of which a variable-renamed variant has not already been in $N$, and in which an atomic constraint $y\triangleleft s$ and a free atom $Q(\ldots, y, \ldots)$ occur with $y$ in the $j$-th argument position. By \ref{enum:proofComplexity:Three:One}, $\widehat{D}$ must be a variable-renamed variant of $C\subst{x}{c}$ for some constant symbol $c$. But according to \ref{enum:proofComplexity:Three:Two}, we must have $(y\triangleleft s) = (x = c)$ modulo variable renaming. But this constraint can only lead to an instantiation point that is contained in $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ -- a contradiction. Hence, we also have $\mathcal{I}_{[\<P,i\>], \widehat{N}_x} \subseteq \mathcal{I}_{\text{\upshape{ap}}_N(x), N}$.
Put together, we just derived $\mathcal{I}_{[\<P,i\>], N} = \mathcal{I}_{[\<P,i\>], \widehat{N}_x}$ for this case.
\end{description}
\noindent\ref{enum:IterativeInstantiation:Four}:
The construction of $\widehat{N}_x$ from $N$ preserves the normal form property for each clause individually, since we start from a clause in normal form, copy the free part of the clause and no new variables are introduced to the constraint part which do not also occur in the free part. Afterwards, consistent renaming of variables ensures pairwise variable disjointness of the clauses in $\widehat{N}_x$.
$\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(\widehat{N}_x), \text{\upshape{bconsts}}(\widehat{N}_x)} \subseteq \widehat{N}_x$ is a consequence of $\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(N), \text{\upshape{bconsts}}(N)} \subseteq N\setminus \{C\}$ (true by assumption and the fact that $C$'s free part is not empty but the free parts of axioms are) and $\text{\upshape{Ax}}_{\mathcal{I}_{\text{\upshape{ap}}_N(x), N}, \text{\upshape{bconsts}}(N)} \subseteq \widehat{N}_x$ (true by construction). Moreover, $\text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(\widehat{N}_x), \text{\upshape{bconsts}}(\widehat{N}_x)} = \text{\upshape{Ax}}_{\text{\upshape{$\alpha$consts}}(N), \text{\upshape{bconsts}}(N)} \cup \text{\upshape{Ax}}_{\mathcal{I}_{\text{\upshape{ap}}_N(x), N}, \text{\upshape{bconsts}}(N)}$ is ensured by \ref{enum:IterativeInstantiation:Two} and the fact that $\text{\upshape{$\alpha$consts}}(\widehat{N}_x) \subseteq \text{\upshape{$\alpha$consts}}(N)\cup\mathcal{I}_{\text{\upshape{ap}}_N(x),N}$.
Suppose there is a constant symbol $\alpha_{e+\varepsilon}$ in $\widehat{N}_x$ so that $e$ does not occur as constant symbol in $\widehat{N}_x$. Hence, $\alpha_{e+\varepsilon}$ must have been an instantiation point in $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$. Due to \ref{enum:IterativeInstantiation:Two}, $e$ cannot occur in $N$ either. Consequently, $\alpha_{e+\varepsilon}$ has been introduced to $\mathcal{I}_{\text{\upshape{ap}}_N(x), N}$ by the occurrence of an atomic constraint $y = \alpha_{e+\varepsilon}$ in some clause in $N$. But we assumed $e' \in \text{\upshape{bconsts}}(N)$ for every $\alpha_{e'+\varepsilon} \in \text{\upshape{$\alpha$consts}}(N)$ -- a contradiction.
\end{proof}
\section{Instantiation of Free-Sort Variables}\label{section:FreeModelTheory}
Checking satisfiability of a finite clause set $N$ over the classical Bernays-Sch\"onfinkel fragment can be done na\"ively by trying all Herbrand interpretations (we assume $N$ contains at least one constant symbol). One key argument is that the domain of these interpretations is the set of all constant symbols occurring in $N$ and is thus finite.
If we add equality to the fragment, the canonical interpretations are Herbrand interpretations modulo congruences on the occurring constant symbols. Although this means we have to consider more than one domain, there are still only finitely many of them. Moreover, their size is upper bounded by the number of constant symbols in $N$.
This idea of canonical domains can easily be transferred to hierarchic models over the BSR fragment with simple bounds. More precisely, we can prove that any clause set $N$ over the BSR fragment with simple bounds is satisfiable w.r.t.\ $\mathfrak{M}$ if and only if there exists a hierarchic model $\mathcal{A}$ so that its free domain is the set of free-sort constant symbols in $N$ modulo a congruence relation $\sim$, i.e.\ $\S^\mathcal{A} = \text{\upshape{fconsts}}(N)/_\sim$, and so that every free-sort constant symbol $c$ is interpreted by the corresponding congruence class $[c]_\sim$.
\begin{lemma}\label{lemma:FreeDomain}
Let $N$ be a clause set in normal form and let $\mathcal{A}$ be a hierarchic model of $N$. We assume $\text{\upshape{fconsts}}(N)$ to contain at least one constant symbol (otherwise we may add the tautology $\|\to c\approx c$ to $N$).
\noindent By $\widehat{\S^\mathcal{A}}$ we denote the restricted domain $\{a \in \S^\mathcal{A} \mid \text{there is a $d\in\text{\upshape{fconsts}}(N)$ such that $a = d^\mathcal{A}$}\}$.
\noindent Let $\sim_\mathcal{A}$ be the binary relation on $\text{\upshape{fconsts}}(N)$ satisfying $c \sim_\mathcal{A} d$ if and only if $c^\mathcal{A} = d^\mathcal{A}$.
\noindent We can construct a hierarchic model $\mathcal{B}$ of $N$ which meets the following requirements:
\begin{enumerate}[label=(\roman{*}), ref=(\roman{*})]
\item\label{enum:FreeDomain:One} the domain $\S^\mathcal{B}$ is the set $\text{\upshape{fconsts}}(N)/_{\sim_\mathcal{A}}$ of equivalence classes w.r.t.\ $\sim_\mathcal{A}$;
\item\label{enum:FreeDomain:Two} $d^\mathcal{B} = [d]_{\sim_\mathcal{A}}$ for every free-sort constant symbol $d\in\text{\upshape{fconsts}}(N)$;
\item\label{enum:FreeDomain:Three} $c^\mathcal{B} = c^\mathcal{A}$ for every base-sort constant symbol $c\in\text{\upshape{bconsts}}(N)\cup\text{\upshape{$\alpha$consts}}(N)$; and
\item\label{enum:FreeDomain:Four} for every atom $A$ in $N$ and every variable assignment $\beta : V_\S \cup V_\mathcal{R} \to \S^\mathcal{A}\cup \mathbb{R}$ for which $\beta(u) \in \widehat{S^\mathcal{A}}$ for all $u\in V_\S$, we have $\mathcal{A}, \beta\models A$ if and only if $\mathcal{B}, \beta_{\sim_\mathcal{A}} \models A$, where for any variable $v \in V_\mathcal{R}\cup V_\S$ we set
\[ \beta_{\sim_\mathcal{A}}(v) :=
\begin{cases}
[d]_{\sim_\mathcal{A}} &\text{if $v\in V_\S$ and $\beta(v) = d^\mathcal{A}$}\\
&\text{for some $d\in\text{\upshape{fconsts}}(N)$,}\\
\beta(v) &\text{if $v\in V_\mathcal{R}$.}
\end{cases}
\]
\end{enumerate}
\end{lemma}
\begin{proof}
We construct the hierarchic interpretation $\mathcal{B}$ as follows:
\begin{itemize}
\item $S^\mathcal{B} := \text{\upshape{fconsts}}(N)/_{\sim_\mathcal{A}}$.
\item For every constant symbol $d\in\text{\upshape{fconsts}}(N)$, we set $d^\mathcal{B} := [d]_{\sim_\mathcal{A}}$.
\item For every constant symbol $c\in\text{\upshape{bconsts}}(N)\cup\text{\upshape{$\alpha$consts}}(N)$, we set $c^\mathcal{B} := c^\mathcal{A}$.
\item To help the formulation of the interpretation of predicate symbols under $\mathcal{B}$ we first define the function $\psi : \S^\mathcal{B} \cup \mathbb{R} \to \S^\mathcal{A} \cup \mathbb{R}$ by
\[ \psi(a) :=
\begin{cases}
d^\mathcal{A} &\text{if $a = [d]_{\sim_\mathcal{A}} \in \S^\mathcal{B}$ for some free-sort}\\
&\text{constant symbol $d$,}\\
a &\text{if $a\in \mathbb{R}$.}
\end{cases}
\]
(Please note that this is well-defined, since $d \sim_\mathcal{A} d'$ entails $d^\mathcal{A} = d'^\mathcal{A}$.)
For every predicate symbol $P/m$ that occurs in $N$ and for all arguments $a_1, \ldots, a_m$ of appropriate sorts, we define the interpretation of $P$ under $\mathcal{B}$ such that $\bigl\<a_1, \ldots, a_m\bigr\> \in P^{\mathcal{B}}$ if and only if $\bigl\< \psi(a_1), \ldots, \psi(a_m) \bigr\>\in P^\mathcal{A}$.
\end{itemize}
Obviously, requirements \ref{enum:FreeDomain:One}, \ref{enum:FreeDomain:Two} and \ref{enum:FreeDomain:Three} are satisfied.
Due to the fact that we find a $\beta$ whose image is a subset of $\widehat{S^\mathcal{A}}\cup\mathbb{R}$ (as described in \ref{enum:FreeDomain:Four}) for every variable assignment $\gamma : V_\mathcal{R}\cup V_\S \to \S^\mathcal{B}\cup\mathbb{R}$ so that $\beta_{\sim_\mathcal{A}} = \gamma$, a proof of \ref{enum:FreeDomain:Four} entails $\mathcal{B}\models N$. Hence, it remains to show \ref{enum:FreeDomain:Four}.
\begin{description}
\item Suppose $A$ is an atomic constraint $s\triangleleft t$. The definitions of $\mathcal{B}$ and $\beta_{\sim_\mathcal{A}}$ immediately imply the equivalence of $\mathcal{A},\beta\models A$ and $\mathcal{B},\beta_{\sim_\mathcal{A}}\models A$.
\item Suppose $A$ is of the form $s \approx t$, $s$ and $t$ being variables or constant symbols of the free sort, respectively. Because of $\mathcal{A}(\beta)(s), \mathcal{A}(\beta)(t)\in \widehat{S^\mathcal{A}}$, there exist constant symbols $d_s, d_t \in \text{\upshape{fconsts}}(N)$ such that $d_s^\mathcal{A} = \mathcal{A}(\beta)(s)$ and $d_t^\mathcal{A} = \mathcal{A}(\beta)(t)$. Consequently, $\mathcal{A}(\beta)(s) = \mathcal{A}(\beta)(t)$ holds if and only if $d_s \sim_\mathcal{A} d_t$, i.e.\ if and only if $[d_s]_{\sim_\mathcal{A}} = [d_t]_{\sim_\mathcal{A}}$. Put differently, $\mathcal{A}, \beta\models s \approx t$ is equivalent to $\mathcal{B},\beta_{\sim_\mathcal{A}}\models s \approx t$.
\item Suppose $A$ is of the form $P(t_1, \ldots, t_m)$ for some predicate symbol $P/m$ occurring in $N$. As we have already argued in the previous case, we can find for every $t_i$ of the free sort a constant symbol $d_i$ fulfilling $d_i^\mathcal{A} = \mathcal{A}(\beta)(t_i)$.
\begin{description}
\item If $t_i = c\in\text{\upshape{fconsts}}(N)$, then $\psi(\mathcal{B}(\beta_{\sim_\mathcal{A}})(c)) = \psi([c]_{\sim_\mathcal{A}}) = c^\mathcal{A} = \mathcal{A}(\beta)(c)$.
\item If $t_i = u\in V_\S$, then $\psi(\mathcal{B}(\beta_{\sim_\mathcal{A}})(u)) = \psi(\beta_{\sim_\mathcal{A}}(u)) = \psi([d_i]_{\sim_\mathcal{A}}) = d_i^\mathcal{A} = \beta(u) = \mathcal{A}(\beta)(u)$.
\end{description}
On the other hand, for all $t_i$ of the base sort, we get $\psi(\mathcal{B}(\beta_{\sim_\mathcal{A}})(t_i)) = \mathcal{B}(\beta_{\sim_\mathcal{A}})(t_i) = \mathcal{A}(\beta)(t_i)$.
Altogether, this leads to $ \bigl\< \psi\bigl(\mathcal{B}(\beta_{\sim_\mathcal{A}})(t_1) \bigr), \ldots,$ $\psi\bigl(\mathcal{B}(\beta_{\sim_\mathcal{A}})(t_m)\bigr)\bigr\> = \bigl\<\mathcal{A}(\beta)(t_1), \ldots, \mathcal{A}(\beta)(t_m)\bigr\>$ and, consequently, also to $\bigl\< \mathcal{B}(\beta_{\sim_\mathcal{A}})(t_1), \ldots, \mathcal{B}(\beta_{\sim_\mathcal{A}})(t_m) \bigr\> \in P^\mathcal{B}$ if and only if $\bigl\< \mathcal{A}(\beta)(t_1), \ldots,$ $\mathcal{A}(\beta)(t_m) \bigr\> \in P^\mathcal{A}$.
\end{description}
\end{proof}
Having such canonical models, it is easy to argue that we can eliminate free-sort variables by exhaustive instantiation with all occurring free-sort constant symbols.
\begin{corollary}\label{corollary:EquisatisfiabilityFreeInstantiation}
Let $N$ be a clause set in normal form that contains at least one constant symbol of the free sort.
Suppose there is a clause $C$ in $N$ which contains a free-sort variable $u$. Let the clause set $\widetilde{N}_u$ be constructed as follows:
$\widetilde{N}_u := \bigl(N\setminus\{C\}\bigr) \;\cup\; \bigl\{ C\subst{u}{c} \bigm| c\in\text{\upshape{fconsts}}(N) \bigr\}$.
The original clause set $N$ is satisfiable if and only if $\widetilde{N}_u$ is satisfiable.
\end{corollary}
\begin{proof}~
While the ``only if''-part holds because of $\widetilde{N}_u$ containing only instances of clauses in $N$, the ``if''-part is slightly more complicated. Let $\mathcal{A}$ be a hierarchic model of $\widetilde{N}_u$ so that $\S^\mathcal{A} = \text{\upshape{fconsts}}(\widetilde{N}_u)/_\sim$ for some congruence relation $\sim$ on $\text{\upshape{fconsts}}(N)$, and for every free-sort constant symbol $c$ we have $c^\mathcal{A} = [c]_\sim$. But then $\mathcal{A}\models \bigl\{ C\subst{u}{c} \bigm| c\in\text{\upshape{fconsts}}(N) \bigr\}$ entails $\mathcal{A}\models C$, since the set $\{[c]_\sim \mid c\in\text{\upshape{fconsts}}(N)\}$ covers the whole domain $\S^\mathcal{A}$ due to $\text{\upshape{fconsts}}(\widetilde{N}_u) = \text{\upshape{fconsts}}(N)$. Hence, we obtain $\mathcal{A}\models N$.
\end{proof}
\section{The Complexity of Deciding Satisfiability}\label{section:Complexity}
In the last two sections we have seen how to eliminate base-sort and free-sort variables by means of finite instantiation. In this section, we put these instantiation mechanisms together to obtain a nondeterministic algorithm that decides the hierarchic satisfiability problem for the BSR fragment with simple bounds and investigate its complexity. As a measure of the length of clause sets, clauses, atoms and multisets thereof, we use the number of occurrences of constant symbols and variables in the respective object, and denote it by $\text{len}(\cdot)$.
\begin{theorem}\label{theorem:Complexity}
Let $N$ be a clause set (of length at least $2$) that does not contain any constant symbol from $\Omega_\text{\upshape{LA}}^\alpha$ but at least one free-sort constant symbol. Satisfiability of $N$ w.r.t.\ $\mathfrak{M}$ can be decided in nondeterministic exponential time. To put it more precisely: the problem lies in $\text{NTIME}\bigl(\text{len}(N)^{\mathfrak{c}\cdot\text{len}(N)}\bigr)$ for some constant $\mathfrak{c} > 1$.
\end{theorem}
\begin{proof}
We devise a na\"ive algorithm that decides a given problem instance as follows.
As input we assume a finite clause set $N$ such that $\text{\upshape{$\alpha$consts}}(N) = \emptyset$ and $\text{\upshape{bconsts}}(N)\neq\emptyset$.
\begin{enumerate}[label=(\Roman{*}), ref=\Roman{*}]
\item\label{enum:DecisionAlgorithm:One} Transform the input clause set $N$ into normal form by applying three steps to every clause $C = \Lambda \;\|\; \Gamma\to\Delta$ in $N$ that contains exactly $k > 0$ distinct base-sort variables $x_1, \ldots, x_k$ occurring in $\Lambda$ but not in $\Gamma\to\Delta$:
\begin{enumerate}[label=(\ref{enum:DecisionAlgorithm:One}.\Roman{*}), ref=(\ref{enum:DecisionAlgorithm:One}.\Roman{*})]
\item\label{enum:DecisionAlgorithm:OneOne} Let $\bar{y}$ be the vector of all variables $x_j$ among $x_1, \ldots, x_k$ for which there is an atomic constraint $x_j = c_j$ in $\Lambda$, and let $\bar{c}$ denote the corresponding vector of constant symbols $c_j$. We replace $C$ by $\Lambda\subst{\bar{y}}{\bar{c}}\;\|\;\Gamma\to\Delta$ in $N$.
\item\label{enum:DecisionAlgorithm:OneTwo} If $\Lambda$ contains an atomic constraint of the form $x_j \not= c_j$ after Step \ref{enum:DecisionAlgorithm:OneOne}, then replace $C$ in $N$ by two clauses $\Lambda', x_j < c_j \;\|\; \Gamma\to\Delta$ and $\Lambda', x_j > c_j \;\|\; \Gamma\to\Delta$, where $\Lambda' := \Lambda \setminus \{x_j \not= c_j\}$. Iterate this procedure on the newly added clauses until all atomic constraints concerning one of the variables $x_1, \ldots, x_k$ have the form $x_j \triangleleft c_j$ with $\triangleleft \in \{<, \leq, \geq, > \}$.
\item\label{enum:DecisionAlgorithm:OneThree} Apply Fourier-Motzkin quantifier elimination to every base-sort variable among $x_1, \ldots,$ $x_k$ that is left in $N$.
\end{enumerate}
We call the resulting clause set $N'$. In addition, rename variables so that all clauses in $N'$ are variable disjoint.
\item\label{enum:DecisionAlgorithm:Two} Compute the equivalence classes with respect to the equivalence relation $\leftrightharpoons_{N'}$ induced by $N'$.
\item\label{enum:DecisionAlgorithm:Three} Compute the set $\text{\upshape{bconsts}}(N')$ of all base-sort constant symbols that occur in $N'$.
For every equivalence class $[\<P,i\>]_{\leftrightharpoons_{N'}}$ with $P:\xi_1\times\ldots\times\xi_m \in \Pi$ such that $\xi_i = \mathcal{R}$, collect all instantiation points that are connected to it, i.e.\ compute $\mathcal{I}_{[\<P,i\>], N'}$. Whenever a constant symbol $\alpha_\t$ freshly enters the collection of instantiation points, construct the axiom set $\text{\upshape{Ax}}_{\{\alpha_\t\}, \text{\upshape{bconsts}}(N')}$ according to Definition \ref{definition:BaseInstantiationAxioms} and add all these axioms to a maintained axiom set $\text{\upshape{Ax}}$.
\item\label{enum:DecisionAlgorithm:Four} Compute the set $\text{\upshape{fconsts}}(N')$ of all free-sort constant symbols occurring in $N'$.
\item\label{enum:DecisionAlgorithm:Five} Perform an all-at-once instantiation process on $N'$ that leads to
\begin{align*}
N'' := \Bigl\{ C\subst{\overline{x}, \overline{u}}{\overline{c}, \overline{d}} \Bigm| C \in N' \text{ and }
&\{x_1, \ldots, x_k\} = \text{\upshape{vars}}(C) \cap V_\mathcal{R} \text{ and }\\
&\{u_1, \ldots, u_\ell\} = \text{\upshape{vars}}(C) \cap V_\S \text{ and }\\
&c_i\in\mathcal{I}_{\text{\upshape{ap}}_{N'}(x_i), N'} \text{ and } d_j\in\text{\upshape{fconsts}}(N')\text{ for all $i$ and $j$} \Bigr\} ~.
\end{align*}
\item\label{enum:DecisionAlgorithm:Six}
Nondeterministically construct a hierarchic interpretation $\mathcal{A}$ in three steps:
\begin{enumerate}[label=(\ref{enum:DecisionAlgorithm:Six}.\Roman{*}), ref=(\ref{enum:DecisionAlgorithm:Six}.\Roman{*})]
\item\label{enum:DecisionAlgorithm:SixOne} Nondeterministically choose a total ordering $\preceq_\mathcal{A}$ on the constant symbols in $\text{\upshape{bconsts}}(N'')$ $\cup\, \text{\upshape{$\alpha$consts}}(N'')$ so that for all $c,d \in \text{\upshape{bconsts}}(N'')\cap\mathbb{R}$ it holds $c\preceq_\mathcal{A} d$ if and only if $c \leq d$.
Based on this ordering, construct (in a deterministic fashion) a mapping $\mu_\mathcal{A} : \text{\upshape{bconsts}}(N'')\cup\text{\upshape{$\alpha$consts}}(N'')\to\mathbb{R}$ so that
\begin{itemize}
\item for all $c\in \text{\upshape{bconsts}}(N'')\cap\mathbb{R}$ we get $\mu_\mathcal{A}(c) = c$, and
\item for all $d_1, d_2 \in \text{\upshape{bconsts}}(N'')\cup\text{\upshape{$\alpha$consts}}(N'')$ it holds $\mu_\mathcal{A}(d_1) \leq \mu_\mathcal{A}(d_2)$ if and only if $d_1 \preceq_\mathcal{A} d_2$.
\end{itemize}
Set $c^\mathcal{A} := \mu_\mathcal{A}(c)$ for every $c\in\text{\upshape{bconsts}}(N'')$.
\item\label{enum:DecisionAlgorithm:SixTwo} Nondeterministically choose a mapping $\nu_\mathcal{A} : \text{\upshape{fconsts}}(N'') \to \text{\upshape{fconsts}}(N'')$.
While $\nu_\mathcal{A}$'s image shall induce the free domain $\S^\mathcal{A} := \{ \nu_\mathcal{A}(c) \mid c \in \text{\upshape{fconsts}}(N'')\}$, the value $\nu_\mathcal{A}(c)$ shall be assigned to $c$ under $\mathcal{A}$ for every $c\in\text{\upshape{fconsts}}(N'')$.
\item\label{enum:DecisionAlgorithm:SixThree} Let $\text{At}(N'')$ denote the set of nonequational ground atoms induced by the essentially ground clause set $N''$, formally
\begin{align*}
At(N'') :=\; \bigl\{ A\subst{\bar{x}}{\bar{c}} \bigm| &\text{there is a nonequational free atom $A$ in the free part $\Gamma\to\Delta$}\\
&\text{of a clause $(\Lambda, x_1= c_1, \ldots, x_k= c_k\;\|\;\Gamma\to\Delta)\in N''$}\\
&\text{with $\{x_1, \ldots, x_k\} = \text{\upshape{vars}}(A)$} \bigr\} ~.
\end{align*}
Given $\text{At}(N'')$, we construct the set $\widehat{\At}(N'')$ by syntactically replacing every base-sort constant symbol $c$ in $\text{At}(N'')$ by $\mu_\mathcal{A}(c)$ and every free-sort constant symbol $e$ by $\nu_\mathcal{A}(e)$.
Nondeterministically choose a subset $\widehat{\At}_\mathcal{A}(N'')$ of $\widehat{\At}(N'')$ that represents the atoms in $\widehat{\At}(N'')$ which shall be true under $\mathcal{A}$, i.e.\ construct a Herbrand model over the atoms in $\widehat{\At}(N'')$.
\end{enumerate}
\item\label{enum:DecisionAlgorithm:Seven} Check whether $\mathcal{A}$ is a hierarchic model of $N''\cup \text{\upshape{Ax}}$.
\\[-2ex]\strut\hfill$\Diamond$
\end{enumerate}
Regarding the correctness of the algorithm, there are two crucial points which we need to address, namely (a) the equisatisfiability of $N$ and $N'$, and (b) the equisatisfiability of $N'$ and $N''\cup\text{\upshape{Ax}}$.\medskip
\noindent{Ad (a)}:
It is straightforward to check that \ref{enum:DecisionAlgorithm:OneOne} and \ref{enum:DecisionAlgorithm:OneTwo} lead to equisatisfiable clause sets with respect to the standard semantics of linear arithmetic.
The Fourier-Motzkin elimination step in \ref{enum:DecisionAlgorithm:OneThree} works for existentially quantified real-valued variables (cf.\ \cite{Dantzig1973}, \cite{Kroening2008}). Given a clause $C = \Lambda \;\|\; \Gamma\to\Delta$ that contains a base-sort variable $x$ occurring in $\Lambda$ but not in $\Gamma\to\Delta$, let $\bar{z}$ denote the vector of all variables in $\Lambda$ except for $x$. Recall that $C$ can be informally understood as $\forall \bar{z}\forall x \bigl(\bigl(\bigwedge \Lambda \wedge \bigwedge \Gamma\bigr) \to \bigvee \Delta\bigr)$. Since $x$ does neither occur in $\Delta$ nor in $\Gamma$ and since any formula $\phi\to\psi$ is equivalent to $\neg\phi \vee \psi$, we can equivalently write $\forall \bar{z} \bigl(\bigl((\exists x\bigwedge \Lambda) \wedge \bigwedge \Gamma\bigr) \to \bigvee \Delta\bigr)$. Now Fourier-Motzkin variable elimination can be applied to the part $\exists x\bigwedge \Lambda$ to transform it into a formula $\bigwedge\Lambda'$ over linear arithmetic constraints (each of the form $c\triangleleft d$ or $y\triangleleft e$) so that $\bigwedge\Lambda'$ is equivalent to $\exists x\bigwedge\Lambda$ with respect to the standard semantics of linear arithmetic, $\Lambda'$ does not contain $x$ anymore and all variables and constant symbols in $\Lambda'$ have already occurred in $\Lambda$ (cf.\ \cite{Dantzig1973}).
Treating all base-sort variables that occur in $\Lambda$ but not in $\Gamma\to\Delta$ as described above, we finally obtain a clause $\Lambda'' \;\|\; \Gamma\to\Delta$ that is equivalent to $\Lambda \;\|\; \Gamma\to\Delta$ and is in normal form.\medskip
\noindent{Ad (b)}:
The construction of clause set $N''$ done in Step (\ref{enum:DecisionAlgorithm:Five}) resembles $|\text{\upshape{vars}}(N')\cap V_\mathcal{R}|$ consecutive applications of Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} to the base-sort variables $x_1, \ldots, x_k$ followed by $|\text{\upshape{vars}}(N')\cap V_\S|$ consecutive applications of Corollary \ref{corollary:EquisatisfiabilityFreeInstantiation} to the free-sort variables $u_1,\ldots,u_\ell$ in order to instantiate all variables that occur in $N'$.
For an arbitrary clause $C$ in $N'$ we have\\
$\text{\upshape{fconsts}}\bigl(C\subst{\bar{x}}{\bar{c}}\subst{u_1}{d_1}\ldots\subst{u_j}{d_j}\bigr) \;\;\subseteq\;\; \text{\upshape{fconsts}}(N')$
for every $j$, $0\leq j \leq \ell$. Hence, Corollary \ref{corollary:EquisatisfiabilityFreeInstantiation} entails equisatisfiability of $N'' \cup \text{\upshape{Ax}}$ and the intermediate set $N''' \cup \text{\upshape{Ax}}$ where
$N''' := \bigl\{C\subst{\bar{x}}{\bar{c}} \bigm|
C \in N' \text{ and }
\{x_1, \ldots, x_k\} = \text{\upshape{vars}}(C) \cap V_\mathcal{R} \text{ and }
c_i \in\mathcal{I}_{\text{\upshape{ap}}_{N'}(x_i), N'}$ $\text{ for all $i$} \bigr\}$.
It remains to show equisatisfiability of $N'$ and $N'''\cup \text{\upshape{Ax}}$.
%
In order to do so, we first reduce the problem to showing equisatisfiability of $N'\cup \text{\upshape{Ax}}$ and $N'''\cup \text{\upshape{Ax}}$: If $N'$ is satisfiable, then so is $N'\cup\text{\upshape{Ax}}$ by Proposition \ref{proposition:ImproperAxiomSatisfiability}, since $N'$ does not contain constant symbols from $\Omega_\text{\upshape{LA}}^\alpha$. Conversely, $N'$ is obviously satisfiable whenever $N'\cup\text{\upshape{Ax}}$ is.
Let $x_1, \ldots, x_k$ be a list of pairwise distinct base-sort variables such that $\{x_1, \ldots, x_k\} := \text{\upshape{vars}}(N')\cap V_\mathcal{R} = \text{\upshape{vars}}(N'\cup\text{\upshape{Ax}})\cap V_\mathcal{R}$. We set $N'_0 := N'$ and for every $j$, $1\leq j\leq k$, we define \[N'_j := \bigl\{C\subst{x_j}{c_j} \;\bigm|\; C\in N'_{j-1} \text{ and } c_j\in\mathcal{I}_{\text{\upshape{ap}}_{N'}(x_j), N'}\bigr\} ~.\]
As the $x_1, \ldots, x_k$ are pairwise distinct and since the involved substitution operations only substitute variables with constant symbols, we can equivalently write
\[N'_j = \bigl\{C\subst{x_1}{c_1}\ldots\subst{x_j}{c_j} \bigm| C \in N' \text{ and } c_i \in\mathcal{I}_{\text{\upshape{ap}}_{N'}(x_{i}), N'} \text{ and $1\leq i\leq j$} \bigr\}\]
for every $j$, $1\leq j\leq k$. Variable disjointness of the clauses in $N'$ entails $N'_k = N'''$, since in this case the iterative substitution in the construction of $N'_k$ yields the same result as simultaneous substitution in the construction of $N'''$ does.
Consider the sequence $N'_0, N'_1, \ldots, N'_k$ of clause sets, for which we know $N'_0 = N'$ and $N'_k = N'''$. If we rename the variables in each set in the sequence such that the clauses are variable disjoint, we can apply Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} $k$ times to conclude equisatisfiability of $N'_0\cup\text{\upshape{Ax}}$, $N'_1\cup\text{\upshape{Ax}}$ and so on up to $N'_k\cup\text{\upshape{Ax}}$. To support this claim, we need to show that the prerequisites of Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} are fulfilled along the sequence. They certainly are for the starting point $N'_0\cup\text{\upshape{Ax}}$, since we assume $N'$ to be in normal form and to not contain constant symbols $\alpha_\t$. For the rest of the sequence, we invoke Lemma \ref{lemma:IterativeInstantiation}, since it ensures ${\leftrightharpoons_{N'_0\cup\text{\upshape{Ax}}}} = \ldots = {\leftrightharpoons_{N'_k\cup\text{\upshape{Ax}}}}$ and $\mathcal{I}_{[\<P,i\>],N'_0} = \ldots = \mathcal{I}_{[\<P,i\>],N'_k}$ for all equivalence classes $[\<P,i\>]$ induced by $\leftrightharpoons_{N'_0\cup\text{\upshape{Ax}}}$.
The following observations justify why it is legitimate for the algorithm described in steps (\ref{enum:DecisionAlgorithm:One}) to (\ref{enum:DecisionAlgorithm:Seven}) to mostly ignore the axioms in $\text{\upshape{Ax}}$.
\begin{enumerate}[label=(b.\arabic{*}), ref=(\ref{enum:proofComplexity:Two}.\arabic{*})]
\item ${\leftrightharpoons_{N'}} = {\leftrightharpoons_{N'\cup\text{\upshape{Ax}}}}$, because the relation $\leftrightharpoons_{N'\cup\text{\upshape{Ax}}}$ exclusively depends on the free parts of clauses in $N'\cup\text{\upshape{Ax}}$, but the clauses in $\text{\upshape{Ax}}$ do not have free parts.
\item $\text{\upshape{bconsts}}(N') \!=\! \text{\upshape{bconsts}}(N'\cup\text{\upshape{Ax}})$, since the definition of $\text{\upshape{Ax}}$ entails $\text{\upshape{bconsts}}(\text{\upshape{Ax}})\subseteq \text{\upshape{bconsts}}(N')$.
\item $\text{\upshape{fconsts}}(N') = \text{\upshape{fconsts}}(N'\cup\text{\upshape{Ax}})$ holds because clauses in $\text{\upshape{Ax}}$ do not contain constant symbols of the free sort.
\item $\mathcal{I}_{[\<P,i\>],N'} = \mathcal{I}_{[\<P,i\>],N'\cup\text{\upshape{Ax}}}$, since only atomic constraints that involve one base-sort variable contribute to the set of instantiation points, but $\text{\upshape{Ax}}$ is a set of ground clauses.
\end{enumerate}
Altogether, the iterative application of Lemma \ref{lemma:EquisatisfiabilityBaseInstantiation} shows that $N'\cup \text{\upshape{Ax}}$ has a hierarchic model if and only if $N'''\cup\text{\upshape{Ax}}$ has one. But as we have already argued above, this entails the equisatisfiability of $N'$ and $N'''\cup\text{\upshape{Ax}}$ and even $N''\cup\text{\upshape{Ax}}$.
This finishes our considerations regarding the correctness of the presented algorithm. Next, we investigate its running time. In order to do so, we take a look at every step individually.
\noindent(\ref{enum:DecisionAlgorithm:One}):
While the substitution operations in Step \ref{enum:DecisionAlgorithm:OneOne} take a total amount of time that is polynomial in $\text{len}(N)$, the length of the clause set does not grow.
The second step, however, may blow up the length of the clause set exponentially. Every clause $\Lambda\;\|\;\Gamma\to\Delta$ might be copied up to $2^{|\Lambda|}$ times, since $\Lambda$ can contain at most $|\Lambda|$ constraints of the form $x_j\not= c_j$. Hence, Step \ref{enum:DecisionAlgorithm:OneTwo} increases the length of the clause set to not more than $\text{len}(N)\cdot 2^{\text{len}(N)}$ and in the worst case takes time polynomial in that new length.
Given a multiset $\Lambda$ of atomic constraints, in which variables $x_1, \ldots, x_k \in \text{\upshape{vars}}(\Lambda)$ are supposed to be eliminated one after another by the Fourier-Motzkin procedure, we can partition $\Lambda$ into $k+1$ parts $\Lambda_0, \Lambda_1, \ldots, \Lambda_k$ so that for all $i > 0$ the part $\Lambda_i$ contains all atomic constraints that involve $x_i$ and only those, and $\Lambda_0 := \Lambda \setminus (\Lambda_1 \cup \ldots \cup \Lambda_k)$. It turns out that eliminating a variable $x_i$ from $\Lambda$ results in a multiset $(\Lambda \setminus \Lambda_i) \cup \widehat{\Lambda}_i$ where $|\widehat{\Lambda}_i|$ is at most $|\Lambda_i|^2$. Hence, after eliminating all $x_i$, we end up with a multiset of atomic constraints of size $|\Lambda_0| + |\Lambda_1|^2 + \ldots + |\Lambda_k|^2 \leq |\Lambda|^2$. Consequently, Step \ref{enum:DecisionAlgorithm:OneTwo} may increase the length of the clause set at most quadratically. The time taken for this step is polynomial in that new length.
Overall, we end up with a length of at most $\text{len}(N)^2 \cdot 2^{2\cdot\text{len}(N)} \leq \text{len}(N)^{3\cdot\text{len}(N)}$ for $N'$ (recall that we assumed $\text{len}(N) \geq 2$).\medskip
\noindent(\ref{enum:DecisionAlgorithm:Two}): This step can be performed in time that is polynomial in the length of $N'$ using an efficient union-find data structure.\medskip
\noindent(\ref{enum:DecisionAlgorithm:Three}): The computation of $\text{\upshape{bconsts}}(N')$ and the collection of all relevant instantiation points takes time polynomial in the length of $N'$.
The set of instantiation points for any base-sort variable is a subset of $\text{\upshape{bconsts}}(N') \cup \{\alpha_{d+\varepsilon} \;\mid\; d\in\text{\upshape{bconsts}}(N')\} \cup \{\alpha_{-\infty}\}$. It is worthwhile to note that Steps \ref{enum:DecisionAlgorithm:OneOne} to \ref{enum:DecisionAlgorithm:OneThree} do not lead to a change in the number of instantiation points, since they only modify atomic constraints that do not contribute to instantiation points. The reason is that the variables $x_1, \ldots, x_k$ addressed in Step (\ref{enum:DecisionAlgorithm:One}) do not occur in the free parts of the modified clause.
For every argument position pair $\<P,i\>$, there are at most $\text{len}(N)$ instantiation points, as every atomic constraint $y\triangleleft c$ in $N$ can induce at most one instantiation point (either $c$ of $\alpha_{c+\varepsilon}$). To account for $\alpha_{-\infty}$: if there is a base-sort variable $y$ to be instantiated in $N'$ at all, then $N'$ must also contain a free atom $Q(\ldots,y,\ldots)$ which did already occur in $N$, and which thus also contributes to the length of $N$. In addition, we have $\text{\upshape{bconsts}}(N') \subseteq \text{\upshape{bconsts}}(N)$, leading to $|\text{\upshape{bconsts}}(N')| \leq \text{len}(N)$.
The construction of the required axiom set $\text{\upshape{Ax}}$ can be done in polynomial time in $\text{len}(N') + \text{len}(\text{\upshape{Ax}})$, where we can bound the length of the axiom set from above by $2\cdot2\cdot\text{len}(N) + 2\cdot 4\cdot\text{len}(N)^2 \leq 10 \cdot \text{len}(N)^2$.\medskip
\noindent(\ref{enum:DecisionAlgorithm:Four}): The extraction of $\text{\upshape{fconsts}}(N')$ does not take longer than polynomial time in the length of $N'$.\medskip
\noindent(\ref{enum:DecisionAlgorithm:Five}):
At first, we consider each clause $C = \Lambda\;\|\;\Gamma\to\Delta$ in $N'$ separately. Since $|\text{\upshape{fconsts}}(N')|$ is upper bounded by $\text{len}(N)$ (every free-sort constant symbol in $N'$ did already occur in $N$), instantiation of the free-sort variables yields a factor of at most $\text{len}(N)^{|\text{\upshape{vars}}(\Gamma\to\Delta) \cap V_\S|}$. We have already argued -- when looking at Step (\ref{enum:DecisionAlgorithm:Three}) -- that the number of instantiations points for each base-sort variable is bounded from above by $\text{len}(N)$. Hence, instantiation of all base-sort variables in the clause adds a factor of at most $\text{len}(N)^{|\text{\upshape{vars}}(\Gamma\to\Delta) \cap V_\mathcal{R}|}$.
There are at most $\text{len}(\Gamma\to\Delta)$ different variables in $C$ that need to be instantiated, and as $\Gamma\to\Delta$ did already occur in $N$ (modulo variable renaming), we have $\text{len}(\Gamma\to\Delta) \leq \text{len}(N)$.
When instantiating a clause, the constraint part may increase in length, namely by at most double the number of instantiated variables. In the worst case, we thus get triple the length of the original, e.g.\ in case of instantiating the clause $\|\to P(x)$ with the default instantiation point $\alpha_{-\infty}$ we obtain $x= \alpha_{-\infty}\;\|\;\to P(x)$.
In total, instantiating a single clause $C = \Lambda\;\|\;\Gamma\to\Delta$ taken from $N'$ leads to a clause set of length at most $3\cdot\text{len}(C) \cdot \text{len}(N)^{\text{len}(\Gamma\to\Delta)}$. Consequently, we can upper bound the length of the fully instantiated clause set $N''$ by $3\cdot\text{len}(N') \cdot \text{len}(N)^{\text{len}(N)} \;\leq\; 3\cdot \text{len}(N)^{4\cdot\text{len}(N)}$.
Instantiating the set of clauses needs only time that is bounded by some polynomial in $\text{len}(N'')$.\medskip
\noindent(\ref{enum:DecisionAlgorithm:Six}): The construction of $\mathcal{A}$ can be done nondeterministically in time that is bounded from above by some polynomial in the length of $N''$.\medskip
\noindent(\ref{enum:DecisionAlgorithm:Seven}):
The check whether $\mathcal{A}$ satisfies $N''\cup\text{\upshape{Ax}}$ can be performed in a deterministic fashion in time polynomial in the length of $N''\cup\text{\upshape{Ax}}$.\bigskip
Taking all the above results into account, we can upper bound the running time of the algorithm by some polynomial in $\text{len}(N)^{4\cdot\text{len}(N)}$. Hence, there is some constant $\mathfrak{c} \geq 4$ such that the nondeterministic running time lies in $\O\bigl(\text{len}(N)^{\mathfrak{c}\cdot \text{len}(N)}\bigr)$. Consequently, the problem of deciding whether a finite clause set $N$ is satisfiable, lies in NEXPTIME.
\end{proof}
NEXPTIME-completeness of satisfiablilty for the Bernays-Sch\"onfinkel-Ramsey fragment of first-order logic without equality (cf.\ \cite{Lewis1980}) immediately yields NEXPTIME-hardness of satisfiability for finite clause sets over the BSR fragment with simple bounds. Together with Theorem \ref{theorem:Complexity} we thus obtain NEXPTIME-completeness of the problem.
\begin{corollary}
The problem of deciding satisfiability w.r.t.\ $\mathfrak{M}$ of finite clause sets over the BSR fragment with simple bounds is NEXPTIME-complete.
\end{corollary}
\section{Beyond Simple Bounds}\label{section:Basification}
In this section we long to answer the question how simple our constraints have to be. The most complex atomic constraints we have allowed by now are of the form $x \triangleleft c$. Being able to cope with this kind, we can leverage the idea of \emph{flattening} to deal with more complicated constraints such as $3x + c < 1$.
The basic idea rests on two steps:
\begin{enumerate}[label=(\Roman{*}), ref=(\Roman{*})]
\item Transform this constraint into the equivalent $x < 1-\tfrac{1}{3} c$.
\item Introduce a fresh Skolem constant $b$, transform the constraint into $x < b$ and add a
\emph{defining clause} $C_b$ expressing $b = 1-\tfrac{1}{3} c$ to the clause set.
\end{enumerate}
These two steps already indicate that this technique is restricted to atomic constraints that are either ground or univariate and linear, and in which the standard operations addition, subtraction, multiplication and division on the reals may be involved. But we may even allow free function symbols $g:\xi_1\times \ldots \times\xi_m \to \xi$ with $\xi_1, \ldots, \xi_m, \xi \in \{\mathcal{R}, \S\}$, as long as all subterms occurring below $g$ are ground.
The key insight at this point is that complex ground terms $s$ can be replaced by fresh constant symbols $b_s$ of the corresponding sort, if we add a \emph{defining clause} $C_{b_s}$ that identifies $b_s$ with $s$ (a technique called \emph{basification} in \cite{Kruglov2012}) -- see below for details. Of course, the replacement of $s$ must be done consistently throughout the clause set. Consequently, we can extend the syntax of clauses so that more complex terms are admitted.
\begin{definition}[BSR Clause Fragment with Ground LA Bounds]\label{definition:extBSRwithConstrSyntax}
Let $\Omega'$ collect nonconstant free function symbols equipped with sorting information.
An \emph{atomic constraint} is of the form $s\triangleleft s'$ with $\triangleleft\in\{<, \leq, =, \not=, \geq, >\}$ and well-sorted base-sort terms $s, s'$ over symbols in $\Omega_\text{\upshape{LA}}\cup$ $\{+, -, \cdot, /\} \cup \Omega \cup\Omega'$ in which every subterm $g(s_1, \ldots, s_m)$ with $g \in \Omega'$ is ground and where $s \triangleleft s'$ contains at most one occurrence of a base-sort variable and no free-sort variables.
A \emph{free atom} $A$ is either of the form $s\approx s'$ with $s,s'$ being either free-sort variables or well-sorted ground terms of the free sort over function symbols in $\Omega_\text{\upshape{LA}} \cup \{+,-,\cdot, /\} \cup \Omega \cup \Omega'$, or $A$ is of the form $P(s_1, \ldots, s_m)$, where $P:\xi_1\times\ldots\times\xi_m \in\Pi$ and for each $i\leq m$ the term $s_i$ is of sort $\xi_i$. If $\xi_i = \mathcal{R}$, then $s_i$ is a base-sort variable, and if $\xi_i = \S$, then $s_i$ is either a free-sort variable or a ground free-sort term over function symbols in $\Omega_\text{\upshape{LA}} \cup \{+,-,\cdot, /\} \cup \Omega \cup \Omega'$.
\end{definition}
As we have already sketched, we can easily transform a clause set $N$ that contains clauses over the extended syntax into an equisatisfiable set $N_1\cup N_2$ so that $N_1$ contains clauses according to Definition \ref{definition:BSRwithConstrSyntax} only and all clauses in $N_2$ are of the form $b_s \not= s \;\|\to\Box$ or $\|\to b_s \approx s$. Clearly, we can proceed with $N_1$ according to Steps (\ref{enum:DecisionAlgorithm:One}) to (\ref{enum:DecisionAlgorithm:Five}) of the algorithm given in the proof of Theorem \ref{theorem:Complexity} and thus obtain the essentially ground clause set $N_1''$ that is equisatisfiable to $N_1$ and a set $\text{\upshape{Ax}}$ of axioms. The construction of a hierarchic interpretation $\mathcal{A}$ for $N_1''$ in Step (\ref{enum:DecisionAlgorithm:Six}) can be modified so that it results in an interpretation that also covers all function symbols that occur in $N_2$ but not in $N_1$. Checking whether $\mathcal{A}$ is a hierarchic model of $N_1'' \cup \text{\upshape{Ax}} \cup N_2$ can be done easily. Consequently, the hierarchic satisfiability problem for clause sets over the extension of the BSR fragment with ground LA bounds is decidable, too.
\section{Undecidable Fragments}\label{section:UndecidableFragments}
So far we have described decidable fragments of first-order logic modulo linear arithmetic. In \cite{Fietzke2012} it has been shown that already the Bernays-Sch\"onfinkel-Horn (BSH) fragment with addition and subtraction on the reals is undecidable. In the current section we describe more fine-grained undecidable fragments.
As it turns out, two-counter machines -- of which the halting problem has been proven to be undecidable in \cite{Minsky1967} -- can be encoded exclusively using very restricted syntax on the constraint part, such as difference constraints $x-y \triangleleft c$, additive constraints $x+y \triangleleft c$, quotient constraints $x \triangleleft c\cdot y$ (which could equivalently be written $\tfrac{x}{y} \triangleleft c$, hence the name) or multiplicative constraints $x \cdot y \triangleleft c$. Even more restrictive, in the case of difference constraints and quotient constraints only a single base-sort constant symbol is necessary. In case of quotient and multiplicative constraints, lower and upper bounds on the used variables do not lead to a decidable fragment -- which would be the case if we were using variables over the integers.
\paragraph*{Difference constraints}
We use the predicate symbol $M: \S\times\mathcal{R}\times\mathcal{R}\times\mathcal{R}$ to address the state of the machine as follows: $M(u, x, y, z)$ stands for a machine at instruction $u$ with counter values $i_1 = x - z - 1$ and $i_2 = y - z - 1$, where the last argument $z$ keeps track of an offset relative to which $x$ and $y$ store the values of the counters. Following this principle, the increment instruction for the first counter $i_1$ is encoded by the clause $x'-x= 1 \;\|\; M(b,x,y,z) \to M(b',x',y,z)$, which leaves the offset untouched. The offset is an appropriate tool that allows us to have a uniform syntactic structure for all atomic constraints. It is due to the offset encoding that we can easily use a difference constraint when checking whether a counter is zero or not. The conditional decrement instruction is split up in two clauses: the zero case $x-z= 1 \;\|\; M(b,x,y,z) \to M(b',x,y,z)$ and the non-zero case $x-z>1,\, y'-y= 1,\, z'-z= 1 \;\|\; M(b,x,y,z) \to M(b',x',y,z')$.
Hence, by undecidability of the halting problem for two-counter machines, we may conclude that satisfiability for the BSH fragment with difference constraints (requiring only the constant $1$ besides the input) and a single free $4$-ary predicate symbol is undecidable, too.
\paragraph*{Quotient constraints}
Encoding two-counter machines in the BSH fragment with quotient constraints works very similar. We only need to change the representation of counter values in a state $M(u, x, y, z)$ as follows: $i_1 = -\log_2(\tfrac{2x}{z}) = -\log_2(x) + \log_2(z) - 1$ and $i_2 = -\log_2(\tfrac{2y}{z})$. Incrementing the first counter is encoded by $2\cdot x'= x \;\|\; M(b,x,y,z) \to M(b',x',y,z)$, and the conditional decrement instruction is represented by $2\cdot x= z \;\|\; M(b,x,y,z) \to M(b',x,y,z)$ and $2\cdot x<z,\, 2\cdot y'= y,\, 2\cdot z'= z \;\|\; M(b,x,y,z) \to M(b',x,y',z')$. Analogous to the case of difference constraints, we may now show undecidability of the satisfiability problem for the BSH fragment with quotient constraints (requiring only the constant $2$) and a single free $4$-ary predicate symbol is undecidable.
We have chosen negative exponents for the encoding of the counter values, since this guarantees that the range of the base-sort variables is bounded from below and above. Thus, we could restrict all base-sort variables to values within $(0, 1]$ (by adding appropriate atomic constraints to every clause), and still end up with an undecidable satisfiability problem.
\paragraph*{Additive constraints}
Having additive constraints of the form $x + y \triangleleft c$ at hand, we can simulate subtraction by defining the additive inverse using a constraint $x + x_- = 0$. To keep track of inverses, we adjust the arity of $M$ accordingly. Counter values are represented in the same way as we have done for difference constraints. The increment instruction for the first counter is thus encoded by $x' + x_- = 1,\, x' + x'_- = 0 \;\|\; M(b, x, x_-, y, y_-, z, z_-) \to M(b', x', x'_-, y, y_-, z, z_-)$. It is now straightforward to come up with the encoding of the conditional decrement.
Hence, satisfiability of the BSH fragment with additive constraints and a single free predicate symbol of arity $7$ is undecidable. However, this time we need two constants, namely $1$ and $0$.
\paragraph*{Multiplicative constraints}
In order to complete the picture, we shortly leave the realm of linear arithmetic and consider multiplicative constraints of the form $x \cdot y \triangleleft c$. These relate to quotient constraints as additive constraints relate to difference constraints. Hence, combining the previously used ideas of offsets and inverses, we can encode two-counter machines also with multiplicative constraints:\\ $x \cdot x'_{-1} = 2,\, x' \cdot x'_{-1} = 1 \;\|\; M(b, x, x_{-1}, y, y_{-1}, z, z_{-1}) \to M(b', x', x'_{-1}, y, y_{-1}, z, z_{-1})$ encodes the increment instruction on the first counter, for instance. As in the case of quotient constraints, we could restrict the range of base-sort variables to $(0,1]$ by suitable constraints. Consequently, this leads to another fragment of which the satisfiability problem is undecidable.
\section{Conclusion and Future Work} \label{sec:conclusion}
Our main contribution is a proof showing that satisfiability of the BSR fragment equipped with simple bounds on real arguments is decidable. The key argument describes a satisfiability preserving replacement of universally quantified clauses by finitely many instances which are in essence variable free.
The analysis of a na\"ive decision procedure proves that satisfiability of this fragment is NEXPTIME-complete. The complexity result is of particular interest, since satisfiability of the Bernays-Sch\"onfinkel fragment has already been known to be NEXPTIME-complete for more than three decades \cite{Lewis1980}.
Our approach to proving decidability has been tailored to be well-suited for an integration into reasoning procedures that perform efficiently on practically relevant problem instances. This in particular applies to the way we handle the real-sorted part of clauses. For the free parts we could have adopted more fine-grained techniques inspired by the ones we used for the real-sorted part. However, we did not do so for the sake of simplicity and brevity.
It turned out that we leave the realm of decidability as soon as we add elementary operations $\circ$ on the reals to our constraint language, even if we restrict the free part to Horn clauses over the Bernays-Sch\"onfinkel fragment. Constraints of the form $x\circ y \triangleleft c$ are sufficient to obtain an undecidable fragment, where $\circ$ can stand for addition, subtraction, multiplication and division, $\triangleleft$ represents standard relations $<, \leq, =, \not=, \geq, >$, and $c$ is a real-valued constant.
This observation nicely complements our main result, since it quite clearly highlights the limits of decidability in this context.
Moreover, it reveals some interplay between real-sorted constraints and the free first-order part. For instance, difference logic (boolean combinations of propositional variables and existentially quantified constraints $x - y \triangleleft c$ with $c$ being a real-valued constant and $\triangleleft \in \{<, \leq\}$) is known to be decidable \cite{Mahfoudh2002}. However, we have seen in Section \ref{section:UndecidableFragments} that its combination with the Bernays-Sch\"onfinkel-Horn fragment is sufficient to formalize two-counter machines.
Although the obvious options for further extending the constraint language lead to undecidability there might still be room for improvement. We leave it as future work to investigate the BSR fragment with simple bounds plus constraints of the form $x \triangleleft y$ with $x, y$ being real-valued variables and $\triangleleft \in \{<, \leq, =, \not=, \geq, >\}$. On the other hand, it is conceivable to combine other decidable free first-order fragments with simple bounds, preferably ones satisfying the finite model property such as the monadic fragment.
As we have already pointed out, a natural next step for us will be to devise useful decision procedures for the BSR fragment with simple bounds that perform well in practice.
\renewcommand*{\backref}[1]{}
\renewcommand*{\backrefalt}[4]{%
\ifcase #1 %
(Not cited.)%
\or
(Cited on page~#2.)%
\else
(Cited on pages~#2.)%
\fi}
\renewcommand*{\backrefsep}{, }
\renewcommand*{\backreftwosep}{ and~}
\renewcommand*{\backreflastsep}{ and~}
|
\section{Introduction}
In recent years, the magnetic and transport properties of Ce-based ternary compounds of type CeT$_2$Al$_{10}$ (T = Fe, Ru and Os), which crystalline in the orthorhombic structure (space group No 63 Cmcm)~\cite{Thiede}, have generated strong interest in both theoretical and experimental condensed matter physics~\cite{1,2,3,ddk,4,dta,5,6,7,8}. This interest arose due to the various ground states observed in this family of Ce-compounds. An unusually sharp phase transition near 27 K in the magnetic susceptibility of CeRu$_2$Al$_{10}$ has been attributed to a spin-dimer formation ~\cite{9,10}. The resistivity of CeRu$_2$Al$_{10}$ exhibits a sharp drop near 27 K resembling an insulator-metal transition~\cite{1}. A very similar phase transition, near 29 K, has been observed in CeOs$_2$Al$_{10}$ ~\cite{5,11}, but in this compound the susceptibility (along the $a-$axis) exhibits a broad maximum near 45 K in contrast to a sharp drop at the phase transition (27 K) in CeRu$_2$Al$_{10}$ ~\cite{5,13}. The broad maximum in the susceptibility and its strong anisotropic behavior in the paramagnetic state reveal the presence of strong hybridization between 4$f$ and conduction electrons as well as strong single ion anisotropy arising from the crystal field potential~\cite{kyu,fst}.
\par
Further difference between the two systems appears in the resistivity of the ordered state; the resistivity of CeOs$_2$Al$_{10}$ displays a thermal activation-type temperature dependence below 15 K while the resistivity of CeRu$_2$Al$_{10}$ exhibits a metallic behavior below the phase transition down to 2 K. In spite of the comparable transition temperatures a fundamental contrast in the electronic disposition of the two compounds has been exposed in recent high-pressure studies ~\cite{5}. Under a hydrostatic pressure of 1.75 GPa the electrical resistivity of CeRu$_2$Al$_{10}$ changes in a way as closely to match the overall behavior of the temperature dependent resistivity of CeOs$_2$Al$_{10}$ in zero applied pressure. Applying pressure is a well recognized tuning method of the 4$f-$ band with respect to the Fermi energy (E$_F$) in narrow-band systems. The results obtained for the two iso-electronic compounds are therefore an indication that the center of gravity of the 4$f-$band is lying on opposite sides of the E$_F$: the 4$f-$ spectral weight in CeRu$_2$Al$_{10}$ is most likely close to but below the E$_F$ whereas in CeOs$_2$Al$_{10}$, the 4$f$ band is most likely above the E$_F$ due to the more extended nature of the 5$d$ band of Os compared to the 4$d$ band of Ru. The 3$d$ transition metal compound CeFe$_2$Al$_{10}$ exhibits Kondo insulating behavior with a transport$-$derived gap of 15 K~\cite{2}, while an NMR study reveals a much larger value of the gap, namely 110 K~\cite{14}.
\par
Neutron diffraction studies of CeT$_2$Al$_{10}$ (T=Ru and Os) reveal a very small ordered moments, 0.34 $\mu_B$ and 0.29 $\mu_B$, respectively, along the $c-$axis, which is not the direction expected from the single ion crystal field (CEF) anisotropy ~\cite{ddk,4,hk}. As the single ion crystal field would prefer the moment along the $a-$axis in both compounds, this indicates that the moment direction in these compounds is governed by the anisotropic magnetic exchange and not by the CEF anisotropy. The inelastic neutron scattering (INS) study at 4.5 K on the polycrystalline samples of CeT$_2$Al$_{10}$ (T=Ru, Os and Fe) reveals a clear sign of a spin-gap formation of 8 meV, 11 meV and 12 meV, respectively~\cite{15, 16}. These gaps are nearly temperature independent up to 24 K, but disappear suddenly at 27 K, 39 K and 75 K respectively~\cite{jmmp}. Above these temperatures, the INS response becomes very broad, of quasi-elastic-type. Very recently inelastic neutron scattering investigation on single crystals of CeRu$_2$Al$_{10}$, CeOs$_2$Al$_{10}$ and CeFe$_2$Al$_{10}$ have been performed~\cite{15,16,18}. Well defined gapped spin waves are observed in CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$ that can well be explained by anisotropic exchange interactions. Even in the paramagnetic state of CeFe$_2$Al$_{10}$ (no magnetic ordering observed down to 50 mK) the neutron study reveals a dispersive gapped magnetic excitations having the same propagation vector $k$ = (1, 0, 0) as observed in CeRu$_2$Al$_{10}$, suggesting that these magnetic excitations in the Kondo insulating state have some connection to the spin wave observed in the magnetically ordered state of CeT$_2$Al$_{10}$ (T = Ru and Os) ~\cite{15,16}.
\par
The effects of electron (Ir/Rh) and hole (Re) doping on the transition metal site in CeT$_2$Al$_{10}$ (T = Ru and Os) have been investigated, through magnetization, resistivity, muon spin rotation ($\mu$SR), and neutron scattering (both elastic and inelastic) ~\cite{7, 8, 19}. These studies show the general trend that the hybridization between 4$f-$electrons and conduction electrons increases with hole-doping, while the Ce-4$f$ electrons become more localized with electron-doping. On hole-doping the spin gap and the antiferromagnetic order with an anomalous direction of the magnetic moment (i.e. moment either along $c-$axis or $b-$axis) not governed by the single ion crystal field anisotropy (this prefers moment along $a-$axis) survive with small ordered state moments of 0.18$-$0.23$\mu_B$ ~\cite{4,19, 20}. In contrast to this, electron doping destabilizes the spin gap formation and the antiferromagnetic ordering becomes normal with moment directions along the $a-$axis e.g. governed by the single ion anisotropy, and larger values of the ordered state moment, $\approx$ 1 $\mu_B$~\cite{21}.
\begin{figure}[t]
\vskip -0.6 cm
\centering
\includegraphics[width = 7 cm]{fig1.eps}
\caption {(Color online) (a) and (b) show Rietveld refinement of the neutron powder diffraction pattern of Ce$_{0.9}$La$_{0.1}$T$_2$Al$_{10}$ (T = Ru and Os) at 300 K collected using the D2B diffractometer with wavelength $\lambda$ =1.594 \AA. The circle symbols (black) and solid line (red) represent the experimental and calculated intensities, respectively, and the line below (blue) is the difference between them. Tick marks indicate the positions of Bragg peaks in the Cmcm space group.}
\end{figure}
\par
Koyabasi {\it et. al.}~\cite{23} and Nishioka {\it et. al.}~\cite{5} suggested that the phase transition at $T_N$ is not due to simple RKKY interactions as it is not possible to describe the behavior of Ce$_{1-x}$Gd$_x$Ru$_2$Al$_{10}$ and Ce(Fe$_{1-x}$Ru$_x$)$_2$Al$_{10}$. Kondo {\it et. al.}~\cite{22} found that $T_N$ is suppressed by the application of a magnetic field as well as by La substitution on the Ce site in CeRu$_2$Al$_{10}$. Considering these interesting observations~\cite{pp,24} found with electron and hole doped CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$, it is timely to investigate the effect of chemical pressure (La or Y doping) on the Ce site in CeT$_2$Al$_{10}$ (T = Ru and Os) using microscopic techniques such as neutron diffraction, inelastic neutron scattering and $\mu$SR measurements. We therefore present in this paper our results of such measurements on La-substituted Ce$_{1-x}$La$_x$T$_2$Al$_{10}$ (T = Ru and Os) to shed light on the nature of the phase transition and the ground state of the Ce ion in these compounds. Our study is motivated by the highly unusual magnetic ordering in the Ru and Os compounds of the CeT$_2$Al$_{10}$ series, with questions about the enigmatic behavior first posed by Nishioka {\it et. al.}~\cite{5}. Moreover, in the special class of Kondo insulator materials, the Ru and Os compounds are to date the only cases where such a strongly hybridized and unstable $f-$shell condenses into a long-range magnetic ordered ground state.
\par
The effect of La and Y substitution on the parent compound CeRu$_2$Al$_{10}$ has been investigated before by three groups by studying magnetic and transport properties~\cite{pp, 23, 24}. La (Y) has a bigger (smaller) ionic radius compared to Ce and hence La substitution expands the lattice corresponding to negative chemical pressure and Y substitution contracts the lattice corresponding to that to positive chemical pressure. The results are interesting but unexpected~\cite{pp,23}. With increasing La concentration x in Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$, the transition temperature is progressively shifting to lower temperature and vanishes near the critical composition $x_c$ $\approx$ 0.7 ~\cite{24,yoy}. Surprisingly Y$-$substitution leads as well to a decrease of $T_N$ rather than to an increase as one would expect for positive chemical pressure, and disappears suddenly between $x$ = 0.4 and 0.5~\cite{yoy}. This behavior cannot be understood by a simple magnetic phase transition, suggesting that the change in the valence of Ce ion plays an important role in the mysterious phase transition ~\cite{25,yoy}. The high-field magnetization measurements on Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0 and 0.25) performed by Kondo {\it et. al.} ~\cite{22} revealed that the long-range order disappears at a critical fields $H_c$ = 50 and 37 T for $x$ = 0 and 0.25, respectively.
\par
From the single crystal susceptibility measurements Tanida {\it et. al.} have proposed that 10\% La doping in CeRu$_2$Al$_{10}$ changes the direction of the ordered state moment from the $c-$axis found in undoped system to the $b-$axis, the hard axis of magnetization. Application of a pressure of 0.3 GPa changes the moment back to the $c-$axis ~\cite{26}. However,magnetic susceptibility data can only give indirect information on the ordered state moment direction and can be erroneous if anisotropic exchange interactions are dominating over the single ion crystal field anisotropy. They cannot give a direct measure of the value of the ordered state moment. The unanswered question remains what is happening to the spin gap formation and its energy scale as the direction of the ordered state moment of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ changes to the $b-$axis from the $c-$axis as in CeRu$_2$Al$_{10}$. In order to answer these questions we have carried out neutron diffraction and inelastic neutron scattering measurements on Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$. The neutron diffraction study provides direct information on the direction as on the magnitude of the ordered moment. Inelastic neutron scattering gives direct information about the magnitude of the spin-gap energy, its temperature and wave-vector (Q) dependency.
\par
In order to gain further information on the microscopic change in the magnetism we have performed muon spin rotation measurements on Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0, 0.3, 0.5 and 0.7) alloys. $\mu$SR is an exceptionally sensitive microscopic probe of cooperative magnetic ordering phenomena and is thus ideally suited to our compounds where previous studied had alluded very small magnetic moment values. Our neutron diffraction study reveals a long$-$range magnetically ordered ground state in both Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ compounds. More interestingly the ordered Ce moment of 0.18 $\mu_B$ is along the $b-$axis in the former, but along the $c-$axis with a value of 0.23 $\mu_B$ for the latter. Our INS study reveals the presence of a well defined spin gaps of 7 meV at 2 K in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and 10 meV at 4.5 K with considerable reduced intensity in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$.
\begin{table}[b]
\begin{center}
\caption{Lattice parameters of Ce$_{1-x}$La$_{x}$T$_2$Al$_{10}$ (T = Ru, Os) for $x$ = 0 , 0.1 refined from the neutron diffraction data collected at 300 K in the orthorhombic Cmcm space group.}
\begin{tabular}{lccccccccccccccc}
\hline
Compounds && && a (\AA) && b (\AA) && c (\AA) && V (\AA) &&\\
\hline
CeRu$_2$Al$_{10}$ && && 9.1246 && 10.2806 && 9.1878 && 861.9~\cite{ymuro}&& \\
Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ && && 9.1224 && 10.2749 && 9.1865 && 861.066 &&\\
CeOs$_2$Al$_{10}$ && && 9.138 && 10.2662 && 9.1694 && 861.686~\cite{2}&&\\
Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ && && 9.1412 && 10.2668 && 9.1898 && 862.474 && \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{figure}[t]
\vskip 0.4 cm
\centering
\includegraphics[width = 9 cm]{fig2.eps}
\caption {(Color online) (a-b) Temperature dependence of dc magnetic susceptibility $\chi(T)$ Ce$_{1-x}$La$_x$T$_2$Al$_{10}$ with $x$ = 0.1 (T = Ru and Os). Inset shows the variation of $T_N$ with La composition for Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$~\cite{pp}. (c-d) Semilogarithmic plot of electrical resistivity vs temperature. (e-f) Temperature variation of specific heat $C_P$ divided by temperature for Ce$_{1-x}$La$_x$T$_2$Al$_{10}$ (T = Ru and Os) (open symbol) with non magnetic counterpart LaRu$_2$Al$_{10}$ and LaOs$_2$Al$_{10}$ (solid line).}
\end{figure}
\begin{figure}[t]
\vskip 0.4 cm
\centering
\includegraphics[width = 5 cm]{fig3.eps}
\caption {(Color online) (a) Temperature variation of magnetic specific heat $C_{4f}$ for Ce$_{0.9}$La$_{0.1}$T$_2$Al$_{10}$ (T = Ru, Os). (b) Calculated magnetic entropy as a function of temperature.}
\end{figure}
\section{Experimental Details}
The polycrystalline samples of Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0, 0.1, 0.3, 0.5 and 0.7), Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, LaRu$_2$Al$_{10}$ and LaOs$_2$Al$_{10}$ were prepared by argon arc melting of the stoichiometric constituents with the starting elements, Ce/La 99.9\% (purity), Ru/Os 99.9\% and Al 99.999\%. The samples were annealed at 800 $^\circ$C for seven days in an evacuated quartz ampoule. The samples were characterized using powder X-ray diffraction or neutron diffraction (on D2B diffractometer at ILL, Grenoble) at 300 K and were found to be dominantly single-phase (see Fig. 1). Magnetic susceptibility measurements were made using a Magnetic Property Measurement System (MPMS) superconducting quantum interference device (SQUID) magnetometer (Quantum Design). Electrical resistivity by the four probe method and heat capacity by the relaxation method were performed in a Quantum Design Physical Properties Measurement System (PPMS).
\par
The $\mu$SR experiments were carried out using the MUSR spectrometer in longitudinal geometry at the ISIS muon source, UK. At the ISIS facility, a pulse of muons is produced every 20 ms and has a FWHM of $\approx$ 70 ns. These muons are implanted into the sample and decay with a half-life of 2.2 $\mu$s into a positron which is emitted preferentially in the direction of the muon spin axis. These positrons are detected and time stamped in the detectors which are positioned before, F, and after, B, the sample. The positron counts, N$_{F,B}$(t), have the functional form
\begin{equation}
N_{F,B}(t)=N_{F,B}(0)e^{-t/\tau_\mu}(1\pm G_z(t))
\end{equation}
where G$_z$(t) is the longitudinal relaxation function. G$_z$(t) is determined using
\begin{equation}
G_z(t)=(N_F(t)-\alpha N_B(t))/(N_F(t)+ \alpha N_B(t))
\end{equation}
where $\alpha$ is a calibration constant which was determined at 35 K by applying a small transverse field ($\approx$ 20 Oe) and adjusting its value until the resulting damped cosine signal was oscillating around zero. The powdered samples were mounted onto a 99.995+\% pure silver plate.
\par
The low temperature neutron diffraction measurements at 1.5 K and 35 K on Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ samples were performed using the high neutron flux D20 diffractometer at ILL, Grenoble, France using constant wavelengths of 1.3 \AA~or 2.41 \AA. The powder samples were mounted in a 10 mm diameter vanadium can, which was cooled down to 2 K using a standard He-4 cryostat. The program FULLPROF~\cite{ndr} was used for Rietveld refinements and group theoretical calculations were performed with the aid of the Sarah/ISOTROPY software ~\cite{27,iso}.
\par
The inelastic neutron scattering measurements between 2 K and 35 K on Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$, Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, CeRu$_2$Al$_{10}$, LaRu$_2$Al$_{10}$ and LaOs$_2$Al$_{10}$ (15 g sample) were carried out using the MARI time-of-flight (TOF) chopper spectrometer at ISIS Facility, while on Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and LaRu$_2$Al$_{10}$ additional data were collected on the IN4 TOF chopper spectrometer at ILL, Grenoble, France. On MARI the samples were wrapped in a thin Al-foil and mounted inside a thin-walled cylindrical Al-can, which was cooled down to 4.5 K inside a top-loading closed-cycle-refrigerator (TCCR) with He-exchange gas around the samples. The measurements were performed with an incident neutron energy E$_i$ of 25 (20) meV, with an elastic resolution (at zero energy transfer) of 1.1 meV (0.8 meV) (FWHM). On IN4 the samples were wrapped in a thin Al-foil, which was cooled down to 2 K inside a standard He-4 cryostat with He-exchange gas around the samples. The measurements were performed with an incident neutron energy E$_i$ of 16.9 meV, with an elastic resolution (at zero energy transfer) of 1.1 meV (FWHM).
\par
\begin{figure}[t]
\vskip 0.4 cm
\centering
\includegraphics[width = 9 cm]{fig4.eps}
\caption {(Color online) Rietveld refinements of the magnetic intensity of Ce$_{1-x}$La$_{x}$T$_2$Al$_{10}$ (T = Ru, Os) along with Ce(Ru/Os$_{1-x}$Re$_x$)$_2$Al$_{10}$ ($x$ = 0.03) obtained as a difference between the diffraction patterns collected at 1.5 K and 35 K. The circle symbols (red) and solid line represent the experimental and calculated intensities, respectively, and the line below (blue) is the difference between them. Tick marks indicate the positions of Bragg peaks for the magnetic scattering with the ${k}$= (1, 0, 0) propagation vector. The upper panel shows the magnetic structures of the La-doped (found in this paper), Ce(Ru/Os$_{1-x}$Re$_x$)$_2$Al$_{10}$ ($x$ = 0.03) and CeT$_2$Al$_{10}$ (T = Ru and Os) samples. For clarity, only Ce and Ru/Re atoms are shown (top).}
\end{figure}
\section{Results and discussions}
\subsection{Bulk properties}
To explain the overall bulk properties of the title compounds, we present here their magnetic susceptibility, electric resistivity, and specific heat data with emphasis on the magnetic phase transitions. Figs. 2 (a$-$f) show the temperature dependent magnetic susceptibility, electrical resistivity, and heat capacity of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. The magnetic susceptibility $\chi(T)$ of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ exhibits a clear peak near 23.0 K, which is due to an antiferromagnetic ordering of Ce moments. On the other hand $\chi(T)$ of Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ exhibits a broad maxima near 40 K and a kink near 21.7 K. The former is due to an opening of the spin gap, while the latter is due to the onset of the antiferromagnetic ordering. A very similar behavior of $\chi(T)$~\cite{7} with a broad maximum at 31 K (due to spin gap formation) above $T_N$= 23 K is observed for CeRu$_{1.96}$Re$_{0.06}$Al$_{10}$. The inverse magnetic susceptibilities of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ exhibit Curie-Weiss behavior between 50 K and 300 K. A linear least-squares fit to the data yields an effective magnetic moment $\mu_{eff}$ = 2.12 $\mu_B$ and a paramagnetic Curie temperature $\theta_p$ = $-$95 K for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and $\mu_{eff}$ = 2.58 $\mu_B$ and $\theta_p$ = $-$175 K for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. The value of the magnetic moment suggests that the Ce atoms are in their normal Ce$^{3+}$ valence state in both the compounds. The negative value of $\theta_p$ is in agreement with AFM ordering, a negative sign for the exchange interactions, and/or the presence of the Kondo effect. The larger negative value of $\theta_p$ of Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ compared to that of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ suggests a stronger hybridization in the former. The inset in the Fig. 2(a) shows the La composition dependence of $T_N$ of Ce$_{1-x}$La$_{x}$Ru$_2$Al$_{10}$, which reveals that $T_N$ decreases almost linearly and becomes zero near x$\ge$0.7~\cite{9,pp}. In this critical region of compositions the heat capacity exhibits a rise at low temperature suggesting the presence of non-Fermi-liquid (NFL) behavior close to a quantum critical point (QCP)~\cite{pp}.
\par
Figs. 2(c) and (d) show the electrical resistivity $\rho(T)$ of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ samples, respectively. At high temperature $\rho(T)$ of both samples increases with decreasing temperature up to $T_N$. Then $\rho(T)$ of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ exhibits a peak near $T_N$ and remains metallic at low temperature, while for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ the $\rho(T)$ shows a slope change at $T_N$, but still increases with further decrease in the temperature down to 2 K. This contrasting behavior of the low temperature resistivity is similar to that observed in the undoped compounds ~\cite{11, 13}. It is interesting to note that the resistivity of slightly electron-(8\% Ir) and hole-(2\% Re) doped CeOs$_2$Al$_{10}$ exhibits metallic behavior in all directions below $T_N$ ~\cite{19}.
\par
Figs. 2 (e) and (f) show the heat capacity divided by temperature $C_P/T$ vs $T$ of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ along with their respective nonmagnetic phonon reference compounds, LaRu$_2$Al$_{10}$ and LaOs$_2$Al$_{10}$. The heat capacity of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ exhibits a $\lambda-$type anomaly near $T_N$, while the anomaly is considerably suppressed in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. A rapid suppression of the heat capacity anomaly near $T_N$ was also observed both in CeRu$_{1.94}$Re$_{0.06}$Al$_{10}$ and CeOs$_{1.96}$Re$_{0.04}$Al$_{10}$ ~\cite{8,19}. By fitting the high temperature heat capacity (above $T_N$) to $C_P/T$=$\gamma$+$\beta T^2$, we have estimated the Sommerfeld coefficient $\gamma$ = 0.125 J/mol-K$^2$ and $\beta$= 3.5$\times$10$^{-4}$ J/mol-K$^4$ for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and $\gamma$ =0.121 J/mol-K$^2$ and $\beta$ = 5.6$\times$10$^{-4}$ J/mol-K$^4$ for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. The observed values of $\gamma$ for both the compounds are smaller than those observed for undoped compounds, $\gamma$ = 0.2 J/mol-K$^2$ for CeRu$_2$Al$_{10}$ and $\gamma$ = 0.541 J/mol-K$^2$ for CeOs$_2$Al$_{10}$, indicating heavy fermion behavior in the undoped compounds. From the value of $\beta$ = (12$\pi^4$/5) ($nN_Ak_B/\Theta_D^3$), where $N_A$ and $k_B$ have the usual meaning, and $n$ = 13 is the number of atoms per f.u., we estimated the Debye temperature to $\Theta_D$ =416 K and 355 K for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ respectively. Fig. 3 (a) shows the magnetic heat capacity variation with temperature for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. The weak anomaly for the former at temperature below 5 K may be attributed to impurity contribution. The value of magnetic entropy $S_{mag}$ [Fig. 3 (b)] at 30 K is 3.5 J/mol-K for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and 3.55 J/mol-K for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, which is much smaller than Rln(2) = 5.76 J/mol-K. The reduced magnetic entropy can be explained on the basis of the Kondo effect. We also estimated the gap in the spin wave by fitting $C_{mag(T)}$ data below $T_N$, and we find gaps 50 K for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and 60 K for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, whose values are approximately half of those reported for the undoped compounds~\cite{5}.
\begin{figure}[t]
\centering
\includegraphics[width = 9.1 cm]{fig5.eps}
\caption {(Color online) $Q$-integrated (0$\le$$Q$$\le$2.5~\AA) intensity versus energy transfer of (a-b) CeT$_2$Al$_{10}$ (T = Ru and Os) measured on the MARI spectrometer (c-f) Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ measured on the IN4 spectrometer along with the nonmagnetic phonon reference compounds LaT$_2$Al$_{10}$ (T = Ru and Os) (red half filled circles), measured with respective incident energy of $E_i$ = 20 meV. For IN4 we have made cuts in scattering angle between 13$^\circ$ to 43$^\circ$.}
\end{figure}
\section{Neutron diffraction}
Figs. 1 (a) and (b) show the neutron diffraction patterns for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, collected at 300 K on the D2B instrument in the high resolution mode which both are consistent with the Cmcm symmetry and can be satisfactorily fitted with the structural model proposed earlier~\cite{Thiede}. The structural parameters for both samples are listed in Table I. A comparison of the lattice parameters of the doped compounds with the parent ones shows that 10\% La doping in CeRu$_2$Al$_{10}$ results in a volume contraction of 0.001, while 10\% La doping in CeOs$_2$Al$_{10}$ increases the volume by 0.001. This can be compared to the volume contraction of 0.10\% observed in 3\% Re-doped (or hole doped) CeRu$_2$Al$_{10}$, where the ordered state moment of 0.20 $\mu_B$ is along the $b-$axis. It is generally agreed that hole-doping in both CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$ increases the hybridization~\cite{7, 20}.
\par
To investigate the magnetic structure, we carried out neutron diffraction measurements on the D20 instrument at 1.5 K and 35 K (12 hours each temperature) on Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. At 1.5 K we observed very weak magnetic Bragg peaks at scattering angles away from the nuclear Bragg peaks, which confirmed the long$-$range antiferromagnetic ordering of the Ce moment at 1.5 K in both compounds. To see the magnetic Bragg peaks clearly, we plotted the difference between the 1.5 K and 35 K data for both compounds, which are shown in Fig. 4, along with their 3\% Re doped counterparts (see Refs ~\cite{7, 20} for details of the neutron diffraction experiment of these materials). The important observation is the qualitatively similar diffraction patterns between the La and Re doped compositions. The absence of the magnetic (0, 1, 0) peak near 10.3 \AA~in both Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and CeRu$_{1.94}$Re$_{0.06}$Al$_{10}$ indicates the ordered moments to be along the $b-$axis, in contrast with the parent CeRu$_2$Al$_{10}$ compound as well as with the Ce(Ru$_{1−x}$Fe$_x$)$_2$Al$_{10}$ series, where the moments were found to be along the $c-$axis ~\cite{7}. The conclusion about the $b-$axis moment direction comes directly from the fact that the magnetic neutron diffraction intensity is proportional to the square of the ordered moment component perpendicular to the scattering vector. On the other hand in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ several magnetic Bragg peaks were observed including the (0, 1, 0) one and the relative intensities of these peaks are very similar to those observed in CeRu$_2$Al$_{10}$, indicating that the moments are probably along the $c-$axis. In spite of the different moment directions, all the observed magnetic Bragg peaks in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ can be indexed based on the same propagation vector $k$ = (1, 0, 0), which is identical to that found in pure CeRu$_2$Al$_{10}$ ~\cite{3}. It is to be noted that the propagation vector $k$ = (0, 1, 0) proposed for CeOs$_2$Al$_{10}$ is equivalent to $k$ = (1, 0, 0) as they are both related by the allowed ($h$+$k$ even) reciprocal translation G = (-1, 1, 0).
\par
In the qualitative refinement of the magnetic structures, we employed a method whereby combinations of axial vectors localized on the 4$c$(Ce) site (as Ce is only the magnetic atom) and transforming as basis functions of the irreducible representations of the wave vector group are systematically tested. The symmetry analysis yields that the reducible magnetic representation is decomposed into six one-dimensional representations, labeled Y$^+_i$ ($i$ = 2,3,4) and Y$^-_i$ ($i$ = 1,2,3). The Y$^+_i$ representations result in a ferromagnetic (FM) alignment of the Ce moments within the primitive unit cell, along different crystallographic directions. On the other hand, Y$^-_i$ transform Ce moments which are AFM coupled within the primitive unit cell. We refined the difference data (1.5 K$-$35 K) using the three AFM structures (with the moments along a$-$, b$-$ and c$-$axes) given by Y$^-_i$ representations. The best fit to the data was obtained with a Ce ordered state moment of 0.18(2)$\mu_B$ AFM coupled along the $b-$axis for Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ [see Fig. 4, where the magnetic unit cells are shown], while an ordered moment of 0.23(1) $\mu_B$ along the $c-$axis was obtained for Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$.
\par
It is interesting to compare these values of the ordered state moments with those found in the undoped systems, 0.34(2) $\mu_B$ for CeRu$_2$Al$_{10}$ and 0.29(2) $\mu_B$ for CeOs$_2$Al$_{10}$ both along the $c-$axis ~\cite{7,8}. This comparison shows a reduction of the ordered moment in 10\% La doped systems, which might be due to a change in the Ce valence with La-doping as proposed in Ref.~\cite{25}. The different directions of the ordered state moment in 10\% La doped CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$ is a surprising observation and can be only explained by assuming a different degree of hybridization (weak in the former material). This means that 10\% La-doped CeOs$_2$Al$_{10}$ sees anisotropic exchange interactions which are still similar to those in the undoped compound preserving hence the moment direction. It should be possible to change the moment direction as well in La$-$doped CeOs$_2$Al$_{10}$ to the $b-$axis by increasing the La content up to 20 or 30\% in CeOs$_2$Al$_{10}$. However, the absolute value of the moment would certainly reduce further with increased doping making its experimental detection difficult. The question remains why$-$despite having a smaller hybridization the ordered moment in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ is smaller in comparison to that found in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. A possible explanation could be that the Kondo effect along the $b-$axis is stronger than along the $c-$axis, which might screen the moment value.
\section{Inelastic neutron scattering study}
With the dramatic and contrasting changes observed in the moment direction and its absolute value in 10\% La-doped CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$ it would be interesting to investigate directly the spin gap formation in these compounds using inelastic neutron scattering. Furthermore, Kawabata {\it et. al.}~\cite{19} reported that the suppression of $T_N$ is well correlated with the gap energy $\Delta$ as a function of electron-(Ir) and hole-(Re) doping and they conclude that the presence of the hybridization gap is indispensable for the AFM order at unusually high $T_N$ in CeOs$_2$Al$_{10}$. Thus the information on the spin gap energy scale in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ is very important. Therefore, we briefly report the INS spectra, which give direct information of the spin gap energy, below and above $T_N$ of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ and also of CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$ for comparison in this section. A detailed report on the inelastic neutron scattering investigations on CeT$_2$Al$_{10}$ (T = Fe, Ru and Os) compounds can be found in Ref. ~\cite{2}.
\par
Fig. 5 displays the inelastic neutron scattering spectra of 10\% La-doped compounds with that of undoped compounds at two temperatures at low$-$Q measured on the MARI and IN4 spectrometers. There is a clear magnetic excitation centered around 7 meV and 10 meV in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ and Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$, respectively which may be compared to the 8 meV and 11 meV excitations found in the parent compounds CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$, respectively~\cite{2,7,8}. The value of the peak position can be taken as a measure of the spin gap energy in these compounds. These results show that in both 10\% La doped systems, despite the moment direction being different and the ordered state moments being reduced, a very small change in the spin gap energy scale is observed. On the other hand the intensity of the spin gap does not change much in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$, while a dramatic reduction in the intensity of the spin gap is observed in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. An interesting observation is the existence of a clear low energy response at low Q in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ at 2 K. This signal is not observed in the non-magnetic phonon reference compound LaRu$_2$Al$_{10}$, which confirms its magnetic origin. Further the high resolution INS study on CeRu$_2$Al$_{10}$ by J. Robert {\it et. al.}~\cite{16} did not reveal any clear sign of the low energy or quasi-elastic excitation at 11 K. This reveals that the low energy excitation in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ has some relation with changes in the moment direction from the $c-$axis to the $b-$axis. It is an open question whether this could be the zero frequency mode observed in several magnetically ordered heavy fermion systems~\cite{bk1, nhv}. To understand this, spin wave measurements on single crystals of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ are highly desirable. On the other hand within the resolution of the MARI experiment we could not see any clear sign of a low energy excitation at 4.5 K in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$; this excitation was also absent in CeOs$_2$Al$_{10}$ from the high resolution INS study~\cite{2}.
\par
Now we discuss the temperature dependence of the spin gap excitation. For Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ with increasing temperature to 10 K no dramatic changes were observed in the spectra, but at 20 K the spin gap energy decreases to 5.5 meV and its width increases to 1.26 meV. Further increase in the temperature to 25 K (same response at 30 K) the inelastic response continues to broaden. The data show two components, a low energy/quasi-elastic component with narrow linewidth and a second, distinctly broader component. The Q-dependent integrated intensity of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ between 5 and 8 meV at 2 K nearly follows the Ce$^{3+}$ magnetic form factor squared (F$^2$(Q), figure not shown), very similar to that observed in pure CeRu$_2$Al$_{10}$~\cite{4}. The observed single ion type response of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ in the magnetic ordered state could be due to the fact that the observed spin wave scattering intensity in CeRu$_2$Al$_{10}$ single crystal is stronger near the zone boundary and considering the presence of powder averaging effect could give single-ion type behavior. It is to be noted that the single-ion type response is also observed in the inelastic response of the spin gap system CeRu$_4$Sb$_{12}$ (no magnetic ordering down to 2 K), which does not exhibit any long-range magnetic ordering down to 50 mK~\cite{29}. On the other hand, the deviation from a single-ion response is observed in the spin gap system CeFe$_4$Sb$_{12}$, where it was proposed that the intersite interactions between Ce and Fe are playing an important role ~\cite{30}. As the spin gaps in CeOs$_2$Al$_{10}$ and CeRu$_2$Al$_{10}$ open up below (or just above) the magnetic ordering temperature one would expect that the spin gap energy and its intensity would be strongly Q$-$dependent, but this is not the case. Furthermore, with increasing temperature to 30 K the response of Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ also becomes very broad and the intensity decreases considerably [Fig. 5 (f)] compared with 4.5 K.
\begin{figure}[t]
\centering
\includegraphics[width = 7 cm]{fig6.eps}
\caption {(Color online) The time evolution of the muon spin relaxation in Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ for various temperatures (above and below $T_N$) in zero field. The solid line is a least-squares fit [using Eq. (3) (above $T_N$) and Eq. (4) (below $T_N$)] to the data as described in the text.}
\end{figure}
\begin{figure}[t]
\vskip 0.4 cm
\centering
\includegraphics[width = 9 cm]{fig7.eps}
\caption {(Color online) The temperature dependence of (a, c, e) muon precession frequency/internal field at the muon site in Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$. (b, d, f) the depolarization rate $\sigma$. The solid line in (c, e) is fit to the data using Eq. (4) (see text).}
\end{figure}
\section{Muon Spin relaxation}
Figs. 6(a$-$f) shows the zero-field (ZF) $\mu$SR spectra at various temperatures of Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0, 0.3, 0.5 and 0.7). The left hand side figures show the spectra at low temperatures, while the right hand side figures show the spectra at high temperature. It is interesting to see the dramatic change in the time-evolution of the $\mu$SR spectra with temperature for all compositions, except for $x$ = 0.7. At 35 K (or 4.5 K) we observe a strong damping at shorter time, and the recovery at longer times for all compositions, which is a typical muon response to nuclear moment, known as the Kubo$-$Toyabe~\cite{31}, arising due to a static distribution of the nuclear dipole moment. Here it arises from the La (stable isotope $^{138}$La, I=5, 0.09\% abundance and $^{139}$La, I=7/2, 99.91\% abundance), Ru (stable isotope $^{99}$Ru, I=5/2, 12.76\% abundance, $^{101}$Ru , I=5/2, 17.06\% abundance) and Al (I=5/2) nuclear moment contributions (I = 0 for Ce, i.e. zero contribution). Above the magnetic ordering temperature the $\mu$SR spectra of all compounds ($x$ = 0 to 0.7) can all be described by the following equation (see Fig. 6, right hand figures):
\begin{equation}
G_z (t)=A_0 \big[\frac{1}{3} + \frac{2}{3} (1-(\sigma t)^2 )e^{(-\sigma t)^2/2}\big]e^{-\lambda t}+A_{bg}
\end{equation}
where $A_0$ is the initial asymmetry, $\sigma/\gamma_\mu$ is the local field distribution, $\gamma_\mu$=13.55 MHz/T is the gyromagnetic ratio of the muon, $\lambda$ is the electronic relaxation rate arising from electronic moments and $A_{bg}$ is a constant background. It is assumed that the electronic moments give an entirely independent muon spin relaxation channel in real time. The value of $\sigma$ was found to be 0.32(3) $\mu$s$^{-1}$ for all compositions from fitting the spectra at high temperature (35 K or 4.5/4 K) to Eq. (3), which suggests that muon stopping sites are the same for all compositions. From a simple electrostatic potential calculations of CeRu$_2$Al$_{10}$ Kambe {\it et. al.}~\cite{32} have proposed muon stopping site at 4$a$ (0,0,0), while Khalyavin {\it et. al.}~\cite{ddk} proposed muon stopping site at 4$c$ (0.5,0,0.25) (using electrostatic calculation of CeOs$_2$Al$_{10}$). Further the Density Functional Theoretical (DFT) calculation of muon stopping sites in Ce(Ru$_{1-x}$Rh$_x$)$_2$Al$_{10}$~\cite{33} supports both muon sites equally possible as they are located at local minimum positions of the electrostatic potential. Further the DFT calculation revealed that the muon stopping site (or position) estimated from the potential calculation do not change much with Rh-doping~\cite{34} as there are no big changes in the potential around the suggested muon sites.
\par
Now we discuss the muon spectra at low temperatures. As shown in the left side of Fig. 6, $\mu$SR spectra of $x$ = 0, 0.3 and 0.5 compounds exhibit a clear sign of coherent frequency oscillations confirming the long range magnetic ordering of the Ce moment. On the other hand $\mu$SR spectra of $x$ = 0.7 at 44 mK reveals the same behavior as that observed at 4.0 K, indicates the non-magnetic (or paramagnetic) ground state. The result is in agreement with the proposed phase diagram $T_N$ vs $x$ (La concentration), reveals $T_N \approx$0 K for $x\ge$ 0.6 [see the inset in Fig. 2 (a)]. For the other compositions, the spectra below $T_N$ are best described by two oscillatory terms and an exponential decay, as given by the following equation
\begin{equation}
G_z (t)=\sum_{i=1}^{2}A_i cos(\omega_i t+\phi)e^{-(\sigma t)^2/2}+A_3 e^{-\lambda t}+A_{bg}
\end{equation}
where $\omega$=$\gamma_\mu$$H_{int}$ is the muon precession frequency ($H_{int}$ is the internal field at the muon site) and $\phi$ is the phase. In Fig. 7 (left) we have plotted the internal field (or muon precession frequency) at the muon site as a function of temperature. This shows that the two internal fields (or frequencies) appear just below 27 K in $x$ = 0, 12.5 K in $x$ = 0.3 and 4.2 K in $x$ = 0.5, showing a clear onset of bulk long-range magnetic order in agreement with $T_N$ proposed in the inset of Fig. 2 (a). It is interesting to note that even though the heat capacity of $x$ = 0.3 and 0.5 exhibits a very broad $\lambda-$ type anomaly near $T_N$ ~\cite{35}, the $\mu$SR shows a typical second order phase transition that can be explained by mean field behavior. Further, the associated internal fields are found to be very small in agreement with a small ordered magnetic moment of the Ce$^{3+}$ ion observed in the neutron diffraction for $x$ = 0 and 0.1. The observed two values of internal fields can be explained on the basis of the dipole field calculation. It was found that Kambe's suggested positions correspond to the 4$a$ sites which had the lower field and Khalyavin's suggested positions correspond to the 4$c$ sites which had the higher field. Further the temperature dependence of $\sigma$ and the differences between the two values of $\sigma$ also exhibit very similar behavior to that of the internal fields. The value of $\lambda$ was bound to be nearly temperature independent for all values of $x$.
\par
Now examining the temperature dependence of the internal fields, we can see that there is a dip in the internal field (see Fig. 7 top left), which occurs around 15 K. Moreover, below 15 K the first and the second component of the depolarization rates also increase (Fig. 7 right). In principle this could originate from various phenomena related to a change in distribution of internal fields, but a structural transition is a likely candidate in view of the structural instability reported on this system~\cite{1}. To find out the value of critical exponents and hence get more information on the nature of the magnetic transition, the temperature dependence of the internal field was fitted~\cite{ab}:
\begin{equation}
H_{int}(T)= H_0\left(1-\left(\frac{T}{T_N}\right)^{\alpha}\right)^{\beta}
\end{equation}
Fitting the temperature dependent internal field of $x$ = 0.3 to Eq. (5) we obtained the value of the parameters for higher [and for lower] internal field: $T_N$ =12.9(3) (same for both fields), $H_0$=16.5(3) Oe [11.8(7) Oe], $\alpha$ = 1.65(6) [1.42(4)] and $\beta$ = 0.81(3) [0.89(2)]. Fit for $x$ = 0.5 data fitting higher field we obtained $T_N$ = 4.06(9), $H_0$ = 9.6(20) Oe, $\alpha$ = 0.99(30) and $\beta$ = 1.2(12). It is to be noted that due to limited temperature range the fit to lower field did not converge. It is interesting to compare these values of the exponents with $\alpha$ = 1.47(2) and $\beta$ = 0.96 observed in CeRu$_{1.94}$Re$_{0.06}$Al$_{10}$~\cite{7}. The larger values of beta compared to 0.5, expected from the mean field theory suggests that magnetic interactions are complex in nature.
\par
Now we compare the results of our neutron diffraction and $\mu$SR of Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ with that of the hole doped systems, Ce(Ru$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$ and Ce(Os$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$. The neutron diffraction study of Ce(Ru$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$ shows that the compound orders antiferromagnetically with a propagation vector $k$ = (1,0,0) and the ordered state moment is 0.20(1) $\mu_B$ along the $b-$axis, in sharp contrast with the ordered moment of 0.34$-$0.42$\mu_B$ along the $c-$axis observed in CeRu$_2$Al$_{10}$ ($T_N$ = 27 K)~\cite{20}, which is very similar behaviour observed in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$. The $\mu$SR study on Ce(Ru$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$ reveals the presence of one internal field (frequency) with value of 80 Oe at 1.2 K. The observed single frequency in Ce(Ru$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$ can be explained by the dipolar field calculation with muon stopping site at 4$c$. On the other hand the $\mu$SR study on Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0, 0.3 and 0.5) shows the presence of two frequencies.
\par
Further Ce(Os$_{0.97}$Re$_{0.03}$)$_2$Al$_{10}$ has been studied by muon spin relaxation and neutron diffraction measurements. A long-range antiferromagnetic ordering of the Ce sublattice with a substantially reduced value of the magnetic moment 0.18(1) $\mu_B$ along the $c-$axis (same direction as in the undoped system) has been found below $T_N$ = 21 K. On the other hand the electron doping (i.e Ir and Rh) in CeRu$_2$Al$_{10}$ and (Ir) in CeOs$_2$Al$_{10}$ show a large ordered moment of $\approx$1 $\mu_B$ along the $a-$axis. The obtained result reveals the crucial difference between electron- and hole-doping effects on the magnetic ordering in CeT$_2$Al$_{10}$ (T = Ru and Os). The former suppresses the anisotropic $c-f$ hybridization and promotes localized Ce moments controlled by single ion anisotropy. On the contrary, the latter increases the hybridization, keeping the dominant role of the anisotropic exchange on the direction of the moments and shifts the system towards a delocalized nonmagnetic state~\cite{21}.
\par
Finally, it should be pointed out that the obtained results pose a question about the role of the hybridization and the crystal field effects in the reduced moment nature of the magnetic ground state in the undoped CeT$_2$Al$_{10}$ (T = Ru and Os) compounds. The behavior of the system under the hole doping, electron doping and chemical pressure (positive by Y-doping and negative by La-doping) along with the applied hydrostatic pressure study points to the key role of the hybridization in the anisotropic character of the exchange interactions observed in the undoped compound as well. This also implies the hybridization effect on the moments reduction, attributed by Strigari {\it et. al.}~\cite{36} to the crystal field effects implicitly.
\begin{table}[t]
\begin{center}
\caption{A list of samples studied Ce$_{1-x}$La$_{x}$T$_2$Al$_{10}$ (T = Ru, Os) for $x$ = 0 , 0.1 and their transitions temperatures ($T_N$ = antiferromagnetic ordering temperature), ground state magnetic moment ($\mu_{AF}$) value and moment directions ($\mu_{d}$) obtained from neutron diffraction study.($^*$this work)}
\begin{tabular}{lccccccccccccccc}
\hline
Compounds && && $T_N$ && $\mu_{AF}$ && $\mu_{d}$ & Spin Gap\\
&& && (K) && ($\mu_B$) && &(meV)\\
\hline
CeRu$_2$Al$_{10}$~\cite{2} && && 27.0 && 0.34(2) && $c-$axis~\cite{2} &8.0\\
Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$ && && 23.0 && 0.18(2) && $b-$axis$^*$& 7.0\\
CeOs$_2$Al$_{10}$ && && 28.5 && 0.29(2) && $c-$axis~\cite{2} & 11\\
Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$ && && 21.7 &&0.23(1) && $c-$axis$^*$ &10\\
\hline
\end{tabular}
\end{center}
\end{table}
\section{Conclusions}
We have carried out a comprehensive study on Ce$_{0.9}$La$_{0.1}$T$_2$Al$_{10}$ (T = Ru and Os) using the complementary techniques of magnetization, resistivity, heat capacity, neutron diffraction, inelastic neutron scattering and muon spin relaxation measurements to understand the unusual behavior of the magnetic moment direction and the opening of a spin gap below $T_N$. The neutron diffraction study is unambiguous in confirming the long-range magnetic order in this compound. More interestingly our ND study shows a very small ordered moment of 0.18 $\mu_B$ along the $b-$axis in Ce$_{0.9}$La$_{0.1}$Ru$_2$Al$_{10}$, but a moment of 0.23 $\mu$B along the $c-$axis in Ce$_{0.9}$La$_{0.1}$Os$_2$Al$_{10}$. This contrasting behavior can be explained based on a different degree of hybridization in CeRu$_2$Al$_{10}$ and CeOs$_2$Al$_{10}$: hybridization is stronger in the latter than in the former. Our INS study reveals the presence of a spin gap of 7 meV and 10 meV in the 10\% La-doped T = Ru and Os compounds, respectively. Interestingly the intensity of the spin gap decreases dramatically in La-doped CeOs$_2$Al$_{10}$ compound.
\par
Further we also present muon spin rotation study on Ce$_{1-x}$La$_x$Ru$_2$Al$_{10}$ ($x$ = 0, 0.3, 0.5 and 0.7), which reveals the presence of two coherent frequency oscillations indicating a long$-$range magnetic ground state in $x$ = 0 to 0.5, but almost temperature independent Kubo-Toyabe response between 45 mK and 4 K for $x$ = 0.7. The absence of temperature dependent relaxation rate in $x$ = 0.7 despite of the logarithmic rise in the heat capacity down to mK is an unusual behavior. One would expect that near a quantum phase transition, $\mu$SR will sense the presence of quantum fluctuations even in the paramagnetic NFL state. These $\mu$SR results are very similar to that observed in YFe$_2$Al$_{10}$, where despite of NFL behavior observed in the magnetic susceptibility and heat capacity, the $\mu$SR spectra~\cite{2} are independent of temperature down to mK. A major achievement of this work has been the finding of frequency oscillations in our $\mu$SR spectra up to $x$ = 0.5, which for the first time establishes the onset of long-range magnetic ordering in the La-doped compounds and the robust magnetic ordering in spite of the large and anisotropic $c-f$ hybridization in this system. The temperature dependence of the $\mu$SR frequencies and muon depolarization rates of $x$ = 0 follow an unusual behavior with further cooling of the sample below 18 K, pointing at the possibility of another phase transition below 15 K. On the other hand the temperature dependence of the $\mu$SR frequencies and muon depolarization rates of $x$ = 0.3 follows conventional behavior expected for a second order phase transition in the mean field theory.
\section*{ACKNOWLEDGEMENT}
Some of us DTA/ADH would like to thank CMPC-STFC for financial support. The work at Hiroshima University was supported by KAKENHI (Grant No. 26400363) from JSPS, Japan. AMS thanks the SA-NRF (Grant 78832) and UJ Research Committee for financial support. A.B would like to acknowledge FRC of UJ, NRF of South Africa and ISIS-STFC for funding support. We would like to thank P. Manuel, J. M. Mignot and I. Watanabe for an interesting discussion.
|
\section{Introduction}
The Yangians, defined by Drinfeld \cite{Dr1, Dr2}, are certain non-commutative Hopf algebras that are important examples of quantum groups. They were studied to generate rational solutions of the {\em Yang-Baxter equation} and there were many applications in statistical mechanics and mathematical physics. Nowadays, the study of Yangians gives many new points of view and important applications to classical Lie theory; see the book \cite{Mo} and references therein.
Consider $Y_N=Y(\mathfrak{gl}_N)$, the Yangian associated to the general Lie algebra $\mathfrak{gl}_N$.
Associated to each composition $\mu$ of $N$, Brundan and Kleshchev established a parabolic presentation for $Y_N$ in \cite{BK1}. Roughly speaking, this new presentation of $Y_N$ corresponds to the Levi decomposition of $\mathfrak{gl}_N$ with respect to $\mu$. In the special case when $\mu=(1,\ldots,1)$, the corresponding presentation is equivalent to Drinfeld's presentation (\cite[Remark 5.12]{BK1}). On the other extreme case when $\mu=(N)$, the corresponding presentation is called the {\em RTT presentation}; see \cite{MNO, Mo}.
The parabolic presentations play a fundamental role in their subsequent works. In \cite{BK2}, they established a concrete realization of finite $W$-algebras associated to {\em any} nilpotent element of type A in terms of Yangians, and a key step is to define a subalgebra of $Y_N$, called the {\em shifted Yangian}. Such a subalgebra can only be defined in terms of the new presentation found in \cite{BK1} except some special cases. The connection between Yangians and finite $W$-algebras was observed earlier in \cite{RS} for some particular nilpotent elements (called {\em rectangular} elements) with a different approach. Moreover, by means of such a realization, one may study the representation theory of finite $W$-algebras by studying the representation theory of Yangians; see \cite{BK3}.
The main goal of this article is to obtain the generalization of \cite{BK1} to the super Yangian $Y_{M|N}=Y(\mathfrak{gl}_{M|N})$, the super Yangian associated to the general linear Lie superalgebra $\mathfrak{gl}_{M|N}$. It is defined by Nazarov \cite{Na} in terms of the RTT presentation as a super analogue of $Y_{N}$.
One of the major differences between $\mathfrak{gl}_N$ and $\mathfrak{gl}_{M|N}$ is that, in the case of $\mathfrak{gl}_N$, we may always choose a canonical Borel subalgebra, since all of the Borel subalgebras are conjugated by the action of the Lie group $GL_N$. This Borel subalgebra gives a root system, and for this given root system we have multiple choices of simple systems. It suffices to choose a canonical simple system since they are all conjugate by the Weyl group action. However, the above argument is no longer true in the case of $\mathfrak{gl}_{M|N}$. Therefore, in the study of $\mathfrak{gl}_{M|N}$ and its representation theory, we may want to specifically mention which simple system we are using, and the notion of {\em 01-sequence} (called $\epsilon\delta$-sequences in \cite[Section 1.3]{CW}) is introduced as a parameterizing set of the conjugacy classes of simple systems of $\mathfrak{gl}_{M|N}$ (also for some other types of Lie superalgebras) under the Weyl group action.
For example, if we identify $\mathfrak{gl}_{M|N}$ with the set of $(M+N)\times (M+N)$ matrices and choose the Cartan subalgebra $\mathfrak{h}$ to be the set of diagonal matrices, then the most common choice for $\mathfrak{s}$ is
$$\mathfrak{s}=\mathfrak{s}^{st}=\stackrel{M}{\overbrace{0\ldots0}}\,\stackrel{N}{\overbrace{1\ldots1}},$$
where a representative simple system of $\mathfrak{gl}_{M|N}$ in the corresponding Weyl group orbit is given by
\[
\Pi^{st}=\{ \delta_{i}-\delta_{i+1}, \epsilon_{j}-\epsilon_{j+1}, \delta_{M}-\epsilon_{1}\,|\, 1\leq i\leq M-1, 1\leq j\leq N-1\}.
\]
Here the notation $\delta_i$ and $\epsilon_j$ denote elements in $\mathfrak{h}^*$ such that
$\delta_{i}(X)$ equals to the $i$-th diagonal entry of $X\in\mathfrak{h}$, and $\epsilon_{j}(X)$ equals to the $(M+j)$-th diagonal entry of $X\in\mathfrak{h}$. With this standard choice, there is only one odd simple root and the behavior of $\mathfrak{gl}_{M|N}$ is ``closest" to the classical $\mathfrak{gl}_N$.
Note that the Weyl group is isomorphic to $S_M\times S_N$, which permutes those $\delta_i$'s and those $\epsilon_j$'s, respectively. Therefore, there is exactly one odd simple root, which is of the form $\delta_i-\epsilon_j$ for some $1\leq i\leq M$, $1\leq j\leq N$, in any other simple system in the Weyl group orbit of $\Pi^{st}$. It implies that the following two simple systems of $\mathfrak{gl}_{3|2}$ are in different Weyl group orbits, since $\Pi_1$ contains only one simple odd root and $\Pi_2$ contains 4 odd simple roots:
\begin{align*}
\Pi_1&=\{\delta_1-\delta_2, \delta_2-\delta_3,\delta_3-\epsilon_1,\epsilon_1-\epsilon_2\} \longleftrightarrow \mathfrak{s}_1=00011,\\
\Pi_2&=\{\delta_1-\epsilon_1, \epsilon_1-\delta_2,\delta_2-\epsilon_2,\epsilon_2-\delta_3\} \longleftrightarrow \mathfrak{s}_2=01010.
\end{align*}
We refer the reader to \cite[Chapter 1.]{CW} for more details and further applications of 01-sequences.
It is noticed in \cite{Pe2} that the notion of $01$-sequence can be perfectly equipped to the RTT presentation of $Y_{M|N}$. It turns out that, up to an isomorphism, the definition of $Y_{M|N}$ is independent of the choices of the 01-sequence $\mathfrak{s}$, and Nazarov's definition corresponds to the case when $\mathfrak{s}=\mathfrak{s}^{st}$ is the canonical one. Since the RTT presentation can be thought as merely a special case of the parabolic presentation by taking the composition $\mu=(M+N)$, the above observation suggests that it should be possible to obtain a corresponding parabolic presentation for {\em any} $\mu$, which triggered this work.
To be precise, the main result of this article (Theorem \ref{Pg}) is that for an {\em arbitrary} fixed 01-sequence $\mathfrak{s}$ of $\mathfrak{gl}_{M|N}$ and an {\em arbitrary} fixed composition $\mu$ of $M+N$, a presentation of $Y_{M|N}$ is obtained.
We quickly explain the idea, which is basically generalizing the argument in \cite{BK1} and adapting some techniques in \cite{Go, Pe1} dealing with the sign factors.
Fix a composition $\mu$ of $M+N$ and fix an arbitrary $01$-sequence $\mathfrak{s}$ of $\mathfrak{gl}_{M|N}$. We first define some distinguished elements in $Y_{M|N}$ associated to $\mu$, denoted by $D$'s, $E$'s and $F$'s, by {\em Gauss decomposition} (or {\em quasideterminants}).
Roughly speaking, the elements $D$'s are those elements in the diagonal blocks of the block matrix decomposition of $Y_{M|N}$ with respect to $\mu$, while the $E$'s and the $F$'s are those elements in the upper and lower diagonal blocks, respectively. Note that these elements depend on the shape $\mu$, where their parities are determined by the given $\mathfrak{s}$. These elements form a generating set for $Y_{M|N}$ (Theorem \ref{gendef}), so the next step is to find enough relations to achieve a presentation.
In the case of \cite{BK1}, if the generators are from two different blocks and the blocks are not ``close", then they commute. Fortunately, this phenomenon remains to be true under our general setting (Lemma~\ref{corcommute}) and it enormously reduces the number of the non-vanishing relations. As a consequence, we only have to focus on the supercommutation relations of the elements that are either in the same block, or their belonging blocks are ``close enough". Let $n$ be the length of $\mu$. When $n=2,3,4$, the situations are less complicated so that we may derive various relations among those generators by direct computation.
Next, we take advantage of the homomorphisms $\psi_L$ and $\zeta_{M|N}$ between super Yangians (see Section 4). These maps carry the relations in the special cases (with $n\leq 4$) to the general case, so that we obtain many relations in $Y_{M|N}$. Finally we prove that we have found enough relations for our presentation.
As a matter of fact, there are already a few results \cite{Go, Pe1} on such a generalization focusing on the canonical case when $\mathfrak{s}=\mathfrak{s}^{st}$.
In \cite{Go}, a presentation of $Y_{M|N}$ when $\mu=(1^{M+N})$ is obtained, which is a generalization of Drinfeld's presentation. In \cite{Pe1}, a similar result when $\mu$ is of the form $\mu=(\lambda,\nu)$, where $\lambda$ is a composition of $M$ and $\nu$ is a composition of $N$, is obtained.
However, the results only explained the case when $\mathfrak{s}=\mathfrak{s}^{st}$. Moreover, the compositions $\mu$ therein are very special so that the elements in one block must have the same parity so the super phenomenon only happens at a few specific places.
Under our setting, $\mu$ and $\mathfrak{s}$ are {\em both arbitrary} so that an even element and an odd element could exist in the same block. As a result, the super phenomenon could happen everywhere. Hence in our current consideration, the signs arising from the $\mathbb{Z}_2$-grading are much more involved than \cite{Go, Pe1} and one needs more elaborate notation and extra care when treating the sign issues. Roughly speaking, we need to correctly insert the necessary sign factors in almost every formula. Certainly, our main theorem covers the above results as special cases.
Finally we mention one undergoing application of our result, which can also be thought as the true motivation of this work. Following the classical $\mathfrak{gl}_N$ case, one may try to generalize the argument in \cite{BK2} so that a realization of finite $W$-superalgebras of type A in terms of the super Yangian $Y_{M|N}$ can be obtained. Such a connection was observed in \cite{BR} for {\em rectangular} nilpotent elements, but this is still open for the general nilpotent case.
In fact, based on \cite{Go, Pe1}, there are already some partial results \cite{BBG, Pe2, Pe3} about the the realization of finite $W$-superalgebras when the nilpotent element is {\em principal} or satisfying certain restrictions. As noticed in \cite{BBG, Pe3}, if we want to generalize the argument in \cite{BK2} to the case of $Y_{M|N}$ in full generality, then a more general presentation of $Y_{M|N}$ is required. One of the reasons is that the {\em shifted super Yangian}, which is a subalgebra of $Y_{M|N}$, can be defined only under some nice assumptions as in \cite{BBG, Pe3}. With our new presentations, the shifted super Yangian can be defined for {\em any} given nilpotent element in $\mathfrak{gl}_{M|N}$ so that it is possible to establish the connection in full generality, and this is currently in progress by the author.
This article is organized in the following fashion. In Section 2, we recall some basic properties of $Y_{M|N}$.
In Section 3, we explicitly define the parabolic generators by means of Gauss decomposition and show that they indeed form a generating set.
In Section 4, we define some homomorphisms between super Yangians so that we may reduce the general case to some less complicated special cases. Section 5 and 6 are devoted to further study about these special cases. Our main theorem is formulated and proved in Section 7.
\section{Prelimilaries}
Let $\mathfrak{s}$ be a $0^M1^N$-sequence (or 01-sequence for short) of $\mathfrak{gl}_{M|N}$, which is a sequence consisting of $M$ $0$'s and $N$ $1$'s, arranged in a row with respect to a certain order. It is well-known \cite[Proposition 1.27]{CW} that there is a bijection between the set of $0^M1^N$-sequence and the Weyl group orbits of simple systems of $\mathfrak{gl}_{M|N}$.
For homogeneous elements $A$ and $B$ in a $ \mathbb Z _2$-graded algebra $L$, the {\em supercommutator of $A$ and $B$} is defined by
\[
\big[ A,B \big] = AB-(-1)^{\pa{A}\pa{B}}BA,
\]
where $\pa{A}$ is the $ \mathbb Z _2$-grading of $A$ in $L$, or called the {\em parity} of $A$. By convention, a homogeneous element $A$ is called $even$ if $\pa{A}=0$, and called $odd$ if $\pa{A}=1$. For each $1\leq i\leq M+N$, let $\pa{i}$ denote the $i$-th digit of the fixed $0^M1^N$-sequence $\mathfrak{s}$.
\begin{definition}
For a given $\mathfrak{s}$, the super Yangian associated to the general linear Lie superalgebra $\mathfrak{gl}_{M|N}$, denoted by $Y_{M|N}$ hereafter, is the associative $\mathbb{Z}_2$-graded algebra (i.e., superalgebra) over $\mathbb{C}$ generated by the {\em RTT generators} \cite{Na}
\begin{equation}\label{RTTgen}
\left\lbrace t_{i,j}^{(r)}\,| \; 1\le i,j \le M+N; r\ge 1\right\rbrace,
\end{equation}
subject to following relations:
\begin{equation}\label{RTT}
\big[ t_{i,j}^{(r)}, t_{h,k}^{(s)} \big] = (-1)^{\pa{i}\,\pa{j} + \pa{i}\,\pa{h} + \pa{j}\,\pa{h}}
\sum_{g=0}^{\mathrm{min}(r,s) -1} \Big( t_{h,j}^{(g)}\, t_{i,k}^{(r+s-1-g)} - t_{h,j}^{(r+s-1-g)}\, t_{i,k}^{(g)} \Big),
\end{equation}
where the parity of $t_{i,j}^{(r)}$ for $r>0$ is defined by $\pa{i}+\pa{j}$ (mod 2), and the bracket is understood as the supercommutator. By convention, we set $t_{i,j}^{(0)}:=\delta_{ij}$.
\end{definition}
Similar to the $\mathfrak{gl}_{M|N}$ case, for $r>0$, the element $t_{i,j}^{(r)}$ is called an $even$ ($odd$, respectively) element if its parity is $0$ ($1$, respectively).
The original definition in \cite{Na} corresponds to the case when $\mathfrak{s}$ is the canonical one; that is, $\mathfrak{s}$ is of the form
\[
\mathfrak{s}=\mathfrak{s}^{st}:=\stackrel{M}{\overbrace{0\ldots0}}\,\stackrel{N}{\overbrace{1\ldots1}}.
\]
As observed in \cite{Pe2}, the definition of $Y_{M|N}$ is independent of the choices of $\mathfrak{s}$, up to an isomorphism, so we often omit it in the notation.
For each $1\leq i,j\leq M+N$, define the formal power series
\[
t_{i,j}(u):= \sum_{r\geq 0} t_{i,j}^{(r)}u^{-r} \in Y_{M|N}[[u^{-1}]].
\]
It is well-known \cite[p.125]{Na} that $Y_{M|N}$ is a Hopf-superalgebra, where the comultiplication
$\Delta:Y_{M|N}\rightarrow Y_{M|N}\otimes Y_{M|N}$ is given by
\begin{equation}\label{Del}
\Delta(t_{i,j}^{(r)})=\sum_{s=0}^r \sum_{k=1}^{M+N} t_{i,k}^{(r-s)}\otimes t_{k,j}^{(s)}.
\end{equation}
Moreover, there exists a surjective homomorphism $$\operatorname{ev}:Y_{M|N}\rightarrow U(\mathfrak{gl}_{M|N})$$ called the {\em evaluation homomorphism}, defined by
\begin{equation}\label{ev}
\operatorname{ev}\big(t_{i,j}(u)\big):= \delta_{ij} + (-1)^{|i|} e_{ij}u^{-1},
\end{equation}
where $e_{ij}\in\mathfrak{gl}_{M|N}$ is the elementary matrix.
For homogeneous elements $x_1,\ldots,x_s$ in a superalgebra $A$, a {\em supermonomial} in $x_1,\ldots,x_s$ means a monomial of the form $x_1^{i_1}\cdots x_s^{i_s}$ for some $i_1,\ldots,i_s\in \mathbb{Z}_{\geq 0}$ and $i_j\leq 1$ if $x_j$ is odd. The following proposition is a PBW theorem for $Y_{M|N}$, where the proof in \cite{Go} works perfectly for any fixed $\mathfrak{s}$.
\begin{proposition}\cite[Theorem 1]{Go}\label{PBWSY}
The set of supermonomials in the following elements of $Y_{M|N}$
\[
\left\lbrace t_{i,j}^{(r)}\, |\, 1\leq i,j\leq M+N, r\geq 1 \right\rbrace
\] taken in some fixed order
forms a linear basis for $Y_{M|N}$.
\end{proposition}
Define the $loop$ $filtration$ on $Y_{M|N}$
\begin{equation}\label{filt2}\notag
L_0 Y_{M|N} \subseteq L_1 Y_{M|N} \subseteq L_2 Y_{M|N} \subseteq \cdots
\end{equation}
by setting $\deg t_{ij}^{(r)}=r-1$ for each $r\geq 1$ and let $L_kY_{M|N}$ be the span of all supermonomials of the form
$$t_{i_1j_1}^{(r_1)}t_{i_2j_2}^{(r_2)}\cdots t_{i_sj_s}^{(r_s)}$$
with total degree not greater than $k$. The associated graded superalgebra is denoted by $\operatorname{gr} Y_{M|N}.$
Let $\mathfrak{gl}_{M|N}[x]$ denote the {\em loop superalgebra} $\mathfrak{gl}_{M|N}\otimes \mathbb{C}[x]$ with the standard basis
$$\lbrace e_{ij}x^r \,|\, 1\leq i,j\leq M+N, r\geq 0\rbrace$$
and let $U(\mathfrak{gl}_{M|N}[x])$ denote its universal enveloping algebra. The next corollary follows from Proposition~\ref{PBWSY}.
\begin{corollary}\cite[Corollary 1]{Go}\label{Yloop}
The graded superalgebra $\operatorname{gr} Y_{M|N}$ is isomorphic to $U(\mathfrak{gl}_{M|N}[x])$ by the map
\begin{center}
$\operatorname{gr} Y_{M|N}\rightarrow U(\mathfrak{gl}_{M|N}[x])$\\
$\operatorname{gr}_{r-1}t_{ij}^{(r)}\mapsto (-1)^{\pa{i}}e_{ij}x^{r-1}.$
\end{center}
\end{corollary}
\section{Parabolic generators}
Let $\mu=(\mu_1,\ldots,\mu_n)$ be a given composition of $M+N$ with length $n$ and fix a $0^M1^N$-sequence $\mathfrak{s}$. We break $\mathfrak{s}$ into $n$ subsequences according to $\mu$; that is,
\[\mathfrak{s}=\mathfrak{s}_1\mathfrak{s}_2\ldots\mathfrak{s}_n,\]
where $\mathfrak{s}_1$ is the subsequence consisting of the first $\mu_1$ digits of $\mathfrak{s}$, $\mathfrak{s}_2$ is the subsequence consisting of the next $\mu_2$ digits of $\mathfrak{s}$, and so on.
For example, if we have $\mathfrak{s}=011100011$ and $\mu=(2,4,3)$, then
\[
\mathfrak{s}=\overbrace{01}^{\mathfrak{s}_1} \, \overbrace{1100}^{\mathfrak{s}_2} \, \overbrace{011}^{\mathfrak{s}_3}.
\]
For each $1\leq a\leq n$, let $p_a$ and $q_a$ denote the number of $0$'s and $1$'s in $\mathfrak{s}_a$, respectively. By definition, each $\mathfrak{s}_a$ is a $0^{p_a}1^{q_a}$-sequence of $\mathfrak{gl}_{p_a|q_a}$. Moreover, for all $1\leq i\leq \mu_a$, define the {\em restricted parity} $\pa{i}_a$ by
\begin{center} $\pa{i}_a$:= the $i$-th digits of $\mathfrak{s}_a$. \end{center}
By definition, for each $1\leq a\leq n$ and $1\leq i\leq \mu_a$, we have
\begin{equation}\label{respa}
\pa{i}_a=\pa{\sum_{j=1}^{a-1}\mu_j+i}.
\end{equation}
Basically, the techniques in \cite{BK1, Pe1} work perfectly with the notion of an arbitrary $\mathfrak{s}$ that we just introduced earlier. In order to make this article self-contained, we spend some time explaining the notation precisely.
Define the $(M+N)\times (M+N)$ matrix with entries in $Y_{M|N}[[u^{-1}]]$ by
\[
T(u):=\Big( t_{i,j}(u) \Big)_{1\leq i,j\leq M+N}
\]
Note that for any fixed $\mathfrak{s}$, the leading minors of the matrix $T(u)$ are invertible and hence it possesses a $Gauss$ $decomposition$ \cite{GR} with respect to $\mu$; that is,
\begin{equation}\label{T=FDE}
T(u) = F(u) D(u) E(u)
\end{equation}
for unique {\em block matrices} $D(u)$, $E(u)$ and $F(u)$ of the form
$$
D(u) = \left(
\begin{array}{cccc}
D_{1}(u) & 0&\cdots&0\\
0 & D_{2}(u) &\cdots&0\\
\vdots&\vdots&\ddots&\vdots\\
0&0 &\cdots&D_{n}(u)
\end{array}
\right),
$$
$$
E(u) =
\left(
\begin{array}{cccc}
I_{\mu_1} & E_{1,2}(u) &\cdots&E_{1,n}(u)\\
0 & I_{\mu_2} &\cdots&E_{2,n}(u)\\
\vdots&\vdots&\ddots&\vdots\\
0&0 &\cdots&I_{\mu_{n}}
\end{array}
\right),\:
$$
$$
F(u) = \left(
\begin{array}{cccc}
I_{\mu_1} & 0 &\cdots&0\\
F_{2,1}(u) & I_{\mu_2} &\cdots&0\\
\vdots&\vdots&\ddots&\vdots\\
F_{n,1}(u)&F_{n,2}(u) &\cdots&I_{\mu_{n}}
\end{array}
\right),
$$
where
\begin{align}
D_a(u) &=\big(D_{a;i,j}(u)\big)_{1 \leq i,j \leq \mu_a},\\
E_{a,b}(u)&=\big(E_{a,b;i,j}(u)\big)_{1 \leq i \leq \mu_a, 1 \leq j \leq \mu_b},\\
F_{b,a}(u)&=\big(F_{b,a;i,j}(u)\big)_{1 \leq i \leq \mu_b, 1 \leq j \leq \mu_a},
\end{align}
are $\mu_a \times \mu_a$,
$\mu_a \times \mu_b$
and $\mu_b \times\mu_a$ matrices, respectively, for all $1\le a\le n$ in (3.3)
and all $1\le a<b\le n$ in (3.4) and (3.5).
Also note that all the submatrices $D_{a}(u)$'s are invertible, and we define the $\mu_a\times\mu_a$ matrix
$D_a^{\prime}(u)=\big(D_{a;i,j}^{\prime}(u)\big)_{1\leq i,j\leq \mu_a}$ by
\begin{equation*}
D_a^{\prime}(u):=\big(D_a(u)\big)^{-1}.
\end{equation*}
The entries of these matrices are certain power series
\begin{eqnarray*}
&D_{a;i,j}(u) =& \sum_{r \geq 0} D_{a;i,j}^{(r)} u^{-r},\\
&D_{a;i,j}^{\prime}(u) =& \sum_{r \geq 0}D^{\prime(r)}_{a;i,j} u^{-r},\\
&E_{a,b;i,j}(u) =& \sum_{r \geq 1} E_{a,b;i,j}^{(r)} u^{-r},\\
&F_{b,a;i,j}(u) =& \sum_{r \geq 1} F_{b,a;i,j}^{(r)} u^{-r}.
\end{eqnarray*}
In addition, for $1\leq a\leq n-1$, we set
\begin{eqnarray*}
&E_{a;i,j}(u) :=& E_{a,a+1;i,j}(u)=\sum_{r \geq 1} E_{a;i,j}^{(r)} u^{-r},\\
&F_{a;i,j}(u) :=& F_{a+1,a;i,j}(u)=\sum_{r \geq 1} F_{a;i,j}^{(r)} u^{-r}.
\end{eqnarray*}
We will show that the coefficients of these series forms a generating set for $Y_{M|N}$.
As a matter of fact, one may describe all these series in terms of the RTT generators explicitly by \textit{quasideterminants} \cite{GR}. Here we follow the notation in \cite[(4.3)]{BK1}. Suppose that $A, B, C$ and $D$ are $a \times a$, $a \times b$, $b \times a$ and $b \times b$ matrices respectively with entries in some ring.
Assuming that the matrix $A$ is invertible, we define
\begin{equation*}
\left|
\begin{array}{cc}
A&B\\
C&
\hbox{\begin{tabular}{|c|}\hline$D$\\\hline\end{tabular}}
\end{array}
\right| := D - C A^{-1} B.
\end{equation*}
Write the matrix $T(u)$ in block form according to $\mu$ as
$$
T(u) = \left(
\begin{array}{lll}
{^\mu}T_{1,1}(u)&\cdots&{^\mu}T_{1,n}(u)\\
\vdots&\ddots&\cdots\\
{^\mu}T_{n,1}(u)&\cdots&{^\mu}T_{n,n}(u)\\
\end{array}
\right),
$$ where each ${^\mu}T_{a,b}(u)$ is a $\mu_a \times \mu_b$ matrix.
\begin{proposition}\cite{GR}\label{quasi} We have
\begin{align}\label{quasid}
&D_a(u) =
\left|
\begin{array}{cccc}
{^\mu}T_{1,1}(u) & \cdots & {^\mu}T_{1,a-1}(u)&{^\mu}T_{1,a}(u)\\
\vdots & \ddots &\vdots&\vdots\\
{^\mu}T_{a-1,1}(u)&\cdots&{^\mu}T_{a-1,a-1}(u)&{^\mu}T_{a-1,a}(u)\\
{^\mu}T_{a,1}(u) & \cdots & {^\mu}T_{a,a-1}(u)&
\hbox{\begin{tabular}{|c|}\hline${^\mu}T_{a,a}(u)$\\\hline\end{tabular}}
\end{array}
\right|,\\[4mm]
&E_{a,b}(u) =\label{quasie}
D^{\prime}_a(u)
\left|\begin{array}{cccc}
{^\mu}T_{1,1}(u) & \cdots &{^\mu}T_{1,a-1}(u)& {^\mu}T_{1,b}(u)\\
\vdots & \ddots &\vdots&\vdots\\
{^\mu}T_{a-1,1}(u) & \cdots & {^\mu}T_{a-1,a-1}(u)&{^\mu}T_{a-1,b}(u)\\
{^\mu}T_{a,1}(u) & \cdots & {^\mu}T_{a,a-1}(u)&
\hbox{\begin{tabular}{|c|}\hline${^\mu}T_{a,b}(u)$\\\hline\end{tabular}}
\end{array}
\right|,\\[4mm]
&F_{b,a}(u) =\label{quasif}
\left|
\begin{array}{cccc}
{^\mu}T_{1,1}(u) & \cdots &{^\mu}T_{1,a-1}(u)& {^\mu}T_{1,a}(u)\\
\vdots & \ddots &\vdots&\vdots\\
{^\mu}T_{a-1,1}(u) & \cdots & {^\mu}T_{a-1,a-1}(u)&{^\mu}T_{a-1,a}(u)\\
{^\mu}T_{b,1}(u) & \cdots & {^\mu}T_{b,a-1}(u)&
\hbox{\begin{tabular}{|c|}\hline${^\mu}T_{b,a}(u)$\\\hline\end{tabular}}
\end{array}
\right|D^{\prime}_a(u),
\end{align}
for all $1\leq a\leq n$ in (\ref{quasid}) and $1\leq a<b\leq n$ in (\ref{quasie}), (\ref{quasif}).
\end{proposition}
Let $T_{a,b;i,j}(u)$ be the $(i,j)$-th entry of the $\mu_a\times\mu_b$ matrix $^\mu T_{a,b}(u)$ and let $T_{a,b;i,j}^{(r)}$ denote the coefficient of $u^{-r}$ in $T_{a,b;i,j}(u)$. As a consequence of Proposition \ref{quasi}, we have
\begin{equation}\label{Jacobi1}
E_{b-1;i,j}^{(1)}=T_{b-1,b;i,j}^{(1)}, \qquad F_{b-1;j,i}^{(1)}=T_{b,b-1;j,i}^{(1)},
\end{equation}
for all $2\leq b\leq n$, all $1\leq i\leq \mu_{b-1}, 1\leq j\leq \mu_{b}$.
In particular,
\begin{equation}\label{Dtident}
D_{1;i,j}^{(r)}=T_{1,1;i,j}^{(r)}=t_{i,j}^{(r)},\quad \text{for all}\quad 1\leq i,j\leq \mu_1, r\geq 0.
\end{equation}
Note that the matrix $T(u)$ is always invertible and we define the entries of its inverse by
\[
\big(T(u)\big)^{-1}:=\big(t_{ij}^{\prime}(u)\big)_{i,j=1}^{M+N}.
\]
Taking inverse to the matrix equation (\ref{T=FDE}), we have
\begin{equation}\label{Tp=FDE}
t_{i,j}^{\prime}(u) = \big( F(u)^{-1} D(u)^{-1} E(u)^{-1} \big)_{i,j}
\end{equation}
Applying the technique in \cite[Section 1]{MNO}, we may rewrite equation (\ref{RTT}) into the following series form
$$
(u-v)[t_{ij}(u),t_{hk}(v)]
=(-1)^{\pa{i}\,\pa{j}+\pa{i}\,\pa{h}+\pa{j}\,\pa{h}}
\Big(t_{hj}(u)t_{ik}(v)-t_{hj}(v)t_{ik}(u)\Big).
$$
The next lemma, which will be used frequently later, can be deduced from the above equation by a similar calculation as in \cite[Section 2]{Pe1}.
\begin{lemma}
\begin{multline}\label{usefull}
(u-v)[t_{ij}(u),t'_{hk}(v)]
=(-1)^{\pa{i}\,\pa{j}+\pa{i}\,\pa{h}+\pa{j}\,\pa{h}}\times\\
\Big(\delta_{h,j}
\sum_{g=1}^{M+N}
t_{ig}(u)t'_{gk}(v)-\delta_{i,k}\sum_{g=1}^{M+N}t'_{hg}(v)t_{gj}(u)\Big),
\end{multline}
for all $1\leq i,j,h,k\leq M+N$.
\end{lemma}
\begin{lemma}
For each pair $a$, $b$ such that $1<a+1<b\leq n-1$ and $1\leq i\leq\mu_a$, $1 \leq j \leq \mu_b$, we have
\begin{equation}
E_{a,b;i,j}^{(r)} = (-1)^{\pa{k}_{b-1}}[E_{a,b-1;i,k}^{(r)}, E_{b-1;k,j}^{(1)}],
\,\,\,\,\,
F_{b,a;j,i}^{(r)} = (-1)^{\pa{k}_{b-1}}[F_{b-1;j,k}^{(1)}, F_{b-1,a;k,i}^{(r)}],\label{ter}
\end{equation}
for any fixed $1 \leq k \leq \mu_{b-1}$.
\end{lemma}
\begin{proof}
This can be proved by induction on $b-a>1$. We perform the initial step when $a=1$ and $b=3$ for the $E$'s here, while the general case and the statement for the $F$'s can be established in a similar way.
By (\ref{quasie}) and (\ref{Jacobi1}), we have
$$[E_{1,2;i,h}^{(r)}, E_{2,3;k,j}^{(1)}]=[\sum_{p=1}^{\mu_1}\sum_{s=0}^{r}D_{1;i,p}^{\prime(s)} T_{1,2;p,h}^{(r-s)} , T_{2,3;k,j}^{(1)}].$$
Note that we may express $D_{1;i,p}^{\prime(r)}$ in terms of $t_{\alpha,\beta}^{\prime(r)}$, and the subscriptions $1\leq \alpha,\beta\leq \mu_1$ will never overlap with the subscriptions of $T_{2,3;k,j}^{(1)}=t_{\mu_1+k,\mu_1+\mu_2+j}^{(1)}$, where $1\leq k\leq \mu_2, 1\leq j\leq \mu_3$. Thus they supercommute by (\ref{usefull}), and the right-hand side of the equation equals to
$$\sum_{p=1}^{\mu_1}\sum_{s=0}^{r}D_{1;i,p}^{\prime(s)} [ T_{1,2;p,h}^{(r-s)}, T_{2,3;k,j}^{(1)}].$$
The bracket can be computed by (\ref{RTT}):
$$[ T_{1,2;p,h}^{(r-s)}, T_{2,3;k,j}^{(1)}]=[t_{p,\mu_1+h}^{(r-s)}, t_{\mu_1+k,\mu_1+\mu_2+j}^{(1)}]=(-1)^{\pa{k}_2} \delta_{h,k} t_{p,\mu_1+\mu_2+j}^{(r-s)}.$$
Thus, for any $1\leq k\leq \mu_2$, we have
$$(-1)^{\pa{k}_2}[E_{1,2;i,k}^{(r)}, E_{2,3;k,j}^{(1)}]=\sum_{p=1}^{\mu_1}\sum_{s=0}^{r}D_{1;i,p}^{\prime(s)}t_{p,\mu_1+\mu_2+j}^{(r-s)}=\sum_{p=1}^{\mu_1}\sum_{s=0}^{r}D_{1;i,p}^{\prime(s)}T_{1,3;p,j}^{(r-s)}.$$
By (\ref{quasie}), the right-hand side of the above equation is exactly $E_{1,3;i,j}^{(r)}$. The general case is similar, except that the expression of $E_{a,b-1;i,h}^{(r)}$ by (\ref{quasie}) is more complicated.
\end{proof}
\begin{theorem}\label{gendef}
The superalgebra $Y_{M|N}$ is generated by the following elements
\begin{align*}
&\big\lbrace D_{a;i,j}^{(r)}, D_{a;i,j}^{\prime(r)} \,|\, {1\leq a\leq n,\; 1\leq i,j\leq \mu_a,\; r\geq 0}\big\rbrace,\\
&\big\lbrace E_{a;i,j}^{(r)} \,|\, {1\leq a< n,\; 1\leq i\leq \mu_a, 1\leq j\leq\mu_{a+1},\; r\geq 1}\big\rbrace,\\
&\big\lbrace F_{a;i,j}^{(r)} \,|\, {1\leq a< n,\; 1\leq i\leq\mu_{a+1}, 1\leq j\leq \mu_a,\; r\geq 1}\big\rbrace.
\end{align*}
\end{theorem}
\begin{proof}
Multiplying the matrix equation (\ref{T=FDE}), we see that each $t_{ij}^{(r)}$ can be expressed as a sum of supermonomials in $D_{a;i,j}^{(r)}$, $E_{a,b;i,j}^{(r)}$ and $F_{b,a;i,j}^{(r)}$, in a certain order that all the $F$'s appear before the $D$'s and all the $D$'s appear before the $E$'s. By (\ref{ter}), it is enough to use $D_{a;i,j}^{(r)}$, $E_{a;i,j}^{(r)}$ and $ F_{a;i,j}^{(r)}$ only rather than use all of the $E$'s and the $F$'s.
\end{proof}
\begin{remark}
Actually we don't need those $D_{a;i,j}^{\prime(r)}$ to obtain a generating set.
They only appear in (\ref{p702}) and (\ref{p706}). One can in fact remove (\ref{p702}) from the defining relations and rewrite (\ref{p706}) into a new expression which is free from those $D^{\prime(r)}_{a;i,j}$.
The reason to include them is to make the presentation look concise.
\end{remark}
The generators of $Y_{M|N}$ in Theorem \ref{gendef} above are called {\em parabolic generators}. Note that these generators depend on the shape $\mu$ and their parities depend on the fixed sequence $\mathfrak{s}$ (see (\ref{pad})-(\ref{paf}) later). We will use the notation $Y_\mu$ or $Y_{M|N}(\mathfrak{s})$ or $Y_\mu(\mathfrak{s})$ to emphasize the choice of $\mu$ or $\mathfrak{s}$ or both when necessary. The goal of this article is to write down explicitly a set of defining relations of $Y_\mu(\mathfrak{s})$ with respect to the parabolic generators for {\em any} fixed $\mu$ and {\em any} fixed $\mathfrak{s}$.
\section{Homomorphisms between super Yangians}
To explicitly write down the relations among the parabolic generators in Theorem \ref{gendef}, we start with the special cases when $n$ are either 2 or 3, that are relatively less complicated. The other relations in full generality can be deduced from these special ones by applying certain nice injective homomorphisms that we are about to introduce.
We start with some notation. Let $\mathfrak{s}$ be a fixed $0^M1^N$-sequence. We define
\begin{itemize}
\item $\check{\mathfrak{s}}$:= the $0^N1^M$-sequence obtained by interchanging the 0's and 1's of $\mathfrak{s}$.
\item $\mathfrak{s}^r$:= the reverse of $\mathfrak{s}$.
\item $\mathfrak{s}^\dagger$:= $(\check{\mathfrak{s}})^r$, the reverse of $\check{\mathfrak{s}}$.
\end{itemize}
For instance, if $\mathfrak{s}=0011010$, then $\check{\mathfrak{s}}=1100101$, $\mathfrak{s}^r=0101100$, and $\mathfrak{s}^\dagger=1010011$.
Moreover, if $\mathfrak{s}_1$ and $\mathfrak{s}_2$ are two 01-sequences, then $\mathfrak{s}_1\mathfrak{s}_2$ simply means the concatenation of $\mathfrak{s}_1$ and $\mathfrak{s}_2$.
The following proposition is a generalization of \cite[Proposition 4.1]{Pe1} to an arbitrary 01-sequence $\mathfrak{s}$, where the proof is similar so we omit; see also \cite[Section~4]{Go}
\begin{proposition}
\begin{enumerate}
\item [1.]The map $\rho_{M|N}:Y_{M|N}(\mathfrak{s})\rightarrow Y_{N|M}(\mathfrak{s}^\dagger)$ defined by
\[
\rho_{M|N}\big(t_{ij}(u)\big)=t_{M+N+1-i,M+N+1-j}(-u)
\]
is an isomorphism.
\item [2.]The map $\omega_{M|N}:Y_{M|N}(\mathfrak{s})\rightarrow Y_{M|N}(\mathfrak{s})$ defined by
\[
\omega_{M|N}\big(T(u)\big)=\big(T(-u)\big)^{-1}
\]
is an automorphism.
\item [3.]The map $\zeta_{M|N}:Y_{M|N}(\mathfrak{s})\rightarrow Y_{N|M}(\mathfrak{s}^\dagger)$ defined by
\[
\zeta_{M|N}=\rho_{M|N}\circ\omega_{M|N}
\]
is an isomorphism.
\item [4.] Let $p,q\in\mathbb{Z}_{\ge0}$ be given and let $\mathfrak{s}_1$ be an arbitrary $0^p1^q$-sequence.
Let
$$\varphi_{p|q}:Y_{M|N}(\mathfrak{s})\rightarrow Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$$
be the injective homomorphism sending each $t_{i,j}^{(r)}$ in $Y_{M|N}(\mathfrak{s})$ to $t_{p+q+i,p+q+j}^{(r)}$ in $Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$. Then
the map $\psi_{p|q}:Y_{M|N}(\mathfrak{s})\rightarrow Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$ defined by
\[
\psi_{p|q}=\omega_{p+M|q+N}\circ\varphi_{p|q}\circ\omega_{M|N},
\]
is an injective homomorphism.
\end{enumerate}
\end{proposition}
In fact, only the maps $\zeta_{M|N}$ and $\psi_{p|q}$ will be used later so we write down their images explicitly.
\begin{lemma}
Let $1\leq i,j \leq M+N$.
\begin{enumerate}
\item [1.]For any $p,q\in\mathbb{Z}_{\geq 0}$, we have \begin{equation}\label{psit}
\psi_{p|q}\big(t_{ij}(u)\big)=
\left| \begin{array}{cccc} t_{1,1}(u) &\cdots &t_{1,p+q}(u) &t_{1, p+q+j}(u)\\
\vdots &\ddots &\vdots &\vdots \\
t_{p+q,1}(u) &\cdots &t_{p+q,p+q}(u) &t_{p+q, p+q+j}(u)\\
t_{p+q+i, 1}(u) &\cdots &t_{p+q+i,p+q}(u) &\boxed{t_{p+q+i, p+q+j}(u)}
\end{array} \right|.
\end{equation}
\item [2.] \begin{equation}\label{zetat}
\zeta_{M|N}\big(t_{ij}(u)\big)=t_{M+N+1-i,M+N+1-j}^{\prime}(u).
\end{equation}
\end{enumerate}
\end{lemma}
\begin{proof}
The first one is exactly the same with \cite[Lemma 4.2]{BK1}, while the second one is immediate from the definition.\end{proof}
Set $L=p+q$ for convenience. Note that (\ref{psit}) depends only on $L$ so we may simply write $\psi_{p|q}=\psi_L$ when appropriate. As a consequence of Proposition \ref{quasi}, we have
\begin{eqnarray}\label{psid}
D_{a;i,j}(u)&=&\psi_{\mu_1+\mu_2+\ldots +\mu_{a-1}}\big(D_{1;i,j}(u)\big),\\ \label{psie}
E_{a;i,j}(u)&=&\psi_{\mu_1+\mu_2+\ldots +\mu_{a-1}}\big(E_{1;i,j}(u)\big),\\ \label{psif}
F_{a;i,j}(u)&=&\psi_{\mu_1+\mu_2+\ldots +\mu_{a-1}}\big(F_{1;i,j}(u)\big).
\end{eqnarray}
In addition, the map $\psi_{L}$ sends $t_{i,j}^{\prime}(u)\in Y_{M|N}(\mathfrak{s})$ to $t_{L+i,L+j}^{\prime}(u)\in Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$, which implies that the image of $\psi_{L}\big(Y_{M|N}(\mathfrak{s})\big)$ in $Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$ is generated by the following elements
\[
\{t_{L+i,L+j}^{\prime(r)}\in Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})\,|\,1\leq i,j\leq M+N, r\geq 0\}.
\]
If we pick any element $t_{i,j}^{(r)}$ in the northwestern $L\times L$ corner of $T(u)$, an $(L+M+N)\times(L+M+N)$ matrix with entries in $Y_{p+M|q+N}[[u^{-1}]]$, then the indices will never overlap with those of $\psi_{L}\big(Y_{M|N}\big)$, that are in the southeastern $(M+N)\times(M+N)$ corner of the same $T(u)$. As a result of equation (\ref{usefull}), they supercommute. Clearly, the elements in the northwestern $L\times L$ corner of $T(u)$ (of $Y_{p+M|q+N}$) generate a subalgebra which is isomorphic to $Y_{p|q}(\mathfrak{s}_1)$ by (\ref{RTT}), so we have obtained the following lemma. Roughly speaking, the map $\psi_{L}$ embeds the matrix $T(u)$ into the southeastern corner of a larger matrix $T(u)$.
\begin{lemma}\label{corcommute}
In $Y_{p+M|q+N}(\mathfrak{s}_1\mathfrak{s})$, the subalgebras $Y_{p|q}(\mathfrak{s}_1)$ and $\psi_{L}\big(Y_{M|N}(\mathfrak{s})\big)$ supercommute with each other.
\end{lemma}
Moreover, by equations (\ref{psit}), (\ref{psid}), (\ref{psie}) and (\ref{psif}), the parities of the parabolic generators can be easily obtained as follows:
\begin{eqnarray}
\label{pad}\text{ parity of } D_{a;i,j}^{(r)}&=&\pa{i}_a+\pa{j}_a \,\,\text{(mod 2)},\\
\label{pae}\text{ parity of } E_{b;h,k}^{(s)}&=&\pa{h}_{b}+\pa{k}_{b+1} \,\,\text{(mod 2)},\\
\label{paf}\text{ parity of } F_{b;f,g}^{(s)}&=&\pa{f}_{b+1}+\pa{g}_{b} \,\,\text{(mod 2)},
\end{eqnarray}
for all $r,s\geq 1$, $1\leq a\leq n$, $1\leq b\leq n-1$, $1\leq i,j\leq \mu_a$, $1\leq h,g\leq \mu_b$, $1\leq k,f\leq \mu_{b+1}$.
Next we analyze $\zeta_{M|N}$. Associate to $\mu$, we may define the elements \{$D_{a;i,j}^{(r)};D^{\prime(r)}_{a;i,j}$\}, \{$E_{a;i,j}^{(r)}$\}, \{$F_{a;i,j}^{(r)}$\} in $Y_{M|N}=Y_\mu(\mathfrak{s})$ by Gauss decomposition. Consider
\[\overleftarrow{\mu}:=(\mu_n,\mu_{n-1},\ldots,\mu_{1}),\]
the reverse of $\mu$.
We may similarly define elements \{$\overleftarrow{D}_{a;i,j}^{(r)};\overleftarrow{D}^{\prime(r)}_{a;i,j}$\}, \{$\overleftarrow{E}_{a;i,j}^{(r)}$\}, \{$\overleftarrow{F}_{a;i,j}^{(r)}$\} in $Y_{N|M}=Y_{\overleftarrow{\mu}}(\mathfrak{s}^\dagger)$ with respect to $\overleftarrow{\mu}$ in the same way. Their relations are described in the following proposition, which is a generalization of \cite[Proposition~1]{Go} and \cite[Proposition~4.4]{Pe1}, where the proof is very similar. Roughly speaking, the map $\zeta_{M|N}$ turns the matrix $T(u)$ up side down, so we take $\mathfrak{s}^\dagger$ in the image space in order to keep the parities unchanged.
\begin{proposition}\label{zetadef} For all $1\leq i,j,h\leq \mu_a$, $1\leq k\leq \mu_{a+1}$, we have
\begin{eqnarray}
\zeta_{M|N}\big(D_{a;i,j}(u)\big)&=&\overleftarrow{D}^{\prime}_{n+1-a;\mu_{a}+1-i,\mu_{a}+1-j}(u), \qquad\forall 1\leq a\leq n, \label{zd} \\
\zeta_{M|N}\big(E_{a;h,k}(u)\big)&=&-\overleftarrow{F}_{n-a;\mu_{a}+1-h,\mu_{a+1}+1-k}(u), \quad\forall 1\leq a\leq n-1, \label{sze}\\
\zeta_{M|N}\big(F_{a;k,h}(u)\big)&=&-\overleftarrow{E}_{n-a;\mu_{a+1}+1-k,\mu_{a}+1-h}(u), \quad\forall 1\leq a\leq n-1\label{szf}.
\end{eqnarray}
\end{proposition}
The following proposition follows directly from (\ref{RTT}), (\ref{psid}), and Lemma \ref{corcommute}. In consequence, the relations among the $D$'s are obtained. One should notice that these $D$'s could be even or odd according to (\ref{pad}). This is different with the cases in \cite{Go,Pe1}, in which they are always even.
\begin{proposition}\label{dd0}
The relations among the elements
$\{D_{a;i,j}^{(r)},D_{a;i,j}^{\prime (r)}\}$ for all $r\geq 0$, ${1\leq i,j\leq \mu_a}$, $1\leq a\leq n$
are given by
\[
D_{a;i,j}^{(0)}=\delta_{ij},\]
\[
{\displaystyle \sum_{t=0}^{r}D_{a;i,p}^{(t)}D_{a;p,j}^{\prime (r-t)}=\delta_{r0}\delta_{ij} },
\]
\begin{multline*}
[D_{a;i,j}^{(r)},D_{b;h,k}^{(s)}]=
\delta_{ab}(-1)^{\pa{i}_a\pa{j}_a+\pa{i}_a\pa{h}_a+\pa{j}_a\pa{h}_a}\times\\
\sum_{t=0}^{min(r,s)-1}\big(D_{a;h,j}^{(t)}D_{a;i,k}^{(r+s-1-t)}
-D_{a;h,j}^{(r+s-1-t)}D_{a;i,k}^{(t)}\big).
\end{multline*}
\end{proposition}
It can be observed from the relations that the elements $\{D_{a;i,j}^{(r)},D_{a;i,j}^{\prime (r)}\}$ generate a subalgebra of $Y_{\mu}$, called the $Levi$ $subalgebra$ of $Y_{\mu}$, and denote it by $Y^0_{\mu}(\mathfrak{s})$.
By Lemma \ref{corcommute}, we have
\begin{eqnarray*}
Y^0_{\mu}(\mathfrak{s})&=&Y_{\mu_1}(\mathfrak{s}_1)\psi_{\mu_1}(Y_{\mu_2}(\mathfrak{s}_2))\psi_{\mu_1+\mu_2}(Y_{\mu_3}(\mathfrak{s}_3))\cdots\psi_{\mu_1+\cdots+\mu_{n-1}}(Y_{\mu_n}(\mathfrak{s}_n))\\
&\cong &Y_{\mu_1}(\mathfrak{s}_1)\otimes Y_{\mu_2}(\mathfrak{s}_2)\otimes\cdots \otimes Y_{\mu_n}(\mathfrak{s}_n),
\end{eqnarray*}
where $\mu=(\mu_1,\mu_2,\ldots ,\mu_n)$ and $\mathfrak{s}=\mathfrak{s}_1\mathfrak{s}_2\cdots\mathfrak{s}_n$.
Note that in the special case when all $\mu_i=1$, the subalgebra $Y^0_{(1,\ldots,1)}(\mathfrak{s})$ is purely even and commutative. One may think $Y^0_{(1,\ldots,1)}(\mathfrak{s})$ in $Y_{M|N}$ as an analogue of the Cartan subalgebra consisting of all diagonal matrices in $\mathfrak{gl}_{M|N}$.
\section{Special Case: $n=2$}
In this section, we focus on the very first non-trivial case under our consideration; that is, $\mu=(\mu_1,\mu_2)$ with a fixed 01-sequence $\mathfrak{s}=\mathfrak{s}_1\mathfrak{s}_2$.
We list our parabolic generators as follow:
\begin{align*}
&\big\lbrace D_{a;i,j}^{(r)}, D_{a;i,j}^{\prime(r)} \,|\, {a=1,2;\; 1\leq i,j\leq \mu_a;\; r\geq 0}\big\rbrace,\\
&\big\lbrace E_{1;i,j}^{(r)} \,|\, 1\leq i\leq \mu_1, 1\leq j\leq\mu_{2};\; r\geq 1\big\rbrace,\\
&\big\lbrace F_{1;i,j}^{(r)} \,|\, 1\leq i\leq\mu_{2}, 1\leq j\leq \mu_1;\; r\geq 1\big\rbrace.
\end{align*}
The following proposition gives explicitly the relations among the generators other than those relations already obtained in Proposition \ref{dd0}.
{\allowdisplaybreaks
\begin{proposition}\label{n=2}
Let $\mu=(\mu_1,\mu_2)$ be a composition of $M+N$. The following identities hold in
$Y_{\mu}((u^{-1},v^{-1}))$:
\begin{align}
(u-v)[D_{1;i,j}(u), E_{1;h,k}(v)]\label{p511}
&=(-1)^{\pa{h}_1\pa{j}_1}\delta_{hj}\sum_{p=1}^{\mu_1}D_{1;i,p}(u)\big(E_{1;p,k}(v)-E_{1;p,k}(u)\big),\\[2mm]
(u-v)[E_{1;i,j}(u), D^{\prime}_{2;h,k}(v)]
&=(-1)^{\pa{h}_2\pa{j}_2}\delta_{hj} \sum_{q=1}^{\mu_2}\big(E_{1;i,q}(u)-E_{1;i,q}(v)\big)D^{\prime}_{2;q,k}(v),\label{p512}\\[2mm]
(u-v)[D_{2;i,j}(u), E_{1;h,k}(v)]
&\notag=(-1)^{\pa{h}_1\pa{k}_2+\pa{h}_1\pa{j}_2+\pa{j}_2\pa{k}_2}\times\\
&\qquad\qquad\qquad D_{2;i,k}(u)\big(E_{1;h,j}(u)-E_{1;h,j}(v)\big),\label{p513}\\[2mm]
(u-v)[D_{1;i,j}(u), F_{1;h,k}(v)]
&\notag=(-1)^{\pa{i}_1\pa{j}_1+\pa{h}_2\pa{i}_1+\pa{h}_2\pa{j}_1}\delta_{ik}\times\\
&\qquad\qquad\qquad \sum_{p=1}^{\mu_1}\big(F_{1;h,p}(u)-F_{1;h,p}(v)\big)D_{1;p,j}(u),\label{p514}\\[2mm]
(u-v)[F_{1;i,j}(u), D^{\prime}_{2;h,k}(v)]
&\notag=(-1)^{\pa{h}_2\pa{i}_2+\pa{h}_2\pa{j}_1+\pa{j}_1\pa{k}_2}\delta_{ik}\times\\
&\qquad\qquad\qquad \sum_{q=1}^{\mu_2}D^{\prime}_{2;h,q}(v)\big(F_{1;q,j}(v)-F_{1;q,j}(u)\big),\label{p515}\\[2mm]
(u-v)[D_{2;i,j}(u), F_{1;h,k}(v)]
&\notag=(-1)^{\pa{h}_2\pa{k}_1+\pa{h}_2\pa{j}_2+\pa{j}_2\pa{k}_1}\times\\
&\qquad\qquad\qquad \big(F_{1;i,k}(v)-F_{1;i,k}(u)\big)D_{2;h,j}(u),\label{p516}\\[2mm]
(u-v)[E_{1;i,j}(u), F_{1;h,k}(v)]
&\notag=(-1)^{\pa{h}_2\pa{i}_1+\pa{i}_1\pa{j}_2+\pa{h}_2\pa{j}_2}D_{2;h,j}(u) D^{\prime}_{1;i,k}(u)\\
&\qquad -(-1)^{\pa{h}_2\pa{k}_1+\pa{j}_2\pa{k}_1+\pa{h}_2\pa{j}_2}
D^{\prime}_{1;i,k}(v) D_{2;h,j}(v),\label{p517}\\[2mm]
(u-v)[E_{1;i,j}(u), E_{1;h,k}(v)]
&\notag=(-1)^{\pa{h}_1\pa{j}_2+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2}\times\\
&\qquad \big(E_{1;i,k}(u)-E_{1;i,k}(v)\big)\big(E_{1;h,j}(u)-E_{1;h,j}(v)\big),\label{p518}\\[2mm]
(u-v)[F_{1;i,j}(u), F_{1;h,k}(v)]
&\notag=(-1)^{\pa{i}_2\pa{j}_1+\pa{h}_2\pa{i}_2+\pa{h}_2\pa{j}_1}\times\\
&\qquad \big(F_{1;h,j}(v)-F_{1;h,j}(u)\big)\big(F_{1;i,k}(u)-F_{1;i,k}(v)\big).\label{p519}
\end{align}
The identities hold for all $1\leq i,j\leq \mu_1$ if $D_{1;i.j}(u)$ appears on the left-hand side of the equation,
for all $1\leq h,k\leq \mu_2$ if $D_{2;h,k}^\prime(u)$ appears on the left-hand side of the equation,
for all $1\leq i^\prime \leq \mu_1$, $1\leq j^\prime \leq \mu_2$ if $E_{1;i^\prime,j^\prime}(u)$ appears on the left-hand side of the equation, and
for all $1\leq h^\prime\leq \mu_2$, $1\leq k^\prime\leq \mu_1$ if $F_{1;h^\prime,k^\prime}(u)$ appears on the left-hand side of the equation.
\end{proposition}
}
\begin{proof}
Our approach is similar to those in \cite[Section 6]{BK1} and \cite[Section 5]{Pe1}.
We compute the matrix products (\ref{T=FDE}) and (\ref{Tp=FDE})
with respect to the composition $\mu=(\mu_1,\mu_2)$ and get the following identities.
\begin{eqnarray}
\label{511}t_{i,j}(u)&=&D_{1;i,j}(u), \qquad\qquad\qquad\quad\;\forall 1\le i,j\le \mu_1,\\
\label{512}t_{i,\mu_1+j}(u)&=&D_{1;i,p}E_{1;p,j}(u),\qquad\qquad\quad\;\;\forall 1\le i\le \mu_1, 1\le j\le \mu_2,\\
\label{513}t_{\mu_1+i,j}(u)&=&F_{1;i,p}(u)D_{1;p,j}(u),\qquad\quad\quad\;\forall 1\le i\le \mu_2, 1\le j\le \mu_1,\\
\label{514}t_{\mu_1+i,\mu_1+j}(u)&=&F_{1;i,p}(u)D_{1;p,q}(u)E_{1;q,j}(u)+D_{2;i,j}(u),\,\,\,\forall 1\le i,j\le \mu_2,\\
\label{515}t^{\prime}_{i,j}(u)&=&D^{\prime}_{1;i,j}(u)+E_{1;i,p'}(u)D^{\prime}_{2;p',q'}(u)F_{1;q',j}(u),\forall 1\le i,j\le \mu_1,\\
\label{516}t^{\prime}_{i,\mu_1+j}(u)&=&-E_{1;i,p'}(u)D^{\prime}_{2;p',j}(u),\qquad\quad\forall 1\le i\le \mu_1, 1\le j\le \mu_2,\\
\label{517}t^{\prime}_{\mu_1+i,j}(u)&=&-D^{\prime}_{2;i,p'}(u)F_{1;p',j}(u),\qquad\quad\,\forall 1\le i\le \mu_2, 1\le j\le \mu_1,\\
\label{518}t^{\prime}_{\mu_1+i,\mu_1+j}(u)&=&D^{\prime}_{2;i,j}(u),\qquad\qquad\qquad\quad\;\,\forall 1\le i,j\le \mu_2,
\end{eqnarray}
where the indices $p,q$ (respectively, $p',q'$) are summed over $1,\ldots,\mu_1$ (respectively, $1,\ldots,\mu_2$).
(\ref{p511})--(\ref{p513}) can be proved similar to \cite[Lemma~6.3]{BK1} and \cite[Proposition~5.1]{Pe1}, except for some issues about the sign factors that have to be carefully treated. (\ref{p514})--(\ref{p516}) and (\ref{p519}) follow from applying the map $\zeta_{M|N}$ to (\ref{p511})--(\ref{p513}) and (\ref{p518}) with suitable choices of indices.
We prove (\ref{p517}) and (\ref{p518}) in detail here as illustrating examples about some new phenomenons and how we deal with them.
To show (\ref{p517}), we need other identities. Computing the brackets in (\ref{p512}) and (\ref{p514}) by definition, we have
\begin{multline}\label{2ef1}
(u-v)E_{1;\alpha ,j}(u)D^{\prime}_{2;h,\beta}(v)
=(-1)^{\pa{h}_2\pa{j}_2}\delta_{hj}\big(E_{1;\alpha ,q}(u)-E_{1;\alpha ,q}(v)\big)D^{\prime}_{2;q,\beta}(v)\\
+(-1)^{(\pa{\alpha}_1+\pa{j}_2)(\pa{h}_2+\pa{\beta}_2)}(u-v)D^{\prime}_{2;h,\beta}(v)E_{1;\alpha ,j}(u),
\end{multline}
\begin{multline}\label{2ef2}
(u-v)F_{1;\beta ,k}(v)D_{1;i,\alpha}(u)=
(-1)^{(\pa{i}_1+\pa{\alpha}_1)(\pa{\beta}_1+\pa{k}_1)}(u-v)D_{1;i,\alpha}(u)F_{1;\beta ,k}(v)\\
-(-1)^{\pa{i}_1\pa{k}_1}\delta_{ki}\big(F_{1;\beta ,p}(u)-F_{1;\beta ,p}(v)\big)D_{1;p,\alpha}(u),
\end{multline}
where $\alpha$, $p$ (respectively, $\beta$, $q$) are summed over $1,\ldots,\mu_1$ (respectively, $1,\ldots, \mu_2$).
By (\ref{usefull}), we have
\begin{multline*}
(u-v)[t_{i,\mu_1+j}(u),t^{\prime}_{\mu_1+h,k}(v)]=(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{h}_2+\pa{j}_2\pa{h}_2}\times \\
\big(\delta_{hj}\sum_{g=1}^{M+N}t_{ig}(u)t^{\prime}_{gk}(v)
-\delta_{ki}\sum_{s=1}^{M+N}t^{\prime}_{\mu_1+h,s}(v)t_{s,\mu_1+j}(u)\big).
\end{multline*}
Substituting by (\ref{511})$-$(\ref{518}), we may rewrite the above identity as the following
\begin{align}\notag
&D_{1;i,\alpha}(u)(u-v)E_{1;\alpha ,j}(u)D^{\prime}_{2;h,\beta}(v)F_{1;\beta,k}(v)+
(-1)^{\pa{j}_2\pa{h}_2}\delta_{hj}D_{1;i,\alpha}(u)D^{\prime}_{1;\alpha ,k}(v)\\\notag
&+(-1)^{\pa{j}_2\pa{h}_2}\delta_{hj}D_{1;i,\alpha}(u)\big(E_{1;\alpha ,q}(v)-E_{1;\alpha ,q}(u)\big)D^{\prime}_{2;q,\beta}(v) F_{\beta ,k}(v)\\\notag
&=(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_2+\pa{k}_1)}(u-v)D^{\prime}_{2;h,\beta}(v) F_{1;\beta ,k}(v)D_{1;i,\alpha}(u)E_{1;\alpha,j}(u)\\\notag
&+(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_2+\pa{k}_1)}\delta_{ki}D_{2;h,\beta}^\prime(v)\big(F_{1;\beta ,p}(u)-F_{1;\beta ,p}(v)\big)D_{1;p,\alpha}(u) E_{1;\alpha ,j}(u)\\ \label{2ef3}
&+(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_2+\pa{k}_1)}\delta_{ki}D^{\prime}_{2;h,\beta}(v)D_{2;\beta ,j}(u),
\end{align}
where $\alpha$, $p$ (respectively, $\beta$, $q$) are summed over $1,\ldots,\mu_1$(resp. $1,\ldots,\mu_2$). Substituting (\ref{2ef1}) and (\ref{2ef2}) into (\ref{2ef3}) and simplifying the result, we obtain
\begin{multline*}
D_{1;i,\alpha}(u)D^\prime_{2;h\beta}(v)D_{2;\beta,j}(v)D^{\prime}_{1;\alpha ,k}(v)\\
+(-1)^{\pa{\alpha}_1\pa{h}_2+\pa{\alpha}_1\pa{k}_1+\pa{j}_2\pa{\beta}_2}(u-v)D_{1;i,\alpha}(u) D^{\prime}_{2;h,\beta}(v)E_{1;\alpha ,j}(u)F_{1;\beta ,k}(v)\\
=
(-1)^{\pa{\alpha}_1\pa{h}_2+\pa{\alpha}_1\pa{k}_1+\pa{j}_2\pa{k}_1}(u-v)D_{1;i,\alpha}(u) D^{\prime}_{2;h,\beta}(v) F_{1;\beta ,k}(v)E_{1;\alpha ,j}(u)\\
+(-1)^{\pa{k}_1\pa{j}_2+\pa{h}_2\pa{\alpha}_1+\pa{\beta}_2\pa{\alpha}_1+\pa{\beta}_2\pa{k}_1}D_{1;i,\alpha}(u) D^{\prime}_{2;h,\beta}(v) D_{1;\alpha,k}^\prime(u) D_{2;\beta ,j}(u).
\end{multline*}
Note that in the above equality, the index $i$ is not involved in those sign factors. We may multiply the matrix $D_1^\prime(u)$ from the left to both sides of the equality above so that we have:
\begin{multline*}
D^\prime_{2;h\beta}(v)D_{2;\beta,j}(v)D^{\prime}_{1;i ,k}(v)\\
+(-1)^{\pa{i}_1\pa{h}_2+\pa{i}_1\pa{k}_1+\pa{j}_2\pa{\beta}_2}(u-v)D^{\prime}_{2;h,\beta}(v)E_{1;i ,j}(u)F_{1;\beta ,k}(v)\\
=
(-1)^{\pa{i}_1\pa{h}_2+\pa{i}_1\pa{k}_1+\pa{j}_2\pa{k}_1}(u-v) D^{\prime}_{2;h,\beta}(v) F_{1;\beta ,k}(v)E_{1;i ,j}(u)\\
+(-1)^{\pa{k}_1\pa{j}_2+\pa{h}_2\pa{i}_1+\pa{\beta}_2\pa{i}_1+\pa{\beta}_2\pa{k}_1}D^{\prime}_{2;h,\beta}(v) D_{1;i,k}^\prime(u) D_{2;\beta ,j}(u).
\end{multline*}
Similar to the above computation, we want to multiply $D_2(v)$ from the left to the above identity. However, we can {\em not} do this directly since the index $h$ is involved in some sign factors; such a phenomenon didn't appear in \cite{BK1, Go, Pe1}.
It turns out that we may multiply a suitable sign factor $(-1)^{\pa{i}_1\pa{h}_2}$ to the above identity so that
\begin{multline*}
(-1)^{\pa{i}_1\pa{j}_2}D^\prime_{2;h,\beta}(v)D_{2;\beta,j}(v)D^{\prime}_{1;i ,k}(v)\\
+(-1)^{\pa{i}_1\pa{k}_1+\pa{j}_2\pa{\beta}_2}(u-v)D^{\prime}_{2;h,\beta}(v)E_{1;i ,j}(u)F_{1;\beta ,k}(v)\\
=
(-1)^{\pa{i}_1\pa{k}_1+\pa{j}_2\pa{k}_1}(u-v) D^{\prime}_{2;h,\beta}(v) F_{1;\beta ,k}(v)E_{1;i ,j}(u)\\
+(-1)^{\pa{k}_1\pa{j}_2+\pa{\beta}_2\pa{i}_1+\pa{\beta}_2\pa{k}_1}D^{\prime}_{2;h,\beta}(v) D_{1;i,k}^\prime(u) D_{2;\beta ,j}(u).
\end{multline*}
Observe that in the very first term we have $D^\prime_{2;h\beta}(v)D_{2;\beta,j}(v)$, which is $\delta_{hj}$, so we may replace $\pa{h}_2$ by $\pa{j}_2$ in the sign factor. Now those sign factors in the above result are free from the index $h$ so we may multiply $D_2(v)$ from the left to obtain
\begin{multline*}
(-1)^{\pa{i}_1\pa{j}_2}D_{2;h,j}(v)D^{\prime}_{1;i ,k}(v)
+(-1)^{\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_2}(u-v)E_{1;i ,j}(u)F_{1;h ,k}(v)\\
=
(-1)^{\pa{i}_1\pa{k}_1+\pa{j}_2\pa{k}_1}(u-v) F_{1;h ,k}(v)E_{1;i ,j}(u)\\
+(-1)^{\pa{k}_1\pa{j}_2+\pa{h}_2\pa{i}_1+\pa{h}_2\pa{k}_1} D_{1;i,k}^\prime(u) D_{2;h ,j}(u).
\end{multline*}
Collecting the corresponding terms and using the fact that $D_{2;h,j}(u)$, $D_{1;i,k}^\prime(u)$ supercommute, we derive (\ref{p517}).
Such a technique appears very often in the remaining part of this article.
For (\ref{p518}), we start with $[t_{i,\mu_1+j}(u), t^{\prime}_{h,\mu_1+k}(v)]=0.$
Multiplying $(u-v)^2$ and computing the bracket after substituting by (\ref{512}) and (\ref{516}), we have
\begin{align}\notag
&(u-v)^2D_{1;i,p}(u)E_{1;p,j}(u)E_{1;h,q}(v)D^{\prime}_{2;q,k}(v)\qquad\qquad\qquad\qquad\qquad\qquad\qquad\\
&-(-1)^{\pa{p}_1\pa{h}_1+\pa{p}_1\pa{q}_2+\pa{j}_2\pa{q}_2}(u-v)E_{1;h,q}(v)D_{1;i,p}(u)(u-v)D^{\prime}_{2;q,k}(v)E_{1;p,j}(u)=0,\label{2ee1}
\end{align}
where the indices $p$ and $q$ are summed from 1 to $\mu_1$ and $\mu_2$, respectively.
Computing the brackets in (5.1) and (\ref{p512}), we obtain the following two identities
\begin{multline*}
(u-v)(-1)^{(\pa{i}_1+\pa{p}_1)(\pa{h}_1+\pa{q}_2)}E_{1;h,q}(v)D_{1;i,p}(u)\\
=(u-v)D_{1;i,p}(u)E_{1;h,q}(v)
-\delta_{hp}(-1)^{\pa{h}_1\pa{p}_1}D_{1;i,g_1}(u)\big(E_{1;g_1,q}(v)-E_{1;g_1,q}(u)\big),
\end{multline*}
\begin{multline*}
(u-v)(-1)^{(\pa{p}_1+\pa{j}_2)(\pa{q}_2+\pa{k}_2)}D^{\prime}_{2;q,k}(v)E_{1;p,j}(u)\\
=(u-v)E_{1;p,j}(u)D^{\prime}_{2;q,k}(v)
+ \delta_{jq}(-1)^{\pa{q}_2\pa{j}_2}\big(E_{1;p,g_2}(u)-E_{1;p,g_2}(v)\big)D^{\prime}_{2;g_2,k}(v),
\end{multline*}
where the indices $p, g_1$ (respectively, $q, g_2$) are summed from 1 to $\mu_1$ (respectively, $\mu_2$).
Substituting these two into the second term of (\ref{2ee1}), multiplying some suitable choices of sign factors as in the proof of (\ref{p517}) so that we may multiply $D_1(u)$ from the left and $D_2(v)$ from the right simultaneously, we derive that
\begin{multline}
(u-v)^2[E_{1;i,j}(u),E_{1;h,k}(v)]=\\
(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1}(u-v)E_{1;h,j}(v)\big(E_{1;i,k}(v)-E_{1;i,k}(u)\big)\\
+(-1)^{\pa{j}_2\pa{h}_1+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2}(u-v)\big(E_{1;i,k}(u)-E_{1;i,k}(v)\big)E_{;1h,j}(u)\\
+\big(E_{1;i,j}(v)-E_{1;i,j}(u)\big)\big(E_{1;h,k}(u)-E_{1;h,k}(v)\big).\label{2ee2}
\end{multline}
For a power series $P$ in $Y_{\mu}[[u^{-1}, v^{-1}]]$, we write $\big\lbrace P\big\rbrace_{d}$ for the homogeneous component of $P$ of total degree $d$ in the variables $u^{-1}$ and $v^{-1}$. Then (\ref{p518}) is a consequence of the following claim.
{\bf{Claim:}} For $d\ge 1$, we have
\begin{multline*}
(u-v)\big\lbrace [E_{1;i,j}(u), E_{1;h,k}(v)]\big\rbrace_{d+1}=\\
\big\lbrace (-1)^{\pa{j}_2\pa{h}_1+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2}\big(E_{1;i,k}(u)-E_{1;i,k}(v)\big)\big(E_{1;h,j}(u)-E_{1;h,j}(v)\big)\big\rbrace_d
\end{multline*}
We prove the claim by induction on $d$.
For $d=1$, take $\big\lbrace\:\big\rbrace_0$ on (\ref{2ee2}) so that
\[
(u-v)^2\big\lbrace [E_{1;i,j}(u),E_{1;h,k}(v)]\big\rbrace_2=0.
\]
Note that the right-hand side of (\ref{2ee2}) is zero when $u=v$, hence we may divide both sides by $(u-v)$ and therefore $(u-v)\big\lbrace [E_{1;i,j}(u),E_{1;h,k}(v)]\big\rbrace_2=0$, as desired.
\\
Assuming that the claim is true for some $d>1$, so we have
\begin{multline}
(-1)^{\pa{k}_2\pa{j}_2+\pa{k}_2\pa{h}_1+\pa{j}_2\pa{h}_1+1}\big\lbrace[E_{1;i,k}(u), E_{1;h,j}(v)]\big\rbrace_{d+1}=\\
\Big\lbrace \frac{\big(E_{1;i,j}(v)-E_{1;i,j}(u)\big)\big(E_{1;h,k}(u)-E_{1;h,k}(v)\big)}{u-v}\Big\rbrace_d\label{2ee3}
\end{multline}
Note that the right-hand side of (\ref{2ee3}) is zero when $u=v$, which implies
\begin{equation}
E_{1;i,k}(v)E_{1;h,j}(v)=(-1)^{\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1+\pa{i}_1\pa{j}_2+1}E_{1;h,j}(v)E_{1;i,k}(v).\label{2ee4}
\end{equation}
Take $\big\lbrace\:\big\rbrace_d$ on (\ref{2ee2}) and replace the last term by (\ref{2ee3}):
\begin{align*}
(&u-v)^2\big\lbrace [E_{1;i,j}(u), E_{1;h,k}(v)]\big\rbrace_{d+2}\\
&=(u-v)(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1}\big\lbrace E_{1;h,j}(v)\big(E_{1;i,k}(v)-E_{1;i,k}(u)\big)\big\rbrace_{d+1}\\
& +(u-v)(-1)^{\pa{j}_2\pa{h}_1+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2}\big\lbrace \big(E_{1;i,k}(u)-E_{1;i,k}(v)\big)E_{1;h,j}(u)\big\rbrace_{d+1}\\
& +(u-v)\big\lbrace (-1)^{\pa{k}_2\pa{j}_2+\pa{k}_2\pa{h}_1+\pa{j}_2\pa{h}_1}E_{1;i,k}(u)E_{1;h,j}(v)-\\
&\qquad\qquad\qquad\qquad\qquad\qquad
(-1)^{\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1+\pa{i}_1\pa{j}_2+1}E_{1;h,j}(v)E_{1;i,k}(u)\big\rbrace_d\\
&=(u-v)\big\lbrace (-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1}E_{1;h,j}(v) E_{1;i,k}(v)\big\rbrace_{d+1}\\
& +(u-v)(-1)^{\pa{j}_2\pa{h}_1+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2}\big\lbrace E_{1;i,k}(u)E_{1;h,j}(u)-E_{1;i,k}(v)E_{1;h,j}(u) \big\rbrace_{d+1}\\
& +(u-v)\big\lbrace (-1)^{\pa{k}_2\pa{j}_2+\pa{k}_2\pa{h}_1+\pa{j}_2\pa{h}_1}E_{1;i,k}(u)E_{1;h,j}(v)\big\rbrace_d
\end{align*}
Substituting the term $(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{h}_1+\pa{j}_2\pa{h}_1}E_{1;h,j}(v) E_{1;i,k}(v)$ by (\ref{2ee4}) and simplifying the result, we have
\begin{align*}
&(u-v)^2\big\lbrace [E_{1;i,j}(u), E_{1;h,k}(v)]\big\rbrace_{d+2}=\\
&(u-v)\big\lbrace (-1)^{\pa{j}_2\pa{h}_1+\pa{j}_2\pa{k}_2+\pa{h}_1\pa{k}_2} \big(E_{1;i,k}(u)-E_{1;i,k}(v)\big)\big(E_{1;h,j}(u)-E_{1;h,j}(v)\big)\big\rbrace_{d+1}.
\end{align*}
Dividing both sides by $u-v$ establishes the claim.
\end{proof}
\section{Special Cases: $n=3$ and the super Serre relations}
In this section, we will consider the generators $D$'s, $E$'s and $F$'s in different super Yangians at the same time but using the same notation. It should be clear from the context which super Yangian we are dealing with.
Similar to the proof of Proposition \ref{n=2}, we compute the matrix products
(\ref{T=FDE}) and (\ref{Tp=FDE}) with respect to the composition $\mu=(\mu_1,\mu_2,\mu_3)$ and derive the following identities.
\begin{eqnarray}
\label{610}t_{i,j}(u)&=&D_{1;i,j},\\
\label{611}t_{i,\mu_1+j}(u)&=&\sum_{p=1}^{\mu_1}D_{1;i,p}E_{1;p,j}(u),\\
\label{612}t_{i,\mu_1+\mu_2+j}(u)&=&\sum_{p=1}^{\mu_1}D_{1;i,p}E_{1,3;p,j}(u),\\
\label{613}t^{\prime}_{i,\mu_1+\mu_2+j}(u)&=&\sum_{p^\prime=1}^{\mu_3}\sum_{q=1}^{\mu_2}\big(E_{1;i,q}(u)E_{2;q,p^\prime}(u)-E_{1,3;i,p^\prime}(u)\big)D^{\prime}_{3;p^\prime,j}(u),\\
\label{614}t^{\prime}_{\mu_1+i,\mu_1+\mu_2+j}(u)&=&-\sum_{p^\prime=1}^{\mu_3}E_{2;i,p'}(u)D^{\prime}_{3;p^\prime,j}(u),\\
\label{615}t^{\prime}_{\mu_1+\mu_2+i,\mu_1+j}(u)&=&-\sum_{p^\prime=1}^{\mu_3}D^{\prime}_{3;i,p^\prime}(u)F_{2;p^\prime,j}(u),
\end{eqnarray}
where (\ref{610}) holds for all $1\leq i,j\leq \mu_1$,
(\ref{611}) holds for all $1\leq i\leq \mu_1$, $1\leq j\leq \mu_2$,
(\ref{612}) and (\ref{613}) hold for all $1\leq i\leq \mu_1$, $1\leq j\leq \mu_3$,
(\ref{614}) holds for all $1\leq i\leq \mu_2$, $1\leq j\leq \mu_3$,
and (\ref{615}) holds for all $1\leq i\leq \mu_3$, $1\leq j\leq \mu_2$.
\begin{lemma}\label{3be}
The following identities hold in $Y_{(\mu_1,\mu_2,\mu_3)}((u^{-1},v^{-1}))$:
\begin{equation}\label{61a}
[E_{1;i,j}(u), F_{2;h,k}(v)] = 0,
\end{equation}
\begin{multline}\label{61b}
[E_{1;i,j}(u), E_{2;h,k}(v)] = \\
\dfrac{(-1)^{\pa{j}_2\pa{h}_2}}{u-v}\delta_{hj}\sum_{q=1}^{\mu_2}\big\lbrace\big(E_{1;i,q}(u)-E_{1;i,q}(v)\big)E_{2;q,k}(v)
+ E_{1,3;i,k}(v) - E_{1,3;i,k}(u)\big\rbrace,
\end{multline}
\begin{multline}\label{61c}
[E_{1,3;i,j}(u), E_{2;h,k}(v)] =\\
(-1)^{\pa{i}_1\pa{j}_3+\pa{i}_1\pa{h}_2+\pa{h}_2\pa{j}_3+\pa{g}_2}
E_{2;h,j}(v) [E_{1;i,g}(u), E_{2;g,k}(v)],
\end{multline}
\begin{multline}\label{61d}
\sum_{q=1}^{\mu_2}[E_{1;i,j}(u), E_{1,3;h,k}(v) - E_{1;h,q}(v) E_{2;q,k}(v)] =\\
(-1)^{\pa{h}_1\pa{j}_2+\pa{j}_2\pa{k}_3+\pa{h}_1\pa{k}_3+\pa{g}_2+1}
[E_{1;i,g}(u), E_{2;g,k}(v)] E_{1;h,j}(u).
\end{multline}
Here (\ref{61a}) holds for all $1\leq i\leq \mu_1$, $1\leq j,k\leq \mu_2$, $1\leq h\leq \mu_3$, (\ref{61b}) holds for all $1\leq i\leq \mu_1$, $1\leq j,h\leq \mu_2$, $1\leq k\leq \mu_3$, (\ref{61c}) holds for all $1\leq i\leq \mu_1$, $1\leq h\leq \mu_2$, $1\leq j,k\leq \mu_3$, and (\ref{61d}) holds for all $1\leq i,h\leq \mu_1$, $1\leq j,g\leq \mu_2$, $1\leq k\leq \mu_3$.
\end{lemma}
\begin{proof}
By (\ref{usefull}), we have $[t_{i,\mu_1+j}(u),t^{\prime}_{\mu_1+\mu_2+h,\mu_1+k}(v)]=0$. Substituting by (\ref{611}) and (\ref{615}), we have
\[
[D_{1;i,p}(u)E_{1;p,j}(u),-D^{\prime}_{3;h,q}(v)F_{2;q,k}(v)]=0.
\]
Computing the bracket, we obtain
\begin{multline}\label{3ee1}
D_{1;i,p}(u)E_{1;p,j}(u)D^{\prime}_{3;h,q}(v)F_{2;q,k}(v)-\\
(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_3+\pa{k}_2)}D^{\prime}_{3;h,q}(v)F_{2;q,k}(v)D_{1;i,p}(u)E_{1;p,j}(u)=0,
\end{multline}
where $p$ and $q$ are summed over $1,\ldots,\mu_1$ and $1,\ldots,\mu_3$, respectively. Similarly, by (\ref{usefull}), we have
\[
[t_{ij}(u),t^{\prime}_{\mu_1+\mu_2+h,\mu_1+k}(v)]=[t_{i,\mu_1+j}(u),t^{\prime}_{\mu_1+\mu_2+h,\mu_1+\mu_2+k}(v)]=0,
\]
which implies that
\[
[D_{1;i,j}(u),F_{2;h,k}(v)]=[E_{1;i,j}(u),D^{\prime}_{3;h,k}(v)]=0.
\]
Substituting these into (\ref{3ee1}) and using the fact that $D_{1;i,j}(u)$, $D^{\prime}_{3;h,k}(v)$ supercommute, we have
\begin{multline*}
(-1)^{\pa{j}_2\pa{q}_3}D_{1;i,p}(u)D^{\prime}_{3;h,q}(v)E_{1;p,j}(u)F_{2;q,k}(v)-\\
(-1)^{\pa{j}_2\pa{k}_2+\pa{p}_1\pa{q}_3+\pa{p}_1\pa{k}_2}D_{1;i,p}(u)D^{\prime}_{3;h,q}(v)F_{2;q,k}(v)E_{1;p,j}(u)=0.
\end{multline*}
The sign factors are free from the indices $i$ and $h$. Multiplying $D_{3}(v)D^{\prime}_{1}(u)$ from the left, we obtain (\ref{61a})
By (\ref{usefull}) again, we have
\[
(u-v)[t_{i,\mu_1+j}(u),t^{\prime}_{\mu_1+h,\mu_1+\mu_2+k}(v)]=(-1)^{\pa{j}_2\pa{h}_2}\delta_{jh}\sum_{s=1}^{M+N}t_{is}(u)t_{s,\mu_1+\mu_2+k}(v).
\]
Substituting by (\ref{610})--(\ref{615}), we have
\begin{multline}\label{3ee2}
(u-v)[D_{1;i,p}(u)E_{1;p,j}(u),-E_{2;h,q}(v)D^{\prime}_{3;q,k}(v)]=\\
(-1)^{\pa{j}_2\pa{h}_2}\delta_{jh}D_{1;i,p}(u)\big\lbrace \big(E_{1;p,r}(v)E_{2;r,q}(v)-E_{1,3;p,q}(v)\big)\\
-E_{1;p,r}(u)E_{2;r,q}(v)+E_{1,3;p,q}(u)\big\rbrace D^{\prime}_{3;q,k}(v),
\end{multline}
where the indices $p,q,r$ are summed from 1 to $\mu_1,\mu_3,\mu_2$, respectively.
Using the facts that
\begin{eqnarray*}
\big[E_{1;i,j}(v),D^{\prime}_{3;h,k}(u)\big]=0,& \qquad\big(\text{proved in the proof of (\ref{61a})}\big)\\
\big[E_{2;i,j}(v),D_{1;h,k}(u)\big]=0,& \qquad \big(\text{obtained from}\;\;[t_{i,j}(u),t^{\prime}_{\mu_1+h,\mu_1+\mu_2+k}(v)]=0\big)
\end{eqnarray*}
we may rewrite (\ref{3ee2}) as the following
\begin{multline}\label{3ee3}
(u-v) D_{1;i,p}(u) \lbrace E_{1;p,j}(u)E_{2;h,q}(v)-\\
(-1)^{(\pa{j}_2+\pa{p}_1)(\pa{h}_2+\pa{q}_3)}E_{2;h,q}(v)E_{1;p,j}(u)\rbrace D^{\prime}_{3;q,k}(v)\\
=(-1)^{\pa{j}_2\pa{h}_2+1}\delta_{jh}D_{1;i,p}(u)\big\lbrace \big(E_{1;p,r}(v)E_{2;r,q}(v)-E_{1,3;p,q}(v)\big)\\
-E_{1;p,r}(u)E_{2;r,q}(v)+E_{1,3;p,q}(u)\big\rbrace D^{\prime}_{3;q,k}(v).
\end{multline}
The sign factors are free from the indices $i$ and $k$. Canceling $D_{1}(u)$ from the left and $D^{\prime}_3(v)$ from the right on both sides of (\ref{3ee3}) and dividing both sides by $u-v$, we have deduced (\ref{61b}).
To show (\ref{61c}), the identity (\ref{p512}) in $Y_{(\mu_2,\mu_3)}((u^{-1},v^{-1}))$ reads as
\begin{equation*}
(u-v)[E_{1;h,k}(u),D^{\prime}_{2;i,j}(v)]=(-1)^{\pa{i}_2\pa{k}_2}\delta_{ki}\big(E_{1;h,p}(u)-E_{1;h,p}(v)\big)D^{\prime}_{2;p,j}(v).
\end{equation*}
Applying the map $\psi_{\mu_1}$ to this identity and using (4.3)$-$(4.5), we have the following identity in $Y_{(\mu_1,\mu_2,\mu_3)}((u^{-1},v^{-1}))$
\begin{equation*}
(u-v)[E_{2;h,k}(u),D^{\prime}_{3;i,j}(v)]=(-1)^{\pa{i}_3\pa{k}_3}\delta_{ki}\big(E_{2;h,p}(u)-E_{2;h,p}(v)\big)D^{\prime}_{3;p,j}(v).
\end{equation*}
Taking the coefficient of $u^0$ in the above identity, we obtain
\begin{equation*}
[E_{2;h,k}^{(1)},D^{\prime}_{3;i,j}(v)]=-(-1)^{\pa{i}_3\pa{k}_3}\delta_{ki}E_{2;h,p}(v)D^{\prime}_{3;p,j}(v).
\end{equation*}
Also by (\ref{ter}), we have
\begin{equation*}
E_{1,3;i,j}(u)=(-1)^{\pa{g}_2}[E_{1;i,g}(u),E^{(1)}_{2;g,j}],\; \text{for any} \; 1\leq g\leq\mu_2.
\end{equation*}
By (\ref{612}), (\ref{613}), the super Jacobi identity and the fact that $[E_{1;i,g}(u),D^{\prime}_{3;h,k}(v)]=0$, we have
\begin{align}\notag\label{3ee4}
[E_{1,3;i,j}(u),D^{\prime}_{3;h,k}(v)]&=\big[(-1)^{\pa{g}_2}[E_{1;i,g}(u),E^{(1)}_{2;g,j}],D^{\prime}_{3;h,k}(v)\big]\\\notag
&=(-1)^{\pa{g}_2}\big[E_{1;i,g}(u),[E^{(1)}_{2;g,j},D^{\prime}_{3;h,k}(v)]\big]+0\\\notag
&=(-1)^{\pa{g}_2}\big[E_{1;i,g}(u),-(-1)^{\pa{j}_3\pa{h}_3}\delta_{hj}E_{2;g,p}(v)D^{\prime}_{3;p,k}(v)\big]\\
&=(-1)^{1+\pa{g}_2+\pa{j}_3\pa{h}_3}\delta_{hj}\big[E_{1;i,g}(u),E_{2;g,p}(v)\big]D^{\prime}_{3;p,k}(v).
\end{align}
By (\ref{usefull}) and (\ref{610})--(\ref{615}), we have
\[
[t_{i,\mu_1+\mu_2+j}(u),t^{\prime}_{\mu_1+h,\mu_1+\mu_2+k}(v)]=[D_{1;i,p}(u)E_{1,3;p,j}(u),-E_{2;h,q}(v)D^{\prime}_{3;q,k}(v)]=0,
\]
where $p$ and $q$ are summed from 1 to $\mu_1$ and $\mu_3$, respectively. Multiplying $D_1^{\prime}(u)$ from the left, we have $[E_{1,3;i,j}(u),E_{2;h,q}(v)D^{\prime}_{3;q,k}(v)]=0,$ which may be written as
\begin{multline*}
[E_{1,3;i,j}(u),E_{2;h,q}(v)]D^{\prime}_{3;q,k}(v)+\\
(-1)^{(\pa{i}_1+\pa{j}_3)(\pa{h}_2+\pa{q}_3)}E_{2;h,q}(v)[E_{1,3;i,j}(u),D^{\prime}_{3;q,k}(v)]=0.
\end{multline*}
Substituting the last bracket by (\ref{3ee4}), we have
\begin{multline*}
[E_{1,3;i,j}(u),E_{2;h,q}(v)]D^{\prime}_{3;q,k}(v)=\\
(-1)^{(\pa{i}_1+\pa{j}_3)(\pa{h}_2+\pa{j}_3)+\pa{g}_2+\pa{j}_3\pa{j}_3}E_{2;h,j}(v)[E_{1;i,g}(u),E_{2;g,p}(v)]D^{\prime}_{3;p,k}(v).
\end{multline*}
Multiplying $D_{3}(v)$ from the right to the above equality, we acquire (\ref{61c}).
Taking the coefficient of $u^0$ in (\ref{61b}), we have
\begin{equation}\label{3ee5}
[E_{1;i,j}^{(1)},E_{2;h,k}(v)]=(-1)^{\pa{j}_2\pa{h}_2}\delta_{hj}\big(-E_{1;i,q}(v)E_{2;q,k}(v)+E_{1,3;i,k}(v)\big).
\end{equation}
Taking the coefficient of $v^0$ in (\ref{p511}), we have
\begin{equation}\label{3ee6}
[D_{1;i,j}(u),E_{1;h,k}^{(1)}]=(-1)^{\pa{j}_1\pa{h}_1}\delta_{hj}D_{1;i,p}(u)E_{1;p,k}(u).
\end{equation}
Together with the super Jacobi identity and the fact that $[D_{1;i,j}(u),E_{2;g,k}(v)]=0$, (\ref{3ee5}) and (\ref{3ee6}) imply that
\begin{align*}
[D_{1;i,j}(u),E_{1,3;h,k}(v)&-E_{1;h,q}(v)E_{2;q,k}(v)]=[D_{1;i,j}(u),(-1)^{\pa{g_2}}\big[E_{1;h,g}^{(1)},E_{2;g,k}(v)]\big]\notag\\
&=(-1)^{\pa{g_2}}\big[[D_{1;i,j}(u),E_{1;h,g}^{(1)}],E_{2;g,k}(v)\big]+0\\
&=(-1)^{\pa{g_2}}[(-1)^{\pa{j}_1\pa{h}_1}\delta_{hj}D_{1;i,p}(u)E_{1;p,g}(u),E_{2;g,k}(v)]\\
&=(-1)^{\pa{g_2}+\pa{j}_1\pa{h}_1}\delta_{hj}D_{1;i,p}(u)[E_{1;p,g}(u),E_{2;g,k}(v)].
\end{align*}
Summing $j$ from 1 to $\mu_1$ in the above identity, we derive
\begin{multline}\label{3ee7}
(-1)^{\pa{i}_1(\pa{h}_1+\pa{k}_3)}\big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big)D_{1;i,p}(u)\\
-(-1)^{\pa{p}_1(\pa{h}_1+\pa{k}_3)}D_{1;i,p}(u)\big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big),\\
=(-1)^{\pa{g}_2+\pa{p}_1\pa{k}_3}\delta_{hp}D_{1;i,p'}(u)[E_{1;p',g}(u),E_{2;g,k}(v)]
\end{multline}
where $r,p,p'$ are summed over $\mu_2, \mu_1 ,\mu_1$, respectively.
On the other hand, by (\ref{usefull}) and (\ref{610})--(\ref{615}), we have
\begin{multline}\label{3ee8}
[t_{i,\mu_1+j}(u),t^{\prime}_{h,\mu_1+\mu_2+k}(v)]=\\
[D_{1;i,p}(u)E_{1;p,j}(u),\big(E_{1;h,r}(v)E_{2;r,q}(v)-E_{1,3;h,q}(v)\big)D^{\prime}_{3;q,k}(v)]=0,
\end{multline}
where $p$ and $q$ are summed from 1 to $\mu_1$ and $\mu_3$, respectively. Multiplying $D_{3}(v)$ from the right, note that $D_3$ supercommutes with $E_1$ and $D_1$, and computing the bracket, we obtain
\begin{multline}\label{3ee9}
D_{1;i,p}(u)E_{1;p,j}(u)\big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big)\\
-(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_1+\pa{k}_3)} \big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big)D_{1;i,p}(u)E_{1;p,j}(u) =0
\end{multline}
where $p$ and $r$ are summed from 1 to $\mu_1$ and $\mu_2$, respectively. Substituting (\ref{3ee7}) into the second term of (\ref{3ee9}), we have
\begin{multline*}
D_{1;i,p}(u)E_{1;p,j}(u)\big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big)\\
-(-1)^{(\pa{p}_1+\pa{j}_2)(\pa{h}_1+\pa{k}_3)} \big\lbrace D_{1;i,p}(u)\big(E_{1;h,r}(v)E_{2;r,k}(v)-E_{1,3;h,k}(v)\big)\big\rbrace E_{1;p,j}(u) \\
-(-1)^{(\pa{g}_2+\pa{j}_2)(\pa{h}_1+\pa{k}_3)} \big\lbrace D_{1;i,p}(u)\big[E_{1;p,g}(u),E_{2;g,k}(v)\big] \big\rbrace E_{1;h,j}(u)=0.
\end{multline*}
Multiplying $D^{\prime}_{1}(u)$ from the left, we deduce that
\begin{multline*}
E_{1;i,j}(u)\big(E_{1;h,q}(v)E_{2;q,k}(v)-E_{1,3;h,k}(v)\big)\\
-(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{h}_1+\pa{k}_3)}\big(E_{1;h,q}(v)E_{2;q,k}(v)-E_{1,3;h,k}(v)\big)E_{1;i,j}(u)\\
=(-1)^{\pa{j}_2(\pa{h}_1+\pa{k}_3)+\pa{g}_2+\pa{h}_1\pa{k}_3}\big[E_{1;i,g}(u),E_{2;g,k}(v)\big]E_{1;h,j}(u).
\end{multline*}
Simplifying the above, we obtain (\ref{61d}).
\end{proof}
The following lemma can be deduced by applying the automorphism $\zeta_{N|M}$ to the corresponding identities of Lemma \ref{3be} in $Y_{\overleftarrow{\mu}}=Y_{(\mu_3,\mu_2,\mu_1)}=Y_{N|M}$.
\begin{lemma}\label{3bf}
The following identities hold in $Y_{(\mu_1,\mu_2,\mu_3)}((u^{-1},v^{-1}))$:
\begin{equation}\label{62a}
[F_{1;i,j}(u),E_{2;h,k}(v)]=0,
\end{equation}
\begin{multline}\label{62b}
[F_{1;i,j}(u),F_{2;h,k}(v)]=
\dfrac{(-1)^{\pa{i}_2\pa{j}_1+\pa{i}_2\pa{h}_3+\pa{j}_1\pa{h}_3}}{u-v}\delta_{ik}\times\\
\big\lbrace \sum_{q=1}^{\mu_2}F_{2;h,q}(v)\big(F_{1;q,j}(v)-F_{1;q,j}(u)\big)-F_{3,1;h,j}(v)+F_{3,1;h,j}(u)\big\rbrace,
\end{multline}
\begin{multline}\label{62c}
[F_{3,1;i,j}(u),F_{2;h,k}(v)]=
(-1)^{\pa{i}_3\pa{j}_1+\pa{i}_3\pa{h}_3+\pa{j}_1\pa{h}_3+\pa{g}_2+1}
[F_{2;h,g}(v),F_{1;g,j}(u)]F_{2;i,k}(v),
\end{multline}
\begin{multline}\label{62d}
\sum_{q=1}^{\mu_2}[F_{1;i,j}(u) , F_{2;h,q}(v)F_{1;q,k}(v)-F_{3,1;h,k}(v)]=\\
(-1)^{(\pa{h}_3+\pa{j}_1)(\pa{k}_1+\pa{g}_2)} F_{1;i,k}(u)[F_{1;g,j}(u),F_{2;h,g}(v)].
\end{multline}
Here (\ref{62a}) holds for all $1\leq i,h\leq \mu_2$, $1\leq j\leq \mu_1$, $1\leq k\leq \mu_3$, (\ref{62b}) holds for all $1\leq i,k\leq \mu_2$, $1\leq j\leq \mu_1$, $1\leq h\leq \mu_3$, (\ref{62c}) holds for all $1\leq i,h\leq \mu_3$, $1\leq j\leq \mu_1$, $1\leq j,k\leq \mu_2$, and (\ref{62d}) holds for all $1\leq i,g\leq \mu_2$, $1\leq j,k\leq \mu_1$, $1\leq h\leq \mu_3$.
\end{lemma}
Our next lemma is a generalization of \cite[Lemma 6.3]{Pe1}. It is surprising that there are no sign factors appearing in the resulting identities.
\begin{lemma}\label{3ef0}
The following identities hold in $Y_{(\mu_1,\mu_2,\mu_3)}[[u^{-1},v^{-1},w^{-1}]]$:
\begin{eqnarray}
\label{63a}&\big[[E_{1;i,j}(u),E_{2;h,k}(v)],E_{2;f,g}(v)\big]=0,\\[1mm]
\notag&\big[E_{1;i,j}(u),[E_{1;h,k}(u),E_{2;f,g}(v)]\big]=0,\\[1mm]
\label{63c}&\big[[E_{1;i,j}(u),E_{2;h,k}(v)],E_{2;f,g}(w)\big]+
\big[[E_{1;i,j}(u),E_{2;h,k}(w)],E_{2;f,g}(v)\big]=0,\\[1mm]
\notag&\big[E_{1;i,j}(u),[E_{1;h,k}(v),E_{2;f,g}(w)]\big]+
\big[E_{1;i,j}(v),[E_{1;h,k}(u),E_{2;f,g}(w)]\big]=0,\\[1mm]
\notag&\big[[F_{1;i,j}(u),F_{2;h,k}(v)],F_{2;f,g}(v)\big]=0,\\[1mm]
\notag&\big[F_{1;i,j}(u),[F_{1;h,k}(u),F_{2;f,g}(v)]\big]=0,\\[1mm]
\notag&\big[[F_{1;i,j}(u),F_{2;h,k}(v)],F_{2;f,g}(w)\big]+\big[[F_{1;i,j}(u),F_{2;h,k}(w)],F_{2;f,g}(v)\big]=0,\\[1mm]
\notag&\big[F_{1;i,j}(u),[F_{1;h,k}(v),F_{2;f,g}(w)]\big]+\big[F_{1;i,j}(v),[F_{1;h,k}(u),F_{2;f,g}(w)]\big]=0.
\end{eqnarray}
The identities hold for all $1\leq i\leq \mu_a, 1\leq j\leq \mu_{a+1}$ if $E_{a;i.j}(u)$ appears and
hold for all $1\leq h \leq \mu_{a+1}$, $1\leq k \leq \mu_a$ if $F_{a;h,k}(u)$ appears, where $a$=1 or 2.
\end{lemma}
\begin{proof}
We prove (\ref{63a}) and (\ref{63c}) in detail while the others are similar.
To show (\ref{63a}), we first claim that
\[
[E_{a;i,j}(v),E_{a;h,k}(v)]=0 \quad \text{for}\; a=1,2 \quad\text{in}\quad Y_{(\mu_1,\mu_2,\mu_3)}[[v^{-1}]].
\]
Indeed, the case $a=1$ follows from (\ref{p518}) and $a=2$ follows from applying the map $\psi_{\mu_1}$ to (\ref{p518}) in $Y_{(\mu_2,\mu_3)}[[v^{-1}]]$.
By the super Jacobi identity, together with the above claim and (\ref{61b}), it suffices to prove the case when $j=h=f$. In this case, we compute the following bracket by Lemma~\ref{3be} as below.
{\allowdisplaybreaks
\begin{align*}
(&u-v)\big[[E_{1;i,j}(u),E_{2;j,k}(v)],E_{2;j,g}(v)\big]\\
&=(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{j}_2+\pa{k}_3)}(u-v)\big[E_{2;j,k}(v),[E_{1;i,j}(u),E_{2;j,g}(v)]\big]\\
&=(-1)^{(\pa{i}_1+\pa{j}_2)(\pa{j}_2+\pa{k}_3)+(\pa{i}_1+\pa{g}_3)(\pa{j}_2+\pa{k}_3)}(u-v)\big[\,[E_{1;i,j}(u),E_{2;j,g}(v)] \,, E_{2;j,k}(v)\big]\\
&=(-1)^{(\pa{j}_2+\pa{g}_3)(\pa{j}_2+\pa{k}_3)}\big[(-1)^{\pa{j}_2} E_{1;i,q}(u) E_{2;q,g}(v)-E_{1;i,q}(v) E_{2;q,g}(v)\\
&\qquad\qquad + E_{1,3;i,g}(v) - E_{1,3;i,g}(u)\, , \,E_{2;j,k}(v)\big]\\
&=(-1)^{\pa{j}_2\pa{k}_3+\pa{j}_2\pa{g}_3+\pa{g}_3\pa{k}_3}\big\lbrace[E_{1;i,q}(u)E_{2;q,g}(v),E_{2;j,k}(v)]+[E_{1,3;i,g}(v),E_{2;j,k}(v)]\\
&\qquad -[E_{1;i,q}(v)E_{2;q,g}(v),E_{2;j,k}(v)]-[E_{1,3;i,g}(u),E_{2;j,k}(v)]\big\rbrace\\
&=(-1)^{\pa{j}_2\pa{k}_3+\pa{j}_2\pa{g}_3+\pa{g}_3\pa{k}_3}\big\lbrace
E_{1;i,q}(u)[E_{2;q,g}(v) , E_{2;j,k}(v)] \\
&\,+(-1)^{(\pa{q}_2+\pa{g}_3)(\pa{j}_2+\pa{k}_3)}[ E_{1;i,q}(u), E_{2;j,k}(v)] E_{2;q,g}(v)
-E_{1;i,q}(v)[E_{2;q,g}(v) , E_{2;j,k}(v)]\\
&\,-(-1)^{(\pa{q}_2+\pa{g}_3)(\pa{j}_2+\pa{k}_3)}[ E_{1;i,q}(v), E_{2;j,k}(v)] E_{2;q,g}(v)\\
&\,+(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{g}_3+\pa{j}_2\pa{g}_3+\pa{j}_2} E_{2;j,g}(v) [E_{1;i,j}(v) , E_{2;j,k}(v)]\\
&\,-(-1)^{\pa{i}_1\pa{j}_2+\pa{i}_1\pa{g}_3+\pa{j}_2\pa{g}_3+\pa{j}_2} E_{2;j,g}(v) [E_{1;i,j}(u) , E_{2;j,k}(v)]\big\rbrace\\
&=(-1)^{\pa{j}_2\pa{k}_3+\pa{q}_2\pa{j}_2+\pa{q}_2\pa{k}_3} [E_{1;i,q}(u), E_{2;j,k}(v)] E_{2;q,g}(v) \\
&\,+(-1)^{\pa{j}_2+\pa{i}_1\pa{j}_2+\pa{i}_1\pa{g}_3+\pa{j}_2\pa{k}_3+\pa{g}_3\pa{k}_3} E_{2;j,g}(v)[ E_{1;i,j}(v), E_{2;j,k}(v)]\\
&\,-(-1)^{\pa{j}_2\pa{k}_3+\pa{q}_2\pa{j}_2+\pa{q}_2\pa{k}_3} [E_{1;i,q}(u), E_{2;j,k}(v)] E_{2;q,g}(v) \\
&\,-(-1)^{\pa{j}_2+\pa{i}_1\pa{j}_2+\pa{i}_1\pa{g}_3+\pa{j}_2\pa{k}_3+\pa{g}_3\pa{k}_3} E_{2;j,g}(v)[ E_{1;i,j}(u), E_{2;j,k}(v)]\\
&=(-1)^{\pa{j}_2}\big[ [E_{1;i,j}(u) , E_{2;j,k}(v)] , E_{2;j,g}(v) \big]
-(-1)^{\pa{j}_2}\big[ [E_{1;i,j}(v) , E_{2;j,k}(v)] , E_{2;j,g}(v) \big]
\end{align*}
}
Thus we have
\begin{multline}\label{3eee0}
(u-v-(-1)^{\pa{j}_2})\big[[E_{1;i,j}(u),E_{2;j,k}(v)],E_{2;j,g}(v)\big]=\\
-(-1)^{\pa{j}_2}\big[[E_{1;i,j}(v),E_{2;j,k}(v)],E_{2;j,g}(v)\big]
\end{multline}
Note that the right-hand side of (\ref{3eee0}) is independent of the choice of $u$. Specifying $u=v+(-1)^{\pa{j}_2}$ in (\ref{3eee0}), we have
\begin{equation*}
0=-(-1)^{\pa{j}_2}\big[[E_{1;i,j}(v),E_{2;j,k}(v)],E_{2;j,g}(v)\big],
\end{equation*}
and hence
\begin{equation*}
(u-v-(-1)^{\pa{j}_2})\big[[E_{1;i,j}(u),E_{2;j,k}(v)],E_{2;j,g}(v)\big]=0, \,\,\text{for any } u.
\end{equation*}
Choose $u$ such that $u-v-(-1)^{\pa{j}_2}$ is invertible, and then (\ref{63a}) follows.
To show (\ref{63c}), it suffices to show that
\begin{equation}\label{3eeec1}
(u-w)(v-w)(u-v)\big[[E_{1;i,j}(u),E_{2;h,k}(v)],E_{2;f,g}(w)\big]
\end{equation}
is symmetric in $v$ and $w$. We may further assume $j=h$, as in the proof of (\ref{63a}).
By (\ref{61b}), we have
\begin{multline*}
(u-w)(v-w)(u-v)\big[[E_{1;i,j}(u),E_{2;h,k}(v)],E_{2;f,g}(w)\big]
=(v-w)(u-w)\times\\
(-1)^{\pa{j}_2}\big[ \big(E_{1;i,q}(u)-E_{1;i,q}(v) \big)E_{2;q,k}(v)+E_{1,3;i,k}(v)-E_{1,3;i,k}(u)\,,\,E_{2;f,g}(w)\big].
\end{multline*}
Computing the brackets by Lemma \ref{3be}, we have
{\allowdisplaybreaks
\begin{align*}\notag
(u&-w)(v-w)(u-v)\big[[E_{1;i,j}(u),E_{2;h,k}(v)],E_{2;f,g}(w)\big]\\
=&(v-w)(u-w)(-1)^{\pa{j}_2}\big[ [E_{1;i,q}(u) , E_{2;q,k}(v)] , E_{2;f,g}(w) \big]\\
&-(v-w)(u-w)(-1)^{\pa{j}_2}\big[ [E_{1;i,q}(v) , E_{2;q,k}(v)] , E_{2;f,g}(w) \big]\\
&+(v-w)(u-w)(-1)^{\pa{j}_2}\big[ E_{1,3;i,k}(v) , E_{2;f,g}(w) \big]\\
&-(v-w)(u-w)(-1)^{\pa{j}_2}\big[ E_{1,3;i,k}(u) , E_{2;f,g}(w) \big]\\
=&(v-w)(u-w)(-1)^{\pa{j}_2}\big[ [E_{1;i,q}(u) , E_{2;q,k}(v)] , E_{2;f,g}(w) \big]\\
&-(v-w)(u-w)(-1)^{\pa{j}_2}\big[ [E_{1;i,q}(v) , E_{2;q,k}(v)] , E_{2;f,g}(w) \big]\\
&+(v-w)(u-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{i}_1\pa{f}_2+\pa{i}_1\pa{k}_3+\pa{f}_2\pa{k}_3}
E_{2;f,k}(w)\big[ E_{1;i,\ell}(v) , E_{2;\ell,g}(w) \big]\\
&-(v-w)(u-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{i}_1\pa{f}_2+\pa{i}_1\pa{k}_3+\pa{f}_2\pa{k}_3}E_{2;f,k}(w)\big[ E_{1;i,\ell}(u) , E_{2;\ell,g}(w) \big]\\
=&(u-w)(v-w)(-1)^{\pa{j}_2}\big\lbrace E_{1;i,q}(u)E_{2;q,k}(v)E_{2;f,g}(w)\\
&\qquad\qquad\qquad-(-1)^{(\pa{f}_2+\pa{g}_3)(\pa{i}_1+\pa{k}_3)}E_{2;f,g}(w)E_{1;i,q}(u)E_{2;q,k}(v)\big\rbrace\\
&-(u-w)(v-w)(-1)^{\pa{j}_2}\big\lbrace E_{1;i,q}(u)E_{2;q,k}(v)E_{2;f,g}(w) \\
&\qquad\qquad\qquad-(-1)^{(\pa{f}_2+\pa{g}_3)(\pa{i}_1+\pa{k}_3)}E_{2;f,g}(w)E_{1;i,q}(u)E_{2;q,k}(v)\big\rbrace\\
&+(v-w)(u-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{i}_1\pa{f}_2+\pa{i}_1\pa{k}_3+\pa{f}_2\pa{k}_3}
E_{2;f,k}(w)\big[ E_{1;i,\ell}(v) , E_{2;\ell,g}(w) \big]\\
&-(v-w)(u-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{i}_1\pa{f}_2+\pa{i}_1\pa{k}_3+\pa{f}_2\pa{k}_3}E_{2;f,k}(w)\big[ E_{1;i,\ell}(u) , E_{2;\ell,g}(w) \big]\\
=&(u-w)(v-w)(-1)^{\pa{j}_2} E_{1;i,q}(u)\big[E_{2;q,k}(v),E_{2;f,g}(w)\big]\\
&-(u-w)(v-w)(-1)^{\pa{j}_2} E_{1;i,q}(v)\big[E_{2;q,k}(v),E_{2;f,g}(w)\big]\\
&-(u-w)(v-w)(-1)^{\pa{j}_2+(\pa{f}_2+\pa{g}_3)(\pa{i}_1+\pa{k}_3)} \big[E_{2;f,g}(w),E_{1;i,q}(u)\big]E_{2;q,k}(v)\\
&+(u-w)(v-w)(-1)^{\pa{j}_2+(\pa{f}_2+\pa{g}_3)(\pa{i}_1+\pa{k}_3)} \big[E_{2;f,g}(w),E_{1;i,q}(v)\big]E_{2;q,k}(v)\\
&+(u-w)(v-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{g}_3\pa{f}_2+\pa{g}_3\pa{k}_3+\pa{f}_2\pa{k}_3}\big[E_{1;i,\ell}(v),E_{2;\ell,g}(w)\big]E_{2;f,k}(w)\\
&-(u-w)(v-w)(-1)^{\pa{j}_2+\pa{\ell}_2+\pa{g}_3\pa{f}_2+\pa{g}_3\pa{k}_3+\pa{f}_2\pa{k}_3}\big[E_{1;i,\ell}(u),E_{2;\ell,g}(w)\big]E_{2;f,k}(w).
\end{align*}}
We use (\ref{p518}) and Lemma \ref{3be} to compute these brackets, then (\ref{3eeec1}) equals to
\begin{align*}
&\quad \varepsilon(u-w)E_{1;i,q}(u)\big(E_{2;q,g}(v)-E_{2;q,g}(w)\big)\big(E_{2;f,k}(v)-E_{2;f,k}(w)\big)\\
\,&-\varepsilon(u-w) E_{1;i,q}(v)\big(E_{2;q,g}(v)-E_{2;q,g}(w)\big)\big(E_{2;f,k}(v)-E_{2;f,k}(w)\big)\\
\,&+\varepsilon(v-w) \big\lbrace \big(E_{1;i,q}(u)-E_{1;i,q}(w)\big)E_{2;q,g}(w)+E_{1,3;i,g}(w)-E_{1,3;i,g}(u)\big\rbrace E_{2;f,k}(v)\\
\,&-\varepsilon(u-w) \big\lbrace \big(E_{1;i,q}(v)-E_{1;i,q}(w)\big)E_{2;q,g}(w)+E_{1,3;i,g}(w)-E_{1,3;i,g}(v)\big\rbrace E_{2;q,k}(v)\\
\,&+\varepsilon(u-w)\big\lbrace \big(E_{1;i,q}(v)-E_{1;i,q}(w)\big)E_{2;q,g}(w)+E_{1,3;i,g}(w)-E_{1,3;i,g}(v)\big\rbrace E_{2;f,k}(w)\\
\,&-\varepsilon(v-w)\big\lbrace \big(E_{1;i,q}(u)-E_{1;i,q}(w)\big)E_{2;q,g}(w)+E_{1,3;i,g}(w)-E_{1,3;i,g}(u)\big\rbrace E_{2;f,k}(w),
\end{align*}
where the $\varepsilon$ is a sign factor given by $\varepsilon=(-1)^{\pa{j}_2+\pa{g}_3\pa{f}_2+\pa{g}_3\pa{k}_3+\pa{f}_2\pa{k}_3}$, and the index $q$ is summed over $1,\ldots,\mu_2$.
Opening the parentheses of the above identity, one may check that the resulting expression is indeed symmetric in $v$ and $w$. Therefore, (\ref{3eeec1}) is symmetric in $v$ and $w$ and hence (\ref{63c}) is established.
\end{proof}
Suppose now that $\mu=(\mu_1,\ldots,\mu_n)$ with $n\geq 4$. The next lemma is a generalization of \cite[Lemma 5]{Go} and \cite[Lemma 7.2]{Pe1}, in which the results were proved only for one specific index.
Here we show that they in fact hold {\em everywhere} and we require some of them to obtain the desired defining relations; see (\ref{p715}), (\ref{p716}).
\begin{lemma}\label{extra}
Associated to $\mu=(\mu_1,\mu_2,\ldots\mu_n)$ with $n\geq 4$, we have the following identities in $Y_{\mu}$, called the super Serre relations:
\begin{equation}\label{eeee}
\big[\,[E_{a;i,f_1}^{(r)},E_{a+1;f_2,j}^{(1)}],[E_{a+1;h,g_1}^{(1)},E_{a+2;g_2,k}^{(s)}]\,\big]=0,
\end{equation}
\begin{equation}\label{ffff}
\big[\,[F_{a;f_1,i}^{(r)},F_{a+1;j,f_2}^{(1)}],[F_{a+1;g_1,h}^{(1)},F_{a+2;k,g_2}^{(s)}]\,\big]=0,\\[1mm]
\end{equation}
for all $1\leq a\leq n-3$ and all $1\leq i\leq \mu_a$, $1\leq f_1,f_2,h\leq \mu_{a+1}$, $1\leq g_1,g_2,j\leq \mu_{a+2}$, $1\leq k\leq \mu_{a+3}$.
\end{lemma}
\begin{proof}
It suffices to prove the following special case of (\ref{eeee}) when $n=4$, while the general cases and (\ref{ffff}) can be achieved by applying the maps $\psi$ and $\zeta_{M|N}$:
\begin{equation}\label{e4e}
\big[\,[E_{1;i,f_1}^{(r)},E_{2;f_2,j}^{(1)}]\,,\,[E_{2;h,g_1}^{(1)},E_{3;g_2,k}^{(s)}]\,\big]=0.
\end{equation}
We claim that for all $1\leq i\leq \mu_1$, $1\leq h\leq \mu_2$, $1\leq j\leq \mu_3$, $1\leq k\leq \mu_4$, we have\begin{equation}\label{exeee}
[E_{1,3;i,j}(u)\,,\,E_{2;h,q}(v)E_{3;q,k}(v)-E_{2,4;h,k}(v)\,]=0,
\end{equation}
where the index $q$ is summed over $1,\ldots,\mu_3$.
To prove (\ref{exeee}), we multiply the matrix equalities (\ref{T=FDE}) and (\ref{Tp=FDE}) associated to the composition $(\mu_1,\mu_2, \mu_3,\mu_4)$ and derive the following identities.
\begin{eqnarray}
E_{1,3;i,j}(u)&=&D^{\prime}_{1;i,p}(u)t_{p,\mu_1+\mu_2+j}(u),\notag\\
E_{2;h,q}(v)E_{3;q,k}(v)-E_{2,4;h,k}(v)&=&t^{\prime}_{\mu_1+h,\mu_1+\mu_2+\mu_3+r}(v)D_{4;r,k}(v)\notag,
\end{eqnarray}
for all $1\le i\le\mu_1$, $1\le j\le\mu_3$, $1\le h\le\mu_2$, $1\le k\le\mu_4$, and the indices $p$, $q$, $r$ are summed from 1 to $\mu_1$, $\mu_3$, $\mu_4$, respectively. Substituting these identities into the bracket in (\ref{exeee}) and setting a notation $n_a:=\mu_1+\mu_2+\ldots+\mu_a$ for short, we have
\begin{align*}
&[E_{1,3;i,j}(u),E_{2;h,q}(v)E_{3;q,k}(v)-E_{2,4;h,k}(v)]\\
&=[D^{\prime}_{1;i,p}(u)t_{p,n_2+j}(u),t^{\prime}_{\mu_1+h,n_3+r}(v)D_{4;r,k}(v)]\\
&=D^{\prime}_{1;i,p}(u)t_{p,n_2+j}(u)t^{\prime}_{\mu_1+h,n_3+r}(v)D_{4;r,k}(v)\\
&\qquad\qquad\qquad-(-1)^{(\pa{i}_1+\pa{j}_3)(\pa{h}_2+\pa{k}_4)}t^{\prime}_{\mu_1+h,n_3+r}(v)D_{4;r,k}(v)D^{\prime}_{1;i,p}(u)t_{p,n_2+j}(u)\\
&=D^{\prime}_{1;i,p}(u)t_{p,n_2+j}(u)t^{\prime}_{\mu_1+h,n_3+r}(v)D_{4;r,k}(v)\\
&\qquad\qquad\qquad-(-1)^{(\pa{h}_2+\pa{r}_4)(\pa{p}_1+\pa{j}_3)}D^{\prime}_{1;i,p}(u)t^{\prime}_{\mu_1+h,n_3+r}(v)t_{p,n_2+j}(u)D_{4;r,k}(v)\\
&=D^{\prime}_{1;i,p}(u)[t_{p,n_2+j}(u),
t^{\prime}_{\mu_1+h,n_3+r}(v)]D_{4;r,k}(v)=0,
\end{align*}
and (\ref{exeee}) follows.
Note that in the above computation we have used the facts that
\[
D_{1;i,j}(u)=t_{ij}(u) \qquad \text{and} \qquad D^{\prime}_{4;i,j}(u)=t^{\prime}_{n_3+i,n_3+j}(u),
\]
therefore
$[D_{1;i,j}(u),t^{\prime}_{\mu_1+h,n_3+k}(v)]=0$
and $[D^{\prime}_{4;i,j}(u),t_{h,n_2+k}(v)]=0$ \;by (\ref{usefull}).
To show (\ref{e4e}), by (\ref{61b}), we may assume that $f_1=f_2=f$ and $g_1=g_2=g$.
Computing the following bracket by (\ref{61b}), we have
\begin{align*}
(u&-v)(w-z)\big[\,[E_{1;i,f}(u),E_{2;f,j}(v)]\,,\,[E_{2;h,g}(w),E_{3;g,k}(z)]\,\big]\notag\\
&=\big[\,(-1)^{\pa{f}_2}E_{1;i,q}(u)E_{2;q,k}(v)-E_{1;i,q}(v)E_{2;q,k}(v)+E_{1,3;i,k}(v)-E_{1,3;i,k}(u),\notag\\
&\qquad (-1)^{\pa{g}_3}E_{2;h,p}(w)E_{3;p,k}(z)-E_{2;h,p}(z)E_{3;p,k}(z)+E_{2,4;h,k}(z)-E_{2,4;h,k}(w)\,\big].\notag
\end{align*}
Taking its coefficient of $u^{-r}z^{-s}v^0w^0$, we have
\[
(-1)^{\pa{f}_2+\pa{g}_3}\sum_{t=1}^{s-1}[E_{1,3;i,j}^{(r)}\,,\,E_{2;h,p}^{(s-t)}E_{3;p,k}^{(t)}-E_{2,4;h,k}^{(s)}],
\]
which equals to the coefficient of $u^{-r}z^{-s}$ in
\[
(-1)^{\pa{f}_2+\pa{g}_3}[E_{1,3;i,j}(u)\,,\,E_{2;h,p}(z)E_{3;p,k}(z)-E_{2,4;h,k}(z)],
\]
which is zero by (\ref{exeee}).
Finally, the coefficient of $u^{-r}z^{-s}v^0w^0$ in
\begin{equation*}
(u-v)(w-z)\big[\,[E_{1;i,f}(u),E_{2;f,j}(v)]\,,\,[E_{2;h,g}(w),E_{3;g,k}(z)]\,\big]
\end{equation*}
is precisely $-\big[\,[E_{1;i,f}^{(r)},E_{2;f,j}^{(1)}]\,,\,[E_{2;h,g}^{(1)},E_{3;g,k}^{(s)}]\,\big]$ and (\ref{e4e}) follows.
\end{proof}
\section{The general Case}
Recall that our goal is to obtain the defining relations of $Y_\mu(\mathfrak{s})=Y_{M|N}$ in terms of the parabolic generators $\lbrace D_{a;i,j}^{(r)}, D_{a;i,j}^{\prime'(r)}\rbrace$, $\lbrace E_{a;i,j}^{(r)}\rbrace$, and $\lbrace F_{a;i,j}^{(r)}\rbrace$ associated to an arbitrary fixed composition $\mu$ of $M+N$ and an arbitrary fixed $0^M1^N$-sequence $\mathfrak{s}$. The following proposition summarizes the results that we have established earlier.
\begin{proposition}\label{srlns}
The following relations hold in $Y_\mu(\mathfrak{s})$:
\begin{eqnarray}
\label{p701}D_{a;i,j}^{(0)}&=&\delta_{ij}\,,\\
\label{p702}\sum_{p=1}^{\mu_a}\sum_{t=0}^{r}D_{a;i,p}^{(t)}D_{a;p,j}^{\prime (r-t)}&=&\delta_{r0}\delta_{ij}\,,\\
\big[D_{a;i,j}^{(r)},D_{b;h,k}^{(s)}\big]&=&
\delta_{ab}(-1)^{\pa{i}_a\pa{j}_a+\pa{i}_a\pa{h}_a+\pa{j}_a\pa{h}_a}\times \notag\\
&&\sum_{t=0}^{min(r,s)-1}\big(D_{a;h,j}^{(t)}D_{a;i,k}^{(r+s-1-t)}-D_{a;h,j}^{(r+s-1-t)}D_{a;i,k}^{(t)}\big),
\end{eqnarray}
{\allowdisplaybreaks
\begin{multline}\label{p704}
[D_{a;i,j}^{(r)}, E_{b;h,k}^{(s)}]
=\delta_{a,b}\delta_{hj}(-1)^{\pa{h}_a\pa{j}_a}\sum_{p=1}^{\mu_a}\sum_{t=0}^{r-1} D_{a;i,p}^{(t)} E_{b;p,k}^{(r+s-1-t)}\\
-\delta_{a,b+1}(-1)^{\pa{h}_b\pa{k}_a+\pa{h}_b\pa{j}_a+\pa{j}_a\pa{k}_a} \sum_{t=0}^{r-1} D_{a;i,k}^{(t)} E_{b;h,j}^{(r+s-1-t)},
\end{multline}
\begin{multline}\label{p705}
[D_{a;i,j}^{(r)}, F_{b;h,k}^{(s)}]
=\delta_{a,b}(-1)^{\pa{i}_a\pa{j}_a+\pa{h}_{a+1}\pa{i}_a+\pa{h}_{a+1}\pa{j}_a}\sum_{p=1}^{\mu_a}\sum_{t=0}^{r-1} F_{b;h,p}^{(r+s-1-t)}D_{a;p,j}^{(t)}\\
+\delta_{a,b+1}(-1)^{\pa{h}_a\pa{k}_b+\pa{h}_a\pa{j}_a+\pa{j}_a\pa{k}_b} \sum_{t=0}^{r-1} F_{b;i,k}^{(r+s-1-t)}D_{a;h,j}^{(t)},
\end{multline}
\begin{multline}\label{p706}
[E_{a;i,j}^{(r)} , F_{b;h,k}^{(s)}]
=\delta_{a,b}(-1)^{\pa{h}_{a+1}\pa{k}_a+\pa{j}_{a+1}\pa{k}_a+\pa{h}_{a+1}\pa{j}_{a+1}+1}
\sum_{t=0}^{r+s-1} D_{a;i,k}^{\prime (r+s-1-t)} D_{a+1;h,j}^{(t)},
\end{multline}
\begin{multline}\label{p707}
[E_{a;i,j}^{(r)} , E_{a;h,k}^{(s)}]
=(-1)^{\pa{h}_{a}\pa{j}_{a+1}+\pa{j}_{a+1}\pa{k}_{a+1}+\pa{h}_{a}\pa{k}_{a+1}}\times\\
\big( \sum_{t=1}^{s-1} E_{a;i,k}^{(r+s-1-t)} E_{a;h,j}^{(t)}
-\sum_{t=1}^{r-1} E_{a;i,k}^{(r+s-1-t)} E_{a;h,j}^{(t)} \big),
\end{multline}
\begin{multline}\label{p708}
[F_{a;i,j}^{(r)} , F_{a;h,k}^{(s)}]
=(-1)^{\pa{h}_{a+1}\pa{j}_{a}+\pa{j}_{a}\pa{k}_{a}+\pa{h}_{a+1}\pa{k}_{a}}\times\\
\big( \sum_{t=1}^{r-1} F_{a;i,k}^{(r+s-1-t)} F_{a;h,j}^{(t)}
-\sum_{t=1}^{s-1} F_{a;i,k}^{(r+s-1-t)} F_{a;h,j}^{(t)} \big),
\end{multline}
\begin{equation}
\label{p709}[E_{a;i,j}^{(r+1)}, E_{a+1;h,k}^{(s)}]-[E_{a;i,j}^{(r)}, E_{a+1;h,k}^{(s+1)}]
=(-1)^{\pa{j}_{a+1}\pa{h}_{a+1}}\delta_{h,j}\sum_{q=1}^{\mu_{a+1}}E_{a;i,q}^{(r)}E_{a+1;q,k}^{(s)}\,,
\end{equation}
\begin{multline}
\label{p710}[F_{a;i,j}^{(r+1)}, F_{a+1;h,k}^{(s)}]-[F_{a;i,j}^{(r)}, F_{a+1;h,k}^{(s+1)}]=\\
(-1)^{\pa{i}_{a+1}(\pa{j}_{a}+\pa{h}_{a+2})+\pa{j}_a\pa{h}_{a+2}+1}\delta_{i,k}\sum_{q=1}^{\mu_{a+1}}F_{a+1;h,q}^{(s)}F_{a;q,j}^{(r)}\,,
\end{multline}
\begin{align}
\label{p711}&[E_{a;i,j}^{(r)}, E_{b;h,k}^{(s)}] = 0
\qquad\qquad\text{\;\;if\;\; $|b-a|>1$ \;\;or\;\; \;if\;\;$b=a+1$ and $h \neq j$},\\[3mm]
\label{p712}&[F_{a;i,j}^{(r)}, F_{b;h,k}^{(s)}] = 0
\qquad\qquad\text{\;\;if\;\; $|b-a|>1$ \;\;or\;\; \;if\;\;$b=a+1$ and $i \neq k$},\\[3mm]
\label{p713}&\big[E_{a;i,j}^{(r)},[E_{a;h,k}^{(s)},E_{b;f,g}^{(\ell)}]\big]+
\big[E_{a;i,j}^{(s)},[E_{a;h,k}^{(r)},E_{b;f,g}^{(\ell)}]\big]=0 \quad \text{if}\,\,\, |a-b|\geq 1,\\[3mm]
\label{p714}&\big[F_{a;i,j}^{(r)},[F_{a;h,k}^{(s)},F_{b;f,g}^{(\ell)}]\big]+
\big[F_{a;i,j}^{(s)},[F_{a;h,k}^{(r)},F_{b;f,g}^{(\ell)}]\big]=0 \quad \text{if}\,\,\, |a-b|\geq 1,\\[3mm]
\label{p715}&\big[\,[E_{a;i,f_1}^{(r)},E_{a+1;f_2,j}^{(1)}]\,,\,[E_{a+1;h,g_1}^{(1)},E_{a+2;g_2,k}^{(s)}]\,\big]=0 \;\;\text{when\;\;} n\geq 4\, \text{and\;\;} \pa{h}_{a+1}+\pa{j}_{a+2}=1,\\[3mm]
\label{p716}&\big[\,[F_{a;i,f_1}^{(r)},F_{a+1;f_2,j}^{(1)}]\,,\,[F_{a+1;h,g_1}^{(1)},F_{a+2;g_2,k}^{(s)}]\,\big]=0 \;\;\text{when\;\;} n\geq 4\, \text{and\;\;} \pa{j}_{a+1}+\pa{h}_{a+2}=1.
\end{align}}
If $D_{a;i,j}^{(r)}$ appears on the left-hand side of the equation, then it holds for all $1\leq i,j\leq \mu_{a}$ and all $r\geq 0$; if $E_{a;h,k}^{(s)}$ appears on the left-hand side of the equation, then it holds for all $1\leq h\leq \mu_a, 1\leq k\leq \mu_{a+1}$ and all $s\geq 1$; if $F_{a;f,g}^{(\ell)}$ appears on the left-hand side of the equation, then it holds for all $1\leq g\leq \mu_a, 1\leq f\leq \mu_{a+1}$ and all $\ell\geq 1$.
\end{proposition}
\begin{proof}
The first three relations follow from Proposition \ref{dd0}. By Proposition~\ref{n=2} and Lemma \ref{3be}--Lemma \ref{extra}, one can show that the identities hold in smaller Yangians, for example, $Y_{(\mu_2,\mu_3)}$. Then we apply the injective homomorphisms $\psi_{p_1|q_1}=\psi_{\mu_1}$ to these identities so that the corresponding identities also hold in bigger Yangians, for example, $Y_{(\mu_1,\mu_2,\mu_3)}$. Repeating this process and we may eventually deduce that all these relations, in series forms, hold in $Y_{\mu}$.
Finally, the following identity converts the relations from series form into the desired form:
\[
\dfrac{S(v)-S(u)}{u-v}=\sum_{r,s\ge 1}S^{(r+s-1)}u^{-r}v^{-s},
\]
for any formal series $S(u)=\sum_{r\ge 0}S^{(r)}u^{-r}$.
\end{proof}
Our main theorem is that the above relations are enough for a set of defining relations of $Y_\mu(\mathfrak{s})$.
\begin{theorem}\label{Pg}
Let $\mu=(\mu_1,\ldots,\mu_n)$ be a composition of $M+N$ and $\mathfrak{s}$ be a $0^M1^N$-sequence. Associated to this $\mu$ and $\mathfrak{s}$, the super Yangian $Y_\mu(\mathfrak{s})$ is generated by the parabolic generators
\begin{align*}
&\lbrace D_{a;i,j}^{(r)}, D_{a;i,j}^{\prime(r)} \,|\, 1\leq a\leq n, 1\leq i,j\leq \mu_a, r\geq 0\rbrace,\\
&\lbrace E_{a;i,j}^{(r)} \,|\, 1\leq a< n, 1\leq i\leq \mu_a, 1\leq j\leq\mu_{a+1}, r\geq 1\rbrace,\\
&\lbrace F_{a;i,j}^{(r)} \,|\, 1\leq a< n, 1\leq i\leq\mu_{a+1}, 1\leq j\leq \mu_a, r\geq 1\rbrace,
\end{align*}
subject only to the relations (\ref{p701})$-$(\ref{p716}).
\end{theorem}
The remaining part of this article is devoted to the proof of our main theorem, which is built on several technical propositions and lemmas.
Let $\widehat{Y}_{\mu}$ denote the abstract superalgebra generated by the elements and relations as in the statement of Theorem~\ref{Pg}, where the parities of the generator are given explicitly by (\ref{pad})-(\ref{paf}). We may further define all the other $E_{a,b;i,j}^{(r)}$ and $F_{b,a;i,j}^{(r)}$ in $\widehat{Y}_{\mu}$ by the relations (\ref{ter}), and it is straightforward to check that these definitions are independent of the choices of $k$ as in \cite[p.22]{BK1}. Let $\Gamma$ be the map
\[
\Gamma: \widehat{Y}_{\mu}\longrightarrow Y_{\mu}
\]
sending every element in $\widehat{Y}_{\mu}$ into the element in $Y_{\mu}$ with the same notation. By Theorem~\ref{gendef} and Proposition~\ref{srlns}, the map $\Gamma$ is a surjective superalgebra homomorphism. Therefore, it remains to prove that $\Gamma$ is injective.
The injectivity of $\Gamma$ is proved similar to the arguments in \cite{BK1, Go, Pe1}. We first find a spanning set for $\widehat{Y}_{\mu}$, and then show that the image of this spanning set under $\Gamma$ is linearly independent in $Y_{\mu}$.
Let $\widehat{Y}^0_{\mu}$ (respectively, $\widehat{Y}^+_{\mu}$, $\widehat{Y}^-_{\mu}$) denote the subalgebras of
$\widehat{Y}_{\mu}$ generated by the elements $\lbrace D_{a;i,j}^{(r)}\rbrace$ (respectively, $\lbrace E_{a,b;i,j}^{(r)}\rbrace$, $\lbrace F_{b,a;i,j}^{(r)}\rbrace$).
Define a filtration on $\widehat{Y}_{\mu}$ (on $\widehat{Y}^0_{\mu}$, $\widehat{Y}^+_{\mu}$ and $\widehat{Y}_{\mu}^-$ as well) by setting
\begin{equation*}
\text{deg}(D_{a;i,j}^{(r)})=\text{deg}(E_{a,b;i,j}^{(r)})=\text{deg}(F_{b,a;i,j}^{(r)})=r-1,\qquad\text{for all}\;\; r\ge 1,
\end{equation*}
and denote the associated graded superalgebra by $\operatorname{gr}\widehat{Y}_{\mu}$.
Let $\ovl{E}_{a,b;i,j}^{(r)}$ denote the image of $E_{a,b;i,j}^{(r)}$ in the graded superalgebra $\operatorname{gr}_{r-1}\widehat{Y}_{\mu}^+$.
\begin{lemma} The following identities hold in $\operatorname{gr}\widehat{Y}_{\mu}^+$ for all $r,s,t\geq 1$:
\begin{enumerate}
\item[(a)] \begin{equation}\label{L717}
[\ovl{E}_{a,a+1;i,j}^{(r)},\ovl{E}_{b,b+1;h,k}^{(s)}]=0,\;\text{if}\;|a-b|\ne 1,
\end{equation}
\item[(b)] \begin{equation}\label{L718}
[\ovl{E}_{a,a+1;i,j}^{(r+1)},\ovl{E}_{b,b+1;h,k}^{(s)}]=
[\ovl{E}_{a,a+1;i,j}^{(r)},\ovl{E}_{b,b+1;h,k}^{(s+1)}],\; \text{if}\; |a-b|=1,
\end{equation}
\item[(c)] \begin{equation}\label{L719}
\big[\ovl{E}_{a,a+1;i,j}^{(r)},[\ovl{E}_{a,a+1;h,k}^{(s)},\ovl{E}_{b,b+1;f,g}^{(t)}]\big]=
-\big[\ovl{E}_{a,a+1;i,j}^{(s)},[\ovl{E}_{a,a+1;h,k}^{(r)},\ovl{E}_{b,b+1;f,g}^{(t)}]\big],
\end{equation}
\; \text{if}\; $|a-b|=1$,
\item[(d)] \begin{multline}\label{L720}
\ovl{E}_{a,b;i,j}^{(r)}=(-1)^{\pa{h}_{b-1}}[\ovl{E}_{a,b-1;i,h}^{(r)},\ovl{E}_{b-1,b;h,j}^{(1)}]
=(-1)^{\pa{k}_{a+1}}[\ovl{E}_{a,a+1;i,k}^{(1)},\ovl{E}_{a+1,b;k,j}^{(r)}],
\end{multline}
for all $b>a+1$ and any $1\leq h\leq\mu_{b-1}$, $1\leq k\leq\mu_{a+1}$.
\end{enumerate}
Here (\ref{L717}) and (\ref{L718}) hold for all $1\leq i\leq \mu_a$, $1\leq j\leq \mu_{a+1}$, $1\leq h\leq \mu_b$, $1\leq k\leq \mu_{b+1}$; (\ref{L719}) holds for all $1\leq i,h\leq \mu_a$, $1\leq j,k\leq \mu_{a+1}$, $1\leq f\leq \mu_b$, $1\leq g\leq \mu_{b+1}$; (\ref{L720}) holds for all $1\leq i\leq \mu_a$, $1\leq j\leq \mu_b$.
\end{lemma}
\begin{proof}
(\ref{L717}) and (\ref{L718}) follow from (\ref{p711}) and (\ref{p709}), while (\ref{L719}) follows from (\ref{p713}). The first equality of (\ref{L720}) follows from (\ref{ter}), while the second one can be deduced from the first equality by super Jacobi identity, (\ref{L718}) and induction on $b-a$.
\end{proof}
\begin{lemma}
The following identities hold in $\operatorname{gr}\widehat{Y}_{\mu}^+$ for all $r,s\geq 1$:
\begin{enumerate}
\item[(a)]
\begin{equation}\label{LL725}
[\ovl{E}_{a,a+2;i,j}^{(r)},\ovl{E}_{a+1,a+2;h,k}^{(s)}]=0,\;\; \text{for all}\;\; 1\leq a\leq n-2,
\end{equation}
\item[(b)]
\begin{equation}\label{LL726}
[\ovl{E}_{a,a+1;i,j}^{(r)},\ovl{E}_{a,a+2;h,k}^{(s)}]=0,\;\; \text{for all}\;\; 1\leq a\leq n-2,
\end{equation}
\item[(c)]
\begin{equation}\label{L723}
[\ovl{E}_{a,a+2;i,j}^{(r)},\ovl{E}_{a+1,a+3;h,k}^{(s)}]=0,\;\; \text{for all}\;\; 1\leq a\leq n-3,
\end{equation}
\item[(d)]
\begin{equation}\label{L724}
[\ovl{E}_{a,b;i,j}^{(r)},\ovl{E}_{c,c+1;h,k}^{(s)}]=0,\;\; \text{for all}\;\; 1\leq a<c<b\leq n.
\end{equation}
\end{enumerate}
Here (\ref{LL725}) holds for all $1\leq i\leq \mu_a$, $1\leq h\leq \mu_{a+1}$, $1\leq j,k\leq \mu_{a+2}$; (\ref{LL726}) holds for all $1\leq i,h\leq \mu_a$, $1\leq j\leq \mu_{a+1}$, $1\leq j,k\leq \mu_{a+2}$; (\ref{L723}) holds for all $1\leq i\leq \mu_a$, $1\leq h\leq \mu_{a+1}$, $1\leq j\leq \mu_{a+2}$, $1\leq k\leq \mu_{a+3}$; (\ref{L724}) holds for all $1\leq i\leq \mu_a$, $1\leq j\leq \mu_b$, $1\leq h\leq \mu_c$, $1\leq k\leq \mu_{c+1}$.
\end{lemma}
\begin{proof}
Similar to the proof in \cite[Lemma 8.3]{Pe1} so we only show (c) in detail here since it is the place that we actually use the super Serre relations.
Assume first that $\pa{h}_{a+1}+\pa{j}_{a+2}=0$.
Applying (\ref{L720}) on the left-hand side of (\ref{L723}) and using the super Jacobi identity, we have
\begin{align*}
[\ovl{E}_{a,a+2;i,j}^{(r)}&,\ovl{E}_{a+1,a+3;h,k}^{(s)}]=\\
&(-1)^{\pa{h}_{a+1}+\pa{j}_{a+2}}\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}]
\,,\,[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big]\\
&=(-1)^{\pa{h}_{a+1}+\pa{j}_{a+2}}\Big\lbrace\Big[\,\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}],
\ovl{E}_{a+1,a+2;h,j}^{(1)}\big],\ovl{E}_{a+2,a+3;j,k}^{(s)}\Big]\\
&+\varepsilon\Big[\ovl{E}_{a+1,a+2;h,j}^{(1)},
\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}],\ovl{E}_{a+2,a+3;j,k}^{(s)}\big]\,\Big]\Big\rbrace,
\end{align*}
where $\varepsilon=(-1)^{(\pa{i}_{a}+\pa{h}_{a+1})(\pa{h}_{a+1}+\pa{j}_{a+2})}$.
By (\ref{L719}), the first term is zero. Using the super Jacobi identity, (\ref{L717}) and (\ref{L720}) again, we may deduce that the above equals to
\begin{align*}
&\varepsilon\Big[\ovl{E}_{a+1,a+2;h,j}^{(1)},
\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}],\ovl{E}_{a+2,a+3;j,k}^{(s)}\big]\,\Big]\\
=&\varepsilon\Big[\ovl{E}_{a+1,a+2;h,j}^{(1)},
\big[\ovl{E}_{a,a+1;i,h}^{(r)},[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big]\,\Big]+0\\
=&\varepsilon\big[\,[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a,a+1;i,h}^{(r)}],
[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big]+0\\
=&(-1)\varepsilon^2\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}]
,[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big]\\
=&(-1)^{1+\pa{h}_{a+1}+\pa{j}_{a+2}}[\ovl{E}_{a,a+2;i,j}^{(r)},\ovl{E}_{a+1,a+3;h,k}^{(s)}].
\end{align*}
By our assumption, $\pa{h}_{a+1}+\pa{j}_{a+2}=0$ and we have done.
Assume on the other hand that $\pa{h}_{a+1}+\pa{j}_{a+2}=1$. Similarly, we apply (\ref{L720}) on the left-hand side of (\ref{L723}) to obtain
\begin{align*}
[\ovl{E}_{a,a+2;i,j}^{(r)}&,\ovl{E}_{a+1,a+3;h,k}^{(s)}]=\\
&(-1)^{\pa{h}_{a+1}+\pa{j}_{a+2}}\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}]
\,,\,[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big]\\
=&-\big[\,[\ovl{E}_{a,a+1;i,h}^{(r)},\ovl{E}_{a+1,a+2;h,j}^{(1)}]
\,,\,[\ovl{E}_{a+1,a+2;h,j}^{(1)},\ovl{E}_{a+2,a+3;j,k}^{(s)}]\,\big],
\end{align*}
which is zero directly by (\ref{p715}).
\end{proof}
The following lemma, generalizing \cite[Lemma 6.7]{BK1} and \cite[(8.1)]{Pe1}, plays a crucial role in the proof of Theorem~\ref{Pg}.
\begin{lemma}\label{injeq}
For all $1\leq a\leq b\leq n$, $1\leq c\leq d\leq n$, $r,s\geq 0$ and
all $1\leq i\leq \mu_a$, $1\leq j\leq \mu_b$, $1\leq h\leq \mu_c$, $1\leq k\leq\mu_d$, we have
\begin{multline*}
[\ovl{E}_{a,b;i,j}^{(r)},\ovl{E}_{c,d;h,k}^{(s)}]=(-1)^{\pa{j}_b\pa{h}_c}\delta_{b,c}\delta_{h,j}\ovl{E}_{a,d;i,k}^{(r+s-1)}\\
-(-1)^{\pa{i}_a\pa{j}_b+\pa{i}_a\pa{h}_c+\pa{j}_b\pa{h}_c}\delta_{a,d}\delta_{i,k}\ovl{E}_{c,b;h,j}^{(r+s-1)}.
\end{multline*}
\end{lemma}
\begin{proof}
Without loss of generality, we may assume that $a\leq c$. The proof is divided into 7 cases and we discuss them one by one.
\begin{description}
\item[Case 1.] $a<b<c<d$:\\
It follows directly from (\ref{L717}) and (\ref{L720}) that the bracket in Lemma \ref{injeq} is zero.
\item[Case 2.] $a<b=c<d$:\\
By (\ref{L718}) and (\ref{L720}), we have
\begin{equation}\label{L725}
[\ovl{E}^{(r+1)}_{b-1,b;i_1,j}\,,\,\ovl{E}^{(s+1)}_{b,b+1;h,k_1}]
=[\ovl{E}^{(r+s+1)}_{b-1,b;i_1,j}\,,\,\ovl{E}^{(1)}_{b,b+1;h,k_1}]
=\delta_{h,j}(-1)^{\pa{h}_b}\ovl{E}^{(r+s+1)}_{b-1,b+1;i_1,k_1}.
\end{equation}
Note that when $h\neq j$, the bracket is zero by (\ref{61b}) and hence the term $\delta_{h,j}$ shows up.
Taking brackets on both sides of (\ref{L725}) with the elements
\[
\ovl{E}^{(1)}_{b+1,b+2;k_1,k_2}, \ovl{E}^{(1)}_{b+2,b+3;k_2,k_3}, \cdots , \ovl{E}^{(1)}_{d-1,d;k_{d-b+1},k}
\]
from the right then using (\ref{L717}), (\ref{L720}) and the super Jacobi identity, we deduce that
\begin{equation}\label{L726}
[\ovl{E}^{(r+1)}_{b-1,b;i_1,j}\,,\,\ovl{E}^{(s+1)}_{b,d;h,k}]=\delta_{h,j}(-1)^{\pa{h}_b}\ovl{E}^{(r+s+1)}_{b-1,d;i_1,k}\,.
\end{equation}
Taking brackets on both sides of (\ref{L726}) with the elements
\[
\ovl{E}^{(1)}_{b-2,b-1;i_2,i_1}, \ovl{E}^{(1)}_{b-3,b-2;i_3,i_2},\cdots,\ovl{E}^{(1)}_{a,a+1;i,i_{b-a-1}}
\]
from the left and using exactly the same method as above, we have
\[
[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{b,d;h,k}]=\delta_{h,j}(-1)^{\pa{h}_b}\ovl{E}^{(r+s-1)}_{a,d;i,k},
\;\text{as desired}.
\]
\item[Case 3.] $a<c<b=d$:\\
Using the super Jacobi identity together with (\ref{L720}) and (\ref{L724}), we have
\begin{align*}
[\ovl{E}^{(r)}_{a,b;i,j}&,\ovl{E}^{(s)}_{c,b;h,k}]
=\big[\ovl{E}^{(r)}_{a,b;i,j},(-1)^{\pa{f_1}_{c+1}}[\ovl{E}^{(1)}_{c,c+1;h,f_1},\ovl{E}^{(s)}_{c+1,b;f_1,k}]\,\big]\\
&=(-1)^{\pa{f_1}_{c+1}}\big[\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{c,c+1;h,f_1}],\ovl{E}^{(s)}_{c+1,b;f_1,k}\big]\\
&\qquad\qquad\qquad\pm(-1)^{\pa{f_1}_{c+1}}\big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c+1,b;f_1,k}]\,\big]\\
&=0\pm (-1)^{\pa{f_1}_{c+1}}\big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c+1,b;f_1,k}]\,\big]\\
&=\cdots =\pm \Big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,[\ovl{E}^{(1)}_{c+1,c+2;f_1,f_2},\ldots,
[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{b-1,b;f_{b-1-c},k}]\,\big]\cdots\Big].
\end{align*}
The bracket $[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{b-1,b;f_{b-1-c},k}]$ in the middle is zero by (\ref{L724}).
\item[Case 4.] $a<c<d<b$:\\
Using the same technique as in Case 3, we have
\begin{align*}
[\ovl{E}^{(r)}_{a,b;i,j}&,\ovl{E}^{(s)}_{c,d;h,k}]
=\big[\ovl{E}^{(r)}_{a,b;i,j},(-1)^{\pa{f_1}_{c+1}}[\ovl{E}^{(1)}_{c,c+1;h,f_1},\ovl{E}^{(s)}_{c+1,d;f_1,k}]\,\big]\\
&=(-1)^{\pa{f_1}_{c+1}}\big[\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{c,c+1;h,f_1}],\ovl{E}^{(s)}_{c+1,d;f_1,k}]\,\big]\\
&\quad\pm(-1)^{\pa{f_1}_{c+1}}\big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c+1,d;f_1,k}]\,\big]\\
&=0\pm(-1)^{\pa{f_1}_{c+1}}\big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c+1,d;f_1,k}]\,\big]\\
&=\cdots =\pm \Big[\ovl{E}^{(1)}_{c,c+1;h,f_1}\,,\,\big[\ovl{E}^{(1)}_{c+1,c+2;f_1,f_2}\,,\ldots,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{d-1,d;f_{d-1-c},k}]\,\big]\cdots\Big].
\end{align*}
Following from (\ref{L724}) again, the bracket $[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{d-1,d;f_{d-1-c},k}]$ vanishes.
\item[Case 5.] $a<c<b<d$:\\
We prove this case by induction on $d-b\geq 1$. When $d-b=1$, we have
\begin{multline*}
[\ovl{E}^{(r)}_{a,b;i,j},\ovl{E}^{(s)}_{c,b+1;h,k}]
=\big[\ovl{E}^{(r)}_{a,b;i,j},(-1)^{\pa{j}_{b+1}}[\ovl{E}^{(s)}_{c,b;h,j},\ovl{E}^{(1)}_{b,b+1;j,k}]\,\big]\\
=(-1)^{\pa{j}_{b+1}}\big[\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c,b;h,j}]\,,\,\ovl{E}^{(1)}_{b,b+1;j,k}\big]
\pm \big[\ovl{E}^{(s)}_{c,b;h,j}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{b,b+1;j,k}]\,\big].
\end{multline*}
Now the bracket in the first term is zero by Case 3, and we may rewrite the whole second term as
$\pm[\ovl{E}^{(r)}_{a,b+1;i,k},\ovl{E}^{(s)}_{c,b;h,j}]$, which is zero by Case 4.
Assume that $d-b>1$, then $d-1>b$. By (\ref{L720}), the bracket equals to
\begin{align*}
[&\ovl{E}^{(r)}_{a,b;i,j},\ovl{E}^{(s)}_{c,d;h,k}]
=\big[\ovl{E}^{(r)}_{a,b;i,j}\,,\,(-1)^{\pa{f}_{d-1}}[\ovl{E}^{(s)}_{c,d-1;h,f}\,,\,\ovl{E}^{(1)}_{d-1,d;f,k}]\,\big]\\
&=(-1)^{\pa{f}_{d-1}}\big[\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{c,d-1;h,f}]\,,\,\ovl{E}^{(1)}_{d-1,d;f,k}\big]
\pm\big[\ovl{E}^{(s)}_{c,d-1;h,f}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{d-1,d;f,k}]\,\big].
\end{align*}
The first term is zero by the induction hypothesis, while the second term is zero as well by Case 1.
\item[Case 6.] $a=c<b<d$:
\begin{align*}
[\ovl{E}^{(r)}_{a,b;i,j},\ovl{E}^{(s)}_{a,d;h,k}]
&=\big[\ovl{E}^{(r)}_{a,b;i,j}\,,\,(-1)^{\pa{f}_{a+1}}[\ovl{E}^{(1)}_{a,a+1;h,f}\,,\,\ovl{E}^{(s)}_{a+1,d;f,k}]\,\big]\\
&=(-1)^{\pa{f}_{a+1}}\big[\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]\,,\,\ovl{E}^{(s)}_{a+1,d;h,k}\big]\\
&\quad \pm\big[\ovl{E}^{(1)}_{a,a+1;h,f}\,,\,[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{a+1,d;f,k}]\,\big].
\end{align*}
Note that $[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{a+1,d;f,k}]=0$ by Case 5. Hence it suffices to show that
\begin{equation}\label{L727}
[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]=0, \qquad \text{for all}\quad b>a.
\end{equation}
We prove (\ref{L727}) by induction on $b-a\geq 1$. When $b-a=1$, it follows from (\ref{L717}). Now
assume $b-a>1$. By (\ref{L720}), we have
\begin{align*}
[\,\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}\,]
&= \big[\,(-1)^{\pa{g}_{b-1}}\,[\,\ovl{E}^{(r)}_{a,b-1;i,g}\,,\,\ovl{E}^{(1)}_{b-1,b;g,j}\,]\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}\big]\\
&=(-1)^{\pa{g}_{b-1}}\,\big[\ovl{E}^{(r)}_{a,b-1;i,g}\,,\,[\ovl{E}^{(1)}_{b-1,b;g,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]\,\big]\\
&\quad \pm\,\big[\ovl{E}^{(1)}_{b-1,b;g,j}\,,\,[\ovl{E}^{(r)}_{a,b-1;i,g}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]\,\big].
\end{align*}
Note that $[\ovl{E}^{(r)}_{a,b-1;i,g}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]=0$ by the induction hypothesis. Also by (\ref{L717}),
$[\ovl{E}^{(1)}_{b-1,b;g,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]=0$ unless $b-1=a+1$, in which case,
(\ref{L727}) becomes $[\ovl{E}^{(r)}_{a,a+2;i,j}\,,\,\ovl{E}^{(1)}_{a,a+1;h,f}]$, which is zero by (\ref{LL726}).
\item[Case 7.] $a=c<b=d$:\\
We claim that
\begin{equation}\label{L728}
[\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{a,b;h,k}]=0.
\end{equation}
If $b=a+1$, it follows directly from (\ref{L717}). If $b>a+1$, we may expand one term in the bracket of (\ref{L728}) by (\ref{L720}) to deduce that
\begin{align*}
[\,\ovl{E}^{(r)}_{a,b;i,j}\,,\,\ovl{E}^{(s)}_{a,b;h,k}\,]
&=\big[\,(-1)^{\pa{f}_{b-1}}\,[\ovl{E}^{(r)}_{a,b-1;i,f}\,,\,\ovl{E}^{(1)}_{b-1,b;f,j}]\,,\,\ovl{E}^{(s)}_{a,b;h,k}\big]\\
&=(-1)^{\pa{f}_{b-1}}\,\big[\ovl{E}^{(r)}_{a,b-1;i,f}\,,\,[\ovl{E}^{(1)}_{b-1,b;f,j}\,,\,\ovl{E}^{(s)}_{a,b;h,k}]\,\big]\\
&\quad\pm\,\big[\ovl{E}^{(1)}_{b-1,b;f,j}\,,\,[\ovl{E}^{(r)}_{a,b-1;i,f}\,,\,\ovl{E}^{(s)}_{a,b;h,k}]\,\big].
\end{align*}
Note that $[\ovl{E}^{(1)}_{b-1,b;f,j}\,,\,\ovl{E}^{(s)}_{a,b;h,k}]=0$ by Case 3 and $[\ovl{E}^{(r)}_{a,b-1;i,f}\,,\,\ovl{E}^{(s)}_{a,b;h,k}]=0$ by Case~6, which proves (\ref{L728}).
\end{description}
This completes the proof of Lemma \ref{injeq}.
\end{proof}
\begin{proposition}\label{ind1}
$\widehat{Y}_{\mu}$ is spanned as a vector superspace by supermonomials in the elements $\lbrace D_{a;i,j}^{(r)}, E_{a,b;i,j}^{(r)}, F_{b,a;i,j}^{(r)}\rbrace$ taken in a certain fixed order so that $F$'s appear before $D$'s and $D$'s appear before $E$'s.
\end{proposition}
\begin{proof}
Lemma \ref{injeq} implies that the graded algebra $\operatorname{gr}\widehat{Y}_{\mu}^+$ is spanned by supermonomials in $\lbrace\ovl{E}_{a,b;i,j}^{(r)}\rbrace$ in some fixed order and hence $\widehat{Y}_{\mu}^+$ is spanned by supermonomials in $\lbrace E_{a,b;i,j}^{(r)}\rbrace$ in some fixed order as well.
By applying the automorphism $\zeta_{M|N}$, we see that $\widehat{Y}_{\mu}^-$ is spanned by supermonomials in $\lbrace F_{a,b;i,j}^{(r)}\rbrace$ in a certain fixed order as well.
Moreover, $\operatorname{gr}\widehat{Y}^0_{\mu}$ is supercommutative by Proposition \ref{dd0}, and it follows that $\widehat{Y}^0_{\mu}$ is spanned by supermonomials in $\lbrace D_{a;i,j}^{(r)}\rbrace$ in a certain fixed order.
Finally, by the defining relations in Proposition \ref{srlns} and the argument above, we may interchange the order between those $D$'s, $E$'s and $F$'s in a supermonomial such that all the $F$'s appear before all the $D$'s and all the $D$'s appear before all the $E$'s.
As a result, the multiplication map is surjective:
\[
\operatorname{gr} \widehat{Y}^-_{\mu}\otimes \operatorname{gr}\widehat{Y}^0_{\mu}\otimes \operatorname{gr}\widehat{Y}^+_{\mu}\twoheadrightarrow \operatorname{gr}\widehat{Y}_{\mu}
\]
and our proposition is established.
\end{proof}
\begin{proposition}\label{ind2}
The images of the supermonomials in Proposition~\ref{ind1} under $\Gamma$ are linearly independent.
\end{proposition}
\begin{proof}
By Corollary~\ref{Yloop}, we may identify $\operatorname{gr} Y_{M|N}=\operatorname{gr} Y_{\mu}$ with the loop superalgebra $U(\mathfrak{gl}_{M|N}[x])$ via
\[
\operatorname{gr} _{r-1}t_{ij}^{(r)}\longmapsto (-1)^{\pa{i}}e_{ij}x^{r-1}.
\]
We consider the following composition
\[
\operatorname{gr} \widehat{Y}_{\mu}^-\otimes \operatorname{gr} \widehat{Y}_{\mu}^0\otimes \operatorname{gr} \widehat{Y}_{\mu}^+\twoheadrightarrow \operatorname{gr} \widehat{Y}_{\mu} \xrightarrow{\Gamma} \operatorname{gr} Y_{\mu}\cong U(\mathfrak{gl}_{M|N}[x]).
\]
Let $n_a:=\mu_1+\mu_2+\ldots+\mu_a$ for short. By Proposition~\ref{quasi}, the image of $\ovl{E}_{a,b;i,j}^{(r)}$ (respectively, $\ovl{D}_{a;i,j}^{(r)}$, $\ovl{F}_{b,a;i,j}^{(r)}$) under the above composition map is $(-1)^{\pa{i}_{a}}e_{n_a+i, n_b+j}x^{r-1}$ (respectively, $(-1)^{\pa{i}_{a}}e_{n_a+i,n_a+j}x^{r-1}$, $(-1)^{\pa{i}_{b}}e_{n_b+i, n_a+j}x^{r-1}$ ). By the PBW theorem for $U(\mathfrak{gl}_{M|N}[x])$, the image (under the map $\Gamma$) of the set of all supermonomials in the following set
\begin{align*}
&\qquad\big\lbrace \operatorname{gr}_{r-1}\ovl{D}_{a;i,j}^{(r)} \,|\, 1\leq a\leq n, \; 1\leq i,j\leq\mu_a, \, r\geq 1 \big\rbrace \\
&\cup\big\lbrace \operatorname{gr}_{r-1}\ovl{E}_{a,b;i,j}^{(r)} \,|\, 1\leq a<b\leq n, \; 1\leq i\leq\mu_a, 1\leq j\leq\mu_b, \, r\geq 1 \big\rbrace\\
&\cup\big\lbrace \operatorname{gr}_{r-1}\ovl{F}_{b,a;i,j}^{(r)} \,|\, 1\leq a<b\leq n, \; 1\leq i\leq\mu_b, 1\leq j\leq\mu_a, \, r\geq 1 \big\rbrace
\end{align*}
taken in a certain fixed order must be linearly independent in $\operatorname{gr} Y_{\mu}$ and hence Proposition~\ref{ind2} follows.
\end{proof}
\begin{corollary}
The homomorphism $\Gamma:\widehat{Y}_{\mu}\rightarrow Y_{\mu}$ is injective, and Theorem~\ref{Pg} follows.
\end{corollary}
\begin{proof}
We have known that $\Gamma$ is a surjective homomorphism. Now a spanning set for $\widehat{Y}_{\mu}$ is obtained by Proposition~\ref{ind1}, while the image of this spanning set under $\Gamma$ is linearly independent in $Y_{\mu}$ by Proposition~\ref{ind2}. This shows that $\Gamma$ is injective.
\end{proof}
Let $Y_\mu^0$, $Y_\mu^+$ and $Y_\mu^-$ denote the subalgebras of $Y_{\mu}$ generated by all the $D$'s, $E$'s and $F$'s, respectively. The next result follows from the proof of Proposition~\ref{ind1} and the proof of Proposition~\ref{ind2}.
\begin{corollary} We have the PBW bases for the following superalgebras.
\begin{enumerate}
\item[(1)] The set of supermonomials in $\{ D_{a;i,j}^{(r)}\}_{1\leq a\leq n, 1\leq i,j\leq \mu_a, r\geq 1}$ taken in a certain fixed order forms a basis for $Y_\mu^0$.
\item[(2)] The set of supermonomials in $\{ E_{a,b;i,j}^{(r)}\}_{1\leq a<b\leq n, 1\leq i\leq\mu_a,1\leq j\leq\mu_b, r\geq 1}$ taken in a certain fixed order forms a basis for $Y_\mu^+$.
\item[(3)] The set of supermonomials in $\{ F_{b,a;i,j}^{(r)}\}_{1\leq a<b\leq n, 1\leq i\leq \mu_b,1\leq i\leq\mu_a, r\geq 1}$ taken in a certain fixed order forms a basis for $Y_\mu^-$.
\item[(4)] The set of supermonomials in the union of the elements listed in (1), (2) and (3) taken in a certain fixed order forms a basis for $Y_{\mu}$.
\end{enumerate}
\end{corollary}
\subsection*{Acknowledgements}
The author is grateful to Weiqiang Wang and Shun-Jen Cheng for numerous discussions. This work is partially supported by MOST grant 103-2115-M-008-012-MY2 and NCTS Young Theorist Award 2015.
|
\section{Introduction}
One of the open problems of statistical mechanics is the description of isolated quantum many-body systems far from equilibrium.
Simulations of real time dynamics are quite difficult and requiring large computer resources even for the simplest nontrivial
one-dimensional systems with short range interactions, although a significant progress has been obtained in {the} last ten or so years when
modern versions of Density Matrix Renormalization Group (DMRG) \cite{White92} approaches using Matrix Product States (MPS) have been developed
\cite{Vidal03,Vidal04} - for recent reviews see \cite{Schol05,Schol11}. It has been understood that the growth of entanglement
in the system is a main obstacle in studying long-time dynamics even for initial states being close to the ground state, for which the
entanglement is typically quite low \cite{Schol11}. This growth may be {even} more restrictive for abrupt quenches and study of excited states \cite{Lacki12}.
By comparison, the study of finite temperature dynamics is much less understood. Some results for large systems were obtained assuming adiabatic evolution (with entropy conservation \cite{Ho07,Pollet08}). One possible ``quasi-exact'' approach involves MPS of a system enlarged by an auxiliary environment. By tracing out the auxiliary degrees of freedom, one obtains the density matrix
of the original system \cite{Verstraete04,Zwolak04}. This approach has been applied for dynamics of thermal systems \cite{Feiguin05,Barthel09,Feiguin11}.
Interestingly, there is an important freedom in the approach, namely that the same density matrix of the system may be obtained for different dynamics of the environment.
One may use this freedom in the wavefunction describing the system and the environment, in order to minimize the growth of system-environment entanglement during the temporal evolution.
That in turn allows for reaching longer times of evolution~\cite{Karrasch12,Karrasch13}. This approach has been used to calculate particle currents and Drude weight as well as spin correlations in XXZ spin chains.
Another promising approach is based on the combination of time-dependent DMRG in the Heisenberg picture -- that allows to obtain the real time evolution of the operator~\cite{Prosen07} -- with the density matrix obtained using imaginary time propagation. This allows for an evaluation of expectation values on a grid of temperature/time points simultaneously \cite{Pizorn14}.
We shall use in the present paper yet another approach called Minimally Entangled Typical Thermal State (METTS) approach
\cite{White09,Stoudenmire10,Metts14}. It has been argued that it can simulate finite temperature quantum systems with a computational cost comparable to ground state DMRG \cite{Stoudenmire10}. As its application to real time dynamics requires propagation of excited states for which the entanglement growth may be a serious obstacle \cite{Prosen09}, this method may be costly to implement.
A very recent study \cite{Barthel14} has shown that METTS may be quite efficient for gaped systems at low temperatures. We shall use it for studying the
dynamics of gaped Mott insulators far from equilibrium.
We will mainly consider a system composed of two one-dimensional Mott insulators at different temperatures that are glued together at a given instant of time.
Such a system can, in principle, be realized in the laboratory. A single insulator is routinely observed by placing ultra-cold
atoms in a one-dimensional deep optical lattice with the transverse directions being frozen by a tight laser confinement \cite{Esslinger04}.
The system can be split by an additional laser potential into two separate parts. One may imagine heating both parts differently and then bringing them into contact by a rapid switch-off of the separating laser.
For such a model system, we concentrate on the heat transfer assuming little direct particle current. In Section II, we discuss briefly our METTS implementation for the Bose-Hubbard Hamiltonian and show exemplary results. Section III brings a simple analytical theory for the observables assuming that the heat transport is dominated by quasiparticle motions,
described (in the low tunneling limit) by the Bogoliubov approach. Both approaches are compared in Section IV where we also discuss an alternative approach {in which} the two subsystems
are smoothly glued. We conclude in Section V.
\section{The system studied using METTS}
\subsection{METTS algorithm}
The METTS algorithm as proposed by White \cite{White09} simulates a thermal canonical ensemble with temperature $T=1/\beta$ (from now on we shall assume the convenient units with the Boltzmann constant $k_B=1$, we assume also $\hbar=1$).
Since the method is described in detail elsewhere \cite{Stoudenmire10},
we provide essentials
only. The METTS approach works by alternately applying the imaginary time evolution operator $\exp(-\beta H/2)$ and a projection measurement onto a given basis set.
This defines a random walk which samples the Hilbert space and creates a properly weighted ensemble $\mathcal{M}$ which enables to estimate thermal averages of operators in the canonical ensemble: $ \langle O \rangle = \textrm{Tr} [\exp(-\beta H) O]$ as $ \langle O \rangle = \sum\limits_{\psi \in \mathcal{M}} \langle \psi | O | \psi \rangle.$
The METTS ensemble allows also for real time evolution of the ensemble $ \langle O(t) \rangle=\sum\limits_{\psi \in \mathcal{M}} \langle \psi | O(t) | \psi \rangle=\sum\limits_{\psi \in \mathcal{M}} \langle \psi(t) | O | \psi(t) \rangle.$ Both the imaginary time and the real time evolutions are performed efficiently with the time-evolving block decimation (TEBD) algorithm \cite{Vidal03}, essentially equivalent to a time-dependent DMRG approach.
The Hamiltonian of the one-dimensional Bose-Hubbard (BH) model reads
\begin{equation}
H = \sum_i H_i= \sum\limits_{i=1}^L \left \{ -J_i\left [ a_i a_{i+1}^\dagger + h.c \right ]+ \frac{U}{2}n_i(n_i-1) \right \}
\label{eqn:BH}
\end{equation}
with $a_i$ ($a_i^\dagger$) being the standard boson annihilation (creation) operator at site $i=1,...,L$, $n_i=a_i^\dagger a_i$, $J_i=J$ is the amplitude of hopping (tunneling)
between $i$ and $i+1$ site; $J_L=0$ for open boundary conditions (OBC) considered below
while $U$ denotes the interaction strength.
We consider the insulating regime $U\gg J$ (we shall assume typically that $J/U=0.05$ deep in the Mott regime).
In the METTS approach, the projection measurement can be performed on any basis set. It is here convenient to use a Fock basis
on each lattice site, meaning that we measure the (random) number of particle on each site, an easy task
for a MPS state~new{\cite{Delande:BrightSoliton:NJP13}}. For low temperatures
considered in this work (we take a typical $\beta U$ of the order of 5),
successive evolution steps between measurements are correlated. To
avoid that, we have started the METTS algorithm from the ground state neglecting the first 500 iterations of the METTS sampling procedure. For the final ensemble, we have taken every 200th METTS vector. Such a sampling is quite time consuming. The alternative would be to change the basis in which the measurements are performed. That, however, would significantly slow down
the time evolution using TEBD, because it would break the total particle number conservation (which would of course
be restored after statistical averaging)
\cite{Vidal03}.
\subsection{System preparation}
Our aim is to study the transport in a non-equilibrium system of two insulators at different temperatures brought into contact at a given time. Two
versions of the procedure will be considered:
\begin{itemize}
\item Two uniform systems with different temperatures are merged by building the tensor product of their density matrices.
\item Preparation of a temperature-inhomogeneous system with a smooth temperature gradient across the sample,
using the canonical thermal distribution of an auxiliary Hamiltonian.
\end{itemize}
\subsubsection{Simple tensor product approach}
\label{sec:tp}
To prepare the initial situation, we have considered two lattices of the same length $M=35$ sites that are linked together forming a longer lattice of length $L=2M.$ The Hamiltonian is $H=H_L+H_R+V_{LR}$ where both $H_i$ terms take the form (\ref{eqn:BH}) and
$V_{LR}$ is the hopping term connecting the two lattices.
Let us define a nonstandard but useful convention that the two systems are linked ``symmetrically'' at $i=0$.
Thus the indices of the right hand side lattice of $M$ sites take half integer values $i=1/2,3/2,...,M/2-1/2$ while those of the left hand side lattice are
$i=-M/2+1/2,...-3/2, -1/2$. Then the coupling term between two subsystems reads $-J(a_{-1/2}a^\dagger_{1/2} +h.c.)$. We shall
assume this ``symmetric'' notation from now on.
The density matrix that describes the full system would be just the product of constituents' density matrices: $\rho=\rho_L \otimes \rho_R,$ each prepared separately under open boundary conditions at different temperatures, i.e., $\rho_i=\exp(-\beta_i H_i$) with $i=L,R$.
We compute two appropriate METTS $\mathcal{M}_L$ and $\mathcal{M}_R$ that represent the density matrices $\rho_L$ and $\rho_R.$ The METTS ensemble that represents the density matrix $\rho_L\otimes \rho_R$ is just $\mathcal{M}_L \otimes \mathcal{M}_R := \{ \psi_L \otimes \psi_R | \psi_i \in \mathcal{M}_i, i=L,R\}.$
This is not efficient computationally. For example when estimating $\langle O_i \rangle$ for $i\in\{L, R\}$ the actual average is as follows (in this case for $L$):
$\langle \psi_L | O_i | \psi_L \rangle\langle \psi_R | \psi_R \rangle = \langle \psi_L |O_i | \psi_L \rangle.$ The sum over METTS vectors from $\mathcal{M}_L \otimes \mathcal{M}_R$ will contain repetitions of each average. It will occur as many times as many vectors are in $\mathcal{M}_R.$ On the other hand, if the two METTS ensembles contain exactly the same number of vectors $V$, then the METTS ensemble $\mathcal{M}_L \otimes \mathcal{M}_R$ yields completely equivalent estimates as: $\mathcal{M}_L \oplus \mathcal{M}_R:=\{\psi_L^i\otimes\psi_R^i | i=1,\ldots,V \},$ where $\psi_{L/R}^i$ denotes the $i$-th METTS vector of $\mathcal{M}_{L/R}.$
In typical applications to the BH model, we have found that the size of $\mathcal{M}_L \otimes \mathcal{M}_R$ is several millions of METTS which is computationally prohibitive, while $\mathcal{M}_L \oplus \mathcal{M}_R$ contains only the same number of vectors as each of $\mathcal{M}_L$ and $\mathcal{M}_R.$
We have therefore used the second approach.
\begin{figure}
\includegraphics[width=8cm]{shorttimes.png}
\caption{Dynamical evolution of the variance of the particle number on each site. The initial state is chosen with an inhomogeneous temperature profile, in order to induce an heat current. The black curves are obtained for a tensor product initial state, where the left and right parts of the system are both at thermal equilibrium, but at different temperatures $\beta_L U=6,\beta_R U=8.$
The red curves correspond to "smoothly" glued initial conditions as described in the main text. For $tU=2,$ the cusp in the black curve is a remnant of boundary effects occurring when tunneling between the two subsystems is abruptly turned on at $t=0.$ The short range fluctuations on the curves
are of statistical origin, an unavoidable effect of the METTS approach. They vanish when the number of METTS states tends to infinity.}
\label{fig:fluct}
\end{figure}
The Hamiltonian governing the dynamics of the whole lattice couples the ``left'' and ``right'' parts of the system through the additional tunneling term, mentioned above, $-Ja_{-1/2}a_{1/2}^\dagger + h.c.$.
When using the tensor product of two density matrices as an initial state, the gluing region is subject to violent short-time transient phenomena
as shown in Fig.~\ref{fig:fluct}: see for example the increased variance just on the left of the gluing point at $tU=2$. They are the remnants of the OBC for the two constituent lattices. This region is also the region where temperature gradient should be expected. Note that in particular this merging method implies that the notion of "temperature" in the middle of the system is questionable. On the other hand this approach seems the simplest one. It will be referred to as the tensor product approach.
The data shown in Fig.~\ref{fig:fluct} -- as well as in all results of quasi-exact METTS simulations in the sequel of this paper -- display
site-to-site fluctuations. This is an intrinsic drawback due to the statistical description of the system state in the METTS method; it also
affects all statistical methods \textit{\`a la Monte-Carlo.} These short range fluctuations decrease when the number of METTS states increases. Note also that
they decay in the course of the temporal evolution, as clearly seen in Fig.~\ref{fig:fluct}.
They are responsible for the noisy character of some figures shown below, but they do not affect any of the
conclusion of this paper.
\subsubsection{Smooth Gluing}
\label{sec:smooth}
Another approach is possible which makes the transition region between {the} two subsystems more subtle.
We prepare the initial state in just a single step: the inhomogeneous system containing the ``left'' and ``right'' parts with different temperatures is prepared in a single canonical ensemble simulation \emph{at thermal equilibrium}.
To achieve that, we consider an auxiliary
Hamiltonian
${\cal H}=\sum_i H_i f(i),$
with the scaling function $f$ smoothly interpolating between $\beta_L$ on the left side and $\beta_R$ on the right side.
For example,
we have used:
\begin{equation} \label{glue}
f(i)= \frac{\beta_R-\beta_L}{2} \tanh(i/s)+\frac{\beta_R+\beta_L}{2}
\end{equation}
where $1\ll s\ll M$ denotes the size of the interface.
We then construct the canonical density matrix for the auxiliary
Hamiltonian at $\beta=1$.
Then, provided the correlation length in the system is much shorter than the interface size $s$ it is expected that
the reduced density matrix of the left part $\rho_L=\textrm{Tr}_{R}\exp(-{\cal H}) \approx \exp(-\beta _L H_L),$ while that for the right part reads $\textrm{Tr}_{L}\exp(-{\cal H}) \approx \exp(-\beta _R H_R).$
This agreement requires also achieving a thermodynamic limit in terms of system size.
\subsection{Observables and thermometry}
The typical observables measured in the created ensemble or during the subsequent temporal evolution is the
average number of particles per site $\av{n_i}$ as well as its variance $\Var(n_i)=\av{n_i^2}-\av{n_i}^2$. The variance at finite
temperature contains both the quantum and the thermal fluctuations and can be used as a measure of the temperature.
The advantage of this measure is that it is local, and thus well suited for the situation of interest where there
are a space-dependent temperature profile and heat currents. It is however \textit{a priori} not obvious that the local variance is a faithful measure of
the local temperature. In order to test this assumption (similar in spirit to the well-known local density approximation used for optical lattices
exposed to a slowly spatially varying trapping potential, see e.g. \cite{urba06})
we consider smoothly glued samples -- see section~\ref{sec:smooth}
with $M=35$ (at unit mean density: 70 bosons on 70 sites), $J/U=0.05$ and $\beta_L U=8, \beta_R U=12$.
The Quantum Monte Carlo (QMC) approach as implemented in ALPS~\cite{Bauer:Alps2:JSM11} allows to find the canonical ensemble density of ${\cal H}:$ $\rho\propto \exp(-{\cal H})$ and consequently to compute the
the variance $\Var(n_i)$ for a temperature inhomogeneous system. The function $f(i)$, eq.~(\ref{glue}), is the local (inverse) temperature.
In Fig.~\ref{term}, we show the local variance $\Var(n_i)$ vs. the local inverse temperature. Except for small differences related to finite size effects, all the data collapse on the same curve, regardless of the sharpness of the transition between the hot and the cold parts of the system.
\begin{figure}
\includegraphics[width=8cm]{terma.png}
\includegraphics[width=8cm]{termb.png}
\caption{(color online) Local thermometer based on the particle number variance. (a): local variance on the various sites for two
smoothly glued Mott insulators, with inverse temperatures $\beta_L U=8$ (left part) and $\beta_R U=12$ (right part). The temperature profile
is given by Eq.~(\ref{glue}), with a (smooth) interface size $s=3$ (black solid line), $s=7$ (red, dashed line) and $s=10$ (blue, dot-dashed line).
(b): local variance vs. the (inverse) local temperature, Eq.~(\ref{glue}). Except for small differences, the three curves collapse,
proving that the local variance -- with a good approximation -- depends only on the local temperature, regardless of the sharpness of the temperature profile. This proves that the local variance can be used as a reliable thermometer in a inhomogeneous system.
}
\label{term}
\end{figure}
\subsection{Real time evolution}
Real time evolution of each member of a METTS ensemble is performed using a home-made implementation of the TEBD algorithm, taking advantage of total particle number conservation \cite{Zakrzewski09}.
The ensemble consists typically of about 12000 METTS. By restricting to MPS of maximum bond dimension $\chi=150,$
we have been able to perform numerical evolution up to time $tU=120,$ (corresponding to $tJ=6$) keeping discarded weights at the level of
$10^{-4}$ at most. We have used a fourth order Trotter decomposition in time to control the
time discretization error \cite{Daley04}.
As we are able, see sections~\ref{sec:tp} and \ref{sec:smooth}, to prepare an initial state far from the thermal equilibrium,
and we can measure both its density and temperature profiles
as a function of increasing time, we can {\sl directly} monitor heat
and mass transport in the system. We can moreover study quantitatively transport properties by calculating the energy and particle
currents in the system.
We follow to some extent the approach presented in \cite{Karrasch13} for spin systems.
Let us rewrite first the Bose-Hubbard system in a symmetrized form as
$H=\sum_i h_i$ with
\begin{eqnarray} \label{version-a}
h_i&=&-\frac{J}{2}\left(a_ia_{i+1}^\dagger+a_{i+1}a_{i}^\dagger+a_ia_{i-1}^\dagger +a_{i-1}a_{i}^\dagger \right)\nonumber\\
&& + \frac{U}{2} n_i(n_i-1)
\end{eqnarray}
The currents flowing in and out of site $i$ may be defined via continuity equations. For example the energy current
may be obtained from
\begin{eqnarray}
\partial_t h_i&=& i[H,h_i]\nonumber \\
&=& i[h_{i-2},h_i]+ i[h_{i-1},h_i]+ i[h_{i+1},h_i]+i[h_{i+2},h_i] \nonumber \\
&\equiv& j^E(i-1/2)-j^E(i+1/2),
\end{eqnarray}
where $ j^E(i-1/2)$ is the current coming into site $i$ from site $i-1$.
Since the site index $i$ is in our case a
half-integer number, the current index is an integer. In particular $j^E(0)$ corresponds to a current passing through the center where the two subsystems are glued together.
A little algebra shows that
\begin{equation} \label{curr}
j^E(i-1/2)= i[h_{i-2},h_i]+ i[h_{i-1},h_i]+i[h_{i-1},h_{i+1}],\end{equation}
with the first and last terms being the consequence of our symmetrized form of $h_i$ in (\ref{version-a}) which involves a site energy plus half of the links to neighbors on both sides.
The Hamiltonian for short range interactions on a lattice may be split as a sum of site terms in several ways, leading to slightly
different expressions for the currents. For example, \cite{Karrasch13} uses an asymmetric site Hamiltonian containing a site energy and the full link to the right.
However, the different forms lead to very similar results whenever the system changes smoothly from site to site.
For our Bose-Hubbard system, the energy current operator defined in Eq.~(\ref{curr}) reads explicitly:
\begin{eqnarray} \label{jEa}
j^E(i-1/2)&=\frac{J^2}{4}\left [ {\cal K}(i-2,i)+{\cal K}(i-1,i+1)\right
\nonumber \\
&+\frac{UJ}{2} \left [ {\cal N}(i,i-1) - {\cal N}(i-1,i) \right ]
\end{eqnarray}
with ${\cal K}(i,j)\equiv i( a^\dagger_i a_j - a^\dagger_j a_i)$ and $ {\cal N}(i,j)\equiv i(a^\dagger_i n_i a_j - h.c.) $ (with $n_i= a^\dagger_i a_i$).
Similarly we may define the mass (particle) current $j^n(i=1/2)$ \footnote{Strictly speaking, we are defining a particle current. In order
to avoid any ambiguity with the quasiparticle/quasihole currents defined in Section~\ref{Sec:bogoliubov}, we prefer to use the
unambiguous words ''mass current''.} using the continuity equation for $\partial_t n_i$
\begin{eqnarray}
\partial_t n_i&=& i[H,n_i]\nonumber \\
&=& i[h_{i-1},n_i]+ i[h_{i},n_i]+i[h_{i+1},n_i] \nonumber \\
&\equiv& j^n(i-1/2)-j^n(i+1/2)
\end{eqnarray}
yielding
\begin{equation} \label{jN}
j^n(i-1/2)=J{\cal K}(i,i-1).
\end{equation}
\begin{figure}
\includegraphics[width=8cm]{merge1.png}
\caption{(color online) Dynamical evolution of the average particle number $\av{n_i}$ and average squared particle number
$\av{n_i^2}$ after two Mott insulators at different temperatures $\beta_L U=6,\beta_R U=8$ are sharply connected
at time $t=0.$ Each subsystem has initially 35 particles on 35 sites. Note the spreading
of the central perturbation observed both on the average density $\av{n_i}$ (lower, black curves in all the panels) and on the average squared density $\av{n_i^2}$ (higher, red lines in all the panels).
}
\label{merge1}
\end{figure}
Having all the observables defined, let us take a closer look at exemplary numerical data.
We consider two sharply connected Mott insulators (with $M=35$ each) at $\beta_L U=6$ and $\beta_R U=8$.
Again the system contains 70 particles in total. For short times, the variance for this run has already been shown in Fig.~\ref{fig:fluct}.
Fig.~\ref{merge1} shows the average density, ${\av{n_i}},$ and the averaged squared density, ${\av{n_i^2}},$ for longer times.
Firstly observe that the site to site fluctuations
of both quantities are larger than on the variance $\av{n_i^2}-\av{n_i}^2$ (subtraction reduces fluctuations),
compare with the lowest panel in Fig.~\ref{fig:fluct}).
Apart from statistical
fluctuations one may clearly observe the spreading of the perturbation from the center of the sample with time.
The density shows a {small but} clear excess above one on the right hand side (the cold one), this excess moves to the right with an apparently constant velocity.
This excess must be somewhere compensated by a lack of particles (the particle number is conserved):
indeed a hole moves to the left (hot side).
Similarly bumps and holes in $\av{n_i^2}$ spread out in both directions approximately linearly in time.
\begin{figure}
\includegraphics[width=8cm]{currEmer.png}
\caption{(color online) Energy currents associated with the data of Fig.~\ref{merge1}. The current initially starts flowing at the connecting point, $i=0$, between
the two samples at different temperatures. It then propagates on both sides, reaching sites $i=\pm5$ around $tU=25,$ and site $i=10$ later around $tU=50.$
}
\label{merge2}
\end{figure}
This picture is confirmed by observing the energy currents shown in Fig.~\ref{merge2}.
One expects that, when gluing two subsystems together, the effect appears first {around the gluing point}.
Then the information travels within the system modifying
various observables (including currents). We observe this effect in Fig.~\ref{merge2}. The current at the place of merging, $j^E(0)$ almost immediately picks up non zero values starting from the gluing time $t=0$.
A crucial observation is that, after some transient, the current $j^E(0)$ is constant at long time, although
the spatial profile of $\langle n_i \rangle$ and $\langle n_i^2 \rangle$ becomes smoother and smoother. This definitely rules out the
possibility of diffusive energy transport in the system. Indeed, in such a case, the current would have been proportional
to the local temperature gradient, and would decay at long times. The persistence of a finite current at long times is
a direct signature of ballistic transport.
This is confirmed by looking at the current at other sites. They fluctuate around zero at short time,
and rise only after some position-dependent delay.
The spread of the information is approximately linear in time (ballistic) as visible in Fig.~\ref{merge2} and more apparent while inspecting the
color plots in Fig.~\ref{fig:3D1}.
The dashed lines give the predicted theoretical limits -- derived in the next Section --
for the ballistic spreading of information.
\begin{figure}
\includegraphics[width=8cm]{current_energy_merge_68_mateusz.png}
\caption{(color online) Spatio-temporal development of the energy current for the data
represented in Fig.~\ref{merge1}.
Observe the ballistic spreading of the perturbation, initially born at the connecting
point, {$i=0$}, between the two Mott insulators at different temperatures. See text for further discussion.}
\label{fig:3D1}
\end{figure}
\section{Bogoliubov theory}
\label{Sec:bogoliubov}
\subsection{Reduction of the Hilbert space}
For a moderate value of the tunneling amplitude $J,$ the Bose-Hubbard ground state at integer filling
is a Mott insulator \cite{Mott}, with a non-zero gap. In order to understand heat and mass transport at finite
temperature, it is thus useful to study the properties of the ground state as well as those of the low energy
excitations. In this section, we give a complete quantitative description of these quantities,
in the limit of small tunneling $J\ll U$ and relatively
{low temperature $\beta U \gg 1.$} Note that we
do not put any restriction on the parameter $\beta J,$ so that this approach takes into account
both the quantum fluctuations induced by $J,$ the thermal fluctuations governed by $\beta,$ and mixed contributions.
We show below that this simple theory is able to quantitatively reproduce the
quasi-exact numerical results obtained for heat and mass transport in the fully out of equilibrium
Bose-Hubbard model.
In the limit of vanishing $J,$ the ground state at unit filling is simply a Fock state with exactly
one particle per site. Elementary excitations are particle-hole excitations where a second
particle is added on one site while leaving another site empty. This costs an energy $U.$
If one drops the requirement of a fixed total number of particles -- thus going from a canonical to a grand canonical
ensemble -- elementary excitations are independent particle or hole excitations. The energy cost of the
particle excitation is $U-\mu$, where $\mu$ is the chemical potential, and the
cost of a hole excitation is $\mu.$ The system is thus gaped for any value of $\mu$ in the $[0,U]$ interval.
At finite temperature, particles or holes can be created, with probabilities respectively $\mathrm{e}^{-\beta(U-\mu)}$
and $\mathrm{e}^{-\beta\mu}.$ The average density will be maintained at 1 if $\mu=U/2.$ Higher excitations (several particles
and/or holes) are negligible if $\beta U \ll 1.$
For a small but finite $J,$ the ground state is no longer a Fock state, but it remains gaped.
Using a perturbative approach in powers of $J/U,$ it is possible~\cite{Damski} to obtain accurate estimates
of all interesting quantities such as energy, average occupation number $\langle n_i \rangle$ or $\langle n_i^2 \rangle$ on site $i$.
We will here use a similar approach, restricting to lowest non-vanishing order in $J/U,$ but
including also elementary excitations. While particle/hole excitations on all sites are degenerate for $J=0,$
tunneling between sites lifts this degeneracy producing a band of quasiparticle and quasihole excitations.
At lowest order in $J/U,$ it is enough to consider single
particle and hole excitations, that is to restrict the Hilbert space on each site
to occupation numbers $n_i=0,1,2.$
This problem has been addressed in differents works,see e.g. \cite{Elstner99,Oosten01}; for a recent review with an extensive list of references, see \cite{Krutitsky15}. In our description below, following~\cite{Cucchietti07}, we
define a ''vacuum'' state as the Fock state
with one particle on each site and introduce
particle and hole creation operators on site $i:$ $c^{\dagger}_i$ and $d^{\dagger}_i$.
Their action is best illustrated by mapping the boson occupation number onto two quantum numbers $(n_c,n_d)$
representing particle and hole occupation numbers: on each site, we have $|2\rangle=|(1,0)\rangle$,
$|1\rangle=|(0,0)\rangle$ and $|0\rangle=|(0,1)\rangle$. One then has to implement that there cannot be less
than zero or more than two bosons on each site and that states with one particle and one hole
on the same site are forbidden.
This results in an effective Hamiltonian quadratic in particle/hole operators, see~\cite{Cucchietti07}
for a detailed derivation:
\begin{equation}
H_{\mathrm{eff}} = -J \sum_{\langle i,j\rangle}{\left[ 2c^{\dagger}_i c_j + d^{\dagger}_i d_j + \sqrt{2} (c_i d_j + c^{\dagger}_i d^{\dagger}_j)\right]}
+ U \sum_i{c^{\dagger}_i c_i}
\end{equation}
where $\langle i,j\rangle$ denotes a pair of neighboring sites.
When the density of excitations is small, $\langle c^{\dagger}_i c_i\rangle, \langle c^{\dagger}_i c_i\rangle \ll 1,$
the $c_i,c^{\dagger}_i,d_i,d^{\dagger}_i$ can be approximately considered as standard bosonic annihilation and creation operators.
\subsection{Bogoliubov approach}
For a finite system with periodic boundary conditions (or an infinite system), this effective Hamiltonian can be
diagonalized using the translational invariance, going to the momentum space \cite{Cucchietti07}:
\begin{equation}
c_k = \frac{1}{L} \sum_r{c_r \mathrm{e}^{-\mathrm{i} kr}}\ \ ,\ \ d_k = \frac{1}{L} \sum_r{d_r \mathrm{e}^{-\mathrm{i} kr}}
\end{equation}
where $L$ is the number of sites of the system.
The effective Hamiltonian writes:
\begin{eqnarray}
H_{\mathrm{eff}} &=-2J \sum_k{\cos k \left[ 2 c^{\dagger}_k c_k + d^{\dagger}_k d_k + \sqrt{2} (c_k d_{-k} + c^{\dagger}_k d^{\dagger}_{-k}) \right]} \nonumber\\
&+ U \sum_k{c^{\dagger}_k c_k}
\end{eqnarray}
which can be conveniently rewritten as:
\begin{eqnarray}
H_{\mathrm{eff}}&=&\sum_k[c^{\dagger}_k,-d_{-k}]
\left[
\begin{array}{cc}
U-4J\cos k & -2\sqrt{2}J\cos k \\
2\sqrt{2}J\cos k & 2J\cos k
\end{array}
\right]
\left[
\begin{array}{c}
c_k \\ d^{\dagger}_{-k}
\end{array}
\right].\nonumber\\
\label{H22}
\end{eqnarray}
This Hamiltonian can be diagonalized using a Bogoliubov transformation:
\begin{equation}
c_k = u_k B_k + v_k^* A^{\dagger}_{-k}\ \ ,\ \ d^{\dagger}_{-k}=v_k B_k + u_k^*A^{\dagger}_{-k}
\end{equation}
where $(u_k,v_k)$ is the eigenmode of the Bogoliubov-de Gennes equations:
\begin{eqnarray}
E_k^{(\mathrm{qp})}\!
\left[
\begin{array}{c}
u_k \\ v_k
\end{array}\right]\!=\!
\left[\begin{array}{cc}
U-4J\cos k & -2\sqrt{2}J\cos k \\
2\sqrt{2}J\cos k & 2J\cos k
\end{array}\right]\!
\left[
\begin{array}{c}
u_k \\ v_k
\end{array}\right]
\label{Eq:Bogoliubov_eigenmodes}
\end{eqnarray}
with eigenenergy:
\begin{equation}
E_k^{(\mathrm{qp})} = \frac{U}{2}-J \cos k + \omega_k,
\label{Eq:energy_qp}
\end{equation}
where
\begin{equation}
\omega_k = \frac{\sqrt{U^2-12JU\cos k+4J^2\cos^2 k}}{2}.
\end{equation}
With the normalization $|u_k|^2-|v_k|^2=1,$ the operators $A_k,B_k,A^{\dagger}_k,B^{\dagger}_k$ are standard
bosonic annihilation and creation operators.
The other eigenmode $(-v_k,u_k)$ is associated with eigenenergy:
\begin{equation}
E_k^{(\mathrm{qh})} = -\frac{U}{2}+J \cos k + \omega_k,
\label{Eq:energy_qh}
\end{equation}
so that the Hamiltonian writes:
\begin{equation}
H_{\mathrm{eff}} = \sum_k{\left[E_k^{(\mathrm{qp})}B^{\dagger}_k B_k + E_k^{(\mathrm{qh})}A^{\dagger}_k A_k + \omega_k + 3J \cos k - \frac{U}{2}\right]}.
\label{Eq:H2}
\end{equation}
Eqs.(\ref{Eq:energy_qp}-\ref{Eq:energy_qh}) are special cases of general formulae avaliable for arbitrary integer mean density (see e.g. \cite{Oosten01}) as reviewed by Krutisky \cite{Krutitsky15}, see his Eq.~(325).
\subsection{Ground state}
The ground state of the system is the vacuum of the $A_k,B_k$ operators and its energy is $\sum_k{(\omega_k + 3J \cos k - U/2)}.$
In a finite system with periodic boundary conditions, the $k$ values are discretized as integer multiples of $2\pi/L.$
The limit of large system is obtained with the substitution:
\begin{equation}
\frac{1}{L} \sum_{k} \to \int_{-\pi}^{+\pi} \frac{\mathrm{d} k}{2\pi}.
\end{equation}
In the limit of small $J$ we are interested in, one gets:
\begin{equation}
\omega_k \approx \frac{U}{2} - 3J\cos k - 8 \frac{J^2}{U} \cos^2 k
\label{Eq:omegak}
\end{equation}
so that the ground state energy is approximately equal to $-4J^2 L/U,$ ($-4J^2/U$ per site), in agreement with \cite{Damski}
(note that the first order in $J$ cancels out).
It is also possible to compute expectation values of simple operators in the ground state. The total number of particles is:
\begin{equation}
N = \sum_i{n_i} = \sum_i{(1+ c^{\dagger}_i c_i - d^{\dagger}_i d_i)} = L+ \sum_k{c^{\dagger}_k c_k}.
\end{equation}
It is simply expressed as:
\begin{equation}
N = L + \sum_k{(B^{\dagger}_k B_k - A^{\dagger}_k A_k)}
\label{Eq:N}
\end{equation}
so that $B^{\dagger}_k$ (resp. $A^{\dagger}_k$) appears as the creation operator for a quasiparticle (resp. quasihole) with momentum $k$.
In the ground state, one trivially recovers $\langle N\rangle = L,$ so that there is on average exactly one particle per site.
The variance of the number of particle per site is more interesting. Indeed:
\begin{equation}
N_2 = \sum_i{n_i^2} = \sum_i{(1+ c^{\dagger}_i c_i - d^{\dagger}_i d_i)^2},
\end{equation}
which gives, in the limit of low density of excitations:
\begin{equation}
N_2 = L + \sum_i{(3 c^{\dagger}_i c_i - d^{\dagger}_i d_i)} = L + \sum_k{(3 c^{\dagger}_k c_k - d^{\dagger}_k d_k)}.
\end{equation}
The physical interpretation of the +3 and -1 coefficients is rather transparent:
$c^{\dagger}_i$ creates an additional particle at site $i,$
thus $n_i^2=4,$ i.e. 3 on top of the 1 particle of the background. Similarly, $d^{\dagger}_i$ produces $n_i=0$, that
is 1 below the background.
In turn, the operator $N_2$ can be expressed as a function of the $A_k$ and $B_k$ operators:
\begin{eqnarray}
N_2 =& L + \sum_k{\left[(3|u_k|^2-|v_k|^2)B^{\dagger}_k B_k - (|u_k|^2-3|v_k|^2)A^{\dagger}_k A_k \right.} \nonumber\\
& \left.+ 2 u_k v_k A_k B_k + 2 u_k^* v_k^* A^{\dagger}_k
B^{\dagger}_k + 2 |v_k|^2 \right].
\label{Eq:N2}
\end{eqnarray}
In the limit of small $J,$ the eigenmodes of Eq.~(\ref{Eq:Bogoliubov_eigenmodes}) are:
\begin{equation}
u_k=1\ \ ,\ \ v_k=2\sqrt{2} \cos k\ J/U,
\end{equation}
which, in a more general form can be found also in \cite{Krutitsky15}, see Eq.~(326) there.
The expectation values on the ground state is:
\begin{equation}
\langle N_2 \rangle = L +\sum_k{16 \cos^2 k \frac{J^2}{U^2}} = L \left(1+\frac{8 J^2}{U^2}\right),
\label{Eq:N2_T0}
\end{equation}
recovering the fact that the variance of the number of particle per site is $8 J^2/U^2$ \cite{Damski}.
\subsection{Thermal excitations}
\label{Sec:bogo_thermal}
We now turn to thermal excitations in the system. One can
use either the canonical or the grand canonical ensemble, all expectation values are expected to be equal in the two ensembles in
the large $L$ limit. It turns out that the grand canonical ensemble is more convenient for calculations.
We thus have to consider configurations with a weight scaling like $\mathrm{e}^{-\beta(H_{\mathrm{eff}}-\mu N)}.$
The operator $H_{\mathrm{eff}}-\mu N$ is readily obtained from Eqs.~(\ref{Eq:H2}),(\ref{Eq:N}):
\begin{eqnarray}
H_{\mathrm{eff}} - \mu N &= \sum_k{\left[(E_k^{(\mathrm{qp})}-\mu)B^{\dagger}_k B_k + (E_k^{(\mathrm{qh})}+\mu) A^{\dagger}_k A_k \right.}\nonumber\\
&\left.+ \omega_k + 3J \cos k - \frac{U}{2}\right].
\label{Eq:H2_GC}
\end{eqnarray}
The calculation of expectation values at the thermal equilibrium is thus straightforward. One obtains
$\langle B_k\rangle = \langle B^{\dagger}_k\rangle= \langle A_k\rangle= \langle A^{\dagger}_k\rangle = 0$ and:
\begin{equation}
\langle B^{\dagger}_k B_k \rangle = \frac{\textrm{Tr}\left[B^{\dagger}_k B_k \mathrm{e}^{-\beta(E_k^{(\mathrm{qp})}-\mu)B^{\dagger}_k B_k}\right]}{\textrm{Tr}\left[\mathrm{e}^{-\beta(E_k^{(\mathrm{qp})}-\mu)B^{\dagger}_k B_k}\right]},
\end{equation}
which, in the limit of low density of excitations, reduces to:
\begin{equation}
\langle B^{\dagger}_k B_k \rangle = \mathrm{e}^{-\beta(E_k^{(\mathrm{qp})}-\mu)}.
\label{Eq:NB}
\end{equation}
Similarly:
\begin{equation}
\langle A^{\dagger}_k A_k \rangle = \mathrm{e}^{-\beta(E_k^{(\mathrm{qh})}+\mu)}.
\label{Eq:NA}
\end{equation}
We thus obtain for the average number of particles:
\begin{equation}
\langle N \rangle = L +\sum_k{\left[\mathrm{e}^{-\beta(E_k^{(\mathrm{qp})}-\mu)} - \mathrm{e}^{-\beta(E_k^{(\mathrm{qh})}+\mu)} \right]}.
\end{equation}
In the limit of small $J/U,$ one can use the approximate expression (\ref{Eq:omegak}) at first order and obtain:
\begin{equation}
\langle N \rangle = L +\sum_k{\left[\mathrm{e}^{-\beta(U-4J\cos k-\mu)} - \mathrm{e}^{-\beta(\mu-2J\cos k)} \right]}.
\end{equation}
In the continuous limit of large $L,$ the sum over $k$ becomes an integral giving:
\begin{equation}
\frac{\av{N}}{L} = 1 + \mathrm{e}^{-\beta(U-\mu)} {\cal I}_0(4\beta J) - \mathrm{e}^{-\beta\mu} {\cal I}_0(2\beta J),
\end{equation}
where ${\cal I}_0$ is the modified Bessel function.
The physical interpretation of these equations is simple. At equilibrium, the system has a finite
density of quasiparticle $n^{\mathrm{(qp)}}(k)$ and quasihole $n^{\mathrm{(qh)}}(k)$ excitations, depending on the momentum $k$ and simply given
by Eqs.~(\ref{Eq:NB},\ref{Eq:NA}) which give in the small $J/U$ limit:
\begin{eqnarray}
n^{\mathrm{(qp)}}(k) & = & \mathrm{e}^{-\beta(U-4J\cos k-\mu)} \nonumber\\
n^{\mathrm{(qh)}}(k) & = & \mathrm{e}^{-\beta(\mu-2J\cos k)},
\label{Eq:nqpqh}
\end{eqnarray}
so that one simply has:
\begin{equation}
\frac{\av{N}}{L} = 1+ \int_{-\pi}^{+\pi}{[n^{\mathrm{(qp)}}(k) - n^{\mathrm{(qh)}}(k)]\ \mathrm{d} k/2\pi}.
\label{Eq:avN}
\end{equation}
The chemical potential $\mu$ that ensures unit filling is thus obtained from the $\langle N \rangle = L$ constraint:
\begin{equation}
\mu = \frac{U}{2} + \frac{1}{2\beta} \ln \frac{{\cal I}_0(2\beta J)}{{\cal I}_0(4\beta J)}.
\label{Eq:mu_eq}
\end{equation}
Similarly, one can easily compute the variance of the local occupation number, relying on Eqs.~(\ref{Eq:N2}),(\ref{Eq:NB}),(\ref{Eq:NA}):
At lowest order in $J/U,$ we obtain:
\begin{equation}
\frac{\av{N_2}}{L} = 1 + 8 \frac{J^2}{U^2} + 3 \mathrm{e}^{-\beta(U-\mu)} {\cal I}_0(4\beta J) - \mathrm{e}^{-\beta\mu} {\cal I}_0(2\beta J),
\end{equation}
which, for the value of $\mu$ at unit average filling, gives:
\begin{equation}
\frac{\av{N_2}}{L} = 1 + 8 \frac{J^2}{U^2} + 2 \mathrm{e}^{-\beta U/2} \sqrt{{\cal I}_0(4\beta J) {\cal I}_0(2\beta J)}.
\label{Eq:N2_final}
\end{equation}
In the limit of zero-temperature, one recovers Eq.~(\ref{Eq:N2_T0}) with only quantum fluctuations due to $J.$
In the limit of vanishing $J,$ the modified Bessel functions tend to unity, and one has purely thermal fluctuations.
In the presence of both quantum and thermal fluctuations, the variance is not perfectly additive,
because of crossed thermal-quantum effects.
Again, the interpretation in terms of density of quasiparticles and quasiholes is simple as:
\begin{equation}
\frac{\av{N_2}}{L} = 1 + 8 \frac{J^2}{U^2} + \int_{-\pi}^{+\pi}{[3n^{\mathrm{(qp)}}(k) - n^{\mathrm{(qh)}}(k)]\ \frac{\mathrm{d} k}{2\pi}}.
\label{Eq:avN2}
\end{equation}
The variance is finally obtained by combining this equation with Eq.~(\ref{Eq:avN}):
\begin{eqnarray}
\Var{(n_i)} & = & \av{n_i^2} - \av{n_i}^2 \nonumber \\
& = & 8 \frac{J^2}{U^2} + \int_{-\pi}^{+\pi}{[n^{\mathrm{(qp)}}(k) + n^{\mathrm{(qh)}}(k)]\ \frac{\mathrm{d} k}{2\pi}}.
\end{eqnarray}
which shows that quasiparticles and quasiholes equally contribute to the increase of the variance.
Other expectation values can be computed as well. For example, the energy density is the expectation value of
the operator $H_i,$ Eq.~(\ref{eqn:BH}), and is given by:
\begin{equation}
\av{H_i} = -4 \frac{J^2}{U} + U \int_{-\pi}^{+\pi}{n^{\mathrm{(qp)}}(k)\ \frac{\mathrm{d} k}{2\pi}}.
\end{equation}
It must be emphasized that all these results are valid only when the density of excitations is low, $n^{\mathrm{(qp)}}(k),n^{\mathrm{(qh)}}(k)\ll 1,$ i.e.
when both $\av{n_i}$ and $\av{n_i^2}$ are close to unity.
We have checked using DMRG and Quantum Monte Carlo numerical simulations with ALPS~\cite{Bauer:Alps2:JSM11} the validity
of Eq.~(\ref{Eq:N2_final}), both in the grand canonical ensemble and in the canonical ensemble for large systems.
For example, for $J=0.05U,$ the prediction is that the variance of the occupation number at zero temperature
should be 0.02 per site, while the numerical result is 0.0198453.
The situation is a bit more complicated for thermal excitations, for two independent reasons:
\begin{itemize}
\item Although $J/U=0.05$ seems a rather small value,
the numerical prefactors in e.g. Eq.~(\ref{Eq:omegak}) are such that one has to
compare $3J$ with $U/2.$ For $J/U>3/2-\sqrt{2}\approx 0.086,$
the Bogoliubov eigenmodes become unstable and
the whole perturbative approach breaks down. Thus, even at $J/U=0.05,$ significant corrections are expected for the thermal fluctuations.
For $\beta U=10,$ we find the variance of the occupation number to be 0.0429 while the prediction is 0.0385, a 10\% difference.
At smaller $J/U,$ the relative difference is smaller.
\item Finite size effects. We observed that the variance obtained using the canonical ensemble in a system of size $L$ has a rather
strong dependence on $L,$ much larger than for the grand canonical ensemble, at least for small $J$ and low temperature. We believe that, when
working in the canonical ensemble, fluctuations in the local occupation number are due to particles or holes jumping from other sites.
If the number of sites is very large, they can provide at no cost the additional particles or holes. In contrast, if the number of sites
is too small to provide these additional particles, the fluctuations and thus the variance will be reduced. As a rule of thumb,
this effect is important when the product of the number of sites $L$ by the variance of the occupation number is not much larger than unity.
For $L=100, J/U=0.05, \beta U=10,$ the variance is reduced to 0.0327.
\end{itemize}
In order to take into account these effects, a simple method is to slightly reduce the density of the quasiparticle
and quasihole excitations, Eq.~(\ref{Eq:nqpqh}), by multiplying them by a constant factor (independent of $k$)
slightly smaller than unity, which reduces the thermal fluctuations, leaving the quantum fluctuations unaffected.
This reduction factor is chosen to reproduce the initial number variance (at {$t=0$}) in each sub-sample: for the ''sharp gluing'' scenario where the right and left sub-samples are initially not connected (see e.g. Fig.~{\ref{fig:n_n2_time_60}}), two different reduction factors are used. For the ''smooth gluing'' scenario with an initial inhomogeneous temperature profile (see Fig.~{\ref{varmetts}}), thermal fluctuations can be globally provided by both sub-samples and a single reduction factor is consequently used.
\subsection{Transport in non-equilibrium systems}
We now consider the situation of systems, not at the thermal equilibrium, where we want to compute the
temporal evolution of expectation values of local operators as well as the currents flowing inside the system.
This is a very complicated problem in general; we will restrict to the Bose-Hubbard model within the simplified assumption discussed above, namely when only local occupation numbers 0, 1 and 2 are allowed on each site.
If we further assume that the system is everywhere close to the Mott insulator state with unit filling, so that the
density of excitations is small, we can use the previous Bogoliubov description in terms of quasiparticles and quasiholes. Note that this \emph{does not} require the system to be locally at a thermal equilibrium.
The Bogoliubov theory described in the previous sections explicitly uses the translational invariance in order to obtain uncoupled elementary excitations with a well defined momentum. This is well suited to describe how excitations propagate, as it boils down to the dispersion relation of the excitations, Eqs.~(\ref{Eq:energy_qp},\ref{Eq:energy_qh}). For a system where translational invariance is broken by say a temperature gradient, this is less convenient, as one needs to build ''wavepackets'' coherently superimposing excitations with various $k.$
A mixed position-momentum representation is in such a case more convenient, similar to the Wigner phase-space
representation used to describe an ordinary quantum particle obeying the Schr\"odinger equation~\cite{Wigner}.
The temporal evolution of the Wigner representation is well approximated by the classical dynamics in the semiclassical
limit where the wavelength is much shorter than the typical size over which the Wigner function varies.
In our case, the typical wavelength of Bogoliubov excitations is $2\pi/k,$ of the order of the lattice spacing.
Thus, provided we consider a spatially smooth profile of temperature/excitations, we can use a classical
description of the system in terms of densities of quasiparticles/quasiholes $n^{\mathrm{(qp)}}/n^{\mathrm{(qh)}}$ depending on both the position $x$ (a continuous variable in this approximation) and the momentum $k,$ which obey the classical Liouville equations of evolution under the Hamiltonian $H_{\mathrm{eff}}$:
\begin{eqnarray}
\frac{\partial n^{\mathrm{(qp)}}(x,k,t)}{\partial t} = - \frac{\mathrm{d} E_k^{(\mathrm{qp})}}{\mathrm{d} k}\ \frac{\partial n^{\mathrm{(qp)}}(x,k,t)}{\partial x} \nonumber\\
\frac{\partial n^{\mathrm{(qh)}}(x,k,t)}{\partial t} = - \frac{\mathrm{d} E_k^{(\mathrm{qh})}}{\mathrm{d} k}\ \frac{\partial n^{\mathrm{(qh)}}(x,k,t)}{\partial x}
\label{Eq:evolution_liouville}
\end{eqnarray}
As various $k$ components do not interact, the evolution is straightforward:
\begin{eqnarray}
n^{\mathrm{(qp)}}(x,k,t)& = &n^{\mathrm{(qp)}}\left
(x-v^{\mathrm{(qp)}}_k t,k,0\right)\nonumber\\
n^{\mathrm{(qh)}}(x,k,t)& = &n^{\mathrm{(qh)}}\left(x-v^{\mathrm{(qh)}}_k t,k,0\right)
\label{Eq:propagation}
\end{eqnarray}
where the velocities of quasiparticles and quasiholes are given by:
\begin{eqnarray}
v^{\mathrm{(qp)}}_k& = &\frac{\mathrm{d} E_k^{(\mathrm{qp})}}{\mathrm{d} k} \nonumber\\
v^{\mathrm{(qh)}}_k& = &\frac{\mathrm{d} E_k^{(\mathrm{qh})}}{\mathrm{d} k}.
\end{eqnarray}
In the limit of small $J,$ they are{, see Eqs.~(\ref{Eq:energy_qp}) and (\ref{Eq:energy_qh})}:
\begin{eqnarray}
v^{\mathrm{(qp)}}_k& = &4J \sin k \nonumber\\
v^{\mathrm{(qh)}}_k& = &2J \sin k,
\end{eqnarray}
which means that quasiparticles propagate twice faster than quasiholes.
A key and non-trivial point of this approach is that it predicts that quasiparticle/quasiholes excitations propagate \emph{ballistically},
as suggested by the quasi-exact numerical results using METTS presented above.
Eqs.~(\ref{Eq:propagation}) make it trivial to compute the total density of quasiparticles and quasiholes at a given
position:
\begin{eqnarray}
n^{\mathrm{(qp)}}(x,t)& = & \int_{-\pi}^{+\pi}\frac{\mathrm{d} k}{2\pi}\ n^{\mathrm{(qp)}}\left(x-v^{\mathrm{(qp)}}_kt,k,0\right)\nonumber\\
n^{\mathrm{(qh)}}(x,t)& = & \int_{-\pi}^{+\pi} \frac{\mathrm{d} k}{2\pi}\ n^{\mathrm{(qh)}}\left(x-v^{\mathrm{(qh)}}_kt,k,0\right)
\label{Eq:nqpqht}
\end{eqnarray}
From these formula, one can deduce the expectation values of any local observable, following the derivation in section~\ref{Sec:bogo_thermal}:
\begin{eqnarray}
\av{n(x)} & = & 1 + n^{\mathrm{(qp)}}(x,t) - n^{\mathrm{(qh)}}(x,t) \nonumber\\
\av{n^2(x)} & = & 1 + 8\frac{J^2}{U^2} + 3n^{\mathrm{(qp)}}(x,t) - n^{\mathrm{(qh)}}(x,t) \nonumber\\
\Var{(n(x))} & = & 8\frac{J^2}{U^2} + n^{\mathrm{(qp)}}(x,t) + n^{\mathrm{(qh)}}(x,t) \nonumber\\
\av{H(x)}/U & = & -4\frac{J^2}{U^2} + n^{\mathrm{(qp)}}(x,t)
\label{Eq:occupation_prediction}
\end{eqnarray}
Although the quasiparticle/quasiholes densities are not in principle directly observable, these equations show
that it is enough to measure two independent quantities, for example the average local occupation number and its variance to extract both the quasiparticle and the quasihole densities.
\bigskip
The Bogoliubov approach also makes it possible to compute the currents. From Eq.~(\ref{Eq:evolution_liouville}), it follows that one can define
quasiparticle/quasihole currents:
\begin{eqnarray}
j^{\mathrm{(qp)}}(x,k,t) & = & v^{\mathrm{(qp)}}_k \ n^{\mathrm{(qp)}}(x,k,t) \nonumber\\
j^{\mathrm{(qh)}}(x,k,t) & = & v^{\mathrm{(qh)}}_k \ n^{\mathrm{(qh)}}(x,k,t)
\end{eqnarray}
which automatically satisfy the continuity equation $\partial n/\partial t + \partial j/\partial x = 0.$
The current at a given position is simply summed over all momenta contributions:
\begin{eqnarray}
j^{\mathrm{(qp)}}(x,t) & \equiv & \int_{-\pi}^{+\pi}{v^{\mathrm{(qp)}}_k\ n^{\mathrm{(qp)}}(x,k,t)\ \frac{\mathrm{d} k}{2\pi}} \nonumber\\
j^{\mathrm{(qh)}}(x,t) & \equiv & \int_{-\pi}^{+\pi}{v^{\mathrm{(qh)}}_k\ n^{\mathrm{(qh)}}(x,k,t)\ \frac{\mathrm{d} k}{2\pi}}
\end{eqnarray}
Following Eqs.~(\ref{Eq:occupation_prediction}), the mass and energy currents are thus given by:
\begin{eqnarray}
\label{eqcur}
j^n(x,t) & = & j^{\mathrm{(qp)}}(x,t) - j^{\mathrm{(qh)}}(x,t)\nonumber \\
j^E(x,t)/U & = & j^{\mathrm{(qp)}}(x,t).
\end{eqnarray}
The physical interpretation of these equations is clear: in the small $J$ limit, the dominant contribution
to energy is the interaction brought by sites with double occupation, associated with quasiparticle excitations.
In contrast, a quasihole does not lead to a change of local energy and thus does not contribute
to the energy current. As expected, quasiparticles and quasiholes both contribute to the mass (density) current,
with opposite signs.
The previous set of equations describe the ballistic transport of quasiparticles and quasiholes
in the system. In general, the momentum distribution of quasiparticles and quasiholes at a given position is not
given by a thermal distribution, Eq.~(\ref{Eq:nqpqh}). This implies that our description goes beyond a local
thermal equilibrium.
Even if the initial state at $t=0$ is in a local thermal equilibrium described by Eq.~(\ref{Eq:nqpqh}), for
space dependent (inverse) temperature $\beta(x)$ and chemical potential $\mu(x)$, this property is lost during time
evolution.
While the integrals involved in Eqs.~(\ref{Eq:nqpqht}) have a trivial structure and are easily
numerically computed for arbitrary initial distributions, it is in general difficult to perform the integrals
analytically. There are however simple cases where it is possible.
Let us first consider the situation where two half blocks, each at thermal equilibrium with unit filling, but with different temperatures $(\beta_L$,$\beta_R)$ (and consequently
different chemical potentials given by Eq.~(\ref{Eq:mu_eq})) are connected at time $t=0.$ Taking $x=0$ as the connection point, this implies that
the initial quasiparticle/quasihole distributions are given by:
\begin{eqnarray}
n^{\mathrm{(qp)}}(x,k,0) = & \mathrm{e}^{-\beta_L(U-4J\cos k-\mu_L)} \ \ & \mathrm{for}\ x<0,\nonumber\\
n^{\mathrm{(qp)}}(x,k,0) = & \mathrm{e}^{-\beta_R(U-4J\cos k-\mu_R)} \ \ & \mathrm{for}\ x>0,\nonumber\\
n^{\mathrm{(qh)}}(x,k,0) = & \mathrm{e}^{-\beta_L(\mu_L-2J\cos k)} \ \ & \mathrm{for}\ x<0,\nonumber\\
n^{\mathrm{(qh)}}(x,k,0) = & \mathrm{e}^{-\beta_R(\mu_R-2J\cos k)} \ \ & \mathrm{for}\ x>0.
\end{eqnarray}
The temporal evolution of each $k$ component is trivial: it consists of a step function moving at a velocity
$v^{\mathrm{(qp)}}_k$ (resp. $v^{\mathrm{(qh)}}_k$) for quasiparticles (resp. quasiholes).
In effect the densities have a very simple geometrical structure: they are ''smoothed'' steps connecting
the asymptotic ''left'' density (on the left side) to the asymptotic ''right'' density (on the right side).
This step function is centered around the origin, keeps the same shape at any time, being simply stretched along the $x$-axis proportionally to time. Because quasiparticles (resp. quasiholes) have a maximum velocity $4J$ (resp. $2J$), the smooth step extends only inside a ''light cone'', in a finite range $[-4Jt,4Jt]$ (resp. $[-2Jt,2Jt]$).
The situation is similar for the currents which vanish outside the light cones and keep the same
spatial structure, simply stretching linearly in time.
We could not perform analytically the integral over $k$ for the
density, but could evaluate it for the currents:
\begin{eqnarray}
j^{\mathrm{(qp)}}(x,t) & = &\frac{1}{\pi} \left[ \mathrm{e}^{-\beta_L(U-\mu_L)} \frac{\sinh \beta_Lu}{\beta_L} - \mathrm{e}^{-\beta_R(U-\mu_R)} \frac{\sinh \beta_Ru}{\beta_R}\right ]\nonumber\\
j^{\mathrm{(qh)}}(x,t) & = &\frac{1}{\pi} \left[ \mathrm{e}^{-\beta_L\mu_L} \frac{\sinh \beta_Lv}{\beta_L} - \mathrm{e}^{-\beta_R\mu_R} \frac{\sinh \beta_Rv}{\beta_R}\right ]
\end{eqnarray}
where $u=\sqrt{16J^2-x^2/t^2}, v=\sqrt{4J^2-x^2/t^2}.$ These expressions are valid inside the ''light cones''
$|x|\leq 4Jt$ (resp. $|x|\leq 2Jt$) for the quasiparticles (resp. quasiholes){; outside the light cones, the currents vanish.}
The currents at the origin $x=0$ are especially simple. They do not depend on time and are given:
\begin{eqnarray}
j^{\mathrm{(qp)}}(0) & = &\frac{1}{\pi} \left[ \mathrm{e}^{-\beta_L(U-\mu_L)} \frac{\sinh 4\beta_L J}{\beta_L} - \mathrm{e}^{-\beta_R(U-\mu_R)} \frac{\sinh 4\beta_R J}{\beta_R}\right ]\nonumber\\
j^{\mathrm{(qh)}}(0) & = &\frac{1}{\pi} \left[ \mathrm{e}^{-\beta_L\mu_L} \frac{\sinh 2\beta_L J}{\beta_L} - \mathrm{e}^{-\beta_R\mu_R} \frac{\sinh 2\beta_R J}{\beta_R}\right ]
\end{eqnarray}
The fact that they are time-independent is a signature of ballistic propagation of the excitations, and nicely fits the observations
made in Fig.~\ref{merge2} on our quasi-exact numerical METTS simulations. Both currents are simply of the form $g(\beta_L)-g(\beta_R)$ for some function $g,$
a property already emphasized as a signature of ballistic transport in~\cite{Karrasch13}. In contrast, diffusive transport would be characterized
by currents depending on the local gradients.
In the regime of intermediate temperature $J \ll \frac{1}{\beta_R},\frac{1}{\beta_R} \ll U,$ the expression simplifies and the currents
have a ''semi-circle'' spatial shape:
\begin{eqnarray}
j^{\mathrm{(qp)}}(x,t) & \propto & \sqrt{16J^2-x^2/t^2}\nonumber\\
j^{\mathrm{(qh)}}(x,t) & \propto & \sqrt{4J^2-x^2/t^2}.
\label{eq:semi-circle}
\end{eqnarray}
At lower temperature ($\beta J$ of the order of unity or smaller), the shape is qualitatively similar.
\section{Comparison of the Bogoliubov theory with numerical data}
\begin{figure*}
\includegraphics[width=8cm]{density_mass_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\huge$\langle n_i\rangle$}\includegraphics[width=8cm]{density_mass_merge_68.png}
\includegraphics[width=8cm]{density_variance_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\Large$\Var{(n_i)}$}\includegraphics[width=8cm]{density_variance_merge_68.png}
\includegraphics[width=8cm]{density_quasiparticle_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\huge$n^{\mathrm{(qp)}}_i$}\includegraphics[width=8cm]{density_quasiparticle_merge_68.png}
\includegraphics[width=8cm]{density_quasihole_merge_68_mateusz.png}\parbox[b][4.5cm][t]{1.6cm}{\huge$n^{\mathrm{(qh)}}_i$}\includegraphics[width=8cm]{density_quasihole_merge_68.png}
\caption{(color online) Comparison of {quasi-exact} numerical METTS simulations (left column) with the Bogoliubov theory (right).
From top to bottom: mean site occupation $\langle n_i \rangle$, variance $\langle n_i^2 \rangle-\langle n_i \rangle^2$, quasiparticle density $n^{\mathrm{(qp)}}_i$ and quasihole density $n^{\mathrm{(qh)}}_i$.
The dashed white (resp. black) straight lines correspond to ballistic propagation of quasiparticles
(resp. quasiholes) from the connection point between the two subsamples at the maximal allowed velocity,
i.e. $x^{\mathrm{(qp)}}(t)=\pm 4Jt$ (resp. $x^{\mathrm{(qh)}}(t)=\pm 2Jt$). See text for a detailed discussion.
}
\label{copar1}
\end{figure*}
\begin{figure*}
\includegraphics[width=8cm]{current_mass_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\Large$j^n(i)$}\includegraphics[width=8cm]{current_mass_merge_68.png}
\vspace{1truemm}\includegraphics[width=8.cm]{current_energy_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\Large$j^E(i)$}\includegraphics[width=8.cm]{current_energy_merge_68.png}
\vspace{1truemm}\includegraphics[width=8cm]{current_variance_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\Large$j^{\mathrm{Var}}(i)$}\includegraphics[width=8cm]{current_variance_merge_68.png}
\vspace{1truemm}\includegraphics[width=8.cm]{current_quasihole_merge_68_mateusz.png}\vspace{2mm}\parbox[b][4.5cm][t]{1.6cm}{\Large$j^{(\mathrm{qh})}(i)$}\includegraphics[width=8.cm]{current_quasihole_merge_68.png}
\caption{(color online)
Comparison of the currents computed from the numerical METTS simulations (left column) with the Bogoliubov theory (right).
{The currents are obtained by numerical calculations for the same parameter values as in Fig.~\ref{copar1}.}
From top to bottom: mass current, energy current, heat (variance) current as defined in Eq.~(\ref{varcur}) and quasihole current
as defined in Eqs.~(\ref{eqcur}) (the quasiparticle current coincides with the energy current). See text for discussion.
}
\label{coparcur}
\end{figure*}
\subsection{Sharp gluing}
We now give a more detailed analysis of the simulations already presented in Fig.~\ref{fig:fluct} and Figs.~\ref{merge1}-\ref{fig:3D1}. Recall that we consider two Mott insulators of length $M=35,$ each one at thermal equilibrium, but with different temperatures $\beta_L U=6,\beta_R U=8$. Each insulator is represented by an appropriate thermal density matrix (METTS ensemble), the total state of the system being the tensor product of those density operators. At time $t=0$ the tunneling between touching sites is turned on to a common value $J/U=0.05$. While in Figs.~\ref{fig:fluct} and \ref{merge1} we have shown the
average particle density $\av{n_i}$, squared particle density $\av{n_i^2}$ and variance $\Var{(n_i)}=\av{n_i^2}-\av{n_i}^2$ at selected times, Fig.~\ref{copar1} presents
the same data resulting from a quasi-exact numerical simulation as
color plots (left column) together with the predictions given by the Bogoliubov theory (right column) described in the previous Section.
A good agreement between the quasi-exact METTS simulations and the Bogoliubov prediction is visible already in the top row, where we compare densities, $\langle n_i \rangle$: a bump propagating to the right side, a hole propagating to the left side. On the left side, the
higher temperature implies a higher density of quasiparticles and quasiholes. Once the two parts are connected, the quasiparticles
propagate ballistically at twice the speed of the quasiholes. It results in an excess of quasiparticles -- hence a higher density -- between
the quasihole light cone (dashed black line) and the quasiparticle light cone (white dashed line).
As noted earlier, the variance $\Var{(n_i)}$, shown in the second row, is less sensitive to noise, and a very good agreement
between numerical METTS simulations and the Bogoliubov theory is observed.
While the basic ingredients of the theoretical approach are the quasiparticle and quasiholes densities, these
quantities cannot be directly measured in the METTS simulations. It is, however, possible to invert Eqs.~(\ref{Eq:occupation_prediction})
and to deduce the densities from the measured quantities $\langle n_i \rangle$ and $\langle n_i^2 \rangle$:
\begin{eqnarray}
n^{\mathrm{(qp)}}_i & = & \frac{\Var{(n_i)}+\av{n_i}-1}{2} - \frac{4J^2}{U^2}\nonumber \\
n^{\mathrm{(qh)}}_i & = & \frac{\Var{(n_i)}-\av{n_i}+1}{2} - \frac{4J^2}{U^2}.
\end{eqnarray}
These formulae allow us to determine the distribution in time of quasiparticles and quasiholes for simulations, as well as conversely find the predictions for the particle density and its variance from the quasiparticle distributions.
Clearly the excess of particles is observed on the cold side (with the corresponding hole on the hot side due to conservation of the total number of particles) moving out ballistically from the {junction} point. That is due to the spread of quasiparticles, as described in the previous section.
Observe that while we keep the same scale for the density, we use different scales for the variance (second row in Fig.~\ref{copar1})
and for the quasiparticle (third row) and quasihole (fourth row) densities.
That is due to finite size effects discussed extensively in the previous section. Instead of fitting the correction we have chosen to plot the raw data; the correction is ``automatically'' taken into account by the rescaling of the colorbar. This provides a convincing argument that a single correction factor valid at all times is sufficient to bring the results of METTS simulations and the theoretical predictions together. While of course some differences may be visible, the overall agreement is quite remarkable showing beyond doubt that the Bogoliubov theory describes very well the results of the simulations and thus the physics of heat and mass transfer in this system.
Note especially the clear demonstration that quasiholes propagate more slowly than quasiparticles (by a factor 2), as the
quasihole density is affected only inside the inner light cone -- marked with the black dashed lines -- while the quasiparticle density
is affected inside the outer light cone -- marked with the white dashed lines.
These conclusions are further confirmed by inspection of different possible currents. As before, the simulations give us access to
mass current $j^n$, Eq.~(\ref{jN}) as well as to the energy current $j^E$, Eq.~(\ref{jEa}). Those can be related to quasiparticle and quasihole currents via
Eqs.~(\ref{eqcur}). One can also define the variance current, which following Eq.~(\ref{Eq:occupation_prediction}) can be defined as
\begin{equation} \label{varcur}
j^{\mathrm{Var}}(x,t) = j^{\mathrm{(qp)}}(x,t) + j^{\mathrm{(qh)}}(x,t) = j^n(x,t) + \frac{2j^E(x,t)}{U}
\end{equation}
Since the variance is directly related to temperature, the variance current can be considered as a heat current.
Observe in Fig.~\ref{coparcur} that, as expected, the currents are nonzero only in the ``light-cone'' region around the point of junction.
The agreement between the simulations and the theory is less spectacular for mass and energy currents while it is much better for the heat current and, in particular, quasihole current (bottom row). Indeed the cone both for simulations and theory is then
restricted by the maximal velocity of quasiholes. The differences in results between the two approaches may be again traced back to finite size effects.
\begin{figure}
\includegraphics[width=8cm]{n_n2_time_60.png}
\caption{(color online) Comparison of the average particle density $\langle n_i \rangle$ (lower curves) and the average squared density $\langle n_i^2 \rangle$ (upper curves) at time $tU=60,$ for the quasi-exact METTS simulations (solid black lines) and the Bogoliubov theory. The raw Bogoliubov predictions are the
dotted green curves; rescaled results, taking into account finite size effects, as shown as dashed red lines and agree very well
with the METTS simulations. See text for discussion of the various features.}
\label{fig:n_n2_time_60}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{figure_currents_t60a.png}
\caption{(color online) Comparison of the mass and energy currents for the same data than in Fig.~\ref{fig:n_n2_time_60}.
The energy current is predicted by the Bogoliubov theory to depend only on the quasiparticle current, see Eq.~(\ref{eqcur}).
It displays the characteristic ''semi-circle'' shape predicted by our Bogoliubov theory, Eq.~(\ref{eq:semi-circle}).
In contrast, the mass current is the difference between the quasiparticle and quasihole ''semi-circles'', see Eq.~(\ref{eqcur}),
and presents two maxima propagating at the maximum velocity of the quasiholes $\pm2J.$}
\label{fig_curr_t60}
\end{figure}
Color plots in Figs.~\ref{copar1},\ref{coparcur} show that our Bogoliubov approach is qualitatively correct: heat and mass transport
in the system are conveniently described by quasiparticle/quasihole excitations that propagate ballistically.
In order to make the comparison more
quantitative, we show in Fig.~\ref{fig:n_n2_time_60} the average particle density $\langle n_i \rangle$ and the average squared density $\langle n_i^2 \rangle$ at time $tU=60.$ After the proper rescaling necessary due to finite size effects (see discussion above), the agreement is very good.
The density bump (on the right side) and hole (on the left side) is well predicted. For the average squared density $\langle n_i^2 \rangle,$ one can see three distinct regions between the ''hot'' left plateau (not yet affected) and the ''cold'' right plateau: two rather abrupt cliffs on the edges
of the two plateaus separated by an intermediate much flatter region. The two cliffs correspond to regions already reached by quasiparticles
while the intermediate flat region is affected both by quasiparticles and quasiholes. The kinks at the frontiers between regions
correspond to quasiparticles/quasiholes with maximum velocities $4J/2J,$ respectively. They are smoothed out, but still clearly visible,
in the quasi-exact METTS simulations. Note that all quasiparticles/quasiholes in the system are thermally excited, so that the effects
observed are truly due to non-equilibrium thermodynamical properties.
In Fig.~\protect{\ref{fig_curr_t60}}, we show the comparison between the mass and energy currents as computed from the METTS simulations
and the predictions of the Bogoliubov approach. Not surprisingly, the salient features are quantitatively well predicted:
the energy current, predicted by Eq.~(\ref{eqcur}) to depend only on the quasiparticle current displays a single bump
with the characteristic ''semi-circle'' shape predicted by Eq.~(\ref{eq:semi-circle}); the mass current is the difference between the quasiparticle and quasihole currents and consequently reveals two maxima propagating at the maximum velocity of quasiholes $\pm2J.$
\subsection{Smooth gluing}
We now confront the predictions of our Bogoliubov theory
with the numerical data obtained for a smoothly glued sample, as the one described in section~\ref{sec:smooth}. As previously, the whole system
consists of 70 sites with $\beta_L U=6$ on the hot side and $\beta_R U=8$ on the cold side.
Fig.~\ref{varmetts} shows the variance of the particle density across the system at initial time as well as
after some real time evolution. {In order} to compare the theory and the numerical experiment correctly, one has to introduce
some rescaling, taking care of the finite size effect. While two different rescaling factors were used for the ''sharp gluing'' case
where the two subsamples are initially not connected, we here use a \textit{single} rescaling parameter for the whole sample, and all times.
This is because -- by construction -- a smooth gluing procedure find the equilibrium initial state for the whole sample at the initial time.
As before, real time evolution washes out site to site fluctuations of the variance: starting from, say $tU=10,$
the variance is {perfectly} reproduced by the Bogoliubov theory.
\begin{figure}
\includegraphics[width=8cm]{variance_metts68_t_0_and_60.png}
\caption{(color online) Particle number variance at initial time $t=0$ and at $tU=60,$ for two smoothly glued Mott insulators with
(inverse) temperatures $\beta_L U=6,\beta_R U=8,$ and $J/U=0.05.$ The interface size, Eq.~(\ref{glue}), is $s=15$.
Solid lines correspond to quasi-exact numerical results using the METTS method, while dashed curves are theoretical estimates.
A single parameter -- required to take into account finite size effects -- has been used to adjust the numerical and theoretical curves. The sharpest (black) step corresponds to the initial time $t=0.$ The smoother (red) curve corresponds to $tU=60.$ Note that the real time evolution strongly smooths the initial site-to-site fluctuations. At $tU=60,$ the theoretical prediction is almost indistinguishable from the quasi-exact numerical
simulation.}
\label{varmetts}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{currents_metts68.png}
\caption{(color online) Mass (upper panel) and energy (lower panel) currents at $tU=60$ for the same data as in Fig.~\ref{varmetts}.
Numerical results (solid black curves) are compared with the predictions of the Bogoliubov theory (dashed red curves).
While the magnitude of the currents is not predicted properly, the overall shape is faithfully predicted by the theory.}
\label{curmetts}
\end{figure}
Fig.~\ref{curmetts} presents the mass and the energy currents at $tU=60$ using the same single scaling factor. Clearly some quantitative discrepancies between the numerics and theoretical simulations exist. One may, however, observe quite nice qualitative shape agreement between the two curves.
\begin{figure}
\includegraphics[width=8cm]{variance_metts_810_t_0_10_and_60.png}
\caption{(color online) Particle number variance at initial time $t=0$, $tU=10$ and at $tU=60,$ for two smoothly glued Mott insulators with
(inverse) temperatures $\beta_L U=8,\beta_R U=10,$ and $J/U=0.05.$ The interface size, Eq.~(\ref{glue}), is $s=15$. Solid lines correspond to quasi-exact numerical results using the METTS method, while dashed curves are theoretical estimates.
A single parameter -- required to take into account finite size effects -- has been used to adjust the numerical and theoretical curves. The sharpest (black) step corresponds to the initial time $t=0.$ The smoother (red and green) curves corresponds to $tU=10$ and $tU=60$. The real time evolution strongly smooths the initial site-to-site fluctuations. At $tU=60,$ the theoretical prediction is almost indistinguishable from the quasi-exact numerical
simulation.}
\label{vmetts10}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{currents_metts_810.png}
\caption{(color online) Mass (upper panel) and energy (lower panel) currents at $tU=60$ for the same data as in Fig.~\ref{vmetts10}.
Numerical results (solid black curves) are compared with the predictions of the Bogoliubov theory (dashed red curves).
The agreement is very satisfactory. }
\label{cmetts10}
\end{figure}
As a final example, we consider the same system but at lower temperature. This time the hot part corresponds to $\beta_L U=8$ while the cold part to $\beta_R U=10,$ again with a smooth gluing of the two subsystems.
The results are presented in Fig.~\ref{vmetts10} and Fig.~\ref{cmetts10}. While the fluctuations are notably larger (quantum effects are more visible), the behavior of the system is similar to the previously considered case. Note also that the imperfect quantitative agreement for the currents observed at higher temperature, Fig.~\ref{curmetts}, is here significantly better. A possible explanation is that, at higher temperature, our simple Bogoliubov theory breaks down at shorter
times, either because of additional excitations or because of the finite lifetime of quasiparticle/quasihole excitations.
\section{Conclusions}
\label{conclusions}
We have studied non-equilibrium dynamics occurring when two quantum insulators at different temperatures are brought into contact.
The process, in the long run, should lead to equilibration of temperatures (at least in the thermodynamic limit).
Using Mott insulators of the Bose-Hubbard model as a specific example -- as it is close to experimental realization with ultracold atoms --
we, however, observe that heat transport is initially ballistic: it efficiently transfers energy from the hot to the cold side of the system, but the full thermalization may occur only on a longer time scale, not reachable by our numerical simulations.
Fully numerical, ``quasi-exact'' calculations have been performed using the METTS algorithm~\cite{White09}.
Even with important computer resources, we could only study the dynamics on a rather limited amount of time.
This is due to two primary reasons. First, the METTS ensemble needed to obtain reasonable averages consisted of more than 10 000 vectors; each of them had to be evolved using the TEBD (or t-DMRG) algorithm. The price to pay is a direct $10^4$ increase of the computational time in comparison to $T=0$ studies of e.g. quantum quenches. Second, for $T>0,$ the METTS ensemble includes by construction some significantly excited states; the real time evolution of such states leads to a fast increase of the entanglement, requiring significant extension of the bond dimension in the MPS description. \textit{A posteriori,} one may conclude that this is not the method of choice and other approaches such as a purification scheme \cite{Karrasch13} or methods employing Heisenberg
picture evolution \cite{Pizorn14} could perform much better, at least according to comments in these papers.
A very recent comparison of purification schemes with METTS \cite{Barthel14} reached similar conclusions.
The numerical results have been quantitatively compared with a simple analytic theory based on Bogoliubov excitations.
The theory explains the ballistic transport numerically observed in our system providing good estimates for observable quantities, provided some plausible corrections are included that took into account the finite size effects.
The theory applies also for smoothly glued system: the transport may be fully understood as the movement of quasiparticles and quasiholes of the integrable Bogoliubov theory.
In some sense, this explains the success of the METTS approach for description of this particular system. As conjectured by Prosen \cite{Prosen07}, the entanglement growth in an integrable system is slower and easier to handle.
\acknowledgments
The authors wish to thank J. Dziarmaga for enlightenment on quasiparticles. We acknowledge the use of the ALPS library~\cite{Bauer:Alps2:JSM11} for Quantum Monte-Carlo calculations using the worm
and directed loop algorithms.
This work has been supported by Polish National Science Center within project No. DEC-2012/04/A/ST2/00088. M.\L. acknowledges support of the Polish National Science Center by means of project no. 2013/08/T/ST2/00112 for the PhD thesis. Numerical simulations were performed thanks to the PL-Grid project: contract number: POIG.02.03.00-00-007/08-00 and at Deszno supercomputer (IF UJ) obtained in the framework of the Polish Innovation Economy Operational Program (POIG.02.01.00-12-023/08).
This work was performed within Polish-French bilateral program POLONIUM No. 33162XA.
|
\section{Introduction}
\label{sec:intro}
The ATLAS Pixel Detector~\cite{Aad:2008zz} is constructed from
planar silicon pixel modules arranged in a barrel and disk geometry.
It is centered on the interaction point at the heart of the ATLAS
Detector~\cite{Aad:2008zzm} at the Large Hadron Collider
(LHC). A linear relationship between leakage current and a hadronic particle
fluence is expected to apply to the pixel silicon sensors and is given
by
\begin{equation}
\Delta{I} = \alpha\cdot\flue\cdot{V},
\label{eq:i-phi}
\end{equation}
where \(\Delta{I} \) is the difference in leakage current at fluence
\( \flue\) relative to the value before irradiation of the
physical volume \( V \), and \( \alpha \) is the current-related
damage coefficient~\cite{Moll:1999nh}. The goal of this study is to
measure the leakage currents of a representative sample of sensors in
the ATLAS Pixel Detector in order to monitor and understand the sensor
damage resulting from increasing radiation dose. The measured results
are compared to predictions based on the Hamburg
Model~\cite{damage-prediction} in which reverse-bias current is
generated by the introduction of defect levels in the band gap due to
non-ionizing energy loss (NIEL) of through-going particles. The defects
migrate and recombine through several processes that result in both
short-term beneficial annealing (which reduces the current) and
long-term anti-annealing (which increases the current even in the
absence of further radiation). The model can be calibrated to the
data to predict the lifetime of the sensors and the Pixel Detector as
a whole for various operating scenarios.
\par
The detector is instrumented as three barrels, specifically the
Layer-0 at \(r=50.5\mm\), the Layer-1 at \(r=88.5\mm\), and the Layer-2
at \(r=122.5\mm\), comprising in total 1456 modules.
The endcap areas are instrumented with three disks located at
\(z=\pm495\mm,\,\pm580\mm \) and \(\pm650\mm\) in each of the forward
and backward regions, comprising an additional 288 modules. The Pixel
Detector and the other elements of the ATLAS Inner Detector span a
pseudorapidity range
\(\rapid< 2.5\)~\footnote{ATLAS uses a right-handed coordinate
system with its origin at the nominal interaction point (IP) in the
centre of the detector and the $z$-axis along the beam pipe. The
$x$-axis points from the IP to the centre of the LHC ring, and the
$y$-axis points upward. Cylindrical coordinates $(r,\phi)$ are used
in the transverse plane, $\phi$ being the azimuthal angle around the
beam pipe. The pseudorapidity is defined in terms of the polar angle
$\theta$ as $\eta=-\ln\tan(\theta/2)$.}.
The pixel sensor modules are mounted on mechanical/cooling supports,
called staves, in the barrel region. Thirteen modules are mounted on
a stave with identical layouts for all layers. The temperature is
maintained stable by an evaporative cooling system.
The entire \(\sim\!{1.7}{\m}^{2}\) sensitive area of the ATLAS pixel
detector is covered with 1744 identical modules. Each module has an
active surface of \(6.08\times1.64\cma\).
The details of the pixel sensor geometry and layout can be found
in~\cite{Aad:2008zz}.
\par
The radiation field in the region of the detector has been predicted
with the aid of a {\sc FLUKA} software package~\cite{Baranov:2005ewa}.
The {\sc FLUKA} transport code to cascade the particles through the
ATLAS detector material has been used together with {\sc PYTHIA~6.2}
or {\sc PHOJET} generating proton-proton minimum-bias events as an
input to {\sc FLUKA}~\cite{Aad:2008zz}.
\par
We assume that the dominant radiation damage type is displacement
defects in the bulk of the pixel sensor, initiated by hadronic species
and caused by NIEL. Surface ionization is
neglected in this treatment. Charged pions are expected to dominate
the bulk damage for the radii covered by the Pixel Detector. Albedo
neutrons originating in the outer ATLAS detectors also contribute.
The resulting displacement defects increase the reverse leakage
current, degrade the charge collection efficiency, and change the
effective doping concentration which directly determines the depletion
voltage. We define the effective fluence \flue as the number of
particles causing damage equivalent to that of \( 1\!\MeV \) neutrons
traversing \( 1\cma \) of a sensor's surface. The fluence \( \flue \)
accumulated by ATLAS pixel detector, and measured in units
of \({\cm}^{-2}\), is expected to be proportional to integrated
luminosity \( \IntL \), measured in \( \invpb \).
\section{Measured Quantities}
The parameter \(\alpha\) in Equation~\ref{eq:i-phi} has been measured
under a variety of conditions, for example \cite{Moll:1999nh}. The
linear form applies to the leakage currents drawn by sensors past
their beneficial annealing periods. At the beginning of data-taking
and during beneficial annealing periods, the sensors draw currents at
levels comparable to dark currents;
thus any hardware implementations for leakage current measurement
should provide access to the lowest current range possible.
\par
The reverse bulk generation current depends on the sensor temperature
$T$. The current measurements \( I(T) \) are normalized to the
reference temperature \(T_{R}\) (\(0\degc\))
with
\begin{equation}
I(T) = I(T_{R})/R(T),\,\, {\rm where} \,\,
R(T) = ( T_{R}/T )^{2}\cdot\exp{\left( {-\frac{E_g}{2k_{B}}}(1/T_{R}-1/T) \right), }
\label{eq:i-temp}
\end{equation}
where \( E_{g} = 1.21\ev \) is the energy of the silicon band gap,
\(k_{B} \) is the Boltzmann constant, and \(T\,{\rm and}\,T_{R} \) are
measured in \degk~\cite{Chilingarov}.
\par
The mean pixel module temperature for all barrel pixel modules is
shown versus time in
Figure~\ref{temperature-profile}~\cite{ATL-INDET-PUB-2014-004}.
There were a few exceptions when the temperature rose to ambient due
to cooling interruptions and calibration scans.
Temperature is read out from every module continuously and averaged
over 30-minute intervals. Temperature data (as well as other slow
control data) are committed into the ATLAS Detector Control System
(DCS) database.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=1.0\textwidth]{./figures/hvpp4-mean-temp-profile.eps}
\end{center}
\caption{ The mean temperature (\degc) of all barrel pixel modules,
versus time (days)~\cite{ATL-INDET-PUB-2014-004}.}
\label{temperature-profile}
\end{figure}
\section{The ATLAS Current Measurement Subsystem}
The ATLAS High Voltage Patch Panel 4 (HVPP4) serves as a fan-out point
for the bias voltages delivered by Type II boards from
Iseg~\cite{Iseg-specs}
high voltage power supplies to the pixel modules. The leakage current
is monitored at the pixel module granularity level by a Current
Measurement Board (CMB) system. A CMB
is mounted on every Type~II fan-out board. The bias voltage to the
sensors is provided by the Iseg channels.
\par
In the present data-taking era, when the radiation damage of the
sensors is relatively low, 6 or 7 pixel modules are fed by one Iseg
power supply channel. At some point after inversion of the sensors,
before the current drawn by 6 or 7 pixel modules reaches the Iseg
limit, power supplies can be added until the system provides one Iseg
power supply channel per pair of pixel modules. The CMB system allows
simultaneous measurement of the leakage currents in 4 pixel modules in
the current range from \(0.04\mkamp\) to \(2\mamp\)
%
%
with a precision of better than \(20\,\%\) per module. ATLAS uses
\(21\) CMBs in Layer-0, \(16\) in Layer-1, and \(16\) in Layer-2. The
analog current measurements are digitized by the \(64\)-channel ATLAS
Embedded Local Monitoring (ELMB) board and sent via the Controller Area
Network (CAN) bus to the DCS database (see~\cite{Aad:2008zz} for
details on the Type~II fan-out board, ELMB, CAN, and DCS). The building
block of the CMB circuit is a current to frequency converter optically
coupled to a frequency to voltage converter as shown in
Figure~\ref{CMB-circuit}. Each CMB holds \(4\times2\) current measurement
circuits and provides the current measurement for \(4\) pixel modules.
Two similar circuits with different gains are allocated per CMB
channel serving the same module. The channels are isolated from each other
and from the pixel module readout system.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.50\textwidth]{./figures/CMBcircuit-new.eps}
\end{center}
\caption{The Current Measurement Board circuit}.
\label{CMB-circuit}
\end{figure}
For each pixel module, the leakage current range
\([10^{-8}-10^{-5}]\amp\) is covered by the high-gain channel, and the
range \([10^{-6}-2\cdot10^{-3}]\amp\) is covered by the low-gain
channel. The output CMB voltage range is defined by one of five
available ELMB input voltage ranges, which are \([0-25]\mvolt\),
\([0-100]\mvolt\), \([0-1]\volt\), \([0-2.5]\volt\), and
\([0-5]\volt\). The output range \([0-5]\volt\) is necessary to
comply with the specifications of the digital ELMB board. The 16-bit
ADC of the ELMB provides a resolution of (ELMB
range)/\(({2}^{16}-1)\). The range \([0-1]\volt\) was used for data
taken through 2011; the range was changed to \([0-5]\volt\) in 2012 as
the high-gain channels reached saturation, corresponding to pixel
leakage currents greater than \(10^{-5}\amp\). After LHC shutdown of
2013-2014, the data-taking will resume with the low-gain channels.
Before installation, each CMB is calibrated offline assuming a linear
model. The calibrated gains of the channels are stored in the
database.
The pedestals are re-calibrated {\it in situ}.
\section{Current Measurements, 2011-2012}
Figure~\ref{cmbpaper-lumi} shows the leakage current recorded by the
CMB system from the beginning of ATLAS data-taking through the end of
2012, as a function of integrated luminosity, for modules in all three
Pixel Detector barrel layers~\cite{ATL-INDET-PUB-2014-004}.
Installation of the CMB system proceeded concurrently with early
data-taking, and this is the reason that some Layer-2 data are missing
from the early part of the run. It is foreseen that the disks will be
instrumented with current monitor boards prior to the start of
data-taking in 2015 (Run 2).
The points at which LHC and cooling were off, hence beneficial
annealing of the silicon was unsuppressed, are apparent as
discontinuities in this figure. This effect, producing the sawtooth
features, is due to the increase of the effective doping concentration
of the silicon due to annealing
of acceptors~\cite{Moll:1999nh, damage-prediction}.
\par
Figure~\ref{cmbpaper-lumi} also includes a
prediction~\cite{damage-prediction} of the leakage current for the
detailed ATLAS geometry and LHC luminosity profile based on the
Hamburg Model. The predictions of the fluence \(\flue\) according to
the luminosity profile are made using the {\sc FLUKA} transport
code~\cite{Baranov:2005ewa}. The predictions are uniform w.r.t. the
azimuthal angle \(\phi \) and suppose a forward-backward symmetry
w.r.t. the polar angle \(\theta \), i.e. \(\flue(\eta)=\flue(-\eta)
\).
At the end of Run 1 with \(\IntL=30\ifb \), the Layer-0 modules have
received a total fluence
\(\flue\eqsim7.0\cdot10^{13}{{\rm \,cm}^{-2}}\),
while Layer-1 and Layer-2 have been irradiated to
\(\flue\eqsim3.0\cdot10^{13}{{\rm \,cm}^{-2}}\) and
\(\flue\eqsim1.8\cdot10^{13}{{\rm \,cm}^{-2}}\) respectively.
The uncertainty of the leakage current predictions comprises the
uncertainty of the calculations of the hadronic particle spectra
initiated by \(\propro\) interactions in the ATLAS detector and the
uncertainty of the Hamburg model in describing radiation damage effects
inflicted by NIEL of hadrons crossing the volume of the
silicon pixel sensors.
\begin{SCfigure}
\centering
\captionof{figure}{ ATLAS Pixel module leakage current versus integrated LHC
luminosity~\cite{ATL-INDET-PUB-2014-004}.
The currents are averaged over their layer for
all modules equipped with Current Measurement Boards
within the layer. The current is continuously monitored
by the ATLAS Detector Control System. A prediction based
on the Hamburg Model is included. Discontinuities are due
to annealing during LHC and cooling stops.}
\includegraphics[width=0.5\linewidth]
{./figures/10jul-hvpp4_CorCur_lumi_Jan2013_total-landscape.eps}
\label{cmbpaper-lumi}
\end{SCfigure}
Figure~\ref{fig:phi-map} shows the distribution in azimuthal angle of
modules whose current is sampled by CMBs in the ATLAS Pixel
Detector.
Figure~\ref{fig:phi-current} shows the data for each of four quadrants
in the azimuthal angle as they are defined by
Figure~\ref{fig:phi-map}~\cite{ATL-INDET-PUB-2014-004}.
No systematic difference in a module leakage current is observed
w.r.t. to the quadrant where the module is located.
\begin{figure}[htbp]
\centering
\begin{subfigure}[t]{.4\textwidth}
\centering
\includegraphics[width=1.05\linewidth]{./figures/phi-quadrants-map.eps}
\caption{ A diagramm of modules.}
\label{fig:phi-map}
\end{subfigure}%
\begin{subfigure}[t]{.6\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/10jul-hvpp4_CorCur_phi_lumi_Jan2013_4_total.eps}
\caption{ Module leakage current versus integrated LHC luminosity.}
\label{fig:phi-current}
\end{subfigure}
\caption{(a)~Sampling of 4 quadrants in azimuthal angle with modules
instrumented by CMBs.~(b)~The module leakage current in each of the 4
quadrants in azimuthal angle~\cite{ATL-INDET-PUB-2014-004}. The
currents are averaged over each layer for all modules equipped
with Current Measurement Boards within the particular quadrant of
the layer. Predictions based on the Hamburg Model are included.}
\label{cmbpaper-phi}
\end{figure}
Figure~\ref{cmbpaper-eta-diagram} shows the instrumented module
distribution, segmented into five sectors in pseudorapidity. The
pseudorapidity acceptance of each sector of each layer is indicated.
Plots in Figure~\ref{fig:eta1} through Figure~\ref{fig:eta5}, show the
leakage current data for each of the 5 ranges in
pseudorapidity~\cite{ATL-INDET-PUB-2014-004}.
Figure~\ref{fig:eta6} shows the current data versus pseudorapidity
for the sectors defined by Figure~\ref{cmbpaper-eta-diagram}, for
integrated luminosity 25\invfb~\cite{ATL-INDET-PUB-2014-004}. The
possible slight asymmetry of the data points in Figure~\ref{fig:eta6}
is thought to be due to the asymmetric beam profile. Yet the
distribution is in agreement within the uncertainties with the leakage
current predictions which are symmetrical according to the fluence
simulation~\cite{Baranov:2005ewa}.
\begin{figure}
\centering
\begin{minipage}{.66\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/10jul-eta-horizontal.eps}
\end{minipage}%
\begin{minipage}{.33\textwidth}
\centering
\captionof{figure}{ A diagram of the distribution of instrumented modules in
each of five pseudorapidity ranges. }
\label{cmbpaper-eta-diagram}
\end{minipage}
\end{figure}
\begin{figure}[htbp]
\centering
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/3jul-hvpp4_CorCur_eta_lumi_Jan2013_5_1_total.eps}
\caption{}
\label{fig:eta1}
\end{subfigure}%
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/3jul-hvpp4_CorCur_eta_lumi_Jan2013_5_2_total.eps}
\caption{}
\label{fig:eta2}
\end{subfigure}%
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/3jul-hvpp4_CorCur_eta_lumi_Jan2013_5_3_total.eps}
\caption{}
\label{fig:eta3}
\end{subfigure}
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/3jul-hvpp4_CorCur_eta_lumi_Jan2013_5_4_total.eps}
\caption{}
\label{fig:eta4}
\end{subfigure}%
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/3jul-hvpp4_CorCur_eta_lumi_Jan2013_5_5_total.eps}
\caption{}
\label{fig:eta5}
\end{subfigure}%
\begin{subfigure}{0.32\textwidth}
\centering
\includegraphics[width=1.0\linewidth]{./figures/hvpp4_CorCur_eta_lumi_Jan2013_5_6_total.eps}
\caption{}
\label{fig:eta6}
\end{subfigure}
\caption{ (a) through (e): ATLAS Pixel module leakage current in each
of 5 pseudorapidity sectors, versus integrated LHC
luminosity~\cite{ATL-INDET-PUB-2014-004}. The currents are
averaged over their layer for all modules equipped with Current
Measurement Boards within the layer. The current is continuously
monitored by the ATLAS Detector Control System. Predictions based
on the Hamburg Model are included.
Discontinuities are due to annealing during LHC and cooling stops.
(f): the module leakage current in
defined 5 pseudorapidity bins, for the integrated luminosity equal
to 25\invfb~\cite{ATL-INDET-PUB-2014-004}.}
\label{cmbpaper-eta}
\end{figure}
\section{Precision and Systematic Uncertainties}
The contributions to the uncertainty on the current include
\begin{itemize}
\item the CMB and ELMB precision on current measurements: \(12\%\) for both ranges.
\item the number of current measurements (two per hour).
\item the uncertainty on the luminosity, \(1.8\%\) in 2011 and \(2.8\%\) in
2012~\cite{Aad:2013ucp}.
\item the temperature uncertainty, which, at less than \(0.3\degc\),
contributes to the uncertainty on the current with \(3.4\%\).
\end{itemize}
\section{Lifetime Estimate}
The temperature-corrected current readings per module can be
extrapolated to predict the amount of current the pixel modules will
draw after a certain integrated luminosity has been collected with the
ATLAS pixel detector.
As leakage current can lead to excessive power and thermal runaway,
limits on the module lifetime \emph{due to current} can be determined
from the limit on the bias voltage that can be applied.
A single Iseg power supply channel can sustain a maximum current of
\(\lsim{4000\mkamp} \) for two modules. There are other consequences
of radiation-induced leakage current upon module lifetime, including
the noise, bias, and threshold limitations of the front end chips as
well as the finite capacity of the cooling system; while important,
these are not treated here.
\par
We extrapolate the present rate of current increase with integrated
luminosity, assuming optimistically that the modules will
be exposed for 10 days to $20$\degc and otherwise maintained at
$-13$\degc in future years. The model assumes that the bias voltage is raised
throughout LHC operation as needed to keep the bulk fully depleted,
until a maximum allowed voltage of 600~V is reached. Only the effects
of proton runs are considered. Figure~\ref{lifetime} shows the
implied leakage current per module versus date and versus integrated
luminosity for Layer-0 in this scenario~\cite{ATL-INDET-PUB-2014-004}.
This extrapolation predicts a leakage current of about
$500\mkamp$ per Layer-0 module (of thickness $250\mkm$) after the accumulation of
about 450\invfb of delivered luminosity, which corresponds to January $2024$ in the
scenario proposed here.
\begin{figure}
\centering
\begin{minipage}{.66\textwidth}
\centering
\includegraphics[width=1.0\linewidth]
{./figures/ileak-density-lumi-profiles-daniel.7.eps}
\end{minipage}%
\begin{minipage}{.33\textwidth}
\centering
\captionof{figure}{ Predicted leakage current per Layer-0 module at
$-13$\degc, versus date (upper) and integrated
luminosity (lower) for the scenario in which modules will
be exposed for 10 days to $20$\degc and otherwise
maintained at $-13$\degc~\cite{ATL-INDET-PUB-2014-004}.}
\label{lifetime}
\end{minipage}
\end{figure}
\section{ Summary and Prospects}
\label{sec:summ}
We have described the principles of radiation damage monitoring using
the current measurements provided by the circuits of the ATLAS Pixel
high voltage delivery system. We have observed radiation damage in
the ATLAS Pixel Detector with characteristics in agreement with the
Hamburg Model. The dependence of the leakage current upon the
integrated luminosity has been presented. A linear behavior of the
response has been observed. The current distribution in azimuthal
quadrants is symmetric as expected. The shape of the current
distribution in pseudorapidity is in agreement with a fluence
calculation based on the {\sc FLUKA} transport code within the
assigned systematic uncertainties. Extrapolation of leakage current
with time predicts survival of Layer-0 up to \(450\invfb\) of
data. This current monitoring system will be expanded to encompass
pixel disk modules in Run~2 and beyond.
\acknowledgments
The author expresses his grateful appreciation to the colleagues from
ATLAS Pixel Detector group for friendly cooperation and support
through all phases of the project on pixel sensor leakage current
monitoring and measurement.
|
\section{Introduction}
A~symbolic substitution is a morphism of the free monoid. More generally, a substitution rule acts on a finite collection of tiles by first inflating them, and then subdividing them into translates of tiles of the initial collection.
Substitutions thus generate symbolic dynamical system as well as tiling spaces.
The Pisot substitution conjecture states that any substitutive dynamical system has pure discrete spectrum, under the algebraic assumption that its expansion factor is a Pisot number (together with some extra assumption of irreducibility).
Pure discrete spectrum means that the substitutive dynamical system is measurably conjugate to a rotation on a compact abelian group.
Pisot substitutions are thus expected to produce self-similar systems with long range order.
This conjecture has been proved in the two-letter case; see \cite{Barge-Diamond:02} together with \cite{Host} or~\cite{Hollander-Solomyak:03}.
For more on the Pisot substitution conjecture, see e.g.\ \cite{CANTBST,ABBLS}.
We prove an extension of the two-letter Pisot substitution conjecture to the symbolic $S$-adic framework, that is, for infinite words generated by iterating different substitutions in a prescribed order.
More precisely, an $S$-adic expansion of an infinite word~$\omega$ is given by a sequence of substitutions $\boldsymbol{\sigma} = (\sigma_n)_{n \in \mathbb{N}}$ (called directive sequence) and a sequence of letters $(i_n)_{n \in \mathbb{N}}$, such that $\omega = \lim_{n\to\infty} \sigma_0 \sigma_1 \cdots \sigma_n(i_{n+1})$.
Such expansions are widely studied. They occur e.g.\ under the term `mixed substitutions' or `multi-substitution' \cite{Gahler:2013,PV:2013}, or else as `fusion systems' \cite{PriebeFrank-Sadun11,PriebeFrank-Sadun14}.
They are closely connected to adic systems, such as considered e.g.\ in~\cite{Fisher:09}. Fore more on $S$-adic systems, see \cite{Durand:00a,Durand:00b,Durand-Leroy-Richomme:13,Berthe-Delecroix}.
We consider the (unimodular) Pisot $S$-adic framework introduced in~\cite{Berthe-Steiner-Thuswaldner}, where unimodular means that the incidence matrices of the substitutions are unimodular.
The $S$-adic Pisot condition is stated in terms of Lyapunov exponents: the second Lyapunov exponent associated with the shift space made of the directive sequences
$\boldsymbol{\sigma}$ and with the cocyles provided by the incidence matrices of the substitutions is negative.
For a given directive sequence~$\boldsymbol{\sigma}$, let $\mathcal{L}_{\boldsymbol{\sigma}}^{(k)}$ stand for the language associated with the shifted directive sequence $(\sigma_{n+k})_{n \in \mathbb{N}}$.
The Pisot condition implies in particular a uniform balancedness property for $\mathcal{L}_{\boldsymbol{\sigma}}^{(k)}$ (i.e., uniformly bounded symbolic discrepancy), uniform with respect to some infinite set of non-negative integers~$k$. Note that the balancedness property is proved in \cite{Sadun:15} to characterize topological conjugacy under changes in the lengths of the tiles of the associated ${\mathbb R}$-action on the $1$-dimensional tilings.
One difficulty when working in the $S$-adic framework is that no natural candidate exists for a left eigenvector (that is, for a stable space).
Recall that the normalized left eigenvector (whose existence comes from the Perron-Frobenius Theorem) in the substitutive case provides in particular the measure of the tiles of the associated tiling of the line.
We introduce here a set of assumptions which, among other things, allows us to work with a generalized left eigenvector (that is, a stable space).
We stress the fact that this vector is not canonically defined (contrarily to the right eigenvector).
These assumptions will be implied by the $S$-adic Pisot condition.
We need first to guarantee the existence of a generalized right eigenvector~$\mathbf{u}$ (which yields the unstable space).
The corresponding conditions are natural and are stated in terms of primitivity of the directive sequence~$\boldsymbol{\sigma}$ in the $S$-adic framework.
We also require the directive sequences~$\boldsymbol{\sigma}$ to be recurrent (every finite combination of substitutions in the sequence occurs infinitely often), which yields unique ergodicity and strong convergence toward the generalized right eigendirection.
We then need rational independence of the coordinates of~$\mathbf{u}$.
This is implied by the assumption of algebraically irreducibility, which states that the characteristic polynomial of the incidence matrix of $\sigma_k \sigma_{k+1} \cdots \sigma_{\ell}$ is irreducible for all~$k$ and all sufficiently large~$\ell$.
Lastly, we then will be able to define a vector playing the role of a left eigenvector, by compactness together with primitivity and recurrence.
The starting point of the proofs in the two-letter substitutive case is that strong coincidences hold~\cite{Barge-Diamond:02}.
We also prove an analogous statement as a starting point. However the fact that strong coincidences hold at order $n$ (i.e.,
for the product $\sigma_0 \sigma_1 \cdots \sigma_n$) does not necessarily imply strong coincidences at order $m$ for $m > n$ (which holds in the substitutive case)
makes the extension to the present framework more delicate than it first occurs. The proof heavily relies, among other things, on the recurrence of the directive sequence.
There are then two strategies, one based on making explicit the action of the rotation on the unit circle~\cite{Host}, whereas the approach of \cite{Hollander-Solomyak:03} uses the balanced pair algorithm.
However, there seems to be no natural expression of the balanced pair algorithm in the $S$-adic framework.
Our strategy for proving discrete spectrum thus relies on the use of `Rauzy fractals' and follows~\cite{Host}.
We recall that a Rauzy fractal is a set which is endowed with an exchange of pieces acting on it that allows to make explicit (by factorizing by a natural lattice) the maximal equicontinuous factor of the underlying symbolic dynamical system (see~\cite[Section~7.5.4]{Fog02}). For more on Rauzy fractals, see e.g. \cite{Fog02,ST09,CANTBST}.
More precisely, the strong coincidence condition implies that there is a well-defined exchange of pieces on the Rauzy fractal.
It remains to factorize this exchange of pieces in order to get a circle rotation.
The factorization comes from \cite{Host} and extends directly from the substitutive case to the present $S$-adic framework.
\smallskip
Let us now state our results more precisely.
(Although most of the terminology of the statements was defined briefly in the introduction above, we refer the reader to Section~\ref{sec:ingredients} for exact definitions.)
We just recall here that a sequence of substitutions $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ on the alphabet $\mathcal{A} = \{1,2\}$ satisfies the \emph{strong coincidence condition} if there is $n \in \mathbb{N}$ such that $\sigma_0 \sigma_1 \cdots \sigma_n(1)\in p_1 i \mathcal{A}^*$ and $\sigma_0 \sigma_1 \cdots \sigma_n(2)\in p_2 i \mathcal{A}^*$ for some $i\in\mathcal{A}$ and words $p_1,p_2\in\mathcal{A}^*$ with the same abelianization $\mathbf{l}(p_1) = \mathbf{l}(p_2)$.
\begin{theorem} \label{scc}
Let $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ be a primitive and algebraically irreducible sequence of substitutions over $\mathcal{A} = \{1,2\}$. Assume that there is $C > 0$ such that for each $\ell \in \mathbb{N}$, there is $n \ge 1$ with $(\sigma_{n},\ldots,\sigma_{n+\ell-1}) = (\sigma_{0},\ldots,\sigma_{\ell-1})$ and the language $\mathcal{L}_{\boldsymbol{\sigma}}^{(n+\ell)}$ is $C$-balanced.
Then $\boldsymbol{\sigma}$ satisfies the strong coincidence condition.
\end{theorem}
Note that we do not assume unimodularity of the substitutions for strong coincidences, whereas we do for the following.
\begin{theorem}\label{sadic2}
Let $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ be a primitive and algebraically irreducible sequence of unimodular substitutions over~$\mathcal{A}=\{1,2\}$. Assume that there is $C > 0$ such that for each $\ell \in \mathbb{N}$, there is $n \ge 1$ with $(\sigma_{n},\ldots,\sigma_{n+\ell-1}) = (\sigma_{0},\ldots,\sigma_{\ell-1})$ and the language $\mathcal{L}_{\boldsymbol{\sigma}}^{(n+\ell)}$ is $C$-balanced.
Then the $S$-adic shift $(X_{\boldsymbol{\sigma}},\Sigma,\mu)$, where $\Sigma$ stands for the shift and $\mu$ is the (unique) shift-invariant measure on $X$, is measurably conjugate to a rotation on the circle~$\mathbb{S}^1$; in particular, it has pure discrete spectrum.
\end{theorem}
Let $S$ be a finite set of substitutions on $\mathcal{A}=\{1,2\}$ having invertible incidence matrices, and let $(D, \Sigma, \nu)$ with $D \subset S^\mathbb{N}$ be an (ergodic) shift equipped with a probability measure~$\nu$.
With each $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}} \in D$, associate the cocycle $A(\boldsymbol{\sigma}) = \tr{\!M}_0$ and denote the \emph{Lyapunov exponents} w.r.t.\ this cocycle by $\theta_1, \theta_2$. As in the more general situation of \cite[\S 6.3]{Berthe-Delecroix}, we say that $(D, \Sigma, \nu)$ satisfies the \emph{Pisot condition} if $\theta_1 > 0 > \theta_2$; see also \cite[Section 2.6]{Berthe-Steiner-Thuswaldner} for details.
With this notation, we are able to state the following theorem.
\begin{theorem}\label{th:ae}
Let $S$ be a finite set of unimodular substitutions on two letters, and let $(D, \Sigma, \nu)$ with $D \subset S^\mathbb{N}$ be a sofic shift that satisfies the Pisot condition. Assume that $\nu$ assigns positive measure to each (non-empty) cylinder, and that there exists a cylinder corresponding to a substitution with positive incidence matrix.
Then, for $\nu$-almost all sequences $\boldsymbol{\sigma} \in D$ the $S$-adic shift $(X_{\boldsymbol{\sigma}},\Sigma,\mu)$ is measurably conjugate to a rotation on the circle~$\mathbb{S}^1$; in particular, it has pure discrete spectrum.
\end{theorem}
Let us describe briefly the organization of the paper.
We recall the required definitions in Section~\ref{sec:ingredients}.
Section~\ref{sec:strongcoincidence} is devoted to the proof of the fact that the strong coincidence condition holds (Theorem~\ref{scc}), and we conclude the proof of Theorems~\ref{sadic2} and~\ref{th:ae} in Section~\ref{sec:conjecture}.
\section{Ingredients}\label{sec:ingredients}
The following $S$-adic framework is defined in full detail and for a finite alphabet $\mathcal{A}=\{1,\ldots,d\}$ in \cite{Berthe-Steiner-Thuswaldner}.
We introduce the reader to some of the main notions and results for $d=2$.
\subsection{$S$-adic shifts}
Let $\mathcal{A} = \{1,2\}$ and let $\mathcal{A}^*$ denote the free monoid of finite words over~$\mathcal{A}$, endowed with
the concatenation of words as product operation.
A~\emph{substitution} $\sigma$ over~$\mathcal{A}$ is an endomorphism of~$\mathcal{A}^*$ sending non-empty words to non-empty words.
The \emph{incidence matrix} (or abelianization) of~$\sigma$ is the square matrix $M_\sigma = (|\sigma(j)|_i)_{i,j\in\mathcal{A}} \in \mathbb{N}^{2\times 2}$,
where the notation $|w|_i$ stands for the number of occurrences of the letter~$i$ in $w \in \mathcal{A}^*$.
The substitution $\sigma$ is said to be \emph{unimodular} if $|\!\det M_\sigma| = 1$.
The \emph{abelianization map} is defined by $\mathbf{l}:\ \mathcal{A}^* \to\mathbb{N}^2, \ w \mapsto {}^t(|w|_1,|w|_2)$.
Note that $M_\sigma\circ\mathbf{l} = \mathbf{l}\circ\sigma$.
Let $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ be a sequence of substitutions over the alphabet~$\mathcal{A}$.
For ease of notation we set $M_n = M_{\sigma_n}$ for $n \in \mathbb{N}$, and
\[
\sigma_{[k,\ell)} = \sigma_k \sigma_{k+1} \cdots \sigma_{\ell-1} \quad \mbox{and} \quad M_{[k,\ell)} = M_k M_{k+1} \cdots M_{\ell-1} \quad (0 \le k \le \ell).
\]
The sequence~$\boldsymbol{\sigma}$ is said to be \emph{primitive} if, for each $k \in \mathbb{N}$, $M_{[k,\ell)}$ is a positive matrix for some $\ell > k$.
We say that $\boldsymbol{\sigma}$ is \emph{algebraically irreducible} if, for each $k \in \mathbb{N}$, the characteristic polynomial of $M_{[k,\ell)}$ is irreducible for all sufficiently large~$\ell$.
Recall that $w\in\mathcal{A}^*$ is called a \emph{factor} of a finite or infinite word $v$ if it occurs at some position in $v$; it is a \emph{prefix} if it occurs at the beginning of $v$.
The \emph{language} associated with the sequence $(\sigma_{m+n})_{n\in\mathbb{N}}$ is
\[
\mathcal{L}_{\boldsymbol{\sigma}}^{(m)} = \big\{w \in \mathcal{A}^*:\, \mbox{$w$ is a factor of $\sigma_{[m,n)}(i)$ for some $i \in\mathcal{A}$, $n\in\mathbb{N}$}\big\} \qquad (m \in \mathbb{N}).
\]
We say that a~pair of words $u, v \in \mathcal{A}^*$ with the same length is \emph{$C$-balanced} if
\[
-C \le |u|_j - |v|_j \le C \quad \mbox{for all}\ j \in \mathcal{A}.
\]
A~language~$\mathcal{L}$ is $C$-\emph{balanced} if each pair of words $u, v \in \mathcal{L}$ with the same length is $C$-balanced.
It is said \emph{balanced} if it is $C$-balanced for some $C>0$.
The \emph{shift}~$\Sigma$ maps an infinite word $(\omega_n)_{n\in\mathbb{N}}$ to $(\omega_{n+1})_{n\in\mathbb{N}}$.
A~dynamical system $(X,\Sigma)$ is a \emph{shift space} if $X$ is a closed shift-invariant set of infinite words over a finite alphabet, equipped with the product topology of the discrete topology.
Given a sequence~$\boldsymbol{\sigma}$, let $S = \{\sigma_n:n\in\mathbb{N}\}$.
The \emph{$S$-adic shift} or \emph{$S$-adic system} with sequence~$\boldsymbol{\sigma}$ is the shift space $(X_{\boldsymbol{\sigma}}, \Sigma)$, where $X_{\boldsymbol{\sigma}}$ denotes the set of infinite words~$\omega$ such that each factor of~$\omega$ is an element of~$\mathcal{L}_{\boldsymbol{\sigma}}^{(0)}$.
If $\boldsymbol{\sigma}$ is primitive, then $X_{\boldsymbol{\sigma}}$ is the closure of the $\Sigma$-orbit of any limit word $\omega$ of~$\boldsymbol{\sigma}$, where $\omega\in\mathcal{A}^\mathbb{N}$ is a \emph{limit word} of~$\boldsymbol{\sigma}$ if there is a sequence of infinite words $(\omega^{(n)})_{n\in\mathbb{N}}$ with $\omega^{(0)} = \omega$ and $\omega^{(n)} = \sigma_n(\omega^{(n+1)})$ for all $n \in \mathbb{N}$.
Recall that a shift space $(X,\Sigma)$ is \emph{minimal} if every non-empty closed shift-invariant subset equals the whole set; it is called \emph{uniquely ergodic} if there exists a unique shift-invariant probability measure on~$X$.
Let $\mu$ be a shift-invariant measure defined on $(X,\Sigma)$.
A~measurable eigenfunction of the system $(X,\Sigma,\mu)$ with associated eigenvalue $\alpha \in \mathbb{R}$ is an $L^2(X,\mu)$ function that satisfies $f(\Sigma^n(\omega)) = e^{2\pi i\alpha n} f(\omega)$ for all $n \in \mathbb{N}$ and $\omega \in X$.
The system $(X,\Sigma, \mu)$ has \emph{pure discrete spectrum} if $L^2(X,\mu)$ is spanned by the measurable eigenfunctions.
Furthermore, every dynamical system with pure discrete spectrum is measurably conjugate to a Kronecker system, i.e., a rotation on a compact abelian group; see~\cite{Walters:82}.
\subsection{Generalized Perron-Frobenius eigenvectors}
For a sequence of non-nega\-tive matrices $(M_n)_{n\in\mathbb{N}}$, there exists by \cite[pp.~91--95]{Furstenberg:60} a positive vector $\mathbf{u} \in \mathbb{R}_+^2$ such that
\[
\bigcap_{n\in\mathbb{N}} M_{[0,n)}\, \mathbb{R}^2_+ = \mathbb{R}_+ \mathbf{u},
\]
provided there are indices $k_1 < \ell_1 \le k_2 < \ell_2 \le \cdots$ and a positive matrix~$B$ such that $B = M_{[k_1,\ell_1)} = M_{[k_2,\ell_2)} = \cdots$.
Thus, for primitive and recurrent sequences $\boldsymbol{\sigma}$, this vector exists and we call it the \emph{generalized right eigenvector of} $\boldsymbol{\sigma}$.
The following criterion for $\mathbf{u}$ to have rationally independent coordinates is \cite[Lemma~4.2]{Berthe-Steiner-Thuswaldner}.
\begin{lemma}\label{le:rational}
Let $\boldsymbol{\sigma}$ be an algebraically irreducible sequence of substitutions with generalized right eigenvector~$\mathbf{u}$ and balanced language~$\mathcal{L}_{\boldsymbol{\sigma}}$. Then the coordinates of~$\mathbf{u}$ are rationally independent.
\end{lemma}
Contrary to the cones $M_{[0,n)}\, \mathbb{R}_+^2$, there is no reason for the cones $\tr{(M_{[0,n)})}\, \mathbb{R}_+^2$ to be nested.
Therefore, the intersection of these cones does not define a generalized left eigenvector of~$\boldsymbol{\sigma}$.
However, for a suitable choice of~$\mathbf{v}$, we have a subsequence $(n_k)_{k\in\mathbb{N}}$ such that the directions of $\mathbf{v}^{(n_k)} := \tr{(M_{[0,n_k)})}\, \mathbf{v}$ tend to that of~$\mathbf{v}$; in this case, $\mathbf{v}$ is called a \emph{recurrent left eigenvector}.
Under the assumptions of primitivity and recurrence of~$\boldsymbol{\sigma}$, given a strictly increasing sequence of non-negative integers~$(n_k)$, one can show that there is a recurrent left eigenvector $\mathbf{v} \in \mathbb{R}_{\ge0}^2 \setminus \{\mathbf{0}\}$ such that
\begin{equation}\label{eq:recurrentcandidate}
\lim_{k\in K,\,k\to\infty} \frac{\mathbf{v}^{(n_k)}}{\|\mathbf{v}^{(n_k)}\|} = \lim_{k\in K,\,k\to\infty}\frac{{}^t(M_{[0,n_k)})\mathbf{v}}{\|{}^t(M_{[0,n_k)})\mathbf{v}\|}= \mathbf{v}
\end{equation}
for some infinite set $K \subset \mathbb{N}$; see \cite[Lemma~5.7]{Berthe-Steiner-Thuswaldner}. Here and in the following, $\|\cdot\|$ denotes the maximum norm~$\|\cdot\|_\infty$.
Note that the hypotheses of Lemma~\ref{le:rational} do not guarantee that the coordinates of~$\mathbf{v}$ are rationally independent.
We will work in the sequel with sequences~$\boldsymbol{\sigma}$ satisfying a list of conditions gathered in the following Property PRICE (which stands for Primitivity, Recurrence, algebraic Irreducibility, $C$-balancedness, and recurrent left Eigenvector).
By \cite[Lemma~5.9]{Berthe-Steiner-Thuswaldner}, this property is a consequence of the assumptions of Theorem~\ref{scc}.
\begin{definition}[Property PRICE]\label{def:star}
We say that a sequence $\boldsymbol{\sigma} = (\sigma_n)$ has Property \emph{PRICE} w.r.t.\ the strictly increasing sequences $(n_k)_{k\in\mathbb{N}}$ and $(\ell_k)_{k\in\mathbb{N}}$ and the vector $\mathbf{v} \in \mathbb{R}_{\ge0}^2 \setminus \{\mathbf{0}\}$ if the following conditions hold.
\begin{itemize}
\labitem{(P)}{defP}
There exists $h \in \mathbb{N}$ and a positive matrix~$B$ such that $M_{[\ell_k-h,\ell_k)} = B$ for all $k \in \mathbb{N}$.
\labitem{(R)}{defR}
We have $(\sigma_{n_k}, \sigma_{n_k+1}, \ldots,\sigma_{n_k+\ell_k-1}) = (\sigma_0, \sigma_1, \ldots,\sigma_{\ell_k-1})$ for all $k\in\mathbb{N}$.
\labitem{(I)}{defI}
The directive sequence~$\boldsymbol{\sigma}$ is algebraically irreducible.
\labitem{(C)}{defC}
There is $C > 0$ such that $\mathcal{L}_{\boldsymbol{\sigma}}^{(n_k+\ell_k)}$ is $C$-balanced for all $k\in\mathbb{N}$.
\labitem{(E)}{defE}
We have $\lim_{k\to\infty} \mathbf{v}^{(n_k)}/\|\mathbf{v}^{(n_k)}\|= \mathbf{v}$.
\end{itemize}
We also simply say that $\boldsymbol{\sigma}$ satisfies Property PRICE if the five conditions hold for some not explicitly specified strictly increasing sequences $(n_k)_{k\in\mathbb{N}}$ and $(\ell_k)_{k\in\mathbb{N}}$ and some $\mathbf{v} \in \mathbb{R}_{\ge0}^2 \setminus \{\mathbf{0}\}$.
\end{definition}
\subsection{Rauzy fractals}
For a vector $\mathbf{w} \in \mathbb{R}^2\setminus\{\mathbf{0}\}$, let
\[
\mathbf{w}^\perp = \{ \mathbf{x}\in\mathbb{R}^2:\, \langle \mathbf{x},\mathbf{w}\rangle = 0 \}
\]
be the line orthogonal to~$\mathbf{w}$ containing the origin, equipped with the Lebesgue measure~$\lambda$.
In particular, for $\mathbf{1} = \tr{(1,1)}$, $\mathbf{1}^\perp$ is the line of vectors whose entries sum up to~$0$.
Let $\pi_{\mathbf{u},\mathbf{w}}$ be the projection along the direction~$\mathbf{u}$ onto~$\mathbf{w}^\perp$.
Given a primitive sequence of substitutions~$\boldsymbol{\sigma}$, the \emph{Rauzy fractal} associated with~$\boldsymbol{\sigma}$ over~$\mathcal{A}$ is:
\[
\mathcal{R} = \overline{\{\pi_{\mathbf{u},\mathbf{1}}\, \mathbf{l}(p):\, p \in \mathcal{A}^*,\ \mbox{$p$ is a prefix of a limit word of $\boldsymbol{\sigma}$}\}}.
\]
The Rauzy fractal has natural \emph{refinements} defined by
\[
\mathcal{R}(w) = \overline{\{ \pi_{\mathbf{u},\mathbf{1}} \, \mathbf{l}(p): p \in \mathcal{A}^*,\ \mbox{$pw$ is a prefix of a limit word of $\boldsymbol{\sigma}$} \}} \quad (w\in\mathcal{A}^*).
\]
If $w \in \mathcal{A}$, then $\mathcal{R}(w)$ is called a \emph{subtile}.
The set $\mathcal{R}$ is bounded if and only if $\mathcal{L}_{\boldsymbol{\sigma}}$ is balanced.
If $\mathcal{L}_{\boldsymbol{\sigma}}$ is $C$-balanced, then $\mathcal{R} \subset [-C,C]^2 \cap \mathbf{1}^\perp$; see \cite[Lemma~4.1]{Berthe-Steiner-Thuswaldner}.
Note that $\mathcal{R}$ is not necessarily an interval (however, it is an interval if the language $\mathcal{L}_{\boldsymbol{\sigma}}$ is Sturmian \cite{Fog02}.)
\subsection{Dynamical properties of $S$-adic shifts}\label{dynp}
For $\boldsymbol{\sigma}$ primitive, algebraically irreducible, and recurrent sequence of substitutions with balanced language~$\mathcal{L}_{\boldsymbol{\sigma}}$, the \emph{representation map}
\[
\varphi:\, X_{\boldsymbol{\sigma}} \to \mathcal{R},\quad u_0u_1\cdots \mapsto \bigcap_{n\in\mathbb{N}} \mathcal{R}(u_0u_1\cdots u_n)
\]
is well-defined, continuous and surjective; for more details, see \cite[Lemma~8.3]{Berthe-Steiner-Thuswaldner}.
Suppose that the strong coincidence condition holds.
Then the \emph{exchange of pieces}
\[
E: \mathcal{R}\to\mathcal{R},\quad \mathbf{x}\mapsto\mathbf{x}+\pi_{\mathbf{u},\mathbf{1}}\,\mathbf{e}_i\quad\text{ if } \mathbf{x}\in\mathcal{R}(i),
\]
is well-defined $\lambda$-almost everywhere on $\mathcal{R}$.
The following results appear in \cite[Theorem~1]{Berthe-Steiner-Thuswaldner}.
The assumptions on the directive sequence~$\boldsymbol{\sigma}$ are the ones of Theorem~\ref{sadic2}.
\begin{proposition}\label{sadicres}
Let $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ be a primitive and algebraically irreducible sequence of unimodular substitutions over~$\mathcal{A}=\{1,2\}$.
Assume that there is $C > 0$ such that for each $\ell \in \mathbb{N}$, there is $n \ge 1$ with $(\sigma_{n},\ldots,\sigma_{n+\ell-1}) = (\sigma_{0},\ldots,\sigma_{\ell-1})$ and the language $\mathcal{L}_{\boldsymbol{\sigma}}^{(n+\ell)}$ is $C$-balanced.
Then the following results are true.
\begin{enumerate}
\item\label{min}
The $S$-adic shift $(X_{\boldsymbol{\sigma}},\Sigma)$ is minimal and uniquely ergodic. Let $\mu$ stand for its unique invariant measure.
\item
Each subtile~$\mathcal{R}(i)$, $i \in \mathcal{A}$, of the Rauzy fractal~$\mathcal{R}$ is a compact set that is the closure of its interior; its boundary has zero Lebesgue measure~$\lambda$.
\item\label{1to1}
If $\boldsymbol{\sigma}$ satisfies the strong coincidence condition, then the subtiles $\mathcal{R}(i)$, $i\in\mathcal{A}$, are mutually disjoint in measure, and
the $S$-adic shift $(X_{\boldsymbol{\sigma}},\Sigma,\mu)$ is measurably conjugate to the exchange of pieces $(\mathcal{R},E,\lambda)$ via~$\varphi$.
\end{enumerate}
\end{proposition}
We will consider in the sequel the one-dimensional lattice $\Lambda := \mathbf{1}^\perp \cap \mathbb{Z}^2 = \mathbb{Z}(\mathbf{e}_2 - \mathbf{e}_1)$.
Let $\pi:\, \mathbf{1}^\perp \to \mathbf{1}^\perp/\Lambda$ be the canonical projection.
Since $\pi_{\mathbf{u},\mathbf{1}}\,\mathbf{e}_2 \equiv \pi_{\mathbf{u},\mathbf{1}}\,\mathbf{e}_1 \bmod \Lambda$ holds, the canonical projection of~$E$ onto $\mathbf{1}^\perp/\Lambda \cong \mathbb{S}^1$
is equal to the translation $\mathbf{x}\mapsto\mathbf{x} + \pi_{\mathbf{u},\mathbf{1}}\,\mathbf{e}_1$.
Then we have the following commutative diagram:
\begin{equation}\label{diag}
\begin{gathered}
\xymatrix{X_{\boldsymbol{\sigma}} \ar[r]^\varphi\ar[d]_\Sigma & \mathcal{R} \ar[r]^\pi\ar[d]_E & \mathbf{1}^\perp/\Lambda\ar[d]_{+\pi_{\mathbf{u},\mathbf{1}}\mathbf{e}_1} \\ X_{\boldsymbol{\sigma}} \ar[r]_\varphi & \mathcal{R}\ar[r]_\pi &\mathbf{1}^\perp/\Lambda }
\end{gathered}
\end{equation}
\section{Strong coincidence}\label{sec:strongcoincidence}
We recall the formalism of \emph{geometric substitution} introduced in \cite{Arnoux-Ito:01}.
For $[\mathbf{x},i] \in \mathbb{Z}^2\times\mathcal{A}$, let
\[
E_1(\sigma)[\mathbf{x},i] = \big\{ [M_\sigma \mathbf{x}+\mathbf{l}(p),j]:\, j\in\mathcal{A}, \, p\in\mathcal{A}^* \text{ and } pj \text{ is a prefix of }\sigma(i) \big\}.
\]
Then strong coincidence holds (on two letters) if and only if there exists $n \in \mathbb{N}$ such that $E_1(\sigma_{[0,n)})[\mathbf{0},1] \cap E_1(\sigma_{[0,n)})[\mathbf{0},2] \neq \emptyset$.
We identify each $[\mathbf{x},i] \in\mathbb{Z}^2\times\mathcal{A}$ with the segment $\mathbf{x} + [0,1)\, \mathbf{e}_i$.
Define the \emph{height} (with respect to~$\mathbf{u}$ and~$\mathbf{v}$) of a point $\mathbf{x} = t \mathbf{u} + \pi_{\mathbf{u},\mathbf{v}}\, \mathbf{x} \in\mathbb{R}^2$ by $H(\mathbf{x}) := t \in \mathbb{R}$.
According to the terminology introduced in \cite{Barge-Diamond:02}, a~\emph{configuration} (of segments) $\mathcal{K}$ of size~$m$ with respect to a vector~$\mathbf{w}$ is a collection of $m$ distinct segments $[\mathbf{x},i] \in \mathbb{Z}^2 \times \mathcal{A}$ such that some translate of~$\mathbf{w}^\perp$ intersects the interior of each element of~$\mathcal{K}$ (the corresponding points thus have the same height with respect to $\mathbf{u}$ and~$\mathbf{w}$).
The $n$-th iterate is
\[
\mathcal{K}^{(n)} = \big\{ E_1(\sigma_{[0,n)})[\mathbf{x},i]:\, [\mathbf{x},i] \in \mathcal{K} \big\}.
\]
Note that he $n$-th iterate of a configuration is not a configuration of segments but a union of ``broken lines''.
Observe that, by \cite[Proposition~4.3 and Lemma~4.1]{Berthe-Steiner-Thuswaldner} for a primitive, algebraically irreducible, and recurrent sequence of substitutions~$\boldsymbol{\sigma}$ with $C$-balanced language~$\mathcal{L}_{\boldsymbol{\sigma}}$, we have $\lim_{n\to\infty} \pi_{\mathbf{u},\mathbf{1}}\, M_{[0,n)} \mathbf{x} = \mathbf{0}$ for each $\mathbf{x} \in \mathbb{R}^2$ and $\| \pi_{\mathbf{u},\mathbf{1}}\,\mathbf{l}(p)\|\le C$ for all prefixes~$p$ of limit words of~$\boldsymbol{\sigma}$.
Thus, for each sufficiently large~$n$, the vertices of~$\mathcal{K}^{(n)}$ are in
\[
T_{\mathbf{u},C} := \{\mathbf{x} \in \mathbb{Z}^2:\, \|\pi_{\mathbf{u},\mathbf{1}}\, \mathbf{x}\| < C+1\},
\]
and we may consider only $\mathcal{K} \subset T_{\mathbf{u},C}$. (This corresponds to \cite[Lemma 2]{Barge-Diamond:02} which is stated in the substitutive case.)
In particular, $ \{[\mathbf{0},1], [\mathbf{0},2] \}$ is a configuration as soon as $\mathbf {w}$ has positive entries. Obviously, thus configuration it is contained in $ T_{\mathbf{u},C}$.
We say that $\mathcal{K}$ has an $n$-\emph{coincidence} if there exist
$[\mathbf{x},i], [\mathbf{y},j] \in \mathcal{K}$ such that $E_1(\sigma_{[0,n)})[\mathbf{x},i]\cap E_1(\sigma_{[0,n)})[\mathbf{y},j]\neq\emptyset$.
Given a set $J \subseteq \mathbb{N}$, we say that a configuration~$\mathcal{K}$ is $J$-\emph{coincident} if $\mathcal{K}$ has an $n$-coincidence for some $n\in J$.
Observe that $n$-coincidence does not necessarily imply $m$-coincidence for $m > n$.
However, translating all vertices of a configuration by a fixed vector does not change the property of being $J$-coincident.
We first prove the following proposition, generalizing the proof of~\cite[Theorem 1]{Barge-Diamond:02}.
\begin{proposition} \label{p:1}
Assume that the sequence of substitutions $\boldsymbol{\sigma} = (\sigma_n)_{n\in\mathbb{N}}$ over the alphabet $\mathcal{A} = \{1,2\}$ has Property PRICE w.r.t.\ the sequences $(n_k)_{k\in\mathbb{N}}$ and $(\ell_k)_{k\in\mathbb{N}}$ and the vector~$\mathbf{v}$, and that
\begin{equation}
\mathbf{v}^\perp \cap \mathbb{Z}^2 \cap (T_{\mathbf{u},C} - T_{\mathbf{u},C}) = \{\mathbf{0}\}, \label{e:vrat}
\end{equation}
with $C$ such that $\mathcal{L}_{\boldsymbol{\sigma}}$ is $C$-balanced.
Then $\boldsymbol{\sigma}$ satisfies the strong coincidence condition.
\end{proposition}
Note that (\ref{e:vrat}) holds in particular when~$\mathbf{v}$ has rationally independent coordinates.
\begin{proof}
Let $(n_k)_{k\in\mathbb{N}}$, $(\ell_k)_{k\in\mathbb{N}}$ be the sequences of property PRICE.
Consider the sets
\[
J_h = \big\{ n_{k_0} + n_{k_1} + \cdots + n_{k_s} : s \ge 0,\, n_{k_j} + \ell_{k_j} \leq \ell_{k_{j+1}} \ \forall\, 0 \le j < s,\, k_0 \ge h\big\}.
\]
Note that $k_{j } < k_{j+1}$, for $ 0 \le j < s$, since $n_{k_j} + \ell_{k_j} \leq \ell_{k_{j+1}}$ implies in particular that $\ell_{k_j} < \ell_{k_{j+1}}$.
Given such a sum in~$J_h$, repeatedly applying~\ref{defR} we get
\begin{equation}\label{rec}
\sigma_{[0,n_{k_0} + n_{k_1} + \cdots + n_{k_s})} = \sigma_{[0,n_{k_s})}\cdots \sigma_{[0,n_{k_1})} \sigma_{[0,n_{k_0})}.
\end{equation}
We only consider in this proof configurations with respect to~$\mathbf{v}$. Let $\mathcal{D}_h$ be the set of \emph{not} $J_h$-coincident configurations
that are contained in~$T_{\mathbf{u},C}$, and $\mathcal{D} = \bigcup_{h\in\mathbb{N}} \mathcal{D}_h$.
Since $J_0 \supset J_1 \supset \cdots$, we have $\mathcal{D}_0 \subseteq \mathcal{D}_1\subseteq \cdots$.
As $\mathbf{u} \in \mathbb{R}_+^2$ and $\mathbf{v} \in \mathbb{R}_{\ge0}^2 \setminus \{\mathbf{0}\}$ are not orthogonal (by Lemma~\ref{le:rational}), $\mathcal{D}$ contains only finitely many configurations up to translation.
Moreover, with each configuration $\mathcal{K} \in \mathcal{D}_h$, all translates of~$\mathcal{K}$ that are in~$T_{\mathbf{u},C}$ are also contained in~$\mathcal{D}_h$, by the translation-invariance of $J_h$-coincidence.
Therefore, we have $\mathcal{D}_h = \mathcal{D}$ for all sufficiently large~$h$.
Let $\mathcal{K}$ be a configuration in~$\mathcal{D}$ of maximal size.
There exists an interval~$I$ of positive length such that, for every $t \in I$, $\mathbf{v}^\perp + t \mathbf{u}$ intersects each of the segments of~$\mathcal{K}$ in its interior. Indeed, \eqref{e:vrat} implies that $\mathbf{v} $ has positive coordinates.
By property~\ref{defE}, the same holds for $(\mathbf{v}^{(n_k)})^\perp + t \mathbf{u}$, provided that $k$ is sufficiently large.
Consider now $\mathcal{K}^{(n_k)}$, with $k$ large enough such that all the following hold: $\mathcal{D}_k = \mathcal{D}$, all segments of~$\mathcal{K}$ intersect $(\mathbf{v}^{(n_k)})^\perp + t \mathbf{u}$ for all $t \in I$, and all vertices of $\mathcal{K}^{(n_k)}$ are in~$T_{\mathbf{u},C}$.
Let $\{\mathbf{p}_1, \mathbf{p}_2, \ldots, \mathbf{p}_{r_k}\}$ be the set of vertices of~$\mathcal{K}^{(n_k)}$ that are contained in $M_{[0,n_k)} \big((\mathbf{v}^{(n_k)})^\perp + I\, \mathbf{u}\big) = \mathbf{v}^\perp + I\, M_{[0,n_k)} \mathbf{u}$.
By~\eqref{e:vrat}, no two points in $\mathbb{Z}^2 \cap T_{\mathbf{u},C}$ have the same height. Therefore, we can assume w.l.o.g.\ that $H(\mathbf{p}_1) < H(\mathbf{p}_2) < \cdots < H(\mathbf{p}_{r_k})$.
For all $1 \le j \le r_k$, the configuration
\[
\mathcal{K}^{(n_k)}_j := \{ [\mathbf{x},i] \in\mathcal{K}^{(n_k)}:\, \big(\mathbf{x} + [0,1)\, \mathbf{e}_i\big) \cap \big(\mathbf{v}^\perp + H(\mathbf{p}_j)\, \mathbf{u}\big) \neq \emptyset \}
\]
has the same size as~$\mathcal{K}$ because $\mathcal{K}$ is not $J_k$-coincident and, for each $[\mathbf{x}',i'] \in \mathcal{K}$, the collection of segments in $E_1(\sigma_{[0,n_k)})[\mathbf{x}',i']$ forms a broken line from $M_{[0,n_k)} \mathbf{x}'$ to $M_{[0,n_k)} (\mathbf{x}'+\mathbf{e}_{i'})$ that intersects, for each $t \in I$, the line $\mathbf{v}^\perp + t\, \mathbf{u}$ exactly once. (Observe that the stripe $\mathbf{v}^\perp + I\, M_{[0,n_k)} \mathbf{u}$ has two complementary components, each of which contains one of the endpoints $M_{[0,n_k)} \mathbf{x}'$ and $M_{[0,n_k)} (\mathbf{x}'+\mathbf{e}_{i'})$ because $\mathbf{x}'$ and $\mathbf{x}'+\mathbf{e}_{i'}$ lie in different complementary components of the stripe $(\mathbf{v}^{(n_k)})^\perp + I \mathbf{u}$.)
Moreover, we have $\mathcal{K}^{(n_k)}_j \in \mathcal{D}$.
Indeed, take $h \ge k$ such that $\ell_h \ge n_k + \ell_k$.
If $\mathcal{K}^{(n_k)}_j$ were not in $\mathcal{D} = \mathcal{D}_h$, then $\mathcal{K}^{(n_k)}_j$ would have an $m$-coincidence for some $m \in J_h$.
Write $m= n_{k_0} + n_{k_1} + \cdots + n_{k_s}$. One has $m+n_k \in J_k$, since $ n_k +\ell_k \leq \ell_h \leq \ell_{k_0} $.
But then $\mathcal{K}$ would have an ($m+n_k$)-coincidence because
\[
E_1(\sigma_{[0,m)}) \mathcal{K}^{(n_k)} = E_1(\sigma_{[0,m)}) E_1(\sigma_{[0,n_k)}) \mathcal{K} = E_1(\sigma_{[0,m+n_k)})\mathcal{K},
\]
contradicting that $\mathcal{K} \in \mathcal{D} = \mathcal{D}_k$.
For all $2 \le j \le r_k$, the configurations $\mathcal{K}^{(n_k)}_{j-1}$ and $\mathcal{K}^{(n_k)}_j$ differ only by segments ending and beginning at~$\mathbf{p}_j$, and the number of segments in $\mathcal{K}^{(n_k)}$ ending and beginning at~$\mathbf{p}_j$ is thus the same.
Let $\mathcal{K}'$ be equal to $\mathcal{K}^{(n_k)}_{j-1}$, with the segments ending at~$\mathbf{p}_j$ removed.
We have the following possibilities.
\begin{enumerate}
\item\label{caseA}
Two segments of~$\mathcal{K}^{(n_k)}_{j-1}$ end at~$\mathbf{p}_j$.
Then $\mathcal{K}^{(n_k)}_j = \mathcal{K}' \cup \{[\mathbf{p}_j,1], [\mathbf{p}_j,2]\}$.
\item\label{caseB}
One segment of~$\mathcal{K}^{(n_k)}_{j-1}$ ends at~$\mathbf{p}_j$, and either $\mathcal{K}' \cup \{[\mathbf{p}_j,1]\}$ or $\mathcal{K}' \cup \{[\mathbf{p}_j,2]\}$ is in~$\mathcal{D}$.
Then $\mathcal{K}^{(n_k)}_j = \mathcal{K}' \cup \{[\mathbf{p}_j,i]\}$ with $i$ such that $\mathcal{K}' \cup \{[\mathbf{p}_j,i]\} \in \mathcal{D}$.
\item
One segment of~$\mathcal{K}^{(n_k)}_{j-1}$ ends at~$\mathbf{p}_j$, and both $\mathcal{K}' \cup \{[\mathbf{p}_j,1]\}$, $\mathcal{K}' \cup \{[\mathbf{p}_j,2]\}$ are in~$\mathcal{D}$.
Since the size of~$\mathcal{K}$ is maximal in~$\mathcal{D}$, we have $\mathcal{K}' \cup \{[\mathbf{p}_j,1], [\mathbf{p}_j,2]\} \notin \mathcal{D}$, hence $E_1(\sigma_{[0,n)}) [\mathbf{p}_j,1] \cap E_1(\sigma_{[0,n)}) [\mathbf{p}_j,2] \neq \emptyset$ for some $n \in \mathbb{N}$.
Then also $E_1(\sigma_{[0,n)}) [\mathbf{0},1] \cap E_1(\sigma_{[0,n)}) [\mathbf{0},2] \neq \emptyset$, thus the strong coincidence condition holds.
\end{enumerate}
Assume now that the strong coincidence condition does not hold.
Hence we are always either in case \eqref{caseA} or case~\eqref{caseB}.
Then $\mathcal{D}$ and the relative positions of the segments within~$\mathcal{K}^{(n_k)}_{j-1}$ entirely determine~$\mathcal{K}^{(n_k)}_j$.
(Note that $\mathbf{p}_j$ is the endpoint of the segments in $\mathcal{K}^{(n_k)}_{j-1}$, disregarding~$\mathbf{p}_{j-1}$, with minimal height.)
Recall that $\mathcal{D}$ contains up to translation only finitely many configurations, and denote the number of such configurations by~$c$.
Then we have $\mathcal{K}^{(n_k)}_{a+b} = \mathcal{K}^{(n_k)}_a + \mathbf{t}_k$ for some $1 \le a,b \le c$, and some translation vector~$\mathbf{t}_k = \mathbf{p}_{b+a} - \mathbf{p}_a \in\mathbb{Z}^2$ (provided that $r_k \ge 2c$).
Consequently, we have $\mathcal{K}^{(n_k)}_{j+b} = \mathcal{K}^{(n_k)}_j + \mathbf{t}_k$ for all $a \le j \le r_k-b$, and thus $\mathcal{K}^{(n_k)}_{a+\ell b} = \mathcal{K}^{(n_k)}_a + \ell \,\mathbf{t}_k$, for all $\ell$ such that $a + \ell b \leq r_k$.
Let now $k \to \infty$.
Then the stripe $\mathbf{v}^\perp + I\, M_{[0,n_k)} \mathbf{u}$ becomes wider and wider as~$k$ grows, hence $r_k \to \infty$.
Since there are only finitely many possibilites for~$\mathbf{t}_k$, there exists thus a $\mathbf{t} \in \mathbb{Z}^2$ such that arbitrarily large multiples of~$\mathbf{t}$ are translation vectors of configurations in~$\mathcal{D}$.
Since all configurations in~$\mathcal{D}$ are in~$T_{\mathbf{u},C}$, the vector~$\mathbf{t}$ must be a scalar multiple of~$\mathbf{u}$, in contradiction with Lemma~\ref{le:rational}.
This proves that the strong coincidence condition holds.
\end{proof}
To prove Theorem~\ref{scc}, we show that the conditions of Proposition~\ref{p:1} are fulfilled for some shifted sequence $(\sigma_{n+h})_{n\in\mathbb{N}}$.
\begin{proof}[Proof of Theorem~\ref{scc}]
Let $\boldsymbol{\sigma}$ satisfy the assumptions of Theorem~\ref{scc}.
By \cite[Lemma~5.9]{Berthe-Steiner-Thuswaldner}, property PRICE holds for some sequences $(n_k)$, $(\ell_k)$, a vector~$\mathbf{v}\in\mathbb{R}^2_{\ge 0}$ and a balancedness constant~$C$.
If $\mathbf{v}$ has rationally independent coordinates, then we can apply Proposition~\ref{p:1} directly. To cover the contrary case, assume in the following that $\mathbf{v}$ is a multiple of a rational vector.
Consider $\mathbf{x} \in \mathbb{Z}^2 \setminus \{\mathbf{0}\}$.
We have
\begin{align*}
\langle \mathbf{x},\mathbf{v}^{(h)} \rangle & = \langle \mathbf{x}, \tr{(M_{[0,h)})} \mathbf{v} \rangle = \langle M_{[0,h)} \mathbf{x}, \mathbf{v} \rangle = \langle H(M_{[0,h)} \mathbf{x}) \mathbf{u} + \pi_{\mathbf{u},\mathbf{v}}\, M_{[0,h)} \mathbf{x}, \mathbf{v} \rangle \\
& = H(M_{[0,h)} \mathbf{x})\, \langle \mathbf{u}, \mathbf{v} \rangle.
\end{align*}
As $\mathbf{u}$ has rationally independent coordinates by Lemma~\ref{le:rational} and $\mathbf{v}$ is a multiple of a rational vector, we have $\langle \mathbf{u}, \mathbf{v} \rangle \neq 0$.
We have $\lim_{h\to\infty} \pi_{\mathbf{u},\mathbf{v}}\, M_{[0,h)} \mathbf{x} = \mathbf{0}$ by \cite[Proposition~4.3]{Berthe-Steiner-Thuswaldner} and $M_{[0,h)} \mathbf{x} \in \mathbb{Z}^2 \setminus \{\mathbf{0}\}$ since $\boldsymbol{\sigma}$ is algebraically irreducible, thus $H(M_{[0,h)} \mathbf{x}) \ne 0$ for all sufficiently large~$h$.
We conclude that, for each $\mathbf{x} \in \mathbb{Z}^2 \setminus \{\mathbf{0}\}$, there is $h_0(\mathbf{x})$ such that
\begin{equation}\label{eq:hx}
\mathbf{x} \not\in (\mathbf{v}^{(h)})^\perp \hbox{ for all } h \ge h_0(\mathbf{x}).
\end{equation}
As $\mathbf{u} \in \mathbb{R}_+^2$, the set $\bigcup_{h\in \mathbb{N}} (\mathbf{v}^{(h)})^\perp \cap (T_{\mathbf{u},C} - T_{\mathbf{u},C})$ is bounded and, hence, its intersection with $\mathbb{Z}^2$ is finite. Thus \eqref{eq:hx} implies that $(\mathbf{v}^{(h)})^\perp \cap (T_{\mathbf{u},C} - T_{\mathbf{u},C}) \cap \mathbb{Z}^2 = \{\mathbf{0}\}$ for all sufficiently large~$h$.
Choose now $h = n_k + \ell_k$ sufficiently large.
Then the language $\mathcal{L}_{\boldsymbol{\sigma}}^{(h)}$ is $C$-balanced by Property~\ref{defC}.
By \cite[Lemma~5.10]{Berthe-Steiner-Thuswaldner}, there exists $k_0$ such that the shifted sequence $(\sigma_{n+h})_{n\in\mathbb{N}}$ has property PRICE w.r.t.~$(n_{k+k_0})$ and $(\ell_{k+k_0}-h)$ and the vector $\mathbf{v}^{(h)}$.
Thus $(\sigma_{n+h})_{n\in\mathbb{N}}$ satisfies the strong coincidence condition by Proposition~\ref{p:1}, i.e., there exists $n$ such that $E_1(\sigma_{[h,n+h)}) [\mathbf{0},1] \cap E_1(\sigma_{[h,n+h)}) [\mathbf{0},2] \neq \emptyset$.
This implies that $E_1(\sigma_{[0,n+h)}) [\mathbf{0},1] \cap E_1(\sigma_{[0,n+h)}) [\mathbf{0},2] \neq \emptyset$, which concludes the proof of the theorem.
\end{proof}
\section{Proof of the $S$-adic Pisot conjecture}\label{sec:conjecture}
In this section we will deduce Theorem~\ref{sadic2} from Theorem~\ref{scc} and the following lemma, which is a generalization of a result of \cite{Host}; see also \cite[Section~6.3.3]{Queffelec:10}.
Recall Section~\ref{dynp}, in particular the diagram~\eqref{diag}.
Theorem~\ref{th:ae} follows immediately from Theorem~\ref{sadic2}.
\begin{lemma}\label{host}
Let $\boldsymbol{\sigma}$ be as in Theorem~\ref{sadic2}.
Then the map $\overline{\varphi} = \pi \circ \varphi$ is one-to-one $\mu$-almost everywhere on $X_{\boldsymbol{\sigma}}$.
\end{lemma}
\begin{proof}
Note first that it is sufficient to show that whenever $\overline{\varphi}(u) = \overline{\varphi}(v)$, one can find a non-negative integer $n$ such that $\varphi(\Sigma^n u) = \varphi(\Sigma^n v)$.
The $\mu$-almost everywhere injectivity of~$\overline{\varphi}$ will thus come from the $\mu$-almost everywhere injectivity of~$\varphi$, which holds according to Proposition~\ref{sadicres}~\eqref{1to1}.
Let now $u,v \in X_{\boldsymbol{\sigma}}$ be such that $\overline{\varphi}(u) = \overline{\varphi}(v)$. Consider the set
\[
\big\{n \in \mathbb{N}:\, \mathbf{z}_n := \varphi(\Sigma^n u) - \varphi(\Sigma^n v) = 0\big\}.
\]
By induction, one has $\mathbf{z}_n \in \Lambda = \mathbb{Z}\, (\mathbf{e}_2-\mathbf{e}_1)$ for all~$n$.
Indeed, $\mathbf{z}_0 \in \Lambda$ because $\overline{\varphi}(u) - \overline{\varphi}(v) = 0$.
Using $E\circ\varphi = \varphi\circ\Sigma$ we see, for all~$n$, that
\[
\mathbf{z}_{n+1} - \mathbf{z}_n = \big(E^{n+1}(\varphi(u)) - E^n(\varphi(u))\big) - \big(E^{n+1}(\varphi(v)) - E^n(\varphi(v))\big) \in \big\{\mathbf{0},\pm (\mathbf{e}_2-\mathbf{e}_1)\big\}.
\]
Let $\pi_0 : \mathbf{1}^\perp \to \mathbb{R}$, $(x,-x)\mapsto x$ be the projection on the first coordinate.
Then $z_n := \pi_0(\mathbf{z}_n) \in\mathbb{Z}$ and $z_{n+1}-z_n \in \{0,\pm 1 \}$ for all $n$.
Since $\mathcal{L}_{\boldsymbol{\sigma}}$ is $C$-balanced, we know that $\varphi(X_{\boldsymbol{\sigma}}) = \mathcal{R} \subset [-C,C]^2 \cap \mathbf{1}^\perp$, thus $\pi_0(\varphi(\Sigma^n u)) \in [-C,C]$ for any $u\in X_{\boldsymbol{\sigma}}$.
Minimality of $(X_{\boldsymbol{\sigma}},\Sigma)$, given by Proposition~\ref{sadicres}~\eqref{min}, implies that the sets
\begin{align*}
A &= \{ n \in \mathbb{N} : \pi_0(\varphi(\Sigma^n u)) > C-1 \}, \\
B &= \{ n \in \mathbb{N} : \pi_0(\varphi(\Sigma^n u)) < -C + 1\}
\end{align*}
are relatively dense.
If $n\in A$, then $\pi_0(\varphi(\Sigma^n v))\leq C < \pi_0(\varphi(\Sigma^n u)) + 1$ and $z_n\geq 0$. Analogously, we deduce that, if $n\in B$, then $z_n\leq 0$. The result follows observing that, given $p\in A$, $q\in B$ such that $p\leq q$, we can find~$n$ such that $p\leq n \leq q$ and $z_n = 0$, using the fact that $z_{n+1} - z_n \in \{0,\pm 1\}$ for all $n$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{sadic2}]
By Theorem~\ref{scc}, the strong coincidence condition holds, which implies that $(X_{\boldsymbol{\sigma}},\Sigma,\mu)$ is measurably conjugate to $(\mathcal{R},E,\lambda)$ via~$\varphi$ by Proposition~\ref{sadicres}~\eqref{1to1}. But by Lemma~\ref{host} we even have that $(X_{\boldsymbol{\sigma}},\Sigma,\mu)$ is measurably conjugate to $(\mathbf{1}^\perp/\Lambda,+\pi_{\mathbf{u},\mathbf{1}}\,\mathbf{e}_1,\lambda)$ via~$\overline{\varphi}$, with $\mathbf{1}^\perp/\Lambda \cong \mathbb{S}^1$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th:ae}]
In view of Theorem~\ref{sadic2}, this is an immediate consequence of \cite[Theorem~2]{Berthe-Steiner-Thuswaldner}.
\end{proof}
\bibliographystyle{amsalpha}
|
\section{Introduction}
No real physical systems could be regarded as a purely closed system, as they
are inevitably interacting with environments. Therefore, the dynamic of the
open quantum systems is in the central of the fundamental quantum mechanics
and quantum information
science.~\cite{Breuer2007} Non-Markovianity~\cite{Angel} is the prominent
concept in open systems and attracts a lot of attention from theoretical and
experimental aspects.~\cite{Chin2012,Bylicka2013,
Vsile2011a,Breuer2009,Laine2010,Vasile2011b,
Luo2012,Lorenzo2013} It has been shown that the non-Markovianity
can be exploited as useful resource in quantum technology. For examples: it
may be used to improve efficiency of quantum information processing and
communication;~\cite{Bylicka2013} it is benefit in quantum metrology;
~\cite{Chin2012} and it may be used to improve the security in
continuous-variable quantum key distribution,~\cite{Vasile2011} etc. It is
therefore important to quantify the non-Markovianity of a quantum
system. However, despite of the extensive investigation, characterization and
measurement of the non-Markovianity of an open system is still limited to very
small systems,~\cite{Lu2010,Xu2010,Laine2012} mostly one-qubit system in
experiments.~\cite{Liu2011}
The measurement base on trace distance, proposed by Breuer, Laine and Piilo
(BLP),~\cite{Breuer2009} is one of the most popular definitions for
non-Markovianity, and is widely used in theoretical and experiment
investigations.~\cite{Liu2011,Liu2013}
To quantify the non-Markovianity by BLP measurement, one needs to find a pair
of initial states to maximal a function based on the trace distance. Only when
the interactions between the system and the environment are exactly known (the
dimension of the whole system is not too huge), the optimal state pair and the
measure of the non-Markovianity can be found numerically. For very limited
quantum open systems which can be exactly solved, such as the Jaynes-Cummings
model~\cite{Dalton2001} and quantum Brownian motion
model,~\cite{Intravaia2013} the measure of the non-Markovianity can be
analytically found. However, for general quantum open systems where we have no
exact information about the interaction between the system and the environment
(or the dimension of the whole system is large), the optimal initial state
pair can only be found experimentally by scanning the dynamics of the whole
initial state spaces which is a tough task even for two-dimension systems. In
a typical experiment~\cite{Liu2014} to quantify the non-Markovianity in a
two-dimension quantum open system, in order to achieve a reasonable accuracy,
total 5000 states' dynamics were measured. Even worse, the number of scanning
states grows exponentially with the dimension of the system. For a system
containing two spin-1/2 qubits (which can be viewed as a 4-dimension system),
one need to scan about $D^4$ different initial states to obtain the
non-Markovianity, where $D$ is the number of samples for each degree of
freedom. Typically, $D$ should be 100 or even more to ensure the
accuracy. Therefore even quantification the non-Markovianity of this simple
two qubits system is beyond our current experimental ability.
In this work, inspired by the idea of standard quantum process
tomography,~\cite{Chuang1997} we propose an efficient method, combining
handful experiments (polynomial scaled with the dimension
of the system) and the numerical optimization method to measure the
non-Markovianity for high dimension systems (The systems containing more than
one qubits can be regarded as high dimension systems). This method require no
prior information about the interactions between the system and its
environment. Due to the linearity of the dynamics of the quantum system, we
need only experimentally measure the dynamics of some linearly independent
states of the system, and the dynamics of the whole state space can be rebuilt
by linear combination of these experimental results. We then find the optimal
state pair in quantifying non-Markovianity through numerical calculation based
on these experiment data. The number of the measurement is scaled as $N^2$,
where $N$ is the dimension of the quantum system. Using this method, the
former intractable non-Markovianity measurement of high dimension system can
be easily investigated. After the benchmark on a two-dimension open system, we
demonstrate our algorithm to quantify the non-Markovianity of a high dimension
open quantum random walk system.
\section{Methods}
The BLP measurement of the non-Markovianity~\cite{Breuer2009} is defined on the trace distance of two state $\rho_1$, $\rho_2$, that is,
$D(\rho_1,\rho_2)=\frac{1}{2}{\rm tr}|\rho_1-\rho_2|$ where $|A|=\sqrt{A^\dag A}$. If $A$ is Hermitian, ${\rm tr}|A|=\sum_i |\lambda_i|$ is the sum of the
absolute value of all the eigenvalue of matrix $A$. This quantity describe the distinguishability between the states $\rho_1$ and $\rho_2$:
if it is zero, the two states are indistinguishable, otherwise, they are
distinguishable. Based on this definition, the measure of the non-Markovianity
of an open system can be defined as,
\begin{equation}\label{eq:max_non}
\mathcal{N}={\rm max}_{\rho_{1,2}(0)}\int_{\delta>0}dt\, \delta(t,\rho_{1,2}(0))\, ,
\end{equation}
where $\delta(t,\rho_{1,2}(0))=\frac{d}{dt}D[\rho_{1}(t),\rho_{2}(t)]$ is the
change rate of the trace distance. $\rho_{i}(t)$, $i$=1, 2 is the density
matrix of the open system at time $t$ with the initial state $\rho_{i}(0)$. The time-integration is extended over all time
intervals in which $\delta$ is positive, and the maximum should be optimized
over all pairs of initial states.
Roughly speaking, the integral intervals
stand for the time intervals when the information flows back to the system from the environment.
The most difficult task to measure the non-Markovianity is to find a
pair of states, $\rho_1$, $\rho_2$ that maximize
Eq.~(\ref{eq:max_non}). Generally, it need experimentally scan the state pairs
in the whole parameter space. Some simplifications can be
made.~\cite{Liu2014,Wibmann2012}
It has been rigorously proven that the optimal
states pair should be on the boundary of the physical state space and the
states pair are orthogonal each other. Therefor, we can only scan the boundary
of physical state space (For two-dimension case, these states are pure
states). Another simplification to the measure was demonstrated in
Ref. \onlinecite{Liu2014}, which illustrated that the measure can be obtained efficiently
in an arbitrary neighborhood of any fixed state in the interior of the state
space. That is, it needs only scan one state of the pair in the physical state
space. This can dramatically reduce the experimental work. However, the number
of the experiments is still too large and will exponentially increase with the
dimension of the systems. Therefore, the non-Markovianity in higher dimension
is still intractable to quantify with the current method.
Now we introduce another scheme to simplify the experimental quantification of
non-Markovianity in an open system which make the high dimension system reachable.
Following the idea of the quantum process tomography,~\cite{Chuang1997} the
state of a quantum open system with $N$ dimension can be expressed as
$N$$\times$$N$ density matrix. Any density matrix can be expanded by $N^2$
linear independent bases,
\begin{eqnarray}\label{rho}
\rho^x_{mn}&=& (|m\rangle\langle n|+|n\rangle\langle m|)/2,\;\; (m>n)
\nonumber \\
\rho^y_{mn}&=&i(|m\rangle\langle n|-|n\rangle\langle m|)/2,\;\; (m>n)
\nonumber \\
\rho^0_m&=&|m\rangle\langle m|)\, ,
\end{eqnarray}
where $|m\rangle$ ($m=1,2,\cdots,N$) is the basis vector of the system. The
operators $\rho^x_{mn}$ ($\rho^y_{mn}$) play the similar role of the pauli
matrices $\sigma^x$ ($\sigma^y$) in two-dimension systems. Without loss of
generality, we assume that the system and the environment is in a
product state at the initial time $t=0$,
i.e. $\rho(0)=\rho_s(0)\otimes\rho_{e}(0)$, where $\rho(t)$, $\rho_s(t)$,
$\rho_{e}(t)$ are the density matrices
of the whole system (system+enviroment), the quantum system and the
environment at time $t$, respectively. Using the above introduced bases, the
state of the open system
at any time $t$ can be written as:
\begin{eqnarray}
\label{eq:rho}
\rho_s(t)=\sum_{m>n}a^x_{mn} \rho^x_{mn}(t)+\sum_{m>n}a^y_{mn}\rho^y_{mn}(t)+\sum_m a^0_m \rho^0_m(t)\,,
\end{eqnarray}
where $a^x_{mn}$, $a^y_{mn}$, $a^0_m$ are time independent constants
determined by the initial states.
$\rho^i_{mn}(t)$ $i=x,y,0$ is the dynamics of the bases.
It suggests that the dynamics of the open system can be completely determined by the dynamics of the bases.
In experiments, the dynamics of the $N^2$ bases can be obtained by the following procedure:
\begin{enumerate}
\item Prepare the initial states of the system to $|m\rangle$,
where $m$=1, 2, $\cdots$, $N$.
$|m\rangle$ can be any set of complete and orthogonal vector bases of the system.
By measuring the dynamics of the open system, we obtain $\rho^0_m(t)$.
\item Prepare the initial states of the system to
$(|m\rangle+|n\rangle)/\sqrt{2}$ ($m>n$), measuring the dynamics of the open
system. We obtain $\rho^x_{mn}(t)+\frac{1}{2}(\rho^0_m(t)+\rho^0_n(t))$.
\item Prepare the initial states of the system to
$(|m\rangle+i|n\rangle)/\sqrt{2}$ ($m>n$), measuring the dynamics of the open
system. We obtain $-\rho^y_{mn}(t)+\frac{1}{2}(\rho^0_m(t)+\rho^0_n(t))$.
\end{enumerate}
We therefore have the dynamics of the $N^2$ bases of the open system.
Using the dynamics of these bases, the BLP measure of the non-Markovianity of the open system
can be achieved by numerically optimizing the parameters, $a^x_{mn}$, $a^y_{mn}$ and $a^0_{m}$, of the initial states
by computer using Eq.~(\ref{eq:max_non}),and Eq.~(\ref{eq:rho}) through deepest descent algorithm.
The nice scale of this method make it possible to apply for the high-dimension system which is intractable for the traditional method.
It is worth noting that the basis introduced here is not unique.
Any $N^2$ such linear independent states are enough for the procedure. The
simplifications introduced in Ref.~\onlinecite{Wibmann2012} and Ref.~\onlinecite{Liu2014},
can also be used to reduce the numerical optimization efforts in this method.
\section{Results and discussion}
\subsection{Benchmark for the two-dimension system}
To demonstrate the power of our scheme, we first make benchmark tests on the
well studied two-dimension open system. In the typical
experiment,~\cite{Liu2014}
the two-dimension open quantum system is provided by the
polarization of a photon which coupled to the environment through its
frequency degree. To quantify the non-Markovianity of this system, 5000
different states in the Bloch surface have been scanned to find the optimal
pair in Eq.(1). The conclusion in \cite{Wibmann2012} has been used to simplify
the experiment,
which states that the optimal state pairs are orthogonal pure states, and
therefore, function $\mathcal{N}$ in Eq.(1) only depends on the angles between
two states on the Bloch surface.
\begin{figure}
\centering
\includegraphics[width=3.2in]{fig1.eps}
\caption{(Color online) Comparing the trace distance directly measured from
experiment and the results obtained from our method.
The blue circles show the trace distance as function of time
of the experimentally determined optimal state pair, by scanning
the whole state space. The red squares show the trace distance of the optimal pair as function of time,
obtained by four basis dynamics followed by numerical optimization, as described in the main text.
}
\label{fig:compare}
\end{figure}
A $2\times2$ density matrix can be expanded by identity matrix $I$ and pauli matrixes $\sigma_x$, $\sigma_y$, $\sigma_z$ as
$\rho_s=\frac{1}{2}(I+\vec{a} \cdot \vec{\sigma})$, where vector $\vec{a}$ is
on the surface of the Bloch sphere. Using our method,
non-Markovianity measure of this open system can be determined by the dynamics
of the pauli matrices, i.e. $\sigma_i(t)$, where $i$=$x$, $y$, $z$ which can
obtained from the dynamics of the four initial states, $|1\rangle$,
$|-1\rangle$ , $\frac{1}{\sqrt2}(|1\rangle+|-1\rangle)$ and
$\frac{1}{\sqrt2}(|1\rangle+i|-1\rangle)$ (where $|-1\rangle$ and $|1\rangle$
are two eigenvectors of the polarization photon).
Fortunately, the dynamics of these four initial states can be directly taken
from the experimental data in Ref.~\onlinecite{Liu2014}. Using these data,
we can completely determine the dynamics of any initial states using
Eq.~(\ref{eq:rho}). We then numerically find the optimal initial states pair,
and the non-Markovianity of this system. The trace distance for the optimal
initial states pair
as a function of time obtained using our scheme is compared to the directly measured one in Fig.~\ref{fig:compare}.
As we see they are in excellent agreement.
The non-Markovianity obtained from our method is 0.6, which are very close to
the measured value 0.58, both are in good agreement with the theoretical value
0.59.
The error in our method is due to the error in the measurement of the 4 basis states.
\subsection{Open quantum walk system}
\begin{figure}
\begin{center}
\includegraphics[width=3.2in]{fig2.eps}
\caption{ (color online) Cartoon of the 3-step open quantum walk system, with $X$=2.
The blue blocks denote the polarized beam splitter.
This quantum walk system includes both the lattice (location of the photon)
and the coin (polarization) which is a intrinsically high-dimension system.}
\label{QW2}
\end{center}
\end{figure}
With the confidence of the method in two-dimension open system, we apply this
method to a high-dimension open system which can not be reached
previously. Here, we demonstrate our method to qualify the non-Markovianity of
an one-dimension open quantum walk (QW) system.~\cite{ABK1,ABK2,ABK3}
which is intrinsically a high-dimension system due to the ansatz coin. QW
system is a generally interested system in quantum information, which has a
lot of application in quantum computation.
It has been shown that it is a nice tool to find new quantum algorithm and it
can be used to constitute a universal model of quantum
computation.~\cite{Childs,Venegas2012} In addition, QW has been experimentally
realized in several different
systems.~\cite{Schreiber2010,Karski2009,Peruzzo2010,Schmitz2009,Broome2010} The
open QW has attracted a lot of attention recently.~\cite{ABK1,ABK2,ABK3}
Here we study the discrete-time QW on a one-dimension lattice. The particle is
located at one site at the beginning. At each step, it can move either to the
left or to the right which is determined by the state of a coin: $|L\rangle$ (move
left) or $|R\rangle$ (move right). In quantum walk, the state of the coin can
be a superposed state. Therefore, the state of the whole QW system (including
the particle in the lattice and the coin) is
$|\psi\rangle=\sum_{x,d}C_{x,d}|x\rangle|d\rangle$ (in open QW, it should be a
density matrix) where $x=0,\pm1,\pm2,\cdots$ are the location of the particle
and $d=L,R$ are the state of the coin. The operator to make up a single step
of the QW can be defined as: $W=TC$, where $T$ is the shift operator and
defined as $T=\sum_j|j-1\rangle\langle j|\otimes| L\rangle\langle
L|+|j+1\rangle\langle j|\otimes| R\rangle\langle R|$, $C$ is the Hadamard coin
operator defined as $C=\frac{1}{\sqrt 2}(|L\rangle\langle L|+|L\rangle\langle
R|+|R\rangle\langle L|-|R\rangle\langle R|)$. The state of the system after
$N_{step}$ steps with the initial state $|\psi_0\rangle$ can be obtained as
$W^{N_{step}}|\psi_0\rangle$.
\begin{figure}
\centering
\includegraphics[width=3.2in]{fig3.eps}
\caption
(Color online) The trace-distances as functions of time for the optimal state pairs
obtained using our method for $X$=0 (red line), 1 (blue line) and 2 (green line).}
\label{toy}
\end{figure}
The discrete-time QW can be implemented with single photon through
an array of beam splitters in which the coin states are mimicked by
the polarization degrees of freedom. For the open QW, the environment can be
introduced through the coupling between the frequency and the polarization
degrees of freedom
of the photon similar to the method used in Ref.~\onlinecite{Liu2014}. In this case, the
single step operator in the open QW can be modified as:
\begin{eqnarray}
U_{step} &=& U_{\delta t}TC, \\
U_{\delta t} &=& \int d\omega\sum_{p=H,V}e^{in_{p}\omega\delta
t_p}|p\rangle\langle p|\otimes|\omega\rangle\langle\omega| \, ,
\label{eq:coupling}
\end{eqnarray}
where $T$ and $C$ are shift operator and Hadamard coin operator defined
before. $U_{\delta t}$ couples the polarization($H$ or $V$) and environment to
give non-Markovianity. $n_p$ is the index of refraction for different
polarization state. $\delta t_p=\frac{L}{v_p}$ is the time that operation $TC$
takes place where $L\sim 0.5 mm$ is the thickness of the beam splitters and
$v_p$ is velocity of light with polarization $p$ in the
splitters. $|\omega\rangle$ is the environment state.
For convenience, we take the environment as a delta function , that is
$|\omega\rangle=\frac{1}{\sqrt2}[\delta(\omega-\omega_1)+\delta(\omega-\omega_2)]$. And $\omega_1=\Omega-\omega_0$, $\omega_2=\Omega+\omega_0$.
We take the value of the parameters from the Ref.~\onlinecite{Liu2014}:
$\omega_0=7.2\times 10^{12} s^{-1}$, $\Omega=\frac{2\pi c}{\lambda}$, where
$\lambda$=780 nm, thus $\Omega=2.4166\times10^{15}s^{-1}$. $n_H=1.554$,
$n_V=1.545$, $\delta t_H=Ln_H/c\sim 1.036\times 10^{-11}s$ and $\delta
t_V\sim 1.030\times 10^{-11}s$. In addition, we use the periodic boundary
condition for the system, i.e. $|x_{min}-1\rangle=|x_{max}\rangle,
|x_{max}+1\rangle=|x_{min}\rangle$ (see Fig.2), and we allow the system
evolves for 20 steps.
\begin{table}
\caption{Comparing the number of initial states to measure
between the present method and the direct scanning method
to obtain the non-Markovianity of the quantum walk system.
For the direct method, we assume 100 initial states to measure for each degree of freedom.
The non-Markovianity of the system is calculated using the present method.}
\begin{center}
\begin{tabular}{cccc}
\hline\hline
& \multicolumn{2}{c}{number of initial states} & \\
X & Direc method & Present method & non-Markovianity \\
\hline
0 & 100$^2$ & 4 & 0.9512 \\
1 & 100$^6$ & 36 & 0.9510 \\
2 & 100$^{10}$ & 100 & 0.9428 \\
\hline\hline
\end{tabular}
\label{table}
\end{center}
\end{table}
For convenience, we set the location of the particle $x$$\in$[-$X$, $X$], then the total dimension of the system is $N$=2$(2X+1)$.
To experimentally measure the non-Markovianity of this system, we need choose $N^2$ linearly independent initial states.
As introduced before, we implement experiments to get the dynamics of the following initial states:
\begin{eqnarray}\label{}
|\psi\rangle&=&|x\rangle|d\rangle\, \\
|\psi\rangle&=&\frac{1}{\sqrt2}(|x_1\rangle|d_1\rangle)+|x_2\rangle|d_2\rangle), \\
|\psi\rangle&=&\frac{1}{\sqrt2}(|x_1\rangle|d_1\rangle)+i|x_2\rangle|d_2\rangle)
\end{eqnarray}
here $x$,$x_1$,$x_2$=-$X$,$\cdots$,$X$ and $d$,$d_1$,$d_2$=$H$, $V$. With the experimental data of these initial states,
the dynamics of any initial state can be linearly constructed, and the optimal state pair can be found
by numerically searching the configuration space.
Because we do not have the experimental data at hand, we study the dynamics of
the QW system numerically using the exact diagonalization method.
We first calculate the
dynamics of the bases, then search the optimal initial states pair numerically
described in previous paragraph, and determine the non-Markovianity. We
collect the results of the trace distances of the optimal state pair for
$X$=0, 1, 2 in Fig.~\ref{toy}. The non-Markovianity for $X$=0, 1, 2 are given
in Table \ref{table}. Actually, the result can be further confirmed via
directly exact diagonalization method in which we obtain exactly the same results.
We also compare the number of the initial states experimentally necessary to
determine the non-Markovianity of the open QW system in Table \ref{table}. The
number scales as 100$^{N}$, where $N$=4$X$+2 if previous direct scanning
method is used. Clearly it is impossible to experimentally determine the
non-Markovianity of the system for $X>$0 using this method. However, by using
our method, in which the number of initial states is only $N^2$, we are able
to measure the non-Markovianity of much larger systems.
\section{Summary}
We have introduced a experimental method to quantify the
non-Markovianity of high-dimension open quantum system. In our method, the
scaling of the experiment is only $N^2$ which is dramatically reduced from
exponential scaling of the conventional method. Therefore, the system which is
intractable by the former method can be easily reached with the current
method. After the benchmark with the well studied two-dimension open system,
we demonstrate the method to the high dimension open quantum walk system. This
method therefore opens up a new path to experimentally study the
non-Markovianity of high dimension open quantum system, which was impossible
previously.
The authors acknowledge the support from the Chinese National Fundamental
Research Program 2011CB921200, the National Natural Science Funds for
Distinguished Young Scholars and NSFC11374275, 11105135, 11474267, 11104261,
11374288 and the Central Universities WK2470000006, WJ2470000007.
|
\section{Creation of density-dependent gauge fields}
\subsection{Two-component system}
\label{subsec:2comp}
We consider two hyperfine states, A and B, of a bosonic or fermionic atom.
Both components are confined to the lowest band of a tilted 1D optical lattice along $x$, of spacing $D$, and depth $V_0=sE_R$, with $E_R=\hbar^2 \pi^2/2mD$
the recoil energy. Without tilting there is a hopping rate $J$ to nearest neighbors. The lattice tilting induces an energy shift $\Delta$ from site to site.
We denote as $w(x-j D)$ the Wannier function maximally localized at site $j$. Due to the tilting, it is convenient
to use Wannier-Stark states. For $J\ll \Delta$, the Wannier-Stark state centered at site $l$ may be approximated as
$\psi_{l}(x)\simeq w(x-lD)+\frac{J}{\Delta}(w(x-(l+1)D)-w(x-(l-1)D))$.
The 3D on-site wavefunction at site $j$
is $\Phi_j({\bf r})=\psi_j(x)\varphi(y,z)$, where $\varphi(y,z)$ is given by the strong transversal confinement. For simplicity
we assume below $\varphi(y,z)\simeq w(y)w(z)$.
On-site interactions between $\alpha$ and $\beta$~($\alpha,\beta=$A,B) are
characterized by the coupling constant
$U_{\alpha,\beta}=\frac{4\pi\hbar^2a_{\alpha,\beta}}{m}\int d^3{\mathbf r} |w({\mathbf r})|^4$, with $a_{\alpha,\beta}$ the corresponding scattering length.
For a sufficiently deep lattice, the evaluation of the on-site interactions
is simplified by means of the harmonic approximation: $\Phi({\bf r})\simeq (\sqrt{\pi} l)^{-3/2}e^{-r^2/l^2}$,
where $l=D s^{-1/4}/\pi$. Using this approximation we obtain:
$U_{\alpha,\beta}\simeq \pi (2\pi)^{3/2} s^{3/4}\frac{\hbar^2 a_{\alpha,\beta}}{mD^3}$. Of course, for fermions only inter-species
on-site interactions are possible.
\subsection{Laser transitions}
\label{subsec:Lasers}
Before discussing Raman-assisted hopping, we comment on specific transitions in alkali atoms.
We analyze first the case of bosons, which we illustrate with the example of $^{87}$Rb (below we comment on other species).
We choose A$\equiv |F=1,m_F=-1\rangle$ and B$\equiv |F=2,m_F=-2\rangle$. Note that one could in principle swap the definition of A and B.
The opposite choice would be however problematic, since the scattering lengths fulfill
$a(|F=1,m_F=-1\rangle,|F=2,m_F=-2\rangle)=a(|F=2,m_F=-2\rangle,|F=2,m_F=-2\rangle)$~\cite{footnote-collision}. Hence swapping A and B would
lead to $a_{AA}=a_{AB}$, which would not allow to remove some spurious Raman transitions~(see Subsec.~\ref{subsec:Spurious}).
Note also that since we consider maximally stretched states, hyperfine-changing
collisions are absent.
We consider two lasers~(see Fig.~\ref{fig:S1}), one with linear polarization ($\pi$ laser), and other with $\sigma_-$ polarization~($\sigma$ laser).
We define as $\delta$ the detuning of the $\pi$ laser from the $|F=2,m_F=-2\rangle\to |F'=2,m'_F=-2\rangle$ transition, which is~(up to slight detunings of the order of kHz or smaller)
the same as that of the $\sigma$ laser from the $|F=1,m_F=-1\rangle\to |F'=2,m'_F=-2\rangle$ transition. Both lasers are red detuned.
Note that due to selection rules $|F=2,m_F=-2\rangle$ is dark with respect to the $\sigma$ laser.
In contrast, the $\pi$ laser links as well $|F=1,m_F=-1\rangle$ to $|F'=2,m'_F=-1\rangle$ and $|F'=1,m'_F=-1\rangle$, with detunings, respectively, $\delta+\omega_{HFS}$ and
$\delta+\omega_{HFS}-\omega'_{HFS}$, where $\omega_{HFS}$ and $\omega'_{HFS}$ are the hyperfine splittings of the $^2S_{1/2}$ and $^2P_{1/2}$ manifold,
respectively. The fact that the $\pi$ laser affects $|F=1,m_F=-1\rangle$ may introduce spurious effects~(see Subsec.~\ref{subsec:4laser}).
Fermions may be treated in a similar way. For the specific case of $^{40}$K, we choose A$\equiv |F=9/2,m_F=-9/2\rangle$ and B$\equiv |F=7/2,m_F=-7/2\rangle$.
We may then employ a similar $\pi$--$\sigma$ laser configuration as in Fig.~\ref{fig:S1}. Since $F=9/2$ lies below $F=7/2$ the lasers must be now blue detuned.
The above mentioned choice of A and B is better than the opposite, since it minimizes the spurious effects introduced the the coupling of $|F=7/2,m_F=-7/2\rangle$ by the $\pi$ laser~(see Subsec.~\ref{subsec:4laser}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.0\columnwidth]{FigS1.pdf}
\vspace{-0.5cm}
\end{center}
\caption{(Color online) Left: Transitions considered for the realization of the Raman-assisted hoppings, with $\pi$ polarized lasers~(blue) and
$\sigma_-$ polarized lasers~(red); right: if $\chi\ll 1$~(see text), one may assume in a very good approximation that the $\pi$ lasers affect
just the A state, whereas the $\sigma$ lasers affect just the B state by construction.}
\label{fig:S1}
\end{figure}
\subsection{Raman-assisted hopping}
\label{subsec:Raman}
Direct hopping is prevented for a large-enough tilting, $J\ll\Delta$.
Raman assisted hopping may be induced by the absorption of a
photon of laser $L_m$ and the emission of a photon of laser $L_n$.
The amplitude of the Raman-assisted hopping, $J_{nm}$, from site $j$ to site $j+1$, may be
evaluated following Ref.~\cite{MiyakeThesis}:
\begin{equation}
J_{nm}=\frac{V_{nm}}{4}
\int d^3{\mathbf r}\, \Phi_{j+1}({\mathbf r})^* e^{i\delta{\mathrm k}^{nm}\cdot {\mathbf r}} \Phi_{j}({\mathbf r})
\end{equation}
where
$\delta{\mathbf k}^{nm}={\mathbf k}_n-{\mathbf k}_m$ is the difference between the laser wave vectors of the lasers, and $V_{nm}$ is the amplitude of the Raman coupling~(see Subsec.~\ref{subsec:4laser}).
For $J\ll \Delta$:
\begin{eqnarray}
&&\!\!\!\!\!\!\!\!\!\!\!\int dx \psi_{j+1}(x)^* e^{i\delta k_x^{nm}x} \psi_{j}(x)\simeq \nonumber \\
&&e^{i\delta k_x^{nm}D(j+1/2)} \left [ \int dx w(x-D)^*w(x) e^{i\delta k_x^{nm}x} \right \delimiter 0 \nonumber \\
&& +\left \delimiter 0 2i\frac{J}{\Delta} \sin\left (\frac{\delta k_x^{nm}D}{2} \right ) \int dx |w(x)|^2 e^{i\delta k_x^{nm}x} \right ].
\end{eqnarray}
For $s\gg 1$~($l\ll D$):
\begin{equation}
J_{nm}\simeq i\left (\frac{V_{nm}}{2\Delta} \right ) J \sin \left ( \frac{\delta k_x^{nm} D}{2} \right ) e^{i \delta k_x^{nm} D(j+1/2)}.
\end{equation}
Note that $\delta k_x^{nm}\neq 0$ is necessary to establish a significant hopping~\cite{MiyakeThesis,MiyakePaper,BlochPaper}.
\subsection{Four-laser scheme}
\label{subsec:4laser}
We employ the notation of the main text, $(\eta_j,\eta_{j+1})$, to denote the states of neighboring sites, where $\eta_{j,j+1}$ may be $0$~(empty site),
$A$ or $B$ (singly occupied site with an A or B component), and $AA$, $AB$ and $BB$ for doubly-occupied sites.
\begingroup
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{| c | c | c | c | c |}
\hline
Species & $\omega_{HFS}$~(GHz) & $\omega'_{HFS}$~(GHz) & $\Gamma$~(MHz) & $\chi$ \\
\hline
Na & $1.722$ & $0.189$ & $9.8$ & $0.21$ \\
$^{87}$Rb & $6.834$ & $0.814$ & $5.75$ & $0.059$ \\
$^{85}$Rb & $3.036$ & $0.361$ & $5.75$ & $0.13$ \\
Cs & $9.193$ & $1.167$ & $4.57$ & $0.046$ \\
\hline
\end{tabular}
\end{ruledtabular}
\caption{
Values of $\omega_{HFS}$, $\omega'_{HFS}$, the linewidth $\Gamma$, and $\chi$~(evaluated for $\delta=0.5$ GHz) for different atomic species.
For Na and $^{87}$Rb, A$\equiv |F=2,m_F=-2\rangle$, and B$\equiv |F=1,m_F=-1\rangle$, and $\chi$ is defined as in Eq.~\eqref{eq:chi}.
For $^{85}$Rb, A$\equiv |F=3,m_F=-3\rangle$, and B$\equiv |F=2,m_F=-2\rangle$, and $\chi=6\left [\frac{5}{27} \frac{\delta}{\delta+\omega_{HFS}}+
\frac{4}{27} \frac{\delta}{\delta+\omega_{HFS}-\omega'_{HFS}} \right ]$.
For Cs, A$\equiv |F=4,m_F=-4\rangle$, and B$\equiv |F=3,m_F=-3\rangle$, and $\chi=6\left [ \frac{7}{48} \frac{\delta}{\delta+\omega_{HFS}}+
\frac{3}{16} \frac{\delta}{\delta+\omega_{HFS}-\omega'_{HFS}} \right ]$.
}
\label{table:T1}
\end{table}
\endgroup
\subsubsection{Bosonic case}
Let us first study the bosonic case. We consider the Raman scheme of Fig. 1 of the main text, where $L_{2,3}$ are $\sigma$ lasers, and $L_{1,4}$ are $\pi$ lasers.
The lasers are characterized by their Rabi frequency $\Omega_{j=1,\dots, 4}$ proportional to the square root of their intensity, their frequency $\omega_j$,
and their wave vector ${\bf k}_j$.
As mentioned above, we illustrate the case of bosonic species with the example of $^{87}$Rb,
with A$\equiv |F=2,m_F=-2\rangle$, and B$\equiv |F=1,m_F=-1\rangle$. Let us consider in detail the different Raman-assisted hops we would like to
induce~(the Clebsch-Gordan factors may variate for other alkali species):
\begin{itemize}
\item (i) $(A,0)\to (0,A)$. This transition leads to an energy shift $\Delta E=-\Delta$. For $\omega_2-\omega_3=\omega_1-\omega_4=-\Delta$, the transition is realized
by either $L_1$ and $L_4$, or by $L_2$ and $L_3$, with amplitude
$V_{23}=\frac{1}{2}\frac{\Omega_2\Omega_3^*}{\delta}+\frac{\Omega_1\Omega_4^*}{\delta}\frac{1}{6}\chi$, with
\begin{equation}
\chi=6\left [\frac{1}{4} \frac{\delta}{\delta+\omega_{HFS}}+
\frac{1}{12} \frac{\delta}{\delta+\omega_{HFS}-\omega'_{HFS}} \right ].
\label{eq:chi}
\end{equation}
\item (ii) $(A,A)\to (0,AB)$. This transition has $\Delta E=U_{AB}-\Delta$, and involves absorbing a $\sigma$ photon, and emitting a $\pi$ one.
We assume $\omega_2-\omega_4=U_{AB}-\Delta+U$, where $U$ is a small detuning~(within the Raman resonance width).
As discussed below this tunneling plays the role of effective on-site interactions in the final model.
The transition is quasi-resonantly provided by $L_2$ and $L_4$, with amplitude $V_{24}=\frac{1}{\sqrt{6}}\frac{\Omega_2\Omega_4^*}{\delta}$.
\item (iii) $(AB,0)\to (A,A)$. For this case, $\Delta E=U_{AB}-\Delta$, and is performed by absorbing a $\pi$ photon, and then emitting a $\sigma$ one.
For $\omega_1-\omega_3=-U_{AB}-\Delta-U$, the transition is quasi-resonantly provided by
$L_1$ and $L_3$, with Raman amplitude $V_{13}=\frac{1}{\sqrt{6}}\frac{\Omega_1\Omega_3^*}{\delta}$.
\item (iv) $(AB,A)\to (A,AB)$. For this transition $\Delta E=-\Delta$, and can be only realized
using $\pi$ lasers. For $\omega_1-\omega_4=-\Delta$, the transition is realized by $L_{2,3}$,
with amplitude $V_{14}=\frac{1}{3}\frac{\Omega_1\Omega_4^*}{\delta}$.
\end{itemize}
\noindent For $\frac{|\Omega_1||\Omega_4|}{4}=\frac{|\Omega_2||\Omega_3|}{3}=
\frac{|\Omega_1||\Omega_3|}{2\sqrt{3}}=\frac{|\Omega_2||\Omega_4|}{2\sqrt{3}}=\Omega^2$,
$\Omega_1=|\Omega_1|e^{-i\phi}$, and $\Omega_{j=2,3,4}=|\Omega_j|$, one obtains
$V_{23}=\frac{\Omega^2}{\delta}(1+\chi e^{-i\phi})$, $V_{24}=\sqrt{2}\frac{\Omega^2}{\delta}$,
$V_{13}=\sqrt{2}\frac{\Omega^2}{\delta}e^{-i\phi}$, and
$V_{14}=2\frac{\Omega^2}{\delta}e^{-i\phi}$. The term proportional to $\chi$ in (i) is not discussed in the main text, for simplicity of the exposition.
It results in an additional term in the final Hamiltonian, as shown in Subsec.~\ref{subsec:Heff}. However, $\chi$ can be kept small . Specific examples for different species for $\delta=0.2$ GHz are
shown in Table~\ref{table:T1}~\cite{SteckRb87,SteckRb85,SteckNa,SteckCs}. We show below that for small $\chi$ the spurious term does not modify the
conclusions of the
main text.
\subsubsection{Fermionic case}
For fermions it is better to choose A as the state with larger $F$, e.g. A$\equiv |F=9/2,-9/2\rangle$ and B$\equiv |F=7/2,-7/2\rangle$ for $^{40}$K.
In the scheme of Fig.~1 of the main text, we choose $L_{2,3}$ as $\pi$ lasers, and $L_{1,4}$ as $\sigma$ lasers. We may proceed as for bosons, but choosing $\Omega_1=-|\Omega_1|e^{-i\phi}$.
Due to
the inverted choice of A and B, we obtain: $V_{23}=\frac{\Omega^2}{\delta}$, $V_{24}=\sqrt{2}\frac{\Omega^2}{\delta}$, $V_{13}=\sqrt{2}\frac{\Omega^2}{\delta}e^{-i\phi}$, and
$V_{14}=2\frac{\Omega^2}{\delta}(e^{-i\phi}+\chi) $. For $^{40}K$,
$\chi=\left [ \frac{32}{243} \frac{\delta}{\delta+\omega_{HFS}}+
\frac{49}{243} \frac{\delta}{\delta+\omega_{HFS}-\omega'_{HFS}} \right ]$, with $\omega_{HFS}=1.286$ GHz, $\omega'_{HFS}=0.155$ GHz~(the linewidth is $\Gamma= 5.96$ MHz)~\cite{TieckeK}.
For $\delta=0.2$ GHz, $\chi=0.072$. Note that inverting the choice of A and B, would lead to $V_{nm}$ of the same form as in the bosonic case but with $\chi$ four times larger ($\chi=0.28$ for $\delta=0.2$GHz).
The latter justifies our choice of A and B.
\subsection{Effective Hamiltonian}
\label{subsec:Heff}
We consider that only processes (i) to (iv) are resonant, and postpone to the Subsec.~\ref{subsec:Spurious} the discussion about spurious processes.
We denote as $c_j$ the bosonic operator corresponding to the Fock-state manifold $\{ |0\rangle$, $|1\rangle\equiv |A\rangle$, $|2\rangle\equiv |AB\rangle\}$.
Let us consider ${\bf k}_{1,2}=k{\bf e}_y$, and ${\bf k}_{3,4}=k{\mathbf e}_x$, and $k=\pi/D$.
\subsubsection{Bosonic case}
For the bosonic case, the Raman-assisted hopping of the previous subsection result in the effective
Hamiltonian:
\begin{eqnarray}
H&=&-t \sum_j (-1)^j \left [ c_{j+1}^\dag e^{i\phi n_j} c_{j} +\mathrm{H.c.} \right ] \nonumber \\
&+& \chi \sum_j (-1)^j \left [ e^{i\phi} c_{j+1}^\dag (1-n_j )(1-n_{j+1}) c_{j} +\mathrm{H.c.} \right ] \nonumber \\
&+& \frac{U}{2}\sum_j n_j(n_j-1).
\label{eq:Heff-1}
\end{eqnarray}
with $n_j=c_j^\dag c_j$, and
$t=\left (\frac{\Omega^2/\delta}{2\Delta} \right ) J$.
Typical values of the Raman-assisted hopping rate, $t$, are of the order of few tens of Hz.
Note that the factor $(-1)^j$, which results from the $x$ projection of $\delta {\mathbf k}$, may be easily eliminated by redefining the bosonic operators in the form:
$b_{4l}=c_{4l}$, $b_{4l+1}=c_{4l+1}$, $b_{4l+2}=-c_{4l+2}$, $b_{4l+3}=-c_{4l+3}$, with $l$ an integer. In this way we obtain:
\begin{eqnarray}
H&=&-t \sum_j \left [ b_{j+1}^\dag e^{i\phi n_j} b_{j} +\mathrm{H.c.} \right ] \nonumber \\
&+& \chi \sum_j \left [ e^{i\phi} b_{j+1}^\dag (1-n_j )(1-n_{j+1}) b_{j} +\mathrm{H.c.} \right ] \nonumber \\
&+& \frac{U}{2}\sum_j n_j(n_j-1).
\label{eq:Heff-bosons}
\end{eqnarray}
\subsubsection{Fermionic case}
Proceeding in exactly the same way for the fermionic case, we obtain
\begin{eqnarray}
H&=&-t \sum_j \left [ b_{j+1}^\dag e^{i\phi n_j} b_{j} +\chi b_{j+1}^\dag n_j n_{j+1} b_{j} +\mathrm{H.c.} \right ] \nonumber \\
&+& \frac{U}{2}\sum_j n_j(n_j-1).
\label{eq:Heff-fermions}
\end{eqnarray}
\subsection{Effect of a finite $\chi>0$}
For $\chi = 0$, Hamiltonians~\eqref{eq:Heff-bosons} and~\eqref{eq:Heff-fermions} reduce
to Eq.~(1) in the main text, and hence to
the anyon-Hubbard model (AHM). $\chi > 0$ results in spurious
collisionally-assisted hopping terms. However, as commented
above, $\chi$ can be kept small.
Figure~\ref{fig:S2}~(a) shows the effect of the spurious correlated
tunneling in Eq.~\eqref{eq:Heff-bosons} which basically effects the low-density part
$\rho<1$ and results in a small shift of the SF-PP transition
(indicated by small arrows) to lower $\rho$. Interestingly a PP phase may be observed also for
$\rho>1$ for sufficiently large $\chi$.
Figure~\ref{fig:S2}~(b) shows the effect of the spurious correlated
tunneling in Eq.~\eqref{eq:Heff-fermions}. The effect is negligible for small lattice
fillings $\rho < 1$. Due to its $\rho^2$ -dependence even for much
larger values of $\chi$ as discussed in the previous subsection the
curves of Fig.~\ref{fig:S2}~(b) basically overlap with the pure
AHM-case. For densities $\rho \geq 1$ the phase boundaries are slightly
modified. In particular the size of the Mott-insulator region increases
and the PP phase may now extend to a small region
above unit filling (see arrows in Fig.~\ref{fig:S2}~(b)).
Note that, interestingly, the choice of a larger detuning $\delta$ and thus larger
values of $\chi$ actually favors the observation of the PP phase in
experiments.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{FigS2-2.pdf}
\caption{Equation of state $\rho=\rho(\mu)$ of the AHM ($U=0$,
$\phi=\pi$, $L=48$) for different values of $\chi$ for the spurious
correlated tunneling in (a) Eq. (6) and (b) Eq. (7). Small arrows depict
the kinks in the $\rho(\mu)$ curve for $\chi=0.2$ characterizing the
PP-SF transition. }
\label{fig:S2}
\end{figure}
\subsection{Comparison to the model of Keilmann et al.}
\label{subsec:Comparison}
Note that (i) and (iv) are energetically degenerated, but (apart from the small spurious effect discussed above) they may be addressed with different lasers due to selection rules. This point constitutes the major
drawback of the proposal of Ref.~\cite{Keilmann2011}. In that proposal, a single component (A) was considered, and process (iv) was of the form (AA,A)$\to$(A,AA), which cannot be resolved
from process (i). As a result, both $L_2$ and $L_3$, and $L_1$ and $L_4$ address both (i) and (iv), preventing the realization of the desired density-dependent Peierls phase.
The two processes may be just discerned by considering a very small detuning $\delta\ll U_{AA},\Delta$. Since $U_{AA}$ and $\Delta$ are of the order of kHz, $\delta\lesssim 100$ Hz is necessary.
However $\Gamma$ is typically of the order of few MHz (see e.g. Table~\ref{table:T1}), rendering unfeasible
the proposal of Ref.~\cite{Keilmann2011}.
\begingroup
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{| c | c | c | c |}
\hline
Species &$a_{AA}/a_B$& $a_{AB}/a_B$ & $|\delta U/h|$~(kHz) \\
\hline
Na & $54.53$ & $64.5$ & $0.72$ \\
$^{87}$Rb & $100.36$ & $99.08$ & $0.23$ \\
$^{85}$Rb & $-469$ & $-386$ & $1.40$ \\
Cs & $-1980$ & $2436$ & $8.89$ \\
\hline
\end{tabular}
\end{ruledtabular}
\caption{
Values of $a_{AA}$, $a_{AB}$, and $\delta U/h$ for different bosonic species, for $s=20$ and $D=532$nm~(except for $^{87}$Rb where we employ $s=25$ and $D=266$nm)~\cite{Tiemann-Private}.
For Na and $^{87}$Rb, A$\equiv |F=2,m_F=-2\rangle$ and B$\equiv |F=1,m_F=-1\rangle$.
For $^{85}$Rb, A$\equiv |F=3,m_F=-3\rangle$, and B$\equiv |F=2,m_F=-2\rangle$.
For Cs, A$\equiv |F=4,m_F=-4\rangle$, and B$\equiv |F=3,m_F=-3\rangle$. Note that negative scattering lengths are of no concern in our set-up.
}
\label{table:T2}
\end{table}
\endgroup
\subsection{Spurious processes}
\label{subsec:Spurious}
Spurious Raman processes must be kept out of resonance:
\begin{itemize}
\item (v) (A,0) $\to$ (0,B): $\Delta E=-\Delta$
\item (vi) (A,A) $\to$ (0,AA): $\Delta E=-\Delta+U_{AA}$
\item (vii) (AA,0) $\to$ (A,A): $\Delta E=-\Delta-U_{AA}$
\item (viii) (AB,A) $\to$ (B,AA): $\Delta E=-\Delta+\delta U$, with $\delta U=(U_{AA}-U_{AB})$
\item (ix) (AA,B) $\to$ (A,AB): $\Delta E=-\Delta-\delta U$
\end{itemize}
Note that (vi), (vii), (viii) and (ix) are not possible for fermionic atoms. Process (v) is just possible with
$J_{24}$ or $J_{13}$. But these laser combinations are (quasi-)resonant with $-\Delta\pm U_{AB}$.
For a sufficiently large $U_{AB}$ process (v) is hence far from resonance with either $J_{24}$ or $J_{13}$.
As an illustration we consider the case of $^{40}$K atoms, with $a_{AB}\simeq 170 a_B$, with $a_B$ the Bohr radius, and a lattice with $D=532$ nm, and $s=20$,
$U_{AB}/h\simeq 7$ kHz.
Note that typical values for the width of the
Raman resonance, $W\lesssim 50$ Hz~\cite{MiyakeThesis}. Hence $U_{AB}\gg W$, and hence process (v) may be safely
ignored.
For bosons, (vi), (vii), (viii) and (ix) must be also considered. These processes can be only realized using
$J_{23}$. To neglect the (vi) and (viii) processes one needs $U_{AA}\gg W$. This is the case in experiments, where typically $U_{AA}\sim 1$ to few $10$s kHz. In contrast, to avoid (viii) and (ix) one must
demand $\delta U \gg W$. The latter condition is more strict, but may be attained as well in experiments. Examples for different bosonic species are illustrated in Table~\ref{table:T2}~\cite{Kronjager-Thesis,Tiemann-Private}.
\subsection{Initial preparation}
\label{subsec:Initial}
If the spurious processes can be neglected, only on-site states 0, A, and AB may be reached by Raman-assisted hopping.
However, B, AA, BB, and even sites with occupation higher than two, may initially exist.
Sites with an occupation larger than two are eventually destroyed by three-body processes, and are hence of no major concern.
This is however not the case of B, AA and BB, which should be avoided in the initial preparation, since they would induce immobile spurious defects.
A straightforward way to avoid immobile defects is to consider a low-density gas of maximally singly occupied sites by A
particles~(doubly occupied AA sites may be eliminated using photo association pulses~\cite{Sherson2010}). One may then compress this dilute gas increasing
the density until reaching the desired density. Since A can just move into AB, doubly occupied sites will be guaranteed to have an AB occupation,
and hence the on-site states remain always within the $\{$0, A, AB$\}$ manifold. \\
\begin{figure*}[ht]
\centering
\includegraphics[width=1.0\linewidth]{figS3.pdf}
\caption{Quasi-momentum distribution $n(k)$ of the A component for
the case of fermionic atoms as function of $U/t$
($\phi=\pi$, $\rho=1/2$, $L=60$ sites). (c) and (d) show $n(k)$
including the oscillating factors $(-1)^j$ created by the Raman-assisted
hopping; (a) and (b) show the same data, but excluding the oscillating terms. (b)
and (d) show $n(k)$ for $U=2$ (solid line), $U=0$ (dashed line) and
$U=-2$ (dotted line). In (a) and (c) the dashed lines denote the phase
transitions between PSF, PP and SF phases. }
\vspace*{-0.3cm}
\label{fig:S3}
\end{figure*}
\section{Quasi-momentum distribution}
\label{sec:Quasimomentum}
As discussed above, the AHM may be realized experimentally
with both fermionic and bosonic species. While many
observables, most relevantly the density, do not depend on the particular
choice of constituents, others, in particular the momentum distribution, are not
gauge invariant.
As shown in Fig.~4(b) and (c) of the main text, bosonic
species closely mimic the quasi-momentum distribution of the effective
model. As mentioned above, the oscillating factors $(-1)^j$
resulting from the necessary $x$ projection of $\delta {\mathbf k}$, may be easily eliminated by a gauge transformation.
However, this transformation must be undone in order to recover the momentum distribution of components A and B. As a result,
A and B components exhibit a doubled unit-cell, and the quasi-momentum distribution $n(k)$ is consequently trivially shifted and
duplicated. For clarity we only depict half the Brillouin zone for A and B in Fig.~4 of the main text.
It is interesting to depict as well the momentum distribution for the fermionic case.
Figures~\ref{fig:S3} (a) and (b), which depict $n(k)$ neglecting the $(-1)^j$ factors, show a
single Fermi-surface in the SF, two for the PP, and a
blurred $n(k)$ for the PSF. The oscillating $(-1)^j$ factors obscure
this simple interpretation
~(as shown in Figs. (c) and (d)). However PSF, PP and SF phases
may be still clearly discriminated from each other by their characteristic momentum
distribution function.
\bibliographystyle{prsty}
|
\section*{Introduction}
The {\it Bernstein problem} is one of the traditional problems
of global differential geometry.
The original result due to Bernstein asserts that
every entire minimal graph in Euclidean three-space
$\mathbb{R}^3(x_1,x_2,x_3)$ is a plane.
In other words,
Bernstein's result shows that the only global solution
on the $(x_1, x_2)$-plane to the
so-called {\it minimal surface equation}
\begin{equation*}
\{1+(f_{x_1})^2\}f_{x_2x_2}-2f_{x_1}f_{x_2}f_{x_1x_2}
+\{1+(f_{x_2})^2\}f_{x_1x_1}=0
\end{equation*}
is a linear function of $x_1$ and $x_2$.
The Bernstein problem has been generalized
to a problem basically asking for a classification of all
entire minimal graphs.
On the other hand, when the ambient space is not the Euclidean
three-space, the Bernstein problem often needs to be amended.
For instance, in Minkowski three-space $\mathbb{L}_3$ equipped with the
natural Lorentz metric $dx_1^2+dx_2^2-dx_3^2$,
there are many entire (timelike) minimal graphs over the timelike plane
$\mathbb{L}_2=\mathbb{R}^2(x_2,x_3)$, see for example \cite{Mil1}.
Next, we focus on the three-dimensional \textit{Heisenberg group} ${\rm Nil}_3$
which is one of the model spaces of Thurston geometries \cite{Thurston}.
The space ${\rm Nil}_3$ is realized as Cartesian three-space
$\mathbb{R}^3(x_1,x_2,x_3)$ equipped with the Riemannian metric
\begin{equation*}
dx_1^2+dx_2^2+\left\{dx_3+\frac{1}{2}(x_2dx_1-x_1dx_2)\right\}^2
\end{equation*}
and a nilpotent Lie group structure, see for example \cite{DIK:mini}.
The Riemannian metric is invariant under
the nilpotent Lie group structure and has a 4-dimensional isometry group.
The identity component of the isometry group is a semi-direct
product $\mathrm{Nil}_3\ltimes \mathrm{SO}_2$.
It has been known for a long time that in
$\mathrm{Nil}_3$ nontrivial entire minimal graphs exist, see for example
\cite{IKOS}. Therefore, in ${\rm Nil}_3$ the
Bernstein problem has been phrased more specifically
as the problem to construct
entire minimal graphs over the natural $(x_1, x_2)$-plane with a
prescribed holomorphic quadratic differential.
Under this formulation, Fern{\'a}ndez and Mira studied the Bernstein problem in
${\rm Nil}_3$ \cite{Fer-Mira2}. They proved that for a prescribed
holomorphic quadratic differential $Q\,dz^2$ over the complex plane
$\mathbb C$ with $Q\neq 0$ or the unit disc $\mathbb{D}$,
there exists a two-parameter family of entire
minimal vertical graphs whose Abresch-Rosenberg differential is $Q\,dz^2$.
Their proof relies firstly on the Lawson-type correspondence
(often called \textit{sister correspondence}) between
minimal surfaces in $\mathrm{Nil}_3$ and
surfaces of constant mean curvature ({\sc CMC} in short)
with mean curvature $H =1/2$ in
the product space $\mathbb{H}^2\times\mathbb{R}$,
where $\mathbb H^2$ denotes the hyperbolic two-space.
Secondly, they use the correspondence between harmonic maps into
$\mathbb{H}^2$ and {\sc CMC} surfaces with mean curvature $H =1/2$
in the product space $\mathbb{H}^2\times\mathbb R$.
Finally, they use a result of Wan and Au \cite{Wan, WanAu}
solving the Bernstein problem for spacelike {\sc CMC} surfaces in
$\mathbb{L}_3$ and use that the Gauss map of
those surfaces is also harmonic into $\mathbb{H}^2$.
Our proof is much more direct.
In our previous work \cite{DIK:mini},
we have established a generalized Weierstrass
type representation for
minimal surfaces in $\mathrm{Nil}_3$.
Every simply connected (nowhere vertical) minimal surface
is obtained from an extended frame for a harmonic map
into the hyperbolic two-space $\mathbb{H}^2$.
In this paper we give a short proof of
the solution to the Bernstein problem in $\mathrm{Nil}_3$
by virtue of the generalized Weierstrass
type representation established in \cite{DIK:mini}.
The advantage of our approach is that we can give a direct
relation between minimal graphs in $\mathrm{Nil}_3$ and
spacelike {\sc CMC} surface with mean curvature $H=1/2$
graphs in $\mathbb{L}_3$, Theorem \ref{thm:symcorrespond}.
This relation enables us to give a simple alternative proof
of Fern{\'a}ndez-Mira's theorem, Theorem \ref{thm:Bern}.
Our new proof actually also provides new insights.
In fact our proof clarifies the geometric meaning of
the two-parameter ambiguity of entire minimal graphs with
prescribed Abresch-Rosenberg differential. While it is quite
clear that the two-parameter family is related to the boosts
in ${\rm SU}_{1, 1}$, our argument also shows how the corresponding family of surfaces varies in ${\rm Nil}_3$.
\section{Bernstein problem}
We discuss the Bernstein problem in ${\rm Nil}_3$, that is,
the classification of entire minimal vertical graphs in
${\rm Nil}_3$. We only consider vertical graphs. Therefore we will sometimes omit the word
``vertical''. In Appendix \ref{sc:Preliminaries}, we give
a short review of facts and results of \cite{DIK:mini} which are used
for the solution to the Bernstein problem via loop groups.
From now on, we denote the coordinates of ${\rm Nil}_3$ or $\mathbb L_3$ by $(x_1, x_2, x_3)$.
\subsection{Completeness}\label{sc:Com}
The basic result used in this paper is Theorem \ref{thm:Sym}. It explains the direct relation between
minimal surfaces in ${\rm Nil}_3$ and spacelike {\sc CMC} surfaces in $\mathbb L_3$.
This close relationship is also underlined by a simple relation between the corresponding metrics.
\begin{Lemma}\label{lm:metricrelation}
Let $f^{\l}$ and $f_{\mathbb L_3}^{\l}$ be an associated family of minimal surfaces in ${\rm Nil}_3$
and an associated family of spacelike {\sc CMC} surfaces with mean curvature $H =1/2$
in $\mathbb L_3$ correlated and defined as in Theorem
\ref{thm:Sym}, respectively. Denote the metric of $f^{\l}$ by $e^u dz d\bar z$
and the metric of $f_{\mathbb L_3}^{\l}$ by $e^{u_{\mathbb L_3}}dz d \bar z$, respectively.
Moreover, let $\phi_3^{\l}dz$ be the coefficient of $e_3$ in
$(f^{\l})^{-1} f_{z}^{\l} dz
= \sum_{i =1 }^3 (\phi_{i}^{\l} dz) e_{i}$. Then the following relation holds$:$
\begin{equation*}
e^{u_{\mathbb L_3}} + 4 |\phi_3^{\l}|^2 = e^{u}.
\end{equation*}
\end{Lemma}
\begin{proof}
It is known that the conformal factors $e^{u}$ and $e^{u_{\mathbb L_3}}$
can be computed explicitly
in terms of spinors, see \cite[Section 3.1]{DIK:mini},
Remark \ref{rm:AppRem} and \cite{BRS:Min}:
\begin{equation*}
e^u = 4 (|\psi_1^{\l}|^2 + |\psi_2^{\l}|^2)^2, \;\;\; e^{u_{\mathbb L_3}}= 4 (|\psi_1^{\l}|^2 - |\psi_2^{\l}|^2)^2,
\end{equation*}
where $\psi_j^{\l} \;\;(j =1, 2)$ is a family of spinors for
the associated family $f^{\l}$.
Since $\phi_3^{\l} = 2 \psi_1^{\l} \overline{\psi_2^{\l}}$, the claim follows.
\end{proof}
\begin{Remark}\label{rm:metrics}
It is known that the metrics $e^{u_{\mathbb L_3}} dz d \bar z$ of an associated family
of spacelike {\sc CMC} surfaces $f_{\mathbb L_3}^{\l}$
are independent of $\l$, that is, on a simply connected domain, any
two members of the associated family $\{f_{\mathbb L_3}^{\l}\}_{\l \in \mathbb S^1}$ are isometric. In fact the metric can be computed by the support $h (dz)^{1/2} (d \bar z)^{1/2}$,
see Appendix \ref{sc:Preliminaries} for definition, as
\begin{equation}\label{eq:supportmetric}
h^2 dz d \bar z= e^{u_{\mathbb L_3}}dz d \bar z.
\end{equation}
However, the metrics $e^{u} dz d \bar z$ of an associated family of
minimal surface $f^{\l}$ depend on $\l$, that is,
any two members of the associated family $\{f^{\l}\}_{\l \in \mathbb S^1}$
are, in general, non-isometric.
\end{Remark}
Using the relation above, we have the following theorem.
\begin{Theorem}\label{thm:compl}
Let $f^{\l}$ and $f_{\mathbb L_3}^{\l}$ be an associated family of minimal surfaces in ${\rm Nil}_3$
and an associated family of spacelike {\sc CMC} surfaces with mean curvature $H =1/2$
in $\mathbb L_3$ correlated and defined as in Theorem
\ref{thm:Sym}, respectively. Assume that one member of the associated family
$\{f_{\mathbb L_3}^{\l}\}_{\l \in \mathbb S^1}$ is closed with respect to
the Euclidean topology. Then each member of the associated family $\{f^{\l}\}_{\l \in \mathbb S^1}$ is a complete, entire graph.
\end{Theorem}
\begin{proof}
We denote the spacelike {\sc CMC} surface in $\mathbb L_3$ which is closed with respect to the
Euclidean topology by $f_{\mathbb L_3}^* = f_{\mathbb L_3}^{\l}|_{\l^* \in \mathbb S^1}$.
From the assumption and by \cite[p.~415]{CY}, we conclude
that $f^{*}_{\mathbb L_3}$ is complete. Moreover from \cite[Proposition 2]{Wan},
$f^{*}_{\mathbb L_3}$ is also an entire graph.
Since within the associated family $\{f_{\mathbb L_3}^{\l}\}_{\l \in \mathbb S^1}$
the metric is invariant, see Remark \ref{rm:metrics},
each member of the associated family of spacelike {\sc CMC} surfaces
$\{f_{\mathbb L_3}^{\l}\}_{\l \in \mathbb S^1}$ is also complete.
Then from Lemma \ref{lm:metricrelation}, we have that
each member of the associated family of minimal surfaces
$\{f^{\l}\}_{\l \in \mathbb S^1}$ is complete.
Let us look more closely at the correspondence between
$f_{\mathbb L_3}^{\l}$ and $f^{\l}$.
From formulas \eqref{eq:SymMin} and \eqref{eq:symNil} we infer by inspection
that $f^{\l}$ and $f^{\l}_{\mathbb L_3}$ share the same $x_1$-, $x_2$-components.
Here, as pointed out before, we denote the coordinates
of ${\rm Nil}_3$ or $\mathbb L_3$ by $(x_1, x_2, x_3)$.
Therefore, since $f_{\mathbb L_3}^{\l}$ is an entire graph, thus $f^{\l}$
also is an entire graph. This completes the proof.
\end{proof}
\subsection{Rigid motions}\label{sc:Rigid}
It is known that the isometry group of $\mathbb L_3$ is the six-dimensional Lie group
which is generated by
a one-parameter family of rotations around $x_3$-axis (the timelike axis),
a two-parameter family of boosts and three families of translations.
In contrast, the isometry group of ${\rm Nil}_3$ is only four-dimensional and
is generated by a one-parameter family of rotations around $x_3$-axis
and three families of translations in ${\rm Nil}_3$.
A comparison of the two Sym formulas in Theorem \ref{thm:Sym}
indicates that isometries of Minkowski space will not necessarily
become isometries of ${\rm Nil}_3$. The precise relation will be made
clear in the Lemma below.
\begin{Lemma}\label{lem:isometry}
Let $f_{\mathbb L_3}^{\l}$ and $\tilde{f}_{\mathbb L_3}^{\l}$ be two associated family of
spacelike {\sc CMC} surfaces
with mean curvature $H=1/2$ in $\mathbb L_3$ defined by the Sym-formula
in Theorem \ref{thm:Sym} for
some extended frames $F^{\l}$ and $\tilde F^{\l}$, respectively and set
$f_{\mathbb L_3}= f_{\mathbb L_3}^{\l}|_{\l=1}$ and $\tilde f_{\mathbb L_3}=\tilde f_{\mathbb L_3}^{\l}|_{\l=1}$.
Moreover, let $f^{\l}$ and $\tilde f^{\l}$ denote the two associated families of
minimal surfaces in
${\rm Nil}_3$ \ defined from the same extended frames $F^{\l}$ and $\tilde F^{\l}$,
respectively and set $f= f^{\l}|_{\l=1}$ and $\tilde f=\tilde f^{\l}|_{\l=1}$.
Assume that $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$
are isometric by some rigid motion in $\mathbb L_3$.
Then the following statements hold$:$
\begin{enumerate}
\item If $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$
are isometric by a rotation around $x_3$-axis $($the timelike axis$)$, then
$f$ and $\tilde f$ are isometric by the rotation around
$x_3$-axis $($the same angle$)$ and some translation.
\item If $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$
are isometric by a translation, then
$f$ and $\tilde f$ are isometric by
some translation $($not necessarily the same translation$)$.
\item If $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$
are isometric by a boost, then
$f$ and $\tilde f$ are, in general, not isometric.
\end{enumerate}
\end{Lemma}
\begin{proof}
Since $f_{\mathbb L_3}(= f_{\mathbb L_3}^{\l}|_{\l=1})$ and
$\tilde f_{\mathbb L_3}(= \tilde f_{\mathbb L_3}^{\l}|_{\l=1})$
are isometric by a rigid motion in $\mathbb L_3$, the isometry between
these two surfaces lifts to the level of frames
$F= F^{\l}|_{\l=1}$
and $\tilde F = \tilde F^{\l} |_{\l=1}$
as $\tilde F = MFk$, where $M$ is a $z$-independent ${\rm SU}_{1, 1}$-valued matrix and
$k$ is a ${\rm U}_1$-valued matrix.
After introducing the loop parameter we obtain the relation
\begin{equation}\label{eq:isoF}
\tilde F^\lambda = M^\lambda F^\lambda k.
\end{equation}
Note that $M^{\l}$ is a $(\Lambda {\rm SU}_{1, 1})_\sigma$-valued matrix and satisfies $M^{\l}|_{\l=1} = M$.
Then it is easy to see that $\tilde f_{\mathbb L_3}^{\l}$ and
$f_{\mathbb L_3}^{\l}$ satisfy the relation:
\begin{equation}\label{eq:rigidmotion}
\tilde f_{\mathbb L_3}^{\l} = M^\lambda f_{\mathbb L_3}^{\l} (M^\lambda)^{-1} -i \l (\partial_{\l} M^\lambda) (M^\lambda)^{-1}.
\end{equation}
Now a straightforward computation shows that
the corresponding two minimal surfaces $f^{\l}$ and $\tilde f^{\l}$
have the following relation:
\begin{equation}\label{eq:minirelation}
\tilde f^{\l} = \left(\operatorname{Ad} (M^\lambda) (f^{\l}_{\mathbb L_3}) - X^{\l}\right)^{o}
- \frac{1}{2} \lbrace \operatorname{Ad}(M^\lambda) (i \l \partial_{\l} f^{\l}_{\mathbb L_3})
+ [X^{\l}, \operatorname{Ad} (M^\lambda)(f^{\l}_{\mathbb L_3})] - Y^{\l}\rbrace^{d},
\end{equation}
where we set
\begin{equation}\label{eq:XY}
X^{\l} = i \l (\partial_{\l} M^\lambda)(M^\lambda)^{-1}, \;\;
Y^{\l} = i \l (\partial_{\l} X^{\l}),
\end{equation}
and the superscripts ``$o$'' and ``$d$''
denote the off-diagonal and diagonal part,
respectively.
For simplicity of notation we do not distinguish here
$f^{\l}$ in ${\rm Nil}_3$ and $\hat f^{\l}$ in $\mathfrak{su}_{1, 1}$.
We note that for each fixed $\l \in \mathbb S^1$,
the first part of the right-hand side in \eqref{eq:rigidmotion}
describes a Lorentz transformation and
the second part of the right-hand side in \eqref{eq:rigidmotion}
describes a translation, respectively.
We now consider each of the three types of generators separately:
$(1)$ First suppose that $\tilde f_{\mathbb L_3}$ and $f_{\mathbb L_3}$
are isometric by a rotation around $x_3$-axis (the timelike axis).
Since the original transformation $M$ was a rotation, it follows
\begin{equation*}
M (= M^\lambda|_{\l =1}) = {\rm diag }(e^{i \theta}, e^{-i \theta}),
\;\;\;\;
\partial_{\l} M^\lambda|_{\l =1} = \mathbf 0.
\end{equation*}
Here $2 \theta$ is the angle of rotation.
A straightforward computation shows that
$f$ and $\tilde f$ satisfy the equation
$\tilde{f} = \operatorname{Ad}(M) (f) + \frac{1}{2}Y^d$,
where the translation term $Y$ can be computed as
\begin{equation*}
Y = Y^{\l}|_{\l =1} = - \l^2 (\partial_{\l}^2 M^\lambda) (M^\lambda)^{-1}|_{\l =1}.
\end{equation*}
Therefore, this one-parameter family consists of isometric minimal surfaces
in ${\rm Nil}_3$.
$(2)$ Next suppose that $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$
are isometric by some translation.
Since the original transformation $M$ was a translation, it follows
\begin{equation*}
M(= M^\lambda|_{\l =1} )= {\rm id}, \;\;(\partial_{\l} M^\lambda) |_{\l =1}\neq \mathbf 0.
\end{equation*}
Substituting $\lambda = 1$ into \eqref{eq:minirelation} we see immediately
that $f$ and $\tilde f$
satisfy the relation $\tilde{f} = f + A$,
where $A = A(x_1,x_2)$ is given by
\begin{equation}
A = - X^o - \frac{1}{2} \left([X, f_{\mathbb L_3}] -Y \right)^d,
\end{equation}
where $X= X^{\l}|_{\l=1}$ and $Y= Y^{\l}|_{\l=1}$ for $X^{\l}$ and
$Y^{\l}$ in \eqref{eq:XY}.
It is clear that $Y$ is independent of $x_1$ and $x_2$, the coordinates for
$f$, but $x_1$ and $x_2$ enter the commutator.
An explicit computation, using the bases stated in Appendix \ref{sc:Sym}
and the transformation formula stated in Appendix \ref{sc:isoNil},
now shows that $\tilde{f}$ can be obtained from $f$ by a translation in ${\rm Nil}_3$
(with a constant vector, whose coefficients basically are the components
of $X^{o}$ and of $Y^d$).
$(3)$ Let us finally consider the transformations $M$ given by boosts in $\mathbb L_3$.
These
transformations form a two-parameter family.
Since the original transformation was a boost,
it follows
\begin{equation}\label{eq:blement}
M = M^\lambda|_{\l =1} = \begin{pmatrix} \alpha & \beta \\
\bar{\beta} & \alpha \end{pmatrix},
\;\;\;\;
\partial_{\l} M^{\l}|_{\l =1} = \mathbf 0.
\end{equation}
Here $\alpha \in \mathbb R$, $\beta \in \mathbb C$ and $\alpha^2 - |\beta|^2=1$
and we obtain
\begin{equation*}
\tilde f = \left(\operatorname{Ad} (M) (f_{\mathbb L_3}) \right)^{o}
- \frac{1}{2} \left( \operatorname{Ad} (M) (i \l \partial_{\l} f^{\l}_{\mathbb L_3} |_{\lambda = 1})
- Y \right)^{d},
\end{equation*}
where $Y= Y^{\l}|_{\l=1}$ for $Y^{\l}$ in \eqref{eq:XY}.
Now it follows by a straightforward computation that
\begin{equation}\label{eq:boost}
\tilde f
=\begin{pmatrix}
(\alpha^2 + |\beta|^2) r + \alpha \beta \bar{s} - \alpha \bar{\beta} s & -2 \alpha \beta p
+ \alpha^2 q - \beta^2 \bar q \\
2 \alpha \bar{\beta} p + \alpha^2 \bar{q} - \bar{\beta}^2 q & -(\alpha^2 + |\beta|^2) r
-\alpha \beta \bar {s} + \alpha \bar{\beta} s
\end{pmatrix}
+ \frac{1}{2}Y^{d},
\end{equation}
where $p,r\in i \mathbb R$ and $q, s \in \mathbb C$ are functions defined by
\begin{equation}\label{eq:pqrs}
f_{\mathbb L_3}^{\l}|_{\l=1} =\begin{pmatrix}p & q \\\bar q & -p\end{pmatrix}\;\mbox{and}\;
- \frac{1}{2}i \l (\partial_{\l}f_{\mathbb L_3}^{\l})|_{\l=1}
=\begin{pmatrix}r & s \\\bar s & -r\end{pmatrix}.
\end{equation}
Note that the components of the minimal surface $f$
in the basis of Appendix \ref{sc:Sym} are
given by
\begin{equation*}
(x_1, x_2, x_3) =(2 \Im q, -2 \Re q, -2 \Im r).
\end{equation*}
Thus from \eqref{eq:boost} and the action of the isometry group of ${\rm Nil}_3$
as described in \eqref{eq:action}, it is easy to see that
$f$ and $\tilde f$ are in general not isometric, see Remark \ref{rm:conditions}
in detail.
\end{proof}
\begin{Remark}\label{rm:conditions}
In case $(3)$ of Lemma \ref{lem:isometry},
from \eqref{eq:boost} and the action of the isometry group of ${\rm Nil}_3$ in \eqref{eq:action},
we see that the $f$ and $\tilde f$ are isometric in ${\rm Nil}_3$ if and only if
there exist some $\theta\in \mathbb R, (a_1, a_2, a_3) \in \mathbb R^3$
such that the following two equations hold:
\begin{eqnarray*}
(\alpha^2 + |\beta|^2)x_3 -4\alpha \Im (\beta\bar s) - Y^{11}=
a x_1 + b x_2 + x_3+ a_3,&\\
-4 \alpha \beta p
+ i (\beta^2 + \alpha^2)x_1
+ (\beta^2 - \alpha^2)x_2
=e^{i\theta}(i x_1 -x_2) + i a_1 - a_2,
&
\end{eqnarray*}
where $Y^{d} = \operatorname{diag} (i Y^{11}, -i Y^{11})$,
$a = \frac{1}{2} (a_1 \sin \theta - a_2 \cos \theta),
b =\frac{1}{2} (a_1 \sin \theta + a_2 \cos \theta)$, and
$p, s$ are purely imaginary, complex valued functions, respectively,
defined in \eqref{eq:pqrs}.
From these two equations,
it is easy to see that they are satisfied for very special
minimal surfaces $f$ only.
\end{Remark}
\begin{Remark}
After fixing base points, the Sym-formula establishes a
$1$-$1$-relation between spacelike {\sc CMC} with mean curvature $H=1/2$
surfaces in $\mathbb L_3$ and minimal surfaces in ${\rm Nil}_3$.
Clearly, the Poincar\'e group ${\rm SU}_{1, 1} \ltimes \mathbb L_3$ acts on the family
of spacelike {\sc CMC}
surfaces in $\mathbb L_3$.
If we fix base points, we eliminate the action of the translation part of
the Poincar\'e group, reducing the action to the Lorentz group ${\rm SU}_{1, 1}$.
Via the Sym-formula, the Poincar\'e group also acts on the family of minimal
surfaces in ${\rm Nil}_3$. Since we fix base points, we can also eliminate the translation
part of the isometry group of ${\rm Nil}_3$. So generically, the dimension of the
family of minimal surfaces should be three.
But from Lemma \ref{lem:isometry}, identifying
minimal surfaces which are isometric
by rotations, we see that two is the highest dimension of any orbit. These orbits are realized by the action of boosts.
From \eqref{eq:blement}, the set of boosts $\mathcal B$ can be computed as
\begin{equation*}
\mathcal B = \left\{ \left.X \, {}^t\!\bar X \;\right|\; X \in {\rm SU}_{1, 1} \right\}.
\end{equation*}
From this it is clear that $\mathcal B$ is the symmetric space
${\rm SU}_{1, 1}/{\rm U}_1$.
\end{Remark}
\subsection{Bernstein problem}\label{sc:Bernstein}
We will finally present a short alternative proof of
the Bernstein problem in ${\rm Nil}_3$ using the loop group method.
The heart of the proof is the following simple relation between
spacelike {\sc CMC} graphs in $\mathbb L_3$ and minimal graphs in ${\rm Nil}_3$.
\begin{Theorem}\label{thm:symcorrespond}
Every entire, complete, spacelike {\sc CMC} graph in $\mathbb L_3$
with mean curvature $H =1/2$ and the Hopf differential $Q_{\mathbb L_3} dz^2$ induces,
via the Sym-formula
$($applied to its associated family$)$, an entire,
complete, minimal graph in ${\rm Nil}_3$ with Abresch-Rosenberg differential
$-Q_{\mathbb L_3} dz^2$.
Conversely every entire, complete, minimal graph in ${\rm Nil}_3$ is obtained in this way.
\end{Theorem}
\begin{proof}
Let $g_{\mathbb L_3}$ be an entire complete
spacelike {\sc CMC} graph with mean curvature $H=1/2$
over the $(x_1,x_2)$-plane
in $\mathbb L_3$ whose Hopf differential
is $Q_{\mathbb L_3}dz^2$.
Let $F^{\l}$ be the extended frame of $g_{\mathbb L_3}$,
see Remark \ref{rm:AppRem} for
the definition, and
apply the Sym-formulas of Theorem \ref{thm:Sym} to obtain $f^{\l}_{\mathbb L_3}$
and $f^{\l}$ from the same extended frame $F^{\l}$.
Note that $f^{\l}_{\mathbb L_3}$ and $f^{\l}$ define an associated family of
spacelike {\sc CMC} surfaces in $\mathbb L_3$ and minimal surfaces in ${\rm Nil}_3$, respectively.
Moreover, $f_{\mathbb L_3}= f_{\mathbb L_3}^{\l}|_{\l =1}$ and $g_{\mathbb L_3}$ are isometric by
some rigid motion in $\mathbb L_3$, by the fundamental theorem of surface theory
(the mean curvature $H$, the Hopf differential $Q_{\mathbb L_3} dz^2$ and
the metric $e^{u_{\mathbb L_3}} dz d \bar z$ are the same),
thus $f_{\mathbb L_3}$ is also a entire, complete,
spacelike {\sc CMC} graph, \cite[Proposition 1]{Wan}.
From formulas \eqref{eq:SymMin} and \eqref{eq:symNil}
we infer by inspection that $f^{\l}$ and $f^{\l}_{\mathbb L_3}$ share the
same $x_1$-, $x_2$-components. Thus $f= f^{\l}|_{\l =1}$
is an entire minimal graph as well.
Moreover, from Theorem \ref{thm:compl} we obtain that $f(= f^{\l}|_{\l =1})$
is complete and by Remark \ref{rm:AppRem}
we know that the Abresch-Rosenberg differential is $-Q_{\mathbb L_3} dz^2$.
To verify the second statement,
let $f$ be an entire, complete, minimal graph in ${\rm Nil}_3$ whose
Abresch-Rosenberg differential is $Q \,dz^2$ and let
$F^{\l}$ be the extended frame of $f$ and $f^\lambda$ its associated family
from the extended frame $F^{\l}$.
Then we have $f=f^{\l}|_{\l =1}$ up to translation in ${\rm Nil}_3$.
Moreover, let $f_{\mathbb L_3}^{\l}$
be the spacelike {\sc CMC} surface in $\mathbb L_3$ defined by the same extended
frame $F^{\l}$
in \eqref{eq:SymMin}.
Note that the Hopf differential of $f_{\mathbb L_3}^{\l}$ is $
Q_{\mathbb L_3}^{\l} dz^2 = -\l^{-2} Q \,dz^2$.
Since $f$ and $f_{\mathbb L_3}^{\l}|_{\l =1}$ have the same $x_1$-,$x_2$-components,
the latter surface is an entire {\sc CMC}
graph in $\mathbb L_3$, and thus by \cite[p.~415]{CY} it is complete.
\end{proof}
Using Theorem \ref{thm:symcorrespond},
it is easy to give the proof of the solution to the Bernstein problem.
\begin{Theorem}\label{thm:Bern}
Let $Q \,dz^2$ be a holomorphic quadratic differential
on $\mathbb D$ or $M = \mathbb C$ with $Q \not\equiv 0$. Then the following statements hold$:$
\begin{enumerate}
\item There exists a two-parameter family of entire,
complete, minimal graphs in ${\rm Nil}_3$, whose Abresch-Rosenberg
differential is $Q\, dz^2$.
\item Any two members of this two-parameter family are generically
non-congruent.
\item Each member of this two-parameter family is induced via the Sym-formula
by $($the associated family of$)$ an entire, complete, spacelike {\sc CMC} graph
in $\mathbb L_3$ with the Hopf differential $-Q \,dz^2$.
\end{enumerate}
\end{Theorem}
\begin{proof}
First we note that it is known that for
a given holomorphic quadratic differential
$Q_{\mathbb L_3}dz^2$ on $\mathbb D$ or $\mathbb C$,
there exists a unique entire complete
spacelike {\sc CMC} graph $g_{\mathbb L_3}$ over the $(x_1,x_2)$-plane
in $\mathbb L_3$ whose Hopf differential
is $Q_{\mathbb L_3}dz^2$, \cite{Wan, WanAu}.
Here ``unique'' means that any other such spacelike {\sc CMC}
graph whose Hopf differential
is $Q_{\mathbb L_3} dz^2$ is isometric to $g_{\mathbb L_3}$
by an isometry of $\mathbb L_3$. We normalize the mean curvature of $g_{\mathbb L_3}$
as $H =1/2$ and set $Q_{\mathbb L_3} dz^2= -Q \,dz^2$, where $Q \,dz^2$ is the quadratic
differential satisfying the condition in the Theorem.
Let $F^{\l}$ be the extended frame of $g_{\mathbb L_3}$, see Remark \ref{rm:AppRem} for
the definition,
and apply the Sym formulas of Theorem \ref{thm:Sym} to obtain $f^{\l}_{\mathbb L_3}$
and $f^{\l}$ from the same extended frame $F^{\l}$.
From Theorem \ref{thm:symcorrespond} we know that $f_{\mathbb L_3}=f^{\l}_{\mathbb L_3}|_{\l=1}$
and $f= f^{\l}|_{\l=1}$ are complete entire graphs.
From the construction it is clear that $f$ is a minimal surface.
Moreover the Abresch-Rosenberg differential of $f$ is $Q \, dz^2$.
We now consider a spacelike {\sc CMC} surface $\tilde g_{\mathbb L_3}$ isometric
to $g_{\mathbb L_3}$ in $\mathbb L_3$. Then as explained in the proof of Lemma \ref{lem:isometry},
the extended frame $\tilde F^{\l}$ of $\tilde g_{\mathbb L_3}$
satisfies $\tilde F^{\l} = M^{\l} F^{\l} k$ for
some $z$-independent $(\Lambda {\rm SU}_{1, 1})_\sigma$-valued matrix $M^{\l}$ and a ${\rm U}_1$-valued matrix $k$,
in particular independent of $\lambda$. For the associated family
$\tilde f_{\mathbb L_3}^\lambda$ of $\tilde g_{\mathbb L_3}$ which is defined
by the Sym formula \eqref{eq:SymMin} from the extended frame $\tilde F^{\l}$,
we see that $\tilde f_{\mathbb L_3}= \tilde f^{\l}_{\mathbb L_3}|_{\l =1}$ and $\tilde g_{\mathbb L_3}$
are isometric.
Thus $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$ are isometric,
and again from \cite[Proposition 1]{Wan} we obtain that $\tilde f_{\mathbb L_3}$
is an entire, complete, spacelike {\sc CMC} graph.
Let $\tilde f^{\l}$
be the corresponding associated family of minimal surfaces in ${\rm Nil}_3$ which is defined
by the Sym formula \eqref{eq:symNil} from
the extended frame $\tilde F^{\l}$. Then using the argument in
Theorem \ref{thm:symcorrespond}, we see that
$\tilde f = \tilde f^{\l}|_{\l =1}$ is an entire, complete minimal graph in ${\rm Nil}_3$.
Note that the Abresch-Rosenberg differential of $\tilde f$ is also $Q \, dz^2$.
We now apply Lemma \ref{lem:isometry}.
If the isometry $f_{\mathbb L_3}$ and $\tilde f_{\mathbb L_3}$ is of case $(1)$ or $(2)$,
then $\tilde f (=\tilde f^{\l}|_{\l =1})$ is congruent to $f(=f^{\l}|_{\l=1})$.
However, if the isometry of $f_{\mathbb L_3}$ and
$\tilde f_{\mathbb L_3}$ is of case $(3)$, then $\tilde f$ is
in general non-congruent to $f$. In particular, the case $(3)$ in
Lemma \ref{lem:isometry} corresponds to a two-parameter family of boosts in $\mathbb L_3$.
Therefore, for an entire, complete,
spacelike CMC surface, there exists a two-parameter family of
non-congruent complete minimal graphs in ${\rm Nil}_3$ which have the same
Abresch-Rosenberg differential $Q dz^2$.
\end{proof}
\begin{Remark}
In \cite{Fer-Mira2}, the two-parameter family of
an entire, complete, minimal graph was obtained by the
choice of the initial condition for a nonlinear partial differential equation.
The solution corresponds to $\phi_3$, that is, the $e_3$-component of
$f^{-1} f_z$ and the initial condition is the initial value
$\phi_3(z_*)$ for some base point $z_*$ in $\mathbb C$ or $\mathbb D$.
In our setting, this freedom naturally appears as the
two-parameter family of boosts in $\mathbb L_3$:
As we see from the proof of Theorem \ref{thm:Bern}, two minimal surfaces
$f^{\l}$ and $\tilde f^{\l}$ satisfy the relation \eqref{eq:minirelation}.
Set
\begin{equation*}
M^{\l}|_{\l =1} =
\begin{pmatrix}
\alpha & \beta \\
\bar{\beta} & \alpha
\end{pmatrix}
\in {\rm SU}_{1, 1},\;\;
(f^{\l})^{-1}f^{\l}_{z}|_{\l=1} = \sum_{k= 1}^3 \phi_k e_k, \;\mbox{and}\;
(\tilde f^{\l})^{-1}\tilde f^{\l}_{z}|_{\l=1} = \sum_{k= 1}^3
\tilde \phi_k e_k,
\end{equation*}
where $\alpha$ is real.
Then a straightforward computation (using the proof of \cite[Theorem 6.1]{DIK:mini})
shows
\begin{equation}
\tilde \phi_3 = (\alpha^2 + |\beta|^2) \phi_3 + 2 i\Re (\alpha \beta)\phi_1
+ 2 i \Im (\alpha \beta) \phi_2.
\end{equation}
From this expression, it is clear that our two-parameter family of boosts
induces a freedom of the initial condition of $\phi_3$ which is naturally
parametrized by $\mathbb C$.
\end{Remark}
\begin{Corollary}
The associated family of every entire, complete, minimal graph in ${\rm Nil}_3$
with a given Abresch-Rosenberg differential $Q \,dz^2$
is a family of entire, complete, minimal graphs in ${\rm Nil}_3$
with the Abresch-Rosenberg differential $\l^{-2} Qdz^2\;(\l \in \mathbb S^1)$.
Moreover, within a given associated family, complete minimal graphs
have the same support $h \,(dz)^{1/2} (d \bar z)^{1/2}$.
\end{Corollary}
\begin{proof}
From the proof of Theorem \ref{thm:Bern}, it is clear that for an entire minimal graph
$f^{\l}|_{\l =1}$ all members of its associated family of minimal surfaces
have the Abresch-Rosenberg differential $\l^{-2} Q dz^2 \;(\l \in \mathbb S^1)$
and the same support $h\,(dz)^{1/2} (d \bar z)^{1/2}$.
To prove that the minimal surfaces in the associated family are graphs, we consider the spacelike {\sc CMC} surfaces $f_{\mathbb L_3}^{\l}$ given in the
proof of Theorem \ref{thm:Bern}.
Then, since $f_{\mathbb L_3}^{\l}|_{\l=1}$ is entire,
it is complete \cite[p.~415]{CY} and thus the spacelike {\sc CMC}
surfaces $f_{\mathbb L_3}^{\l}$
in the associated family are also complete. They are in fact isometric to
$f_{\mathbb L_3}^{\l}|_{\l=1}$. Note that the complete metric is given by
$h^2\, dz d \bar z$, see \eqref{eq:supportmetric}.
Therefore, by \cite[Proposition 1]{Wan}, all
$f_{\mathbb L_3}^{\l}$ are entire graphs, and thus the corresponding minimal surfaces
$f^{\l}$ in ${\rm Nil}_3$ are also
entire graphs.
The completeness of the associated family follows from Lemma \ref{lm:metricrelation}.
\end{proof}
\begin{Remark}[Canonical examples]
In \cite{Fer-Mira2}, all entire, complete, minimal vertical graphs
are called the {\it canonical examples}.
\end{Remark}
|
\section{Introduction}
\label{sec:intro}
\IEEEPARstart{T}{ime series} are sequences of real numbers measured at successive, usually regular time intervals. Data in the form of time series pervade science, business, and society. Examples range from economics to medicine, from biology to physics, and from social to computer sciences. Repetitions or recurrences of similar phenomena are a fundamental characteristic of non-random natural and artificial systems and, as a measurement of the activity of such systems, time series often include pairs of segments of strikingly high similarity. These segment pairs are commonly called motifs~\cite{Lin02WTDM}, and their existence is unlikely to be due to chance alone. In fact, they usually carry important information about the underlying system. Thus, motif discovery is fundamental for understanding, characterizing, modeling, and predicting the system behind the time series~\cite{Mueen14WIRES}. Besides, motif discovery is a core part of several higher-level algorithms dealing with time series, in particular classification, clustering, summarization, compression, and rule-discovery algorithms (see, e.g., references in~\cite{Mueen14WIRES,Mueen13ICDM}).
Identifying similar segment pairs or motifs implies examining all pairwise comparisons between all possible segments in a time series. This, specially when dealing with long time series streams, results in prohibitive time and space complexities. It is for this reason that the majority of motif discovery algorithms resort to some kind of data discretization or approximation that allows them to hash and retrieve segments efficiently. Following the works by Lin~et~al.~\cite{Lin02WTDM} and Chiu~et~al.~\cite{Chiu03KDD}, many of such approaches employ the SAX representation~\cite{Lin07DMKD} and/or a sparse collision matrix~\cite{Buhler02JCB}. These allow them to achieve a theoretically low computational complexity, but sometimes at the expense of very high constant factors. In addition, approximate algorithms usually suffer from a number of data-dependent parameters that, in most situations, are not intuitive to set (e.g.,~time/amplitude resolutions, dissimilarity radius, segment length, minimum segment frequency, etc.).
A few recent approaches overcome some of these limitations. For instance, Castro~\&~Azevedo~\cite{Castro10SDM} propose an amplitude multi-resolution approach to detect frequent segments, Li~\&~Lin~\cite{Li10MDM} use a grammar inference algorithm for exploring motifs with lengths above a certain threshold, Wilson et al.~\cite{Wilson08IJAC} use concepts from immune memory to deal with different lengths, and Floratou~et~al.~\cite{Floratou11TKDE} combine suffix trees with segment models to find motifs of any length. Nevertheless, in general, these approaches still suffer from other data-dependent parameters whose correct tuning can require considerable time. In addition, approximate algorithms are restricted to a specific dissimilarity measure between segments (the one implicit in their discretization step) and do not allow easy access to preliminary results, which is commonly known as anytime algorithms~\cite{Zilberstein96AIM}. Finally, to the best of our knowledge, only~\cite{Tanaka05ML,Yankov07KDD,Tang08KBS} consider the identification of motif pairs containing segments of different lengths. This can be considered a relevant feature, as it produces better results in a number of different domains~\cite{Yankov07KDD}.
In contrast to approximate approaches, algorithms that do not discretize the data have been comparatively much less popular, with low efficiency generally. Exceptions to this statement achieved efficiency by sampling the data stream~\cite{Catalano06PKDD} or by identifying extreme points that constrained the search~\cite{Mohammad09NGC}. In fact, until the work of Mueen~et~al.~\cite{Mueen09SDM}, the exact identification of time series motifs was thought to be intractable for even time series of moderate length. In said work, a clever segment ordering was combined with a lower bound based on the triangular inequality to yield the true, exact, most similar motif. According to the authors, the proposed algorithm was more efficient than existing approaches, including all exact and many approximate ones~\cite{Mueen09SDM}. After Mueen~et~al.'s work, a number of improvements have been proposed, the majority focusing on eliminating the need to set a fixed segment length~\cite{Nunthanid11ECTI,Yingcharxxx13ICDM,Mohammad14ACIIDS}.
Mueen himself has recently published a variable-length motif discovery algorithm which clearly outperforms the iterative search for the optimal length using~\cite{Mueen09SDM} and, from the reported numbers, also outperforms further approaches such as~\cite{Nunthanid11ECTI,Yingcharxxx13ICDM,Mohammad14ACIIDS}. This algorithm, called MOEN~\cite{Mueen13ICDM}, is essentially parameter-free, and is believed to be one of the most efficient motif discovery algorithms available nowadays. However, its complexity is still quadratic in the length of the time series~\cite{Mueen13ICDM}, and hence its applicability to large-scale time series streams remains problematic. Furthermore, in order to derive the lower bounds used, the algorithm is restricted with regard to the dissimilarity measure used to compare time series segments (Euclidean distance after z-normalization). In general, exact motif discovery algorithms have important restrictions with regard to the dissimilarity measure, and many of them still suffer from being non-intuitive and tedious to tune parameters. Moreover, few of them allow for anytime versions and, to the best of our knowledge, not one of them is able to identify motif pairs containing segments of different lengths.
In this article, we propose a new standpoint to time series motif discovery by treating the problem as an anytime multimodal optimization task. To the best of our knowledge, this standpoint is completely unseen in the literature. Here, we firstly reason and discuss its multiple advantages (Sec.~\ref{sec:motif}). Next, we present SWARMMOTIF (Sec.~\ref{sec:swarmmotif}), an anytime algorithm for time series motif discovery based on particle swarm optimization (PSO). We subsequently evaluate the performance of the proposed approach using 9~different real-world time series from distinct domains (Sec.~\ref{sec:eval}). Our results show that SWARMMOTIF is extremely competitive when compared to the state-of-the-art, obtaining motif pairs of comparable similarity in considerably less time and with minimum storage requirements (Sec.~\ref{sec:results}). Moreover, we show that SWARMMOTIF is significantly robust against different implementation choices.
To conclude, we briefly comment on the application of multimodal optimization techniques to time series analysis and mining, which we believe has great potential (Sec.~\ref{sec:conclusion}). The data and code used in our experiments will available onlin
.
\section{Time Series Motif Discovery as an Anytime Multimodal Optimization Task}
\label{sec:motif}
\subsection{Definitions and Task Complexity}
\label{sec:motifdefs}
From the work by Mueen et al.~\cite{Mueen09SDM,Mueen13ICDM}, we can derive a formal, generic similarity-based definition~\cite{Mueen14WIRES} of time series motifs. Given a time series $\textbf{z}$ of length $n$, $\textbf{z}=[z_1,\dots z_n]$, a normalized segment dissimilarity measure $D$, and a temporal window of interest between $w_{\text{min}}$ and $w_{\text{max}}$ samples, the top-$k$ time series motifs $\mathcal{M}=\{\textbf{m}_1,\dots \textbf{m}_k\}$ correspond to the $k$ most similar segment pairs $\textbf{z}_a^{w_a}=[z_a,\dots z_{a+w_a-1}]$ and $\textbf{z}_b^{w_b}=[z_b,\dots z_{b+w_b-1}]$, for $w_a,w_b\in[w_{\text{min}},w_{\text{max}}]$, $a\in[1,n-w_a+1]$, and $b\in[1,n-w_b+1]$ where, in order to avoid repeated and trivial matches~\cite{Lin02WTDM}, $a+w_a< b$. Thus, the $i$-th motif can be fully described by the tuple $\textbf{m}_i=\{a,w_a,b,w_b\}$. The motifs in $\mathcal{M}$ are non-overlapping\footnote{Notice that, following~\cite{Mueen13ICDM}, this definition can be trivially extended to different degrees of overlap.} and ordered from lowest to highest dissimilarity such that $D(\textbf{m}_1)\leq D(\textbf{m}_2) \leq \dots \leq D(\textbf{m}_k)$ where $D(\textbf{m}_i)=D(\{a,w_a,b,w_b\})=D(\textbf{z}_a^{w_a},\textbf{z}_b^{w_b})$. An example of a time series motif pair from a real data set is shown in Fig.~\ref{fig:example}.
\begin{figure}[t]
\centering
\includegraphics{example_motif.pdf}
\caption{Example of a time series motif pair found in the \textsc{Wildlife} time series of~\cite{Mueen10KDD} using SWARMMOTIF and normalized dynamic time warping as the dissimilarity measure: $a=248$, $w_a=244$, $b=720$, and $w_b=235$. Note that $w_a\neq w_b$ and, hence, a warping of the two segments needs to take place. This specific solution cannot be found by any of the approaches mentioned in Sec.~\ref{sec:intro}.}
\label{fig:example}
\end{figure}
It is important to stress that $D$ needs to normalize with respect to the lengths of the considered segments. Otherwise, we would not be able to compare motifs of different lengths. There are many ways to normalize with respect to the length of the considered segments. Ratanamahatana~\&~Keogh~\cite{Ratanamahatana04WMTSD} list a number of intuitive normalization mechanisms for dynamic time warping that can easily be applied to other measures. For instance, in the case of a dissimilarity measure based on the $\text{L}_{\text{p}}$ norm~\cite{Serra14KBS}, we can directly divide by the segment length\footnote{The only exception is with $\text{L}_{\infty}$, which could be considered as already being normalized.}, using brute-force upsampling to the largest length~\cite{Ratanamahatana04WMTSD} when $w_a\neq w_b$.
From the definitions above, we can see that a brute-force search in the motif space for the most similar motifs is of $O(n^2{w_\Delta}^2)$, where $w_\Delta=w_{\text{max}}-w_{\text{min}}+1$ (for the final time complexity one needs to further multiply by the cost of calculating $D$). Hence, for instance, in a perfectly feasible case where $n=10^7$ and $w_\Delta=10^3$, we have $10^{20}$ possibilities. Magnitudes like this challenge the memory and speed of any optimization algorithm, specially if we have no clue to guide the search~\cite{Hillinger10BOOK}. However, it is one of our main objectives to show here that time series generally provide some continuity to this search space, and that this continuity can be exploited by optimization algorithms.
\subsection{Continuity}
\label{sec:motifcont}
A fundamental property of time series is autocorrelation, implying that consecutive samples in a time series have some degree of resemblance and that, most of the time, we do not observe extremal differences between them\footnote{If a time series had no autocorrelation, we might better treat it as an independent random process.}. This property, together with the established ways of computing similarity between time series~\cite{Serra14KBS}, is what gives continuity to our search space. Consider a typical dissimilarity measure like dynamic time warping between z-normalized segments and the time series of Fig.~\ref{fig:example}. If we fix the motif starting points $a$ and $b$ to some random values, we can compute $D(\textbf{z}_a^{i},\textbf{z}_b^{j})$ for $i,j=w_{\text{min}},\dots, w_{\text{max}}$ (Fig.~\ref{fig:continuity}A). We see that these two dimensions have a clear continuity, i.e.,~that $D(\textbf{z}_a^{i},\textbf{z}_b^{j}) \sim D(\textbf{z}_a^{i+1},\textbf{z}_b^{j}) \sim D(\textbf{z}_a^{i},\textbf{z}_b^{j+1}) \sim D(\textbf{z}_a^{i+1},\textbf{z}_b^{j+1})$, and so forth. Similarly, if we fix the motif lengths $w_a$ and $w_b$ to some random values, we can compute $D(\textbf{z}_i^{w_a},\textbf{z}_j^{w_b})$ for $i=1,\dots n-w_a$ and $j=1,\dots n-w_b$ (Fig.~\ref{fig:continuity}B). We see that the remaining two dimensions of the problem also have some continuity, i.e.,~$D(\textbf{z}_i^{w_a},\textbf{z}_j^{w_b+j}) \sim D(\textbf{z}_{i+1}^{w_a},\textbf{z}_{j}^{w_b}) \sim D(\textbf{z}_{i}^{w_a},\textbf{z}_{j+1}^{w_b}) \sim D(\textbf{z}_{i+1}^{w_a},\textbf{z}_{j+1}^{w_b})$, and so forth. The result is a four-dimensional, multimodal, continuous but noisy\footnote{We use the term noisy here to stress that the continuity of the space may be altered at some points due to potential noise in the time series. It is not the case that we have a noisy, unreliable dissimilarity measurement $D$ that could change in successive evaluations.} motif space, where the dissimilarity $D$ acts as the fitness measure and the top-$k$ valley peaks (considering dissimilarity) correspond to the top-$k$ motifs in $\mathcal{M}$.
\begin{figure}[t]
\centering
\includegraphics{continuityA.pdf}
\includegraphics{continuityB.pdf}
\caption{Visualizations of the search space obtained with the \textsc{Wildlife} time series~\cite{Mueen10KDD} and dynamic time warping as the dissimilarity measure: (A) fixing $a=110$ and $b=602$ but $i,j=200,\dots 250$ and (B) fixing $w_a=222$ and $w_b=240$ but $i,j=1,\dots 500$. Darker colors corresponds to more similarity.}
\label{fig:continuity}
\end{figure}
\subsection{Anytime Solutions}
\label{sec:motifany}
Finding an optimization algorithm that can locate the global minima of the previous search spaces faster than existing motif discovery algorithms can be a difficult task. However, we have robust and established algorithms for efficiently locating prominent local minima in complex search spaces~\cite{Blum03ACMCS,Jin05TEVC,Bianchi09NC}. Hence, we can intuitively devise a simple strategy: if we keep the best found minima and randomly reinitialize the optimization algorithm every time it stagnates, we should, sooner or later, start locating the global minima. In the meantime, we could have obtained relatively good candidates. This corresponds to the basic paradigm of anytime algorithms~\cite{Zilberstein96AIM}.
Anytime algorithms have recently been highlighted as ``very beneficial for motif discovery in massive [time series] datasets''~\cite{Yingcharxxx13ICDM}. In an anytime algorithm for motif discovery, $D(\textbf{m}_i)$ improves over time, until it reaches the top-$k$ dissimilarity values $D(\textbf{m}_i)^\ast$ obtained by a brute-force search approach. Thus, we gradually improve $\mathcal{M}$ until we reach the true exact solution $\mathcal{M}^\ast$. A good anytime algorithm will quickly find low $D(\textbf{m}_i)$, ideally reaching $D(\textbf{m}_i)^\ast$ earlier than its non-anytime competitors (Fig.~\ref{fig:anytime}).
Note that a sufficiently good $\mathcal{M}$ may suffice in most situations, without the need that $\mathcal{M}=\mathcal{M}^\ast$. This is particularly true for more exploratory tasks, where one is typically interested in data understanding and visual inspection (see~\cite{Mueen14WIRES}), and can also hold for other tasks, as top-$k$ motifs can be very similar among themselves. In the latter situation, given a seed within $\mathcal{M}^\ast$, we can easily and efficiently retrieve further repetitions via common established approaches~\cite{Kahveci04TKDE,Rakthanmanon12KDD}. Thus, only non-frequent or singular motifs may be missed. These can be valuable too, as the fact that they are non-frequent does not imply that they cannot carry important information (think for example of extreme events of interest that perhaps only happen twice in a measurement). For those singular motifs, we can wait longer if using an anytime algorithm, or we can resort to the state-of-the-art if that is able to provide its output within an affordable time limit.
\begin{figure}[t]
\centering
\includegraphics{anytime_motif.pdf}
\caption{Schema of a plot to assess the performance of an anytime motif discovery algorithm (blue curve). Error bars indicate 5~and 95~percentiles of $D(\textbf{m}_i)$, their central marker indicates the median, and the isolated dots indicate maximum and minimum values. The gray area at the bottom denotes the area where $D(\textbf{m}_i)^\ast$ lie. The top-left black error bar acts as a reference and shows the range of $D(\textbf{m}_i)$ obtained by random sampling the motif space. The bottom-right red error bar is placed at the time that the baseline algorithm spent in the calculations. Better performing anytime algorithms have a curve closer to the bottom left corner, quickly entering the gray area as their execution time increases. Notice that both axes are logarithmic.}
\label{fig:anytime}
\end{figure}
\subsection{Particle Swarm Optimization}
\label{sec:motifpso}
The continuity and anytime observations above (Secs.~\ref{sec:motifcont} and~\ref{sec:motifany}) relax the requirements for the optimization algorithm to be employed in the considered motif spaces. In fact, if we do not have to assess the global optimality of a solution, we have a number of approaches that can deal with large, multimodal, continuous but noisy search spaces~\cite{Blum03ACMCS,Jin05TEVC,Bianchi09NC}. Among them, we choose PSO~\cite{Clerc06BOOK,Poli07SI,Banks07aNC,Blum08BOOK,Parsopoulos10BOOK}. PSO is a population-based stochastic approach for solving continuous and discrete optimization problems~\cite{Blum08BOOK} which has been applied to multimodal problems~\cite{Barrera09BOOKCHAP}. It is a metaheuristic~\cite{Bianchi09NC}, meaning that it cannot guarantee whether the found solution corresponds to a global optimum. The original PSO algorithm cannot even guarantee the convergence to a local optimum, but adapted versions of it have been proven to solve this issue~\cite{VanDerBerg10FI}. Other versions guarantee the convergence to the global optimum, but only with the number of iterations approaching infinity~\cite{VanDerBerg10FI}.
PSO has gained increasing popularity among researchers and practitioners as a robust and efficient technique for solving difficult optimization problems. It makes few or no assumptions about the problem being optimized, does not require it to be differentiable, can search very large spaces of candidate solutions, and can be applied to problems that are irregular, incomplete, noisy, dynamic, etc.\ (see~\cite{Blum08BOOK, Clerc06BOOK, Poli07SI, Banks07aNC, Parsopoulos10BOOK, Barrera09BOOKCHAP} and references therein). PSO iteratively tries to improve a candidate solution with regard to a given measure of quality or fitness function. Hence, furthermore, it can be considered an anytime algorithm.
\subsection{Advantages of an Optimization-Based Solution Using Particle Swarms}
\label{sec:motifadv}
Notice that treating time series motif discovery as an optimization problem naturally yields several advantages:
\begin{enumerate}
\item We do not require much memory, as we can basically store only the stream time series and preprocess the required segments at every fitness evaluation.
\item We are able to achieve a certain efficiency, as optimization algorithms do not usually explore the full solution space and perform few fitness evaluations~\cite{Hillinger10BOOK}.
\item We can employ any dissimilarity measure $D$ as our fitness function. Its only requirements are segment length independence and a minimal search space continuity. Intuitively, this holds for the high majority of time series dissimilarity measures that are currently used (Secs.~\ref{sec:motifdefs} and~\ref{sec:motifcont}). Additionally, we can straightforwardly incorporate notions of `interestingness', hubness, or complexity (see references in~\cite{Serra14KBS}). This flexibility is very uncommon in current time series motif discovery algorithms (Sec.~\ref{sec:intro}).
\item We do not need to force the two segments of the motif to be of the same length. The dissimilarity function $D$ can expressly handle segments of different lengths or we can simply upsample to the largest length (see~\cite{Ratanamahatana04WMTSD}). Although considering different segment lengths has been highlighted as an objectively better approach, practically none of the current time series motif discovery algorithms contemplates this option (Sec.~\ref{sec:intro}).
\item Since we search for the optimal $w_a$ and $w_b$, together with $a$ and $b$, we do not need to set the exact segment lengths as a parameter. Instead, we can use a more intuitive and easier to set range of lengths $w_a,w_b\in[w_{\text{min}},w_{\text{max}}]$.
\item We can easily modify our fitness criterion to work with different task settings. Thus, just by replacing $D$, we are able to work with multi-dimensional time series~\cite{Hao14IV}, detect sub-dimensional motifs~\cite{Minnen07ICDM}, perform a constrained motif discovery task~\cite{Mohammad09NGC}, etc.
\item We can incorporate notions of motif frequency to our fitness function and hence expand our similarity-based definition of motif to incorporate both notions~\cite{Mueen14WIRES}. For instance, instead of optimizing for individual motifs $\textbf{m}_i$, we can optimize sets of motifs $\mathcal{M}'_i$ of size $r_i$ such that $\frac{1}{r_i} \sum_{\textbf{m}_j\in \mathcal{M}'_i} D(\textbf{m}_j)$ is minimal. We can choose $r_i$ to be a minimum frequency of motif appearance or we can even decide to optimize it following any suitable criterion.
\end{enumerate}
In addition, using PSO has a number of interesting properties, some of which may be shared with other metaheurisics:
\begin{enumerate}
\item We have a straightforward mapping to the problem at hand (Sec.~\ref{sec:swarmmotifmain}).
\item By construction, we have an anytime algorithm (Sec.~\ref{sec:motifpso}).
\item We can obtain accurate and much faster solutions, as compared to the state-of-the-art in time series motif discovery (Sec.~\ref{sec:resfinal}).
\item We have an essentially parameter-free algorithm~\cite{Blum08BOOK}. As will be shown, all our parameter choices turn out to be non-critical to achieve the most competitive performances (Secs.~\ref{sec:resparam} and~\ref{sec:resvariants}).
\item We have an easily parallelizable algorithm. The agent-based nature of PSO naturally yields to parallel implementations~\cite{Banks07aNC}.
\item We still have the possibility to apply lower bounding techniques to $D$ in order to reduce its computational cost~\cite{Mueen14WIRES,Rakthanmanon12KDD}. Among others, we may exploit the particles' best-so-far values or spatially close dissimilarities.
\item All of these use a simple, easy to implement algorithm requiring low storage capabilities (Sec.~\ref{sec:swarmmotifdetail}).
\end{enumerate}
\section{SWARMMOTIF}
\label{sec:swarmmotif}
\subsection{Main Algorithm}
\label{sec:swarmmotifmain}
Our PSO approach to time series motif discovery is based on the combination of two well-known extensions to the canonical PSO~\cite{Poli07SI}. On one hand, we employ multiple reinitializations of the swarm on stagnation~\cite{Eberhart01CEC}. On the other hand, we exploit the particles' ``local memories'' with the intention of forming stable niches across different local minima~\cite{Li10TEVC}. The former emulates a parallel multi-swarm approach~\cite{Barrera09BOOKCHAP} without the need of having to define the number of swarms and their communication. The latter, when combined with the former, results in a low-complexity niching strategy~\cite{Barrera09BOOKCHAP} that does not require niching parameters (see the related discussion in~\cite{Bonyadi14SEC,Bonyadi14BOOKCHAP}). SWARMMOTIF, the implementation of the two extensions, is detailed in Algorithm~\ref{alg:pso}.
\begin{algorithm}[t]
\caption{\textsc{SWARMMOTIF}($\textbf{z}$,$D$,$w_{\text{min}}$,$w_{\text{max}}$,$k$,$t_{\text{max}}$)}
\label{alg:pso}
\algsetup{indent=0.5cm}
\begin{algorithmic}[1]
\vspace*{0.1cm}
\REQUIRE Time series $\textbf{z}$ of length $n$, dissimilarity measure $D$, minimum and maximum segment length $w_{\text{min}}$ and $w_{\text{max}}$, number of motifs $k$, and maximum amount of time (number of iterations) $t_{\text{max}}$. \\
{\renewcommand{\algorithmicrequire}{\textbf{Require:}}
\REQUIRE Number of particles $\kappa$, topology $\theta$, constriction constant $\phi$, and maximum amount of time at stagnation (number of iterations) $\tau$. \\
}
\ENSURE A set of motifs $\mathcal{M}$. \\
\vspace*{0.1cm}
\STATE $c_0,c_1,c_2 \leftarrow$ \textsc{GetConstants}($\phi$)
\STATE $\mathcal{X},\mathcal{V},\mathcal{S},\mathcal{P} \leftarrow$ \textsc{InitializeSwarm}($n$,$w_{\text{min}}$,$w_{\text{max}}$,$\kappa$)
\STATE $\Theta \leftarrow$ \textsc{InitializeTopology}($\theta$,$\kappa$)
\STATE $s^\ast \leftarrow \infty$
\STATE $\mathcal{M} \leftarrow$ \textsc{EmptyPriorityQueue}()
\FOR{$t=1,\dots t_{\text{max}}$}
\FOR{$i=1,\dots \kappa$}
\IF{ \textsc{ValidPosition}($\textbf{x}_i$)}
\STATE $d \leftarrow D(\textbf{x}_i)$
\IF{$d < s_i$}
\STATE $s_i \leftarrow d$
\STATE $\textbf{p}_i \leftarrow \textbf{x}_i$
\STATE $\mathcal{M}$.\textsc{Push}($d$,$\textbf{x}_i$)
\IF{$d < s^\ast$}
\STATE $s^\ast \leftarrow d$
\STATE $t_{\text{update}} \leftarrow t$
\ENDIF
\ENDIF
\ENDIF
\ENDFOR
\FOR{$i=1,\dots \kappa$}
\STATE $g \leftarrow i$
\FOR{$j$ \textbf{in} $\Theta_i$}
\IF{$s_j \leq s_g$}
\STATE $g \leftarrow j$
\ENDIF
\ENDFOR
\STATE $\textbf{v}_i \leftarrow c_0\textbf{v}_i + c_1\textbf{u}_1\otimes(\textbf{p}_i-\textbf{x}_i) + c_2\textbf{u}_2\otimes(\textbf{p}_g-\textbf{x}_i)$
\STATE $\textbf{x}_i \leftarrow \textbf{x}_i + \textbf{v}_i$
\ENDFOR
\IF{$t-t_{\text{update}} = \tau$}
\STATE $\mathcal{X},\mathcal{V},\mathcal{S},\mathcal{P} \leftarrow$ \textsc{InitializeSwarm}($n$,$w_{\text{min}}$,$w_{\text{max}}$,$\kappa$)
\STATE $s^\ast \leftarrow \infty$
\ENDIF
\ENDFOR
\RETURN \textsc{NonOverlapping}($\mathcal{M}$,$k$) \\
\end{algorithmic}
\end{algorithm}
SWARMMOTIF takes a time series $\textbf{z}$ of length $n$ as input, together with a segment dissimilarity measure $D$, and the range of segment lengths of interest, limited by $w_{\text{min}}$ and $w_{\text{max}}$. The user also needs to specify $k$, the desired number of motifs, and $t_{\text{max}}$, the maximum time spent by the algorithm (in iterations\footnote{The number of iterations is easy to infer from the available time as, for the same input, the elapsed time will be roughly directly proportional to the number of iterations.}). SWARMMOTIF outputs a set of $k$ non-overlapping motifs $\mathcal{M}$. We implement $\mathcal{M}$ as a priority queue, which typically stores more than $k$ elements to ensure that it contains $k$ non-overlapping motifs. This way, by sorting the motif candidates as soon as they are found, we allow potential queries to $\mathcal{M}$ at any time during the algorithm's execution. In that case, we only need to dynamically check the candidates' overlap (Sec.~\ref{sec:swarmmotifdetail}). Notice that $n$, $D$, $w_{\text{min}}$, $w_{\text{max}}$, $k$, and $t_{\text{max}}$ are not parameters of the algorithm, but requirements of the task (they depend on the data, the problem, and the available time). The only parameters to be set, as specified in Algorithm~\ref{alg:pso}'s requirements, are the number of particles $\kappa$, the topology $\theta$, the constriction constant $\phi$, and the maximum amount of iterations at stagnation $\tau$. Nevertheless, we will show that practically none of the possible parameter choices introduces a significant variation in the reported performance (Sec.~\ref{sec:resparam}).
Having clarified SWARMMOTIF's input, output, and requirements, we now elaborate on its procedures. Algorithm~\ref{alg:pso} starts by computing the velocity update constants (line~1) following Clerc's constriction method~\cite{Clerc02TEVC}, i.e.,
\begin{equation*}
c_0 = \frac{2}{\left\vert 2-\phi-\sqrt{\phi^2-4\phi} \right\vert}
\end{equation*}
and
\begin{equation}
c_1=c_2=c_0\phi/2 .
\label{eq:c1c2}
\end{equation}
Next, a swarm with $\kappa$ particles is initialized (line~2). The swarm is formed by four data structures: a set of particle positions $\mathcal{X}=\{\textbf{x}_1,\dots \textbf{x}_\kappa\}$, a set of particle velocities $\mathcal{V}=\{\textbf{v}_1,\dots \textbf{v}_\kappa\}$, a set of particle best scores $\mathcal{S}=\{s_1,\dots s_\kappa\}$, and a set of particle best positions $\mathcal{P}=\{\textbf{p}_1,\dots \textbf{p}_\kappa\}$ (the initialization of these four data structures is detailed in Algorithm~\ref{alg:swarminit}). Particles' positions $\textbf{x}_i$ and $\textbf{p}_i$ completely determine a motif candidate, and have a direct correspondence with $\textbf{m}_i$ (see Sec.~\ref{sec:swarmmotifdetail}). A further data structure $\Theta$ indicates the indices of the neighbors of each particle according to a given social topology $\theta$ (line~3). Apart from the swarm, we also initialize a global best score $s^\ast$ (line~4) and the priority queue $\mathcal{M}$ (line~5). We then enter the main loop (lines~6--26). In it, we perform three main actions. Firstly, we compute the particles' fitness and perform the necessary updates (lines~7--16). Secondly, we modify the particles' position and velocity using their personal and neighborhood best positions (lines~17--23). Thirdly, we control for stagnation and reinitialize the swarm if needed (lines~24--26). Finally, when we exit the loop, we return the first $k$ non-overlapping motif candidates from $\mathcal{M}$ (line~27).
\begin{algorithm}[t]
\caption{\textsc{InitializeSwarm}($n$,$w_{\text{min}}$,$w_{\text{max}}$,$\kappa$)}
\label{alg:swarminit}
\algsetup{indent=0.5cm}
\begin{algorithmic}[1]
\vspace*{0.1cm}
\REQUIRE Time series length $n$, minimum and maximum segment length $w_{\text{min}}$ and $w_{\text{max}}$, and number of particles $\kappa$. \\
\ENSURE Particle positions $\mathcal{X}$, velocities $\mathcal{V}$, best scores $\mathcal{S}$, and best positions $\mathcal{P}$. \\
\vspace*{0.1cm}
\FOR{$i=1,\dots \kappa$}
\STATE $x_{i,2} \leftarrow w_{\text{min}}+(w_{\text{max}}+1-w_{\text{min}})u$
\STATE $x_{i,4} \leftarrow w_{\text{min}}+(w_{\text{max}}+1-w_{\text{min}})u$
\STATE $x_{i,1} \leftarrow 1+(n-x_{i,2})(1-\sqrt{u})$
\STATE $x_{i,3} \leftarrow 1+(n-x_{i,4}-(x_{i,1}+x_{i,2}))u$
\STATE $ \textbf{x}' \leftarrow$ \textit{As in lines 2--5}
\STATE $\textbf{v}_i \leftarrow \textbf{x}'-\textbf{x}_i$
\STATE $s_i \leftarrow \infty$
\STATE $\textbf{p}_i \leftarrow \textbf{x}_i$
\ENDFOR
\RETURN $\mathcal{X},\mathcal{V},\mathcal{S},\mathcal{P}$ \\
\end{algorithmic}
\end{algorithm}
The particles' fitness loop (lines~7--16) can be described as follows. For the particles that have a valid position within the ranges used for particle initializations (line~8; see also Algorithm~\ref{alg:swarminit} for initializations), we calculate their fitness $D$ (line~9) and, if needed, update their personal bests $s_i$ and $\textbf{p}_i$ (lines 10--12). As mentioned, $D$ needs to be independent of the segments' lengths, which is typically an easy condition for time series dissimilarity measures (Sec.~\ref{sec:motifdefs}). In the case that the particles find a new personal best, we save the motif dissimilarity $d$ and its position $\textbf{x}_i$ into $\mathcal{M}$ (line~13). Next, we update $t_{\text{update}}$, the last iteration when an improvement of the global best score $s^\ast$ has occurred (lines~14--16).
The particles' update loop (lines~17--23) is straightforward. We first select each particle's best neighbor $g$ using the neighborhood personal best scores $s_j$ (lines~18--21). Then, we use the positions of the best neighbor's personal best $\textbf{p}_g$ and the particle's personal best $\textbf{p}_i$ to compute its new velocity and position (lines~22--23). We employ component-wise multiplication, denoted by $\otimes$, and two random vectors $\textbf{u}_1$ and $\textbf{u}_2$ whose individual components $u_{i,j}=U(0,1)$, being $U(l,h)$ a uniform real random number generator such that $l\leq U(l,h) < h$. Note that by considering the particles' neighborhood personal bests $\textbf{p}_g$ we follow the aforementioned local neighborhood niching strategy~\cite{Li10TEVC}. At the end of the loop we control for stagnation by counting the number of iterations since the last global best update and applying a threshold $\tau$ (line~24). Note that this is the mechanism responsible for the aforementioned multiple reinitialization strategy~\cite{Eberhart01CEC}.
The initialization of the swarm used in Algorithm~\ref{alg:pso} (lines~2 and~25) is further detailed in Algorithm~\ref{alg:swarminit}. In it, for each particle, two random positions $\textbf{x}_i$ and $\textbf{x}'$ are drawn (lines 2--6) and the initial velocity is computed as the subtraction of the two (line 7). To obtain $\textbf{x}_i$ and $\textbf{x}'$, uniform real random numbers $u=U(0,1)$ are subsequently generated. The personal best score $s_i$ is set to infinite (line 8) and $\textbf{x}_i$ is taken as the current best position $\textbf{p}_i$ (line 9). Note that $\sqrt{u}$ (line 4) is used to ensure a uniform distribution of the particles across the triangular subspace formed by $x_{i,1}$ and $x_{i,3}$ (line 5; see also Sec.~\ref{sec:motifdefs}).
\subsection{Implementation Details}
\label{sec:swarmmotifdetail}
Some implementation details are missing in Algorithms~\ref{alg:pso} and~\ref{alg:swarminit}. Firstly, positions $\textbf{x}_i$ are floored component-wise inside \textsc{ValidPosition}, $D$, and $\mathcal{M}$ (thus obtaining motif $\textbf{m}_i$). Secondly, the motif priority queue $\mathcal{M}$ is implemented as an associative container (logarithmic insertion time) that sorts its elements according to $d$ and stores $\textbf{m}_i$. Thirdly, the last visited positions are cached into a hash table (constant lookup time) in order to avoid some of the possible repeated dissimilarity computations. Fourthly, we incorporate the option to constrain the motif search by specifying a maximum segment stretch in Algorithm~\ref{alg:swarminit} and \textsc{ValidPosition}. Finally, the function that returns the non-overlapping top-$k$ motifs employs a boolean array of size $n$ in order to avoid $O(k^2)$ comparisons between members of the queue (cf.~\cite{Mueen13ICDM}). Notice that we have a memory-efficient implementation, as we basically only need to store $\textbf{z}$ and the boolean array (both of $O(n)$ space), $\mathcal{M}$ (of $O(k)$ space, $k\ll n$), and $\mathcal{X}$, $\mathcal{V}$, $\mathcal{S}$, $\mathcal{P}$, and $\Theta$ (all of them of $O(\kappa)$ space, $\kappa\ll n$). The aforementioned hash table (optional) can be allocated in any predefined, available memory segment. For the sake of brevity, the interested reader is referred to the provided code for a full account of the outlined implementation details.
\subsection{Variants}
\label{sec:swarmmotifvariants}
Given the main Algorithm~\ref{alg:pso}, we consider a number of variations that may potentially improve SWARMMOTIF's performance without introducing too much algorithmic complexity:
\begin{itemize}
\item Sociability: We study whether a ``cognitive-only'' model, a ``social-only'' one, or different weightings of the two yield to some improvements~\cite{Kennedy97CEC}. To do so, we just need to introduce a parameter $\alpha\in[0,1]$ controlling the degree of `sociability' of the particles, and implement lines~1--2 of Algorithm~\ref{alg:variants} instead of Eq.~\ref{eq:c1c2}.
\item Stochastic: We investigate the use of a random inertia weight~\cite{Eberhart01CEC}. This may alleviate the need of using the same $c_0$ in different environments, providing a potentially more adaptive trade-off between exploration and exploitation (also controlled by $\alpha$ in the previous point). To consider this variant, we just need to replace line~22 in Algorithm~\ref{alg:pso} by line~3 in Algorithm~\ref{alg:variants}.
\item Velocity clamping: In addition to constriction, we study limiting the maximum velocity of the particles~\cite{Kennedy95ICNN}. Empirical studies have shown that the simultaneous consideration of a constriction factor and velocity clamping results in improved performance on certain problems~\cite{Eberhart00CEC}. To apply velocity clamping we add lines~4--6 of Algorithm~\ref{alg:variants} between lines~22 and~23 of Algorithm~\ref{alg:pso}.
\item Craziness: We introduce so-called ``craziness'' or ``perturbation'' in the particles' velocities, as initially suggested by Kennedy~\&~Eberhart~\cite{Kennedy95ICNN}. In such variant, inspired by the sudden direction changes observed in flocking birds, the particles' velocity is altered with a certain probability $\rho$, with the aim of favoring exploration by increasing directional diversity and discouraging premature convergence~\cite{Fourie00IWMDO}. We coincide with~\cite{Fourie00IWMDO} in that, in some sense, this can be seen as a mutation operation. To implement craziness we add lines 7--10 of Algorithm~\ref{alg:variants} between lines~22 and~23 of Algorithm~\ref{alg:pso}.
\end{itemize}
\begin{algorithm}[t]
\caption{Variations to Algorithm~\ref{alg:pso}: sociability (lines~1--2), stochastic (line~3), velocity clamping (lines~4--6), and craziness (lines~7--10).}
\label{alg:variants}
\algsetup{indent=0.5cm}
\begin{algorithmic}[1]
\vspace*{0.1cm}
\STATE $c_1 \leftarrow c_0\phi(1-\alpha)$
\STATE $c_2 \leftarrow c_0\phi\alpha$
\\ \vspace*{-0.1cm}
\hrulefill
\vspace*{0.1cm}
\STATE $\textbf{v}_i \leftarrow (1-2(1-c_0))u\textbf{v}_i + c_1\textbf{u}_1\otimes(\textbf{p}_i-\textbf{x}_i) + c_2\textbf{u}_2\otimes(\textbf{p}_g-\textbf{x}_i)$
\\ \vspace*{-0.1cm}
\hrulefill
\vspace*{0.1cm}
\STATE $\textbf{v}^{\text{range}} \leftarrow [n,w_\Delta,n,w_\Delta]/2$
\FOR{$v_j^{\text{range}}$ \textbf{in} $\textbf{v}^{\text{range}}$}
\STATE $v_{i,j} \leftarrow \text{min}(v_j^{\text{range}},\text{max}(-v_j^{\text{range}},v_{i,j}))$
\ENDFOR
\vspace*{-0.1cm}
\hrulefill
\vspace*{0.1cm}
\STATE $\textbf{v}'_i \leftarrow$ \textit{As in Algorithm~\ref{alg:swarminit}}
\FOR{$v_{i,j}$ \textbf{in} $\textbf{v}_i$}
\IF{$u<\rho$}
\STATE $v_{i,j} \leftarrow v'_{i,j}$.
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
\section{Evaluation Methodology}
\label{sec:eval}
To evaluate SWARMMOTIF's speed and accuracy we consider plots like the one presented in Fig.~\ref{fig:anytime}. As a reference, we draw uniform random samples from the motif search space and compute their dissimilarities (we take as many samples as the length $n$ of the time series). As a baseline, we use the top-25 motifs found by MOEN~\cite{Mueen13ICDM}, which we will denote by $\mathcal{M}^\ast$. Existing empirical evidence~\cite{Mueen09SDM,Mueen13ICDM} suggests that MOEN is the most efficient algorithm to retrieve the top exact similarity-based motifs in a range of lengths\footnote{Besides, we could not find any other promising exact or anytime approach with some available code, nor with sufficient detail to allow a reliable implementation.} (Sec.~\ref{sec:intro}). Notice furthermore that here we are not that interested in obtaining all top-25 true exact motifs, but more concerned on obtaining good seed motifs within these using an anytime approach (Sec.~\ref{sec:motifany}).
As its competitors, MOEN has however some limitations (Sec.~\ref{sec:intro}). Thus, to fairly compare results, we have to apply some constrains to our algorithm. Since MOEN can only use the Euclidean distance between z-normalized segments, here we also adopt this formulation for $D$. In addition, as MOEN only considers pairs of segments of the same length (without resampling), we have to constrain SWARMMOTIF so that $x_{i,2}=x_{i,4}$. Therefore, the reported motif dissimilarities $D(\textbf{m}_i)$ correspond to the Euclidean distance between two z-normalized segments of the same length (we divide by the length of the segments to compare different segment lengths, Sec.~\ref{sec:motifdefs}). In the reported experiments, SWARMMOTIF is run 10 times with $k=10$. We stop its execution when we find 95\% of $D(\textbf{m}_i)$ within $\mathcal{M}^\ast$. This way, we assess the time taken to retrieve any 10 motifs from those with at least 95\% confidence. All experiments are performed using a single core of an Intel\textsuperscript{\textregistered} Xeon\textsuperscript{\textregistered} CPU E5-2620 at 2.00 GHz.
To demonstrate that SWARMMOTIF is not biased towards a particular data source, time series length, or motif length, we consider 9~different time series of varying length, coming from distinct domains, and a number of arbitrary but source-consistent motif lengths (Table~\ref{tab:datasets}). As mentioned, we make these time series and our code available online (Sec.~\ref{sec:intro}). Four of the time series have been used for motif discovery in previous studies~\cite{Mueen09SDM,Mueen09ICDM}, while the other five are employed here for the first time for this task:
\begin{enumerate}
\item \textsc{DowJones}: The daily closing values of the Dow Jones average in the USA from May 2, 1885 to April 22, 2014~\cite{Williamson12WEB}.
\item \textsc{CarCount}: The number of cars measured for the Glendale on ramp for the 101~North freeway in Los Angeles, CA, USA~\cite{Ihler06KDD}. The measurement was carried out by the Freeway Performance Measurement System\footnote{http://pems.dot.ca.gov} and the data was retrieved from the UCI Machine Learning Repository~\cite{UCI13WEB}. Segments of missing values were manually interpolated or removed.
\begin{table}[!t]
\renewcommand{\arraystretch}{1.1}
\caption{Characteristics of the time series used for evaluation, together with the considered segment length ranges.}
\label{tab:datasets}
\renewcommand{\tabcolsep}{4.5pt}
\centering
\begin{tabular}{l|ccccc}
\hline\hline
Name & Duration & Sample Rate & $n$ & $w_{\text{min}}$ & $w_{\text{max}}$ \\
\hline
\textsc{DowJones} & 129\,y & 1\,d$^{-1}$ & 35503 & 30 & 300 \\
\textsc{CarCount} & 175\,d & 1/5\,min$^{-1}$ & 47534 & 270 & 320 \\
\textsc{Insect} & 32\,min & 38\,Hz & 73929 & 300 & 500 \\
\textsc{EEG} & 1\,h & 50\,Hz & 180214 & 150 & 250 \\
\textsc{FieldRecording} & 3\,h & 21.5\,Hz & 226383 & 50 & 150 \\
\textsc{Wind} & 4\,y & 1/6\,min$^{-1}$ & 369138 & 200 & 280 \\
\textsc{Power} & 32\,months & 1/5\,min$^{-1}$ & 410608 & 250 & 620 \\
\textsc{EOG} & 9\,h & 25\,Hz & 809948 & 150 & 200 \\
\textsc{RandomWalk} & - & - & 1000000 & 100 & 250 \\
\hline\hline
\end{tabular}
\end{table}
\item \textsc{Insect}: The electrical penetration graph of a beet leafhopper (\textit{circulifer tenellus})~\cite{Mueen09SDM}. The time series was retrieved from Mueen's website\footnote{\url{http://www.cs.ucr.edu/~mueen/MK}}.
\item \textsc{EEG}: A one hour electroencephalogram (in $\mu$V) from a single channel in a sleeping patient~\cite{Mueen09SDM}. The time series was retrieved from Mueen's website\footnote{\url{http://www.cs.ucr.edu/~mueen/OnlineMotif}} and, according to the authors, was smoothed and filtered using domain-standard procedures.
\item \textsc{FieldRecording}: The spectral centroid (in Hz) of a field recording retrieved from Freesound\footnote{\url{http://www.freesound.org/people/JeffWojo/sounds/121250}}~\cite{Font13ACMMM}. We used the mean of the stereo channels and the spectral centroid (linear frequency) Vamp SDK example plugin from Sonic Visualizer~\cite{Cannam10ACMMM}. We used a Hann window of 8192~samples at 44.1~KHz with 75\% overlap. \item \textsc{Wind}: The wind speed (in m/s) registered in the buoy of Rincon del San Jose, TX, USA, between January 1, 2010 and April 11, 2014. The time series was retrieved from the Texas Coastal Ocean Observation Network website\footnote{\url{http://lighthouse.tamucc.edu/pq}}. Segments of missing values were manually interpolated or removed.
\item \textsc{Power}: The electric power consumption (in KW) of an individual household\footnote{\url{http://archive.ics.uci.edu/ml/datasets/Individual+household+electric+power+consumption}}. The data was retrieved from the UCI Machine Learning Repository~\cite{UCI13WEB}. We took the global active power, removed missing values, and downsampled the original time series by a factor of 5 using averaging and 50\% overlap.
\item \textsc{EOG}: An electrooculogram tracking the eye movements of a sleeping patient~\cite{Goldberger00CIRCULATION}. We took the downsampled time series~\cite{Mueen09ICDM} from Mueen's web page\footnote{\url{http://www.cs.ucr.edu/~mueen/DAME}}.
\item \textsc{RandomWalk}: A random walk time series. This was artificially synthesized using $z_{i+1}=z_i+N(0,1)$ for $i=2,\dots n$ and $z_1=0$, where $N(0,1)$ is a real Gaussian random number generator with zero mean and unit variance.
\end{enumerate}
To assess the statistical significance of the differences between alternative parameter settings, we employ a two stage approach. First, we consider all settings at the same time and perform the Friedman's test~\cite{Hollander99BOOK}, which is a non-parametric statistical test used to detect differences in treatments across multiple test attempts. We use as inputs the median values for all settings for 25 equally-spaced time steps. In the case some difference between settings is detected (i.e.,~we reject the null hypothesis that the settings' performances come from the same distribution), we proceed to the second stage. In it, we perform all possible pairwise comparisons between settings using the Wilcoxon signed-rank test~\cite{Hollander99BOOK}, another non-parametric statistical hypothesis test used for comparing matched samples. To counteract the problem of multiple comparisons and control the so-called family-wise error rate, we employ the Holm-Bonferroni correction~\cite{Holm79SJS}. In all statistical tests, we consider a significance level of 0.01.
\section{Results}
\label{sec:results}
\subsection{Configuration}
\label{sec:resparam}
In pre-analysis, and according to common practice, we set $\kappa=100$, $\phi=4.1$, and $\tau=2000$. We then experimented with 6~different static topologies $\theta$~\cite{Mendes04THESIS}: global best, local best (two neighbors), Von Neumann, random (three neighbors), wheel, and binary tree. The results showed the qualitative equivalence of all topologies except, perhaps, global best and wheel (Fig.~\ref{fig:restopology}). In some data sets, these two turned out to yield slightly worse performances for short-time runs of the algorithm (small $t$), although for longer runs they gradually became equivalent to the rest. However, in general, no systematic statistically significant difference was detected between topologies. With this in mind, we chose the local best topology to further favor exploration and parallelism, and to be more consistent with our local neighborhood design principle of Sec.~\ref{sec:swarmmotifmain}.
\begin{figure}[t]
\centering
\includegraphics{eeg_logy2.pdf}
\caption{Effect of $\theta$ on the \textsc{EEG} time series. Equivalent results were observed with the other time series.}
\label{fig:restopology}
\end{figure}
Next, we studied the effect of the number of particles $\kappa$ and the stagnation threshold $\tau$. To do so, we kept the previous configuration with the local best topology and subsequently evaluated $\kappa=\{20,40,80,160,320\}$ and $\tau=\{500,1000,2000,4000,8000\}$. Essentially, we observed almost no performance changes under these alternative settings (Figs.~\ref{fig:resnparticles} and~\ref{fig:resiterconv}). We only found a statistically significant difference in the case of the \textsc{CarCount} data set. Specifically, the performance with $\tau=500$ was found to be statistically significantly worse than $\tau\geq 2000$. Regarding $\kappa$, and after considering different $n$, $w_\Delta$ and $k$, a partial tendency seemed to emerge: an increasing number of particles $\kappa$ was slightly beneficial for increasing lengths $n$, increasing $w_\Delta$, and increasing $k$. Unfortunately, we could not obtain strong empirical evidence nor formal proof for this statement. Nonetheless, in subsequent experiments, we decided to use a value for $\kappa$ and $\tau$ that dynamically adapts SWARMMOTIF's configuration to such predefined task parameters. We arbitrarily set $\tau$ proportional to $\kappa$, and $\kappa$ proportional to $n$ and in direct relation to $w_\Delta$ and $k$ (we refer to the provided code for the exact formulation).
\begin{figure}[t]
\centering
\includegraphics{wind_cles2.pdf}
\caption{Effect of $\kappa$ on the \textsc{Wind} time series. Equivalent results were observed with the other time series.}
\label{fig:resnparticles}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics{fieldrecording_conv2.pdf}
\caption{Effect of $\tau$ on the \textsc{FieldRecording} time series. Equivalent results were observed with the other time series.}
\label{fig:resiterconv}
\end{figure}
To conclude our pre-analysis, we studied the influence of the constriction constant $\phi$. Following common practice, we considered $\phi=\{4.02,4.05,4.1,4.2,4.4,4.8\}$. In this case, we saw that high values had a negative impact on performance (Fig.~\ref{fig:resphi}). In particular, values of $\phi\geq 4.2$ or $\phi\geq 4.4$, depending on the data set, statistically significantly increased the motif dissimilarities at a given $t$. Contrastingly, values $4<\phi<4.2$ yielded stable dissimilarities with no statistically significant variation (in some data sets, this range could be extended to $4<\phi\leq4.4$). It is well-known that higher $\phi$ values favor exploitation rather than exploration~\cite{Clerc02TEVC}. Hence, it is not strange to observe that low $\phi$ values are more appropriate for searching the large motif spaces we consider here. We finally chose $\phi=4.05$.
\begin{figure}[t]
\centering
\includegraphics{insect_-phi2.pdf}
\caption{Effect of $\phi$ on the \textsc{Insect} time series. Equivalent results were observed with the other time series.}
\label{fig:resphi}
\end{figure}
Overall, the result of our pre-analysis suggests a high degree of robustness with respect to the possible configurations. The topology $\theta$, the number of particles $\kappa$, the stagnation threshold $\tau$, and the constriction constant $\phi$ have, in general, no significant influence on the obtained results. The only consistent exception is observed for values of $\phi\geq 4.4$, which are not the most common practice~\cite{Parsopoulos10BOOK}. The global best and wheel topologies could also constitute a further exception. However, as we have shown, these become qualitatively equivalent to the rest as execution time $t$ increases, yielding no statistically significant difference. We believe that the reported stability of SWARMMOTIF against the tested configurations and data sets justifies the use of our setting for finding motifs in diverse time series coming from further application domains.
\subsection{Variants}
\label{sec:resvariants}
Using the configuration resulting from the previous section, we subsequently assessed the performance of the variations considered in Sec.~\ref{sec:swarmmotifvariants}. We started with the sociability variant, experimenting with social-only models, $\alpha=1$, cognitive-only models, $\alpha=0$, and intermediate configurations, $\alpha=\{0.2,0.33,0.66,0.8\}$. Apart from the fact that no clear tendency could be observed, none of the previous settings was able to consistently reach the performance achieved by the original variant ($\alpha=1/2$, Eq.~\ref{eq:c1c2}) in all time series. That is, none of the previous settings could statistically significantly outperform $\alpha=1/2$ in the majority of the data sets.
Next, we experimented with the stochastic and the velocity clamping variants. While the former did not improve our results, the latter led to a statistically significant improvement for some time series. Because of that, we decided to discard the use of a stochastic variant but to incorporate velocity clamping to our main algorithm. The former could be difficult to justify while the latter has empirical evidence behind it (Sec.~\ref{sec:swarmmotifvariants}).
Finally, we experimented with craziness and its probability $\rho$. The results showed a similar performance for $0\leq\rho<0.001$, a slightly better performance for $0.001\leq\rho\leq 0.01$, and an increasingly worse performance for $\rho>0.01$ (Fig.~\ref{fig:rescrazy}). A statistically significant difference was found between $\rho\leq 0.01$ and $\rho> 0.1$, being $\rho> 0.1$ a consistently worse setting. These results were expected, as the swarm performs a more random search with increasing $\rho$, being completely random in the limiting case of $\rho=1$. The slightly better performance for $0.001\leq\rho\leq 0.01$ was not found to be statistically significant under our criteria. However, it was visually noticeable for some data sets. For instance, with the \textsc{EEG} data set, we see that curves 33 and 34 hit the dissimilarities of the true exact motif set $\mathcal{M}^\ast$ (gray area) two or three times earlier than curves 30, 31, and 32 (Fig.~\ref{fig:rescrazy}). With these results, and seeing that $\rho$ values between 0.001 and 0.01 never harmed the performance of the algorithm, we chose $\rho=0.002$.
\begin{figure}[t]
\centering
\includegraphics{eeg_crazy.pdf}
\caption{Effect of $\rho$ on the \textsc{EEG} time series. Equivalent results were observed with the other time series.}
\label{fig:rescrazy}
\end{figure}
\subsection{Final Performance}
\label{sec:resfinal}
After extending SWARMMOTIF with velocity clamping and craziness, we assess its performance on all considered time series using the default parameter combination resulting from the previous two sections. As can be seen, the obtained motif dissimilarities are far from the random sampling in all cases (Fig.~\ref{fig:resfinal}; notice the logarithmic axes). In addition, we see that SWARMMOTIF is able to already obtain dissimilarities within $\mathcal{M}^\ast$ as soon as its execution begins. Specifically, motif dissimilarities in $\mathcal{M}$ start to overlap the ones in $\mathcal{M}^\ast$ at $t<10$\,s for practically all time series. The only exceptions are \textsc{EOG} and \textsc{RandomWalk}, where $\mathcal{M}$ starts to overlap with $\mathcal{M}^\ast$ at $t<100$\,s. We hypothesize that taking a smaller number of particles $\kappa$ could make $\mathcal{M}$ overlap with $\mathcal{M}^\ast$ earlier, but leave the formal assessment of this hypothesis for future work.
\begin{figure*}[t]
\centering
\includegraphics{dowjones_final.pdf} \,\,
\includegraphics{dodgers_final.pdf} \,\,
\includegraphics{insect_final.pdf} \\ \vspace*{0.3cm}
\includegraphics{eeg_final.pdf} \,\,
\includegraphics{fieldrecording_final.pdf} \,\,
\includegraphics{wind_final.pdf} \\ \vspace*{0.3cm}
\includegraphics{power_final.pdf} \,\,
\includegraphics{eog_final.pdf} \,\,
\includegraphics{rw_final.pdf}
\caption{SWARMMOTIF performance on the considered time series: (A) \textsc{DowJones}, (B) \textsc{CarCount}, (C) \textsc{Insect}, (D) \textsc{EEG}, (E), \textsc{FieldRecording}, (F) \textsc{Wind}, (G) \textsc{Power}, (H) \textsc{EOG}, and (I) \textsc{RandomWalk}.}
\label{fig:resfinal}
\end{figure*}
Finally, as execution time $t$ progresses, we see that SWARMMOTIF consistently retrieves lower dissimilarities, up to the point that $\mathcal{M}\simeq\mathcal{M}^\ast$ (Fig.~\ref{fig:resfinal} and Table~\ref{tab:result}). Following the condition we specify in Sec.~\ref{sec:eval}, this means that the distances in the motif set obtained by SWARMMOTIF are not statistically worse than the ones of the true exact motif set. With respect to MOEN's execution time, this happens 1487 (\textsc{DowJones}), 184 (\textsc{CarCount}), 50 (\textsc{Insect}), 179 (\textsc{EEG}), 100 (\textsc{FieldRecording}), 241 (\textsc{Wind}), 286 (\textsc{Power}), 74 (\textsc{EOG}), and 287 (\textsc{RandomWalk}) times faster. This implies between one and three orders of magnitude speedups (more than that for \textsc{DowJones}). Overall, we believe this is an extremely competitive performance for an anytime motif discovery algorithm.
\begin{table*}
\centering
\caption{Comparison between MOEN and SWARMMOTIF results. The latter are taken at the time when the defined stopping criterion is met (Sec.~\ref{sec:eval}). Results for other time steps are available online (Sec.~\ref{sec:intro}).}
\label{tab:result}
\resizebox{1\linewidth}{!}{
\begin{tabular}{ll|ccccccccc}
\hline\hline
Approach & Result & \textsc{DowJones} & \textsc{Dodgers} & \textsc{Insect} & \textsc{EEG} & \textsc{FieldRecording} & \textsc{Wind} & \textsc{Power} & \textsc{EOG} & \textsc{RandomWalk} \\
\hlin
MOEN & $t$ [s] & 2933.9 & 838.0 & 9051.1 & 21596.0 & 76415.0 & 70727.2 & 672161.9 & 430790.3 & 430676.6 \\
& $D(\textbf{m}_i)^\ast$ median & 0.0147 & 0.0260 & 0.0140 & 0.0119 & 0.0301 & 0.0117 & 0.0199 & 0.0206 & 0.0091 \\
& $D(\textbf{m}_i)^\ast$ min & 0.0084 & 0.0225 & 0.0102 & 0.0081 & 0.0240 & 0.0102 & 0.0102 & 0.0137 & 0.0080 \\
& $D(\textbf{m}_i)^\ast$ max & 0.0191 & 0.0279 & 0.0168 & 0.0131 & 0.0347 & 0.0124 & 0.0232 & 0.0218 & 0.0094 \\
\hline
SWARMMOTIF & $t$ [s] & 2.0 & 4.6 & 180.0 & 120.6 & 764.8 & 293.1 & 2346.5 & 5807.7 & 1497.3 \\
& $D(\textbf{m}_i)$ median & 0.0132 & 0.0259 & 0.0135 & 0.0119 & 0.0315 & 0.0117 & 0.0214 & 0.0195 & 0.0090 \\
& $D(\textbf{m}_i)$ 5\textsuperscript{th} percentile & 0.0087 & 0.0234 & 0.0121 & 0.0090 & 0.0239 & 0.0102 & 0.0142 & 0.0163 & 0.0079 \\
& $D(\textbf{m}_i)$ 95\textsuperscript{th} percentile & 0.0182 & 0.0278 & 0.0162 & 0.0130 & 0.0341 & 0.0122 & 0.0230 & 0.0212 & 0.0092 \\
\hline\hline
\end{tabular}
}
\end{table*}
\section{Conclusion}
\label{sec:conclusion}
In this article, we propose an innovative standpoint to the task of time series motif discovery by formulating it as an anytime multimodal optimization problem. After a concise but comprehensive literature review, we reason out the new formulation and the development of an approach based on evolutionary computation. We then highlight the several advantages of this new formulation, many of which relate to a high degree of flexibility of the solutions that come from it. To the best of our knowledge, such a degree of flexibility is unseen in previous works on time series motif discovery.
We next present SWARMMOTIF, an anytime multimodal optimization algorithm for time series motif discovery based on particle swarms. We show that SWARMMOTIF is extremely competitive when compared to the best approach we could find in the literature. It obtains motifs of comparable similarities, in considerably less time, and with minimum memory requirements. This is confirmed with 9~independent real-world time series of increasing length coming from a variety of domains. Besides, we find that the high majority of the possible implementation choices lead to non-significant performance changes in all considered time series. Thus, given this robustness, we can think about the proposed solution as being parameter-free from the user's perspective. Overall, if we add the aforementioned, unprecedented degree of flexibility, SWARMMOTIF stands out as one of the most prominent choices for motif discovery in long time series streams. Since the used data and code are available online (Sec.~\ref{sec:intro}), the research presented here is fully reproducible, and SWARMMOTIF is freely available to researchers and practitioners.
We believe that the consideration of multimodal optimization algorithms is a relevant direction for future research in the field of time series analysis and mining. Not only with regard to motif discovery, but also in other tasks such as querying for segments of unknown length~\cite{Kahveci04TKDE} or determining optimal alignments and similarities~\cite{Serra14KBS}. With regard to the latter, we envision powerful approaches to variable-length local similarity calculations, in the vein of existing dynamic programming approaches~\cite{Smith81JMB,Serra09NJP}. Finally, we believe that considering the search spaces and the time constraints derived from time series problems can be a challenge for the evolutionary computation community. We look forward to exploring all these topics in forthcoming works, together with other multimodal optimization techniques that could be easily mapped to the problem of time series motif discovery.
\section*{Acknowledgment}
We would like to thank all the people who contributed the data sets used in this study and Abdullah Mueen for additionally sharing his code. We would also like to thank Xavier Anguera for useful discussions that motivated the present work. This research has been funded by 2009-SGR-1434 from Generalitat de Catalunya, JAEDOC069/2010 from Consejo Superior de Investigaciones Cient\'ificas (JS), TIN2012-38450-C03-03 from the Spanish Government, and E.U.\ Social and FEDER funds (JS).
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
Analysis of the flux variability of active galactic nuclei (AGNs) allows constraints to be put on the associated emission mechanisms and jet parameters of the AGN jet. Flux variability in AGNs is seen across the entire electromagnetic spectrum with the fastest variability seen in the higher energy regions, in particular the keV to TeV regions. Of all AGNs, the strongest and fastest variability is often seen in a class of AGNs known as blazars, which are AGNs that have a jet pointed in a direction that is very close to our line of sight. Analysis of the flux variability of blazars in the high-energy region of the electromagnetic spectrum is a useful method for limiting the size of the emission region in an AGN, thereby limiting the emission mechanisms.
Due to light propagation time and geometry effects, fast variability can limit the size of the emission region R of a blazar by the inequality $R<(cT\delta)/(1+z)$ where T is the variability timescale of the flux, $\delta$ is the Doppler factor, and z is cosmological redshift. The Doppler factor $\delta$ is defined as $\delta^{-1}=\gamma(1-\beta cos\theta)$, where $\gamma$ is the Lorentz factor in the jet, $\beta$ is the ratio of the jet velocity to the speed of light in vacuum, and $\theta$ is the angle of the jet with respect to the observer. The variability timescale, T, is typically described by the flux doubling time, which will be referred to as T$_{double}$ or the flux halving time, which will be referred to as T$_{1/2}$.
There has been repeated evidence of fast flux variability in the TeV $\gamma$-ray region with doubling or halving times as short as 224$\pm$60 s from PKS 2155-304 (Aharonian et al., 2007), 15 minutes from Mrk 421 (Gaidos et al., 1996) and $\sim$ 9 min from BL Lac (Arlen et al., 2012). Results from simultaneous TeV and X-ray studies probing variability on the order of hours have shown evidence that X-rays and TeV $\gamma$-rays may originate from the same region (e.g. Maraschi et al. 1999). This correlation at larger timescales, in addition to observed fast variability of TeV $\gamma$-rays, has motivated the search for fast variability in blazars at X-ray wavelengths.
In general, X-ray observations have not revealed variability with doubling timescales less than approximately 15 min. There has been isolated evidence for much faster variability in X-rays from a single blazar, H0323+022 (Feigelson et al., 1986), with a halving time less than 30 seconds and a quoted statistical significance of P $<10^{-13}$. However, this fast variability in X-rays was a single, isolated event that has not been seen again. The authors did consider systematic and/or instrumental effects, but none could be identified.
With the launch of the \textit{Swift} X-Ray Telescope (Burrows et al., 2005), X-ray data in the 0.2-10 keV region has been obtained on many AGNs on timescales ranging from seconds to over 8 years.
We report on our search for fast variability in blazar flares. We have searched our large \textit{Swift} database for blazar flares with doubling or halving times less than 15 minutes. The data taken from the \textit{Swift} database for analysis are outlined in Section \ref{sec:data}. The analysis techniques, including the necessary constraints set on the data in the search for fast blazar variability, is outlined in Section \ref{sec:analysis}. We provide the results of the analysis in Section \ref{sec:results} with discussion of the results in Section \ref{sec:discussion}.
\section{Data}
\label{sec:data}
These data were obtained with the X-Ray Telescope (XRT) on the \textit{Swift} observatory (Burrows et al. 2005), and the reduced light curves are all available on a public web site and database described by Stroh \& Falcone (2013). The \textit{Swift}-XRT is one of three instruments onboard the \textit{Swift} Gamma Ray Burst Explorer (Gehrels et al., 2004). In particular, the \textit{Swift}-XRT is an X-ray imaging spectrometer with the ability to obtain data for lightcurves with high timing resolution. The Swift-XRT data in this study are from two different modes, window timing mode (WT) and photon counting mode (PC). All data are screened to ensure that no individual observations, and therefore no potential flares, include a switch from one mode to the other. See Burrows et al., 2005 for an in depth discussion of the XRT modes.
The data are comprised of 7.43\e{6} seconds of AGN observations from 12544 continuous observations of the 143 AGN observed at the time of analysis. These data were obtained between 2004 December 17 and 2014 May 28. The data were processed using the most up to date version of \textit{Swift} tools at the time of analysis: \textit{Swift} Software version 4.1 and FTOOLS version 6.14 (Blackburn, 1995) with observations processed using xrtpipeline version 0.12.8. For WT mode data, we additionally filtered out the first 150 seconds of data after a \textit{Swift} telescope slew since the telescope pointing was not always settled well enough to allow accurate standard WT mode data analysis. Greater detail on how these data were obtained and processed is available in Stroh \& Falcone (2013).
The number of observations for each object range from as many as 1192 observations in the case of Mrk 421 to as little as 1 observation in the case of PKS 0244-470. The average continuous observation time is about 592 s, with a maximum observation time of 2565 s and a minimum observation time of 21 s.
\afterpage{%
\begin{figure*}[ht!]
\begin{center}
\subfigure{%
\label{fig:1st}
\includegraphics[width=0.3\textwidth]{0954+658_00035381001_7_new.eps}
}%
\subfigure{%
\label{fig:2nd}
\includegraphics[width=0.3\textwidth]{1ES2344+514_00035031050_2_new.eps}
}%
\subfigure{%
\label{fig:3rd}
\includegraphics[width=0.3\textwidth]{3C279_00035019131_8_new.eps}
} \\
\subfigure{%
\label{fig:4th}
\includegraphics[width=0.3\textwidth]{3C279_00053516001_1_new.eps}
}
\subfigure{%
\label{fig:5th}
\includegraphics[width=0.3\textwidth]{3C345_00036390005_5_new.eps}
}%
\subfigure{%
\label{fig:6th}
\includegraphics[width=0.3\textwidth]{3C454_3_00031493027_1_new.eps}
} \\
\subfigure{%
\label{fig:7th}
\includegraphics[width=0.3\textwidth]{PKS0235+164_00030880008_6_new.eps}
}
\subfigure{%
\label{fig:8th}
\includegraphics[width=0.3\textwidth]{PKS0537-441_00030138004_6_new.eps}
}%
\subfigure{%
\label{fig:9th}
\includegraphics[width=0.3\textwidth]{PKS0537-441_00050150003_2_new.eps}
}\\%
\subfigure{%
\label{fig:10th}
\includegraphics[width=0.3\textwidth]{PKS1424+240_00041539012_1_new.eps}
}
\subfigure{%
\label{fig:11th}
\includegraphics[width=0.3\textwidth]{PKS1510-089_00031173069_1_new.eps}
}
\subfigure{%
\label{fig:12th}
\includegraphics[width=0.3\textwidth]{RXJ0324_6+3410_00036533033_5_new.eps}
}\\%
\subfigure{%
\label{fig:13th}
\includegraphics[width=0.3\textwidth]{S50716+714_00035009047_13_new.eps}
}
\subfigure{%
\label{fig:14th}
\includegraphics[width=0.3\textwidth]{S50716+714_00035009048_5_new.eps}
}%
\subfigure{%
\label{fig:15th}
\includegraphics[width=0.3\textwidth]{WCom_00035018040_2_new.eps}
}
\end{center}
\caption{%
The light curves of the 15 observations flagged for a potential quick time X-ray flare are shown. The observation ID specified in the title of each plot is the observation ID assigned by \textit{Swift}.
}%
\label{fig:plots}
\end{figure*}
\clearpage
}
\section{Analysis}
\label{sec:analysis}
The goal of the analysis is to determine if fast blazar variability with a rate doubling or halving time less than 15 minutes exists within these data.
The data were analyzed on an observation by observation basis to test for this fast variability. Various constraints were needed to: (1) define the data set of observations that would have enough counts and a long enough duration to provide the potential to find significant flaring, and (2) define the observations within the aforementioned data set that would have variability, and in particular, the potential for flux doubling or halving. The necessary constraints to define our data set include: a maximum count rate of at least 0.1 cts/s, at least three data points in each observation, and a length of observation of at least 5 minutes. The constraints that define a possible flare include: restricting the maximum count rate to be at least 1.8 times greater than the minimum count rate during the observation, 3 consecutive data points either increasing or decreasing in count rate, and a reduced $\chi^2$ value from a straight line fit greater than 3.0. In addition to these pre-defined constraints, manual inspection was required for the observations that passed the constraints to ensure that flux variations from systematic effects (e.g. mode switching) were not flagged as possible flares. The necessity of each of these constraints is discussed in more detail below.
\subsection{Constraints That Define Data Set}
\label{subsec:definedataset}
\subsubsection{Max Count Rate}
A maximum count rate of at least 0.1 cts/s during an observation is necessary to allow for a statistically significant flare to occur in the timescales being probed. Taking into account light curve error bars at this count rate and applying $\chi^2$ statistics, a statistically significant flare cannot be detected at rates lower than 0.1 cts/s over the timescales of concern. Therefore, if the maximum count rate of an observation is not at least 0.1 cts/s then it would be impossible to detect a fast flare from that observation, thus we eliminate such observations from the data sample.
\subsubsection{Data Point Requirement}
The requirement of at least three data points within an observation as a means of defining our data set is closely related to the discussion in Section \ref{subsub:consecutive}. We cannot include observations that have less than three data points as part of our data set because they do not have the ability to have three increasing or decreasing data points and therefore would not possibly be able to be flagged as a possible flare.
\subsubsection{Orbit Length}
We also require observations to be at least 5 minutes in length. This is to allow enough time for a significant flare to occur within a typical observation. Our data set also has max observation time of 2565 seconds, which implicitly sets an upper limit to the length of the doubling or halving time observed of a potential flare, though no explicit cut was set
\subsection{Constraints That Define Flaring}
\label{subsec:defineflare}
\subsubsection{Doubling/Halving Factor}
We require the maximum count rate to be 1.8 times greater than the minimum count rate to ensure that any potential flares at least approach a flux doubling or halving since the purpose of the analysis is to search for AGN flares with fast doubling or halving timescales.
\subsubsection{Consecutive Points}
\label{subsub:consecutive}
We require at least 3 consecutive data points to be either increasing or decreasing in count rate to minimize random noise in the count rate and guarantee a possible steady increase or decrease in the flare.
\subsubsection{Reduced $\chi^2$ Value}
\label{subsub:chi2}
Finally, a reduced $\chi^2$ value from a straight line fit of the weighted mean of the count rate greater than 3.0 is necessary to ensure sufficient variability in the count rate of the source. This constraint ensures that the selected observation is not consistent with a quiescent state of the AGN. This reduced $\chi^2$ value was also used in determining the p-value of the null hypothesis of a straight line fit.
\begin{figure}[ht!]
\begin{center}
\subfigure{%
\label{fig:afirst}
\includegraphics[width=0.4\textwidth]{Mrk421_00030352205_1_new.eps}
} \\
\subfigure{%
\label{fig:asecond}
\includegraphics[width=0.4\textwidth]{1156+295_00036381002_4_new.eps}
}
\end{center}
\caption{%
The light curve from Mrk 421 (top) shows a sharp break in the light curve that results in spurious data points as the result of the \textit{Swift}-XRT switching observing modes from PC mode to WT mode. The light curve of 1156+295 (bottom) shows a single spurious data point that was potentially a result of fluctuations at the beginning of an observation that could not be considered a potential quick time flare.
}%
\label{fig:spurious}
\end{figure}
\subsection{Manual Analysis}
\label{subsec:manual}
Manual inspection was required to ensure that all observations flagged as possible flares did not have any issues that could not be flagged systematically. One of these issues includes the \textit{Swift}-XRT mode switching from PC to WT or vice versa, which often caused an abrupt break in the light curve and spurious points to occur in the data as shown in light curve on the top of Figure \ref{fig:spurious}. This issue would occasionally cause observations to be flagged as a possible flare and would have to be discarded manually. Other observations that were processed out through manual inspection had a single spurious data point responsible for the identification of the observation as a potential flare, with the single point occurring either at the beginning or end of an observation with no flaring evidence throughout the rest of the light curve. These observations (e.g. Figure \ref{fig:spurious}, bottom) also had to be discarded from the data set.
\section{Results}
\label{sec:results}
Of the initial 12544 AGN observations in the large data set, 8606 AGNs observations had the possibility of having a detectable quick time flare based on the count-rate constraints and the a priori assumptions that define our data set as discussed in Section \ref{subsec:definedataset}. Of the remaining 8606 AGN observations, 31 observations passed the constraints that defined a possible flare as discussed in Section \ref{subsec:defineflare}. Once the manual analysis was done, 21 observations remained that did not have any mode switching issues or were reliant on a single spurious data point, as discussed in Section \ref{subsec:manual}. Of the remaining 21 observations, six were flagged as being clearly due to random fluctuations, without any clearly discernible flare, and were therefore not considered to be potential flares, and one observation from PKS 0537-441 showed signs of two flaring episodes within a single observation. Therefore, 16 potential flares from 15 observations and 12 different AGN remained. Details from these 16 potential flares are shown in Table \ref{table:orbits}, and each light curve is shown in the mosaic of Figure \ref{fig:plots}.
We used two methods to estimate the significance of the quick time soft X-ray flares; one method compares the light curve data to an assumption that the X-ray rate was constant with a value equal to the mean rate during the observation, and a second method compares the light curve data to an assumption that the X-ray rate was constant with a value equal to a best guess at the quiescent source rate near the time of the potential flare. In the first method, the chance probability p-value was determined based on the null hypothesis probability of a straight line fit of the weighted mean and using the reduced $\chi^2$ value that was found as discussed in Section \ref{subsub:chi2}. In the second method the chance probability p-value was determined based on the null hypothesis probability of the light curve data being described by a constant, straight-line fit to the minimum rate surrounding each potential flare, which was used as a best-guess estimate of the "quiescent" state of emission surrounding the potential flare. We computed the reduced $\chi^2$ value of the data during the potential flare relative to this "quiescent" constant rate light curve and then determined the p-value. These values are shown in Table \ref{table:orbits} under "Weighted Mean p-value" and "Quiescent p-value" respectively.
Once the p-value was determined using the two different methods, we had to take into account a trials factor based on the number of observations in the data set that had the potential to provide a detection of a rapid soft X-ray flare. The number for the trials factor was determined by counting the number of observations that passed the set constraints as discussed in Section \ref{subsec:defineflare} and Section \ref{subsec:manual}, as well as the visual inspection described above. Therefore, there were 16 trials that were taken into account, and the resultant p-values for each method are shown in Table \ref{table:orbits} in the "Post-trials" column next to each respective method.
\begin{figure}[ht!]
\begin{center}
\subfigure{%
\label{fig:bfirst}
\includegraphics[width=0.4\textwidth]{aaa_flare_check_RXJ0324_6+3410_00036533033_orbit5.eps}
} \\
\subfigure{%
\label{fig:bsecond}
\includegraphics[width=0.4\textwidth]{aaa_flare_check_1ES2344+514_00035031050_orbit2.eps}
}
\end{center}
\caption{\label{fig:double}These plots show the selected section of the observations from RX J0324.6+3410 on 2013 Jan 15 (top) and 1ES 2344+514 on 2012 January 3 (bottom) that were fit with either increasing or decreasing lines to estimate the doubling or halving time.}
\label{fig:RXJ0324}
\end{figure}
In Table \ref{table:orbits} we also include the reduced $\chi^2$ values that were found on the section of each observation that was deemed part of the potential quick time flare as discussed earlier for the weighted mean and quiescent straight line fits. We also include the doubling times and halving times when applicable to each flagged observation. These times were determined by fitting a straight line to the sections of each observation deemed to be part of a possible flare and calculating how much time would be required for half of the maximum count rate to double or for the maximum count rate to halve when appropriate. An example of this fitting can be seen in Figure \ref{fig:double}. We also include the date of observation of each potential flare as well as the start time of the observation containing the possible flare as a reference to compare the data to the light curves available on the \textit{Swift}-XRT monitoring site (Stroh \& Falcone, 2013).
\section{Discussion}
\label{sec:discussion}
Analysis of our large \textit{Swift}-XRT database indicates that our data set does not show strongly significant evidence for soft X-ray flaring on the timescales probed, i.e. less than 15 minutes. Of the low-significance potential flares that were found and outlined in Table \ref{table:orbits} and Figure \ref{fig:plots}, we find potential doubling times in 6 different AGNs between 2.2 min and 5.7 min as well as halving times in 10 different AGNs between 2.2 min and 9.1 min. All of the potential flares in Table \ref{table:orbits} come from the blazar class (FSRQs or BL Lacs), which are more likely to exhibit rapid flaring due to their aligned relativistic jets. However, none of these potential flares represents a truly significant detection after accounting for trials, and therefore, no strong statements can be made about the existence, or nonexistence, of short timescale flares in the data set.
The potential flare from the RX J0324.6+3410 observation is shown in more detail in Figure \ref{fig:RXJ0324} (top). This potential flare has a doubling time, calculated as discussed in Section \ref{sec:results}, of 5.7 minutes. If the variability is truly from flaring, the potential emission region size can be calculated by applying the inequality $R<(cT\delta)/(1+z)$, which was discussed in Section \ref{sec:intro}. The doubling time variabilitiy as well as z=0.061 (Linford et al. 2012) equate to an upper limit on the emission region size of 9.7\e{12}$\delta$ cm, where $\delta$ is the unknown doppler factor that was discussed in Section \ref{sec:intro}.
If the potential flares outlined in Table \ref{table:orbits} were indeed quick time soft X-ray flares, these would be the fastest doubling and halving times for X-ray AGN flares ever observed, with the exception of the single, isolated event observed by Feigelson et al. (1986). As stated previously, variability in AGNs at timescales approximately this short had previously only been seen repeatedly with statistical significance in the TeV region (Aharonian et al., 2007; Gaidos et al., 1996, Albert et al., 2007).
Simultaneous multiwavelength studies of blazars in TeV $\gamma$-rays and X-rays are vital to determine if very fast X-ray variability is coincident with very fast TeV variaiblity as has been seen previously on longer timescales (Maraschi et al., 1999). Further X-ray observations of these sources with more sensitive instruments for longer continuous integrations to determine flare timescales will allow us to perform better searches for fast X-ray variability in the future. This will enable more complete studies of the connection between X-ray and TeV emission regions in blazar jets. \\
\begin{sidewaystable*}[!ht]
{%
\centering
\caption{The 16 potential quick time soft X-ray flares flagged for more analysis.\label{table:orbits}}
\begin{tabular}{c c c c c c c c c c c}
\hline\hline
Source & Date & Obs. Time & \multicolumn{1}{p{1.5cm}}{\centering Weighted \\ Mean \\ $\chi^2_{red}$} & \multicolumn{1}{p{1.5cm}}{\centering Quiescent \\ $\chi^2_{red}$} & T$_{double}$ & T$_{1/2}$ & \multicolumn{1}{p{1.5cm}}{\centering Weighted \\ Mean \\ p-value} & Post-Trials & \multicolumn{1}{p{1.5cm}}{\centering Quiescent \\ p-value} & Post-Trials \\ [0.3ex]
\hline
\\
TXS 0954+658 & 04-Jul-06 & 53920.442 & 3.23 & 4.38 & - & 9.1 & 3.96\e{-2} & 6.34\e{-1} & 1.25\e{-2} & 1.99\e{-1} \\%7
1ES 2344+514 & 03-Jan-10 & 55199.419 & 4.26 & 6.70 & - & 2.8 & 5.13\e{-3} & 8.20\e{-2} & 1.61\e{-4} & 2.58\e{-3} \\%2
3C 279 & 31-Dec-13 & 56657.480 & 4.45 & 11.96& 3.4 & - & 3.95\e{-3} & 6.32\e{-2} & 7.91\e{-8} & 1.27\e{-6} \\%8
3C 279 & 10-Jul-06 & 53926.208 & 3.05 & 4.69 & - & 3.9 & 2.73\e{-2} & 4.36\e{-1} & 2.81\e{-3} & 4.49\e{-2} \\%1
3C 345 & 23-Aug-09 & 55066.809 & 3.19 & 4.29 & - & 7.0 & 4.10\e{-2} & 6.56\e{-1} & 1.37\e{-2} & 2.19\e{-1} \\%5
3C 454.3 & 02-Oct-13 & 56567.783 & 3.63 & 11.52& 4.2 & - & 2.66\e{-2} & 4.26\e{-1} & 9.90\e{-6} & 1.58\e{-4} \\%1
PKS 0235+164 & 06-Feb-07 & 54137.292 & 5.95 & 7.79 & 2.5 & - & 2.60\e{-3} & 4.16\e{-2} & 4.13\e{-4} & 6.61\e{-3} \\%6
PKS 0537-441, 1 & 08-Oct-08 & 54747.866 & 4.37 & 6.10 & - & 3.6 & 1.27\e{-2} & 2.03\e{-1} & 2.25\e{-3} & 3.60\e{-2} \\%6
PKS 0537-441, 2 & 08-Oct-08 & 54747.869 & 5.05 & 6.76 & 3.2 & - & 6.41\e{-3} & 1.02\e{-1} & 1.16\e{-3} & 1.85\e{-2} \\%6
PKS 0537-441 & 26-Jan-05 & 53396.605 & 5.68 & 8.08 & 2.2 & - & 3.40\e{-3} & 5.44\e{-2} & 3.11\e{-4} & 4.97\e{-3} \\%2
PKS 1424+240 & 13-Apr-13 & 56395.315 & 3.50 & 5.45 & - & 4.9 & 1.47\e{-2} & 2.35\e{-1} & 9.68\e{-4} & 1.55\e{-2} \\%1
PKS 1510-089 & 28-Apr-11 & 55679.860 & 5.94 & 6.96 & - & 5.8 & 2.63\e{-3} & 4.21\e{-2} & 9.46\e{-4} & 1.51\e{-2} \\%1
RX J0324.6+3410 & 15-Jan-13 & 56307.859 & 3.91 & 9.46 & 5.7 & - & 3.57\e{-3} & 5.71\e{-2} & 1.21\e{-7} & 1.93\e{-6} \\%5
S5 0716+714 & 14-Dec-09 & 55179.704 & 4.05 & 6.62 & - & 2.2 & 1.74\e{-2} & 2.78\e{-1} & 1.33\e{-3} & 2.13\e{-2} \\%13
S5 0716+714 & 15-Dec-09 & 55180.299 & 5.51 & 10.98& - & 3.2 & 4.05\e{-3} & 6.48\e{-2} & 1.71\e{-5} & 2.73\e{-4} \\%5
W Com & 09-Jun-10 & 55356.738 & 3.78 & 5.36 & - & 8.1 & 2.28\e{-2} & 3.65\e{-1} & 4.68\e{-3} & 7.49\e{-2} \\%2
[1ex]
\hline
\end{tabular}
}
\* Notes.---Col. (1): Object name referred to on \textit{Swift}-XRT monitoring site. Col. (2): Date of observation. Col. (3): Start time of observation in MJD. Col. (4): Reduced $\chi^2$ of weighted mean straight line fit. Col. (5): Reduced $\chi^2$ of simulated quiescent line fit. Col. (6): Estimated doubling time of potential flare in minutes. Col. (7): Estimated halving time of potential flare in minutes. Col. (8): Estimated p-value significance of weighted mean straight line fit before trials factor. Col. (9): Estimated p-value significance of weighted mean straight line fit after trials factor. Col. (10): Estimated p-value significance of quiescent line fit before trials factor. Col. (11): Estimated p-value significance of quiescent line fit after trials factor.
\end{sidewaystable*}
|
\section{Introduction}
Upon compressions that are sufficiently rapid to avoid crystallization, a fluid of hard spheres (HS)
first turns sluggish and then forms a glass~\cite{BB11,Ca09}.
This glass can then be further compressed until the system jams~\cite{LN98}, which occurs under the application of an infinite confining pressure~\cite{KL07,PZ10}.
Glass formation is entropic, i.e., particles vibrate and thus cage each other in place, while jamming is mechanical, i.e., no motion is possible and particles are held steady through direct contacts with each other. Over the last decade, this two-transition scenario has been broadly validated,
both numerically and theoretically~\cite{DTS05,SDST06,SLN06,CBS09,HD09,PZ10,CIPZ11,IBS12,IBS13}. Interestingly, recent
advances predict that -- at least in the mean-field, infinite-dimensional ($d\to\io$) limit -- there exists
a third transition, a so-called {\it Gardner transition}, that is intermediate in density and pressure between glass formation and jamming~\cite{KPUZ13,CKPUZ14,Charbonneau2014B,Rainone2014}.
First discovered in spin-glass models~\cite{Ga85,GKS85,MR03,MR04,Ri13}, the Gardner transition corresponds to
a single glass {metabasin}
splitting into a complex hierarchy of marginally stable {sub-basins}.
The transition is thus akin to the spin-glass transition of the Sherrington-Kirkpatrick (SK) model, wherein a critical temperature separates a
paramagnetic phase, in which a single thermodynamic state exists, from a marginal phase, in which a large number of distinct spin-glass states appear~\cite{MPV87}.
In structural glasses, however, the high-temperature phase corresponds to a given glass {metabasin}
that has been dynamically selected by a quenching protocol; it is this {metabasin} that
then undergoes a spin-glass-like transition~\footnote{
The Gardner transition is akin to that of the SK model in presence of a random magnetic field, because the system is confined in a given glass
metabasin and the self-induced disorder characteristic of this glass metabasin acts as a self-induced
``external'' random field. At the mean-field level, this distinction does not make any
difference, but the presence of a random field is important in finite dimensions~\cite{UB14}.
}.
The discovery of a
Gardner transition in glasses has already markedly advanced our theoretical understanding of jamming
by providing analytical predictions for the critical {jamming} exponents~\cite{CKPUZ14,LDW13,Charbonneau2014B,CCPZ15,MW15}.
It further suggests an explanation for the abundance of soft vibrational modes in glasses~\cite{WNW05,MW15},
for the peculiar behavior of the specific heat in quantum glasses~\cite{LDW13,KPUZ13},
and various other transport and thermodynamic properties in this regime.
Before these fascinating problems can be {tackled, however,}
a crucial question is whether the Gardner transition itself,
whose existence is well established in the $d\rightarrow\io$ limit, exists in finite (low) dimensions.
Renormalization group results indicate
that the transition might disappear or dramatically change of nature in low $d$~\cite{UB14}.
Yet a similar line of inquiry has been pursued for decades in the context of spin glasses~\cite{BY86},
leading to the conclusion that, whatever the ultimate fate of the phase transition in the thermodynamic limit may be,
the $d\to\io$ scenario provides a very
good description of the system over the relevant experimental length and time scales~\cite{Janus}.
At present, the {most direct}
way to assess the relevance of the Gardner transition for the description of experimental glasses is through
numerical simulations.
It is therefore important to {first} identify the observable consequences of this transition {in well-controlled model systems}.
This study primarily aims to develop procedures and to identify observables in order to reliably detect the Gardner transition.
To that effect, we consider a simple structural glass former, the infinite-range Mari-Kurchan (MK) model~\cite{MK11,mari:09}.
The model is quite abstract and in some ways far from realistic models of glasses, but
{\it (i)} it is a mean-field model by construction;
{\it (ii)} it shares, in any finite dimension $d$, the same qualitative phase diagram as infinite-dimensional hard spheres, provided one neglects {the effect of hopping on the glassy dynamics}~\cite{Charbonneau2014};
{\it (iii)} it can be studied analytically in great detail, using the methods of~\cite{Rainone2014,Charbonneau2014}, that we further developed
for this work;
{\it (iv)} {and,} most importantly, it can be easily simulated in any finite dimensions $d$, {including $d=3$ as we do here.}
{Discerning
the signatures of the Gardner transition in this well-controlled setting,
where we are certain that the transition exists, shall later on enable us and others to study
more realistic glass models for which the existence of the transition is not a priori guaranteed.}
{The plan of this paper is as follows. In Sect.~\ref{sec:II} we describe the MK model and its glassy behavior, and in Sec.~\ref{sec:numerics} we detail the numerical procedures we use for the study. In Sec.~\ref{sec:GardnerDetection} we discuss several quantities that bear the signature of the transition, as suggested by the analogy with spin glasses.}
For instance, at both the spin-glass and the Gardner transitions, the ``spin-glass susceptibility'' diverges and the distribution of overlaps (distances) between different replicas becomes non-trivial. {Bringing the various estimates of the transition together in Sec.~\ref{sec:summary} reveals that}
the Gardner transition can be reliably and reproducibly located through numerical methods in the MK model, and {that} the results {are} fully consistent
with theoretical expectations.
{We conclude in Sec.~\ref{sec:conclusion} with a description of}
other possible measurements to detect and characterize the Gardner transition, which may be more appropriate for numerical simulations of more realistic model glass formers as well as for experiments.
{Before embarking on this program, let us make a note of warning to the reader.}
Because the aim of work is
to identify numerical methods to detect the Gardner
transition, we {have attempted}
to make the numerical part as self-contained as possible.
For what concerns the theoretical part, {however, we have chosen to be more succinct and have instead}
relied on previous work both on spin glasses and structural glasses, which is {here} only briefly recalled. We {expect} that the reader unfamiliar
with spin-glass theory will {nonetheless} be able to read and understand the numerical part without much difficulty.
\section{Model and basic physical picture}
\label{sec:II}
We consider a simple glass-former, the infinite-range Mari-Kurchan model~\cite{MK11,mari:09} -- initially proposed by Kraichnan~\cite{Kr62} --
in which $N$ hard spheres of diameter $\sigma$ interact through a pair potential {that} is a function of distance
shifted by a quenched random vector $\mathbf{\L}_{ij}$. The total interaction energy is thus
\begin{equation}
U = \sum_{i<j}^{N} u\left(|\mathbf{r}_{ij}|\right),
\end{equation}
where for particles at positions $\{\mathbf{r}_i\}$ the shifted distance $\mathbf{r}_{ij}$ is defined as $\mathbf{r}_{ij} =\mathbf{r}_i - \mathbf{r}_j + \mathbf{\L}_{ij}$,
and $u(r )$ is the HS potential, i.e., $e^{-u(r)} = \th(r -\s)$ with $r = |\mathbf{r}|$. The random {shifts}, which are uniformly distributed over the system volume $V$,
induce a quenched disorder that suppresses both crystallization and nucleation between metastable glassy states~\cite{MK11,mari:09,Charbonneau2014}.
The model further enables {\it planting}, which is a
simple process for generating equilibrated liquid configurations at all densities~\cite{KZ09,Charbonneau2014} (see Sec.~\ref{sec:planting}).
In all spatial dimensions, the MK model has a mean-field structure \emph{by construction},
due to the infinite-range random shifts~\cite{MK11},
and exhibits a jamming transition in the same universality class as standard HS \cite{mari:09}.
The MK model is also fully equivalent to standard HS in the limit $d \to \infty$, where both models can be solved
exactly using the mean-field methods {described in} Refs.~\cite{PZ10,MK11,KPUZ13,CKPUZ14,Charbonneau2014B,Rainone2014,Rainone2666}.
In the rest of this section, we briefly describe the phase diagram in the infinite-dimensional limit (Fig.~\ref{fig:theory}),
we explain how it can be applied to the MK model in $d=3$, and we discuss
some of the finite-dimensional corrections that have thus far been considered~\cite{Charbonneau2014}.
\begin{figure}[t]
\centerline{\hbox{\includegraphics [width = \columnwidth] {dinfstatefo_new.eps}}}
\caption{(Color online).
Phase diagram of the HS and MK models in the limit $d\to\io$.
Results are partially derived from Ref.~\cite{Rainone2014}, complemented by new results obtained
for the Gardner phase{~\cite{Rainone2666}}.
The liquid EOS given by Eq.~\eqref{eq:liquid_eos} (full black line)
gives $p/d = (2^d \f/d)/2$ in the limit $d\to\io$. The black dot denotes the dynamical glass transition at $2^d\varphi_{\rm d}/d=4.8$. Glass EOSs from state following (SF) for $2^d \varphi_{\rm 0}/d = 5, 6, 6.667, 7, 8$ are also reported (thin colored lines, from left to right). The envelope for the Gardner transition for each SF (colored dot) is given by the dashed line.
}
\label{fig:theory}
\end{figure}
\subsection{Equilibrium states (liquid phase)}
The liquid phase of the MK model ergodically samples equilibrium configurations following the Gibbs distribution and
has a remarkably simple structure. Its pair correlation function is given by
\begin{equation}
\begin{split}
g_2(\mathbf{r}) & \equiv
\frac{V}{N(N-1)} \overline{\left\langle \sum_{i \neq j} \d( \mathbf{r}_{ij}- \mathbf{r} ) \right\rangle} \\
& = e^{-\beta u(r )} =
\theta (r - \sigma) ,
\end{split}
\label{eq:g2}
\end{equation}
where $\beta = 1/T$ is the inverse temperature, $\langle \cdots \rangle$ denotes thermal averaging, and $\overline{\cdots}$ denotes averaging over
quenched disorder, i.e., over {$\mathbf{\L}_{ij}$}. The second virial coefficient is
\begin{equation}
B_2 =- \frac{1}{2V} \iint f(\mathbf{r}_{12}) d \mathbf{r}_1 d\mathbf{r}_2 = \frac{V_d \sigma^d}{2},
\end{equation}
where the Mayer function $f(\mathbf{r}) = e^{-\beta u(r)} - 1$, and $V_d$ is the volume of a $d$-dimensional ball of unit radius.
Because no indirect correlations exist, higher-order correlation functions can be factorized in a trivial way and the corresponding virial coefficients are zero. For example, the three-body correlation function
\begin{equation}
\begin{split}
g_3(\mathbf{r}, \mathbf{r}') & \equiv
\frac{V^2}{N(N-1)(N-2)} \overline{\left\langle \sum_{i \neq j \neq k} \d( \mathbf{r}_{ij} - \mathbf{r} ) \d( \mathbf{r}_{ik} - \mathbf{r}' ) \right\rangle} \\
& = g_2(\mathbf{r}) g_2(\mathbf{r}'),
\end{split}
\end{equation}
and the third virial coefficient
\begin{equation}
\begin{split}
B_3 & = -\frac{1}{3V}\iiint f(\mathbf{r}_{12})f(\mathbf{r}_{13})f(\mathbf{r}_{23}) d\mathbf{r}_1 d\mathbf{r}_2 d\mathbf{r}_3 \\
&= -\frac{1}{3V}\iiint f(\mathbf{r}_{12}) f(\mathbf{r}_{13}) \times \\
& \times f(\mathbf{r}_{13} - \mathbf{r}_{12} + \mathbf{\L}_{12} + \mathbf{\L}_{23} - \mathbf{\L}_{13}) d\mathbf{r}_1 d\mathbf{r}_2 d\mathbf{r}_3 \\
& = 0.
\end{split}
\end{equation}
Note that if $|\mathbf{r}_{12}|, |\mathbf{r}_{13}| < \s$, then
$f(\mathbf{r}_{13} - \mathbf{r}_{12} + \mathbf{\L}_{12} + \mathbf{\L}_{23} - \mathbf{\L}_{13})=0$ in the thermodynamic limit because random shifts are uncorrelated and typically of the system size, {and thus} $|\mathbf{r}_{13} - \mathbf{r}_{12} + \mathbf{\L}_{12} + \mathbf{\L}_{23} - \mathbf{\L}_{13}| \gg \sigma$. It is straightforward to generalize this argument to show that all higher-order virial coefficients are also zero~\cite{MK11}.
Because only the second virial coefficient is non-zero, the reduced pressure $p$ equation of state (EOS) for the liquid {is}
\begin{equation}
p \equiv \beta P/\rho = 1 + B_2 \rho = 1 + 2^{d-1} \ph,
\label{eq:liquid_eos}
\end{equation}
where the combination of inverse temperature $\beta $, pressure $P$, and number density $\rho=N/V$ gives a unitless quantity $p$ whose only dependence is on the liquid volume fraction $\ph=\rho V_d (\sigma/2)^d$.
\subsection{Dynamical glass transition}
Although the structure and thermodynamics of the liquid are trivial, its dynamics is not.
We will focus here on the equilibrium dynamics, i.e. {\it starting from an equilibrium initial condition}~\footnote{
If one considers instead the dynamics starting from an out-of-equilibrium initial condition, then the system is able to equilibrate
for $\f < \varphi_{\rm d}$ while {\it aging} effects, which persist for infinite times, are observed for $\f > \varphi_{\rm d}$~\cite{CK93,CC05}.
}.
In infinite dimension, a dynamical glass transition $\varphi_{\rm d}$ separates two distinct dynamical regimes.
For $\f < \varphi_{\rm d}$, the dynamics is diffusive at long times, as expected of any liquid.
Upon approaching $\varphi_{\rm d}$, however,
the dynamics grows increasingly sluggish, and
above $\varphi_{\rm d}$, each particle is fully confined within a cage formed by its neighbors. The typical size of that cage is the cage order parameter $\Delta_1$
(the meaning of the suffix will become clear below),
which, in that regime, can be extracted from the long-time limit of the mean-squared displacement (MSD)
\begin{equation}\label{eq:delta}
\begin{split}
\Delta (t)&= \frac{1}{N} \sum _{i=1}^{N} \av{|{\vec r}_i(t) - {\vec r}_i(0)|^2} \ , \\
\Delta_1 &= \lim_{t\to\io , \ \f > \varphi_{\rm d}} \Delta(t) \ .
\end{split}
\end{equation}
From the $d\to\io$ solution, we know that the equilibrium distribution of the order parameter, $P_{\rm eq}(\Delta)$, has two peaks for $\f>\varphi_{\rm d}$ (see Fig.~\ref{fig:fractal})~\cite{FP95,CC05}. The first characterizes the distance between two glass configurations within a same {metabasin}. It is centered around $\Delta_1$, which is the typical size of this basin. The second characterizes the inter-basin distance. It is centered around $\Delta_0 = \infty$, because states that belong to different metabasins are completely uncorrelated.
In technical terms, this situation is described by a 1-step {replica symmetry breaking} (1RSB) scheme~\cite{MPV87}.
Note, however, that the peak at $${\bf D\,}$_1$ has an exponentially small weight in $N$, because there exists an exponentially large number of distinct glass
states~\cite{CC05}. Hence, in the thermodynamic limit, $P_{\rm eq}(\Delta) = \d(${\bf D\,}$ - ${\bf D\,}$_0)$ everywhere in the liquid phase, i.e. even for $\f > \varphi_{\rm d}$.
\begin{figure}[t]
\centerline{\hbox{\includegraphics [width = 0.85\columnwidth] {fractal.pdf}}}
\caption{(Color online). (a) Organization of glass states (blue dots) for $\varphi_{\rm 0}>\varphi_{\rm d}$; the typical intra-basin MSD
is $\Delta_1$, while the inter-basin MSD is $\Delta_0 = \infty$.
(b) In the SF-fullRSB phase, $\varphi_{\rm G} < \f < \varphi_{\rm J}$, glass {metabasins subdivide} into a hierarchical structure of sub-basins with typical innermost MSD
$\Delta_{\rm EA}$ and outermost {(meta)basin} MSD $\Delta_1$. {Schematics of the equilibrium $P_{\rm eq}(\Delta)$ and restricted equilibrium $P_{\rm SF}(\Delta)$ (blue area) distributions are given for (c) the SF-RS (or 1RSB) and (d) the fullRSB phases.}}
\label{fig:fractal}
\end{figure}
\subsection{Glass state following and the Gardner transition}
\label{sec:gardner}
In $d\to\io$, each equilibrium configuration at density $\varphi_{\rm 0} > \varphi_{\rm d}$ is {forever} trapped into one of {the} exponentially many glass metabasins.
The pressure of an equilibrium configuration at $\varphi_{\rm 0}$ is given by
the liquid EOS, Eq.~\eqref{eq:liquid_eos}, but if one compresses (or decompresses) such a configuration up (or down) to a density
$\f$, the system remains
within the metabasin that was initially selected. {The system} thus falls out of equilibrium in the sense that it
cannot visit the ensemble of all possible distinct glass metabasins and thus follows an EOS different from {that of} the liquid.
{The infinite lifetime of these $d\rightarrow\infty$ glass states implies that}
one can
adiabatically follow the EOS of a single glass state through a {\it restricted equilibrium} approach,
a construction also {known as}
state
following (SF)~\cite{Franz1995,barrat1997,ZK10,Rainone2014}.
SF consists in using a first configuration {that is equilibrated}
at initial density $\varphi_{\rm 0}$, and a second configuration that
is in a restricted equilibrium at a different density $\f$~\cite{Franz1995,barrat1997,ZK10,Rainone2014}. {The restriction is that that second configuration must be part of the same metabasin as the first one.}
The state of the second configuration {thus} describes the evolution of the
glass {metabasin with $\f$}.
Dynamically, this {process} corresponds to preparing a system at equilibrium at initial density $\varphi_{\rm 0}$,
then compressing it at density $\f$, and {\it assuming that it is able to equilibrate inside the glass {metabasin, but} without escaping it}. {In simpler words,}
structural relaxations are frozen
and particles {only} rattle inside their cages.
Upon compression, these cages shrink until the jamming density, $\varphi_{\rm J}(\varphi_{\rm 0})$, is reached.
Particles are then mechanically in contact, which makes the system mechanically rigid.
In the following, we call the restricted equilibrium EOS of a given glass metabasin simply its glass EOS.
Examples of {different} $d\to\io$ glass EOS obtained by SF are given in Fig.~\ref{fig:theory}~\cite{Rainone2014,Rainone2666}.
In $d\to\io$, a compressed state under restricted equilibrium undergoes a Gardner transition at $\varphi_{\rm G}(\varphi_{\rm 0})$,
at which point the glass metabasin {subdivides} into a hierarchy of sub-basins.
Technically,
before the Gardner transition, i.e., for $\f < \varphi_{\rm G}(\varphi_{\rm 0})$, the glass {metabasin} is {obtained} by a replica symmetric SF (SF-RS) computation~\footnote{A replica symmetric SF computation is {quite} similar to a standard 1RSB computation.}, while
in the Gardner phase, i.e., for $\f > \varphi_{\rm G}(\varphi_{\rm 0})$,
a full replica symmetry breaking SF (SF-fullRSB) {computation} is needed~\cite{Ga85,GKS85,MPV87,CKPUZ14,Charbonneau2014B,Rainone2014}.
The Gardner transition is therefore akin to the {spin glass transition} one
of the SK model, but {restricted}
to a {single} glass metabasin.
The high temperature phase of the SK model {thus} corresponds to a simple glass metabasin for $\f < \varphi_{\rm G}(\varphi_{\rm 0})$ while the low temperature phase
corresponds to a fractured and marginal metabasin for $\f > \varphi_{\rm G}(\varphi_{\rm 0})$.
{Note that we here complement} the SF-RS computation of Ref.~\cite{Rainone2014} with a SF-fullRSB computation,
{in order} to obtain
the complete EOS reported in Fig.~\ref{fig:theory}. The SF-fullRSB equations are similar to the ones reported in Ref.~\cite{Charbonneau2014B}, {but the full}
details {of this work are} given elsewhere~\cite{Rainone2666}.
Following a {glass} state in restricted equilibrium gives
for $\f < \varphi_{\rm G}(\varphi_{\rm 0})$ a distribution $P_{\rm SF}(\Delta)$ that has a single peak at $\Delta_1$.
Note that because the system is confined to a single glass basin, the peak at $${\bf D\,}$_0$ is absent.
At $\varphi_{\rm G}(\varphi_{\rm 0})$ the glass basin fractures in a SF-fullRSB structure, and correspondingly
the single peak in the distribution $P_{\rm SF}(\Delta)$ fractures into two peaks centered around
$\Delta_{\rm EA}$ and $\Delta_1$, connected by a wide continuous band (see Fig.~\ref{fig:fractal} and~\cite{MPV87}).
Here, $\Delta_{\rm EA}$ is the typical size of the innermost sub-basins at the lowest hierarchical level,
and $\Delta_{\rm 1}$ is the typical {distance between}
the outermost {sub-basins, i.e., the size of the metabasin}.
For the same reason as above, $P_{\rm SF}(\Delta)$ does not show the $${\bf D\,}$_0$ peak.
Because the cage order parameter changes continuously at $\varphi_{\rm G}$, the Gardner transition is a continuous critical transition~\cite{Ga85,GKS85}.
The Gardner phase is also marginally stable, in the sense that a zero mode is always present in the stability matrix of the free energy~\cite{MPV87}.
Because jamming is located within the Gardner phase, its marginal stability and critical scaling behaviors can consequently be obtained from a
fullRSB thermodynamic calculation~\cite{CKPUZ14,Charbonneau2014B}.
\subsection{Timescales}
\label{sec:timescales}
The timescales that characterize the dynamics in the different phases of the MK model are predicted based on the general correspondence between
statics and dynamics in spin glasses~\cite{CK93,CK94}. Here {again, we only consider the system behavior in the} limit $d\to\io$.
{Beyond} the microscopic timescale $\t_0$, over which dynamics is essentially ballistic, one gets the following picture.
\begin{itemize}
\item
In the liquid phase below $\f_{\rm d}$, dynamics is characterized by two timescales: the $\b$-relaxation timescale $\t_\b$, over which particles
explore their {transient cage}, and a longer timescale $\t_\a$, over which dynamics is diffusive.
Both timescales are finite for $\f < \f_{\rm d}$, and diverge at $\f_{\rm d}$ according to the scaling
predicted by mode-coupling theory (MCT)~\cite{Go09}. {We do not discuss this regime further because it is not directly related to the Gardner transition}.
\item
In the liquid phase above $\f_{\rm d}$ and {in the SF-RS (simple glass) phase}, the same two timescales exist:
$\t_\a \sim \exp(N)$ is the timescale for jumping from one glass {metabasin} to another, which {being infinite in the thermodynamic limit properly defines the glass metabasins,} and
$\t_\b$ is the timescale for equilibrating {within} a glass {metabasin, i.e., for particles to explore their cage,} which is finite for $\f > \varphi_{\rm d}$
{but diverges}
upon approaching $\varphi_{\rm d}$ from densities above it, again following a MCT-like scaling form~\cite{Go09}.
\item
Upon approaching the Gardner transition, $\t_{\b}$
again diverges, {scaling as} $\t_{\b} \sim (\varphi_{\rm G} - \f)^\g$.
{It does so for the exact} same reason as {at near the spin-glass transition of} the SK model
{and corresponds to the} critical slowing down close to a second-order
phase transition.
\item
In the Gardner phase, {dynamics is described by three timescales}:
$\t_\b$ is the timescale for equilibrating {within} a single glass sub-basin ($${\bf D\,}$ < ${\bf D\,}$_{\rm EA}$),
{$\t_{\rm meta}$ is the timescale over which system explores the structure of the sub-basins within a given glass metabasin
($${\bf D\,}$_{\rm EA} < ${\bf D\,}$ < ${\bf D\,}$_1$), and $\t_\a \sim \exp(N)$ remains the timescale for jumping from one glass metabasin to another ($${\bf D\,}$ \sim ${\bf D\,}$_0$)}. To be more {clear}, none of these processes correspond to a simple exponential with a single timescale.
{Everywhere in the Gardner phase $\t_\b=\infty$, because of the phase's marginality.
The relaxation inside a single sub-basin is thus expected to scale as a power-law in time.}
The exploration of sub-basins is
characterized
by a complex distribution of free energy barriers and relaxation times, and {thus} $\t_{\rm meta} \sim \exp(N^\a)$ with $\a<1$
({$\a=1/3$} is expected in the SK model~\cite{MPV87}). Note that
barriers between sub-basins are much lower than {those} between metabasins, hence $\t_{\rm meta} \ll \t_\a$.
\end{itemize}
Let us now situate {the scheme for} our work in this complex $d\rightarrow\infty$ dynamic phase diagram. We {are conveniently} {\it not} concerned with the scalings
of the MCT regime.
Thanks to the planting procedure (see Sec.~\ref{sec:planting}), we are {indeed} able to {easily} generate
{\it equilibrium} configurations {of the MK model for arbitrary $\varphi_{\rm 0}$, including $\varphi_{\rm 0} > \varphi_{\rm d}$}. These configurations are well equilibrated in a glass {metabasins} and therefore
have an infinite $\t_\a$ (in the thermodynamic limit) and a finite $\t_\b$. {Because $\t_\a=\infty$ in this density regime,} we can simply forget
about its existence and consider that we are forever restricted into {a given} glass {metabasin}. From this point of view, we are in
a situation similar to that of a spin glass
where one is able to start at equilibrium in the paramagnetic, high temperature phase.
If we compress the system slowly enough to a final density $\f \in (\varphi_{\rm 0}, \varphi_{\rm G})$, {then equilibration within
the glass {metabasin} is possible, because $\t_\b$ remains finite. In this situation, the system {behavior} is described by the SF-RS {computation}.
To study the Gardner transition, however, we need to compress the system up to $\varphi_{\rm G}$, at which $\t_\b$ diverges.
Hence, upon approaching $\varphi_{\rm G}$, $\t_\b$ eventually becomes larger than the simulation timescale}, and equilibration (even in the restricted SF
sense) becomes impossible. The system {thus} falls out of (restricted) equilibrium and is not described anymore by the SF computation.
For $\f > \varphi_{\rm G}$ the situation is even worse {both} as $\t_\b$ and $\t_{\rm meta}$ are infinite. The SF-fullRSB computation, which gives
the restricted equilibrium properties in this regime, is only an approximation even at long (but not divergent with {$\exp(N)$}) times. {The situation is akin to that}
in the SK model where the fullRSB computation is only
an approximation of the states reached dynamically at long times in the spin-glass phase.
{Because} in this regime the planting technique does not work (it only works on the liquid line in Fig.~\ref{fig:theory}), we cannot use it
to study the restricted equilibrium in the Gardner phase.
In the following we will {therefore} present two kinds of data:
\begin{itemize}
\item for $\varphi_{\rm 0} \leq \f < \varphi_{\rm G}$, far enough from $\varphi_{\rm G}$ such that $\t_\b$ is smaller than the simulation timescale, we
{obtain} restricted
equilibrium data
\item for $\f \sim \varphi_{\rm G}$ and $\f > \varphi_{\rm G}$, the system is out of (restricted) equilibrium, shows {\it aging} effects, and, at long times, {we obtain states
that are qualitatively similar to the SF-fullRSB ones, but not exactly equal to them}.
\end{itemize}
In Sec.~\ref{sec:numerics} we give a more precise definition of the numerical protocol we use to study these different regimes.
Note that, according to the above discussion, from now on we will refer to the restricted equilibrium simply as ``equilibrium", given that
we always work inside a glass {metabasin}.
\subsection{Finite-dimensional MK model}
\label{sec:finiteD}
To conclude this section, we discuss the additional effects that appear when one considers the MK model in finite dimensions, and that affect
the above discussion.
From a theoretical point of view, a finite $d$ affects quantitatively the phase diagram, but to a lowest degree of approximation
one can simply
take the phase diagram of Fig.~\ref{fig:theory} {and fix the value of $d$ (e.g. $d=3$) to obtain result for the EOS}. However, there are
several systematic corrections that impact the accuracy of this result.
\begin{itemize}
\item
The infinite-dimensional results are obtained within a Gaussian structure of the cage,
which provides the exact results in the limit $d\to\io$~\cite{KPZ12}. The Gaussian equations
are slightly different in finite $d$~\cite{PZ10}. The corrections have the form of a series in $1/d$ and
are quite small. For example, in $d=3$ one has {$2^3 \varphi_{\rm d}/3 \approx 4.74$} instead of the infinite-dimensional
result $2^d \varphi_{\rm d}/d \approx 4.8$. These corrections are negligible and could be easily taken into account if needed.
\item
A more important problem is {the inexactitude} of the Gaussian assumption for the MK model in finite dimensions. One should {instead}
optimize the free energy over a generic cage function, {but this computation is technically quite difficult}. In Ref.~\cite{Charbonneau2014}, {however,} two different {cage functions
were studied and found to give similar results}. The corrections coming from the non-Gaussianity of the cage
are {thus} also rather small, {although they are} more difficult to estimate.
\item
In finite $d$, cages are heterogeneous~\cite{Charbonneau2014}. {Carefully taking this effect into account} would require a cavity calculation
similar to the one performed in Ref.~\cite{Charbonneau2014}, which is beyond the scope of this work, {but is also likely a small correction for $\varphi_{\rm 0}>\varphi_{\rm d}$~\cite{Charbonneau2014}}.
\end{itemize}
For all these reasons, the results of Fig.~\ref{fig:theory} are only an approximation to the finite-dimensional phase diagram.
In the following, {we will nonetheless take the results of the simplest theoretical approximation}, namely take the $d\to\io$ results
of Fig.~\ref{fig:theory} and use them in $d=3$, {after properly rescaling the axes}.
We expect (and check) that the corrections are quite small. Sometimes, {however, in order to take effective account of} these corrections,
we will {further} rescale the {results, so as to} obtain a better {quantitative} agreement between theory and simulation.
The most important genuinely finite-dimensional effect is {\it hopping}.
As discussed in Ref.~\cite{Charbonneau2014}, slightly above $\varphi_{\rm d}$, particles are not perfectly confined in their cages,
as one would expect based on the $d\to\io$ picture. Instead, each particle is allowed to explore a network of cages,
connected by narrow pathways. Hopping consists in particles jumping between distinct cages. The corresponding time
scale is finite at $\varphi_{\rm d}$ and thus the $d\to\io$ divergence of the relaxation time is washed away in finite $d$.
Hopping effects also change deeply the dynamics of the system with respect to the $d\to\io$ prediction, and
the configurations at
$\varphi_{\rm 0} \gtrsim \varphi_{\rm d}$ are not able to constrain the dynamics for infinitely-long times. {Because the timescale for hopping increases quickly upon increasing both density and dimension, considering values of $\varphi_{\rm 0}$ that are slightly above $\varphi_{\rm d}$ suffices}.
In practice, {based on the analysis of} Ref.~\cite{Charbonneau2014}, in $d=3$ and
for $\varphi_{\rm 0} \geq 2.5$ hopping is strongly suppressed, hence in that regime the mean-field, {$d\rightarrow\infty$} scenario should apply {reasonably} well.
\section{Numerical Approach}
\label{sec:numerics}
In this section, we provide the numerical details used in the simulations of the glass states of the MK model.
\subsection{Planting}
\label{sec:planting}
An important algorithmic advantage of the MK model is that planting can be used to generate equilibrium liquid configurations at any $\varphi_{\rm 0}$~\cite{KZ09,Charbonneau2014}. This procedure sidesteps the tedious and time-consuming work of first preparing dense equilibrium configurations, as would be needed for typical glass formers, such as HS. The basic idea is to switch the order in which initial particle positions $\{\mathbf{r}_i\}$ and random shifts $\{\mathbf{\Lambda}_{ij}\}$ are determined.
One first chooses the particle positions $\{\mathbf{r}_i\}$ independently and uniformly in the volume $V$, and then for each particle pair one chooses
a random shift $\mathbf{\Lambda}_{ij}$ uniformly under the sole constraint that the two particles should not overlap, which is quite straightforward to satisfy.
As long as the quenched and the annealed averages of the free energy are the same
(see Ref.~\cite{KZ09} for a more detailed discussion),
a planted state is a true equilibrium state and
automatically satisfies the liquid EOS, Eq.~\eqref{eq:liquid_eos}. This condition is met along the replica symmetric phase for $\varphi_{\rm 0} < \f_{\rm K}$, where $\f_{\rm K}$ is the Kauzmann point at which the configurational entropy vanishes~\cite{PZ10,MK11}. Because in the MK model $\f_{\rm K} = \infty$~\cite{MK11}, planting a liquid configuration is thus possible at any density, which dramatically reduces the computational cost of the initial equilibration.
In our notation, a given
$\{\mathbf{r}_i\}$ and $\{\boldsymbol{\Lambda}_{ij}\}$ defines a {\em sample}.
A sample thus identifies a given system (defined by
$\{\boldsymbol{\Lambda}_{ij}\}$) and, for this system, one of its glass {metabasins}
(selected by $\{\mathbf{r}_i\}$).
\subsection{Molecular dynamics (MD) simulations}
We use event-driven molecular dynamics (MD) to
simulate MK particles in $d=3$~\cite{SDST06,Charbonneau2014}. Periodic boundary conditions with the minimum
image convention are implemented on the shifted distances $|\mathbf{r}_i - \mathbf{r}_j +
\boldsymbol{\Lambda}_{ij}|$. Time $t$ is expressed in units of $\sqrt{\beta m
\sigma^2}$, where the particle mass $m$ and diameter $\sigma$ as well as the
inverse temperature $\beta$ are set to unity. Systems consist of $N= 800$
particles unless otherwise specified. This system size is large enough to
contain a first full shell of neighbors around each particle, and to keep the
periodic boundary effects on caging to a minimum~\cite{Charbonneau2014}.
Finite-size effects are studied for some of the observables for the initial
liquid density $\varphi_{\rm 0}= 2.5$ (see Sec.~\ref{sec:static}).
To simulate SF, for a given sample, we start from the planted equilibrium configuration at a
packing fraction $\varphi_{\rm 0}$, and grow the spheres following the
Lubachevsky-Stillinger algorithm~\cite{DTS05,SDST06} at constant growth rate
$\gamma = 0.001$, unless otherwise specified, up to a desired $\f$.
Once
compression is stopped at the target density, the origin of time is set. {We will thus typically (although not always) define
the waiting time $t_w$, as the time that has elapsed since the end of the compression}.
From that moment on
we start measuring observables, keeping
density and temperature (and thus energy) constant.
Note that because $\gamma$ is finite and rather small, part of the equilibration happens already during compression,
so provided we are not too close to $\varphi_{\rm G}$ the system is stationary at all $t_w$ (we come back on this point later).
This procedure
is repeated over $N_\mathrm{s}$ samples in order to average over thermal and
quenched disorders. Errors are computed using the jack-knife method~\cite{amit2005field}. Depending on the statistical convergence of the different observables, $N_{\mathrm{s}}$ is varied from 500 to 75,000, as specified in the discussion of the various measurements.
\subsection{Observables}
The pressure evolution along SF is reasonably well described by a free-volume EOS
\begin{equation}
\frac{1}{p} = C \left( 1 - \frac{\ph}{\varphi_{\rm J}}\right),
\label{eq:free_volume}
\end{equation}
where $C$ is a fitting parameter. Fitting Eq.~\eqref{eq:free_volume} to the compression results provides an estimate of $\varphi_{\rm J}$ (see Table~\ref{tab:phiG}).
If a sufficiently small $\gamma$ is chosen, no aging is observed in the pressure, and
using slower compression rates gives only negligible corrections to the glass EOS (see Fig.~\ref{fig:state_following}).
Interestingly, upon decompression, the state follows the same EOS up to a threshold density at which it melts into a liquid phase. This phenomenon has been recently predicted by the theory~\cite{MPR14,Rainone2014}, and observed numerically in simulated ultrastable glasses~\cite{Singh2013, Hocky2014}.
To obtain more structural information about the free energy landscape, we also simulate a {\it cloning} procedure. The approach consists of taking two exact copies (clones) $A$ and $B$ of the same planted configuration at $\varphi_{\rm 0}$, and assigning them different initial velocities, randomly drawn from the Maxwell--Boltzmann distribution. These two copies are then independently compressed up to $\f$, before measuring the mean-squared distance between them
\begin{equation}\label{eq:deltaAB}
\Delta_{AB} (t)= \frac{1}{N} \sum _{i=1}^{N} \av{|{\vec r}_i^A(t) - {\vec r}_i^B(t)|^2},
\end{equation}
where ${\vec r}_i^A(t)$ and ${\vec r}_i^B(t)$ are the positions of particle $i$ at time $t$ in clones $A$ and $B$, respectively.
Although the two clones start from the same initial configuration, {their compression histories are different} once $\f$ is reached.
The detailed behavior of the clones will be discussed in Sec.~\ref{sec:time},
but let us explain here briefly why the cloning procedure is useful to detect
the Gardner transition.
In the SF-RS phase,
the two clones are uncorrelated in the glass basin and $\Delta_{AB}(t)$ converges quickly
(on a time scale $t \sim \t_\b$ if the two clones are
not sufficiently well equilibrated along the compression)
to the equilibrium value $\Delta_{AB} = ${\bf D\,}$_1$. Hence $\Delta_{AB} = \lim_{t\to\io} ${\bf D\,}$(t)$ in the SF-RS phase.
{By contrast,} if the end point of the compression falls {within} the {SF-fullRSB} (or Gardner) phase, clones
most likely fall into different sub-basins. Their mean square distance can then be described by a non-trivial
time-dependent probability distribution, $P_{AB}(t,${\bf D\,}$)$,
that depends on the way sub-basins are sampled. Calculating these weights is difficult, because the two clones are generally out of equilibrium.
Because the probability that the two clones fall in the same state is very small, {however,} and therefore
$\Delta_{AB}(t) > \Delta_{\rm EA}$ for all $t$. {Hence, in the Gardner
phase $\Delta_{AB}(t)$ at all $t$ is strictly larger than the long-time limit of the MSD.}
The long time limit of $\Delta_{AB} - ${\bf D\,}$$, {being} zero in the SF-RS phase {and} non-zero in the SF-fullRSB phase,
thus provides an {\it order parameter} for the Gardner transition.
\begin{figure}[t]
\centerline{\hbox{\includegraphics [width = \columnwidth] {state_following.pdf}}}
\caption{(Color online). (a)
Compressions ($\gamma>0$) and decompressions ($\gamma<0$) of an initial equilibrium state at $\varphi_{\rm 0} = 2.5$ and 4.0. The results are averaged over $N_{\mathrm{s}}=100$ samples. The theoretical curves uses data from Ref.~\cite{Rainone2014} and from this work~\cite{Rainone2666}.
(b)~Compression data are indistinguishable for $\gamma \leq 0.01$.
}
\label{fig:state_following}
\end{figure}
\section{Detecting the Gardner transition}
\label{sec:GardnerDetection}
In this section, we describe different means of detecting the Gardner transition through numerical simulations designed to follow the evolution
of glassy states. Let us stress once again that the aim of this work is to understand, in a controlled setting, how well numerical simulations and experiments
can detect the Gardner
transition. {We can} therefore test different strategies to understand their advantages and limitations, and use the analytical results to assess the quality
of the numerical results.
Based on the discussion in Sec.~\ref{sec:II},
we follow two complementary approaches. The first is based on dynamics. From the long-time dependence of the mean square displacement,
we determine {$\t_\b$},
whose divergence in the glass {metabasin} signals the Gardner transition.
The second is based on the properties of the distribution function $P_{AB}(${\bf D\,}$)$, which becomes non-trivial in the Gardner phase. We investigate
different moments of this distribution to {detect} signatures of the transition.
\subsection{Dynamics}
\label{sec:time}
As discussed in Sec.~\ref{sec:timescales}, the Gardner transition is a second order phase transition
associated with a diverging characteristic relaxation time $\t_\b$ that controls the dynamics of the MSD.
To detect this relaxation time we make use of the following observables already briefly discussed
in Sec.~\ref{sec:numerics}. Recall that the origin of time $t=t_w=0$ is set at the end of the compression when
a given final density $\f$ {is} reached.
\subsubsection{Definition of the relevant observables}
\label{sec:IVA1}
We define the MSD between two configurations at
different times:
\begin{equation}
\begin{split}
\widehat\Delta (t,t_w) = \frac{1}{N} \sum _{i=1}^{N} |{\vec r}_i(t + t_w) - {\vec r}_i(t_w)|^2 \ ,
\end{split}
\end{equation}
and the MSD between two different clones
\begin{equation}
\begin{split}
\widehat\Delta_{AB} (t) = \frac{1}{N} \sum _{i=1}^{N} |{\vec r}^A_i(t) - {\vec r}^B_i(t)|^2 \ .
\end{split}
\end{equation}
From these two instantaneous quantities we can define different observables. The statistical
average (over compressions and samples) gives
\begin{equation}
${\bf D\,}$(t,t_w) = \av{
\widehat\Delta (t,t_w)
} \ ,
\hskip20pt
\Delta_{AB}(t) =\av{
\widehat\Delta_{AB}(t)
} \ ,
\end{equation}
{and we define}
\begin{equation}
\d ${\bf D\,}$(t,t_w) = \Delta_{AB}(t + t_w) - ${\bf D\,}$(t,t_w) \ .
\end{equation}
We also define a time-dependent caging susceptibility as the normalized variance of the MSD
\begin{equation}\label{eq:chi1}
\chi(t,t_w) =N\frac{\av{\widehat${\bf D\,}$^2(t,t_w)}-{\av{\widehat${\bf D\,}$(t,t_w)}^2}}{\av{\widehat${\bf D\,}$(t,t_w)}^2},
\end{equation}
and its counterpart of cloned configurations
\begin{equation}\label{eq:chi2}
\chi_{AB}(t) =N\frac{\av{\widehat${\bf D\,}$_{AB}^2(t)}-{\av{\widehat${\bf D\,}$_{AB}(t)}^2}}{\av{\widehat${\bf D\,}$_{AB}(t)}^2}.
\end{equation}
In the SF-RS phase in {restricted} equilibrium, {i.e., at large enough $t_w$ and $t> 0$,}
$${\bf D\,}$(t,t_w) = ${\bf D\,}$(t)$ gives back Eq.~\myref{eq:delta},
while
$\Delta_{AB}(t) = ${\bf D\,}$_1$ does not depend on time.
In this case
\begin{equation}\label{eq:Dasy}
${\bf D\,}$_1 = \Delta_{AB}( \forall t) = \lim_{t\to\io} ${\bf D\,}$(t)
\end{equation}
gives the average cage radius of the glass basin.
Similarly, $\chi(t,t_w) = \chi(t)$ and $\chi_{AB}(t) = \chi$, where
\begin{equation}
\chi = \chi_{AB}(\forall t) = \lim_{t\to\io} \chi(t)
\end{equation}
is the average susceptibility of the glass basin.
Finally, note that
based on Eq.~\eqref{eq:Dasy}, we have
\begin{equation}\label{eq:longdD}
\lim_{t\to\io} \lim_{t_w\to\io} \d${\bf D\,}$(t,t_w) = 0 \ .
\end{equation}
{By contrast,} in the SF-fullRSB phase, equilibrium is not reached even for very large $t_w$ ({recall that we do not consider here
times that are comparable to $\exp(N)$}). Therefore, $\d${\bf D\,}$(t,t_w)$ remains non-zero even for large $t_w$ and $t$, {which makes the long time limit
of $\d${\bf D\,}$(t,t_w)$} a dynamic order parameter for the Gardner transition.
\subsubsection{Qualitative change in caging and susceptibility}
\begin{figure}[t]
\includegraphics [width =\columnwidth] {D_250.pdf}
\caption{(Color online). SF from $\varphi_{\rm 0}= 2.50$ gives
(a) $\Delta_{AB}(t)$, (b) $\Delta(t, t_w = 0)$, and (c) $\delta \Delta(t, t_w = 0)$ at different $\varphi$.
Solid lines are fits to Eq.~\eqref{eq:fit_tau} (bottom panel).
}
\label{fig:delta-t}
\end{figure}
\begin{figure}[t]
\includegraphics [width =\columnwidth] {aging.pdf}
\caption{(Color online). (a) Long-time results for $\Delta_{AB}(t)$.
The data show that, for $\f < \varphi_{\rm G}$, $\Delta_{AB}(t)$ saturates at its asymptotic value (dotted line) in the long-time limit (see, e.g., $\varphi = 2.88, 2.90$). (b) $\Delta(t, t_w)$ and (c) $\delta \Delta(t, t_w)$ are
plotted as a function of $t$ for {various} $\varphi$ and $t_w$.}
\label{fig:delta-ttw}
\end{figure}
Figure~\ref{fig:delta-t} shows the
time dependence of
$\Delta_{AB}(t)$, $${\bf D\,}$(t,t_w=0)$ and $\d${\bf D\,}$(t,t_w=0)$
for $\varphi_{\rm 0}=2.5$, averaged over $N_{\mathrm{s}}=15,000$ to $N_{\mathrm{s}}=500$
from the lowest and highest $\f$, respectively.
Both $\Delta$ and $\Delta_{AB}$ are observed to behave slightly differently below and above $\varphi_{\rm G}$
($\varphi_{\rm G}\approx 3.00$ for $\varphi_{\rm 0}=2.50$, a more precise estimate {is obtained} below).
For $\ph<\varphi_{\rm G}$, $\Delta(t,t_w=0)$
first (and up to a microscopic time $\tau_0 \sim 10^{-1}$) grows quickly because of the ballistic motion of particles~\cite{Charbonneau2014} and then more slowly in the $\b$ relaxation regime, before eventually reaching the plateau $${\bf D\,}$ = ${\bf D\,}$_1$ that defines the cage size. This plateau coincides with the (almost time-independent) results for $\Delta_{AB}(t)$, as is qualitatively expected for a system in a SF-RS phase.
For $\ph \gtrsim \varphi_{\rm G}$, the situation is a bit more convoluted.
As discussed above, beyond the Gardner transition each of the original metabasins is expected to subdivide into a hierarchical
distribution of glassy states.
Equilibration within the glass metabasin is now, {however,} impossible, and we thus observe that $\Delta_{AB}(t)$ depends on time for all observable
times while $${\bf D\,}$(t,t_w=0)$ remains always strictly smaller than $\Delta_{AB}(t)$ and never reaches a plateau.
Correspondingly $\d ${\bf D\,}$(t,t_w=0)$
does not decay to zero.
In Fig.~\ref{fig:delta-ttw}, we {select a few densities and consider} the evolution of $\Delta_{AB}(t)$ over much larger times than in Fig.~\ref{fig:delta-t}
{as well as} the dependence of $\Delta(t,t_w)$ on both $t$ and $t_w$.
{Note that} these data are averaged over {many fewer}
samples, $N_{\rm s} \approx 300$.
{We consider the detailed behavior of these two quantities}:
\begin{itemize}
\item
{In general,} $\Delta_{AB}(t)$ (Fig.~\ref{fig:delta-ttw}a)
should correspond to the average distance between {configurations restricted to} within a {given} metabasin,
but for {$t<\t_\b$}
the basins are sampled with non-equilibrium weights, and
hence $\Delta_{AB}(t)$ {slowly} drifts.
For $\f < \varphi_{\rm G}$, we {indeed observe} that {once} $t \sim \t_\b$,
$\Delta_{AB}(t)$ reaches
a stationary value. Upon approaching $\varphi_{\rm G}$ and for $\f > \varphi_{\rm G}$, {however,} $\t_\b$ is {becomes so large that}
the drift of $\Delta_{AB}(t)$ persists at all {simulated} times.
Note that the drift {can be positive or negative},
depending on density.
\item
The behavior of $${\bf D\,}$(t,t_w)$ (Fig.~\ref{fig:delta-ttw}b) is naturally described as a function of $t$ for fixed $t_w$.
Below $\varphi_{\rm G}$ (e.g., for $\f=2.85$), $${\bf D\,}$(t,t_w)$ is independent of $t_w$ and behaves as in Fig.~\ref{fig:delta-t}.
Beyond $\varphi_{\rm G}$ (e.g., for $\f=3.10$), {however,}
the system is initially
trapped into a sub-basin,
and thus $\Delta(t, t_w)$
grows until reaching a plateau corresponding to the size of the sub-basin, $\Delta_{\rm EA}$, for $t \ll \t_{\rm meta}(t_w)$.
For $t \sim \t_{\rm meta}(t_w)$, the system {can} explore the structure of sub-basins, hence
$\Delta(t, t_w)$ keeps increasing,
and no clear first plateau can be detected.
A second plateau should be reached in the limit $t \gg \t_{\rm meta}(t_w)$, when the metabasin is fully explored,
but this regime is here beyond computational reach.
The timescale $\t_{\rm meta}(t_w)$ {indeed}
increases with $t_w$~\cite{CK93,CK94}, {as} is clearly visible in
Fig.~\ref{fig:delta-ttw}, {where} the drift shifts towards larger times upon increasing $t_w$.
Note that in the limit $N\to\io$, $\t_{\rm meta}(t_w)$ should
diverge for $t_w\to\io$, {but} a finite $N$ acts as a cutoff and $\t_{\rm meta}(t_w)$ should {instead} saturate to a value $\sim \exp(N^{1/3})$ for
large $t_w$.
\item
In the Gardner phase $${\bf D\,}$(t,t_w) < \Delta_{AB}(t)$ for all accessible times $t,t_w$.
Therefore, {although} $\d${\bf D\,}$(t,t_w)$ goes to zero at large times for $\f < \varphi_{\rm G}$, for
$\f \gtrsim \varphi_{\rm G}$ it does not {fully} decay in the accessible $t$ regime (for all $t_w$) and instead converges
to a plateau (Fig.~\ref{fig:delta-ttw}c). Note also that the evolution of
$\d${\bf D\,}$(t,t_w)$ with increasing density is qualitatively identical to {that} of the overlap with decreasing temperature
in spin glasses, as reported in Ref.~\cite[Fig.~8]{Ogielski1985} ({compare also} to Fig.~\ref{fig:delta-t}c).
{These results further support the strong analogy between the Gardner and the spin-glass transitions}.
\end{itemize}
\begin{figure}[t]
\includegraphics [width =\columnwidth] {chi_t.pdf}
\caption{(Color online). (a) Time evolution of the caging susceptibility $\chi(t,t_w=0)$ (scatters) and $\chi_{AB}(t)$ (lines).
(b)~For three selected densities, we report the evolution of $\chi_{AB}(t)$ for much longer times (dotted lines) and the dependence
of $\chi(t,t_w)$ on both $t$ and $t_w$ (see legend).
}
\label{fig:chi-t}
\end{figure}
The dynamical behavior of the caging susceptibility is qualitatively similar {that} of the MSD (Fig.~\ref{fig:chi-t}).
As for the cage order parameters, {in the SF-RS phase (for $\ph< \varphi_{\rm G}$)} the two susceptibilities become identical in the long time limit, i.e., $\chi(t \to \infty,t_w) = \chi_{AB}(t \to \infty)$. By contrast, in the {SF-fullRSB} phase (for $\ph>\varphi_{\rm G}$), on a timescale $t \lesssim \t_{\rm meta}(t_w)$, we generally observe that $\chi(t,t_w)<\chi_{AB}(t)$. Note that the magnitude of the susceptibility increases by more than a decade in the density range considered here, which is a clear signature of the Gardner transition. A detailed analysis of this increase is discussed in Sec.~\ref{sec:susceptibility}.
\subsubsection{Computation of timescales}
Based on the qualitative picture of the dynamics presented above, we now attempt
a more quantitative analysis based
on the {classic work of Ogielski on spin glasses}~\cite{Ogielski1985}
{and its recent extension} to the study of the dynamical transition in the $d=3$ Edwards-Anderson model under
an external field~\cite{janus:campodyn}.
The idea consists in obtaining a relaxation timescale $\tau(t_w)$ from the decay of $\d\Delta(t,t_w)$ (Fig.~\ref{fig:delta-t}c).
Note, however, that this scheme is only well defined in the SF-RS phase, where equilibrium can be reached and
Eq.~\eqref{eq:longdD} holds. In the SF-fullRSB phase, $\d\Delta(t,t_w)$ does not decay to zero and evolves continuously over
a broad range of time scales. Extracting a single timescale is {then not so} straightforward.
We will focus in the following on the data for $t_w=0$, {for which} we have more statistics. {For $\f \lesssim \varphi_{\rm G}$}, these data
are {also} representative of all $t_w$.
In order to facilitate the numerical analysis, we make use of analytical results. According to the general theory of
critical glassy dynamics developed in Refs.~\cite{SZ82,CFLPRR12,PR12}, upon approaching the Gardner point in (restricted) equilibrium
from the SF-RS
phase, hence when $\d${\bf D\,}$(t,t_w) = \d${\bf D\,}$(t)$ does not depend on $t_w$, one has:
\begin{equation}
\d${\bf D\,}$(t) \sim \d\f \FF( t /\t_\b ) \ , \hskip20pt \t_\b \sim \d\f^{-\g} \ ,
\end{equation}
where $\d\f = |\f - \varphi_{\rm G}|$ and the function $\FF(x)$ is such that $\FF(x\ll 1) \sim x^{-a}$ while $\FF(x\gg 1)$ decays exponentially.
Here the exponents $a$ and $\g = 1/a$ are related to the so-called MCT exponent parameter $\l$ by the relation
\begin{equation}
\l = \frac{\G(1-a)^2}{\G(1-2a)} \ .
\end{equation}
The parameter $\l$ can be computed analytically within the replica method by analizing the cubic terms of the replica
action~\cite{CFLPRR12,PR12,KPUZ13}.
{The results of this computation} are reported in Table~\ref{tab:ana}
(details {are} given in Refs.~\cite{Rainone2666}).
In order to estimate $\tau_\b$, we fit the results for $\d${\bf D\,}$(t,t_w=0)$ using an empirical form $\FF(x) \propto x^{-a}e^{-x^b}$ that has been
used for spin glasses~\cite{Ogielski1985,janus:campodyn}, {which gives}
\begin{equation}
\d\Delta(t, t_w=0)=
c\ \frac{\exp{[-(t/\tau'_\b)^b]}}{t^{a}} \ ,
\label{eq:fit_tau}
\end{equation}
where the parameters $a$, $b$ and $\tau'_\b$ depend on $\ph$, $\varphi_{\rm 0}$, {and}
$\tau'_\b$ offers a first estimate of $\tau_\b$.
Note that we fit the exponent $a$ instead of using the analytical result, because the critical regime {over which}
the exponent coincides with $a$ is narrow and away from $\varphi_{\rm G}$ the effective exponent is quite different (see Ref.~\cite[Fig.~12]{Ogielski1985}).
With this choice
all the fits are very good, as reflected by the Pearson
$\chi_{\rm P}^2$ per degree of freedom (d.o.f.)~\footnote{Recall that
$\chi_{\rm P}^2=\sum_{i=1}^{N_T}[y_i-f(t_i)]^2/\sigma_i^2$,
where $N_T$ is the numbers of times $t_i$ and $y_i=\d\Delta(t_i,t_w=0)$, $\sigma_i$ is the error of $y_i$, and $f$ is the fitting
function, Eq.~\eqref{eq:fit_tau}} being much less than 1
(examples of the quality of the fit are given in Fig.~\ref{fig:delta-t}c).
As $\ph$ approaches $\varphi_{\rm G}$ for a given $\varphi_{\rm 0}$, $\tau_\b$ is expected to diverge {as}
\begin{equation}\label{fit:tau}
\tau_\b \sim |\ph - \varphi_{\rm G}^\tau|^{-\gamma}.
\end{equation}
{In order} to obtain a more constrained value of $\varphi_{\rm G}^\t$, we fix the exponent $\gamma$ to its analytic value given in Table~\ref{tab:ana}
and only fit $\varphi_{\rm G}^\t$ and the prefactor.
This time the fits are not excellent, with values of $\chi_{\rm P}^2/\text{d.o.f.}$ of the order of 1 or larger
(see Fig.~\ref{fig:timescale}). The results for $\varphi_{\rm G}^\t$ are reported in Table~\ref{tab:phiG}.
\begin{table}[t]
\begin{tabular}{c@{\hspace{1cm}} c@{\hspace{1cm}}c @{\hspace{1cm}}c}
\hline
$2^d \varphi_0 /d$ & $2^d\varphi_{\rm G}/d$ & $\l$ & $\g = 1/a$\\
\hline
4.8 & 4.8 & 0.7027 & 3.069 \\
4.9 & 5.34 & 0.5607 & 2.651\\
5 & 5.64 & 0.5091 & 2.547\\
5.25 & 6.18 & 0.4378 & 2.427\\
5.5 & 6.61 & 0.3938 & 2.363\\
6 & 7.33 & 0.3398 & 2.295\\
6.667 & 8.16 & 0.2957 & 2.245\\
7 & 8.54 & 0.2801 & 2.228 \\
8 & 9.63 & 0.2469 & 2.194 \\
10.667 & 12.36 & 0.2042 & 2.154 \\
\hline
\end{tabular}
\caption{Analytical results (in the limit $d\to\io$) for $\varphi_{\rm G}$ and the exponents $\l$, $\g$ and
$a$ for several $\varphi_{\rm 0}$~\label{table:lambdaMCT}.}
\label{tab:ana}
\end{table}
\begin{figure}[t]
\includegraphics [width = \columnwidth] {tau.pdf}
\caption{(Color online). (a) Growth of $\tau'_\b$ with $\varphi_\mathrm{G}^\tau-\varphi$ for $\varphi_\mathrm{G}^\tau$ from Table~\ref{tab:phiG}.
(b)~The two estimates of $\tau_\b$, $\tau'_\b$ and $\tau''_\b$, as a function of $\varphi_\mathrm{G}^\tau-\varphi$ for $\varphi_{\rm 0}=2.5$.
(c)~Evolution of $a$ and $b$ with $\varphi$ for $\varphi_{\rm 0}=2.5$.
}
\label{fig:timescale}
\end{figure}
An alternative estimate of $\tau_\b$ can be obtained from the logarithmic
scaling of $\d\Delta(t,t_w=0)$ at long times (see
Fig.~\ref{fig:timescale-alt}). For
$\ph<\varphi_{\rm G}$, the fitting form
\begin{equation}\label{eq:fit_tau2}
\d \Delta(t,t_w=0)= k \left[1-\frac{\log (t)}{\log (\tau''_\b)}\right]
\end{equation}
with a density-dependent constant $k$ gives $\tau''_\b$~\cite{janus:campodyn}.
Comparing $\tau'_\b$ and $\tau''_\b$ suggests that the divergence of the two
timescales is compatible with a same $\varphi_{\rm G}$ and a very
similar power-law exponent (see
Fig.~\ref{fig:timescale}b). The insensibility of the estimator of $\tau_\b$ to its
precise definition adds support to our claim that the observed
divergence is due to a true thermodynamic transition.
Also, we repeated the analysis for $t_w > 0$ with very similar results.
To conclude the discussion, note that the results for $\Delta_{AB}(t)$ and $\Delta(t,t_w)$ at low
densities, i.e., $\varphi_{\rm d}\leq\varphi_{\rm 0} \lesssim 2.2$, may be affected by hopping (Sec.~\ref{sec:finiteD} and~\cite{Charbonneau2014}). In these systems the timescale for leaving a metabasin (albeit only through local hopping processes) is comparable to $\tau_\b$ even near $\varphi_{\rm G}$. The estimate of $\varphi_{\rm G}$ in this regime is therefore subject to a larger error, which explains the bigger difference between $\varphi_\mathrm{G}^{\tau}$
and other $\varphi_{\rm G}$ estimates (Table~\ref{tab:phiG}). For the limit case $\varphi_{\rm 0} = 1.8$, we do not even attempt to fit the data because
no clear power-law regime can be distinguished. By contrast, for $\varphi_{\rm 0} \ge 2.5$, hopping is negligible on the timescales achieved numerically.
\begin{figure}[t]
\includegraphics [width = \columnwidth] {tau_alt.pdf}
\caption{(Color online).
Logarithmic decay of $\d\Delta(t,t_w=0)$ with time for $\varphi_{\rm 0}=2.5$. The results are fitted to Eq.~\eqref{eq:fit_tau2} (dashed lines).}
\label{fig:timescale-alt}
\end{figure}
\subsection{Static functions}
\label{sec:static}
As we have shown in Sec.~\ref{sec:time}, at the Gardner transition two-clones observables such
as $\Delta_{AB}(t)$ become different from the long-time limit of dynamic observables such as $${\bf D\,}$(t,t_w)$.
In this subsection, we thus estimate the location of the Gardner transition using an approach based on the study of
two-clone static observables.
{These observables are} static because,
as we showed in Sec.~\ref{sec:time}, they are time independent for
$\varphi < \varphi_{\rm G}$. {We can thus arbitrarily choose any $t$} to compute them.
Here, we choose $t_{\rm s}=0.2V^{1/3} \sim 2$, such that
$\tau_0 < t_{\rm s} \ll \tau_{\rm meta}$, {which we abbreviate below as}
$\Delta\equiv\Delta(t_{\rm s},t_w=0)$ and
$\Delta_{AB}\equiv\Delta_{AB}(t_{\rm s})$. {We use a similar} notation for all other observables
unless otherwise specified.
\subsubsection{Average mean square displacement}
\label{sec:avmsd}
{We consider the averages of $${\bf D\,}$$ and $\Delta_{AB}$ and compare them with the theoretical SF results (see Fig.~\ref{fig:delta_phi}
for $\varphi_{\rm 0}=2.5$)}.
In order to take into account the corrections discussed in Sec.~\ref{sec:finiteD}, both the numerical and the theoretical
datasets have been rescaled. For the vertical axis, we rescale {the results to $${\bf D\,}$_1(\varphi_{\rm 0})$, which} is known both
in theory and in simulation, {such} that both datasets are equal to 1 for $\f = \varphi_{\rm 0}$.
For the horizontal axis, we rescale $\f$ to $\varphi_{\rm G}$
{with the theoretical $\varphi_{\rm G}$ for the SF-RS curve} (Table~\ref{table:lambdaMCT}) {and a fit factor
$\varphi_{\rm G}^{\rm T}$ for the numerical results} (Table~\ref{tab:phiG}).
Figure~\ref{fig:delta_phi}
clearly shows the difference between the two expected regimes (Sec.~\ref{sec:gardner}):
for $\f < \varphi_{\rm G}$, we {obtain} $${\bf D\,}$ \sim \Delta_{AB} \sim ${\bf D\,}$_1$, while for $\f>\varphi_{\rm G}$, we {obtain} $${\bf D\,}$ \sim \Delta_{\rm EA}$ and $\Delta_{AB} \sim ${\bf D\,}$_1$,
and hence $\Delta_{AB} > ${\bf D\,}$$.
{Although} at short times we observe $\Delta_{AB}(t) \approx ${\bf D\,}$_1$,
on much longer timescales we expect $\Delta_{AB}(t)$ to evolve slowly towards its equilibrium value $\Delta_{AB}(t) = \langle ${\bf D\,}$ \rangle_{\rm SF}$,
which is however only approached for $t$ that diverge with $N$.
\begin{figure}[t]
\includegraphics [width = \columnwidth]{compare_theory.pdf}
\caption{(Color online). Density evolution of the averages of $\Delta$ and $\Delta_{AB}$ for $\varphi_{\rm 0}=2.50$.
Numerical results (points) are compared with theoretical predictions for $${\bf D\,}$_1$, $${\bf D\,}$_{\rm EA}$ and $\langle ${\bf D\,}$ \rangle_{\rm SF}$ (lines).
Both datasets are scaled to reference values (see Sec.~\ref{sec:avmsd} for details) to obtain a better agreement between theory and simulation.
}
\label{fig:delta_phi}
\end{figure}
\subsubsection{Probability distribution functions}
{We next consider the probability distribution function (pdf)}
of the cage order parameters by computing $\Delta$ and $\Delta_{AB}$ for each
sample and constructing the histogram over samples.
Note that because
$\Delta$ is the mean square displacement restricted to a single {sub-basin and the waiting time is fairly short}, its distribution
represents $P_{\rm SF}(\Delta)$ in
the SF-RS phase, {and only}
the peak around $\Delta_{\rm EA}$ in the SF-fullRSB
phase (see Fig~\ref{fig:fractal}).
Figure~\ref{fig:pdf} shows $P(\Delta)$ and $P(\Delta_{AB})$ for $\varphi_{\rm 0}=2.5$ calculated
from
$N_{\mathrm{s}}=40,000$ -- 75,000 samples. The shape of $P(\Delta)$ is Gaussian-like
at all $\ph$, and {its} mean value monotonically
decreases with increasing $\ph$. The shape of $P(\Delta_{AB})$, however, changes
considerably over that same regime. For $\f < \varphi_{\rm G}$, it is Gaussian
and analogous to that of $P(\Delta)$, but
near $\varphi_{\rm G}$ it develops an exponential tail
akin to a Gumbel distribution. If
$\varphi$ is further increased, $P(\Delta_{AB})$ then becomes broader, which is consistent with the
theoretical expectation
(see Fig.~\ref{fig:fractal}).
The development of an exponential tail at the critical point has been
observed and studied for spin glasses in a
field~\cite{Pa12,Janus14}. The effect is thought to be due to disorder.
Whereas the results for most samples fall within Gaussian
fluctuations around a given mean value, a few rare samples have
much larger $\Delta_{AB}$ than the mean, giving an
exponential tail to the distribution. The smaller the system, the stronger the effect (Fig.~\ref{fig:SKpdf}).
These rare fluctuations are hypothesized to originate from the sample-to-sample fluctuations of the critical
point~\cite{Pa12}, which then translates into significant sample-to-sample
fluctuations of some of the measured observables.
We come back to
this point in Sec.~\ref{sec:sample}.
The connection between the changing shape of the distribution and criticality suggests that we can determine the critical transition from $P(\Delta_{AB})$ alone. We propose below two alternative procedures for detecting the Gardner transition using standard moments of the distribution.
\begin{figure}[t]
\begin{center}
\includegraphics [width = \columnwidth] {pdf.pdf}
\end{center}
\caption{(Color online). (a) $P(\Delta)$ and (b) $P(\Delta_{AB})$ at different $\ph$ for $\varphi_{\rm 0}=2.5$.
} \label{fig:pdf}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics [width = \columnwidth] {sk_pdf_N.pdf}
\end{center}
\caption{(Color online). (a) Finite-size behavior of $P(\Delta_{AB})$ at $\ph = 3.0 \approx \varphi_{\rm G}$ and (b) density evolution of the rescaled skewness $\Gamma$ for three different $N$ and for $\varphi_{\rm 0}=2.5$.
}\label{fig:SKpdf}
\end{figure}
\subsubsection{Caging susceptibility}
\label{sec:susceptibility}
We first define a caging susceptibility from the normalized
variance of $P(\Delta_{AB})$
\begin{equation}\label{eq:chi}
\chi=N\frac{\av{\Delta_{AB}^2}-{\av{\Delta_{AB}}^2}}{\av{\Delta_{AB}}^2},
\end{equation}
where the denominator corrects for the fact that $\av{\Delta_{AB}}$ changes with $\ph$. As in the vicinity of any critical point, the susceptibility is expected to diverge as
\begin{equation}\label{eq:divergence}
\chi\propto (\varphi_{\rm G}^\chi-\varphi)^{-1},
\end{equation}
where the critical exponent 1 is due to the fact that the MK model is mean-field in nature.
\begin{figure}[t]
\begin{center}
\includegraphics [width = \columnwidth] {finite-size-chi.pdf}
\end{center}
\caption{(Color online). Finite-size behavior of the two terms of the definition of the
susceptibility in Eq.~\eqref{eq:chi}: (a) $\av{\Delta_{AB}}$, and (b)
$N(\av{\Delta_{AB}^2}-{\av{\Delta_{AB}}^2})$, for $\varphi_{\rm 0}=2.50$. The dashed lines are
fits of a second-order polynomial in
$1/N$ to the data. The $1/N\to0$ extrapolation of these two fits have been used to extract the
${N\to\infty}$ limit of $\chi$, given in Fig.~\ref{fig:chiSK} by a
dark solid line.}\label{fig:finite-size-chi}.
\end{figure}
The definition of $\chi$ involves taking the quotient of two quantities that
both suffer from strong finite-size corrections. In order to control for this effect
we study the behavior of both terms as function of $1/N$ (Fig.~\ref{fig:finite-size-chi}). The denominator, $\av{\Delta_{AB}}$, behaves as a regular
series in $1/N$, decreases smoothly with $\ph$, and eventually saturates above the
Gardner point. The numerator,
$N(\av{\Delta_{AB}^2}-{\av{\Delta_{AB}}^2})$ has, however, a more complex behavior. While it follows a
nearly $1/N$ behavior for $\ph<\varphi_{\rm G}$ with a small dependence on $\ph$, it
grows significantly faster both with $N$ and $\ph$ for $\ph>\varphi_{\rm G}$.
We attempt to
extract the value of both quantities at the thermodynamic limit using a second-order
polynomial fit in $1/N$, even {though} for $\ph > \varphi_{\rm G}$ the estimate of $N(\av{\Delta_{AB}^2}-{\av{\Delta_{AB}}^2})$
obtained in this way is not reliable and {larger systems would be needed} to obtain a better extrapolation.
The results for the susceptibility for $N=800$ are reported in
Fig.~\ref{fig:chiSK}a for different values of $\varphi_{\rm 0}$ (and thus $\varphi_{\rm G}$).
For $\varphi_{\rm 0}=2.50$ we also include $\chi^{N\to\infty}$ obtained using the
extrapolations of Fig.~\ref{fig:finite-size-chi}. The comparison suggests that although the finite-size
effects in both $\av{\Delta_{AB}}$ and $N(\av{\Delta_{AB}^2}-{\av{\Delta_{AB}}^2})$ are still very strong for this system size,
the determination of $\chi$ is fairly well controlled, at least for $\f < \varphi_{\rm G}$.
{The numerical data in Fig.~\ref{fig:chiSK}a suggest} a roughly linear behavior of $\chi^{-1}$ for all $\varphi_{\rm 0}$ except
$\varphi_{\rm 0}=1.8$, where the spacing between $\varphi_{\rm 0}$ and $\varphi_{\rm G}$ is narrowest, and where hopping most likely obfuscates the critical regime (see Fig.~\ref{fig:phase_diagram}).
For the other $\varphi_{\rm 0}$, we {estimate} the Gardner transition by fitting Eq.~\eqref{eq:divergence} in the $\ph<\varphi_{\rm G}$ region (Table~\ref{tab:phiG} and Fig.~\ref{fig:phase_diagram}).
In Fig.~\ref{fig:chiSK_compare} we compare the numerical {results} with the theoretical predictions.
{Note, however, that the theoretical curves} do not correspond to the full $\chi$ {value},
but only {include} the leading divergent term, namely the inverse of the replicon eigenvalue
(see~\cite{Charbonneau2014B,Rainone2666} for details).
{As discussed for Fig.~\ref{fig:delta_phi},
both the theoretical and numerical
datasets are {also} rescaled. The vertical axis is normalized to {1 at} $\f=\varphi_{\rm 0}$, while the horizontal $\f$
axis is scaled to $\varphi_{\rm G}$, using the theoretical
values in Table~\ref{table:lambdaMCT} and the fit values
$\varphi_{\rm G}^{\rm T}$ for the numerics (Table~\ref{tab:phiG})}.
{Interestingly,} the theoretical curves are not linear {over} the whole density regime. {They instead} bend towards $\varphi_{\rm G}$,
before {vanishing linearly, but only in a fairly} small region around $\varphi_{\rm G}$. {The linear scaling observed in the simulation data is thus likely}
due to finite-size effects.
The numerical estimates of $\varphi_{\rm G}^\c$ obtained {from the} linear fit of the whole set of data for $\f < \varphi_{\rm G}$
{thus} slightly
overestimate the true transition point. {Yet the effect is quite small -- approximately $2\%$, based on Fig.~\ref{fig:chiSK_compare}}.
For $\f \sim \varphi_{\rm G}$ and above, we {observe} that finite-size effects dominate the numerical determination
of $\chi$, which remains finite instead of diverging.
\begin{figure}[t]
\begin{center}
\includegraphics [width = \columnwidth] {chi_sk_phi_old.pdf}
\end{center}
\caption{(Color online). (a) Inverse caging susceptibility, $\chi^{-1}$, and (b)
skewness, $\Gamma$, as functions of $\varphi$ for different $\varphi_{\rm 0}$. Upon
approaching $\varphi_{\rm G}$, $\chi$ apparently diverges as
$|\varphi-\varphi_\mathrm{G}^\chi|^{-1}$, except for $\varphi_{\rm 0}=1.8$. Note
that the divergence of $\chi$ coincides with the maximum of $\Gamma$.
The $N\to\infty$ extrapolation of the curves in
Fig.~\ref{fig:finite-size-chi} were used to obtain the dotted line.}
\label{fig:chiSK}
\end{figure}
\subsubsection{Caging skewness}
Near the Gardner transition, large sample-to-sample fluctuations give rise to a strong exponential tail in $P(\Delta_{AB})$. This effect can be quantified by the skewness of the distribution
\begin{equation}\label{eq:skewness}
\Gamma=\frac{\av{\left(\Delta_{AB}-{\av{\Delta_{AB}}}\right)^3}}{\av{\left(\Delta_{AB}-\av{\Delta_{AB}}\right)^2}^{3/2}}\ .
\end{equation}
Recall that the skewness is a measure of a distribution's asymmetry
and that a Gaussian distribution would have $\Gamma=0$. Sample-to-sample fluctuations are expected to be maximal at the critical point (see Sec.~\ref{sec:sample}), which provides an estimate of the Gardner transition, $\varphi_{\rm G}^{\Gamma}$ (see Fig.~\ref{fig:chiSK}b and Table~\ref{tab:phiG}).
For all $\varphi_{\rm 0}$, we see that
$\varphi_{\rm G}^{\Gamma}$ is very close to the fitted divergence of the susceptibility, $\varphi_{\rm G}^\chi$.
Finite-size analysis further shows
that the rescaled skewness, $\Gamma \sqrt{N}$, collapses the data for different $N$
(Fig.~\ref{fig:SKpdf}). The peak
of $\Gamma$ is thus expected to persist all the
way to the thermodynamic limit, consistently with comparable
observations in spin-glass models~\cite{Pa12}.
\begin{figure}[t]
\begin{center}
\includegraphics [width = \columnwidth] {chi_phi_theory_phi0.pdf}
\end{center}
\caption{(Color online).
The numerical susceptibilities (scatters) are compared with theoretical predictions (lines).
Both datasets are scaled to reference values (see Sec.~\ref{sec:susceptibility} for details) to obtain a better agreement between theory and simulation.
We also report the linear fits to the numerical data used to determine $\varphi_{\rm G}^\c$ (dashed lines).
This linear fit overestimates the transition point by
roughly~2\%.
}\label{fig:chiSK_compare}
\end{figure}
\subsection{Sample-to-sample fluctuations}
\label{sec:sample}
We have assumed above that the abnormal behavior of $\Gamma$ around the Gardner transition is due to sample-to-sample fluctuations,
and we further motivate this hypothesis here.
Recall that in our notation a sample is defined by a given configuration
$\{\mathbf{r}_i\}$ and a set of random shifts $\{\boldsymbol{\Lambda}_{ij}\}$ (see Sec.~\ref{sec:planting}).
To compute the moments $\chi$ and $\Gamma$ of each sample,
we perform $N_{\mathrm s}=10,000$ independent clonings for each sample.
In Fig.~\ref{fig:sample_dependence}, the data for the ensemble of samples (same data as in Fig.~\ref{fig:chiSK})
are compared with those of four individual samples. We note that the density evolution of $\Gamma$ for the individual samples can have a very different behavior from the ensemble one. In particular, a peak around $\varphi_{\rm G}$ is generally not seen.
Sample-to-sample fluctuations have a smaller effect on $\chi$ (Fig.~\ref{fig:sample_dependence}). While the magnitude and the critical density exhibit some fluctuations, they all display a divergence similar to that of Eq.~\eqref{eq:divergence}. Because each realization of disorder corresponds to different SF-RS metabasins, our results suggest that the metabasins themselves have slightly different properties. In particular, they exhibit different $\varphi_{\rm G}$, which is likely the physical origin of the exponential tail of $P(\Delta_{\rm AB})$ and thus of its anomalous skewness~\cite{Pa12}.
\begin{figure}[t]
\begin{center}
\includegraphics [width = 0.8\columnwidth] {sample_dependence.pdf}
\end{center}
\caption{(Color online). (a) The susceptibility $\chi$ and (b) the skewness $\Gamma$ averaged over the ensemble of clones are compared with those of four individual samples, for $\varphi_{\rm 0}=2.50$.
}\label{fig:sample_dependence}
\end{figure}
\section{Summary of results}
\label{sec:summary}
The Gardner transition at $\varphi_{\rm G}$ was independently and quantitatively identified from: {\it (i)}~the power-law divergence of the characteristic time $\tau$ in Fig.~\ref{fig:timescale}, {\it (ii)}~the linear vanishing of the inverse susceptibility $\chi^{-1}$ in Fig.~\ref{fig:chiSK}a,
and {\it (iii)}~the maximum of the skewness $\Gamma$ in Fig.~\ref{fig:chiSK}b.
Table~\ref{tab:phiG} contains these results and
Figure~\ref{fig:phase_diagram} {presents} them in a phase diagram. The different estimates of $\varphi_{\rm G}$ are generally quantitatively consistent with each other and qualitatively consistent with the mean-field SF results from $d\to\infty$.
The agreement with the mean-field calculation improves with density, likely because the Gaussian caging approximation used in the theory becomes a better approximation at higher densities~\cite{KPZ12,Charbonneau2014}.
A couple of reasons underlie the discrepancy between numerical estimates and theory in the vicinity of $\varphi_{\rm d}$. First, caging becomes imperfect at such low densities, which allows particles to hop between neighboring cages on a timescale comparable with the simulation time~\cite{Charbonneau2014}. Hopping thus affects the dynamics of the system ($\Delta(t,t_w)$ and $\Delta_{AB}(t)$) and also transforms the transitions (at either $\varphi_{\rm d}$ or $\varphi_{\rm G}$) in crossovers. Second, because $\varphi_{\rm G}$ is expected to converge to $\varphi_{\rm d}$ upon approaching the dynamical glass transition (see Figs.~\ref{fig:theory} and~\ref{fig:phase_diagram}), the critical regime becomes too small to make any fit to a critical power-law scaling.
\begin{figure}[t]
\centerline{\hbox{ \includegraphics [width = \columnwidth] {PD.pdf}}}
\caption{(Color online). Inverse reduced pressure $1/p$ density $\f$ phase diagram of the MK model in $d = 3$. The liquid EOS follows Eq.~\eqref{eq:liquid_eos}, while a specific state follows a glass EOS from $\varphi_{\rm 0}$ up to jamming at $\varphi_{\rm J}$. Numerical estimates of the Gardner transition evolve with $\varphi_{\rm 0}$ similarly as the theoretical predictions~\cite{Rainone2014}.
}
\label{fig:phase_diagram}
\end{figure}
\begin{table}[t]
\caption{
Results for $\varphi_\mathrm{G}$ and $\varphi_{\rm J}$
for various $\varphi_{\rm 0}$.
Errors are estimated as follows.
For $\varphi_\mathrm{G}^{\Gamma}$, the error is the average distance between the susceptibility maximum and the next largest point.
For $\varphi_\mathrm{G}^{\mathrm{\chi}}$ and $\varphi_\mathrm{G}^{\mathrm{\tau}}$, fitting $\f$
intervals are varied to obtain the smallest and largest $\varphi_{\rm G}$ for which a fit is possible. The distance between the two values is the error.
}
\centering
\begin{tabular}{ |c | c | c | c|c| c|}
\hline
$\varphi_{\rm 0}$ & $\varphi_\mathrm{G}^{\Gamma}$ & $\varphi_\mathrm{G}^{\mathrm{\chi}}$&$\varphi_\mathrm{G}^{\mathrm{\tau}}$&$\varphi_\mathrm{G}^{\mathrm{T}}$&$\varphi_{\rm J}$\\
\hline\hline
1.8 & 2.10(5) & 2.11(3) & -- & -- & 2.534(3)\\
2.2 & 2.64(5) & 2.670(11) & 2.72(4)& -- & 2.876(4)\\
2.5 & 2.995(15) & 3.006(6)&3.06(2)&2.96 &3.151(3)\\
3.0 & 3.54(2) & 3.537(3)&3.554(17)&3.49 &3.622(5)\\
4.0 & 4.550(12) & 4.551(5) &4.57(2)&4.48&4.584(4)\\
\hline
\end{tabular}
\label{tab:phiG}
\end{table}
In addition to {\it (i)-(iii)}, a {more indirect} way {\it (iv)} to determine $\varphi_{\rm G}$ has been reported in Fig.~\ref{fig:delta_phi} {by comparing}
the density evolution of the cage order parameters with the theoretical predictions in $d\to\io$.
However, while the effect is qualitatively {(and visually) clear, obtaining a quantitative measure of $\varphi_{\rm G}$ from this approach is somewhat ill-defined.}
{For $\f > \varphi_{\rm G}$ one also should treat $\Delta$ results with caution.} As discussed above, because of the finite-size and out-of-equilibrium nature of the system, the value of $\Delta$ drifts with time as different sub-basins are explored.
{Because in practice} we evaluate $\Delta$ at a relatively short time $t_{\rm s}$ (Sec.~\ref{sec:static}), {we likely} obtain a reasonable estimate of the size of a single sub-basin, {but this procedure is also somewhat uncontrolled.}
Overall, {we find that} the most reliable way to determine $\varphi_{\rm G}$ is the divergence of $\chi$ (procedure {\it (i)}). This {effect} is clearly
the most spectacular signature of the transition, and sample-to-sample fluctuations do not much affect {its detection}
(Fig.~\ref{fig:sample_dependence}). Interestingly, because $\chi_{AB}$ is almost independent of time, one can determine $\chi$ reliably
using $\Delta_{AB}$ at short times, as we did in this paper.
Once $\varphi_{\rm G}$ is determined in this way, a useful test is to check that this value is consistent
with the behavior of $${\bf D\,}$$ and $\Delta_{AB}$, as in Fig.~\ref{fig:delta_phi}.
Procedure {\it (iii)}, i.e. finding the maximum of the skewness, requires averaging
over a large number of samples which will surely be difficult in numerical simulations of realistic models of glasses as well as in experiments,
where producing equilibrium configurations is extremely difficult. Procedure {\it (i)}, i.e. the divergence of $\tau_\b$, is also difficult because
the determination of $\tau_\b$ is subject to some ambiguity, but the study of the dynamics is useful because aging effects are manifest
for $\f > \varphi_{\rm G}$ (Fig.~\ref{fig:delta-ttw}).
Procedure {\it (iv)} is used here as a consistency check with the theory,
rather than a method to detect $\varphi_{\rm G}$.
\section{Conclusions}
\label{sec:conclusion}
The Gardner transition separates a stable glass {metabasin} at low densities (high temperatures) from a complex {hierarchy of marginally stable sub-basins, at high densities (low temperatures)}. Its existence, which is proven
in mean-field glass models~\cite{Ga85,GKS85}, has deep consequences on the low-temperature physics of glasses {and on jamming}~\cite{CKPUZ14,CCPZ15}. It is therefore
extremely interesting to check whether such a transition exists in realistic models of glasses and in experiments.
In this work we have investigated and compared several numerical procedures for detecting the Gardner
transition in the MK model, which belongs to the universality class of mean-field {spin glasses} while {remaining fairly close} to realistic glass formers~\cite{MK11}.
We have presented three independent
approaches for locating the transition,
all of which show that the transition
exists and is found in a region that is roughly consistent with theoretical predictions. We have discussed the advantages and drawbacks of each of these strategies
and the {importance} of finite-size effects.
This work paves the way for studying the Gardner transition in more
realistic numerical models of glasses, where the very existence of the Gardner transition is
debated~\cite{UB14}.
Our approach is also suitable to be
reproduced in experiments.
SF, for instance, corresponds to a straightforward annealing, and some of the observables
should be readily available through standard microscopy or scattering techniques.
One key hurdle to generalizing our methodology to other systems is the
need to
equilibrate, without planting (and thus through slow annealing), a glass state well above the (avoided) dynamic glass transition.
To follow adiabatically a glass state, one should be able to prepare
initial configurations such that the $\a$-relaxation time $\t_\a$ is very large ($\t_\a \gg \t_\b$),
so that the SF experiment can be performed on time scale $\t_\b \ll \t \ll \t_\a$, as discussed in Sec.~\ref{sec:timescales}.
For numerical simulations this requirement can be particularly computationally onerous, but it may be more easily achievable in experimental systems, where longer timescales can typically be reached.
In particular, it would be very interesting to investigate the existence of the Gardner transition in ultrastable glasses that
can be prepared through vapor deposition and have an extremely large $\t_\a$~\cite{Singh2013, Hocky2014,ultra1,ultra2}.
In experiments, the bigger challenge would be to substitute
the cloning procedure with a (potentially very) large number of experimental replicates.
Finite-dimensional non-mean-field glass formers
display features that are not observable in the
MK model. In particular, we expect a diverging length scale to be
associated with the Gardner transition in these systems. This length
scale is expected to capture static heterogeneity, which represents the
spatial inhomogeneity of cage sizes around and above $\varphi_{\rm G}$. In
principle, this kind of static heterogeneity should be different from
both the dynamic heterogeneity around the dynamic glass transition,
and the heterogeneity close to jamming, which is related to soft
relaxation modes~\cite{BW09b}.
Understanding the relevance of marginal stability for glassy dynamics, and the relation with the
ideas of Ref.~\cite{BW09b}, would open a new window on the property of low-temperature glasses.
\acknowledgments
We wish to thank L.~Berthier, O.~Dauchot, W.~Kob, J.~P.~Bouchaud, G.~Biroli, J.~Kurchan, S.~Franz,
T.~Rizzo, F.~Ricci-Tersenghi, and M.~Wyart for very useful discussions, and especially
P.~Urbani and H.~Yoshino for collaborating with two of us in the theoretical part of this project~\cite{Rainone2014,Rainone2666}.
P.C. acknowledges support from the Alfred P. Sloan Foundation and NSF support No. NSF DMR-1055586. B.S. acknowledges the support by MINECO (Spain) through research contract No. FIS2012-35719-C02. The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP7/2007-2013)/ERC grant agreement no. [247328].
|
\section{Introduction}
\subsection{The Darboux porism}
Let $ABCD$ be a planar quadrilateral. Denote by $D'$ the image of $D$ under reflection in the line $AC$. We call the transformation
\[
F_D \colon ABCD \mapsto ABCD'
\]
the \emph{$D$-folding}. It acts on the set of all quadrilaterals $ABCD$ with $A \ne C$. Define the $C$-folding in a similar way, see Figure \ref{fig:CDFold}.
\begin{figure}[htb]
\begin{center}
\begin{picture}(0,0)%
\includegraphics{CDFold.pdf}%
\end{picture}%
\setlength{\unitlength}{2072sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(11088,1753)(-1544,199)
\put(7201,659){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$C'$}%
}}}}
\put(7651,1289){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$D'$}%
}}}}
\put(9451,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$B$}%
}}}}
\put(6391,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$A$}%
}}}}
\put(-269,1784){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$C$}%
}}}}
\put(-989,1694){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$D$}%
}}}}
\put(2476,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$A$}%
}}}}
\put(3736,1829){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$C$}%
}}}}
\put(3691,1064){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$D'$}%
}}}}
\put(5536,299){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$B$}%
}}}}
\put(1531,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$B$}%
}}}}
\put(-1529,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$A$}%
}}}}
\put(1936,929){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$F_D$}%
}}}}
\put(5896,929){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$F_C$}%
}}}}
\end{picture}%
\end{center}
\caption{Composition of $C$- and $D$-foldings.}
\label{fig:CDFold}
\end{figure}
Alternate the $C$- and the $D$-foldings, so that the vertices $A$ and $B$ stay in their places, while $C$ and $D$ jump. If we are lucky, then after several iterations the points $C$ and $D$ return to their initial positions.
\begin{dfn}
A quadrilateral $ABCD$ is called \emph{$n$-periodic}, if it is invariant under the $n$-fold iteration of the composition of $C$- and $D$-foldings:
\[
(F_C \circ F_D)^n (ABCD) = ABCD
\]
\end{dfn}
The quadrilateral from Figure \ref{fig:CDFold} is $3$-periodic, see Figure \ref{fig:Darboux6}. Surprizingly enough, the periodicity depends only on the side lengths.
\begin{figure}[htb]
\begin{center}
\begin{picture}(0,0)%
\includegraphics{DarbouxPorism.pdf}%
\end{picture}%
\setlength{\unitlength}{2072sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(10667,5278)(-2833,-3371)
\put(-269,1784){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$C$}%
}}}}
\put(-989,1694){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$D$}%
}}}}
\put(1531,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$B$}%
}}}}
\put(-1529,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$A$}%
}}}}
\end{picture}%
\end{center}
\caption{A $3$-periodic quadrilateral. Side lengths are $1, 3, 3\sqrt{5}, 5$.}
\label{fig:Darboux6}
\end{figure}
\begin{thm}[Darboux \cite{Dar79}]
\label{thm:Darboux}
If a quadrilateral is $n$-periodic, then every quadrilateral with the same side lengths is $n$-periodic.
\end{thm}
The clue to the Darboux porism is the fact that the space of congruence classes of quadrilaterals with fixed side lengths is an elliptic curve, and the foldings $F_C$ and $F_D$ act on it as involutions. Hence $F_C \circ F_D$ is a translation, just as well as $(F_C \circ F_D)^n$. If a translation has a fixed point (one quadrilateral is $n$-periodic), then it is an identity (all quadrilaterals with the same side lengths are periodic).
Unlike the Poncelet porism \cite{DR14}, its relative, the Darboux porism is much less known and was rediscovered several times in the recent decades.
In the present article we take a closer look at the configuration space $\Qor^{\mathbb C}(a)$ of quadrilaterals with side lengths $a \in {\mathbb R}^4$.
We study also the space $Q^{{\mathbb C}}(a)$ of non-oriented quadrilaterals, which is an elliptic curve doubly covered by $\Qor^{\mathbb C}(a)$. There are some surprising algebraic identities that result in a natural isomorphism between the spaces $Q^{{\mathbb C}}(a)$ and $Q^{{\mathbb C}}(\bar a)$, where $a \mapsto \bar a$ is a certain involution on ${\mathbb R}^4$. Geometrically this leads to ``conjugate pairs'' of quadrilaterals. Finally, we express $n$-periodicity of a quadrilateral in terms of its side lengths.
\subsection{Euler-Chasles correspondence and Jacobi elliptic functions}
Introduce the variables
\[
z_1 = \tan\frac{\varphi_1}2, \quad z_2 = \tan\frac{\varphi_2}2
\]
where $\varphi_1$ and $\varphi_2$ are adjacent angles of a quadrilateral. Then the space of congruence classes of quadrilaterals, respecting the orientation, with fixed side lengths becomes identified with the algebraic curve
\begin{equation}
\label{eqn:Curve}
c_{22} z_1^2 z_2^2 + c_{20} z_1^2 + c_{02} z_2^2 + 2 c_{11} z_1 z_2 + c_{00} = 0
\end{equation}
Here coefficients $c_{ij}$ depend on the side lengths. This is a special biquadratic equation in two variables. Biquadratic equations are also known under the name of Euler-Chasles correspondences \cite{BV96}. In the past decades they attracted a lot of attention: as a basis for $QRT$-maps in the theory of discrete integrable systems \cite{Dui10}; as a solution of the Yang-Baxter equation \cite{Bax89, Kri81}; as an approach to flexible polyhedra \cite{Sta10, Gai13, Izm_Koko}.
It is known that in a generic case the curve \eqref{eqn:Curve} can be parametrized as
\[
z_1 = p_1 \varphi(t), \quad z_2 = p_2 \varphi(t+\tau)
\]
where $\varphi$ is a second-order elliptic function with simple poles and zeros. Since any pair of adjacent angles is related by a biquadratic equation, all angles turn out to be scaled shifts of the same elliptic function.
If $a = (a_1, a_2, a_3, a_4)$ are the side lengths of a quadrilateral, then the non-degeneracy condition ensuring that the configuration space $Q_{or}(a)$ is an elliptic curve is
\[
a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0
\]
for all choices of signs. Put differently, this means $a_{\min} + a_{\max} \ne s$, where $a_{\min}$ and $a_{\max}$ are the lengths of the shortest and the longest side, and $s = \frac{a_1+a_2+a_3+a_4}2$ is the half-perimeter.
In kinematics, the inequality $a_{\min} + a_{\max} < s$ is knows as the Grashof condition. It means that the shortest side can make a full turn with respect to each of its neighbors; also it means that the real part of the configuration space $Q_{or}(a)$ has two components, that is the quadrilateral cannot be deformed into its mirror image. The following theorem describes a parametrization of $\Qor^{\mathbb C}(a)$ and underlines the difference between Grashof and non-Grashof quadrilaterals.
\begin{thm}
\label{thm:Param}
If $a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0$, then the complexified configuration space $\Qor^{\mathbb C}(a)$ of quadrilaterals with side lengths $a$ is an elliptic curve ${\mathbb C}/\Lambda$.
\begin{enumerate}
\item
If $a_{\min} + a_{\max} < s$, then the lattice $\Lambda$ is rectangular, and the cotangents of the halves of the exterior angles can be parametrized as
\begin{align*}
\cot\frac{\varphi_1}2 &= q_1 \dn(t+t_1; k), &\cot\frac{\varphi_2}2 &= q_2 \dn(t+t_2; k),\\
\cot\frac{\varphi_3}2 &= q_3 \dn(t+t_3; k), &\cot\frac{\varphi_4}2 &= q_4 \dn(t+t_4; k)
\end{align*}
Here $k = \sqrt{\frac{a_1a_2a_3a_4}{\bar a_1 \bar a_2 \bar a_3 \bar a_4}}$, the amplitudes $q_i$ are real or purely imaginary, and the shifts $t_i$ satisfy
\[
\Im t_i \in \{0, 2K'\}, \quad t_1 - t_3, t_2 - t_4 \in \{\pm K\}
\]
The real part of the configuration space consists of two components corresponding to $\Im t \in \{0, 2K'\}$.
\item
If $a_{\min} + a_{\max} > s$, then the lattice $\Lambda$ is rhombic, and the cotangents of the halves of the exterior angles can be parametrized as
\begin{align*}
\cot\frac{\varphi_1}2 &= q_1 \cn(t+t_1; k), &\cot\frac{\varphi_2}2 &= q_2 \cn(t+t_2; k),\\
\cot\frac{\varphi_3}2 &= q_3 \cn(t+t_3; k), &\cot\frac{\varphi_4}2 &= q_4 \cn(t+t_4; k)
\end{align*}
Here $k = \sqrt{\frac{\bar a_1 \bar a_2 \bar a_3 \bar a_4}{a_1a_2a_3a_4}}$, each of the amplitudes $q_i$ is real or purely imaginary, and the shifts $t_i$ satisfy
\[
\Im t_i \in \{0, 2K'\}, \quad t_1 - t_3, t_2 - t_4 \in \{\pm K\}
\]
The real part of the configuration space has one component corresponding to $t \in {\mathbb R}$.
\end{enumerate}
\end{thm}
The proof of Theorem \ref{thm:Param} is contained in Sections \ref{sec:ParamThm} and \ref{sec:EllReal}. There one can also find exact values of the amplitudes and shifts.
The numbers $\bar a_i$ appearing in the formula for the Jacobi modulus are defined as follows.
\begin{equation}
\label{eqn:BarA}
\bar a_1 = \frac{-a_1 + a_2 + a_3 + a_4}2 = s-a_1, \quad \bar a_i = s - a_i, \ i = 2, 3, 4
\end{equation}
The pair of opposite angles $\varphi_1$, $\varphi_3$ in a quadrilateral is also subject to a relation of the form \eqref{eqn:Curve}, but with a vanishing coefficient at $z_1z_3$. A scaling of variables brings the equation into the form
\begin{equation}
\label{eqn:Opp}
u^2 + v^2 = 1 + mu^2v^2
\end{equation}
with a real $m<1$.
This curve can be parametrized as
\[
u = f(t;k), \quad v = f(t+K;k)
\]
where $f = \dn$ if $m>0$ and $f = \cn$ if $m<0$. Equation \eqref{eqn:Opp} is closely related to the equation
\[
u^2 + v^2 = a^2 + a^2u^2v^2
\]
suggested by Edwards \cite{Edw07} as a new normal form for elliptic curves. Edwards constructs a parametrization of $u$ and $v$ ``from the scratch'' using a version of theta-functions.
\subsection{The space of non-oriented quadrilaterals}
\label{sec:Conj}
Let $Q^{{\mathbb C}}(a)$ be the space of congruence classes, disregarding the orientation, of quadrilaterals with side lengths $a$. Clearly, the map $\Qor^{\mathbb C}(a) \to Q^{{\mathbb C}}(a)$ is a double cover, easy to describe in terms of the holomorphic parameter of Theorem \ref{thm:Param}. There is an unexpected natural isomorphism between the spaces $Q^{{\mathbb C}}(a)$ and $Q^{{\mathbb C}}(\bar a)$ with $\bar a$ as in \eqref{eqn:BarA}: for every quadrilateral with the side lengths $a$ there is a quadrilateral with the same diagonal lengths and the side lengths $\bar a$.
\begin{thm}
\label{thm:NonOr}
If $a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0$, then the complexified configuration space $Q^{{\mathbb C}}(a)$ of non-oriented quadrilaterals with the side lengths $a$ is an elliptic curve ${\mathbb C}/\Lambda$ with a rectangular lattice $\Lambda$.
There is a natural isomorphism
\[
Q^{{\mathbb C}}(a) \to Q^{{\mathbb C}}(\bar a)
\]
that identifies two quadrilaterals with the same diagonal lengths.
If $\Qor^{\mathbb C}(a)$ has a rectangular lattice, then $\Qor^{\mathbb C}(\bar a)$ has a rhombic lattice. In particular, the above isomorphism does not lift to the spaces of oriented quadrilaterals. The double covers
\[
\Qor^{\mathbb C}(a) \to Q^{{\mathbb C}}(a) \cong Q^{{\mathbb C}}(\bar a) \leftarrow \Qor^{\mathbb C}(\bar a)
\]
look as shown on Figure \ref{fig:Covers}.
\end{thm}
Theorem \ref{thm:NonOr} is proved in Section \ref{sec:NonOr}.
\begin{figure}[ht]
\centering
\includegraphics{DoubleCovers.pdf}
\caption{Spaces $\Qor^{\mathbb C}(a)$ and $\Qor^{\mathbb C}(\bar a)$ covering $Q^{{\mathbb C}}(a) \cong Q^{{\mathbb C}}(\bar a)$.}
\label{fig:Covers}
\end{figure}
The isomorphism $Q^{{\mathbb C}}(a) \cong Q^{{\mathbb C}}(\bar a)$ turns out to be equivalent to the Ivory theorem, see Figure \ref{fig:Ivory}. Theorem \ref{thm:NonOr} holds also for spherical and hyperbolic quadrilaterals. This extends the Ivory theorem to the sphere and the hyperbolic plane, with ellipses and hyperbolas defined geodesically, see Theorem \ref{thm:IvorySph}.
Computations that allow to express in a particularly nice way the Jacobi modulus, amplitudes, and shifts in Theorem \ref{thm:Param}, and those leading to Theorem \ref{thm:NonOr} are based on a number of identities for dual quadruples $(a,b,c,d)$ and $(\bar a, \bar b, \bar c, \bar d)$. These are collected below.
\begin{gather*}
ab - \bar c \bar d = (s-a-c)(s-b-c), \quad ab - \bar a \bar b = s(s-c-d)\\
abcd - \bar a \bar b \bar c \bar d = s(s-a-b)(s-b-c)(s-a-c)\\
ab + cd = \bar a \bar b + \bar c \bar d, \quad
a^2 + b^2 + c^2 + d^2 = \bar a^2 + \bar b^2 + \bar c^2 + \bar d^2\\
ab - cd = \frac12 (\bar c^2 + \bar d^2 - \bar a^2 - \bar b^2)
\end{gather*}
\subsection{The periodicity condition}
Benoit and Hulin \cite{BH04} studied the periodicity condition by constructing a pair of circles whose Poncelet dynamics is equivalent to the folding dynamics of the quadrilateral.
The following theorem deals with the periodicity disregarding the orientation. Proposition \ref{prp:Period} describes how the period lengths on the curves $Q^{{\mathbb C}}(a)$ and $\Qor^{\mathbb C}(a)$ are related.
\begin{thm}
\label{thm:Period}
A quadrilateral with the side lengths $a$ is $n$-periodic disregarding the orientation if and only if the following condition is satisfied.
\begin{align*}
&\begin{vmatrix}
A_2 & A_3 & \ldots & A_{k+1}\\
A_3 & A_4 & \ldots & A_{k+2}\\
\vdots & \vdots & \ddots & \vdots\\
A_{k+1} & A_{k+2} & \ldots & A_{2k}
\end{vmatrix}
= 0, \quad \text{ if } n = 2k+1\\
&\begin{vmatrix}
A_3 & A_4 & \ldots & A_{k+1}\\
A_4 & A_5 & \ldots & A_{k+2}\\
\vdots & \vdots & \ddots & \vdots\\
A_{k+1} & A_{k+2} & \ldots & A_{2k-1}
\end{vmatrix}
= 0, \quad \text{ if } n = 2k
\end{align*}
Here $A_i$ are the coefficients of the expansion
\[
\sqrt{(x - \Delta(a)^2)(x - \delta(a)^2)(x - \delta(\bar a)^2)} = \sum_{i=0}^\infty A_i x^i
\]
where
\[
\Delta(a) = a_1a_3 + a_2a_4, \quad \delta(a) = a_1a_3 - a_2a_4
\]
\end{thm}
Note that $\Delta(\bar a) = \Delta(a)$ due to the identities from Section \ref{sec:Conj}, so that the quadrilaterals with the side lengths $\bar a$ have the same folding period (disregarding the orientation) as the quadrilaterals with the side lengths $a$. This is also obvious from the isomorphism $Q^{{\mathbb C}}(a) \cong Q^{{\mathbb C}}(\bar a)$ via equal diagonal lengths.
\subsection{Acknowledgments}
Parts of this work were done during the author's visits to IHP Paris and to the Penn State University. The author thanks both institutions for hospitality. Also he wishes to thank Arseniy Akopyan, Udo Hertrich-Jeromin, Boris Springborn, and Yuri Suris for useful discussions.
\section{The space of oriented quadrilaterals in terms of their angles}
\label{sec:Eq}
\subsection{Notation}
A \emph{planar quadrilateral} is for us an ordered quadruple of points $(A, B, C, D)$ in the euclidean plane such that
\[
A \ne B,\quad B \ne C,\quad C \ne D,\quad D \ne A
\]
Two quadrilaterals $ABCD$ and $A'B'C'D'$ are called \emph{directly congruent} if there exists an orientation-preserving isometry $f \colon {\mathbb R}^2 \to {\mathbb R}^2$ such that
\[
f(A) = A',\quad f(B) = B',\quad f(C) = C',\quad f(D) = D'
\]
In this article, we study the set of (direct) congruence classes of quadrilaterals with fixed side lengths. A mechanical interpretation of this is the configuration space of a four-bar linkage with the positions of two adjacent joints fixed and whose bars are allowed to cross each other.
For a quadruple $a := (a_1, a_2, a_3, a_4)$ of real numbers there exists a
quadrilateral with side lengths $a$ if and only if the following inequalities hold:
\begin{subequations}
\begin{equation}
\label{eqn:QuadIneqA}
a_i > 0 \quad \forall i
\end{equation}
\begin{equation}
\label{eqn:QuadIneqB}
a_i < a_j + a_k + a_l \quad \forall i,
\end{equation}
\end{subequations}
where $\{j,k,l\} = \{1,2,3,4\} \setminus \{i\}$ in the second line.
\begin{dfn}
For $a \in {\mathbb R}^4$ satisfying the conditions \eqref{eqn:QuadIneqA} and \eqref{eqn:QuadIneqB}, denote by $Q_{or}(a)$ the set of direct congruence classes of quadrilaterals $ABCD$ with
\[
AB = a_1, \quad BC = a_2, \quad CD = a_3, \quad DA = a_4
\]
\end{dfn}
\begin{rem}
We require a congruence to preserve the marking of the vertices (or, equivalently, the marking of the sides). This does matter only if the sequence $a$ is symmetric under the action of an element of the dihedral group on $(1, 2, 3, 4)$.
\end{rem}
Denote by $\varphi_i$ the angle between the sides marked by $i-1$ and $i (\mathrm{mod}\
4)$. More exactly, $\varphi_i$ are the turning angles for the velocity vector of a point that
runs along the perimeter in the direction given by the cyclic order $(1,2,3,4)$, see Figure
\ref{fig:SidesAngles}, left.
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{NotV.pdf}%
\end{picture}%
\setlength{\unitlength}{2072sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(10059,2588)(-191,-1333)
\put(2044,144){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(7156,-1169){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_1$}%
}}}}
\put(3511,-16){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(9136,164){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$v_2$}%
}}}}
\put(6976, 29){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$v_4$}%
}}}}
\put(1126,-961){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(4681,-961){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(8056,-1006){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$v_1$}%
}}}}
\put(9541,-601){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_2$}%
}}}}
\put(633,757){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(-166,-285){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(2521,-561){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_2$}%
}}}}
\put(131,-1200){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_1$}%
}}}}
\put(-52,300){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_4$}%
}}}}
\put(1222,1054){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_3$}%
}}}}
\put(3906,-649){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\pi-\varphi_1$}%
}}}}
\put(4554,-37){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$y$}%
}}}}
\put(4392,838){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(5710,-225){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(4525,360){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\pi-\varphi_3$}%
}}}}
\end{picture}%
\caption{Notations; finding a relation for opposite
angles; finding a relation for adjacent angles.}
\label{fig:SidesAngles}
\end{figure}
\subsection{Equations relating the angles of a quadrilateral}
\begin{lem}
\label{lem:CosLinDep}
The cosines of opposite angles of a quadrilateral are subject to a linear dependence, with coefficients depending on the side lengths.
\end{lem}
\begin{proof}
By expressing the diagonal length $y$ on Figure \ref{fig:SidesAngles} with the
help of the cosine
law first through $a_1, a_2, \varphi_1$ and then through $a_3, a_4, \varphi_3$,
we obtain
\begin{equation}
\label{eqn:LinCos}
a_1^2 + a_4^2 + 2a_1a_4 \cos\varphi_1 = a_2^2 + a_3^2 + 2a_2a_3 \cos\varphi_3
= 0
\end{equation}
\end{proof}
This result is classical. Bricard \cite{Bri97} mentions it as well-known. The following substitution also appears in \cite{Bri97}.
\begin{equation}
\label{eqn:XTan}
z_i := \tan \frac{\varphi_i}2
\end{equation}
\begin{prp}
\label{prp:OppAngles}
The tangents of the opposite half-angles of a quadrilateral satisfy the following algebraic relation
\begin{equation}
\label{eqn:OppX}
b_{22} z_1^2 z_3^2 + b_{20} z_1^2 + b_{02} z_3^2 + b_{00} = 0, \quad
\text{where}
\end{equation}
\begin{equation*}
\begin{aligned}
b_{22} &= (a_1 + a_2 - a_3 - a_4)(a_1 - a_2 + a_3 - a_4)\\
b_{20} &= (a_1 + a_2 + a_3 - a_4)(a_1 - a_2 - a_3 - a_4)\\
b_{02} &= (a_1 - a_2 + a_3 + a_4)(a_1 + a_2 - a_3 + a_4)\\
b_{00} &= (a_1 - a_2 - a_3 + a_4)(a_1 + a_2 + a_3 + a_4)
\end{aligned}
\end{equation*}
\end{prp}
\begin{proof}
Substitute $\cos\varphi_1 = \frac{1 - z_1^2}{1 + z_1^2}$
and $\cos\varphi_3 = \frac{1 - z_3^2}{1 + z_3^2}$ in \eqref{eqn:LinCos}. A simple computation yields
\eqref{eqn:OppX}.
\end{proof}
With a bit more work one finds a relation between pairs of adjacent angles.
\begin{prp}
\label{prp:AdjCoord}
The tangents of the adjacent half-angles of a quadrilateral satisfy the equation
\begin{equation}
\label{eqn:AdjX}
c_{22} z_1^2 z_2^2 + c_{20} z_1^2 + c_{02} z_2^2 + 2 c_{11} z_1 z_2 + c_{00} = 0,
\quad \text{where}
\end{equation}
\begin{equation*}
\begin{aligned}
c_{22} &= (a_1 - a_2 - a_3 - a_4)(a_1 - a_2 + a_3 - a_4)\\
c_{20} &= (a_1 + a_2 + a_3 - a_4)(a_1 + a_2 - a_3 - a_4)\\
c_{02} &= (a_1 - a_2 + a_3 + a_4)(a_1 - a_2 - a_3 + a_4)\\
c_{11} &= -4a_2a_4\\
c_{00} &= (a_1 + a_2 - a_3 + a_4)(a_1 + a_2 + a_3 + a_4)
\end{aligned}
\end{equation*}
\end{prp}
\begin{proof}
Denote by $v_i$ the vector running along the $i$-th side of the quadrilateral, see Figure \ref{fig:SidesAngles}, right. Expanding the scalar
product on the left
hand side of $\|v_4 + v_1 + v_2\|^2 = \|v_3\|^2$ we obtain
$$
(a_1^2 + a_2^2 - a_3^2 + a_4^2) + 2 a_1 a_4 \cos \varphi_1 + 2 a_1 a_2 \cos
\varphi_2 + 2 a_2 a_4
\cos (\varphi_1 + \varphi_2) = 0
$$
The substitution and a tedious computation produce \eqref{eqn:AdjX}.
\end{proof}
\subsection{Bihomogeneous equations and algebraic curves in ${\mathbb C}{\mathrm P}^1 \times {\mathbb C}{\mathrm P}^1$}
The substitution \eqref{eqn:XTan} identifies ${\mathbb R}/2\pi$, which is the range of $\varphi_i$, with ${\mathbb R}{\mathrm P}^1$, which is the range of $z_i$. Thus it is geometrically reasonable to consider equations \eqref{eqn:OppX} and \eqref{eqn:AdjX} as equations on $({\mathbb R}{\mathrm P}^1)^2$. The proper setting for this are bihomogeneous polynomials.
That is, we introduce projective variables $(z_1 : w_1)$ and $(z_2 : w_2)$ and rewrite equation \eqref{eqn:AdjX} as
\begin{equation}
\label{eqn:Bihom}
c_{22} z_1^2 z_2^2 + c_{20} z_1^2 w_2^2 + c_{02} w_1^2 z_2^2 + 2 c_{11} z_1w_1 z_2w_2 + c_{00} w_1^2 w_2^2 = 0
\end{equation}
This point of view doesn't affect the affine part of the curve but may change the number of points at infinity (the infinity of ${\mathbb R}{\mathrm P}^1 \times {\mathbb R}{\mathrm P}^1$ is the union of two projective lines instead of one for ${\mathbb R}{\mathrm P}^2$). Indeed, while the usual projectivization of \eqref{eqn:AdjX} has two points $(1 : 0 : 0)$ and $(0 : 1 : 0)$ at infinity, the curve \eqref{eqn:Bihom} has four:
\[
\left( \pm
\sqrt{\frac{-c_{02}}{c_{22}}}, \infty \right), \quad \left( \infty, \pm
\sqrt{\frac{c_{20}}{c_{22}}}
\right)
\]
For a generic choice of coefficients $c_{ij}$ the curve \eqref{eqn:Bihom} is non-singular, as opposed to the usual projectivization of \eqref{eqn:AdjX}.
\subsection{A system of six equations}
By Propositions \ref{prp:OppAngles} and \ref{prp:AdjCoord}, the angles of every quadrilateral satisfy a system of six equations: two of the form \eqref{eqn:OppX} and four of the form \eqref{eqn:AdjX}. In this section we show that, vice versa, under a certain genericity assumption every solution of the system corresponds to a quadrilateral.
\begin{prp}
Assume that the quadruple $a$ is not made of two pairs of equal adjacent numbers:
\begin{equation}
\label{eqn:NonDeg}
\text{neither } a_1 = a_2, a_3 = a_4 \text{ nor } a_1 = a_4, a_2 = a_3
\end{equation}
Then every solution $(z_1, z_2, z_3, z_4)$ of the system of six equations on the pairs of angles corresponds to a unique quadrilateral in $Q_{or}(a)$.
\end{prp}
\begin{proof}
By reverting the argument in the proof of Proposition \ref{prp:AdjCoord} one sees that for every solution of \eqref{eqn:AdjX} there is a unique quadrilateral in $Q_{or}(a)$ with angles $\varphi_1$ and $\varphi_2$. The same is true for the other three equations relating adjacent angles. For every solution of \eqref{eqn:OppX} there is also a quadrilateral with angles $\varphi_1$ and $\varphi_3$. This quadrilateral can be non-unique only if the second and the fourth vertices coincide, for which $a_1=a_4$ and $a_2=a_3$ is needed.
Thus for every solution $(z_1, z_2, z_3, z_4)$ of the system of six equations there are six quadrilaterals $C_{ij} \in Q_{or}(a)$ with the property that in $C_{ij}$ the angles $\varphi_i$ and $\varphi_j$ have the correct values. Our goal is to show that there is $C$ with all correct angles. Consider $C_{12}$, $C_{13}$, and $C_{14}$. They all have the same $\varphi_1$. Assumption \eqref{eqn:NonDeg} implies that there are at most two quadrilaterals in $Q_{or}(a)$ with a given $\varphi_1$. Thus at least two of the three quadrilaterals must coincide. This yields a quadrilateral $C_1$ with three correct angles, one of which is $\varphi_1$. If the fourth angle $\varphi_i$ is incorrect, then again, consider the three quadrilaterals where this angle is correct and obtain a quadrilateral $C_i$ with three correct angles, one of which is $\varphi_i$. As $C_1$ and $C_i$ have two angles in common, they coincide, and $C = C_1 = C_i$ is the desired quadrilateral.
\end{proof}
\subsection{Birational equivalence between \eqref{eqn:OppX} and \eqref{eqn:AdjX}}
We have just seen that the configuration space $Q_{or}(a)$ is an algebraic curve in $({\mathbb R}{\mathrm P}^1)^4$ given by six equations. It will often be convenient to restrict our attention to a single equation in two variables. The following statement allows us to do so.
\begin{prp}
\label{prp:AlgConfSpace}
The projection of $Q_{or}(a) \subset ({\mathbb R}{\mathrm P}^1)^4$ to every coordinate plane $(z_i, z_{i+1})$ is
a birational equivalence.
The projection of $Q_{or}(a)$ to the coordinate plane $(z_i, z_{i+2})$ is a birational
equivalence unless $a_i = a_{i-1}$ and $a_{i+1} = a_{i+2}$. (All indices are taken in $\{1,2,3,4\}$ modulo $4$.)
\end{prp}
\begin{proof}
It suffices to consider the case $i=1$.
As noted before, the projection of $Q_{or}(a)$ to the $(z_1, z_2)$-plane is injective. It suffices to show that for every $(z_1, z_2, z_3, z_4) \in Q_{or}(a)$ the values of $z_3$ and $z_4$ are rational functions of $z_1$ and $z_2$. By projecting the quadrilateral to its second side and to the line orthogonal to it, we obtain
\begin{equation}
\label{eqn:Birat}
\begin{aligned}
a_1 \cos\varphi_2 + a_2 + a_3 \cos\varphi_3 + a_4 \cos(\varphi_1+\varphi_2) = 0\\
a_1 \sin\varphi_2 - a_3 \sin\varphi_3 + a_4 \sin(\varphi_1+\varphi_2) = 0
\end{aligned}
\end{equation}
It follows that
\[
z_3 = \frac{1-\cos\varphi_3}{\sin\varphi_3} = \frac{a_1 \cos\varphi_2 + a_2 + a_3 + a_4 \cos(\varphi_1+\varphi_2)}{a_1 \sin\varphi_2 + a_4 \sin(\varphi_1+\varphi_2)}
\]
which is rational in $z_1$ and $z_2$.
Similarly, if either $a_1 \ne a_4$ or $a_2 \ne a_3$, then the projection of $Q_{or}(a)$ to the $(z_1, z_3)$-plane is injective. To show that the inverse map is rational, rewrite \eqref{eqn:Birat} as
\[
\begin{aligned}
(a_1 + a_4 \cos\varphi_1) \cos\varphi_2 - a_4 \sin\varphi_1 \cdot \sin\varphi_2 = -a_2 - a_3\cos\varphi_3\\
a_4\sin\varphi_1 \cdot \cos\varphi_2 + (a_1 + a_4 \cos\varphi_1) \sin\varphi_2 = a_3 \sin\varphi_3
\end{aligned}
\]
This is a system of linear equations on $\cos\varphi_2$ and $\sin\varphi_2$ with the determinant $(a_1 + a_4 \cos\varphi_1 )^2 + a_4^2 \sin^2\varphi_1$. The determinant vanishes only for $\varphi_1 = \pi$ and $a_1 = a_4$, which also implies $a_2 = a_3$. If it does not vanish, then by solving the system we can express $\cos\varphi_2$ and $\sin\varphi_2$ as rational functions of $z_1$ and $z_3$.
\end{proof}
Now that the algebraic structure of $Q_{or}(a)$ became clear, we can define the complexified configuration space.
\begin{dfn}
The \emph{complexified configuration space} $\Qor^{\mathbb C}(a)$ of oriented quadrilaterals with edge lengths $a$ is the algebraic curve in $({\mathbb C}{\mathrm P}^1)^4$ defined by the six equations of the form \eqref{eqn:OppX} and \eqref{eqn:AdjX}.
\end{dfn}
Due to Proposition \ref{prp:AlgConfSpace}, $\Qor^{\mathbb C}(a)$ is birationally equivalent to the curve \eqref{eqn:AdjX} and is birationally equivalent to \eqref{eqn:OppX} unless $a_1 = a_4$ and $a_2 = a_3$.
\section{Parametrizations of the configuration spaces of oriented quadrilaterals}
\label{sec:Param}
\subsection{Classifying linkages by their degree of degeneracy}
The shape of the configuration space $\Qor^{\mathbb C}(a)$ turns out to depend on the number of solutions of the equation
\begin{equation}
\label{eqn:Grashof}
a_1 \pm a_2 \pm a_3 \pm a_4 = 0
\end{equation}
It is easy to see that if \eqref{eqn:Grashof} has at least two solutions, then $(a_1, a_2, a_3, a_4)$ consists of two pairs of equal numbers. This leads to the following classification of quadrilaterals.
\begin{dfn}
\label{dfn:Types}
A quadrilateral with side lengths $a_1, a_2, a_3, a_4$ (in this cyclic order) is said to be
\begin{itemize}
\item of \emph{elliptic} type, if equation \eqref{eqn:Grashof} has no solution;
\item of \emph{conic} type, if the equation \eqref{eqn:Grashof} has exactly one solution;
\item an \emph{isogram}, if its opposite sides are equal: $a_1 = a_3$, $a_2 = a_4$;
\item a \emph{deltoid}, if two pairs of adjacent sides are equal: $a_1 = a_2, a_3 = a_4$ or $a_1 = a_4, a_2 = a_3$;
\item a \emph{rhombus}, if all sides are equal.
\end{itemize}
\end{dfn}
Let us look at the last three simple cases before we proceed to the more interesting conic and elliptic types.
\subsubsection{The rhombus}
Equation \eqref{eqn:AdjX} becomes $z_1^2 = z_3^2$, and \eqref{eqn:OppX} becomes (in the bihomogeneous form)
\[
w_1w_2(z_1z_2 - w_1w_2) = 0
\]
The other equations can be obtained by cyclically permuting the indices. The configuration space consists of three lines
\[
z_1 = \infty, \quad z_2 = \infty, \quad z_1z_2 = 1
\]
The first two consist of ``folded'' configurations, when two opposite vertices are at the same point, and the edges rotate around this point; the third line corresponds to actual rhombi with $\varphi_1 + \varphi_2 = \pi$.
Note that the condition \eqref{eqn:NonDeg} is violated, and the system has ``fantom'' solutions that don't correspond to any quadrilateral, e.~g. $z_1 = z_3 = \infty$, $z_2 = z_4$.
\subsubsection{The deltoid}
Assume $a_1 = a_2 \ne a_3 = a_4$. Equation \eqref{eqn:AdjX} becomes
\[
(a_1-a_3)z_1^2 - 2a_3z_1z_2 + (a_1+a_3) = 0
\]
Besides, in the bihomogeneous form the factor $w_2$ appears, so that the configuration space consists of two lines
\[
\Qor^{\mathbb C}(a) = \{z_2 = \infty\} \cup \left\{ z_2 = \frac{(a_1-a_3)z_1^2 + (a_1+a_3)}{2a_3z_1} \right\}
\]
Again, the first line corresponds to a folded deltoid, and the second expresses an angle at a ``peak'' of the deltoid through the angle at its base. The fantom solution $z_1 = z_3 = \infty$, $z_2 = z_4$ is also present in this case.
\subsubsection{The isogram}
Under assumption $a_1 = a_3 \ne a_2 = a_4$ equation \eqref{eqn:AdjX} becomes
\[
(a_1-a_2)z_1^2z_2^2 + 2a_2 z_1z_2 - (a_1+a_2) = 0
\]
which factorizes as
\[
(z_1z_2 - 1)((a_1-a_2)z_1z_2 + (a_1+a_2)) = 0
\]
The configuration space consists of two lines: $z_2 = z_1^{-1}$, consisting of parallelograms, and $z_2 = \frac{a_1+a_2}{a_1-a_2} z_2^{-1}$, consisting of antiparallelograms.
\subsection{Conic quadrilaterals: parametrization by trigonometric functions}
If equation \eqref{eqn:Grashof} has exactly one solution, then two cases must be distinguished: either the two pairs of opposite sides add up to the same total length, or two pairs of adjacent sides do so.
In what follows we will be repeatedly using the fact that $a+b = c+d$ implies
\begin{equation*}
ac - bd = (a+b)(a-d), \quad cd - ab = (a-c)(a-d)
\end{equation*}
\subsubsection{Circumsribable quadrilaterals: $a_1 + a_3 = a_2 + a_4$}
\begin{prp}
\label{prp:Param1+3=2+4}
Assume that $a_1 + a_3 = a_2 + a_4$ is the unique solution of the equation \eqref{eqn:Grashof} and that $a_3 = a_{\min}$. Then the configuration space $\Qor^{\mathbb C}(a)$ is isomorphic to the one-point compactification of ${\mathbb C}/2\pi{\mathbb Z}$, and a bijection ${\mathbb C}/2\pi{\mathbb Z} \to \Qor^{\mathbb C}(a) \setminus \{\infty\}$ can be established by the parametrization
\begin{gather*}
z_1 = p_1 \sin t, \quad z_2 = p_2 \sin (t + i\sigma), \\
z_3 = p_3 \sin \left(t - \frac{\pi}2\right), \quad
z_4 = p_4 \sin \left(t + \frac{\pi}2 + i\sigma\right)
\end{gather*}
where
$$
p_1 = \sqrt{\frac{(a_1+a_3)(a_1-a_2)}{a_2a_3}} = \sqrt{\frac{a_1a_4}{a_2a_3} - 1}
$$
with $p_2$, $p_3$, $p_4$ obtained by cyclically permuting the indices, and
$$
\sigma = \ln \left( \sqrt{\frac{a_2a_4}{a_2a_4 - a_1a_3}} + \sqrt{\frac{a_1a_3}{a_2a_4 - a_1a_3}} \right)
$$
The square roots of negative numbers are assumed to take value in $i{\mathbb R}_+$.
\end{prp}
\begin{proof}
Due to $a_1+a_3 = a_2+a_4$ equations \eqref{eqn:OppX} and \eqref{eqn:AdjX} become
\begin{subequations}
\begin{equation}
\label{eqn:OppCon}
-a_2a_3 z_1^2 + a_1a_4 z_3^2 + (a_1a_4 - a_2a_3) = 0
\end{equation}
\begin{equation}
\label{eqn:Conic1}
a_2(a_1-a_4) z_1^2 + a_4(a_1-a_2) z_2^2 - 2a_2a_4 z_1z_2 + a_1(a_1+a_3) = 0
\end{equation}
\end{subequations}
Equation \eqref{eqn:OppCon} rewrites as
\[
\frac{z_1^2}{p_1^2} + \frac{z_3^2}{p_3^2} = 1
\]
with
\[
p_1 = \sqrt{\frac{a_1a_4}{a_2a_3} - 1}, \quad p_3 = \sqrt{\frac{a_2a_3}{a_1a_4} - 1}
\]
so that the coordinates $(z_1, z_3)$ can be parametrized as
\[
z_1 = p_1 \sin t, \quad z_3 = p_3 \cos t
\]
The substitution of $z_1 = p_1 u_1$ and $z_2 = p_2 u_2$ in \eqref{eqn:Conic1} under the assumption $a_1 = a_{\max}$ results in
\[
u_1^2 + u_2^2 - 2 \sqrt{\frac{a_2a_4}{a_2a_4-a_1a_3}} u_1u_2 + \frac{a_1a_3}{a_2a_4 - a_1a_3} = 0
\]
which has the form
\begin{equation}
\label{eqn:SinShift}
u^2 + v^2 - 2 \cos \tau uv - \sin^2 \tau = 0
\end{equation}
with $\tau = i\sigma$ and $\sigma$ as stated in the theorem. On the other hand, $u$ and $v$ in \eqref{eqn:SinShift} can be parametrized as $u(t) = \sin t$ and $v(t) = \sin(t+\tau)$. This leads to
\[
z_1 = p_1 \sin t, \quad z_2 = p_2 \sin(t+i\sigma)
\]
The relation between each pair of opposite angles imply that the above parametrization is completed by
\[
z_3 = p_3 \sin\left(t \pm \frac{\pi}2\right), \quad z_4 = p_4 \sin\left(t+i\sigma \pm \frac{\pi}2\right)
\]
for some choices of the signs. To determine the signs, one may analyze the equations between $z_2$ and $z_3$ and between $z_3$ and $z_4$ in a similar way. Alternatively, one looks at special configurations of quadrilaterals.
\end{proof}
\subsubsection{Edge lengths satisfy $a_1+a_2 = a_3+a_4$}
Equation \eqref{eqn:OppX} relating $z_1$ and $z_3$ takes the same form as in the case of curcumscribable quadrilaterals:
\[
-a_2a_3 z_1^2 + a_1a_4 z_3^2 + (a_1a_4 - a_2a_3) = 0
\]
so that we have again a parametrization
\[
z_1 = p_1 \sin t, \quad z_3 = p_3 \cos t
\]
But for the pair $\varphi_2, \varphi_4$ we have
\[
(a_3a_4-a_1a_2) z_2^2z_4^2 - a_1a_2 z_2^2 + a_3a_4 z_4^2 = 0
\]
or, equivalently,
\[
-a_3a_4 z_2^{-2} + a_1a_2 z_4^{-2} + (a_1a_2 - a_3a_4) = 0
\]
which leads to
\[
z_2^{-1} = p_2 \sin t, \quad z_4^{-1} = p_4 \cos t
\]
Similarly to the preceding section, we obtain the following (note a sign change for $z_3$).
\begin{prp}
\label{prp:Param1+2=3+4}
Assume that $a_1 + a_2 = a_3 + a_4$ is the unique solution of the equation \eqref{eqn:Grashof} and that $a_3 = a_{\min}$. Then the configuration space $\Qor^{\mathbb C}(a)$ is isomorphic to a one-point compactification of ${\mathbb C}/2\pi{\mathbb Z}$, and the bijection ${\mathbb C}/2\pi{\mathbb Z} \to \Qor^{\mathbb C}(a) \setminus \{\infty\}$ can be established by the parametrization
\begin{gather*}
z_1 = p_1 \sin t, \quad z_2^{-1} = p_2 \sin (t + i\sigma), \\
z_3 = p_3 \sin \left(t + \frac{\pi}2\right), \quad
z_4^{-1} = p_4 \sin \left(t + \frac{\pi}2 + i\sigma\right)
\end{gather*}
where
$$
p_1 = \sqrt{\frac{a_1a_4}{a_2a_3} - 1}
$$
with $p_2$, $p_3$, $p_4$ obtained by cyclically permuting the indices, and
$$
\sigma = \ln \left( \sqrt{\frac{a_2a_4}{a_2a_4 - a_1a_3}} + \sqrt{\frac{a_1a_3}{a_2a_4 - a_1a_3}} \right)
$$
\end{prp}
\subsubsection{The real part of the configuration space}
In Theorems \ref{prp:Param1+3=2+4} and \ref{prp:Param1+2=3+4} we have $p_1, p_2 \in {\mathbb R}$ and $p_3,p_4 \in i{\mathbb R}$. It follows that the variables $z_i$ take real values if and only if $\Re t \equiv \frac{\pi}2 (\mathrm{mod} \pi)$. The parameter change $t = \frac\pi{2} + iu$ leads to a parametrization that makes the real part of $\Qor^{\mathbb C}$ more tangible.
\begin{cor}
Assume that $a_1 + a_3 = a_2 + a_4$ is the unique solution of the equation \eqref{eqn:Grashof} and that $a_3 = a_{\min}$. Then the configuration space $\Qor^{\mathbb C}$ can be parametrized as
\begin{equation}
\begin{aligned}
z_1 &= p_1 \cosh u\\
z_2 &= p_2 \cosh (u+\sigma)\\
z_3 &= ip_3 \sinh u\\
z_4 &= -ip_4 \sinh (u+\sigma)
\end{aligned}
\end{equation}
where
$$
p_1 = \sqrt{\frac{a_1a_4}{a_2a_3} - 1} \quad \text{etc.}
$$
and
$$
\cosh\sigma = \sqrt{\frac{a_2a_4}{a_2a_4 - a_1a_3}}, \quad \sinh\sigma = \sqrt{\frac{a_1a_3}{a_2a_4 - a_1a_3}}
$$
See Figure \ref{fig:RealConic}.
If $a_1 + a_2 = a_3 + a_4$ is the unique solution of \eqref{eqn:Grashof}, and $a_3 = a_{\min}$, then there is a parametrization
\begin{equation}
\begin{aligned}
z_1 &= p_1 \cosh u\\
z_2^{-1} &= p_2 \cosh(u+\sigma)\\
z_3 &= -ip_3 \sinh u\\
z_4^{-1} &= -ip_4 \sinh(u+\sigma)
\end{aligned}
\end{equation}
with $p_i$ and $\sigma$ computed by the same formulas. See Figure \ref{fig:RealConic2}.
\end{cor}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{RealConic.pdf}%
\end{picture}%
\setlength{\unitlength}{3729sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(5219,2203)(-21,-1151)
\put(5086,-871){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Re$}%
}}}}
\put(2746,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$0$}%
}}}}
\put(4051,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sigma$}%
}}}}
\put(1441,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-\sigma$}%
}}}}
\put(136,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-2\sigma$}%
}}}}
\put(2746,-196){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\pi i$}%
}}}}
\put(4006,-196){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sigma+\pi i$}%
}}}}
\put(136,-151){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-2\sigma + \pi i$}%
}}}}
\put(1441,-151){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-\sigma + \pi i$}%
}}}}
\put(2746,749){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2\pi i$}%
}}}}
\put(2476,929){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Im$}%
}}}}
\end{picture}%
\caption{The real part of the configuration space in the case $a_1 + a_3 = a_2 + a_4$.}
\label{fig:RealConic}
\end{figure}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{RealConic2.pdf}%
\end{picture}%
\setlength{\unitlength}{3729sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(5219,2203)(-21,-1151)
\put(5086,-871){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Re$}%
}}}}
\put(2746,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$0$}%
}}}}
\put(4051,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sigma$}%
}}}}
\put(1441,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-\sigma$}%
}}}}
\put(136,-1051){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-2\sigma$}%
}}}}
\put(2746,-196){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\pi i$}%
}}}}
\put(4006,-196){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sigma+\pi i$}%
}}}}
\put(136,-151){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-2\sigma + \pi i$}%
}}}}
\put(1441,-151){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$-\sigma + \pi i$}%
}}}}
\put(2746,749){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2\pi i$}%
}}}}
\put(2476,929){\makebox(0,0)[lb]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Im$}%
}}}}
\end{picture}%
\caption{The real part of the configuration space in the case $a_1 + a_2 = a_3 + a_4$.}
\label{fig:RealConic2}
\end{figure}
\subsection{Elliptic quadrilaterals: parametrization by elliptic functions}
\label{sec:EllParam}
\subsubsection{Special biquadratic equations and elliptic curves}
\label{sec:Modulus}
A biquadratic polynomial is a polynomial of degree $2$ in each of its two variables. Here we will look at a special class of biquadratic polynomials.
\begin{lem}
Let $m \in {\mathbb C} \setminus \{0, 1\}$. Then the equation
\begin{equation}
\label{eqn:Biquad}
u^2 + v^2 = 1 + mu^2v^2
\end{equation}
defines an elliptic curve.
\end{lem}
\begin{proof}
The substitution
\[
x = u,\, y = v(1-mu^2) \Leftrightarrow u = x,\, v = \frac{y}{1-mx^2}
\]
transforms equation \eqref{eqn:Biquad} into
$$
y^2 = (1 - x^2)(1 - m x^2)
$$
If $m \notin \{0,1\}$, then the polynomial on the right hand side has four distinct roots, thus the equation defines an elliptic curve.
\end{proof}
It is possible to transform equation \eqref{eqn:OppX} to the form \eqref{eqn:Biquad} by a substitution $z_1 = p_1 u_1, z_3 = p_3 u_3$. In order to make the formulas more compact we need some handy notations.
Let $a, b, c, d \in {\mathbb R}$. Put
\begin{equation}
\label{eqn:abar}
\begin{aligned}
s := \frac{a+b+c+d}2, \quad \bar a &:= s-a = \frac{-a+b+c+d}2\\
\bar b &:= s-b, \quad \bar c := s-c, \quad \bar d := s-d
\end{aligned}
\end{equation}
We have $a+b = \bar c + \bar d$, $s-a-b = -(s-c-d)$ etc. Other, less obvious identities, are collected in the following lemma.
\begin{lem}
\label{lem:SIdent}
In the above notations we have
\begin{equation*}
\begin{aligned}
ab - \bar c \bar d &= (s-a-c)(s-b-c)\\
ab - \bar a \bar b &= s(s-c-d)\\
abcd - \bar a \bar b \bar c \bar d &= s(s-a-b)(s-b-c)(s-a-c)
\end{aligned}
\end{equation*}
\end{lem}
Now equation \eqref{eqn:OppX} can be rewritten in any of the following two ways:
\begin{subequations}
\begin{equation}
\label{eqn:OppZinS}
(s - a_3 - a_4)(s - a_2 - a_4) z_1^2 z_3^2 - \bar a_1 \bar a_4 z_1^2 + \bar a_2 \bar a_3 z_3^2 + s(s - a_2 - a_3) = 0
\end{equation}
\begin{equation}
\label{eqn:OppZinSa}
(a_2a_3 - \bar a_1\bar a_4) z_1^2 z_3^2 - \bar a_1 \bar a_4 z_1^2 + \bar a_2 \bar a_3 z_3^2 + (a_1a_4 - \bar a_1 \bar a_4) = 0
\end{equation}
\end{subequations}
\begin{prp}
\label{prp:QEll}
The substitution $z_1 = p_1 u$, $z_3 = p_3 v$ with
\begin{equation}
\label{eqn:p1p3}
p_1 = \sqrt{\frac{s(s-a_2-a_3)}{\bar a_1 \bar a_4}} = \sqrt{\frac{a_1a_4}{\bar a_1 \bar a_4} - 1}, \quad p_3 = \sqrt{\frac{s(s-a_1-a_4)}{\bar a_2 \bar a_3}} = \sqrt{\frac{a_2a_3}{\bar a_2 \bar a_3} - 1}
\end{equation}
transforms equation \eqref{eqn:OppX} to \eqref{eqn:Biquad} with
\begin{equation}
\label{eqn:m}
m = -\frac{s(s-a_2-a_3)(s-a_2-a_4)(s-a_3-a_4)}{\bar a_1 \bar a_2 \bar a_3 \bar a_4} = 1 - \frac{a_1a_2a_3a_4}{\bar a_1 \bar a_2 \bar a_3 \bar a_4}
\end{equation}
In particular, if \eqref{eqn:Grashof} has no solutions, then $\Qor^{\mathbb C}(a)$ is an elliptic curve.
\end{prp}
\begin{proof}
An equation of the form
\[
Az^2 w^2 + Bz^2 + Cw^2 + D = 0
\]
with $D \ne 0$ can be brought into the form \eqref{eqn:Biquad} by the substitution
\[
z^2 = -\frac{D}{B} u^2, \quad w^2 = -\frac{D}{C} v^2
\]
This yields $m = \frac{AD}{BC}$. Applying this to the equation \eqref{eqn:OppZinS} and taking into account the third identity of Lemma \ref{lem:SIdent} gives us the value of $m$ as stated in the proposition. Note that $B, C \ne 0$ due to \eqref{eqn:QuadIneqB}, and $D \ne 0$, since otherwise $a_1 + a_4 = a_2 + a_3$.
Further, $m \ne 1$ because $a_i \ne 0$, and $m \ne 0$ because this would result in $s = a_i + a_j$ for some $i, j$.
\end{proof}
\subsubsection{Biquadratic equation and Jacobi elliptic functions}
Jacobi's elliptic functions can be defined for any complex value $k \notin \{0, \pm 1\}$ of the Jacobi modulus, \cite{McKM99, AE06}.
\begin{lem}
\label{lem:EllParam}
Let $m \in {\mathbb C} \setminus \{0,1\}$. Then the curve
\begin{equation}
\label{eqn:SnShiftK}
x^2 + y^2 = m x^2 y^2 + 1
\end{equation}
possesses the parametrization
$$
x(t) = \sn (t;k), \qquad y(t) = \sn(t+K;k) \quad \text{with } k = \sqrt{m}
$$
as well as the parametrization
$$
x(t) = \cn (t;k), \qquad y(t) = \cn(t+K;k) \quad \text{with } k = \sqrt{\frac{m}{m-1}}
$$
\end{lem}
\begin{proof}
The $\sn$-parametrization follows from the identity
\[
\sn(t+K;k) = \frac{\cn(t;k)}{\dn(t;k)} = \pm \sqrt{\frac{1-\sn^2(t;k)}{1-k^2\sn^2(t;k)}}
\]
The $\cn$-parametrization follows from the identity
\[
\cn(t+K;k) = -k' \frac{\sn(t;k)}{\dn(t;k)} = \pm k' \sqrt{\frac{1-\cn^2(t;k)}{(k')^2 + k^2\cn^2(t;k)}}
\]
\end{proof}
\begin{rem}
By a parameter change or by symmetry reasons one can obtain alternative parametrizations of \eqref{eqn:SnShiftK}:
\[
x = \sn t, \quad y = \sn(t-K) \quad \text{and} \quad x = \cn t, \quad y = \cn(t-K)
\]
\end{rem}
The value of $m$ in our case is given by \eqref{eqn:m}, so that $m \in (-\infty, 0) \cup (0,1)$. Let us make a case distinction in order to bring the Jacobi parameter $k$ into its usual range $(0,1)$.
\begin{prp}
Let $a_{\min}$ and $a_{\max}$ denote the minimum, respectively the maximum value taken by $a_1, a_2, a_3, a_4$, and let $a_{\min} + a_{\max} \ne s$.
\begin{enumerate}
\item
If $a_{\min} + a_{\max} < s$, then the complexified configuration space $\Qor^{\mathbb C}(a)$ is isomorphic to the quotient of ${\mathbb C}$ by a rectangular lattice, which is the period lattice of the elliptic sine function with the parameter
\[
k = \sqrt{1 - \frac{a_1a_2a_3a_4}{\bar a_1 \bar a_2 \bar a_3 \bar a_4}} \in (0,1)
\]
The tangents of halves of the opposite angles $\varphi_1$ and $\varphi_3$ can be parametrized as
\[
z_1 = p_1 \sn(t;k), \quad z_3 = p_3 \sn(t+K;k)
\]
with $p_1$ and $p_3$ as in \eqref{eqn:p1p3}.
\item
If $a_{\min} + a_{\max} > s$, then the complexified configuration space $\Qor^{\mathbb C}(a)$ is isomorphic to the quotient of $C$ by a rhombic lattice, which is the period lattice of the elliptic cosine function with the parameter
\[
k = \sqrt{1 - \frac{\bar a_1 \bar a_2 \bar a_3 \bar a_4}{a_1a_2a_3a_4}} \in (0,1)
\]
The tangents of halves of the opposite angles $\varphi_1$ and $\varphi_3$ can be parametrized as
\[
z_1 = p_1 \cn(t;k), \quad z_3 = p_3 \cn(t+K;k)
\]
with $p_1$ and $p_3$ as in \eqref{eqn:p1p3}.
\end{enumerate}
\end{prp}
\begin{proof}
This is a direct consequence of Proposition \ref{prp:QEll} and Lemma \ref{lem:EllParam}. The only things to observe are that $a_{\min} + a_{\max} \ne s$ if and only if \eqref{eqn:Grashof} has no solutions, and
\begin{multline*}
a_{\min} + a_{\max} < s \Leftrightarrow s(s-a_1-a_2)(s-a_2-a_3)(s-a_3-a_1) < 0 \\
\Leftrightarrow a_1a_2a_3a_4 < \bar a_1 \bar a_2 \bar a_3 \bar a_4
\end{multline*}
\end{proof}
\subsubsection{Shift parametrizations of more general biquadratic equations}
A general symmetric biquadratic equation can be parametrized by a shift of an elliptic function: $x = f(t), y = f(t+\tau)$. This is essentially due to Euler \cite{Eul68} who used biquadratic equations to integrate the relation
$$
\frac{dx}{\sqrt{P(x)}} + \frac{dy}{\sqrt{P(y)}} = 0, \quad \deg P = 4
$$
Euler's argument in a special case is reproduced in \cite{Mar66}.
The following lemma is a special case that we need in order to parametrize equation \eqref{eqn:AdjX}.
\begin{lem}
\label{lem:SnShift}
The curve
\begin{equation}
\label{eqn:SnShift}
x^2 + y^2 = A x^2 y^2 + 2B xy + C
\end{equation}
possesses the parametrization
$$
x(t) = \sn(t;k), \qquad y(t) = \sn(t+\tau;k)
$$
if $k \in (0,1)$ and $\tau \in {\mathbb C}$ satisfy
\begin{equation}
\label{eqn:SnSubst}A = k^2 \sn^2\tau, \quad B = \cn\tau \dn\tau, \quad C = \sn^2\tau
\end{equation}
and the parametrization
$$
x(t) = \cn(t), \qquad y(t) = \cn(t+\tau)
$$
if $k \in (0,1)$ and $\tau \in {\mathbb C}$ satisfy
\begin{equation}
\label{eqn:CnSubst}
A = -k^2 \frac{\sn^2\tau}{\dn^2\tau}, \quad B = \frac{\cn\tau}{\dn^2\tau}, \quad C = (k')^2 \frac{\sn^2\tau}{\dn^2\tau}
\end{equation}
Here $k' = \sqrt{1-k^2}$ is the conjugate modulus.
\end{lem}
\begin{proof}
The addition law
$$
\sn(t+\tau) = \frac{\sn t \cn\tau \dn\tau + \cn t \dn t \sn\tau}{1 - k^2 \sn^2 t \sn^2\tau}
$$
can be rewritten as
$$
(1 - k^2 \sn^2 t \sn^2\tau) \sn(t+\tau) - \cn\tau \dn\tau \sn t = \sqrt{(1 - \sn^2 t)(1 - k^2 \sn^2 t)}
$$
By taking the squares of both parts and substituting $x = \sn t$, $y = \sn(t+\tau)$ we obtain
\begin{multline*}
k^4 \sn^4\tau \ x^4y^2 - k^2 \sn^2\tau \ x^4 - 2k^2 \sn^2\tau \ x^2y^2 + 2k^2 \sn^2\tau \cn\tau \dn\tau \ x^3y \\
+ (1 + k^2 \sn^4\tau) x^2 - 2 \cn\tau \dn\tau \ xy + y^2 - \sn^2 \tau = 0
\end{multline*}
The same equation holds with $x$ and $y$ exchanged. After antisymmetrizing and dividing by $k^2 \sn^2\tau (x^2 - y^2)$ we obtain the equation \eqref{eqn:SnShift}.
In the $\cn$ case the proof is similar.
\end{proof}
A biquadratic equation \eqref{eqn:SnShift} can be viewed as an implicit form of addition formulas for $\sn$ or $\cn$.
\subsubsection{Parametrization theorems}
\label{sec:ParamThm}
We are ready to derive a simultaneous parametrization of all four angles of a quadrilateral with fixed side lengths. With the help of notation \eqref{eqn:abar} equation
\eqref{eqn:AdjX} can be rewritten as
\begin{multline}
\label{eqn:AdjZinS}
\bar a_1 (s - a_1 - a_3) z_1^2 z_2^2 + \bar a_4 (s - a_3 - a_4) z_1^2 + \bar a_2 (s - a_2 - a_3) z_2^2 \\
- 2 a_2 a_4 z_1 z_2 + s \bar a_3 = 0
\end{multline}
Cyclic shifts of indices produce equations for other pairs of adjacent angles.
\begin{prp}
\label{prp:SnEuc}
Let $a = (a_1, a_2, a_3, a_4)$ be such that $a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0$ for all choices of the signs. Define
\[
p_1 = \sqrt{\frac{s(s-a_2-a_3)}{\bar a_1 \bar a_4}} = \sqrt{\frac{a_1 a_4}{\bar a_1 \bar a_4} - 1} \in {\mathbb R}_+ \cup i{\mathbb R}_+
\]
and similarly $p_2$, $p_3$, and $p_4$ through a cyclic shift of indices.
The configuration space $\Qor^{\mathbb C}(a)$ has one of the following parametrizations in terms of the tangents of the half-angles of the quadrilateral.
\begin{enumerate}
\item
If $a_{\min} + a_{\max} < s$ and $a_3 = a_{\min}$, then
\[
z_1 = p_1 \sn t, \quad z_2 = p_2 \sn(t+\tau), \quad z_3 = p_3 \sn(t-K), \quad z_4 = p_4 \sn(t+K+\tau)
\]
where the modulus of $\sn$ is
$$
k = \sqrt{1 - \frac{a_1a_2a_3a_4}{\bar a_1 \bar a_2 \bar a_3 \bar a_4}} \in (0,1),
$$
and the phase shift $\tau$ is determined by
\[
\dn \tau = \sqrt{\frac{a_2a_4}{\bar a_2 \bar a_4}}, \quad \tau \in (0, iK')
\]
\item
If $a_{\min} + a_{\max} > s$ and $a_1 = a_{\max}$, then
\[
z_1 = p_1 \cn t, \quad z_2 = p_2 \cn(t+\tau), \quad z_3 = p_3 \cn(t-K), \quad z_4 = p_4 \cn(t+K+\tau)
\]
where the modulus of $\cn$ is
$$
k = \sqrt{1 - \frac{\bar a_1 \bar a_2 \bar a_3 \bar a_4}{a_1a_2a_3a_4}} \in (0,1),
$$
and the phase shift $\tau$ is determined by
\[
\dn \tau = \sqrt{\frac{\bar a_2 \bar a_4}{a_2a_4}}, \quad \tau \in (0, iK')
\]
\end{enumerate}
\end{prp}
\begin{proof}
Substituting $z_1 = p_1 u_1$ and $z_2 = p_2 u_2$ and using the identities from Lemma \ref{lem:SIdent}, we transform \eqref{eqn:AdjZinS} to
\begin{equation}
\label{eqn:SnPar}
u_1^2 + u_2^2 = \frac{\bar a_2 \bar a_4 - a_2 a_4}{\bar a_2 \bar a_4} u_1^2 u_2^2 - 2 \frac{a_2 a_4}{\sqrt{\bar a_2 \bar a_4}\sqrt{a_2a_4 - \bar a_1 \bar a_3}} u_1 u_2 - \frac{\bar a_1 \bar a_3}{a_2 a_4 - \bar a_1 \bar a_3}
\end{equation}
The range of $\dn$ on $(0, iK')$ is $(1, +\infty)$. Due to Lemma \ref{lem:SIdent} we have $a_2a_4 > \bar a_2 \bar a_4$, thus there exists $\tau \in (0, iK')$ with $\dn \tau = \sqrt{\frac{a_2a_4}{\bar a_2 \bar a_4}}$. From this we compute
\[
\sn^2 \tau = - \frac{\bar a_1 \bar a_3}{a_2 a_4 - \bar a_1 \bar a_3}, \quad \cn^2 \tau = \frac{a_2a_4}{a_2 a_4 - \bar a_1 \bar a_3}
\]
Besides, $\cn \tau > 0$ for $\tau \in (0, iK')$. It follows that the coefficients of \eqref{eqn:SnPar} coincide with the coefficients of \eqref{eqn:SnShift} for given values of $k$ and $\tau$. A similar check can be made for other pairs of adjacent angles (where one should be careful with extraction of square roots of negative numbers), which finishes the proof of the $sn$-parametrization.
In the case $a_{\min} + a_{\max} > s$ use the second part of Lemma \ref{lem:SnShift}.
\end{proof}
\subsubsection{The real part of the configuration space}
\label{sec:EllReal}
Under the assumptions of Theorem \ref{prp:SnEuc} we have $p_1, p_2 \in {\mathbb R}_+$, $p_3, p_4 \in i{\mathbb R}_+$. It follows that all of the variables $z_i$ take real values if and only if
\begin{align*}
\Re t \equiv K (\mathrm{mod} 2K), &\text{ for the } \sn\text{-parametrization}\\
\Re t \equiv 0 (\mathrm{mod} 2K), &\text{ for the } \cn\text{-parametrization}
\end{align*}
As in the trigonometric case, we make a variable change so that real values of parameter yield real angles $\varphi_i$.
In the $\sn$-case put $t = iu + K$, and in the $\cn$-case put $t = iu$. Jacobi's imaginary transformations \cite[Chapter 2]{AE06} allow to rewrite the parametrizations of Theorem \ref{prp:SnEuc} in terms of elliptic functions with the conjugate modulus.
\begin{prp}
\label{prp:RealEll}
\begin{enumerate}
\item
Let $a_{\min} + a_{\max} < s$, and assume that $a_3 = a_{\min}$. Then the configuration space $\Qor^{\mathbb C}$ can be parametrized as
\begin{gather*}
z_1 = \frac{p_1}{\dn(u;k')}, \quad z_2 = \frac{p_2}{\dn(u+\sigma;k')}\\
z_3 = \frac{p_3}{\dn(u+iK;k')} = ip_3 \frac{\sn(u;k')}{\cn(u;k'}\\
z_4 = \frac{p_4}{\dn(u-iK+\sigma;k')} = -ip_4 \frac{\sn(u+\sigma;k')}{\cn(u+\sigma;k')}
\end{gather*}
where
$$
k' = \sqrt{\frac{a_1a_2a_3a_4}{\bar a_1 \bar a_2 \bar a_3 \bar a_4}} \in (0,1),
$$
the amplitudes $p_i$ are given by
\[
p_1 = \sqrt{\frac{s(s-a_2-a_3)}{\bar a_1 \bar a_4}} = \sqrt{\frac{a_1 a_4}{\bar a_1 \bar a_4} - 1}, \quad \text{etc.}
\]
and the phase shift $\sigma$ is determined by
\[
\sn(\sigma;k') = \sqrt{\frac{\bar a_1 \bar a_3}{a_2 a_4}}, \quad \sigma \in (0, K')
\]
\item
Let $a_{\min} + a_{\max} > s$, and assume that $a_1 = a_{\max}$. Then the configuration space $\Qor^{\mathbb C}$ can be parametrized as
\begin{gather*}
z_1 = \frac{p_1}{\cn(u;k')}, \quad z_2 = \frac{p_2}{\cn(u+\sigma;k')}\\
z_3 = \frac{p_3}{\cn(u+iK;k')} = ik'p_3 \frac{\sn(u;k')}{\dn(u;k')}\\
z_4 = \frac{p_4}{\cn(u-iK+\sigma)} = -ik'p_4 \frac{\sn(u+\sigma;k')}{\dn(u+\sigma;k')}
\end{gather*}
where
$$
k' = \sqrt{\frac{\bar a_1 \bar a_2 \bar a_3 \bar a_4}{a_1a_2a_3a_4}} \in (0,1),
$$
the amplitudes $p_i$ are given by
\[
p_1 = \sqrt{\frac{s(s-a_2-a_3)}{\bar a_1 \bar a_4}} = \sqrt{\frac{a_1 a_4}{\bar a_1 \bar a_4} - 1}, \quad \text{etc.}
\]
and the phase shift $\sigma$ is determined by
\[
\sn(\sigma;k') = \sqrt{\frac{a_1 a_3}{\bar a_2 \bar a_4}}, \quad \sigma \in (0, K')
\]
\end{enumerate}
\end{prp}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{RealSn.pdf}%
\end{picture}%
\setlength{\unitlength}{3315sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(6057,2430)(-284,-1333)
\put( 46,974){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Im$}%
}}}}
\put(1576,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$K'-\sigma$}%
}}}}
\put(5671,-826){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Re$}%
}}}}
\put(-269,794){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$4iK$}%
}}}}
\put(-269,-106){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2iK$}%
}}}}
\put( 46,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$0$}%
}}}}
\put(2701,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$K'$}%
}}}}
\put(4276,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2K'-\sigma$}%
}}}}
\put(5401,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2K'$}%
}}}}
\end{picture}%
\caption{The real part of the configuration space in the case $a_{\min} + a_{\max} < s$.}
\label{fig:RealSn}
\end{figure}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{RealCn.pdf}%
\end{picture}%
\setlength{\unitlength}{3315sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(6073,2880)(-210,-1873)
\put(-89,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$0$}%
}}}}
\put(5356,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$4K'$}%
}}}}
\put(541,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$K'-\sigma$}%
}}}}
\put(1351,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$K'$}%
}}}}
\put(1846,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2K'-\sigma$}%
}}}}
\put(2611,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2K'$}%
}}}}
\put(4546,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$4K' - \sigma$}%
}}}}
\put(3196,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$3K'-\sigma$}%
}}}}
\put(3961,-646){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$3K'$}%
}}}}
\put(5761,-466){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Re$}%
}}}}
\put( 46,884){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\Im$}%
}}}}
\put(2746,839){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$2K'+2iK$}%
}}}}
\end{picture}%
\caption{The real part of the configuration space in the case $a_{\min} + a_{\max} > s$.}
\label{fig:RealCn}
\end{figure}
The shapes of quadrilaterals corresponding to special values of parameter $t$ are shown on Figures \ref{fig:RealSn} and \ref{fig:RealCn}. Note that the real part consists of one component (a quadrilateral can be transformed to its mirror image by a continuous deformation) if $a_{\min} + a_{\max} > s$, and of two components (no continuous deformation between a quadrilateral and its mirror image) if $a_{\min} + a_{\max} < s$.
Since $z_i^{-1} = \cot\frac{\varphi_i}2$, Proposition \ref{prp:RealEll} implies Theorem \ref{thm:Param}.
\section{The configuration spaces of non-oriented quadrilaterals}
\label{sec:NonOr}
\subsection{Double covers}
\label{sec:DoubCov}
For $a = (a_1, a_2, a_3, a_4)$ satisfying the conditions \eqref{eqn:QuadIneqA} and \eqref{eqn:QuadIneqB}, let $Q(a)$ denote the space of congruence classes of quadrilaterals in ${\mathbb R}^2$ with side lengths $a$ in this cyclic order. This time, orientation-reversing congruences are allowed, so that to almost every element of $Q(a)$ there correspond two elements of $Q_{or}(a)$. Exceptions are quadrilaterals with collinear vertices.
If $a$ is elliptic, that is $a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0$ for every choice of signs, then there are no exceptional quadrilaterals and the natural map $Q_{or}(a) \to Q(a)$ is a double cover (disconnencted for non-Grashof quadrilaterals). The equivalence relation on $Q_{or}(a)$ looks very simple in terms of the angles of a quadrilateral:
\[
(\varphi_1, \varphi_2, \varphi_3, \varphi_4) \sim (-\varphi_1, -\varphi_2, -\varphi_3, -\varphi_4)
\]
and has the same form in the variables $z_i = \tan \frac{\varphi}2$. By extending this to $\Qor^{\mathbb C}(a)$ we define the complexified configuration space of non-oriented quadrilaterals
\[
Q^{{\mathbb C}}(a) := \Qor^{\mathbb C}(a)/(z_1, z_2, z_3, z_4) \sim (-z_1, -z_2, -z_3, -z_4)
\]
The quotient map
\[
\Qor^{\mathbb C}(a) \to Q^{{\mathbb C}}(a)
\]
is a double cover between elliptic curves.
In terms of the holomorphic parameter $t$ of Proposition \ref{prp:SnEuc} we have $t \sim t+2K$ both in the $\sn$ and in the $\cn$ case. This identifies $Q(a)$ with ${\mathbb C}/2K{\mathbb Z} + 2K'iZ$ and proves the first part of Theorem \ref{thm:NonOr}.
\subsection{Diagonal lengths as coordinates on the configuration space}
\label{sec:DiagLengths}
The cosines $\cos \varphi_i$ are meromorphic functions on the elliptic curve $Q^{{\mathbb C}}(a)$. On the other hand, by the cosine law they are linearly related to the squares of diagonal lengths. Since a quadrilateral with fixed edge lengths is uniquely determined up to congruence by the lengths of its diagonals, this provides a natural embedding of $Q^{{\mathbb C}}(a)$ in $({\mathbb C}{\mathrm P}^1)^2$.
\begin{lem}
Let $x$ and $y$ be the lengths of diagonals ($x$ is separating $a_1$ and $a_2$ from $a_3$ and $a_4$), and
\[
u = x^2, \quad v = y^2
\]
Then the configuration space $Q(a)$ is an algebraic curve given by the equation
\begin{equation}
\label{eqn:DiagCoord}
u^2 v + u v^2 + 2d_{11}uv + d_{10}u + d_{01}v + d_{00} = 0, \quad \text{where}
\end{equation}
\begin{equation*}
\begin{aligned}
d_{11} &= -\frac12(a_1^2 + a_2^2 + a_3^2 + a_4^2)\\
d_{10} &= (a_1^2 - a_4^2)(a_2^2 - a_3^2)\\
d_{01} &= (a_1^2 - a_2^2)(a_4^2 - a_3^2)\\
d_{00} &= (a_1^2 - a_2^2 + a_3^2 - a_4^2)(a_1^2 a_3^2 - a_2^2 a_4^2)
\end{aligned}
\end{equation*}
\end{lem}
\begin{proof}
Viewing a planar quadrilateral as a tetrahedron of zero volume, we equate to zero the Gram determinant of the vectors (two sides and a diagonal) issued from one of its vertices.
\[
\det
\begin{pmatrix}
a_1^2 & \frac{a_1^2 + x^2 - a_2^2}2 & \frac{a_1^2 + a_4^2 - y^2}2\\
\frac{a_1^2 + x^2 - a_2^2}2 & x^2 & \frac{x^2 + a_4^2 - a_3^2}2\\
\frac{a_1^2 + a_4^2 - y^2}2 & \frac{x^2 + a_4^2 - a_3^2}2 & a_4^2
\end{pmatrix}
= 0
\]
A simple computation leads to the equation in the lemma.
\end{proof}
\subsection{Conjugate quadrilaterals}
Recall the following notations.
\begin{dfn}
For $a \in {\mathbb R}^4$ the \emph{conjugate quadruple} $\bar a \in {\mathbb R}^4$ is defined as
$$
\begin{aligned}
\bar a_1 &:= s-a_1 = \frac{-a_1 + a_2 + a_3 + a_4}2, \quad &\bar a_2 &:= s-a_2 = \frac{a_1 - a_2 + a_3 + a_4}2,\\
\bar a_3 &:= s-a_3 = \frac{a_1 + a_2 - a_3 + a_4}2, \quad &\bar a_4 &:= s-a_4 = \frac{a_1 + a_2 + a_3 - a_4}2
\end{aligned}
$$
\end{dfn}
\begin{lem}
The map $a \mapsto \bar a$ is an involution on the space of quadruples $a \in {\mathbb R}^4$ satisfying inequalities \eqref{eqn:QuadIneqA} and \eqref{eqn:QuadIneqB}.
\end{lem}
\begin{proof}
We have $\bar{\bar a} = a$ since $\sum \bar a_i = \sum a_i$. The quadrangle inequality $s > a_i$ is equivalent to the positivity of the length $\bar a_i > 0$.
\end{proof}
Thus a quadruple conjugate to the edge lengths of a quadrilateral can itself be used as edge lengths of another quadrilateral.
Note that $\bar a = a$ if and only if all $a_i$ are equal, and $\bar a$ coincides with $a$ up to the dihedral group action on $\{1, 2, 3, 4\}$ if and only if $a_{\min} + a_{\max} = s$. Thus for $a_1 \pm a_2 \pm a_3 \pm a_4 \ne 0$ quadrilaterals in $Q(\bar a)$ are not congruent to quadrilaterals in $Q(a)$, even if we ignore the marking of sides.
The more surprising is the second part of Theorem \ref{thm:NonOr}: for every quadrilateral with side lengths $a$ and diagonal lengths $x$ and $y$ there is a quadrilateral with side lengths $\bar a$ and diagonal lengths $x$ and $y$.
For this, it suffices to show that the coefficients $d_{ij}$ in \eqref{eqn:DiagCoord} don't change if we replace $a$ by $\bar a$. This follows from $a+b = \bar c + \bar d$, $a-b = \bar b - \bar a$, which implies that the coefficients $d_{01}$ and $d_{10}$ don't change and from Lemma \ref{lem:SIdent2} below, which takes care of $d_{11}$ and $d_{00}$.
\begin{lem}
\label{lem:SIdent2}
Let
\[
s = \frac{a+b+c+d}2, \quad \bar a = s-a, \quad \ldots
\]
Then we have
\begin{gather*}
a^2 + b^2 + c^2 + d^2 = \bar a^2 + \bar b^2 + \bar c^2 + \bar d^2\\
ab + cd = \bar a \bar b + \bar c \bar d\\
ab - cd = \frac12 (\bar c^2 + \bar d^2 - \bar a^2 - \bar b^2)
\end{gather*}
\end{lem}
It is easy to see that
\[
a_{\min} + a_{\max} < s \Leftrightarrow \bar a_{\min} + \bar a_{\max} > \bar s = s
\]
Thus $\Qor^{\mathbb C}(a)$ is a quotient of ${\mathbb C}$ by a rectangular lattice if and only if $\Qor^{\mathbb C}(\bar a)$ is a quotient of ${\mathbb C}$ by a rhombic lattice. Thus the orientation forgetting covers over $Q^{{\mathbb C}}(a) \cong Q^{{\mathbb C}}(\bar a)$ look as shown on Figure \ref{fig:Covers}, and Theorem \ref{thm:NonOr} is proved.
\section{Periodic quadrilaterals}
\label{sec:Periodic}
\subsection{Folding of conic quadrilaterals}
\begin{prp}
For a quadrilateral of conic type, the composition of foldings
\[
F_3 \circ F_4 \colon \Qor^{\mathbb C}(a) \to \Qor^{\mathbb C}(a)
\]
has a unique fixed point, namely the ``flattened'' quadrilateral with collinear vertices.
For any initial position $q \in \Qor^{\mathbb C}(a)$, the sequence $q_n = (F_3 \circ F_4)^n(q)$ converges to the flattened quadrilateral, with the angles of $q_n$ tending to $0$ or $\pi$ exponentially.
\end{prp}
\begin{proof}
Proposition \ref{prp:Param1+3=2+4} implies that in terms of the parameter $t$ the folding maps take the form
\[
F_3(t) = \pi - t, \quad F_4(t) = \pi - 2i\sigma - t
\]
so that $(F_3 \circ F_4)(t) = t + 2i\sigma$. Thus
\[
(F_3 \circ F_4)^n(t) \to \infty
\]
and $\infty$ is a unique fixed point of $F_4 \circ F_3$. The exponential decay of the angles follows from
\[
z_1((F_3 \circ F_4)^n(t)) = p_1 \sin(t + 2ni\sigma) \sim e^{2n\sigma}
\]
and $z_1 = \tan \frac{\varphi_1}2 \sim \frac2{\pi - \varphi_1}$ as $\varphi_1 \to \pi$.
\end{proof}
\subsection{Folding of elliptic quadrilaterals}
\begin{proof}[Proof of the Darboux porism]
Proposition \ref{prp:SnEuc} implies that the folding maps $F_i \colon \Qor^{\mathbb C}(a) \to \Qor^{\mathbb C}(a)$ are involutions of the form $t \mapsto t_i - t$. Thus the composition $F_3 \circ F_4$ is a translation $t \mapsto t + (t_3-t_4)$. The $n$-th iteration $(F_3 \circ F_4)^n$ is a translation by $n(t_3-t_4)$. If it has a fixed point, then $n(t_3-t_4)$ belongs to the lattice of periods, and consequently $(F_3 \circ F_4)^n = \mathrm{id}$.
\end{proof}
The definition of $n$-periodicity given in the Introduction depends on the choice of an ordered pair of adjacent vertices. Let us clarify this dependence.
\begin{lem}
If the map $F_3 \circ F_4 \colon \Qor^{\mathbb C}(a) \to \Qor^{\mathbb C}(a)$ has order $n$, then the maps $F_4 \circ F_3$, $F_1 \circ F_2$, and $F_2 \circ F_1$ also have order $n$. The map $F_2 \circ F_3$ also has a finite order, but possibly different from $n$.
\end{lem}
\begin{proof}
The maps $F_4 \circ F_3$ and $F_3 \circ F_4$ have the same order because they are conjugate to each other. Proposition \ref{prp:SnEuc} implies that the maps $F_3 \circ F_4$ and $F_1 \circ F_2$ are translations by the same vector (modulo the lattice of periods), therefore have the same order. Finally, the same theorem shows that
\[
F_3 \circ F_4(t) = t + 2\tau, \quad F_2 \circ F_3(t) = t - 2K - 2\tau
\]
Thus one translation is commensurable with the lattice of periods if and only if the other is.
\end{proof}
\begin{dfn}
A quadrilateral is called \emph{$n$-periodic}, if both $F_3 \circ F_4$ and $F_2 \circ F_3$ have order $n$. It is called \emph{$(n,m)$-periodic}, if one of the maps $F_3 \circ F_4$ and $F_2 \circ F_3$ has order $n$, and the other has order $m$.
\end{dfn}
On the other hand, the folding period \emph{modulo orientation}, that is the order of the natural map $\hat F_3 \circ \hat F_4 \colon Q^{{\mathbb C}}(a) \to Q^{{\mathbb C}}(a)$ does not depend on the choice of a pair of adjacent vertices because $\hat F_1 = \hat F_3$ and $\hat F_2 = \hat F_4$.
\begin{prp}
\label{prp:Period}
Let $\sigma \in (0,K')$ be the basic shift as defined in Theorem \ref{prp:RealEll}. Assume that $\sigma = \frac{m}{n}K'$ with $m$ and $n$ coprime. Then all quadrilaterals with side lengths $(a_1,a_2,a_3,a_4)$ are $n$-periodic up to symmetry. Besides,
\begin{enumerate}
\item in the $\sn$-case
\[
\begin{aligned}
n \text{ is even} &\Rightarrow \text{the period length is }n\\
n \text{ is odd} &\Rightarrow \text{the period lengths are }(n,2n)
\end{aligned}
\]
\item in the $\cn$-case
\[
\begin{aligned}
n \text{ is even} &\Rightarrow \text{the period length is }2n\\
n \text{ is odd} &\Rightarrow \text{the period lengths are }(n,2n)
\end{aligned}
\]
\end{enumerate}
\end{prp}
\begin{proof}
Let us use the parameter change made in Section \ref{sec:EllReal}, see Figures \ref{fig:RealSn} and \ref{fig:RealCn} for illustration. In terms of the real parameter $u$ the two shifts performed by different pairs of adjacent foldings are
\[
2\sigma \text{ and } 2iK - 2\sigma
\]
The $n$-fold iteration results in the shifts by
\[
n \cdot 2\sigma = 2mK', \quad n(2iK - 2\sigma) = 2niK - 2mK'
\]
These are the smallest multiples of $2\sigma$ that belong to the lattice $2K'{\mathbb Z} + 2iK{\mathbb Z}$, which is the period lattice for $Q^{{\mathbb C}}(a)$, see Section \ref{sec:DoubCov}.
More details are needed to determine the period on $\Qor^{\mathbb C}(a)$.
If $n$ is even, then $m$ is odd. Thus both of the above shifts are periods of $\dn$ and half-periods of $\cn$.
If $n$ is odd, then the first shift is a period of $\dn$, but the second only a half-period. If $m$ is also odd, then the first one is a half-period for $\cn$, and the second a period; vice versa if $m$ is even.
\end{proof}
\subsection{Quadrilaterals with small periods}
\subsubsection{Period $2$: orthodiagonal quadrilaterals}
The composition of foldings $F_3 \circ F_4$ has order $2$ if and only if $F_3$ and $F_4$ commute. This happens if and only if the third and the fourth vertices stay on their respective diagonals after the folding. This, in turn, is equivalent to diagonals being orthogonal. Pythagorean theorem implies that diagonals of a quadrilateral are orthogonal if and only if $a_1^2 + a_3^2 = a_2^2 + a_4^2$. This proves the period $2$ case of the Darboux porism without using elliptic functions.
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{Orthodiag.pdf}%
\end{picture}%
\setlength{\unitlength}{3108sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(3855,2338)(-239,-1331)
\put(3601,-286){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$A$}%
}}}}
\put(1126,884){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$B$}%
}}}}
\put(-224,-331){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$C$}%
}}}}
\put(1081,-1276){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$D$}%
}}}}
\end{picture}%
\caption{Folding of an orthodiagonal quadrilateral.}
\end{figure}
\begin{prp}
\label{prp:Period2}
A quadrilateral is $2$-periodic if and only if its diagonal are orthogonal to each other. This property depends only on the side lengths of the quadrilateral, and is equivalent to
\begin{equation}
a_1^2 + a_3^2 = a_2^2 + a_4^2
\end{equation}
\end{prp}
Note that in the orthodiagonal case equation \eqref{eqn:AdjX} can be rewritten as
\[
(A_1 z_1 + B_1 z_1^{-1})(A_2 z_2 + B_2z_2^{-1}) = 1
\]
which also makes the commutation of foldings apparent.
\subsubsection{Period $4$}
Introduce the notation
\[
A = a_1a_3, \quad B = a_2 a_4, \quad \bar A = \bar a_1 \bar a_3, \quad \bar B = \bar a_2 \bar a_4
\]
Then by Theorem \ref{prp:RealEll} we have
\begin{subequations}
\begin{equation}
\label{eqn:DataSn}
(k')^2 = \frac{AB}{\bar A \bar B}, \quad \sn^2 \sigma = \frac{\bar A}B, \quad \cn^2\sigma = 1 - \frac{\bar A}B, \quad \dn^2 \sigma = 1 - \frac{A}{\bar B}
\end{equation}
\begin{equation}
\label{eqn:DataCn}
(k')^2 = \frac{\bar A \bar B}{AB}, \quad \sn^2 \sigma = \frac{A}{\bar B}, \quad \cn^2\sigma = 1 - \frac{A}{\bar B}, \quad \dn^2 \sigma = 1 - \frac{\bar A}{B}
\end{equation}
\end{subequations}
in the $\sn$- and in the $\cn$-case, respectively. (All elliptic functions have elliptic parameter $k'$.) Note that the numbers $A, B, \bar A, \bar B$ satisfy the identities
\[
A+B = \bar A + \bar B, \quad A\bar A - B \bar B = (A+B)(A-\bar B)
\]
Also we have $a_{\min} + a_{\max} < s \Leftrightarrow AB < \bar A \bar B$.
From \eqref{eqn:DataSn} we compute
\[
\sn 2\sigma = \frac{2\sn\sigma\cn\sigma\dn\sigma}{1-k^2\sn^4\sigma} = \frac{2\sqrt{\bar A \bar B}}{\bar A + \bar B}
\]
This gives another proof of Proposition \ref{prp:Period2}, since $\sigma \in (0,K')$ equals $\frac{K'}2$ if and only if $\sn(2\sigma;k') = 1$, which is equivalent to
\[
2 \sqrt{\bar A \bar B} = \bar A + \bar B
\]
that is, to $\bar A = \bar B$, which by Lemma \ref{lem:SIdent2} is equivalent to $a_1^2 + a_3^2 = a_2^2 + a_4^2$.
\begin{prp}
A quadrilateral is $4$-periodic if and only if one of the following takes place:
\begin{enumerate}
\item
$a_{\min} + a_{\max} > s$ and $a_1a_3 = a_2a_4$;
\item
$a_{\min} + a_{\max} < s$ and
\[
(a_1a_3 - a_2a_4)(a_1a_3 + a_2a_4) = \pm 2 \sqrt{\bar a_1 \bar a_2 \bar a_3 \bar a_4}(a_1^2 + a_3^2 - a_2^2 - a_4^2)
\]
\end{enumerate}
\end{prp}
\begin{proof}
According to Proposition \ref{prp:Period}, a quadrilateral is $4$-periodic if either $a_{\min} + a_{\max} < s$ and $\sigma \in \{\frac{K'}4, \frac{3K'}4\}$ or $a_{\min} + a_{\max} > s$ and $\sigma = \frac{K'}2$.
In the latter case by symmetry with the $\sn$-case we have
\[
\sigma = \frac{K'}2 \Leftrightarrow \frac{2\sqrt{AB}}{A + B} = 1 \Leftrightarrow A = B
\]
In the former case we compute
\[
\sn 2\sigma = \frac{2\sqrt{\bar A \bar B}}{\bar A + \bar B} \Rightarrow \cn 2\sigma = \pm \frac{|\bar A - \bar B|}{A + B}, \quad \dn 2\sigma = \frac{|A-B|}{A+B}
\]
where $\cn 2\sigma > 0$ if and only if $\sigma \in (0, \frac{K'}2)$. Using once again the formula for $\sn 2x$, we obtain
\[
\sn 4\sigma = \pm 4 \frac{\sqrt{\bar A \bar B} |\bar A - \bar B|}{(A+B)|A-B|}
\]
Since $\sigma \in \{\frac{K'}4, \frac{3K'}4\} \Leftrightarrow \sn 4\sigma = \pm 1$, we obtain the second condition.
\end{proof}
\subsubsection{Period 3}
\begin{prp}
\label{prp:Period3}
A quadrilateral is $(3,6)$-periodic if and only if
\[
(a_1a_3 + a_2a_4)^2 \in \{4a_1a_3\bar a_2 \bar a_4, 4a_2a_4\bar a_1 \bar a_3, 4a_1a_3\bar a_1 \bar a_3, 4 a_2a_4\bar a_2\bar a_4\}
\]
\end{prp}
\begin{proof}
Due to Proposition \ref{prp:Period} we have to find the necessary and sufficient conditions for $\sigma \in \{\frac{K'}3, \frac{2K'}3\}$.
For $\sigma \in (0,K')$ we have
\begin{gather*}
\sigma = \frac{K'}3 \Leftrightarrow \sn 2\sigma = \sn(K'-\sigma) = \frac{\cn \sigma}{\dn \sigma}\\
\sigma = \frac{2K'}3 \Leftrightarrow \sn 2\sigma = \sn(2K'-\sigma) = \sn\sigma
\end{gather*}
This leads to the equations $(A+B)^2 = 4\bar A B$ and $(A+B)^2 = 4B\bar B$ in the $\sn$-case, and to $(A+B)^2 = 4A \bar B$ and $(A+B)^2 = 4 A \bar A$ in the $\cn$-case. Since we were making assumptions $a_3 = a_{\min}$ and $a_1 = a_{\max}$ in the $\sn$- and $\cn$-cases respectively, omitting them leads to symmetrizing with respect to $A$ and $B$. Thus all four conditions are valid both in the $\sn$- and in the $\cn$-case.
\end{proof}
\begin{exl}
For $a_1=a_2=a_3=a$ and $a_4=b$ we have $\bar a_1 = \bar a_2 = \bar a_3 = \frac{a+b}2$ and $\bar a_4 = \frac{3a-b}2$, so that
\[
(a_1a_3+a_2a_4)^2 = a^2(a+b)^2 = 4a_1a_3\bar a_1 \bar a_3
\]
and hence every quadrilateral with side lengths $(a,a,a,b)$ is $(3,6)$-periodic. For a geometric proof of this fact, see \cite[Figure 7]{Zvo97}. Note that the quadrilateral is Grashof for $a>b$ and non-Grashof for $a<b$.
\end{exl}
\subsection{The general periodicity condition}
We have $\sigma = \frac{m}{n} K'$ (and thus the quadrilateral is periodic) if and only if
\[
\cn(n\sigma; k') \in \{0, -1, 1\}
\]
It is possible to express $\cn(n\sigma)$ as a rational function of $\cn\sigma$, so that $\sigma = \frac{m}{n}K'$ becomes an algebraic conditions on the side lengths. These rational functions can be computed recursively, with the help of the following formulas that generalize the Chebyshev polynomials.
\[
R_0(t) = 1, \quad R_1(t) = t, \quad R_{n+1}(t) = \frac{2t R_n(t)}{1-(k')^2(1-t^2)(1-R_n^2(t))} - R_{n-1}(t)
\]
Benoist and Hulin \cite{BH04} relate the dynamics of quadrilateral folding to the Poncelet dynamics on the circles of radii $R = a_2a_4$ and $r = \frac12(a_4^2 - a_3^2 + a_2^2 - a_1^2)$ whose centers are at the distance $a = a_1a_3$ from each other. From the known conditions on short trajectories they obtain characterization of quadrilaterals with small folding period. One may then use Cayley's explicit conditions on the periodicity of Poncelet trajectories to obtain the periodicity condition for quadrilateral folding. We choose a more direct approach, following the general principle derived by Griffiths and Harris \cite{GH78} that also leads to a modern proof of Cayley's conditions. This results in our Theorem \ref{thm:Period}, which we prove below.
\begin{thm}[Griffiths--Harris \cite{GH78}]
\label{thm:GH}
Let $E$ be an elliptic curve and $p, q \in E$. For an integer $n \ge 3$ we have $n(p-q) = 0$ with respect to the additive structure on $E$ if and only if the following condition is satisfied.
Choose a meromorphic function $x$ of order $2$ on $E$ such that $x$ has a double pole at $p$ and a zero at $q$. Let $a, b, c \in {\mathbb C}$ be the values of $x$ at its branch points other than $p$, so that $E$ is isomorphic to the Riemann surface of the function
\[
y = \sqrt{(x-a)(x-b)(x-c)} = \sum_{i=0}^\infty A_i x^i
\]
The condition on $p-q$ being of order $n$ is
\begin{align*}
&\begin{vmatrix}
A_2 & A_3 & \ldots & A_{k+1}\\
A_3 & A_4 & \ldots & A_{k+2}\\
\vdots & \vdots & \ddots & \vdots\\
A_{k+1} & A_{k+2} & \ldots & A_{2k}
\end{vmatrix}
= 0, \quad \text{ if } n = 2k+1\\
&\begin{vmatrix}
A_3 & A_4 & \ldots & A_{k+1}\\
A_4 & A_5 & \ldots & A_{k+2}\\
\vdots & \vdots & \ddots & \vdots\\
A_{k+1} & A_{k+2} & \ldots & A_{2k-1}
\end{vmatrix}
= 0, \quad \text{ if } n = 2k
\end{align*}
\end{thm}
Our first goal is to construct a function $x$ as described in the Griffiths-Harris theorem.
Consider the function $u$ equal to the squared length of the diagonal separating the sides $a_1, a_2$ from $a_3, a_4$. At the branch points, $u$ takes the values
\[
(a_1 + a_2)^2, \quad (a_1 - a_2)^2, \quad (a_3 + a_4)^2, \quad (a_3 - a_4)^2
\]
Let $p \in Q(a)$ be such that $u(p) = (a_3 + a_4)^2$. If $a_1+a_2 > a_3+a_4$, then $p$ lies in the real part of $Q(a)$ and corresponds to the quadrilateral in which the sides $a_3$ and $a_4$ are aligned to form a segment of length $a_3 + a_4$. The quadrilateral $p$ is a fixed point for the folding $\hat F_4$, and the folding $\hat F_3$ maps it to the quadrilateral on Figure \ref{fig:pq}, right.
\begin{figure}
\centering
\begin{picture}(0,0)%
\includegraphics{pq.pdf}%
\end{picture}%
\setlength{\unitlength}{4144sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(4334,1528)(328,-658)
\put(2088, 71){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(739, 31){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(3669,-598){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(3943,-97){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(3701,489){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(3397,-125){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(4388, 16){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sqrt{v(q)} = \sqrt{v(p)}$}%
}}}}
\put(1200,-598){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(1594,697){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(1473,-82){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sqrt{v(p)}$}%
}}}}
\put(2880, 26){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\sqrt{u(q)}$}%
}}}}
\end{picture}%
\caption{The quadrilaterals $p$ and $q = F_3 \circ F_4(p)$.}
\label{fig:pq}
\end{figure}
Denote $q = \hat F_3 \circ \hat F_4(p)$. We have $(\hat F_3 \circ \hat F_4)^n = \mathrm{id}$ if and only if $n(p-q) = 0$. The function
\begin{equation}
\label{eqn:x}
x = \frac{u - u(q)}{u - u(p)}
\end{equation}
has a double pole at $p$ and a zero at $q$, and the same branch points as $u$.
Now it remains to compute the values of $x$ at the three remaining branch points:
\[
\frac{(a_1+a_2)^2 - u(q)}{(a_1+a_2)^2 - (a_3 + a_4)^2}, \quad \frac{(a_1-a_2)^2 - u(q)}{(a_1-a_2)^2 - (a_3 + a_4)^2}, \quad \frac{(a_3-a_4)^2 - u(q)}{(a_3-a_4)^2 - (a_3 + a_4)^2}
\]
\begin{lem}
\label{lem:uq}
We have
\begin{gather*}
v(p) = v(q) = \frac{a_3(a_1^2 - a_4^2) + a_4(a_2^2 - a_3^2)}{a_3+a_4}\\
u(q) = \frac{(a_4-a_3)(a_1^2-a_2^2)}{a_3+a_4} + \frac{(a_1^2 - a_2^2 + a_3^2 - a_4^2)(a_1^2a_3^2 - a_2^2a_4^2)}{(a_3+a_4)(a_3(a_1^2 - a_4^2) + a_4(a_2^2 - a_3^2))}
\end{gather*}
\end{lem}
\begin{proof}
By the cosine law we have
\begin{gather*}
v(p) = a_1^2 + a_4^2 + 2a_1a_4 \cos\varphi_1(p)\\
a_2^2 = a_1^2 + (a_3+a_4)^2 + 2 a_1(a_3+a_4) \cos\varphi_1(p)
\end{gather*}
Expressing $\cos\varphi_1(p)$ from the second equation and substituting into the first gives us the formula for $v(p)$.
To compute $u(q)$, rewrite equation \eqref{eqn:DiagCoord} as a quadratic equation in $u$:
\[
v u^2 + (v^2 + 2d_{11} v + d_{10}) u + (d_{01}v + d_{00}) = 0
\]
At $v = v(p)$ this equation has two solutions $u = u(p) = (a_3+a_4)^2$ and $u = u(q)$. Therefore
\[
(a_3+a_4)^2 u(q) = \frac{d_{01} v(p) + d_{00}}{v(p)}
\]
and we arrive at the formula for $u(q)$ stated in the lemma.
\end{proof}
Finally, a lengthy calculation yields the following formulas for the values of $x$ at the branch points:
\[
\frac{(a_1a_3+a_2a_4)^2}{u(p)v(p)}, \quad \frac{(a_1a_3-a_2a_4)^2}{u(p)v(p)}, \quad \frac{(\bar a_1 \bar a_3-\bar a_2 \bar a_4)^2}{u(p)v(p)}
\]
Scaling $x$ by an appropriate functions we obtain a function with the same poles and zeros and with the values
\[
(a_1a_3+a_2a_4)^2, \quad(a_1a_3-a_2a_4)^2, \quad (\bar a_1 \bar a_3-\bar a_2 \bar a_4)^2
\]
at its branch points. This proves Theorem \ref{thm:Period}.
\section{Applications}
\label{sec:App}
\subsection{Ivory's theorem}
\label{sec:Ivory}
The second part of Theorem \ref{thm:NonOr} says that for every quadrilateral with the side lengths $a$ and diagonal lengths $x$ and $y$ there is a quadrilateral with the side lengths $\bar a$ and the same diagonal lengths $x$ and $y$. We call these quadrilaterals \emph{conjugate}.
See Figure \ref{fig:IsomConj} for an example, where we have chosen
\[
a = (10, 5, 6, 3), \quad \bar a = (2, 7, 6, 9)
\]
\begin{figure}
\centering
\begin{picture}(0,0)%
\includegraphics{IsomConj.pdf}%
\end{picture}%
\setlength{\unitlength}{4144sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(4749,1545)(-11,-1156)
\put(991,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(1216, 29){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(361,-466){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(3556,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_3$}%
}}}}
\put(3871,254){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_1$}%
}}}}
\put(3061,-331){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_2$}%
}}}}
\put(811,-556){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$x$}%
}}}}
\put(1396,-556){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$y$}%
}}}}
\put(3331,-601){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$x$}%
}}}}
\put(1981,-421){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(4456,-241){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_4$}%
}}}}
\put(4006,-376){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$y$}%
}}}}
\end{picture}%
\caption{A pair of conjugate quadrilaterals.}
\label{fig:IsomConj}
\end{figure}
Arseniy Akopyan has noted \cite{Akop12} that the existence of a conjugate to each quadrilateral is equivalent to the Ivory theorem.
\begin{thm}[Ivory's Theorem]
Take two ellipses and two hyperbolas from a confocal family of conics. Then the diagonals of any curvilinear quadrilateral cut out by them have equal lengths.
\end{thm}
\begin{proof}
Consider a quadrilateral with vertices at the foci and at the endpoints of one of the diagonals of the given curvilinear quadrilateral. The order of vertices is chosen so that the foci form a pair of opposite vertices. The quadrilateral formed in the same way but using the second diagonal of the curvilinear quadrilateral has side lengths conjugate to those of the first quadrilateral. Since this pair of conjugate quadrilaterals has one diagonal in common (the segment between the foci), the other diagonals have equal lengths. See Figure \ref{fig:Ivory}.
\end{proof}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{Ivory.pdf}%
\end{picture}%
\setlength{\unitlength}{3315sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(5919,2274)(-4556,-298)
\put(991,929){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put( 91,704){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_3$}%
}}}}
\put(-134,1244){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put( 91,974){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(-314,299){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_2$}%
}}}}
\put(766,119){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_1$}%
}}}}
\put(631,1019){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(946,659){\makebox(0,0)[lb]{\smash{{\SetFigFont{7}{8.4}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\bar a_4$}%
}}}}
\end{picture}%
\caption{Ivory's theorem and its proof using conjugate quadrilaterals.}
\label{fig:Ivory}
\end{figure}
With the same argument one can prove the Ivory theorem on the sphere and in the hyperbolic plane, see Section \ref{sec:SphHyp}.
\subsection{Dixon's angle condition in the Burmester mechanism}
The linkage on Figure \ref{fig:Dixon} consists of two small quadrilaterals sharing an edge and of one big quadrilateral having an angle in common with each of the small ones. It is easy to see that a linkage of this form is rigid in general. The following is a necessary condition of its flexibility.
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{Dixon.pdf}%
\end{picture}%
\setlength{\unitlength}{4144sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(4974,1329)(-11,-523)
\put(3378,-138){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\theta_1$}%
}}}}
\put(3139,-408){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\psi_1$}%
}}}}
\put(4204,-357){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_2$}%
}}}}
\put(3665,-334){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\varphi_1$}%
}}}}
\put(4647,-346){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\psi_2$}%
}}}}
\put(4331, -2){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\theta_2$}%
}}}}
\end{picture}%
\caption{An overconstrained, that is generically rigid, linkage.}
\label{fig:Dixon}
\end{figure}
\begin{prp}
If the linkage on Figure \ref{fig:Dixon} is flexible and both small quadrilaterals are of elliptic type, then the angles marked at the left remain either equal modulo $\pi$ or complement each other modulo $\pi$ during the flex:
\[
\theta_1 \pm \theta_2 = n\pi
\]
\end{prp}
\begin{proof}
Forget the big quadrilateral and require only that the horizontal sides of the small quadrilaterals remain aligned. The configuration space $Q_{12}$ of this coupling is a fiber product of the configuration spaces $Q_1$ and $Q_2$ of the small quadrilaterals, see the diagram \eqref{eqn:Z12CD}.
\begin{equation}
\label{eqn:Z12CD}
\xymatrix {
&& Q_{12} \ar[dl] \ar[dr] && \\
& Q_1 \ar[dl] \ar[dr]_{f_1} && Q_2 \ar[dr] \ar[dl]^{f_2} \\
w_1 \in {\mathbb C}{\mathrm P}^1 && z \in{\mathbb C}{\mathrm P}^1 && {\mathbb C}{\mathrm P}^1 \ni w_2
}
\end{equation}
Here $z = z_1 = \tan\frac{\varphi_1}2 = \cot\frac{\varphi_2}2 = z_2^{-1}$ and $w_i = \tan\frac{\psi_i}2$.
If the branch sets of the maps $f_1$ and $f_2$ on the diagram don't coincide, then $Q_{12}$ is a hyperelliptic curve.
On the other hand, if the coupling remains flexible when restrained by the big quadrilateral, then $Q_{12}$ is the configuration space of the big quadrilateral, thus cannot be hyperelliptic.
Branch points of $f_i$ correspond to $\theta_i \in \{0,\pi\}$. Thus for flexibility it is necessary that
\[
\theta_1 \in \{0,\pi\} \Leftrightarrow \theta_2 \in \{0, \pi\}
\]
Applying Lemma \ref{lem:CosLinDep} to both small quadrilaterals we obtain that $\cos \theta_1$ and $\cos \theta_2$ must be linearly dependent. It follows that either $\cos\theta_1 = \cos \theta_2$, that is $\theta_1 - \theta_2 = n\pi$, or $\cos\theta_1 = -\cos\theta_2$, that is $\theta_1 + \theta_2 = n\pi$.
\end{proof}
It turns out that if the linkage on Figure \ref{fig:Dixon} is flexible, then another bar can be added without restraining its motion, see \cite{Wun68}. This results in a linkage on Figure \ref{fig:BurHart}, left, called the Burmester focal mechanism. Of course, points on the sides of the big quadrilateral may also lie on the extensions of these sides. In particular, if one of them lies at infinity, then the trajectory of the central joint is a line orthogonal to the corresponding side. The result is a straight-line mechanism of Hart \cite{Har77,Kem78,Dar79a}, see Figure \ref{fig:BurHart}, right.
\begin{figure}[ht]
\centering
\includegraphics{BurHart.pdf}
\caption{Burmester's linkage and its special case, Hart's straight line mechanism.}
\label{fig:BurHart}
\end{figure}
The problem of characterizing all flexible configurations of the form shown on Figure \ref{fig:BurHart}, left, was studied by Krause \cite{Kra08} using Darboux's parametrization of the configuration space. He found three families: in the first case two opposite small quadrilaterals are parallelograms while the other two are similar; in the second all small quadrilaterals are deltoids and the big one is orthodiagonal; in the third family all quadrilaterals are conic or elliptic and the lengths of 12 segments are subject to 7 polynomial relations.
More examples of overconstrained flexible linkages can be found in \cite{Sta13}.
\subsection{Flexible polyhedra}
\label{sec:FlexPol}
Consider a polyhedral cone made of four faces meeting along four edges. If the edges are viewed as ideal hinges, and the faces are allowed to cross each other, then the configuration space of the cone is that of a spherical quadrilateral obtained by intersecting the cone with a sphere centered at the apex. This allows to study the flexibility of polyhedral surfaces with vertices of degree $4$ by means of elliptic functions.
The only closed polyhedron with all vertices of degree $4$ is the octahedron. Bricard \cite{Bri97} studied the system of biquadratic equations between the tangents of dihedral half-angles and described three classes of flexible (self-intersecting) octahedra. In fact, it suffices to look at the three dihedral angles adjacent to the same face, so Bricard's method was to find when the resultant of two biquadratic polynomials $P_1(z_2, z_3)$ and $P_2(z_1, z_3)$ is divisible by a biquadratic polynomial $P_3(z_1,z_2)$.
Recently, Gaifullin \cite{Gai13} classified flexible cross-polytopes (analogs of the octahedron in higher dimensions) in the spaces of constant curvature. Unexpectedly enough (polyhedra of higher dimensions tend to be ``more rigid''), flexible cross-polytopes exist in all dimensions, besides, in ${\mathbb S}^3$ there is a type that is not present in Bricard's classification for ${\mathbb R}^3$. Gaifullin used parametrization of dihedral angles by Jacobi elliptic functions, similar to the one in Section \ref{sec:EllReal}.
\begin{figure}[ht]
\centering
\includegraphics[width=.8\textwidth]{Koko4.pdf}
\caption{A Kokotsakis polyhedron and the associated spherical linkage.}
\label{fig:Koko4}
\end{figure}
There are plenty of non-closed $3$-dimensional polyhedra with all vertices of degree $4$. The neighborhood of a face of such a polyhedron consists of an $n$-gon surrounded by a ``belt'' of triangles and quadrilaterals, and is called a \emph{Kokotsakis polyhedron}, \cite{Kok33}. See Figure \ref{fig:Koko4}, left, for a polyhedron with a quadrangular base. All polygons are required to remain flat during a deformation. Thus the polyhedron is flexible if and only if the spherical framework with scissors-like joints on Figure \ref{fig:Koko4}, right, deforms in such a way that the two marked angles remain equal.
The configuration space of a Kokotsakis polyhedron is described by a system of $n$ biquadratic equations in $n$ variables. Kokotsakis \cite{Kok33}, Graf and Sauer \cite{SG31}, and Stachel \cite{Sta10} described several types of flexible polyhedra. In \cite{Izm_Koko} a complete classification in the case of a quadrangular base was obtained. Parametrization by elliptic functions is used in an essential way; among the new flexible types there is an irreducible one, where the resultant of $P(z_1,z_2)$ and $Q(z_2,z_3)$ is an irreducible polynomial and thus proportional to the resultant of $R(z_3,z_4)$ and $S(z_4,z_1)$.
\section{Configuration spaces of spherical and hyperbolic quadrilaterals}
\label{sec:SphHyp}
Qualitatively, all results of this article apply for quadrilaterals on the sphere or on the hyperbolic plane. Configuration spaces of spherical quadrilaterals play a role in the study of flexible polyhedra, see Section \ref{sec:FlexPol}.
For spherical quadrilaterals, the admissible side lengths are subject to inequalities
\begin{subequations}
\begin{equation}
\label{eqn:QuadIneqASph}
0 < a_i < \pi \quad \forall i
\end{equation}
\begin{equation}
\label{eqn:QuadIneqBSph}
a_i < a_j + a_k + a_l < a_i + 2\pi \quad \forall i,
\end{equation}
\end{subequations}
while in the hyperbolic case the conditions are the same \eqref{eqn:QuadIneqA}, \eqref{eqn:QuadIneqB} as in the euclidean. Similarly, in place of equation \eqref{eqn:Grashof} in the spherical case we have
\[
a_1 \pm a_2 \pm a_3 \pm a_4 \equiv 0 (\mathrm{mod} 2\pi)
\]
This adds ``antideltoids'' and ``antiisograms'' to the respective cases in Definition \ref{dfn:Types}.
In terms of the tangents of half-angles the configuration space of oriented quadrilaterals is described by equations \eqref{eqn:OppX} and \eqref{eqn:AdjX}, where in the spherical case we have
\begin{equation*}
\begin{aligned}
b_{22} &= \sin\frac{a_1 + a_2 - a_3 - a_4}2 \sin\frac{a_1 - a_2 + a_3 - a_4}2\\
b_{20} &= \sin\frac{a_1 + a_2 + a_3 - a_4}2 \sin\frac{a_1 - a_2 - a_3 - a_4}2\\
b_{02} &= \sin\frac{a_1 - a_2 + a_3 + a_4}2 \sin\frac{a_1 + a_2 - a_3 + a_4}2\\
b_{00} &= \sin\frac{a_1 - a_2 - a_3 + a_4}2 \sin\frac{a_1 + a_2 + a_3 + a_4}2
\end{aligned}
\end{equation*}
and similarly for $c_{ij}$ with $c_{11} = -\sin a_2 \sin a_4$. In the hyperbolic case $\sinh$ takes the place of $\sin$.
Further, orthodiagonal spherical quadrilaterals are characterized by
\[
\cos a_1 \cos a_3 = \cos a_2 \cos a_4
\]
and these are the only $2$-periodic ones.
There are analogs of identities from Lemmas \ref{lem:SIdent} and \ref{lem:SIdent2}. Some of them are collected in the next lemma.
\begin{lem}
\label{lem:SIdentSph}
For any $a, b, c, d$ the following identities hold.
\begin{gather*}
\cos a \cos b + \cos c \cos d = \cos \bar a \cos \bar b + \cos \bar c \cos \bar d\\
\sin a \sin b + \sin c \sin d = \sin \bar a \sin \bar b + \sin \bar c \sin \bar d\\
\cos a \cos b - \cos c \cos d = \sin \bar a \sin \bar b - \sin \bar c \sin \bar d\\
\sin a \sin b \sin c \sin d + \cos a \cos b \cos c \cos d = \sin \bar a \sin \bar b \sin \bar c \sin \bar d + \cos \bar a \cos \bar b \cos \bar c \cos \bar d
\end{gather*}
Here $\bar a = \frac12(-a + b + c + d)$ etc.
The same identities hold when $\sin$ and $\cos$ are replaced by $\sinh$ and $\cosh$.
\end{lem}
Parametrization of the configuration space in terms of Jacobi elliptic functions (or trigonometric functions, for conic quadrilaterals) are derived in the spherical case in \cite{Izm_Koko}.
Equation in terms of diagonal lengths, describing the space of quadrilaterals up to possibly orientation changing congruences undergoes a more significant change.
\begin{lem}
\label{lem:DiagCoordSph}
Let $x$ and $y$ be the lengths of diagonals in a spherical quadrilateral with side lengths $a_1, a_2, a_3, a_4$ (in this cyclic order). Denote
\[
u = \cos x, \quad v = \cos y
\]
Then the configuration space of quadrilaterals up to congruence is an algebraic curve given by the equation
\begin{equation*}
\label{eqn:DiagCoordSph}
u^2 v^2 - u^2 - v^2 + 2d_{11} uv + d_{10}x + d_{01}y + d_{00} = 0, \quad \text{where}
\end{equation*}
\begin{equation*}
\begin{aligned}
d_{11} &= - (\cos a_1 \cos a_3 + \cos a_2 \cos a_4)\\
d_{10} &= 2(\cos a_1 \cos a_2 + \cos a_3 \cos a_4)\\
d_{01} &= 2(\cos a_1 \cos a_4 + \cos a_2 \cos a_3)\\
d_{00} &= 1 - \cos^2 a_1 - \cos^2 a_2 - \cos^2 a_3 - \cos^2 a_4 + (\cos a_1 \cos a_3 - \cos a_2 \cos a_4)^2
\end{aligned}
\end{equation*}
The same holds for the configuration space of hyperbolic quadrilaterals, with $\cos$ and $\sin$ replaced by $\cosh$ and $\sinh$.
\end{lem}
\begin{proof}
Consider the Gram matrix of the vectors from the center of the sphere (or center of the hyperboloid, in the hyperbolic case) to the vertices of the quadrilateral. Its determinant must be equal to zero.
\end{proof}
In the hyperbolic case, a similar equation is derived by considering the Minkowski space model of the hyperbolic plane.
\begin{thm}[Ivory's theorem on the sphere and in the hyperbolic plane]
\label{thm:IvorySph}
Let $A_1$ and $A_2$ be two points on the sphere, respectively in the hyperbolic plane.
Define an ellipse with the foci $A_1$ and $A_2$ as the set of points with a constant sum of geodesic distances to $A_1$ and $A_2$, and a hyperbola as the set of points with a constant absolute value of the difference of these distances.
Then in every curvilinear quadrilateral cut out by a confocal family of ellipses and hyperbolas the diagonals have equal geodesic lengths.
\end{thm}
Note that a spherical hyperbola consists of two spherical ellipses (replace one of the foci by its antipode).
\begin{proof}
Coefficients of the equation \eqref{eqn:DiagCoordSph} don't change when $a_i$ are replaced by $\bar a_i$. For the coefficients $d_{11}$, $d_{10}$, $d_{01}$ this is immediate from Lemma \ref{lem:SIdentSph}, while $d_{00}$ must first be shown to be equal to
\begin{multline*}
-1 + (\sin a_1 \sin a_3 + \sin a_2 \sin a_4)^2\\
- 2 \sin a_1 \sin a_2 \sin a_3 \sin a_4 - 2 \cos a_1 \cos a_2 \cos a_3 \cos a_4
\end{multline*}
(with hyperbolic trigonometric functions in the hyperbolic case).
Now the Ivory's theorem follows by the same argument as in the euclidean case, see Section \ref{sec:Ivory}.
\end{proof}
It is not clear to us whether Theorem \ref{thm:IvorySph} is related to a version of Ivory's theorem for hyperbolic spaces presented in \cite{SW04}.
Finally, note that while in the euclidean case the scaling of the sides
\[
(a_1, a_2, a_3, a_4) \mapsto (\lambda a_1, \lambda a_2, \lambda a_3, \lambda a_4)
\]
doesn't change the configuration space, in the spherical and hyperbolic cases it does.
\section{Open questions}
\begin{prb}
The area $S$ is a meromorphic function on the space of quadrilaterals with fixed side lengths. The inscribed quadrilateral $I(a)$ is a branch point of $A$, and we have $S(I(a)) = \sqrt{\bar a_1 \bar a_2 \bar a_3 \bar a_4}$ (Brahmagupta's formula). What are the other branch points and other critical values of $A$? Is there a nice theory in the spherical and hyperbolic case?
\end{prb}
\begin{prb}
Derive periodicity conditions for foldings of spherical and hyperbolic quadrilaterals.
\end{prb}
\begin{prb}
Study flexible bipyramids (or suspensions) with the help of parametrizations of the configuration spaces of spherical quadrilaterals.
\end{prb}
A bipyramid over a quadrangle is an octahedron, thus this is a generalization of Bricard's problem of flexible octahedra. Flexible bipyramids were studied in \cite{Con74,AC11}. It is not known whether there exist non-trivial flexible bipyramids over the pentagon; in the articles cited examples of flexible bipyramids over the hexagon were found. By cutting off the two apices of a bipyramid one obtains a polyhedron with vertices of degree four. This leads to a system of biquadratic equations, see Section \ref{sec:FlexPol}. Besides this system, there is also a holonomy condition (since the polyhedron is not simply connected).
\begin{prb}
The dihedral group action on $a = (a_1, a_2, a_3, a_4)$ does not change the configuration space $\Qor^{\mathbb C}(a)$. What happens when the cyclic order of sides is changed in one of the following two ways?
\[
a' = (a_1, a_3, a_2, a_4), \quad a'' = (a_1, a_2, a_4, a_3)
\]
\end{prb}
\begin{figure}[ht]
\centering
\begin{picture}(0,0)%
\includegraphics{Permute.pdf}%
\end{picture}%
\setlength{\unitlength}{4144sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(2004,1992)(-11,-1468)
\put(451,-466){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_1$}%
}}}}
\put(181,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put( 91,-61){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(766,389){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\put(1666,-16){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(901,-196){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_2$}%
}}}}
\put(766,-1096){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_4$}%
}}}}
\put(1441,-736){\makebox(0,0)[lb]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$a_3$}%
}}}}
\end{picture}%
\caption{Linking the configuration spaces $\Qor^{\mathbb C}(a)$, $\Qor^{\mathbb C}(a')$, and $\Qor^{\mathbb C}(a'')$.}
\label{fig:Permute}
\end{figure}
The curves $\Qor^{\mathbb C}(a)$, $\Qor^{\mathbb C}(a')$, and $\Qor^{\mathbb C}(a'')$ are isomorphic, as follows from Theorem \ref{thm:Param} or as is indicated on Figure \ref{fig:Permute}. What does change, are the parametrizations. As formulas in Proposition \ref{prp:SnEuc} suggest, the shifts start mixing up with the amplitudes. The configuration space of the linkage on Figure \ref{fig:Permute} seems to be a four-fold covering of each of the spaces $\Qor^{\mathbb C}(a)$, $\Qor^{\mathbb C}(a')$, and $\Qor^{\mathbb C}(a'')$.
|
\section{Introduction}
The analysis of neutrino oscillation experiments
requires having under control all the nuclear
effects, which are inherent to any $\nu$-nucleus
scattering event. These effects play a major role
in modifying the free $\nu$-nucleon cross section,
and a good understanding of them is clearly mandatory
in order to reduce the systematic uncertainty of the models
employed in the experimental analysis.\\
There is strong theoretical evidence \cite{Martini:2009uj,
Nieves:2011pp,Amaro:2010sd} about a significant
contribution from multi-nucleon knockout to the inclusive
Charged Current (CC) cross section around and above the
quasielastic (QE) peak region.\\
There are, at least, three different microscopic models
\cite{Martini:2009uj,Nieves:2011pp,Amaro:2010sd,Amaro:2011aa}
which are based in the relativistic Fermi Gas. But they differ
from each other in several assumptions and different nuclear
ingredients for the interaction. Therefore, it is really
difficult to disentangle the origin of any discrepancy in
the final results.\\
Other models \cite{Lalakulich:2012ac,Sobczyk:2012ms}
are based on the available phase space just
assuming a constant transition matrix element and fitting
it to the experimental cross section.\\
Finally, our goal is to reduce the computational time
needed in this kind of calculations, or at least
to establish what assumptions previously made by other
authors are really good enough in order to estimate
accurately and fast the contribution of these
multi-nucleon processes to the inclusive channel.
\section{2p-2h phase space in the Relativistic Fermi
Gas model}
The hadron tensor for the 2p-2h channel in a fully
relativistic framework is given by:
\begin{eqnarray}
W^{\mu\nu}_{2p2h}&=&\frac{V}{(2\pi)^9}\int d^3p^\prime_1\,
d^3p^\prime_2\, d^3h_1\, d^3h_2\, \frac{m^4_N}{E_1E_2E^\prime_1E^\prime_2}
r^{\mu\nu}(\mathbf{p}^\prime_1,\mathbf{p}^\prime_2,\mathbf{h}_1,
\mathbf{h}_2) \delta(E^\prime_1+E^\prime_2-E_1-E_2-\omega)
\nonumber\\
&&\Theta(p^\prime_1,p^\prime_2,h_1,h_2)
\delta^3(\mathbf{p}^\prime_1+
\mathbf{p}^\prime_2-\mathbf{h}_1-\mathbf{h}_2-\mathbf{q})
\label{hadron_tensor}
\end{eqnarray}
where $m_N$ is the nucleon mass, $V$ is the volume of the system
and we have defined the product of step functions,
\begin{equation}
\Theta(p^\prime_1,p^\prime_2,h_1,h_2)\equiv\theta(p^\prime_1-k_F)
\theta(p^\prime_2-k_F)\theta(k_F-h_1)\theta(k_F-h_2)
\end{equation}
which encodes the nuclear model.\\
Finally, the function
$r^{\mu\nu}(\mathbf{p}^\prime_1,\mathbf{p}^\prime_2,\mathbf{h}_1,
\mathbf{h}_2)$ is the elementary ``hadron'' tensor for
the transition of a nucleon pair with given initial $(\mathbf{h}_1,
\mathbf{h}_2)$ and final $(\mathbf{p}^\prime_1,\mathbf{p}^\prime_2)$
momenta, summed up over spin and isospin, given schematically in
terms of the antisymmetrized two-body currents by:
\begin{equation}
r^{\mu\nu}(\mathbf{p}^\prime_1,\mathbf{p}^\prime_2,\mathbf{h}_1,
\mathbf{h}_2)=\frac14\sum_{\sigma,\tau}
j^\mu(1^\prime,2^\prime;1,2)^{*}_A j^\nu(1^\prime,2^\prime;1,2)_A
\end{equation}
The above multidimensional integral (\ref{hadron_tensor})
can be done either numerically or its dimensions can be
further reduced under some approximations \cite{Van81,
Alberico:1983zg,Gil:1997bm}. We do not know exactly
the origin of the discrepancies between the available models.
These could be due to different two-body currents,
to local Fermi Gas model rather than global
one, or other effects. It is difficult
to make any concluding assessment
right now, but all those models should agree at the level
of phase space function $F(\omega,q)$, obtained assuming a
constant $r^{\mu\nu}$:
\begin{eqnarray}
F(\omega,q)\equiv\int d^3h_1\, d^3h_2\, d^3p^\prime_1\,
\frac{m^4_N}{E_1E_2E^\prime_1E^\prime_2}
\Theta(p^\prime_1,p^\prime_2,h_1,h_2)
\delta(E^\prime_1+E^\prime_2-E_1-E_2-\omega)
\end{eqnarray}
where $r^{\mu\nu}=1$ and $\mathbf{p}^\prime_2=\mathbf{h}_1+
\mathbf{h}_2+\mathbf{q}-\mathbf{p}^\prime_1$ integrating out the
delta function of momentum conservation.\\
The remaining delta function enables analytical integration
over the modulus of $\mathbf{p}^\prime_1$:
\begin{eqnarray}
F(\omega,q)=2\pi \int d^3h_1\, d^3h_2\, d\theta^\prime_1\,
\sin\theta^\prime_1 \frac{m^2_N}{E_1E_2}
\sum_{\alpha=\pm}\frac{m^2_N p^{\prime2}_1}{
|E^\prime_2 p^\prime_1-E^\prime_1\mathbf{p}^\prime_2\cdot
\hat{\mathbf{p}}^\prime_1|}\Theta(p^\prime_1,p^\prime_2,h_1,h_2)\Bigg|_{
p^\prime_1=p^{\prime(\alpha)}_1}
\end{eqnarray}
and the sum inside the integral sign runs over the two solutions
$p^{\prime(\pm)}_1$ of the energy conservation delta function
(see appendix C on Ref. \cite{Rui14}).
\subsection{Frozen nucleon approximation}
The frozen nucleon approximation is just a particular
case of the mean-value theorem in several variables
\begin{eqnarray}
\int_{a}^{b} f(x)dx&=&f(c)(b-a) \quad \text{with} \quad
c\in\left[a,b\right]\\
\int_{\mathcal{V}}f(\mathbf{r})d^n\mathbf{r}&=&f(\mathbf{c})
\int_{\mathcal{V}}d^n\mathbf{r}=f(\mathbf{c})
\mathcal{V}\quad \text{with} \quad
\mathbf{c}\in\mathcal{V}\label{theorem_n_dimensions}
\end{eqnarray}
In our case we are going to skip the two 3D integrations
over the holes momenta $(\mathbf{h}_1,\mathbf{h}_2)$ by
fixing them to some, in principle unknown, values inside
the Fermi sphere, while keeping the integration over the
emission angle and constraining it by energy conservation.
Of course, the obvious particularization of the mean-value
theorem (\ref{theorem_n_dimensions}) is:
\begin{eqnarray}
F(\omega,q)&=&\int d^3h_1 d^3h_2 d^3p^\prime_1\; f(\mathbf{h}_1,
\mathbf{h}_2,\mathbf{p}^\prime_1)=
\left(\frac43\pi k^3_F
\right)^2 \int d^3p^\prime_1\; f(\langle\mathbf{h}_1\rangle,
\langle\mathbf{h}_2\rangle,\mathbf{p}^\prime_1)
\end{eqnarray}
where $(\langle\mathbf{h}_1\rangle,
\langle\mathbf{h}_2\rangle)$ are the
two unknown hole momenta inside
the Fermi sphere.\\
Up to now the whole discussion is exact. What makes
the difference between an excellent approximation
to the true result or a poor one is just
the selection of the ``average''
hole momenta $(\langle\mathbf{h}_1\rangle,
\langle\mathbf{h}_2\rangle)$. Our choice to call it
\emph{frozen nucleon approximation} will be
$(\vec{0},\vec{0})$. There are mainly two arguments
which favor this choice:
\begin{itemize}
\item For high $q>>k_F>h_i$, one can assume both initial
nucleons at rest.
\item If one does not assume the above statement, one has to
determine the average angles for the two holes. And
this cannot be done without computing the full 7D integral
(see section \ref{sec:other_configurations}).
\end{itemize}
Now, we can define the \emph{frozen approximation}
phase-space function $\overline{F}(\omega,q)$:
\begin{equation}
\overline{F}(\omega,q)=\left(\frac43\pi k^3_F\right)^2
\int d^3p^\prime_1\,\delta(E^\prime_1+E^\prime_2
-\omega-2m_N)\,\Theta(p^\prime_1,p^\prime_2,0,0)
\frac{m^2_N}{E^\prime_1 E^\prime_2}\label{Fbar}
\end{equation}
where now $\mathbf{p}^\prime_2=\mathbf{q}-\mathbf{p}^\prime_1$.\\
The above phase-space function (\ref{Fbar}) allows us to
define the angular distribution in the \emph{frozen
approximation}:
\begin{eqnarray}
\overline{F}(\omega,q)&=&\left(\frac43\pi k^3_F\right)^2
\,2\pi \int^{\pi}_{0} d\theta^\prime_1\,\Phi(\theta^\prime_1)\\
\Phi(\theta^\prime_1)&=&\sin\theta^\prime_1\int dp^\prime_1\,
p^{\prime\,2}_1\,\delta(E^\prime_1+E^\prime_2
-\omega-2m_N)\,\Theta(p^\prime_1,p^\prime_2,0,0)
\frac{m^2_N}{E^\prime_1 E^\prime_2}
\end{eqnarray}
\vspace{-2cm}
\begin{figure}[h]
\centering
\begin{subfigure}{.5\textwidth}
\hspace{-2.5cm}
\includegraphics[height=16cm,width=12cm]{fig13.ps}
\vspace{-6cm}
\caption{}\label{fig:angular_distribution_500mev}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\hspace{-2.25cm}
\includegraphics[height=16cm,width=12cm]{fig10.ps}
\vspace{-6cm}
\caption{}\label{fig:angular_distribution_3gev}
\end{subfigure}
\caption{Angular distributions $\Phi(\theta^\prime_1)$
for $q=500$ MeV/c (left panel) and for $q=3$ GeV/c
(right panel) in the frozen nucleon approximation, for
three different values of $\omega$ as a function of the
emission angle $\theta^\prime_1$.}
\label{fig:angular_distribution}
\end{figure}
\\
The problem stands on the divergence of the angular
distribution for some kinematics. This can be seen in
some of the panels of figure \ref{fig:angular_distribution}
for the frozen nucleon approximation. For other values of
the holes momenta, the position of the pole is
different, but it is still present. This makes
crucial to determine analytically the angular
interval where the integration has to be performed.
This was done explicitly on section VI of Ref.
\cite{Rui14}.\\
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\hspace{-2.5cm}
\includegraphics[height=16cm,width=12cm]{fig26.ps}
\vspace{-6cm}
\caption{}\label{fig:comparison_lowq}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\hspace{-2.25cm}
\includegraphics[height=16cm,width=12cm]{fig27.ps}
\vspace{-6cm}
\caption{}\label{fig:comparison_highq}
\end{subfigure}
\caption{Comparison between frozen nucleon approximation
and full integral for low
and intermediate momentum transfers (left panel) and
for high momentum transfers (right panel).}
\label{fig:comparison_phase_space}
\end{figure}
Once the problem of the divergence has been correctly
addressed, the results for the phase space function in
the frozen approximation can be compared with the full
integral in figure \ref{fig:comparison_phase_space}.
Here we can appreciate the quality of the approximated
result for a wide range of momentum transfers. The main
discrepancies arise on the low energy-transfer region
in each plot. But the phase space function is
well-reproduced in the rest of the interval and, furthermore,
we have skipped 6 additional integrals (over holes momenta),
thus reducing significantly the computation time.
These results indicate that this approximation is
especially well-suited for Montecarlo
event generators, particularly if the goal is
to estimate quite accurately the cross section for
multi-nucleon knockout in the shortest time as possible.
\subsection{Other initial configurations}\label{sec:other_configurations}
We have chosen up to now the
\emph{frozen nucleon approximation}, but other initial
$(\langle\mathbf{h}_1\rangle,
\langle\mathbf{h}_2\rangle)$ configurations could
have been selected as well. We have considered six
configurations depicted in fig. \ref{fig:configurations_a}.
The configurations
with total initial momentum of the pair equal to zero
(U,D or T,$-$T) would, in principle, yield a result very close to
that of the \emph{frozen approximation}. This can be
easily observed in figure
\ref{fig:comparison_configurations_zero}.
\begin{figure}[h]
\centering
\begin{subfigure}{.5\textwidth}
\hspace{-1.5cm}
\vspace{5cm}
\includegraphics[height=15cm,width=10.5cm]{geometria6.ps}
\vspace{-10cm}
\caption{}\label{fig:configurations_a}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\hspace{-1.25cm}
\vspace{3cm}
\includegraphics[height=16cm,width=12cm]{fig28.ps}
\vspace{-8.5cm}
\caption{}\label{fig:configurations_b}
\end{subfigure}
\caption{Geometry employed for emission of a
pair of nucleons with initial parallel momenta (UU,DD),
anti-parallel (UD,T-T) and perpendicular (TT',-T-T') on
the left panel. In the right panel a
comparison between
average momentum approximation and full
integral is shown. Initial momenta are both
200 MeV/c.}\label{fig:configurations}
\end{figure}
In the first comparison, Fig. \ref{fig:configurations_b},
we show the contribution
of two pairs of nucleons with the same momentum
$h_1=h_2=200$ MeV/c, and both parallel, pointing
upward (U) and downward (D) with respect to the z
axis, that is, the direction of $\mathbf{q}$.
The contribution of the
UU configuration is smaller than average, while the DD
is larger. This is so because in the UU case the total
momentum $p^\prime$ in the final state is large. By momentum
conservation, the momenta $p^\prime_1$ and $p^\prime_2$
must also be large.
Therefore, these states need a large excitation energy, and
they start to contribute for high $\omega$ transfer. In the DD
configuration, the total momentum $p^\prime$ is smaller, so the final
momenta $p^\prime_1$ and $p^\prime_2$ can also be small,
with small required excitation energy. Therefore, they start
to contribute at lower $\omega$.\\
In the example of Fig. \ref{fig:comparison_configurations_zero},
two anti-parallel configurations
are shown. In the UD case, one nucleon is moving upward
and the other downward the z axis with total momentum
zero of the pair. This situation is similar to that of a pair of
highly correlated nucleons with large relative momentum
\cite{Korover:2014dma}. Since the total momentum
is zero, the final 2p-2h state
has total momentum q, exactly the same that it would have
in the \emph{frozen nucleon approximation}. Therefore,
the contribution of this configuration is similar to the average.
The same conclusions can be drawn in the case of the
configuration $T$, $-T$, with one nucleon moving along the $x$
axis (transverse direction) and the other along $-x$ with
opposite momentum. The contribution of this pair is
exactly the same as that of the UD configuration in the
total phase-space function.\\
Finally, we show in Fig.
\ref{fig:comparison_configurations_orthogonal} two
intermediate cases that
are neither parallel nor anti-parallel configurations. They
consist of two pairs of transverse nucleons moving along
mutually perpendicular directions. In the first case, we
consider a $T$ nucleon and a second $T^\prime$ nucleon moving in
the y axis out of the scattering plane. The contribution of
the $TT^\prime$ pair is large, while the
one of the opposite case, $-T$,
$-T^\prime$, is small. On the average, they are close to the total
result.
\begin{figure}
\centering
\begin{subfigure}{.5\textwidth}
\hspace{-2.5cm}
\includegraphics[height=16cm,width=12cm]{fig29.ps}
\vspace{-6cm}
\caption{}\label{fig:comparison_configurations_zero}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\hspace{-2.25cm}
\includegraphics[height=16cm,width=12cm]{fig30.ps}
\vspace{-6cm}
\caption{}\label{fig:comparison_configurations_orthogonal}
\end{subfigure}
\caption{Comparison between average momentum
approximation and full integral. In the left panel
the initial momenta
are both 200 MeV/c with total momentum of the pair
equal to zero. In the right one the initial momenta
are both 200 MeV/c pointing in orthogonal directions.}
\label{fig:comparison_configurations}
\end{figure}
\section{Conclusions}
We have performed a detailed study of the two-particle-
two-hole phase-space function, which is proportional to the
nuclear two-particle emission response function for
constant current matrix elements. Our final goal was
to obtain accurate enough results without calculating the
7D integral. The frozen nucleon approximation (1D integral),
that is,
neglecting the momenta of the initial nucleons for high
momentum transfer, seems to be a quite promising approach
to reduce the computation time without missing significant
accuracy. The CPU time of the 7D
integral has been reduced significantly.
We are presently working on an implementation of
the present method with a complete model of the MEC
operators.
\section{Acknowledgements}
This work was supported by DGI (Spain), Grants
No. FIS2011-24149 and No. FIS2011-28738-C02-01; by
the Junta de Andaluc\'ia (Grants No. FQM-225 and
No. FQM-160); by the Spanish Consolider-Ingenio 2010
programmed CPAN, in part (M. B. B.) by the INFN project
MANYBODY and in part (T. W. D.) by U.S. Department of
Energy under Cooperative Agreement No. DE-FC02-
94ER40818. C. A. is supported by a CPAN postdoctoral
contract.
|
\section{Introduction}
Let $\Phi=(\phi_1,\ldots,\phi_4):[0,1]\to\R^4$ be a nondegenerate $C^4$ curve, in the sense that
\begin{equation}
\label{fe5}
\inf _{t_i\in[0,1]}|\operatorname{det}[\phi_j^{(i)}(t_i)]_{1\le i,j\le 4}|>0.
\end{equation}We will always write $$e(z)=e^{2\pi i z},$$
and for a positive weight $v$
$$\|f\|_{L^p(v)}=(\int|f(x)|^pv(x)dx)^{1/p}.$$
Also, for each ball $B$ centered at $c$ and with radius $R$, $w_{B}$ will denote the weight
$$w_{B}(x)= \frac{1}{(1+\frac{|x-c|}{R})^{100}}.$$
The following result was proved in \cite{Bo}.
\begin{theorem}
\label{thm1}
Let $I_1,I_2$ be intervals in $[0,1]$ with $\dist(I_1,I_2)\sim 1$ and consider a partition of $[0,1]$ into intervals $I_\tau$ of length $N^{-1/2}$. Then for each $h_i:I_i\to\C$ and each ball $B_N\subset \R^4$ with radius $N$
$$\|(\prod_{i=1}^2|\int_{I_i}h_i(t)e(x\cdot\Phi(t))dt|)^{1/2}\|_{L^{12}(B_{N})}\lesssim_\epsilon N^{-1/6+\epsilon}(\prod_{i=1}^2\sum_{I_\tau}\|\int_{I_\tau}h_i(t)e(x\cdot\Phi(t))dt\|_{L^{6}(w_{B_{N}})}^6)^{\frac{1}{12}}.$$
\end{theorem}
The inequality in Theorem \ref{thm1} played a key role in \cite{Bo} in the estimates on the Riemann zeta function on the critical line. In this paper we offer a different perspective on Theorem \ref{thm1} and show that it is a consequence of the decoupling theory for surfaces in $\R^4$ that we will develop here.
More precisely, consider a compact $C^3$ surface in $\R^4$
$$\Psi(t,s)=(\psi_1(t,s),\ldots,\psi_4(t,s)),\;\;\;\;\psi_i:[0,1]^2\to \R$$
which is assumed to satisfy the nondegeneracy condition
\begin{equation}
\label{fe1}
\operatorname{rank}[\Psi_t(t,s),\Psi_s(t,s),\Psi_{tt}(t,s),\Psi_{ss}(t,s),\Psi_{ts}(t,s)]=4,
\end{equation}
for each $t,s$. Our main new result is the following theorem.
\begin{theorem}
\label{ft1}For each $p\ge 2$, $g:[0,1]^2\to\C$ and each ball $B_N\subset\R^4$ with radius $N$
$$
\|\int_{[0,1]^2}g(t,s)e(x\cdot\Psi(t,s))dtds\|_{L^p(w_{B_{N}})}\le $$
\begin{equation}
\label{fe3}
\le D(N,p)(\sum_{\Delta\subset [0,1]^2\atop{l(\Delta)=N^{-1/2}}}\|\int_{\Delta}g(t,s)e(x\cdot\Psi(t,s))dtds\|_{L^p(w_{B_{N}})}^p)^{1/p}
\end{equation}
where the sum on the right hand side is over a partition of $[0,1]^2$ into squares $\Delta$ with side length $l(\Delta)=N^{-1/2}$, and
$$D(N,p)\lesssim_\epsilon N^{\frac12-\frac1p+\epsilon},\,2\le p\le 6$$
$$D(N,p)\lesssim_\epsilon N^{1-\frac4p+\epsilon},\,p\ge 6$$
\end{theorem}
\bigskip
A standard computation with $g=1_{[0,1]^2}$ reveals that the Theorem is essentially sharp, more precisely
\begin{equation}
\label{fe30}
D(N,p)\gtrsim N^{\frac12-\frac1p}\text{ for }2\le p\le 6,\;\;D(N,p)\gtrsim N^{1-\frac4p}\text{ for }6\le p\le \infty.
\end{equation}
It is worth recording the following trivial upper bound that follows from the Cauchy--Schwartz inequality
\begin{equation}
\label{fe3008}
D(N,p)\lesssim N^{1-\frac1p},\text{ for } p\ge 1.
\end{equation}
We will prove that $D(N,6)\lesssim_\epsilon N^{\frac16+\epsilon}$. The estimates for other $p$ will follow by interpolation with the trivial $p=2$ and $p=\infty$ results.
\bigskip
An equivalent way of describing a surface that satisfies \eqref{fe1} is the fact that it is locally non flat, in the following sense. For each $(t_0,s_0)\in [0,1]^2$ there is no unit vector $\gamma\in\R^4$ such that
$$|\langle\gamma,\Psi(t_0+\Delta t,s_0+\Delta s)-\Psi(t_0,s_0)\rangle|=O(|(\Delta t,\Delta s)|^3), \;\;\;\text{when } |(\Delta t,\Delta s)|\;\to 0.$$
This is easily seen by using Taylor's formula with third order error terms.
\bigskip
Near every $(t_0,s_0)\in [0,1]^2$, the surface can be represented with respect to an appropriate system of coordinates as
\begin{equation}
\label{fe11}
(t,s, A_1t^2+2A_2ts+A_3s^2, A_4t^2+2A_5ts+A_6s^2)+O(|(t,s)|^3).
\end{equation}
It suffices to choose two perpendicular axes in the tangent plane at $(t_0,s_0)$ and to apply the Taylor expansion for the surface with respect to these variables. It is easy to see that $\Psi$ satisfies \eqref{fe1} at $(t_0,s_0)$ if and only if
\begin{equation}
\label{fe12}
\operatorname{rank}\left[ \begin{array}{ccc}
A_1 & A_2 & A_3 \\
A_4 & A_5 & A_6 \end{array} \right]=2
\end{equation}
\bigskip
Let us now place \eqref{fe3} in the context of the more general decoupling theory from \cite{BD3} and \cite{BD4}. These papers completely settle the case of hypersurfaces in all dimensions, whose Gaussian curvature is nonzero everywhere. In two dimensions, the sharp inequality takes the form
$$\|\int_{[0,1]}h(t)e(x_1t+x_2\gamma(t))dt\|_{L^6(w_{B_N})}\lesssim_\epsilon $$
\begin{equation}
\label{fe6}
N^\epsilon(\sum_\tau\|\int_{I_\tau}h(t)e(x_1t+x_2\gamma(t))dt\|_{L^6(w_{B_N})}^2)^{1/2},
\end{equation}
where $I_\tau$ are intervals of length $N^{-1/2}$ that partition $[0,1]$ and $\gamma:[0,1]\to\R$ satisfies the curvature condition $$\inf_{0\le t\le 1}|\gamma''(t)|>0.$$
The paper \cite{BD4} also analyzes the decoupling theory for curves satisfying \eqref{fe5} in arbitrary dimensions $n$, but the picture for $n\ge 3$ is rather incomplete. The natural scaling for curves dictates that the intervals $I_\tau$ should have larger length $N^{-1/n}$, in order to run the machinery from \cite{BD3}. However, Theorem \ref{thm1} goes against this principle and uses intervals $I_\tau$ of length $N^{-1/2}$, rather than $N^{-1/4}$. One of the goals of this paper is to clarify why it is possible to work with such a scale.
While \cite{BD3} and \cite{BD4} deal with manifolds of codimension one and dimension one, respectively, inequality \eqref{fe3} is the first example that does not fall in either category. We will argue that the natural scale to work with in the context of surfaces in $\R^4$ is $N^{-1/2}$, the same as for hypersurfaces in any dimension. This will allow for the argument in \cite{BD3} to be carried over in this context. Moreover, Theorem \ref{thm1} will be seen to be a rather immediate consequence of Theorem \ref{ft1}. See Section \ref{fs2} below.
\bigskip
Inequality \eqref{fe3} at the critical index $p=6$ is an $l^6(L^6)$ decoupling, a bit weaker than the $l^2(L^6)$ decoupling
$$
\|\int_{[0,1]^2}g(t,s)e(x\cdot\Psi(t,s))dtds\|_{L^6(w_{B_{N}})}\lesssim_\epsilon
$$\begin{equation}
\label{fe2}
N^{\epsilon}(\sum_{\Delta\subset [0,1]^2\atop{l(\Delta)=N^{-1/2}}}\|\int_{[0,1]^2}g(t,s)e(x\cdot\Psi(t,s))dtds\|_{L^6(w_{B_{N}})}^2)^{1/2}
\end{equation}
since \eqref{fe2} automatically implies \eqref{fe3} via H\"older's inequality. We will see below however that \eqref{fe2} is false for some $\Psi$ satisfying \eqref{fe1}. Indeed, a standard discretization argument as in \cite{BD3} shows that if \eqref{fe2} were true, it would imply the following estimate for exponential sums
\begin{equation}
\label{fe4}
\frac{1}{|B_N|^{1/6}}\|\sum_{n=1}^Na_ne(x\cdot\Psi(\xi_n))\|_{L^6(B_N)}\lesssim_\epsilon N^{\epsilon}\|a_n\|_{l^2},
\end{equation}
for arbitrary $a_n\in \C$ and $N^{-1/2}$ separated $\xi_n\in[0,1]^2$. Choose now
$$\Psi(t,s)=(t,s,t^2,ts),$$
which is easily seen to satisfy \eqref{fe1}. This surface contains the line $$\{(0,s,0,0):s\in\R\}.$$ Choose $N^{1/2}$ equidistributed points $\xi_n$ on the line segment $\{0\}\times [0,1]$ which are $N^{-1/2}$ separated, and let $a_n=1$. It is easy to see that \eqref{fe4} will fail for this choice. We point out that \eqref{fe2} holds true for other $\Psi$ such as
$$\Psi(t,s)=(t,s,t^2,s^2),$$
a consequence of \eqref{fe6} (applied twice) and Fubini. This easy argument is not applicable for more ``twisted" $\Psi$, but at least explains the $L^6$ numerology, and shows that \eqref{fe3} generalizes \eqref{fe6}. It will later become clear that the proof of \eqref{fe6} in \cite{BD3} is a good template for our proof of \eqref{fe3}. In particular, both inequalities rely crucially on a bilinear estimate.
Theorem \ref{ft1} has natural extensions to the case when $\Psi:[0,1]^k\to\R^{2k}$, $k\ge 3$, but we will not pursue them here. The decoupling constants will depend on the dimension of the largest affine subspace of the manifold $\Psi$. This is similar to the role played by signature in the decoupling theory for hyperbolic paraboloids developed in \cite{BD4}.
\bigskip
\section{The proof of Theorem \ref{thm1}}
\label{fs2}
This section will be devoted to proving Theorem \ref{thm1} assuming Theorem \ref{ft1}, for $p=6$.
Given a curve $\Phi$ satisfying \eqref{fe5} and intervals $I_1,I_2$ as in Theorem \ref{thm1}, we construct the surface
\begin{equation}
\label{fe8}
\Psi_\Phi(t,s)=(\phi_1(t)+\phi_1(s),\ldots, \phi_4(t)+\phi_4(s)),\;\;t\in I_1,\;s\in I_2.
\end{equation}
We prove that $\Psi_\Phi$ satisfies \eqref{fe1} on $I_1\times I_2$. It will suffice if we show that
\begin{equation}
\label{fe7}
\operatorname{det}\left[ \begin{array}{cccc}
\phi_1'(t) & \phi_2'(t) & \phi_3'(t) & \phi_4'(t) \\ \phi_1'(s) & \phi_2'(s) & \phi_3'(s) & \phi_4'(s)\\ \phi_1''(t) & \phi_2''(t) & \phi_3''(t) & \phi_4''(t) \\ \phi_1''(s) & \phi_2''(s) & \phi_3''(s) & \phi_4''(s)
\end{array} \right]\not =0,\,\,\text{for }t\in I_1,\;s\in I_2.
\end{equation}
Note that $$\operatorname{det}\left[ \begin{array}{cccc}
\phi_1'(t) & \phi_2'(t) & \phi_3'(t) & \phi_4'(t) \\ \phi_1'(s) & \phi_2'(s) & \phi_3'(s) & \phi_4'(s)\\ \phi_1''(t) & \phi_2''(t) & \phi_3''(t) & \phi_4''(t) \\ \phi_1''(s) & \phi_2''(s) & \phi_3''(s) & \phi_4''(s)
\end{array} \right]=\lim_{\epsilon\to 0}\frac1{\epsilon^2}\operatorname{det}\left[ \begin{array}{cccc}
\phi_1'(t) & \phi_2'(t) & \phi_3'(t) & \phi_4'(t) \\ \phi_1'(s) & \phi_2'(s) & \phi_3'(s) & \phi_4'(s)\\ \phi_1'(t+\epsilon) & \phi_2'(t+\epsilon) & \phi_3'(t+\epsilon) & \phi_4'(t+\epsilon) \\ \phi_1'(s+\epsilon) & \phi_2'(s+\epsilon) & \phi_3'(s+\epsilon) & \phi_4'(s+\epsilon)
\end{array} \right].$$
A generalization of the Mean-Value Theorem (see \cite{PS}, Voll II, part V, Chap 1, No. 95) guarantees that
$$\operatorname{det}\left[ \begin{array}{cccc}
\phi_1'(t) & \phi_2'(t) & \phi_3'(t) & \phi_4'(t) \\ \phi_1'(s) & \phi_2'(s) & \phi_3'(s) & \phi_4'(s)\\ \phi_1'(t+\epsilon) & \phi_2'(t+\epsilon) & \phi_3'(t+\epsilon) & \phi_4'(t+\epsilon) \\ \phi_1'(s+\epsilon) & \phi_2'(s+\epsilon) & \phi_3'(s+\epsilon) & \phi_4'(s+\epsilon)
\end{array} \right]=$$$$\epsilon^2(t-s)^2(t+\epsilon-s)(s+\epsilon-t)\operatorname{det}\left[ \begin{array}{cccc}
\phi_1'(\tau_1) & \phi_2'(\tau_1) & \phi_3'(\tau_1) & \phi_4'(\tau_1) \\ \phi_1''(\tau_2) & \phi_2''(\tau_2) & \phi_3''(\tau_2) & \phi_4''(\tau_2)\\ \phi_1'''(\tau_3) & \phi_2'''(\tau_3) & \phi_3'''(\tau_3) & \phi_4'''(\tau_3) \\ \phi_1''''(\tau_4) & \phi_2''''(\tau_4) & \phi_3''''(\tau_4) & \phi_4''''(\tau_4)
\end{array} \right]$$
for some $\tau_i\in[0,1]$ depending on $t,s,\epsilon$. It is now clear that \eqref{fe7} follows from \eqref{fe5}.
\bigskip
For an arbitrary $S\subset [0,1]^2$, $g:S\to\C$ and $\Psi$ satisfying \eqref{fe1} define the extension operator
$$E_{S,\Psi}g(x)=\int_S g(t,s)e(x\cdot \Psi(t,s))dtds.$$
Going back to Theorem \ref{thm1} we note that
$$\prod_{i=1}^2\int_{I_i}h_i(t)e(x\cdot\Phi(t))dt=E_{I_1\times I_2,\Psi_\Phi}g(x)$$
with $g(t,s)=h_1(t)h_2(s).$ By applying Theorem \ref{ft1} it follows that
$$\|(\prod_{i=1}^2|\int_{I_i}h_i(t)e(x\cdot\Phi(t))dt|)^{1/2}\|_{L^{12}(B_{N})}\lesssim_\epsilon N^{\frac16+\epsilon}(\sum_{\Delta\subset [0,1]^2\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta,\Psi_\Phi}g\|_{L^6(w_{B_{N}})}^6)^{1/12}.$$
\bigskip
Fix $\Delta=J_1\times J_2$ from the summation above. To prove Theorem \ref{thm1}, it will suffice to argue that
$$\|E_{\Delta,\Psi_\Phi}g\|_{L^6(w_{B_{N}})}\lesssim$$
\begin{equation}
\label{fe9}
\lesssim N^{-\frac23}\|\int_{J_1}h_1(t)e(x\cdot\Phi(t))dt\|_{L^6(w_{B_{N}})}\|\int_{J_2}h_2(t)e(x\cdot\Phi(t))dt\|_{L^6(w_{B_{N}})}.
\end{equation}
This inequality will follow from the following transversality result.
\begin{lemma}Let $T_1,T_2$ be two cylindrical tubes in $\R^4$ with length $\sim N^{-1/2}$ in the direction $\v_i$ and radius $\sim N^{-1}$. Assume the angle between $\v_1$ and $\v_2$ is $\sim 1$. Let $f_i$ be a function which is Fourier supported in $T_i$. Then
$$\|f_1f_2\|_{L^2(\R^4)}\lesssim N^{-2}\|f_1\|_{L^2(\R^4)}\|f_2\|_{L^2(\R^4)}.$$
\end{lemma}
\begin{proof}
Start with a wave packet decomposition
$$f_i=\sum_{P\in\P_i}a_P\phi_P,$$
where $\P_i$ is a finitely overlapping cover of $\R^4$ with $N\times N\times N\times N^{1/2}$ rectangular plates $P_i$ orthogonal to $\v_i$. In particular, we may assume that
$$\int|\phi_P|^2=1,\;\;P\in\P_1\cup\P_2$$
$$\|f_i\|_2\sim(\sum_{P\in\P_i}|a_P|^2)^{1/2},$$
and also that $\phi_P$ has rapid decay away from $P$. The transversality guaranteed by the angle between $\v_1,\v_2$ forces
$$\int|\phi_{P_1}\phi_{P_1'}\phi_{P_2}\phi_{P_2'}|\le N^{-4}c(P_1,P_1')c(P_2,P_2'), \;\;P_i,P_i'\in\P_i,$$
where $c(P_i,P_i')$ are weights that decrease rapidly with the distance between $P_i,P_i'$. We conclude that
$$\|f_1f_2\|_{L^2(\R^4)}^2\le N^{-4}\sum_{P_1,P_1'\in\P_1}|a_{P_1}a_{P_1'}c(P_1,P_1')|\sum_{P_2,P_2'\in\P_2}|a_{P_2}a_{P_2'}c(P_2,P_2')|.$$
By a few applications of H\"older's inequality, this is further bounded by
$$N^{-4}\sum_{P_1\in\P_1}|a_{P_1}|^2\sum_{P_2\in\P_2}|a_{P_2}|^2,$$
as desired.
\end{proof}
\bigskip
To see \eqref{fe9}, consider a positive weight $v_{B_N}$ in the Schwartz class with Fourier support in the $N^{-1}$ neighborhood of the origin in $\R^4$ and which is $\ge 1$ on $B_N$. Note that the functions
$$f_i(x)=(v_{B_N}(x)\int_{J_i}h_i(t)e(x\cdot\Phi(t))dt)^3$$
satisfy the requirements of the lemma from above. We get
$$\|E_{\Delta,\Psi_\Phi}g\|_{L^6(B_{N})}\lesssim \|f_1f_2\|_{L^2(\R^4)}^{1/3}\lesssim$$$$\lesssim N^{-2/3}\|\int_{J_1}h_1(t)e(x\cdot\Phi(t))dt\|_{L^6(v^6_{B_{N}})}\|\int_{J_2}h_2(t)e(x\cdot\Phi(t))dt\|_{L^6(v^6_{B_{N}})}.$$
Inequality \eqref{fe9} now follows from standard manipulations, by using the fact that $v_{B_N}$ has Schwartz decay, while $w_{B_N}$ has prescribed polynomial decay.
\section{Reduction to quadratic surfaces}
This section will clarify why in Theorem \ref{ft1} the decoupling intervals $\Delta$ have scale $N^{-1/2}$. The key will be the approximation of nondegenerate surfaces by quadratic ones.
For a $\Psi$ satisfying \eqref{fe1} we let $\A_N=\A_N(\Psi)$ be the $N^{-1}$ neighborhood of the surface $\Psi([0,1]^2)$. Consider a fixed finitely overlapping cover of $[0,1]^2$ with squares $\Delta$ of side length $N^{-1/2}$, and let $\P_N$ be the associated cover of $\A_N$ with $N^{-1}$ neighborhoods $\theta$ of $\Psi(\Delta)$. Note that each $\theta$ is essentially a rectangular region with dimensions $N^{-1},N^{-1}, N^{-1/2}, N^{-1/2}$. We will denote by $f_\theta$ an appropriate smooth Fourier restriction of $f$ to $\theta$ so that
$$f=\sum_{\theta\in \P_N}f_\theta.$$
Let $K_{p,\Psi}(N)$ denote the best constant such that
\begin{equation}
\label{fe18}
\|f\|_{L^p(\R^4)}\le K_{p,\Psi}(N)(\sum_{\theta\in \P_N}\|f_\theta\|_{L^p(\R^4)}^p)^{1/p}
\end{equation}
holds for each $f$ Fourier supported in $\A_N(\Psi)$.
We first observe that Theorem \ref{ft1} is equivalent with proving that
\begin{equation}
\label{fe10}
K_{6,\Psi}(N)\lesssim_\epsilon N^{\frac13+\epsilon}.
\end{equation}
This can be seen by foliating $\A_N(\Psi)$ into translates of $\Psi$. We refer the reader to \cite{BD3} for details for a similar statement in a related context.
Next, we will prove that \eqref{fe10} for a given $\Psi$ follows if we assume \eqref{fe10} for quadratic surfaces of the form
$$\Psi_{\bf{A}}(t,s)=(t,s, A_1t^2+2A_2ts+A_3s^2, A_4t^2+2A_5ts+A_6s^2),$$
with ${\bf A}=(A_1,\ldots,A_6)$ satisfying \eqref{fe12}. To see this, let $f$ be Fourier supported in $\A_N(\Psi)$. Since $\widehat{f}$ is also Fourier supported in $\A_{N^{\frac23}}(\Psi)$, we have the initial decoupling
\begin{equation}
\label{fe14}
\|f\|_{L^p(\R^4)}\le K_{p,\Psi}(N^{\frac23})(\sum_{\tau\in \P_{N^{\frac23}}}\|f_\tau\|_{L^p(\R^4)}^p)^{1/p}.
\end{equation}
Note that each $f_\tau$ is Fourier supported on $\tau\cap \A_N(\Psi)$. Also, \eqref{fe11} shows that after a rotation $\tau\cap \A_N(\Psi)$ is a subset of $\A_{O(N)}(\Psi_{\bf A})$ for some ${\bf A}$ satisfying \eqref{fe12}. By applying the linear rescaling
$$(t_1,\ldots,t_4)\mapsto(N^{1/3}t_1,N^{1/3}t_2,N^{2/3}t_3,N^{2/3}t_4)$$
with respect to the local system of coordinates,
$\tau\cap \A_N(\Psi)$ is mapped into $\A_{O(N^{1/3})}(\Psi_{\bf A})$.
By rescaling back, it follows that
\begin{equation}
\label{fe15}
\|f_\tau\|_{L^p(\R^4)}\le K_{p,\Psi_{\bf A}}(N^{1/3})(\sum_{\theta\in \P_N\atop{\theta\subset \tau}}\|f_\theta\|_{L^p(\R^4)}^p)^{1/p}.
\end{equation}
The vector ${\bf A}$ will of course depend on $\tau$. Combining \eqref{fe14} and \eqref{fe15} we conclude that there exists $C_\Psi$ depending only on $\Psi$, such that for each $N$
\begin{equation}
\label{fe13}
K_{p,\Psi}(N)\le K_{p,\Psi}(N^{\frac23})\sup_{{\bf A}\in\L}K_{p,\Psi_{\bf A}}(N^{\frac13})
\end{equation}
with $$\L=$$$$\left\{{\bf A}:\;|A_i|\le C_\Psi, \;\max\left\{\left|\operatorname{det}\left[ \begin{array}{ccc}
A_1 & A_2 \\
A_4 & A_5 \end{array} \right]\right|, \left|\operatorname{det}\left[ \begin{array}{ccc}
A_1 & A_3 \\
A_4 & A_6 \end{array} \right]\right|, \left|\operatorname{det}\left[ \begin{array}{ccc}
A_3 & A_2 \\
A_6 & A_5 \end{array} \right]\right| \right\}\ge C_\Psi^{-1}\right\}.$$
Thus, assuming
$$\sup_{{\bf A}\in\L}K_{6,\Psi_{\bf A}}(N)\lesssim_{\epsilon,C_\Psi} N^{\frac13+\epsilon},$$
\eqref{fe10} will follow by iterating \eqref{fe13}.
From now on we will work with the surface $\Psi_{\bf A}$, for a fixed ${\bf A}\in \L$, and will show that
$$K_{6,\Psi_{\bf A}}(N)\lesssim_\epsilon N^{\frac13+\epsilon},$$
A careful analysis of the forthcoming argument will show that the implicit constant will depend on $C_\Psi$. We will never specify the exact dependence.
\bigskip
\section{Transversality and a bilinear theorem}
Fix ${\bf A}\in \L$. The extension operator $E$ will from now on be implicitly understood to be with respect to $\Psi_{\bf A}$. Define
$$c_{1,{\bf A}}=\operatorname{\det} \left[ \begin{array}{ccc}
A_1 & A_2 \\
A_4 & A_5 \end{array} \right], \;\;c_{2,{\bf A}}=\operatorname{det}\left[ \begin{array}{ccc}
A_1 & A_3 \\
A_4 & A_6 \end{array} \right], \;\;c_{3,{\bf A}}=\operatorname{det}\left[ \begin{array}{ccc}
A_6 & A_5 \\
A_3 & A_2 \end{array} \right].$$
\begin{definition}
Let $\nu\le 1$. We say that two sets $S_1,S_2\subset [0,1]^2$ are $\nu$-transverse if
\begin{equation}
\label{fe16}
c_{1,{\bf A}}(t_1-t_2)^2+c_{2,{\bf A}}(t_1-t_2)(s_1-s_2)+ c_{3,{\bf A}}(s_1-s_2)^2\ge \nu
\end{equation}
for each $(t_i,s_i)\in S_i.$
\end{definition}
The following bilinear theorem will play a key role in our approach.
\begin{theorem}
\label{ft3}
Let $R_1$, $R_2$ be two $\nu$-transverse squares in $[0,1]^2$. Then for each $g_i:R_i\to\C$ we have
$$\||E_{R_1}g_1E_{R_2}g_2|^{1/2}\|_{L^4(\R^4)}\lesssim_\nu (\|g_1\|_{L^2(R_1)}\|g_2\|_{L^2(R_2)})^{1/2}.$$
\end{theorem}
\begin{proof}
We will perform the following change of variables
$$(t_1,s_1,t_2,s_2)\in R_1\times R_2\mapsto (u_1,u_2,u_3,u_4)=$$$$(t_1+t_2,s_1+s_2,A_1(t_1^2+t_2^2)+2A_2(t_1s_1+t_2s_2)+A_3(s_1^2+s_2^2), A_4(t_1^2+t_2^2)+2A_5(t_1s_1+t_2s_2)+A_6(s_1^2+s_2^2))$$
whose Jacobian is
$$c_{1,{\bf A}}(t_1-t_2)^2+c_{2,{\bf A}}(t_1-t_2)(s_1-s_2)+ c_{3,{\bf A}}(s_1-s_2)^2\ge \nu.$$
It follows that
$$|E_{R_1}g_1(x)E_{R_2}g_2(x)|=\widehat{GJ}(x)$$
where $G(u_1,\ldots,u_4)=g_1(t_1,s_1)g_2(t_2,s_2)$ and $|J(u)|\le \nu^{-1}$.
Using Plancherel's identity we get
$$\||E_{R_1}g_1E_{R_2}g_2|^{1/2}\|_{L^4(\R^4)}=\|GJ\|_{L^2(\R^4)}^{1/2}\le$$$$\le \nu^{-\frac14}(\int_{\R^4}|G^2(u)J(u)|du)^{\frac14}=\nu^{-\frac14}(\|g_1\|_{L^2(R_1)}\|g_2\|_{L^2(R_2)})^{1/2}.$$
\end{proof}
\begin{corollary}
\label{fc1}
Let $R_1,R_2\subset [0,1]^2$ be $\nu$-transverse squares. Then for each $4\le p\le \infty$ and $g_i:R_i\to\C$ we have
\begin{equation}
\label{we3}
\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}} |E_{\Delta}g_i|^2)^{1/4}\|_{L^{p}(w_{B_N})}\lesssim_\nu N^{-\frac4{p}}(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{p/2}(w_{B_N})}^2)^{\frac1{4}}.
\end{equation}
\end{corollary}
\begin{proof}
A standard consequence of Theorem \ref{ft3} and Plancherel's identity is the following local inequality
$$\|(\prod_{i=1}^2 |E_{R_i}g_i|)^{1/2}\|_{L^{4}(w_{B_N})}\lesssim_\nu N^{-1}(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{2}(w_{B_N})}^2)^{\frac1{4}}.$$
A randomization argument further leads to the inequality
$$\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}} |E_{\Delta}g_i|^2)^{1/4}\|_{L^{4}(w_{B_N})}\lesssim_\nu N^{-1}(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{2}(w_{B_N})}^2)^{\frac1{4}}.$$
It now suffices to interpolate this with the trivial inequality
$$\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}} |E_{\Delta}g_i|^2)^{1/4}\|_{L^{\infty}(w_{B_N})}\le (\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{\infty}(w_{B_N})}^2)^{\frac1{4}}.$$
We refer the reader to \cite{BD3} for how this type of interpolation is performed.
\end{proof}
\bigskip
For the argument in the following sections, it will be important that transversality is quite generic.
\begin{proposition}
\label{fp1}
For $K=2^m\ge 1$, consider the collection $Col_K$ of the $K^2$ dyadic squares in $[0,1]^2$ with side length $K^{-1}$. For each $R\in Col_K$, there are $O(K)$ squares $R'\in Col_K$ which are $K^{-2}$-transverse to $R$.
\end{proposition}
\begin{proof}
Let $\lambda_1,\lambda_2$ be the eigenvalues of $$M_{\bf A}=\left[ \begin{array}{ccc}
c_{1,{\bf A}} & \frac{c_{2,{\bf A}}}2 \\
\frac{c_{2,{\bf A}}}2 & c_{3,{\bf A}} \end{array} \right]$$
and let $\v_1,\v_2$ be a corresponding orthonormal eigenbasis.
Assume $|\lambda_1|\ge |\lambda_2|$. The condition ${\bf A}\in \L$ guarantees that $|\lambda_1|\gtrsim 1$.
If $(\beta_1,\beta_2)$ are the coordinates of $(t_1-t_2,s_1-s_2)$ with respect to the $(\v_1,\v_2)$ basis, then \eqref{fe16} becomes
\begin{equation}
\label{fe17}
|\beta_1^2\lambda_1+\beta_2^2\lambda_2|\ge K^{-2}.
\end{equation}
If $\lambda_1\lambda_2\ge 0$, then $|\beta_1|\ge |\lambda_1|^{-1/2}K^{-1}$ will automatically force \eqref{fe17}. $R$ and $R'$ will be $K^{-2}$-transverse as soon as $R'$ is outside the $O(K^{-1})$ wide strip containing $R$ and stretching in the direction of $\v_2$. There are $O(K)$ squares in this strip.
If $\lambda_1\lambda_2<0 $, then $$\min\left\{\left|\beta_1+\beta_2\sqrt{-\frac{\lambda_2}{\lambda_1}}\;\right|,\left|\beta_1-\beta_2\sqrt{-\frac{\lambda_2}{\lambda_1}}\;\right|\right\}\ge |\lambda_1|^{-1/2}K^{-1}$$ will again force \eqref{fe17}. $R$ and $R'$ will be $K^{-2}$-transverse as soon as $R'$ is outside the two $O(K^{-1})$ wide strips containing $R$ and stretching in the directions of $\pm\sqrt{-\frac{\lambda_2}{\lambda_1}}\v_1+\v_2$.
\end{proof}
\section{Linear versus bilinear decoupling}
We will make use of the following ``trivial" decoupling to treat the non transverse contribution in the Bourgain--Guth decomposition.
\begin{lemma}
\label{fl1}
Let $R_1,\ldots,R_K$ be pairwise disjoint squares in $[0,1]^2$ with side length $K^{-1}$. Then for each $2\le p\le \infty$
$$ \|\sum_iE_{R_i}g\|_{L^p(w_{B_K})}\lesssim_p K^{1-\frac2p}(\sum_i\|E_{R_i}g\|_{L^p(w_{B_K})}^p)^{1/p}.$$
\end{lemma}
\begin{proof}
The key observation is the fact that if $f_1,\ldots,f_K:\R^4\to\C$ are such that $\widehat{f_i}$ is supported on a ball $B_i$ and the dilated balls $(2B_i)_{i=1}^K$ are pairwise disjoint, then
\begin{equation}
\label{fe36}
\|f_1+\ldots+f_K\|_{L^p(\R^4)}\lesssim_p K^{1-\frac2p}(\sum_i\|f_i\|_{L^p(\R^4)}^p)^{\frac1p}.
\end{equation}
In fact more is true. If $T_i$ is a smooth Fourier multiplier adapted to $2B_i$ and equal to 1 on $B_i$, then the inequality
$$\|T_1(f_1)+\ldots+T_K(f_K)\|_{L^p(\R^4)}\lesssim_p K^{1-\frac2p}(\sum_i\|f_i\|_{L^p(\R^4)}^p)^{\frac1p}$$
for arbitrary $f_i\in L^p(\R^4)$
follows by interpolating the immediate $L^2$ and $L^\infty$ estimates. Inequality \eqref{fe36} is the best one can say in general, if no further assumption is made on the Fourier supports of $f_i$. Indeed, if $\widehat{f_i}=1_{B_i}$ with $B_i$ equidistant balls of radius one with collinear centers, then the reverse inequality will hold.
Let now $v_{B_K}$ be a Schwartz function with Fourier support in the $K^{-1}$ neighborhood of the origin in $\R^4$ and which is $\ge 1$ on $B_K$. It suffices to note that the Fourier supports of the functions $f_i=v_{B_K}E_{R_i}g$ have bounded overlap.
\end{proof}
\bigskip
For $2\le p<\infty$ and $N\ge 1$, recall that $D(N,p)$ is the smallest constant such that the decoupling
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_N})}\le D(N,p)(\sum_{l(\Delta)=N^{-1/2}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p)^{1/p}$$
holds true for all $g$ and all balls $B_N$ or radius $N$. The sum on the right is over a partition of $[0,1]^2$ into dyadic squares $\Delta$ of side length $N^{-1/2}$.
We now introduce a bilinear version of $D(N,p) $. Given also $\nu\le 1$, let $D_{multi}(N,p,\nu)$ be the smallest constant such that the bilinear decoupling
$$\||E_{R_1}g_1E_{R_2}g_2|^{\frac12}\|_{L^p(w_{B_N})}\le D_{multi}(N,p,\nu)(\prod_{i=1}^2\sum_{l(\Delta)=N^{-1/2}}\|E_{\Delta}g_i\|_{L^p(w_{B_N})}^p)^{\frac1{2p}}$$
holds true for all $\nu$-transverse squares $R_1,R_2\subset [0,1]^{2}$ with arbitrary side lengths, all $g_i:R_i\to\C$ and all balls $B_N\in\R^4$ with radius $N$.
H\"older's inequality shows that $D_{multi}(N,p,\nu)\le D(N,p)$. The rest of the section will be devoted to proving that the reverse inequality is also essentially true. This will follow from a variant of the Bourgain--Guth induction on scales in \cite{BG}. More precisely, we prove the following result.
\begin{theorem}
\label{ft2}
For each $\nu\le \frac1{10}$ and $p\ge 2$ there exists $C_{\nu}>0$ and $\epsilon(\nu,p)$ with $\lim_{\nu\to 0}\epsilon(\nu,p)=0$ such that for each $N\ge 1$
\begin{equation}
\label{fe31}
D(N,p)\le C_{\nu}N^{\epsilon(\nu,p)}\sup_{1\le M\le N}(\frac{M}{N})^{\frac1p-\frac12}D_{multi}(M,p,\nu).
\end{equation}
\end{theorem}
The key step in achieving this result is the following inequality.
\begin{proposition}
\label{fp2}
For $2\le p<\infty$, there is a constant $C_p$ which only depends on $p$ so that for each $g$ and $N,K\ge 1$ we have
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_N})}^p\le $$$$ C_p\left(K^{p-2}\sum_{l(R)=\frac1K}\|E_{R}g\|_{L^p(w_{B_N})}^p+K^{4p}D_{multi}(N,p,K^{-2})^p\sum_{l(\Delta)=N^{-1/2}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p\right).$$
\end{proposition}
\bigskip
The exponent $4p$ in $K^{4p}$ is not important and could easily be improved, but the exponent $p-2$ in $K^{p-2}$ is sharp and will play a critical role in the rest of the argument.
\bigskip
\begin{proof}
Following the standard formalism from \cite{BG}, we may assume that $|E_{R}g(x)|$ is essentially constant on each ball $B_K$ of radius $K$, and will we denote by $|E_{R}g(B_K)|$ this value. Write
$$E_{[0,1]^2}g(B_K)=\sum_{l(R)=\frac1K}E_{R}g(B_K).$$
Fix $B_K$. Let $R^*$ be a square which maximizes the value of $|E_{R}g(B_K)|$. We distinguish two cases.
First, if there is some $R^{**}$ which is $K^{-2}$-transverse to $R^*$ and such that
$|E_{R^{**}}g(B_K)|\ge K^{-2}|E_{R^*}g(B_K)|$, then
$$|E_{[0,1]^2}g(B_K)|\le K^3(|E_{R^{*}}g(B_K)E_{R^{**}}g(B_K)|)^{1/2}.$$
Otherwise, $|E_{R}g(B_K)|<K^{-2}|E_{R^*}g(B_K)|$ whenever $R$ is $K^{-2}$-transverse to $R^*$. In this case we can write
$$|E_{[0,1]^2}g(B_K)|\le 2|E_{R^*}g(B_K)|+|\sum_{R\text{ not } K^{-2}-\text{transverse to }R^* }E_{R}g(B_K)|.$$
Using Lemma \ref{fl1} and Proposition \ref{fp1} we get
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_K})}\lesssim_p \|E_{R^*}g\|_{L^p(w_{B_K})}+ K^{1-\frac2{p}}(\sum_{R\text{ not }K^{-2}-\text{transverse to }R^*}\|E_{R}g\|_{L^p(w_{B_K})}^p)^{1/p}\le $$
$$\lesssim K^{1-\frac2{p}}(\sum_{\text{all }R}\|E_{R}g\|_{L^p(w_{B_K})}^p)^{1/p}.$$
To summarize, in either case we can write
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_K})}\lesssim_p $$$$ K^3\max_{R_1,R_2:\;K^{-2}-\text{transverse}}\|(|E_{R_1}gE_{R_2}g|)^{1/2}\|_{L^p(w_{B_K})}+ K^{1-\frac2{p}}(\sum_{\text{all }R}\|E_{R}g\|_{L^p(w_{B_K})}^p)^{1/p}\le$$
$$\le K^3(\sum_{R_1,R_2:\;K^{-2}-\text{transverse}}\|(|E_{R_1}gE_{R_2}g|)^{1/2}\|_{L^p(w_{B_K})}^p)^{1/p}+ K^{1-\frac2{p}}(\sum_{\text{all }R}\|E_{R}g\|_{L^p(w_{B_K})}^p)^{1/p}.$$
Raising to the power $p$ and summing over $B_K\subset B_N$ leads to the desired conclusion.
\end{proof}
\bigskip
It is worth mentioning that in general, one can not do better than the trivial decoupling to estimate the contribution from the non transverse terms. This is illustrated by the example
$$\Psi(t,s)=(t,s,t^2,ts).$$
The squares $R_1,\ldots,R_K\subset [0,1]^2$ containing the line segment $\{0\}\times [0,1]$ will be pairwise non transverse. It is easy to see that
$$\|\sum_iE_{R_i}1\|_{L^p(w_{B_K})}\sim_p K^{1-\frac2p}(\sum_i\|E_{R_i}1\|_{L^p(w_{B_K})}^p)^{1/p}.$$
This is due to the fact that the sets $\Psi(R_i)$ intersect the line $$\{(0,s,0,0):s\in\R\}.$$
The lack of curvature prevents anything better than the trivial decoupling to hold.
\bigskip
\bigskip
Using a form of parabolic rescaling, the result in Proposition \ref{fp2} leads to the following general result.
\begin{proposition}
\label{fp3}
Let $R\subset[0,1]^2$ be a square with side length $\delta$. For $2\le p<\infty$, there is a constant $C_p$ which only depends on $p$ so that for each $g$, $K\ge 1$ and $N>\delta^{-2}$ we have
$$\|E_{R}g\|_{L^p(w_{B_N})}^p\le $$$$ C_p\left(K^{p-2}\sum_{R'\subset R\atop{l(R')=\frac\delta{K}}}\|E_{R'}g\|_{L^p(w_{B_N})}^p+K^{4p}D_{multi}(N\delta^{2},p,K^{-2})^p\sum_{\Delta\subset R\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p\right).$$
\end{proposition}
\begin{proof}
Assume $R=[a,a+\delta]\times [b,b+\delta]$. The affine change of variables
$$(t,s)\in R\mapsto(t',s')=\eta(t,s)=(\frac{t-a}{\delta},\frac{s-b}{\delta})\in[0,1]^2$$ shows that
$$|E_Rg(x)|=\delta^2|E_{[0,1]^2}g^{a,b}(\bar{x})|,$$
$$|E_{R'}g(x)|=\delta^2|E_{R''}g^{a,b}(\bar{x})|,$$
where
$R''=\eta(R')$ is a square with side length $\frac1K$,
$$g^{a,b}(t',s')=g(\delta t'+a,\delta s'+b),$$
$$\bar{x}_1=\delta(x_1+x_3(2aA_1+2bA_2)+x_4(2aA_4+2bA_5))$$
$$\bar{x}_2=\delta(x_2+x_3(2bA_3+2aA_2)+x_4(2bA_6+2aA_5))$$
$$\bar{x}_3=\delta^2x_3,\;\;\bar{x}_4=\delta^2x_4.$$
Note that $\bar{x}$ is the image of $x$ under a shear transformation. Call $C_N$ the image of the ball $B_N$ in $\R^4$ under this transformation. Cover $C_N$ with a family $\F$ of balls $B_{\delta^2N}$ with $O(1)$ overlap. Write
$$\|E_{R}g\|_{L^p(w_{B_N})}=\delta^{2-\frac6p}\|E_{[0,1]^2}g^{a,b}\|_{L^p(w_{C_N})}$$
for an appropriate weight $w_{C_N}$. The right hand side is bounded by
$$\delta^{2-\frac6p}(\sum_{B_{\delta^2N}\in\F}\|E_{[0,1]^2}g^{a,b}\|_{L^p(w_{B_{\delta^2N}})}^p)^{1/p}.$$
Apply Proposition \ref{fp2} to each of the terms $\|E_{[0,1]^2}g^{a,b}\|_{L^p(w_{B_{\delta^2N}})}$ and then rescale back.
\end{proof}
\bigskip
We are now in position to prove Theorem \ref{ft2}. By iterating Proposition \ref{fp3} $n$ times we get
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_N})}^p\le (C_pK^{p-2})^n\sum_{l(R)=\frac1{K^n}}\|E_{R}g\|_{L^p(w_{B_N})}^p+$$$$+C_pK^{4p}\sum_{l(\Delta)=N^{-1/2}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p\sum_{j=0}^{n-1}(C_pK^{p-2})^{j}D_{multi}(NK^{-2j},p,K^{-2})^p.$$
Given $\nu\le \frac1{10}$, we apply this with $K^{-2}=\nu$ and $n$ such that $K^n= N^{\frac12}$. We get
$$\|E_{[0,1]^2}g\|_{L^p(w_{B_N})}\le $$$$ N^{\frac1{4p}\log_{\nu^{-1}}C_p}N^{^{\frac12-\frac1p}}(1+\nu^{-2}\sum_{j=0}^{n-1}(N\nu^{j})^{\frac1p-\frac12}D_{multi}(N\nu^{j},p,\nu))(\sum_{l(\Delta)={N^{-1/2}}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p)^{1/p}
\le$$
$$ N^{\frac1{4p}\log_{\nu^{-1}}C_p}N^{^{\frac12-\frac1p}}(1+\frac{\log_{\nu^{-1}}N}{4\nu^2}\max_{1\le M\le N}M^{\frac1p-\frac12}D_{multi}(M,p,\nu))(\sum_{l(\Delta)={N^{-1/2}}}\|E_{\Delta}g\|_{L^p(w_{B_N})}^p)^{1/p}.$$
The proof of Theorem \ref{ft2} is now complete.
\bigskip
\section{The final argument}
In this section we finish the proof of Theorem \ref{ft1}, by showing that
$$D(N,6)\lesssim_\epsilon N^{\frac13+\epsilon}.$$
For $p\ge 4$ define $\kappa_p$ such that
$$\frac2p=\frac{1-\kappa_p}{2}+\frac{\kappa_p}{p},$$
in other words, $$\kappa_p=\frac{p-4}{p-2}.$$
\begin{proposition}
Let $R_1,R_2$ be $\nu$-transverse squares in $[0,1]^2$ with arbitrary side lengths.
We have that for each radius $R\ge N$, $p\ge 4$ and $g_i:R_i\to \C$
$$\|(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}|E_{\tau}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_R})}\lesssim_{\nu,p}$$
$$
\lesssim_{\nu,p}\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_R})}^{1-\kappa_p}(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p}(w_{B_R})}^2)^{\frac{\kappa_p}{4}}.
$$
\end{proposition}
\begin{proof}
Let $B$ be an arbitrary ball of radius $N^{1/2}$. We start by recalling that \eqref{we3} on $B$ gives
\begin{equation}
\label{fe20}
\|(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}} |E_{\tau}g_i|^2)^{1/4}\|_{L^{p}(w_{B})}\lesssim_\nu N^{-\frac2{p}}(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p/2}(w_{B})}^2)^{\frac1{4}}.
\end{equation}
Write using H\"older's inequality
\begin{equation}
\label{we7}
(\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p/2}(w_{B})}^2)^{\frac1{2}}\le (\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{2}(w_{B})}^2)^{\frac{1-\kappa_p}{2}}(\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p}(w_{B})}^2)^{\frac{\kappa_p}{2}}.
\end{equation}
The next key element in our argument is the almost orthogonality specific to $L^2$, which will allow us to pass from scale $N^{-1/4}$ to scale $N^{-1/2}$. Indeed, since $(E_\Delta g_i)w_{B}$ are almost orthogonal for $l(\Delta)=N^{-1/2}$, we have
$$(\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{2}(w_{B})}^2)^{1/2}\lesssim (\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{2}(w_{B})}^2)^{1/2}.$$
We can now rely on the fact that $|E_{\Delta}g_i|$ is essentially constant on balls $B'$ of radius $N^{1/2}$ to argue that
$$(\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{2}(B')}^2)^{\frac1{2}}\sim |B'|^{1/2}(\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{2}}|_{B'}$$
and thus
\begin{equation}
\label{we8}
(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{2}(w_{B})}^2)^{\frac1{4}}\lesssim |B|^{\frac12-\frac1p}\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B})}.
\end{equation}
Combining \eqref{fe20}, \eqref{we7} and \eqref{we8} we get
$$
\|(\prod_{i=1}^2\sum_{\atop{l(\tau')=N^{-1/4}}}|E_{\tau}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B})}\lesssim_\nu \|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B})}^{1-\kappa_p}(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p}(w_{B})}^2)^{\frac{\kappa_p}{4}}.
$$
Summing this up over a finitely overlapping family of balls $B\subset B_R$ we get the desired inequality.
\end{proof}
\medskip
We will iterate the result of this proposition in the following form, a consequence of the Cauchy--Schwartz inequality
$$\|(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}|E_{\tau}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_R})}\le$$
\begin{equation}
\label{fe21}
\le C_{p,\nu}N^{\frac{\kappa_p}{2}(\frac12-\frac1p)}\|(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_R})}^{1-\kappa_p}(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/4}}}\|E_{\tau}g_i\|_{L^{p}(w_{B_R})}^p)^{\frac{\kappa_p}{2p}}.
\end{equation}
\bigskip
We will also need the following immediate consequence of the Cauchy--Schwartz inequality. While the exponent $2^{-s}$ in $N^{2^{-s}}$ can be improved by using the bilinear Theorem \ref{ft3}, the following trivial estimate will suffice for our purposes.
\begin{lemma}
\label{wlem0081}Consider two rectangles $R_1,R_2\subset [0,1]^2$ with arbitrary side lengths. Assume $g_i$ is supported on $R_i$.
Then for $1\le p\le\infty$ and $s\ge 2$
$$\|(\prod_{i=1}^2|E_{R_i}g_i|)^{1/2}\|_{L^{p}({w_{B_N}})}\le N^{2^{-s}}\|(\prod_{i=1}^2\sum_{\atop{l(\tau_s)=N^{-2^{-s}}}}|E_{\tau_s}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_N})}.$$
\end{lemma}
\bigskip
Using parabolic rescaling as in the proof of Theorem \ref{fp3}, we get that for each square $R\subset[0,1]^2$ with side length $N^{-\rho}$, $\rho\le \frac12$
\begin{equation}
\label{fe22}
\|E_Rg\|_{L^p(w_{B_N})}\le D(N^{1-2\rho},p)(\sum_{\Delta\subset R\atop{l(\Delta)=N^{-1/2}}}\|E_\Delta g\|_{L^p(w_{B_N})}^p)^{1/p}.
\end{equation}
\bigskip
Fix a pair of $\nu$-transverse rectangles $R_1,R_2\subset [0,1]^2$ with arbitrary side lengths and assume $g_i$ is supported on $R_i$. Start with Lemma \ref{wlem0081}, continue with iterating \eqref{fe21} $s-1$ times, and invoke \eqref{fe22} at each step to write
$$\|(\prod_{i=1}^2|E_{R_i}g_i|)^{1/2}\|_{L^{p}({B_N})}\le N^{2^{-s}}C_{p,\nu}^{s-1}(\prod_{i=1}^2\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{p}(w_{B_N})}^p)^{\frac{1}{2p}}\times$$
$$\times N^{\frac{\kappa_p}{2}(\frac12-\frac1p)(1-\kappa_p)^{s-2}}\cdot\ldots\cdot N^{\frac{\kappa_p}{2^{s-2}}(\frac12-\frac1p)(1-\kappa_p)}N^{\frac{\kappa_p}{2^{s-1}}(\frac12-\frac1p)}\|(\prod_{i=1}^2\sum_{\atop{l(\tau)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{4}}\|_{L^{p}(w_{B_N})}^{(1-\kappa_p)^{s}}\times$$
\begin{equation}
\label{fe34}
\times D(N^{1-2^{-s+1}},p)^{\kappa_p}D(N^{1-2^{-s+2}},p)^{\kappa_p(1-\kappa_p)}\cdot\ldots\cdot D(N^{1/2},p)^{\kappa_p(1-\kappa_p)^{s-2}}.
\end{equation}
Note that
$$\|(\sum_{\atop{l(\Delta)=N^{-1/2}}}|E_{\Delta}g_i|^2)^{\frac1{2}}\|_{L^{p}(w_{B_N})}\le N^{\frac12-\frac1p}(\sum_{\atop{l(\Delta)=N^{-1/2}}}\|E_{\Delta}g_i\|_{L^{p}(w_{B_N})}^p)^{1/p},$$
is an immediate consequence of Minkowski's and H\"older's inequalities. Using this, \eqref{fe34} can be rewritten as follows
$$D_{multi}(N,p,\nu)\le C_{p,\nu}^{s-1} N^{2^{-s}}N^{\kappa_p 2^{-s}(1-\frac2p)\frac{1-(2(1-\kappa_p))^{s-1}}{2\kappa_p-1}}\times$$
\begin{equation}
\label{fe23}
\times D(N^{1-2^{-s+1}},p)^{\kappa_p}D(N^{1-2^{-s+2}},p)^{\kappa_p(1-\kappa_p)}\cdot\ldots\cdot D(N^{1/2},p)^{\kappa_p(1-\kappa_p)^{s-2}} N^{O_p((1-\kappa_p)^s)}.
\end{equation}
\bigskip
Let $\gamma_p$ be the unique positive number such that
$$\lim_{N\to\infty}\frac{D(N,p)}{N^{\gamma_p+\epsilon}}=0,\;\text{for each }\epsilon>0$$
and
\begin{equation}
\label{fe27}
\limsup_{N\to\infty}\frac{D(N,p)}{N^{\gamma_p-\epsilon}}=\infty,\;\text{for each }\epsilon>0.
\end{equation}
The existence of such $\gamma_p$ is guaranteed by \eqref{fe30} and \eqref{fe3008}.
Recall that our goal is to prove that $\gamma_6=\frac13.$
By using the fact that $D(N,p)\lesssim_\epsilon N^{\gamma_p+\epsilon}$ in \eqref{fe23}, it follows that for each $\nu$ , $\epsilon>0$ and $s\ge 2$
\begin{equation}
\label{fe32}
\limsup_{N\to\infty}\frac{D_{multi}(N,p,\nu)}{N^{\gamma_{p,\epsilon,s}}}<\infty
\end{equation}
where
$$\gamma_{p,\epsilon,s}=2^{-s}+\kappa_p(\gamma_p+\epsilon)(\frac{1-(1-\kappa_p)^{s}}{\kappa_p}-2^{-s+1}\frac{1-(2(1-\kappa_p))^{s}}{2\kappa_p-1})+$$$$+\kappa_p 2^{-s}(1-\frac2p)\frac{1-(2(1-\kappa_p))^{s-1}}{2\kappa_p-1}+O_p((1-\kappa_p)^s).$$
We will show now that if $p>6$ then
$$\gamma_p\le \frac{p-6}{2p-8}+\frac12-\frac1p.$$
If we manage to do this, it will suffice to let $p\to 6$ to get $\gamma_6\le \frac13$, hence $\gamma_6=\frac13$, as desired. We first note that
\begin{equation}
\label{fe24}
2(1-\kappa_p)=\frac4{p-2}<1
\end{equation}
Assume for contradiction that
\begin{equation}
\label{fe26}
\gamma_p>\frac{p-6}{2p-8}+\frac12-\frac1p.
\end{equation}
A simple computation using \eqref{fe24} and \eqref{fe26} shows that for $s$ large enough, and $\epsilon$ small enough we have
$\gamma_{p,\epsilon,s}<\gamma_p.$
Choose now $\nu$ so small that
\begin{equation}
\label{fe33}
\gamma_{p,\epsilon,s}':=\gamma_{p,\epsilon,s}+\epsilon(\nu,p)<\gamma_p,
\end{equation}
and
\begin{equation}
\label{fe35}
\frac12-\frac1p+\epsilon(\nu,p)<1-\frac4p
\end{equation}
where $\epsilon(\nu,p)$ is from Theorem \ref{ft2}. The values of $\epsilon,s,\nu$ are fixed for the rest of the argument.
We will argue that
$$\limsup_{N\to\infty}\frac{D(N,p)}{N^{\max\{\gamma_{p,\epsilon,s}',\frac12-\frac1p+\epsilon(\nu,p)\}}}<\infty.$$
This will contradict either \eqref{fe27} (in light of \eqref{fe33}) or \eqref{fe30} (in light of \eqref{fe35}). We distinguish two cases.
If $\gamma_{p,\epsilon,s}\le \frac12-\frac1p$, then \eqref{fe32} and \eqref{fe31} lead to
$$\limsup_{N\to\infty}\frac{D(N,p)}{N^{\frac12-\frac1p+\epsilon(\nu,p)}}<\infty.$$
If $\gamma_{p,\epsilon,s}\ge \frac12-\frac1p$, then \eqref{fe32} and \eqref{fe31} lead to
$$\limsup_{N\to\infty}\frac{D(N,p)}{N^{\gamma_{p,\epsilon,s}'}}<\infty.$$
In conclusion, inequality \eqref{fe26} can not hold, and the proof of Theorem \ref{ft1} is complete.
\label{fslast}
|
\section{Introduction}
Frames are redundant bases which turn out in certain applications
to be more flexible than the better known orthonormal bases (ONBs)
in Hilbert space; see, e.g., \cite{MR2837145,MR3167899,MR2367342,MR2147063}.
The frames allow for more symmetries than ONBs do, especially in the
context of signal analysis, and of wavelet constructions; see, e.g.,
\cite{CoDa93,BDP05,Dutkay_2006}. Since frame bases (although containing
redundancies) still allow for efficient algorithms, they have found
many applications, even in finite dimensions; see, for example, \cite{BeFi03,CaCh03,Chr03,Eld02,FJKO05}.
As is well known, when a vector $f$ in a Hilbert space $\mathscr{H}$
is expanded in an orthonormal basis $B$, there is then automatically
an associated Parseval identity. In physical terms, this identity
typically reflects a \emph{stability} feature of a decomposition based
on the chosen ONB $B$. Specifically, Parseval's identity reflects
a conserved quantity for a problem at hand, for example, energy conservation
in quantum mechanics.
The theory of frames begins with the observation that there are useful
vector systems which are in fact not ONBs but for which a Parseval
formula still holds. In fact, in applications it is important to go
beyond ONBs. While this viewpoint originated in signal processing
(in connection with frequency bands, aliasing, and filters), the subject
of frames appears now to be of independent interest in mathematics.
Recently the solution to the Kadison-Singer question, and its equivalent
forms, has proved to take a form involving frames in Hilbert space.
While we will not discuss the Kadison-Singer question here, we mention
it as an illustration of the usefulness of frames in problems which
in their original form were not stated in the language of frames.
The Kadison-Singer question is about uniqueness of extensions of pure
states, where a pure state is a quantum theoretical notion. For interested
readers, we refer to \cite{MR3210706,MR2964016,2014arXiv1407.4768C,2013arXiv1306.3969M}.
This entails systems of vectors $\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}\subset\mathscr{H}$,
called \emph{frames}, where there are two constants $a,b$, s.t. $0<a\leq b<\infty$,
satisfying:
\begin{equation}
a\left\Vert f\right\Vert _{\mathscr{H}}^{2}\leq\sum_{n=1}^{\infty}\left|\left\langle \varphi_{n},f\right\rangle _{\mathscr{H}}\right|^{2}\leq b\left\Vert f\right\Vert _{\mathscr{H}}^{2},\quad\forall f\in\mathscr{H}.\label{eq:d1}
\end{equation}
The two constants $a$ and $b$ in (\ref{eq:d1}) are called \emph{frame
bounds}. (\emph{Parseval frame} means $a=b=1$.)
The purpose of this paper is to associate canonical \emph{projection
valued measures} (PVM) to system of vectors $\left\{ \varphi_{n}\right\} $
which are not frames but such that
\begin{equation}
\sum_{j\in\mathbb{N}}\left|\left\langle \varphi_{j},\varphi_{n}\right\rangle _{\mathscr{H}}\right|^{2}<\infty,\quad\forall n\in\mathbb{N};
\end{equation}
and, from this, then to generate systems of \emph{frames} in suitable
canonical closed subspaces of $\mathscr{H}$, with the subspaces derived
from the PVM, and with a variety of choices of \emph{frame bounds}.
As an application, we give natural and explicit conditions on these
frame systems to generate reproducing kernel Hilbert spaces (RKHS).
The RKHSs have been studied extensively since the pioneering papers
by Aronszajn in the 1940ties, see e.g., \cite{Aro43,Aro48}. They
further play an important role in the theory of partial differential
operators (PDO); for example as Green's functions of second order
elliptic PDOs; see e.g., \cite{Nel57,HKL14}. Other applications include
engineering, physics, machine-learning theory (see \cite{KH11,SZ09,CS02}),
stochastic processes (e.g., Gaussian free fields), numerical analysis,
and more. See, e.g., \cite{AD93,ABDdS93,AD92,AJSV13,AJV14}, and also
\cite{MR2913695,MR2975345,MR3091062,MR3101840}.
\section{Gramians}
Let $\mathscr{H}$ be a Hilbert space, and $\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}$
an indexed system of vectors in $\mathscr{H}$; then we say that the
$\infty\times\infty$ matrix
\begin{equation}
G=\left(\left\langle \varphi_{i},\varphi_{j}\right\rangle _{\mathscr{H}}\right)_{i,j=1}^{\infty}\label{eq:m1}
\end{equation}
is a \emph{Gramian}. Gramians are so named since they were first used
in the Gram-Schmidt orthogonalization; but, more recently, they have
played an important role in the analysis of frames. In order to get
frames, it is natural to further assume that
\begin{equation}
\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}^{\perp}=\left\{ 0\right\} ,\label{eq:m2}
\end{equation}
i.e., that the vectors span a dense subspace in $\mathscr{H}$.
In order to understand frame properties indexed by such systems, it
is necessary to interpret $G$ in (\ref{eq:m1}) as an operator in
$l^{2}$. This works if we further assume:
\begin{equation}
\text{For \ensuremath{\forall n\in\mathbb{N},\quad}\ensuremath{\sum_{j\in\mathbb{N}}\left|\left\langle \varphi_{j},\varphi_{n}\right\rangle _{\mathscr{H}}\right|^{2}<\infty}.}\label{eq:m3}
\end{equation}
In that case, there is an operator $T_{G}$ as follows:
\begin{equation}
\left(T_{G}c\right)_{n}=\sum_{j=1}^{\infty}\left\langle \varphi_{n},\varphi_{j}\right\rangle _{\mathscr{H}}c_{j},\quad\forall c=\left(c_{j}\right)\in l^{2},\label{eq:m4}
\end{equation}
or $c$ in a dense subspace of $l^{2}$. More precisely, $T_{G}$
defines a Hermitian semi-bounded operator $T_{G}:l^{2}\rightarrow l^{2}$
with dense domain (finite sequences) in $l^{2}$. But in general,
$T_{G}$ (from (\ref{eq:m4})) is unbounded, its domain is only dense
in $l^{2}$; and it may not even be essentially selfadjoint.
By von Neumann's theory \cite{DS88b}, $T_{G}$ is essentially selfadjoint
if and only if the following holds:
\begin{equation}
\left\{ \xi\in l^{2}\:\big|\: T_{G}^{*}\xi=-\xi\right\} =0.\label{eq:m5}
\end{equation}
One checks that (\ref{eq:m5}) holds precisely when
\begin{equation}
\Big[\sum_{j=1}^{\infty}\left\langle \varphi_{j},\varphi_{n}\right\rangle _{\mathscr{H}}\xi_{j}=-\xi_{n}\Big]\Longrightarrow\xi=0\;\mbox{in}\; l^{2}.\label{eq:m6}
\end{equation}
Using a theorem from \cite{MR0438178,MR507913}, one further checks
that (\ref{eq:m5}), or equivalently, (\ref{eq:m6}), holds if the
Gramian satisfies:
$\exists\left\{ b_{n}\right\} \subset\mathbb{R}_{+}$ such that
\begin{equation}
\begin{split} & \sum_{j\neq n}\left|\left\langle \varphi_{j},\varphi_{n}\right\rangle _{\mathscr{H}}\right|=O\left(b_{n}\right),\;\mbox{and}\\
& \sum_{n}\frac{1}{\sqrt{b_{n}}}=\infty.
\end{split}
\label{eq:m7}
\end{equation}
\begin{rem}
Using Aronszajn's theorem, one shows that the following two (\ref{enu:m1})\&(\ref{enu:m2})
are the ``same'':
\begin{enumerate}
\item \label{enu:m1}$p:\mathbb{Z}_{+}\times\mathbb{Z}_{+}\rightarrow\mathbb{C}$
positive definite function; and
\item \label{enu:m2}a system $(\mathscr{H},\left\{ \varphi_{i}\right\} _{i\in\mathbb{Z}_{+}}\subset\mathscr{H})$
s.t. $\left\{ \varphi_{i}\right\} ^{\perp}=0$, and
\begin{equation}
p\left(i,j\right)=\left\langle \varphi_{i},\varphi_{j}\right\rangle _{\mathscr{H}},\quad\left(i,j\right)\in\mathbb{Z}_{+}\times\mathbb{Z}_{+}.\label{eq:mm1}
\end{equation}
\end{enumerate}
Let $\mathscr{F}\subset l^{2}$ denote the (dense) subspace of all
\emph{finite} sequence: If
\begin{equation}
\sum_{j}\left|p\left(i,j\right)\right|^{2}=\sum_{j}\left|\left\langle \varphi_{i},\varphi_{j}\right\rangle \right|^{2}<\infty,\quad\forall i\in\mathbb{Z}_{+},\label{eq:mm2}
\end{equation}
then we get an operator $T_{p}$ in $l^{2}$ as follows:
\begin{equation}
\left(T_{p}c\right)_{i}:=\sum_{j}p\left(i,j\right)c_{j},\quad\forall c\in\mathscr{F}\label{eq:mm3}
\end{equation}
defines a Hermitian semibounded operator (generally unbounded) in
$l^{2}$ with dense domain $dom(T_{p})=\mathscr{F}$, i.e.,
\begin{equation}
\left\langle c,T_{p}c\right\rangle _{l^{2}}\geq0,\quad\forall c\in\mathscr{F}.\label{eq:mm4}
\end{equation}
\end{rem}
\section{Main theorems}
\begin{defn}
Let $\mathscr{H}$ be a Hilbert space, and let $\left\{ \varphi_{n}\right\} _{n=1}^{\infty}\subset\mathscr{H}$
be a system of vectors s.t. $\left\{ \varphi_{n}\right\} ^{\perp}=0$,
i.e., the span is dense in $\mathscr{H}$. Let $a,b\in\mathbb{R}_{+}$,
be given such that $0<a\leq b<\infty$; and let $\mathscr{K}\subset\mathscr{H}$
be a closed subspace. We say that $\mathscr{K}$ is an $\left(a,b\right)$
\emph{frame subspace} iff (Def)
\begin{equation}
a\left\Vert f\right\Vert ^{2}\leq\sum_{n=1}^{\infty}\left|\left\langle \varphi_{n},f\right\rangle \right|^{2}\leq b\left\Vert f\right\Vert ^{2},\label{eq:f1}
\end{equation}
holds for all vectors $f\in\mathscr{K}$. Here $\left\langle \cdot,\cdot\right\rangle $
denotes the inner product in $\mathscr{H}$, and $\left\Vert f\right\Vert =\left\langle f,f\right\rangle ^{\frac{1}{2}}$
the corresponding norm. \end{defn}
\begin{thm}
\label{thm:mspace}Let $\mathscr{H}$, $\left\{ \varphi_{n}\right\} $
be a pair, $\mathscr{H}$ a Hilbert space, and $\left\{ \varphi_{n}\right\} _{1}^{\infty}\subset\mathscr{H}$,
satisfying $\left\{ \varphi_{n}\right\} ^{\perp}=0$. Assume further
that
\begin{equation}
\sum_{j=1}^{\infty}\left|\left\langle \varphi_{n},\varphi_{j}\right\rangle \right|^{2}<\infty,\quad\forall n\in\mathbb{N},\;\mbox{and}\label{eq:f2a}
\end{equation}
\begin{equation}
\Big[\xi\in l^{2},\;\sum_{j}\left\langle \varphi_{n},\varphi_{j}\right\rangle \xi_{j}=-\xi_{n}\Big]\Longrightarrow\xi=0\;\mbox{in}\; l^{2}.\label{eq:f2b}
\end{equation}
Then, for every $a,b\in\mathbb{R}_{+}$, s.t. $0<a\leq b<\infty$,
there is a unique \uline{maximal} subspace $\mathscr{K}=\mathscr{H}\left(a,b\right)$
which satisfies the frame condition (\ref{eq:f1}).\end{thm}
\begin{proof}
We begin with a specification of a pair of dense subspaces in $\mathscr{H}$,
and in $l^{2}$, $\mathscr{D}_{\mathscr{H}}$, and $\mathscr{D}_{l^{2}}$:
\begin{itemize}
\item[] $\mathscr{D}_{\mathscr{H}}=span\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}$,
i.e., all finite linear combination of these vectors, and
\item[] $\mathscr{D}_{l^{2}}=span\left\{ \varepsilon_{n}\right\} $, where
$\varepsilon_{n}\left(j\right)=\delta_{nj}$; i.e., $\mathscr{D}_{l^{2}}$
consists of all finitely supported sequences.
\end{itemize}
We introduce two densely defined operators
\begin{equation}
\mathscr{H}\xrightarrow{\; L\;}l^{2},\quad\mbox{and}\quad l^{2}\xrightarrow{\; M\;}\mathscr{H}
\end{equation}
as follows: Set the respective domains,
\begin{equation}
dom\left(L\right)=\mathscr{D}_{\mathscr{H}},\quad\mbox{and}\quad dom\left(M\right)=\mathscr{D}_{l^{2}},
\end{equation}
and set
\begin{eqnarray}
Lf & = & \left(\left\langle \varphi_{n},f\right\rangle \right)_{n},\quad\forall f\in\mathscr{D}_{\mathscr{H}};\label{eq:f3}\\
M\xi & = & \sum_{n=1}^{\infty}\xi_{n}\varphi_{n},\quad\forall\xi\in\mathscr{D}_{l^{2}}.\label{eq:f4}
\end{eqnarray}
From the definition of these two operators, we get the following symmetry:
\begin{equation}
\left\langle Lf,\xi\right\rangle _{l^{2}}=\left\langle f,M\xi\right\rangle _{\mathscr{H}},\quad\forall f\in\mathscr{D}_{\mathscr{H}},\:\forall\xi\in\mathscr{D}_{l^{2}},\label{eq:f5}
\end{equation}
where the respective inner products are indicated with subscripts,
$\left\langle \cdot,\cdot\right\rangle _{\mathscr{H}}$ and $\left\langle \cdot,\cdot\right\rangle _{l^{2}}$.
If $L^{*}$ and $M^{*}$ denote the respective adjoint operators,
then
\begin{equation}
l^{2}\xrightarrow{\; L^{*}\;}\mathscr{H},\quad\mbox{and}\quad\mathscr{H}\xrightarrow{\; M^{*}\;}l^{2}.\label{eq:f6}
\end{equation}
We will need the following:\end{proof}
\begin{lem}
\label{lem:LM}We have
\begin{equation}
L\subseteq M^{*},\quad\mbox{and}\quad M\subseteq L^{*},\label{eq:f7}
\end{equation}
where the containments in (\ref{eq:f7}) refer to the respective graphs,
i.e., (\ref{eq:f7}) states that
\begin{equation}
G\left(L\right)\subseteq G\left(M^{*}\right),\quad\mbox{and}\quad G\left(M\right)\subseteq G\left(L^{*}\right).\label{eq:f8}
\end{equation}
By the graph of a linear operator $T:\mathscr{H}_{1}\rightarrow\mathscr{H}_{2}$
between Hilbert spaces, we mean
\begin{equation}
G\left(T\right)=\left\{ \begin{pmatrix}h_{1}\\
Th_{1}
\end{pmatrix}\:\Big|\: h_{1}\in dom\left(T\right)\right\} \subset\begin{pmatrix}\underset{\bigoplus}{\mathscr{H}_{1}}\\
\mathscr{H}_{1}
\end{pmatrix}.\label{eq:f9}
\end{equation}
\end{lem}
\begin{proof}
The conclusion in the lemma is immediate from (\ref{eq:f5}).\end{proof}
\begin{cor}
Both of the operators $L$ and $M$ in (\ref{eq:f3})-(\ref{eq:f4})
are \uline{closable}, i.e., the closure of $G\left(L\right)$ is
a densely defined closed operator $\mathscr{H}\xrightarrow{\;\overline{L}\;}l^{2}$,
and the closure of $G\left(M\right)$ is a densely defined closed
operator $l^{2}\xrightarrow{\;\overline{M}\;}\mathscr{H}$.\end{cor}
\begin{proof}
Follows from the fact that a densely defined operator $\mathscr{H}_{1}\xrightarrow{\; T\;}\mathscr{H}_{2}$
is closable if and only if the domain of the adjoint $T^{*}$ is dense
in $\mathscr{H}_{2}$. The assertion that both $L^{*}$ and $M^{*}$
have dense domains is immediate from (\ref{eq:f7}).\end{proof}
\begin{lem}
\label{lem:closure}Let $L$ and $M$ be the two operators in Lemma
\ref{lem:LM}, then
\begin{equation}
\overline{L}=M^{*},\quad\mbox{and}\quad\overline{M}=L^{*}.\label{eq:f10}
\end{equation}
\end{lem}
\begin{proof}
From (\ref{eq:f7}), we automatically get containments:
\begin{equation}
\overline{L}\subseteq M^{*},\quad\mbox{and}\quad\overline{M}\subseteq L^{*},\label{eq:f11}
\end{equation}
but the conclusion (\ref{eq:f10}) in the lemma, is that both containments
(\ref{eq:f11}) are equality.
Returning to the definitions (\ref{eq:f3}) and (\ref{eq:f4}), we
note that (\ref{eq:f10}) follows if we prove the following two implications:
\begin{equation}
\left\{ \begin{split} & f\in dom(M^{*}),\;\mbox{and}\\
& \left\langle \varphi_{n},f\right\rangle _{\mathscr{H}}+\sum_{j=1}^{\infty}\left\langle \varphi_{n},\varphi_{j}\right\rangle _{\mathscr{H}}\left(M^{*}f\right)_{j}=0
\end{split}
\right\} \Longrightarrow f=0\;\mbox{in}\;\mathscr{H}.\label{eq:f12}
\end{equation}
And
\begin{equation}
\left\{ \begin{split} & \xi\in dom(L^{*}),\;\mbox{and}\\
& \sum_{j=1}^{n}\left\langle \varphi_{n},\varphi_{j}\right\rangle _{\mathscr{H}}\xi_{j}=-\xi_{n},\;\forall n\in\mathbb{N}
\end{split}
\right\} \Longrightarrow\xi=0\;\mbox{in}\; l^{2}.\label{eq:f13}
\end{equation}
Using the symmetry (\ref{eq:f11}), we note that either one of the
two implications (\ref{eq:f12}), or (\ref{eq:f13}), implies the
other.
We will prove (\ref{eq:f13}). For this, we make use of assumption
(\ref{eq:f2a}) and (\ref{eq:f2b}) from the premise in Theorem \ref{thm:mspace}.
Here the $\infty\times\infty$ matrix $\left(\left\langle \varphi_{n},\varphi_{j}\right\rangle _{\mathscr{H}}\right)$
defines a Hermitian symmetric operator $T$ with dense domain in $l^{2}$,
\begin{equation}
\left(Tc\right)_{n}:=\sum_{j}\left\langle \varphi_{n},\varphi_{j}\right\rangle _{\mathscr{H}}c_{j},\quad\forall c=\left(c_{j}\right)\in dom\left(T\right).\label{eq:f14}
\end{equation}
Since $T$ is semibounded, i.e.,
\begin{equation}
\left\langle c,Tc\right\rangle _{l^{2}}\geq0,\quad\forall c\in dom\left(T\right),\label{eq:f15}
\end{equation}
it follows from assumption (\ref{eq:f2b}) that $T$ has deficiency
indices $\left(0,0\right)$; and in particular that the implication
(\ref{eq:f13}) holds.
This completes the proof of Lemma \ref{lem:closure}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:mspace}, continued.]
We conclude from the lemmas that $L_{1}:=\overline{L}$ (closure)
is a closed operator with dense domain in $\mathscr{H}$, and $\mathscr{H}\xrightarrow{\;\overline{L}\;}l^{2}$.
By a theorem of von Neumann \cite{DS88b}, we further conclude that
the two operators $L^{*}L_{1}$, and $L_{1}L^{*}$, are both selfadjoint
(s.a.); the first in $\mathscr{H}$, and the second in $l^{2}$ (not
merely Hermitian symmetric), but both have deficiency indices $\left(0,0\right)$,
and both are semibounded, i.e., with spectrum in $\left[0,\infty\right]$.
Since $L^{*}L_{1}$ is s.a. with dense domain in $\mathscr{H}$, it
has a unique spectral resolution in the form of a projection valued
measure $P\left(\cdot\right)$, i.e.,
\begin{enumerate}
\item $P\left(A\right)=P\left(A\right)^{*}=P\left(A\right)^{2}$, $\forall A\in\mathscr{B}:=\mathscr{B}\left([0,\infty)\right)$;
\item $P\left(A\cap B\right)=P\left(A\right)P\left(B\right)$, $A,B\in\mathscr{B}$;
\item $P\left(\cdot\right)$ is sigma-additive;
\item $P\left([0,\infty)\right)=I_{\mathscr{H}}$, the identity operator
in $\mathscr{H}$; and
\item $L^{*}L_{1}=\int_{0}^{\infty}\lambda P\left(d\lambda\right)$.
\end{enumerate}
More generally, if $\varphi$ is a Borel function, then by functional
calculus the operator $\varphi\left(L^{*}L_{1}\right)$ is:
\begin{equation}
\varphi\left(L^{*}L_{1}\right)=\int_{0}^{\infty}\varphi\left(\lambda\right)P\left(d\lambda\right);\label{eq:g2}
\end{equation}
and the following are equivalent:
\begin{equation}
f\in dom\left(\varphi\left(L^{*}L_{1}\right)\right)\Longleftrightarrow\int_{0}^{\infty}\left|\varphi\left(\lambda\right)\right|^{2}\left\Vert P\left(d\lambda\right)f\right\Vert _{\mathscr{H}}^{2}<\infty.\label{eq:g3}
\end{equation}
If $\alpha=\left[a,b\right]$, $0<a\leq b<\infty$, i.e., a fixed
compact interval in $\left(0,\infty\right)$, set
\begin{equation}
\mathscr{H}_{\alpha}=P\left(\alpha\right)\mathscr{H}.\label{eq:g4}
\end{equation}
Setting function $\varphi\left(\lambda\right)=\lambda\chi_{\alpha}\left(\lambda\right)$
in (\ref{eq:g2}), we get
\begin{equation}
a\left\Vert f\right\Vert _{\mathscr{H}}^{2}\leq\left\langle f,L^{*}L_{1}f\right\rangle _{\mathscr{H}}\leq b\left\Vert f\right\Vert _{\mathscr{H}}^{2},\quad\forall f\in\mathscr{H}_{\alpha}.\label{eq:g5}
\end{equation}
Since $\left\langle f,L^{*}L_{1}f\right\rangle _{\mathscr{H}}=\sum_{n=1}^{\infty}\left|\left\langle \varphi_{n},f\right\rangle _{\mathscr{H}}\right|^{2}$,
it follows that (\ref{eq:g5}) is the desired frame-estimate, i.e.,
\begin{equation}
a\left\Vert f\right\Vert _{\mathscr{H}}^{2}\leq\sum_{n=1}^{\infty}\left|\left\langle \varphi_{n},f\right\rangle _{\mathscr{H}}\right|^{2}\leq b\left\Vert f\right\Vert _{\mathscr{H}}^{2},\quad\forall f\in\mathscr{H}_{\alpha}.\label{eq:g6}
\end{equation}
Moreover, if $\mathscr{K}\subset\mathscr{H}$ is a closed subspace
in $\mathscr{H}$ s.t. (\ref{eq:g6}) holds for $\forall f\in\mathscr{K}$,
then it follows from (\ref{eq:g2})-(\ref{eq:g4}) that $\mathscr{K}\subset\mathscr{H}_{\alpha}$,
$\alpha=\left[a,b\right]$. This concludes the proof of the theorem.
\end{proof}
The following result is implied by Lemma \ref{lem:closure}, and is
of independent interest:
\begin{prop}
Let $\left(\mathscr{H},\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}\right)$
satisfy the conditions (\ref{eq:f2a})-(\ref{eq:f2b}) in Theorem
\ref{thm:mspace}, and let
\begin{equation}
L_{1}=U\left(L^{*}L_{1}\right)^{\frac{1}{2}}=\left(L_{1}L^{*}\right)^{\frac{1}{2}}U\label{eq:g7}
\end{equation}
be the corresponding polar decomposition of $\mathscr{H}\xrightarrow{\; L_{1}\;}l^{2}$,
then
\begin{equation}
U=T_{G}^{-\frac{1}{2}}L_{1},\;\mbox{and}\label{eq:g8}
\end{equation}
$U$ is an isometry of $\mathscr{H}$ into $l^{2}$.\end{prop}
\begin{proof}
By the properties of the polar decomposition (\ref{eq:g7}), $U$
is a partial isometry with initial space = $\mathscr{H}\ominus\ker\left(L_{1}\right)$;
so we only need to prove that $\ker\left(L_{1}\right)=0$. But from
(\ref{eq:f3}) we get $\ker\left(L_{1}\right)=\left\{ \varphi_{n}\right\} ^{\perp}=0$,
by (\ref{eq:m2}).\end{proof}
\begin{defn}
Let $\Omega$ be a set, and let $\mathscr{H}$ be a Hilbert space
of functions on $\Omega$. We say that $\mathscr{H}$ is a \emph{reproducing
kernel Hilbert} space (RKHS) iff (Def): $\mathscr{H}\ni f\longmapsto f\left(t\right)\in\mathbb{C}$
is norm-continuous for all $t\in\Omega$; or, equivalent, if for all
$t\in\Omega$, there exists a $K_{t}\in\mathscr{H}$ s.t.
\begin{equation}
f\left(t\right)=\left\langle K_{t},f\right\rangle _{\mathscr{H}},\quad\forall f\in\mathscr{H}.\label{eq:e0}
\end{equation}
We say that (\ref{eq:e0}) is the reproducing property. (See, e.g.,
\cite{Aro43,Aro48}.)\end{defn}
\begin{thm}
\label{thm:rkhs}Let $\mathscr{H}$ be a Hilbert space, and $\left\{ \varphi_{n}\right\} _{n\in\mathbb{N}}$
be a \uline{frame} with frame bounds $a,b$, where
\begin{equation}
0<a\leq b<\infty.\label{eq:e1}
\end{equation}
Suppose further that there is a set $\Omega$ s.t. each $\varphi_{n}$
is a function on $\Omega$, and that
\begin{equation}
\left(\varphi_{n}\left(t\right)\right)_{n\in\mathbb{N}}\in l^{2},\quad\forall t\in\Omega.\label{eq:e2}
\end{equation}
Then $\mathscr{H}$ is a reproducing kernel Hilbert space (RKHS) of
functions on $\Omega$. \end{thm}
\begin{proof}
By (\ref{eq:e1})\&(\ref{eq:e2}) we conclude that
\begin{equation}
\sum_{n\in\mathbb{N}}\varphi_{n}\left(t\right)\varphi_{n}\in\mathscr{H},\label{eq:e3}
\end{equation}
and that
\begin{equation}
a\sum_{n}\left|\varphi_{n}\left(t\right)\right|^{2}\leq\left\Vert \sum_{n=1}^{\infty}\varphi_{n}\left(t\right)\varphi_{n}\right\Vert _{\mathscr{H}}^{2}\leq b\sum_{n}\left|\varphi_{n}\left(t\right)\right|^{2}\label{eq:e4}
\end{equation}
holds for all $t\in\Omega$.
For $t\in\Omega$, set
\begin{equation}
l\left(t\right)=\begin{bmatrix}\varphi_{1}\left(t\right)\\
\varphi_{2}\left(t\right)\\
\vdots\\
\vdots
\end{bmatrix}\in l^{2}\label{eq:e5}
\end{equation}
and set
\begin{eqnarray}
K^{G}\left(s,t\right) & := & \left\langle l\left(s\right),G^{-1}l\left(t\right)\right\rangle _{2}\label{eq:e6}\\
& = & \underset{\left(n,m\right)\in\mathbb{N}\times\mathbb{N}}{\sum\sum}\overline{\varphi_{n}\left(s\right)}\left(G^{-1}\right)_{nm}\varphi_{m}\left(t\right).\nonumber
\end{eqnarray}
We claim that $K^{G}\left(\cdot,\cdot\right)$ in (\ref{eq:e6}) turns
$\mathscr{H}$ into a RKHS.
To see this, set
\begin{eqnarray}
Lf & = & \left(\left\langle \varphi_{j},f\right\rangle _{\mathscr{H}}\right)_{j\in\mathbb{N}},\;\mbox{and}\label{eq:e7}\\
L^{*}c & = & \sum_{j=1}^{\infty}c_{j}\varphi_{j}\label{eq:e8}
\end{eqnarray}
then by standard frame theory $L^{*}L$ and $LL^{*}$ are both bounded
and selfadjoint. Moreover, $LL^{*}=T_{G}=$ the operator in $l^{2}$
defined from the Gramian matrix
\begin{equation}
G=\left(\left\langle \varphi_{i},\varphi_{j}\right\rangle _{\mathscr{H}}\right)_{i,j=1}^{\infty}.\label{eq:e9a}
\end{equation}
Since $a,b$ are the frame bounds, we get:
\begin{align}
aI_{\mathscr{H}} & \leq L^{*}L\leq bI_{\mathscr{H}},\;\mbox{and}\label{eq:e9b}\\
aI_{l^{2}} & \leq LL^{*}\leq bI_{l^{2}}\label{eq:e10}
\end{align}
where $I$ denotes the respective identity operators in $\mathscr{H}$,
and in $l^{2}$; and where ``$\leq$'' denotes the natural ordering
of Hermitian (bounded) operators. It further follows that the system,
\begin{equation}
\psi_{n}:=\left(L^{*}L\right)^{-\frac{1}{2}}\varphi_{n},\quad n\in\mathbb{N}
\end{equation}
is a \emph{Parseval} frame in $\mathscr{H}$, i.e.,
\begin{equation}
\left\Vert f\right\Vert _{\mathscr{H}}^{2}=\sum_{n=1}^{\infty}\left|\left\langle \psi_{n},f\right\rangle _{\mathscr{H}}\right|^{2}\label{eq:e12}
\end{equation}
holds for all $f\in\mathscr{H}$.
Using (\ref{eq:e3}), we see that every $f\in\mathscr{H}$ identifies
with a function on $\Omega$; and we shall use the same notation for
this function representations. By (\ref{eq:e12}),
\begin{equation}
f=\sum_{n=1}^{\infty}\left\langle \psi_{n},f\right\rangle _{\mathscr{H}}\psi_{n}\label{eq:e13}
\end{equation}
holds for $f\in\mathscr{H}$, with norm-convergence. Hence (using
the function-representation):
\begin{eqnarray*}
f\left(t\right) & = & \sum_{n=1}^{\infty}\left\langle \psi_{n},f\right\rangle _{\mathscr{H}}\psi_{n}\left(t\right)\\
& = & \left\langle \sum_{n=1}^{\infty}\overline{\psi_{n}\left(\cdot\right)}\psi_{n},f\right\rangle _{\mathscr{H}}\\
& = & \left\langle \sum_{n=1}^{\infty}\overline{\left(\left(L^{*}L\right)^{-\frac{1}{2}}\varphi_{n}\right)}\left(\cdot\right)\left(\left(L^{*}L\right)^{-\frac{1}{2}}\varphi_{n}\right)\left(t\right),f\right\rangle _{\mathscr{H}}\\
& = & \left\langle \left\langle T_{G}^{-\frac{1}{2}}l\left(\cdot\right),T_{G}^{-\frac{1}{2}}l\left(t\right)\right\rangle _{2},f\right\rangle _{\mathscr{H}}\\
& = & \left\langle K^{G}\left(\cdot,t\right),f\right\rangle _{\mathscr{H}},
\end{eqnarray*}
valid for $\forall t\in\Omega$, and all $f\in\mathscr{H}$. This
is the desired RKHS-property for $\mathscr{H}$.
\end{proof}
\section{Applications}
\subsection{Hilbert matrix}
In Theorem \ref{thm:rkhs} we assume that the given Hilbert space
$\mathscr{H}$ has a frame $\left\{ \varphi_{n}\right\} \subset\mathscr{K}$
consisting of functions on a set $\Omega$. So this entails both a
lower, and an upper bound, i.e., $0<B_{1}\leq B_{2}<\infty$. The
following example shows that the conclusion in the theorem is false
if there is \emph{not} a positive lower frame bound.
Set $\mathscr{H}=L^{2}\left(0,1\right)$, $\Omega=\left(0,1\right)$,
the open unit-interval, and $\varphi_{n}\left(t\right)=t^{n}$, $n\in\left\{ 0\right\} \cup\mathbb{N}=\mathbb{N}_{0}$.
In this case, the Gramian
\begin{equation}
G_{n,m}=\int_{0}^{1}x^{n+m}dx=\frac{1}{n+m+1}
\end{equation}
is the $\infty\times\infty$ Hilbert matrix, see \cite{Ros58,Kat57,Tau49}.
In this case it is known that there is an upper frame bound $B_{2}=\pi$,
i.e.,
\begin{equation}
\sum_{n=0}^{\infty}\left|\int_{0}^{1}f\left(x\right)x^{n}dx\right|^{2}\leq\pi\int_{0}^{1}\left|f\left(x\right)\right|^{2}dx;
\end{equation}
in fact, for the operator-norm, we have $\left\Vert G\right\Vert _{l^{2}\rightarrow l^{2}}=\pi$.
Moreover, $G$ defines a selfadjoint operator in $l^{2}\left(\mathbb{N}_{0}\right)$
with spectrum $\left[0,\pi\right]$, the closed interval. This implies
that there cannot be a positive lower frame-bound.
Further, it is immediate by inspection that $\mathscr{H}=L^{2}\left(0,1\right)$
is \emph{not} a RKHS.
\subsection{Random fields}
\begin{defn}[see \cite{MR2509983}]
Let $V$ be a countable discrete set (for example the vertex-set
in an infinite graph \cite{MR3195830,MR3203830}.) Let $\left(\Omega,\mathscr{F},P\right)$
be a probability space, and let $\mathscr{H}:=L^{2}\left(\Omega,P\right)$
be the corresponding Hilbert space of $L^{2}$-random variables, i.e.,
$f:\Omega\rightarrow\mathbb{R}$, measurable s.t.
\begin{equation}
\mathbb{E}(\left|f\right|^{2})=\int_{\Omega}\left|f\right|^{2}dP<\infty.
\end{equation}
A system $\left(\varphi_{x}\right)_{x\in V}\subset\mathscr{H}$ is
said to be a \emph{random frame} if
\begin{equation}
\sum_{y\in V}\left|\mathbb{E}\left(\varphi_{x}\varphi_{y}\right)\right|^{2}<\infty,\quad\forall x\in V;\label{eq:r1}
\end{equation}
and the following implication holds:
\begin{equation}
\Big[\sum_{y\in V}\mathbb{E}\left(\varphi_{x}\varphi_{y}\right)c_{y}=-c_{x},\;\left(c_{x}\right)\in l^{2}\left(V\right)\Big]\Longrightarrow\left(c_{x}\right)=0\;\mbox{in}\; l^{2}\left(V\right).\label{eq:r2}
\end{equation}
\end{defn}
\begin{cor}
Let $\left(\Omega,\mathscr{F},P\right)$ be a probability space and
let $\left(\varphi_{x}\right)_{x\in V}$ be a random frame in $L^{2}\left(\Omega,P\right)$.
Then
\begin{enumerate}
\item \label{enu:r1}the \uline{generalized frame operator}
\begin{equation}
L^{*}L_{1}f=\sum_{x\in V}\mathbb{E}\left(\varphi_{x}f\right)\varphi_{x}\label{eq:rr1}
\end{equation}
is selfadjoint in $L^{2}\left(\Omega,P\right)$, generally unbounded;
and
\item Let $a,b\in\mathbb{R}_{+}$ satisfying $0<a\leq b<\infty$, set $\alpha=\left[a,b\right]$.
Let $P_{E}\left(\cdot\right)$ be the projection valued measure of
$L^{*}L_{1}$, and $\mathscr{H}\left(\alpha\right)=P_{E}\left(\alpha\right)\mathscr{H}$,
then the estimate
\begin{equation}
a\:\mathbb{E}(\left|f\right|^{2})\leq\sum_{x\in V}\left|\mathbb{E}\left(\varphi_{x}f\right)\right|^{2}\leq b\:\mathbb{E}(\left|f\right|^{2})\label{eq:rr2}
\end{equation}
holds for all $f\in\mathscr{H}\left(\alpha\right)$. (i.e., If $P_{E}\left(\cdot\right)$
denotes the spectral resolution of the operator $L^{*}L_{1}$ in (\ref{enu:r1}),
we may take $\mathscr{H}\left(\alpha\right)=P_{E}\left(\alpha\right)\mathscr{H}\left(=P_{E}\left(\left[a,b\right]\right)\mathscr{H}\right)$
in (\ref{eq:rr2}).)
\end{enumerate}
\end{cor}
\begin{acknowledgement*}
The co-authors thank the following colleagues for helpful and enlightening
discussions: Professors Daniel Alpay, Sergii Bezuglyi, Ilwoo Cho,
Ka Sing Lau, Paul Muhly, Myung-Sin Song, Wayne Polyzou, Keri Kornelson,
and members in the Math Physics seminar at the University of Iowa.
\bibliographystyle{amsalpha}
|
\section{Non-split almost complex supermanifolds}
\noindent
Let $(\mathcal M,J)$ be an almost complex supermanifold with sheaf of superfunctions $\mathcal C^\infty_\mathcal M$. Denote the $\mathcal C^\infty_\mathcal M(M)$-module of global superderivations of $\mathcal C^\infty_\mathcal M$ by $\mathcal V_{\mathcal M}=\mathcal V_{\mathcal M,\bar 0}\oplus \mathcal V_{\mathcal M,\bar 1}$. Furthermore fix a Batchelor model $\mathcal M \to (M,\Gamma_{\Lambda E^\ast}^\infty)$ yielding $\mathbb Z$-gradings (denoted by lower indexes) and filtrations (denoted by upper indexes in brackets) on $\mathcal C_\mathcal M^\infty$, $\mathcal V_{\mathcal M}$ and $End(\mathcal V_{\mathcal M})$, the last denoting $\mathcal C^\infty_\mathcal M(M)$-linear maps. The even automorphism of $\mathcal C_\mathcal M^\infty(M)$-modules $J \in End(\mathcal V_{\mathcal M})_{\bar 0}$ can be uniquely decomposed into $J=J_R(Id+J_N)$ with invertible $J_R=J_0$ and nilpotent $J_N$. The finite exp and log series yield a unique representation $Id+J_N=\exp(Y)$ with $Y \in End^{(2)}(\mathcal V_{\mathcal M})_{\bar 0}$.
\begin{lem}\label{le1}
The tensor $J$ is an almost complex structure if and only if $J_R$ is an almost complex structure and $Y J_R+J_R Y=0$.
\end{lem}
\begin{proof}
From $J^2=-Id$ we obtain $J_R^2=-Id$ and $\exp(Y) J_R \exp(Y)=J_R$. For reasons of degree $Y_2 J_R+J_R Y_2=0$. Assume that $Y_{2k} J_R+J_R Y_{2k}=0$ holds for all $k< n$. Set $Y_{[2k]}:=\sum_{j=1}^kY_{2j}$. It is $\exp(Y)=\exp(Y_{[2n-2]})+{Y_{2n}}$ up to terms of degree $>2n$. Hence $\exp(Y) J_R \exp(Y)=\exp(Y_{[2n-2]})\exp(-Y_{[2n-2]})J_R+{Y_{2n} J_R+J_RY_{2n}}$ up to terms of degree $>2n$. This completes the induction. The converse implication follows directly.
\end{proof}
We call $J_R$ the \textit{reduction} of $J$, deforming $J_R$ by $t\mapsto J_R\exp(tY)$. In particular $J_R$ yields an almost complex structure on $M$ and an almost complex structure on the vector bundle $E \to M$. Hence even and odd dimension of $\mathcal M$ are even. Further topological conditions on $M$ and $E$ for the existence of an almost complex structure can be obtained from \cite{Mas} and e.g.~\cite{TE}. Adapted to our considerations the almost complex supermanifold $(\mathcal M,J)$ is {split} if there is a Batchelor model, such that the almost complex structure $J$ has nilpotent component $Y=0$. Note that this problem is completely local since Lemma \ref{le1} allows cutting off the nilpotent $Y$ in $J=J_R\exp(Y)$.
\bigskip\noindent
Let $\Phi=(\varphi,\varphi^\ast)$ be an automorphism of the supermanifold $\mathcal M$. The global even isomorphism of superalgebras $\varphi^\ast \in Aut(\mathcal C_\mathcal M^\infty(M))_{\bar 0}$ over $\varphi$ is decomposable into $\varphi^\ast=\exp(\zeta)\varphi_0^\ast $ with $\zeta \in \mathcal V_{\mathcal M,\bar 0}^{(2)}$ and $\varphi_0^\ast$ preserving the $\mathbb Z$-degree induced by the Batchelor model (see e.g.~\cite{Ro1}). Denote by $Aut(E^\ast)$ the bundle automorphisms over arbitrary diffeomorphisms of $M$, then $\varphi_0^\ast$ is induced by an element $\varphi_0 \in Aut(E^\ast)$ over $\varphi$.
The automorphism $\varphi^\ast$ transforms $J$ into $\varphi^\ast.J$ given by $(\varphi^\ast.J)(\chi):=\varphi^\ast(J((\varphi^\ast)^{-1}\chi \varphi^\ast))(\varphi^\ast)^{-1}$.
Denoting $ad(\zeta):=[\zeta,\cdot]$, assuming $\zeta \in \mathcal V_{\mathcal M,\bar 0}^{(2k)}$ and applying $\varphi^\ast=(Id+\zeta)\varphi_0^\ast$ up to terms in $\mathcal V_{\mathcal M,\bar 0}^{(4k)}$, it is
$\varphi^\ast.J=\varphi_0^\ast.J+[ad(\zeta),\varphi_0^\ast.J]$ up to terms in $End^{(4k)}(\mathcal V_{\mathcal M})_{\bar 0}$.
Comparing both sides with respect to the degree yields:
\begin{prop}\label{lem2}
The almost complex supermanifolds $(\mathcal M,J)$ and $(\mathcal M,J^\prime)$ with structures $J=J_R \exp(Y)$, $J^\prime=J_R^\prime \exp(Y^\prime) $, $Y,Y^\prime \in End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0}$ are isomorphic up to error terms in $End^{(4k)}(\mathcal V_{\mathcal M})_{\bar 0}$ via an automorphism $\varphi^\ast$ with $\varphi^\ast(\varphi_0^\ast)^{-1} \in \exp(\mathcal V_{\mathcal M,\bar 0}^{(2k)})$ if and only if there exist $\varphi_0 \in Aut(E^\ast)$ and $\zeta \in \mathcal V_{\mathcal M,\bar 0}^{(2k)}$ such that $J_R^\prime=\varphi_0^\ast.J_R$ and:
\begin{align*}
Y_{2j}^\prime=\varphi_0^\ast.Y_{2j}-ad(\zeta_{2j})-J_R^\prime ad(\zeta_{2j})J_R^\prime, \quad k\leq j<2k
\end{align*}
\end{prop}
From now on we fix the reduction $J_R$ and hence assume that for an automorphism $\psi^\ast$ of $\mathcal M$, the map $\psi_0^\ast$ is pseudo-holomorphic with respect to $J_R$, denoted $\psi_0^\ast\in Hol(\mathcal M,J_R)$. Let $Hol(\mathcal M,J_R,2k)$ be the automorphisms $\psi^\ast=\exp(\xi)\psi^\ast_0$ of $\mathcal M$ such that $J_R=\psi^\ast.J_R$ up to terms in $End^{(2k)}(\mathcal V_\mathcal M)_{\bar 0}$. Note that $\psi^\ast\in Hol(\mathcal M,J_R,2k)$ includes $\psi_0^\ast \in Hol(\mathcal M,J_R)$ and that $\exp(\mathcal V_{\mathcal M,\bar 0}^{(2k)})\subset Hol(\mathcal M,J_R,2k)$ is a normal subgroup.
\bigskip\noindent
Define on the endomorphisms of real vector spaces $End_{\mathbb R}(\mathcal V_{\mathcal M})$ the $\mathcal C^\infty_\mathcal M(M)$-linear $\mathbb Z$-degree preserving map: $$F_{J_R}:End_{\mathbb R}(\mathcal V_{\mathcal M})\to End_{\mathbb R}(\mathcal V_{\mathcal M}), \quad F_{J_R}(\gamma):=\gamma+J_R \gamma J_R$$
The set $F_{J_R}(End^{(2k)}(\mathcal V_{\mathcal M}))$ is by Lemma \ref{le1} exactly the nilpotent parts $Y$ of almost complex structures $J=J_R\exp(Y)$ deforming $J_R$ in degree $2k$ and higher. Note further that $F_{J_R} (ad(\mathcal V_{\mathcal M})) \subset End(\mathcal V_{\mathcal M})$ and more precisely $F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{(2k)})) \subset End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0}$.
\begin{de} Let the upper index $2k\in 2\mathbb N$ in curly brackets denote the sum of terms of $\mathbb Z$-degree $2k$ up to $4k-2$. For $J=J_R\exp(Y)$, $Y\in End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0}$ we call the class $[Y^{\{2k\}}]$ in the quotient of vector spaces $F_{J_R} ( End^{\{2k\}}(\mathcal V_{\mathcal M})_{\bar 0})/F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{2k\}}))$ the $2k$-th split obstruction class of $J$.
\end{de}
The $Hol(\mathcal M,J_R,2k)$-action on $F_{J_R} ( End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0})$ is given up to terms in $End^{(4k)}(\mathcal V_\mathcal M)_{\bar 0}$ by $(\psi^\ast,Y) \mapsto J_R(J_R-\psi^\ast.J_R)+\psi^\ast.Y$. Since $\psi^\ast.F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{2k\}}))\subset F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{2k\}}))$, it is well-defined on $F_{J_R} ( End^{\{2k\}}(\mathcal V_{\mathcal M})_{\bar 0})/F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{2k\}}))$. By
Proposition \ref{lem2} it induces an action of $PHol(\mathcal M,J_R,2k):=Hol(\mathcal M,J_R,2k)/\exp(\mathcal V_{\mathcal M,\bar 0}^{(2k)})$ on $F_{J_R} ( End^{\{2k\}}(\mathcal V_{\mathcal M})_{\bar 0})/F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{2k\}}))$.
It follows that for an almost complex supermanifold that is split up to terms of degree $2k$ and higher, the $2k$-th split obstruction class is well-defined up to the $PHol(\mathcal M,J_R,2k)$-action.
Note that for a given almost complex structure $J=J_R\exp(Y)$ the obstructions can be checked starting with $j=1$ iteratively: if
$Y_{2j}=ad(\zeta_{2j})+J_R ad(\zeta_{2j})J_R$ can be solved for a $\zeta_{2j}\in \mathcal V_{\mathcal M,2j}$ then there is an automorphism of the supermanifold $\mathcal M$ such that $J=J_R \exp(Y^\prime)$ with $Y^\prime\in End^{(2(j+1))}(\mathcal V_{\mathcal M})_{\bar 0}$. In the non-split case this procedure ends with a well-defined $2k$ and associated orbit of $2k$-th split obstruction classes. We note as a special case:
\begin{prop}
Let $(\mathcal M,J_R)$ be a split almost complex supermanifold of odd dimension $2(2m+r)$, $m\geq 0$, $r \in \{0,1\}$. The almost complex supermanifolds $(\mathcal M,J)$ with reduction $J_R$ that are split up to terms of degree $(2m+r)+1$ and higher, correspond bijectively to the $PHol(\mathcal M,J_R,2(m+1))$-orbits on $F_{J_R} ( End^{(2(m+1))}(\mathcal V_{\mathcal M})_{\bar 0})/F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{(2(m+1))}))$.
\end{prop}
As a technical tool we note an identification for the quotient appearing in the split obstruction classes. Denote by $\mathcal E^1_\mathcal M=\mathcal E^1_{\mathcal M,\bar 0}\oplus \mathcal E^1_{\mathcal M,\bar 1}$ the global super-1-forms on $\mathcal M$ and by $\mbox{d}_\mathcal M$ the de Rham operator on the algebra $\mathcal E_\mathcal M$ of superforms. It is $End(\mathcal V_\mathcal M)=\mathcal V_\mathcal M \otimes_{\mathcal C^\infty_\mathcal M(M)} \mathcal E^{1}_\mathcal M $.
\begin{prop} \label{prop1}
For all $k$, the map
\begin{align*}
\Theta_{J_R}:\ F_{J_R}\big(\mathcal V_\mathcal M \otimes_{\mathcal C^\infty_\mathcal M(M)} \mathcal E^{1}_\mathcal M \big)^{(2k)}_{\bar 0} & \longrightarrow F_{J_R} ( End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0})/F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{(2k)}))
\end{align*}
locally for homogeneous arguments defined by $$F_{J_R}(\chi\otimes {\rm d}_\mathcal M f) \longmapsto (-1)^{|f||\chi|}f\cdot F_{J_R}( ad(\chi))+F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{(2k)}))$$
is a well-defined, surjective morphism of $\mathbb Z$-filtered super vector spaces. For any element $\psi^\ast \in Hol(\mathcal M,J_R,2k)$ and $[\psi^\ast] \in PHol(\mathcal M,J_R,2k)$ it is $\Theta_{J_R}(\psi^\ast.Z)=[\psi^\ast].(\Theta_{J_R}(Z))$.
\end{prop}
\begin{proof}
For homogeneous components of $\chi\otimes {\rm d}_\mathcal M f \in \mathcal V_\mathcal M \otimes_{\mathcal C^\infty_\mathcal M(M)} \mathcal E^{1}_\mathcal M$ the decomposition
$\chi\otimes {\rm d}_\mathcal M f=(-1)^{|f||\chi|}(f\cdot ad(\chi)-ad(f\chi))$
is well-defined up to terms in $ad(\mathcal V_{\mathcal M})$.
\end{proof}
\section{Non-Split Riemannian supermanifolds}
\noindent
Let $(\mathcal M,g)$ be a Riemannian supermanifold with even non-degenerate supersymmetric form $g\in Hom(\mathcal V_\mathcal M \otimes_{\mathcal C^\infty_\mathcal M(M)} \mathcal V_\mathcal M,\mathcal C^\infty_\mathcal M(M))_{\bar 0}$. Here we will mostly regard $g$ as an isomorphism of $\mathcal C^\infty_\mathcal M$-modules $g \in Hom(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$ with $g(X)(Y)=(-1)^{|X||Y|}g(Y)(X)$ for homogeneous arguments. The context will fix which point of view is used. For a given Batchelor model $\mathcal M \to (M,\Gamma_{\Lambda E^\ast}^\infty)$ decompose $g=g_R(Id+g_N)$ with invertible $g_R=g_0+g_2$ and nilpotent $g_N\in End^{(2)}(\mathcal V_\mathcal M)_{\bar 0}$ such that $g_0g_N \in Hom^{(4)}(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$. With the finite log and exp series we write $g=g_R\exp(W)$ with $W \in End^{(2)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$, where $End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$ denotes those $W \in End^{(2k)}(\mathcal V_\mathcal M)_{\bar 0}$ such that $g_0W \in Hom^{(2k+2)}(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$.
\begin{lem}\label{le2}
If the tensor $g=g_R\exp(W)$, $W \in End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$ is a Riemannian metric then $g_R$ is a Riemannian metric and $g_R(W(\cdot),\cdot)=g_R(\cdot,W(\cdot))$ up to terms of degree $4k+2$ and higher.
\end{lem}
\begin{proof}
Due to supersymmetry $g_R(\exp(W)(\cdot),\cdot)=g_R(\cdot,\exp(W)(\cdot))$. The approximation $\exp(W)=1+W$ holds up to terms of degree $4k$ with error term $\frac{1}{2}W_{2k}^2$ in degree $4k$. Since $W \in End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$ it is $g_RW_{2k}^2 \in Hom^{(4k+2)}(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$.
\end{proof}
We call $g_R$ the \textit{reduction} of $g$. Here the metric $g$ appears as a deformation of the underlying Riemannian metric $g_{0}$ on $M$ via $t\mapsto (g_{0}+t\cdot g_{2})\exp(\sum_{j=1}^\infty t^jW_{2j})$.
Note that $g_R$ also yields a non-degenerate alternating form on the bundle $E$. So in contrast to the non-graded case there is a true condition for the existence of a Riemannian metric: the existence of a nowhere vanishing section of $E\wedge E\to M$. In particular the odd dimension of $\mathcal M$ has to be even. A Riemannian supermanifold $(\mathcal M,g)$ is {split}, if there is a Batchelor model, such that the Riemannian metric $g$ has nilpotent component $W=0$. Again the appearing deformations are essentially local via cutting off $g_R\exp(W)$ by $(g_{0}+f\cdot g_{2})\exp(\sum_{j=1}^\infty f^jW_{2j})$ with cut-off function $f$.
\bigskip\noindent
As before let $\Phi=(\varphi,\varphi^\ast)$, $\varphi^\ast=\exp(\zeta)\varphi_0^\ast$ be an automorphism of the supermanifold $\mathcal M$.
We obtain $\varphi^\ast.g$ given by $(\varphi^\ast.g)(\chi)(\chi^\prime)=\varphi^\ast\big( g((\varphi^\ast)^{-1}\chi \varphi^\ast,(\varphi^\ast)^{-1}\chi^\prime \varphi^\ast)\big)$. Assuming $\zeta \in \mathcal V_{\mathcal M,\bar 0}^{(2k)}$ this yields:
\begin{align}\label{asd}
\varphi^\ast.g=\varphi_0^\ast.g-(\varphi_0^\ast.g)(ad(\zeta)\otimes Id + Id \otimes ad(\zeta))+\zeta (\varphi_0^\ast.g)
\end{align}
in $Hom(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$ up to terms in $Hom^{(4k)}(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$. Note that for the term $\zeta (\varphi_0^\ast.g)$, the metric is regarded as an element in $Hom(\mathcal V_\mathcal M\otimes_{\mathcal C^\infty_\mathcal M(M)}\mathcal V_\mathcal M ,\mathcal C^\infty_\mathcal M(M))_{\bar 0}$. Define $\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)}$ to be the elements in $\zeta \in \mathcal V_{\mathcal M,\bar 0}^{(2k)}$ satisfying $g_0(ad(\zeta)\otimes Id + Id \otimes ad(\zeta))+\zeta g_0\in Hom^{(2k+2)}(\mathcal V_\mathcal M,\mathcal V_\mathcal M^\ast)_{\bar 0}$. Comparing the terms in (\ref{asd}) with respect to the degree yields:
\begin{prop}\label{lem3}
The Riemannian supermanifolds $(\mathcal M,g)$ and $(\mathcal M,g^\prime)$ with Riemannian metrics $g=g_R\exp(W)$, $g^\prime=g_R^\prime\exp(W^\prime)$, $W \in End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$, $W^\prime \in End^{(2k)}_{g_0^\prime}(\mathcal V_\mathcal M)_{\bar 0}$ are isomorphic up to error terms in $Hom^{(4k)}(\mathcal V_\mathcal M, \mathcal V_\mathcal M^\ast)_{\bar 0}$ via an automorphism $\varphi^\ast$ with $\varphi^\ast(\varphi_0^{\ast})^{-1}\in\exp(\mathcal V_{\mathcal M,g_0^\prime,\bar 0}^{(2k)})$ if and only if there exist $\varphi_0 \in Aut(E^\ast)$ and $\zeta \in \mathcal V_{\mathcal M,g_0^\prime,\bar 0}^{(2k)}$ such that
$ g_R^\prime=\varphi_0^\ast.g_R$ and:
\begin{align*}
W_{2j}^\prime=\varphi_0^\ast.W_{2j}-ad(\zeta_{2j})-\Big((g_R^\prime)^{-1} (ad^\ast(\zeta)-\zeta) g_R^\prime\Big)_{2j}, \quad k\leq j<2k
\end{align*}
Here $\varphi_0^\ast.W$ is defined by $(\varphi_0^\ast.W)(\chi):=\varphi^\ast_0(W((\varphi_0^\ast)^{-1}\chi \varphi_0^\ast))(\varphi_0^\ast)^{-1}$ and the homomorphism $ad^\ast:\mathcal V_\mathcal M \to End_\mathbb R(End_\mathbb R(\mathcal V_\mathcal M,\mathcal C^\infty_\mathcal M(M)))$ denotes the representation dual to $ad$.
\end{prop}
Fix $g_R$ from now on and denote by $Iso(\mathcal M,g_R,2k+2)$ the automorphisms $\psi^\ast=\exp(\xi)\psi^\ast_0$ of $\mathcal M$ such that $g_R=\psi^\ast.g_R$ up to a term $S:=g_R-\psi^\ast.g_R \in Hom^{(2k+2)}(\mathcal V_\mathcal M,\mathcal V_\mathcal M^\ast)_{\bar 0}$. Note that this forces $g_0g_R^{-1}S \in Hom^{(2k+2)}(\mathcal V_\mathcal M,\mathcal V_\mathcal M^\ast)_{\bar 0}$. Further $\exp(\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})\subset Iso(\mathcal M,g_R,2k+2)$ is a normal subgroup.
\bigskip\noindent
Parallel to the analysis of the almost complex structures we define the maps
\begin{align*}
&F_{g_R}: End(\mathcal V_\mathcal M) \to End(\mathcal V_\mathcal M),&& \ F_{g_R}(\gamma):=\gamma+g_R^{-1} \gamma^\ast g_R\\
&G_{g_R}: \quad \mathcal V_\mathcal M \quad\quad\to End(\mathcal V_\mathcal M),&& \ G_{g_R}(\zeta):=ad(\zeta)+g_R^{-1} (ad^\ast(\zeta)-\zeta) g_R&&&&&&&&&
\end{align*}
denoting by $\gamma^\ast$ the induced element in $End(\mathcal V_\mathcal M^\ast)$ and $ad^\ast$ as above. By Lemma \ref{le2} the elements in $F_{g_R}(End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0})$ are up to degree $\geq 4k+2$ the appearing $W$s in Riemannian metrics $g=g_R\exp(W)$ that are split up to degree $\geq 2k$. Further $G_{g_R}(\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})$ lies in $F_{g_R}(End^{(2k)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0})$.
\begin{de} For $g=g_R\exp(W)$ with $W\in End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0}$ we call the class $[W^{\{2k\}}]$ in the quotient of vector spaces $F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0})^{\{2k\}}/G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})^{\{2k\}}$ the $2k$-th split obstruction class of $g$.
\end{de}
The $Iso(\mathcal M,g_R,2k+2)$-action on $F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0})$ is given up to terms in $End^{(4k)}(\mathcal V_\mathcal M)_{\bar 0}$ by $(\psi^\ast,W)\mapsto g_R^{-1}(\psi^\ast.g_R-g_R)+ \psi^\ast.W$. It is $\psi^\ast.G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})\subset G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})$ by direct calculation. Analog to the almost complex case using Proposition \ref{lem3}, the action of $Iso(\mathcal M,g_R,2k+2)$ induces a $PIso(\mathcal M,g_R,2k+2):=Iso(\mathcal M,g_R,2k+2)/\exp(\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})$-action on the quotient $F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0})^{\{2k\}}/G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})^{\{2k\}}$. Hence the $2k$-th split obstruction class is well-defined up to the $PIso(\mathcal M,g_R,2k+2)$-action for a Riemannian supermanifold that is split up to terms of degree $2k+2$ and higher. We have in particular analogously to the almost complex case:
\begin{prop}
Let $(\mathcal M,g_R)$ be a split Riemannian supermanifold of odd dimension $2(2m+r)$, $m\geq 0$, $r \in\{0,1\}$. The Riemannian supermanifolds $(\mathcal M,g)$ with reduction $g_R$ that are split up to terms of degree $(2m+r)+3$ and higher, correspond bijectively to the $Iso(\mathcal M,g_R,2(m+2))$-orbits on $F_{g_R} ( End^{(2(m+1))}(\mathcal V_{\mathcal M})_{\bar 0})/G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2(m+1))})$.
\end{prop}
Also an analogy to Proposition \ref{prop1} holds:
\begin{prop} \label{prop2}
The map
\begin{align*}
\Theta_{g_R}:\ F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0}) & \longrightarrow F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0})/G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})
\end{align*}
locally defined by $$F_{g_R}(\chi\otimes {\rm d}_\mathcal M f) \longmapsto (-1)^{|f||\chi|}f \cdot G_{g_R}( \chi)+G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)})$$
is a well-defined surjective morphism of $\mathbb Z$-filtered vector spaces. For any element $\psi^\ast$ in $Iso(\mathcal M,g_R,2k+2)$ and $[\psi^\ast] \in PIso(\mathcal M,g_R,2k+2)$ it is $\Theta_{g_R}(\psi^\ast.Z)=[\psi^\ast].(\Theta_{g_R}(Z))$.
\end{prop}
\begin{proof}
Apply $F_{g_R}$ to $\chi\otimes {\rm d}_\mathcal M f=(-1)^{(|f||\chi|}(f\cdot ad(\chi)-ad(f\chi))$ and add $f \chi-f\chi$ in the bracket. This yields a map $ F_{g_R} ( End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0}) \to F_{g_R} ( End^{(2k)}(\mathcal V_{\mathcal M})_{\bar 0})/G_{g_R} (\mathcal V_{\mathcal M,\bar 0}^{(2k)}) $. Since $\chi\otimes {\rm d}_\mathcal M f$ is in $ End^{(2k)}_{g_0}(\mathcal V_{\mathcal M})_{\bar 0}$, its degree $2k$ term is of the form $\sum \tilde f_i \frac{\partial}{\partial \xi_i}\otimes d_\mathcal M \hat f_i$ for an odd coordinate system $(\xi_i)$. This forces $f \chi \in \mathcal V_{\mathcal M,g_0,\bar 0}^{(2k)}$ by direct calculation.
\end{proof}
\section{Examples of global nowhere split structures}\label{sec3}
Here explicit examples of non-split almost complex structures, resp. non-split Riemannian metrics are given. The constructed tensors are nowhere split.
\bigskip\noindent
Let $(M, J_M)$ be an almost complex manifold of dimension $2n$ and let $\mathcal M$ be the supermanifold defined by differential forms, i.e.~$\mathcal C^\infty_\mathcal M=\mathcal E_M$. The vector fields in $\mathcal V_M$ act on $\mathcal C^\infty_\mathcal M$ by Lie derivation. Let further $\pi:\mathcal V_\mathcal M \to \mathcal V_\mathcal M$ be the odd $\mathcal C^\infty_\mathcal M$-linear operator well-defined by $\pi^2=Id$ and $\pi(\chi)(\omega):=\iota_\chi \omega$ for $\chi \in \mathcal V_M\subset\mathcal V_{\mathcal M,\bar 0} $ and $\omega \in \mathcal C^\infty_\mathcal M$.
\bigskip\noindent
By \cite[prop. 4.1]{MS} there exist non-degenerate $2$-forms $\eta \in \mathcal C^\infty_\mathcal M$ compatible with $J_M$. We fix one and denote by $g^\prime$ the $J_M$-invariant Riemannian metric $\eta(\cdot,J_M(\cdot))$ on $M$. Furthermore we embed $$End(\mathcal V_M)\cong( \mathcal V_M \otimes_{\mathcal C^\infty_M(M)} \mathcal E^1_M) \hookrightarrow \mathcal V_\mathcal M\quad \mbox{by} \quad \chi \otimes \alpha \mapsto \xi_{\chi\otimes \alpha}:=\alpha\cdot \chi$$ and obtain $\xi_{Id},\xi_{J_M} \in \mathcal V_{\mathcal M, 1}$ and $\pi(\xi_{Id}),\pi(\xi_{J_M}) \in \mathcal V_{\mathcal M, 0}$.
\bigskip\noindent
We define on $\mathcal M$ by $\mathcal C^\infty_\mathcal M$-linear continuation to $End(\mathcal V_\mathcal M)$, resp. $Hom(\mathcal V_\mathcal M,\mathcal V_\mathcal M^\ast)$:
\begin{itemize}
\item[(i)] the split almost complex structure $J_R$ by $J_M \oplus (\pi\circ J_M\circ\pi) \in End_{\mathcal C^\infty_M}(\mathcal V_M \oplus \pi(\mathcal V_M))$
\item[(ii)] the split Riemannian metric $g_R$ by $g^\prime + (\eta\circ (\pi\otimes\pi)) \in Hom_{\mathcal C^\infty_M}(\mathcal V_M^{\otimes2}\oplus \pi(\mathcal V_M)^{\otimes2},\mathcal C^\infty_M)$ and $g_R(\mathcal V_M \otimes \pi(\mathcal V_M))=0$
\end{itemize}
Note the following technical lemma for later application:
\begin{lem}\label{tech}
Let $f,g \in End(\mathcal V_M)$, $\omega \in \mathcal C^\infty_{\mathcal M,2}$. Then:
\begin{itemize}
\item[a)] $\quad{\pi(\xi_{f})} (\omega) =\frac{1}{2}(\omega(f(\cdot),\cdot)+\omega(\cdot,f(\cdot)))$
\item[b)] $\quad[\pi(\xi_f),\pi(\xi_g)]=-\pi(\xi_{[f,g]})$
\item[c)] $\quad J_R(\xi_{J_M})=-\xi_{Id}\quad$ and $\quad J_R(\pi(\xi_{J_M}))=-\pi(\xi_{Id})$
\end{itemize}
\end{lem}
Further fix in the almost complex, resp. Riemannian case the tensors:
\begin{itemize}
\item[(i)] $J=J_R\exp(\eta \cdot Y_\eta)$ with $Y_{\eta}\in End^{(2)}(\mathcal V_\mathcal M)_{\bar 0}$ by $Y_{\eta}=F_{J_R}(\pi(\xi_{J_M})\otimes {\rm d}_\mathcal M \eta) $
\item[(ii)] $g=g_R\exp(\eta \cdot W_\eta)$ with $W_\eta\in End^{(2)}(\mathcal V_\mathcal M)_{\bar 0}$ by $W_\eta=F_{g_R}(\pi(\xi_{J_M})\otimes d_\mathcal M\eta)$
\end{itemize}
We prove:
\begin{lem}\label{t2}
Assume that $n>1$. The endomorphisms $Y_\eta$ and $W_\eta$ are nowhere vanishing.
\end{lem}
\begin{proof}
With Lemma \ref{tech} c) follows
$Y_{\eta}=\pi(\xi_{J_M})\otimes {\rm d}_\mathcal M \eta-\pi(\xi_{Id})\otimes (({\rm d}_\mathcal M \eta)\circ J_R) $.
Applying $(Y_\eta(\pi(\xi_{J_M})))(\eta)$ we obtain $(\pi(\xi_{J_M})(\eta))^2+(\pi(\xi_{Id})(\eta))^2$.
By Lemma \ref{tech} a) it is $\pi(\xi_{J_M})(\eta)=0$ since $\eta$ is compatible with $J_M$, and $(Y_\eta(\pi(\xi_{J_M})))(\eta)=\eta^2$. So $Y_{\eta}$ is nowhere vanishing.
\smallskip\noindent
For the second statement note
$W_\eta=\pi(\xi_{J_M})\otimes d_\mathcal M\eta+g_R^{-1}(d_\mathcal M \eta)\cdot g_R(\pi(\xi_{J_M}))$.
Further $(W_\eta(\pi(\xi_{J_M})))(\eta)=(\pi(\xi_{J_M})(\eta))^2+(g_R^{-1}(d_\mathcal M \eta))(\eta)\cdot \eta(\xi_{J_M},\xi_{J_M})$. Due to the compatibility of $\eta$ and $J_M$, it is $\eta(\xi_{J_M},\xi_{J_M})=\eta$ and as before, $\pi(\xi_{J_M})(\eta)=0$. Further a calculation yields $(g_R^{-1}(d_\mathcal M \eta))(\eta)=g_R(g_R^{-1}(d_\mathcal M \eta),g_R^{-1}(d_\mathcal M \eta))=\eta$ up to terms of degree $4$ and higher. Hence $(W_\eta(\pi(\xi_{J_M})))(\eta)=\eta^2$ up to terms of degree $6$ and higher. So $W_\eta$ is nowhere vanishing.
\end{proof}
Finally it follows:
\begin{theo}
Assume that $n>2$. The almost complex structure $J=J_R\exp(\eta \cdot Y_\eta)$ and the Riemannian metric $g=g_R\exp(\eta \cdot W_\eta)$ on $\mathcal M$ are nowhere split.
\end{theo}
\begin{proof}
For $\psi^\ast=\exp(\xi)\psi_0^\ast \in Hol(\mathcal M,J_R,4)$ we obtain $[ad(\xi_2),J_R]=0$. With the identity $\exp(\xi_{2}).J_R=\exp(ad(\xi_{2}))J_R \exp(ad(\xi_{2}))=\exp([ad(\xi_{2}),\cdot])(J_R)=J_R$ it follows that $\psi^\ast$ maps $Y\in F_{J_R} ( End^{\{4\}}(\mathcal V_{\mathcal M})_{\bar 0})$ to $F_{J_R}(ad(\xi_4))+\psi^\ast_0.Y$ up to terms of degree $\geq 6$. Hence $F_{J_R} (ad(\mathcal V_{\mathcal M,\bar 0}^{\{4\}}))$ is a $PHol(\mathcal M,J_R,4)$-orbit up to terms of degree $\geq 6$. Using Proposition \ref{prop1} and $\eta Y_\eta=F_{J_R}(\eta \pi(\xi_{J_M})\otimes {\rm d}_\mathcal M \eta)$ it is sufficient to check that $\eta F_{J_R}(ad(\eta \pi(\xi_{J_M})))$ does not vanish in degree $4$. From the proof of Proposition \ref{prop1} we know the identity $F_{J_R}(ad(\eta\pi(\xi_{J_M})))=\eta\cdot F_{J_R}(ad(\pi(\xi_{J_M})))-Y_\eta$. Now Lemma \ref{tech} b) and c) yields that $ F_{J_R}(ad(\pi(\xi_{J_M})))(\pi(\xi_{J_M}))=0$. Hence it is $F_{J_R}(ad(\eta\pi(\xi_{J_M})))(\pi(\xi_{J_M}))(\eta)=-Y_\eta(\pi(\xi_{J_M}))(\eta)$ which is $\eta^2$ as it was shown in the proof of Lemma \ref{t2}. This proves the first statement.
\smallskip\noindent
It is by direct calculation $\eta W_\eta \in End^{(4)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$.
Let $\psi^\ast=\exp(\xi)\psi_0^\ast \in Iso(\mathcal M,g_R,6)$, then the degree $4$ term of $\psi^\ast.g$ vanishes while the degree $6$ term is $\psi_0^\ast.(g_2W_4+g_0W_6)+(\psi^\ast.g_R)_6$. Further $(1+\xi_2)\psi_0^\ast$ preserves $g_R$ and so does $\exp(\xi_2)\psi_0^\ast$. The term $(\exp(\xi_4+\xi_6).g_R)_6$ equals $(g_R(ad(\xi_4+\xi_6)\otimes Id + Id \otimes ad(\xi_4+\xi_6))-(\xi_4+\xi_6) g_R)_6$. So $\psi^\ast$ maps $W \in End^{(4)}_{g_0}(\mathcal V_\mathcal M)_{\bar 0}$ to $G_{g_R}(\xi_4)+\psi_0^\ast(W_4)$ up to terms of degree $\geq 6$. Since $\psi^\ast$ preserves the vanishing degree four term of $g_R$, it follows that $\xi_4 \in \mathcal V_{\mathcal M,g_0,\bar 0}^{\{4\}}$. Hence $G_{g_R} (\mathcal V_{\mathcal M,g_0,\bar 0}^{\{4\}})$ is a $PIso(\mathcal M,g_R,6)$-orbit up to terms of degree $\geq 6$.
Analogue to the almost complex case and following Proposition \ref{prop2} it is sufficient to show that $\eta G_{g_R}(\eta \pi(\xi_{J_M}))$ is nowhere vanishing. We have the identity $G_{g_R}(\eta \pi(\xi_{J_M}))=\eta\cdot G_{g_R}(\pi(\xi_{J_M}))-W_\eta$. Note that for $\alpha \in \mathcal V_\mathcal M^\ast$ it is $(g_R^{-1}(\alpha))(\eta)=\alpha(g_R^{-1}(d_\mathcal M \eta))$ and $g_R^{-1}(d_\mathcal M\eta)=\pi(\xi_{Id})$ up to terms of degree two and higher. Using these details, Lemma \ref{tech} b) and the definition of $g_R$ one obtains $G_{g_R}(\pi(\xi_{J_M}))(\pi(\xi_{J_M}))(\eta)=-\pi(\xi_{J_M})(\eta (\pi(\xi_{Id}) ,\pi(\xi_{J_M})))$ up to terms of degree four and higher. A direct calculation using the graded Leibniz rule and $J_M$-invariance of $\eta$ shows that $G_{g_R}(\pi(\xi_{J_M}))(\pi(\xi_{J_M}))(\eta)$ vanishes up to terms of degree four and higher. Hence $G_{g_R}(\eta \pi(\xi_{J_M}))(\pi(\xi_{J_M}))(\eta)=-W_\eta (\pi(\xi_{J_M}))(\eta)$ up to terms of degree 6 and higher. In the proof of Lemma \ref{t2} it was shown that the degree 4 term of this expression is $\eta^2$. This proves the second statement.
\end{proof}
|
\section*{Notation}
Most of the notation used in this paper is completely
standard. For the convenience of the reader, to avoid
ambiguities, we state it clearly
from the beginning.
The standard Euclidean basis of~$\mathbb R^n$ is denoted by~$e_1,\cdots,e_n$
(so that, in particular, $e_1=(1,0,\cdots,0)$).
If~$E\subseteq\mathbb R^n$ and~$v\in\mathbb R^n$, we use the notation
$$ E+v:=\{ p+v,\ p\in E\}.$$
Also,~$|E|$ is the Lebesgue measure of~$E$.
The $(n-1)$-dimensional Hausdorff measure is denoted by~${\mathcal{H}}^{n-1}$.
The notation~$\Delta$ will be used for the symmetric difference,
i.e.~$E\Delta F:= (E\setminus F)\cup (F\setminus E)$.
We denote by~$\chi_E$ the characteristic function of a set~$E$, i.e.
$$ \chi_E(x):=\left\{
\begin{matrix}
1 & {\mbox{ if }} x\in E,\\
0 & {\mbox{ if }} x\not\in E.
\end{matrix}
\right.$$
Also, throughout the paper, the world ``decreasing''
stands simply for ``non increasing''.
\section{Introduction}
The main goal of this paper is to
construct a nonlocal analogue of
the classical Delaunay surfaces (see~\cite{Del}), i.e.
surfaces that minimize a fractional perimeter functional among cylindrically
decreasing symmetric
competitors that are periodic
in a given direction.
The notion of perimeter that we take into account
is a periodic functional of fractional type,
whose critical points are related to axially symmetric objects.
We also study the main geometric properties of the minimizers,
such as dislocation of mass and closeness to periodic array of balls.
For this scope, we will introduce a new fractional perimeter functional
that takes into account the periodicity of the surfaces
and we develop a fine analysis of the functional in order to
obtain suitable compactness properties.
The setting we work in is the following.
We consider a fractional parameter~$s\in(0,1)$. We
use coordinates $x=(x_1,x')\in\mathbb R\times\mathbb R^{n-1}=\mathbb R^n$, with~$n\ge2$,
and deal with the slab
$$S:=[-1/2,\,1/2]\times\mathbb R^{n-1}.$$
We consider the kernel~$K:\mathbb R^n\setminus(\Z\times\mathbb R^{n-1})\to\mathbb R$,
$$ K(x):=\sum_{k\in\Z}\frac{1}{|x+ke_1|^{n+s}}$$
and, given a set~$E\subseteq \mathbb R^n$, we define
$$ P_S(E):=\int_{E\cap S} \int_{S\setminus E} K(x-y)\,dx\,dy=
\int_{E\cap S} \int_{S\setminus E}
\sum_{k\in\Z}\frac{dx\,dy}{|x-y+ke_1|^{n+s}}.$$
This fractional functional is related to, but quite different from,
the nonlocal perimeter introduced in~\cite{CRS} (namely,
it shares with it some nonlocal features, but it has different scaling
behaviors and periodicity properties).
More precisely,
on the one hand, the functional studied here may be considered
as a periodic version (in the horizontal direction) of
the fractional perimeter in~\cite{CRS}. On the other hand,
the kernel that we consider is non-standard, since it
has different scaling properties in the different coordinate directions.
We consider the class of our competitors\label{456scdfvg}
${\mathcal{K}}$, that is given by the sets~$F\subseteq S$
of the form
$$ F=\big\{ (x_1,x')\in S {\mbox{ with }} |x'|\le f(x_1)\big\},$$
for a given even function~$f:[-1/2,\,1/2]\to[0,+\infty]$ that is
decreasing in~$[0,\,1/2]$.
In this setting, we prove the existence
of volume constrained minimizers of~$P_S$ in~${\mathcal{K}}$:
\begin{theorem}\label{THE EX}
For any~$\mu>0$ there exists a minimizer for~$P_S$ in~${\mathcal{K}}$
with volume
constraint equal to~$\mu$.
More explicitly, for any~$\mu>0$ there exists a set~$F_*\in{\mathcal{K}}$
such that~$|F_*|=\mu$ and, for any~$F\in{\mathcal{K}}$
such that~$|F|=\mu$, we have that~$P_S(F_*)\le P_S(F)$.
\end{theorem}
Recently, in the literature, there has been an intense
effort towards the construction of geometric object of nonlocal
nature that extend classical (i.e. local) ones, see e.g.~\cite{Dav, I5, I4}.
In some cases, the nonlocal objects inherit strong geometric
properties from the classical case, but also important
differences arise.
In our setting, we think it is an interesting problem to determine whether
cylinders are minimizers for large volume.
As for small volumes,
the next result points out (in a quantitative way)
that in this case the minimizing
set does not put a considerable proportion of mass
close to the boundary of the slab (in particular, it is ``far from
being a cylinder''):
\begin{theorem}\label{TH DX}
Let
$$ F_*=
\Big\{ (x_1,x')\in S {\mbox{ s.t. }} |x'|\le f(x_1)\Big\}$$
be a minimizer with volume constraint~$\mu>0$, as given
in Theorem~\ref{THE EX}. Then
\begin{equation}\label{density 0}
\frac{f(1/4)}{ \mu^{\frac{1}{n-1}} }\le C\,\mu^{\frac{s}{n^2(n-1)}},\end{equation}
for some~$C>1$.
In particular, for any~$\delta\in(0,1)$, if~$\mu\in(0, C^{-1}\delta^{\frac{n^2}{s}})$,
we have that
\begin{equation}\label{density 1}
\frac{\big|F_*\cap \{|x_1|\ge 1/4\}\big|}{|F_*|}\le \delta
\end{equation}
for a suitable constant~$C>1$.
\end{theorem}
In case of small volumes, we also show that minimizers are close
to balls.
The notion of closeness will be measured by the so-called
Fraenkel asymmetry (or symmetric deficit) of a set~$E$, which
is defined as
$$ {\rm Def}(E):=\inf
\frac{|E\Delta B|}{|E|},$$
where the infimum is taken
over every ball $B\subset\mathbb R^n$ with $|B|=|E|$.
Roughly speaking, the Fraenkel asymmetry measures the $L^1$ distance
of~$E$ from being a ball of the same volume (the ball
may be conveniently translated in order to cover the set~$E$ as much
as possible, and the quantity above is normalized with
respect to the volume in order to be scale invariant).
In this setting we have:
\begin{theorem}\label{YES BALL}
Let~$F_*\subseteq S$ be a minimizer according to Theorem~\ref{THE EX},
with volume constraint~$\mu$. Then, if~$\mu$ is small enough,
$F_*$ is close to a ball.
More precisely, for any~$\mu\in(0,1)$, we have that
$$ {\rm Def}(F_*)\le C\,
\mu^{\frac{n^2-s^2}{2 n^2(n-1)}}.$$
\end{theorem}
We observe that the control in the sense of deficit obtained in
Theorem~\ref{YES BALL}
may lead to further uniform (though perhaps less explicit)
asymptotic bounds, also
in the $C^1$-sense, by exploiting suitable density estimates
and approximation results, see e.g. Corollary~3.6 in~\cite{I5}
and Theorem 6.1 in~\cite{I4}.
\medskip
We point out that, after the present paper was completed
and submitted for publication, the very interesting
article~\cite{cabre} appeared, dealing
with surfaces of constant fractional mean curvature
(in the sense of~\cite{CRS, AV}).
This paper also
provides very fascinating fractional counterparts of
the Delaunay surfaces (in a different setting than the one
considered here,
and in the planar case, with an announcement of
the $n$-dimensional results to come).
\medskip
The rest of the paper is organized as follows. In Section~\ref{KD}
we study the decay properties of the kernel. Then, in Section~\ref{REL}
we give a detailed comparison between our functional and
the one in~\cite{CRS} (this is not only interesting
for seeing similarities and differences with the existing literature,
but it is also useful for constructing competitors and deriving
estimates).
As a matter of fact, the proof of Theorem~\ref{THE EX}
also requires a careful energy analysis
and ad-hoc compactness arguments in order to
use the direct minimization method: these arguments are collected
in Sections~\ref{EB} and~\ref{CI2}.
With this, all the preliminary work will be completed, and we
will be able to prove
Theorems~\ref{THE EX}, \ref{TH DX}
and~\ref{YES BALL} in Sections~\ref{S1}, \ref{S4}
and~\ref{S5}, respectively.
The paper ends with two appendices. First, in Appendix~\ref{APP90A},
we show that the limit as~$s\nearrow1$ of our fractional functional
converges to the classical ``periodic'' perimeter
(i.e. to the perimeter on the cylinder obtained by identifying the
``sides'' of the slab~$S$). Then, in Appendix~\ref{S rad},
we remark that the assumption of cylindrical symmetry for the competitors
in~${\mathcal{K}}$ can be relaxed (in the sense that our
fractional functional decreases under cylindrical rearrangements).
\section{Kernel decay}\label{KD}
First, we point out that our functional is compatible with the
periodic structure in the horizontal direction.
For this, if~$F\subseteq S$, we define
the periodic extension of~$F$ as
$$ F_{\rm per}:=\bigcup_{k\in\Z}(F+ke_1),$$
and we have:
\begin{lemma}\label{period}
For any~$\tau\in \mathbb R$, it holds that~$P_S
(F_{\rm per}+\tau e_1)=P_S(F_{\rm per})$.
\end{lemma}
\begin{proof} The function~$x_1\mapsto \chi_{ F_{\rm per} }(x_1,x')$
is $1$-periodic, and so is~$x_1\mapsto K(x_1,x')$, for any fixed~$x'\in\mathbb R^{n-1}$.
Therefore, for any~$\tau\in \mathbb R$,
\begin{eqnarray*}&& \int_{[-1/2,\,1/2]} \,dx_1 \int_{[-1/2,\,1/2]} \,dy_1
\,\chi_{ F_{\rm per} +\tau e_1}(x_1,x')
\chi_{ \mathbb R^n\setminus (F_{\rm per}+\tau e_1) }(y_1,y') K(x-y)
\\ &=&
\int_{[-1/2,\,1/2]-\tau} \,dx_1 \int_{[-1/2,\,1/2]-\tau} \,dy_1
\,\chi_{ F_{\rm per} }(x_1,x')
\chi_{ \mathbb R^n\setminus F_{\rm per} }(y_1,y') K(x-y)
\\ &=&\int_{[-1/2,\,1/2]} \,dx_1 \int_{[-1/2,\,1/2]} \,dy_1
\,\chi_{ F_{\rm per} }(x_1,x')
\chi_{ \mathbb R^n\setminus F_{\rm per} }(y_1,y') K(x-y) .\end{eqnarray*}
Thus the desired result follows by integrating over~$x'$ and~$y'$.
\end{proof}
Now we prove a useful decay estimate on our kernel.
We remark that the scaling properties of our kernel
are quite different from the ones of many nonlocal problems
that have been studied in the literature: as a matter of
fact, the kernel that we study is not homogeneous
and it has quite different singular behaviors
locally and at infinity. Indeed, close to the origin
the dominant term is of the order of~$|x|^{-n-s}$,
but at infinity the $x_1$ direction ``averages out'',
as detailed in the following result:
\begin{lemma}\label{DEA}
For any~$x=(x_1,x')\in\mathbb R^n$ with~$|x'|\ge1$, we have
that~$K(x)\le C\,|x'|^{1-n-s}$,
for some~$C>0$.
\end{lemma}
\begin{proof} Fix~$x=(x_1,x')\in\mathbb R^n$ with~$|x'|\ge1$.
Let~$a_\pm :=\pm |x'|-x_1+1$ and~$b_\pm:=\pm3|x'|-x_1$. Let also~$y_\pm$
be the integer part of~$b_\pm$.
We observe that
$$ b_+=3|x'|-x_1 \ge |x'|+2-x_1=a_+ +1.$$
This says that at least one integer lies in the segment~$[a_+,b_+]$
and so~$y_+\ge a_+$. Therefore
$$ y_++x_1-1\ge a_+ +x_1-1 =|x'|.$$
Moreover
$$ y_-+x_1+1\le b_- +x_1+1\le -|x'|-1.$$
Consequently
\begin{eqnarray*}
\sum_{{k\in\Z}\atop{k\in (-\infty,b_-]\cup[b_+,+\infty) }} \frac{1}{|x_1+k|^{n+s}}
&\le&
\sum_{{k\in\Z}\atop{k\ge y_+}} \frac{1}{(x_1+k)^{n+s}}
+
\sum_{{k\in\Z}\atop{k\le y_-+1}} \frac{1}{(-x_1-k)^{n+s}}
\\ &=&
\sum_{{k\in\Z}\atop{k\ge y_+}} \frac{1}{(x_1+k)^{n+s}}
+
\sum_{{k\in\Z}\atop{k\ge -y_- -1}} \frac{1}{(-x_1+k)^{n+s}}
\\ &\le&
\sum_{{k\in\Z}\atop{k\ge y_+}} \int_{x_1+k-1}^{x_1+k}\frac{dt}{t^{n+s}}
+
\sum_{{k\in\Z}\atop{k\ge -y_- -1}} \int_{-x_1+k-1}^{-x_1+k}
\frac{dt}{t^{n+s}}
\\ &=&
\int_{x_1+y_+-1}^{+\infty}\frac{dt}{t^{n+s}}
+
\int_{-x_1-y_--2}^{+\infty}\frac{dt}{t^{n+s}}
\\ &\le&
\int_{|x'|}^{+\infty}\frac{dt}{t^{n+s}}
+
\int_{|x'|}^{+\infty}\frac{dt}{t^{n+s}}
\\ &=&
\frac{C}{|x'|^{n+s-1}}.
\end{eqnarray*}
This says that
\begin{equation}\label{ES-1}
\sum_{{k\in\Z}\atop{k\not\in (-3|x'|-x_1,\,3|x'|-x_1)}} \frac{1}{|x_1+k|^{n+s}}
\le \frac{C}{|x'|^{n+s-1}}.
\end{equation}
Now, we observe that the interval~$[-3|x'|-x_1,\,3|x'|-x_1]$
has length~$6|x'|$ and so it contains at most~$6|x'|+1\le 7|x'|$
integers. This implies that
\begin{equation}\label{ES-3}
\sum_{{k\in\Z}\atop{ k\in [-3|x'|-x_1,\,3|x'|-x_1]}} \frac{1}{|x'|^{n+s}}
\le \frac{7}{|x'|^{n+s-1}}.\end{equation}
Moreover,
$$ |x+ke_1|^{n+s}
= \Big( |x_1+k|^2+|x'|^2\Big)^{\frac{n+s}{2}} \ge
\max \big\{ |x_1+k|^{n+s},\, |x'|^{n+s}\big\}.$$
Thus, recalling~\eqref{ES-1} and~\eqref{ES-3},
we conclude that
\begin{eqnarray*}
K(x)&\le&
\sum_{{k\in\Z}\atop{{k\not\in (-3|x'|-x_1,\,3|x'|-x_1)}}}
\frac{1}{|x+ke_1|^{n+s}}
+\sum_{{k\in\Z}\atop{ k\in [-3|x'|-x_1,\,3|x'|-x_1]}} \frac{1}{|x+ke_1|^{n+s}}
\\ &\le&
\sum_{{k\in\Z}\atop{{k\not\in (-3|x'|-x_1,\,3|x'|-x_1)}}}
\frac{1}{|x_1+k|^{n+s}}
+\sum_{{k\in\Z}\atop{ k\in [-3|x'|-x_1,\,3|x'|-x_1]}} \frac{1}{|x'|^{n+s}}
\\ &\le&\frac{C+7}{|x'|^{n+s-1}},\end{eqnarray*}
which gives the desired claim up to renaming~$C$.
\end{proof}
\begin{corollary}\label{DEA-1}
Fix~$M\in\mathbb N$. We define
\begin{equation}\label{h M}
h_M(x):=\min\{M,\, K(x)\}.\end{equation}
Then, for any~$x\in\mathbb R^n$,
$$h_M(x)\le C_M \min\{1,\,|x'|^{1-n-s}\},$$
for some~$C_M>0$ possibly depending on~$M$.
\end{corollary}
\begin{proof} If~$|x'|\ge1$, we use Lemma~\ref{DEA} to see that
$$ \min\{1,\,|x'|^{1-n-s}\}=|x'|^{1-n-s} \ge C^{-1} K(x)
\ge C^{-1} h_M(x).$$
On the other hand, if~$|x'|<1$, we have that
$$ \min\{1,\,|x'|^{1-n-s}\}= 1\ge M^{-1} h_M(x).$$
Combining these two estimates we obtain the desired result.
\end{proof}
\section{Relation with the fractional perimeter}\label{REL}
The aim of this section is to point out the relation between
our functional and the fractional perimeter~${\rm Per}_s$
introduced in~\cite{CRS}. That is, we set
$$ {\rm Per}_s(F):=\int_{F} \int_{\mathbb R^n\setminus F}
\frac{dx\,dy}{|x-y|^{n+s}}$$
and we show that: on the one hand, our functional
is always below the fractional perimeter~${\rm Per}_s$,
on the other hand, our functional
is always above the fractional perimeter~${\rm Per}_s$,
up to a correction that depends on higher order volume terms,
and on a volume term coming from the boundary of the slab~$S$.
The precise statement goes as follows:
\begin{proposition}\label{EST PER-s}
Let~$F\subseteq S$.
Then
\begin{equation}
\label{Per s 1} P_S(F)\le {\rm Per}_s(F).\end{equation}
More precisely,
\begin{equation}\label{90}
{\rm Per}_s(F)-P_S(F)=
\sum_{{k\in\Z\setminus\{0\}}}
\int_{F} \int_{F}
\frac{dx\,dy}{|x-y+ke_1|^{n+s}}.\end{equation}
In addition,
if we set
\begin{equation}\label{hat} \begin{split}
&\widetilde F:= F\cap \{x_1\in [1/4,\,1/2]\},\\
&\widehat F:= (F+e_1)\cap \{x_1\in [1/2,\,3/4]\}\\
{\mbox{and }}\ & \Pi_S (F):= \int_{\widetilde F}\int_{\widehat F}
\frac{dx\,dy}{|x-y|^{n+s}},\end{split}\end{equation}
we have that
\begin{equation}\label{Per s 2} {\rm Per}_s(F)\le P_S(F)
+C\,\Big( |F|^2 + \Pi_S(F)\Big),\end{equation}
for some~$C>0$.
\end{proposition}
\begin{proof} We use the change of variable~$\widetilde y=y+ke_1$ to see that
$$ P_S(F)=
\int_{F} \int_{S\setminus F}
\sum_{k\in\Z}\frac{dx\,dy}{|x-y+ke_1|^{n+s}}
=
\int_{F} \int_{(S\setminus F)+ke_1}
\sum_{k\in\Z}\frac{dx\,d\widetilde y}{|x-\widetilde y|^{n+s}}
=\int_{F} \int_{(S\setminus F)_{\rm per}}
\frac{dx\,d\widetilde y}{|x-\widetilde y|^{n+s}}.$$
We observe that~$(S\setminus F)_{\rm per}\subseteq\mathbb R^n\setminus F$,
so we obtain that
$$ P_S(F)\le \int_{F} \int_{(S\setminus F)_{\rm per}}
\frac{dx\,dy}{|x-y|^{n+s}}\le
\int_{F} \int_{\mathbb R^n\setminus F}
\frac{dx\,dy}{|x-y|^{n+s}}={\rm Per}_s(F).$$
This establishes~\eqref{Per s 1}. More generally, we see that
$$ (\mathbb R^n\setminus F)\,\setminus\, \big( (S\setminus F)_{\rm per}\big)
=\bigcup_{{k\in\Z\setminus\{0\}}} (F+ke_1),$$
therefore, with another change of variable, we have
\begin{equation}\label{ifuhc4rfofdsasdfg}
{\rm Per}_s(F)-P_S(F)=
\sum_{{k\in\Z\setminus\{0\}}}
\int_{F} \int_{F+ke_1}
\frac{dx\,dy}{|x-y|^{n+s}}=
\sum_{{k\in\Z\setminus\{0\}}}
\int_{F} \int_{F}
\frac{dx\,dy}{|x-y+ke_1|^{n+s}}.\end{equation}
This proves~\eqref{90}.
Now we observe that, if~$|k|\ge2$ and~$x$, $y\in S$, then
$$ |x-y+ke_1|\ge |x_1-y_1+k|\ge |k|-|x_1-y_1|\ge |k|-1\ge \frac{|k|}{2},$$
therefore
\begin{equation} \label{yuioghddds}
\sum_{{k\in\Z\setminus\{0\}}\atop{|k|\ge2}}
\int_{F} \int_{F}
\frac{dx\,dy}{|x-y+ke_1|^{n+s}}\le
\sum_{{k\in\Z\setminus\{0\}}\atop{|k|\ge2}}
\int_{F} \int_{F}
\frac{dx\,dy}{(|k|/2)^{n+s}}\le C\,|F|^2,\end{equation}
for some~$C>0$.
Moreover,
\begin{equation}\label{84}
\int_{F} \int_{F+e_1} \chi_{[1/4,\,+\infty)}(|x_1-y_1|)
\frac{dx\,dy}{|x-y|^{n+s}} \le
\int_{F} \int_{F+e_1}
\frac{dx\,dy}{(1/4)^{n+s}}\le C\,|F|^2,\end{equation}
for some~$C>0$.
Also, if~$x\in S$, $y\in S+e_1$ and~$|x_1-y_1|\le 1/4$, we have that
\begin{eqnarray*}&&
x_1\ge y_1 - |x_1-y_1|\ge -\frac{1}{2} + 1 -\frac14 =\frac14\\
{\mbox{and }}&& y_1-1\le |y_1-x_1|+x_1-1\le \frac14 +\frac12 -1 =-\frac14.
\end{eqnarray*}
As a consequence, if~$x\in F\subseteq S$ and~$y_1\in F+e_1\subseteq
S+e_1$, with~$|x-y|\le1/4$, we have that~$x\in \widetilde F$
and~$y\in \widehat F$, where the notation in~\eqref{hat} is here in use,
therefore
$$ \int_{F} \int_{F+e_1} \chi_{[0,\,1/4]}(|x_1-y_1|)
\frac{dx\,dy}{|x-y|^{n+s}} \le\Pi_S(F).
$$
This and~\eqref{84} give that
\begin{equation}\label{84bis}
\int_{F} \int_{F+e_1}
\frac{dx\,dy}{|x-y|^{n+s}} \le
C\,\Big(|F|^2+\Pi_S(F)\Big).\end{equation}
Thus, using the change of variable~$\bar x:=x+e_1$
and~$\bar y:=y+e_1$, we also have that
\begin{equation}\label{85}
\int_{F} \int_{F-e_1}
\frac{dx\,dy}{|x-y|^{n+s}} =
\int_{F+e_1} \int_{F}
\frac{dx\,dy}{|x-y|^{n+s}}\le
C\,\Big(|F|^2+\Pi_S(F)\Big).\end{equation}
Putting together~\eqref{84bis} and~\eqref{85} we obtain
$$ \sum_{{k\in\Z\setminus\{0\}}\atop{|k|\le 1}}
\int_{F} \int_{F}
\frac{dx\,dy}{|x-y+ke_1|^{n+s}}=
\int_{F} \int_{F+e_1}
\frac{dx\,dy}{|x-y|^{n+s}}+
\int_{F} \int_{F-e_1}
\frac{dx\,dy}{|x-y|^{n+s}}\le
C\,\Big(|F|^2+\Pi_S(F)\Big).$$
This and~\eqref{yuioghddds} imply that
$$ \sum_{{k\in\Z\setminus\{0\}}}
\int_{F} \int_{F}
\frac{dx\,dy}{|x-y+ke_1|^{n+s}}
\leq C\,\Big(|F|^2+\Pi_S(F)\Big).$$
Recalling~\eqref{ifuhc4rfofdsasdfg},
we see that this ends the proof of~\eqref{Per s 2}.
\end{proof}
As a consequence of our preliminary computations, we obtain that
cylinders have finite energy:
\begin{corollary}\label{finite}
The functional attains a finite value on cylinders.
Namely, for any~$R>0$, let~${\mathcal{C}} := \{(x_1,x')\in \mathbb R\times\mathbb R^{n-1}
{\mbox{ s.t. }}|x'|\le R\}$. Then~$P_S({\mathcal{C}}\cap S)<+\infty$.
\end{corollary}
\begin{proof}
We take a set~$\widetilde{\mathcal{C}}$ with smooth boundary
and contained in~$[-1,\,1]\times\mathbb R^{n-1}$
such that~$\widetilde{\mathcal{C}}\cap S={\mathcal{C}}\cap S$.
Then, we use~\eqref{Per s 1}
and we obtain that
$$ +\infty > {\rm Per}_s (\widetilde{\mathcal{C}})\ge
P_S(\widetilde{\mathcal{C}}\cap S)=P_S({\mathcal{C}}\cap S),$$
as desired.
\end{proof}
Next result computes the term~$\Pi_S$ in~\eqref{Per s 2}
in the special case of small cylinders (this will play a role
in the proof of Theorem~\ref{YES BALL}).
\begin{lemma}\label{play}
For any~$r\in(0,\,1/4)$, let~${\mathcal{C}} := \{(x_1,x')\in S
{\mbox{ s.t. }}|x'|\le r\}$. Then~$\Pi_S({\mathcal{C}})\le C\,r^{n-s}$.
\end{lemma}
\begin{proof} We first translate in the first coordinate
and then change variable~$X:=x/r$ and~$Y:=y/r$, so that we obtain
\begin{equation}\label{87}\begin{split}
\Pi_S({\mathcal{C}})\,&=\int_{1/4}^{1/2}\,dx_1 \int_{1/2}^{3/4}\,dy_1
\int_{|x'|\le r} \,dx'\int_{|y'|\le r} \,dy' \frac{1}{|x-y|^{n+s}}
\\ &=\int_{-1/4}^{0}\,dx_1 \int_{0}^{1/4}\,dy_1
\int_{|x'|\le r} \,dx'\int_{|y'|\le r} \,dy' \frac{1}{|x-y|^{n+s}}
\\ &= r^{n-s}
\int_{-1/(4r)}^{0}\,dX_1 \int_{0}^{1/(4r)}\,dY_1
\int_{|X'|\le 1} \,dX'\int_{|Y'|\le 1} \,dY' \frac{1}{|X-Y|^{n+s}}
.\end{split}\end{equation}
Now we take a
bounded set~${\mathcal{C}}^\star\subset [-1,0]\times\mathbb R^{n-1}$
with smooth boundary that contains~$\{(x_1,x')\in\mathbb R^n {\mbox{ s.t. }}
x_1\in [-1,0] {\mbox{ and }} |x'|\le 1\}$. Then we have that
\begin{equation*}
\int_{-1}^{0}\,dX_1 \int_{0}^{1}\,dY_1
\int_{|X'|\le 1} \,dX'\int_{|Y'|\le 1} \,dY' \frac{1}{|X-Y|^{n+s}}
\le \int_{{\mathcal{C}}^\star}\int_{\mathbb R^n\setminus {\mathcal{C}}^\star}
\frac{dX\,dY}{|X-Y|^{n+s}} \le {\rm Per}_s ({\mathcal{C}}^\star)\le C,
\end{equation*}
for some~$C>0$, thus~\eqref{87} becomes
\begin{equation}\label{88}\begin{split}
\Pi_S({\mathcal{C}})
\le C\, r^{n-s} & \left( 1 +
\int_{-1/(4r)}^{-1}\,dX_1 \int_{0}^{1/(4r)}\,dY_1
\int_{|X'|\le 1} \,dX'\int_{|Y'|\le 1} \,dY' \frac{1}{|X-Y|^{n+s}}
\right.\\&\left.
+\int_{-1/(4r)}^{0}\,dX_1 \int_{1}^{1/(4r)}\,dY_1
\int_{|X'|\le 1} \,dX'\int_{|Y'|\le 1} \,dY' \frac{1}{|X-Y|^{n+s}}
\right).\end{split}\end{equation}
Notice that we can change variable~$(x,y):=(-Y,-X)$
and see that
\begin{eqnarray*}&& \int_{-1/(4r)}^{0}\,dX_1 \int_{1}^{1/(4r)}\,dY_1
\int_{|X'|\le 1} \,dX'\int_{|Y'|\le 1} \,dY' \frac{1}{|X-Y|^{n+s}}
\\ &&\qquad=
\int_0^{1/(4r)}\,dy_1 \int_{-1/(4r)}^{-1}\,dx_1
\int_{|y'|\le 1} \,dy'\int_{|x'|\le 1} \,dx' \frac{1}{|x-y|^{n+s}}.\end{eqnarray*}
As a consequence, we can write~\eqref{88} as
\begin{equation}\label{89}
\Pi_S({\mathcal{C}})
\le C\, r^{n-s}\left( 1+2
\int_{-1/(4r)}^{-1}\,dx_1 \int_{0}^{1/(4r)}\,dy_1
\int_{|x'|\le 1} \,dx'\int_{|y'|\le 1} \,dy' \frac{1}{|x-y|^{n+s}}
\right).\end{equation}
Now we observe that
\begin{eqnarray*}
&& \int_{-1/(4r)}^{-1}\,dx_1 \int_{0}^{1/(4r)}\,dy_1
\int_{|x'|\le 1} \,dx'\int_{|y'|\le 1} \,dy' \frac{1}{|x-y|^{n+s}} \\
&\le&
\int_{-1/(4r)}^{-1}\,dx_1 \int_{0}^{1/(4r)}\,dy_1
\int_{|x'|\le 1} \,dx'\int_{|y'|\le 1} \,dy' \frac{1}{|x_1-y_1|^{n+s}}\\
&=&
C \,\int_{-1/(4r)}^{-1}\,dx_1 \int_{0}^{1/(4r)}\,dy_1
\frac{1}{|x_1-y_1|^{n+s}} \\
&\le& C \,\int_{-1/(4r)}^{-1}\,dx_1 (-x_1)^{1-n-s} \\
&\le& C\, \int_{1}^{+\infty}\,d\tau \,\tau^{1-n-s}
\\ &\le& C.
\end{eqnarray*}
The desired result thus follows by plugging this estimate into~\eqref{89}.
\end{proof}
\section{Energy bounds}\label{EB}
We consider here an auxiliary energy functional
and we prove that the functional~$P_S$ is bounded from below
by it. The proof requires a very careful analysis of
the different contributions and the result,
together with the one in the subsequent Proposition~\ref{converge},
will play a crucial role for the proof of Theorem~\ref{THE EX},
since it will lead to the compactness of the minimizing sequences.
\begin{proposition}\label{energy}
Let~$F\subseteq S$. Suppose that there exists
an even
function~$f:[-1/2,\,1/2]\rightarrow [0,+\infty]$,
with $f$ decreasing in~$[0,\,1/2]$,
such that
$$ F:= \Big\{ (x_1,x')\in S {\mbox{ s.t. }} |x'|\le f(x_1)\Big\}.$$
Let~$\varepsilon_*\le \alpha_*\in [0,\,1/2]$ and suppose that
\begin{equation}\label{eps star}
f(x_1)\ge 4 {\mbox{ for all }} x_1\in [0,\varepsilon_*)
\end{equation}
and
\begin{equation}\label{a star}
f(x_1)\ge 2{f(y_1)} {\mbox{ for all }} x_1\in [0,\varepsilon_*)
{\mbox{ and }} y_1\in (\alpha_*,\,1/2].
\end{equation}
Then
$$ P_S(F)\ge C \,
\int_{0}^{\varepsilon_*} \left[
\int_{\alpha_*}^{1/2}
\frac{f^{n-1}(x_1)}{|x_1-y_1|^{1+s}}\,dy_1\right]\,dx_1,$$
for a suitable constant~$C>0$.
\end{proposition}
\begin{proof} We have
\begin{eqnarray*}
P_S(F)&=&\int_{F} \int_{S\setminus F} K(x-y)\,dx\,dy \\
&\ge&\int_{F} \int_{S\setminus F} \frac{dx\,dy}{|x-y|^{n+s}}
\\ &=& \int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1 \int_{\{ |x'|\le f(x_1)\}} dx'
\int_{\{ |y'|> f(y_1)\}} dy' \frac{1}{|x-y|^{n+s}}.
\end{eqnarray*}
Now we introduce cylindrical coordinates by writing~$x'= \rho\theta$
and~$y'= r\omega$, with~$\theta$, $\omega\in S^{n-2}$. We obtain
\begin{equation}\label{uisgfdsajhgfd3456} P_S(F) \geq
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1 \int_{S^{n-2}}\,d\theta
\int_{S^{n-2}}\,d\omega
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|\rho\theta-r\omega|^2\big)^{\frac{n+s}{2}} }
.\end{equation}
Here and in the sequel, $d\theta$ and~$d\omega$
are short notations for~$d{\mathcal{H}}^{n-1}(\theta)$
and~$d{\mathcal{H}}^{n-1}(\omega)$, respectively.
Now, for any~$\theta\in S^{n-2}$, we consider a rotation~$R_\theta$
on~$S^{n-2}$
such that~$\theta =R_\theta e_2$.
Then, we can rotate~$\omega=R_\theta\widetilde\omega$,
and obtain that
\begin{eqnarray*}
&& \int_{S^{n-2}}\,d\omega \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|\rho\theta-r\omega|^2\big)^{\frac{n+s}{2}} }=
\int_{S^{n-2}}\,d\omega \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|\rho R_\theta e_2 -r\omega|^2\big)^{\frac{n+s}{2}} }\\
&&\qquad=
\int_{S^{n-2}}\,d\widetilde\omega \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|R_\theta(\rho e_2 -r\widetilde\omega)|^2\big)^{\frac{n+s}{2}} }
=\int_{S^{n-2}}\,d\widetilde\omega \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|\rho e_2 -r\widetilde\omega|^2\big)^{\frac{n+s}{2}} }
.\end{eqnarray*}
Notice that the latter integral in now independent of~$\theta$.
Then we can write~\eqref{uisgfdsajhgfd3456} as
\begin{equation}\label{uisgfdsajhgfd3456-2} P_S(F) \geq C
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1
\int_{S^{n-2}}\,d\omega
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr \frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2 +|\rho e_2-r\omega|^2\big)^{\frac{n+s}{2}} }
,\end{equation}
for some~$C>0$.
Now we observe that
$$ |\rho e_2-r\omega|^2 = \rho^2 +r^2 -2\rho r \omega_2
= |\rho-r|^2 +2\rho r(1-\omega_2).$$
where~$\omega_2=\omega\cdot e_2$ is the second component of the vector~$\omega
\in S^{n-2}\subset \{0\}\times\mathbb R^{n-1}$.
Now\footnote{For concreteness we suppose in this part that~$n\geq3$.
In the very special case~$n=2$,
one does not have any component~$(\omega_3,\cdots,\omega_n)$,
so she or he can just disregard~$\underline\omega$ and go directly to~\eqref{uisgfdsajhgfd3456-3}.}
we define~$\underline\omega:=(\omega_3,\cdots,\omega_n)$.
Then~$\omega=(0,\omega_2,\underline\omega)$
and
\begin{equation}\label{67d8cv9bbfrff}
\omega_2^2+|\underline \omega|^2=1.
\end{equation}
We also set
$$ S^{n-2}_\star:=
\left\{ \omega =(0,\omega_2,\underline\omega)\in S^{n-2}
{\mbox{ s.t. }} \omega_2\ge \frac{9}{10}\right\}.$$
Using~\eqref{67d8cv9bbfrff}, we see that,
$$ \left\{ |\underline\omega|\le \frac{\sqrt{19}}{10}\right\}
\subseteq S^{n-2}_\star.$$
Also, in~$S^{n-2}_\star$,
$$ 2\rho r (1-\omega_2)=
\frac{2\rho r (1-\omega_2^2)}{1+\omega_2} \le 2\rho r (1-\omega_2^2)=
2\rho r|\underline\omega|^2.$$
Therefore, fixed any~$a\ge0$, we have
\begin{eqnarray*}
&& \int_{S^{n-2}}
\frac{ d\omega }{
\big( a^2 +|\rho e_2-r\omega|^2\big)^{\frac{n+s}{2}} }
=
\int_{S^{n-2}}
\frac{ d\omega }{
\big( a^2 + |\rho-r|^2 +2\rho r(1-\omega_2)\big)^{\frac{n+s}{2}} }\\
&&\qquad \ge \int_{S^{n-2}_\star}
\frac{ d\omega }{
\big( a^2 + |\rho-r|^2 +2\rho r(1-\omega_2)\big)^{\frac{n+s}{2}} }\ge
\int_{S^{n-2}_\star}
\frac{ d\omega }{
\big( a^2 + |\rho-r|^2 +2\rho r |\underline\omega|^2\big)^{\frac{n+s}{2}} }
\\ &&\qquad \ge C\,
\int_{\{|\underline\omega|\le \sqrt{19}/10\}}
\frac{ d\underline\omega }{
\big( a^2 + |\rho-r|^2 +2\rho r |\underline\omega|^2\big)^{\frac{n+s}{2}} },
\end{eqnarray*}
for some~$C>0$, possibly different from line to line.
So we use polar coordinates~$\mathbb R^{n-2}\ni \underline\omega = R \varphi$,
with~$\varphi\in S^{n-3}$ and obtain
from the latter estimate that
\begin{equation} \label{jkcvbndssfghrewefg}
\int_{S^{n-2}}
\frac{ d\omega }{
\big( a^2 +|\rho e_2-r\omega|^2\big)^{\frac{n+s}{2}} }
\ge
C\,
\int_0^{\sqrt{19}/10}
\frac{ R^{n-3}\, dR}{
\big( a^2 + |\rho-r|^2 +2\rho r R^2\big)^{\frac{n+s}{2}} }.\end{equation}
Now we observe that, for any~$X$, $Y\ge0$, we have that
\begin{equation} \label{789dfvgjhgfdbvcxw}
\int_0^{\sqrt{19}/10}
\frac{ R^{n-3}\, dR}{
\big( X^2+Y^2 R^2\big)^{\frac{n+s}{2}} }
=\left(\frac{X}{Y}\right)^{n-2}\, \frac{1}{X^{n+s}}\,
\int_0^{ \frac{ \sqrt{19}\,Y }{10\,X}}
\frac{ t^{n-3}\, dt}{
\big( 1+t^2\big)^{\frac{n+s}{2}} },\end{equation}
where the change of variable~$R=Xt/Y$ was performed.
Now we denote, for any~$x\ge0$,
$$ \Gamma(x):=
\int_0^{x}
\frac{ t^{n-3}\, dt}{
\big( 1+t^2\big)^{\frac{n+s}{2}} }.$$
We observe that if~$x\in[0,1]$ and~$t\in[0,x]$, then~$1+t^2\le 2$ and so
$$ \Gamma(x)\ge C\,\int_0^x t^{n-3}\,dt = C\,x^{n-2},$$
up to renaming~$C>0$.
Moreover, if~$x\ge1$, then
$$ \Gamma(x)\ge
\int_0^{1}
\frac{ t^{n-3}\, dt}{
\big( 1+t^2\big)^{\frac{n+s}{2}} }\ge C.$$
Summarizing, for any~$x\ge0$, we have that
$$ \Gamma(x)\ge C\,\min\{ x^{n-2},\, 1\}.$$
Thus, going back to~\eqref{789dfvgjhgfdbvcxw},
\begin{equation} \label{78009dfvgjhgfdbvcxw-78}
\begin{split} &\int_0^{\sqrt{19}/10}
\frac{ R^{n-3}\, dR}{
\big( X^2+Y^2 R^2\big)^{\frac{n+s}{2}} }
=\left(\frac{X}{Y}\right)^{n-2}\, \frac{1}{X^{n+s}}\,
\Gamma \left(\frac{\sqrt{19}\,Y}{10\,X}\right)
\\ &\qquad \ge C\, \left(\frac{X}{Y}\right)^{n-2}\, \frac{1}{X^{n+s}}\,
\min\left\{ 1,\, \left(\frac{Y}{X}\right)^{n-2}\right\}
= C\, \frac{1}{X^{n+s}}\,
\min\left\{ \left(\frac{X}{Y}\right)^{n-2},\, 1\right\}
.\end{split}\end{equation}
Now we take~$X:=\sqrt{a^2+|\rho-r|^2}$ and~$Y:=\sqrt{2\rho r}$
and we plug~\eqref{78009dfvgjhgfdbvcxw-78} into~\eqref{jkcvbndssfghrewefg}.
In this way we obtain
\begin{equation} \label{000jkcvbndssfghrewefg}
\int_{S^{n-2}}
\frac{ d\omega }{
\big( a^2 +|\rho e_2-r\omega|^2\big)^{\frac{n+s}{2}} }
\ge
C\, \frac{1}{\big( a^2+|\rho-r|^2\big)^{\frac{n+s}{2}}}\,
\min\left\{
\left(\frac{a^2+|\rho-r|^2}{\rho r}\right)^{\frac{n-2}{2}},\, 1\right\}
.\end{equation}
Hence we take~$a:=|x_1-y_1|$ and we insert~\eqref{000jkcvbndssfghrewefg}
into~\eqref{uisgfdsajhgfd3456-2}, obtaining that
\begin{equation}\label{uisgfdsajhgfd3456-3} \begin{split}
P_S(F) \,&\ge C\,
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr \\
&\qquad\quad\frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2+|\rho-r|^2\big)^{\frac{n+s}{2}} }\,
\min\left\{
\left(\frac{|x_1-y_1|^2+|\rho-r|^2}{\rho r}\right)^{\frac{n-2}{2}},\, 1\right\}
.\end{split}\end{equation}
Now we observe that
\begin{eqnarray*}
&& \int_{-1/2}^{1/2} dx_1\,\int_{-1/2}^{1/2} dy_1
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr \\
&&\qquad\chi_{\{ |x_1-y_1|^2+|\rho-r|^2\le \rho r\}}\,
\frac{ \rho^{n-2} r^{n-2} }{
\big( |x_1-y_1|^2+|\rho-r|^2\big)^{\frac{n+s}{2}} }\,
\min\left\{
\left(\frac{|x_1-y_1|^2+|\rho-r|^2}{\rho r}\right)^{\frac{n-2}{2}},\, 1\right\}\\
&=&
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr
\, \chi_{ \{ |x_1-y_1|^2+|\rho-r|^2\le \rho r\} }\,
\frac{ \rho^{\frac{n-2}{2}} r^{\frac{n-2}{2}} }{
\big( |x_1-y_1|^2+|\rho-r|^2\big)^{1+\frac{s}{2}} }
.\end{eqnarray*}
This and~\eqref{uisgfdsajhgfd3456-3} give that
\begin{equation}\label{uisgfdsajhgfd3456-45}
P_S(F) \ge C\,
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1
\int_0^{f(x_1)} d\rho
\int_{f(y_1)}^{+\infty} dr
\, \chi_{ \{ |x_1-y_1|^2+|\rho-r|^2\le \rho r\} }\,
\frac{ \rho^{\frac{n-2}{2}} r^{\frac{n-2}{2}} }{
\big( |x_1-y_1|^2+|\rho-r|^2\big)^{1+\frac{s}{2}} }.
\end{equation}
Accordingly, we perform the change of variable~$\rho=|x_1-y_1|\,\alpha$
and~$r=|x_1-y_1|\,\beta$, so that~\eqref{uisgfdsajhgfd3456-45}
becomes
\begin{eqnarray*} P_S(F) &\ge &C\,
\int_{-1/2}^{1/2} dx_1\,
\int_{-1/2}^{1/2} dy_1
\int_0^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\frac{f(y_1)}{|x_1-y_1|}}^{+\infty} d\beta
\, \chi_{ \{ 1+|\alpha-\beta|^2\le \alpha\beta\} }\,
\frac{ |x_1-y_1|^{n-2-s}\,\alpha^{\frac{n-2}{2}} \beta^{\frac{n-2}{2}} }{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} }
\\ &\ge& C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\frac{f(y_1)}{|x_1-y_1|}}^{+\infty} d\beta
\, \chi_{ \{ 1+|\alpha-\beta|^2\le \alpha\beta\} }\,
\frac{ |x_1-y_1|^{n-2-s}\,\alpha^{\frac{n-2}{2}} \beta^{\frac{n-2}{2}} }{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} } ,
\end{eqnarray*}
where~$\varepsilon_*$ and~$\alpha_*$
were introduced in~\eqref{eps star}
and~\eqref{a star}.
As a matter of fact, using~\eqref{a star}, we obtain that,
in the domain above,
$$ \frac{f(y_1)}{|x_1-y_1|}\le
\frac{f(x_1)}{2\,|x_1-y_1|} < \frac{9 f(x_1)}{10\,|x_1-y_1|}\le \alpha.$$
As a consequence
\begin{equation}\label{yuidfghvnmnbvcf} \begin{split} P_S(F)\, &\ge C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\alpha}^{+\infty} d\beta
\, \chi_{ \{ 1+|\alpha-\beta|^2\le \alpha\beta\} }\,
\frac{ |x_1-y_1|^{n-2-s}\,\alpha^{\frac{n-2}{2}} \beta^{\frac{n-2}{2}} }{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} }\\
&\ge C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\alpha}^{\alpha+1} d\beta
\, \chi_{ \{ 1+|\alpha-\beta|^2\le \alpha\beta\} }\,
\frac{ |x_1-y_1|^{n-2-s}\,\alpha^{\frac{n-2}{2}} \beta^{\frac{n-2}{2}} }{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} } .
\end{split}\end{equation}
Now we observe that
in the domain above~$0\le\beta-\alpha\le1$, therefore, recalling~\eqref{eps star},
$$ 1+|\alpha-\beta|^2\le 2 \le \frac{f^2(x_1)}{8} \le
\frac{f^2(x_1)}{8\,|x_1-y_1|^2} $$
and
\begin{equation}\label{mngfd45678} \alpha\beta \ge \alpha^2 \ge
\frac{81\,f^2(x_1)}{100\,|x_1-y_1|^2} >\frac{f^2(x_1)}{8\,|x_1-y_1|^2},\end{equation}
that is
$$ 1+|\alpha-\beta|^2 < \alpha\beta.$$
Accordingly, \eqref{yuidfghvnmnbvcf} boils down to
$$ P_S(F)\ge C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\alpha}^{\alpha+1} d\beta
\; \frac{ |x_1-y_1|^{n-2-s}\,\alpha^{\frac{n-2}{2}} \beta^{\frac{n-2}{2}} }{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} }
.$$
Hence, using again~\eqref{mngfd45678},
$$ P_S(F)
\ge C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\int_{\alpha}^{\alpha+1} d\beta
\; \frac{ |x_1-y_1|^{n-2-s}}{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} }\cdot
\frac{f^{n-2}(x_1)}{|x_1-y_1|^{n-2}}.$$
Now we point out that
$$ \int_\alpha^{\alpha+1}
\frac{ d\beta}{
\big( 1+|\alpha-\beta|^2\big)^{1+\frac{s}{2}} }=
\int_0^{1}
\frac{ d\tau}{
\big( 1+\tau^2\big)^{1+\frac{s}{2}} }\ge C,$$
and thus we get that
\begin{eqnarray*} P_S(F)
&\ge& C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\int_{\frac{9 f(x_1)}{10 \,|x_1-y_1|}}^{\frac{f(x_1)}{|x_1-y_1|}} d\alpha
\; |x_1-y_1|^{n-2-s}
\cdot\frac{f^{n-2}(x_1)}{|x_1-y_1|^{n-2}}\\
&\ge&
C\,
\int_{0}^{\varepsilon_*} dx_1\,
\int_{\alpha_*}^{1/2} dy_1
\; |x_1-y_1|^{n-2-s}
\cdot\frac{f^{n-2}(x_1)}{|x_1-y_1|^{n-2}}\cdot
\frac{f(x_1)}{|x_1-y_1|},\end{eqnarray*}
which completes the proof of Proposition~\ref{energy}.
\end{proof}
\section{Convergence issues}\label{CI2}
Here we show that uniform energy bounds, as the ones obtained
in Proposition~\ref{energy}, joined with volume constraints,
imply a suitable compactness.
\begin{proposition}\label{converge}
Let~$C_1$, $C_2>0$.
Consider a sequence of
even functions~$f_k:[-1/2,\,1/2]\rightarrow [0,+\infty]$,
with $f_k$ decreasing in~$[0,\,1/2]$.
Assume that
for any~$\varepsilon_{*,k}\le \alpha_{*,k}\in [0,\,1/2]$ such that
\begin{equation}\label{LAL1}
f_k(x_1)\ge 4 {\mbox{ for all }} x_1\in [0,\varepsilon_{*,k})
\end{equation}
and
\begin{equation}\label{LAL2}
f_k(x_1)\ge 2{f_k(y_1)} {\mbox{ for all }} x_1\in [0,\varepsilon_{*,k})
{\mbox{ and }} y_1\in (\alpha_{*,k},\,1/2]
\end{equation}
it holds that
\begin{equation}\label{LAL3}
\int_{0}^{\varepsilon_{*,k}} \left[
\int_{\alpha_{*,k}}^{1/2}
\frac{f_k^{n-1}(x_1)}{|x_1-y_1|^{1+s}}\,dy_1\right]\,dx_1\le C_1.
\end{equation}
Suppose also that
\begin{equation}\label{5.3bis} \int_{0}^{1/2} f_k^{n-1}(x_1)\,dx_1 = C_2.\end{equation}
Then, there exists a function~$f$ such that,
up to a subsequence, $f_k\to f$ a.e. in $[-1/2,\,1/2]$ as~$k\to+\infty$, and
$$ \int_{0}^{1/2} f^{n-1}(x_1)\,dx_1 = C_2.$$
\end{proposition}
\begin{proof} First
we point out that, for any~$r\in[0,1/2]$,
$$ C_2 \ge \int_{0}^{r} f^{n-1}_k(x_1)\,dx_1
\ge r\,\inf_{ [0,\,r] } f_k^{n-1},$$
that is
\begin{equation}\label{inf f}
f_k(r)=\inf_{ [0,\,r] } f_k\le\left(\frac{C_2}{r}\right)^{\frac{1}{n-1}}.
\end{equation}
In particular~$f_k(1/2)\le (2C_2)^{\frac{1}{n-1}}$.
So, for any~$M>4+(2C_2)^{\frac{1}{n-1}}$, we are allowed to define
\begin{equation}\label{5.4bis}
\varepsilon_k(M):= \inf \{ r \in[0,\,1/2] {\mbox{ s.t. }}
f_k(r)<M\}.\end{equation}
We claim that
\begin{equation}\label{inf f eps}
\varepsilon_k(M)\le \frac{C_2}{M^{n-1}}.
\end{equation}
Indeed, if~$\varepsilon_k(M)=0$ we are done, so we assume~$\varepsilon_k(M)>0$.
Then, for any~$k$ and~$M$ fixed,
for any~$j\in\mathbb N$, we can take~$r_j\in[(1-2^{-j})\,\varepsilon_k(M),\,\varepsilon_k(M)]$
such that~$f_k(r_j)\ge M$. Hence,
from~\eqref{inf f}, we have that
$$ M\le f_k(r_j)\le \left(\frac{C_2}{r_j}\right)^{\frac{1}{n-1}}.$$
So we pass~$j\to+\infty$ and we obtain
$$ M\le \left(\frac{C_2}{\varepsilon_k(M)}\right)^{\frac{1}{n-1}},$$
that proves~\eqref{inf f eps}.
Now we define~$\alpha_k(M):=\varepsilon_k(M/2)$.
We point out that, if~$x_1\in [0,\varepsilon_{k}(M))$ then
$f_k(x_1)\ge M$, and therefore~$f_k(x_1)\ge 4$. Also,
if~$y_1\in (\alpha_{k}(M),\,1/2]=(\varepsilon_{k}(M/2),\,1/2]$, we have that~$
f_k(y_1)\le M/2$. Accordingly,
if~$x_1\in [0,\varepsilon_{k}(M))$ and~$y_1\in (\alpha_{k}(M),\,1/2]$,
we have that~$f(x_1)\ge M\ge 2f(y_1)$.
These considerations show that~\eqref{LAL1} and~\eqref{LAL2}
are satisfied by~$\varepsilon_{*,k}:=\varepsilon_k(M)$ and~$\alpha_{*,k}:=
\alpha_k(M)$. As a consequence
\begin{equation}\label{LAL4}
{\mbox{ formula \eqref{LAL3} holds true with
$\varepsilon_{*,k}:=\varepsilon_k(M)$ and $\alpha_{*,k}:=
\alpha_k(M)$.}}
\end{equation}
Now we claim that there exists a constant~$C_\star>1$ such that,
if~$M\ge C_\star$ then
\begin{equation}\label{FUNF INT}
\int_0^{\varepsilon_k(M)} f_k^{n-1}(x_1)\,dx_1\le \frac{C_\star}{M^{s(n-1)}}.
\end{equation}
To prove this, we use~\eqref{LAL4} to notice that
\begin{equation}\label{id8ocugdsqwww} {C_1}\ge
\int_{0}^{\varepsilon_k(M)} \left[ \int_{\alpha_k(M)}^{1/2}
f_k^{n-1}(x_1)\,\frac{1}{|x_1-y_1|^{1+s}}\,dy_1\right]\,dx_1.\end{equation}
Now, fixed~$x_1<\varepsilon_k(M)\le \varepsilon_k(M/2)=\alpha_k(M)$, we compute
\begin{equation}\label{7dfisq1df}
\int_{\alpha_k(M)}^{1/2} \frac{1}{|x_1-y_1|^{1+s}}\,dy_1
=\int_{\alpha_k(M)}^{1/2} \frac{1}{(y_1-x_1)^{1+s}}\,dy_1
= \frac{1}{s} \left[\frac{1}{(\alpha_k(M)-x_1)^s}
-\frac{1}{((1/2)-x_1)^s}
\right].\end{equation}
Now, if~$x_1<\varepsilon_k(M)$, we have that
$$ \frac12 -x_1 \ge \frac12 -\varepsilon_k(M)\ge \frac12 -
\frac{C_2}{M^{n-1}} \ge\frac14,$$
if~$M$ is sufficiently large (independently on~$k$), thanks to~\eqref{inf f eps}.
Moreover, using again~\eqref{inf f eps}, we see that
$$ \alpha_k(M)-x_1\le \alpha_k(M)=\varepsilon_k(M/2)\le
\frac{C_3}{M^{n-1}},$$
for some~$C_3>0$.
Therefore~\eqref{7dfisq1df}
implies that
\begin{equation*}
\int_{\alpha_k(M)}^{1/2} \frac{1}{|x_1-y_1|^{1+s}}\,dy_1
\ge \left[ {C_4}\,{M^{s(n-1)}}
-C_5
\right]\ge C_6 M^{s(n-1)},\end{equation*}
for suitable constants~$C_i>0$, as long as~$M$ is large
enough, independently on~$k$.
So we plug this information into~\eqref{id8ocugdsqwww}, and we conclude that
$$ C_7\ge M^{s(n-1)}
\int_{0}^{\varepsilon_k(M)} f_k^{n-1}(x_1)\,dx_1,$$
and this proves~\eqref{FUNF INT}.
Now we claim that
\begin{equation}\label{c.a.e.}
f_k\to f {\mbox{ a.e. in }}[0,1/2],{\mbox{ as }}k\to+\infty,
\end{equation}
up to subsequences, for some function~$f$. To prove this,
we use the compactness of the functions with bounded variation,
joined with a diagonal trick.
We fix~$M\in\mathbb N$, $M\ge1$, and we use \eqref{inf f eps}
to see that~$(\varepsilon_k(M),\,1/2)\supseteq (C_2M^{1-n},\,1/2)$,
for any~$k$,
and so, by~\eqref{5.4bis},
$$ \sup_{ (C_2M^{1-n},\,1/2) } f_k\le \sup_{(\varepsilon_k(M),\,1/2)}f_k\le M.$$
Since~$f_k$ is monotone, this gives that~$\| f_k\|_{BV(C_2M^{1-n},\,1/2)}\le 2M$.
As a consequence of this
and of the
compactness of bounded variations functions
(see e.g. Theorem~3.23 in~\cite{AFP}) we get that, for any fixed~$M\in\mathbb N$,
$f_k\to f^{(M)}$ a.e. in~$(C_2M^{1-n},\,1/2)$, up to a subsequence,
for some function~$f^{(M)} :(C_2M^{1-n},\,1/2)\to[0,+\infty]$.
More explicitly, we write this subsequence by introducing
an increasing function~$\phi_M:\mathbb N\to\mathbb N$, and by stating that
\begin{equation}\label{5.13}
f_{\phi_M\circ\dots\circ\phi_1(k)}\to f^{(M)} {\mbox{ a.e. in }} (C_2M^{1-n},\,1/2).
\end{equation}
As a matter of fact, for a.e. $x_1\in (C_2M^{1-n},\,1/2)$, we have that
$$ f^{(M+1)}(x_1)=\lim_{k\to+\infty} f_{\phi_{M+1}\circ
\phi_M\circ\dots\circ\phi_1(k)}(x_1)=
\lim_{k\to+\infty} f_{\phi_M\circ\dots\circ\phi_1(k)}(x_1)=f^{(M)}(x_1),$$
so we can define~$f:(0,1/2)\to[0,+\infty]$ by setting~$f(x_1):=
f^{(M)}(x_1)$, for some~$M$ so large that~$C_2M^{1-n}<x_1$.
Hence, we consider the diagonal subsequence
$f_{\phi_k\circ\dots\circ\phi_1(k)}$ and we prove that it converges to~$f$
a.e. in~$(0,\,1/2)$. For this scope, we fix~$\varepsilon>0$,
$x_1\in (0,\,1/2)$
(possibly outside a set of measure zero) and~$M_{x_1}
\in\mathbb N$ such that~$C_2M_{x_1}^{1-n}<x_1$ and
we use~\eqref{5.13}
to find~$k(\varepsilon,x_1)$ such that, if~$k\ge k(\varepsilon,x_1)$, then
$$ |f_{\phi_{M_{x_1}}\circ\dots\circ\phi_1(k)} (x_1)-f(x_1)|\le \varepsilon.$$
Now, if~$k\ge M_{x_1}+k(\varepsilon,x_1)$,
we have that~$\phi_k\circ\dots\circ\phi_1(k)$
is a subsequence of~$\phi_{M_{x_1}}\circ\dots\circ\phi_1(k)$ and thus
$$ |f_{\phi_{k}\circ\dots\circ\phi_1(k)} (x_1)-f(x_1)|\le \varepsilon.$$
Since~$\varepsilon$ is arbitrary, this shows that~$f_{\phi_k\circ\dots\circ\phi_1(k)}(x_1)\to
f(x_1)$, which in turn completes the proof of~\eqref{c.a.e.}.
Now we identify $f_k$ with the subsequence constructed
in~\eqref{c.a.e.} and prove that
\begin{equation}\label{c.L1}
f_k^{n-1}\to f^{n-1} {\mbox{ in }}L^1(0,1/2),{\mbox{ as }}k\to+\infty.
\end{equation}
For this scope, we fix~$\delta>0$
and we use~\eqref{c.a.e.} and Egoroff's Theorem to find~$E_\delta\subseteq(0,1/2)$
such that~$|(0,1/2)\setminus E_\delta|\le\delta$ and~$f_k^{n-1}\to f^{n-1}$
uniformly on~$E_\delta$. We choose~$M_\delta:= \delta^{-1/2}$
and we use~\eqref{FUNF INT} to conclude that
\begin{equation*}
\int_0^{\varepsilon_k(M_\delta)} f_k^{n-1}(x_1)\,dx_1\le
\frac{C_\star}{M_\delta^{s(n-1)}}= C_\star \delta^{\frac{s(n-1)}{2}},\end{equation*}
for every~$k$.
Therefore
\begin{equation}\label{POI1}\begin{split}
& \int_0^{1/2} |f^{n-1}_k(x_1)-f^{n-1}_h(x_1)|\,dx_1\le
\int_{(0,1/2)\setminus E_\delta} \big(
f^{n-1}_k(x_1)+f^{n-1}_h(x_1)\big)\,dx_1
+ \|f^{n-1}_k-f_h^{n-1}\|_{L^\infty(E_\delta)} |E_\delta|\\
&\quad\le 2C_\star \delta^{\frac{s(n-1)}{2}}+
\int_{(\varepsilon_k(M_\delta),1/2)\setminus E_\delta} f^{n-1}_k(x_1)\,dx_1
+\int_{(\varepsilon_h(M_\delta),1/2)\setminus E_\delta} f^{n-1}_h(x_1)\,dx_1
+\|f^{n-1}_k-f_h^{n-1}\|_{L^\infty(E_\delta)}
.\end{split}\end{equation}
Now we recall that, by~\eqref{5.4bis}, $f_k\le M_\delta$
in~$(\varepsilon_k(M_\delta),1/2)$, thus
$$ \int_{(\varepsilon_k(M_\delta),1/2)\setminus E_\delta} f^{n-1}_k(x_1)\,dx_1
\le M_\delta\, \big|(\varepsilon_k(M_\delta),1/2)\setminus E_\delta\big|
\le M_\delta \delta=\sqrt\delta.$$
The same holds with~$h$ instead of~$k$. Consequently, formula~\eqref{POI1}
gives that
$$ \int_0^{1/2} |f^{n-1}_k(x_1)-f^{n-1}_h(x_1)|\,dx_1\le
2C_\star \delta^{\frac{s(n-1)}{2}} +2\sqrt\delta+
\|f^{n-1}_k-f_h^{n-1}\|_{L^\infty(E_\delta)}.$$
So, if we choose~$h$, $k$ so large that~$\|f^{n-1}_k-f_h^{n-1}\|_{L^\infty(E_\delta)}\le\sqrt\delta$,
we obtain
$$ \int_0^{1/2} |f^{n-1}_k(x_1)-f^{n-1}_h(x_1)|\,dx_1\le
2C_\star \delta^{\frac{s(n-1)}{2}} +3\sqrt\delta.$$
Since $\delta$ was arbitrarily fixed, we have just shown that~$f_k$
is a Cauchy sequence on~$L^1(0,1/2)$, which implies~\eqref{c.L1}.
The desired claim now follows from~\eqref{c.L1} and an even reflection
in~$(-1/2,0)$.
\end{proof}
\section{Proof of Theorem~\ref{THE EX}}\label{S1}
The proof uses the direct methods of the calculus of variations,
combined with
the fine estimates of Propositions~\ref{energy}
and~\ref{converge}. That is, we take a minimizing sequence
of sets~$F_k\in{\mathcal{K}}$ with~$|F_k|=\mu$ and
\begin{equation}\label{sequence}
\lim_{k\to+\infty} P_S(F_k) =\inf_{{F\subseteq S}\atop{|F|=\mu}} P_S(F).\end{equation}
Since~$F_k\in{\mathcal{K}}$,
we have that~$F_k$
has the form
$$ F_k= \Big\{ (x_1,x')\in S {\mbox{ s.t. }} |x'|\le f_k(x_1)\Big\},$$
with~$f_k:[-1/2,\,1/2]\to[0,+\infty]$ even and decreasing in~$[0,1/2]$.
We remark that, for any~$k\in\mathbb N$,
$$ \mu = |F_k| = C_o \,\int_0^{1/2} f_k^{n-1}(x_1)\,dx_1,$$
for some dimensional constant~$C_o>0$.
Also, we can fix a set~$F_o$ with~$P_S(F_o)<+\infty$
(recall for instance Corollary~\ref{finite}),
and we may assume that
\begin{equation}\label{6.1bis}
{\mbox{$P_S(F_k)\le
P_S(F_o)$ for every~$k\in\mathbb N$.}}
\end{equation}
We want to prove that
\begin{equation}\label{-CONV-1}
{\mbox{up to a subsequence, $f_k\to f$ a.e. in $[-1/2,\,1/2]$,
with }} C_o\,\int_0^{1/2} f^{n-1}(x_1)\,dx_1=\mu.
\end{equation}
Indeed, if there is a sequence along which
$$ \sup_{k\in\mathbb N} \|f_k\|_{L^\infty([-1/2,\,1/2])} <+\infty$$
then
$$ \sup_{k\in\mathbb N} \|f_k^{n-1}\|_{BV([-1/2,\,1/2])} <+\infty$$
and so, again up to a subsequence,
we can pass to the limit in $L^1([-1/2,\,1/2])$
and a.e. in~$[-1/2,\,1/2]$, see e.g. Theorem~3.23 in~\cite{AFP},
and obtain~\eqref{-CONV-1}. Thus, we can suppose that
$$ \sup_{k\in\mathbb N} \|f_k\|_{L^\infty([-1/2,\,1/2])} =+\infty.$$
In this case, we check that the assumptions of Proposition~\ref{converge}
are satisfied. For this, let~$\varepsilon_{*,k}\le \alpha_{*,k}\in [0,\,1/2]$ such that
\begin{equation}\label{LAL1bis}
f_k(x_1)\ge 4 {\mbox{ for all }} x_1\in [0,\varepsilon_{*,k})
\end{equation}
and
\begin{equation}\label{LAL2bis}
f_k(x_1)\ge 2{f_k(y_1)} {\mbox{ for all }} x_1\in [0,\varepsilon_{*,k})
{\mbox{ and }} y_1\in (\alpha_{*,k},\,1/2].\end{equation}
We observe that~\eqref{LAL1bis} and~\eqref{LAL2bis}
say that~\eqref{eps star}
and~\eqref{a star} are satisfied (for all the indices~$k$),
hence we can use
Proposition~\ref{energy} and conclude that, for every~$k\in\mathbb N$,
$$ P_S(F_k)\ge C \,
\int_{0}^{\varepsilon_{*,k}} \left[
\int_{\alpha_{*,k}}^{1/2}
\frac{f^{n-1}_k(x_1)}{|x_1-y_1|^{1+s}}\,dy_1\right]\,dx_1,$$
for a suitable~$C>0$.
Therefore, for every~$k\in\mathbb N$,
$$ P_S(F_o)\ge C \,
\int_{0}^{\varepsilon_{*,k}} \left[
\int_{\alpha_{*,k}}^{1/2}
\frac{f^{n-1}_k(x_1)}{|x_1-y_1|^{1+s}}\,dy_1\right]\,dx_1,$$
thanks to~\eqref{6.1bis},
and this gives that condition~\eqref{LAL3}
is satisfied in this case. Consequently,
\eqref{-CONV-1} follows from Proposition~\ref{converge}.
Thus, we define
$$ F_* :=\Big\{ (x_1,x')\in S {\mbox{ s.t. }} |x'|\le f(x_1)\Big\},$$
and we show that~$F_*$ is the desired minimizer.
First of all, $|F_*|=\mu$, thanks to the integral constraint in~\eqref{-CONV-1}.
Furthermore, as~$k\to+\infty$,
\begin{equation}\label{sequence2}
{\mbox{$\chi_{F_k}\to\chi_{F_*}$ a.e. in $S$.}}
\end{equation}
To check this, we recall that, by~\eqref{-CONV-1},
$f_k\to f$ in~$S\setminus Z_1$, with~$|Z_1|=0$.
Moreover, we have that, for any fixed~$x_1\in[-1/2,\,1/2]$,
the set~$\Lambda^{x_1}:=
\{x'\in\mathbb R^{n-1}
{\mbox{ s.t. }} |x'|=f(x_1)\}$ is a sphere in~$\mathbb R^{n-1}$
and so it is of measure zero in~$\mathbb R^{n-1}$ (in symbols, $|\Lambda^{x_1}|=0$).
Thus, if~$Z_2:= \{(x_1,x')\in S {\mbox{ s.t. }}|x'|=f(x_1) \}$,
we have, by Fubini's Theorem, that
$$ |Z_2| =\int_{-1/2}^{1/2}\,dx_1 \int_{\mathbb R^{n-1}}\,dx'\chi_{\{f(x_1)=|x'|\}}(|x'|)
=\int_{-1/2}^{1/2}\,dx_1 |\Lambda^{x_1}|=0.$$
Therefore, \eqref{sequence2} would follow if we show that~$f_k\to f$
in~$S\setminus (Z_1\cup Z_2)$. For this, fix~$x\in S\setminus (Z_1\cup Z_2)$.
Since~$x\not\in Z_2$ we have that either~$|x'|<f(x_1)$
or~$|x'|>f(x_1)$. Since~$x\not\in Z_1$ we have that~$f_k(x)\to f(x)$,
so that for large~$k$, either~$|x'|<f_k(x_1)$
or~$|x'|>f_k(x_1)$, respectively. This shows that, for large~$k$,
$x\in F_*$ if and only if~$x\in F_k$, therefore~$\chi_{F_k}(x)=\chi_{F}(x)$,
and this proves~\eqref{sequence2}.
Consequently, using~\eqref{sequence}
and~\eqref{sequence2}, we have, by Fatou Lemma,
\begin{eqnarray*}
&&\inf_{{F\subseteq S}\atop{|F|=\mu}} P_S(F)=
\lim_{k\to+\infty} P_S(F_k) =
\lim_{k\to+\infty}
\int_{F_k} \int_{S\setminus F_k} K(x-y)\,dx\,dy \\ &&\qquad=
\lim_{k\to+\infty}
\int_{S} \int_{S} \chi_{F_k}(x)\,\chi_{S\setminus F_k}(y)\,K(x-y)\,dx\,dy
\ge \int_{S} \int_{S} \chi_{F_*}(x)\,\chi_{S\setminus F_*}(y)\,K(x-y)\,dx\,dy
=P_S(F_*).
\end{eqnarray*}
This shows the desired minimization property and it ends the proof
of Theorem~\ref{THE EX}.
\section{Proof of Theorem~\ref{TH DX}}\label{S4}
We suppose that~$f(1/4)\ge \beta \mu^{\frac{1}{n-1}}$, for some~$\beta\ge0$,
and we obtain an estimate on~$\beta$. For this we use the volume constraint
and we observe that
$$ C_o \mu =\int_{0}^{1} f^{n-1}(x_1)\,dx_1\ge
\int_{0}^{1/8} f^{n-1}(x_1)\,dx_1\ge \frac{f^{n-1}(1/8)}{8},$$
for some~$C_o>0$.
That is, by monotonicity, we have that
$$ \beta \mu^{\frac{1}{n-1}} \le f(1/4) \le f(x_1)\le f(1/8)\le
C_1\mu^{\frac{1}{n-1}}$$
for every~$x_1\in[1/8,\,1/4]$ (here and in the sequel~$C_i>0$
is an appropriate dimensional constant). Notice that this already says that
\begin{equation}\label{beta}
\beta\le C_1.
\end{equation}
As a consequence, we see that~\eqref{density 0}
is obvious if~$\mu\ge 1/16^{n-1}$, so we suppose from now on
that
\begin{equation}\label{16}\mu\in(0,1/16^{n-1}).\end{equation}
We let~$B$ the ball with volume~$\mu$
(say, centered at the origin) and we
use the minimality property of~$F_*$ and~\eqref{Per s 1}
to see that
\begin{eqnarray*}
C_2\mu^{\frac{n-s}{n}} &=& {\rm Per}_s(B)
\\ &\ge& P_S(B)\\
&\ge& P_S(F_*) \\
&\ge& C_3 \int_{1/8}^{1/4} \,dx_1
\int_{1/8}^{1/4} \,dy_1 \int_{|x'|\le f(x_1)}\,dx'
\int_{|y'|>f(y_1)}\,dy' \frac{1}{|x-y|^{n+s}} \\
&\ge& C_3 \int_{1/8}^{1/4} \,dx_1
\int_{1/8}^{1/4} \,dy_1 \int_{|x'|\le \beta \mu^{\frac{1}{n-1}}}\,dx'
\int_{|y'|> C_1 \mu^{\frac{1}{n-1}}}\,dy' \frac{1}{|x-y|^{n+s}} .
\end{eqnarray*}
Now we let~$M:=\mu^{-\frac{1}{n-1}}$. Notice that~$M\ge 16$,
thanks to~\eqref{16},
thus, if~$N$ is the integer part of~$M/8$, we have that~$N\le M/8$
and
\begin{equation}\label{8.2bis} N\ge \frac{M}{8}-1\ge\frac{M}{16}.\end{equation}
Hence we change variables~$X:= Mx$ and~$Y:=My$
and then we translate in the first coordinate, and we obtain
\begin{equation}\label{83}\begin{split}
C_4\mu^{\frac{n-s}{n}}\,&\ge M^{s-n}
\int_{M/8}^{M/4} \,dX_1
\int_{M/8}^{M/4} \,dY_1 \int_{|X'|\le \beta M \mu^{\frac{1}{n-1}}}\,dX'
\int_{|Y'|> C_1 M \mu^{\frac{1}{n-1}}}\,dY' \frac{1}{|X-Y|^{n+s}}
\\ &=
M^{s-n}
\int_{0}^{M/8} \,dx_1
\int_{0}^{M/8} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}}
\\ &\ge
M^{s-n}
\int_{0}^{N} \,dx_1
\int_{0}^{N} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}}
\\ &\ge
M^{s-n}\sum_{k=1}^N
\int_{k-1}^{k} \,dx_1
\int_{k-1}^{k} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}}
\\ &=
M^{s-n}\sum_{k=1}^N
\int_{0}^{1} \,dx_1
\int_{0}^{1} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}}
\\ &= M^{s-n} N \int_{0}^{1} \,dx_1
\int_{0}^{1} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}}
.\end{split}\end{equation}
Now we observe that
if~$x_1$, $y_1\in[0,\,1]$,
$|x'|\le\beta$ and~$|y'|> C_1+1$, then
$$ |x_1-y_1|\le 1 \le C_1+1-\beta\le |y'|-|x'|\le |x'-y'|,$$
thanks to~\eqref{beta}, thus in this case
$$ |x-y|\le C_5 |x'-y'|\le C_5\big( |x'|+|y'|\big)\le
C_5 \big( \beta+|y'|\big)\le C_5 \big( C_1+|y'|\big)\le 2C_5\,|y'|.$$
Accordingly
\begin{eqnarray*} && \int_{0}^{1} \,dx_1
\int_{0}^{1} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1}\,dy' \frac{1}{|x-y|^{n+s}} \ge
\int_{0}^{1} \,dx_1
\int_{0}^{1} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1+1}\,dy' \frac{1}{|x-y|^{n+s}} \\
&&\quad \ge C_6 \int_{0}^{1} \,dx_1
\int_{0}^{1} \,dy_1 \int_{|x'|\le \beta }\,dx'
\int_{|y'|> C_1+1}\,dy' \frac{1}{|y'|^{n+s}} = C_7 \beta^{n-1}.
\end{eqnarray*}
By inserting this information into~\eqref{83} and using~\eqref{8.2bis}
we obtain that
$$ \mu^{\frac{n-s}{n}}\ge C_8 \,M^{s-n}N\beta^{n-1}\ge \frac{C_8\,M^{1+s-n}\beta^{n-1}}{16}=
C_9 \mu^{\frac{n-s-1}{n-1}}\beta^{n-1},$$
and this gives that~$\beta\le C_{10} \mu^{\frac{s}{n^2(n-1)}}$.
This proves~\eqref{density 0}.
Now we prove~\eqref{density 1}. For this we use the monotonicity of~$f$,
the volume constraint
and~\eqref{density 0} to compute
$$\frac{\big|F_*\cap \{|x_1|\ge 1/4\}\big|}{|F_*|}= C_{11}\mu^{-1}
\int_{1/4}^1 f^{n-1}(x_1)\,dx_1 \le C_{12} \mu^{-1} f^{n-1}(1/4)
\le C_{13} \mu^{\frac{s}{n^2}}, $$
which implies~\eqref{density 1}.
The proof of Theorem~\ref{TH DX}
is thus complete.
\section{Proof of Theorem~\ref{YES BALL}}\label{S5}
We take $B$ the ball of volume~$\mu$ (say, centered at the origin).
Using the minimality of~$F_*$ and Proposition~\ref{EST PER-s},
we see that
\begin{equation}\label{678d9f033ehgfsdn}
0\le \frac{P_S(B)-P_S(F_*)}{\mu^{\frac{n-s}{n}}}
\le\frac{ {\rm Per}_s(B)
-{\rm Per}_s(F_*)
+C\,\Big( |F_*|^2 + \Pi_S(F_*)\Big) }{\mu^{\frac{n-s}{n}}}.
\end{equation}
Now we use Theorem~\ref{TH DX}, so we write
$$ F_*=
\Big\{ (x_1,x')\in S {\mbox{ s.t. }} |x'|\le f(x_1)\Big\}$$
with
$$ \frac{f(1/4)}{ \mu^{\frac{1}{n-1}} }\le C\,\mu^{\frac{s}{n^2(n-1)}}.$$
In particular, using the monotonicity of~$f$, we have that, for any~$x_1\in
[1/4,\,1/2]$
$$ f(x_1)\le f(1/4) \le C\,\mu^{\frac{s}{n^2(n-1)}} \times \mu^{\frac{1}{n-1}}
=:r.$$
This (in the notation of~\eqref{hat}) says that~$\widetilde F_*$
and~$\widehat F_*$ are contained in the cylinder of radius~$r$, and therefore,
by Lemma~\ref{play},
$$ \Pi_S(F_*)\le C\,r^{n-s} = C \mu^{\frac{s(n-s)}{n^2(n-1)}}
\times \mu^{\frac{n-s}{n-1}}.$$
Using this and the fact that~$|F_*|=\mu=|B|$, we write~\eqref{678d9f033ehgfsdn}
as
\begin{equation}\label{1-11678d9f033ehgfsdn}
0\le\frac{ {\rm Per}_s(B) }{|B|^{\frac{n-s}{n}}}
-\frac{ {\rm Per}_s(F_*) }{|F_*|^{\frac{n-s}{n}}}
+C\,\Big( \mu^{\frac{n+s}{n}}
+
\mu^{\frac{n^2-s^2}{n^2(n-1)}}
\Big) .
\end{equation}
Now we use the quantitative isoperimetric inequality in
Theorem~1.1 of~\cite{I5}, according to which
$$ \frac{ {\rm Per}_s(F_*) }{|F_*|^{\frac{n-s}{n}}}
\ge
\frac{ {\rm Per}_s(B) }{|B|^{\frac{n-s}{n}}}
\,\Big( 1+ C\,{\rm Def}^2(F^*)\Big).$$
By inserting this into~\eqref{1-11678d9f033ehgfsdn}
we conclude that
\begin{equation}\label{9sicx3ed} 0\le -\frac{ {\rm Per}_s(B) }{|B|^{\frac{n-s}{n}}}\,
C\,{\rm Def}^2(F^*)
+C\,\Big( \mu^{\frac{n+s}{n}}
+
\mu^{\frac{n^2-s^2}{n^2(n-1)}}\Big).\end{equation}
Since
$$ \frac{n^2-s^2}{n^2(n-1)} =
\frac{n-s}{n (n-1)}
\frac{(n+s)}{n} \le
\frac{n}{n (n-1)}
\frac{(n+s)}{n} \le \frac{(n+s)}{n},$$
we see that~\eqref{9sicx3ed}
implies the thesis of Theorem~\ref{YES BALL}.
\section*{Acknowledgements}
J. D{\'a}vila and M. del Pino
have been supported by grants Fondecyt 1130360 and 1150066, Fondo
Basal
CMM and by Millenium Nucleus CAPDE NC130017.
S. Dipierro has been supported by Alexander von Humboldt-Stiftung.
E. Valdinoci
has been supported by PRIN grant
201274FYK7 ``Critical Point Theory
and Perturbative Methods for Nonlinear Differential Equations''
and ERC
grant 277749 ``EPSILON Elliptic Pde's and Symmetry of Interfaces and
Layers for Odd Nonlinearities''.
|
\subsection*{Extended abstract} We introduce
certain f\/inite-dimensional algebras denoted by ${\cal{PC}}_n$ and
${\cal{PF}}_{n,m}$ which are certain quotients of the plactic algebra
${\cal{P}}_{n}$, which had been introduced by A.~Lascoux and
\mbox{M.-P.}~Sch\"{u}t\-zen\-berger \cite{LS1}. We show that
$\dim({\cal{PF}}_{n,k})$ is equal to the number of symmetric plane partitions
f\/itting inside the box $n \times k \times k$,
$\dim({\cal{PC}}_{n})$ is equal to the number of alternating sign matrices of
size $n \times n$, moreover,
\begin{gather*}
\dim({\cal{PF}}_{n,n}) = {\rm TSPP}(n+1) \times {\rm TSSCPP}(n),\\
\dim({\cal{PF}}_{n,n+1})= {\rm TSPP}(n+1) \times {\rm TSSCPP}(n+1),\\
\dim({\cal{PF}}_{n+2,n})= \dim({\cal{PF}}_{n,n+1}),\qquad
\dim({\cal{PF}}_{n+3,n})={\frac{1}{2}} \dim({\cal{PF}}_{n+1,n+1}),
\end{gather*}
and study decomposition of the Cauchy kernels corresponding to the algebras
${\cal{PC}}_n$ and ${\cal{PF}}_{n,m}$;
as well as introduce polynomials which are common generalizations of the
(double) Schubert, $\beta$-Grothendieck, Demazure (known also as {\it key polynomials}), (plactic) key-Grothendieck, (plactic) Stanley and stable $\beta$-Grothendieck polynomials. Using a family of the Hecke type divided dif\/ference operators
we introduce polynomials which are common generalizations of the Schubert,
$\beta$-Grothendieck, dual $\beta$-Grothendieck, $\beta$-Demazure--Grothendieck, and Di Francesco--Zinn-Justin polynomials.
We also introduce and study some properties of the double af\/f\/ine
nilCoxeter algebras and related polynomials,
put forward a $q$-deformed version of the Knuth relations and
plactic algebra.
\section{Introduction}
The Grothendieck polynomials had been introduced by A.~Lascoux and M.-P.~Sch\"{u}t\-zen\-berger in~\cite{LS6} and studied in detail in~\cite{L33}. There are two equivalent versions of the Grothendieck polynomials depending on a choice
of a basis in the Grothendieck ring $K^{\star}({\cal{F}}l_n)$ of the
complete f\/lag variety ${\cal{F}}l_n$. The basis $\{\exp(\xi_{1}), \ldots,\exp(\xi_{n}) \}$ in $K^{*}({\cal{F}}l_n)$ is one choice, and another choice is the
basis $\{1-\exp(- {\xi}_{j}), \,1 \le j \le n\}$, where $\{{\xi}_{j},
\, 1 \le j \le n \}$ denote the Chern classes of the tautological linear
bundles $L_{j}$ over the f\/lag variety ${\cal{F}}l_n$. In the present paper we
use the basis in a deformed Grothendieck ring $K^{*, \beta}({\cal{F}}_n)$ of
the f\/lag variety ${\cal{F}}l_n$ generated by the set of elements $\{x_{i} =
x_{i}^{(\beta)} = 1-\exp(\beta {\xi}_{i}),\, i=1, \ldots, n \}$. This basis has
been introduced and used for construction of the $\beta$-Grothendieck
polynomials in~\cite{FK1,FK}.
A basis in the classical Grothendieck ring of the f\/lag variety in question
corresponds to the choice $\beta = -1$. For arbitrary $\beta$ the ring
generated by the elements $\big\{x_{i}^{(\beta)},\, 1 \le i \le n \big\}$ has been
identif\/ied with the Grothendieck ring corresponding to the generalized
cohomology theory associated with the multiplicative formal group law
$F(x,y)=x+y+\beta x y$, see \cite{Hu}. The Grothendieck polynomials
corresponding to the classical $K$-theory ring $K^{\star}({\cal{F}}l_{n})$,
i.e., the case $\beta = -1$, had been studied in depth by A.~Lascoux and
M.-P.~Sch\"{u}t\-zen\-berger in \cite{LS7}. The $\beta$-Grothendieck
polynomials has been studied in~\cite{FK1,FK2,Hu}.
The {\it plactic monoid} over a f\/inite totally ordered set $\mathbb{A}=
\{a < b <c < \cdots < d \}$ is the quotient of the free monoid generated by
elements from $\mathbb{A}$ subject to the elementary Knuth transformations~\cite{Kn}
\begin{gather}\label{equation1.1}
bca = bac\quad \& \quad acb = cab, \qquad \text{and}\qquad bab=bba \quad \& \quad aba=baa,
\end{gather}
for any triple $\{a < b < c \} \subset \mathbb{A}$.
To our knowledge, the concept of ``plactic monoid'' has its origins in a
paper by C.~Schensted~\cite{Sc}, concerning the study of the longest
increasing subsequence of a permutation, and a~paper by D.~Knuth~\cite{Kn},
concerning the study of combinatorial and algebraic properties of the
Robinson--Schensted correspondence\footnote{See, e.g., \url{https://en.wikipedia.org/wiki/Robinson-Schensted_correspondence}.}.
As far as we know, this monoid and the (unital) algebra $\cal{P}(\mathbb{A})$ corresponding to that monoid\footnote{If $\mathbb{A} = \{1 < 2 < \cdots < n \}$, the elements of the
algebra ${\cal{P}}(\mathbb{A})$ can be identif\/ied with semistandard Young
tableaux. It was discovered by D.~Knuth~\cite{Kn} that modulo {\it Knuth
equivalence} the equivalence classes of semistandard Young tableaux form an
algebra, and he has named this algebra by {\it tableaux algebra}. It is easily
seen that the tableaux algebra introduced by D.~Knuth is isomorphic to the
algebra introduced by M.-P.~Sch\"{u}tzenberger~\cite{Sch}.}, had been introduced, studied and used by M.-P.~Sch\"{u}tzenberger, see
\cite[Section~5]{Sch}, to give the f\/irst complete proof of the famous
\textit{Littlewood--Richardson rule} in the theory of symmetric functions. A~bit
later this monoid, was named the ``mono\"{i}de plaxique'' and studied in depth
by A.~Lascoux and M.-P.~Sch\"{u}tzenberger~\cite{LS1}. The algebra
corresponding to the plactic monoid is commonly known as
{\it plactic algebra}.
One of the basic properties of the plactic algebra~\cite{Sch} is that it
contains the distinguished commutative subalgebra which is generated by
noncommutative elementary (quasi-symmetric) polynomials\footnote{See, e.g., \cite{KT} for def\/inition of noncommutative quasi-symmetric functions and polynomials.}
\begin{gather*
e_k({\mathbb{A}}_n) = \sum_{i_1 > i_2 > \cdots > i_k} a_{i_{1}} a_{i_{2}}
\cdots a_{i_{k}}, \qquad k= 1, \ldots,n,
\end{gather*}
see, e.g., \cite[Corollary~5.9]{Sch} and~\cite{FG}.
We refer the reader to nice written overview~\cite{LLT} of the basic
properties and applications of the plactic monoid in combinatorics.
It is easy to see that the plactic relations for two letters $a < b$, namely,
\begin{gather*}
aba=baa,\qquad bab=bba,
\end{gather*}
imply the commutativity of noncommutative elementary polynomials in two
variables. In other words, the plactic relations for two letters imply that
\begin{gather*}
ba (a+b)=(a+b) ba, \qquad a < b.
\end{gather*}
It has been proved in~\cite{FG} that these relations together with the Knuth
relations~\eqref{equation1.1} for three letters $a < b < c$, imply the commutativity
of noncommutative elementary quasi-symmetric polynomials for any number of
variables.
In the present paper we prove that in fact the commutativity of noncommutative
elementary quasi-symmetric polynomials for $n=2$ and $n=3$ implies the
commutativity of that polynomials for all~$n$, see Theorem~\ref{theorem2.23}\footnote{Let \looseness=-1 us stress that conditions necessary and suf\/f\/icient to
assure the commutativity of noncommutative elementary polynomials for the
number of variables equals $n=2$ and $n=3$ turn out to be weaker then that
listed in~\cite{FG}.}.
One of the main objectives of the present paper is to study combinatorial
properties of the generalized plactic Cauchy kernel
\begin{gather*
{\cal C}({\mathfrak{P}}_n,U)= \prod_{i=1}^{n-1}
\left \{{\prod_{j=n-1}^{i}(1+ p_{i,j-i+1} u_{j})} \right\},
\end{gather*}
where ${\mathfrak{P}}_n$ stands for the set of parameters $\{p_{ij},\, 2 \le i+j \le n+1, \,i >1,\,j>1 \}$, and $U := U_{n}$ stands for a certain noncommutative
algebra we are interested in, see Section~\ref{section5}.
We also want to bring to the attention of the reader on some
interesting combinatorial properties of {\it rectangular} Cauchy kernels{\samepage
\begin{gather*
{\cal{F}}({\mathfrak{P}}_{n,m}, U) = \prod_{i=1}^{n-1} \left\{\prod_{j=m-1}^{1} (1 + p_{i,\overline{i-j+1}^{(m)}} u_{j}) \right\},
\end{gather*}
where ${\mathfrak{P}}_{n,m} = \{p_{ij} \}_{1 \le i \le n \atop 1\le j \le m}$;
see Def\/inition~\ref{def6.6} for the meaning of symbol~$\overline{a}^{(m)}$.}
We treat these kernels in the (reduced) plactic algebras~${\cal{PC}}_n$ and
${\cal{PF}}_{n,m}$ correspondingly. The algebras~${\cal{PC}}_n$ and~${\cal{PF}}_{n,m}$ are f\/inite-dimensional and have bases parameterized by
certain Young tableaux described in Sections~\ref{section5.1} and~\ref{section6} correspondingly.
Decomposition of the rectangular Cauchy kernel with respect to the basis in
the algebra ${\cal{PF}}_{n,m}$ mentioned above, gives rise to a~set of
polynomials which are common generalizations of the (double) Schubert.
$\beta$-Gro\-thendieck, Demazure and Stanley polynomials. To be more precise,
the polynomials listed above correspond to certain quotients of the plactic
algebra ${\cal{PF}}_{n,m}$ and appropriate specializations of parameters~$\{p_{ij} \}$ involved in our def\/inition of polynomials
$U_{\alpha}(\{p_{ij} \})$, see Section~\ref{section6}.
As it was pointed out in the beginning of Introduction, the Knuth (or plactic)
relations~\eqref{equation1.1} have been discovered in~\cite{Kn} in the course of the
study of algebraic and combinatorial properties of the Robinson--Schensted
correspondence. Motivated by the study of basic properties of a~{\it quantum} version of the tropical/geometric Robinson--Schensted--Knuth correspondence~-- work in progress, but see~\cite{BK,AK,KB,NY,OP} for def\/inition and basic properties of the tropical/geometric RSK,~-- the author of the present paper came to a discovery that certain
deformations of the Knuth relations {\it preserve} the Hilbert series
(resp. the Hilbert polynomials) of the plactic algebras~${\cal{P}}_n$ and~${\cal{F}}_n$ (resp.\ the algebras~${\cal{PC}}_n$ and~${\cal{PF}}_n$).
More precisely, let $\{q_2,\ldots,q_{n} \}$ be a set of (mutually commuting)
parameters, and ${\mathbb{U}}_n:=\{u_1,\ldots,u_{n} \}$ be a set of
generators of the free associative algebra over~$\mathbb{Q}$ of rank~$n$. Let $Y, Z
\subset [1,n]$ be subsets such that $Y \cup Z=[1,n]$ and $Y \cap Z=\varnothing$. Let us set $p(a)=0$, if $a \in Y$ and $p(a)=1$, if $a \in Z$.
Def\/ine ${q}$-deformed super Knuth relations among the generators
$u_1, \ldots,u_{n}$ as follows:
\begin{align*}
{\rm SPL}_q \colon \ & (-1)^{p(i)p(k)} q_{k} u_j u_i u_k=u_j u_k u_i, \qquad i < j \le k,\\
& (-1)^{p(i)p(k)} q_{k} u_i u_k u_j=u_k u_i u_j, \qquad i \le j <k.
\end{align*}
We def\/ine
\begin{itemize}\itemsep=0pt
\item $q$-deformed superplactic algebra ${\cal{SQP}}_n$ to
be the quotient of the free associative algebra $\mathbb{Q} \langle u_1,\ldots,u_{n}
\rangle$ by the two-sided ideal generated by the set of {\bf q}-deformed
Knuth rela\-tions~$({\rm SPL}_{\bf q})$,
\item reduced $q$-deformed superplactic algebras
${\cal{SQPC}}_n$ and ${\cal{SQPF}}_{n,m}$ to be the quotient of the algebra
${\cal{SQP}}_n$ by the two-sided ideals described in Def\/initions~\ref{def5.15} and~\ref{def6.7} cor\-respon\-dingly.
\end{itemize}
We state
\begin{con}\label{conjecture1.0}
The algebra ${\cal{SQP}}_n$ and the algebras ${\cal{SQPC}}_n$ and
${\cal{SQPF}}_{n,m}$, are flat deformations of the algebras
${\cal{P}}_n$, ${\cal{PC}}_n$ and ${\cal{PF}}_{n,m}$ correspondingly.
\end{con}
In fact one can consider more general deformation of the Knuth relations, for example take a set of parameters ${\bf Q}:=\{q_{ik},\, 1 \le i < k \le n \}$
and impose on the set of generators $\{u_1,\ldots,u_n \}$ the following
relations
\begin{gather*
q_{ik} u_j u_i u_k=u_j u_k u_i, \qquad i < j \le k,\qquad q_{ik} u_i u_k u_j=u_k u_i u_j, \qquad i \le j <k.
\end{gather*}
However we don't know how to describe a set of conditions on parameters~${\bf Q}$ which imply the f\/latness of the corresponding quotient algebra(s),
as well as we don't know an interpretation and dimension of the algebras
${\cal{SQPC}}_n$ and ${\cal{SQPF}}_{n,m}$ for a ``generic'' values of parameters~${\bf Q}$. We {\it expect} the dimension of algebras ${\cal{SQPC}}_n$ and
${\cal{SQPF}}_{n,m}$ each depends piece-wise polynomially on a set of
parameters $\{q_{ij} \in \mathbb{Z}_{\ge 0},\, 1 \le i < j \le n \}$, and pose a~problem to describe its poly\-no\-mia\-lity chambers.
We also mention and leave for a separate publication(s), the case of algebras
and polynomials associated with {\it superplactic} monoid~\cite{LNS, LT}, which corresponds to the relations ${\rm SPL}_q$ with $q_i =1$, $\forall\, i$. Finally we point out an interesting and important paper~\cite{Li} wherein
the case $Z= \varnothing$, and the all deformation parameters are equal to
each other, has been independently introduced and studied in depth.
Let us repeat that the important property of plactic algebras ${\cal{P}}_n$
is that the noncommutative elementary polynomials
\begin{gather*}
e_{k}(u_{1}, \ldots,n_{n-1}):= \sum_{n-1 \ge a_{1} \ge a_{2} \ge a_{k} \ge1}
u_{a_{1}} \cdots u_{a_{k}}, \qquad k=1,\ldots,n-1,
\end{gather*}
generate a commutative subalgebra inside of the plactic algebra
${\cal{P}}_n$, see, e.g.,~\cite{FG, LS1}. Therefore all our f\/inite-dimensional algebras introduced in the present paper, have a~distinguished
f\/inite-dimensional commutative subalgebra. We have in mined to describe
these algebras explicitly in a separate publication.
In Section~\ref{section2} we state and prove necessary and suf\/f\/icient conditions in order
the elementary noncommutative polynomials form a mutually commuting family.
Surprisingly enough to check the commutativity of noncommutative elementary
polynomials for any $n$, it's enough to check these
conditions only for $n= 2,3$. However a combinatorial meaning of a
generalization of the Lascoux--Sch\"{u}tzenberger plactic algebra ${\cal{P}}_n$
obtained in this way, is still missing.
The plactic algebra ${\cal{PF}}_{n,m}$ introduced in Section~\ref{section6}, has a monomial
basis parametrized by the set of Young tableaux of shape
$ \lambda \subset (n^{m})$ f\/illed by the numbers from the set
$\{1,\ldots,m \}$. In the case $n=m$ it is
well-known~\cite{G,KGV,Ma}, that this number is equal to the
number of symmetric plane partitions f\/itting inside the cube $n \times n \times n$. Surprisingly enough this number admits a~factorization in the product of
the number of totally symmetric plane partitions (${\rm TSPP}$) by the number of
totally symmetric self-complementary plane partitions~(${\rm TSSCPP}$) f\/it inside
the same cube. A similar phenomenon happens if $|m-n| \le 2$, see Section~\ref{section6}.
More precisely, we add to the well-known equalities
\begin{gather*}
\#|{\cal{B}}_{1,n}| =2^n, \qquad \#|{\cal{B}}_{2,n}| = {2 n +1 \choose n},\qquad
\#|{\cal{B}}_{3,n}| =2^n \operatorname{Cat}_{n+1} \qquad \cite[A003645]{SL}, \\
\#|{\cal{B}}_{4,n}| = {\frac{1}{2}} \operatorname{Cat}_{n+1} \operatorname{Cat}_{n+2} \qquad \cite[A000356]{SL},\\
\#|{\cal{B}}_{n,5}| =
\frac{{n+5 \choose 5} {n+7 \choose 7} {n+9 \choose 9}}{{n+2
\choose 2}{n+4 \choose 4}} \qquad \cite[A133348]{SL},
\end{gather*}
the following relations
\begin{gather*}
\#|{\cal{B}}_{n,n}| = {\rm TSPP}(n+1) \times {\rm ASM}(n), \qquad \#|{\cal{B}}_{n,n+1}| = {\rm TSPP}(n+1) \times {\rm ASM}(n+1), \\
\#|{\cal{B}}_{n+2,n}| =\#|{\cal{B}}_{n,n+1}|, \qquad \#|{\cal{B}}_{n+3,n}|= {\frac{1}{2}} \#|{\cal{B}}_{n+1,n+1}|,\\
\# |{\rm PP}(n)| =\#|{\rm TSSCPP}(n)| \times \#|{\rm ASMHT}(2n)|= \#|{\rm CSSCPP}(2n)| \times \#|{\rm CSPP}(n)|, \\
\#|{\rm CSPP}(2n)| = \#|{\rm TSPP}(2n)| \times \# |{\rm CSTCPP}(2n)|, \\
\#|{\rm CSPP}(2n+1)| = 2^{2 n} \#|{\rm TSPP}(2n+1)| \times \#|{\rm TSPP}(2n)|,
\end{gather*}
where ${\rm PP}(n)$ stands for the set of plane partitions f\/it in a cub of size
$n \times n \times n$; ${\rm AMSHT}(2n)$ denotes the set of alternating sign
matrices of size $2n \times 2n$ invariant under a half-turn and ${\rm CSSPP}(2n)$
denotes the set of cyclically symmetric self-complementary plane partitions
f\/itting inside a cub of size $2n \times 2n \times 2n$, see, e.g.,~\cite{B};
${\rm CSTCPP}(n)$ stands for the set of cyclically symmetric transpose complementary plane partitions f\/itting inside a cub of size $2n \times 2n \times 2n$, see, e.g.,~\cite[$A051255$]{SL}.
See Section~\ref{section6} for the def\/inition of the sets ${\cal{B}}_{n,m}$
and examples.
In Exercise~\ref{exer6.3} we state some (new) divisibility properties of
the numbers $\#|B_{n+4,n}|$.
It is well-known that ${\rm ASMHT}(2n)= {\rm ASM}(n) \times {\rm CSPP}(n)$, where ${\rm CSPP}(n)$
denotes the number of cyclically symmetric plane partitions f\/itting inside
$n$-cube, and ${\rm CSSCPP}(2n)={\rm ASM}(n)^2$, see,~e.g.,~\cite{B,Ku} and \cite[$A006366$]{SL}.
\begin{prb} \quad
\begin{itemize}\itemsep=0pt
\item Construct bijection between the set of plane partitions fit
inside $n$-cube and the set of $($ordered$)$ triples $(\pi_1,\pi_2, \wp)$,
where $ (\pi_1,\pi_2)$ is a pair of ${\rm TSSCPP}(n)$ and $\wp$ is a cyclically
symmetric plane partition fitting inside $n$-cube.
\item Describe the involution $\kappa \colon {\rm PP}(n) \longrightarrow {\rm PP}(n)$
which is induced by the involution $(\pi_1,\pi_2, \wp)$
$\longrightarrow (\pi_2,\pi_1,\wp)$ on the set ${\rm TSSCPP}(n) \times {\rm TSSCP}(n)
\times {\rm CSPP}(n)$, and its fixed points. Clearly one has $\# | {\rm Fix}(\kappa)| =
{\rm ASMHT}(2n)$.
\item Characterize pairs of plane partitions $(\Pi_1, \Pi_2) \in
{\rm PP}(n) \times {\rm PP}(n)$ such that
\begin{gather*}
(a) \quad \wp(\Pi_1)=\wp(\Pi_2); \qquad (b) \quad (\pi_1(\Pi_1),\pi_2(\Pi_1))=
(\pi_1(\Pi_2),\pi_2(\Pi_2)).
\end{gather*}
\end{itemize}
\end{prb}
These relations have straightforward proofs based on the explicit product
formulas for the numbers
\begin{gather*}
\#|{\rm SPP}(n)| = \prod_{1 \le i \le j \le k} \frac{n+i+j+k-1}{i+j+k-1},\\
\#|{\rm TSPP}(n)| =\prod_{i=1}^{n} \prod_{j=i}^{n} \prod_{k=j}^{n} \frac{i+j+k-1}{i+j+k-2}, \\
\#|{\rm PP}(n)|= \prod_{i=0}^{n} \prod_{j=1}^{n-1}{\frac{3 n -i -j}{2 n -i -j}}=
\prod_{i=1}^{n}{\frac{{2 n +i \choose n}}{{n +i \choose n}}},
\end{gather*}
but bijective proofs of these identities are an open problem.
It follows from \cite{AL, LS1} that the dimension of the (reduced)
plactic algebra ${\cal{PC}}_n$ is equal to the number of alternating sign
matrices of size $n \times n$ (note that ${\rm ASM}(n)={\rm TSSCPP}(n)$). Therefore the
key-Grothendieck polynomials can be obtained from $U$-polynomials (see
Section~\ref{section6}, Theorem~\ref{theorem6.12}) after the specialization
$p_{ij}=0$, if $i+j > n+1$.
In Section~\ref{section4} following \cite{Ki1} we introduce and study a
family of polynomials
which are a~common generalization of the Schubert, $\beta$-Grothendieck, dual
$\beta$-Grothendieck, $\beta$-Demazure, $\beta$-key-Grothendieck, Bott--Samelson and $q$-Demazure polynomials, Whittaker functions (see~\cite{BBL} and
Lemma~\ref{lem4.20}) and Di~Francesco--Zinn-Justin polynomials (see
Section~\ref{section4}).
Namely, for any permutation $w \in {\mathbb{S}}_n$ and composition $\zeta
\subset \delta_n := (n-1,n-2,\ldots,2,1)$, we introduce polynomials
\begin{gather*}
{\cal{KN}}_{w}^{(\beta,\alpha,\gamma,h)}(X_n)= h^{\ell(w)} T_{s_{i_{1}}}
\cdots T_{s_{i_{\ell}}} \big(x^{\delta_n}\big), \\
{\rm KD}_{\zeta}^{(\beta,\alpha,\gamma,h)}(X_n)= h^{\ell(v_{\zeta})} T_{s_{i_{1}}}
\cdots T_{s_{i_{\ell}}} \big(x^{\zeta^{+}}\big),
\end{gather*}
where
\begin{gather*}
\begin{split}
& T_{i}:=T_{i}^{(\beta,\alpha,\gamma,h)} =- \alpha + ((\alpha + \beta + \gamma) x_{i} +\gamma x_{i+1} +h \\
& \hphantom{T_{i}:=T_{i}^{(\beta,\alpha,\gamma,h)} =}{} +
h^{-1} (\alpha+\gamma)(\beta + \gamma) x_{i} x_{i+1}) \partial_{i,i+1},\qquad
i=1,\dots,n-1,
\end{split}
\end{gather*}
denote a collection of divided dif\/ference operators which satisfy the Coxeter
and Hecke relations
\begin{gather*}
T_{i} T_{j} T_{i} = T_{j} T_{i} T_{j}, \qquad \text{if}\quad |i-j|=1;\qquad
T_{i} T_{j}=T_{j} T_{i}, \qquad \text{if} \quad |i-j| \ge 2, \\
T_{i}^2 = (\beta -\alpha) T_{i} +\beta \alpha, \qquad i=1,\ldots,n-1;
\end{gather*}
by def\/inition for any permutation $w \in {\mathbb{S}}_n$ we set
\begin{gather*}
T_{w} := T_{s_{i_{1}}} \cdots T_{s_{i_{\ell}}},
\end{gather*}
for any reduced decomposition $w =s_{i_{1}} \cdots s_{i_{\ell}}$ of a permutation in question; $\zeta^{+}$ denotes a unique partition obtained from $\zeta$
by ordering its parts, and~$v_{\zeta} \in {\mathbb{S}}_n$ denotes the minimal
length permutation such that $v_{\zeta}(\zeta)=\zeta^{+}$.
Assume that $h=1$.\footnote{Clearly that if $h \not= 0$, then after rescaling parameters
$\alpha$, $\beta$ and $\gamma$
one can assume that $h=1$. However, see, e.g., \cite[Section~5]{Ki1}, the
parameter~$h$ plays important role in the study of dif\/ferent specializations
of the variab\-les~$x_i$, $1 \le i \le n-1$ and parameters~$\alpha$,~$\beta$ and~$\gamma$.} If $\alpha=\gamma = 0$, these polynomials coincide with
the $\beta$-Grothendieck polynomials~\cite{FK1}, if $\beta=\alpha =1$,
$\gamma=0$ these polynomials coincide with the Di Francesco--Zinn-Justin
polynomials~\cite{DZ1}, if $\beta=\gamma=0$, these polynomials coincide
with dual $\alpha$-Grothendieck polynomials ${\cal{H}}_{w}^{(\alpha)}(X_n)$,\footnote{To avoid the reader's confusion, let us explain that in our paper we use the letter~$\alpha$ either as the lower index to denote a composition, or
as the upper index to denote a parameter which appears in certain polynomials
treated in our paper. For example ${\cal{H}}_{\alpha}^{(\alpha)}(X_n)$ denotes
the dual $\alpha$-Grothendieck polynomial corresponding to a composition~$\alpha$. Note that the $\alpha$-Grothendieck polynomial ${\mathfrak{G}}_{w}^{(\alpha)}(X_n)$ can be obtained from the polynomial ${\mathfrak{G}}_{w}^{(\beta)}(X_n)$ by replacing~$\beta$ by~$\alpha$.} where by def\/inition we set $X_n:= (x_1,\ldots,x_n)$.
\begin{con} For any permutation $w \in {\mathbb{S}}_n$ and any composition
$\zeta \subset \delta_n$, polynomials ${\cal{KN}}_{w}^{(\beta,\alpha,\gamma,h)}(X_n)$ and ${\rm KD}_{\zeta}^{(\alpha,\beta,\gamma,h)}(X_n)$ have nonnegative coefficients, i.e.,
\begin{gather*}
{\cal{KN}}_{w}^{(\beta,\alpha,\gamma,h)}(X_n) \in \mathbb{N}[\alpha,\beta,\gamma,h]
[X_n], \qquad {\rm KD}_{\zeta}^{(\beta,\alpha,\gamma,h)}(X_n) \in \mathbb{N}[\alpha,\beta,\gamma,h] [X_n].
\end{gather*}
\end{con}
We {\it expect} that these polynomials have some geometrical meaning
to be discovered.
More generally we study divided dif\/ference type operators of the form
\begin{gather*}
T_{ij}:= T_{ij}^{(a,b,c,h,e)}= a + (b x_i +c x_j + h +e x_i x_j) \partial_{ij},
\end{gather*}
depending on parameters $a$, $b$, $c$, $h$, $e$ and satisfying the $2D$-Coxeter relations
\begin{gather*}
T_{ij} T_{jk} T_{ij}= T_{jk} T_{ij} T_{jk}, \qquad 1 \le i < j < k \le n,\\
T_{ij} T_{kl} =T_{kl} T_{ij}, \qquad \text{if}\quad \{i,j\} \cap \{k,l \}=\varnothing.
\end{gather*}
We f\/ind that the necessary and suf\/f\/icient condition which ensure the validity
of the $2D$-Coxeter relations is the following relation among the parameters\footnote{In other words, the divided dif\/ference operators $ \{T_{ij}:= T_{ij}^{(a,b,c,h,e)}\}$ which obey the $2D$-Coxeter relations, have the following
form:
\begin{gather*}
T_{ij}^{(a,b,c,h)}= \left(1+\frac{(a+b)}{h} x_i\right)\left(1+\frac{c-a}{h} x_{j}\right) \partial_{ij} +a \sigma_{ij},
\qquad \text{if}\quad h \not= 0, \quad \text{$a$, $b$, $c$ arbitrary}.
\end{gather*}
If $h=0$, then either $T_{ij}= ((c-a) x_j+e x_i x_j) \partial_{ij} +
a \sigma_{ij}$, or $T_{ij} =(a+b) x_i + e x_i x_j) \partial_{ij} +a \sigma_{ij}$,
where $\sigma_{ij}$ stands for the exchange operator: $\sigma_{ij}(F(z_i,z_j)) = F(z_j,z_i)$.}:
\begin{gather*}
(a+b)(a-c)+h e = 0.
\end{gather*}
Therefore, if the above relation between parameters $a$, $b$, $c$, $h$, $e$ holds, then
for any permutation $w \in {\mathbb{S}}_n$ the operator
\begin{gather*}
T_{w}:= T_{w}^{(a,b,c,h,e)}= T_{i_{1}}^{(a,b,c,h,e)} \cdots T_{i_{\ell}}^{(a,b,c,h,e)},
\end{gather*}
where $w=s_{i_{1}} \cdots s_{i_{\ell}}$ is any reduced decomposition of~$w$, is
\textit{well-defined}. Hence under the same assumption on parameters,
for any permutation $w \in {\mathbb{S}}_n$ one can attach the well-def\/ined
polynomial
\begin{gather*}
G_{w}^{(a,b,c,h,e)}(X,Y) := {T_{w}^{(x)}}^{(a,b,c,h,e)}\bigg(\prod_{i \ge 1,
j \ge 1 \atop i+j \le n+1}(x_i+y_j)\bigg),
\end{gather*}
and in much the same fashion to def\/ine polynomials
\begin{gather*}
D_{\alpha}^{(a,b,c,h,e)}(X,Y) := {T_{w_{\alpha}}^{(x)}}^{(a,b,c,h,e)}\big(x^{\alpha^{+}}\big)
\end{gather*}
for any composition $\alpha$ such that $\alpha_{i} \le n-i$, $\forall\, i$.
We have used the notation $ {T^{(x)}}_{w}^{(a,b,c,h,e)}$ to point out
that this operator acts only on the variables $X =(x_1,\ldots,x_n)$; for any
composition $\alpha \in \mathbb{Z}_{\ge 0}^n$, $\alpha^{+}$~denotes a unique
partition obtained from~$\alpha$ by reordering its parts in (weakly)
decreasing order, and~$w_{\alpha}$ denotes a unique minimal length
permutation in the symmetric group~$\mathbb{S}_{n}$ such that
$w_{\alpha}(\alpha)=\alpha^{+}$.
In the present paper we are interested in to list a conditions on
parameters $A:= \{a,b,c,h,e \}$ with the constraint
\begin{gather*
(a+b)(a-c)+h e=0,
\end{gather*}
which ensure that the above polynomials
$G_{w}^{(a,b,c,h,e)}(X)$ and $D_{\alpha}^{(a,b,c,h,e)}(X)$ or their
specialization $x_i=1$, $\forall\, i$, have nonnegative coef\/f\/icients.
We state the following conjectures:
\begin{itemize}\itemsep=0pt
\item ${\cal{KN}}_{w}^{(\beta,\alpha,\gamma)}(X_n) \in \mathbb{N}[\alpha, \beta, \gamma] [X_n]$,
\item $G_{w}^{(-b,a+b+c,c,1,(b+c)(a+c)}(X_n) \in \mathbb{N} [a,b,c] [X_n]$,
\item $G_{w}^{(-b,a+b+c,c+d,1,(b+c+d)(a+c)}(x_i=1,\, \forall\, i) \in
\mathbb{N}[a,b,c,d]$, where $a$, $b$, $c$, $d$ are free parameters.
\end{itemize}
In the present paper we treat the case
\begin{gather}\label{equation1.7}
A= (- \beta,\beta+\alpha+\gamma, \gamma,1,(\alpha+\gamma)(\beta+\gamma)).
\end{gather}
As it was pointed above, in this case polynomials $G_{w}^{A}(X)$ are common
generalization of Schubert, $\beta$-Grothendieck and dual $\beta$-Grothendieck,
and Di Francesco--Zinn-Justin polynomials. We expect a certain interpretation
of the polynomials $G_{w}^{A}$ for general $\beta$, $\alpha$ and $\gamma$.
As it was pointed out earlier, one of the basic properties of the plactic
monoid ${\cal{P}}_{n}$ is that the noncommutative elementary symmetric
polynomials $\{e_{k}(u_{1},\ldots,u_{n-1}) \}_{1 \le k \le n-1}$ generate
a~commutative subalgebra in the plactic algebra in question. One can
reformulate this statement as follows. Consider the generating function
\begin{gather*
A_{i}(x): = \prod_{a=n-1}^{i}(1+x u_{a}) = \sum_{a=0}^{i}e_{a}(u_{n-1}, \ldots,u_{i}) x^{i-a},
\end{gather*}
where we set $e_{0}(U)=1$. Then the commutativity property of noncommutative
elementary symmetric polynomials is equivalent to the following commutativity
relation in the plactic as well as in the generic plactic, algebras ${\cal{P}}_n$ and ${\mathfrak{P}}_n$~\cite{FG}, and Theorem~\ref{theorem2.23},
\begin{gather*}
A_{i}(x) A_{i}(y)= A_{i}(y) A_{i}(x), \qquad 1 \le i \le n-1.
\end{gather*}
Now let us consider the Cauchy kernel
\begin{gather*
{\cal{C}}({\mathfrak{P}}_n,U) = A_{1}(z_1) \cdots A_{n-1}(z_{n-1}),
\end{gather*}
where we assume that the pairwise commuting variables $z_1, \ldots,z_{n-1}$
commute with the all generators of the algebras~${\cal{P}}_n$ and
${\mathfrak{P}}_n$. In what follows we consider the natural completion
${\widehat{\mathfrak{P}}}_n$ of the plactic algebra~${\mathfrak{P}}_n$ to
allow consider elements of the form $(1 +x u_{i})^{-1}$. Elements of this
form exist in any Hecke type quotient of the plactic algebra
${\widetilde{\mathfrak{P}}}_n$. Having in mind
this assumption, let us compute the
action of divided dif\/ference operators $\partial_{i,i+1}^{z}$ on the Cauchy
kernel. In the computation below, the commutativity property of the elements
$A_{i}(x)$ and $A_{i}(y)$ plays the key role. Let us start computation of
$\partial_{i,i+1}^{z}({\cal{C}}({\mathfrak{P}}_n,U)) =
\partial_{i,i+1}^{z}(A_1(z_1) \cdots A_{n-1}(z_{n-1}))$.
First of all write $A_{i+1}(z_{i+1})= A_{i}(z_{i+1})(1+z_{i+1}
u_i)^{-1}$. According to the basic property of the elements~$A_i(x)$, one sees
that the expression $A_{i}(z_{i}) A_{i}(z_{i+1})$ is symmetric with respect
to $z_i$ and $z_{i+1}$, and hence is invariant under the action of divided
dif\/ference operator $\partial_{i,i+1}^{z}$. Therefore,
\begin{gather*}
\partial_{i,i+1}^{z}({\cal{C}}({\mathfrak{P}}_n,U)) = A_1(z_1) \cdots
A_{i}(z_i) A_{i}(z_{i+1}) \partial_{i,i+1}^{z}\big((1+z_{i+1} u_{i})^{-1}\big)\\
\hphantom{\partial_{i,i+1}^{z}({\cal{C}}({\mathfrak{P}}_n,U)) =}{}\times A_{i+2}(z_{i+2}) \cdots A_{n-1}(z_{n-1}).
\end{gather*}
It is clearly seen that $\partial_{i,i+1}^{z}((1+z_{i+1} u_{i})^{-1})=
(1+z_{i} u_{i})^{-1} (1+z_{i+1} u_{i})^{-1} u_{i}$. Therefore,
\begin{gather*}
\partial_{i,i+1}^{z}({\cal{C}}({\mathfrak{P}}_n,U)) =
A_1(z_1) \cdots
A_{i}(z_i) A_{i+1}(z_{i+1}) (1+z_{i} u_{i})^{-1} u_{i} A_{i+2}(z_{i+2}) \cdots A_{n-1}(z_{n-1}).
\end{gather*}
It is easy to see that if one adds Hecke's type relations on the generators
\begin{gather*}
u_{i}^2= (a+b) u_i + a b, \qquad i=1,\ldots,n-1,
\end{gather*}
then
\begin{gather*
(1+ z u_{i})^{-1} u_{i} =\frac{u_{i} - z a b}{(1+b z)(1-a z)}.
\end{gather*}
Therefore in the quotient of the plactic algebra~${\mathfrak{P}}_n$ by the
Hecke type relations listed above and by the ``locality'' relations
\begin{gather*}
u_i u_j = u_j u_i, \qquad \text{if} \quad |i-j| \ge 2,
\end{gather*}
one obtains
\begin{gather*}
(-b+(1+z_{i} b)) \partial_{i,i+1}^z \left(A_{1}(z_1) \cdots A_{n-1}(z_{n-1}) \right) = \left(A_{1}(z_1) \cdots A_{n-1}(z_{n-1}) \right) \left(
\frac{e_{i} - b}{1-a z_{i}}\right).
\end{gather*}
Finally, if $a=0$, then the above identity takes the following form
\begin{gather*
\partial_{i,i+1}^z \left((1+z_{i+1} b) A_{1}(z_1) \cdots A_{n-1}(z_{n-1}) \right) = \left(A_{1}(z_1) \cdots A_{n-1}(z_{n-1}) \right) ( e_{i} - b).
\end{gather*}
In other words the above identity is equivalent to the statement~\cite{FK} that
in the idCoxeter algebra~${\cal{IC}}_n$ the Cauchy kernel
${\cal{C}}({\mathfrak{P}}_n,U)$ is the generating function for the
$b$-Grothendieck polynomials. Moreover, each (generalized) double
$b$-Grothendieck polynomial is a positive linear combination of the key-Grothendieck polynomials.
A proof of this statement is a corollary of the more
general statement which will be frequently used throughout the present paper,
namely, if an equivalence relation~${\approx}_2$ is a ref\/inement of that~${\approx}_1$, that is if assumption $a \, {\approx}_2 \, b \Longrightarrow a \, {\approx}_1 \, b$
holds $\forall\, a,\, b$, then each equivalence class w.r.t.\ relation~${\approx}_1$
is disjoint union of the equivalence classes w.r.t.\ relation~${\approx}_2$.
In the special case $b= -1$ and $P_{ij}= x_i+ y_j$
if $ 2 \le i+j \le n+1$, $p_{ij}=0$, if $i+j > n+1$, this result had been stated
in~\cite{L2}.
As a possible mean to def\/ine {\it affine versions} of polynomials treated in
the present paper, we introduce the {\it double affine nilCoxeter algebra of
type A} and give construction of a generic family of Hecke's type elements\footnote{Remind that by the name {\it a family of
Hecke's type elements} we mean a set of elements $ \{e_1,\dots,e_n \}$ such
that
$e_i^2 = A e_i +B$, $A$, $B$ are parameters (Hecke type relations),
$e_i e_j =e_j e_i$, if $|i-j| \ge 2$, $e_i e_j e_i=e_j e_i e_j$, if $|i-j|=1$ (Coxeter relations).}
we will be put to use in the present paper.
In Section~\ref{section5.1} we suggest a~common generalization of some combinatorial formulas from~\cite{CV} and~\cite{Gou}. Namely, we give explicit formula
\begin{gather}\label{equation1.3}
{\prod_{1 \le i \le k, \, 1 \le j \le n \atop j-i \le n-k } \frac{N-i-j+1}{i+j-1}} {\prod_{1 \le i \le k,\, 1 \le j \le n \atop j-i > n-k} \frac{N+i+j-1}{i+j-1}}
\end{gather}
for the number of $k$-tuples of noncrossing Dyck paths connecting the points
$(0,0)$ and $(N,N-n-k)$. Interpretations of the number~\eqref{equation1.3} as
the number of certain $k$-triangulations of a convex $(N+1)$-gon, or that of
certain alternating sign matrices of size $N \times N$, are interesting
tasks.
In the case $N=n+k$ we recover the \cite[r.h.s.\ of formula~(2)]{CV}. In the
case $k=2$ our formula~\eqref{equation1.3} is equivalent to that obtained in~\cite{Gou}. Our proof that the number~\eqref{equation1.3} counts certain $k$-tuples of noncrossing Dyck paths is
based on the study of combinatorial properties of the so-called column
multi-Schur functions $s_{\lambda}^{*} (X_n)$ introduced in Theorem~\ref{theorem5.6}, cf.~\cite{Ma, Wa}. In particular we show that for rectangular partition $\lambda =(n^k)$ the polynomial $s_{\lambda}^{*}(X_{n+k})$ is essentially coincide
with the Schubert polynomial corresponding to the Richardson permutation
$1^{k} \times w_{0}^{(n)}$. We introduce also a multivariable deformation of
the numbers $\operatorname{ASM}(n)$, namely,
\begin{gather*
\operatorname{ASM}(X_{n-1};t):= \sum_{\lambda \subset \delta_n} s_{\lambda}^{*}(X) t^{|\lambda|}.
\end{gather*}
Finally, in Section~\ref{section5.1} we give combinatorial interpretations of
rectangular and staircase components of the refined ${\rm TSSCP}$ vector~\cite{DZ1} in terms of $k$-fans of noncrossing Dyck paths in rectangular case and Gandhi--Dumont polynomials and Genocchi numbers in staircase case.
As Appendix we include several examples of polynomials studied in the present
paper to illustrate results obtained in these notes. We also include an
expository text concerning the {\it MacNeille completion} of a poset to draw
attention of the reader to this subject. It is an exami\-nation of the
MacNeille completion of the poset associated with the (strong) Bruhat order on the
symmetric group, that was one of the main streams of the study in the present
paper. Namely, our concern was the challenge how to attach to each edge $e$ of
the MacNeille completion~${\cal{MN}}({\mathbb{S}}_n)$ of the Bruhat order poset on the symmetric group ${\mathbb{S}}_n$ an operator $\partial_{e}$ acting on the ring of polynomials $\mathbb{Z}[X_n]$, such that $\partial_{e}(K_{h(e)})= K_{t(e)}$
together with compatibility conditions among the set of operators
$\{\partial_{e}\}_{e \in {\cal{MN}}({\mathbb{S}}_{n})}$, that is for any two vertices of~${\cal{MN}}({\mathbb{S}}_{n})$, say~$\alpha$ and~$\beta$, and a~path
$p_{\alpha,\beta}$ in the MacNeille completion which connects these vertices, the naturally def\/ined operator $\partial_{p_{\alpha,\beta}}$ depends only on the vertices~$\alpha$ and~$\beta$ taken, and doesn't depend on a path~$p_{\alpha,\beta}$ selected. As far as I know, this problem is still open.
\section{Plactic, nilplactic and idplactic algebras}\label{section2}
\begin{de}[\cite{LS1}]\label{definition2.1} The {\it plactic algebra ${\cal P}_n$} is an
(unital) associative algebra over $\mathbb{Z}$ generated by elements $\{u_1,
\dots, u_{n-1} \}$ subject to the set of relations
\begin{gather*}
\text{(PL1)} \quad u_j u_i u_k=u_j u_k u_i, \qquad u_i u_k u_j=u_k u_i u_j, \qquad \text{if} \quad i < j < k, \\
\text{(PL2)} \quad u_i u_j u_i=u_j u_i u_i, \qquad u_j u_i u_j=u_j u_j u_i, \qquad \text{if}\quad i < j.
\end{gather*}
\end{de}
\begin{pr}[\cite{LS1}]\label{prop2.2} Tableau words\footnote{For the reader convenience we recall a def\/inition of a
{\it tableau word}. Let $T$ be a (regular shape) semistandard Young
tableau. The tableau word $w(T)$ associated with~$T$ is the {\it reading
word} of~$T$ is the sequence of entries of~$T$ obtained by concatenating the
columns of~$T$ bottom to top consecutively starting from the f\/irst column.
For example, take
\begin{gather*}
T= \begin{matrix}1 & 2 &3 &3 \cr 2 & 3 & 4\cr 3 & 4 \cr 5
\end{matrix}
\end{gather*}
The corresponding {\it tableau word} is $w(T)= 5321432433$. By def\/inition, a
tableau word is the tableau word corresponding to some (regular shape)
semistandard Young tableau. It is well-known~\cite{LS2} that the number of
tableau subwords contained in the staircase word
$I_{0}^{(n)}:= \underbrace{u_{n-1}u_{n-2}\cdots u_{2}u_{1}}\underbrace{u_{n-1}u_{n-2}\cdots u_{2}}\cdots \underbrace{u_{n-1}u_{u-2}}\underbrace{u_{n-1}}$
is equal to the number of alternating
sign matrices ${\rm ASM}(n)$.\label{footnote9}}
in the alphabet $U=\{u_1,\dots,u_{n-1} \}$ form a~basis in the plactic algebra~${\cal P}_n$.
\end{pr}
In other words, each plactic class contain a unique tableau word.
In particular,
\begin{gather*}
\operatorname{Hilb}({\cal P}_{n+1},t)=(1-t)^{-n}\big(1-t^2\big)^{-{n \choose 2}}.
\end{gather*}
\begin{rem}\label{rem2.3} There exists another algebra over $\mathbb{Z}$ which
has the same Hilbert series as that of the plactic algebra ${\cal P}_{n}$.
Namely, def\/ine
algebra ${\cal L}_n$ to be an associative algebra over $\mathbb{Z}$ generated by the
elements $ \{e_1,e_2,\ldots,e_{n-1}\}$, subject to the set of relations
\begin{gather*}
(e_i,(e_j,e_k)):= e_i e_j e_k -e_j e_i e_k -e_j e_k e_i +e_k e_j e_i = 0,
\end{gather*}
for all\looseness=-1 $1 \le i,j, k \le n-1$, $j < k$. Observe that the number of def\/ining relations in the algebra~${\cal L}_n$
is equal to $2 {n \choose 3}$. Note that elements $e_1+e_2$ and
$e_2 e_1$ do not commute in the algebra~${\cal{L}}_3$, but do commute if are considered as elements in the plactic algebra~${\cal{P}}_3$. See Example~\ref{exam5.25}(C) for some details.
\end{rem}
\begin{exer}\label{exer2.4} \samepage\quad
\begin{itemize}\itemsep=0pt
\item Show that the dimension of the degree $k$ homogeneous
component ${{\cal L}}^{(k)}_n$ of the algebra~${\cal L}_n$ is equal to the
number semistandard Young tableaux of the size~$k$ f\/illed by the numbers from
the set $\{1,2,\ldots,n-1 \}$.
\item Let us set $e_{ij}:= (e_i,e_j):= e_i e_j-e_j e_i$, $i < j$.
Show that the elements $\{e_{ij} \}_{1 \le i < j \le n-1}$ generate
the {\it center} of the algebra~${\cal{L}}_n$.
\end{itemize}
\end{exer}
\begin{de}\label{def2.5} \quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] The {\it local plactic algebra} ${\cal LP}_n$, see, e.g.,~\cite{FG},
is an associative algebra over $\mathbb{Z}$ generated by elements $\{u_1,\ldots,u_{n-1} \}$ subject to the set of relations
\begin{gather*}
u_i u_j=u_j u_i, \qquad \text{if}\quad |i-j| \ge 2, \\ u_{j} u_i^2=u_i u_{j} u_i, \qquad u_{j}^2 u_i
=u_j u_i u_j, \qquad \text{if}\quad |i-j|=1.
\end{gather*}
One can show (A.K.) that
\begin{gather*}
\operatorname{Hilb}({\cal LP}_n,t)= \prod_{j=1}^{n} \left(
{1 \over 1-t^j} \right)^{n+1-j}.
\end{gather*}
\item[$(b)$] The {\it affine local plactic algebra} $\widehat{{\cal{PL}}}_n$, see~\cite{KoS}, is an associative algebra over~$\mathbb{Q}$ generated by the elements
$\{e_{0},\ldots,e_{n-1} \}$ subject to the set of relations listed in item $(a)$, where all indices are understood modulo~$n$.
\item[$(c)$] Let $q$ be a parameter, the {\it affine quantum local plactic algebra} $\widehat{{\cal{PL}}}_n^{(q)}$ is an associative algebra over $\mathbb{Q}[q]$ generated by the elements $\{e_{0},\ldots,e_{n-1} \}$ subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(i)$] $e_i e_j = e_j e_i$, if $|i-j| > 1$,
\item[$(ii)$] $(1+q)e_i e_j e_i = q e_i^2 e_j + e_j e_i^2$, $(1+q)e_j e_i e_j = q e_i e_j^2 + e_j^2 e_i$, if $|i-j|=1$,
where all indices appearing in the relation~$(ii)$ are understood modulo~$n$.
\end{enumerate}
\end{enumerate}
\end{de}
It was observed in~\cite{Kor} that the relations listed in~$(ii)$ are the Serre
relations of $U_{q}^{\ge}\big(\hat{\mathfrak{gl}}(n)\big)$ rewritten\footnote{For the reader convenient, we recall, see, e.g., \cite{Kor} the
def\/inition of algebra $U_{q}^{\ge}(\hat{\mathfrak{gl}}(n))$. Namely, this
algebra is an unital associative algebra over~$\mathbb{Q}(q^{\pm 1})$ generated by
the elements
$\{k_{i}^{\pm 1},E_i \}_{i=0,\ldots,n-1}$, subject to the set of relations
\begin{enumerate}\itemsep=0pt
\item[$(R1)$] $k_i k_j = k_j k_i$, $k_i k_i^{-1}= k_i^{-1}=1$, $\forall\, i,\, j$,
\item[$(R2)$] $k_i E_j = q^{\delta_{i,j}-\delta_{i,j+1}} E_j k_i$,
\item[$(R3)$] (Serre relations)
\begin{gather*}
E_i^2 E_{i+1} -\big(q+q^{-1}\big) E_i E_{i+1} E_{i}+ E_{i+1} E_i^2 = 0,\qquad
E_{i+1}^2 E_{i} -\big(q+q^{-1}\big) E_{i+1} E_{i} E_{i+1}+ E_{i} E_{i+1}^2 = 0,
\end{gather*}
where all indices are understood modulo $n$, e.g., $e_{n}=e_{0}$.
\end{enumerate}
It is clearly seen that after rewriting the Serre relations $(R3)$ in terms of
the elements $\{e_i:=k_i E_i\}_{0 \le i \le n-1}$, one comes to the relations~$(ii)$ which have been listed in~$(c)$.}
in generators~$k_i E_i$. Some interesting properties and applications of the
local and af\/f\/ine local plactic algebras one can f\/ind in~\cite{Kor,KoS,KT}. It seems an interesting problem to investigate properties and
applications of the af\/f\/ine {\it pseudoplactic} algebra which is the quotient
of the af\/f\/ine local plactic algebra ${\widehat{\cal{LP}}}_n$ by the two-sided ideal generated by the set of elements
\begin{gather*}
\{(e_j,(e_i,e_k)), \, 0 \le i < j < k \le n-1 \}.
\end{gather*}
\begin{de}[nil Temperley--Lieb algebra]\label{def2.6}
Denote by ${\cal TL}_n^{(0)}$ the quotient of the local plactic algebra
${\cal LP}_n$ by the two-sided ideal generated by the elements
$\big\{u_1^2,\ldots,u_{n-1}^2 \big\}$.
\end{de}
It is well-known that $\dim {\cal{TL}}_n = C_n$,
the $n$-th Catalan number. One also has
\begin{gather*}
\operatorname{Hilb}\big({\cal{TL}}_{4}^{(0)},t\big)= (1,3,5,4,1), \qquad \operatorname{Hilb}\big({\cal{TL}}_{5}^{(0)},t\big)=(1,4,9,12,10,4,2), \\ \operatorname{Hilb}({\cal{TL}}_{6},t)=(1,5,14,25,31,26,16,9,4,1).
\end{gather*}
\begin{pr}\label{prop2.7}
The Hilbert polynomial $\operatorname{Hilb}({\cal{TL}}_{n},t)$ is equal to the generating
function for the number or $321$-avoiding permutations of the set
$\{1,2,\ldots,n\}$
having the inversion number equals to $k$, see~{\rm \cite[$A140717$]{SL}}, for other
combinatorial interpretations of the polynomials $\operatorname{Hilb}({\cal{TL}}_{n},t)$.
\end{pr}
\begin{exer}\label{exer2.8} \quad
\begin{itemize}\itemsep=0pt
\item Show that $\deg_{t} \operatorname{Hilb}({\cal{TL}}_{n},t) = \big\lbrack {\frac{n^2}{4}} \big\rbrack$.
\item Show that
\begin{gather*}
\lim_{k \rightarrow \infty} t^{k^{2}} \operatorname{Hilb}\big({\cal{TL}}_{2k},t^{-1}\big) =
\prod_{n \ge 1} {\frac{\big(1-t^{4 n-2}\big)}{\big(1-t^{2 n -1}\big)^{4} \big(1-t^{4 n}\big)}} = \sum_{n \ge 0} a_{n} t^{n},
\end{gather*}
where $a_{n}$ is equal to the number of $2$-colored generalized Frobenius
partitions, see, e.g., \cite[$A051136$]{SL} and the literature quoted therein\footnote{The second equality in the above formula is due to G.~Andrews~\cite{An}. The second formula for the generating function of the numbers $\{a_n\}_{n \ge 0}$ displayed in~\cite[$A051136$]{SL}, either contains misprints or counts something else.}.
\item Show that
\begin{gather*}
\lim_{k \rightarrow \infty} t^{k(k+1)} \operatorname{Hilb}\big({\cal{TL}}_{2k+1},t^{-1}\big) =
2 \prod_{n \ge 1} {\frac{(1+t^{n}) \big(1+t^{2n}\big)^2}{1-t^n}} = 2 \sum_{n \ge 0}
b_{n} t^{n}.
\end{gather*}
See \cite[$A201078$]{SL} for more details concerning relations of this exercise
with the Ramanujan theta functions\footnote{See, e.g., \url{http://en.wikipedia.org/wiki/Ramanujan_theta_function}.}.
\end{itemize}
\end{exer}
We denote by ${\cal{TH}}_{n}^{(\beta)}$ the quotient of the local plactic algebra ${\cal{LP}}_n$ by the two-sided ideal generated by the elements
$\big\{u_i^2- \beta u_i,\, i=1,\ldots,n-1 \big\}$.
\begin{de} \label{def2.9} The modif\/ied \textit{plactic algebra} ${\cal{MP}}_n$ is an
associative algebra over $\mathbb{Z}$ generated by $\{u_1,\ldots,u_{n-1} \}$ subject
to the set of relations (PL1) and that
\begin{gather*}
u_j u_j u_i = u_j u_i u_i, \qquad \text{and} \qquad u_i u_j u_i=u_j u_i u_j, \qquad \text{if} \quad 1 \le i < j \le n-1.
\end{gather*}
\end{de}
\begin{de}\label{def2.10} The \textit{nilplactic algebra} ${\cal{NP}}_n$ is an associative algebra over $\mathbb{Q}$ generated by $\{u_1, \ldots,u_{n-1} \}$,
subject to set of relations
\begin{gather*
u_i^2=0, \qquad u_{i} u_{i+1} u_{i}= u_{i+1} u_{i} u_{i+1},
\end{gather*}
the set of relations (PL1), and that $u_i u_j u_i= u_j u_i u_j$, if $|i-j| \ge 2$.
\end{de}
\begin{pr}[\cite{LS3}]\label{prop2.11}
Each nilplactic class not containing zero, contains one and only one tableau word.
\end{pr}
\begin{pr}\label{prop2.12} The nilplactic algebra ${\cal{NP}}_n$ has finite dimension, its
Hilbert polynomial $\operatorname{Hilb}({\cal{NP}}_n,t)$ has degree ${n \choose 2}$, and
$\dim_{\mathbb{Q}}({\cal{NP}}_n)_{{n \choose 2}} =1$.
\end{pr}
\begin{ex}\label{exer2.13}
\begin{gather*}
\operatorname{Hilb}({\cal{NP}}_3,t)= (1,2,2,1), \qquad \operatorname{Hilb}({\cal{NP}}_4,t)=(1,3,6,6,5,3,1), \\
\operatorname{Hilb}({\cal{NP}}_5,t)=(1,4,12,19,26,26,22,15,9,4,1), \qquad \dim_{\mathbb{Q}}({\cal{NP}}_5)=139,\\
\operatorname{Hilb}({\cal{NP}}_6,t)=(1,5,20,44,84,119,147,152,140,224,81,52,29,14,5,1),\\
\dim_{\mathbb{Q}}({\cal{NP}}_6) = 1008.
\end{gather*}
\end{ex}
\begin{de}\label{def2.14} The \textit{idplactic algebra} ${\cal{IP}}_{n}:= {\cal{IP}}_{n}^{(\beta)}$ is an
associative algebra over $\mathbb{Q}[\beta]$ ge\-ne\-rated by elements $\{u_1,\ldots,u_{n-1} \}$ subject to the set of relations
\begin{gather}\label{equation2.2}
u_i^2= \beta u_i, \qquad u_{i} u_{j} u_{i}= u_{j} u_{i} u_{j}, \qquad i < j,
\end{gather}
and the set of relations~(PL1).
\end{de}
In other words,the idplactic algebra ${\cal{IP}}_n^{(\beta)}$ is the quotient
of the modif\/ied plactic algebra~${\cal{MP}}_n$ by the two-sided ideal generated by the elements $\big\{u_i^2- \beta u_i, \, 1 \le i \le n\big\}$.
\begin{pr}\label{prop2.15} Each idplactic class not containing zero, contains a unique tableau word associated with a row strict semistandard Young tableau\footnote{Recall that a row strict semistandard Young tableau, say~$T$, is a tableau
such that the numbers in each row and each column of $T$ are strictly
increasing. For example,
\begin{gather*}
\begin{matrix}
1 & 2 & 3 & 5 & 6\\
2 & 3 & 6 & 8\\
4 & 5 & 7
\end{matrix}
\end{gather*}}.
\end{pr}
For each word $w$ denote by $\operatorname{rl}(w)$ the length of a unique tableau word of
minimal length which is idplactic equivalent to $w$.
\begin{ex}\label{exam2.16} Consider words in the alphabet $\{a < b < c < d \}$. Then
\begin{gather*}
\operatorname{rl}(dbadc)=4=\operatorname{rl}(cadbd), \qquad \operatorname{rl}(dbadbc)=5=\operatorname{rl}(cbadbd).
\end{gather*}
Indeed,
\begin{gather*}
dbadc \sim dbdac \sim dbdca \sim ddbca \sim dbac,\\
dbadbc \sim dbabdc \sim dabadc \sim adbdac \sim abdbca \sim abbdca
\sim dbabc.
\end{gather*}
\end{ex}
Note that according to our def\/inition, tableau words $w=31$, $w=13$ and
$w=313$ belong to dif\/ferent idplactic classes.
\begin{pr}\label{2.17} The idplactic algebra ${\cal {IP}}_n^{(\beta)}$ has finite
dimension, and its Hilbert polynomial has degree~${n \choose 2}$.
\end{pr}
\begin{ex}\label{exam2.18}
\begin{gather*}
\operatorname{Hilb}({\cal {IP}}_3,t)=(1,2,2,1), \qquad \operatorname{Hilb}({\cal {IP}}_4,t)=
(1,3,6,7,5,3,1), \qquad \dim({\cal {IP}}_4)=26, \\
\operatorname{Hilb}({\cal {IP}}_5,t)=(1,4,12,22,30,32,24,15,9,4,1), \qquad \dim({\cal {IP}}_5) = 154, \\
\operatorname{Hilb}({\cal {IP}}_6,t)=(1,5,20,50,100,156,188,193,173,126,84,52,29,14,5,1), \\
\dim({\cal {IP}}_6)=1197, \qquad
\dim({\cal {IP}}_7)= 9401.
\end{gather*}
\end{ex}
Note that for a given $n$ some words corresponding to strict semistandard
Young tableaux of size between~5 and ${n-1 \choose 2}$ are idplactic equivalent to zero. For example,
\begin{gather*}
\begin{matrix}
1 & 2 & 4\\
3 & 4 \end{matrix} \sim 31 42 4 \sim 31242 \sim 13242 \sim 13422 \sim 0,
\\
\begin{matrix}
1 & 2 & 3 \\
2 & 4 \\
4
\end{matrix} \sim 421 42 3 \sim 421243 \sim 412143 \sim 142413 \sim 124213 \sim
122431 \sim 0.
\end{gather*}
It seems an interesting {\it problem} to count the number of all row strict
semistandard Young tableaux contained in the staircase $(n,n-1, \ldots,2,1)$
and bounded by~$n$, as well as count the number the number of such tableaux
which are idplactic equivalent to zero. For $n=2$ these numbers are $(6,6)$, for
$n=3$ these numbers are $(26,26)$, for $n=4$ these numbers are $(160,154)$ and
for $n=5$ they are $(1427,1197)$.
\begin{de}\label{def2.19} The \textit{idplactic Temperly--Lieb algebra} ${\cal PTL}_n^{(\beta)}$
is def\/ine to be the quotient of the idplactic algebra ${\cal IP}_n^{(\beta)}$
by the two-sided ideal generated by the elements
\begin{gather*}
\{u_i u_j u_i, \, \forall \, i \not= j \}.
\end{gather*}
\end{de}
For example,
\begin{gather*}
\operatorname{Hilb}\big({\cal{PTL}}_4^{(0)},t\big) =(1,3,6,4,1)_{t},\qquad
\operatorname{Hilb}\big({\cal{PTL}}_5^{(0)},t\big) = (1,4,12,16,14,4,2)_t,\\
\operatorname{Hilb}\big({\cal{PTL}}_6^{(0)},t\big) = (1,5,20,40,60,46,32,10,4,1)_{t},\\
\operatorname{Hilb}\big({\cal{PTL}}_7^{(0)},t\big)=
(1,6,30,80,170,216,238,152,96,44,14,4,2)_{t}.
\end{gather*}
One can show that
\begin{gather*}
\deg_{t}
\operatorname{Hilb}\big({\cal{PTL}}_n^{(0)},t\big) = \left[\frac{n^2}{4} \right],
\end{gather*}
and
\begin{gather*}
{\rm Coef\/f}_{t^{\max}}
\operatorname{Hilb}({\cal{PTL}}_n,t) =\begin{cases} 1, & \text{if $n$ is even},\\
2, & \text{if $n$ is odd}.
\end{cases}
\end{gather*}
\begin{de}\label{def2.20} The \textit{nilCoxeter algebra} ${\cal {NC}}_n$ is def\/ined to be the
quotient of the nilplactic algebra ${\cal {NP}}_n$ by the two-sided ideal
generated by elements $\{u_i u_j-u_j u_i,\, |i-j| \ge 2 \}$.
\end{de}
Clearly the nilCoxeter algebra ${\cal NC}_n$ is a quotient of the modif\/ied
plactic algebra ${\cal MP}_n$ by the two-sided ideal generated by the elements
$\{u_i u_j -u_j u_i, \, |i-j| \ge 2 \}$.
\begin{de}\label{def2.21} The {\it idCoxeter algebra} ${\cal {IC}}_n^{(\beta)}$ is
def\/ined to be the quotient of the idplactic algebra
${\cal {IP}}_n^{(\beta)}$ by the two-sided ideal generated by the
elements $\{u_i u_j-u_j u_i,\, |i-j| \ge 2 \}$.
\end{de}
It is well-known, see, e.g., \cite{AL}, that the algebra
${\cal NC}_n$ and ${\cal IC}_n^{(\beta)}$
has dimension $n !$, and the elements $\{u_w := u_{i_{1}} \cdots u_{i_{\ell}} \}$, where
$w =s_{i_{1}} \cdots s_{i_{\ell}}$ is any reduced decomposition of
$w \in {\mathbb S}_n$, form a {\it basis} in the nilCoxeter and idCoxeter
algebras ${\cal NC}_n$ and ${\cal IC}_n^{(\beta)}$.
\begin{rem}\label{2.22} There is a common generalization of the algebras def\/ined
above which is due to S.~Fomin and C.~Greene~\cite{FG}.
Namely, def\/ine generalized plactic algebra
${\widetilde {\cal P}}_n$ to be an associative algebra generated by elements
$u_1,\dots,u_{n-1}$, subject to the relations (PL2) and relations
\begin{gather}\label{equation2.3}
u_ju_i(u_i+u_j)=(u_i+u_j)u_ju_i,\qquad i < j.
\end{gather}
The relation \eqref{equation2.3} can be written also in the form
\begin{gather*}
u_j(u_iu_j-u_ju_i)=(u_iu_j-u_ju_i)u_i, \qquad i < j.
\end{gather*}
\end{rem}
\begin{Theorem}[\cite{FG}]\label{theorem2.23} For each pair of numbers
$1 \le i < j \le n$ define
\begin{gather*}
A_{i,j}(x)= \prod_{k=j}^{i}(1+x u_k).
\end{gather*}
Then the elements $A_{i,j}(x)$ and $A_{i,j}(y)$ commute in the generalized
plactic algebra ${\widetilde {\cal P}}_n$.
\end{Theorem}
\begin{cor} \label{cor2.24} Let $1 \le i < j \le n$ be a pair of numbers.
Noncommutative elementary polynomials
$e_a^{ij}:= \sum\limits_{j \ge i_1 \ge \cdots \ge i_k \ge i}u_{i_1} \cdots u_{i_a}$,
$i \le a \le j$, generate a commutative subalgebra ${\cal C}_{i,j}$ of
rank $j-i+1$ in the plactic algebra ${\cal P}_n$.
Moreover, the algebra ${\cal C}_{1,n}$ is a maximal commutative subalgebra of
${\cal P}_n$.
\end{cor}
To establish Theorem~\ref{theorem2.23}, we are going to prove more general result. To start
with, let us def\/ine {\it generic plactic algebra} ${\mathfrak{P}}_n$.
\begin{de}\label{def2.25} The {\it generic plactic algebra ${\mathfrak{P}}_n$} is an
associative algebra over $\mathbb{Z}$ generated by
$\{e_1,\dots, e_{n-1} \}$ subject to the set of relations
\begin{alignat}{3}\label{equation2.4}
& e_j (e_i,e_j) = (e_i,e_j) e_i, \qquad && \text{if}\quad i < j,&
\\
\label{equation2.5}
&(e_j,(e_i,e_k)) = 0, \qquad && \text{if} \quad i < j < k,&
\\
\label{equation2.6}
& (e_j,e_k)(e_i,e_k)=0,\qquad && \text{if}\quad i < j < k.&
\end{alignat}
Hereinafter we shell use the notation $(a,b):= [a,b]:=a b-b a$.
\end{de}
{\sloppy Clearly seen that relations \eqref{equation2.4}--\eqref{equation2.6} are consequence of the plactic
relations~(PL1) and~(PL2), but not vice versa.
}
\begin{Theorem}\label{theorem2.26} Define
\begin{gather*}
A_{n}(x)= \prod_{k=j}^{1}(1+x e_k).
\end{gather*}
Then the elements $A_{n}(x)$ and $A_{n}(y)$ commute in the generic
plactic algebra ${\mathfrak{P}}_n$.
Moreover the elements $A_n(x)$ and $A_n(y)$ commute if and only if the
generators $\{e_1,\ldots,e_{n-1} \}$ satisfy the relations
\eqref{equation2.4}--\eqref{equation2.6}.
\end{Theorem}
\begin{proof} For $n=2,3$ the statement of Theorem \ref{theorem2.26} is obvious. Now assume
that the statement of Theorem \ref{theorem2.26} is true in the algebra $\mathfrak{P}_n$. We
have to prove that the commutator $[A_{n+1}(x),A_{n+1}(y)]$ is equal to zero.
First of all, $A_{n+1}(x)=(1+ x e_n) A_{n}(x)$. Therefore
\begin{gather*}
[A_{n+1}(x),A_{n+1}(y)] = (1+x e_n) [A_{n}(x),1+y e_n] A_n(y) -
[A_n(y), 1+x e_n] A_n(x).
\end{gather*}
Using the standard identity $[a b,c] =a [b,c]+[a,c] b$, one f\/inds that
\begin{gather*}
\frac{1}{x y} [A_{n}(x), 1+y e_n] = \sum_{i=}^{n-1} \prod_{a=n-1}^{i+1}
(1+x e_a) (e_i,e_n) \prod_{a=i-1}^{1} (1+x e_a).
\end{gather*}
Using relations \eqref{equation2.4} we can move the commutator $(e_i,e_n)$ to
the left, since $i < a< n$, till we meet the term $(1+ x e_n)$. Using
relations \eqref{equation2.5} we see that $(1+x e_n) (e_i,n)= (e_i,n)(1+ x e_i)$. Therefore we come to the
following relation
\begin{gather*}
\frac{1}{x y} [A_{n}(x), 1+y e_n] \\
\qquad{} =\sum_{i=n-1}^{1} (e_i,e_n) \left(
(1+ x e_i) \prod_{a=n-1 \atop a \not= i}^{1} (1+x e_a) A_n(y) -
(1+ y e_i) \prod_{a=n-1 \atop a \not= i}^{1} (1+y e_a) A_n(x) \right).
\end{gather*}
Finally let us observe that according to the relation \eqref{equation2.6},
\begin{gather*}
(e_i,e_n)\bigl( (1+x e_i)(1+x e_{n-1}) -(1+ x e_{n-1})(1+x e_i)\bigr) =
x^2 (e_i,e_n) (e_i, e_{n-1}) = 0.
\end{gather*}
Indeed,
\begin{gather*}
(e_i,e_n)(e_i,e_{n-1})= (e_i,e_{n}) e_i e_{n-1} - (e_i,e_{n}) e_{n-1}
e_{i} = e_n e_{n-1} (e_i,e_n) - e_{n-1} e_{n} (e_i,e_n) =0.
\end{gather*} Therefore
\begin{gather*}
\frac{1}{x y} [A_{n}(x), 1+y e_n] = \left(\sum_{i={n-1}}^{1} (e_i,e_n)
\right) [A_n(x),A_n(y)] =0
\end{gather*}
according to the induction assumption.
Finally, if $i < j$, then $(e_i+e_j,e_j e_i) =0 \Longleftrightarrow \eqref{equation2.4}$;
if $i < j < k$, and the relations~\ref{equation2.4} hold, then $(e_i+e_j+e_k,
e_j e_i+e_k e_j+e_k e_i)=0 \Longleftrightarrow \eqref{equation2.5}$;
if $i < j < k$, and relations~\eqref{equation2.4} and~\eqref{equation2.5} hold, then $(e_i+e_j+e_k,
e_k e_j e_i) =0 \Longleftrightarrow \eqref{equation2.6}$;
the relations $(e_j e_i+e_k e_j+e_k e_i,e_k e_j e_i)=0$ are consequences of
the above ones.
\end{proof}
Let $T$ be a semistandard tableau and $w(T)$ be the column reading word
corresponding to the tableau~$T$. Denote by $R(T)$ (resp. $IR(T)$) the set of
words which are plactic (resp.\ idplactic) equivalent to $w(T)$.
Let ${\bf a}=(a_1,\dots,a_n) \in R(T)$, where $n:=|T|$
(resp. ${\bf a}=(a_1,\dots,a_m) \in IR(T)$, where $m \ge |T|$).
\begin{de}[compatible sequences ${\bf b}$]\label{def2.27} Given a word ${\bf a} \in R(T)$
(resp.~${\bf a} \in IR(T)$), denote by
$C(\bf a)$ (resp.~$IC(\bf a)$) the set of sequences of positive integers,
called compatible sequences,
${\bf b}:=(b_1 \le b_2 \le \cdots \le b_m)$ such that
\begin{gather*
b_i \le a_i, \qquad \text{and if} \quad a_i \le a_{i+1}, \qquad \text{then}\quad b_i < b_{i+1}.
\end{gather*}
\end{de}
Finally, def\/ine the set $C(T)$ (resp.~$IC(T)$) to be the union
$\bigcup C({\bf a})$ (resp. the union $\bigcup IC({\bf a})$),
where ${\bf a}$ runs over all words which are plactic (resp. idplactic)
equivalent to the word~$w(T)$.
\begin{ex}\label{exam2.28} Take $ T= \begin{matrix}2 &3 \cr 3 & \cr
\end{matrix}$.
The corresponding tableau
word is $w(T)=323$. We have $R(T)=\{232, 323 \}$ and $IR(T)= R(T) \bigcup
\{2323, 3223, 3232, 3233, 3323, 32323,\dots \}$. Moreover,
\begin{gather*}
C(T)= \left \{\begin{matrix} {\bf a}\colon & 232 & 323 & 323 & 323 & 323 \cr
{\bf b}\colon & 122 & 112 & 113 & 123 & 223
\end{matrix} \right \}, \\
IC(T)= C(T) \bigcup \left
\{\begin{matrix} {\bf a}\colon & 2323 & 3223 & 3232 & 3233 & 3323 & 32323 \cr
{\bf b}\colon & 1223 & 1123 & 1122 & 1123 & 1223 & 11223
\end{matrix}\right\}.
\end{gather*}
\end{ex}
Let $ {\mathfrak P}:= {\mathfrak{P}}_n := \{p_{i,j}, \, i \ge 1,\,j \ge 1,\, 2 \le i+j \le n+1 \}$
be the set of (mutually commuting) variables.
\begin{de}\label{def2.29}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Let $T$ be a semistandard tableau, and
$n:=|T|$. Def\/ine the double key polynomial ${\cal K}_{T}({\mathfrak P})$
corresponding to the tableau~$T$ to be
\begin{gather*
{\cal K}_{T}(\mathfrak{P})= \sum_{{\bf b} \in C(T)} \prod_{i=1}^{n}
p_{b_{i},a_i-b_i+1}.
\end{gather*}
\item[(2)] Let $T$ be a semistandard tableau, and $n:=|T|$. Def\/ine the
double key Grothendieck polynomial ${\cal {GK}}_{T}(\mathfrak{P})$
corresponding to the tableau $T$ to be
\begin{gather*
{\cal {GK}}_{T}({\mathfrak P})= \sum_{{\bf b} \in IC(T)} \prod_{i=1}^{m}
p_{b_{i},a_i-b_i+1}.
\end{gather*}
\end{enumerate}
\end{de}
In the case when $p_{i,j}= x_i+y_j$, $\forall\, i,j$, where $X= \{x_1,\ldots,x_n \}$ and $Y = \{y_1, \ldots,y_n \}$ denote two sets of variables, we will
write ${\cal K }_{T}(X,Y)$, ${\cal GK}_{T}(X,Y),\dots$, instead of
${\cal K}_{T}({\mathfrak P})$,
${\cal GK}_{T}({\mathfrak P}), \ldots $.
\begin{de}\label{def2.30} Let $T$ be a semistandard tableau, denote by $\alpha(T)=
(\alpha_1,\dots,\alpha_{n})$ the exponent of the {\it smallest} monomial
in the set $\big\{x^{\bf b}:= \prod\limits_{i=1}^{m} x_{i}^{b_{i}}, \,{\bf b} \in C(T) \big\}$
with respect to the lexicographic order.
\end{de}
We will call the composition $\alpha(T)$ to be the bottom code of tableau $T$.
\section{Divided dif\/ference operators}\label{section3}
In this subsection we remind some basic properties of divided dif\/ference
operators will be put to use in subsequent sections. For more details,
see \cite{Ma1}.
Let $f$ be a function of the variables $x$ and $y$ (and possibly other
variables), and $\eta \not= 0$ be a~parameter. Def\/ine the {\it divided
difference operator $\partial_{xy}(\eta)$} as follows
\begin{gather*
\partial_{xy}(\eta) f(x,y)={f(x,y)-f(\eta^{-1} y,\eta x)
\over x-\eta^{-1} y}.
\end{gather*}
Equivalently, $(x-\eta^{-1} y) \partial_{xy}(\eta)= 1-s_{xy}^{\eta}$,
where the operator $s_{xy}^{\eta}$ acts on the variables $(x,y, \ldots)$
according to the rule: $s_{xy}^{\eta}$ transforms the pair $(x,y)$ to
$(\eta^{-1} y,\eta x)$, and f\/ixes all other variables. We set by def\/inition,
$s_{yx}^{\eta}:=s_{xy}^{{\eta^{-1}}}$.
The operator $\partial_{xy}(\eta)$ takes polynomials to polynomials and has
degree~$-1$. The case $\eta =1$ corresponds to the {\it Newton divided
difference operator} $\partial_{xy}:=\partial_{xy}(1)$.
\begin{lem}\label{lem3.1} \quad
\begin{enumerate}\itemsep=0pt
\item[$(0)$] $s_{xy}^{\eta} s_{xz}^{\eta \xi}=s_{yz}^{\xi} s_{xy}^{\eta}$,
$s_{xy}^{\eta} s_{xz}^{\eta \xi} s_{yz}^{\xi}=
s_{yz}^{\xi} s_{xz}^{\eta \xi} s_{xy}^{\eta}$,
\item[$(1)$] $\partial_{yx}(\eta)= -\eta \partial_{xy}(\eta^{-1})$,
$s_{xy}^{\eta} \partial_{yz}(\xi)=
\eta^{-1} \partial_{xz}(\eta \xi) s_{xy}^{\eta}$,
\item[$(2)$] $\partial_{xy}(\eta)^2= 0$,
\item[$(3)$] $($three term relation$)$
$\partial_{xy}(\eta) \partial_{yz}(\xi)=
\eta^{-1} \partial_{xz}(\eta \xi) \partial_{xy}(\eta)+
\partial_{yz}(\xi) \partial_{xz}(\eta \xi)$.
\item[$(4)$] $($twisted Leibniz rule$)$
$ \partial_{xy}(\eta) (fg)=
\partial_{xy}(\eta) (f) g+s_{xy}^{\eta}(f) \partial_{xy}(\eta) (g)$,
\item[$(5)$] $($crossing relations, cf. {\rm \cite[formula~(4.6)]{FK})}
\begin{itemize}\itemsep=0pt
\item $x \partial_{xy}(\eta)=\eta^{-1} \partial_{xy}(\eta) y+1,
y \partial_{xy}(\eta)= \eta \partial_{xy}(\eta) x-\eta$,
\item $\partial_{xy}(\eta) y \partial_{yz}(\xi)=
\partial_{xz}(\eta \xi) x \partial_{xy}(\eta)+
\xi^{-1}\partial_{yz}(\xi) z \partial_{xz}(\eta \xi)$,
\end{itemize}
\item[$(6)$] $\partial_{xy} \partial_{xz}
\partial_{yz} \partial_{xz}=0$.
\end{enumerate}
\end{lem}
Let $x_1,\ldots,x_n$ be independent variables, and let
$P_n:= \mathbb{Q}[x_1,\ldots,x_n]$. For each $i < j $ put
$\partial_{ij}:=\partial_{x_i x_j}(1)$ and $\partial_{ji}=-\partial_{ij}$.
From Lemma~\ref{lem3.1} we have
\begin{gather*}
\partial_{ij}^2=0, \qquad
\partial_{ij} \partial_{jk}+\partial_{ki} \partial_{ij}+\partial_{jk}
\partial_{ki}=0, \\
\partial_{ij} x_j \partial_{jk}+\partial_{ki} x_i \partial_{ij}+
\partial_{jk} x_k \partial_{ki},
\qquad \text{if} \quad \text{$i$, $j$, $k$ are distinct}.
\end{gather*}
It is interesting to consider also an {\it additive or affine} analog
$\partial_{xy}[k]$
of the divided dif\/ference operators $\partial_{xy}(\eta)$, namely,
\begin{gather*
\partial_{xy}[k](f(x,y))={f(x,y)-f(y-k,x+k) \over x-y+k}.
\end{gather*}
One has $\partial_{yx}[k] = -\partial_{xy}[-k]$, and
$\partial_{xy}[p] \partial_{yz}[q]=
\partial_{xz}[p+q] \partial_{xy}[p]+
\partial_{yz}[q] \partial_{xz}[p+q]$.
{\it A short historical comments in order.}
As far as I know, the divided dif\/ference operator had been invented
by I.~Newton in/around the year~$1687$, see, e.g., \cite[Ref.~\protect{[142]}]{AL}. Since that time the literature concerning divided dif\/ferences, a plethora of its
generalizations and applications in dif\/ferent f\/ields of mathematics and
physics, grows exponentially and essentially is immense. To the best of our
knowledge, the f\/irst systematic use of the isobaric divided dif\/ference operators, namely $ \pi_i(f)= \partial_i(x_{i} f)$, and $\overline{\pi}_{i}:=\pi_{i}-1$, goes back to papers by M.~Demazure concerning the study of desingularization of Schubert varieties and computation of characters of certain modules which
is nowadays called {\it Demazure modules}, see, e.g.,~\cite{De} and the literature quoted therein;
the f\/irst systematic use and applications of divided dif\/ference operators to
the study of cohomology rings of partial f\/lag varieties $G/P$ and
description of Schubert classes inside the former, goes back to a paper by
I.N.~Berstein, I.M.~Gelfand and S.I.~Gelfand~\cite{BGG};
it is A.~Lascoux and M.-P.~Sch\"{u}tzenberger who had discovered
{\it a polynomial representative} of a Schubert class in the cohomology ring
of the type~$A$ complete f\/lag varieties, and developed a rich and beautiful
combinatorial theory of these polynomials, see~\cite{LS9,LS7,LSZ,LS4,LS3} for acquainting the reader with basic ideas and
results concerning the Lascoux--Sch\"{u}tzenber Schubert polynomials.
\section{Schubert, Grothendieck and key polynomials}\label{section4}
Let $w \in {\mathbb S}_n$ be a permutation, $X=(x_1,\dots,x_n)$ and
$Y=(y_1,\dots,y_n)$ be two sets of variables. Denote by
$w_0 \in {\mathbb S}_n$ the longest
permutation, and by $\delta_n= (n-1,n-2,\dots,1)$ the staircase
partition. For each partition $\lambda$ def\/ine
\begin{gather}\label{equation4.1}
R_{\lambda}(X,Y):=
\prod_{(i,j) \in \lambda} (x_i+y_j).
\end{gather}
For $i=1,\dots,n-1$, let $s_i=(i,i+1) \in {\mathbb S}_n$ denote the simple
transposition that interchanges $i$ and $i+1$ and f\/ixes all other elements of
the set $\{1,\dots,n \}$. If $\alpha=
(\alpha_1,\dots,\alpha_i,\alpha_{i+1},\dots,\alpha_n)$ is a~composition, we
will write
\begin{gather*
s_i\alpha = (\alpha_1,\dots,\alpha_{i+1},\alpha_{i},\dots,\alpha_n).
\end{gather*}
\begin{de}[cf.~\cite{AL} and the literature quoted therein]\label{def4.1}\quad
\begin{itemize}\itemsep=0pt
\item For each permutation $w \in {\mathbb S}_n$ {\it the double
Schubert polynomial} ${{\mathfrak S}}_w(X,Y)$ is def\/ined to be
\begin{gather*
\partial_{w^{-1} w_{0}}^{(x)}(R_{\delta_{n}}(X,Y)).
\end{gather*}
\item
Let $\alpha$ be a composition.
\textit{The key polynomials} $K[\alpha](X)$ are def\/ined recursively
as follows:
if $\alpha$ is a partition, then $K[\alpha](X)= x^{\alpha}$;
otherwise, if $\alpha$ and $i$ are such that $\alpha_i < \alpha_{i+1}$, then
\begin{gather*
K[s_i(\alpha)](X)= \partial_{i} \big( x_i K[\alpha](X) \big).
\end{gather*}
\item \textit{The reduced key polynomials}
${\widehat K}[\alpha](X)$ are def\/ined
recursively as follows:
if $\alpha$ is a partition, then ${\widehat K}[\alpha](X)= K[\alpha](X)=
x^{\alpha}$;
otherwise, if $\alpha$ and $i$ are such that $\alpha_i < \alpha_{i+1}$, then
\begin{gather*
{\widehat K}[s_i(\alpha)](X)=
x_{i+1} \partial_{i} \big( {\widehat K}[\alpha](X) \big).
\end{gather*}
\item For each permutation $w \in {\mathbb S}_n$ the {\it double
$\beta$-Grothendieck polynomial} ${\cal G}_w^{\beta}(X,Y)$ is def\/ined
recursively as follows:
if $w=w_0$ is the longest element, then ${\cal G}_{w_0}(X,Y)=
R_{\delta_{n}}(X,Y)$;
if $w$ and $i$ are such that $w_i > w_{i+1}$, i.e., $l(ws_i)=l(w)-1$, then
\begin{gather*
{\cal G}_{ws_i}^{\beta}(X,Y)=
\partial_i^{(x)} \big( {(1+ \beta x_{i+1}) {\cal G}_w^{\beta} (X,Y) } \big).
\end{gather*}
\item For each permutation $w \in {\mathbb S}_n$ the {\it double dual
$\beta$-Grothendieck polynomial} ${\cal H}_w^{\beta}(X,Y)$ is def\/ined
recursively as follows:
if $w=w_0$ is the longest element, then ${\cal H}_{w_0}(X,Y) =
R_{\delta_{n}}(X,Y)$;
if $w$ and $i$ are such that $w_i > w_{i+1}$, i.e., $l(ws_i)=l(w)-1$, then
\begin{gather*
{\cal H}_{ws_i}^{\beta}(X,Y)=
(1+ \beta x_{i}) \partial_i^{(x)} \big( {\cal H}_w^{\beta}(X,Y) \big).
\end{gather*}
\item \textit{The key $\beta$-Grothendieck polynomials} $\operatorname{KG}[\alpha](X;\beta)$ are def\/ined recursively as follows\footnote{In the case $\beta= -1$ divided dif\/ference operators
$D_i:=\partial_{i}(x_i-x_i x_{i+1})$ \cite[formula~(6)]{L3} had been
used by A.~Lascoux to describe the transition on Grothendieck polynomials, i.e.,
stable decomposition of any Grothendieck polynomial corresponding to a~permutation $w \in {\mathbb{S}}_n$. into a sum of Grasmannian ones
corresponding to a collection of {\it Grasmannin} permutations
$v_{\lambda} \in {\mathbb{S}}_{\infty}$, see~\cite{L3} for details. The above
mentioned operators~$D_i$ had been used in~\cite{L3} to construct a basis
$\Omega_{\alpha} \vert \alpha \in \mathbb{Z}_{\ge 0}$ that deforms the basis which
is built up from the Demazure (known also as key) polynomials. Therefore
polynomials $\operatorname{KG}[\alpha](X;\beta= -1)$ coincide with those introduced by A.~Lascoux in~\cite{L3}.
In \cite{RY} the authors give a conjectural construction for polynomials
$\Omega_{\alpha}$ based on the use of {\it extended Kohnert moves}, see, e.g., \cite[Appendix by N.~Bergeron]{Ma} for def\/inition of the Kohnert moves. We
state {\it conjecture} that
\begin{gather*
J_{\alpha}^{(\beta)} = \operatorname{KG}[\alpha](X;\beta),
\end{gather*}
where polynomials $J_{\alpha}^{\beta)}$ are def\/ined in \cite{RY} using the
$K$-theoretic versions of the Kohnert moves. For $\beta = -1$ this Conjecture
has been stated by C.~Ross and A.~Yong in \cite{RY}. It seems an interesting
problem to relate the $K$-theoretic Kohnert moves with certain moves of f\/irst
introduced in~\cite{FK1}.}:
if $\alpha$ is a partition, then $\operatorname{KG}[\alpha](X;\beta)= x^{\alpha}$;
otherwise, if $\alpha$ and $i$ are such that $\alpha_i < \alpha_{i+1}$, then
\begin{gather*
\operatorname{KG}[s_i(\alpha)](X;\beta)=
\partial_i \big( (x_i+\beta x_i x_{i+1}) \operatorname{KG}[\alpha](X;\beta) \big).
\end{gather*}
\item \textit{The reduced key $\beta$-Grothendieck polynomials}
${\widehat {\operatorname{KG}}}[\alpha](X;\beta)$ are def\/ined recursively as follows:
if $\alpha$ is a partition, then ${\widehat {\operatorname{KG}}}[\alpha](X;\beta)= x^{\alpha}$;
otherwise, if~$\alpha$ and~$i$ are such that $\alpha_i < \alpha_{i+1}$, then
\begin{gather*
{\widehat {\operatorname{KG}}}[s_i(\alpha)](X;\beta)=
(x_{i+1}+\beta x_i x_{i+1}) \partial_i
\big( {\widehat {\operatorname{KG}}}[\alpha](X;\beta) \big).
\end{gather*}
\end{itemize}
\end{de}
For brevity, we will write $\operatorname{KG}[\alpha](X)$ and ${\widehat {\operatorname{KG}}}[\alpha](X)$ instead of $\operatorname{KG}[\alpha](X;\beta)$ and ${\widehat {\operatorname{KG}}}[\alpha](X;\beta)$.
\begin{rem}\label{rem4.2} We can also introduce polynomials ${\cal Z}_w$, which are
def\/ined recursively as follows:
if $w=w_0$ is the longest element, then ${\cal Z}_{w_0}(X)=
x^{\delta_{n}}$;
if~$w$ and~$i$ are such that $w_i > w_{i+1}$, i.e., $l(ws_i)=l(w)-1$, then
\begin{gather*
{\cal Z}_{ws_i}(X)=
\partial_i \big( (x_{i+1}+x_ix_{i+1}) {\cal Z}_w(X) \big).
\end{gather*}
However, one can show that
\begin{gather*}
{\cal Z}_w(x_1,\dots,x_n)=
(x_1 \cdots x_n)^{n-1}{\cal G}_{w_{0}w w_{0}}\big(x_n^{-1},\dots,x_1^{-1}\big).
\end{gather*}
\end{rem}
\begin{Theorem}\label{theorem4.3}\sloppy The polynomials ${{\mathfrak S}}_w(X,Y)$, $K[\alpha](X)$,
$\widehat {K}[\alpha](X)$, ${\cal G}_w(X,Y)$, ${\cal H}_w(X,Y)$,
$\operatorname{KG}[\alpha](X)$ and $\widehat {\operatorname{KG}}[\alpha](X)$
have nonnegative integer coefficients.
\end{Theorem}
The key step in a proof of Theorem~\ref{theorem4.3} is an observation that for a~given~$n$
the all algebras involved in the def\/inition of the polynomials listed in that theorem, happen to be a suitable quotients of the reduced plactic algebra~${\cal{P}}_n$, and can be extracted from the Cauchy kernel associated with the
algebra ${\cal{P}}_n$ (or that~${\cal{P}}_{n,n}$).
We will use notation ${{\mathfrak S}}_w(X)$, ${\cal G}_w(X)$, \dots, for
polynomials ${{\mathfrak S}}_w(X,0)$, ${\cal G}_w(X,0)$, \dots.
\begin{de}\label{def4.4} For each permutation $w \in {\mathbb S}_n$ the
\textit{Di~Francesco--Zinn-Justin polynomials} ${\cal {DZ}}_w(X)$ are def\/ined
recursively as follows:
if $w$ is the longest element in ${\mathbb S}_n$, then
${\cal {DZ}}_w(X)= R_{\delta_{n}}(X,0)$;
otherwise, if $w$ and $i$ are such that $w_i > w_{i+1}$, i.e.,
$l(ws_i)=l(w)-1$, then
\begin{gather*
{\cal {DZ}}_{ws_i}(X)= \big( (1+x_i) \partial_{i}^{(x)} +
\partial_i^{(x)}(x_{i+1}+x_ix_{i+1}) \big) {\cal {DZ}}_w(X).
\end{gather*}
\end{de}
\begin{con}\label{conj4.5} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Polynomials ${\cal {DZ}}_w(X)$ have nonnegative integer
coefficients.
\item[$(2)$] For each permutation $w \in {\mathbb S}_n$ the polynomial
${\cal {DZ}}_w(X)$ is a linear combination of key polynomials
$K[\alpha](X)$ with nonnegative integer coefficients.
\end{enumerate}
\end{con}
As for def\/i
nition of the double Di Francesco--Zinn-Justin polynomials
${\cal {DZ}}_w(X,Y)$ they are well def\/ined, but may have negative coef\/f\/icients.
Let $\beta$ and $\alpha$ be two parameters, consider divided
dif\/ference operator
\begin{gather*
T_i :=T_{i}^{\beta,\alpha} = -\beta +((\beta+\alpha) x_i +1 +\beta \alpha x_{i} x_{i+1}) \partial_{i,i+1}.
\end{gather*}
\begin{de}\label{def4.6}
Let $w \in {\mathbb{S}}_n$, def\/ine {\it Hecke--Grothendieck} polynomials
${\cal{KN}}_{w}^{\beta,\alpha}(X_n)$ to be
\begin{gather*}
{\cal{KN}}_{w}^{(\beta,\alpha)}(X_n) := T_{w}^{\beta,\alpha}
\big(x^{\delta_{n}}\big),
\end{gather*}
where as before $x^{\delta_{n}}:= x_{1}^{n-1} x_{2}^{n-2} \cdots x_{n-1}$;
if $u \in {\mathbb{S}}_n$, then set
\begin{gather*
T_{u}^{\beta,\alpha} := T_{i_{1}}^{\beta,\alpha} \cdots T_{i_{\ell}}^{\beta,\alpha},
\end{gather*}
where $w =s_{i_{1}} \cdots s_{i_{\ell}}$ is any reduced decomposition of a
permutation taken.
More generally, let $\beta$, $\alpha$ and $\gamma$ be parameters,
consider divided dif\/ference operators
\begin{gather*}
T_{i}:=T_{i}^{\beta,\alpha,\gamma} = - \beta +((\alpha+\beta+\gamma)
x_i +\gamma x_{i+1} + 1 \\
\hphantom{T_{i}:=T_{i}^{\beta,\alpha,\gamma} =}{} +(\beta+\gamma)(\alpha+ \gamma) x_{i} x_{i+1})
\partial_{i,i+1},\qquad i=1,\ldots,n-1.
\end{gather*}
For a permutation $w \in {\mathbb{S}}_n$ def\/ine polynomials
\begin{gather*
{\cal{KN}}_{w}^{(\beta,\alpha,\gamma)}(X_n):= T_{i_{1}}^{\beta,\alpha,\gamma} \cdots T_{i_{\ell}}^{\beta,\alpha,\gamma} \big(x^{\delta_{n}}\big),
\end{gather*}
where $w=s_{i_{1}} \cdots s_{i_{\ell}}$ is any reduced decomposition of~$w$.
\end{de}
\begin{rem}\label{rem4.7} A few comments in order.
\begin{enumerate}\itemsep=0pt
\item[$(a)$] The divided dif\/ference operators $\big\{T_{i}:=T_{i}^{(\beta,\alpha,\gamma)}, \,i=1,
\dots,n-1 \big\}$ satisfy the following relations
\begin{itemize}\itemsep=0pt
\item (Hecke relations)
\begin{gather*}
T_i^2 =(\alpha- \beta) T_{i} + \alpha \beta,
\end{gather*}
\item (Coxeter relations)
\begin{gather*}
T_{i} T_{i+1}T_{i}= T_{i+1} T_{i} T_{i+1},
T_{i}T_{j}=T_{j} T_{i},\qquad \text{if}\quad
|i-j| \ge 2.
\end{gather*}
\end{itemize}
Therefore the elements $T_{w}^{\beta,\alpha, \gamma}$ are well def\/ined for any
$w \in {\mathbb{S}}_n$.
\begin{itemize}\itemsep=0pt
\item (Inversion)
\begin{gather*}
(1+x T_i)^{-1} = \frac{1+(\alpha - \beta) x - x T_{i}}{(1-\beta x)(1 + \alpha x)}.
\end{gather*}
\end{itemize}
\item[$(b)$] Polynomials ${\cal{KN}}_{w}^{(\beta,\alpha,\gamma)}$ constitute a
common generalization of
\begin{itemize}\itemsep=0pt
\item the $\beta$-Grothendieck polynomials, namely, ${\cal{G}}_{w}^{(\beta)} = {\cal{KN}}_{w_{0} w^{-1}}^{(\beta, \alpha =0,\gamma=0)}$,
\item the Di Francesco--Zinn-Justin polynomials,
namely, ${\cal{DZ}}_{w} ={\cal{KN}}_{w}^{(\beta=\alpha=1,\gamma=0)}$,
\item the dual $\alpha$-Grothendieck polynomials,
namely, ${\cal{KN}}_{w_{0} w^{-1}}^{(\beta=0, \alpha,\gamma=0)} ={\cal{H}}_{w}^{\alpha}(X)$.
\end{itemize}
\end{enumerate}
\end{rem}
\begin{pr}\label{prop4.8} \quad
\begin{itemize}\itemsep=0pt
\item $($Duality$)$ Let $w \in {\mathbb{S}}_n$, $\ell = \ell(w)$ denotes
its length, then $(\alpha \beta \not= 0)$
\begin{gather*
{\cal{KN}}_{w}^{(\beta,\alpha)}(1) = (\beta \alpha)^{\ell}
{\cal{KN}}_{{w^{-1}}}^{(\alpha^{-1},\beta^{-1})}(1).
\end{gather*}
\item $($Stability$)$ Let $w \in \mathbb{S}_n$ be a permutation and
$w=s_{i_{1}} s_{i_{2}} \cdots s_{i_{\ell}}$ be any its reduced decomposition.
Assume that $ i_{a} \le n-3$, $\forall\, 1\le a \le \ell$, and define permutation $\widetilde{w} := s_{i_{1}+1} s_{i_{2}+1} \cdots s_{i_{\ell}+1} \in \mathbb{S}_n$. Then
\begin{gather*}
{\cal{KN}}_{w}^{(\beta,\alpha)}(1) = {\cal{KN}}_{\widetilde{w}}^{(\beta, \alpha)}(1).
\end{gather*}
\end{itemize}
\end{pr}
It is well-known that
\begin{itemize}\itemsep=0pt
\item the number ${\cal{KN}}_{w_{0}}^{(\beta=1,\alpha=1)}(1)$
is equal to the degree of the variety of pairs commuting matrices of size $n \times n$ \cite{DZ2, Knu}.
\item the bidegree of the af\/f\/ine homogeneous variety $V_w$, $w \in {\mathbb{S}}_n$ \cite{DZ1} is equal to
\begin{gather*}
A^{{n \choose 2}-\ell(w)} B^{{n \choose 2} +\ell(w)} {\cal{KN}}_{w}^{(\beta =\alpha = A/B)}(1),
\end{gather*}
see \cite{GN, DZ1,Knu} for more details and applications.
\end{itemize}
\begin{con}\label{conj4.9} \quad
\begin{itemize}\itemsep=0pt
\item Polynomials ${\cal {KN}}_{w}^{(\beta,\alpha,\gamma)}(X)$ have
nonnegative integer coefficients
\begin{gather*}
{\cal {KN}}_{w}^{(\beta,\alpha,\gamma)}(X) \in \mathbb{N}[\beta,\alpha, \gamma ] [X_n].
\end{gather*}
\item Polynomials ${\cal {KN}}_{w}^{(\beta,\alpha,\gamma )}(x_1=1,\,
\forall\, i)$ have nonnegative integer coefficients
\begin{gather*
{\cal {KN}}_{w}^{(\beta,\alpha,\gamma)}(x_i=1,\forall i) \in
\mathbb{N}[\beta,\alpha, \gamma ].
\end{gather*}
\item Double polynomials
\begin{gather*}
{\cal{KN}}_{w}^{(\beta = 0,\alpha, \gamma)}(X,Y) = T_{w}^{\beta=0,\alpha,\gamma}(x) \prod_{i+j \le n+1 \atop i \ge 1, \, j \ge 1}(x_i+y_j)
\end{gather*}
are well defined and have nonnegative integer coefficients\footnote{Note that the assumption $\beta=0$ is necessary.}.
\item Consider permutation $w=[n,1,2,\dots,n-1] \in \mathbb{S}_n$.
Clearly $w= s_{n-1} s_{n-2} \cdots s_{2} s_{1}$.
The number ${\cal{KN}}_{w}^{(\beta=1,\alpha=1,\gamma=0)}(1)$ is equal to
the number of Schr\"{o}der paths of semilength~$(n-1)$ in which the
$(2,0)$-steps come in $3$ colors and with no peaks at level~$1$,
see {\rm \cite[$A162326$]{SL}} for further properties of these numbers.
\end{itemize}
\end{con}
It is well-known, see, e.g., \cite[$A126216$]{SL}, that the polynomial
${\cal{KN}}_{w}^{(\beta,\alpha=0)}(1)$ counts the number of {\it dissections} of a convex $(n+1)$-gon according the number of diagonals involved, where as
the polynomial ${\cal{KN}}_{w}^{(\beta,\alpha)}(1)$ (up to a normalization) is
equal to the {\it bidegree} of certain algebraic varieties introduced and
studied by A.~Knutson~\cite{Knu}.
A few comments in order.
\begin{enumerate}\itemsep=0pt
\item[$(a)$] One can consider more general family of polynomials
${\cal{KN}}_{w}^{(a,b,c,d)}(X_n)$ by the use of the divided dif\/ference
operators
$
T_{i}^{a,b,c,d}:= -b+((b+d) x_i + c x_{i+1}+1+d (b+c) x_i x_{i+1}) \partial_{i,i+1}^{x}
$
instead of that $T_{i}^{\beta,\alpha,\gamma}$. However the
polynomials ${\cal{KN}}_{w}^{(a,b,c,d)}(1) \in \mathbb{Z}[a,b,c,d]$ may have negative coef\/f\/icients in general. {\it Conjecturally}, to ensure the positivity of
polynomials ${\cal{KN}}_{w}^{(a,b,c,d)}(X_n)$, it is necessary take $d:=a+c+r$. In this case we state conjecture
\begin{gather*}
{\cal{KN}}_{w}^{(a,b,c,a+c+r)}(X_n) \in \mathbb{N}[a,b,c,r]
\end{gather*}
We state more general Conjecture~\ref{conjecture1.0} in introduction. In the present paper we
treat only the case $r=0$, since a combinatorial
meaning of polynomials ${\cal{KN}}_{w}^{(a,b,c,a+c+r)}(1)$ in the case
$r \not= 0$ is missed for the author.
\item[$(b)$] If $\gamma \not= 0$, the polynomials ${\cal {KN}}_{w}^{(\beta,\alpha,\gamma)}(X_n) \in \mathbb{Z} [\alpha, \beta,\gamma] [X_n]$ may have negative coef\/f\/icients in general.
\end{enumerate}
\begin{Theorem}\label{theorem4.10} Let $T$ be a semistandard tableau and $\alpha(T)$ be its
bottom code, see Definition~{\rm \ref{def2.27}}. Then
\begin{gather*}
{\cal K}_{T}(X)=K[\alpha(T)](X), \qquad
{\cal {KG}}_{T}(X)=\operatorname{KG}[\alpha(T)](X).
\end{gather*}
\end{Theorem}
Let $\alpha=(\alpha_1 \le \alpha_2 \le \cdots \le \alpha_r)$ be a
composition, def\/ine partition $\alpha^{+}=(\alpha_r \ge \cdots \ge \alpha_1)$.
\begin{pr}\label{prop4.11} If $\alpha=(\alpha_1 \le \alpha_2 \le \cdots \le \alpha_r)$ is
a composition and $n \ge r$, then
\begin{gather*}
K[\alpha](X_n)=s_{\alpha^{+}}(X_r).
\end{gather*}
\end{pr}
For example, $K[0,1,2,\dots,n-1]= \prod\limits_{1 \le i < j \le n}(x_i+x_j)$. Note
that
${\widehat K}[0,1,2,\dots,n-1]= \prod\limits_{i=2}^{n}x_i^{i-1}$.
\begin{pr}\label{4.12} If $\alpha=(\alpha_1 \le \alpha_2 \le \cdots \le \alpha_r)$ is
a composition and $n \ge r$, then
\begin{gather*}
\operatorname{KG}[\alpha](X_n)={\cal G}[{\alpha^{+}}](X_r).
\end{gather*}
\end{pr}
For example, $\operatorname{KG}[0,1,2,\dots,n-1]=
\prod\limits_{1 \le i < j \le n}(x_i+x_j+x_ix_j)$.
Note that
\begin{gather*}
{\widehat {\operatorname{KG}}}[0,1,2,\dots,n-1]=
\prod_{i=2}^{n}x_i^{i-1} \prod_{i=1}^{n-1}(1+x_i)^{n-i}.
\end{gather*}
\begin{de}\label{def4.14} Def\/ine degenerate af\/f\/ine $2D$ nil-Coxeter algebra
${\cal{ANC}}_n^{(2)}$ to be an associative algebra over $\mathbb{Q}$ generated by the
set of elements $\{\{u_{i,j} \}_{1 \le i <j \le n},\, x_1,\ldots,
x_n \}$ subject to the set of relations
\begin{itemize}\itemsep=0pt
\item $x_i x_j = x_j x_i$ for all $i \not= j$, $x_i u_{j,k}=u_{j,k}
x_i$, if $i \not= j, k$,
\item $u_{i,j} u_{k,l} = u_{k,l} u_{i,j}$, if $i$, $j$, $k$, $l$ are pairwise
distinct,
\item ($2D$-Coxeter relations) $u_{i,j} u_{j,k} u_{i,j} =u_{j,k} u_{i,j} u_{j,k}$, if $1 \le i < j < k \le n$,
\item $x_i u_{i,j}=u_{i,j} x_j + 1$, $x_j u_{i,j}=u_{i,j} x_i - 1$.
\end{itemize}
\end{de}
Now for a set of parameters\footnote{By def\/inition, a \textit{parameter} is assumed to be belongs to the
center of the algebra in question.}
$ A:=(a,b,c,h,e)$ def\/ine elements
\begin{gather*}
T_{ij}:= a+ (b x_i+ c x_j +h+e x_i x_j) u_{i,j}, \qquad i < j.
\end{gather*}
Throughout the present paper we set $T_i:= T_{i,i+1}$.
\begin{lem}\label{lem4.15} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] $T_{i,j}^2 =(2a+b-c) T_{i,j} -a(a+b-c)$,
if $a=0$, then $T_{ij}^2 = (b-c) T_{ij}$.
\item[$(2)$] $2D$-Coxeter relations
$T_{i,j} T_{j,k} T_{i,j} = T_{j,k} T_{i,j} T_{j,k}$
are valid, if and only if the following relation among parameters
$a$, $b$, $c$, $e$, $h$ holds\footnote{The relation~\eqref{equation4.1} between parameters $a$, $b$, $c$, $e$, $h$ def\/ines a
{\it rational} four-dimensional hypersurface. Its open chart $ \{e h \not= 0
\}$ contains, for example, the following set (cf.~\cite{L3}): $\{a=p_1
p_4-p_2 p_3, \, b= p_2 p_3, \, c= p_1 p_4, \, e=p_1 p_3$, $h= p_2 p_4 \}$, where $(p_1,p_2,p_3,p_4)$ are arbitrary parameters. However the points $(-b,a+b+c,c,1,(a+c)(b+c)$, $(a,b,c) \in \mathbb{N}^3\}$ do not belong to this set.}
\begin{gather}\label{equation4.19}
(a+b)(a-c)+h e =0.
\end{gather}
\item[$(3)$] Yang--Baxter relations
$T_{i,j} T_{i,k} T_{j.k}= T_{j,k} T_{i,k} T_{i,j}$
are valid if and only if $b=c=e=0$, i.e., $T_{ij}=a+d u_{ij}$.
\item[$(4)$] $T_{ij}^2=1$ if and only if $a= \pm 1$, $c=b \pm 2$, $he= (b\pm 1)^2$.
\item[$(5)$] Assume that parameters $a$, $b$, $c$, $h$, $e$ satisfy the conditions~\eqref{equation4.19} and
that $b c+1=h e$. Then
\begin{gather*}
T_{ij} x_i T_{ij} = (h e - b c) x_j+ (h+(a+b)(x_i+x_j)+e x_i x_j) T_{ij}.
\end{gather*}
\end{enumerate}
Some special cases
\begin{itemize}\itemsep=0pt
\item $($representation of affine modified Hecke algebra~{\rm \cite{TO})}
if $A=(a,-a,c,h,0)$, then $T_{ij} x_{i} T_{ij}$ $ = a c x_{j} + h T_{ij}$, $i < j$,
\item if $A=(-a,a+b+c,c,1,(a+c)(b+c)$, then
\begin{gather*}
T_{ij} x_{i} T_{ij}= a b x_{j}+ (1 +(b+c)(x_i+x_j)+(a+c)(b+c) x_{i} x_{j}) T_{ij}.
\end{gather*}
\end{itemize}
\begin{enumerate}\itemsep=0pt
\item[$(6)$] $($Quantum Yang--Baxter relations, or baxterization of Hecke's
algebra generators.$)$ Assume that parameters $a$, $b$, $c$, $h$, $e$
satisfy the conditions~\eqref{equation4.19} and that $\beta:=2 a+b-c \not= 0$. Then
$($cf.~{\rm \cite{IK, LLT}} and the literature quoted therein$)$
the elements $R_{ij}(u,v):= 1+ \frac{\lambda-\mu}
{\beta \mu} T_{ij}$ satisfy the twisted quantum Yang--Baxter relations
\begin{gather*}
R_{ij}(\lambda_i,\mu_j) R_{jk}(\lambda_i,\nu_k) R_{ij}(\mu_j,\nu_k)= R_{jk}(\mu_j,\nu_k) R_{ij}(\lambda_i,\nu_k) R_{jk}(\lambda_i,\mu_j), \qquad i < j < k,
\end{gather*}
where $\{\lambda_i,\mu_i,\nu_i \}_{1 \le i \le n}$ are parameters.
\end{enumerate}
\end{lem}
\begin{cor}\label{cor4.16}
If $(a+b)(a-c)+h e=0$, then for any permutation
$w \in {\mathbb{S}}_n$ the element
\begin{gather*}
T_w:=T_{i_{1}} \cdots T_{i_{l}} \in {\cal{ANC}}_n^{(2)},
\end{gather*}
where $w=s_{i_{1}} \cdots s_{i_{l}}$ is any reduced decomposition of~$w$, is well-defined.
\end{cor}
\begin{ex}\label{exam4.17} \quad
\begin{itemize}\itemsep=0pt
\item Each of the set of elements
\begin{gather*}
s_i^{( h)}=1+(x_{i+1}-x_i + h) u_{i,i+1}\qquad \text{and} \\
t_i^{(h)}=-1+(x_i-x_{i+1}+ h(1+x_i)(1+x_{i+1})) u_{ij}, \qquad i=1,\ldots,n-1,
\end{gather*}
by itself generate the symmetric group ${\mathbb{S}}_n$.
\item If one adds the af\/f\/ine elements $s_{0}^{(h)} :=\pi s_{n-1}^{(h)} \pi^{-1}$ and $t_{0}^{(h)}:=\pi t_{n-1}^{(h)} \pi^{-1}$, then each of the set of elements $\big\{s_{j}^{(h)},\, j \in \mathbb{Z}/n \mathbb{Z} \big\}$ and $\big\{t_{j}^{(h)},\, j
\in \mathbb{Z}/n\mathbb{Z}\big\}$ by itself generate the af\/f\/ine symmetric group
${\mathbb{S}}_n^{\text{af\/f}}$, see \eqref{comm4.21pi} for a def\/inition of the transformation~$\pi$.
\item It seems an interesting problem to classify all rational,
trigonometric and elliptic divided dif\/ference operators satisfying the Coxeter
relations. A general divided dif\/ference ope\-ra\-tor with {\it polynomial
coefficients} had been constructed in~\cite{LSZ}, see also Lemma~\ref{lem4.15}, relation~\eqref{equation4.1}. One can construct a family of {\it rational} representations of the
symmetric group (as well as its af\/f\/ine extension) by ``iterating'' the
transformations $s_j^{(h)}$, $j \in \mathbb{Z}/n\mathbb{Z}$. For example, take parameters
$a$ and $b$, def\/ine \textit{secondary} divided dif\/ference operator
\begin{gather*}
\partial_{xy}^{[a,b]}:= -1 +(b+y-x) \partial_{xy}^{[a]},
\end{gather*}
where
\begin{gather*}
\partial_{xy}^{[a]}:= \frac{1- {{\overline{s}}_{xy}^{(a)}}}{a-x+y},\qquad
{\overline{s}}_{xy}^{(a)}:=-1+(a+x-y) \partial_{xy}.
\end{gather*}
Observe that the set of operators $ \big\{s_{i}^{[a,b]}:= s_{x_{i},x_{i+1}}^{[a,b]},\,
i \in \mathbb{Z}/n\mathbb{Z} \big\}$ gives rise to a rational representation of the af\/f\/ine
symmetric group ${\mathbb{S}}_n^{\text{af\/f}}$ on the f\/ield of rational functions
$\mathbb{Z}[a,b](X_n)$.
In the special case $a:= A$, $b:= A/h$, $h:=1-\beta/2$ the operators
$s_{i}^{[a,b]}$ coincide with operators $\Theta_i$, $i \in \mathbb{Z}/n\mathbb{Z}$ have been
introduced in~\cite[equation~(4.17)]{KnZ}.
\end{itemize}
\end{ex}
Let $A=(a,b,c,h,e)$ be a sequence of integers satisfying the conditions~\eqref{equation4.1}. Denote by $\partial_i^{A}$ the divided dif\/ference operator
\begin{gather*}
\partial_{i}^{A}= a+(b x_i+c x_{i+1}+h+e x_i x_{i+1}) \partial_i, \qquad i=1,\ldots,n-1.
\end{gather*}
It follows from Lemma~\ref{lem4.15} that the operators $\big\{\partial_{i}^{A}\big\}_{1 \le i
\le n}$ satisfy the Coxeter relations
\begin{gather*}
\partial_{i}^{A} \partial_{i+1}^{A} \partial_{i}^{A}=\partial_{i+1}^{A} \partial_{i}^{A} \partial_{i+1}^{A}, \qquad i=1,\ldots,n-1.
\end{gather*}
\begin{de}\label{def4.19}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Let $w \in \mathbb{S}_n$ be a permutation. Def\/ine the generalized
Schubert polynomial corresponding to permutation $w$ as follows
\begin{gather*}
\mathfrak{S}_w^{A}(X_n) = \partial_{w^{-1} w_{0}}^{A} x^{\delta_n}, \qquad
x^{\delta_n}:= x_1^{n-1} x_2^{n-2} \cdots x_{n-1},
\end{gather*}
and $w_0$ denotes the longest element in the symmetric group $\mathbb{S}_{n}$.
\item[(2)] Let $\alpha$ be a composition with at most $n$ parts, denote by
$w_{\alpha} \in \mathbb{S}_n$ the permutation such that $w_{\alpha}(\alpha)=
\alpha^{+}$. Let us recall that $\alpha^{+}$ denotes a unique partition
corresponding to composition~$\alpha$.
\end{enumerate}
\end{de}
\begin{lem}\label{lem4.20} Let $w \in \mathbb{S}_n$ be a permutation.{\samepage
\begin{itemize}\itemsep=0pt
\item If $A=(0,0,0,1,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the Schubert polynomial $\mathfrak{S}_{w}(X_n)$.
\item If $A=(-\beta,\beta,0,1,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is
equal to the $\beta$-Grothendieck polynomial $\mathfrak{G}_w^{({\beta})}(X_n)$
introduced in~{\rm \cite{FK1}}.
\item If $A=(0,\beta,0,1,0)$ then $\mathfrak{S}_{w}^{A}(X_n)$ is equal
to the dual $\beta$-Grothendieck polynomial ${\cal{H}}_{w}^{(\beta)}(X_n)$,
studied in depth for $\beta=-1$ and in the basis $\{x_i:=\exp(\xi_{l})\}$ in~{\rm \cite{L33}}.
\item If $A=(-1,2,0,1,1)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the Di Francesco--Zinn-Justin polynomials introduced in~{\rm \cite{DZ1}}.
\item If $A=(1,-1,1, h,0)$, then $\mathfrak{S}_{w}^{A}(X_n)$ is equal to
the $h$-Schubert polynomials.
\end{itemize}
In all cases listed above the polynomials $\mathfrak{S}_{w}^{A}(X_n)$ have
non-negative integer coefficients.}
\end{lem}
Def\/ine the generalized key or Demazure polynomial corresponding to a
composition~$\alpha$ as follows
\begin{gather*}
K_{\alpha}^{A}(X_n)= \partial_{w_{\alpha}}^{A} x^{\alpha^{+}}.
\end{gather*}
\begin{itemize}\itemsep=0pt
\item If $A=(1,0,1,0,0)$, then $K_{\alpha}^{A}(X_n)$ is equal to key (or
Demazure) polynomial corresponding to~$\alpha$.
\item If $A=(0,0,1,0,0)$, then $K_{\alpha}^{A}(X_n)$ is equal to the
reduced key polynomial.
\item If $A=(1{,}0{,}1{,}0{,}\beta)$, then $K_{\alpha}^{A}(X_n)$ is equal to the
key $\beta$-Grothendieck polynomial $\operatorname{KG}_{\alpha}^{(\beta)}\!(X_n)$.
\item If $A=(0,0,1,0,\beta)$, then $K_{\alpha}^{A}(X_n)$ is equal to
the reduced key $\beta$-Grothendieck polynomials.
\end{itemize}
In all cases listed above the polynomials $\mathfrak{S}_{w}^{A}(X_n)$ have
non-negative integer coef\/f\/icients.
\begin{itemize}\itemsep=0pt
\item If $A=(-1,q^{-1},-1,0,0)$ and $\lambda$ is a partition, then (up
to a scalar factor) polynomial $K_{\lambda}^{A}(X_n)$ can be identify with a
certain {\it Whittaker function} (of type $A$), see~\cite[Theorem~A]{BBL}.
Note that operator $T_{i}^{A}:= -1 +(q^{-1} x_i-x_{i+1}) \partial_i$, $1 \le i
\le n-1$, satisfy the Coxeter and Hecke relations, namely $(T_{i}^{A})^{2}=
(q^{-1}-1) T_{i}^{A}+q^{-1}$. In~\cite{BBL} the operator $T_{i}^{A}$ has been
denoted by~${\mathfrak{T}}_i$.
\item Let $w \in {\mathbb{S}}_n$ be a permutation and $\mathfrak{m}=(i_1,\ldots,i_{\ell})$ be a reduced word for $w$, i.e., $w=s_{i_{1}} \cdots s_{i_{\ell}}$ and $\ell(w)= \ell$. Denote by $Z_{\mathfrak{m}}$ the {\it Bott--Samelson} nonsingular variety corresponding to the reduced word~$\mathfrak{m}$.
It is well-known that the Bott--Samelson variety $Z_{\mathfrak{m}}$ is
birationally isomorphic to the Schubert variety $X_{w}$ associated with
permutation $w$, i.e., the Bott--Samelson variety $Z_{\mathfrak{m}}$ is a~desingularization of the Schubert variety~$X_{w}$. Following~\cite{BBL} def\/ine
the Bott--Samelson polynomials $Z_{\mathfrak{m}}(x,\lambda,v)$ as follows
\begin{gather*}
Z_{\mathfrak{m}}(x,\lambda,v) = \big(1+T_{i_{1}}^{A}\big) \cdots \big(1+T_{i_{\ell}}^{A}\big) x^{\lambda},
\end{gather*}
where $A=(-v,-1,1,0)$. Note that $\big(1+T_{i}^{A}\big)^2= (1+v)\big(1+T_{i}^{A}\big)$, and the
divided dif\/ference operators $1+T_{i}^{A}=1+v+(x_{i+1}-v x_{i}) \partial_i$
do {\it not} satisfy the Coxeter relations.
\item If $A=(- \beta, \beta + \alpha,0,1,\beta \alpha)$, then
${\mathfrak S}_{w}^{A}(X_n)$ constitutes a common generalization of the $\beta$-Grothendieck and the Di Francesco--Zinn-Justin polynomials.
\item If $A=(t,-1,t,1,0)$, then the divided dif\/ference operators
\begin{gather*
T_i^{A}:= t+(-x_i+t x_{i+1} +1) \partial_i, \qquad 1 \le i \le n-1,
\end{gather*}
their \textit{baxterizations} and the \textit{raising operator}
\begin{gather*
\phi:= (x_n-1) \pi,
\end{gather*}
where $\pi$ denotes the {\it $q^{-1}$-shift operator}, namely $\pi(x_1,\ldots,x_n)=(x_n/q,x_1,\ldots,x_{n-1})$,
can be used to \textit{generate} the
\textit{interpolation Macdonald polynomials} as well as the
\textit{nonsymmetric Macdonald polynomials}, see~\cite{LRW} for details.
\end{itemize}
In similar fashion, relying on the operator $\phi$, operators
\begin{gather*}
T_i^{\beta,\alpha,\gamma,h}:=
-\beta+((\alpha+\beta+\gamma) x_i+\gamma x_{i+1}+h\\
\hphantom{T_i^{\beta,\alpha,\gamma,h}:=}{}
+h^{-1} (\alpha+\gamma)(\beta+\gamma) x_i x_{i+1}) \partial_{i}, \qquad 1 \le i \le n-1,
\end{gather*}
and their \textit{baxterization},
one (A.K.) can introduce polynomials $M_{\delta}^{\beta,\alpha,\gamma,q}(X_n)$, where
$\delta$ is a \textit{composition}. These polynomials are \textit{common
generalizations} of the interpolation Macdonald polynomials
$M_{\delta}(X_n;q,t)$ (the case $\beta=-t$, $\alpha=-1$, $\gamma =t$), as well as
the Schubert, $\beta$-Grothendieck and its \textit{dual}, Demazure and
Di Francesco--Zinn-Justin polynomials, and \textit{conjecturally} their
{\it affine} analogues/versions. Details will appear elsewhere.
{\bf Double af\/f\/ine nilCoxeter algebra.}
Let $t$, $q$, $a$, $b$, $c$, $h$, $d$ be
parameters.
\begin{de}\label{def4.22} Def\/ine double af\/f\/ine nil-Coxeter algebra $\operatorname{DANC}_{n}$ to be (unital)
associative algebra over $\mathbb{Q}\big(q^{\pm 1},t^{\pm 1}\big)$ with the set of generators $ \big\{e_1,\ldots,e_{n-1}, x_{1},\ldots,x_n, \pi^{\pm 1} \big\}$ subject to relations
\begin{itemize}\itemsep=0pt
\item (nilCoxeter relations)
\begin{gather*}
e_i e_j =e_j e_i,\quad \text{if} \quad |i-j| \ge 2, \qquad e_i^2=0, \quad \forall\, i, \qquad e_i e_j e_i =e_j e_i e_j, \quad \text{if}\quad |i-j|=1;
\end{gather*}
\item (crossing relations)
\begin{gather*
x_i e_k =e_k x_i, \quad \text{if} \quad k \not=i, i+1, \qquad x_i e_i - e_{i} x_{i+1} =1, \qquad
e_i x_i - x_{i+1} e_i =1;
\end{gather*}
\item (af\/f\/ine crossing relations)
\begin{gather*}
\pi x_i =x_{i+1} \pi, \quad \text{if}\quad i < n,\qquad \pi x_{n} = q^{-1} x_1 \pi, \\
\pi e_{i} = e_{i+1} \pi, \quad \text{if}\quad i < n-1, \qquad \pi^2 e_{n-1} = q e_{1} \pi^2.
\end{gather*}
\end{itemize}
\end{de}
Now let us introduce elements $e_{0}:= \pi e_{n-1} \pi^{-1}$ and
\begin{gather*
T_{0}:= T_{0}^{a,b,c,h,d} =
\pi T_{n-1} \pi^{-1} = a+\big(b x_{n}+ q^{-1} c x_1 + h +
q^{-1} d x_1 x_n\big) e_{0}.
\end{gather*}
It is easy to see that $\pi e_{0} =q e_{1} \pi$,
\begin{gather*}
\pi T_{0}^{a,b,c,h,d} = T_{1}^{a,b,c,q h,q^{-1} d}
e_{1} \pi = T_{1}^{a,b,c,h,d} +\big((1-q) h+\big(1-q^{-1}\big) d x_1 x_2\big) e_1.
\end{gather*}
Now let us assume that $a = t$, $b=-t$, $d=e=0$, $c=1$. Then,
\begin{gather*}
T_i= t+(x_{i+1} -t x_{i}) e_i , \quad i=1,\ldots,n-1 , \qquad T_0= t+\big(q^{-1} x_1 -t x_{n}\big) e_{0},
\\
T_i^2 =(t-1) T +t , \quad 0 \le i < n, \qquad T_i x_i T_i =t x_{i+1}, \quad 1 \le i <
n, \qquad T_0 x_{n} T_{0}= t q^{-1} x_1,
\\
T_0 T_1 T_0=T_1 T_0 T_1 , \qquad T_{n-1} T_0 T_{n-1} = T_0 T_{n-1} T_0, \qquad T_0 T_i =T_i T_0, \qquad \text{if} \quad 2 \le i < n-1.
\end{gather*}
The operators $T_{i}:= T_{i}^{t,-t,1,0,0}$, $0 \le i \le n-1$ have been used
in~\cite{LRW} to give an ``elementary'' construction of nonsymmetric Macdonald polynomials. Indeed, one can realize the operator~$\pi$ as
follows:
\begin{gather}\label{comm4.21pi}
\pi(f)= f(x_{n} /q, x_1,x_2,\ldots,x_{n-1}), \qquad \text{so that} \quad
\pi^{-1}(f)=(x_2,\ldots,x_{n},q x_1),
\end{gather}
and introduce the raising operator~\cite{LRW} to be
\begin{gather*}
\phi(f(X_n)) = (x_n-1) \pi (f(X_n)).
\end{gather*}
It is easily seen that $\phi T_i = T_{i+1} \phi$, $i=0,\dots,n-2$,
and $\phi^2 T_{n-1}= T_1 \phi^2$. It has been established in~\cite{LRW}
how to use the operators $\phi$, $T_1,\ldots,T_{n-1}$ to to give formulas for
the \textit{interpolation Macdonald polynomials}. Using operators $\phi,
T_{i}^{(a,b,c,h,d)}, i=1,\ldots,n-1$ instead of $\phi$, $T_1,\ldots,T_{n-1}$, $1 \le i \le n-1$, one get a $4$-parameter generalization of the interpolation
Macdonald polynomials, as well as the nonsymmetric Macdonald polynomials.
It follows from the nilCoxeter relations listed above, that the
Dunkl--Cherednik elements, cf.~\cite{CH},
\begin{gather*
Y_i:= \left( \prod_{a=i-1}^{1} T_{a}^{-1} \right) \pi \left( \prod_{a=n-1}^{i+1} T_{a} \right), \qquad i=1,\ldots,n,
\end{gather*}
where $T_i=T_{i}^{t,-t,1,0,0}$, generate a commutative subalgebra in the
double af\/f\/ine nilCoxeter algebra $\operatorname{DANC}_{n}$. Note that the algebra $DANC_n$
contains lot of other interesting commutative subalgebras, see, e.g.,~\cite{IK}.
It seems interesting to give an interpretation of polynomials generated by
the set of opera\-tors~$T_{i}^{t,-t,1,h,e}$, $i=0,\dots,n-1$ in a~way similar to
that given in~\cite{LRW}. We expect that these polynomials provide an~af\/f\/ine
version of polynomials ${\cal{KN}}_{w}^{(-t,-1,1,1,0)}(X)$, $w \in
{\mathbb{S}}_n\subset {\mathbb{S}}_n^{\text{af\/f}}$, see Remark~\ref{rem4.7}.
Note that for any af\/f\/ine permutation $ v \in {\mathbb{S}}_{n}^{\text{af\/f}}$, the
operator
\begin{gather*}
T_{v}^{(a,b,c,h,d)}=T_{i_{1}}^{(a,b,c,h,d)} \cdots
T_{i_{\ell}}^{(a,b,c,h,d)},
\end{gather*}
where $v=s_{i_{1}} \cdots s_{i_{\ell}}$ is any reduced decomposition
of $v$, is {\it well-defined} up to the sign~$\pm 1$. It seems an
interesting {\it problem} to investigate properties of polynomials
$L_{v}[\alpha](X_n)$, where $v \in {\mathbb{S}}_n^{\text{af\/f}}$ and $\alpha \in
\mathbb{Z}_{\ge 0}^{n}$, and f\/ind its algebro-geometric interpretations.
\section{Cauchy kernel}\label{section5}
Let $u_1,u_2,\dots,u_{n-1}$ be a set of generators of the free algebra
${\cal F}_{n-1}$, which are assumed also to be commute with the all variables
$ {\mathfrak{P}}_n:=\{p_{i,j}, \, 2 \le i+j \le n+1,\, i \ge 1, j \ge 1 \}$.
\begin{de}\label{def5.1} The Cauchy kernel ${\cal C}({\mathfrak{P}}_n,U)$ is def\/ined to be
as the ordered product
\begin{gather}\label{equation5.1}
{\cal C}({\mathfrak{P}}_n,U)= \prod_{i=1}^{n-1}
\left \{{\prod_{j=n-1}^{i}(1+ p_{i,j-i+1} u_{j})} \right\}.
\end{gather}
\end{de}
For example,
\begin{gather*}
{\cal C}({\mathfrak{P}}_4,U)=(1+p_{1,3} u_3)(1+p_{1,2} u_2)
(1+p_{1,1} u_1)(1+p_{2,2} u_3)(1+p_{2,1} u_2)(1+p_{3,1} u_3).
\end{gather*}
In the case $\{p_{ij}=x_i, \, \forall \, j \}$ we will write ${\cal{C}}_{n}(X,U)$
instead of ${\cal C}({\mathfrak{P}}_n,U)$.
\begin{lem}\label{lem5.2}
\begin{gather}\label{equation5.2}
{\cal C}({\mathfrak{P}}_n,U) = \sum_{({\bf a},{\bf b}) \in {\cal S}_n}
\prod_{j=1}^{p} p_{\{{a_{j},b_{j} \}}} w({\bf a},{\bf b}),
\end{gather}
where ${\bf a} =(a_1,\ldots,a_p)$, ${\bf b}=(b_1,\ldots,b_p)$,
$w({\bf a},{\bf b})= \prod\limits_{j=1}^{p} u_{a_{j}+b_{j} -1}$, and the sum in
\eqref{equation5.2} runs over the set
\begin{gather*}
{\cal S}_n := \big\{({\bf a},{\bf b}) \in \mathbb{N}^p \times \mathbb{N}^p \,|\, {\bf a}=
(a_1 \le a_2 \le \cdots \le a_p),\\
\hphantom{{\cal S}_n := \{({\bf a},{\bf b}) \in \mathbb{N}^p \times \mathbb{N}^p \,|\,}{} a_i+b_i \le n, \ \text{and if} \ a_i=a_{i+1}
\Longrightarrow b_i > b_{i+1} \big\}.
\end{gather*}
\end{lem}
We denote by ${\cal S}_n^{(0)}$ the set $\{({\bf a},{\bf b}) \in {\cal S}_n
\,|\, w({\bf a},{\bf b}) \ \text{is a tableau word}\}$.
The number of terms in the r.h.s.\ of~\eqref{equation5.1} is equal to
$2^{{n \choose 2}}$, and therefore is equal to the number $\# |\operatorname{STY}(\delta_n,
\le n)|$ of semistandard Young tableaux of the staircase shape
$\delta_n:=(n-1$, $n-2,\ldots,2,1)$ f\/illed by the numbers from the set
$\{1,2,\ldots,n \}$. It is also easily seen that the all terms appearing in
the r.h.s.\ of~\eqref{equation5.2} are dif\/ferent, and thus $\# | {\cal S}_n |= \# |\operatorname{STY}(\delta_n,
\le n)|$.
We are interested in the decompositions of the Cauchy kernel
${\cal C}({\mathfrak{P}}_n,U)$ in the algebras ${\cal P}_{n}$,
${\cal NP}_{n}$, ${\cal IP}_{n}$, ${\cal NC}_{n}$ and ${\cal IC}_{n}$.
\subsection[Plactic algebra ${\cal P}_n$]{Plactic algebra $\boldsymbol{{\cal P}_n}$}\label{section5.1}
Let $\lambda$ be a partition and $\alpha$ be a composition of the same size.
Denote by ${\widetilde {\operatorname{STY}}}(\lambda,\alpha)$ the set of semistandard Young
tableaux~$T$ of the shape $\lambda$ and content $\alpha$ which must satisfy
the following conditions:
for each $k=1,2,\dots$, the all numbers~$k$ are located in the
f\/irst~$k$ columns of the tableau~$T$. In other words, the all entries~$T(i,j)$ of a semistandard tableau $T \in {\widetilde {\operatorname{STY}}}(\lambda,\alpha)$
have to satisfy the following conditions: $T_{i,j} \ge j$.
For a given (semi-standard) Young tableau $T$ let us denote by~$R_i(T)$ the
set of numbers placed in the $i$-th row of~$T$, and denote by
${\widetilde {\operatorname{STY}}}_{0}(\lambda,\alpha)$ the subset of the set
${\widetilde {\operatorname{STY}}}(\lambda,\alpha)$ involving only tableaux $T$ which
satisfy the following constrains
\begin{gather*}
R_1(T) \supset R_2(T) \supset R_3(T) \supset \cdots.
\end{gather*}
To continue, let us denote by ${\cal A}_n$ (respectively by
${\cal A}^{(0)}_n$) the union of the sets
${\widetilde {\operatorname{STY}}}(\lambda, \alpha)$ (resp.\ that of
${\widetilde {\operatorname{STY}}}_{0}(\lambda, \alpha)$)
for all partitions $\lambda$ such that $\lambda_i \le n-i$ for $i=1,2,\dots,
n-1$, and all compositions $\alpha$, $l(\alpha) \le n-1$. Finally, denote by
${\cal A}_n(\lambda)$ (resp.~${\cal A}^{(0)}_n(\lambda)$) the subset of
${\cal A}_n$ (resp.~${\cal A}^{(0)}_n(\lambda)$) consisting of all tableaux
of the shape~$\lambda$.
\begin{lem}\label{lem5.3} \quad
\begin{itemize}\itemsep=0pt
\item $|{\cal A}_n(\delta_n)|=1$,
$|{\cal A}_n(\delta_{n-1})|= (n-1)!$, $|{\cal A}_n((n-1))|= C_{n-1}$ the
$(n-1)$-th Catalan number. More generally,
\begin{gather*}
\big|{\cal A}_n\big(\big(1^{k}\big)\big)\big|= {n -1 \choose k}, \qquad |{\cal A}_n((k))|= \frac{n-k}{n} {n+k-1 \choose k} = \dim {\mathbb{S}}_{n+k-1}^{(n - 1,k)},
\end{gather*}
$k=1 \dots,n-1$, cf.~{\rm \cite[$A009766$]{SL}}; here $\dim {\mathbb{S}}_{n}^{\lambda}$ stands for the dimension of the irreducible representation of the
symmetric group ${\mathbb{S}}_{n}$ corresponding to a~partition~$\lambda \vdash n$.
\item Let $k \ge \ell \ge 2$, $n \ge k+2$, then
\begin{gather*}
|{\cal A}_n((k,\ell))| = \frac{(n-k)(n-\ell+1)(k-\ell+1)(n^2-n-\ell(k+1))}
{\ell ! (k+1) k (n+k)} {n+k \choose k-1} \prod_{i=1}^{\ell-2}(n+i),
\end{gather*}
$ k=1, \ldots,n$. The case $k= \ell$ has been studied in~{\rm \cite{Gou}} where one
can find a combinatorial interpretation of the numbers $|{\cal A}_n((k,k))|$
for all positive integers~$n$; see also~{\rm \cite[$A005701$, $A033276$]{SL}} for
more details concerning the cases $k=2$ and $k=3$. Note that in the case $k=\ell$ one has $n^2-n-k(k+1)= (n+k)(n-k-1)$, and the above formula can be rewritten as follows $( k \ge 2)$
\begin{gather*}
|{\cal A}_n((k,k))| = \frac{(n-k-1)(n-k)(n-k-1)}{(k+1) k^2 (k-1)}~{n+k-2 \choose k-2}~{n+k-1 \choose k-1}.
\end{gather*}
\item Boundary case: the number $|{\cal{A}}_{N}((n^k))|$ for $N=n+k$,\footnote{So far as we know, the third equality has been proved for the f\/irst
time in~\cite{CV}. The both sides of the third identity have a big variety of
combinatorial interpretations such as the number of $k$-tuples of noncrossing
Dyck paths; that of $k$-tringulations of a convex $(n+k+1)$-gon; that of semistandard Young tableaux with entries from the set $\{1,\ldots,n\}$ having only
columns of an even length and bounded by height~$2 k$~\cite{CV}; that of
pipe dreams ({\it or} compatible sequences) associated with the Richardson
permutation $1^{k} \times w_{0}^{(n)} \in {\mathbb{S}}_{n+k}$, etc.,
see, e.g.,~\cite[$A078920$]{SL} and the literature quoted therein. It seems an interesting task to read of\/f an alternating sign matrix of size~$(n+k) \times (n+k)$ from a given $k$-triangulation of a convex $n+k+1)$-gon.}
\begin{gather*}
\big|{\cal{A}}_{k+n}\big(\big(n^k\big)\big)\big| =
\big|{\cal{A}}_{k+n}\big(\big((n+1)^k\big)\big)| = \det|\operatorname{Cat}_{n+i+j-1}|_{1 \le i, j \le k}=
\!\!\prod_{1 \le i \le j \le n} \!\!\frac{i+j+2 k}{i+j}.
\end{gather*}
\item More generally $($A.K.$)$,
\begin{gather*}
\big|{\cal{A}}_{N}\big(\big(n^k\big)\big)| = \prod_{1 \le i \le k,\, 1 \le j \le n \atop j-i \le n-k }\frac{N-i-j+1}{i+j-1} {\prod_{1 \le i \le k,\, 1 \le j \le n \atop j-i > n-k} \frac{N+i+j-1}{i+j-1}}.
\end{gather*}
Moreover, the number $|{\cal{A}}_{N}((n^k))|$ also is equal to the number of
$k$-tuples of noncrossing Dyck paths staring from the point $(0,0)$ and ending at the point $(N, N-n-k)$.\footnote{So far as we know, for the case $k=2$ an equivalent formula for the number of pairs of noncrossing Dyck paths connecting the points~$(0,0)$ and $(N,N-n-2)$, has been obtained for the f\/irst time in~\cite{Gou}.}
\item There exists a bijection
$\rho_n \colon {\cal A}_n \longrightarrow \operatorname{ASM}(n)$ such that the image
$\operatorname{Im} \big({\cal A}^{(0)}_n\big)$ contains the set of $n \times n$ permutation matrices.
\item The number of row strict $($as well as column strict$)$
diagrams $\lambda \subset \delta_{n+1}$ is equal to~$2^n$.
\end{itemize}
\end{lem}
Recall that a row-strict diagram\footnote{Known also as a strict partition.}
$\lambda$ is on such that $\lambda_i - \lambda_{i+1} \ge 1$, $\forall\, i$;
$\delta_{n}:= (n-1$, $n-2,\ldots,2,1)$.
\begin{ex}\label{exam5.4} Take $n= 5$ so that $\operatorname{ASM}(5)=429$ and $\operatorname{Cat}(5)=42$. One has
\begin{gather*}
\big|{\cal A}^{(0)}_5\big|= \sum_{\lambda \subset \delta_4:=(4,3,2,1) \atop \lambda_i-\lambda_{i+1} \ge 1\, \forall\, i} \big|{\cal{A}}_{5}^{(0)}(\lambda)\big| =
\big|{\cal A}^{(0)}_5(\varnothing)\big|+\big|{\cal A}^{(0)}_5((1))\big|+
\big|{\cal A}^{(0)}_5((2))\big|+\big|{\cal A}^{(0)}_5((3))\big|\\
\hphantom{\big|{\cal A}^{(0)}_5\big|=}{}
+\big|{\cal A}^{(0)}_5((2,1))\big|+
\big|{\cal A}^{(0)}_5((4))\big|+\big|{\cal A}^{(0)}_5((3,1))\big|+\big|{\cal A}^{(0)}_5((3,2))\big|+
\big|{\cal A}^{(0)}_5((4,1))\big|\\
\hphantom{\big|{\cal A}^{(0)}_5\big|=}{}
+\big|{\cal A}^{(0)}_5((4,2))\big|+\big|{\cal A}^{(0)}_5((3,2,1))\big|+\big|{\cal A}^{(0)}_5((4,3))\big|+\big|{\cal A}^{(0)}_5((4,2,1))\big|\\
\hphantom{\big|{\cal A}^{(0)}_5\big|=}{}
+
\big|{\cal A}^{(0)}_5((4,3,1))\big|+\big|{\cal A}^{(0)}_5((4,3,2))\big|+
\big|{\cal A}^{(0)}_5((4,3,2,1))\big| \\
\hphantom{\big|{\cal A}^{(0)}_5\big|}{}
=
1+4+9+14+6+14+16+4+21+14+4+1+9+2+1+1 =121 \\
\hphantom{\big|{\cal A}^{(0)}_5\big|=}{}\text{(sum of $16$ terms)},\\
\sum_{k=0}^{4}|{\cal A}_{5}((k))| =1 +4+ 9+ 14 +14 = 42.
\end{gather*}
\end{ex}
We expect that the image $\rho_n \big( \bigcup_{k=0}^{n-1} {\cal A}_{n}((k))\big)$ coincides with the set of $n \times n$ permutation matrices corresponding to either $321$-avoiding (or 132-avoiding) permutations.
Now we are going to def\/ine a statistic $n(T)$ on the set ${\cal A}_n$.
\begin{de}\label{def5.5} Let $\lambda$ be a partition, $\alpha$ be a composition of the
same size. For each tableau
$T \in {\widetilde {\operatorname{STY}}}(\lambda,\alpha) \subset {\cal A}_n(\lambda)$ def\/ine
\begin{gather*}
n(T)=\alpha_n= \# | \{(i,j) \in \lambda | T(i,j)=n \}|.
\end{gather*}
Clearly, $n(T) \le \lambda_1$.
Def\/ine polynomials
\begin{gather*}
{\cal A}_{\lambda}(t):=
\sum_{T \in {\cal A}_n(\lambda)}t^{\lambda_1-n(T)}.
\end{gather*}
\end{de}
It is instructive to display the numbers
$\{{\cal A}_n(\lambda), \lambda \subset \delta_n \}$ as a vector of the
length equals to the $n$-th Catalan number. For example,
\begin{gather*}
\begin{split}
& {\cal A}_4(\varnothing{,}(1),(2),(1,1),(3),(2,1),(1,1,1),(3,1),(2,2),(2,1,1),
(3,2),(3,1,1),(2,2,1),(3,2,1))\!\\
& \qquad{} =
(1,3,5,3,5,6,1,6,3,2,3,2,1,1).
\end{split}
\end{gather*}
It is easy to see that the above data, as well as the corresponding data for
$n=5$, coincide with
the list of ref\/ined totally symmetric self-complementary plane partitions that
f\/it in the box $2n \times 2n \times 2n$ ($\operatorname{TSSCPP}(n)$ for short)
listed for $n=1,2,3,4,5$ in \cite[Appendix~D]{DZ1}.
In fact we have
\begin{Theorem}\label{theorem5.6} The sequence $\{{\cal A}_n(\lambda),
\lambda \subset \delta_n \}$ coincides with the set of refined
$\operatorname{TSSCPP}(n)$ numbers
as defined in~{\rm \cite{DZ1}}. More precisely,
\begin{itemize}\itemsep=0pt
\item $|{\cal A}_n(\lambda;N)|=
\det \big|{N-i \choose \lambda'_{j}-j+i} \big|_{1 \le i,j \le
\ell(\lambda)}$,
\item we have
\begin{gather*}
{\cal A}_{\lambda}(N;t):=
\det \left|{N-i-1 \choose \lambda'_{j}-j+i-1}+t {N-i-1 \choose
\lambda'_{j}-j+i} \right|_{1 \le i,j \le \ell(\lambda)},
\end{gather*}
\item let $\lambda$ be a partition, $|\lambda|=n$, consider the column
multi-Schur polynomial and $t$-deformation thereof
\begin{gather*} s_{\lambda}^{*}(X_N):= \det |
e_{\lambda'_{j} -j+i}(X_{N-i}) |_{1 \le i,j \le \ell(\lambda)}, \qquad \text{cf.
{\rm \cite[Chapter~III]{Ma}, \cite{Wa}}}, \quad \text{and}
\\
s_{\lambda}^{*}(X_N;t):= \det |x_{N-i} e_{\lambda'_{j}-j+i-1}(X_{N-i-1})+t e_{\lambda'_{j}-j+i}(X_{N-i-1}) |_{1 \le i,j \le \ell(\lambda)},
\end{gather*}
then, assuming that $\lambda \subset \delta_n$ and $N \ge \lambda_1+\lambda'_{\ell(\lambda)}$, the
polynomial $s_{\lambda}^{*}(X_N)$ has nonnegative $($integer$)$
coefficients,
\item polynomial ${\cal A}_{\lambda}(t)$ is equal to a $t$-analog of
refined $\operatorname{TSSCPP}(n)$ numbers
$P_n(\lambda'_{n-1}+1,\dots$, $\lambda'_{n-i}+i,\dots,\lambda'_1+n-1 | t)$
introduced by means of recurrence relations in {\rm
\cite[relation~(3.5)]{DZ1}},
\item one has
\begin{gather*}
s_{((n^{k}))}^{*}(X_{n+k}) = M_{n,k}(X_{n+k}) {\mathfrak S}_{1^{k} \times
w_{0}^{(n)}}
\big(x_{1}^{-1},\ldots,x_{k+n}^{-1}\big),
\end{gather*}
where ${\mathfrak S}_{{1^{k} \times w_{0}^{(n)}}}(X_{n+k})$ denotes the Schubert
polynomial corresponding to the permutation
\begin{gather*}
1^k \times w_{0}^{(n)}= [1,2,\ldots,k, n+k,n+k-1,\ldots,k+1],
\end{gather*}
and $ M_{n,k}(X_{n+k}) =
\prod\limits_{a=1}^{k}x_{a}^{-1} \prod\limits_{a=1}^{n+k}~x_{a}^{\min(n+k-a+1,n)}$.
\end{itemize}
\end{Theorem}
In particular, $ \sum\limits_{\lambda \subset \delta_n} {\cal A}_{\lambda}(t) =
\sum\limits_{1 \le j \le n-1} A_{n,j} t^{j-1}$, where $A_{n,j}$ stands for the
number of alternating sign matrices ($\operatorname{ASM}_n$ for short) of size $n \times n$
with a~1 on top of the $j$-th column.
\begin{cor}[\cite{L1, LS1}]\label{cor5.7} The number of different tableau subwords in the word
\begin{gather*}
w_0:= \prod _{j=1}^{n-1} \left\{\prod_{a=n-1}^{j} a \right\}
\end{gather*}
is equal to the number of alternating sign matrices of size $n \times n$, i.e.,
\begin{gather*}
|{\cal A}_n| = |\operatorname{TSSCPP}(n)|=|\operatorname{ASM}_n|.
\end{gather*}
\end{cor}
It is well-known~\cite{B} that
\begin{gather*}
A_{n,j}={n+j-2 \choose j-1} {(2n-j-1)! \over (n-j)!} \prod_{i=0}^{n-2}
{(3i+1)! \over (n+i)!},
\end{gather*}
and the total number $A_n$ of $\operatorname{ASM}$ of size $n \times n$ is equal to
\begin{gather*}
A_n \equiv A_{n+1,1}=\sum_{j=1}^{n} A_{n,j}=
\prod_{i=0}^{n-1}{(3i+1)! \over (n+i)!}.
\end{gather*}
\begin{Theorem}[the case $\lambda =\delta_n:=(n-1,n-2,\ldots,2,1)$]\label{theorem 5.A}\quad
\begin{itemize}\itemsep=0pt
\item One has
\begin{gather*}
{\cal A}_{\delta_{n}}(n+1;t) = \prod_{j=2}^{n}~(1+ j t).
\end{gather*}
\item Gandhi--Dumont polynomials $($see {\rm \cite{Du}} and {\rm \cite[$A036970$]{SL})},
\begin{gather*}
{\cal A}_{\delta_{n}}(n+2;t) = \sum_{k=2}^{n} B_{n,k} t^k,
\qquad
B_{n,k}= \sum_{\{k_{j}\}} \prod_{j=2}^{n-1} {2 j- k_{j-1} \choose 2 j-
k_{j}},
\end{gather*}
where the sum runs over set of sequences $ \{1 \le k_{1} < k_2 < \cdots <
k_{n-2} < k_{n-1}= 2 n -k \}$.
\item In particular,
\begin{gather*}
{\cal A}_{\delta_{n}}(n+2;0) = G_{2n}, \qquad {\cal A}_{\delta_{n}}(n+2;1)=
G_{2n+2},
\end{gather*}
where $G_{2n} = 2 (2^{2n} -1) B_{2n}$ and $B_{2n}$ denotes the unsigned
Genocchi numbers\footnote{Recall that the unsigned Genocchi numbers are defined through
the
generting function
\begin{gather*}
\frac{2 t}{e^t +1} =\sum_{n \ge 1} \frac{G_{2n}}{(2n) !}(- 1)^n t^{2n},
\end{gather*}
see, e.g., \cite{Du}, or \url{https://en.wikipedia.org/wiki/Genocchi_number}.},
and $B_{2n}$ denotes the Bernoulli number, see, e.g.,
{\rm \cite[$A027642$]{SL}}.
\item ${\cal A}_{\delta_{n}}(n+2; -1) = (-1)^n$.
\item Let ${\cal A}_{\delta_{n}}(N;t,q)$ denote the principal
specialization $x_i:=q^{i-1}$, $i \ge 1$, of the polynomial
$s_{\delta_{n}}^{*}
(X_{N};t)$, and write ${\cal A}_{\delta_{n}}(n+2;t,q)= \sum\limits_{k=2}^{n-1}
B_{n,k}(q) t^k$. Then $B_{n,k}(q) \in \mathbb{N}[q]$.
\end{itemize}
\end{Theorem}
For example, ${\cal A}_{\delta_{6}}(8;t)=
(2073,8146,12840,10248,4200,720)_{t}$, $2073 = G_{12}$, ${\cal
A}_{\delta_{6}}(8;1)= 38227 =G_{14}$, cf.~\cite[$A036970$]{SL};
${\cal A}_{(2,1)}(5;t,q)= q {4 \brack 2}_{q} +t (q+q^2)^3 + t^2 q^5
{3 \brack 1}_{q}$. The last example shows that the polynomials ${\cal
A}_{(\delta_n)}(n+2;t,q)$ give rise to a~$q$-de\-formation of the
polynomials $p_{n+2}(t;\delta_n)$ associated with ref\/ined {\rm
${\rm TSSCPP}(n)$} introduced in~\cite{DZ1} which appeared to be coincide $($A.K.$)$
with the Gandhi polynomials introduced, e.g., in~\cite{Du}. However, the
polynomials ${\cal A}_{(\delta_n)}(n+2;t,q)$ are
dif\/ferent from a $q$-deformation of Gandhi polynomials defined in~\cite{HZ}.
It is easy to see that ${\cal{A}}_{\delta_{n}}(X_{n+2};t)=
{\cal{A}}_{\delta_{n}}^{(1)}(X_{n};t)+ x_{n+1}
{\cal{A}}_{\delta_{n}}^{(2)}(X_{n};t)$ for some polynomials depending on
the variables $\{x_1,\ldots,x_n\}$ with nonnegative integer coef\/f\/icients.
\begin{Theorem}[the Genocchi numbers of the second kind, \protect{\cite[$A005439$]{SL}}]\label{theorem 5.B} \quad
\begin{itemize}\itemsep=0pt
\item ${\cal{A}}_{\delta_n}^{(1)}(t;\, x_i=1,\, \forall \,i \in [1,n])$ is a
polynomial of the following form
\begin{gather*}
{\cal{A}}_{\delta_n}^{(1)}(t;\, x_i=1,\, \forall \, i \in [1,n])= t\/
G_{n-2}^{(2)}+\cdots+ (n-1) ! t^{n-2}, \\
{\cal{A}}_{\delta_n}^{(1)}(t=1;\, x_i=1, \, \forall\, i \in [1,n])=
G_{n-2}^{(2)},
\end{gather*}
where $G_{n}^{(2)}$ stands for the $n$-th Genocchi number of the second
kind.
It is well known that $G_{n}^{(2)}= 2^{n-1} G_{n}^{(m)}$, where
$G_n^{(m)}$
denotes the so-called $n$-th median Genocchi number~{\rm \cite[$A000366$]{SL}}.
\item ${\cal{A}}_{\delta_{n}}^{(2)}(t=0;\,x_i=1,\, \forall\, i \in [1,n])=
G_{n-1}$,
where as before, $G_{n}$ denotes the $n$-th Genocchi number $($of the
first kind$)$ {\rm \cite[$A036970$]{SL}}.
\end{itemize}
\end{Theorem}
\begin{ex}\label{example5.10}\quad
\begin{enumerate}\itemsep=0pt
\item[(1)] Take $\lambda= \delta_3$ and $n=5$, then
\begin{gather*}
{\cal{A}}_{\delta_{3}}(X_4;t)=t(x_1+x_2)\big(x_3^2+t (x_1 x_2 +x_1 x_3+x_2 x_3)\big)\\
\hphantom{{\cal{A}}_{\delta_{3}}(X_4;t)=}{}
+x_4(x_1 x_2+x_1 x_3+x_3 x_3 +t(x_1+x_1)(x_1+x_2+x_3)).
\end{gather*}
Therefore, ${\cal{A}}_{\delta_{3}}(x_i=1, \,\forall \,i \in [1,3];t) = 2 t+6
t^2 +x_4( 3+6 t)$.
\item[(2)] Take $\lambda= \delta_4$ and $n=6$, then
\begin{gather*}
{\cal{A}}_{\delta_{4}}(x_i=1,\, \forall\, i \in [1,4];t)= 8 t\big(1+3 t+3
t^2\big)+ x_5 \big( 17+ 46 t +36 t^2\big).
\end{gather*}
\item[(3)] Take $\lambda= \delta_5$ and $n=7$, then
\begin{gather*}
{\cal{A}}_{\delta_{5}}(x_i=1,\, \forall \, i \in [1,5];t)= t(56,192,240,120)_{t}+
5 x_6 (31,100,114,48)_{t}.
\end{gather*}
\end{enumerate}
\end{ex}
Therefore, the polynomials ${\cal{A}}_{\delta_{n}}(X_{n+1};t)$ and
${\cal{A}}_{\delta_{n}}^{(1)}(X_{n};t)$ def\/ine multi-parameter
deformations of the Genocchi numbers of the f\/irst and the second types
correspondingly. It is an
interesting task to relate these polynomials with those have been
studied in~\cite{HZ}, if so.
\begin{prb} Give combinatorial interpretations of polynomials ${\cal
A}_{p \delta_{n}}\!(N{;}t{,}q)$ and ${\cal A}_{(n^k)}\!(N{;}t{,}q)$ for all $N \ge p n -1$.
\end{prb}
Let as before $\operatorname{STY}(\delta_n \le n):={\cal ST}_n$ denotes
the set of all semistandard Young tableaux of the staircase shape
$\delta_n=(n-1,n-2,\ldots,2,1)$ f\/illed by the numbers from the set
$\{1,\ldots,n \}$. Denote by ${\cal ST}_n^{(0)}$ the subset of ``anti-diagonally'' increasing tableaux, i.e.,
\begin{gather*}
{\cal ST}_n^{(0)} = \{T \in \operatorname{STY}(\delta_n, \le n) \,|\, T_{i,j}
\ge T_{i-1,j+1} \, \text{for all}\, 2 \le i \le n-1, \,1 \le j \le n-2 \}.
\end{gather*}
One (A.K.) can construct bijections
\begin{gather*}
\iota_n \colon \ {\cal S}_n \sim {\cal ST}_n, \qquad \zeta_n \colon \ {\cal A}_n \sim {\cal ST}_n^{(0)}
\end{gather*}
such that $\operatorname{Im}(\iota_n)=\operatorname{Im}(\zeta_n)$.
\begin{pr}\label{prop5.9}
\begin{gather*}
\sum_{\lambda=(\lambda_1,\ldots,\lambda_n) \atop \rho_n \ge \lambda}
K_{\rho_{n}, \lambda}
\begin{pmatrix}
& n \\
m_0(\lambda), & m_1(\lambda), & \ldots, & m_{n}(\lambda)
\end{pmatrix} = 2^{{n \choose 2}},
\\
\sum_{\lambda=(\lambda_1,\ldots,\lambda_n) \atop \rho_n \ge \lambda}
\begin{pmatrix}
& n \\
m_0(\lambda), & m_1(\lambda), & \ldots, & m_n(\lambda)
\end{pmatrix} = {\cal{F}}_n,
\end{gather*}
where ${\cal {F}}_n$ denotes the number of forests of trees on $n$ labeled
nodes;
$K_{\rho_{n},\lambda}$ denotes the Kostka number, i.e., the number of
semistandard Young tableaux of the shape $\rho_{n}:= (n-1,n-2,$ $\ldots,1)$ and
content/weight~$\lambda$;
for any partition $\lambda =(\lambda_1 \ge \lambda_2 \ge \cdots \ge \lambda_n \ge 0)$ we set \mbox{$m_i(\lambda) = \{j \,|\, \lambda_j =i\}$}.
\end{pr}
Let $\alpha$ be a composition, we denote by~$\alpha^{+}$ the partition
obtained from $\alpha$ by reordering of its parts. For example, if
$\alpha= (0,2,0,3,1,0)$ then $\alpha^{+}= (3,2,1)$. Note that $\ell(\alpha)=6$,
but~$\ell(\alpha^{+})=3$.
Now let $\alpha$ be a composition such that $\rho_{n} \ge \alpha^{+}$,
$\ell(\alpha) \le n$, that is $\alpha_j = 0$, if $j > \ell(\alpha)$,
$|\alpha| = {n \choose 2}$ and
\begin{gather*}
\sum_{k \le j} {(\rho_n)}_k \ge \sum_{k \le j} (\alpha^{+})_k, \qquad \forall \, j.
\end{gather*}
There is a unique semistandard Young tableau $T_n(\alpha)$ of shape $\rho_n$
and content~$\alpha$ which corresponds to the maximal conf\/iguration of type
$(\rho_n;\alpha)$ and has all quantum numbers (riggings) equal to zero. It
follows from Proposition~\ref{prop5.9} that $\# \{\alpha \,|\, \ell(\alpha) \le n, \, \rho_n
\ge \alpha^{+} \} = {\cal {F}}_n$. Therefore there is a~natural embedding of the set of forests on $n$ labeled nodes to the set of
semistandard Young tableaux of shape $\rho_n$ f\/illed by the numbers from the
set $[1,\ldots,n]$. We denote by ${\cal {FT}}_n \subset
\operatorname{STY}( \rho_n, \le n)$ the subset $ \{T_n(\alpha) \,|\, \rho_n \ge \alpha^{+}, \, \ell(\alpha) \le n \}$. Note that the set ${\cal{K}}_n:= \{\alpha \,|\, \ell(\alpha)=
n,\, (\alpha)^{+} = \rho_n \}$ contains $n !$ compositions, and under the
rigged conf\/iguration bijection the elements of the set ${\cal{K}}_n$
correspond to the {\it key} tableaux~\cite{LS3} of shape~$\rho_n$. See also~\cite{Av} for connections of the Lascoux--Sch\"{u}tzenberger \textit{keys} and~$\operatorname{ASM}$.
Let us say a few words about the Kostka numbers $K_{\rho_{n}, \alpha}$. First
of all, it's clear that if $\alpha =(\alpha_1,\alpha_2,\dots)$ is a~composition such that $\alpha_1=n-1$, then $K_{\rho_n,\alpha}=
K_{\rho_{n-1},\alpha[1]}$, where we set $\alpha[1]:=(\alpha_2,\ldots)$.
Now assume that $n=2k+1$ is an odd integer, and consider partitions
$\nu_n:=(k^n)$ and $\mu_n:= ((k+1)^{k},k^{k})$. Then
\begin{gather*}
K_{\rho_n,\nu_n} = \operatorname{Coef\/f}_{(x_{1} x_{2} \cdots x_{n})^k} \left( \prod_{1 \le i < j \le n} (x_i+x_j) \right),
\qquad {2k \choose k} K_{\rho_n,\mu_n} =K_{\rho_n,\nu_n}.
\end{gather*}
It is well-known that the number $K_{\rho_n,\nu_n}$ is equal to number of
labeled regular tournaments with $n:=2k+1$ nodes, see, e.g.,~\cite[$A007079$]{SL}.
In the case when $n=2k$ is an even number, one can show that
\begin{gather*}
K_{\rho_n,\nu_n} = K_{\rho_{n-1},{\nu_{n-1}}}, \qquad K_{\rho_n,\mu_n} =
K_{\rho_{n+1},{\mu_{n+1}}}.
\end{gather*}
Note that the rigged conf\/iguration bijection gives rise to an embedding of
the set of labeled regular tournaments with $n:=2k+1$ nodes
\begin{gather*}
\text{to the set $\operatorname{STY}(\rho_n, \le n)$, if $n$ is an odd integer, and} \\
\text{to the set $\operatorname{STY}(\rho_{n-1}, \le n-1)$, if $n$ is even integer}.
\end{gather*}
\begin{Theorem}\label{theorem5.10} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] In the plactic algebra ${\cal P}_{n}$ the Cauchy kernel
has the following decomposition
\begin{gather*
{\cal C}_n(\mathfrak{P},U) = \sum_{T \in {\cal A}_n}
{\cal K}_{T}(\mathfrak{P}) u_{w(T)}.
\end{gather*}
\item[$(2)$] Let $T \in {\cal A}_n$, and $\alpha(T)$ be its bottom code. Then
\begin{gather*}
{\cal K}_{T}(\mathfrak{P}) -\prod_{(i,j) \in T} p_{\{i,T(i,j)-j+1\}} \ge 0,
\end{gather*}
and equality holds if and only if the bottom code~$\alpha(T)$ is a partition.
\end{enumerate}
\end{Theorem}
Note that the number of {\it different shapes} among the tableaux in the set
${\cal A}_n$ is equal to the Catalan number
$C_n:={1 \over {n+1}} {2n \choose n}$.
\begin{prb}\label{prob5.11} Construct a bijection between the set ${\cal A}_n$ and the set
of alternating sign matrices $\operatorname{ASM}_n$.
\end{prb}
\begin{ex}\label{exam5.12} For $n=4$ one has
\begin{gather*}
{\cal C}_{4}(X,U)=
K[0]+
K[1]u_1+
K[01]u_2+
K[001]u_3+
K[11](u_{12}+u_{22})+
K[2](u_{21}+u_{31})\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[101]u_{13}+
K[02]u_{32}+
K[011](u_{23}+u_{33})+
K[3]u_{321}+
K[12](u_{312}+u_{322})\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[21]u_{212}+
K[111](u_{123}+u_{133}+u_{233}+u_{223}+u_{333})+
K[021]u_{323}\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[201](u_{313}+u_{213})+
K[31]u_{3212}+
K[301]u_{3213}\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[22](u_{3132}+u_{2132}+u_{3232})+
K[121](u_{3123}+u_{3233}+u_{3223})\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[211](u_{2123}+u_{2133}+u_{3133})+
K[32]u_{32132}+
K[311](u_{32123}+u_{32133})\\
\hphantom{{\cal C}_{4}(X,U)= }{}
+
K[221](u_{21323}+u_{31323}+u_{32323})+
K[321]u_{321323}.
\end{gather*}
\end{ex}
Let $w \in \mathbb{S}_n$ be a permutation with the Lehmer code $\alpha(w)$.
\begin{de}\label{def5.13} Def\/ine the {\it plactic polynomial}
${\cal {PL}}_{w}(U)$ to be
\begin{gather*
{\cal {PL}}_{w}(U) = \biggl \{\sum_{T \in {\cal A}_n,\, \alpha(T)=
\alpha(w)} u_{w(T)} \biggr \}.
\end{gather*}
\end{de}
\begin{comm}\label{com5.14} It is easily seen from a def\/inition of the Cauchy kernel
that
\begin{gather*
{\cal{C}}_{n}(X,U)= \sum _{\alpha \subset \delta_{n}} K[\alpha](X)
{\cal {PL}}_{w_{0} w_{\alpha}^{-1}} (U),
\end{gather*}
where $w_{\alpha}$ denotes a unique permutation in $\mathbb{S}_{n}$ with the
Lehmer code equals $\alpha$; $K[\alpha](X)$ denotes the key polynomial corresponding to composition $\alpha \subset \delta_n$. The polynomials ${\cal {PL}}_{w_{0} w_{\alpha}^{-1}}$ can be treated as a~plactic version of noncommutative Schur and
Schubert polynomials introduced and studied in \cite{FG,Kn,AL,LS3,Le}.
Now let $X=\{x_1,\ldots,x_n \}$ be a set of mutually commuting variables, and
\begin{gather*}
I_{0}^{(n)}:= \big\{\underbrace{n-1,n-2,\ldots,2,1}_{n-1},
\ldots, \underbrace{n-1,,n-2,\ldots,k+1,k}_{n-k},\ldots, \underbrace{n-1,n-2}_{2},n-1 \big\}
\end{gather*} be
lexicographically maximal reduced expression for the longest element $w_{0}
\in \mathbb{S}_n$. Let $I$ be a~tableau subword of the set $I_{0}:=I_{0}^{(n)}$. One can show (A.K.) that under the specialization
\begin{gather*}
u_i= \begin{cases}x_i,&\text{if} \ \ i \in I_{0} {\setminus} I, \\
1,& \text{if} \ \ i \in I
\end{cases}
\end{gather*}
the polynomial ${\cal {PL}}_{w_{0} w_{\alpha}^{-1}} (U)$ turns into the
Schubert polynomial ${\mathfrak S}_{w_{\alpha}}(X)$. In a similar fashion, consider
the decomposition
\begin{gather*
{\cal{C}}_{n}(X,U)= \sum _{\alpha \subset \delta_{n}} \operatorname{KG}[\alpha](X;-\beta)
{\cal {PL}}_{w_{0} w_{\alpha}^{-1}} (U; \beta).
\end{gather*}
One can show (A.K.) that under the same specialization as has been listed above, the polynomial ${\cal {PL}}_{w_{0} w_{\alpha}^{-1}} (U; \beta)$ turns into
the $\beta$-Grothendieck polynomial ${\cal{G}}_{w_{\alpha}}^{\beta}(X)$.
\end{comm}
\begin{de}\label{def5.15} Def\/ine algebra ${\cal{PC}}_n$ to be the quotient of the plactic algebra ${\cal{P}}_n$ by the two-sided ideal $J_n$ by the set of
monomials
\begin{gather*
\{u_{i_{1}} u_{i_{2}} \cdots u_{i_{n}} \}, \qquad 1 \le i_{1} \le i_{2} \le
\cdots \le i_n \le n, \qquad \#\{a \vert i_a=j\} \le j, \qquad \forall\, j=1,\ldots,n.
\end{gather*}
\end{de}
\begin{Theorem}\label{theorem5.16} {\samepage \quad
\begin{itemize}\itemsep=0pt
\item The algebra ${\cal{PC}}_n$ has dimension equals to $\operatorname{ASM}(n)$,
\item $\operatorname{Hilb}({\cal{PC}}_n,q)= \sum\limits_{\lambda \in \delta_{n-1}} |{\cal{A}}_{n}(\lambda)| q^{|\lambda|}$,
\item $\operatorname{Hilb}(({\cal{PC}}_{n+1})^{ab},q)= \sum\limits_{k=0}^{n} \frac{n-k+1}{n+1} {n+k \choose n} q^k$, cf.~{\rm \cite[$A009766$]{SL}}.
\end{itemize}}
\end{Theorem}
\begin{de}\label{def5.17} Denote by ${\cal{PC}}_n^{\sharp}$ the quotient of the algebra
${\cal{PC}}_n$ by the two-sided ideal ge\-ne\-rated
by the elements $\{u_i u_j -u_j u_i,\, |i-j| \ge 2 \}$.
\end{de}
\begin{pr}\label{prop5.18}
Dimension $\dim {\cal{PC}}_n^{\sharp}$ of the algebra ${\cal{PC}}_n^{\sharp}$
is equal to the number of Dyck paths whose ascent lengths are exactly
$\{1,2,\ldots,n+1 \}$.
\end{pr}
See \cite[$A107876$, $A107877$]{SL} where the f\/irst few of these numbers are
displayed.
\begin{ex}\label{exam5.19}
\begin{gather*}
\operatorname{Hilb}\big({\cal{PC}}_5^{\sharp},t\big) =(1,4,12,27,48,56,54,38,20,7,1)_{t} ,\qquad
\dim {\cal{PC}}_5^{\sharp} = 268,\\
\operatorname{Hilb}\big({\cal{PC}}_6^{\sharp},t\big) =
(1,5,18,50,116,221,321,398,414,368,275,175,89,35,9,1)_{t},\\
\dim {\cal{PC}}_6^{\sharp} = 2496 , \qquad
\dim {\cal{PC}}_7^{\sharp} = 28612.
\end{gather*}
\end{ex}
\begin{ex}\label{exam5.20}
\begin{gather*}
\operatorname{Hilb}({\cal{PC}}_3,q)=(1,2,3,1)_{q} , \qquad
\operatorname{Hilb}({\cal{PC}}_4,q)=(1,3,8,12,11,6,1)_{q}, \\
\operatorname{Hilb}({\cal{PC}}_5,q)=(1,4,15,35,69,91,98,70,35,10,1)_{q}, \\
\operatorname{Hilb}({\cal{PC}}_6,q)=(1,5,24,74,204,435,783,1144,1379,1346,1037,628,275,85,15,1)_{q}, \\
\operatorname{Hilb}({\cal{PC}}_7,q)= (1, 6, 35, 133, 461, 1281, 3196, 6686,
12472, 19804, 27811, 33271, 34685,\\
\hphantom{\operatorname{Hilb}({\cal{PC}}_7,q)=(}{} 30527, 22864, 14124, 7126,2828,840,
175, 21, 1)_{q}.
\end{gather*}
\end{ex}
\begin{prb}\label{prob5.21} Denote by ${\mathfrak{A}}_n$ the algebra generated by the
curvature of $2$-forms of the tautolo\-gi\-cal Hermitian linear bundles
$\xi_{i}$, $1 \le i \le n$, over the flag variety~${\cal{F}}l_{n}$~{\rm \cite{SS}}.
It is well-known~{\rm \cite{PSS}} that the Hilbert polynomial of the algebra~${\mathfrak{A}}_n$ is equal to
\begin{gather*}
\operatorname{Hilb}({\mathfrak{A}}_{n},t) = \sum_{F \in {\cal{F}}_{n}} t^{\operatorname{inv}(F)} =
\sum_{F \in {\cal{F}}_{n}} t^{\operatorname{maj}(F)},
\end{gather*}
where the sum runs over the set ${\cal{F}}_{n}$ of forests $F$ on the $n$
labeled vertices, and $\operatorname{inv}(F)$ $($resp.\ $\operatorname{maj}(F))$ denotes the inversion index $($resp.\ the major index$)$ of a forest~$F$.\footnote{For the readers convenience we recall def\/initions of statistics
$\operatorname{inv}(F)$ and $\operatorname{maj}(F)$. Given a forest $F$ on $n$ labeled vertices, one can
construct a tree~$T$ by adding a new vertex (root) connected with the maximal
vertices in the connected components of~$F$.
The inversion index $\operatorname{inv}(F)$ is
equal to the number of pairs $(i,j)$ such that $1 \le i < j \le n$, and the
vertex labeled by~$j$ lies on the shortest path in $T$ from the vertex
labeled by~$i$ to the root.
The major index $\operatorname{maj}(F)$ is equal to $\sum\limits_{x \in \operatorname{Des}(F)} h(x)$; here for
any vertex $x \in F$, $h(x)$ is the size of the subtree rooted at~$x$;
the descent set $\operatorname{Des}(F)$ of $F$ consists of the vertices $x \in F$ which have
the labeling strictly greater than the labeling of its child.}
Clearly that
\begin{gather*}
\dim({\mathfrak{A}}_n)_{{n \choose 2}} =\dim({\cal{PC}}_n)_{{n \choose 2}}
= \dim(H^{\star}({\cal{F}}l_{n},\mathbb{Q}))_{{n \choose 2}} = 1.
\end{gather*}
For example,
\begin{gather*}
\operatorname{Hilb}({\cal{PC}}_6,t)=(1,5,24,74,204,435,783,1144,1379,1346,1037,628,275,85,15,1)_{t}, \\
\operatorname{Hilb}({\mathfrak{A}}_6,t)=(1,5,15,35,70,126,204,300,405,490,511,424,245,85,15,1)_{t}, \\
\operatorname{Hilb}(H^{\star}({\cal{F}}l_{n},\mathbb{Q}),t) = (1,5,14,29,49,71,90,101,101,90,71,49,29,14,5,1)_{t}.
\end{gather*}
We \textit{expect} that $\dim({\cal{PC}}_n)_{{n\choose 2}-1} =
{n \choose 2}$ and $\dim({\cal{PC}}_n)_{{n\choose 2}-2} = {\frac{3 n+5}{4}} {n+2 \choose 3} =s(n+2,2)$, where $s(n,k)$ denotes the Stirling number of the
first kind, see, e.g., {\rm \cite[$A000914$]{SL}}.
\end{prb}
\begin{prb}\quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] Is it true that $\operatorname{Hilb}({\cal{PC}}_n,t) - \operatorname{Hilb}({\mathfrak{A}}_n,t)
\in \mathbb{N} [t]$?
If so, as we \textit{expect}, does there exist an embedding of sets
$ \iota\colon {\cal{F}}(n) \hookrightarrow {\cal{A}}_n$ such that $\operatorname{inv}(F) =
n(\iota(F))$ for all $F \in {\cal{F}}_n$?
See Section~{\rm \ref{section5.1}}, Definition~{\rm \ref{def5.5}}, for definitions of the
set ${\cal{A}}_n$ and statistics $n(T)$, $T \in {\cal{A}}_n$.
\item[$(2)$] \textit{Define} a ``natural'' bijection $ \kappa\colon \operatorname{STY}(\delta_n, \le n) \longleftrightarrow 2^{\delta_{n}} $ such that the set $\kappa(\operatorname{MT}(n))$
admits a ``nice'' combinatorial description.
\end{enumerate}
\end{prb}
Here $\operatorname{MT}(n)$ denotes the set of (increasing) monotone triangles, namely, a
subset of the set $ \operatorname{STY}(\delta_n, \le n)$ consisting of tableaux
$\{T= (t_{i,j})\, |\, i+j \le n+1,\, i \ge 1, j \ge 1 \}$ such that
$t_{i,j} \ge t_{i-1,j+1}$, $2 \le i \le n$, $1 \le j < n$, cf.~\cite{ST1};
$\delta_n=(n-1,n-2,\ldots,2,1)$;
$ \operatorname{STY}(\delta_n, \le n)$ denotes the set of semistandard Young tableaux of
shape $\delta_n$ with entries bounded by $n$;
$2^{\delta_{n}}$ stands for the set of all subsets of boxes of the staircase
diagram~$\delta_{n}$.
It is well-known that $\#|\operatorname{STY}(\delta_n, \le n)| = 2^{\delta_{n}} =
2^{{n \choose 2}}$.
\begin{comm}\label{com5.22} One can ask a natural question:
when do noncommutative elementary polynomials $e_1(\mathbb{A}),\dots,
e_n(\mathbb{A})$ form a \textit{$q$-commuting family}, i.e.,
$e_i(\mathbb{A}) e_j(\mathbb{A})= q e_j(\mathbb{A}) e_{i}(\mathbb{A})$, $1 \le i < j \le n$?
Clearly in the case of two variables one needs to necessitate the
following relations
\begin{gather*}
e_i e_j e_i+e_j e_j e_i=q e_j e_i e_i+q e_j e_i e_j,\qquad i < j.
\end{gather*}
Having in mind to construct a $q$-deformation of the plactic
algebra~${\cal{P}}_n$ such that the wanted $q$-commutativity conditions are
fulf\/illed, one would be forced to add the following relations
\begin{gather*}
q e_j e_i e_j=e_j e_j e_i \qquad \text{and}\qquad q e_j e_i e_i=e_i e_i e_j e_i, \qquad i < j.
\end{gather*}
It is easily seen that these two relations are compatible if\/f $q^2=1$.
Indeed,
\begin{gather*}
e_j \underline{e_j e_i e_j} = q e_j \underline{e_i e_j e_i} = q^2 e_j e_j
e_i e_i \quad \Longrightarrow \quad q^2=1.
\end{gather*}
In the case $q=1$ one comes to the Knuth relations $({\rm PL1})$ and $({\rm PL2})$.
In the case $q= -1$ one
comes to the ``odd'' analogue of the Knuth relations, or ``odd'' plactic
relations (${\rm OPL}_n$), i.e., $({\rm OPL}_n):$
\begin{gather*}
u_j u_i u_k= - u_j u_k u_i, \qquad \text{if}\quad i < j \le k \le n, \qquad \text{and}\\
u_i u_k u_j = - u_k u_i u_j, \qquad \text{if} \quad i \le j < k \le n.
\end{gather*}
\end{comm}
\begin{pr}[A.K.] \label{prop5.23} Assume that the elements $\{u_1,\ldots,u_{n-1} \}$ satisfy the odd plactic relations $({\rm OPL}_n)$. Then the noncommutative elementary
polynomials $e_1(U),\ldots,e_n(U)$ are mutually anticommute.
\end{pr}
More generally, let ${\bf {\cal{Q}}_n}:= \{q_{ij}\}_{1 \le i < j \le n-1}$ be a
set of parameters.
Def\/ine \textit{generalized plactic algebra} ${\cal{QP}}_n$ to be (unital)
associative algebra over the ring $\mathbb{Z}[\{q_{ij}^{\pm 1} \}_{1 \le i < j \le
n-1}]$ generated by elements $u_1, \ldots, u_{n-1} $ subject to the set of
relations
\begin{gather}
q_{ik} u_{j} u_{i} u_{k} = u_j u_{k} u_i, \qquad \text{if}\quad i < j \le k, \qquad \text{and}\nonumber\\
q_{ik} u_i u_k u_j =u_k u_i u_j, \qquad \text{if}\quad i \le j < k.\label{equation5.8}
\end{gather}
\begin{pr}\label{5.24} Assume that $q_{ij}:= q_j$, $\forall\, 1 \le i < j$ be a set of invertible parameters.
Then the reduced generalized plactic algebra ${\cal{QPC}}_n$ is a~free $\mathbb{Z}[q_2^{\pm 1},\ldots,q_{n-1}^{\pm 1}]$-module of rank equals to the
number of alternating sign matrices $\operatorname{ASM}(n)$.
Moreover,
\begin{gather*}
\operatorname{Hilb}({\cal{QPC}}_n,t) = \operatorname{Hilb}({\cal{PC}}_n,t), \operatorname{Hilb}({\cal{QP}}_n,t)=
\operatorname{Hilb}({\cal{P}}_n,t).
\end{gather*}
\end{pr}
Recall that \textit{reduced generalized plactic algebra} ${\cal{QPC}}_n$ is the quotient of the generalized plactic algebra by the two-sided ideal~$J_n$
introduced in Def\/inition~\ref{def5.15}.
\begin{ex}\label{exam5.25} \quad\samepage
\begin{enumerate}\itemsep=0pt
\item[(A)] Super plactic monoid \cite{LNS, LT}. Assume that the
set of generators $U:=\{u_1,\ldots,u_{n-1} \}$ is divided on two non-crossing
subsets, say~$Y$ and~$Z$, $Y \cup Z = U, Y \cap Z =\varnothing$. To each
element $u \in U$ let us assign the weight $wt(u)$ as follows: $wt(u) =0$
if $u \in Y$, and $wt(u)=1$ if $u \in Z$. Finally, def\/ine parameters of the
generalized plactic algebra ${\cal{QP}}_n$ to be $q_{ij}= (-1)^{wt(u_{i}) wt(u_{j})}$. As a result we led to conclude that the generalized plactic
algebra ${\cal{QP}}_n$ in question coincides with the super plactic algebra
${\cal{PS}}(V)$ introduced in~\cite{LT}. We will denote this algebra by
${\cal{SP}}_{k,l}$, where $k=|Y|$, $l=|Z|$. We refer the reader to papers~\cite{LT} and~\cite{LNS} for more details about connection of the super plactic
algebra and super Young tableaux, and super analogue of the Robinson--Schensted--Knuth correspondence. We are planning to report on some properties of the
Cauchy kernel in the (reduced) super plactic algebra elsewhere.
\item[(B)] $q$-analogue of the plactic algebra.
Now let $q \not= 0, \pm 1$ be a parameter, and assume that $q_{ij}=q$, $\forall\, 1 \le i < j \le n-1$. This case has been treated recently in \cite{Li}. We
expect that the generalized Knuth relations~\eqref{equation5.8} are related
with \textit{quantum version} of the tropical/geometric RSK-correspondence (work in progress), and, as expected, with a $q$-weighted version of the
Robinson--Schensted algorithm, presented in~\cite{OP}. Another interesting
\textit{problem} is to understand a meaning of ${\cal{Q}}$-plactic
polynomials coming from the decomposition of the (plactic) Cauchy kernels
${\cal{C}}_n$ and ${\cal{F}}_n$ in the reduced generalized plactic algebra
${\cal{QPC}}_n$ (work in progress).
\item[(C)] Quantum pseudoplactic algebra $\operatorname{PPL}_{n}^{(q)}$ \cite{KT}.
By def\/inition, the quantum pseudoplactic algebra $\operatorname{PPL}_{n}{(q)}$ is an
associative algebra, generated, say over $\mathbb{Q}$, by the set of elements $\{e_1,\ldots,e_{n-1}\}$ subject to the set of def\/ining relations
\begin{gather*}
(a) \quad (1+q) e_i e_j e_i -q e_i^2 e_j - e_j e_i^2 =0, \qquad\! (1+q) e_j e_i e_j -e_j^2 e_j - q e_i e_j^2 =0, \qquad\! i < j,\\
(b) \quad (e_j,(e_i,e_k)):=e_j e_i e_k - e_j e_k e_i - e_i e_k e_j + e_k e_i e_j = 0, \qquad i < j < k.
\end{gather*}
\end{enumerate}
\end{ex}
\textit{Note} that if $q=1$, then the relations $(a)$ can be written in the
form $(e_i,(e_i,e_j))=0$ and $(e_2,(e_j,(e_i,e_j))=0$ correspondingly.
Therefore, $\operatorname{PPL}_{3}^{(q=1)}$ is the universal enveloping algebra over $\mathbb{Q}$ of the Lie algebra ${\mathfrak{sl}}^{+}_{3}$. The quotient of the algebra $\operatorname{PPL}_{n}^{q=1}$ by the two-sided ideal generated by the elements $(e_i,(e_j,e_k))$, $i$, $j$, $k$ are distinct, is isomorphic to the algebra from Remark~\ref{rem2.3}.
Def\/ine noncommutative $q$-elementary polynomials $\Lambda_{k}(q;X_n)$, cf.~\cite{KT}, as follows
\begin{gather}\label{equation5.9}
\Lambda_{k}(X;q):= \sum_{n \ge i_1 >i_2 > \cdots > i_k \ge 1} (x_{i_{1}},(x_{i_{2}},( \ldots,(x_{i_{k-1}},x_{i_{k}})_{q}) \cdots )_{q})_{q}.
\end{gather}
\begin{pr}[\cite{KT}] \label{prop5.26} The noncommutative $q$-elementary
polynomials
$ \{\!\Lambda_{k}(E_{n{-}1}{;}q)\}_{1 {\le} k {\le} n{-}1}\! \}$
are pairwise commute in the algebra $\operatorname{PPL}_{n}^{(q)}$.
\end{pr}
\subsection[Nilplactic algebra ${\cal {NP}}_n$]{Nilplactic algebra $\boldsymbol{{\cal {NP}}_n}$}\label{section5.2}
Let $\lambda$ be a partition and $\alpha$ be a composition of the same size.
Denote by ${\widehat {\operatorname{STY}}}(\lambda,\alpha)$ the set of columns and rows
strict Young tableaux $T$ of the shape $\lambda$ and content $\alpha$ such that
the corresponding tableau word~$w(T)$ is reduced, i.e., $l(w(T))=|T|$.
Denote by ${\cal B}_n$ the union of the sets ${\widehat {\operatorname{STY}}}(\lambda,\alpha)$
for all partitions $\lambda$ such that $\lambda_i \le n-i$ for $i=1,2,\dots,
n-1$, and all compositions~$\alpha$, $\alpha \subset \delta_n$.
For example, $|{\cal B}_n|=1,2,6,25,139,1008,\dots $, for
$n=1,2,3,4,5,6,\dots$.
{\samepage \begin{Theorem}\label{theorem5.27} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$] In the nilplactic algebra ${\cal {NP}}_{n}$ the Cauchy kernel
has the following decomposition
\begin{gather*
{\cal C}_n(\mathfrak{P},U) = \sum_{T \in {\cal B}_n} {\cal K}_{T}(\mathfrak{P})
u_{w(T)}.
\end{gather*}
\item[$(2)$] Let $T \in {\cal B}_n$ be a tableau, and assume that its bottom code is a~partition. Then
\begin{gather*
{\cal K}_{T}(\mathfrak{P})= \prod_{(i,j) \in T} p_{\{i,T(i,j)-j+1 \}}.
\end{gather*}
\end{enumerate}
\end{Theorem}}
\begin{ex}\label{exam5.28} For $n=4$ one has
\begin{gather*}
{\cal C}_{4}(X,U)= K[0]+K[1]u_1+K[01]u_2+K[001]u_3+K[11]u_{12}+K[2](u_{21}+u_{31})\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+K[101]u_{13}+
K[02]u_{32}+K[011]u_{23}+K[3]u_{321}+K[12]u_{312}+K[21]u_{212}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+K[111]u_{123}+K[021]u_{323}+K[201]u_{213}+K[31]u_{3212}+
K[301]u_{3213}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+K[22]u_{2132}+K[121]u_{3123}+K[211]u_{2123}+K[32]u_{32132}+
K[311]u_{32123}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+K[221]u_{21323}+K[321]u_{321323}.
\end{gather*}
\end{ex}
\subsection[Idplactic algebra ${\cal {IP}}_n$]{Idplactic algebra $\boldsymbol{{\cal {IP}}_n}$}\label{section5.3}
Let $\lambda$ be a partition and $\alpha$ be a composition of the same size.
Denote by ${\overline{\operatorname{STY}}}(\lambda,\alpha)$ the set of columns and rows
strict Young tableaux $T$ of the shape $\lambda$ and content $\alpha$ such that
$l(w(T))=\operatorname{rl}(w(T))$, i.e., the tableau word\footnote{See page~\pageref{footnote9} for the def\/inition of {\it tableau word}.}
$w(T)$ is a unique tableau word of
minimal length in the idplactic class of~$w(T)$, cf.\ Example~\ref{exam2.16}.
Denote by ${\cal D}_n$ the union of the sets
${\overline{\operatorname{STY}}}(\lambda,\alpha)$
for all partitions $\lambda$ such that $\lambda_i \le n-i$ for $i=1,2,\dots,n-1$, and all compositions $\alpha$, $l(\alpha) \le n-1$.
For example, $\# |{\cal D}_n|=1,2,6,26,154,1197,\dots $, for $n=1,2,3,4,5,6,\dots$.
\begin{Theorem}\label{theorem5.29} \quad
\begin{enumerate}\itemsep=0pt
\item[$(1)$]
In the idplactic algebra ${\cal {IP}}_{n}$ the Cauchy kernel
has the following decomposition
\begin{gather*
{\cal C}_n(X,Y,U) =
\sum_{T \in {\cal D}_n} {\cal {\operatorname{KG}}}_{T}(X,Y)
u_{w(T)}.
\end{gather*}
\item[$(2)$] Let $T \in {\cal D}_n$ be a tableau, and assume that its bottom code is a
partition. Then
\begin{gather*
{\cal {\operatorname{KG}}}_{T}(X,Y)= {\cal K}_{T}(X,Y)= \prod_{(i,j) \in T} (x_i+y_{T(i,j)-j+1}).
\end{gather*}
\end{enumerate}
\end{Theorem}
\begin{ex}\label{exam5.30} For $n=4$ one has
\begin{gather*}
{\cal C}_{4}(X,U)=
\operatorname{KG}[0]+\operatorname{KG}[1]u_1+\operatorname{KG}[01]u_2+\operatorname{KG}[001]u_3+\operatorname{KG}[11]u_{12}
+\operatorname{KG}[2](u_{21}+u_{31})\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+
\operatorname{KG}[101]u_{13}+
\operatorname{KG}[02]u_{32}+\operatorname{KG}[011]u_{23}+\operatorname{KG}[3]u_{321}+\operatorname{KG}[12]u_{312}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+\operatorname{KG}[21]u_{212}+
\operatorname{KG}[111]u_{123}+\operatorname{KG}[021]u_{323}+\operatorname{KG}[201](u_{313}+u_{213})\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+K[31]u_{3212}+
\operatorname{KG}[301]u_{3213}+\operatorname{KG}[22]u_{2132}+\operatorname{KG}[121]u_{3123}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+\operatorname{KG}[211]u_{2123}+
\operatorname{KG}[32]u_{32132}+
\operatorname{KG}[311]u_{32123}+\operatorname{KG}[221]u_{21323}\\
\hphantom{{\cal C}_{4}(X,U)=}{}
+\operatorname{KG}[321]u_{321323}.
\end{gather*}
\end{ex}
\begin{Theorem}\label{theorem5.31} For each composition~$\alpha$ the key Grothendieck
polynomial $\operatorname{KG}[\alpha](X)$ is a linear combination of key
polynomials $K[\beta](X)$ with nonnegative integer coefficients.
\end{Theorem}
\subsection[NilCoxeter algebra ${\cal {NC}}_n$]{NilCoxeter algebra $\boldsymbol{{\cal {NC}}_n}$}\label{section5.4}
\begin{Theorem}\label{theorem5.32}
In the nilCoxeter algebra ${\cal {NC}}_{n}$ the Cauchy kernel
has the following decomposition
\begin{gather*
{\cal C}_n(X,Y,U) =
\sum_{w \in {\mathbb S}_n} {{\mathfrak S}}_w(X,Y) u_{w}.
\end{gather*}
\end{Theorem}
Let $w \in {\mathbb S}_n$ be a permutation, denote by $R(w)$ the set of all its
reduced decompositions. Since the nilCoxeter algebra ${\cal {NC}}_n$ is the
quotient of the nilplactic algebra ${\cal {NP}}_n$, the set $R(w)$ is the
union of nilplactic classes of some tableau words $w(T_i)$:
$R(w)= \bigcup C(T_i)$.
Moreover, $R(w)$ consists of only one nilplactic class if and only if $w$ is
a {\it vexillary} permutation. In general case we see that the set of
compatible sequences $CR(w)$ for permutation~$w$ is the union of sets~$C(T_i)$.
\begin{cor}\label{cor5.33} Let $w \in {\mathbb S}_n$ be a permutation of length~$l$, then
\begin{enumerate}\itemsep=0pt
\item[$(1)$] ${{\mathfrak S}}_{w}(X,Y) = \sum\limits_{{\bf b} \in CR(w)} x_{b_1} \cdots x_{b_{l}}$.
\item[$(2)$] Double Schubert polynomial ${{\mathfrak S}}_{w}(X,Y)$ is a linear combination of
double key polynomials ${\cal K}_{T}(X,Y)$, $T \in {\cal B}_n, w=w(T)$,
with nonnegative integer coefficients.
\end{enumerate}
\end{cor}
\subsection[IdCoxeter algebras ${\cal {IC}}_n^{\pm}$]{IdCoxeter algebras $\boldsymbol{{\cal {IC}}_n^{\pm}}$}\label{section5.5}
\begin{Theorem}\label{theorem5.34}
In the IdCoxeter algebra ${\cal {IC}}_{n}^{+}$ with $\beta=1$,
the Cauchy kernel has the following decomposition
\begin{gather*
{\cal C}_n(X,Y,U) =
\sum_{w \in {\mathbb S}_n} {\cal G}_w(X,Y) u_{w}.
\end{gather*}
\end{Theorem}
\begin{Theorem}\label{theorem5.35}
In the IdCoxeter algebra ${\cal {IC}}_{n}^{-}$ with $\beta=-1$,
one has the following decomposition
\begin{gather*
\prod_{i=1}^{n-1} \left\{{\prod_{j=n-1}^{i} ((1+x_{i})
(1+y_{j-i+1})+(x_{i}+y_{j-i+1}) u_{j})} \right\} =
\sum_{w \in {\mathbb S}_n} {\cal H}_w(X,Y) u_{w}.
\end{gather*}
\end{Theorem}
A few remarks in order.
\begin{enumerate}\itemsep=0pt
\item[$(a)$] The (dual) Cauchy identity \eqref{equation5.9} is still valid in the idplactic
algebra with constrain $u_{i}^2 = - \beta u_i$, $i=1,\ldots, n-1$.
\item[$(b)$] The left hand side of the identity \eqref{equation5.9} can be written in the
following form
\begin{gather*
\prod_{1 \le i,j \le n \atop i+j \le n} (x_i+y_j)
\prod_{i=1}^{n-1} \left \{{\prod_{j=n-1}^{i}} \frac{1}{1-(x_{i}+ y_{j-i+1}+ \beta x_{i} y_{j-i+1}) u_{j}} \right\}.
\end{gather*}
Indeed, $(1+ \beta x +x u_{i})(1-x u_{i}) = 1+ \beta x$, since $u_{i}^2= - \beta u_{i}$.
\end{enumerate}
Let $w \in {\mathbb S}_n$ be a permutation, denote by $\operatorname{IR}(w)$ the
set of all decompositions in the idCoxeter algebra ${\cal IC}_n$ of the
element $u_w$ as the product of the generators $u_{i}$, $1 \le i \le n-1$,
of the algebra~${\cal IC}_n$.
Since the idCoxeter algebra ${\cal {IC}}_n$ is the
quotient of the idplactic algebra ${\cal {IP}}_n$, the set $\operatorname{IR}(w)$
is the union of idplactic classes of some tableau words~$w(T_i)$:
$\operatorname{IR}(w)= \bigcup \operatorname{IR}(T_i)$.
Moreover, the set of
compatible sequences $\operatorname{IC}(w)$ for permutation $w$ is the union
of sets $\operatorname{IC}(T_i)$.
\begin{cor}\label{cor5.36} Let $w \in {\mathbb S}_n$ be a permutation of length~$l$, then
\begin{enumerate}\itemsep=0pt
\item[$(1)$] ${\cal G}_{w}(X,Y) =
\sum\limits_{{\bf b} \in IC(w)}
\prod\limits_{i=1}^{l}(x_{b_i}+y_{a_i-b_i+1})$.
\item[$(2)$] Double Grothendieck polynomial ${\cal G}_{w}(X,Y)$ is a linear
combination of double key Grothen\-dieck polynomials ${\cal {\operatorname{KG}}}_{T}(X,Y)$,
$ T \in {\cal B}_n$, $w=w(T)$, with nonnegative integer coefficients.
\end{enumerate}
\end{cor}
\section[${\cal F}$-kernel and symmetric plane partitions]{$\boldsymbol{{\cal F}}$-kernel and symmetric plane partitions}\label{section6}
Let us f\/ix natural number $n$ and $k$, and a partition $\lambda
\subset (n^k)$. Clearly the
number of such partitions is equal to ${n +k \choose n}$; note that in the
case $n=k$ the number ${2 n \choose n}$ is equal to the {\it Catalan number
of type~$B_n$}.
Denote by ${\cal B}_{n,k}(\lambda)$ the set of semistandard Young tableaux of
shape $\lambda \subset (n^k)$ f\/illed by the numbers from the set $\{1,2,\ldots,n \}$. For a
tableau $T \in {\cal B}_{n,k}$ set as before,
\begin{gather*}
n(T):= \operatorname{Card} \{(i,j) \in \lambda
\mid T(i,j)=n \},
\end{gather*}
and def\/ine polynomial
\begin{gather}\label{equation6.1}
{\cal B}_{n,k}(\lambda)(q):= \sum_{T \in {\cal B}_{n,k}(\lambda)}
q^{\lambda_{1}-n(T)}.
\end{gather}
Denote by ${\cal B}_{n,k} := \bigcup_{\lambda \subset (n^{k})}
{\cal B}_{n,k}(\lambda)$.
\begin{lem}[\cite{G,KGV}]\label{lem6.1} The number of elements in the set
${\cal B}_{n,k}$ is equal to
\begin{gather*}
\# | {\cal B}_{n,k} | = \prod_{1 \le i \le j \le k} {i+j+n-1 \over i+j-1} = \prod_{0 \le 2 a \le k -1} \frac{{n+2 k- 2a -1 \choose n}}
{{n+2 a\choose n}}.
\end{gather*}
\end{lem}
See also \cite[$A073165$]{SL} for other combinatorial interpretations of the
numbers $ \# | {\cal B}_{n,k} |$. For example, the number
$ \# | {\cal B}_{n,k} |$ is equal to the number of symmetric plane partitions
that f\/it inside the box $n \times k \times k$. Note that $B_{n,k}= T(n+k,k)$, where the triangle of positive integers $\{T(n+k,k)\}$ can be found in \cite[$A102539$]{SL}.
\begin{pr}\label{prop6.2} One has
\begin{gather*}
\bullet \quad \# |{\cal B}_{n,n} | := \operatorname{SPP}(n+1)= \operatorname{TSPP}(n+1) \times \operatorname{ASM}(n),\\
\hphantom{\bullet} \quad \# |{\cal B}_{n,n+1} | = \operatorname{TSPP}(n+1) \times \operatorname{ASM}(n+1),
\end{gather*}
where $\operatorname{TSPP}(n)$ denotes the number of totally symmetric plane partitions f\/it
inside the $n \times n \times n$-box, see, e.g., {\rm \cite[$A005157$]{SL}}, whereas
$\operatorname{ASM}(n)=\operatorname{TSSCPP}(2n)$ denotes the of
$n \times n$ alternating sign matrices, and $\operatorname{TSSCPP}(2n)$ denotes the number of
totally symmetric self-complimentary plane partitions fit inside the
$2n \times 2n \times 2n$-box.
\begin{gather*}
\bullet \quad \# | {\cal B}_{n+2,n} | = \# | {\cal B}_{n,n+1} |.
\end{gather*}
\end{pr}
Note that in the case $n=k$ the number ${\cal{B}}_{n} := {\cal{B}}_{n,n}$ is
equal to the number of symmetric plane portions f\/itting inside the
$n \times n \times n$-box, see~\cite[$A049505$]{SL}. Let us point out that in
general it may happen that the number $\# |{\cal B}_{n,n+2} |$ is not
divisible by any $\operatorname{ASM}(m)$, $m \ge 3$. For example, ${\cal{B}}_{3,5}= 4224=
2^5 \times 3 \times 11$. On the other hand, it's possible that the number
$\# |{\cal B}_{n,n+2} |$ is divisible by $\operatorname{ASM}(n+1)$, but does not divisible by
$\operatorname{ASM}(n+2)$. For example, ${\cal{B}}_{4,6}=306735 = 715 \times 429$, but $
306735\, {\nmid} \, 7436 =\operatorname{ASM}(6)$.
\begin{exer}\label{exer6.3} \quad
\begin{enumerate}\itemsep=0pt
\item[$(a)$] Show that $B_{n+4,n}$ is divisible by
\begin{gather*}
\begin{cases}
\operatorname{TSPP}(n+2), & \text{if} \ \ n \equiv 1 \ (\operatorname{mod}~2), \ n \ge 3,\\
\operatorname{ASM}(n+2), & \text{if} \ \ n \equiv 2 \ (\operatorname{mod}~8),\\
\operatorname{ASM}(n+1) \ \text{and} \ \operatorname{ASM}(n+2), & \text{if} \ \ n \equiv 4 \ (\operatorname{mod}~8), \ n \neq 4; \ B_{8,4}=\operatorname{ASM}(5)^2,\\
\operatorname{ASM}(n+1) \operatorname{and} \ \operatorname{ASM}(n+2), & \text{if} \ \ n \equiv 6 \ (\operatorname{mod}~8),\\
\operatorname{ASM}(n+1), & \text{if} \ \ n \equiv 0 \ (\operatorname{mod}~8), \ n \ge 1.
\end{cases}
\end{gather*}
\item[$(b)$] Show that $B_{n,n+4}$ is divisible by
\begin{gather*}
\begin{cases}
\operatorname{ASM}(n+1), & \text{if} \ \ n \equiv 0 \ (\operatorname{mod}~2), \\
\operatorname{TSPP}(n+1), & \text{if} \ \ n \equiv 1 \ (\operatorname{mod}~2).
\end{cases}
\end{gather*}
\end{enumerate}
\end{exer}
In all cases listed in Exercise~\ref{exer6.3}, it is an open {\it problem} to give
combinatorial interpretations of the corresponding ratios.
\begin{prb}\label{prob6.4} Let $a$ is equal to either $0$ or $1$. Construct
bijection between the set $\operatorname{SPP}(n,n+a,$ $n+a)$ of symmetric plane partitions
fitting inside the box $n \times n+a \times n+a$ and the set
of pairs $(P,M)$ where $P$ is the totally symmetric plane partitions
fitting inside the box $n \times n \times n$ and $M$ is an alternating sign
matrix of size $n+a \times n+a$.
\end{prb}
\begin{ex}\label{exam6.5} Take $n=3$. One has $\# |{\cal{B}}_{3}|= 112 = 16 \times 7$. The number of partitions $\lambda \subset (3^3)$ is
equal to 20, namely, the following partitions
\begin{gather*}
\big\{\varnothing,(1),(2),(1,1),(3),(2,1),\big(1^3\big),(3,1),(2,2),\big(2,1^2\big),(3,2),\big(3,1^2\big),\big(2^2,1\big),\big(3^2\big), \\
(3,2,1),\big(2^3\big),\big(3^2,1\big),\big(3,2^2\big),\big(3^2,2\big),\big(3^3\big) \big\},
\end{gather*}
and
\begin{gather*}
\begin{split}
& {\cal{B}}_3(q):= \sum_{\lambda \subset (3^3)}
\# | {\cal{B}}_3(\lambda)| q^{|\lambda|} = (1,3,9,19,24,24,19,9,3,1)\\
& \hphantom{{\cal{B}}_3(q):= \sum_{\lambda \subset (3^3)}
\# | {\cal{B}}_3(\lambda)| q^{|\lambda|}}{} = (1+q)^3(1+q^2)\big(1+5 q^2 +q^4\big).
\end{split}
\end{gather*}
Note, however, that
\begin{gather*}
\sum_{\lambda \subset (4^4)} \# | {\cal B}_4(\lambda)| q^{|\lambda|} =
(1,4,16,44,116,204,336,420,490,420,336,204,116,44,16,4,1)
\end{gather*}
is an irreducible polynomial, but its value at $q=1$ is equal to $2772=66
\times 42$.
\end{ex}
Let ${\bf p}=(p_{i,j})_{1 \le i \le n, 1 \le j \le k}$ be a
$n \times k$ matrix of variables.
\begin{de}\label{def6.6} Def\/ine the kernel ${\cal F}_{n,k}({\bf p}, U)$ as follows
\begin{gather*
{\cal F}_{n,k}({\bf p}, U) = \prod_{i=1}^{k-1} \prod_{j=
n-1}^{1} (1+ p_{i, {\overline {j-i+1}}^{(n)}} u_j),
\end{gather*}
where for a f\/ixed $n \in \mathbb{N}$ and an integer $a \in \mathbb{Z}$, we set
\begin{gather*}
\overline{a}=\overline{a}^{(n)}: = \begin{cases}
a, & \text{if} \ \ a \ge 1,\\
n+a-1, & \text{if} \ \ a \le 0.
\end{cases}
\end{gather*}
\end{de}
For example,
\begin{gather*}
{\cal F}_{3}({\bf p},U)= (1+p_{1,2} u_2)(1+p_{1,1} u_1)(1+p_{2,1} u_2)(1+p_{2,2} u_1).
\end{gather*}
In the plactic algebra ${\cal{FP}}_{3,3}$ one has
\begin{gather*}
{\cal F}_{3,3}({\bf p},U) = 1
+ (p_{1,1}+p_{2,2}) u_1
+ (p_{1,2}+p_{2,1}) u_2
+p_{1,1} p_{2,1} u_{11}
+p_{1,1} p_{2,1} u_{12} \\
\hphantom{{\cal F}_{3,3}({\bf p},U) =}{}
+ (p_{1,2} p_{1,1}+p_{1,2} p_{2,2} +p_{2,1} p_{2,2}) u_{21}
+p_{1,2} p_{2,1} u_{22} \\
\hphantom{{\cal F}_{3,3}({\bf p},U) =}{}
+(p_{1,1} p_{1,2} p_{2,2}+ p_{1,2} p_{2,2} p_{2,1}) u_{212}
+(p_{1,1} p_{1,2} p_{2,2}+p_{1,1} p_{2,2} p_{2,1}) u_{211} \\
\hphantom{{\cal F}_{3,3}({\bf p},U) =}{}
+p_{1,1} p_{1,2} p_{2,1} p_{2,2} u_{2121}.
\end{gather*}
\begin{de}\label{def6.7} Def\/ine algebra ${\cal{PF}}_{n,k}$ to be the quotient of the plactic
algebra ${\cal{P}}_n$ by the two-sided ideal $I_n$ generated by the set of
monomials
\begin{gather*}
\{u_{i_{1}} u_{i_{2}} \cdots u_{i_{k}} \}, \qquad 1 \le i_{1} \le i_{2} \le
\cdots \le i_k \le n -1.
\end{gather*}
\end{de}
\begin{Theorem}\label{theorem6.8} \quad
\begin{gather*}
\operatorname{Hilb}({\cal{PF}}_{n,k},q)= {\cal{B}}_{n -1 ,k-1}(q),
\end{gather*}
In particular,
\begin{itemize}\itemsep=0pt
\item The algebra ${\cal{PF}}_{n,n}$ has dimension equals to the number
of symmetric plane partitions $\operatorname{SPP}(n - 1)$,
\begin{gather*
\operatorname{Hilb}({\cal{PF}}_{n,k},q) = q^{\frac{k n}{2}} {\mathfrak{so}}_{(\frac{k}{2})^n } \big(\underbrace{q^{\pm 1}, \dots, q^{\pm 1}}_{k},1\big),
\end{gather*}
where ${\mathfrak{so}}_{(\frac{k}{2})^n} \big(\underbrace{q^{\pm 1}, \dots, q^{\pm 1}}_{k},1\big)$ denotes the specialization $x_{2j}=q$, $x_{2j-1}=q^{-1}$, $1 \le j \le k$, of the character ${\mathfrak{so}}_{\lambda} \big(x_1, x_1^{-1}\ldots, x_{k},x_{k}^{-1},1\big)$ of the odd orthogonal Lie algebra
${\mathfrak{so}}(2k+1)$ corresponding to the highest weight $\lambda = \big(\underbrace{\tfrac{k}{2},\ldots,\tfrac{k}{2}}_{n}\big)$.
\item $\deg_{q} \operatorname{Hilb}({\cal{PF}}_{n,k},q) =(n-1)(k-1)$, and
$\dim ({\cal{PF}}_{n,k})_{(n-1)(k-1)}=1$.
\item The Hilbert polynomial $\operatorname{Hilb}({\cal{PF}}_{n,k},q)$ is symmetric
and unimodal polynomial in the va\-riab\-le~$q$.
\item $\operatorname{Hilb}(({\cal{PF}}_{n,k})^{ab},q) =\sum\limits_{j=0}^{k-1} {n+j-2 \choose n-2} q^{j}$, $\dim {(\cal{PF}}_{n,k})^{ab}= {n+k-2 \choose k-1}$.
\end{itemize}
\end{Theorem}
The key step in proofs of Lemma~\ref{lem6.1} and Theorem~\ref{theorem6.8} is based on the following
identity
\begin{gather*
\sum_{\lambda \subset (n^{k})} s_{\lambda} (x_1,\ldots,x_k) =
(x_1 \cdots x_k)^{n/2} {\mathfrak{so}}_{(\frac{k}{2})^n}\big(x_{1},x_{1}^{-1}, \ldots,x_{k},x_{k}^{-1},1\big),
\end{gather*}
see, e.g., \cite[Chapter~I, Section~5, Example~19]{Ma}, \cite{KGV} and the literature quoted
therein.
\begin{prb}\label{prob6.9}
Let $\Gamma:= {\Gamma_{n,m}}^{k,\ell} = (n^{k},m^{\ell})$, $n \ge m$ be a
``fat hook''. Find generalizations of the identity~\eqref{equation6.1} and those listed
in~{\rm \cite[p.~71]{Kin}}, to the case of fat hooks, namely to find ``nice''
expressions for the following sums
\begin{gather*}
\sum_{\lambda \subset \Gamma} s_{\lambda} (X_{k+\ell}),\qquad
\sum_{\lambda \subset \Gamma} s_{\lambda} (X_{k+\ell}) s_{\lambda}(Y_{k+\ell}).
\end{gather*}
Find ``bosonic'' type formulas for these sum at the limit
$n \longrightarrow \infty$, $\ell \longrightarrow \infty$, $m$, $k$ are fixed.
\end{prb}
\begin{ex}\label{exam6.10}
\begin{gather*}
\operatorname{Hilb}({\cal{PF}}_{2,3},q)= (1,3,9,9,9,3,1)_{q}, \qquad \dim ({\cal{PF}}_{2,4}) = 35 = 5 \times 7, \\
\dim ({\cal{PF}}_{2,5}) = 126 =3 \times 42, \qquad \dim {\cal{PF}}_{2,n}= {2 n+1 \choose n}= (2 n+1) \operatorname{Cat}_{n} \\
(\text{see, e.g., \cite[$A001700$]{SL}}),\\
\operatorname{Hilb}({\cal{PF}}_{3,4},q)= (1,4,16,44,81,120,140,120,81,44,16,4,1)_{q}, \\
\dim ({\cal{PF}}_{3,4}) = 672 = 16 \times 42,\\
\operatorname{Hilb}({\cal{PF}}_5,q)=(1,4,16,44,116,204,336,420,490,420,336,204,116,44,16,4,1)_{q}, \\
\dim({\cal{PF}}_{5,5}) =2772=66 \times 42.
\end{gather*}
\end{ex}
\begin{pr}\label{prop6.11}
\begin{gather*}
\operatorname{Hilb} ({\cal{PF}}_{2,n},q) = \sum_{k=0}^{2 n} {n \choose \big[\frac{k}{2} \big]} {n \choose \big[\frac{k+1}{2} \big]} q^k;
\qquad \text{recall} \qquad \dim ({\cal{PF}}_{2,n}) = {2 n + 1 \choose n}.
\end{gather*}
Therefore, $\operatorname{Hilb}({\cal{PF}}_{2,n},q)$ is equal to the generating function
for the number of symmetric Dyck paths of semilength $2 n-1$ according to the
number of peaks, see~{\rm \cite[$A088855$]{SL}},
\begin{gather*}
\dim ({\cal{PF}}_{3,n}) = 2^n \operatorname{Cat}_{n+1}, \qquad \text{if} \quad n \ge 1.
\end{gather*}
\end{pr}
For example,
\begin{gather*}
\dim ({\cal{PF}}_{3,6}) =27456 =64 \times 429,\\
\operatorname{Hilb}({\cal{PF}}_{3,6},q) = (1,6,36,146,435,1056,2066,3276,4326,4760,4326,3276,2066,1056,\\
\hphantom{\operatorname{Hilb}({\cal{PF}}_{3,6},q) = (}{} 435,146,36,6,1).
\end{gather*}
Several interesting interpretations of these numbers are given in
\cite[$A003645$]{SL}.
{\samepage \begin{Theorem}\label{theorem6.12} \quad
\begin{itemize}\itemsep=0pt
\item Symmetric plane partitions and Catalan numbers:
\begin{gather*}
\# |{\cal{B}}_{4,n}| = {\frac{1}{2}} \operatorname{Cat}_{n+1} \times \operatorname{Cat}_{n+2}.
\end{gather*}
\item Symmetric plane partitions and alternating sign matrices:
\begin{gather*}
\# |{\cal{B}}_{n+3,n}| = {\frac{1}{2}} \operatorname{TSPP}(n+1) \times \operatorname{ASM}(n+1) = \frac{1}{2} \# |{\cal{B}}_{n+1,n+1}|.
\end{gather*}
\item Plane partitions and alternating sign matrices invariant under
a half-turn:
\begin{gather*}
\# |\operatorname{PP}(n)| = \operatorname{ASM}(n) \times \operatorname{ASMHT}(2n),
\end{gather*}
where $PP(n)$ denotes the number of plane partitions fitting inside an $n \times n \times n$ box, see, e.g., {\rm \cite{B,Ku,Ma}}, {\rm \cite[$A008793$]{SL}} and the literature quoted theirin; $\operatorname{ASMHT}(2n)$ denotes the number
of alternating sign $2n \times 2n$-matrices invariant under a half-turn, see,
e.g., {\rm \cite{B, Ku,Ok,ST1}}, {\rm \cite[$A005138$]{SL}}.
\item Plactic decomposition of the ${\cal{F}}_n$-kernel:
\begin{gather}\label{equation6.5}
{\cal{F}}_{n,m}({\bf p},U) =\sum_{T} u_{T} U_{T}(\{p_{ij} \}),
\end{gather}
where summation runs over the set of semistandard Young tableaux $T$ of shape
$\lambda \subset (n)^m$ filled by the numbers from the set $\{1,\ldots, m\}$.
\item $U_{T}(\{p_{ij}=1,\, \forall \, i,j\}) =
\dim V_{\lambda'}^{{{\mathfrak{gl}}(m)}}$, where $\lambda$ denotes the
shape of a tableau~$T$, and~$\lambda'$ denotes the conjugate/transpose of a
partition $\lambda$.
\end{itemize}
\end{Theorem}}
\begin{exer}\label{exer6.13}
It is well-known~\cite{Gou} that that the number $\operatorname{Cat}_{n+1} \operatorname{Cat}_{n+2}$ counts the number~$S_{n+1}^{(4)}$ of standard Young tableaux having $2 n +1$ boxes and at most four rows.
Give a~bijective proof of the equality $\# |{\mathcal{B}}_{4,n}| = {\frac{1}{2}} S_{n+1}^{(4)}$.
\end{exer}
|
\section{Introduction}
Since cooperative relaying can improve both spectral efficiency and spatial diversity, it is a promising core technology for next-generation wireless communication networks.
So far, most studies have considered half-duplex (HD) relaying based on two-phase operation where a source transmits data to relays at the first time slot and the relays forward it to a destination at the second time slot \cite{LTW04TIT,LW04TIT}. However, such HD relaying causes a \emph{loss of multiplexing gain} expressed as an one-half pre-log factor.
To overcome the loss of multiplexing gain, several practical full-duplex (FD) relaying solutions have been studied \cite{DS10Asilomar,CJS+10MC,JCK+11MC,DDS12TWC,AKS+12MC,BMK13SC,DSA+14TVT,BK14NSDI,BJK14SC,BK14SC}.
Since \emph{strong self-interference} is a main problem which has to be resolved in FD relaying, the previous work primarily focused on self-interference cancellation based on antenna separation techniques in the wireless propagation domain and signal cancellation techniques in the analog circuit and digital domains.
Although these studies showed feasibility of FD relaying using small-scale wireless communication devices such as WiFi and IEEE 802.15.4, the technology is still premature for cellular communications, which require additional cancellation gains due to practical limitations such as varying center frequencies, bandwidth, and circuit imperfections.
In order to mitigate the loss of multiplexing gain in HD relaying, \emph{successive relaying} protocols have been proposed for a two-relay network \cite{RW07JSAC,FWT+07TWC,RGK10TIT,MK10TC,WFT+10TWC} and multiple-relay networks \cite{TaN08JSAC,HLT12TWC,KCJ+13TWC}.
In these protocols, two relays take turns acting as receivers and transmitters successively and a source and a transmitting relay transmit their own information simultaneously. Here, the source transmits new information and the relay transmits previously received information.
The main issue for such successive relaying protocols is to efficiently handle \emph{inter-relay interference} (IRI) from the transmitting relay to the receiving relay.
Towards this end, successive interference cancellation (SIC) and/or sophisticated coding and joint decoding techniques have been employed in the literature.
However, the SIC requires strong interference scenarios and the joint decoding requires high computational complexity.
Furthermore, although the successive relaying asymptotically achieves the spectral efficiency of the FD relaying with respect to the number of channel uses, it requires a sufficiently long block length (equivalently, coherence time) over slow fading channels.
Employing a buffer at the relay, such long block length constraints can be relaxed. Focusing on these advantages, \emph{buffer-aided relaying} has been proposed in a three-node network \cite{XFT+08TWC,ZSP11GC,ZlS13TIT,ZSP13JSAC}. The key idea is an opportunistic relaying mode selection (buffering or forwarding) according to channel conditions.
HD buffer-aided relaying can achieve up to two-fold spectral efficiency under asymmetric channel conditions between $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links, compared to HD relaying without buffer.
Additionally, bidirectional buffer-aided relaying with two-way traffic \cite{LPC+13CL,JZI+13EUSIPCO,JZI+13GC,SZA+13ISIT}, buffer-aided relaying over dual-hop broadcast channels \cite{ZSA+13TWC} and a shared relay channel with two source-destination pairs \cite{ZSA+13TC} have been studied.
By extending to multiple-relay networks, several \emph{opportunistic relaying} schemes, which exploit the best HD buffer-aided relay, have been proposed \cite{IMS12TWC,KCT12TWC,IKS12TVT}.
Ikhlef \emph{et al.} \cite{IMS12TWC} have proposed a $\max-\max$ relay selection (MMRS) scheme, which selects the best $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ relays with the maximum channel gains.
However, the MMRS scheme does not fully take advantage of the benefits of buffer-aided relaying since it maintains the two-phase operation.
Therefore, Krikidis \emph{et al.} \cite{KCT12TWC} have proposed a $\max-\mathrm{link}$ relay selection (MLRS) scheme, which selects the best relaying mode as well as the maximum channel gain.
Most recently, Ikhlef \emph{et al.} \cite{IKS12TVT} have proposed a space full-duplex $\max-\max$ relay selection (SFD-MMRS) scheme, which mimics the FD relaying by utilizing the best receiving and transmitting relays operating simultaneously.
In this scheme, they did not consider IRI by assuming fixed-relays with highly directional antennas.
However, this assumption does not always hold and it is hard to be practically realized as the number of relays increases.
With consideration of IRI, Kim and Bengtsson \cite{KB13PIMRC} proposed a virtual FD buffer-aided relaying scheme based on opportunistic relay selection (RS) with zero-forcing beamforming (BF) for IRI cancellation in order to maximize average end-to-end rate assuming \emph{adaptive rate transmission}.
Nomikos \emph{et al.} \cite{NCK+13PIMRC,NVC+13TETT} have proposed a buffer-aided successive opportunistic relaying (BA-SOR) scheme employing SIC at the receiving relay for \emph{fixed rate transmission}.
In \cite{NCK+13PIMRC}, even if it partially overcame the strong interference requirement of SIC through power allocation at the source and relays, the main objective was to minimize the total energy expenditure.
In \cite{NVC+13TETT}, the average end-to-end rate of the BA-SOR scheme, which has been originally devised for fixed rate transmission, is numerically shown for adaptive rate transmission. However, a fixed low SIC threshold, $r_0$ (e.g., 2 bps/Hz), has been applied even for adaptive rate transmission whereas the threshold value should be set to the information rate of the $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link.
Thus, this yields an optimistic result in terms of the average throughput.
In this paper, our main goal is to approach the average end-to-end rate of ideal FD relaying even in the presence of IRI.
To this end, we propose transmission schemes based on a joint RS and BF design utilizing multiple buffer-aided relays and multiple antennas at the relays.
For the joint RS and BF design, we first propose a weighted sum-rate maximization using instantaneous channel and buffer states for achieving the average end-to-end rate maximization.
Then, we separately design linear BF for each (receiving and transmitting) relay pair, which cancels or suppresses IRI, and optimal RS for maximizing the weighted sum-rate based on the beamformers found for each relay pair.
To focus on maximizing the average end-to-end rate, we employ adaptive rate transmission at the source and relays (i.e., channel state information at transmitter for both nodes) and consider delay-tolerant applications.
Our main contributions in this work are summarized as follows:
\begin{itemize}
\item A new RS criterion based on a weighted sum of instantaneous rates is proposed to maximize the average end-to-end rate in a virtual FD buffer-aided relaying network with adaptive rate transmission.
\item Various transmit and receive BF design strategies at the multiple antenna relays are proposed in order to cancel or suppress IRI.
\item We show that joint RS and BF schemes achieve the ideal FD relaying bound in terms of the average end-to-end rate asymptotically with increasing the number of antennas and/or the number of relays.
\item Compared to our previous work \cite{KB13PIMRC}, we propose more practical joint RS and BF schemes which can support non-identical channel conditions, including an iterative optimal BF-based RS scheme which achieves the best average end-to-end rate. Moreover, we provide extensive numerical results including average end-to-end rate, average delay, effect of IRI intensity, behavior of optimal weight factors, and effect of finite buffer size.
\end{itemize}
The rest of this paper is organized as follows. In Section~\ref{sec:system_model}, the system model is presented. The instantaneous rates and average end-to-end rate of a buffer-aided relaying network are described in Section~\ref{sec:capacity}. Buffer-aided joint RS and BF schemes considering IRI are proposed in Section~\ref{sec:relay_selection}.
In Section~\ref{sec:performance_evaluation}, the performance of the proposed schemes are evaluated through simulations. Finally, conclusive remarks and future work are provided in Section~\ref{sec:conclusion}.
\section{System Model} \label{sec:system_model}
In this paper, we consider a source, $\mathcal{S}$, and a destination, $\mathcal{D}$, which have a single antenna, and $K$ buffer-aided relays with $M$ antennas each (e.g., in Fig.~{\ref{fig:sys_model}}, $M=2$).
Denote the set of HD buffer-aided decode-and-forward relays by $\mathcal{K}=\{1,\ldots,K\}$.
We assume that there is no direct path between the source and destination as in the related literature \cite{TaN08JSAC,HLT12TWC,KCJ+13TWC,XFT+08TWC,ZSP11GC,ZlS13TIT,ZSP13JSAC,IMS12TWC,KCT12TWC,IKS12TVT,KB13PIMRC,NCK+13PIMRC,NVC+13TETT}.
This system model can be regarded as an example of relay-assisted device-to-device communications where the source and destination are low-cost devices with some limitations such as a single antenna.
The source is supposed to always have data traffic to transmit.
In addition, let $\mathbf{h}_{\mathcal{S}i}$, $\mathbf{h}_{j\mathcal{D}}$, and $\mathbf{H}_{ji}$, $i,j\in\mathcal{K}$ denote the channel coefficient vectors and matrices of $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$, $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$, and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ links, respectively.
The channel fading is assumed to be stationary and ergodic and in the asymptotic analysis and numerical examples, we will make the additional assumption that all the channel coefficients follow circular symmetric complex Gaussian distributions, $\mathbf{h}_{\mathcal{S}i}\sim\mathcal{CN}(0,\sigma_{\mathcal{S}\mathcal{R}_i}^2 \mathbf{I})$, $\mathbf{h}_{j\mathcal{D}}\sim\mathcal{CN}(0,\sigma_{\mathcal{R}_j\mathcal{D}}^2 \mathbf{I})$, and $\mathrm{vec}[{\mathbf{H}_{ji}}]\sim\mathcal{CN}(0,\sigma_{\mathcal{R}_j\mathcal{R}_i}^2 \mathbf{I})$ where $\mathrm{vec}[\cdot]$ denotes the vectorization of a matrix.
\begin{figure}
\centering
\includegraphics[width=6cm]{./fig/fig1_sys_model.eps}
\caption{System Model: a single source, a single destination, and multiple relays with buffer and multiple antennas (e.g., $M=2$).}
\label{fig:sys_model}
\vspace{-3mm}
\end{figure}
In order to mimic FD relaying, $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ transmissions are performed simultaneously by using the best pair of receiving and transmitting relays as in \cite{IKS12TVT,NCK+13PIMRC,NVC+13TETT,KB13PIMRC}.
To this end, the receiving relay decodes the data received from the source and stores it in its buffer, while the transmitting relay encodes data from its buffer and sends it to the destination.
For a given selected relay pair $(i,j)$, $i\neq j$, the received signal vector at the receiving relay $i$ is
\begin{align}
\mathbf{y}_i^{(i,j)} &= \mathbf{h}_{\mathcal{S}i} x_{\mathcal{S}} + \mathbf{H}_{ji} \mathbf{x}_{j} + \mathbf{n}_i = \mathbf{h}_{\mathcal{S}i} x_{\mathcal{S}} + \mathbf{H}_{ji} \mathbf{w}_{j} x_j + \mathbf{n}_i,
\label{eq:recv_sig_vec_sr}
\end{align}
where
$\mathbf{h}_{\mathcal{S}i}\in\mathbb{C}^{M \times 1}$ denotes the channel vector from the source to the $i$-th relay,
$\mathbf{H}_{ji}\in\mathbb{C}^{M \times M}$ denotes the inter-relay channel matrix from the $j$-th relay to the $i$-th relay, $x_{\mathcal{S}}$ denotes the transmitted data symbol from the source, and $\mathbf{x}_{j} = \mathbf{w}_{j} x_j$ denotes the transmitted data symbol vector from the $j$-th relay where $\mathbf{w}_{j}=[w_{j}^{1},\ldots,w_{j}^{M}]^T$ and $x_j$ represent the transmit BF vector of the $j$-th relay and the transmitted data symbol of the $j$-th relay, respectively. Here, $\mathbb{E}\left[|x_{\mathcal{S}}|^2\right] \leq {P}_{\mathcal{S}}$ and $\mathbb{E}\left[|x_j|^2\right] \leq {P}_{\mathcal{R}}$ where ${P}_{\mathcal{S}}$ and ${P}_{\mathcal{R}}$ denote the maximum transmit powers of the source and the relay, respectively, and $\| \mathbf{w}_{j} \| = 1$ for $i,j\in\mathcal{K}$ where $\|\cdot\|$ denotes the 2-norm. $\mathbf{n}_i$ denotes an additive white Gaussian noise (AWGN) vector with zero mean and covariance $\sigma_n^2\mathbf{I}$, i.e., $\mathbf{n}_i\sim\mathcal{CN}(0,\sigma_n^2\mathbf{I})$.
For a given selected relay pair $(i,j)$, $i \neq j$, the received signal at the destination is
\begin{align}
y_{\mathcal{D}}^{(i,j)} &= \mathbf{h}_{j\mathcal{D}}^H\mathbf{x}_{j} + n_{\mathcal{D}}
= \mathbf{h}_{j\mathcal{D}}^H\mathbf{w}_{j} x_j + n_{\mathcal{D}},
\label{eq:recv_sig_rd}
\end{align}
where
$\mathbf{h}_{j\mathcal{D}}\in\mathbb{C}^{M \times 1}$ denotes the channel vector from the $j$-th relay to the destination\footnote{For notational convenience, we define $\mathbf{h}_{j\mathcal{D}}$ as the complex conjugate channel vector differently from the definition of $\mathbf{h}_{\mathcal{S}i}$, i.e., $\mathbf{h}_{j\mathcal{D}}^H=[h_{j\mathcal{D}}^{1},\ldots,h_{j\mathcal{D}}^{M}]$ while $\mathbf{h}_{\mathcal{S}i}=[h_{\mathcal{S}i}^{1},\ldots,h_{\mathcal{S}i}^{M}]^T$.}, $(\cdot)^H$ denotes the Hermitian transpose,
and $n_{\mathcal{D}}$ denotes AWGN with zero mean and variance $\sigma_n^2$, i.e., $n_{\mathcal{D}}\sim\mathcal{CN}(0,\sigma_n^2)$.
\section{Instantaneous and Average Data Rates of a Buffer-Aided Relaying Network} \label{sec:capacity}
Using a linear receive BF vector $\mathbf{u}_{i}=[u_{i}^{1},\ldots,u_{i}^{M}]^T$ at the $i$-th receiving relay, with $\|\mathbf{u}_i\|=1$, the received signal after the receive BF becomes
\begin{align}
r_{i}^{(i,j)} = \mathbf{u}_i^H \mathbf{y}_{i}^{(i,j)}
&=\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i} x_{\mathcal{S}} + \mathbf{u}_i^H \mathbf{H}_{ji} \mathbf{w}_{j} x_j + \tilde{n}_i,
\label{eq:recv_sig_sr}
\end{align}
for given relay pair $(i,j)$, where $\tilde{n}_i= \mathbf{u}_i^H \mathbf{n}_i\sim \mathcal{CN}(0,\sigma_n^2)$.
From \eqref{eq:recv_sig_sr} and \eqref{eq:recv_sig_rd}, the instantaneous received SINR/SNR for the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links at time slot $t$ are expressed, respectively, as:
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t)
&= \frac{\rho_{\mathcal{S}} | \mathbf{u}_{i}^{H}\mathbf{h}_{\mathcal{S}i} |^2 }{ 1 + \rho_{\mathcal{R}} |\mathbf{u}_{i}^{H} \mathbf{H}_{ji}\mathbf{w}_{j} |^2 },\label{eq:SINRi}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t)
&= \rho_{\mathcal{R}} | \mathbf{h}_{j\mathcal{D}}^H \mathbf{w}_{j} |^2,
\label{eq:SNRD}
\end{align}
where $\rho_{\mathcal{S}} = P_{\mathcal{S}} / \sigma_n^2$ and $\rho_{\mathcal{R}} = P_{\mathcal{R}} / \sigma_n^2$.
Let $B_i(t)$ denote the number of bits in the buffer normalized by the number of channel uses of the $i$-th relay at the end of time slot $t$.
Assuming a Gaussian codebook and information theoretic capacity achieving coding scheme, and taking the buffer state at the receiving relay $i$ into consideration, the instantaneous rate of the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ link is
\begin{align}
C_{\mathcal{S}i}^{(i,j)}(t) = \min\left\{ \log_2\left(1+\gamma_{\mathcal{S}i}^{(i,j)}(t)\right), B_{max}-B_i(t-1) \right\},
\label{eq:C_si}
\end{align}
where $\min\{\cdot, \cdot\}$ denotes the minimum value of arguments, $B_{max}$ denotes the maximum buffer size, and the buffer of the $i$-th relay is updated by
\begin{align}
B_i(t) = B_i(t-1) + C_{\mathcal{S}i}^{(i,j)}(t).\nonumber
\end{align}
For the transmitting relay $j$, the instantaneous rate of link $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ in time slot $t$ is
\begin{align}
C_{j\mathcal{D}}^{(i,j)}(t) = \min\left\{ \log_2\left(1+\gamma_{j\mathcal{D}}^{(i,j)}(t)\right), B_j(t-1) \right\},
\label{eq:C_jd}
\end{align}
where the buffer of the $j$-th relay is updated by
\begin{align}
B_j(t) = B_j(t-1) - C_{j\mathcal{D}}^{(i,j)}(t).\nonumber
\end{align}
The average received data rate at the destination, over a time window of length $T$, is given by
\begin{equation}
\label{eq:dest_rate}
\bar{C}_{\mathcal{D}} = \frac{1}{T} \sum_{t=1}^T C_{j(t)\mathcal{D}}^{(i(t),j(t))}(t)\;,
\end{equation}
where $i(t)$ and $j(t)$ denote the indices of the selected receiving and transmitting relay indices in time slot $t$, respectively. Similarly, the average transmitted data rate at the source is
\begin{equation}
\label{eq:source_rate}
\bar{C}_{\mathcal{S}} = \frac{1}{T} \sum_{t=1}^T C_{\mathcal{S}i(t)}^{(i(t),j(t))}(t)\;.
\end{equation}
\section{Buffer-Aided Joint Relay Selection and Beamforming Schemes in the Presence of Inter-Relay Interference} \label{sec:relay_selection}
Our objective is to maximize the average data rate given in~\eqref{eq:dest_rate}, through a joint RS and BF design. Optimizing the minimum of the instantaneous rates, which corresponds to maximizing a lower bound on $\min\{\bar{C}_{\mathcal{S}},\bar{C}_{\mathcal{D}}\}$ \cite{XFT+08TWC}, was proposed in \cite{BKR+06JSAC} as an HD best relay selection criterion. Since this is suboptimal in the presence of IRI, we here propose an alternative approach, by formulating the following optimization problem,
\begin{subequations}
\label{eq:C_D_opt}
\begin{IEEEeqnarray}{cl}
\hspace{-8mm}\underset{\{\mathbf{u}_{i(t)},\mathbf{w}_{j(t)}, i(t), j(t)\}}{\mbox{max}}~ & \bar{C}_{\mathcal{D}} \\
\hspace{-12mm}\mbox{s. t.} & \frac{1}{T} \sum_{t=1}^T \Big[ C_{\mathcal{S}k}^{(i(t),j(t))}(t)\nonumber\\
\hspace{-12mm}& \hspace{8mm}- C_{k\mathcal{D}}^{(i(t),j(t))}(t) \Big] \geq 0, ~\forall k \in \mathcal{K}, \label{eq:non-decreasing_buffers} \\
\hspace{-12mm}& \|\mathbf{u}_{i(t)}\| \leq 1, \|\mathbf{w}_{j(t)}\| \leq 1,\label{eq:BF_norms}\\
\hspace{-12mm}& i(t) \neq j(t).\label{eq:constraint_neq}
\end{IEEEeqnarray}
\end{subequations}
In order for the relay buffers to stay non-empty, we have here added the constraints~\eqref{eq:non-decreasing_buffers} (the buffer is absorbing). Obviously, we want~\eqref{eq:non-decreasing_buffers} to hold with equality (buffer stability), but the inequality formulation is exploited below to determine constrains on the weighting parameters. Problem~\eqref{eq:C_D_opt} is non-convex in the beamforming vectors and combinatorial in the relay selection. For tractability, we therefore use a Lagrange relaxation~\cite{Fisher:2004}, studying the partial Lagrangian with dual variables $\alpha_k$ corresponding to the constraints~\eqref{eq:non-decreasing_buffers},
\begin{align}
\label{eq:Lagr_C_D}
&\mathcal{L}(\mathbf{u}_{i(t)},\mathbf{w}_{j(t)}, i(t), j(t),\{\alpha_k\})\nonumber\\
&= \bar{C}_{\mathcal{D}} +
\sum_{k\in\mathcal{K}} \alpha_k \left( \frac{1}{T} \sum_{t=1}^T \left[ C_{\mathcal{S}k}^{(i(t),j(t))}(t) - C_{k\mathcal{D}}^{(i(t),j(t))}(t) \right] \right).
\end{align}
Similarly to~\cite{XFT+08TWC}, and for the moment assuming unbounded buffers, it can be shown that the probability that $B_j(t-1) < \log_2\left(1+\gamma_{j\mathcal{D}}^{(i,j)}(t)\right)$ goes to zero, if the channel fading is stationary and~\eqref{eq:non-decreasing_buffers} holds. Therefore, the RS and BF can be optimized separately for each time step. Collecting the terms corresponding to time step $t$, the primal variables optimizing $\mathcal{L}(\cdot)$ are given by
\begin{subequations}
\label{eq:weighted_opt}
\begin{IEEEeqnarray}{cl}
\underset{\{\mathbf{u}_{i},\mathbf{w}_{j}, i, j\}}{\mbox{max}}~ & \alpha_i C_{\mathcal{S}i}^{(i,j)} + (1-\alpha_j) C_{j\mathcal{D}}^{(i,j)} \\
\mbox{s. t.} & \eqref{eq:BF_norms}\quad \&\quad \eqref{eq:constraint_neq}
\end{IEEEeqnarray}
\end{subequations}
and it follows from the inequality in~\eqref{eq:non-decreasing_buffers} that $\alpha_k\geq 0$ at the optimum. A similar argument, instead optimizing the average transmit rate at the source, under constraints that the average relay buffers are non-increasing, with Lagrange multipliers $1-\alpha_k$, gives exactly the same criterion~\eqref{eq:weighted_opt}, but with the constraints that $1-\alpha_k\geq 0$. The conclusion of this Lagrange relaxation is therefore that the optimal RS and BF strategy can be done separately for each time step based on instantaneous information and has the form~\eqref{eq:weighted_opt}, for some choice of weight factors (dual variables) $\alpha_k$ with $0\leq \alpha_k\leq 1$.
One possible approach to determine the optimal $\alpha_k$ is to use subgradient optimization. As shown in~\cite{Bertsekas:95}, a subgradient for $\alpha_k$ is given by the left hand side of~\eqref{eq:non-decreasing_buffers}. However, to avoid having to run the system $T$ time steps between each update of $\alpha_k$, we propose to replace the time average in~\eqref{eq:non-decreasing_buffers} by an exponentially weighted sum
\begin{equation}
\label{eq:exp_aver_queue_change}
\Delta_{B_k}(t) = (1-\lambda) \sum_{\tau=0}^t \lambda^{t-\tau} \left[ C_{\mathcal{S}k}^{(i(\tau),j(\tau))}(\tau) - C_{k\mathcal{D}}^{(i(\tau),j(\tau))}(\tau) \right],
\end{equation}
where the forgetting factor $\lambda$ is close to, but less then one, and update $\alpha_i$ at each time step using
\begin{equation}
\alpha_k(t) = \min\left\{1,\max\{0,\alpha_k(t-1) - \mu(t) \Delta_{B_k}(t)\}\right\} \label{eq:alpha_subgrad_update}
\end{equation}
for some suitable choice of step size $\mu(t)$. Note that $\Delta_{B_k}(t)$ only will be an approximation of the true subgradient, since it is influenced by several earlier values of $\alpha_k(t)$ and also since it is stochastic. Still, numerical experiments have always shown convergence.
Since keeping all the buffers stable is a necessary optimality condition, an alternative approach is to use the back-pressure algorithm~\cite{Tassiulas:1992,GeorgiadisNT:2006}. Assume a stationary stochastic arrival process with a rate less than or equal to the achievable source-destination rate and let $B_{\mathcal{S}}(t)$ denote the source buffer occupancy at time $t$. Then, a standard derivation of the back-pressure algorithm, using the Lyapunov function $B_{\mathcal{S}}^2(t) + B_i^2(t)$, gives the following design criterion,
\begin{subequations}
\label{eq:backpressure}
\begin{IEEEeqnarray}{cl}
\underset{\{\mathbf{u}_{i},\mathbf{w}_{j}, i, j\}}{\mbox{max}}~ & (B_{\mathcal{S}}(t)-B_i(t)) C_{\mathcal{S}i}^{(i,j)} + B_j(t) C_{j\mathcal{D}}^{(i,j)} \\
\mbox{s.t. } & \eqref{eq:BF_norms}\quad \&\quad \eqref{eq:constraint_neq}
\end{IEEEeqnarray}
\end{subequations}
Dividing this criterion by $B_{\mathcal{S}}(t)$ gives an expression of exactly the same form as~\eqref{eq:weighted_opt}, if we set $\alpha_i=1-B_i(t)/B_{\mathcal{S}}(t)$. Therefore, an alternative approach to determine the optimal $\alpha_i$, is to run the back-pressure algorithm until it reaches stationarity, and use a time average of $1-B_i(t)/B_{\mathcal{S}}(t)$ as $\alpha_i$.
Throughout this paper, we consider an exhaustive search under global channel state information (CSI) and buffer state information (BSI) for all proposed schemes to obtain the optimal performance in RS.
A discussion on the complexity of our proposed schemes for both RS and BF aspects is provided in Section~\ref{subsec:complexity}
The conventional centralized/distributed RS approaches \cite{BKR+06JSAC,BSW07TWC,NMV+13CAMAD} can be applied for implementation of the proposed RS schemes.
To reduce the amount of CSI feedback\footnote{Since the proposed schemes require up-to-date CSI, they are primarily applicable in scenarios with low mobility, relative to the symbol duration, such as typical WiFi deployments.}, distributed RS approaches are more desirable than the centralized approach in practice. For example, first of all, the relays can estimate channels based on orthogonal pilot signals from the source and the destination at the same time. Then, the relays can estimate inter-relay channels in a round robin manner using orthogonal MIMO pilot signals.
If BSIs and CSIs for $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ links (i.e., $\mathbf{h}_{\mathcal{S}i}$'s) are shared among the relays, each relay is able to select the local-best receiving relay by regarding itself as the transmitting relay.
Similarly to the timer-based distributed RS in \cite{BKR+06JSAC,DYH12TVT}, each relay sets a timer based on an inverse of its local-best objective function value and sends a request-to-send message after timer expiration. Then, the destination sends back a clear-to-send message for the earliest access relay.
If the relay receives the message, it broadcasts the RS information to all the nodes. Afterwards, the source and the transmitting relay start to transmit their own packets.
Through this procedure, the best relay pair can be determined in a distributed manner.
The detailed implementation issues are beyond the scope of this work.
The instantaneous rates in \eqref{eq:backpressure} are determined by effective SINR/SNRs at the receiving and transmitting relays, depending on the transmit and receive beamformers.
Thus, the transmit and receive beamformers have to be determined separately for each candidate pair of relays.
However, finding the optimal BF vectors for every given relay pair is non-convex and therefore we propose an iterative optimal BF-based RS scheme.
Since the iterative solution requires a high computational complexity,
several low-complexity suboptimal BF-based RS schemes are also proposed in the rest of this section.
Note that the suboptimal schemes have less complexity in the BF design than the optimal scheme but the same RS protocol complexity. (See Section~\ref{subsec:complexity}.)
For simplicity, we omit the time slot index $t$ in the relay indices $i(t)$ and $j(t)$ hereafter.
\subsection{Proposed Optimal Beamforming-based Relay Selection Scheme}
Obviously, if there is no IRI, maximal ratio combining (MRC) at the receiving relay and maximal ratio transmit (MRT) BF at the transmitting relay, named \emph{IRI-free BF}, are optimal. However, this idealized IRI-free BF is not optimal in the presence of IRI.
In this subsection, we propose an iterative optimal BF-based RS scheme to maximize the average end-to-end rate.
Denoting $f(\gamma) = \log_2(1+\gamma)$,
the optimization problem \eqref{eq:weighted_opt} for given $\{\alpha_i\}$ and for each given relay pair (i.j) is
\begin{align}
\begin{array}{cl}
\underset{\{\mathbf{u}_{i},\mathbf{w}_{j}, i, j\}}{\mbox{max}} & \quad \alpha f\left(\frac{\rho_{\mathcal{S}}|\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i}|^2}{\mathbf{u}_i^H \left(\rho_{\mathcal{R}}\mathbf{H}_{ji}\mathbf{w}_{j}\mathbf{w}_{j}^H\mathbf{H}_{ji}^H + \mathbf{I}\right) \mathbf{u}_i}\right)\\
&\quad+ (1-\alpha) f\left(\rho_{\mathcal{R}}\mathbf{h}_{j\mathcal{D}}^H\mathbf{w}_{j}\mathbf{w}_{j}^H \mathbf{h}_{j\mathcal{D}}\right)\\
\mbox{s. t.} & \quad \|\mathbf{u}_i\| \leq 1, \|\mathbf{w}_j\| \leq 1.
\end{array}
\end{align}
Since this optimization problem cannot be solved directly due to non-convexity, we propose an alternating optimization which iterates between (i) \emph{for fixed $\mathbf{w}_j$, optimizing $\mathbf{u}_i$}, and (ii) \emph{for fixed $\mathbf{u}_i$, optimizing $\mathbf{w}_j$}.
For given $\mathbf{w}_j$, the optimal receive beamformer $\mathbf{u}_i$ is obviously given by the MMSE solution,
\begin{align}
\mathbf{u}_i = c_u \left(\rho_{\mathcal{R}}\mathbf{H}_{ji}\mathbf{w}_{j}\mathbf{w}_{j}^H\mathbf{H}_{ji}^H + \mathbf{I}\right)^{-1} \mathbf{h}_{\mathcal{S}i},
\end{align}
where the scaling factor $c_u$ is selected such that $\|\mathbf{u}_i\|=1$.
For given $\mathbf{u}_i$, optimizing the transmit beamformer $\mathbf{w}_j$
is a non-convex problem. Denote $\mathbf{g}_{ji} \triangleq \mathbf{H}_{ji}^H \mathbf{u}_{i}$. The Lagrangian of the optimization problem is
\begin{align}
&\mathcal{L}(\mathbf{w}_j, \lambda) = \alpha f\left( \frac{\rho_{\mathcal{S}}|\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i}|^2}{ \rho_{\mathcal{R}}\mathbf{g}_{ji}^H \mathbf{w}_j\mathbf{w}_j^H \mathbf{g}_{ji} + \|\mathbf{u}_i\|^2 } \right) \nonumber\\
&\quad + (1-\alpha) f\left(\rho_{\mathcal{R}}\mathbf{h}_{j\mathcal{D}}^H \mathbf{w}_j \mathbf{w}_j^H \mathbf{h}_{j\mathcal{D}} \right) + \lambda \left( 1 - \mathbf{w}_j^H \mathbf{w}_j \right),
\end{align}
where the gradient with respect to $\mathbf{w}_{j}$ is obtained by
\begin{align}
&\bigtriangledown_{\mathbf{w}_j}\mathcal{L}(\mathbf{w}_{j}, \lambda) \nonumber\\
&= - \frac{\alpha f^{\prime}(\gamma_{\mathcal{S}i})\rho_{\mathcal{S}}|\mathbf{u}_i^H\mathbf{h}_{\mathcal{S}i}|^2}{ \left( \rho_{\mathcal{R}}\mathbf{g}_{ji}^H \mathbf{w}_j\mathbf{w}_j^H \mathbf{g}_{ji} + \|\mathbf{u}_i\|^2 \right)^2 } 2\rho_{\mathcal{R}}{\mathbf{g}_{ji}}\mathbf{g}_{ji}^H \mathbf{w}_j \nonumber\\
&\quad+ (1-\alpha) f^{\prime}(\gamma_{j\mathcal{D}}) 2\rho_{\mathcal{R}}\mathbf{h}_{j\mathcal{D}}\mathbf{h}_{j\mathcal{D}}^H \mathbf{w}_j - 2\lambda\mathbf{w}_j,
\end{align}
where $\gamma_{\mathcal{S}i} = \frac{\rho_{\mathcal{S}}|\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i}|^2}{ \rho_{\mathcal{R}}\mathbf{g}_{ji}^H \mathbf{w}_j\mathbf{w}_j^H \mathbf{g}_{ji} + \|\mathbf{u}_i\|^2 }$ and $\gamma_{j\mathcal{D}} = \rho_{\mathcal{R}}\mathbf{h}_{j\mathcal{D}}^H\mathbf{w}_j\mathbf{w}_j^H \mathbf{h}_{j\mathcal{D}}$.
Hence, the KKT conditions give
\begin{align}
\left(\lambda\mathbf{I} + \mu {\mathbf{g}_{ji}}\mathbf{g}_{ji}^H\right)\mathbf{w}_j
= (1-\alpha_j) f^{\prime}(\gamma_{j\mathcal{D}}) \rho_{\mathcal{R}} \mathbf{h}_{j\mathcal{D}} \mathbf{h}_{j\mathcal{D}}^H \mathbf{w}_j,
\label{eq:Lambda}
\end{align}
where $\mu \triangleq \frac{\alpha_i f^{\prime}(\gamma_{\mathcal{S}i})\rho_{\mathcal{S}}|\mathbf{u}_i^H\mathbf{h}_{\mathcal{S}i}|^2}{ \left( \rho_{\mathcal{R}}\mathbf{g}_{ji}^H \mathbf{w}_j{\mathbf{w}_j}^H \mathbf{g}_{ji} + \|\mathbf{u}_i\|^2 \right)^2 }\rho_{\mathcal{R}}$.
Thus, the optimal $\mathbf{w}_j$ has the form
\begin{align}
\mathbf{w}_j = c_w \left(\lambda\mathbf{I} + \mu {\mathbf{g}_{ji}}\mathbf{g}_{ji}^H\right)^{-1} \mathbf{h}_{j\mathcal{D}},
\label{eq:sol_opt_w}
\end{align}
for some values of the positive real-valued parameters $\lambda$ and $\mu$, and scaling constant $c_w$.
Using the matrix inversion lemma,\footnote{$(A-UD^{-1}V)^{-1} = A^{-1} + A^{-1}U\left(D-VA^{-1}U\right)^{-1}VA^{-1}$ \cite{HJ13MA}.} \eqref{eq:sol_opt_w} is rewritten as
\begin{align}
\mathbf{w}_j &= c_w \left( \lambda^{-1}\mathbf{I} - \lambda^{-1}\mathbf{g}_{ji} \left( \mathbf{g}_{ji}^H \mathbf{g}_{ji} + \lambda/\mu \right)^{-1} \mathbf{g}_{ji}^H\right) \mathbf{h}_{j\mathcal{D}} \nonumber\\
&= \frac{c_w}{\lambda} \left( \mathbf{h}_{j\mathcal{D}} - \frac{\mathbf{g}_{ji}^H \mathbf{h}_{j\mathcal{D}}}{\mathbf{g}_{ji}^H \mathbf{g}_{ji} + \lambda/\mu} \mathbf{g}_{ji}\right) \nonumber\\
&= \frac{c_w}{\lambda} \Bigg( \underbrace{ \mathbf{h}_{j\mathcal{D}} - \frac{\mathbf{g}_{ji}^H \mathbf{h}_{j\mathcal{D}}}{\mathbf{g}_{ji}^H \mathbf{g}_{ji}} \mathbf{g}_{ji} }_{\triangleq \tilde{\mathbf{w}}^{\perp}_{j} } \nonumber\\
&\hspace{10mm}+ \left(1 - \frac{\mathbf{g}_{ji}^H \mathbf{g}_{ji}}{\mathbf{g}_{ji}^H \mathbf{g}_{ji} + \lambda/\mu} \right)\underbrace{ \frac{\mathbf{g}_{ji}^H \mathbf{h}_{j\mathcal{D}}}{\mathbf{g}_{ji}^H \mathbf{g}_{ji}} \mathbf{g}_{ji} }_{\triangleq \tilde{\mathbf{w}}^{\parallel}_{j} } \Bigg),
\label{eq:sol_opt_w_2}
\end{align}
where $ \tilde{\mathbf{w}}^{\perp}_{j} $ is defined as the projection of $\mathbf{h}_{j\mathcal{D}}$ onto the orthogonal subspace of $\mathbf{g}_{ji}$ and $ \tilde{\mathbf{w}}^{\parallel}_{j}$ is defined as the projection of $\mathbf{h}_{j\mathcal{D}}$ onto $\mathbf{g}_{ji}$.
Let us define the corresponding normalized vectors ${\mathbf{w}}_j^{\parallel}\triangleq\frac{\tilde{\mathbf{w}}_j^{\parallel}}{\|\tilde{\mathbf{w}}_j^{\parallel}\|}$ and ${\mathbf{w}}_j^{\perp}\triangleq\frac{\tilde{\mathbf{w}}_j^{\perp}}{\|\tilde{\mathbf{w}}_j^{\perp}\|}$.
Since $\|\mathbf{w}_j\|=1$ at the optimum and the scaling constants in \eqref{eq:sol_opt_w_2} are real valued and positive, the optimal $\mathbf{w}_j$ has the form
\begin{align}
\mathbf{w}_j = \beta {\mathbf{w}}_j^{\parallel} + \sqrt{1-\beta^2}{\mathbf{w}}_j^{\perp},
\label{eq:opt_w}
\end{align}
for some value of $\beta\in[0,1]$. Thus, the optimization problem at hand is reduced to find the optimal $\beta$ parameter for given $(i,j)$ relay pair.
Substituting \eqref{eq:opt_w} into the original objective function gives
\begin{align}
&\underset{\beta\in[0,1]}{\mbox{max}} \quad \alpha f\left(\frac{\rho_{\mathcal{S}}|\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i}|^2}{\rho_{\mathcal{R}} |\beta \mathbf{g}_{ji}^H {\mathbf{w}}_{j}^{\parallel} |^2 + \|\mathbf{u}_i\|^2}\right)\nonumber\\
&\hspace{12mm} + (1-\alpha) f\left(\rho_{\mathcal{R}} |\beta \mathbf{h}_{j\mathcal{D}}^H {\mathbf{w}}_j^{\parallel} + \sqrt{1-\beta^2} \mathbf{h}_{j\mathcal{D}}^H {\mathbf{w}}_j^{\perp} |^2 \right),
\label{eq:obj_function}
\end{align}
where $f(\gamma)=\log_2\left(1+\gamma\right)$.
In general, this objective function may have multiple optima in the interval $[0,1]$ with respect to $\beta$. However, a rough grid search in this interval and a few Gauss-Newton steps starting at the optimum grid point after the grid search can quickly find the global optimum. Note that \eqref{eq:obj_function} is easily vectorized which can speed up the grid search.
Equations \eqref{eq:sol_opt_w_2} and \eqref{eq:obj_function} are iterated to find the optimal beamformers pair until a fixed number of iterations is reached or a certain stopping condition is satisfied.\footnote{In numerical results, we used a stopping condition where the differences in vector norms should be under a certain error tolerance ($\varepsilon_t$), i.e., $\|\mathbf{u}_i(n)-\mathbf{u}_i(n-1)\|<\varepsilon_t$ and $\|\mathbf{w}_j(n)-\mathbf{w}_j(n-1)\|<\varepsilon_t$ for the $n$-th iteration. For all results, we set $\varepsilon_t=10^{-4}$ and then the number of iterations is about several hundreds depending on channel realizations.}
Let the optimized beamformers and $\beta$ parameter be denoted by $\mathbf{u}_i^{\star}$, $\mathbf{w}_j^{\star}$, and $\beta^{\star}$. Then, the instantaneous SNRs for the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links are expressed, respectively, as:
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t) &= \frac{\rho_{\mathcal{S}}|\mathbf{u}_i^{\star H} \mathbf{h}_{\mathcal{S}i}|^2}{\rho_{\mathcal{R}} |\beta^{\star} \mathbf{g}_{ji}^H {\mathbf{w}}_{j}^{\parallel \star} |^2 + \|\mathbf{u}_i^{\star}\|^2},\label{eq:sinr_si_optbf}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t) &=\rho_{\mathcal{R}} \left|\beta^{\star} \mathbf{h}_{j\mathcal{D}}^H {\mathbf{w}}_j^{\parallel \star} + \sqrt{1-\beta^{\star 2}} \mathbf{h}_{j\mathcal{D}}^H {\mathbf{w}}_j^{\perp \star} \right|^2,\label{eq:sinr_jd_optbf}
\end{align}
where $\mathbf{w}_j^{\star} = \beta^{\star} {\mathbf{w}}_j^{\parallel \star} + \sqrt{1-\beta^{\star 2}} {\mathbf{w}}_j^{\perp \star}$.
Hence, substituting \eqref{eq:sinr_si_optbf} and \eqref{eq:sinr_jd_optbf} into \eqref{eq:C_si} and \eqref{eq:C_jd}, respectively, the best relay pair is selected by \eqref{eq:backpressure}.
\subsection{Proposed Relay Selection Scheme with Zero-Forcing Beamforming (ZFBF)-based IRI Cancellation}
In this subsection, we propose to optimize a transmit beamformer based on zero-forcing (ZF) at the transmitting relay.
First of all, we use the MRC beamformer for the receiving relay $i$, i.e., $\mathbf{u}_i = \frac{{\mathbf{h}_{\mathcal{S}i}}}{\| \mathbf{h}_{\mathcal{S}i} \|}$ and then maximize the effective channel power gain of the $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link under a ZF condition.
Therefore, for a given relay pair $(i,j)$, the following optimization problem is formulated:
\begin{subequations}
\begin{IEEEeqnarray}{cl}
\mbox{max} &\quad |\mathbf{w}_{j}^H {\mathbf{h}_{j\mathcal{D}}}|^2 \label{eq:opt_null_space}\\
\mbox{s. t.} &\quad \mathbf{u}_i^H\mathbf{H}_{ji}\mathbf{w}_{j} = 0,\label{eq:opt_null_space_c1}\\
&\quad \|\mathbf{w}_{j}\|=1. \label{eq:opt_null_space_c2}
\end{IEEEeqnarray}
\end{subequations}
Let $\mathbf{V}_{ji}\in\mathbb{C}^{M\times(M-1)}$ be a matrix whose columns span the null-space of $\mathbf{g}_{ji} \triangleq \mathbf{H}_{ji}^H\mathbf{u}_i$.
Then, any BF vector ${\mathbf{w}_{j}}$ fulfilling the first constraint in \eqref{eq:opt_null_space_c1} can be written as ${\mathbf{w}_{j}}=\mathbf{V}_{ji} \mathbf{z}$, where $\mathbf{z} \in C^{(M-1)\times 1}$.
Hence, the optimization problem is reformulated by
\begin{subequations}
\begin{IEEEeqnarray}{cl}
\underset{\mathbf{z}}{\mbox{max}} & \quad |{\mathbf{z}}^H \mathbf{V}_{ji}^H {\mathbf{h}_{j\mathcal{D}}}|^2 \label{eq:opt_null_space_re}\\
\mbox{s. t.} & \quad \|\mathbf{V}_{ji} \mathbf{z}\|=1,
\end{IEEEeqnarray}
\end{subequations}
which is equivalent to
\begin{subequations}
\begin{IEEEeqnarray}{cl}
\underset{\mathbf{z}}{\mbox{max}} &\quad \displaystyle \frac{\mathbf{z}^H \mathbf{V}_{ji}^H {\mathbf{h}_{j\mathcal{D}}} \mathbf{h}_{j\mathcal{D}}^H {\mathbf{V}_{ji}} \mathbf{z}}{\mathbf{z}^H \mathbf{V}_{ji}^H{\mathbf{V}_{ji}} \mathbf{z}} \label{eq:opt_null_space_final}\\
\mbox{s. t.} &\quad \|\mathbf{V}_{ji} \mathbf{z}\|=1.
\end{IEEEeqnarray}
\end{subequations}
The solution of this problem is $\mathbf{z}^{\star}=c_z (\mathbf{V}_{ji}^H\mathbf{V}_{ji})^{-1}\mathbf{V}_{ji}^H {\mathbf{h}_{j\mathcal{D}}}$, resulting in $\mathbf{w}_j^{\star} = c_z \mathbf{V}_{ji} (\mathbf{V}_{ji}^H \mathbf{V}_{ji})^{-1}$ $\cdot\mathbf{V}_{ji}^H \mathbf{h}_{j\mathcal{D}} = c_z \tilde{\mathbf{w}}_{j}^{\star} $, where the scalar $c_z$ is chosen so that $\|\mathbf{w}_j^{\star}\|=1$, i.e. $ \mathbf{w}_j^{\star} = \frac{\tilde{\mathbf{w}}_{j}^{\star}}{\|{\tilde{\mathbf{w}}_{j}}^{\star}\|} $ in which
\begin{align}
\tilde{\mathbf{w}}_{j}^{\star} &= \mathbf{V}_{ji}(\mathbf{V}_{ji}^H\mathbf{V}_{ji})^{-1}\mathbf{V}_{ji}^H {\mathbf{h}_{j\mathcal{D}}}
= \left( \mathbf{I} - \frac{\mathbf{g}_{ji} \mathbf{g}_{ji}^H}{\mathbf{g}_{ji}^H\mathbf{g}_{ji}} \right){\mathbf{h}_{j\mathcal{D}}} \nonumber\\
&= {\mathbf{h}_{j\mathcal{D}}} - c \mathbf{H}_{ji}^H\mathbf{h}_{\mathcal{S}i},
\label{eq:q_ZFBF}
\end{align}
where the scalar value $c=\frac{ \mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji} {\mathbf{h}_{j\mathcal{D}}}}{ \|\mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji}\|^2 }$ since $\mathbf{u}_{i} = \frac{\mathbf{h}_{\mathcal{S}i}}{\|\mathbf{h}_{\mathcal{S}i}\|}$.
This optimum solution implies a projection of ${\mathbf{h}_{j\mathcal{D}}}$ onto the null-space $\mathbf{V}_{ji}$.
Therefore, substituting $\mathbf{u}_i$ and $\mathbf{w}_j$ into \eqref{eq:SINRi} and \eqref{eq:SNRD}, the instantaneous SNRs for the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links are expressed, respectively, as:
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t) &= \rho_{\mathcal{S}} \|\mathbf{h}_{\mathcal{S}i}\|^2,\label{eq:sinr_si_zfbf}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t) &=\frac{ \rho_{\mathcal{R}} \left| \|{\mathbf{h}_{j\mathcal{D}}}\|^2 - \tilde{c} \right|^2}{\|{\mathbf{h}_{j\mathcal{D}}} - c \mathbf{H}_{ji}^H\mathbf{h}_{\mathcal{S}i}\|^2},\label{eq:sinr_jd_zfbf}
\end{align}
where the scalar values $c=\frac{ \mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji} {\mathbf{h}_{j\mathcal{D}}}}{ \|\mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji}\|^2 }$ and $\tilde{c}=\frac{ |\mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji} {\mathbf{h}_{j\mathcal{D}}}|^2 }{ \|\mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji} \|^2 }$.
As a result, substituting \eqref{eq:sinr_si_zfbf} and \eqref{eq:sinr_jd_zfbf} into \eqref{eq:C_si} and \eqref{eq:C_jd}, respectively, the best relay pair is selected by \eqref{eq:backpressure}.
\begin{proposition
\label{prop:ZFBF_antennas}
The proposed ZFBF-based RS scheme asymptotically achieves the average end-to-end rate of ideal FD relaying with probability one as the number of antennas ($M$) goes to infinity.
\begin{proof}
Let $(\mathbf{H})_m$ denote the $m$-th column vector of the matrix $\mathbf{H}$ and $(\mathbf{H})_{m,l}$ denote the $(m,l)$-th element of the matrix $\mathbf{H}$.
Denoting $\mathbf{g}_{ji} = [g_{ji}^{1},g_{ji}^{2},\ldots,g_{ji}^{M}]^T$ in \eqref{eq:q_ZFBF}, $g_{ji}^{m}=(\mathbf{H}_{ji})_m^H \mathbf{u}_{i}\sim\mathcal{CN}(0,\sigma_{\mathcal{RR}}^2)$ since $(\mathbf{H}_{ji})_{m,l}\sim\mathcal{CN}(0,\sigma_{\mathcal{RR}}^2)$ and $\|\mathbf{u}_i\|=1$. Assuming $\sigma_{\mathcal{RR}}^2=1$ without loss of generality, $g_{ji}^{m}\sim\mathcal{CN}(0,1)$ and therefore $\mathbf{g}_{ji}^H \mathbf{g}_{ji} = \sum_{m=1}^{M} |g_{ji}^{m}|^2$ follows a chi-squared distribution with $2M$ degrees of freedom, i.e., $\mathbf{g}_{ji}^H \mathbf{g}_{ji} \sim \chi_{2M}^{2}$. Meanwhile, the diagonal elements, $(\mathbf{g}_{ji}\mathbf{g}_{ji}^H)_{m,m}\triangleq |g_{ji}^{m}|^2$, follow an exponential distribution with parameter one, i.e., $(\mathbf{g}_{ji}\mathbf{g}_{ji}^H)_{m,m}\sim\mathrm{Exp}(1), \forall m\in\{1,\ldots,M\}$ and the off-diagonal elements, $(\mathbf{g}_{ji}\mathbf{g}_{ji}^H)_{m,l}, \forall m\neq l$, follow the distribution of a product of two independent Gaussians with zero mean and unit variance (see \cite{OM12TSP}). The distribution of each element in $\mathbf{g}_{ji}\mathbf{g}_{ji}^H$ is not varying with respect to $M$ while only the size of matrix grows according to $M$. As a result, as $M\rightarrow\infty$, the denominator $\mathbf{g}_{ji}^H \mathbf{g}_{ji}$ goes to infinity, while all the elements in the numerator $\mathbf{g}_{ji}\mathbf{g}_{ji}^H$ remain as constant values with respect to $M$. Hence, $\frac{\mathbf{g}_{ji}\mathbf{g}_{ji}^H}{\mathbf{g}_{ji}^H \mathbf{g}_{ji}} \rightarrow \mathbf{0}$ as $M\rightarrow \infty$. Accordingly, $\mathbf{w}_{j}^{\star} \rightarrow \frac{\mathbf{h}_{j\mathcal{D}}}{\|\mathbf{h}_{j\mathcal{D}}\|}$ and $\gamma_{j\mathcal{D}}^{(i,j)}(t) \rightarrow \rho_{\mathcal{R}}\|\mathbf{h}_{j\mathcal{D}}\|^2$ as $M \rightarrow \infty$, which completes the proof.
\end{proof}
\end{proposition}
\begin{remark}\label{rm:ZFBF_relays}
The proposed ZFBF-based RS scheme also approach the average end-to-end rate of ideal FD relaying as the number of relays ($K$) goes to infinity due to increased selection diversity. However, increasing the number of relays cannot guarantee to achieve the performance of the ideal FD relaying as in Proposition~\ref{prop:ZFBF_antennas} since its selection diversity is always less than or equal to that of the ideal FD relaying. In other words, the best transmitting relay should meet $\frac{\mathbf{g}_{ji}\mathbf{g}_{ji}^H}{\mathbf{g}_{ji}^H \mathbf{g}_{ji}} = \mathbf{0}$ for achieving the same performance as the ideal FD relaying. Although there exist certain relays satisfying $\frac{\mathbf{g}_{ji}\mathbf{g}_{ji}^H}{\mathbf{g}_{ji}^H \mathbf{g}_{ji}} \approx \mathbf{0}$ with high probability as $K$ goes to infinity, they are a subset of the set of relays while the ideal FD relaying can always take the full selection diversity due to no IRI assumption.
\end{remark}
\begin{remark}\label{rm:optBF}
If the ZFBF solution in \eqref{eq:q_ZFBF} is set to the initial vector for $\mathbf{w}_j$ in the first step of the proposed iterative optimal BF, it always yields a better solution than the ZFBF, regardless of the number of iterations, since $\underset{\mathbf{u}_i}{\mathrm{argmax}}\frac{\rho_{\mathcal{S}}\mathbf{u}_i^H\mathbf{h}_{\mathcal{S}i} \mathbf{h}_{\mathcal{S}i}^H {\mathbf{u}_i}}{\mathbf{u}_i^H \left(\rho_{\mathcal{R}}\mathbf{H}_{ji}\mathbf{w}_{j}\mathbf{w}_{j}^H\mathbf{H}_{ji}^H + \mathbf{I}\right) {\mathbf{u}_i}}=\underset{\mathbf{u}_i}{\mathrm{argmax}}\frac{\rho_{\mathcal{S}}\mathbf{u}_i^H\mathbf{h}_{\mathcal{S}i} \mathbf{h}_{\mathcal{S}i}^H {\mathbf{u}_i}}{\mathbf{u}_i^H{\mathbf{u}_i}}=\frac{\mathbf{h}_{\mathcal{S}i}}{\|\mathbf{h}_{\mathcal{S}i}\|}$ which yields exactly the same BF pair of the ZFBF-based RS scheme and an additional iteration gives a better solution.
We therefore propose to initialize the alternating optimization in this way, even though there are no guarantees that this will provide the global optimum.
\end{remark}
\subsection{Proposed Relay Selection Scheme with Minimum Mean Square Error (MMSE)-based IRI Suppression}
In this subsection, we propose a receive beamformer based on minimum mean square error (MMSE) for maximizing the effective SINR at the receiving relay.
First of all, we use the MRT beamformer for the transmitting relay $j$, i.e., $\mathbf{w}_j = \frac{{\mathbf{h}_{j\mathcal{D}}}}{\| \mathbf{h}_{j\mathcal{D}} \|}$, and then we find a receive beamformer for maximizing the effective SINR at the receiving relay.
Therefore, for given relay pair $(i,j)$ and $\mathbf{w}_j$, the following optimization problem is formulated:
\begin{subequations}
\label{eq:OPT_maxSNR}
\begin{IEEEeqnarray}{cl}
\underset{\mathbf{u}_i}{\mbox{max}} &\quad \frac{{ \rho_{\mathcal{S}} }\mathbf{u}_i^H \mathbf{h}_{\mathcal{S}i} \mathbf{h}_{\mathcal{S}i}^H {\mathbf{u}_i}}{\mathbf{u}_i^H \left({ \rho_{\mathcal{R}} } \mathbf{H}_{ji}\mathbf{w}_{j}\mathbf{w}_{j}^H\mathbf{H}_{ji}^H + { \mathbf{I} }\right) {\mathbf{u}_i}}\\
\mbox{s. t.} & \quad \|\mathbf{u}_{i}\|=1.
\end{IEEEeqnarray}
\end{subequations}
The solution of \eqref{eq:OPT_maxSNR} is given by the scaled MMSE as $\mathbf{u}_i = c_{m}\left(\rho_{\mathcal{R}}\mathbf{H}_{ji}\mathbf{w}_{j}\mathbf{w}_{j}^H\mathbf{H}_{ji}^H + \mathbf{I}\right)^{-1}\mathbf{h}_{\mathcal{S}i}$ where the scaling factor $c_{m}$ is chosen such that $\|\mathbf{u}_{i}\|=1$.
Therefore, substituting $\mathbf{u}_i$ and $\mathbf{w}_j$ into \eqref{eq:SINRi} and \eqref{eq:SNRD}, the instantaneous SNRs for the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links are expressed, respectively, as:
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t)
&= {\rho_{\mathcal{S}} \mathbf{h}_{\mathcal{S}i}^H\left(\frac{\rho_{\mathcal{R}}}{\|\mathbf{h}_{j\mathcal{D}}\|^2}\mathbf{H}_{ji}\mathbf{h}_{j\mathcal{D}}\mathbf{h}_{j\mathcal{D}}^H\mathbf{H}_{ji}^H + \mathbf{I}\right)^{-1}\mathbf{h}_{\mathcal{S}i},} \label{eq:sinr_si_mmse}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t) &=\rho_{\mathcal{R}} \|\mathbf{h}_{j\mathcal{D}}\|^2.\label{eq:sinr_jd_mmse}
\end{align}
Substituting \eqref{eq:sinr_si_mmse} and \eqref{eq:sinr_jd_mmse} into \eqref{eq:C_si} and \eqref{eq:C_jd}, respectively, the best relay pair is selected by \eqref{eq:backpressure}.
\begin{proposition
\label{prop:MMSE_lowSNR}
The proposed MMSE-based RS scheme asymptotically achieves the average end-to-end rate of ideal FD relaying at low SNR.
\begin{proof}
As $\rho_R \to 0$, $\gamma_{\mathcal{S}i}^{(i,j)}(t) \to \rho_{\mathcal{S}}\mathbf{h}_{\mathcal{S}i}^H \mathbf{I}^{-1}\mathbf{h}_{\mathcal{S}i}=\rho_{\mathcal{S}}\|\mathbf{h}_{\mathcal{S}i}\|^2$, which completes the proof.
\end{proof}
\end{proposition}
\subsection{Proposed Relay Selection Scheme with Orthonormal Basis (OB)-based IRI Cancellation}
In this subsection, we propose a perfect IRI cancellation BF based on orthonormal basis vectors.
To this end, we first generate two random orthonormal vectors $\mathbf{u}$ and $\mathbf{q}$, i.e., $\mathbf{u}^H\mathbf{q}=0$, $\Vert\mathbf{u}\Vert=1$, and $\Vert\mathbf{q}\Vert=1$.
Then, we use $\mathbf{u}$ as the receive beamformer at the receiving relay and $\mathbf{w}_{j}=\frac{\mathbf{H}_{ji}^{-1}\mathbf{q}}{\Vert\mathbf{H}_{ji}^{-1}\mathbf{q}\Vert}$ as the transmit beamformer at the transmitting relay, respectively.
Since $\mathbf{u}^H \mathbf{H}_{ji} \mathbf{w}_{j}=\mathbf{u}^H \mathbf{H}_{ji} \frac{\mathbf{H}_{ji}^{-1} \mathbf{q}}{\Vert\mathbf{H}_{ji}^{-1} \mathbf{q}\Vert}=0$, $\|\mathbf{u}\|=1$, and $\|\mathbf{w}_j\|=1$, substituting $\mathbf{u}$ and $\mathbf{w}_j$ into \eqref{eq:SINRi} and \eqref{eq:SNRD}, the instantaneous SNRs for the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links are expressed as:
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t) &= \rho_{\mathcal{S}} |\tilde{h}_{\mathcal{S}i}|^2,\label{eq:sinr_si_ob}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t) &= \rho_{\mathcal{R}}|\tilde{h}_{j\mathcal{D}}|^2,\label{eq:sinr_jd_ob}
\end{align}
where $\tilde{h}_{\mathcal{S}i} = \mathbf{u}^H \mathbf{h}_{\mathcal{S}i} \sim \mathcal{CN}(0,\sigma_{\mathcal{S}\mathcal{R}}^2)$ and $\tilde{h}_{j\mathcal{D}} = \mathbf{h}_{j\mathcal{D}}^H\mathbf{w}_{j} \sim \mathcal{CN}(0,\sigma_{\mathcal{R}\mathcal{D}}^2)$.
As a result, substituting \eqref{eq:sinr_si_ob} and \eqref{eq:sinr_jd_ob} into \eqref{eq:C_si} and \eqref{eq:C_jd}, respectively, the best relay pair is selected by \eqref{eq:backpressure}.
\begin{remark}
From \eqref{eq:sinr_si_ob} and \eqref{eq:sinr_jd_ob}, the proposed OB-based RS scheme achieves the average end-to-end rate of ideal FD relaying with a single antenna at the relays.
\end{remark}
\begin{remark}\label{rm:ZFBF_OB}
The proposed ZFBF-based RS scheme is always better than the proposed OB-based RS scheme in terms of the average end-to-end rate while both of the schemes perfectly cancel IRI. However, it is not always better in the viewpoint of instantaneous RS and BF, since the randomly chosen beamformers in the proposed OB-based RS scheme might be better at a certain RS instance for not the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link rate but the weighted sum-rate.
\end{remark}
\subsection{Proposed SINR-based Relay Selection Scheme with Beamforming Neglecting IRI}
Although the IRI-free BF is not optimal in the presence of IRI, we propose to use them and utilize effective SINR/SNR measures after BF in RS as the simplest joint RS and BF scheme.
Accordingly, for relay pair $(i,j)$, the receive BF vector is given by $\mathbf{u}_i = \frac{{\mathbf{h}_{\mathcal{S}i}}}{\| \mathbf{h}_{\mathcal{S}i} \|}$
and the transmit BF vector is given by $\mathbf{w}_j = \frac{{\mathbf{h}_{j\mathcal{D}}}}{\Vert \mathbf{h}_{j\mathcal{D}} \Vert}$.
Substituting $\mathbf{u}_i$ and $\mathbf{w}_j$ into \eqref{eq:SINRi} and \eqref{eq:SNRD},
the instantaneous SINR and SNR of both the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} i\}}$ and $\mathop{\{j \hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links are obtained, respectively, by
\begin{align}
\gamma_{\mathcal{S}i}^{(i,j)}(t)
&=\frac{ \Vert\mathbf{h}_{\mathcal{S}i}\Vert^2 \rho_{\mathcal{S}}}{1 + \displaystyle\frac{|\mathbf{h}_{\mathcal{S}i}^H\mathbf{H}_{ji} {\mathbf{h}_{j\mathcal{D}}} |^2}{\Vert\mathbf{h}_{\mathcal{S}i}\Vert^2\Vert\mathbf{h}_{j\mathcal{D}}\Vert^2}\rho_{\mathcal{R}} },
\label{eq:sinr_cap_si}\\
\gamma_{j\mathcal{D}}^{(i,j)}(t) &= \rho_{\mathcal{R}} \Vert\mathbf{h}_{j\mathcal{D}}\Vert^2.
\label{eq:sinr_cap_jd}
\end{align}
Substituting \eqref{eq:sinr_cap_si} and \eqref{eq:sinr_cap_jd} into \eqref{eq:C_si} and \eqref{eq:C_jd}, respectively, the best relay pair is selected by \eqref{eq:backpressure}.
\subsection{Discussion on Complexity of the Proposed Joint Relay Selection and Beamforming}\label{subsec:complexity}
For the RS protocol, an exhaustive search within $K\times(K-1)$ combinations with global CSI and BSI is required to obtain the optimal performance. Accordingly, the complexity of the optimal relay pair selection is $\mathcal{O}(K^2)$ for all the schemes. However, through the distributed RS approach, CSIs for $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ links can be directly estimated at each relay from pilot signals and thus the amount of feedback on CSI can be reduced from $(2KM + M^2K(K-1)/2)$ to $KM$. It is also worth mentioning that the number of available (fixed) relays in practical network scenarios would not be so large due to geographical limitations. On the other hand, the proposed schemes can still be effective even for a two-relay network if sufficient number of antennas are available at the relays. Moreover, as will be shown in Section \ref{subsec:rate}, the performance enhancement from more than five relays would be marginal.
\Sumin{For BF design for given relay pair, if we denote the complexity for computing a BF vector including $M$ dimensional matrices and/or vectors by $C_{\mathrm{BF}}$,
the complexity of all the suboptimal BF schemes becomes $2C_{\mathrm{BF}}$
since they require to compute two BF vectors.
However, they can have different computational times in practice according to degree of matrix and/or vector computations. For instance, the MMSE-based scheme requires to compute $M$ dimensional matrix inversion, while the SINR-based scheme only requires vector normalizations.
On the other hand, the complexity of the optimal BF scheme is derived as $L(2C_{\mathrm{BF}} + C_{\beta})
, where $L$ denotes the number of iterations and $C_{\beta}$ denotes the complexity to find the optimal $\beta$ including both a rough line search and some Gauss-Newton steps.}
Consequently, the optimal BF scheme requires additional complexity for iteration process and an additional optimization parameter, compared to the suboptimal BF schemes.
After all, the suboptimal BF schemes are useful for online operation and a network which consists of nodes with low computation power.
\section{Performance Evaluation} \label{sec:performance_evaluation}
In this section, we evaluate the proposed joint buffer-aided RS and BF schemes in terms of the average end-to-end rate and average delay through Monte-Carlo simulations, compared to conventional HD RS schemes and SFD-MMRS scheme representing state-of-the-art in the literature. For multiple-antenna extension of the conventional schemes, we suppose that they use the IRI-free beamformers.
As an upper bound of the average end-to-end rate, we consider the optimal joint RS and BF in \eqref{eq:weighted_opt} assuming no IRI.
The optimal weight factors $\{\alpha_k^{\star}\}$ are adaptively obtained based on the back-pressure algorithm during a pre-training phase and applied to each relay pair selection.
We assume half-full initial buffer state at all relays for the pre-training phase but zero initial buffer state for data transmission phase.
We consider Rayleigh block fading channels with average gains of $\sigma_{\mathcal{S}\mathcal{R}_i}^2$, $\sigma_{\mathcal{R}_j\mathcal{D}}^2$, and $\sigma_{\mathcal{R}_j\mathcal{R}_i}^2$ for $i,j\in\mathcal{K}$, 10000 packet transmissions from the source, and assume $P_{\mathcal{S}} = P_{\mathcal{R}}$ throughout all simulations.
\subsection{Benchmarks}
\subsubsection{HD Best Relay Selection (BRS) Scheme \cite{BKR+06JSAC}}
In the HD-BRS scheme without buffering at relays, the best relay is determined by
\begin{align}
i^{\star} &= \underset{i \in \mathcal{K}}{\mathrm{argmax}}~\min\left\{C_{\mathcal{S}i}(t), C_{i\mathcal{D}}(t)\right\},
\end{align}
where $C_{\mathcal{S}i}(t) = \frac{1}{2}\log_2\left(1 + \gamma_{\mathcal{S}i}(t)\right)$ and $C_{i\mathcal{D}}(t) = \frac{1}{2}\log_2\left(1 + \gamma_{i\mathcal{D}}(t)\right)$.
\subsubsection{HD $\mathrm{max-max}$ Relay Selection (MMRS) Scheme \cite{IMS12TWC}}
In the HD-MMRS scheme maintaining the two-phase operation, the best relay at first time slot, $i^{\star}$, and the best relay at second time slot, $j^{\star}$, are selected as follows:
\begin{align}
i^{\star} &= \underset{i\in \mathcal{K}}{\mathrm{argmax}}~C_{\mathcal{S}i}(2t),\quad
j^{\star} = \underset{j\in \mathcal{K}}{\mathrm{argmax}}~C_{j\mathcal{D}}(2t+1)
\end{align}
where $C_{\mathcal{S}i}(2t) = \min \big\{ \frac{1}{2}\log_2\left(1 + \gamma_{\mathcal{S}i}(2t)\right), B_{max} - B_i(2t-1) \big\}$ and $C_{j\mathcal{D}}(2t+1) = \min \big\{ \frac{1}{2}\log_2(1 + \gamma_{j\mathcal{D}}(2t+1)),$ $B_j(2t) \big\}$.
\subsubsection{HD $\mathrm{max-link}$ Relay Selection (MLRS) Scheme \cite{KCT12TWC}}
Since the HD-MLRS scheme has been developed for fixed rate transmission, we slightly modify it to adaptive rate transmission by adding a link selection parameter $d_i$ for the $i$-th relay.
In the conventional HD-MLRS scheme releasing the two-phase operation condition, the best relay $i^{\star}$ and the best link $d_{i^{\star}}$ (integer variable; receiving 1 or transmitting 0) at each time slot are determined by
\begin{align}
\left(i^{\star}, d_{i^{\star}}\right) = \underset{i\in \mathcal{K},d_i\in\{0,1\}}{\mathrm{argmax}} \left\{ d_i C_{\mathcal{S}i}(t) + (1-d_i) C_{i\mathcal{D}}(t) \right\},
\end{align}
where $C_{\mathcal{S}i}(t) = \min \left\{ \frac{1}{2}\log_2\left(1 + \gamma_{\mathcal{S}i}(t)\right), B_{max} - B_i(t-1) \right\}$ and $C_{i\mathcal{D}}(t) = \min \big\{ \frac{1}{2}\log_2\left(1 + \gamma_{i\mathcal{D}}(t)\right),$ $B_i(t-1) \big\}$.
\subsubsection{SFD-MMRS Scheme \cite{IKS12TVT}}
In the SFD-MMRS scheme, each $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ or $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link selects the best relay and the second best relay based on channel gains without consideration of IRI. Denote the relay indices of the best and second best relays by $i_1$ and $j_2$ for the receiving relay and $i_1$ and $j_2$ for the transmitting relay, respectively. Then they are selected as follows:
\begin{align}
i_1 &= \underset{i\in\mathcal{K}}{\mathrm{argmax}}~C_{\mathcal{S}i}(t),
\quad i_2 = \underset{i\in\mathcal{K}\backslash\{i_1\}}{\mathrm{argmax}}~C_{\mathcal{S}i}(t), \nonumber\\
j_1 &= \underset{j\in\mathcal{K}}{\mathrm{argmax}}~C_{j\mathcal{D}}(t),
\quad j_2 = \underset{j\in\mathcal{K}\backslash\{j_1\}}{\mathrm{argmax}}~C_{j\mathcal{D}}(t), \nonumber
\end{align}
where $C_{\mathcal{S}i}(t) = \min \left\{ \log_2\left(1 + \gamma_{\mathcal{S}i}(t)\right), B_{max} - B_i(t-1) \right\}$ and $C_{j\mathcal{D}}(t) = \min \big\{ \log_2\left(1 + \gamma_{j\mathcal{D}}(t)\right),$ $B_j(t-1) \big\}$.
If the best relays for both links are same, it finds the best relay pair among combinations with the second best relays based on a minimum of achievable rates, i.e.,
\begin{align}
(i^{\star}, j^{\star}) = \left\{
\begin{array}{lll}
(i_1, j_1), & \mbox{\text{if $i_1\neq j_1$}}\\
(i_2, j_1), & \mbox{\text{if $i_1= j_1$ and $\min\{C_{\mathcal{S}i_2}(t), C_{j_1\mathcal{D}}(t)\}$}}\\
& \hspace{18mm}\mbox{\text{$> \min\{C_{\mathcal{S}i_1}(t), C_{j_2\mathcal{D}}(t)\}$}}\\
(i_1, j_2), & \mbox{\text{otherwise.}}
\end{array}
\right.
\end{align}
Since the SFD-MMRS scheme assumes a single antenna at the relays and no IRI, we extend it to multiple antennas using the IRI-free BF, but include the actual IRI in the results marked by ``SFD-MMRS-IRI''.
The performance degradation of the non-ideal SFD-MMRS scheme due to IRI is shown as numerical results in the following subsections.
\subsection{Average End-to-End Rate}\label{subsec:rate}
\subsubsection{i.i.d. Channel Case ($\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
Fig.~\ref{fig:C_K2_M2} shows the average end-to-end rate for varying SNR when $K=2$, $M=2$, $B_{max}\rightarrow\infty$, and the average channel qualities of all the links are identical.
The ideal SFD-MMRS scheme performs close to the upper bound obtained by \eqref{eq:weighted_opt}.
This validates that our weighted sum-rate maximization based on instantaneous rates works well to maximize the average end-to-end rate.
If we impose IRI into the SFD-MMRS scheme, its performance is significantly degraded with increasing SNR, i.e., in the interference-limited regime.
Although the proposed SINR-based RS scheme improves the average end-to-end rate, its contribution is not significant at medium/high SNR.
On the contrary, the average end-to-end rates of the other proposed schemes still increase with increasing SNR due to IRI cancellation/suppression.
Since the proposed OB-based RS scheme achieves the single antenna upper bound, the proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes outperform the single antenna upper bound.
While the optimal BF optimizes both transmit and receive beamformers iteratively, the ZFBF and MMSE BF optimize just one of the beamformers fixing the other beamformer.
As a result, the ZFBF-based and MMSE-based RS schemes achieve lower average end-to-end rates than the optimal BF-based RS scheme. Moreover, the MMSE-based RS scheme achieves better performance at low/medium SNR than the ZFBF-based RS scheme, since the MMSE BF can achieve the ideal upper bound at low SNR regime as proved in Proposition~\ref{prop:MMSE_lowSNR}.
Regarding the conventional HD RS schemes, the HD-MLRS and HD-MMRS schemes outperform the HD-BRS scheme since they additionally utilize buffering at relays. Furthermore, the HD-MLRS scheme outperforms the HD-MMRS scheme since it obtains more diversity gain by releasing the two-phase operation condition.
Compared to the HD RS schemes, the slopes of curves of the proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes are almost double. For instance, at $\mathrm{SNR}=30$ dB, the proposed optimal BF-based RS scheme excesses twice the average end-to-end rate by the HD-BRS scheme and the proposed ZFBF-based and MMSE-based RS schemes achieve slightly less performance than it.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig2.eps}
\vspace{-3mm}
\caption{Average end-to-end rate vs. SNR for i.i.d. channel case ($K=2$, $M=2$, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:C_K2_M2}
\vspace{-5mm}
\end{figure}
Fig.~\ref{fig:C_varying_M} shows the average end-to-end rates for varying number of antennas at the relays when $K=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, and i.i.d. IRI channel conditions.
Except for the proposed OB-based RS scheme,
the other schemes basically increase the average end-to-end rate as the number of antennas increases.
The proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes converge to the upper bound when increasing the number of antennas, as shown in Proposition~\ref{prop:ZFBF_antennas} for the ZFBF-based RS scheme and stated in Remark~\ref{rm:optBF} for the optimal BF-based RS scheme.
Eventually, it is worth noting that the proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes can fully recover the loss of multiplexing gain caused by the HD relaying as the number of antennas increases.
Even if the number of antennas increases, the average end-to-end rates achieved by the SINR-based RS scheme and the non-ideal SFD-MMRS scheme are still less than those by the HD RS schemes due to the effects of strong IRI and limited selection diversity.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig3.eps}
\vspace{-3mm}
\caption{Effect of the number of antennas ($K=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:C_varying_M}
\vspace{-5mm}
\end{figure}
Fig.~\ref{fig:C_varying_K} shows the average end-to-end rate for varying the number of relays.
All the schemes achieve improved average end-to-end rates due to the increased selection diversity gain as the number of relays increases.
Differently from in Fig.~\ref{fig:C_varying_M}, the proposed SINR-based RS scheme obtains significant gains with increasing number of relays.
Accordingly, it outperforms the conventional HD RS schemes when $K>3$.
In contrast, the non-ideal SFD-MMRS scheme is still the worst with marginal improvements.
This is because it never considers IRI at all while the SINR-based RS scheme takes IRI into account at the RS.
The optimal BF-based, ZFBF-based, and MMSE-based RS schemes also approach the upper bound as the number of relays increases.
Thus, they can asymptotically recover the loss of multiplexing gain with respect to the number of relays even if the exact convergence is not guaranteed in this case.
However, it is shown that the rate improvement is much slower than when increasing the number of antennas.
When $K=10$, all the proposed schemes except for the SINR-based RS scheme achieve greater than or equal to double the average end-to-end rate of the HD-BRS scheme.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig4.eps}
\vspace{-3mm}
\caption{Effect of the number of relays ($M=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:C_varying_K}
\vspace{-5mm}
\end{figure}
\subsubsection{i.i.d. Channel Cases with Different IRI Intensities}
To investigate the effect of average IRI intensity, we consider two different average IRI channel conditions: (i) \emph{weak IRI case} ($\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=-10~\mathrm{dB}, \sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=0~\mathrm{dB}, \forall i,j\in\mathcal{K}$); (ii) \emph{strong IRI case} ($\sigma_{\mathcal{R}\mathcal{R}}^2=10~\mathrm{dB}, \sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=0~\mathrm{dB}, \forall i,j\in\mathcal{K}$).
Fig.~\ref{fig:nid_C_K3_M4}~(a) shows the average end-to-end rate for weak IRI case when $K=3$, $M=4$, and $B_{max}\rightarrow\infty$.
When $K=3$ and $M=4$, the proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes already almost approach the upper bound regardless of SNR.
The optimal BF-based RS scheme outperforms the other schemes and the MMSE-based RS scheme slightly outperforms the ZFBF-based RS scheme at low SNR.
While the OB-based RS scheme still achieves the single antenna upper bound regardless of the average IRI intensity, the SINR-based RS scheme and the non-ideal SFD-MMRS scheme always outperform the HD RS schemes and outperform the OB-based RS scheme at low/medium SNR. Especially, when $\mathrm{SNR}=0$ dB, the non-ideal SFD-MMRS scheme achieves almost the same performance as the ideal upper bound since this channel condition yields very weak interference which is negligible.
Fig.~\ref{fig:nid_C_K3_M4}~(b) shows the average end-to-end rate for strong IRI case in the same setup.
The proposed ZFBF-based and OB-based RS schemes yield exactly the same performance as in Fig.~\ref{fig:nid_C_K3_M4}~(a) since they do not depend on the intensity of IRI due to perfect IRI cancellation. The proposed MMSE-based RS scheme achieves almost identical performance as the proposed ZFBF-based RS scheme even at low SNR since the intensity of IRI is already strong compared to the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ channel conditions. The optimal BF-based RS scheme performs between the upper bound and the MMSE-based and ZFBF-based RS schemes regardless of the intensity of IRI and SNR. In contrast, the non-ideal SFD-MMRS scheme significantly degrades the average end-to-end rate which is always worse than those of the HD RS schemes.
\begin{figure}[tp]
\centering
\subfigure[]{
\includegraphics[width=8.6cm]{./fig/fig5a.eps}}
\subfigure[]{
\includegraphics[width=8.6cm]{./fig/fig5b.eps}}
\caption{Effect of average IRI intensity ($K=3$, $M=4$, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$) (a) Weak IRI ($\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=-10~\mathrm{dB},\forall i,j\in\mathcal{K}$) (b) Strong IRI ($\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=10~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:nid_C_K3_M4}
\vspace{-5mm}
\end{figure}
\subsubsection{non-i.i.d. Channel Case}
As a more practical network scenario, we consider a network with three relays where the links have different average channel gains as $[\sigma_{\mathcal{S}\mathcal{R}_1}^2,\sigma_{\mathcal{S}\mathcal{R}_2}^2,\sigma_{\mathcal{S}\mathcal{R}_3}^2]=[1,0,-1]~\mathrm{dB}$, $[\sigma_{\mathcal{R}_1\mathcal{D}}^2,\sigma_{\mathcal{R}_2\mathcal{D}}^2,\sigma_{\mathcal{R}_3\mathcal{D}}^2]=[-1,0,1]~\mathrm{dB}$, and $\sigma_{\mathcal{R}_1\mathcal{R}_2}^2=\sigma_{\mathcal{R}_2\mathcal{R}_1}^2=\sigma_{\mathcal{R}_2\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_2}^2=0~\mathrm{dB},\sigma_{\mathcal{R}_1\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_1}^2=-1~\mathrm{dB}$.
In this setup, $\{\mathcal{S}-\mathcal{R}_2-\mathcal{D}\}$ are equally spaced, $\mathcal{R}_1$ is closer to $\mathcal{S}$ and $\mathcal{R}_3$ is closer to $\mathcal{D}$.
$\{\mathcal{R}_1-\mathcal{R}_2-\mathcal{R}_3\}$ are equally spaced and therefore, $\{\mathcal{R}_1-\mathcal{R}_3\}$ becomes a less interfered channel than $\{\mathcal{R}_1-\mathcal{R}_2\}$ and $\{\mathcal{R}_2-\mathcal{R}_3\}$.
Fig.~\ref{fig:C_nid} shows the average end-to-end rate with various SNR values for the non-i.i.d. channel case.
The basic trend is the same as that for i.i.d. channel cases shown previously. Hence, our proposed weighted sum-rate maximization approach works well also under non-identical channel condition.
Furthermore, the proposed optimal BF-based, ZFBF-based, MMSE-based RS schemes are still effective for generalized asymmetric network topologies.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig6.eps}
\vspace{-3mm}
\caption{Average end-to-end rate vs. SNR for non-i.i.d. channel case ($K=3$, $M=2$, $[\sigma_{\mathcal{S}\mathcal{R}_1}^2,\sigma_{\mathcal{S}\mathcal{R}_2}^2,\sigma_{\mathcal{S}\mathcal{R}_3}^2]=[1,0,-1]~\mathrm{dB}$, $[\sigma_{\mathcal{R}_1\mathcal{D}}^2,\sigma_{\mathcal{R}_2\mathcal{D}}^2,\sigma_{\mathcal{R}_3\mathcal{D}}^2]=[-1,0,1]~\mathrm{dB}$, $\sigma_{\mathcal{R}_1\mathcal{R}_2}^2=\sigma_{\mathcal{R}_2\mathcal{R}_1}^2=\sigma_{\mathcal{R}_2\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_2}^2=0~\mathrm{dB},\sigma_{\mathcal{R}_1\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_1}^2=-1~\mathrm{dB}$)}
\label{fig:C_nid}
\vspace{-5mm}
\end{figure}
\subsection{Delay Performance}\label{subsec:delay}
Basically, the buffer-aided relaying obtains additional selection diversity gain by sacrificing delay performance. Even if we mainly focus on the average end-to-end rate of delay-tolerant applications, we evaluate the average delay performance through simulations.
Due to a full-queue assumption at the source, the average delay is defined as the average queueing delay at relays, which implies the time difference between the arrival time of a single packet at a relay and the successfully received time of the packet at the destination in number of time slots (number of channel uses).
A single packet is transmitted from the source at each time slot and its size is determined by the selected $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ channel gain according to adaptive rate transmission.
Therefore, a delay of one means that a packet is stored at the relay in a certain time slot and all information bits contained in the packet are successfully forwarded to the destination in the next time slot. \Sumin{As stated in Section~\ref{sec:performance_evaluation}, we also consider 10000 packet transmissions from the source for the delay performance in following.}
In Fig.~\ref{fig:delay_varying_M}, we show the average delay for varying number of antennas when $K=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, and $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$.
Except for the proposed OB-based RS scheme, the average delays of all the schemes decrease as the number of antennas increases since the effective channel gain increases with the number of antennas. In the proposed OB-based RS scheme, the average delay is varying with the values slightly less than 50 time slots because its effective channel gain is same as the single antenna case.
The ideal SFD-MMRS scheme has moderate average delay and the non-ideal SFD-MMRS scheme and the SINR-based RS scheme have very short delays approaching one. The reason is that both schemes have the bottleneck in $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link due to uncoordinated IRI under i.i.d. channel condition and thus the size of packets transmitted from the source to a relay is much smaller than average $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rate.
On the contrary, the proposed optimal BF-based, ZFBF-based, and MMSE-based RS schemes have longer delays than that of the ideal SFD-MMRS scheme. In particular, the proposed ZFBF-based RS scheme has a significant delay when $M=2$ since its bottleneck is the $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link while the bottleneck of the MMSE-based RS scheme is the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link in average sense. After all, the average packet size from the source in the ZFBF-based RS scheme is larger but its average $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rate is smaller than those of the MMSE-based RS scheme.
However, the ZFBF-based RS scheme recovers the loss in the effective SNR at the destination and thus the average delay is significantly reduced when $M\geq 4$.
Similarly, in the optimal BF-based RS scheme, both average $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rates are more balanced than those in the MMSE-based RS scheme.
This results in a larger average $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link rate and a smaller average $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rate for the optimal BF-based RS scheme, compared to the MMSE-based RS scheme.
Accordingly, the average delay of the optimal BF-based RS scheme becomes worse than that of the MMSE-based RS scheme, since its average packet size is larger but its average $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rate is smaller than those of the MMSE-based RS scheme.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig7.eps}
\vspace{-3mm}
\caption{Average delay for varying number of antennas ($K=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:delay_varying_M}
\vspace{-5mm}
\end{figure}
Fig.~\ref{fig:delay_varying_K} shows the average delay for varying number of relays in the same setup.
The basic trend of the ideal SFD-MMRS scheme increases as the number of relays increases since the more relays exist, each relay is less likely to be selected.
The SINR-based RS scheme follows the similar trend while the non-ideal SFD-MMRS scheme has still very short delays approaching one. This is because the SINR-based RS scheme can provide much larger average link rates as the number of relays increases as shown in Fig.~\ref{fig:C_varying_K}.
The other proposed schemes optimizing beamformers have a different trend where the average delay decreases until $K=3$ ($K=5$ in the ZFBF-based RS scheme) and increases again as the number of relays increases further.
When $K=2$ and $M=2$, the minimum network setup, the degrees of freedom optimizing transmit and/or receive beamformers is too restricted but one additional relay gives an additional degree of freedom improving link rates. Hence, the average delays of the proposed schemes are rather decreased when $K=3$, but increased again after then, since the additional gain is almost saturated after a certain moment compared to the reduced selection opportunity per relay.
\Sumin{From Figs.~\ref{fig:delay_varying_M} and \ref{fig:delay_varying_K}, since the proposed ZFBF-based RS scheme requires too much delay at the minimum network setup ($K=2, M=2$), the MMSE-based RS scheme is more desirable for this setup in the perspective of the average delay performance.}
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig8.eps}
\vspace{-3mm}
\caption{Average delay for varying number of relays ($M=2$, $\mathrm{SNR}=20$ dB, $B_{max}\rightarrow\infty$, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:delay_varying_K}
\vspace{-5mm}
\end{figure}
\subsection{Behavior of the Optimal Weight Factors}\label{subsec:opt_alpha}
In this subsection, we investigate the behavior of the optimal $\alpha_k$ parameters for the proposed schemes in two different channel cases: \emph{i.i.d.} and \emph{non-i.i.d.} channel cases with three relays.
In simulations, to find the optimal $\alpha_k$ values, we employ the back-pressure algorithm approach using a pre-training phase as stated in \eqref{eq:backpressure}.
Fig.~\ref{fig:opt_a_K3_M2} shows the optimal $\alpha_k$ values obtained after the pre-training phase when $K=3$ and $M=2$.
Fig.~\ref{fig:opt_a_K3_M2}~(a) shows an i.i.d. channel case where all the links have an identical average channel quality as $0$ dB.
In the figure, all $\alpha_k$ values for each scheme are identical due to the i.i.d. channel condition.
In principle, if $\alpha_k$ is close to one, the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link rate dominates the cost function and if $\alpha_k$ is close to zero, the $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link rate dominates the cost function in the relay pair selection.
The upper bound and the optimal BF-based RS scheme balance both link rates as $\alpha_k^{\star}\approx 0.5$ since the optimum is achieved when $\mathbb{E}[C_{\mathcal{SR}}(t)] = \mathbb{E}[C_{\mathcal{RD}}(t)]$.
In contrast, the optimal $\alpha_k$ values of the proposed MMSE-based and SINR-based RS schemes are close to one, while that of the proposed ZFBF-based RS scheme is close to zero. This is because the bottleneck of the MMSE-based and SINR-based RS schemes is the $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ link while the bottleneck of the ZFBF-based RS scheme is the $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ link.
Fig.~\ref{fig:opt_a_K3_M2}~(b) shows a non-i.i.d. channel case, the same setup in Fig.~\ref{fig:C_nid}. For all the schemes, $\alpha_k$'s have different values according to average channel quality and $\alpha_2$'s have almost the same values as those for the i.i.d. channel case since $\mathcal{R}_2$ has the same average channel quality as the i.i.d. channel case. In the figure, it holds that $\alpha_1 < \alpha_2 < \alpha_3$ because $\sigma_{\mathcal{SR}_1}^2 > \sigma_{\mathcal{R}_1\mathcal{D}}^2$, $\sigma_{\mathcal{SR}_2}^2 = \sigma_{\mathcal{R}_2\mathcal{D}}^2$, and $\sigma_{\mathcal{SR}_3}^2 < \sigma_{\mathcal{R}_3\mathcal{D}}^2$, and the optimal $\alpha_k$ parameter tends to be determined in order to balance both $\mathop{\{\mathcal{S}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{R}\}}$ and $\mathop{\{\mathcal{R}\hspace{-0.09cm} \rightarrow \hspace{-0.09cm} \mathcal{D}\}}$ links.
As SNR increases, the differences among $\alpha_k$ values are reduced since imbalance effect among links is diminished at high SNR.
Similarly to the i.i.d. channel case, the upper bound and the optimal BF-based RS scheme are biased at $\alpha_k=0.5$, and the proposed MMSE-based and SINR-based RS schemes are biased to be close to one while the ZFBF-based RS scheme is biased to be close to zero.
\begin{figure}[tp]
\centering
\subfigure[]{
\includegraphics[width=8.6cm]{./fig/fig9a.eps}}
\subfigure[]{
\includegraphics[width=8.6cm]{./fig/fig9b.eps}}
\caption{Optimal $\alpha$ value after a training phase ($K=3$, $M=2$) (a) i.i.d. channel case ($\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$) (b) non-i.i.d. channel case ($[\sigma_{\mathcal{S}\mathcal{R}_1}^2,\sigma_{\mathcal{S}\mathcal{R}_2}^2,\sigma_{\mathcal{S}\mathcal{R}_3}^2]=[1,0,-1]~\mathrm{dB}$, $[\sigma_{\mathcal{R}_1\mathcal{D}}^2,\sigma_{\mathcal{R}_2\mathcal{D}}^2,\sigma_{\mathcal{R}_3\mathcal{D}}^2]=[-1,0,1]~\mathrm{dB}$, $\sigma_{\mathcal{R}_1\mathcal{R}_2}^2=\sigma_{\mathcal{R}_2\mathcal{R}_1}^2=\sigma_{\mathcal{R}_2\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_2}^2=0~\mathrm{dB},\sigma_{\mathcal{R}_1\mathcal{R}_3}^2=\sigma_{\mathcal{R}_3\mathcal{R}_1}^2=-1~\mathrm{dB}$)}
\label{fig:opt_a_K3_M2}
\vspace{-5mm}
\end{figure}
\subsection{Effects of Finite Buffer Size}
In practice, the buffer size at relays is finite and thus it restricts the performance since a full-buffer relay cannot be selected as the receiving relay.
We investigate the performance in terms of average end-to-end rate and average delay according to the finite buffer size.
Fig.~\ref{fig:C_avg_finite_buffer_size} shows the average end-to-end rate for varying buffer size when $K=3$, $M=2$, $\mathrm{SNR}=20$ dB, and $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$.
All the schemes rapidly converge to their own performance upper limits with infinite buffer size as the buffer size increases.
Accordingly, a buffer size $B_{\max}\geq 50$ bits/Hz is sufficient to obtain the performance upper limits with infinite buffer size.
For instance, when the bandwidth is 10 MHz, the buffer size is required to be about 60 MB in order to achieve the average end-to-end rate with infinite buffer size.
This value is allowable at relays in the viewpoint of present memory size.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig10.eps}
\vspace{-3mm}
\caption{Average end-to-end rate with finite buffer size ($K=3$, $M=2$, $\mathrm{SNR}=20$ dB, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:C_avg_finite_buffer_size}
\end{figure}
Fig.~\ref{fig:delay_finite_buffer_size} shows the average delay performance with finite buffer size.
For all the schemes, the average delays converge to their own limits as the buffer size increases although the convergence speed is different between the schemes.
Interestingly, the finite buffer size can be helpful to reduce the average delay without a loss in the average end-to-end rate. For example, even if the proposed ZFBF-based RS scheme has the longest delay, the average delay can be less than 15 time slots with achieving the average end-to-end rate limit if the buffer size is set to 50 bits/Hz.
Therefore, if the buffer size is set to an appropriate value, it is enough to achieve near-optimal average end-to-end rate under reasonable average delay.
\begin{figure}[tp]
\centering
\includegraphics[width=8.6cm]{./fig/fig11.eps}
\vspace{-3mm}
\caption{Average delay with finite buffer size ($K=3$, $M=2$, $\mathrm{SNR}=20$ dB, $\sigma_{\mathcal{S}\mathcal{R}_i}^2=\sigma_{\mathcal{R}_j\mathcal{D}}^2=\sigma_{\mathcal{R}_j\mathcal{R}_i}^2=0~\mathrm{dB},\forall i,j\in\mathcal{K}$)}
\label{fig:delay_finite_buffer_size}
\vspace{-5mm}
\end{figure}
\section{Conclusion} \label{sec:conclusion}
In this paper, we proposed virtual FD buffer-aided joint RS and BF schemes taking IRI into account in a buffer-aided multiple relays network, where each relay is equipped with multiple antennas. We first formulated a weighted sum-rate maximization based on instantaneous rates maximizing the average end-to-end rate. Based on the alternative objective function, we proposed various RS schemes based on optimal and suboptimal BF designs to cancel or suppress IRI taking a trade-off between computational complexity and performance into consideration. Through simulations, the proposed joint RS and BF schemes were evaluated in terms of the average end-to-end rate and average delay, compared to several conventional HD RS and SFD-MMRS schemes.
In numerical results, asymptotic trends were investigated with respect to the number of relays and the number of antennas at relays. The proposed joint RS and BF schemes recover the loss of multiplexing gain in the HD relaying even in the presence of IRI as the number of antennas and/or the number of relays increase.
Although the complexities of the proposed joint RS and BF schemes are high due to global CSI required, they can be useful benchmarks in system simulations.
In addition, the behavior of the optimal weight factor and the effects of finite buffer size were shown in various different network setups.
Basic trend in the optimal weight factor is moving toward to reduce the link rate gap between the bottleneck link and the other link.
Although the finite buffer size limits the average end-to-end rate, it can help to bound the average delay with achieving near-optimal average end-to-end rate if it is set to an appropriate value.
For future studies, it is possible to extend to multiple source-destination pairs with multiple antennas, to apply for non-full queue traffic at the source, to consider imperfect CSI and BSI, to develop a low complexity and limited feedback RS scheme, and to apply for other applications such as cognitive radio and physical layer security.
\section*{Acknowledgement}
\Sumin{We would like to thank the anonymous reviewers for constructive comments that helped us improve the quality of this paper including more practical concerns.}
\balance
\bibliographystyle{./style/IEEEtran_v111}
|
\section{Introduction}
It is easy to motivate the study of soft matter systems. Everyday life provides us with the most immediate examples. When we go shopping, we carry home our groceries in plastic bags made from polymers. Most flat screen displays exploit the switching behavior of liquid crystals in an electric field. We use solutions of surfactants when we soap ourselves while taking a shower. And finally, a major part of ourselves is actually soft matter, as almost all biological cells are. Considering the shear amount of soft matter distributed commercially, research activities in this field naturally extend far beyond purely academic purposes.
It is much less straightforward to give a definite, unambiguous, and precise definition of what the term ``soft matter'' actually means or which states of matter it actually includes. Maybe one should proceed the other way around and explain which materials it \textit{ex}cludes by comparison to common other states of matter.
On the one hand, we interpret the term ``softness'' as a strong response to a comparatively weak force \cite{doi2013soft}. This becomes most illustrative when we compare the elastic behavior of metal solids and ordinary rubbery materials. The Young modulus of a rubber can for example be ten thousand times lower than that of steel \cite{treloar1975physics}. I.e.\ applying a tensile force to that piece of rubber can stretch it about ten thousand times further than when applying it to a corresponding piece of steel. In general, due to these huge responses to external forces, the observed behavior of soft materials is markedly nonlinear \cite{doi2013soft}.
On the other hand, the response of soft matter systems to external stimuli is generally comparatively slow \cite{doi2013soft}. The relaxation times in polymers can be hours or days (or even more) \cite{strobl1997physics}, whereas it takes nanoseconds in simple liquids \cite{hansen2006theory}.
It is therefore easy to drive and maintain the systems out of equilibrium, and non-equilibrium effects play a central role in the study of soft materials \cite{doi2013soft}.
One reason for these properties is the size of the building blocks \cite{doi2013soft}. These sizes typically fall into the range of several nanometers up to about a hundred micrometers.
Some examples together with a rough classification of the possible associated dimensions are schematically depicted in Fig.~\ref{fig_lengths}.
Clearly, small rod-like liquid crystalline molecules like PAA with a length of about two nanometers
are on the lower end
\cite{degennes1993physics}. Colloids, i.e.\ mesoscopic particles stabilized in a permanent dispersion, cover a broad range of the considered sizes by definition \cite{dhont1996introduction,palberg1999crystallization,ivlev2012complex}.
On the contrary, a rubber made of sufficiently chemically crosslinked polymer chains can actually be considered as one giant molecule and reach macroscopic dimensions \cite{strobl1997physics}. At this point a strict definition via the size of the constituents becomes more involved.
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure01.pdf}}
\caption{Schematic illustration of typical examples of soft-matter building blocks and materials, together with a rough estimate of their characteristic sizes (horizontal bars indicate a broader range of possible occurring length scales). The rod-like molecule PAA is a common representative of low-molecular-weight liquid crystals \cite{degennes1993physics} that can show an orientationally ordered anisotropic nematic phase. Surfactant molecules combine two antagonistic features on one building block: their head is usually hydrophilic, whereas their tail(s) is (are) mostly hydrophobic \cite{witten2004structured}. This leads, for instance, to the formation of closed bilayer membranes in the form of vesicles \cite{jones2002soft}, with giant unilamellar vesicles (GUVs) as an extreme example (in reality, unlike the schematic, the layer thickness is orders of magnitude smaller than the radius) \cite{hub1982preparation,mueller1983formation,sakuma2011model}. Polymers \cite{strobl1997physics, doi2007theory} in the simplest case can be treated as flexible linear chain-like objects. When covalently chemically crosslinked in the form of rubbers and elastomers (crosslinks are indicated by dots) a macromolecule of macroscopic dimension can be obtained \cite{treloar1975physics,strobl1997physics}. Block copolymers feature chemically different parts (marked by different colors) that can micro-phase separate into regularly ordered structures \cite{hamley1998physics}, here indicated for a lamellar texture. In colloids, the constituting particles are stabilized to form a permanent dispersion \cite{dhont1996introduction,palberg1999crystallization,ivlev2012complex}. Again, different sides of the particles can be functionalized with different chemical properties so that the resulting Janus particles \cite{walther2008janus} form clusters or even regular lattices \cite{chen2012janus,jiang2010janus}. Finally, biological cells such as motile bacteria \cite{harshey2003bacterial,darnton2007torque,kearns2010field} combine a multitude of functionalities. Collective arrangements of biological cells in the form of biological tissue reaches macroscopic dimensions.}
\label{fig_lengths}
\end{figure}
Nevertheless, we can often say that soft matter building blocks are large enough so that quantum effects can be neglected, but small enough that thermal fluctuations play a significant role \cite{hamley2007introduction, jones2002soft}. For example, most of the elastic behavior of rubbers must be attributed to entropy, i.e.\ thermal fluctuations \cite{strobl1997physics, doi2007theory}. Entropy also supports the stabilization of colloidal dispersions as a complement to the necessary artificial stabilization mechanisms \cite{dhont1996introduction,ivlev2012complex}.
Apart from that, the size of the building blocks allows us to use coarse-grained descriptions to characterize their collective physical behavior \cite{jones2002soft}. Chemical details can often be neglected, which leads to a certain degree of ``universality'' in the characterization. This point of view got particularly famous in the field of polymer physics \cite{degennes1979scaling}.
Finally, the size of the constituents makes it possible to arrange different or even antagonistic properties on one building block. Some examples are depicted in Fig.~\ref{fig_lengths}. These are block copolymers, where chemically different connected parts of the molecules tend to micro-phase separate into
regular spatial textures \cite{hamley1998physics}; surfactant molecules composed of hydrophobic (``water-fearing'') and hydrophilic (``water-loving'') parts that self-assemble for instance into
bilayer membranes and vesicles in an aqueous environment \cite{jones2002soft,witten2004structured}; and colloidal Janus particles \cite{walther2008janus} that can form finite clusters or regular lattices \cite{chen2012janus,jiang2010janus}.
With all these facts in mind, it is not surprising that the rheological behavior of such systems is relatively complex when compared to simple liquids \cite{larson1999structure}. Many soft matter systems are therefore also classified as ``complex fluids''. Sometimes the terms ``soft matter'' and ``complex fluids'' are even used interchangeably \cite{degennes1992soft}.
In the following, we will concentrate on soft matter systems out of their equilibrium ground state. It is impossible to cover all aspects and areas of this subfield, which was expanding so rapidly over the past decades. Instead, we will mainly focus on typical representatives in each part and embed these examples into the broader context. Roughly, we will increase the degree of activation that leads to non-equilibrium states through the subsequent sections.
First, we consider the switching or tuning of the states of soft matter systems by external fields. For example, as noted above, in display devices the orientation of the optical axis in liquid crystals is switched by external electric fields. Through a static external field, the system is forced out of its initial equilibrium ground state. However, it can reach a new static equilibrium state in the presence of the external field. \label{mod_intro} We will mainly concentrate on liquid crystalline as well as superparamagnetic or ferromagnetic gels and elastomers, the optical and mechanical properties of which can be tuned by external fields. They can serve as soft actuators, and even their possible role as ``artificial muscles'' was pointed out.
Next, we focus on systems permanently driven out of equilibrium by steady external shear flows. A brief account of the rheological behavior of liquid crystals and block copolymers is included. Coming back to the introductory cases mentioned above, an example situation from everyday life would be soap or body lotion while we are spreading it on our skin. Related to this topic, we address the progress made on the dynamics of vesicles when driven by an external shear flow.
The broadest section is devoted to the recently exploding field of ``active'' soft matter systems. We understand the word ``active'' in the sense that individual building blocks have a sort of propulsion mechanism, which leads to motion. Crawling cells or swimming bacteria can run this mechanism for instance by consuming fuel or food, whereas colloidal particles can self-propel via self-phoretic effects. In general, a key point is that the direction of motion is not prescribed from outside. These ingredients lead to new and fascinating properties of the individual constituents and when they collectively act together. There is a significant overlap between the fields of soft matter and biological physics in this area.
Finally,
a short summary is appended. We conclude by giving a brief impression of the current state of the field and its evolution.
\newpage
\section{Tuned soft matter}
By ``tuning'' we understand an adjustment of the material properties to a state that serves the current purpose or convenience. The crucial point with the soft matter systems introduced below comes with the non-invasive way of this adjustment. An external field is applied to the closed cell or piece of material to achieve the new setting.
In general, a static external field can force the system to acquire a new static equilibrium state in the presence of the field. If the external field is increased or decreased in small steps, after each adjustment the system finds itself in a weakly non-equilibrium situation. It has to find the new equilibrium state in the presence of the switched external field. Frequently, this process can be described by simple relaxation dynamics following the minimum of the modified free energy landscape. If the external field is adjusted slowly enough, a quasi-static switching can often be achieved, avoiding genuinely non-equilibrium situations.
If our focus is on the process of conversion of external field energy to reach a new state of the system, we may speak of ``actuation''.
\subsection{Low-molecular-weight liquid crystals}\label{lmwlc}
As a starting point, let us briefly recall the well-known Fr\'{e}edericksz transition in nematic liquid crystals \cite{degennes1993physics}. This transition can be induced from outside, for instance by applying an external electric field. As will be explained in more detail below, it involves a collective reorientation on the molecular level, changing the optical birefringence properties of a sample cell. In this way, the light transmittance of the cell can be reversibly tuned from outside.
In the most illustrative case, the liquid crystalline building blocks are small rod-like organic molecules. The molecular structure of one characteristic example, namely PAA, was depicted in Fig.~\ref{fig_lengths}. It shows the typical dimensions, which in this case are about two nanometers in length and about half a nanometer in thickness \cite{degennes1993physics}.
At high enough temperatures, the material is in an isotropic disordered liquid state as schematically illustrated in Fig.~\ref{fig_lc-phases}~(a).
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure02.pdf}}
\caption{Schematic illustration of typical phases of common low-molecular-weight liquid crystals consisting of elongated rod-like molecules. (a) The disordered isotropic liquid-like state shows neither long-ranged orientational nor long-ranged positional order. (b) In the nematic phase, there is no long-ranged positional order. However, apart from thermal fluctuations, the long axes of the molecules on average orient along a common direction called the director. (c) Smectic phases are characterized by a layer-like positional arrangement of the molecules, while in the depicted smectic-C state there is no long-ranged positional order within each layer. (d) Finally, crystalline solids show the highest degree of order with long-ranged positional order in the form of stacking of the molecules on regular lattice sites and long-ranged orientational order. Transitions between the depicted phases can be induced in corresponding liquid crystalline materials for example by temperature changes.}
\label{fig_lc-phases}
\end{figure}
The system does not feature any long-ranged positional nor orientational order of the constituents. It behaves in an isotropic way. In contrast to that, we find both positional and orientational order at low enough temperatures when the material forms a crystal. Then the molecules are stacked on regular lattice sites and orient their long axis along a common direction, see Fig.~\ref{fig_lc-phases}~(d). Between these two extremes of a disordered liquid and a crystalline solid, liquid crystals feature further intermediate phases, the so-called mesophases. These liquid crystalline phases show a higher degree of order (corresponding to a lower degree of symmetry) than the fluid but a lower degree of order (corresponding to a higher degree of symmetry) than the crystal. For example, in the nematic state, the one that we will refer to in the following, there is no long-ranged positional order. However, on average, the molecules orient along a common direction, see Fig.~\ref{fig_lc-phases}~(b). In this way, the material becomes anisotropic. Smectic states, which we will come back to later in this review, feature a regular stacking in layers, see Fig.~\ref{fig_lc-phases}~(c). More precisely, they show a quasi-long-ranged positional order in one spatial dimension.
When the system is in the nematic phase, the building blocks tend to collectively order their long axes along a common direction. This average orientation is called the nematic director $\mathbf{\hat{n}}$ \cite{degennes1993physics}. Genuinely nematic phases are non-polar, that is the directions $+\mathbf{\hat{n}}$ and $-\mathbf{\hat{n}}$ cannot be distinguished. The degree of orientational order is measured by a scalar order parameter
\begin{equation}\label{eq:s}
s = \frac{1}{2}\left\langle(3\cos^2\vartheta-1)\right\rangle,
\end{equation}
where the average $\langle...\rangle$ is taken over all molecules in the sample, and $\vartheta$ for each individual molecule measures the angle between its long axis and the average orientation $\mathbf{\hat{n}}$. We obtain $s=0$ in the disordered isotropic liquid state, and $s=1$ in a perfectly ordered nematic state. A combined symmetric traceless order parameter tensor $\mathbf{S}$ can be introduced, which contains both information, the degree of orientational order and the average orientation $\mathbf{\hat{n}}$,
\begin{equation}\label{eq:nematicOPtensor}
\mathbf{S} = s\left(\mathbf{\hat{n}}\mathbf{\hat{n}}-\frac{1}{d}\mathbf{I}\right).
\end{equation}
Here, $\mathbf{\hat{n}}\mathbf{\hat{n}}$ denotes a dyadic product, $d$ is the dimension of space (usually $d=3$), and $\mathbf{I}$ represents the unity matrix\footnote{Also another notation is common, in which the letters $S$ and $\mathbf{Q}$ are used instead of $s$ and $\mathbf{S}$, respectively.}.
On the one hand, since $\mathbf{\hat{n}}$ corresponds to the average orientation of the ordered molecules, it is an axis of anisotropy. For example, the dielectric constants $\epsilon_{\|}$ and $\epsilon_{\bot}$ measured parallel and perpendicular to $\mathbf{\hat{n}}$, respectively, generally differ from each other. Let us consider the case in which the dielectric anisotropy $\epsilon_a=\epsilon_{\|}-\epsilon_{\bot}$ is positive, $\epsilon_a>0$. Then a static external electric field $\mathbf{E}$ tends to orient the director $\mathbf{\hat{n}}$ along itself. This tendency is reflected by a corresponding contribution to the effective thermodynamic potential
\begin{equation}\label{F:elec}
F_{elec} = - \frac{1}{2}\epsilon_a(\mathbf{\hat{n}}\cdot\mathbf{E})^2.
\end{equation}
In this expression, possible constant prefactors due to the system of measure are absorbed into the dielectric anisotropy $\epsilon_a$.
On the other hand, spatial variations of the director, i.e.\ spatial distortions, increase the thermodynamic potential and are therefore suppressed. One way of quantifying this thermodynamic penalty is by contracting the tensor $\nabla\mathbf{\hat{n}}$,
\begin{equation}\label{F:dist}
F_{dist} = \frac{1}{2}K\,(\nabla\mathbf{\hat{n}})\!:\!(\nabla\mathbf{\hat{n}}),
\end{equation}
with $K>0$ another material-dependent parameter\footnote{We here use a simplified version of this thermodynamic expression. Due to the anisotropy of the system, the coefficient $K$ in general is replaced by a fourth-rank tensor \cite{degennes1993physics}.}. As an example, the nematic liquid crystal can be confined in a thin cell between two parallel plates as depicted in Fig.~\ref{fig_lc_frederiks}.
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure03.pdf}}
\caption{Reorientation of the nematic director induced by an external electric field in a thin cell of prescribed boundary conditions. The nematic director is indicated by the double-headed arrow. At the upper and lower confining plate it is rigidly anchored horizontally. On the one hand, if the sample is thin enough, this leads to a homogeneous horizontal alignment of the director to avoid spatial distortions in the director orientations (left). On the other hand, for positive dielectric anisotropy of the nematic molecules, a vertically applied external electric field $\mathbf{E}$ tends to orient the nematic molecules and the director in vertical direction (center). Both effects -- horizontal alignment due to the boundary conditions and vertical alignment through the external electric field -- compete with each other. Balancing the two resulting torques leads to a finite angle of director reorientation with respect to the horizontal direction (right) above a non-vanishing critical threshold amplitude of the electric field.}
\label{fig_lc_frederiks}
\end{figure}
The surfaces of the cell can be prepared such that they anchor the orientation of the director, for instance in a parallel way, as indicated on the left-hand side of Fig.~\ref{fig_lc_frederiks}. This, via Eq.~(\ref{F:dist}), sets a uniform director orientation in the whole cell.
If now an external electric field is switched on with an orientation perpendicular to the cell surfaces, see the center of Fig.~\ref{fig_lc_frederiks}, it tends to reorient the director and align it parallel to itself as given by Eq.~(\ref{F:elec}). However, via Eq.~(\ref{F:dist}), such a reorientation induces thermodynamically penalized distortions because of the surface anchoring on the plates. The competition between the two effects leads to a critical electric field amplitude that has to be exceeded for director reorientations to become possible. Above this threshold field amplitude, director reorientations are induced. From Eqs.~(\ref{F:elec}) and (\ref{F:dist}) this critical field amplitude follows directly as
\begin{equation}
E_c=\frac{\pi}{d}\sqrt{\frac{K}{\epsilon_a}},
\end{equation}
where $d$ is the thickness of the cell and the impact of possible electric currents has been neglected. In the new static reoriented state, the two effects of the reorienting electric field and the counteracting surface anchoring balance each other. This is indicated on the right-hand side of Fig.~\ref{fig_lc_frederiks}. When the electric field is switched off, the director rotates back into its initial state due to the surface anchoring.
In this way, the anisotropy properties of the presented sample cell can be reversibly tuned from outside. Naturally, this anisotropy in a real sample is connected to optical birefringence. Thus the optical transmission properties of the cell can be reversibly switched and adjusted by the external electric field.
This is the underlying principle extensively used to construct optical display devices that are based on liquid crystals \cite{kawamoto2002history}. However, the technical realizations today involve much more complicated geometries than the simple Fr\'{e}edericksz cell outlined above.
\subsection{Liquid crystalline elastomers}
Commonly used liquid crystals have relatively low molecular weight and are classified as complex fluids. As suggested by the name, they typically flow away when not maintained by a containing cell. Another promising material would be one that is stable and self-standing without supporting container walls, but still features the optical transmission properties of common liquid crystals. Such materials were actually realized by stepping towards higher molecular weights in the form of liquid crystalline polymers and elastomers.
In general, polymers are synthesized by covalently binding together from hundreds up to millions identical molecular repeat units, the so-called monomers. In this way, new huge macromolecules are formed \cite{strobl1997physics}. Different chemical routes for this procedure are available. The number of monomers on the resulting polymeric molecule defines the degree of polymerization. In the simplest case, long linear chain-like molecules are obtained, but several further molecular architectures can be realized.
The simplest theoretical model describing linear polymer chains is the freely-jointed chain model \cite{strobl1997physics}. It already gives us enough qualitative insight to understand the discussion below. In this model, as illustrated in Fig.~\ref{fig_freely-jointed-chain}, the polymeric molecule is reduced to a chain of $N$ straight segments that are linked to each other at their ends.
\begin{figure}
\centerline{\includegraphics[width=4.5cm]{figure04.pdf}}
\caption{
Illustration of the freely-jointed chain model for a linear polymer chain. The chain is represented by $N$ freely-jointed segments of end-to-end vectors $\mathbf{r}_n=\mathbf{R}_n-\mathbf{R}_{n-1}$ ($n=1,...,N$), where $\mathbf{R}_n$ ($n=1,...,N-1$) give the positions of the joints. All segments are of identical length, $\|\mathbf{r}_n\|=b$ for all $n$, and their orientations are completely independent of each other. $\mathbf{R}=\mathbf{R}_N-\mathbf{R}_0$ is the end-to-end vector of the whole polymer chain.
}
\label{fig_freely-jointed-chain}
\end{figure}
Each segment has an identical length $b$ much larger than the size of a single monomer. We denote the positions of the joints between the segments as $\mathbf{R}_n$ for $n=1,...,N-1$, while $\mathbf{R}_0$ and $\mathbf{R}_N$ give the positions of the two ends of the whole resulting chain, see Fig.~\ref{fig_freely-jointed-chain}. The chain is ``freely'' jointed in the sense that each segment orientation is independent of the orientation of all other segments. Thus, for each segment, the probability distribution $\psi(\mathbf{r}_n)$ of its end-to-end vector $\mathbf{r}_n=\mathbf{R}_n-\mathbf{R}_{n-1}$ is given by
\begin{equation}
\psi(\mathbf{r}_n) = \frac{1}{4\pi b^2}\,\delta\left(\|\mathbf{r}_n\|-b\right) \qquad (n=1,...,N),
\end{equation}
where $\delta(\cdot)$ denotes the Dirac $\delta$-function. It serves here to prescribe the length $b$ of the segment. In this minimum model, the probability distribution to find a certain chain configuration is simply given by the product of these individual single-segment probability distributions $\psi(\mathbf{r}_n)$.
On this basis, it is straightforward to determine the probability distribution for the end-to-end vector $\mathbf{R}=\mathbf{R}_N-\mathbf{R}_0$ of the whole chain. For $N\gg1$, we obtain
\begin{equation}\label{eq:end-to-end}
\psi(\mathbf{R}) = \left(\frac{3}{2\pi Nb^2}\right)^{3/2}\exp\left\{ -\frac{3\mathbf{R}^2}{2Nb^2} \right\}.
\end{equation}
Obviously, this is of Gaussian form.
From a statistical mechanics point of view, we expect the probability distribution to be of the form $\psi(\mathbf{R})\sim\exp[-F(\mathbf{R})/k_BT]$. Here $F(\mathbf{R})$ denotes the energy of the chain configuration with end-to-end vector $\mathbf{R}$, $k_B$ is Boltzmann's constant, and $T$ the temperature. Comparison with Eq.~(\ref{eq:end-to-end}) shows that
\begin{equation}\label{eq:spring}
F(\mathbf{R}) = \frac{1}{2} \frac{3k_BT}{Nb^2} \mathbf{R}^2.
\end{equation}
Thus the polymer chain behaves like an elastic spring between its two end points $\mathbf{R}_0$ and $\mathbf{R}_N$. In the unstrained state the length of this spring vanishes. We find $3k_BT/Nb^2$ for the magnitude of its elastic constant.
Based on this simple model, we see that polymeric materials show an elastic response. The important point is that these elastic forces mainly do not origin from an internal energy. They are mostly entropic in nature. Everything that enters Eq.~(\ref{eq:end-to-end}) is the entropy resulting from the single-chain configuration. The elastic constant in Eq.~(\ref{eq:spring}) is set by the temperature $T$.
As a next step, one would like to obtain an elastic material of macroscopic dimension. For this purpose, many polymer chains are chemically crosslinked, i.e.\ covalently bound, to form one chemical unit. In this way, an elastomer or rubber is obtained. The main source of elasticity is again of entropic origin and referred to as ``rubber elasticity'' \cite{treloar1975physics}.
To make the connection to the field of liquid crystals, first liquid crystalline polymers were synthesized. Two principal architectures are available: on the one hand, liquid crystalline molecules can be part of the polymer backbone chains in liquid crystalline main-chain polymers \cite{percec1991liquid}, see Fig.~\ref{fig_mcsc}~(a); on the other hand, the liquid crystalline molecules can be attached as side-groups to a polymer backbone in liquid crystalline side-chain polymers \cite{finkelmann1978model}, see Fig.~\ref{fig_mcsc}~(b).
\begin{figure}
\centerline{\includegraphics[width=9.cm]{figure05.pdf}}
\caption{Schematic illustration of two architectures of liquid crystalline polymers. Again the liquid crystalline rod-like components are indicated by ellipsoids as in Fig.~\ref{fig_lc-phases}. Non-liquid-crystalline parts of the polymer ``backbone'' chains are depicted by the bright strings. (a) On the one hand, in main-chain liquid crystalline polymers, the liquid crystalline units are part of the polymer backbones. (b) On the other hand, the liquid crystalline units are attached as side groups to the polymer chains in side-chain liquid crystalline polymers. Usually this is achieved via shorter hydrocarbon chains, so-called spacer groups, which here are indicated in darker color.}
\label{fig_mcsc}
\end{figure}
When in a later step such liquid crystalline polymers were chemically crosslinked, rubbery liquid crystalline elastomers were obtained \cite{finkelmann1981liquid,bergmann1997liquid}. In this way, the properties of liquid crystals and those of elastic rubbers could be combined in one material. The coupling provides new interesting features as described below.
There are different protocols available to obtain different kinds of liquid crystalline elastomers concerning the nematic orientational alignment in the samples. Without special aligning mechanisms
it is natural for nematic bulk materials in general not to feature a globally aligned nematic director. Already for low-molecular-weight liquid crystals, bulk samples in the nematic phase usually appear turbid. They strongly scatter light \cite{degennes1993physics}. The reason are thermal fluctuations in the director orientations. As a consequence, the nematic director is not macroscopically aligned, but continuously varies over space.
Additional effort, such as the prescribed alignment on the sample surfaces depicted in Fig.~\ref{fig_lc_frederiks}, is necessary to achieve a global alignment of the director orientation.
Naturally, thermal orientational fluctuations also occur in polymeric substances.
Therefore, if during the crosslinking process of synthesizing a liquid crystalline elastomer no special action is taken, polydomain samples are obtained \cite{finkelmann1994liquid}. The director in these polydomain materials is not globally uniform in the nematic state but continuously varies over space. However, protocols to manufacture monodomain samples featuring a macroscopically aligned director are available. For this purpose, during the final crosslinking step of the manufacturing process, the sample must be maintained in the liquid crystalline phase and the director must be homogeneously aligned in the polymer solution. In practice, this was achieved by external magnetic fields \cite{legge1991memory,mitchell1993strain,rogez2011influence}, external electric fields \cite{rogez2011influence}, aligning boundary conditions \cite{urayama2005electrooptical,komp2005versatile,urayama2007stretching}, or stretching a pre-crosslinked sample \cite{kupfer1991nematic,kupfer1994liquid}. A corresponding memory of the director orientation during crosslinking is then imprinted into the material during the crosslinking procedure. When the liquid crystalline molecules were reoriented by some external action, they were observed to rotate back into the memorized orientations after the external influence is switched off again. This was experimentally verified even for polydomain samples crosslinked in the liquid crystalline state \cite{urayama2006slow,urayama2009polydomain}. Apparently the orientations are stored in the architecture of the polymer network.
The most interesting properties of liquid crystalline elastomers arise from their orient\-ational-deformational coupling. On the one hand, a mechanical deformation like stretching or compression can tune the state of birefringence and the optical properties. As one specific example, liquid crystalline elastomers were used to experimentally demonstrate the possibility to construct mirrorless lasers of a wave-length that is tunable by mechanical deformations \cite{finkelmann2001tunable,schmidtke2005probing}. More in general, it was shown for various different samples that mechanical deformations can reorient the director \cite{kupfer1991nematic,mitchell1993strain,kupfer1994liquid,kundler1995strain,roberts1997single, kundler1998director,urayama2007stretching}. This is indicated on the left-hand side of Fig.~\ref{fig_lc_elastomer}:
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure06.pdf}}
\caption{Coupling between director reorientations and elastic deformations in nematic liquid crystal elastomers. The nematic director is indicated by the double-headed arrow, whereas the rectangular box illustrates the deformations of the elastomer. An initial ground state is depicted in the center. When an external stretch (marked by the vertical broad arrows on the left) is applied in the direction perpendicular to the initial director orientation, the coupling to the director can induce its reorientation into the stretching direction. For practical purposes, the inverse effect would be beneficial, i.e.\ mechanical deformations induced by applying an electric field that reorients the director, as illustrated on the right. Unfortunately, the necessary field amplitudes turn out to be too high for practical purposes, except for materials swollen with conventional small-molecule liquid crystals.}
\label{fig_lc_elastomer}
\end{figure}
when a nematic liquid crystalline elastomer is stretched perpendicularly to its initial director orientation, a reorientation of the director into the stretching direction can result.
The reorientation process involved a pronounced nonlinearity in the corresponding stress-strain curves \cite{kupfer1991nematic,kupfer1994liquid,urayama2007stretching,krause2009nematic}. On the other hand, it was demonstrated that the inverse effect is also within reach experimentally: reorientations of the director can induce mechanical deformations. From an application point of view, this would allow the construction of soft actuators.
One could use an external electric field to reorient the director, which is transformed into mechanical actuation. Unfortunately, so far the necessary field amplitudes were found to be too high for practical purposes.
However, for samples swollen with a common low-molecular-weight (small-molecule) liquid crystal, a sizeable amount of so-called electro-mechanical coupling was observed \cite{yusuf2004swelling,urayama2005electrically,yusuf2005low,urayama2005electrooptical, urayama2006swelling,urayama2006deformation,cho2006electrooptical, urayama2007selected,cho2007trifunctionally, fukunaga2008dynamics, hashimoto2008multifunctional}. In this case the director can be reoriented by an external electric field, which can result in deformations of up to the lower double-digit per-cent regime.
Prestretching a material to the point where director reorientation sets in may allow to observe the electric-field-induced effect also in conventional non-swollen samples, as outlined theoretically \cite{menzel2009response}. Close to this critical point, the reorienting external stretching deformation on the one hand and the imprinted memory that anchors the director on the other hand approximately balance each other. Then a relatively weak additional external mechanical or electric field can result in a comparatively large response in the director orientation.
Apart from that, other routes of regulation or actuation were demonstrated experimentally, for example photo- \cite{finkelmann2001new,sanchez2011opto} or temperature-induced \cite{yusuf2004hystereses,cho2006thermo,sawa2010thermally,fleischmann2012one} deformations. Possible applications as soft actuators or even ``artificial muscles'' were discussed \cite{hebert1997dynamics,yu2006soft,ohm2010liquid,yang2011micron,jiang2013actuators}. Large length changes of several $100$~\% were observed for loaded and unloaded samples under reversible heating and cooling through the isotropic-nematic transition \cite{wermter2001liquid,krause2009nematic}. For many applications, however, control parameters different from temperature are more desirable.
On the theoretical side, we mention two principal routes of approach that reproduce the measured nonlinear stress-strain behavior as well as the detected deformations due to director reorientations. The first one is based on the concept of rubber elasticity \cite{treloar1975physics}. It assumes Gaussian statistics for the step lengths along the polymer strands \cite{warner2003liquid}. As a starting point, it further assumes that the main role of the liquid crystalline building blocks is to introduce an anisotropy direction into these Gaussian statistics. Thus, instead of Eq.~(\ref{eq:end-to-end}), the end-to-end vectors of the polymer strands between network points of the polymer mesh are assumed to follow the distorted distribution
\begin{equation}
\psi(\mathbf{R}) \propto \exp\left\{ -\frac{3}{2\tilde{L}^2}\,\mathbf{R}\cdot(\mathbf{I}-r_a\mathbf{\hat{n}}\mathbf{\hat{n}})\cdot\mathbf{R} \right\},
\end{equation}
where $\mathbf{\hat{n}}$ again indicates the nematic director, $r_a$ is an anisotropy parameter, $\mathbf{I}$ indicates the unity matrix, and $\tilde{L}$ is an effective length connected to the actual contour lengths of the polymer strands.
Open issues concerning this picture are how valid the assumption of the distorted Gaussian statistics generally is \cite{rogez2011influence}, and particularly how the orientational memory influences the statistics.
A different microscopic approach was proposed in the form of randomly crosslinked anisotropically interacting rigid dimers \cite{xing2008nematic}, which reproduces the expressions for the theory of rubber elasticity on a more coarse-grained level.
The second route is macroscopic and based on symmetry arguments. It couples the continuum descriptions of liquid crystals and elastic materials \cite{degennes1980weak,brand1994electrohydrodynamics,lubensky2002symmetries,muller2005undulation,menzel2006rotatoelectricity, menzel2007cholesteric,menzel2008instabilities,ye2007semisoft,menzel2007nonlinear,menzel2009nonlinear,ye2009phase}.
We noted above that a memory of the director orientation during synthesis is imprinted into the network structure of the polymer mesh.
One central ingredient to many of the macroscopic studies is therefore the observation that rotations of the director out of its memorized equilibrium orientation cost energy. Such ``relative rotations'' between the director and the polymer network environment are included as additional variables \cite{degennes1980weak,brand1994electrohydrodynamics,muller2005undulation,menzel2006rotatoelectricity, menzel2007cholesteric,menzel2008instabilities}. First proposed for small deviations from the initial state, this concept was later generalized to the nonlinear regime \cite{menzel2007nonlinear,menzel2009nonlinear}.
It is schematically illustrated in Fig.~\ref{fig_relrot}.
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure07.pdf}}
\caption{Illustration of the concept of relative rotations in liquid crystalline elastomers. In the ground state (center) the director takes the orientation memorized by the architecture of the polymer network from the time of crosslinking. Relative rotations as a macroscopic variable describe how the director is rotated with respect to that memorized orientation. This can occur by a reorientation of the director relatively to the polymeric network (left), or a rotation of the network together with the imprinted memorized direction with respect to the actual director orientation (right). The variables of relative rotations test whether the results from both processes differ from each other and allow a formal inclusion in the corresponding energetic considerations. In all pictures, the strong-colored double-headed arrow indicates the actual director orientation, and the bright box symbolizes the polymeric network. On the left- and on the right-hand side, the light double-headed arrow represents the imprinted ground-state director orientation memorized by the network structure at the moment of crosslinking.}
\label{fig_relrot}
\end{figure}
\subsection{Ferrogels and magnetic elastomers}
Another example of externally tunable polymeric soft matter systems is given by magnetic gels and elastomers \cite{filipcsei2007magnetic}. Again, these materials combine the features of two different classes of complex fluids: those of ferro- and magnetorheological fluids \cite{rosensweig1985ferrohydrodynamics,odenbach2002ferrofluids,odenbach2003ferrofluids,odenbach2003magnetoviscous, huke2004magnetic,odenbach2004recent,fischer2005brownian,ilg2005structure,klapp2005dipolar,holm2005structure, embs2006measuring,ilg2006structure,gollwitzer2007surface,vicente2011magnetorheological} with those of crosslinked polymers \cite{strobl1997physics}.
The systems can be classified as composite magnetic hybrid materials \cite{jolivet2002synthesis,garcia2003mesoporous,yuan2011one,sarkar2012polymer,kao2013toward}. They consist of superparamagnetic or ferromagnetic particles that are embedded in a crosslinked polymer matrix. Depending on the degree of swelling of such a magnetic elastomer with a solvent it may rather be called a magnetic gel \cite{jolivet2002synthesis,messing2011cobalt,frickel2011magneto}. In the dry state, the materials can become hard and glass-like \cite{jolivet2002synthesis}.
Different sizes of the embedded magnetic particles, typically in the range of nano- to micrometers \cite{filipcsei2007magnetic}, can lead to qualitatively different material behavior. If the particles are smaller than ten to fifteen nanometers, the direction of their magnetic moment is not fixed with respect to the particle axes. Instead, thermal fluctuations can lead to a reorientation of the magnetic moment \cite{neel1949theorie}, which is called N\'{e}el mechanism. For larger particles, reorientations of the magnetic moments mainly occur via the Brownian or Debye mechanism \cite{coffey1993ferromagnetic}, i.e.\ rigid rotations together with the whole particle. Still larger particles, from around a hundred nanometers, can feature multidomains of internal magnetization \cite{frenkel1930spontaneous,brown1968fundamental,aharoni1988elongated,seynaeve2001transition,hergt2006magnetic}.
In the absence of an embedding polymer matrix, suspensions of such magnetic particles in a carrier liquid were stabilized in the form of ferro- and magnetorheological fluids \cite{rosensweig1985ferrohydrodynamics,odenbach2002ferrofluids,odenbach2003ferrofluids,odenbach2003magnetoviscous, huke2004magnetic,odenbach2004recent,fischer2005brownian,ilg2005structure,klapp2005dipolar,holm2005structure, embs2006measuring,ilg2006structure,gollwitzer2007surface,vicente2011magnetorheological}.
Such fluids show properties of high practical relevance \cite{raj1990commercial,raj1995advances}. In particular, their dynamic flow behavior determined by their macroscopic viscosity can be tuned reversibly through an external magnetic field. This feature, commonly referred to as ``magnetoviscous effect'', was observed experimentally and explained theoretically \cite{rosensweig1969viscosity,mctague1969magnetoviscosity, zubarev2002rheological,thurm2002magnetic,thurm2002magnetoviscous,thurm2003particle, odenbach2003ferrofluids,pop2004microstructure,odenbach2004recent,ilg2005anisotropy,pop2006investigation}. It is attributed on the one hand to a hindrance of particle rotations that would reorient the magnetization axes of the particles away from the aligning external magnetic field direction. On the other hand, it results from the formation of micro-aggregates in the presence of external magnetic fields. Typically, the second contribution, when present, exceeds the first one. The important point is that the adjustment of the viscosity can be achieved in a non-invasive way reversibly from outside.
We turn back to the presence of an embedding polymer matrix in magnetic gels and elastomers.
In analogy to the viscous behavior of magnetic fluids, for magnetic gels and elastomers the resistance to elastic deformations, quantified by the elastic moduli \cite{landau1986elasticity}, was experimentally probed. It turned out that the elastic moduli can be tuned reversibly from outside by external magnetic fields \cite{deng2006development,filipcsei2007magnetic,stepanov2007effect, chen2007investigation,bose2009magnetorheological,evans2012highly,borin2013tuning}.
This is one of the most outstanding properties of these materials. From an application point of view, it is for example interesting for the construction of novel externally tunable damping devices \cite{sun2008study} or vibration absorbers \cite{deng2006development}.
The analysis of theoretical minimum models demonstrated that the spatial distribution of the magnetic particles plays a central role and qualitatively determines the nature of this effect \cite{ivaneyko2011magneto,ivaneyko2012effects,ivaneyko2014mechanical}. It was found that the tensile elastic modulus decreases or increases with increasing amplitude of an external magnetic field applied along the tension axis, depending on the particle arrangement. For instance, a regular simple-cubic or bcc lattice arrangement of the particle positions implied a decreasing tensile elastic modulus \cite{ivaneyko2011magneto,ivaneyko2012effects}. In contrast to that, the modulus increased for an fcc structure \cite{ivaneyko2012effects}.
To a certain degree, the particle distribution can in fact be influenced during the synthesis of the materials. When a strong external magnetic field is applied before and during the crosslinking procedure, the magnetic particles were observed to form chain-like aggregates \cite{collin2003frozen,varga2003smart,filipcsei2007magnetic,gunther2012xray, borbath2012xmuct,gundermann2013comparison}. After the crosslinking has been completed, the field can be switched off with the anisotropic particle distribution persisting. The anisotropy has been locked into the materials.
It has been demonstrated in experiments that this imprinted anisotropy can support the tunability of the elastic modulus by an external magnetic field \cite{filipcsei2007magnetic}. In that case, the largest effect of tunability was found when all three directions were aligned: the imprinted anisotropy direction, the additional external magnetic field to tune the mechanical properties, as well as the direction of mechanical deformation to measure the elastic modulus \cite{filipcsei2007magnetic}. This trend could be reproduced in a numerical study that assumed affine deformations of the material \cite{camp2011modeling,camp2011effects}. Although the degree of anisotropy in this case was significantly lower than that of chain-like aggregates, the effect was clearly observed.
We remark that the assumption of affine deformations, which maps the macroscopic length changes linearly to all distances within the system, constitutes a major simplification. Yet it is often necessary to allow theoretical progress at all. Nevertheless, the more irregular the distributions of the magnetic particles within the samples become, the more this can lead to increasingly erroneous and even qualitatively incorrect results. Care has to be taken when this approximation is applied to realistic systems as has recently been pointed out on the basis of a simple dipole-spring model
\cite{pessot2014structural}.
Allowing for non-affine deformations, an explicit numerical investigation on the impact of chain-like particle arrangements was performed using finite element simulations \cite{han2013field}. It showed that deviations from a straight chain architecture in the form of a zig-zag structure can qualitatively influence the behavior under external magnetic fields: the zig-zag angle determines whether the compressive elastic modulus decreases or increases with increasing amplitude of the external magnetic field. Even non-monotonic behavior is possible.
The behavior within one linear chain was studied in a minimum dipole-spring model of hard spheres connected by elastic springs \cite{annunziata2013hardening}. When the magnetic moment of the spheres is increased by an external magnetic field such that the hard spheres attract each other, they will come closer until they finally touch. Neglecting thermal fluctuations of the spheres, this process has the signature of a phase transition. It is of second order when the initial separation is small. If the initial separation is large, it is of first order, with the spheres clashing together in the final part of approach. A critical point separates both scenarios. Since the compressive elastic modulus along the chain diverges when the spheres touch each other, the process was called ``hardening transition''. At the critical point, just before the spheres touch each other, the compressive elastic modulus along the chain vanishes.
Also the behavior of free-standing filamental chains was studied numerically \cite{sanchez2013filaments,cerda2013phase}. In this case, various different states like compact, helicoidal, partially collapsed, simply closed, and extended open configurations were detected.
A second fascinating property of magnetic elastomers and gels arises from their reversible deformations that can be induced by external magnetic fields. The spatial arrangement of the magnetic particles can qualitatively impact whether elongation or contraction occurs along the field direction, as pointed out by theory and simulations \cite{stolbov2011modelling,zubarev2013effect}. An example is illustrated in Fig.~\ref{fig_ferrogel}.
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure08.pdf}}
\caption{In magnetic gels and elastomers the spatial distribution of the magnetic particles can qualitatively influence the macroscopic behavior. Here, this is illustrated for the magnetostrictive behavior of a two-dimensional sheet in a homogeneous external magnetic field (see Ref.~\cite{stolbov2011modelling}). Left-hand side: for a random ``gas-like'' distribution of magnetic particles switching on the external magnetic field was observed to result in a contraction along the field direction \cite{stolbov2011modelling}. Right-hand side: in contrast to that, for randomly oriented doublets of magnetic particles, the same set-up was found to lead to an elongation along the field direction \cite{stolbov2011modelling}.
}
\label{fig_ferrogel}
\end{figure}
Apart from the magnetostrictive behavior \cite{guan2008magnetostrictive,stolbov2011modelling,gong2012full, zubarev2012theory,zubarev2013magnetodeformation,allahyarov2014magnetomechanical}, i.e.\ deformations induced by homogeneous external magnetic fields, particularly the ``actuation'' by inhomogeneous external magnetic fields \cite{zrinyi1996deformation,zrinyi1997direct} was investigated in experiments and by modeling. To minimize the free energy, the materials are drawn into regions of higher magnetic field. This can lead to elastic deformations. After switching off the external field, the elastic energy is released and the probe can reversibly switch back to its initial state. It was demonstrated for a real sample that frequencies around $40\mbox{~Hz}$ can easily be followed \cite{filipcsei2007magnetic}. Consequently these materials again are ideal candidates for the construction of soft actuators \cite{zimmermann2006modelling,filipcsei2007magnetic} or ``artificial muscles'' \cite{zrinyi2000intelligent}.
These applications involve dynamic properties, which so far have been addressed in only a few studies on the theoretical side \cite{jarkova2003hydrodynamics,bohlius2004macroscopic,tarama2014tunable}, requiring further investigation in the future.
Apart from that, the use of magnetic gels as sensors indicating magnetic fields and field gradients by deformation was experimentally outlined \cite{szabo1998shape,ramanujan2006mechanical, liu2006magnetic}. Other studies discuss their application to control by external magnetic fields the amount and rate of drug release \cite{liu2006magnetic,brazel2009magnetothermally}. A further subject is the combat against cancer cells. For this purpose, magnetic particles are embedded into the corresponding cell tissue. A quickly alternating external magnetic field can be applied that continuously remagnetizes the particles. Through hysteretic losses during the remagnetization cycles, heat is generated. Such hyperthermic treatment \cite{jordan1999magnetic,babincova2001superparamagnetic, lao2004magnetic,hergt2006magnetic} can destroy the cancer cells.
For all these purposes, the degree of the so-called magneto-mechanical coupling should be maximized. A recent attempt in this direction was to not only embed the magnetic particles in the polymer matrix, but to directly crosslink them into the polymer network \cite{frickel2009functional,frickel2011magneto,messing2011cobalt}, see Fig.~\ref{fig_kinds-ferrogels}.
\begin{figure}
\centerline{\includegraphics[width=9.cm]{figure09.pdf}}
\caption{Schematic illustration of two qualitatively different kinds of ferrogels implying different degrees of magneto-mechanical coupling. (a) In the classical materials the magnetic particles are simply embedded in the surrounding mesh of polymer chains that form the polymer matrix. For spherical particles, there may not be any permanent restoring torque induced by particle rotations. (b) In a new class of ferrogels, the polymer chains are covalently bound to the surface of the magnetic particles \cite{frickel2011magneto,messing2011cobalt}. Thus the magnetic particles serve as crosslinkers. Rotations of the spherical particles inevitably induce deformations of the surrounding polymer mesh, leading to permanent restoring torques. In both cases, the density of polymer chains is significantly higher than indicated in the simplified schematics.}
\label{fig_kinds-ferrogels}
\end{figure}
During sample preparation, this was achieved via surface-functionalization of the magnetic particles by attaching reactive groups on their surfaces. As a consequence, the magnetic particles themselves can serve as crosslinkers. Rotating the particles, the polymer chains attached to their surfaces get ``rolled up'' around the particles as investigated by a simulation study \cite{weeber2012deformation}. This ``rolling-up'' leads to restoring torques.
An orientational memory of the particle orientations arises, implying an energetic penalty when the particles rotate relatively to their environment \cite{annunziata2013hardening}. It can have a significant impact on the appearance of the system. For example, depending on the orientational memory, qualitatively different energetic ground states were obtained for a linear chain of ferromagnetic particles in a dipole-spring model \cite{annunziata2013hardening}: ferromagnetic, with the magnetic moments aligned along the chain axis; antiferromagnetic-like, with neighboring dipoles rotated by 180 degrees around the chain axis; or a spiral-like magnetization, with neighboring dipolar moments rotated by an intermediate angle around the chain axis.
Even stronger magneto-mechanical coupling might be achieved through the use of elongated instead of spherical magnetic particles \cite{roeder2012shear}.
\subsection{Links between liquid crystalline and magnetic elastomers}
At the end of this section, we include a brief qualitative comparison between the two classes of materials considered above, i.e.\ liquid crystalline and magnetic elastomers and gels. We only concentrate on a few selected points that may be important when a mutual support of the two fields is considered. With the macroscopic theories for the two classes of material being partially mappable onto each other, also their observable behavior in analogous experiments may be so.
Both materials combine the elastic behavior of polymeric substances with the anisotropy arising from an orientational ordering of their constituents. In the case of liquid crystalline elastomers, this anisotropy results from the average orientation of the liquid crystalline molecules along the director. For magnetic elastomers and gels, it is identified with the orientation of the magnetic moments in the equilibrium configuration.
Depending on the procedure of synthesis, a more or less pronounced memory of the initial orientations at the crosslinking time is locked into the samples. Then relative rotations between this locked-in memory direction and the actual orientation
cost energy. We outlined above the use of such relative rotations as a variable in the macroscopic characterization of liquid crystalline elastomers \cite{degennes1980weak,brand1994electrohydrodynamics, menzel2007nonlinear,menzel2009nonlinear,menzel2009response}, see Fig.~\ref{fig_relrot}, and in a mesoscopic model of magnetic gels and elastomers \cite{annunziata2013hardening}. Furthermore, relative rotations have already been considered in a macroscopic characterization of anisotropic magnetic gels \cite{bohlius2004macroscopic}.
A next step will be to connect the mesoscopic and the macroscopic descriptions of magnetic gels by an appropriate coarse-graining procedure.
Similarities between the two classes of materials, i.e.\ between liquid crystalline elastomers and magnetic gels, were pointed out on the macroscopic scale \cite{brand2011physical}.
Generally, it is an interesting question, how the memory actually gets locked into the architecture on the molecular level. In other words, what structural changes occur for a polymer network in comparison to a conventional material without memory. The differences between the two classes of materials are of course large, already in view of the corresponding length scales. On the one hand, the liquid crystalline molecules are relatively small and can even be part of the polymer backbones in main-chain liquid crystalline elastomers. Therefore one would expect that an orientational memory should be reflected on the molecular level by the structure of the crosslinked polymer network. On the other hand, magnetic elastomers are real composite materials of mesoscopic colloidal magnetic particles embedded in a polymer matrix. Here, the spatial arrangement of the mesoscopic magnetic particles within the elastic matrix can be the source of the orientational memory. So the mechanism of memorizing may be quite different in the two cases.
Nevertheless, based on the similarities in the macroscopic theories, it would be worthwhile to test the materials in those experiments in which the respective other material class features its outstanding properties. The experience acquired for one of the two material classes may help to guide the way also for the other material class. On the one hand, this means to clarify the question, whether an orientational coupling between strain deformations and the anisotropy direction also exists within anisotropic magnetic gels. In a next step, this would lead to the question, whether this property is likewise connected to a marked nonlinear stress-strain behavior as observed for liquid crystalline elastomers. On the other hand, in analogy to the elastic moduli that are reversibly tunable by external magnetic fields in the case of magnetic gels, a similar effect may be observable for liquid crystalline elastomers and gels. Here, rather an externally applied electric field would serve to tune the elastic properties. It would be satisfying to observe a mutual benefit between the two fields of liquid crystalline elastomers and magnetic gels.
Moreover, instead of only tuning their properties, the materials could also be utilized as model systems for driven and active soft matter in genuinely non-equilibrium situations. We mentioned that both liquid crystalline elastomers and magnetic gels can be used as soft actuators by switching an appropriate external field \cite{yusuf2005low,urayama2006deformation,urayama2007selected,cho2007trifunctionally, fukunaga2008dynamics,hashimoto2008multifunctional,urayama2005electrooptical,urayama2005electrically, finkelmann2001new,sanchez2011opto,cho2006thermo,sawa2010thermally,fleischmann2012one, hebert1997dynamics,yu2006soft,ohm2010liquid,yang2011micron,jiang2013actuators,filipcsei2007magnetic, stolbov2011modelling,guan2008magnetostrictive,gong2012full,zubarev2012theory,zubarev2013magnetodeformation, allahyarov2014magnetomechanical,zrinyi1996deformation,zrinyi1997direct,zimmermann2006modelling, zrinyi2000intelligent}. In a next step, these materials could be steadily driven by external fields so that they cannot reach an equilibrium state any more. Examples of driven soft matter will be addressed in the subsequent part of this review. In another step, they can be continuously ``activated'' and then be categorized as active soft materials. For instance, the potential of liquid crystalline elastomers for the construction of light-activated swimmers was experimentally outlined \cite{camacho2004fast}. Similarly, an efficient production of light-activatable microscopic artificial cilia was demonstrated via inkjet printing \cite{vanoosten2009printed}; induced deformations of these slender artificial fibers \cite{vanoosten2009printed} mimic the deformation cycle of natural cilia that are found on bacterial surfaces and serve for self-propulsion of the microorganisms \cite{brennen1977fluid}. Likewise, magnetically activated elastic filaments were used to construct propelling artificial microswimmers \cite{dreyfus2005microscopic}. We will come back to the migration of self-propelled particles and microswimmers in the later part of this review.
Apart from that, on the theoretical side, the macroscopic characterizations of gels featuring orientational degrees of freedom can be extended to include active stresses that result for example from chemical reactions. Such approaches were used to identify and describe possible mechanisms for the motility of crawling biological cells \cite{kruse2004asters,kruse2005generic,juelicher2007active, carlsson2011mechanisms,recho2013contraction,recho2014optimality}.
\newpage
\section{Driven soft matter}
Naturally, in each system thermal fluctuations lead to a sort of ``internal drive''. For instance, colloidal particles are subject to Brownian motion, which adds to prevent them from sedimentation in a gravitational field and keeps them suspended \cite{dhont1996introduction}. Another example are experiments on colloidal particles thermally driven along the top of aligned membrane tubes that adhere to a substrate \cite{wang2009anomalous,wang2012brownian}. An interesting crossover between exponential and Gaussian forms of the displacement statistics was observed in the latter context. Generally, when asymmetric ratchet-like conditions direct the internal thermal drive, net motion can arise \cite{peskin1993cellular,faucheux1995optical}.
In contrast to that, we here focus on soft matter systems driven from outside. In our case this is achieved by externally imposed shear flows. The simplest situation is the one of an imposed planar linear shear profile as the one schematically indicated in Fig.~\ref{fig_linear-shear-flow}. This geometry is characterized by three different directions: the flow direction, here horizontally oriented; the direction of the shear gradient, here vertically oriented; and the vorticity direction perpendicular to the shear plane, i.e.\ here pointing into the plane of the figure.
\begin{figure}
\centerline{\includegraphics[width=4.5cm]{figure10.pdf}}
\caption{The flow profile corresponding to a planar linear shear flow is depicted on the left. It reveals the three different directions indicated on the right: the flow direction oriented along the flow velocity $\mathbf{v}$ (horizontal); the direction of the shear gradient $\mathbf{\nabla}\|\mathbf{v}\|$ (vertical); and the vorticity direction along $\mathbf{\nabla}\times\mathbf{v}$, here oriented perpendicular to the shear plane.}
\label{fig_linear-shear-flow}
\end{figure}
In practice, the strategy to approximately realize this flow field in simple situations is to horizontally drive the top and/or bottom plate(s) of the confining cell with constant speed.
We start by repeating simple effects that emerge for bulk-filling low-molecular-weight nematic liquid crystals. Increasing the complexity, we then briefly review aspects occurring for bulk-filling periodically modulated phases such as they appear in block copolymer systems. Finally we address localized structures in the form of vesicles.
\subsection{Flow alignment and director tumbling in nematic liquid crystals}
Similar to the previous chapter, we first turn to conventional low-molecular-weight liquid crystals deep in the nematic phase. Locally, the liquid crystalline molecules align on average along an ordering direction, the so-called director $\mathbf{\hat{n}}$ \cite{degennes1993physics}, see also Sec.~\ref{lmwlc}. In the hydrodynamic equations describing the macroscopic dynamic behavior of a nematic liquid crystal, the director orientation $\mathbf{\hat{n}}$ couples to the hydrodynamic flow field $\mathbf{v}$ \cite{degennes1993physics,pleiner1996pattern}.
For an imposed flow field as the one depicted in Fig.~\ref{fig_linear-shear-flow}, we now consider a bulk state of vanishing director distortions $\nabla\mathbf{\hat{n}}=\mathbf{0}$.
Then the hydrodynamic equation for the director orientation becomes \cite{pleiner1996pattern,rienacker1999orientational}
\begin{equation}\label{eq:dyn-n}
\frac{\partial\mathbf{\hat{n}}}{\partial t} =
\mathbf{\Omega}\cdot\mathbf{\hat{n}} +
\lambda \left[
\mathbf{A}\cdot\mathbf{\hat{n}}-(\mathbf{\hat{n}}\cdot\mathbf{A}\cdot\mathbf{\hat{n}})\,\mathbf{\hat{n}}
\right].
\end{equation}
In general, the two gradient tensors $\mathbf{\Omega}=[(\nabla\mathbf{v})^T-(\nabla\mathbf{v})]/2$ and $\mathbf{A}=[(\nabla\mathbf{v})^T+(\nabla\mathbf{v})]/2$
contain the rotational and elongational effect of the flow field $\mathbf{v}$, respectively. Here ${\nabla}\mathbf{v}$ again denotes a dyadic product and the superscript $^T$ its transpose. A convective contribution $\mathbf{v}\cdot\nabla\mathbf{\hat{n}}$ vanishes in our case due to the assumed spatially homogeneous director orientation $\nabla\mathbf{\hat{n}}=\mathbf{0}$.
We can see from Eq.~(\ref{eq:dyn-n}) that, besides the shear rate set by the imposed flow field, only one parameter $\lambda$ determines the dynamic behavior. It can be written as the ratio between two viscosities and thus represents material properties \cite{degennes1993physics,rienacker1999orientational}.
Depending on the magnitude of this parameter, the director can show qualitatively different types of behavior in the shear flow. On the one hand, for $|\lambda|\geq1$, the director reaches a steady-state orientation within the shear plane \cite{leslie1968some,degennes1993physics,gahwiller1972temperature,pieranski1974two}. This situation is typically called ``flow alignment'' or ``shear alignment'' and is depicted in Fig.~\ref{fig_lc_shear}~(a).
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure11.pdf}}
\caption{Illustration of three dynamic modes of a nematic liquid crystal under an imposed planar shear. The nematic director is indicated by the double-headed arrow, whereas the orientation of the shear is given by the two arrows in the top right and bottom left in each panel. (a) In the case of ``flow alignment'' or ``shear alignment'' the director takes a steady orientation within the shear plane and with a finite tilt angle with respect to the flow direction. (b) When ``wagging'', the director oscillates within the shear plane around an orientation of usually non-vanishing tilt with respect to the flow direction. (c) During ``tumbling'' motions the director makes full rotations in the shear plane.}
\label{fig_lc_shear}
\end{figure}
The orientation angle with respect to the flow direction, i.e.\ the ``flow alignment angle'', remains constant and is determined by the parameter $\lambda$. On the other hand, the director does not find a steady-state orientation, but continuously rotates in the shear plane, for $|\lambda|<1$. This dynamic state is usually referred to as ``tumbling'' \cite{gahwiller1972temperature,pieranski1974two}, see Fig.~\ref{fig_lc_shear}~(c), and consequently $\lambda$ is often referred to as the tumbling parameter.
Using the Fokker-Planck approach, these two dynamic regimes were connected to properties on the molecular level \cite{risken1996fokker,zwanzig2001nonequilibrium,archer1995molecular,kroger1995viscosity}.
The above considerations imply that the shear rate of the imposed flow field is low enough so that the degree of nematic ordering, see Eq.~(\ref{eq:s}), is not significantly altered from its equilibrium value.
Director tumbling was also observed and discussed for mixtures of liquid crystals \cite{ternet1999flow} and for liquid crystalline polymers \cite{burghardt1991role,rienacker1999orientational}. For certain parameter values, a transition was found with increasing shear rate from tumbling to flow alignment via an intermediate third dynamic state called ``wagging'' \cite{larson1990arrested,rienacker1999orientational}. In this third state, which is included in Fig.~\ref{fig_lc_shear}~(b), the director oscillates back and forth instead of performing full tumbling rotations. Still further dynamic modes were identified later, including chaotic types of motion \cite{rienacker2002chaotic}.
\subsection{Alignment and reorientations in periodically modulated phases}
At our next level of complexity, we concentrate on materials that in their ground states cannot be considered as spatially homogeneous any more. Instead, they feature regular periodic modulations in their density or in the concentrations of their constituents. An illustrative example are block copolymer melts or solutions.
To understand the situation in block copolymer melts, it is instructive to recall the problem of mixing of two polymers of a different kind. This scenario is addressed by the Flory-Huggins theory \cite{strobl1997physics}. For low-molecular-weight liquids, a spontaneous mixing is usually driven by a gain in translational entropy: each molecule is now provided with the extended volume of the whole mixture, not only with the one of its initial single component liquid. In polymers, however, this effect is strongly reduced. The monomers are chemically trapped in the polymer chains. Only each polymer chain as a whole can increase its translational entropy. Therefore the gain in translational entropy under mixing is significantly lower than for low-molecular-weight liquids. Instead, the inter-species interactions between the monomers of different kinds in comparison to the intra-species interactions between monomers of the same kind determine whether the polymers mix or form two separated phases. In most cases, this competition favors the separated state. Most pairs of polymers do not form homogeneous mixtures at ambient temperatures \cite{strobl1997physics}.
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure12.pdf}}
\caption{Schematic illustration of the two-dimensional patterns shown by micro-phase separated linear diblock copolymers. The temperature must be low enough so that micro-phase separation can take place. In the series from left to right, the fraction of one kind of polymer on the diblock copolymer chain with respect to the other one increases. This is depicted by a typical chain in each panel. If one species strongly dominates by its volume fraction, the state is disordered, i.e.\ not micro-phase separated, as in the most left and right examples. When the volume fractions become more balanced but still one species dominates, hexagonal textures can be observed as in the second examples from left and right. In three dimensions these would correspond to hexagonally arranged cylinders elongated perpendicular to the plane of the picture. Lamellar structures (center) appear for approximately equal volume fractions of the two blocks and are likewise obtained in three dimensions.
}
\label{fig_microphase}
\end{figure}
The situation becomes interesting in block copolymers. Here, the different polymer chains that tend to demix under the previous conditions are combined into one molecule.
For example, a diblock copolymer chain is composed of two chains of chemically different polymers that are covalently bound to each other at one of their ends.
Due to the covalent bonds between the blocks, a macroscopic demixing is not possible for block copolymers. Consequently, demixing can only occur on the length scale of a single block in the form of a micro-phase separation. These demixed blocks can then arrange into spatially periodic patterns. Depending mostly on temperature and the volume fractions of the two different blocks, lamellar phases, hexagonally arranged cylinders, bcc or fcc (in solutions) structures, and more complicated textures are obtained \cite{meier1969theory, leibler1980theory,ohta1986equilibrium,fredrickson1987fluctuation,bates1990block, fredrickson1996dynamics,strobl1997physics,hamley1998physics,hamley2001structure, teramoto2002double,nonomura2003formation,yamada2004kinetics}.
For approximately equal volume fractions of the two blocks, a lamellar micro-phase separated state emerges. It features a layered arrangement of the different blocks and was depicted in Fig.~\ref{fig_lengths}. Patterns expected for two-dimensional arrangements of linear diblock copolymers are depicted in Fig.~\ref{fig_microphase}.
When such materials are exposed to shear, an orientation of the periodic structures with respect to the shear velocity and its gradient has been observed in numerous experimental studies \cite{koppi1992lamellae,winey1993interdependence,balsara1994shear, balsara1994insitu,patel1995shear,zhang1995frequency,zhang1995symmetric, zhang1996annealing,fredrickson1996dynamics,zhang1997symmetric,maring1997threshold, wiesner1997lamellar,chen1997pathways,chen1998flow,pople1999shear,wang1999ordering, leist1999double,zipfel1999shear,wang2000birefringence,hamley2000effect, hamley2001structure,langela2002microphase}; see Fig.~\ref{fig_linear-shear-flow} for the shear geometry. The most illustrative example is the one of lamellar layered textures \cite{koppi1992lamellae,winey1993interdependence,balsara1994shear, balsara1994insitu,patel1995shear,zhang1995frequency,zhang1995symmetric, zhang1996annealing,fredrickson1996dynamics,zhang1997symmetric,maring1997threshold, wiesner1997lamellar,chen1997pathways,chen1998flow,pople1999shear,leist1999double, zipfel1999shear,hamley2000effect,hamley2001structure,langela2002microphase}.
For a simple planar linear shear flow, three principal orientations of the lamellae are possible as illustrated in Fig.~\ref{fig_bcp_shear}:
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure13.pdf}}
\caption{Illustration of the three principal orientations of a lamellar sheet-like structure under an imposed planar shear: (a) parallel, (b) transverse, and (c) perpendicular.}
\label{fig_bcp_shear}
\end{figure}
a ``parallel'' orientation with the lamellae normals parallel to the direction of the shear gradient; a ``transverse'' orientation with the lamellae normals parallel to the direction of the shear flow; and a ``perpendicular'' orientation with the lamellae normals perpendicular to the two previous cases, i.e.\ parallel to the vorticity direction.
Interestingly, when a steady linear shear flow
was applied to a lamellar block copolymer sample, the signature of the transverse orientation was observed at low shear rates \cite{pople1999shear}. This is surprising because it leads to a significant distortion of the lamellar layers. Switching to higher shear rates, a reorientation process occurred with the layer normals now lying in the plane spanned by the directions of the shear gradient and the vorticity \cite{pople1999shear}. This implies a combination between the parallel and the perpendicular alignment, which was also observed in other studies \cite{balsara1994shear,balsara1994insitu}. However, the steady shear could not induce a perfect reorientation. Even more, the steady shear led to defects in previously well-aligned lamellar samples for example in the form of focal conical textures \cite{winey1993interdependence}. It turned out that large-amplitude oscillatory shear was much more effective to align these samples than steady shear flow. Thus the orientational effects of oscillatory shear were studied extensively.
For large-amplitude oscillatory shear, reorientation transitions in lamellar phases were observed by varying the shear frequency or the shear amplitude \cite{koppi1992lamellae,balsara1994insitu,patel1995shear,zhang1995frequency,zhang1995symmetric, zhang1996annealing,zhang1997symmetric,maring1997threshold, wiesner1997lamellar,chen1997pathways,chen1998flow,leist1999double, zipfel1999shear,hamley2001structure,langela2002microphase}. Even a ``double-flip'' reorientation process can be found for increasing shear frequency. At low frequencies, a parallel orientation is often obtained \cite{balsara1994shear,balsara1994insitu,zhang1995frequency,zhang1996annealing,zhang1997symmetric, maring1997threshold,wiesner1997lamellar,leist1999double,zipfel1999shear}. An increase in the shear frequency can lead to a flip to the perpendicular orientation \cite{zhang1995frequency,zhang1996annealing,zhang1997symmetric, maring1997threshold,wiesner1997lamellar,leist1999double,zipfel1999shear}. Very high frequencies, however, can induce a second flip back into the parallel geometry \cite{patel1995shear,zhang1995frequency,maring1997threshold,wiesner1997lamellar,leist1999double}. Despite the significant distortion of the lamellae, the transverse orientation was also observed in the case of oscillatory shear \cite{zhang1995symmetric,wiesner1997lamellar,chen1997pathways}.
Experimentally, the investigations can be performed in a Couette cell or using rheometric devices as displayed in Fig.~\ref{fig_rheometer}.
\begin{figure}
\centerline{\includegraphics[width=9.cm]{figure14.pdf}}
\caption{Schematic illustration of two principal types of shear rheometer. (a) In the plate-plate geometry, the sample is sandwiched between two parallel plates, one of them rotating relatively with respect to the other around their common axis. The displacements on the plate surface and thus the shear amplitude increase linearly from the rotation axis in the center to the outside. (b) A cone-plate geometry counteracts this effect because the sample thickness likewise increases linearly from the central rotation axis to the outside. In both cases, steady shear (darker single-headed arrow) or oscillatory shear (brighter double-headed arrow) can be applied. The arrows also coincide with the local direction of the shear velocity, see Fig.~\ref{fig_linear-shear-flow}.
}
\label{fig_rheometer}
\end{figure}
In the plate-plate geometry, the sample is sandwiched between two parallel plates that rotate relatively with respect to each other around their common axis. At the surface of the rotating plate, the displacement increases linearly with increasing distance from the rotation axis in the center. Thus effects of different shear amplitude, defined here as displacement versus sample thickness, can be investigated simultaneously. A cone-plate geometry counteracts this effect. Here, one plate is replaced by a cone, leading to a linearly increasing sample thickness from the rotation axis in the center to the outside. Generally, the principal lamellar orientation in the sample cell can for instance be determined by small-angle x-ray or neutron scattering \cite{koppi1992lamellae,winey1993interdependence,balsara1994shear, balsara1994insitu,patel1995shear,zhang1995frequency,zhang1995symmetric, zhang1996annealing,fredrickson1996dynamics,zhang1997symmetric,maring1997threshold, wiesner1997lamellar,chen1997pathways,chen1998flow,wang1999ordering, leist1999double,zipfel1999shear,pople1999shear,hamley2000effect, hamley2001structure,langela2002microphase}. From the results of different beam directions, conclusions can be drawn about the orientational state.
Understanding the underlying reasons for the orientational behavior under shear is a complicated task and also material dependent. Several different effects connected to various length scales play together and determine the overall behavior. For example, the boundary of the sample cavity can have an aligning effect on the structure \cite{boker2002large}; the micro-phase separated textures are distorted by shear deformations, leading to elastic strains \cite{kawasaki1986phase,yamada2006elastic,tamate2008structural}; structural fluctuations can couple to shear deformations \cite{bruinsma1992shear}; viscosity contrasts are present in each sample due to the stacking of the molecules and due to the material properties of the different blocks \cite{patel1995shear,zhang1995frequency}; the molecules are elongated along the layer normal, which sets a preferred path of molecular motion in particular for entangled polymers at higher molecular weights \cite{zhang1995symmetric}; shear can induce a tilting of the elongated chains away from the layer normal, leading to a restoring torque \cite{chen1997pathways,auernhammer2000undulation,auernhammer2002shear,soddemann2004shear}; and as we have seen in the previous section on low-molecular-weight liquid crystals, elongated molecules by themselves can already show various types of dynamic alignment behavior. Apparently, different frequencies and amplitudes of the applied shear can be used to preferably address some of these ingredients. In this way, their relative impact on the collective dynamic response can be modified and the overall appearance of the sample can be tuned.
Other materials like common smectic (layered) low-molecular-weight liquid crystals \cite{safinya1991nematic,bruinsma1992shear,larson1993rheology,panizza1995effects}, see Fig.~\ref{fig_lc-phases}~(c), or lamellar phases of surfactant solutions \cite{bruinsma1992shear,berghausen1998shear,zipfel1999influence, zipfel2001cylindrical,mortensen2001structural,richtering2001rheology} can similarly feature orientational effects due to their ``sheet-like'' structure.
Reorientations under shear can also be observed in structural phases different from the lamellar one \cite{fredrickson1996dynamics,daniel2000effect,wang2000birefringence,hamley2000effect, hamley2001structure,angelescu2004macroscopic}. Besides, for hexagonally arranged cylinders as well as for spheres ordered in bcc- and fcc-textures, a sliding of whole structural sheets over each other was suggested \cite{doi1993anomalous,ohta1993anomalous,koppi1994epitaxial, berret1996shear,molino1998identification,daniel2000effect, daniel2001nonlinear,hamley2001structure}.
Applying shear to block copolymer systems is an issue of high practical relevance. Due to the reorientation effects, it can serve to heal multi-domain textures and obtain a monodomain structure \cite{winey1993morphology,winey1993interdependence,larson1993rheology, molino1998identification,chen1998flow,wang1999ordering}. Apart from shear, alignments can also be induced by other routes, for example by applying external electrical fields to materials that show a dielectric contrast between their blocks \cite{amundson1991effect,amundson1993alignment,amundson1994alignment, gurovich1994microphase,gurovich1995why,boker2002microscopic,boker2002large, kyrylyuk2002lamellar,zvelindovsky2003comment,boker2003electric, xu2005electric,olszowka2006large,lyakhova2006kinetic,olszowka2008control,liedel2012beyond}. From a technological point of view, block copolymers possess a high potential for applications in the nano-sciences \cite{hamley2003nanostructure,park2003enabling,darling2007directing}: they provide regular periodic arrays of tunable material properties on the molecular length scales of the different blocks. This can be exploited for nano-lithographic processes, to produce regular arrays of nano-dots or nano-wires, or to synthesize nanoporous materials. Monodomain structures are often beneficial for this purpose. Therefore understanding and clarifying the possible routes to generate them is a task of central importance.
\subsection{Vesicles in shear flow}
In the above cases, we considered the properties of bulk-filling phases under shear flow. This is different from the following examples, where we focus our attention on localized finitely-sized objects such as
closed membranes and vesicles. In contrast to a bulk-filling periodic phase, complete rotations of a limited object in a liquid environment is easily possible. Together with the deformability of the considered entities, this leads to new dynamic states.
The systems that we focus our attention on are based on amphiphilic molecules in an aqueous environment \cite{lipowsky1991conformation,seifert1997configurations}. Such molecules typically feature a hydrophilic (``water-loving'') head group and one or more hydrophobic (``water-fearing'') tails, usually based on hydrocarbon chains; see Figs.~\ref{fig_lengths} and \ref{fig_micellar-aggregates}~(a).
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure15.pdf}}
\caption{Schematic examples of amphiphilic (surfactant) molecules and illustration of some of the aggregates that they can form in an aqueous solution. (a) Two extreme examples of typical structures of amphiphilic molecules are displayed. Both have a hydrophilic (``water-loving'') head group, typically in the form of some polar or charged molecular groups. The tails are usually based on hydrocarbon chains and hydrophobic (``water-fearing''). In the first example, the head group is relatively bulky and there is only one shorter tail. In the second example, there are two tails of approximately the same diameter as the head group. The formation of aggregates in an aqueous environment in both cases is possible above a certain concentration. It leads to a screening of the hydrophobic tails from the aqueous environment by the hydrophilic head groups. (b) Due to their geometric properties, the molecules of the first kind favor the formation of micellar aggregates of high surface curvature. (c) The molecules of the second kind, due to their geometric shape, are suitable to generate flat bilayer membranes. Thermal fluctuations are of course present but not indicated here. To close the membranes at their outer ends, some arrangements of high surface curvature and thus higher energy are necessary. (d) For large membranes, this can be avoided by forming closed surfaces in the form of vesicles. The ratio of their radius versus the membrane thickness in reality can be much larger than depicted here, so that locally the surface curvature of the vesicle becomes negligible. See also the indicated sizes in Fig.~\ref{fig_lengths}.}
\label{fig_micellar-aggregates}
\end{figure}
To reduce the contact with the surrounding water molecules, the hydrophobic chains above a critical concentration tend to form intermolecular aggregates. This is again a sort of micro-phase separation. A simple example is depicted in Fig.~\ref{fig_micellar-aggregates}~(b) in the form of a micelle. In this aggregate, the hydrophobic chains can hide away from the water molecules and form the core of the micelle. They are screened from the water molecules by the outward-pointing hydrophilic head groups that search the contact to the aqueous environment.
Both, the concentration and the geometry of the single amphiphilic molecules determine which sort of intermolecular aggregates form.
If the head groups are very bulky and if there is only one short chain per molecule, the formation of spherical micelles as depicted in Fig.~\ref{fig_micellar-aggregates}~(b) is supported. They have a high curvature of their surface.
In contrast to that, if the hydrophilic head and the hydrophobic tail groups have approximately the same diameter, the formation of aggregates with low curvature of the surface is beneficial. This situation can for example be observed for bilayer membranes. In this case, the amphiphilic molecules are arranged in two adjacent layers stacked on each other as illustrated in Figs.~\ref{fig_lengths} and \ref{fig_micellar-aggregates}~(c). Here, the hydrophobic chains point to the inside of the two stacked layers and expel the water molecules. The hydrophilic heads point to the outside and screen the inner region from the aqueous environment.
At its lateral rims, the bilayer membrane would either have to expose some of the hydrophobic chains to the aqueous environment or show an unnaturally high curvature of its surface.
To avoid such high-energy regions at the lateral rims, the bilayer can bend on length scales large compared to its thickness and form a closed surface. The resulting closed bilayer membrane is called a vesicle and also indicated in Figs.~\ref{fig_lengths} and \ref{fig_micellar-aggregates}~(d). Typically, an aqueous solution is found on both sides of the bilayer membrane, i.e.\ outside the vesicle and also enclosed on the inside.
Interesting dynamic states arise due to the typical constraints following from the architecture of vesicles. On experimental time scales, their inner volume is generally conserved \cite{deuling1976curvature,seifert1997configurations}. Most importantly, also the surface area of vesicles remains practically constant \cite{deuling1976curvature,seifert1997configurations}. This is in contrast to simple liquid drops and qualitatively affects their shape and dynamics \cite{danker2007rheology,danker2008rheology}. Apart from that, a possible viscosity contrast between the inner encapsulated fluid and the outer fluid impacts the dynamic behavior
\cite{biben2002advected,biben2003tumbling,beaucourt2004steady,rioual2004analytical,biben2005phase, kantsler2005orientation,kantsler2006transition,mader2006dynamics, misbah2006vacillating,noguchi2007swinging,lebedev2007dynamics, danker2007dynamics,danker2007rheology,mader2007coupling, finken2008two,kessler2009elastic,messlinger2009dynamical,deschamps2009phase, noguchi2009swinging,ghigliotti2010rheology, finken2011micro,farutin2011symmetry,biben2011three,zabusky2011dynamics, farutin2012squaring,gires2012hydrodynamic,farutin2013analytical,lamura2013dynamics}.
The same is true for a possible friction between the two layers of amphiphilic molecules that form the bilayer membrane and may slide over each other during certain dynamic modes \cite{seifert1997configurations}.
On the one hand, the membrane may correspond to a two-dimensional liquid that does not sustain in-plane shear stresses, see Fig.~\ref{fig_inplaneshear}.
\begin{figure}
\centerline{\includegraphics[width=6.cm]{figure16.pdf}}
\caption{A schematic vesicle is half-cut to show the membrane location. Within the plane of the membrane, a shear stress is applied to the membrane as indicated by the two arrows. If the membrane does not sustain an in-plane shear stress but behaves like a two-dimensional liquid, the vesicle is called a ``fluid vesicle''. Likewise, it is referred to as a ``viscous vesicle'' when the effect of the intra-membrane viscosity is emphasized. In contrast to that, if the membrane does sustain in-plane shear stresses, the vesicle is called an ``elastic vesicle''.
}
\label{fig_inplaneshear}
\end{figure}
Then the vesicles are called ``fluid vesicles'' \cite{lipowsky1991conformation,seifert1997configurations}. We concentrate on the dynamics of such objects in simple linear planar shear flow as introduced in Fig.~\ref{fig_linear-shear-flow}.
If there are no viscosity contrasts between the liquid on the inside and on the outside of the vesicle, a dynamic state called ``tank-treading'' is forecast by modeling and observed experimentally \cite{kraus1996fluid,haas1997deformation,biben2002advected, biben2003tumbling,beaucourt2004steady,kantsler2005orientation,mader2006dynamics, noguchi2007swinging,lebedev2007dynamics,danker2007dynamics,danker2008rheology, messlinger2009dynamical,deschamps2009phase,biben2011three,zabusky2011dynamics}. It is depicted in Fig.~\ref{fig_vesicle_shear}~(a): due to the elongational part of the shear flow the vesicle becomes oriented within the shear plane, but with a constant inclination angle to the flow direction; the rotational part of the shear flow induces a rotational motion of the membrane around the vesicle interior.
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure17.pdf}}
\caption{Illustration of three dynamic modes of a vesicle under an imposed planar shear flow. The vesicle is indicated by the ellipsoid. (a) While ``tank-treading'', the vesicle takes a steady finitely inclined orientation with respect to the flow direction within the shear plane. Only the vesicle membrane performs a continuous rotational motion along its contour. (b) During ``swinging'' motions (also ``trembling'' or ``vacillating breathing'') the long axis of the vesicle oscillates within the shear plane around an orientation that is usually finitely inclined with respect to the flow direction. (c) When ``tumbling'' the vesicle performs full rotations in the shear plane.}
\label{fig_vesicle_shear}
\end{figure}
Increasing the viscosity contrast at not too high shear rates, a bifurcation to a different type of motion called ``tumbling'' is predicted from modeling and observed experimentally \cite{biben2002advected,biben2003tumbling,beaucourt2004steady, rioual2004analytical,kantsler2006transition,mader2006dynamics, noguchi2007swinging,lebedev2007dynamics,danker2007dynamics,danker2008rheology, messlinger2009dynamical,deschamps2009phase,biben2011three,zabusky2011dynamics}. It is illustrated in Fig.~\ref{fig_vesicle_shear}~(c). Here, the vesicle as a whole rotates in the shear flow, instead of only its membrane tank-treading around its interior.
Thermal fluctuations can induce tumbling already for vanishing viscosity contrast at low shear rates \cite{abreu2012effect}.
At higher shear rates, ``swinging'' (also ``trembling'' or ``vacillating breathing'') was observed instead of pure tumbling at least for intermediate viscosity contrasts \cite{kantsler2006transition,misbah2006vacillating,noguchi2007swinging, danker2007dynamics,lebedev2007dynamics,danker2008rheology, messlinger2009dynamical,deschamps2009phase,biben2011three, zabusky2011dynamics}. In this state, the long axis of the vesicle does not perform full rotations as in the tumbling mode, but only oscillates up and down as indicated in Fig.~\ref{fig_vesicle_shear}~(b).
Summarizing, a tendency towards the right-hand side in the series of panels (a)--(c) in Fig.~\ref{fig_vesicle_shear} is generally supported by an increasing viscosity contrast between the outside and the inside of the vesicle and a decreasing shear rate. Yet, the situation can be more involved in each individual case.
More complex dynamic modes were identified recently \cite{biben2011three,farutin2012squaring}, and the qualitative influence that thermal fluctuations can have were pointed out \cite{deschamps2009dynamics,zabusky2011dynamics,levant2012amplification,abreu2013noisy}.
Apart from that, the impact of an explicitly viscous membrane was investigated \cite{noguchi2004fluid,noguchi2005dynamics,noguchi2005vesicle,noguchi2007swinging}. These ``viscous vesicles'' generally show a reduced inclination angle during tank-treading. Increasing the membrane viscosity can further induce a transition from tank-treading to tumbling or swinging.
On the other hand, the membrane may for example be polymerized so that it does sustain in-plane shear stresses \cite{lipowsky1991conformation,seifert1997configurations}, see Fig.~\ref{fig_inplaneshear}. Such objects are referred to as ``elastic vesicles'' \cite{noguchi2005shape}.
The basic features of the dynamics of fluid vesicles in linear shear flow are recovered both from modeling and experiments: a tumbling motion at low shear rate and a tank-treading motion at high shear rate emerge \cite{abkarian2007swinging,kessler2008swinging,kessler2009elastic, noguchi2009swinging,abreu2012effect,koleva2012deformation}. However, during the tank-treading motion, the orientation angle with respect to the flow direction oscillates due to the shape-memory \cite{abkarian2007swinging,kessler2008swinging,kessler2009elastic, noguchi2009swinging,abreu2012effect,koleva2012deformation}. This is referred to as ``swinging''. Furthermore, if the viscosity inside becomes large compared to the outside viscosity, an intermittent type of motion is observed as a combination between the swinging tank-treading and the tumbling dynamics \cite{abkarian2007swinging,kessler2008swinging,kessler2009elastic, noguchi2009swinging,abreu2012effect}.
From a biological point of view, elastic vesicles form an important topic because red blood cells fall into this category \cite{abkarian2007swinging,dupire2010chaotic}. To understand the transport of red blood cells in blood vessels, the dynamics of elastic vesicles subjected to Poiseuille flows in cylindrical tubes, see Fig.~\ref{fig_poiseuille}~(a), is studied extensively.
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure18.pdf}}
\caption{Schematic illustration of different possible shapes adopted by vesicles in a Poiseuille flow. The snapshots have to be considered as two-dimensional cross-sectional views through the center of a tube-like channel. (a) Profile of the Poiseuille flow across the channel. (b) Elongation of a discoidal vesicle along the fluid flow. (c) Parachute-like shape and orientation. (d) Slipper-like configuration. Gray bars indicate the channel walls.}
\label{fig_poiseuille}
\end{figure}
Whereas fluid vesicles of initially discoidal shape typically elongate in the Poiseuille flow, see Fig.~\ref{fig_poiseuille}~(b), or show axisymmetric ``parachute'' shapes in strong flow fields, see Fig.~\ref{fig_poiseuille}~(c), elastic vesicles usually adopt a parachute-like deformation or an asymmetric ``slipper''-like shape as indicated in Fig.~\ref{fig_poiseuille}~(d) \cite{noguchi2005shape,kaoui2009why}. Many vesicles together in a tube were found to adopt disordered distributions with discoidal shapes, ordered states of aligned parachutes, as well as slippers arranged in a zigzag-like fashion \cite{mcwhirter2009flow,mcwhirter2011deformation,mcwhirter2012ordering} depending on the flow velocity and packing fraction. Hydrodynamic lift forces keep the vesicles away from the walls \cite{seifert1999hydrodynamic,cantat1999lift,lorz2000weakly,sukumaran2001influence, abkarian2002tank,callens2008hydrodynamic,messlinger2009dynamical}.
In addition to that, further dynamic effects have been discovered, for example a wrinkling deformation of the elastic membrane in shear flows \cite{walter2001shear,finken2006wrinkling,finken2011micro}.
Analytical approaches included the above-mentioned constraints following from the vesicle architecture and were able to analyze the resulting dynamic properties \cite{keller1982motion,seifert1999hydrodynamic,rioual2004analytical, finken2006wrinkling,misbah2006vacillating, noguchi2007swinging,lebedev2007dynamics,skotheim2007red,mader2007coupling, danker2007dynamics,danker2007rheology,finken2008two,danker2008rheology, kessler2009elastic,noguchi2009swinging,danker2009vesicles,noguchi2009swinging, farutin2011symmetry,finken2011micro,abreu2012effect, gires2012hydrodynamic,abreu2013noisy}.
Due to the curved vesicle surfaces, elements from differential geometry are necessary to perform these calculations. Apart from the analytical approaches, two simulation routes were particularly successful to investigate the dynamic behavior of vesicles in a liquid environment. One of them is particle-based and referred to as multi-particle collision dynamics \cite{noguchi2004fluid,noguchi2005dynamics,noguchi2005shape,noguchi2005vesicle, noguchi2006dynamics,finken2008two,messlinger2009dynamical,mcwhirter2009flow, noguchi2010dynamics,mcwhirter2011deformation, mcwhirter2012ordering,lamura2013dynamics}. It includes the surrounding and encapsulated fluid particles, modeling the many-particle interactions in an effective way. Still, it correctly obeys the physical conservation laws. The second route is a phase field model \cite{biben2002advected,biben2003tumbling,beaucourt2004steady,biben2005phase, jamet2007towards,jamet2008toward,ghigliotti2010rheology}. It is based on an auxiliary continuous scalar order parameter field. Two different constant values of this scalar order parameter are used to distinguish between the inside and the outside of the vesicle, for example $-1$ and $+1$, respectively. At the location of the membrane, the order parameter field rapidly but continuously changes between these two values. In this way, the boundary of the vesicle can easily be identified. It can be effectively tracked by simply advecting the order parameter field with the hydrodynamic flow of the system.
A third route to the problem was outlined in the form of a density-field approach \cite{menzel2011density}. For periodic bulk-filling textures, such a procedure had for example been introduced for block copolymer structures \cite{ohta1986equilibrium,kawasaki1986phase,bahiana1990cell, ohta1993anomalous,doi1993anomalous,drolet1999lamellae,nonomura2001growth, chen2002lamellar,yamada2006elastic,yamada2007interface, tamate2008structural} or for crystalline materials via the phase field crystal approach \cite{elder2002modeling,elder2004modeling,berry2006diffusive, athreya2007adaptive,provatas2007using,tupper2008phase, chan2009molecular,chan2010plasticity}.
These kinds of formalism are based on a single scalar real order parameter naturally given by the density field. They apply on diffusive time scales.
A corresponding free energy density to obtain localized nonperiodic structures was introduced in a phenomenological way and the resulting density-field description was coupled to hydrodynamic flow fields \cite{menzel2011density}.
In this way, in two spatial dimensions, characteristics of tank-treading were observed in linear shear flow, and folding into the parachute form as well as an elongation in a Poiseuille channel flow were obtained \cite{menzel2011density}.
A challenging task for the future is to derive such density-field approaches starting from a more microscopic and particle-resolved basis.
\newpage
\section{Active soft matter}
So far we have been concerned with ``passive'' systems that at most were driven from outside. We now turn to ``active'' or ``self-driven'' materials.
It is not always straightforward to clearly distinguish between such ``active'' soft matter \cite{ebbens2010pursuit,ramaswamy2010mechanics,romanczuk2012active, cates2012diffusive,marchetti2013hydrodynamics} and ``externally driven'' soft matter as considered in the previous section.
The term ``active'' generally implies that at least some components of the system feature an individual ``internal drive''.
Unfortunately, this definition requires a sort of restricted point of view. There is always a residual coupling of the internal driving mechanism to the outside world. We understand this statement by considering examples of active self-propelled particles.
One case of particularly well studied artificial self-propelled microswimmers
is given by dispersed colloidal Janus particles \cite{paxton2004catalytic,howse2007self,jiang2010active,volpe2011microswimmers,buttinoni2012active, theurkauff2012dynamic,buttinoni2013dynamical}.
In general, colloids are made of mesoscopic particles or droplets of sizes between $1$~nm and $10-100$~$\mu$m, see also Fig.~\ref{fig_lengths}, dispersed in another substance \cite{dhont1996introduction,palberg1999crystallization,ivlev2012complex}. Everyday examples are ink, paint, and milk. To stabilize colloidal suspensions, coagulation of the particles must be hindered. For this purpose, the refractive index of the colloidal particles and their environment should be matched to each other to reduce the van-der-Waals interaction. Furthermore, electric charges on the particles can lead to stabilization due to repulsion. Apart from that, also steric stabilization is possible, when the colloidal particles are covered by polymer brushes: close contact between the particles would restrict the possible configurations of the polymer chains, which induces an effective entropic short-range repulsion. Micron-sized colloidal particles offer the advantage that they can easily be tracked by optical methods.
There are different ways to cover one side of the surface of a colloidal particle by a material of significantly different properties than the other side \cite{hong2006simple,walther2008janus,jiang2010janus,chen2012janus}. Such colloidal particles are then called Janus particles. This can be exploited, e.g., to selectively heat the stronger absorbing side by laser light irradiated from outside \cite{jiang2010active}. As a result, a self-propelled motion can be induced via thermophoretic effects: the particle via its asymmetric light absorption generates a temperature gradient around itself; this leads to hydrodynamic stress gradients that can set the surrounding fluid into motion, which effectively leads to particle motion \cite{golestanian2007designing}. A further possibility is to enforce demixing of a binary fluid on the heated side \cite{volpe2011microswimmers,buttinoni2012active,buttinoni2013dynamical} or to selectively catalyze chemical reactions only on the surface of one side of the particles \cite{howse2007self,theurkauff2012dynamic}. The arising concentration gradients can lead to self-propulsion via diffusiophoretic effects analogous to the case of thermophoresis \cite{golestanian2007designing}. However, the fuel to drive these chemical reactions needs to be provided from outside, as must be the irradiated laser light in the case of thermophoretic motion.
Granular hoppers on a vibrating plate form another model system to study self-propelled motions \cite{narayan2007long,kudrolli2008swarming,deseigne2010collective,deseigne2012vibrated}. Here, typically the motion in the plane parallel to the surface of the plate is observed. However, this motion is initiated and driven from outside by the externally tuned vibration.
Finally, even such units as swimming bacteria \cite{berg1972chemotaxis,wada2007model,suematsu2011localized,drescher2011fluid}, crawling cells and amoebae \cite{rappel1999self,szabo2006phase,peruani2012collective}, or molecular motors that can set filaments into motion \cite{schaller2010polar,sanchez2012spontaneous} are not completely independent of their surroundings. At some point they need to take up the food or fuel provided by their environment.
We thus further narrow our present definition of active motion. The internal drive of the active units may be induced from outside. However, their motion must break a symmetry of the system. In other words, there must be a degree of freedom in the direction of induced motion that is not prescribed by the external driving. For example, a vertical vibration of the substrate can induce the motion of granular hoppers on a horizontal plate. However, the direction of motion within the plane of the plate surface is not specified. In that sense, the active particles are free to choose their migration direction.
In the following, we are mainly interested in collective effects that arise when many active and self-propelled particles act together. Several aspects of this complex behavior seem to have general character and can already be understood from minimal model systems. Therefore, we increase the level of complexity of the systems under investigation.
We start from the simplest case of point-like self-propelled particles featuring a constant self-propulsion velocity. Although this is an idealized situation, basic principles of the collective behavior are revealed. In a next step, we include systems where steric interactions occur, e.g.\ dry granular systems. After that, we outline the consequences of replacing the constant self-propulsion velocity by an active driving force. Artificial microswimmers like self-propelling colloidal Janus particles and biological swimmers in the form of bacteria are addressed. The influence of hydrodynamic interactions on their behavior is discussed. We then outline the role of particle deformability. Finally, we give examples of studies dedicated to the collective behavior of animals.
\subsection{Point-like self-propelled particles}
The most prominent and fundamental model introduced in the context of idealized point-like particles is the one by Vicsek et al.\ \cite{vicsek1995novel}, which is illustrated in Fig.~\ref{vicsekmodel} and explained in the following.
\begin{figure}[t]
\centerline{\includegraphics[width=.9\textwidth]{figure19.pdf}}
\caption{Illustration of the Vicsek model \cite{vicsek1995novel} in two spatial dimensions. All particles, marked by the dark bullets, self-propel with the same speed $v_0$. This is indicated by the velocity vectors attached to the bullets that all have the same magnitude. At each discrete numerical time $t$, the velocity orientation $\vartheta_i(t)$ of each particle $i$ is updated: it is set to the average velocity orientation of all particles located within a distance $r_0$ from the $i$th particle; furthermore some orientational noise is added. During each time step $\Delta t$, the particles move along their velocity orientations by a discrete distance $v_0\Delta t$, here shown for a unit time step $\Delta t=1$.}
\label{vicsekmodel}
\end{figure}
In this model, $N$ particles self-propel in a two-dimensional plane of periodic boundary conditions. As a key ingredient and major simplification, the magnitude of the individual velocities of all particles, i.e.\ their speed, is assumed to be equal to a constant $v_0$. It remains identical and unchanged for all times. The current orientation of the velocity vector of the $i$th particle can be parameterized by an angle $\vartheta_i$, $i=1,...,N$. This orientational angle $\vartheta_i$ is adjusted at each time step to the mean of the velocity orientations of all particles $j$ that are located within a spherical environment of radius $r_0$ around the $i$th particle:
\begin{equation}\label{vicsek_vartheta}
\vartheta_i(t+\Delta t) = \langle \vartheta_j(t) \rangle_{r_0} + \eta_i(t).
\end{equation}
Here, $\langle...\rangle_{r_0}$ describes this average over all particles within the sphere of radius $r_0$ around the $i$th particle. $t$ marks the time and $\Delta t$ sets the discrete time step. Furthermore, the angular noise $\eta_i(t)$ is originally taken as a random number with uniform probability out of a centered interval at each time step. After each time step $\Delta t$, the particle positions $\mathbf{r}_i$ are updated as
\begin{equation}\label{eq:vicsek_rupdate}
\mathbf{r}_i(t+\Delta t) = \mathbf{r}_i(t)+\Delta t\, v_0
\left(\begin{array}{c} \cos\left[\vartheta_i(t)\right] \\ \sin\left[\vartheta_i(t)\right] \end{array}\right).
\end{equation}
As a result of the competition between the alignment with the local environment and the stochastic noise, a phase transition is observed. This is a transition between a phase of disordered motion of zero net particle flux on the one hand and a phase of ordered collective motion on the other hand.
In the second case, all self-propelled particles migrate on average collectively into the same direction. The two phases are schematically indicated in the bottom insets of Fig.~\ref{fig_do-phase-transition}.
\begin{figure}
\centerline{\includegraphics[width=9.cm]{figure20.pdf}}
\caption{Schematic illustration of the order-disorder transition in the Vicsek model and its variants. The transition can be induced by increasing the characteristic noise amplitude or by decreasing the mean particle density. In the ordered state (bottom left inset) the particles on average migrate collectively into a common direction, whereas in the disordered state (bottom right inset) coherent particle motion does not occur. The order parameter $P$ is calculated from the magnitude of the sample-averaged velocity orientations, see Eq.~(\ref{eq:Pvicsek}). Close to the transition, spatial inhomogeneities are usually observed in the form of density bands that travel perpendicularly to their elongation direction through a diluted disordered background (top right inset).}
\label{fig_do-phase-transition}
\end{figure}
An order parameter to quantify this order-disorder transition is given by the magnitude of the sample-averaged velocity orientations,
\begin{equation}\label{eq:Pvicsek}
P(t)
= \left\|\frac{1}{N}\sum_{i=1}^N \left(\begin{array}{c} \cos\left[\vartheta_i(t)\right] \\ \sin\left[\vartheta_i(t)\right] \end{array}\right) \right\| ,
\end{equation}
where the velocity orientations are parameterized
by the orientation angles $\vartheta_i$ ($i=1,...,N$) [see Eqs.~(\ref{vicsek_vartheta}) and (\ref{eq:vicsek_rupdate}) as well as Fig.~\ref{vicsekmodel}]. The magnitude of the order parameter is $P=1$ in a completely ordered state of all particles collectively migrating into the same direction, and $P=0$ in the disordered state. As indicated in Fig.~\ref{fig_do-phase-transition}, the phase transition from the ordered to the disordered state of motion can be induced by
decreasing the mean particle density or increasing the characteristic noise amplitude.
This transition was studied in detail over the past few years \cite{bertin2006boltzmann,chate2008collective,bertin2009hydrodynamic}. In simulations of large systems, it was found that the transition is of first order \cite{gregoire2004onset,bertin2006boltzmann,chate2008collective,bertin2009hydrodynamic}. The discontinuity in the transition is apparently related to spatial inhomogeneities that arise in the particle density around the transition point. Density bands emerge that tend to collectively migrate perpendicularly to their elongation \cite{chate2008collective,bertin2009hydrodynamic,mishra2010fluctuations,peruani2011polar}
as indicated by the top inset in Fig.~\ref{fig_do-phase-transition}. Due to this directed collective migration, they can pick up further particles from the environment of disordered motion. Such a process further increases the density within the band. In this sense, a kind of self-supporting mechanism develops \cite{gopinath2012dynamical}. Also in real experiments, traveling density bands have been observed \cite{schaller2010polar}.
The nature of these traveling density bands was discussed in the framework of solitons \cite{chate2008collective,bertin2009hydrodynamic}. Typically the density bands feature a sharp front and an extended tail. Recently, their behavior under head-on collisions is increasingly investigated for point- and non-point-like particles \cite{ihle2013invasion,yamanaka2014formation,tarama2014individual}. Penetration of colliding density bands and recovery after collision have been observed \cite{ihle2013invasion,yamanaka2014formation}. It has been demonstrated that they are obtained as different propagating solutions of conventional continuum models for self-propelled particle crowds \cite{toner1995long,toner1998flocks} in the form of multiple parallel density bands, single solitary bands, and single active droplets \cite{caussin2014emergent}.
Several variants of the Vicsek model were pointed out and analyzed \cite{chate2008modeling}. For example, the effects of metric-free alignment interactions \cite{ginelli2010relevance} were discussed. In three spatial dimensions, the traveling density bands were recovered in the form of migrating density planes \cite{chate2008collective}. Other studies replaced the discrete nature of the dynamic equations Eqs.~(\ref{vicsek_vartheta}) and (\ref{eq:vicsek_rupdate}) by differential equations. In particular, the discrete averaging process $\langle...\rangle_{r_0}$ in Eq.~(\ref{vicsek_vartheta}) was replaced by a more continuous functional form \cite{peruani2008mean,peruani2010cluster,menzel2012collective}:
\begin{equation}
\frac{d\vartheta_i(t)}{dt} = -\frac{\partial U}{\partial\vartheta_i}+\Gamma_i(t), \qquad \frac{d\mathbf{r}_i}{dt}=v_0\left(\begin{array}{c} \cos\left[\vartheta_i(t)\right] \\ \sin\left[\vartheta_i(t)\right] \end{array}\right), \qquad i=1,\dots,N.
\end{equation}
For simplicity, the orientational noise $\Gamma_i(t)$ is assumed to result from a Gaussian white process. The continuous function $U$ is based on pairwise alignment interactions between the particles. For example, the functional form
\begin{equation}\label{eq_U}
U(\mathbf{r}_1,\dots,\mathbf{r}_N,\vartheta_1,\dots,\vartheta_N) = -\sum_{\substack{i,j=1\\ i<j}}^N \Theta\left(r_0-\|\mathbf{r}_i-\mathbf{r}_j\|\right)\cos(\vartheta_i-\vartheta_j)
\end{equation}
again leads to pairwise velocity alignment for particles closer to each other than the distance $r_0$. Here, $\Theta$ represents the Heaviside step function.
The snapshot in the top inset of Fig.~\ref{fig_do-phase-transition} was obtained from a numerical calculation following a procedure along these lines.
Using such an approach, the situation of a binary mixture was considered for different rules of the inter-species velocity alignment \cite{menzel2012collective}. Starting from the particle picture, continuum equations for the one-particle probability densities were derived \cite{menzel2012collective} within the Fokker-Planck framework \cite{risken1996fokker,zwanzig2001nonequilibrium,lee2010fluctuation,savel2003controlling}.
Also macroscopic hydrodynamic-like continuum equations for the macroscopic order parameters were obtained and analyzed \cite{menzel2012collective}.
Interestingly, when the above-mentioned density bands appear in one species, they can induce spatial heterogeneities in the other species via the inter-species coupling \cite{menzel2012collective}.
These results may be interesting for the dynamics of biofilm formation. Biofilms are surface- or interface-attached communities of microorganisms \cite{otoole2000biofilm,stoodley2002biofilms}, in nature usually composed of more than one species \cite{an2006quorum,elias2012multi}. Often at least part of the microorganisms in a film adopts a motile state \cite{kearns2005cell,an2006quorum,vlamakis2008control,veening2010gene}. Their collective behavior should be influenced by inter-species interactions.
Other variants of the Vicsek model, such as polar particles of apolar alignment interactions \cite{chate2008modeling,peruani2008mean,ginelli2010large,peruani2010cluster,peruani2011polar} or nematic particles that randomly reverse their migration direction \cite{chate2006simple,chate2008modeling} were investigated.
For the first kind of systems, density bands with particle migration along the contour of the band were observed \cite{chate2008modeling,ginelli2010large,peruani2011polar}, in contrast to the above-mentioned density bands that migrate perpendicularly to their elongation direction. An example of apolar alignment interaction is depicted in Fig.~\ref{fig_apolar}.
\begin{figure}
\centerline{\includegraphics[width=11.cm]{figure21.pdf}}
\caption{
Schematic illustration of polar alignment rules on the left and apolar alignment rules on the right. Self-propelled particles are indicated by rectangular boxes, while the arrows inside the boxes mark the direction of active drive. For polar alignment rules, the particles always tend to align their velocity orientations into the same direction. In the case of apolar alignment rules, which may be induced, for example, by steric interactions of elongated particles, an antiparallel orientation of the velocity vectors is equally preferred.
}
\label{fig_apolar}
\end{figure}
A further characteristic feature of self-propelled particle systems has been outlined for the active nematic case \cite{ramaswamy2003active}. We denote the average number of active particles in a subregion of the whole system by $N$ and the fluctuations of this number by $\delta\! N$. Typically a relation of the form ${\langle(\delta\! N)^2\rangle}^{1/2}\propto N^{\zeta}$ is obtained, with $\zeta=0.5$ in equilibrium systems. In the two-dimensional case of active nematics, an exponent $\zeta=1$ was predicted \cite{ramaswamy2003active,toner2005hydrodynamics} and later confirmed through simulations \cite{chate2006simple}. These large density fluctuations were termed ``giant number fluctuations''. Likewise, in the case of active polar particles an elevated exponent $\zeta>0.5$ was predicted \cite{toner1995long,toner1998flocks} and confirmed in simulations \cite{chate2008modeling,ginelli2010large,peruani2011polar}. Such large number fluctuations were also observed in experiments \cite{narayan2007long,deseigne2010collective,deseigne2012vibrated}. As mentioned below, in systems featuring steric interactions, large number fluctuations can result from a cluster transition during which self-propelled particles mutually block their active migration \cite{fily2012athermal,redner2013structure,buttinoni2013dynamical, bialke2013microscopic,fily2014freezing,stenhammar2013continuum,stenhammar2014phase}.
Finally, we note that two-dimensional systems of self-propelled particles can feature a long-ranged orientational order of their migration directions \cite{toner1995long,toner1998flocks,toner2005hydrodynamics}. This is in marked contrast to equilibrium systems \cite{mermin1966absence}. Long-ranged order can emerge in the active case because information about the orientations is additionally propagated by the self-propelled motion of the particles.
\subsection[Self-propelled particles featuring steric interactions]{\hspace{-.1cm}Self-propelled particles featuring steric interactions}
Originally, the point-like self-propelled particles in the Vicsek model \cite{vicsek1995novel} only interact via the alignment of their velocity orientations. In principle, this can lead to artifacts such as an unbounded growth of the local particle density.
As a first step to solve this problem, modifications of the Vicsek approach were introduced. For example, the migration directions of particles that come too close to each other can be reversed to bound the density \cite{czirok1996formation}. Or a pairwise alignment of the velocity vectors into opposite directions is induced for particles that are not well separated \cite{romenskyy2013statistical,menzel2013unidirectional,weber2014defect}. Another route adjusts the migration direction according to an inter-particle potential \cite{gregoire2003moving}. If a finite intermediate distance is preferred, crystal-like arrangements can arise within the particle swarm \cite{gregoire2003moving}. Such a situation can be interpreted as a combination of steric repulsive and cohesive attractive interactions \cite{gregoire2004onset}.
In addition to steric interactions between the particles, also steric interactions with confining walls can be considered. An example is a channel geometry made by two parallel confining plates \cite{menzel2013unidirectional}. Using a functional form for the alignment interactions as in Eq.~(\ref{eq_U}) in combination with discrete time steps, pattern formation was observed at intermediate overall particle densities in the channel \cite{menzel2013unidirectional}. On the one hand, lanes emerge along the channel direction that are either directly supported by the walls or ``free-standing'' in the interior \cite{menzel2013unidirectional}, see Fig.~\ref{fig_unidirectional}~(a). These lanes are ``unidirectional'' in the sense that all particles in the channel self-propel on average into the same direction. On the other hand, at slightly higher particle densities, active cluster crystals form that collectively migrate along the channel \cite{menzel2013unidirectional}, see Fig.~\ref{fig_unidirectional}~(b). The mechanism behind the formation of these structures results from a combination of two ingredients: first discrete migration steps of the self-propelled particles, and second the possibility for overreactions in the velocity alignment.
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure22.pdf}}
\caption{
Pattern formation in a modified Vicsek model of self-propelled particles confined in a channel between two parallel horizontal plates \cite{menzel2013unidirectional}. The patterns appear at intermediate overall particle densities. As indicated by the arrows, the structures in both cases collectively migrate along the channel direction. (a) Increasing the overall particle density from below, first a state of {unidirectional} laning appears \cite{menzel2013unidirectional}: all particles on average migrate into the \textit{same} direction. (b) At slightly higher particle densities, migrating cluster crystals emerge \cite{menzel2013unidirectional}. Each lattice point in this state is occupied by several self-propelled particles. The mechanism leading to the pattern formation is based on discrete migration steps of the self-propelled particles and the possibility of an overreaction in their velocity alignment. Channel walls are indicated by the gray bars and periodic boundary conditions are applied in the horizontal direction.
}
\label{fig_unidirectional}
\end{figure}
From an experimental point of view, dry systems of manifestly sterically interacting self-propelled particles can be realized by granular hoppers \cite{blair2003vortices,volfson2004anisotropy,galanis2006spontaneous,narayan2007long, aranson2007swirling,kudrolli2008swarming,deseigne2010collective,deseigne2012vibrated}. For this purpose, the granular particles are typically set into motion by a vertical vibration of the horizontal substrate. Only the in-plane displacements are recorded. Both, the migrations and ordering of apolar \cite{blair2003vortices,galanis2006spontaneous,narayan2007long,aranson2007swirling} and polar \cite{kudrolli2008swarming} rod-like objects were investigated, as well as the motion of polar disks \cite{deseigne2010collective,deseigne2012vibrated,weber2013long}. In the latter case, the steric interactions between the particles are isotropic due to the disk-like shape and do not explicitly provide an alignment interaction. Interestingly, the collective behavior of these vibrated polar disks could nevertheless remarkably well be mapped onto the Vicsek model \cite{deseigne2012vibrated}.
A polar granular hopper features an anisotropic density distribution along its body to break the forward-backward symmetry. Under vibration it thus shows an imposed preferred migration direction \cite{kudrolli2008swarming}. In contrast to that, for apolar hoppers the symmetry must be broken from outside by an additional horizontal mode of vibration \cite{aranson2007swirling}, or the symmetry must be broken spontaneously to perform steps of propagation. Vibrated apolar rods were observed to incline with respect to the substrate surface and to preferentially migrate into the inclination direction \cite{blair2003vortices,volfson2004anisotropy}. The mechanism of spontaneous symmetry breaking was investigated in more detail for a dimer model particle \cite{dorbolo2005dynamics}.
In the case of vibrated polar disks, the polarity results from breaking the symmetry of contact with the vibrating substrate by attaching two asymmetric legs \cite{deseigne2012vibrated}.
As revealed by these studies \cite{volfson2004anisotropy,dorbolo2005dynamics}, the ``dry'' friction between the particles and the substrate plays a major role for the propagation mechanism. The displacement statistics of ``dry'' objects on vibrated substrates were analyzed experimentally, also as a function of the substrate inclination \cite{goohpattader2009experimental,goohpattader2010diffusive,goohpattader2011stochastic}. Interestingly, the displacement distribution functions appear to have exponential tails \cite{goohpattader2009experimental,goohpattader2010diffusive,goohpattader2011stochastic}. Similar results were obtained for vibrated water droplets \cite{goohpattader2009experimental,mettu2010stochastic}
Typically the interactions with the substrate are modeled by a ``dry'' friction term of the Coulomb type \cite{persson2000sliding}. Generally, a simple model equation for the one-dimensional motion of a single granular particle exposed to dry friction with the substrate and an additional dynamic (viscous) friction with the remaining environment can be written in the form of a Langevin equation as \cite{gennes2005brownian,hayakawa2005langevin,baule2010path,baule2011stick, menzel2011effect,baule2013rectification}
\begin{equation}\label{Coulomb}
m\frac{dv}{dt} = -m\frac{v}{\tau}-\sigma(v)\Delta+\gamma(t), \qquad\frac{dx}{dt}=v.
\end{equation}
Here, $m$ is the effective mass of the particle, $v$ its (generally non-constant) velocity, and $x$ its position. The first term on the right-hand side of the velocity equation denotes the dynamic (viscous) friction with a relaxation time $\tau$. $\sigma(v)$ returns the sign of $v$, i.e.\ $\sigma(v>0)=+1$, $\sigma(v=0)=0$, and $\sigma(v<0)=-1$. Thus the second term on the right-hand side includes the dry Coulomb friction of constant strength $\Delta$. It implies that a moving particle is always slowed down by the same amount of deceleration, independently of its current speed. Finally, $\gamma(t)$ describes a stochastic force, which may arise from fluctuations of the environment and stochastic noise in the vibration mechanism.
Transferring Eq.~(\ref{Coulomb}) to the Fokker-Planck framework \cite{risken1996fokker,zwanzig2001nonequilibrium, kawarada2004non,gennes2005brownian,hayakawa2005langevin,menzel2011effect}, one finds that the steady-state velocity distribution function features a cusp at $v=0$ due to Coulomb friction \cite{kawarada2004non,hayakawa2005langevin,gennes2005brownian, baule2010path,baule2011stick,menzel2011effect,baule2012singular}. Interestingly, the irreversible part of the resulting Fokker-Planck equation can be mapped onto the case of a quantum mechanical harmonic oscillator with a pinning delta-potential at its center \cite{avakian1987spectroscopy,janke1988statistical,patil2006harmonic,touchette2010brownian,menzel2011effect}.
Starting from a sharp initial state, the spatial particle distributions develop non-Gaussian tails on intermediate time-scales \cite{menzel2011effect}. Multiple time-scales emerge in the evolution of the distribution functions due to Coulomb friction \cite{menzel2011effect}. Nevertheless, despite such a non-Gaussian behavior, the mean-square displacement of the particles still grows linearly in time \cite{wang2009anomalous,menzel2011effect,wang2012brownian}.
It will be an interesting problem for the future to study in more detail the effect of aspects of the friction mechanism on the collective behavior of granular hoppers.
\subsection{Active driving force instead of constant velocity}
\label{sec-active-driving-force}
In the previous sections, we reported about the success of the Vicsek model, which, despite its simplicity, could even reflect the collective behavior in an example system of granular hoppers \cite{deseigne2012vibrated}. One of the key assumptions in the Vicsek model is the constant self-propulsion speed of each individual particle \cite{vicsek1995novel}. This is an adequate simplification in a dilute system. However, it becomes problematic in dense systems of interacting particles. In particular, jammed situations where particles mutually block their ways cannot be described adequately in this model.
It is then reasonable to switch to the picture of an active driving force \cite{hagen2011brownian,henkes2011active,zottl2012nonlinear,fily2012athermal,bialke2012crystallization, wittkowski2012self,redner2013structure,ni2013pushing,zottl2013periodic,ferrante2013collective,ferrante2013elasticity, abkenar2013collective,bialke2013microscopic,buttinoni2013dynamical,hagen2014gravitaxis} instead of a constant self-propulsion velocity. In the simplest general case, the particles are now characterized by their position $\mathbf{r}$, by their velocity $\mathbf{v}$, and by a polar unit vector $\mathbf{\hat{p}}$ that gives the current orientation of the active driving force. Here, situations can arise, in which the direction of the velocity $\mathbf{v}$ and the orientation $\mathbf{\hat{p}}$ of the driving force are not parallel. The magnitude of the velocity $\mathbf{v}$ can change over time, whereas the strength of the active driving force $F_{d}$ is usually kept constant for simplicity.
In many cases, the motion of the particles is overdamped. This is true for instance for colloidal Janus particles \cite{paxton2004catalytic,howse2007self,jiang2010active,volpe2011microswimmers,buttinoni2012active, theurkauff2012dynamic,buttinoni2013dynamical} or bacteria \cite{berg1972chemotaxis,wada2007model,suematsu2011localized,drescher2011fluid} that self-propel in an aqueous environment at low Reynolds numbers \cite{purcell1977life}. Overdamped Brownian dynamics is appropriate to describe this kind of motion and the velocity variable becomes redundant. The structure of the rescaled dynamic equations for an isotropic self-propelled particle of constant strength $F_d$ of its active driving force can be written as
\begin{equation}
\frac{d\mathbf{r}}{dt}=-\nabla U+F_{d}\,\mathbf{\hat{p}}+\mbox{\boldmath$\xi$\unboldmath},
\qquad\frac{d\mathbf{\hat{p}}}{dt}=(\mbox{\boldmath$\omega$\unboldmath}+\mbox{\boldmath$\eta$\unboldmath})\times\mathbf{\hat{p}}.
\label{constantdriving}
\end{equation}
Here, the potential $U$ contains for example the influence of external fields like gravity,
the steric interactions with other particles, or the steric interactions with confining walls. The angular velocity $\mbox{\boldmath$\omega$\unboldmath}$ results from the torque on the particle for example due to external alignment by a magnetic field. Both, $\mbox{\boldmath$\xi$\unboldmath}$ and $\mbox{\boldmath$\eta$\unboldmath}$, include the impact of stochastic fluctuations.
So far, most of the corresponding studies have been restricted to two spatial dimensions.
First, the stochastic motion of single isolated self-propelled particles was analyzed \cite{howse2007self,hagen2011brownian}. In the long-time limit, the stochastic noise determines the statistical properties. Consequently, in the long-time regime, the motion appears diffusive. However, the magnitude of the corresponding overall diffusion constant can be significantly increased by self-propulsion when compared to the case of an analogous passive particle \cite{howse2007self,hagen2011brownian,zheng2013non}. In contrast to that, self-propulsion significantly alters an intermediate time regime. For example, non-Gaussian displacement statistics can emerge at intermediate times in spite of a Gaussian form of the stochastic noise \cite{hagen2011brownian}. These results were confirmed by comparison to experimental systems of self-propelling colloidal Janus particles \cite{zheng2013non}. Also the impact of an external flow field was studied \cite{zottl2012nonlinear,zottl2013periodic}.
Apart from that, Eqs.~(\ref{constantdriving}) were generalized and investigated for the case of rod-like particles, where the diffusion and friction matrices become uniaxially anisotropic \cite{teeffelen2009clockwise}. For general particle shapes, aligning torques can arise due to steric interactions between anisotropic particles or because of interactions with confining boundaries. These aligning torques enter the equations via the angular velocity $\mbox{\boldmath$\omega$\unboldmath}$. Furthermore, a torque that leads to a continuous reorientation of the polar direction $\mathbf{\hat{p}}$ can arise from the mechanism of self-propulsion in combination with the particle shape or be actively generated. The stochastic motions of such circle swimmers were analyzed \cite{teeffelen2008dynamics,teeffelen2009clockwise} and compared with experiments \cite{kummel2013circular}. Self-organization in array-like vortex textures was found \cite{kaiser2013vortex}. Moreover, it was demonstrated that in three spatial dimensions the asymmetry of biaxial self-propelled particles can lead to helical and even to superhelical trajectories \cite{wittkowski2012self}.
Concerning the collective behavior of many interacting particles, we first focus on the case of spherical objects \cite{henkes2011active,fily2012athermal,bialke2012crystallization, bialke2013microscopic,buttinoni2013dynamical,ni2013pushing}. Due to their isotropic shape, they do not feature a steric alignment interaction. This promotes the emergence of new truly non-equilibrium effects. A spherical particle that self-propels towards a hard wall is blocked in its motion and slowed down \cite{elgeti2013wall}. It can only escape again when the direction of active drive has reoriented away from the wall. The time scale for this process is set by rotational diffusion and thus indirectly by the temperature of the system. During this time, other particles can hit and additionally get blocked, if the density and self-propulsion speed are high enough.
\begin{figure}
\centerline{\includegraphics[width=9.cm]{figure23.pdf}}
\caption{
Schematic illustration of the blocking mechanism of isotropically sterically interacting self-propelled particles. The self-driven particles are of spherical shape. Arrows indicate the direction of active drive which can reorient by rotational diffusion. A two-dimensional motion of the particles within the plane of the figure is assumed. (a) Two particles block each other, supported by a rigid wall. (b) Self-supported blocking can also occur in the bulk. If rotational diffusion is slow enough in comparison to the active drive and the density is high enough, both situations can serve as seeds for the formation of larger clusters.
}
\label{fig_trapped}
\end{figure}
Through such a mutual blocking, see Fig.~\ref{fig_trapped}~(a), wall-supported clusters can emerge\footnote{This was demonstrated for rod-like particles \cite{wensink2008aggregation}. In that case, after a certain time, the clusters can slip off the wall, supported by the reorienting steric interactions with the wall \cite{elgeti2013wall}. On the contrary, due to their isotropic steric interactions, self-propelled spheres are more effectively blocked by walls than rod-like particles \cite{elgeti2013wall}.}.
Interestingly, the analogous process is also observed in a self-supported way in the bulk \cite{fily2012athermal,redner2013structure} in two-dimensional systems, see Fig.~\ref{fig_trapped}~(b). Here, the blocking mechanism was analyzed in more detail \cite{bialke2013microscopic} and also investigated experimentally \cite{buttinoni2013dynamical}. A real phase separation into a clustered and a gas-like phase can be obtained as a steady state of the system \cite{redner2013structure,cates2013active,stenhammar2013continuum}. For clarity, it is stressed that these clusters form in the absence of any attractive forces and despite purely repulsive interactions between the particles. The clusters emerge solely from mutual steric blocking of the migration due to the non-equilibrium active drive.
Continuum equations were derived that trace the phase separation into the gas-like and clustered phases back to a force imbalance arising from the mutual steric blocking \cite{bialke2013microscopic}, and a weakly nonlinear stability analysis \cite{cross1993pattern} was carried out \cite{speck2014effective}. In this way, the nature of the transition of phase separation could be investigated in more detail. Interestingly, the onset of the clustering transition can be mapped onto the phase separation of passive particles described by the famous Cahn-Hilliard model \cite{cahn1958free}. In particular, this implies the existence of an effective free energy density that can characterize the non-equilibrium clustering transition \cite{speck2014effective}.
The analysis suggests a change of the clustering transition from continuous at high particle densities to discontinuous at lower particle densities. This change in the nature of the clustering transition was confirmed by numerical investigations \cite{speck2014effective}.
Naturally, the clusters imply large fluctuations of the local particle density \cite{fily2012athermal}. A reentrance of the fluid phase is observed as a function of the driving force \cite{bialke2013microscopic}.
At high densities, the active systems become jammed \cite{henkes2011active,ni2013pushing} or crystallize \cite{bialke2012crystallization,redner2013structure}. Active crystals were also observed experimentally \cite{palacci2013living}, with attractive particle interactions typically being involved in the crystallization process \cite{palacci2013living,mognetti2013living}.
On the one hand, active crystals have been studied numerically by particle simulations \cite{bialke2012crystallization,redner2013structure,ferrante2013collective,ferrante2013elasticity, weber2014defect}. The self-organization of purely repulsively interacting actively driven particles into crystal-like structures was investigated \cite{bialke2012crystallization}. Furthermore, the collective behavior of self-propelled particles that are ordered in crystal-like arrangements and interact by harmonic-spring potentials has been analyzed in detail \cite{ferrante2013collective,ferrante2013elasticity}. In the latter case, there is no explicit alignment interaction between neighboring particles. Nevertheless, the elastic interactions between them channel the individual attempts of self-propulsion towards one global orientation of motion. Finally, the whole structure collectively migrates into one common direction.
On the other hand, a field-theoretic approach was introduced to study the collective behavior of active crystals \cite{menzel2013traveling,menzel2014active}. It can be derived microscopically from classical dynamical density functional theory \cite{marconi1999dynamic,archer2004dynamical, elder2007phase,teeffelen2009derivation,tegze2009diffusion,jaatinen2009thermodynamics, lowen2010phase,wittkowski2010derivation,wittkowski2011polar} applying appropriate assumptions.
Two microscopic order parameter fields were used to characterize the state of the system composed of many self-driven particles.
The first one is a particle-resolved density field $\psi(\mathbf{r},t)$;
the second one is a polar order parameter field $\mathbf{P}(\mathbf{r},t)$ that characterizes the locally preferred orientation of the active driving direction $\mathbf{\hat{p}}$, see Eqs.~(\ref{constantdriving}). Coupled dynamic equations describe the time evolution of these order parameter fields \cite{menzel2013traveling,menzel2014active}.
Resulting periodic structures are displayed in Fig.~\ref{fig_activecrystal}. Concepts from two prominent continuum descriptions were unified in this approach. The first one is the phase field crystal model characterizing periodic modulations in the density field.
This model was successfully used to reproduce solidification and crystallization phenomena in conventional passive crystalline structures on particle-resolved length and diffusive time scales \cite{elder2002modeling,elder2004modeling,stefanovic2006phase,elder2007phase, teeffelen2009derivation,tegze2009diffusion,jaatinen2009thermodynamics, stefanovic2009phase,chan2010plasticity,ramos2010dynamical,tegze2011faceting}.
The second prominent approach contained here stems from the macroscopic Toner-Tu model for the collective motion of non-crystallized self-propelled particles \cite{toner1995long,toner1998flocks}.
It can distinguish between systems that feature spontaneous directed self-propulsion, for example emerging from explicit alignment interactions, and systems without spontaneous alignment of their self-propulsion directions.
Starting from initially disordered active systems, the field-theoretic approach shows that an active single crystal can form via the coarsening of multi-domain structures \cite{menzel2013traveling}. If a local alignment mechanism for the active driving directions of the particles exists,
the crystal finally collectively travels into one common migration direction. If there is no such alignment,
the active crystal remains at rest for small active drive, see Fig.~\ref{fig_activecrystal}~(a). However, beyond a threshold value of the active drive, the crystal still finds a common migration direction and starts to collectively travel. This is similar to the above-mentioned results from the particle-resolved simulations of active crystalline structures \cite{ferrante2013collective,ferrante2013elasticity}. With increasing active drive, a transition from traveling hexagonal to traveling rhombic, quadratic, and lamellar structures was observed \cite{menzel2013traveling}. Such a transition series is displayed in Fig.~\ref{fig_activecrystal}~(b)--(d).
\begin{figure}
\centerline{\includegraphics[width=\textwidth]{figure24.pdf}}
\caption{Active crystals as modeled and displayed in Refs.~\cite{menzel2013traveling,menzel2014active}. The particle-resolved density field $\psi(\mathbf{r},t)$ is encoded by the color, where brighter color means higher density. Thin needles, pointing from the thicker to the thinner ends, indicate the local orientation and magnitude of the polar orientational order parameter field $\mathbf{P}(\mathbf{r},t)$. Thick arrows mark directions of collective motion. Panels from left to right were obtained for an increasing strength of the active drive. Here, explicit alignment mechanisms for the self-propulsion directions were not included. Therefore the hexagonal active crystal in panel (a) at low active drive remains at rest. Beyond a threshold magnitude of the active drive, however, this structure can nevertheless start to migrate, as shown for the traveling hexagonal crystal in panel (b). Increasing the active drive further, structural transitions to traveling rhombic and quadratic (c) as well as traveling lamellar textures (d) were observed.
}
\label{fig_activecrystal}
\end{figure}
In addition to that, using the active phase field crystal theory \cite{menzel2013traveling}, a linear stability analysis of the traveling active single crystalline structures was performed \cite{menzel2014active}. This analysis is complicated by the fact that the density field in the crystalline ground state is already spatially modulated. It was found that the investigated collectively traveling active single crystals are linearly stable. The impact of hydrodynamic interactions on the stability of such active crystals was considered and is discussed in the next section.
Turning now to non-spherical self-propelled particles, a well-studied example is given by self-propelled rods actively driven parallel to their long axis \cite{peruani2006nonequilibrium,wensink2008aggregation,baskaran2008enhanced, baskaran2008hydrodynamics,elgeti2009self,yang2010swarm, wensink2012meso,baskaran2010nonequilibrium,wensink2012emergent, mccandlish2012spontaneous,kaiser2012how,abkenar2013collective,kaiser2013vortex,kaiser2013capturing}. The biological motivation arises for instance from experimental investigations on elongated bacteria
crawling or gliding on a substrate \cite{harshey2003bacterial,peruani2006nonequilibrium,aranson2007model,kearns2010field,peruani2012collective}. Another biological example are filaments actively driven by molecular motors. These motors can be fixed on a substrate at one end \cite{schaller2010polar,schaller2011frozen} or directly work between the filamental rods \cite{sanchez2011cilia,sanchez2012spontaneous}. Non-biological artificial experimental systems can in particular be realized by vibrating polar granular rods \cite{kudrolli2008swarming}. In all these cases, the anisotropy of the steric interactions provides a mechanism of particle alignment. It turns out that the self-propulsion supports nematic ordering of rod-like particles by lowering the transition density from the isotropic to the nematic state \cite{baskaran2008enhanced}. This is in line with the prediction of long-ranged orientational order in two-dimensional self-propelled particle systems \cite{toner1995long,toner1998flocks,toner2005hydrodynamics}. Apart from that, various different collective states of migration have been revealed for interacting self-propelled rods \cite{peruani2006nonequilibrium,wensink2008aggregation,yang2010swarm,wensink2012meso,wensink2012emergent,mccandlish2012spontaneous,abkenar2013collective}. Among those are disordered, jammed, clustered,
and swarming states, where in the latter case the rods organize in localized packets of common migration directio
\footnote{The notation in this context is not unambiguous. Sometimes polar swarms are also denoted as clusters, whereas sometimes non-polar clusters are denoted as swarms.}.
Laning states were identified \cite{wensink2012meso,wensink2012emergent,mccandlish2012spontaneous,abkenar2013collective} that feature parallel stripes or lanes within which all particles propel into the same direction. However, neighboring lanes in these systems show antiparallel migration directions, see Fig.~\ref{fig_laning} for a schematic illustration.
\begin{figure}
\centerline{\includegraphics[width=7.5cm]{figure25.pdf}}
\caption{
Schematic illustration of a laning state. Darker particles move upwards, whereas brighter particles migrate downwards. The particles organize themselves into neighboring lanes of opposite migration directions. Either the opposite migration directions of the individual particles are explicitly imposed from outside for particles driven by an external force; or the opposite migration directions emerge due to self-organization in systems of identical self-driven particles. In reality the lane thickness is not necessarily as regular as depicted here and can comprise different numbers of particles.
}
\label{fig_laning}
\end{figure}
This is in contrast to the above-mentioned case of unidirectional laning \cite{menzel2013unidirectional}. Laning as in Fig.~\ref{fig_laning} was previously observed in systems of externally driven particles, where for different particles a driving force into opposite directions was imposed from outside
\cite{dzubiella2002lane,chakrabarti2004reentrance,rex2008influence,wysocki2009oscillatory, lowen2010particle,vissers2011lane,ikeda2012instabilities,glanz2012nature}.
In the case of self-propelled particles, laning appears spontaneously without explicitly superimposing the opposite migration directions. Finally, even turbulent states were discovered in simulations, theory, and experiments on swimming bacteria \cite{cisneros2007fluid,wensink2012meso,dunkel2013fluid}, although these organisms swim in an environment of low Reynolds numbers. The source of this turbulence at low Reynolds numbers is the continuous energy input on the length scale of the individual self-driven particles.
Apart from that, the behavior of active particles featuring various shapes more complicated than rod-like was studied \cite{wensink2014controlling,nguyen2014emergent}. In particular, forward-backward symmetries can be broken through the particle shape \cite{wensink2014controlling}. Identical shapes, but differently applied active driving forces can lead to phase separation \cite{wensink2014controlling}. Depending on the relative orientation of the self-propulsion direction with respect to the particle shape, concave particles can self-organize into micro-rotors \cite{wensink2014controlling}. Finally, for binary mixtures of non-spherical active rotors, a phase separation into domains of clockwise and counter-clockwise rotations was identified \cite{nguyen2014emergent}.
\subsection{Role of hydrodynamic interactions}
Many realizations of self-propelled particles exist in a liquid environment. Examples are
artificial microswimmers as for instance the colloidal Janus particles addressed above \cite{paxton2004catalytic,howse2007self,jiang2010active,volpe2011microswimmers,buttinoni2012active, theurkauff2012dynamic,buttinoni2013dynamical}, and swimming bacteria \cite{berg1972chemotaxis,wada2007model,suematsu2011localized,drescher2011fluid}.
Any self-propelling swimmer sets its surrounding fluid into motion. The motion of the other swimmers, which are suspended in the fluid, is naturally influenced by this fluid flow. In turn, also the motion of all the other swimmers acts back onto the first one. These fluid-mediated interactions are referred to as hydrodynamic interactions. They are long-ranged.
To include hydrodynamic interactions, we must take into account the fluid flow. It is typically a very good approximation to consider the -- often aqueous -- surrounding liquid as incompressible, i.e.\ to set its mass density $\rho$ equal to a constant. Then the central equation of motion in fluid dynamics that determines the fluid flow, the Navier-Stokes equation \cite{landau1987fluid}, can be written in the form
\begin{equation}\label{eq:NS}
\rho\,\frac{\partial\mathbf{v}(\mathbf{r},t)}{\partial t} + \rho \left[\mathbf{v}(\mathbf{r},t)\cdot\nabla\right]\mathbf{v}(\mathbf{r},t) = -\nabla p(\mathbf{r},t) + \eta\nabla^2\mathbf{v}(\mathbf{r},t) + \mathbf{f}(\mathbf{r},t).
\end{equation}
It describes the influence of volume force densities acting on a fluid element at position $\mathbf{r}$ and time $t$. $\mathbf{v}(\mathbf{r},t)$ is the resulting flow velocity field of the fluid at position $\mathbf{r}$ and time $t$. The left-hand side of Eq.~(\ref{eq:NS}) gives the inertial contribution;
here the second term expresses the fact that the actual motion of the fluid elements must be taken into account to correctly determine the change of momentum at a spatial position $\mathbf{r}$.
On the right-hand side, the first term denotes the force density due to gradients in the pressure field $p(\mathbf{r},t)$, the second term includes dissipation with $\eta$ the viscosity of the fluid, and the last term takes into account any additional force density field $\mathbf{f}(\mathbf{r},t)$ acting on the fluid. In our case, $\mathbf{f}(\mathbf{r},t)$ contains the force densities that each swimmer exerts on its liquid environment.
Rescaling all lengths by a characteristic length scale $L$, all velocities by a characteristic velocity $V$, time by $L/V$, pressure by $\rho V^2$, and force density by $\rho V^2/L$, Eq.~(\ref{eq:NS}) takes the form
\begin{equation}\label{eq:NSresc}
\frac{\partial\mathbf{v}'}{\partial t'} + \left[\mathbf{v}'\cdot\nabla'\right]\mathbf{v}' = -\nabla' p' + \mbox{Re}^{-1}\nabla'^2\mathbf{v}' + \mathbf{f}',
\end{equation}
with space and time dependencies not marked explicitly any more. We find
\begin{equation}\label{eq:Re}
\mbox{Re}=\frac{\rho L V}{\eta}
\end{equation}
for the famous dimensionless Reynolds number.
Typical microswimmers in the form of colloidal Janus particles and swimming bacteria have dimensions $L$ in the range between $10$~nm and $10$~$\mu$m. A characteristic swimming speed would for example be $10$~$\mu$m$/$s. Thus, we can see from Eq.~(\ref{eq:Re}) that their swimming in an aqueous environment usually occurs at low Reynolds numbers $\mbox{Re}\ll1$ \cite{purcell1977life}.
We can infer the consequences of this estimate from Eq.~(\ref{eq:NSresc}). For low Reynolds numbers, the inertial terms on the left-hand side can be neglected when compared to the dissipative contribution on the right-hand side.
Hydrodynamics at low Reynolds numbers is thus described by Stokes' equation \cite{happel1983low}:
\begin{equation}\label{eq:Stokes}
\mathbf{0} = -\nabla' p' + \mbox{Re}^{-1}\nabla'^2\mathbf{v}' + \mathbf{f}'.
\end{equation}
This equation leads to some peculiarities for swimming at low Reynolds numbers that are markedly different from our every-day experiences. In particular, the equation is linear in the velocity field and instant in time. Consequently, for balanced pressure gradients, flow immediately stops when no further force density $\mathbf{f}'$ is applied to the fluid. This has qualitative consequences when we compare to the situation at higher Reynolds numbers.
At high enough Reynolds numbers, a swimmer can move by reciprocal shape changes, see also Fig.~\ref{figure_reciprocal}: first a quick shape change is performed in the form of a power stroke; then a slow recovery stroke follows in the exactly inverse way of the power stroke, only that it is performed significantly more slowly.
\begin{figure}
\centerline{\includegraphics[width=.8\textwidth]{figure26.pdf}}
\caption{Illustration of the qualitatively different situation for swimming at high and low Reynolds number (``scallop theorem'') \cite{purcell1977life}. The simple swimmer (black) performs a reciprocal shape change, i.e.\ the shape change during the power stroke is just the inverse of the shape change during the recovery stroke. Only the speed of shape change is significantly higher during the power stroke than during the recovery stroke. At high Reynolds numbers this leads to a net displacement after the cycle is completed due to inertial effects. At low Reynolds number, inertial effects are negligible and a net displacement cannot be achieved in this way. }
\label{figure_reciprocal}
\end{figure}
So the central difference between the two strokes lies in the speed of execution. In this situation, the inertial terms on the left-hand side of Eq.~(\ref{eq:NS}) can lead to a symmetry breaking and net motion.
However, reciprocal shape changes of the swimmer cannot lead to such net motion at low Reynolds numbers.
The symmetry-breaking inertial terms are missing in the Stokes equation Eq.~(\ref{eq:Stokes}).
These facts are often referred to as the ``scallop theorem'' \cite{purcell1977life}.
Furthermore, in the absence of pressure gradients, the swimmer motion completely stops together with the fluid flow $\mathbf{v}'$ as soon as the swimmer does not apply any force density $\mathbf{f}'$ to the fluid any longer. Apart from that, the sum of all instant forces exerted by a low Reynolds number swimmer on the surrounding fluid vanishes because each stroke that it performs is balanced by a counter-acting drag force due to the motion of its body. It can be shown that, with increasing distance from such a ``force-free'' swimmer, the induced flow field decays significantly more quickly than for an object that exerts a net force on the surrounding fluid \cite{yeomans2014introduction}. Furthermore, the swimmer must provide a suitable mechanism to break the symmetry and achieve a net forward motion \cite{lauga2011life}.
Different routes to reach this goal were pointed out. One example are non-reciprocal cycles of contraction and expansion of a straight three-linked-sphere swimmer \cite{najafi2004simple,golestanian2008analytic} as displayed in Fig.~\ref{fig_swimmertypes}~(a).
\begin{figure}[t]
\centerline{\includegraphics[width=\textwidth]{figure27.pdf}}
\caption{Illustration of different self-propulsion mechanisms for swimming at low Reynolds numbers. Panel (a) shows the three-linked-sphere swimmer suggested in Ref.~\cite{najafi2004simple}. Its deformation cycle (snapshots from top to bottom) is nonreciprocal. Due to hydrodynamic interactions, contraction or expansion of one bond leads to a different net displacement depending on whether the other bond is in the contracted or expanded state. The vertical bar is added for clarity to highlight that a net propulsion to the right has occurred during the cycle. Panel (b) depicts the nonreciprocal deformation cycle (again snapshots from top to bottom) of the two-linked-sphere swimmer introduced in Ref.~\cite{avron2005pushmepullyou}. In the completely contracted or expanded state of the bond, the swimmer exchanges the volume of the two spheres, for example by pumping the contents from one side to the other through the bond. Since the viscous drag is higher for the sphere of larger radius, the swimmer features a net propulsion to the right (again the vertical bar is added to make the net motion visible). Finally, panel (c) displays the minimal model swimmer suggested in Ref.~\cite{aditi2002hydrodynamic}. The outer spheres are small and experience a negligible viscous drag by the fluid. However, they represent centers of forces acting on the fluid and setting it into motion (indicated by the bright arrows). Due to the antiparallel alignment of the forces, the total applied force vanishes. In the upper case, the swimmer pushes the fluid outwards, which is why it is referred to as a ``pusher'', while in the lower case, it pulls the fluid inwards and would be called a ``puller''. The main swimmer body is asymmetrically located with respect to the force centers. Thus it experiences a net viscous drag by the self-induced fluid flow.}
\label{fig_swimmertypes}
\end{figure}
Another example are volume changes of the spheres of a two-linked-sphere swimmer during contraction and expansion \cite{avron2005pushmepullyou}, see Fig.~\ref{fig_swimmertypes}~(b). Apart from that, the rigid structure of a swimming object can already break the symmetry by itself. Ideal candidates for this purpose are helical objects that have a certain handedness \cite{berg2003rotary,yonekura2003complete,wada2007model,ghosh2009controlled,zhang2009artificial, spagnolie2011comparative,tottori2012magnetic}. And as mentioned above, Janus particles intrinsically break the forward-backward symmetry, which can be exploited to start a net-propulsion mechanism \cite{paxton2004catalytic,howse2007self,jiang2010active,volpe2011microswimmers,golestanian2007designing, buttinoni2012active,theurkauff2012dynamic,buttinoni2013dynamical}.
One swimming strategy for a swimmer at low Reynolds numbers is to exploit the drag resulting from its self-induced fluid flow. To understand this principle, it is most illustrative to consider two force centers as shown in Fig.~\ref{fig_swimmertypes}~(c). In the upper case, the forces applied to the fluid are oriented in a way that the swimmer pushes the fluid outwards. It is therefore called a pusher. The bottom case shows the opposite situation, in which the swimmer pulls the fluid inwards. Thus the swimmer is referred to as a puller. If the swimmer body is located asymmetrically with respect to the centers of force acting on the surrounding fluid, it will experience a drag from the self-induced fluid flow \cite{aditi2002hydrodynamic,hatwalne2004rheology,baskaran2009statistical}. Thus the swimmer will show a net propulsion. This propulsion principle can be generalized to more realistic swimmer shapes.
We now turn to the situation of more than one swimmer suspended in the fluid. Then one of them feels the drag due to the resulting fluid flow induced by all other swimmers, and vice versa.
The impact that these hydrodynamic interactions between active swimmers has on their net interactions and on their collective behavior is still under investigation. For the three-linked-sphere swimmer mentioned above \cite{najafi2004simple}, the pairwise hydrodynamic interaction was analyzed explicitly \cite{pooley2007hydrodynamic}. A complicated dependence on the relative distance, orientation, and phase of the shape change was obtained. Another study focused on the averaged flow fields around circle swimmers \cite{fily2012cooperative}. For particles that propel by pushing the fluid, a repulsive interaction was found on average, whereas for swimmers that pull the fluid, the average interaction was attractive. A pair of counterrotating pullers performs a net translational motion \cite{leoni2010dynamics,fily2012cooperative}. Straight-swimming pullers that were forced to swim parallel to each other were observed to repel each other \cite{gotze2010mesoscale,molina2013hydrodynamic}, whereas pushers in the same set-up were found to attract each other \cite{cisneros2007fluid,gotze2010mesoscale}. However, when released from the constraint of parallel swimming, no permanently bound state due to hydrodynamic interaction was found in spite of the initially attractive configuration \cite{gotze2010mesoscale}.
Several cases of motion in a liquid environment were identified in which hydrodynamic effects seem not to play a crucial role.
For bacterial cells, it was demonstrated that generally cell-cell interactions and cell-surface interactions are dominated by orientational diffusion, steric interactions, and lubrication forces, and that hydrodynamic interactions play a minor role \cite{drescher2011fluid}. Apart from that, different collective effects observed experimentally for microswimmers were successfully reproduced in simulations without including hydrodynamic interactions. Examples are the previously addressed observations of clusters and turbulent states \cite{buttinoni2013dynamical,wensink2012meso}.
Nevertheless, there are many situations in which hydrodynamic interactions were shown to be important, both for the single-swimmer and for the collective behavior.
For instance, hydrodynamic interactions with confining walls can modify the migration properties \cite{blake1971note,or2009dynamics,elgeti2010hydrodynamics, brotto2013hydrodynamics,li2014hydrodynamic, zottl2014hydrodynamics}. As demonstrated numerically, the nature of the propulsion mechanism can determine whether active microswimmers preferentially orient towards a confining surface, or whether they turn away and leave the wall \cite{li2014hydrodynamic}. Under narrow confinement between two walls, the propulsion mechanism can support orientations towards one of the walls or orientations parallel to their surfaces, which supports blocking or planar swimming motion, respectively \cite{li2014hydrodynamic,zottl2014hydrodynamics}. Naturally, this influences the formation of the clusters mentioned in Sec.~\ref{sec-active-driving-force} \cite{buttinoni2013dynamical,li2014hydrodynamic,zottl2014hydrodynamics}. Apart from that, a speed-up of microswimmers through hydrodynamic interactions with a flexible confining tube was predicted \cite{ledesma2013enhanced}. Microswimmers confined in a harmonic trap were found to orientationally order due to hydrodynamic interactions and induce a net fluid flow \cite{hennes2014self}. Moreover, in experiments, a transient capturing of microswimmers in circular trajectories around passive colloidal spheres was observed \cite{takagi2014hydrodynamic}.%
Generally, hydrodynamic interactions can induce the formation of vortices and swirls \cite{hernandez2005transport}. In that sense, hydrodynamic interactions can destroy long-ranged orientationally ordered collective motion, but also more localized orientational order in swarms \cite{lushi2014fluid}. Likewise, in the case of two-dimensional active traveling crystals, destabilizing effects of hydrodynamic interactions were observed numerically \cite{menzel2014active}. The self-propelling particles forming the active crystal were considered to migrate on a substrate covered by a thin fluid film. As a consequence of the resulting flow fields, the traveling single crystal can break up into different domains, or the crystalline lattice structure can vanish altogether. The stability of different lattice structures under hydrodynamic interactions was analyzed \cite{desreumaux2012active}.
Recently, it became clear from theoretical and numerical investigations that hydrodynamic far-field and near-field interactions can differ in their influence on the orientational order and stability of the suspension. Naturally, far-field interactions prevail in dilute suspensions whereas near-field interactions become important in concentrated solutions of microswimmers. Most of these studies were performed using yet another swimmer model called ``squirmer'' \cite{lighthill1952squirming,blake1971spherical,ishikawa2006hydrodynamic}: a non-vanishing flow field is prescribed as a boundary condition for the surrounding fluid on the surface of each swimmer; the swimmers themselves are often considered as rigid objects, but the non-vanishing surface flow drives them through the liquid environment
Following these lines, we first consider dilute suspensions of swimmers that solely interact hydrodynamically. For those, it was demonstrated that orientationally ordered states are always unstable \cite{saintillan2008instabilities,saintillan2008pattern,saintillan2013active}. Furthermore, for elongated swimmers featuring a puller propulsion mechanism, it was found that isotropic and spatially homogeneous suspensions are stable \cite{saintillan2008instabilities,saintillan2008pattern,saintillan2013active}. In contrast to that, homogeneous isotropic suspensions of elongated pushers are unstable and show an interesting nonlinear dynamics including enhanced mixing \cite{saintillan2008instabilities,saintillan2008pattern,saintillan2013active}.%
When concentrated suspensions or the close proximity of confining boundaries are addressed, hydrodynamic near-field interactions become important. It is then often mandatory to explicitly resolve the induced flow fields in the vicinity of the swimmers. For this purpose, squirmer models are ideal candidates \cite{gotze2010mesoscale,evans2011orientational,alarcon2013spontaneous, zottl2014hydrodynamics,li2014hydrodynamic}. In the case of concentrated suspensions of spherical squirmers in the bulk, net polar orientational ordering of the swimming directions due to hydrodynamic interactions was numerically observed \cite{evans2011orientational,alarcon2013spontaneous}. This is in contrast to the above-mentioned results for dilute swimmer suspensions \cite{saintillan2008instabilities,saintillan2008pattern,saintillan2013active}. For pullers the degree of order was higher than for pushers \cite{evans2011orientational,alarcon2013spontaneous}. Likewise, as indicated above, hydrodynamic near-field interactions between two confining walls determine the relative orientation with respect to the boundaries \cite{li2014hydrodynamic,zottl2014hydrodynamics}. An enhanced cluster formation results for pusher squirmers when compared to pullers \cite{li2014hydrodynamic,zottl2014hydrodynamics}.%
Finally, hydrodynamic interactions can lead to synchronization. This particularly applies when active rotors or driven filaments are considered \cite{reichert2005synchronization,kim2006pumping,vilfan2006hydrodynamic,yang2008cooperation, polin2009chlamydomonas,kotar2010hydrodynamic,uchida2010synchronizationprl,uchida2010synchronization, uchida2011generic,uchida2011many,golestanian2011hydrodynamic,osterman2011finding, reigh2012synchronization,reigh2013synchronization,elgeti2013emergence, theers2013synchronization,theers2014effects}. It was summarized that rotor pairs can only synchronize if the system is not symmetric under exchanging the two rotors \cite{lenz2006collective,golestanian2011hydrodynamic}. Carpets of driven hydrodynamically coordinated active rotors were suggested to be used to pump or mix fluid in microfluidic devices \cite{uchida2010synchronizationprl}. Furthermore, carpets of actively driven deformable filaments are found on the surfaces of certain bacteria and can be used for self-propulsion \cite{brennen1977fluid,vilfan2006hydrodynamic,osterman2011finding}. Their coordinated motion can provide a swimming mechanism for these cells at low Reynolds numbers.
\subsection{Deformable self-propelled particles}
On our way of increasing complexity in active particle systems we now add a further degree of complication. Several self-propelled objects are not rigid. They deform as it is observed for instance for certain cells that crawl on substrates \cite{rappel1999self,maeda2008ordered,kaindl2012spatio}. Particularly obvious examples are self-propelled droplets on interfaces \cite{lee2002chemical,nagai2005mode,chen2009self} or in bulk fluid \cite{thutupalli2011swarming,kitahata2011spontaneous,kitahata2012spontaneous}. These droplets maintain chemical reactions \cite{thutupalli2011swarming,kitahata2011spontaneous,kitahata2012spontaneous} that lead to concentration gradients along their surfaces. A similar mechanism works by depositing chemicals from their inside to the outside \cite{lee2002chemical,nagai2005mode,chen2009self}. In effect, all these processes induce and maintain gradients of surface tension along their surfaces \cite{chen2009self,kitahata2011spontaneous,kitahata2012spontaneous, schmitt2013swimming,yoshinaga2014spontaneous}. These lead to stress gradients along the surfaces. On rigid substrates, such stress gradients on the droplet surface can directly push or pull on the droplet and lead to a net drive. In fluid environments, the surface gradients can induce convective flows on the inside and outside of the droplet surface. These convective flows in turn propel the droplets \cite{kitahata2011spontaneous,schmitt2013swimming,yoshinaga2014spontaneous}. If no special action is taken from the outside, the droplet motion has to start by spontaneous symmetry breaking; for instance by a sufficiently strong concentration fluctuation along its surface. In the ideal case, the concentration fluctuation leading to the net motion is maintained in a self-supported way by the resulting motion \cite{yoshinaga2012drift}, for example when the concentration gradients are enhanced by the induced fluid flows. It was demonstrated that self-propelled droplets can be relatively robust deformable objects. In an experiment, they were observed to deform and squeeze themselves through steric barriers \cite{chen2009self} and recover afterwards.
To study the impact of deformability on the particle motion, shape deviations from a spherical undeformed state can be taken into account in a systematic way \cite{hiraiwa2010dynamics}. The lowest order deformations are of axially symmetric elliptic shape.
Such elliptic deformations can be described by a symmetric traceless tensor $\mathbf{S}$ of second rank \cite{ohta2009deformation}. It is formally equivalent to the order parameter tensor introduced in Eq.~(\ref{eq:nematicOPtensor}) to characterize nematic liquid crystalline phases \cite{degennes1993physics}. Now, however, $s$ parameterizes the degree of deformation, while $\mathbf{\hat{n}}$ represents the orientation of the symmetry axis of the elliptic shape changes, see also Fig.~\ref{fig_deformable}.
\begin{figure}[t]
\centerline{\includegraphics[width=4.cm]{figure28.pdf}}
\caption{A self-propelled particle that deviates from its spherical ground state by axially symmetric elliptic deformations. The elliptically deformed state is described by the degree of deformation $s$ and by the orientation of the resulting ellipsoid. Since there is no deformational forward-backward asymmetry along the symmetry axis of the ellipsoid, we indicate its orientation by a double-headed arrow referred to as $\mathbf{\hat{n}}$. The self-propulsion contributes to the velocity $\mathbf{v}$ of the particle. Although representing separate degrees of freedom for each particle, the velocity $\mathbf{v}$ and the deformation tensor $\mathbf{S}=s(\mathbf{\hat{n}}\mathbf{\hat{n}}-\mathbf{I}/d)$ are generally coupled to each other, see Eqs.~(\ref{eq:ohtaohkuma1}) and (\ref{eq:ohtaohkuma2}).}
\label{fig_deformable}
\end{figure}
The relation to the nematic order parameter tensor comes from the fact that axially symmetric elliptic deformations are forward-backward symmetric. In Fig.~\ref{fig_deformable} this is indicated by the double-headed arrow that marks the symmetry axis of the ellipsoid.
Coupled dynamic equations between the elliptic deformation tensor $\mathbf{S}$ and the velocity vector $\mathbf{v}$ of the particle were derived on symmetry grounds and evaluated \cite{ohta2009deformable}. To lowest order in the coupling terms, they read \cite{ohta2009deformable}:
\begin{eqnarray}
\frac{\mbox{d}\mathbf{v}}{\mbox{d} t} &=& \gamma\mathbf{v}-\mathbf{v}^2\mathbf{v} -a\,\mathbf{S}\cdot\mathbf{v},
\label{eq:ohtaohkuma1}
\\[.1cm]
\frac{\mbox{d}\mathbf{S}}{\mbox{d} t} &=& -\kappa\mathbf{S}+b\!\left(\mathbf{v}\mathbf{v}-\frac{1}{d}\mathbf{v}^2\,\mathbf{I}\right).
\label{eq:ohtaohkuma2}
\end{eqnarray}
In passive systems, the coefficient $\gamma$ would be negative and the term would correspond to linear viscous friction with the environment. On the contrary, in active systems, this coefficient can become positive, $\gamma>0$. Then, the first two terms on the right-hand side of Eq.~(\ref{eq:ohtaohkuma1}) can lead to a stationary solution of non-vanishing velocity $\mathbf{v}\neq\mathbf{0}$, i.e.\ self-propulsion. An analogous approach can be found in the famous macroscopic continuum characterization of flocks of self-propelled particles by Toner and Tu \cite{toner1995long,toner1998flocks}. The orientation of the resulting non-vanishing velocity vector $\mathbf{v}$ is not fixed a priori and results from spontaneous breaking of the continuous rotational symmetry. In the second equation, i.e.\ in Eq.~(\ref{eq:ohtaohkuma2}), the coefficient $\kappa>0$ expresses that the shape tends to relax back to the undeformed ground state for vanishing velocity $\mathbf{v}$. As before, $\mathbf{v}\mathbf{v}$ represents a dyadic product, $d$ the dimensionality of the system, and $\mathbf{I}$ the unity matrix. Most importantly, the terms with the coefficients $a$ and $b$ include the leading-order coupling terms between the dynamic variables $\mathbf{v}$ and $\mathbf{S}$. These contributions are to a big extent responsible for the peculiar behavior of deformable self-propelled particles as it is briefly summarized in the following.
In two spatial dimensions, the deformability can induce a bifurcation from straight to circular motion with increasing active drive \cite{ohta2009deformable}. Also quasiperiodic motions as well as array-like trajectories featuring rectangular edges were observed \cite{hiraiwa2010dynamics}. Applying an external electric field to polarizable particles or an external gravitational field, further types of trajectories such as different kinds of zig-zag motions and cycloidal motions appear \cite{tarama2011dynamics}. In the absence of external fields but in three spatial dimensions, additional helical trajectories are obtained \cite{shitara2011deformable,hiraiwa2011dynamics}.
The collective behavior in two spatial dimensions was studied extensively \cite{ohta2009deformable,ohkuma2010deformable}. First, a global coupling was introduced that forces the deformed particles to align along their globally averaged deformation axis \cite{ohta2009deformable}. This coupling can lead to chaotic motion \cite{ohta2009deformable,ohkuma2010deformable}. After that, instead of the global coupling, a pairwise alignment interaction between deformed particles was imposed \cite{itino2011collective,itino2012dynamics}. In addition to that, as a steric interaction, a soft pairwise repulsive Gaussian potential was applied. Starting from random initial conditions at low or vanishing orientational noise amplitudes, the particles collectively order their migration directions as a function of time as expected \cite{vicsek1995novel}. All particles then collectively propel into the same direction. Moreover, they also tend to locally positionally order in a hexagonal way. Interestingly, it was observed that under compression of the system, implying an increase in the particle density, the ordered collective migration breaks down at a critical density \cite{itino2011collective,itino2012dynamics}. The system becomes disordered and fluid-like. This is just the opposite effect as to that expected from the Vicsek model, in which the orientational order in the motion increases continuously with the particle density \cite{vicsek1995novel}.
In a later study that used an anisotropic pairwise repulsive Gaussian potential, it was concluded that the disorder transition from a collectively traveling hexagonal crystal to a fluid-like state is enabled by the soft Gaussian interaction potential \cite{menzel2012soft}. The effect is well known in corresponding equilibrium systems. At low densities, these equilibrium systems are in a fluid phase, they crystallize at intermediate densities, but they return to a fluid state at high densities \cite{stillinger1976phase,prestipino2005phase, prestipino2011hexatic,prestipino2007phase}. A central ingredient for this reentrant fluidization is the boundedness of the Gaussian interaction potential, even when two particles are placed at the same position \cite{stillinger1976phase}. It is concluded that the analogous effect is observed in active systems of self-propelled particles interacting sterically by a pairwise Gaussian potential \cite{menzel2012soft} and can be supported by their deformability \cite{itino2011collective,itino2012dynamics,menzel2012soft}.
Apart from that, mutual deformations of the particles due to steric interactions between them were considered \cite{menzel2012soft}. This process can provide an implicit alignment mechanism for their self-propulsion directions as illustrated in Fig.~\ref{fig_deformable_alignment} for two particles. In many-particle systems, hexagonal lattices of self-propelled particles featuring a common migration direction can emerge that represent active collectively traveling crystals. Introducing a cut-off for the steric Gaussian interaction potential, traveling crystals of rectangular lattice structure are obtained. Likewise, active resting cluster crystals with more than one particle on each lattice site are observed, as they were previously reported for equilibrium systems \cite{likos1998freezing,schmidt1999an}. In cases of strong particle deformations, laning patterns can appear from randomly initialized systems \cite{menzel2012soft}. Adding orientational noise, the Vicsek transition \cite{vicsek1995novel} from collectively ordered to disordered motion is recovered with increasing noise amplitude \cite{menzel2012soft}.
\begin{figure}
\centerline{\includegraphics[width=4.8cm]{figure29.pdf}}
\caption{Aligning steric interactions between deformable self-propelled particles \cite{menzel2012soft}. Depicted is a time series from bottom to top with dotted lines indicating the particle paths. Two deformable self-propelled particles migrate towards a common point. At this point, their active drive pushes them together. Due to steric interactions, this leads to deformations. Eqs.~(\ref{eq:ohtaohkuma1}) and (\ref{eq:ohtaohkuma2}) imply that the deformations can reorient the self-propulsion directions, here towards the common axis of elongation of the particles. The outgoing angle between the particles is smaller than the ingoing angle, which means that an effective alignment mechanism is at work.}
\label{fig_deformable_alignment}
\end{figure}
Finally, the formation of propagating fronts was reported for two different cases. First, large systems of repulsive Gaussian interactions and reentrant fluidity become inhomogeneous at the reentrance density \cite{itino2012dynamics,tarama2014individual}. They show regions of collectively ordered and regions of disordered motion. In contrast to the traveling density bands observed in the Vicsek model \cite{chate2008collective,bertin2009hydrodynamic,peruani2011polar,menzel2012collective}, here a front of disordered motion propagates into the region of collectively ordered migration. Furthermore, the disordered region has a higher density than the area of ordered collective motion. Second, and in analogy to the results from the Vicsek model \cite{chate2008collective,bertin2009hydrodynamic,peruani2011polar,menzel2012collective},
traveling high-density bands of ordered collective motion emerge in systems of density-dependent active drive \cite{yamanaka2014formation}. The active drive in this case is set to increase with density. Such traveling density bands were shown to survive and penetrate through each other under repeated head-on collisions \cite{yamanaka2014formation}.
Apart from that, the model was complexified in two different directions. On the one hand, the next-higher mode of deformation was included \cite{hiraiwa2010dynamics}. It can be described by a third-rank tensor \cite{ohta2009deformation} that is familiar from the study of liquid crystals of tetrahedratic order \cite{fel1995tetrahedral,brand2005tetrahedratic}. This uneven mode of deformation breaks the forward-backward symmetry of the particle shape. In this case, a zig-zag trajectory and chaotic motions can be obtained already for a single particle in the absence of an external field \cite{hiraiwa2010dynamics}. The description was further extended to include the lowest four modes of deformation \cite{tarama2013oscillatory}. On the other hand, in addition to self-propulsion and deformability, an active rotational (spinning) motion was experimentally observed \cite{nagai2013rotational} and theoretically considered \cite{tarama2012spinning}.
Different kinds of circular and quasi-periodic orbital trajectories result from this modification \cite{tarama2012spinning}. Furthermore, period-doubling and chaotic behavior is observed in the trajectories and in the state variables such as in the degree of deformation \cite{tarama2012spinning}. Helical types of motion are found in three spatial dimensions, among them also one with a superhelical trajectory \cite{tarama2013dynamics}.
In a next step, the two-dimensional dynamics of a deformable self-propel\-led particle in a swirl flow \cite{tarama2014deformable} and in a planar linear shear flow were analyzed \cite{tarama2013dynamicsshear}. When exposed to shear, the elongational part of the flow tends to additionally deform the particle as depicted in Fig.~\ref{fig_deformable_elongation}.
\begin{figure}
\centerline{\includegraphics[width=8.5cm]{figure30.pdf}}
\caption{Elongation of a deformable particle by an externally imposed shear flow \cite{tarama2013dynamicsshear}. Initially spherical deformable particles like droplets (left) are deformed due to the elongational component of the flow (right). In addition to that, the rotational component of the shear flow tends to induce further dynamical processes, see Figs.~\ref{fig_lc_shear} and \ref{fig_vesicle_shear}.}
\label{fig_deformable_elongation}
\end{figure}
Moreover, in the case of the planar linear shear flow, the rotational contribution of the flow field tends to reorient the particle axes and can complexify the dynamic behavior.
For this situation, an active straight motion, a winding motion, as well as different types of cycloidal motion were predicted \cite{tarama2013dynamicsshear}. An interesting issue for future studies is to find out, whether these kinds of motion represent the active analogues to the flow-aligned or tank-treading, wagging or swinging, and tumbling modes, respectively, displayed in Figs.~\ref{fig_lc_shear} and \ref{fig_vesicle_shear}. Experimentally, the shear geometry could be realized by placing self-propelling droplets on the surface of a liquid \cite{nagai2005mode,takabatake2011spontaneous} that is confined between two parallel counterpropagating walls.
The situation gets even more complex, when an additional active rotational spinning of the particle is considered in the shear flow \cite{tarama2013dynamicsshear}. In this case, active spinning with the same sense of rotation as the externally imposed flow field must be distinguished from active spinning with the opposite sense. When the active and the externally imposed rotations counteract and balance each other, a kind of active straight trajectory is recovered. Other types of observed dynamics are periodic, quasi-periodic, different kinds of cycloidal and undulated cycloidal, different kinds of winding, as well as different kinds of chaotic motions.
To summarize the situation, we can say that the additional degrees of freedom arising from the deformability lead to quite complex dynamics, even for isolated active particles. Combining deformability, active propulsion, active rotation, and the various types of external influences -- in our example the imposed shear flow -- opens the way to a nearly unbounded number of dynamical states.
\subsection{Collective behavior of animals}
At the end of this review, we now reach the uppermost spring of our ladder of complexity in the field of active systems.
By construction this is a highly non-trivial topic and we can only touch the surface of it in the present framework. The main question that needs to be answered in this context is how much the characterization of each of the following systems can be simplified. The physicist's goal is naturally to map the animal behavior onto a minimum model that captures the central ingredients but can still be evaluated efficiently. How will this work for living creatures that have their own mind and make their own decisions?
Already for microorganisms the situation is quite difficult \cite{cates2012diffusive}. For example, reversals or changes in their migration directions are often observed \cite{wada2013bidirectional,tailleur2008statistical,berg1972chemotaxis, macnab1972gradient,polin2009chlamydomonas,min2009high}. The latter can lead to a type of run-and-tumble motion \cite{berg1972chemotaxis,macnab1972gradient,polin2009chlamydomonas,min2009high,bennett2013emergent,cates2013active}. On the one hand, this behavior may result from the complicated propulsion mechanism \cite{bennett2013emergent}. On the other hand, it can be the manifestation of an ``active decision'' \cite{yi2000robust}. If the change in migration direction occurs in response to an external stimulus, the reaction is generally referred to as ``taxis'' or ``tactic'' behavior. For instance, microorganisms can react in a chemotactic way when they detect chemicals that indicate nutrition, or they use chemicals for communication between each other without physical contact \cite{berg1972chemotaxis,macnab1972gradient,cluzel2000ultrasensitive, miller2001quorum,waters2005quorum,an2006quorum,taktikos2012collective}. Further examples concern the reaction to light (phototaxis) \cite{hader1987polarotaxis,suematsu2011localized,garcia2013light}, gravitation (gravitaxis) \cite{yoshimura2003gravitaxis,roberts2006mechanisms}, flow fields (rheotaxis) \cite{marcos2012bacterial}, adhesion gradients (haptotaxis) \cite{kolmakov2010designing,kolmakov2011designing}, and other external stimuli. Nevertheless, aspects of the collective behavior
can often be described by simple particle descriptions in the form of extended Vicsek approaches \cite{czirok1996formation,szabo2006phase} or by minimum continuum models \cite{aranson2007model,wensink2012meso,dunkel2013fluid,svensek2013collective}.
Grasshoppers (or locusts) form a class of insects the collective dynamics of which has attracted attention for at least one reason: cannibalism is observed for these insects and seems to influence their individual and collective behavior \cite{simpson2006cannibal,hansen2011cannibalism,bazazi2011nutritional}. On the one hand, an individual tries to rush after grasshoppers ahead of itself. On the other hand, it tries to flee from individuals behind itself. This was demonstrated by restricting the visual field through paintings on part of their eyes \cite{bazazi2008collective}. A restricted ability to detect approaching members in a group reduces motion and promotes cannibalism \cite{bazazi2008collective}, which indicates that cannibalism supports migration. The situation was modeled in a pursuit-escape approach by finite-size random walkers that experience friction with the substrate and social interactions \cite{romanczuk2009collective,romanczuk2012swarming}: other individuals moving away in the front are hunted after, whereas they are fled from if they approach from behind. In particular, the pursuit interaction tends to increase the orientational order in the collective motion. It increases the mean migration speed, which is also the case for the escape interaction at higher densities. A further study analyzed the radial distribution of other grasshoppers around one individual \cite{buhl2012using}. It revealed an isotropic shape in contrast to the anisotropic one obtained from minimum pursuit-escape approaches. The situation might still be more complex.
Another group of animals repeatedly studied in the context of collective motion are fish. Data can be obtained by video-tracking the individuals of a swarm in artificial water tanks \cite{hensor2005modelling,becco2006experimental,gautrais2009analyzing,herbert2011inferring, katz2011inferring,gautrais2012deciphering,tunstrom2013collective}. It was observed that fish swarms were of smaller size when food sources were detected \cite{hoare2004context} and of larger size in reaction to predator attacks \cite{hoare2004context}.
The swarms are generally elongated and denser at the front than at the back \cite{katz2011inferring,hemelrijk2012schools}. Increasing the mean density of fish in a container, a transition to cooperative collective motion could be detected at a threshold density \cite{becco2006experimental,cambui2012density} similarly to the predictions of the Vicsek model \cite{vicsek1995novel}. However, the interaction rules deduced for individuals appear to be quite different from those assumed in the Vicsek model: direct orientational alignment seems to be weak \cite{herbert2011inferring,katz2011inferring}, interactions are restricted to the one single nearest neighbor \cite{herbert2011inferring} or at least are not a simple average over all pairwise interactions \cite{katz2011inferring}, and variations of swimming speed are important \cite{herbert2011inferring,katz2011inferring}.
In a recent study, three modes of collective swimming motion were identified \cite{tunstrom2013collective}: a swarming state of relatively low order, a polarized state, and a milling state of high global rotation. Transitions were frequently observed.
These forms of collective behavior should be compared to the motion of single isolated fish. In the latter case, relatively continuous trajectories without sharp bends or kinks were observed, corresponding to persistently turning walkers rather than conventional random walks
\cite{gautrais2009analyzing,gautrais2012deciphering}.
While the swimming motion of small fish can still be effectively confined to a quasi two-dimensional geometry in the lab, the study of fish and bird swarms in nature must typically consider three-dimensional textures. An exception are linear string-like and V-shaped arrangements (skeins) of, e.g., migrating geese, the structure and dynamics of which was recorded and analyzed \cite{hayakawa2010spatiotemporal,hayakawa2012group}. To empirically study real three-dimensional flocks of birds, however, new tools had to be developed to effectively reconstruct the individual positions within the swarm from recorded two-dimensional image data \cite{cavagna2008starflag1,cavagna2008starflag2,cavagna2013diffusion}.
In contrast to the observations made for fish, starling swarms were denser at their outer boundary than in the center \cite{ballerini2008empirical,cavagna2013diffusion} without a systematic front-back asymmetry \cite{ballerini2008empirical}. A correlation between the density and number of individuals in the swarm could not be identified \cite{ballerini2008empirical}.
On average, the probability to find a nearest-neighboring bird in the direction of motion was markedly depleted compared to other directions \cite{ballerini2008empirical,ballerini2008interaction}.
Maybe most importantly, the interactions between the individuals were observed to follow topological rather than metric rules \cite{ballerini2008interaction,bialek2012statistical}: the interactions occur with a fixed number of nearest neighbors instead of all neighbors within a fixed interaction distance.
Qualitative differences arise from this subtlety, for example the swarms are more cohesive under predator attacks \cite{ballerini2008interaction}.
Furthermore, in the famous Vicsek model \cite{vicsek1995novel}, a change to non-metric topological interactions renders the order-disorder transition for collective motion continuous, without the emergence of density bands \cite{ginelli2010relevance}. This example demonstrates the importance of intensified recording and comparing to empirical data -- a task that is highly non-trivial for such complex systems as animal swarms.
\newpage
\section{Conclusions}
In the above, we took a tour through various subtopics concerning the physics of soft matter out of its equilibrium ground state. The field of soft matter has grown far too broad for all its aspects to be touched in such a brief account. Only selected issues could be included. We were guided by two central aspects characteristic for soft matter systems: they typically show large responses to external stimuli; and they are easily driven out of equilibrium. The degree of non-equilibrium served to order the different topics, where in each case we considered systems of increasing complexity.
Static external fields can switch the state of a system from its initial equilibrium ground state to a new static equilibrium state. We considered external electric fields that reorient the director in cells of low-molecular-weight liquid crystals as well as in swollen or prestretched liquid crystalline elastomers. In the latter materials, also mechanical fields can be applied to change the director orientation. Furthermore, the electric fields can induce mechanical deformations in swollen liquid crystalline elastomers, which makes these materials candidates for the use as soft actuators. The same is true for ferrogels or magnetic elastomers when they are exposed to external magnetic fields. Apart from that, the external fields can be used to tune the mechanical properties. For example, the elastic moduli can be reversibly adjusted in a non-invasive way by applying an external field.
After that, we addressed externally imposed shear flows that lead to steady and dynamic states of motion. In such shear flows, the director of nematic liquid crystals can be observed to ``flow align'', i.e.\ to take a steady inclination angle with respect to the flow within the shear plane; it can oscillate in a ``wagging'' motion; or it can describe continuous full ``tumbling'' rotations. In certain cases, these states can also be observed successively with decreasing shear rate. Orientational effects under shear are further found in bulk-filling, but spatially periodically modulated systems. We considered the example of micro-phase-separated block copolymer melts or solutions. Especially under large-amplitude oscillatory shear, reorientations of the structures with respect to the shear directions occur as a function of the shear rate and amplitude. Furthermore, for localized objects such as vesicles that are composed of closed bilayer membranes a picture similar to the one for the director of nematic liquid crystals emerges under steady shear flow. The steady state of constant inclination angle is called ``tank-treading'', with the bilayer membrane running around the interior of the vesicle; an oscillating inclination angle is referred to as ``swinging'', ``trembling'', or ``vacillating breathing''; and, again, full rotations occur in the ``tumbling'' mode. A non-vanishing viscosity contrast between the liquid on the inside and on the outside of the vesicle is necessary to obtain these states successively with decreasing shear rate.
Finally, we turned to systems composed of constituents that themselves are considered to be active. They feature a mechanism of self-propulsion and possibly also of active rotations. Our focus was on the collective non-equilibrium states arising from the interactions between these constituents. Examples are ordered collective motion, traveling density bands, laning, resting and traveling crystals, or non-equilibrium clustering. We increased the complexity from point-like particles to finitely-sized objects including steric interactions; from particles propelling with constant velocity to those featuring a constant strength of active drive; from dry systems to those interacting hydrodynamically; from rigid particles to those being deformable; and, eventually, by briefly addressing recent studies on living creatures. Quite remarkably, basic aspects, such as the formation of flocks of individuals that order their migration directions at high enough density, can be found on all levels of complexity. This supports the idea that it is possible to develop basic underlying concepts and systematic methods to describe also such kinds of genuinely non-equilibrium systems.
As a concluding remark, we would like to append a central argument in favor of this field. Although often emerging from topics of daily life or biology, many of its subareas are relatively young. The boundaries are rapidly being pushed forward, thus continuously providing new interesting problems of research. Far from equilibrium, i.e.\ inherently for active systems, the tools of understanding still need to be established, although often an intuitive picture can straightforwardly be developed. Therefore, especially for young researchers, soft matter physics provides a promising opportunity in a meanwhile well-recognized environment.
\newpage
\section{Acknowledgments}
In the first place, the author thanks Hartmut L\"owen for his continuous support.
Apart from him, the author thanks
Takeaki Araki, G\"unter Auernhammer, Stefan Bohlius, Alexander B\"oker, Dmitry Borin, Helmut Brand, Peet Cremer, Burkhard D\"unweg, Heino Finkelmann, Paul Goldbart, Nigel Goldenfeld, Steve Granick, Hisao Hayakawa, Marco Heinen, Sascha Hilgenfeldt, Christian Holm, J\"urgen Horbach, Thomas Ihle, Shoichi Kai, Andreas Kaiser, Toshihiro Kawakatsu, Ken Nagai, Stefan Odenbach, Takao Ohta, Yoshitsugu Oono, Harald Pleiner, Miha Ravnik, Eric Roeben, Lisa Roeder, Takahiro Sakaue, Masaki Sano, Annette Schmidt, Matthias Schmidt, Michael Schmie\-de\-berg, Maksim Sipos, Thomas Speck, Torsten St\"uhn, Daniel Sven\v{s}ek, Yuka Tabe, Kazumasa Take\-uchi, Mitsusuke Tarama, Nariya Uchida, Kenji Urayama, Axel Voigt, Hirofumi Wada, Bo Wang, Rudolf Weeber, Rik Wensink, Fangfu Ye, Natsuhiko Yoshinaga, and many others
for scientific discussions and interactions.
Over the past years, support leading to the above presentation was received from the Deutsche Forschungsgemeinschaft through the FOR 608 ``Nonlinear dynamics of complex continua'', a DFG research fellowship, the Japanese-German core-to-core collaboration program ``Non-equilibrium phenomena in Soft Matter'', the SPP 1681 ``Feldgesteuerte Partikel-Matrix-Wechselwirkungen: Erzeugung, skalen\"ubergreifende Mo\-del\-lie\-rung und Anwendung magnetischer Hybridmaterialien'', and the SPP 1726 ``Microswimmers -- from single particle motion to collective behavior''.
\newpage
\section{References}
|
\section{Introduction}
Suppose $\tilde{\pi}(\mathbf{x})$ is a target density kernel where
$\mathbf{x}\in\mathbb{R}^{d}$ so that $\pi(\mathbf{x})=\tilde{\pi}(\mathbf{x})/\int\tilde{\pi}(\mathbf{x})d\mathbf{x}$
is a probability density function. Bayesian analysis of many statistical
models necessitates the calculation of integrals with respect to $\pi(\mathbf{x})$
when $\pi(\mathbf{x})$ is analytically intractable and only the unnormalized
density kernel $\tilde{\pi}(\mathbf{x})$ is available for evaluation
\citep[see e.g.][]{GelmanBDA3}. To approximate such integrals, Markov
Chain Monte Carlo (MCMC) has seen widespread use, and the development
of better MCMC algorithms that can help researchers tackle even larger
and more challenging models is still a highly active field \citep[see e.g.][]{RSSB:RSSB736,mcmc_handbook,girolami_calderhead_11}.
MCMC relies on constructing a discrete time Markov process $\left\{ \mathbf{x}_{t}\right\} $
that has $\pi(\mathbf{x})$ as its stationary distribution. Consequently,
an ergodic theorem ensures that, under some technical conditions,
integrals on the form $\int g(\mathbf{x})\pi(\mathbf{x})d\mathbf{x}$
can approximated as $\hat{\mu}_{T}(g)=\frac{1}{T}\sum_{t=1}^{T}g(\mathbf{x}_{t})$
for some large $T$ \citep[e.g.][chapter 6]{robert_casella}. A very
general way of constructing such a Markov process, and that serves
as the foundation of most MCMC methods, is the Metropolis Hastings
(MH) algorithm \citep{Metropolis53,Hastings70}. Suppose the current
state of the Markov process is $\mathbf{x}_{t}$ and that $q(\cdot|\mathbf{x}_{t})$
is some proposal probability density function. Then the MH algorithm
proceeds by proposing $\mathbf{x}^{*}\sim q(\cdot|\mathbf{x}_{t})$,
and accepting the new state $\mathbf{x}_{t+1}$ to be equal to $\mathbf{x}^{*}$
with probability $\alpha(\mathbf{x}_{t},\mathbf{x}^{*})=\min\left[1,\tilde{\pi}(\mathbf{x}^{*})q(\mathbf{x}_{t}|\mathbf{x}^{*})/(\tilde{\pi}(\mathbf{x}_{t})q(\mathbf{x}^{*}|\mathbf{x}_{t}))\right]$.
With remaining probability $1-\alpha(\mathbf{x}_{t},\mathbf{x}^{*})$,
$\mathbf{x}_{t+1}$ is set equal to $\mathbf{x}_{t}$. The MH algorithm
is extremely general as only minimal restrictions on the proposal
density are necessary for $\left\{ \mathbf{x}_{t}\right\} $ to satisfy
detailed balance with $\pi(\mathbf{x})$ as the stationary distribution
\citep{robert_casella}.
These minimal restrictions have lead to a large number of strategies
for choosing proposal distributions \citep[see e.g.][]{liu_mc_2001,robert_casella}.
The most widespread strategy is the Gaussian random walk proposal,
corresponding to $q(\cdot|\mathbf{x}_{t})=\mathcal{N}(\cdot|\mathbf{x}_{t},\Sigma)$
where $\mathcal{N}(\mathbf{x}|\mu,\Sigma)$ denotes the $N(\mu,\Sigma)$
density evaluated at $\mathbf{x}$. The covariance matrix $\Sigma$
should be chosen according to the target distribution at hand. As
a rule of thumb, a scale of $q(\cdot|\mathbf{x}_{t})$ that is ``large''
relative to the scale of the target density leads to a low acceptance
probability, and thereby a slow exploration of target distribution.
At the other end of the spectrum, if the scale of $q(\cdot|\mathbf{x}_{t})$
is ``small'' relative to the scale of the target, one typically
obtain a high acceptance probability but the cumulative distance traversed
by many accepted steps will still be small. Provided that the variance
of $\hat{\mu}_{T}(g)$ exists, both extreme cases lead to the variance
of $\hat{\mu}_{T}(g)$ being large for fixed $T$, whereas intermediate
choices of scale typically lead to more efficient sampling in the
sense that the variance of $\hat{\mu}_{T}(g)$ is smaller for fixed
$T$ than the variance at the extreme cases. It should be noted that
``large'' and ``small'' scales relative to the target are not
absolute or even well-defined magnitudes, and depend among many other
aspects, on the dimension $d$ of the target \citep[see e.g.][]{Gelman95,roberts2001}.
An additional complication is that many target distributions of interest,
e.g. joint posterior distributions of latent variables and variance
parameter of the latent variables, contain significantly different
scaling properties in different regions of the support of the target.
Consequently, there is often a need for proposal distributions to
adapt to local properties of the target distribution to achieve efficient
sampling.
Many researchers have explored MCMC methods that exploit local derivative
information from the target distribution to obtain more efficient
sampling. Early contributions in this strand of the literature includes
the Metropolis adjusted Langevin algorithm (MALA) \citep[see e.g.][]{roberts1996,roberts_stramer_02}.
MALA is a special case of the MH algorithm where the proposal distribution
is equal to a time-discretization of a Langevin diffusion, where the
Langevin diffusion has the target distribution as its stationary distribution.
The drift term in such diffusion, and consequently the mean of the
proposal distribution, will depend on the gradient of $\log\tilde{\pi}(\mathbf{x})$.
To achieve reasonable acceptance rates, it is often necessary to choose
the (time-) step size of the proposal distribution rather small, and
it has been argued that the MALA in this case will revert to a behavior
similar to that of the random walk MH \citep{1206.1901}.
Another way of constructing MCMC samplers that exploit derivative
information is Hamiltonian (or Hybrid) Monte Carlo (HMC) \citep{Duane1987216,liu_mc_2001,1206.1901,beskos2013,1411.6669,JMLR:v15:hoffman14a}.
HMC has the potential of producing proposals that are far from the
current state, while retaining arbitrarily high acceptance rates.
The proposals originate from numerically integrating a set of ordinary
differential equations that involve the gradient of $\log\tilde{\pi}(\mathbf{x})$
and it is often necessary to evaluate this gradient hundreds of times
per proposed state to achieve reasonable acceptance rates while retaining
proposals that are far from the current state.
Most implementations of MALA and HMC involve a user-specified scaling
matrix (in addition to other tuning parameters) to achieve reasonable
efficiency. In a landmark paper, \citet{girolami_calderhead_11} proposes
to recast MALA and HMC on a Riemann manifold that respects the local
scaling properties of the target, and consequently alleviates the
need to choose a global scaling matrix. Suppose $\mathbf{x}$ are
the parameters to be sampled and $\mathbf{y}$ the observations associated
with a Bayesian statistical model that admit the explicit factorization
into data likelihood $p(\mathbf{y}|\mathbf{x})$ and prior distribution
$p(\mathbf{x})$, i.e. $\tilde{\pi}(\mathbf{x})\propto p(\mathbf{y}|\mathbf{x})p(\mathbf{x})$.
Then \citet{girolami_calderhead_11} took their metric tensor to be
the matrix
\begin{equation}
G_{GC}(\mathbf{x})=-E_{p(\mathbf{y}|\mathbf{x})}\left[\nabla_{\mathbf{x}}^{2}\log p(\mathbf{y}|\mathbf{x})\right]-\nabla_{\mathbf{x}}^{2}\log p(\mathbf{x}),\label{eq:GC-tensor}
\end{equation}
namely the Fisher information matrix associated with data likelihood
plus the negative Hessian of the log-prior. This choice was demonstrated
to be highly suited for Langevin and Hamiltonian based algorithms
that take local scaling properties into account. However, due to the
expectation in (\ref{eq:GC-tensor}), the Fisher information matrix
is rarely available in closed form, and the potential for automating
the implementation of MCMC samplers based on such a metric seems a
formidable task.
In this paper, an alternative approach based on deriving a metric
tensor for simplified manifold MALA directly from the negative Hessian
of $\log\tilde{\pi}(\mathbf{x})$ is explored. As noted by \citet{girolami_calderhead_11}
and several of the discussants of that paper \citep[see e.g.][]{disc_sanzserna,disc_NTNU,disc_Jasra},
this approach admit a high degree of automation via e.g. the usage
of automatic differentiation software \citep{grie:2000}, but is complicated
by the fact that the negative Hessian is typically not positive definite
over the complete support of the target. Obtaining positive definite
local scaling matrices from potentially indefinite Hessian matrices
for application in modified Newton methods has seen substantial treatment
in non-convex numerical optimization literature \citep[see section 6.3 of][ for an overview]{noce:wrig:1999}.
Such approaches typically rely on either modification of the eigenvalues
via a full spectral decomposition of the Hessian, or some form of
modified Cholesky algorithm that produces a factorization of a symmetric
positive definite matrix that is close to the Hessian in some metric.
In the MCMC context, the former approach was taken by \citet{martin_etal2012,1212.4693},
whereas in the present paper I advocate the latter as it has potential
to exploit sparsity of the Hessian.
The added generality with respect to target densities that can be
handled by Hessian-based metric tensors (based either on the spectral
decomposition or modified Cholesky factorizations) relative to (\ref{eq:GC-tensor})
comes at the cost of potential pathological behavior in regions of
the log-target with near zero curvature in some direction. To counteract
such effects, the primary contribution of this paper is a new adaptive
step size procedure for simplified manifold MALA that chooses the
step size locally based how well the Hessian-based metric tensor reflects
the local scaling of the target. The adaptive step size procedure
remains active throughout the MCMC simulation and at the same time
retains the target as the stationary distribution. The adaptive step
size procedure exploits the well known fact that a MALA update is
equivalent to a particular HMC method using one time integration step
\citep{1206.1901}, and therefore admits selection of step sizes for
simplified manifold MALA based on the (dimensionless) energy error
of the associated Hamiltonian system. An additional pro of the adaptive
step size procedure is that it appears to alleviate the need for different
tuning parameters in transient and stationary regimes that is typical
for standard MALA implementations \citep{RSSB:RSSB500}.
The remaining of the paper is laid out as follows. Section 2 describes
the applied metric tensor and the adaptive step size selection, and
illustrates the proposed methodology using a pilot example. Section
3 illustrates the proposed methodology on two realistic example models,
and compares the proposed methodology to alternative MCMC methods.
Finally section 4 provides some discussion.
\section{Adaptive step size modified Hessian MALA}
Assume that $\tilde{\pi}(\mathbf{x})$ is sufficiently smooth to admit
continuous and bounded derivatives up to order 2. I denote the gradient
of the log-kernel $g(\mathbf{x})=\nabla_{\mathbf{x}}\log\tilde{\pi}(\mathbf{x})$
and the Hessian of the log-kernel as $H(\mathbf{x})=\nabla_{\mathbf{x}}^{2}\log\tilde{\pi}(\mathbf{x})$.
The $d\times d$ identity matrix is denoted $I_{d}$.
I take as vantage point the simplified manifold MALA (sMMALA) or position
specific preconditioned MALA of \citet{girolami_calderhead_11} \citep[see also][]{e16063074,Xifara201414},
which is a MH method characterized by the proposal distribution
\begin{equation}
q(\cdot|\mathbf{x})=\mathcal{N}\left(\cdot|\mathbf{x}+\frac{\varepsilon^{2}}{2}\left[G(\mathbf{x})\right]^{-1}g(\mathbf{x})\;,\;\varepsilon^{2}\left[G(\mathbf{x})\right]^{-1}\right).\label{eq:sMMALA_prop_distr}
\end{equation}
Here $G(\mathbf{x})$ is a $d\times d$ symmetric and positive definite
matrix, from now on referred to as the metric tensor, and $\varepsilon$
is a tunable step size.
\subsection{Modified Cholesky based metric tensor\label{sub:The-GMW-factorization}}
This section explains a metric tensor based on an adaptation of the
modified Cholesky decomposition of \citet{gill_murray_1974}, \citet{gilletal},
from now on referred to as GMW. Suppose $A\in\mathbb{R}^{d\times d}$
is a symmetric, not necessarily positive definite matrix. Given a
user-specified small positive scale parameter $u$, our adaptation
of GMW (details and algorithm in Appendix \ref{sec:A-version-of-chol})
aims at finding a Cholesky decomposition $LL^{T}$ where $L$ is lower
triangular so that
\begin{equation}
\hat{A}=LL^{T}=A+J,\text{ where }L_{i,i}^{2}\geq u\max\left(1,\max_{1\leq i,j\leq d}|A_{i,j}|\right)>0,\; i=1,\dots,d.\label{eq:Def_LLchol}
\end{equation}
The matrix $J$ is a diagonal matrix with non-negative elements chosen
to ensure that $\hat{A}$ is positive definite, and in particular
the design of the GMW Cholesky algorithm ensures that if $A$ is sufficiently
positive definite, then $J=0$. As $J$ is a diagonal matrix and assuming
that $A$ has non-zero diagonal elements, the sparsity structures
of $A$ and $A+J$ are identical. This implies that whenever $A$
has a sparsity structure that can be exploited using sparse numerical
Cholesky factorizations \citep{davis_sparse}, the same applies for
the GMW factorization.
The GMW factorization has seen widespread use within the numerical
optimization literature as a key ingredient in practical Newton optimization
algorithms for non-convex optimization problems \citep{noce:wrig:1999},
and has also served as the basis for further refinements as in \citet{doi:10.1137/0911064,doi:10.1137/S105262349833266X}.
For simplicity, I employ a variant of the algorithm given in \citet{gilletal}.
In the remainder of this paper I work with the modified Hessian metric
tensor
\[
G_{MC}(\mathbf{x})=L(\mathbf{x})L(\mathbf{x})^{T},
\]
where $L(\mathbf{x})$ is the output from the GMW Cholesky factorization
(\ref{eq:Def_LLchol}) applied to $-H(\mathbf{x})$. In itself, this
approach for finding a metric tensor is not fail-safe. To see this,
consider for instance the $d=1$ case, for which the proposed metric
tensor reduces to
\begin{equation}
G_{MC}(\mathbf{x})=\max(u,|H(\mathbf{x})|).\label{eq:1D-metric}
\end{equation}
\begin{figure}
\centering{}\includegraphics[scale=0.8]{fixed_trace-crop}\protect\caption{\label{fig:fixed-trace-t}50,000 iterations of sMMALA with metric
tensor (\ref{eq:1D-metric}), $u=0.001$ and $\varepsilon=0.75$ applied
to a $t$-distribution with 4 degrees of freedom. It is seen that
the chain rarely enters the regions near $|\mathbf{x}|$=2, and when
it does, it tends to get stuck. An example of this is seen starting
in iteration 21347 where the chain is stuck for 340 iterations at
approximately 1.9529.}
\end{figure}
Typically I take $u$ to be rather small in order not to disturb the
proposal mechanism of sMMALA (\ref{eq:sMMALA_prop_distr}) in regions
where $|H(\mathbf{x})|$ is small but still contains useful information
on the local scaling properties of the target distribution. However
if left unmodified, a sMMALA method based on (\ref{eq:1D-metric})
would very infrequently move into regions near roots of the second
derivative (inflection points). To see this, observe that the backward
transition density $q(\mathbf{x}_{t}|\mathbf{x}^{\text{*}})$ occurring
in the nominator of the acceptance probability will be
\begin{equation}
q(\mathbf{x}_{t}|\mathbf{x}^{\text{*}})=O(\sqrt{u}/\varepsilon)\text{ when }G(\mathbf{x}^{\text{*}})=u.\label{eq:alpha_order}
\end{equation}
When the method has moved into such a region, the chain will tend
to be stuck for many iterations as the proposal distribution then
has too large variance.
To illustrate this phenomenon, consider a $t$-distributed target
with 4 degrees of freedom. In this case there are two inflection points
of the log-density located at $|\mathbf{x}|=2$. Figure \ref{fig:fixed-trace-t}
shows 50,000 MCMC iterations using sMMALA based on (\ref{eq:1D-metric})
for this target. It is seen that the chain rarely enters the regions
close to the inflection points, and when it does, the chain tend to
get stuck. It is worth noticing that these effects are not unique
to the modified Cholesky approach. Specifically, (\ref{eq:1D-metric})
would also be the result if the metric tensor was computed by modifying
the eigenvalues of a full spectral decomposition as in \citet{martin_etal2012},
and the metric tensor would be a soft/regularized absolute value applied
to $H(\mathbf{x})$ in the methodology of \citet{1212.4693}. Rather
the effects are the cost of the added generality associated with using
second order derivative information directly rather than via the Fisher
information matrix as in \citet{girolami_calderhead_11}. To mitigate
these undesirable effects, (\ref{eq:alpha_order}) suggest either
that $u$ or $\varepsilon$ must be chosen adaptively. In this work,
I focus on the latter as a simple and computationally efficient method
can be implemented for this purpose.
\subsection{sMMALA with randomized adaptive selection of $\varepsilon$\label{sub:sMMALA-with-randomized}}
In this work, I propose to select the proposal step size as a function
of the current state $\mathbf{x}_{t}$ and an easily simulated auxiliary
random variable $\mathbf{w}_{t}\sim\pi(\mathbf{w})$, namely $\varepsilon=\varepsilon(\mathbf{x}_{t},\mathbf{w}_{t})$.
The role of $\mathbf{w}$ is essentially to randomize the step size,
and a further motivation for including it will be clear from section
\ref{sub:Line-search-mechanisms} when the functional forms of $\varepsilon(\cdot,\cdot)$
advocated here are discussed. Before that, I introduce the an overarching
sampling algorithm in the form of a Metropolis within Gibbs sampler
based on the sMMALA update where invariant distribution will be shown
to be the augmented target
\begin{equation}
\pi(\mathbf{x},\mathbf{w})=\pi(\mathbf{x})\pi(\mathbf{w}).\label{eq:augmented_target}
\end{equation}
Notice that the algorithm with target (\ref{eq:augmented_target})
is primarily a construct to make the method time-homogenous and thereby
enabling easy establishment of convergence results when an adaptive
step size selection $\varepsilon(\mathbf{x},\mathbf{w})$ is employed.
In practice only the $\mathbf{x}$-subchain is of interest.
Given the current configuration $(\mathbf{x}_{t},\mathbf{w}_{t})\sim\pi(\mathbf{x},\mathbf{w})$,
one step of the proposed Metropolis within Gibbs sampler \citep[see e.g.][Algorithm A.43]{robert_casella}
based on sMMALA for the augmented target may then be summarized by
the steps
\begin{enumerate}
\item Compute forward step size $\varepsilon_{f}=\varepsilon(\mathbf{x}_{t},\mathbf{w}_{t})$.
\item Draw proposal $\mathbf{x}_{t+1}^{*}\sim N(\mathbf{x}_{t}+\frac{\varepsilon_{f}^{2}}{2}G^{-1}(\mathbf{x}_{t})g(\mathbf{x}_{t}),\varepsilon_{f}^{2}G^{-1}(\mathbf{x}_{t}))$.
\item Compute backward step size $\varepsilon_{b}=\varepsilon(\mathbf{x}_{t+1}^{*},\mathbf{w}_{t}).$
\item Compute MH accept probability
\[
\alpha(\mathbf{x}_{t},\mathbf{x}_{t+1}^{*})=\min\left(1,\frac{\tilde{\pi}(\mathbf{x}_{t+1}^{*})\mathcal{N}\left(\mathbf{x}_{t}|\mathbf{x}_{t+1}^{*}+\frac{\varepsilon_{b}^{2}}{2}G^{-1}(\mathbf{x}_{t+1}^{*})g(\mathbf{x}_{t+1}^{*}),\varepsilon_{b}^{2}G^{-1}(\mathbf{x}_{t+1}^{*})\right)}{\tilde{\pi}(\mathbf{x}_{t})\mathcal{N}\left(\mathbf{x}_{t+1}^{*}|\mathbf{x}_{t}+\frac{\varepsilon_{f}^{2}}{2}G^{-1}(\mathbf{x}_{t})g(\mathbf{x}_{t}),\varepsilon_{f}^{2}G^{-1}(\mathbf{x}_{t})\right)}\right).
\]
Let $\mathbf{x}_{t+1}=\mathbf{x}_{t+1}^{*}$ with probability $\alpha(\mathbf{x}_{t},\mathbf{x}_{t+1}^{*})$
and let $\mathbf{x}_{t+1}=\mathbf{x}_{t}$ with remaining probability.
\item Draw $\mathbf{w}_{t+1}\sim\pi(\mathbf{w})$.
\end{enumerate}
Notice in particular that $\varepsilon_{f},\varepsilon_{b}$ are computed
using the same $\mathbf{w}$-argument and therefore steps 1-4 constitute
a reversible MH step for each $\mathbf{w}$. Based on this observation,
I prove following result: \\
\textbf{Proposition 1: }\emph{Provided that $0<\varepsilon(\mathbf{x},\mathbf{w})<\infty$
for each $(\mathbf{x},\mathbf{w})$ in the support of $\pi(\mathbf{x},\mathbf{w})$,
the Metropolis within Gibbs sampler in steps 1-5 is $\pi(\mathbf{x},\mathbf{w})$-irreducible,
aperiodic and has $\pi(\mathbf{x},\mathbf{w})$ as invariant distribution.}
The proof is given in Appendix \ref{sec:Proof-of-Proposition} and
relies on the fact that for a Metropolis within Gibbs sampler, the
parameters of the proposal distribution in the MH step for updating
one block ($\mathbf{x}$) may depend on the current state of the remaining
blocks ($\mathbf{w}$) \citep[see e.g.][]{gilks_adaptive_dir,robert_casella,zhang11quasi}.
Note in particular that steps 1-5 do not fulfill detailed balance
for target $\pi(\mathbf{x},\mathbf{w})$. On the other hand, a (time-inhomogenous
due to changing $\mathbf{w}$) detailed balance holds for each transition
of the $\mathbf{x}$-subchain. Having established that such adaptive
selection of step size does not disturb the invariant distribution
of the $\mathbf{x}$-subchain for a large class of functions $\varepsilon(\cdot,\cdot)$,
I now consider the specific functional forms of $\varepsilon(\cdot,\cdot)$
I advocate.
\subsection{Adaptive step size selection based on Hamiltonian dynamics\label{sub:Line-search-mechanisms}}
The method for selection of $\varepsilon$ uses a well-known duality
between MALA and HMC, namely that the proposal of MALA corresponds
to a single time-integration step of Euclidian metric HMC starting
at $\mathbf{x}_{t}$ when the leap frog integrator \citep{Leimkuhler:2004}
is employed \citep{1206.1901}. Here I propose to choose the step
size for sMMALA so that the energy error of a single trial step of
the corresponding (Euclidian metric) HMC method with mass matrix $G(\mathbf{x}_{t})$
is below a tunable threshold. The rationale for such an approach is
that the absolute energy error provides a measure of time integration
error that is comparable across different areas of the target density,
and that will be large if the local scaling properties of $\log\pi(\mathbf{x})$
are poorly reflected by $G(\mathbf{x})$, as in the student $t$-pilot
example discussed above. Computing the absolute energy error of the
trial step is inexpensive relative to e.g. doing complete trial sMMALA
steps for different values of $\varepsilon$, with the computational
cost typically dominated by a few additional gradient evaluations.
To implement such energy error based selection $\mathbf{\varepsilon}(\cdot,\cdot)$,
I first take $\pi(\mathbf{w})$ to be a $d$-dimensional standard
Gaussian variable. Given the current configuration of $(\mathbf{x}_{t},\mathbf{w}_{t})$,
I define the position specific dummy Hamiltonian
\begin{equation}
\mathcal{H}(\mathbf{q}(\tau),\mathbf{p}(\tau)|\mathbf{x}_{t})=-\log\tilde{\pi}(\mathbf{q}(\tau))+\frac{1}{2}\mathbf{p(\tau)}^{T}G(\mathbf{x}_{t})^{-1}\mathbf{p}(\tau),\label{eq:dummy_hamiltonian}
\end{equation}
where $\mathbf{q}(\tau)$ denotes position and $\mathbf{p}(\tau)$
denotes momentum at fictional time $\tau$. Then a single leap frog
step of (time-) size $\varepsilon$ starting from $\mathbf{q}(0)=\mathbf{x}_{t}$
and $\mathbf{p}(0)=L(\mathbf{x}_{t})\mathbf{w}_{t}$ is performed:
\begin{eqnarray}
\mathbf{p}(\varepsilon/2) & = & \mathbf{p}(0)+\frac{\varepsilon}{2}g(\mathbf{q}(0))=L(\mathbf{x}_{t})\mathbf{w}_{t}+\frac{\varepsilon}{2}g(\mathbf{x}_{t}),\nonumber \\
\mathbf{q(\varepsilon)} & = & \mathbf{q}(0)+\varepsilon G(\mathbf{x}_{t})^{-1}\mathbf{p}(\varepsilon/2)=\mathbf{x}_{t}+\frac{\varepsilon^{2}}{2}G^{-1}(\mathbf{x}_{t})g(\mathbf{x}_{t})+\varepsilon L(\mathbf{x}_{t})^{-T}\mathbf{w}_{t}=\mathbf{x}^{*}(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t}),\label{eq:proposal_eq}\\
\mathbf{p}(\varepsilon) & = & \mathbf{p}(\varepsilon/2)+\frac{\varepsilon}{2}g(\mathbf{q}(\varepsilon))=L(\mathbf{x}_{t})\mathbf{w}_{t}+\frac{\varepsilon}{2}\left(g(\mathbf{x}_{t})+g(\mathbf{x}^{*}(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t}))\right).\nonumber
\end{eqnarray}
The trial proposal $\mathbf{x}^{*}(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t})$
would occur if the standard normal vector $\mathbf{w}_{t}$ was used
to generate a proposal from (\ref{eq:sMMALA_prop_distr}) with current
state equal to $\mathbf{x}_{t}$ and time step size $\varepsilon$.
The energy error associated with this trial time integration step
is given as
\begin{eqnarray*}
\Delta(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t}) & = & \mathcal{H}(\text{\ensuremath{\mathbf{q}}}(0),\mathbf{p}(0)|\text{\ensuremath{\mathbf{x}}}_{t})-\mathcal{H}(\mathbf{q}(\varepsilon),\mathbf{p(\varepsilon)}|\mathbf{x}_{t})\\
& = & -\log\tilde{\pi}(\mathbf{x}_{t})+\log\tilde{\pi}(\mathbf{x}^{*}(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t}))\\
& & -\frac{\varepsilon}{2}\mathbf{w}_{t}^{T}\mathbf{r}-\frac{\varepsilon^{2}}{8}\mathbf{r}^{T}\mathbf{r}
\end{eqnarray*}
where $\mathbf{r}=L(\mathbf{x}_{t})^{-1}\left(g(\mathbf{x}_{t})+g(\mathbf{x}^{*}(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t}))\right)$.
Based on the expression for $\Delta(\varepsilon|\mathbf{x}_{t},\mathbf{w}_{t})$,
a possible method for choosing $\varepsilon=\varepsilon(\mathbf{x},\mathbf{w}$)
would involve the following steps:
Suppose $\gamma>0$ is the tunable maximal allowable trial energy
error, with lower values of $\gamma$ corresponding to higher acceptance
rates and smaller $\varepsilon$. Let $0<\bar{\varepsilon}<\infty$
be the maximal step size and let $\varepsilon_{0}$ denote the smallest
positive root in $\varepsilon$ of $|\Delta(\varepsilon|\mathbf{x},\mathbf{w})|=\gamma$.
An idealized candidate for choosing $\varepsilon$ is then $\varepsilon(\mathbf{x},\mathbf{w})=\min(\bar{\varepsilon},\varepsilon_{0})$.
The $\varepsilon>0$ bound required in Proposition 1 is automatically
fulfilled as consequence smoothness assumptions on $\log\tilde{\pi}(\mathbf{x})$,
and thus such a step size function would lead the sampling algorithm
outlined in section \ref{sub:sMMALA-with-randomized} to have the
correct invariant distribution. In practice, locating $\varepsilon_{0}$
would amount to solving an equation that involves the gradient of
the log-target numerically, and for computational effectivity reasons
I therefore propose a method for only approximating $\varepsilon_{0}$
in Section \ref{sub:Practical-implementation-and}.
It is worth noticing that the Hamiltonian (\ref{eq:dummy_hamiltonian})
should be interpreted in the Euclidian metric sense as a tool to measure
the time integration error around $\mathbf{q}(0)=\mathbf{x}_{t}$,
$\mathbf{p}(0)=L(\mathbf{x}_{t})\mathbf{w}_{t}$ when $G(\mathbf{x}_{t})$
is the fixed mass matrix. This is relevant for sMMALA as the distribution
of proposed $\mathbf{q}(\varepsilon)$-states is (\ref{eq:sMMALA_prop_distr})
in this case. Moreover, it is in contrast to interpreting (\ref{eq:dummy_hamiltonian})
in a Riemann manifold HMC sense with Hamiltonian say $-\log\tilde{\pi}(\mathbf{q})+\frac{1}{2}\mathbf{p}^{T}G(\mathbf{q})^{-1}\mathbf{p}$
(which is off target due to a missing $\frac{1}{2}\log|G(\mathbf{q})|$
term \citep{girolami_calderhead_11,1212.4693}). Under the latter
interpretation, the energy error is irrelevant for sMMALA as the proposed
$\mathbf{q}(\varepsilon)$ will not be distributed according to (\ref{eq:sMMALA_prop_distr})
regardless of the numerical integrator being employed, as the time
evolution of $\mathbf{p}$ will involve the gradient of $\frac{1}{2}\mathbf{p}^{T}G(\mathbf{q})^{-1}\mathbf{p}$
with respect to $\mathbf{q}$. In the same vein, I notice that $\min(1,\exp(\Delta))$
cannot be interpreted as a trial acceptance probability for the sMMALA
as the sMMALA deviates from being a reversible and volume-preserving
discretization of either Hamiltonian when the $\mathbf{x}_{t}$-dependence
of the mass matrix is taken into account in (\ref{eq:dummy_hamiltonian}).
Rather, $\gamma$ should be interpreted as a tolerance of local (irreversible)
time integration energy error, and in particular its relationship
to the acceptance probability of the overall (reversible) sMMALA step
is non-linear.
The role of $\mathbf{w}$ is to act as typical (standardized) momentum
variable that changes in each iteration of the overarching MCMC procedure
to reflect the variation in the distribution of the proposal (\ref{eq:sMMALA_prop_distr}).
A further potential candidate would be to integrate out $\mathbf{w}$,
e.g. $\varepsilon(\mathbf{x})=\int\varepsilon(\mathbf{x},\mathbf{w})\pi(\mathbf{w})d\mathbf{w}$
so that steps 1-4 of section \ref{sub:sMMALA-with-randomized} would
constitute a time-homogenous and reversible MH step. However even
for moderate dimension $d$ this becomes computationally intractable,
as this integral has to be computed numerically. Further, it is not
obvious that a deterministic (conditional on $\mathbf{x}$) step size
$\varepsilon(\mathbf{x})$ has clear advantages over a stochastic
step size $\varepsilon(\mathbf{x},\mathbf{w})$.
\begin{figure}
\centering{}\includegraphics[scale=0.8]{dynamic_trace-crop}\protect\caption{\label{fig:dyn_trace-t}50,000 iterations of sMMALA with metric tensor
(\ref{eq:1D-metric}), $u=0.001$, $\varepsilon(\mathbf{x},\mathbf{w})=\min(1,\varepsilon_{0})$
as defined in the text and $\gamma=1.0$ applied to a $t$-distribution
with 4 degrees of freedom. It is seen that the adaptive selection
of step size resolves sampling in the problematic regions around the
inflection points.}
\end{figure}
\begin{figure}
\centering{}\includegraphics[scale=0.5]{dynamic_epsilon-crop}\protect\caption{\label{fig:Values-of-the}Values of the (forward) step size $\varepsilon(\mathbf{x}_{t},\mathbf{w}_{t})=\min(1,\varepsilon_{0})$
as a function of $\mathbf{x}_{t}$ for the simulation presented in
Figure \ref{fig:dyn_trace-t} are given as dots. The solid line indicates
the average adaptive step size calculated as $\int\varepsilon(\mathbf{x},\mathbf{w})p(\mathbf{w})d\mathbf{w}$
as a function of $\mathbf{x}$.}
\end{figure}
To illustrate how the adaptive step size selection $\varepsilon(\mathbf{x},\mathbf{w})=\min(1,\varepsilon_{0})$
resolves the shortcomings of sMMALA near inflection points, I consider
again the $t_{4}-$distribution target considered in section \ref{sub:The-GMW-factorization}
with metric tensor (\ref{eq:1D-metric}). Figure \ref{fig:dyn_trace-t}
presents 50,000 MCMC iterations using adaptive step size sMMALA using
a maximal absolute energy error $\gamma=1.0$. It is seen that the
adaptive step size selection resolves the inefficient sampling near
the inflection points seen for the fixed step size implementation
in Figure \ref{fig:fixed-trace-t}. Figure \ref{fig:Values-of-the}
shows the actual values of the (forward) adaptive step sizes $\varepsilon(\mathbf{x}_{t},\mathbf{w}_{t})$
as a function of $\mathbf{x}_{t}$, along with their expected value
found by integrating out $\mathbf{w}$. It is seen that the energy
error criterion appear to discriminate very well that the metric tensor
(\ref{eq:1D-metric}) shows pathological behavior around the inflection
points at $|\mathbf{x}|=2.$
As pointed out by a reviewer, the above energy error argument can
also be applied in non-adaptive step size sMMALA methods to locate
regions of the support where deficiencies in the metric tensor lead
to inefficient sampling analogous to what is seen in Figure \ref{fig:fixed-trace-t}.
This is in particularly of interest for $d\geq3$, as then regions
with inefficient sampling could be difficult to spot using e.g. trace
or scatter plots. In practice, using energy error for such an end
would amount to calculating the energy error associated with each
proposed state. Specifically, suppose $\mathbf{x}_{t}$ is the current
state of a non-adaptive sMMALA method and $\mathbf{z}\sim N(0,I_{d})$
is the standardized variable used to generate a proposal $\mathbf{x}^{*}$
according to (\ref{eq:sMMALA_prop_distr}). The forward (backward)
energy error $\Delta_{t}^{f}$ ($\Delta_{t}^{b}$) associated with
the proposal $\mathbf{x}^{*}$ are given by
\begin{eqnarray*}
\Delta_{t}^{f} & = & -\log\tilde{\pi}(\mathbf{x}_{t})+\log\tilde{\pi}(\mathbf{x}^{*})-\frac{\varepsilon}{2}\mathbf{z}^{T}\mathbf{r}_{f}-\frac{\varepsilon^{2}}{8}\mathbf{r}_{f}^{T}\mathbf{r}_{f}\\
\Delta_{t}^{b} & = & -\log\tilde{\pi}(\mathbf{x}^{*})+\log\tilde{\pi}(\mathbf{x}_{t})-\frac{\varepsilon}{2}\mathbf{s}^{T}\mathbf{r}_{b}-\frac{\varepsilon^{2}}{8}\mathbf{r}_{b}^{T}\mathbf{r}_{b}\\
\mathbf{s} & = & \varepsilon^{-1}L(\mathbf{x}^{*})^{T}\left[\mathbf{x}_{t}-\mathbf{x}^{*}-\frac{\varepsilon^{2}}{2}G^{-1}(\mathbf{x}^{*})g(\mathbf{x}^{*})\right]
\end{eqnarray*}
where $\mathbf{r}_{f}=L(\mathbf{x}_{t})^{-1}\left(g(\mathbf{x}_{t})+g(\mathbf{x}^{*})\right)$
and $\mathbf{r}_{b}=L(\mathbf{x}^{*})^{-1}\left(g(\mathbf{x}_{t})+g(\mathbf{x}^{*})\right)$.
A large $|\Delta_{t}^{f}|$ ($|\Delta_{t}^{b}|$) (e.g. $|\Delta_{t}^{k}|>5,\; k=f,b$)
will indicate that local scaling properties in the region around $\mathbf{x}_{t}$
($\mathbf{x}^{*}$) are poorly reflected by $G(\mathbf{x}_{t})$ ($G(\mathbf{x}^{*})$)
and the user should consider revising the specification of $G$ or
at least reduce $\varepsilon$. Note in particular that $\Delta_{t}^{k},\; k=f,b$
can be computed in each iteration with minimal additional cost as
the involved log-kernels, gradients and metric tensors must be computed
anyway to form the acceptance probability for proposal $\mathbf{x}^{*}$.
\subsection{Practical implementation and choosing $\gamma$\label{sub:Practical-implementation-and}}
\begin{algorithm}
\begin{ntabbing}
\hspace{2em} \= \hspace{2em} \= \hspace{2em} \= \\
Set $\varepsilon_1\leftarrow 1$ (or some reasonable $\bar \varepsilon$) \label{}\\
Set $\beta \leftarrow 10$ (or some reasonable energy error threshold). \label{}\\
Set $\rho \leftarrow 0.5$ (or some reasonable step size decrement). \label{}\\
{\bfseries for}($s=1,2,\dots$ ) \label{}\\
\> Compute $|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|$.\label{}\\
\> {\bfseries if}($|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|>\beta$)\label{}\\
\> \> $\varepsilon_{s+1} \leftarrow \rho \varepsilon_s$.\label{}\\
\> {\bfseries else} \label{}\\
\> \> {\bfseries if}($|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|<\gamma$)\label{}\\
\> \> \> {\bfseries Return} $\varepsilon(\mathbf{x},\mathbf{w}) = \varepsilon_s$.\label{}\\
\> \> {\bfseries else}\label{}\\
\> \> \> $\varepsilon_{s+1} \leftarrow 0.95 \left( \frac{\gamma}{|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|} \right)^{1/3} \varepsilon_{s}$.\label{}\\
\> \> {\bfseries end if}\label{}\\
\> {\bfseries end if}\label{}\\
{\bfseries end for}\label{}\\
\end{ntabbing}
\protect\caption{\label{alg:line_search}Practical backtracking line search algorithm. }
\end{algorithm}
For practical applications it is not necessary, nor desirable from
a computational perspective, to find $\varepsilon_{0}$ with high
precision. Rather a backtracking iterative scheme as exemplified in
Algorithm \ref{alg:line_search} can be used. This algorithm is inspired
by line searches commonly used numerical optimization \citep{noce:wrig:1999}
and the overarching idea is to generate a sequence trial step sizes
$\bar{\varepsilon}=\varepsilon_{1}>\varepsilon_{2}>\dots$ until the
criterion $|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|<\gamma$
is fulfilled. Algorithm \ref{alg:line_search} has two types of decrements,
where in the case that absolute energy error is greater than $\beta$,
the next trial step size is a factor $\rho<1$ times the incumbent
one. If the absolute energy error is smaller than $\beta$, but greater
than $\gamma$, the choice of the next trial step size is informed
by the fact that $\Delta(\varepsilon|\mathbf{x},\mathbf{w})=O(\varepsilon^{3})$
for small $\varepsilon$. More specifically, I let next trial step
$\varepsilon_{s+1}$ be 0.95 times the root in $\varepsilon$ of $\hat{\Delta}_{s}(\varepsilon)-\gamma$
where $\hat{\Delta}_{s}(\varepsilon)=(\varepsilon/\varepsilon_{s})^{3}|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|$
is the third order monomial that interpolates the observed absolute
energy error at $\varepsilon_{s}$. The factor 0.95 is included to
counteract slow convergence infrequently seen when $|\Delta(\varepsilon_{s}|\mathbf{x},\mathbf{w})|$
is close to $\gamma$. Of course Algorithm \ref{alg:line_search}
constitute only an example of how to implement $\varepsilon(\mathbf{x},\mathbf{w})$,
and may be further tuned and refined for each model instance without
disturbing the stationary distribution as long as it remains the same
throughout the MCMC simulation and fulfills the assumptions in Proposition
1. The cost of each iteration will typically be dominated by the gradient
evaluation needed for computing $\mathbf{r}$ for each trial $\varepsilon$.
For non-Gaussian targets, $\gamma$ needs to be tuned to obtain e.g.
the desired acceptance rate or to maximize some more direct performance
measure such as the expected jumping distance or effective sample
size. The tuning could be done by e.g. dual averaging during the burn
in iterations as in \citet{JMLR:v15:hoffman14a}. However, I have
found that as a rule of thumb, values of $\gamma$ between 1 and 2
tends to produce acceptable results for low to moderate-dimensional
problems as the ones considered below. In this case, the adaptive
step size selection typically produces long step sizes ($\varepsilon\sim\bar{\varepsilon}$)
when $G$ contains useful scaling information, but acts as safeguard
and reduces the step size substantially when $G$ shows some form
of pathological behavior.
It is interesting to consider the impact of dimension $d$ on the
proposed adaptive step size selection, and consequently on the overall
performance of the proposed methodology. For this purpose, I consider
any $d$-dimensional non-degenerate $N(\mu,\Sigma)$ target, as it
is rather straightforward to show that
\[
E_{(\mathbf{x},\mathbf{w})}\left(\Delta(\varepsilon|\mathbf{x},\mathbf{w})\right)=d\left(\frac{1}{4}\varepsilon^{4}-\frac{1}{32}\varepsilon^{6}\right),
\]
when $G(\mathbf{x})=-H(\mathbf{x})=\Sigma^{-1}$. For any fixed $\gamma$,
which in the Gaussian case translates to a fixed acceptance rate invariant
of $d$, the adaptive step size has leading term behavior $\varepsilon=O(d^{-1/4})$.
Thus for high-dimensional target distributions, it is inevitable that
adaptive step size sMMALA will revert to random walk behavior, and
as a consequence I consider primarily low to moderate-dimensional
applications as the primary scope.
\section{Illustrations}
This section considers example models and compares the proposed Adaptive
step size Modified Hessian MALA (AMH-MALA) methodology to alternatives.
The GARCH(1,1) model with $t$-distributed innovations considered
first is included to highlight the behavior of AMH-MALA for a posterior
distribution that exhibit strong non-linear dependence structures
and has indefinite Hessian in substantial parts of the relevant parameter
space. The Bayesian binary response regressions considered afterwards
are included to allow for easy comparison with the methodology presented
in \citet{girolami_calderhead_11}.
To compare the performance of different MCMC methods, I follow \citet{girolami_calderhead_11}
in estimating effective sample size (ESS) via the initial monotone
sequence estimator of \citet{geyer1992}, and in particular compare
the minimum ESS (across the $d$ dimensions of $\pi$) per second
CPU-time spent obtaining the samples. All computations were carried
out on a 2014 macbook pro with a 2 GHz Intel Core i7 processor and
8 GB of memory. The code used for the simulations is available at
http://www.ux.uis.no/\textasciitilde{}tore/code/adaptive\_langevin/,
including a C++ function implementing the dense modified Cholesky
factorization that can be called from Matlab.
\subsection{GARCH(1,1) model with $t$-distributed innovations}
\begin{table}
\begin{centering}
\begin{tabular}{lcccccc}
\hline
Method & CPU time & $\alpha_{0}$ & $\alpha_{1}$ & $\beta$ & $\nu$ & minimum ESS\tabularnewline
& (s) & ESS & ESS & ESS & ESS & per second\tabularnewline
\hline
bayesGARCH & 18.7 & 80.0 & 52.9 & 48.2 & 63.9 & \textbf{(2.46)}\tabularnewline
AMH-MALA & 140 & 283 & 310 & 252 & 398 & \textbf{1.79}\tabularnewline
AMH-MALA(eig) & 146 & 251 & 262 & 229 & 424 & 1.51\tabularnewline
HMC & 2653 & \textbf{4738} & \textbf{2284} & \textbf{2503} & \textbf{4993} & 0.86\tabularnewline
\hline
\end{tabular}\protect\caption{\label{tab:GARCH-Effective-samples-sizes}Effective samples sizes
and minimum effective samples sizes per second CPU time for different
MCMC algorithms applied to the GARCH(1,1) model with $t$-distributed
innovations. Best performances are indicated in bold font. All figures
are calculated over 5000 MCMC iterations after burn in. Note that
the bayesGARCH Gibbs sampler is partially written in C, and thus the
CPU-time and minimum ESS per CPU time are not directly comparable
to the remaining figures. }
\par\end{centering}
\end{table}
\begin{figure}
\centering{}\includegraphics[scale=0.9]{garchplot-crop}\protect\caption{\label{fig:garch-Diagnostics}Diagnostics for a typical run of AMH-MALA
with initial parameters set to the unrealistic values $\log(\alpha_{0})=-10$,
$\log(\alpha_{1})=-1$, $\log(\beta)=-3$ and $\nu=20$. The upper
4 panels are trace plots of parameters $\alpha_{0},\alpha_{1},\beta$
and $\log(\nu)$ where the logarithm is applied to latter for visual
reasons. The 5. panel is a trace plot of the forward step size $\varepsilon_{f}$.
The 6. panel depicts the smallest eigenvalue of $-H(\mathbf{x}_{t})$,
with smallest eigenvalue for the first 211 iterations being smaller
than -25. The last panel shows the log-target target density at the
current iteration, i.e. $\log\tilde{\pi}(\mathbf{x}_{t})$.}
\end{figure}
As the first realistic illustration I consider sampling the posterior
parameters of a GARCH(1,1) model with $t$-distributed innovations
\citep[see e.g.][Chapter 4]{QRM_mcneil} for log-return observations
$y_{1},\dots,y_{T}$ on the form
\begin{eqnarray}
y_{i} & = & \sqrt{h_{i}}\eta_{i},\;\eta_{i}=\sqrt{\frac{\nu-2}{\nu}}\tilde{\eta}_{i},\;\tilde{\eta}_{i}\sim\text{iid }t(\nu),\; i=1,\dots,T,\label{eq:garch1}\\
h_{i} & = & \alpha_{0}+\alpha_{1}y_{i-1}^{2}+\beta h_{i-1},\; i=2,\dots,T,\; h_{1}=\alpha_{0}.\label{eq:garch2}
\end{eqnarray}
The priors are taken from \citet{RJournal_2010-2_Ardia+Hoogerheide}
and are the default priors used in the associated \texttt{bayesGARCH}
R-package for the same model. Specifically, $\alpha_{0},\alpha_{1},\beta$
have independent truncated normal priors $p(\alpha_{k})\propto\exp(-\frac{1}{2}\frac{\alpha_{k}^{2}}{1000})\mathbf{1}_{\left\{ \alpha_{k}>0\right\} },\; k=0,1$
and $p(\beta)\propto\exp(-\frac{1}{2}\frac{\beta^{2}}{1000})\mathbf{1}_{\left\{ \beta>0\right\} }$.
The prior for the degrees of freedom parameter $\nu$ is a truncated
exponential, namely $p(\nu)\propto\exp(-\nu/100)\mathbf{1}_{\left\{ \nu>2\right\} }$.
For our illustration, I employ a data set of Deutschmark vs British
Pound (DEM/GBP) foreign exchange rate log-returns at daily frequency,
which is also included in the \texttt{bayesGARCH} package. The sample
covers January 3, 1985 to December 31, 1991, and constitute $T=1974$
log-return observations.
The proposed AMH-MALA methodology is compared to the Gibbs sampler
implemented in the \texttt{bayesGARCH} package \citep{RJournal_2010-2_Ardia+Hoogerheide}
and HMC with mass matrix fixed to the identity matrix. For HMC I follow
\citet{girolami_calderhead_11} and choose 100 time integration steps
and the integration step size was taken to be $\varepsilon=0.0075\pm10\%$
uniform noise to attain acceptance rates of around 80\%. In addition
I also consider AMH-MALA implemented with a metric tensor similar
to the one proposed by \citet{1212.4693} and denote this by AMH-MALA(eig).
Specifically, let $\lambda_{i}$ denote the $i$th eigenvalue of $-H$.
Then $G$ has the same eigenvectors as $-H$ but the $i$th eigenvalue
is taken as $\max(|\lambda_{i}|,0.001)$. For both AMH-MALA methods,
I use $\gamma=1.0$ and Algorithm \ref{alg:line_search} with $\beta=10.0$,
$\rho=0.5$ for adaptive step size selection. For AMH-MALA based on
the modified Cholesky, I use $u=0.001$. HMC and the AMH-MALA methods
are applied in $\log$-transformed parameters $\alpha_{0}^{\prime}=\log(\alpha_{0})$,
$\alpha_{1}^{\prime}=\log(\alpha_{1})$, $\beta^{\prime}=\log(\beta)$
and $\nu^{\prime}=\log(\nu-2)$ to remove potential numerical problems
related to the truncations imposed by the priors.
Table \ref{tab:GARCH-Effective-samples-sizes} provides mean CPU times,
mean ESSes and mean minimum ESS per computing time for the parameters
of the GARCH model (\ref{eq:garch1}-\ref{eq:garch2}) across 10 independent
repeated runs of 5000 samples each. For \texttt{bayesGARCH} I used
5000 burn in iterations, whereas the remaining methods used 1000 burn
in iterations. The reported CPU times excludes burn in. It should
be noted that the Gibbs sampler in the \texttt{bayesGARCH} package
is written partially in C and the computing times are therefore not
directly comparable to the remaining methods, which are implemented
fully in Matlab. It is seen that AMH-MALA produces the most effective
samples per unit computing time for the methods written in Matlab,
and also produces substantially better ESSes per iteration than the
\texttt{bayesGARCH}. AMH-MALA and AMH-MALA(eig) produces similar results,
indicating that there is little added value to using full spectral
decomposition over the modified Cholesky factorization for this situation.
Figure \ref{fig:garch-Diagnostics} depicts various diagnostic output
for a typical run of AMH-MALA applied to the parameters of (\ref{eq:garch1}-\ref{eq:garch2}).
The initial parameters are set to highly ill informed values to illustrate
the different behavior of AMH-MALA in transient and stationary regimes.
It is seen that AMH-MALA takes shorter forward steps in the transient
regime up to approximately iteration 220. The negative Hessian is
severely indefinite for these iterations. This shows that choosing
step size based on the energy error enables to AMH-MALA to make progress
in the transient regime with the same tuning parameters as are used
in the stationary regime. To contrast this, fixing $\varepsilon_{f}=\varepsilon_{b}=1$
but otherwise keeping the setup identical to that reported in Figure
\ref{fig:garch-Diagnostics} results in a single accepted move in
the course of the 2000 iterations, indicating the need for different
tuning parameters in stationary and transient regimes \citep{RSSB:RSSB500}.
Visual inspection of Figure \ref{fig:garch-Diagnostics} indicates
that small values of $\varepsilon_{f}$ are often associated with
near zero minimum eigenvalues. This can be statistically confirmed
as the correlation between $\log|\min_{i}\lambda_{i}|$ and $\varepsilon_{f}$
is 0.63 for iterations 220-2000. This indicates that the adaptive
step size procedure on average sees both large positive- and large
negative curvature information as useful and consequently takes long
step sizes, whereas it takes shorter step sizes when very little curvature
information is available in some direction. Still there is a small
positive correlation of 0.24 between $\min_{i}\lambda_{i}$ and the
acceptance probability, indicating that AMH-MALA performs slightly
better in regions of the support where the target is log-concave.
\subsection{Bayesian binary response regression}
\begin{table}
\begin{centering}
\begin{tabular}{llccccc}
\hline
Model & Method & CPU time & minimum & median & maximum & minimum ESS\tabularnewline
& & (s) & ESS & ESS & ESS & per second\tabularnewline
\hline
\multicolumn{7}{c}{Australian credit data set ($d=15$, $n=690$)}\tabularnewline
\hline
logit & AMH-MALA & 2.6 & \textbf{ 436} & 579 & 725 & 167\tabularnewline
logit & sMMALA & 1.6 & \textbf{ 436} & \textbf{ 591} & \textbf{ 745} & \textbf{271}\tabularnewline
\hline
probit & AMH-MALA & 3.4 & \textbf{ 575} & \textbf{ 774} & \textbf{ 919} & 167\tabularnewline
probit & sMMALA & 1.8 & 491 & 665 & 869 & \textbf{ 280}\tabularnewline
probit & Adaptive sMMALA & 3.3 & 475 & 652 & 809 & 142\tabularnewline
\hline
\multicolumn{7}{c}{German credit data set ($d=25,$ $n=1000$)}\tabularnewline
\hline
logit & AMH-MALA & 3.6 & 417 & 570 & 707 & 117\tabularnewline
logit & sMMALA & 2.1 & \textbf{ 432} & \textbf{ 616} & \textbf{ 751} & \textbf{ 202}\tabularnewline
\hline
probit & AMH-MALA & 4.5 & \textbf{ 579} & \textbf{ 747} & \textbf{ 886} & 129\tabularnewline
probit & sMMALA & 2.4 & 534 & 700 & 846 & \textbf{ 226}\tabularnewline
probit & Adaptive sMMALA & 4.4 & 488 & 649 & 785 & 111\tabularnewline
\hline
\multicolumn{7}{c}{Heart data set ($d=14,$ $n=270$)}\tabularnewline
\hline
logit & AMH-MALA & 2.0 & 362 & 468 & 579 & 176\tabularnewline
logit & sMMALA & 1.3 & \textbf{ 364} & \textbf{ 473} & \textbf{ 587} & \textbf{ 273}\tabularnewline
\hline
probit & AMH-MALA & 2.5 & \textbf{ 612} & \textbf{ 760} & \textbf{ 883} & 247\tabularnewline
probit & sMMALA & 1.4 & 451 & 593 & 734 & \textbf{331}\tabularnewline
probit & Adaptive sMMALA & 2.4 & 475 & 590 & 719 & 197\tabularnewline
\hline
\multicolumn{7}{c}{Pima Indian data set ($d=8$, $n=532$)}\tabularnewline
\hline
logit & AMH-MALA & 2.2 & \textbf{ 1043} & \textbf{ 1184} & 1296 & 479\tabularnewline
logit & sMMALA & 1.4 & 1010 & 1174 & \textbf{ 1314} & \textbf{ 732}\tabularnewline
\hline
probit & AMH-MALA & 2.7 & \textbf{ 1143} & \textbf{ 1282} & 1428 & 425\tabularnewline
probit & sMMALA & 1.4 & 1089 & 1253 & \textbf{ 1480} & \textbf{ 752}\tabularnewline
probit & Adaptive sMMALA & 2.6 & 1116 & 1242 & 1454 & 431\tabularnewline
\hline
\multicolumn{7}{c}{Ripley data set ($d=7$, $n=250$)}\tabularnewline
\hline
logit & AMH-MALA & 1.3 & 285 & 375 & \textbf{467} & 227\tabularnewline
logit & sMMALA & 0.6 & \textbf{ 291} & \textbf{ 387} & 464 & \textbf{ 506}\tabularnewline
\hline
probit & AMH-MALA & 1.8 & \textbf{ 387} & \textbf{ 536} & \textbf{ 623} & 212\tabularnewline
probit & sMMALA & 0.8 & 287 & 379 & 490 & \textbf{ 381}\tabularnewline
probit & Adaptive sMMALA & 1.8 & 293 & 391 & 506 & 161\tabularnewline
\hline
\end{tabular}
\par\end{centering}
\protect\caption{\label{tab:Effective-sample-sizes-binary}Effective sample sizes and
CPU times for the Bayesian binary response regressions. Best performances
are indicated with bold font. ``logit'' correspond models with logistic
link function. In this case the negative Hessian and Fisher information
matrix coincides. ``probit'' correspond to the standard normal cumulative
distribution function as the inverse link function. In this case the
negative Hessian is different from the Fisher information matrix.
sMMALA replicates the simplified manifold MALA of \citet{girolami_calderhead_11}
whereas Adaptive sMMALA uses the metric tensor of \citet{girolami_calderhead_11}
along with adaptive step size selection.}
\end{table}
This section considers Bayesian inference for two types of binary
response generalized linear models. The models are included in order
to compare the proposed methodology with that of \citet{girolami_calderhead_11}.
Specifically, I consider models for observed binary responses $\mathbf{y}=(y_{1},\dots,y_{n})$
on the form
\begin{eqnarray}
P(y_{i} & = & 1)=\rho\left[(X\beta)_{i}\right],\; i=1,\dots,n,\label{eq:bin_reg_obs}\\
\beta & \sim & N(0,100I_{d}).\nonumber
\end{eqnarray}
Here $X\in\mathbb{R}^{n\times d}$ is a design matrix and the inverse
link function $\rho$ is specified either as a logit link corresponding
to $\rho(x)=\exp(x)/(1+\exp(x))$ or the probit link corresponding
to $\rho(x)=\Phi(x)$ where $\Phi$ denotes the $N(0,1)$ distribution
function. For both link functions, the Fisher information matrix is
available in closed form, and for the logit link the negative Hessian
of log-likelihood function associated with (\ref{eq:bin_reg_obs})
coincides with the Fisher information matrix, whereas this is not
the case for the probit model. However, the negative Hessian is still
positive definite in the relevant region for the probit model.
I consider the collection of 5 data sets used by \citet{girolami_calderhead_11}
where $n$ ranges between 250 and 1000 and $d$ ranges between 7 and
25. The sMMALA method using $G_{GC}$ is used as a reference. For
the logit model, this amounts to a replication of the simplified MMALA-part
of the Monte Carlo experiment of \citet{girolami_calderhead_11},
and therefore admits calibration of their results against those presented
here. In particular, I recoded the method of \citet{girolami_calderhead_11}
so that all codes use the same routines for evaluating likelihoods,
gradients and Hessians to make the comparison as fair as possible.
For AMH-MALA I employed $\gamma=2.0$ and Algorithm \ref{alg:line_search}
with $\beta=20.0$, $\rho=0.7$. In this setting, Algorithm \ref{alg:line_search}
works mainly as a safeguard against numerical problems occurring when
some $\rho\left[(X\beta)_{i}\right]\rightarrow0$ or $\rho\left[(X\beta)_{i}\right]\rightarrow1$.
The results are presented in Table \ref{tab:Effective-sample-sizes-binary}.
Through out, I collect 5000 samples after 5000 iterations of burn
in, and the timings are for producing the post burn in samples. All
experiments are repeated 10 times and reported numbers are averages
across these replica. In \citet{girolami_calderhead_11}, the simplified
MMALA was found to be the best method for the logit model for 4 out
5 data sets when other contenders included the Riemann manifold HMC
and full manifold MALA and the performance measure was the minimum
(over $\beta$) ESS produced per unit of time.
For the logit model, AMH-MALA may be interpreted as an adaptive step
size version of \citet{girolami_calderhead_11}'s simplified MMALA,
I see that the line search mechanism leads to approximately a factor
2 increase computing time, whereas the number of effective samples
produced per iteration are approximately the same. Looking at the
results reported in Table 3-7 in \citet{girolami_calderhead_11} using
simplified MMALA as a reference, I find that AMH-MALA performance
roughly on par with Riemann manifold HMC for this model and these
data sets.
For the probit model, where $-H(\beta)$ and $G_{GC}$ do not coincide,
I see slightly larger differences in terms of effective sample size
in favor of AMH-MALA, whereas the relative consumption of CPU time
is roughly the same as in the logit case. To further investigate this
find, I also implemented a method based on $G_{GC}$ but otherwise
identical to AMH-MALA and denote this method Adaptive sMMALA. It is
seen that the improved ESSes appear to be a feature of the application
of the Hessian rather than adaptive step size selection as the ESSes
of sMMALA and Adaptive sMMALA are roughly the same. From this I may
conclude that the negative Hessian may be a better alternative than
the Fisher information-based metric tensor for models where the Hessian
is positive definite in all relevant regions.
\section{Discussion }
This paper makes usage of the \citet{gilletal} modified Cholesky
factorizations for producing positive definite metric tensors from
the Hessian matrix for simplified manifold MALA methods. A new adaptive
step size procedure that resolves the shortcomings of metric tensors
derived from the Hessian in regions where the log-target has near
zero curvature in some direction is proposed. The adaptive step size
selection also appears to alleviate the need for different tuning
parameters in transient and stationary regimes. The combination of
the two constitute a large step towards developing reliable manifold
MCMC methods that can be implemented for models with unknown or intractable
Fisher information, and even for targets that does not admit a factorization
into prior and likelihood. Through examples of low to moderate dimension,
it is shown that proposed methodology performs very well relative
to alternative MCMC methods.
To handle high-dimensional problems, it is likely that a HMC variant
of the proposed methodology is needed. One avenue would be to make
$G(\mathbf{x})$ a smooth function via the usage of soft-max functions
\citep{1212.4693} in Algorithm \ref{alg:Left-looking-square-root}
and implement full Riemann manifold HMC along the lines of \citet{girolami_calderhead_11}.
This approach would enable the exploitation of sparsity of the Hessian
not afforded by methods based on spectral decompositions \citep{1212.4693},
but would still require computing third derivatives of $\log\tilde{\pi}$.
An interesting alternative that is currently being investigated involves
embedding HMC into a population MCMC framework where each member of
the population has the same target. In such a setup, one member of
the population is expended in each MCMC iteration for calculating
a position-specific mass matrix and time integration step size using
the modified Cholesky factorization and adaptive step size selection
procedure proposed here. These parameters are then applied in standard
HMC updates of the remaining population members to mimic the effects
of Riemann manifold HMC while retaining a computationally attractive
separable Hamiltonian.
Another avenue for further work is to extend the adaptive step size
methodology via energy error arguments of Section \ref{sub:Line-search-mechanisms}
to other MCMC methods, via the observation that other, possibly non-symplectic,
numerical integration schemes applied to Hamilton's equations associated
with (\ref{eq:dummy_hamiltonian}) leads to different known proposal
distributions. In particular, a symplectic Euler type B integrator
\citep[page 26]{Leimkuhler:2004} lead to a $N(\mathbf{x}+\varepsilon^{2}G(\mathbf{x})^{-1}g(\mathbf{x}),\varepsilon^{2}G(\mathbf{x})^{-1})$
proposal which nests (for $\varepsilon=1$) the stochastic Newton
MCMC method of \citet{martin_etal2012}. A standard Euler integrator
lead to a position specific scale random walk proposal $N(\mathbf{x},\varepsilon^{2}G(\mathbf{x})^{-1})$.
\bibliographystyle{chicago}
|
\section{Introduction}
During the first three years of data sampling in full
configuration, the IceCube Neutrino Telescope based at the South Pole has
registered a sample of very high-energy neutrinos (30~TeV -- 2~PeV) of astrophysical origin \citep{aartsen13, aartsen14}.
With a sample of 37 events the background only hypothesis with no astrophysical component has been rejected with
a significance of more than 5~$\sigma$. A new era for high-energy astrophysics just started.
Excluding, in this context, the possible
connection with dark matter \citep{cherryetal14, esmailietal14},
the IceCube discovery calls for astrophysical sites, where cosmic-ray acceleration at the ``knee'' energy scale and above
is efficiently at work, and the target, be it gas or photons, is sufficiently dense for
interactions with the cosmic rays.
A firm association between the detected neutrinos
and a specific class (or classes) of astrophysical sources
is still lacking, notwithstanding the numerous scenarios that have been
proposed so far. Models suggesting a pure Galactic origin of
the observed neutrino signal
can be strongly constrained by
diffuse TeV-PeV $\gamma$-ray limits
(see e.g.~\citealt{ahlersmurase13, joshi14} and references therein), given
that the production of very high energy (VHE) $\gamma$-rays is an inevitable
outcome of the hadronic interactions that lead to
the neutrino production
in the first place.
However, a Galactic component contributing to the IceCube neutrino flux cannot be excluded at the moment.
Proton-proton ($pp$) collisions in galaxy groups and/or star forming galaxies
have also been invoked to explain the diffuse neutrino flux (e.g. \citealt{loebwaxman06, muraseahlers13}).
Other extragalactic scenarios that include neutrino production through photohadronic ($ {\rm p}\pi$) interactions
in active galactic nuclei (AGN) \citep{stecker91, mannheim95,
halzenzas97,rachen98, muecke99, atoyandermer01, muecke03}
and gamma-ray bursts (GRBs) \citep{waxmanbahcall97, murase08, zhangkumar13, asanomeszaros14, reynoso14, petroetal14, winter14} have been
extensively discussed in the literature.
A commonly adopted approach for the calculation of the neutrino flux at Earth
(see e.g. \citealt{mannheimprotheroe01, baerwald13, kistler13, murase14})
is the following: (i) a generic distribution of accelerated
protons in the source is assumed; (ii) a generic photon distribution that acts
as a target field for $ {\rm p}\pi$ interactions is also assumed; (iii) the neutrino spectrum is then calculated; and
(iv) the neutrino flux is integrated over redshift assuming a luminosity function for the particular astrophysical source.
If many sources of the same type
contribute to the observed flux, then the approach
described above is the most relevant one.
A different approach is to
focus on a few individual sources of the same class and derive
the respective neutrino fluxes using proton and photon distributions that
are unique to each object \citep{muecke03, dpm14}.
In this regard, information about the directions
of the neutrino events on the sky is important.
This approach may still lead to a study of the diffuse neutrino flux,
from a certain class of sources, but with the caveat that it may
not represent the total diffuse flux; the latter could
be arrived at only upon an estimate of the percentage of that class's contribution to the total.
\cite{padovaniresconi14} -- henceforth PR14, have recently searched
for plausible astrophysical counterparts
within the error circles of IceCube neutrinos using high-energy $\gamma$-ray (GeV-TeV) catalogs.
Assuming that each neutrino event is associated with one astrophysical source and using
a model-independent method they derived the most probable counterparts for 9 out of the 18 neutrino events
of their sample. Interestingly, these include 8 BL~Lac objects, amongst which the nearest blazar, Mrk~421,
and two pulsar wind nebulae. Although Flat Spectrum Radio Quasars (FSRQs) are believed to be more efficient PeV neutrino emitters
than BL~Lacs (e.g. \citealt{dermermurase14}; see, however, \cite{tavecchioetal14}),
these do not appear in the list of most probable counterparts.
This is not necessarily related to the physics of neutrino emission from different
types of blazars but could be a consequence of
the ``energetic'' test applied by PR14 (for more details, see \S 4 therein).
Motivated by the aforementioned results, we investigate
the BL~Lac--PeV neutrino connection within a specific theoretical framework
for blazar emission where the $\gamma$-rays are of photohadronic origin.
In contrast to the majority of works on neutrino emission
from blazars, we do not assume a generic SED but we fit instead the multiwavelength (MW) photon spectra
using the radiation of particles whose distributions have been accordingly modified
by the cooling and escape processes.
In our model the low-energy emission of the blazar SED
is attributed to synchrotron radiation of
relativistic electrons, whereas the observed high-energy (GeV-TeV) emission is
the result of synchrotron radiation from pairs produced by charged pion
decays. Pions are the by-product of $ {\rm p}\pi$ interactions
of co-accelerated protons with the internally produced synchrotron photons.
Our approach allows us to identify the observed $\gamma$-ray emission as the emission from the $ {\rm p}\pi$ component
and associate it with a very high-energy ($\sim2-20$~PeV) neutrino emission. Analytical expressions for
the typical energies
of neutrinos and photons produced through photohadronic processes, the proton cooling rate due to
photopion and photopair interactions, as well as the dependence of all aforementioned quantities on
observable quantities of blazar emission can be found in \cite{petromast14}.
First results of the model were presented in \cite{dpm14} -- henceforth DPM14, where
the neutrino flux expected from the prototype blazar Mrk~421 in a low state was derived.
The prediction of the model muon neutrino flux
was tantalizingly close to the IC-40 sensitivity limit
calculated by \cite{tchernin2013} for Mrk~421 (see Fig.~1 in DPM14).
We extend our previous analysis to include those BL~Lacs
identified as the most probable counterparts of the IceCube neutrinos by PR14 (see Table~4 therein).
Our aim is threefold: (i) to examine whether the working model (the so-called LH$\pi$ model, see DPM14) can fit
the SED of the aforementioned blazars -- there was no guarantee before our attempts, since
we did not select {\sl a priori} sources with specific spectral characteristics; (ii) to calculate the neutrino flux
relative to the $\gamma$-ray one, should acceptable fits be obtained, and (iii)
to compare the model-derived neutrino fluxes with those calculated for each single neutrino event.
We show that Mrk~421 does not constitute a unique case. Instead, all
BL~Lacs in our sample can be modelled by the leptohadronic model
and for most of them the model-predicted neutrino flux is close to the observed one, at least within the $3\sigma$ error bars.
This paper is structured as follows. We start with a short description of the theoretical
model and the fitting method we used in Sect.~\ref{modelmethod}. In Sect.~\ref{case} we provide
general information about the observational datasets used for each BL~Lac.
In Sect.~\ref{results} we present the combined MW photon and neutrino spectra
for each of the case studies, focusing on the ratio of the model-predicted neutrino luminosity to the
$\gamma$-ray luminosity and its dependence on observable quantities.
We also comment on the model-derived energetics for each source. We discuss our results in
Sect.~\ref{discussion} and conclude with a summary in Sect.~\ref{summary}.
Throughout this work we have adopted a cosmology
with $\Omega_{\rm m}=0.3$, $\Omega_{\Lambda}=0.7$ and
$H_0=70$ km s$^{-1}$ Mpc$^{-1}$. For the attenuation of the VHE part of the
spectra due to photon-photon pair production on the extragalactic background light (EBL) we used the
model of \cite{franceschini08}.
\section{Model and method}
\label{modelmethod}
\subsection{Model description}
\label{model}
The physical model we use has been described, in general terms,
in \cite{DMPR2012} - henceforth DMPR12, and, in the context of
modelling Mrk~421 in particular, in \cite{mastetal13} and DPM14.
We consider a spherical blob of radius $R$, containing a
tangled magnetic field of strength B and moving towards us
with a Doppler factor $\delta$.
Protons and electrons are accelerated by some mechanism whose
details lie outside the immediate scope of this work, and
are subsequently injected isotropically in the volume of the blob with a constant rate.
Their interaction with the magnetic
field and with secondary particles leads to the development
of a
system where five stable particle populations
are at work: protons, which lose energy by synchrotron radiation,
Bethe-Heitler (pe) pair production, and photopion ($ {\rm p}\pi$) interactions;
electrons (including positrons), which lose energy by
synchrotron radiation and inverse Compton scattering;
photons, which gain and lose energy in a variety of ways;
neutrons, which can escape almost unimpeded from the
source region, with a small probability of photopion
interactions; and neutrinos, which escape completely unimpeded.
The interplay of the processes governing the evolution
of the energy distributions of those five populations is
formulated with a set of time-dependent kinetic equations.
Through them, energy is conserved in a self-consistent manner,
since all the energy gained by a particle type has to come
from an equal amount of energy lost by another particle type.
To simultaneously solve the coupled kinetic equations for all
particle types we use the time-dependent code described in DMPR12.
Photopion interactions are modelled using the results of the
Monte Carlo event generator SOPHIA \citep{SOPHIA2000},
while Bethe--Heitler
pair production is similarly modelled with the Monte Carlo
results of \cite{Protheroe1996} and \cite{mastetal05}.
The only particles not modelled with kinetic equations are muons,
pions and kaons. Their treatment is described in DPM14 and
\cite{petroetal14} but is of negligible importance for the
range of parameter values used in the present work.
The prime mover of all processes is the rate at which energy is injected
in relativistic protons and electrons. This, in terms of
compactnesses, is expressed as:
\begin{eqnarray}
\ell_{\rm i, inj}={{L'_{\rm i, inj} \sigma_{\rm T}}\over{4\pi R m_{\rm i} c^3}},
\label{lpinj}
\end{eqnarray}
where i denotes protons or electrons (i=e,p),
$\sigma_{\rm T}$ is the Thomson cross section, and $L'_{\rm i, inj}$ denotes the injection luminosity as measured in
the rest frame of the emitting region.
In principle, the radius $R$ and Doppler factor $\delta$ can be
constrained using variability arguments. However, we
prefer to treat them as free parameters, since
information for high-energy variability is not available for
all sources in the sample.
The escape times for protons and electrons are assumed to be
the same ($t_{\rm p,esc}=t_{\rm e,esc}$), and complete the set of free parameters
dependent on the geometry and kinematics of the blob.
What remains to be defined are the shapes of the proton and primary
electron injection functions, which are described as power-laws with index
$s_{\rm i}$ extending from $\gamma_{\rm i, \min}$ to $\gamma_{\rm i, \max}$.
In some cases, modelling of the SED requires the primary electron distribution to be
a broken power-law, with the break energy $\gamma_{\rm e, br}$ and the
power-law index above the break $s_{\rm e, h}$ being two additional free parameters.
As long as the escaping protons and neutrons are energetic enough, they
are susceptible to photopion
interactions with ambient photons in Galactic and intergalactic
space, such as the cosmic microwave and infrared backgrounds, producing
additional high-energy neutrinos \citep{stecker73}.
Neutrons also rapidly decay into protons
\citep{sikoraetal87, kirkmast89, giovanonikazanas90, atoyandermer01},
leading also to neutrino production.
In this work we focus on the neutrino emission
from the objects themselves, and will not consider any such
additional contributions from escaping high-energy nucleons.
The reader should note that we placed our emphasis on fitting the MW
data with a variant of the leptohadronic model where the $\gamma$-ray emission is
attributed to synchrotron radiation of the secondaries arising from charged pion decay.
Clearly, this is not the only way of fitting the SED,
as it is possible to fit the $\gamma$-rays with proton synchrotron radiation (e.g. \citealt{aharonian00}). In that case,
however, the model produces a very low neutrino flux at $\sim$EeV energies (see LHs model in DPM14), which makes it
irrelevant to the IceCube TeV-PeV detections.
Moreover, scenarios where the high-energy blazar emission is the result
of $\gamma$-rays produced through $pp$ collisions within the blazar jet, constitute another alternative\footnote{
To our knowledge, direct application of such models to MW observations of blazars is still lacking in the literature.}. Such models
would require, however, uncomfortably high densities of non-relativistic (cold) protons to be present
in the blazar jet (see e.g.~\citealt{atoyandermer04} and references therein).
Last but not least, there is always the possibility of assuming a
purely leptonic model \citep{boettcherreimer13} which, however, does not produce any neutrinos.
\subsection{Fitting method}
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{fitting12706.eps}
\caption{Flow diagram illustrating the steps of the fitting method.}
\label{fitting}
\end{figure}
The multi-parameter nature of the problem (11-13 free parameters; see Sect.~\ref{results}) and
the long simulation time, which is required for every parameter set (of the order of $\sim0.5$ d),
prevents us from a blind search of the whole available parameter space.
Moreover, the DMPR12 code in its current form has not been developed for parallel computing.
Yet, the application of the model to actual observations is still viable (e.g. \citealt{mastetal13, petro14})
thanks to the insight we gain from an analytical treatment of the problem.
Thus, instead of automatically scanning the parameter space,
we start the fitting procedure with a set of values, which are physically motivated (see below).
Then, we perform a series of numerical simulations
with parameter values lying close to the initial parameter set,
until a reasonably good fit to the SED is obtained. The term ``reasonably good'' fit
refers to a model SED that passes through most observational points. This is not necessarily the best fit
in the strict sense, i.e. the one with the lowest $\chi^2$ value. However,
the main results of our study, such as the $\gamma$-ray/neutrino connection,
are not affected by the exact value of the $\chi^2$ statistics.
The steps of the fitting method
are illustrated in Fig.~\ref{fitting}.
The parameters of the problem can be divided into two groups; those related to the properties of the
emission region, i.e. $B$, $R$ and $\delta$ and those
related to the primary particle distributions (see two columns in Fig.~\ref{fitting}).
As benchmark values for the magnetic field we chose $B=0.1, 1, 10$~G. These cover the range
of low-to-moderate magnetic fields. Higher values, e.g. $B\gtrsim 20$~G, were not included to the benchmark set, since
they tend to favour proton synchrotron radiation against photohadronic processes (for proton synchrotron models, see
\citealt{aharonian00, muecke03, boettcherreimer13, dpm14, cerrutietal14}).
For each value of $B$, we then adopted $R =10^{15}$~cm and $10^{16}$~cm and the
set of benchmark values is completed with $\delta=15$ and $30$.
We did not search in the first place for fits having
even higher values of $\delta$ and $R$, since they would decrease
the optical depth ($f_{ {\rm p}\pi}$) for photopion interactions and for neutrino production.
If the low-energy hump of the SED has peak spectral index $\beta$, peak frequency $\nu_{\rm s}$
and total luminosity $L_{\rm syn}$ the optical depth for neutrino production is given by (e.g. PM15)
\begin{eqnarray}
f_{ {\rm p}\pi} (\gamma_{\rm p})
\simeq 4.4 \times 10^{-3} \frac{L_{\rm syn, 45} \lambda(\beta, \nu_{\rm s}) }{R_{15} \delta_1^3 \nu_{\rm s, 16} (1+z)}
\left(\frac{2 \gamma_{\rm p}}{\gamma_{\rm p}^{( {\rm p}\pi)}}\right)^{\beta}, \gamma_{\rm p} > \gamma_{\rm p}^{( {\rm p}\pi)}\frac{\nu_{\rm s}}{2 \nu_{\min}},
\label{eq1}
\end{eqnarray}
where we introduced the notation $q_{x}=q/10^x$ in cgs units,
$\lambda$ is a function of $\beta$ and $\nu_{\rm s}$ (see equations~(23) and (27) in PM15), $z$ is the redshift of the source,
$\nu_{\min}$ is the low cutoff frequency of the low-energy spectrum (for a continuous power-law), $\gamma_{\rm p}$
is the proton Lorentz factor and $\gamma_{\rm p}^{( {\rm p}\pi)}$ is the minimum Lorentz factor that satisfies the threshold
condition for $ {\rm p}\pi$ interactions with the peak energy photons. This is given by (for details, see PM15)
\begin{eqnarray}
\gamma_{\rm p}^{( {\rm p}\pi)} \simeq 3.5 \times 10^7 \delta_1 \nu_{\rm s, 16}^{-1} (1+z)^{-1}.
\label{eq2}
\end{eqnarray}
Protons with the aforementioned Lorentz factor lead to the production of neutrinos with energy, as measured in the observer's frame, given by
\begin{eqnarray}
\epsilon_{\nu} \simeq 17.5 \ {\rm PeV} \delta_1^2(1+z)^{-2} \nu_{\rm s, 16}^{-1}.
\label{Ev}
\end{eqnarray}
In most cases, the above expressions is a safe estimate of the peak energy of the neutrino spectrum.
We use equation~(\ref{eq2}) as the minimum requirement for the maximum proton energy of
the proton distribution, i.e. $\gamma_{\rm p, \max} = \kappa \gamma_{\rm p}^{( {\rm p}\pi)}$ with $\kappa$ a numerical factor of $\sim 2-5$.
On the other hand, the maximum (or break) Lorentz factor of the primary electron distribution is related
to the peak frequency of the low-energy hump as
\begin{eqnarray}
\gamma_{\rm e, \max/\rm br} \simeq 6 \times 10^4 \left(\frac{\nu_{\rm s, 16} (1+z)}{\delta_1 B_{-1}}\right)^{1/2}.
\label{eq3}
\end{eqnarray}
For the lower cutoffs of the proton and primary electron injection functions we use as default values
$\gamma_{\rm e,p \min}=1$ (grey colored parameters in Fig.~\ref{fitting}).
If we find large deviations of the model SED from the observations
at $\nu < \nu_{\rm s}$, we search for $\gamma_{\rm e, \min} > 1$.
We search for different $\gamma_{\rm p,\min}$ than the default one
only if the proton distribution has $s_{\rm p} \ge 2$ and extension to $\gamma_{\rm p, \min}=1$
leads to extremely high injection luminosities. We note, however, that this is not a typical case.
The power-law index of the primary electron distribution ($s_{\rm e}$ or $s_{\rm e, h}$) is directly
related to the observed spectral index of the low-energy component in the SED.
However, it is not straightforward to relate $s_{\rm p}$ with the observed photon
index, e.g. in the Fermi/LAT regime, where the emission from secondaries produced in photohadronic interactions dominates.
Thus, as a first estimate we use $s_{\rm p} \simeq s_{\rm e}$.
The injection luminosity of primary electrons is also directly related to the observed luminosity of the low-energy hump
as (e.g. \citealt{mastkirk97})
\begin{eqnarray}
\ell_{\rm e, inj} = 2\times10^{-4}\frac{L_{\rm syn, 45}}{R_{15} \delta_1^4}.
\label{eq4}
\end{eqnarray}
To derive a first estimate of the proton injection luminosity we assume that the $\gamma$-ray ($> 10$~GeV)
luminosity is totally explained by the synchrotron emission of $ {\rm p}\pi$ pairs:
\begin{eqnarray}
\label{eq00}
L_{\gamma} & \approx & \delta^4 \epsilon'_{\gamma} L'_{ {\rm p}\pi}(\epsilon'_{\gamma})\big|_{{\rm peak}} \\
\epsilon'_{\gamma} L'_{ {\rm p}\pi}(\epsilon'_{\gamma}) & \approx & \frac{1}{8}
\label{eq01}
f_{ {\rm p}\pi}\left(\epsilon'_{\rm p}\right)\epsilon'_{\rm p} L'_{\rm p, inj}\left(\epsilon'_{\rm p} \right) \approx \frac{1}{8}
f_{ {\rm p}\pi}\left(\epsilon'_{\rm p}\right) L'_{\rm p, inj},
\label{eq02}
\end{eqnarray}
where $f_{ {\rm p}\pi}$ is given by equation~(\ref{eq1}). The factor $1/8$ in equation~(\ref{eq01}) comes
from the assumption that half of the produced pions produce electrons/positron pairs, and each of them gains $\sim 1/4$ of the parent
proton energy:
\begin{eqnarray}
\label{eq03}
\epsilon'_{\rm p} & = & \gamma_{\rm p} m_{\rm p} c^2.
\end{eqnarray}
The characteristic photon energy is given by
\begin{eqnarray}
\label{eq04}
\epsilon'_{\gamma} & \simeq & m_{\rm e} c^2 \frac{B}{B_{\rm cr}} \gamma_{\rm e, {\rm p}\pi}^2,
\end{eqnarray}
where
\begin{eqnarray}
\gamma_{\rm e, {\rm p}\pi} & \simeq & \kappa_{ {\rm p}\pi} \frac{m_{\rm p}}{4 m_{\rm e}} \gamma_{\rm p}, \ \kappa_{ {\rm p}\pi}=0.2
\label{eq05}
\end{eqnarray}
Using equations~(\ref{eq00})-(\ref{eq05}), (\ref{lpinj}), and (\ref{eq1}) we derive a rough estimate of the proton injection compactness:
\begin{eqnarray}
\ell_{\rm p, inj} \simeq 2.3 \times 10^{-3}
\frac{L_{\gamma, 46}}{L_{\rm syn, 45}}\frac{\nu_{\rm s,16}(1+z)}{\delta_1 2^{\beta} \lambda(\beta, \nu_{\rm s})}.
\label{eq5}
\end{eqnarray}
Summarizing, for each set of benchmark values for $B, R$ and $\delta$ and equipped with the analytical estimates of
equations~(\ref{eq1})-(\ref{eq5})
and observables, we may significantly reduce the parameter space we need to scan (Fig.~\ref{fitting}).
\section{Case studies}
\label{case}
Our modelling targets are BL Lacs that are spatially and energetically correlated
with neutrino detections. BL~Lacs
are known for their variable emission
at almost all energies. In high-states they usually exhibit major
flares, especially in X-rays and $\gamma$-rays.
Modelling of their broadband emission in low and high states
requires a different parameterization. For example,
a high-flux state might need a
stronger magnetic field (for modelling of blazar flares, see \citealt{mastkirk97, coppiaharonian99, sikoraetal01, krawczynski02}),
which in turn would affect the resulting $\gamma$-ray and neutrino
fluxes.
Thus, it is important to build an SED using simultaneous data or, if this is not possible, using
MW observations that have been obtained at least within the same year.
Out of the 8 BL Lacs reported by
PR14, two have unknown redshifts: SUMSS J014347-584550
and PMN J0816-1311, which are the probable counterparts of neutrinos 20 and 27, respectively. Since redshift is a crucial
parameter in determining the luminosity of the emitting regions
and, thus, the interplay of processes at work there, we have
chosen to exclude those two BL Lacs from this study.
Therefore, we are left with six targets, which are:
\begin{itemize}
\item Mrk~421 at $z=0.031$ \citep{Sbarufatti2005_Mrk421z}.
This is the closest high-frequency peaked blazar (HBL) and the target of many MW campaigns.
Mrk~421 was also the first source which our model (see Sect.~\ref{model}) was applied to and the
main focus of an earlier work (DPM14).
We first quote the results of DPM14, that
used the simultaneous X-ray (RXTE) and TeV $\gamma$-ray (Whipple and HEGRA)
data of the 2001 MW campaign that lasted 6 days \citep{Fossati2008}.
In particular, DPM14 fitted the SED of March 22nd/23rd 2001 that corresponds to a pre-flare state.
During this period no GeV observations were available. Thus,
the Fermi/LAT data during the period of the IC-40 configuration \citep{Abdo2011} were included for comparison reasons only.
As a second step, we use the MW observations of \cite{Abdo2011} together with the
neutrino event ID 9 in order to build a {\sl hybrid} SED.
To avoid repetition and to emphasize the prediction of DPM14 regarding the muon neutrino flux,
we do not attempt to
fit the new dataset. Instead, we use the same parameters as in DPM14, apart from the Doppler factor, and
compare the model results against the \cite{Abdo2011} observations and the neutrino flux for the event ID 9.
\item PG~1553+113 at $z=0.4$ \citep{aleksic12}. This is not only the most distant source of this sample, but
also belongs to the class of most distant TeV blazars in general. Its redshift has not been firmly measured,
yet the value adopted here can be considered
as a safe estimate (for different techniques and results, see \citealt{aleksic12}).
Similarly to 1ES~1011+496 below, it has been only recently detected in VHE $\gamma$-rays by
both H.E.S.S. \citep{aharonian06} and MAGIC \citep{albert07b}.
The data from the second year Fermi-LAT
Sources Catalog at $\sim 1$ GeV (2FGL\_lc) \footnote{http://www.asdc.asi.it/fermi2fgl/} indicate a variable high-energy flux.
The 2005-2009 MAGIC observations also suggest variability at the VHE part of the spectrum (see Fig.~1 in \citealt{aleksic12}).
In 2005 the source exhibited large flux variations in X-rays, yet there were no simultaneous data, neither at GeV nor at TeV energies.
We adopted the data shown in Fig.~6 of \cite{aleksic12}.
For the SED fitting we used, in particular,
the X-ray data that correspond to the intermediate flux
level of 2005 along with the time-averaged 2008-2009 Fermi/LAT data and the MAGIC observations
for the periods
2005-2006 and 2007-2009. We
include in our figures the high state X-ray data of 2005 only for illustrative reasons.
In principle, one could try to fit the time-averaged $\gamma$-ray data
together with the 2005 high state in X-rays.
However, we find less plausible a scenario where the 2005 X-ray high-state is not accompanied by an increased
$\gamma$-ray flux but, instead, is related to the average $\gamma$-ray flux level. There are two reasons for this:
the $\gamma$-ray and X-ray emission in blazars are usually correlated (e.g. for Mrk~421,
see \citealt{Fossati2008}) and the $\gamma$-ray emission of PG~1553+113 is itself
variable (see e.g. \citealt{aharonian06}).
\item 1ES 1011+496 at $z=0.212$ \citep{albert07}. This is an HBL source at indermediate
redshift, that has been detected for the first time in TeV $\gamma$-rays in 2007 with MAGIC \citep{albert07}. The
discovery was the result of a follow up observation of the optical high state reported by the Tuorla blazar monitoring program\footnote{http://users.utu.fi/kani/1m}.
The previous non-detections in TeV $\gamma$-rays suggest that 1ES~1011+496 is variable
both in optical and VHE $\gamma$-rays. Its X-ray emission is described by a
steep spectrum in both low and high states (see \citealt{reinthal12} and references therein).
Because of the variability observed across the MW spectrum, we chose to apply our model to simultaneous data, if possible.
We therefore used data from the first MW campaign in 2008 as presented in \cite{reinthal12}, despite the fact
that they are characterized as preliminary.
We note that the MAGIC data presented in \cite{reinthal12} have been de-absorbed using
the EBL model of \cite{kneiske10}. Using the same model we retrieved the observed MAGIC spectrum, which
we show in the respective figure.
We also took into account the Fermi/LAT data from the first (1FGL) \citep{abdo10}
and second (2FGL) \citep{nolan12} catalogs. In the respective figures we include 2FGL\_lc data, in order to show the variability in the GeV regime.
\item H~2356-309 at $z=0.165$ \citep{falomo91}. This is one of the brightest HBLs in X-rays, which led
to its early X-ray detection with UHURU \citep{forman78}.
Its X-ray emission is variable both in flux and in spectral shape (for more information, see \citealt{abramowski10}).
TeV $\gamma$-ray emission from H 2356-309 was observed for the first time in 2004 by H.E.S.S. \citep{aharonian06b}.
Using the SED builder tool\footnote{http://www.asdc.asi.it/SED} of the ASI Science Data Centre (ASDC)
we chose those X-ray observations that fall within the period
2004-2007 of H.E.S.S. observations. We also included data from the 1FGL and 2FGL catalogs, although
they refer to the period 2008-2010, as well as hard X-ray data obtained by Swift/BAT between 2004 and 2010.
Finally, we over-plotted the historical X-ray high state observed with BeppoSaX, for comparison
reasons.
\item 1H~1914-194 at $z=0.137$ \citep{Carangelo2003_1H1914z}. This is the first BL~Lac of the sample that has not yet
been detected in TeV $\gamma$-rays, and the least well covered in different energy bands among the sources of this sample.
In order to apply our model we constructed a SED using available data from the ASDC, which
is comprised mainly of NED archival data. Given the quality of the available dataset, our main focus was
to describe as well as possible the Fermi/LAT data, while roughly describing the observations in other energy bands.
\item 1~RXS~J054357.3-553206 \ at $z=0.374$ \citep{Stadnik2014_1RXSz}. This is the second most distant object of the sample, after blazar PG~1553+113.
It was first detected and classified as a BL~Lac by ROSAT \citep{Fischer1998AN}.
Following radio observations by \cite{Anderson2009} it was classified as an HBL,
and it was soon thereafter detected in $\gamma$-rays by Fermi/LAT \citep{abdo10}.
It remains still undetected in TeV energies. Its redshift was
estimated by \cite{Stadnik2014_1RXSz}, based on the assumption that the flux of the host elliptical galaxies of
BL Lacs can be used as a standard candle, and also by \cite{Pita2014}, based on spectroscopy on a weak
Ca II H/K doublet and a Na~I absorption line. While the \cite{Pita2014} estimate
($z=0.237$) is somewhat lower than the \cite{Stadnik2014_1RXSz} one, adopting it would
not significantly alter out fit.
Using the SED builder tool of ASDC we combined infrared (WISE), X-ray (Swift/XRT and XMM-Newton) and $\gamma$-ray
(Fermi/LAT) observations that were obtained within the same year (2010).
We also included older XMM-Newton observations performed in 2005 and 2007, for comparison reasons. Finally, the
2FGL\_lc data illustrate the variability in the GeV energy band.
\end{itemize}
Before closing this section and for completeness reasons, we
summarize in Table~\ref{tab-00},
the coordinates of the five IceCube neutrinos and the six most probable blazar counterparts, as well as
their angular offset. We also included the median angular error for each neutrino detection and the
respective detection time in Modified Julian Days (MJD)\footnote{We note that all
data listed in Table~\ref{tab-00} are adopted from Tables 1 and 2 in PR14.}. A sky map (in equatorial coordinates)
of the five IceCube neutrinos and their respective counterparts is shown in Figure~\ref{map}.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{map2.eps}
\caption{Sky map in equatorial coordinates of the five IceCube neutrino events (crosses) that have as probable astrophysical counterparts
the six BL~Lac sources (stars) mentioned in text. The red circles correspond to the median angular
error (in degrees) for each neutrino event (see Table~\ref{tab-00}).}
\label{map}
\end{figure}
\begin{table*}
\caption{List of the most probable counterparts of selected high-energy IceCube neutrinos used in our modelling.}
\begin{tabular}{ccc ccl ccc}
\hline
IceCube ID & Ra(deg) & Dec(deg) & Median angular error (deg) & Time (MJD) & Counterpart & Ra(deg) & Dec(deg) & Angular offset (deg) \\
\hline
9 & 151.25 & 33.6 & 16.5 & 55685.6629638 & Mrk~421 & 166.08 & 38.2 & 12.8 \\
\multicolumn{5}{c}{\phantom{a}} & 1ES~1011+496 & 155.77 & 49.4 & 15.9 \\
10 & 5.00 & -29.4 & 8.1 & 55695.2730442 & H~2356-309 & 359.78 & -30.6 & 4.7 \\
17 & 247.40 & 14.5 & 11.6 & 55800.3755444 & PG~1553+113 & 238.93& 11.2 & 8.9 \\
19 & 76.90 & -59.7 & 9.7 & 55925.7958570 & 1RXS~J054357.3-553206 & 85.98 & -55.5& 6.4\\
22 & 293.70 & -22.1 & 12.1 & 55941.9757760 & 1H~1914-194& 289.44 & -19.4 & 4.8\\
\hline
\end{tabular}
\label{tab-00}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=0.45\textwidth]{Mrk421_v2.eps}
\includegraphics[width=0.46\textwidth]{Mrk421_abdo2011_v2.eps}
\caption{Left panel: SED of blazar Mrk~421 as modelled in DPM14. The Fermi/LAT points are shown only for illustrative reasons, as they
were not included in the fit of the original paper DPM14. Right panel: SED of Mrk~421 averaged over the period
19 January 2009 - 1 June 2009 (Abdo et al. 2011); all data points are from Fig.~4 in Abdo et al. (2011).
The model SED (black line) and neutrino spectra (red lines)
are obtained for the same parameters as in DPM14, except for the Doppler factor, i.e. $\delta=20$.
The neutrino event ID 9 is also shown. In both panels, the model spectra are not corrected for absorption on the EBL.
For comparison reasons, the unattenuated $\gamma$-ray emission from the $\pi^0$ decay is over-plotted with dash-dotted grey lines.}
\label{mrk421}
\end{figure*}
\section{Results}
\label{results}
We first present the results of our SED modelling along with the predicted neutrino flux
for each source, i.e. we build model {\sl hybrid} SEDs. The parameter values
derived for each source are summarized in Table~\ref{tab-0}. Various aspects of the
adopted theoretical framework were investigated in detail in a
recent work (\citealt{petromast14} -- hereafter PM15). For details and analytical expressions
that give insight to the results that follow, we refer the reader to PM15.
In the last two subsections we discuss in more detail the neutrino emission and the energetics
for each source.
\begin{table*}
\caption{Parameter values for the SED modelling of blazars shown in Table~4 in PR14. The ordering of presentation is the same as
in Sect.~\ref{results}.
The parenthesis below the name of each source encloses the ID of the associated neutrino event. The double horizontal line
separates parameters used as an input to the numerical code (upper table) from
those that are derived from it (lower table).}
\begin{threeparttable}
\begin{tabular}{cccc ccc}
\hline\hline
Parameter & Mrk~421 & PG~1553+113 &1ES~1011+496 & H~2356-309 & 1H~1914-194 & 1RXS~J054357.3-553206 \\
& (ID 9) & (ID 17) & (ID 9) & (ID 10) & (ID 22) & (ID 19)\\
\hline\hline
z & 0.031 & 0.4 & 0.212& 0.165 & 0.137 & 0.374 \\
B (G) & 5 & 0.05 & 0.1 & 5 & 5 & 0.1\\
$R$ (cm) & $3\times 10^{15}$& $2\times 10^{17}$ & $3\times 10^{16}$ & $3\times 10^{15}$ & $3 \times 10^{15}$ & $3\times 10^{16}$\\
$\delta$ & 26.5 & 30 & 33 & 30 & 18 & 31\\
\hline
$\gamma_{\rm e, \min}$ & 1 & 1 & 1 & $1.2 \times 10^3$ & $1.2\times10$ & 1 \\
$\gamma_{\rm e, \max}$ & $8\times 10^4$& $2\times 10^6$ & $10^5$\tnote{a} & $10^5$ & $10^5$ & $1.2 \times 10^5$\tnote{a} \\
$\gamma_{\rm e, br}$ & -- & $2.5 \times 10^5$ & -- & $2\times 10^{4}$ & $10^3$ & -- \\
$s_{\rm e, l}$ & 1.2 & 1.7 & 2.0 & 1.7 & 2.0 & 1.7\\
$s_{\rm e, h} $ & -- & 3.7 & -- & 2.0 & 3.0 &-- \\
$\ell_{\rm e, inj}$ & $3.2 \times 10^{-5}$ & $2.5\times 10^{-5}$ & $4\times 10^{-5}$ &$8\times 10^{-6}$ & $6\times 10^{-5}$ & $2.5 \times 10^{-5}$\\
\hline
$\gamma_{\rm p, \min}$ & 1 & $10^3$ & 1 & 1 & 1 & 1 \\
$\gamma_{\rm p, \max}$ & $3.2\times 10^6$ & $6\times 10^6$ & $1.2\times 10^7$& $10^7$ & $2\times 10^6$ & $1.2 \times 10^7$\\
$s_{\rm p}$ & 1.3 & 2.0 & 2.0 & 2.0 & $1.7$ & 2.0\\
$\ell_{\rm p, inj}$& $2\times 10^{-3}$ & $3\times 10^{-4}$ & $4\times 10^{-3}$ & $2.5\times 10^{-3}$ & $10^{-2}$ & $8 \times 10^{-4}$ \\
\hline
\hline
$L'_{\rm e, inj}$~(erg/s)\tnote{b} & $4.4\times 10^{40}$ & $2.3\times 10^{42}$ & $5.6 \times 10^{41}$ & $1.1\times 10^{40}$ & $8.3\times 10^{40}$& $3.5 \times 10^{41}$ \\
$L'_{\rm p, inj}$~(erg/s)\tnote{b} & $5.1 \times 10^{45}$& $5.1 \times 10^{46}$ & $10^{47}$ & $6.4 \times 10^{45}$ &
$2.5\times 10^{46}$& $2 \times 10^{46}$\\
$L_{\gamma, \rm TeV}$~(erg/s)\tnote{c}& $1.7\times10^{45}$ & $7.9 \times 10^{46}$ & $5\times 10^{45}$ & $5\times10^{44}$ & $ 10^{45}$ & $6.3\times 10^{45}$ \\
$L_{\nu}$~(erg/s)\tnote{d} & $2.5\times 10^{45}$ & $8.1 \times 10^{45}$ & $2.5\times 10^{45}$ & $4\times 10^{44}$ & $2 \times 10^{45}$ & $6.3\times 10^{44}$ \\
$Y_{\nu \gamma}$\tnote{e} &1.5 & 0.1 & 0.5\tnote{f}& 0.8& $2.0$ & 0.1\\
$f_{ {\rm p}\pi}$\tnote{g} & $2\times 10^{-5}$ & $6\times 10^{-6}$ & $1.3\times 10^{-6}$ & $3.1\times 10^{-6}$ & $1.2\times10^{-5}$ & $1.1\times 10^{-5}$\\
\hline
\end{tabular}
\begin{tablenotes}
\item[a] An exponential cutoff of the form $e^{-\gamma/\gamma_{\rm e, \max}}$ was used.
\item[b] Proton and electron injection luminosities are given in the comoving frame.
\item[c] Integrated $0.01-1$~TeV $\gamma$-ray luminosity of the model SED. Basically, this coincides with the observed value.
\item[d] Observed total neutrino luminosity, i.e. $(\nu_{\mu} + \overline{\nu}_{\mu}) + (\nu_{\rm e} + \overline{\nu}_{\rm e})$.
\item[e] $Y_{\nu \gamma}$ is defined in equation~(\ref{Ygn}).
\item[f] This is the value derived without including in our modelling
the upper limit in hard X-rays (see Sect.~\ref{1es}). For this reason, it should be considered only
as an upper limit.
\item[g] Estimate of the $ {\rm p}\pi$ optical depth (or efficiency). We define it as $f_{ {\rm p}\pi} \simeq 8 L'_{\gamma, \rm TeV}/L'_{\rm p, inj}$,
where $L'_{\gamma, \rm TeV}$ is the integrated $0.01-1$~TeV $\gamma$-ray luminosity as measured in the comoving frame. The values are only an upper limit because of the assumption that the observed $\gamma$-ray luminosity is totally explained
by $ {\rm p}\pi$ interactions.
\end{tablenotes}
\end{threeparttable}
\label{tab-0}
\end{table*}
\subsection{SED modelling}
\subsubsection{Mrk~421}
As already pointed out in the introduction, the most intriguing result of the
leptohadronic modelling of Mrk~421 was the predicted muon neutrino flux, which
was close to the IC-40 sensitivity limit calculated for the particular source by \cite{tchernin2013}.
This is shown in Fig.~\ref{mrk421} (left panel).
We note that in the original paper DPM14, whose results we quote here, the Fermi/LAT data \citep{Abdo2011}
shown as grey open symbols in Fig.~\ref{mrk421} were not included in the fit, since they were not simultaneous with the rest of the observations.
The total and muon neutrino spectra are plotted with dotted and solid red lines, respectively.
We use the same convention in all figures that follow, unless stated otherwise.
The bump that appears two orders of magnitude below
the peak of the neutrino spectrum in energy originates from neutron decay; we refer the reader to
DPM14 for more details.
According to PR14, Mrk~421 is the most probable counterpart of the corresponding
IceCube event (ID 9). This is now included in the right panel of Fig.~\ref{mrk421}, where
the time-averaged SED during the
MW campaign of January 2009 - June 2009 \citep{Abdo2011} is also shown\footnote{All data points are taken from Fig.~4 in \cite{Abdo2011}.}.
As we have already pointed out in Sect.~\ref{case}, in order to avoid repetition,
we did not attempt to fit the new data set of \cite{Abdo2011}.
Thus, the model photon (black line) and neutrino (red line) spectra
depicted in the right panel of Fig.~\ref{mrk421},
were obtained for the same parameters used in DPM14 (see also Table~\ref{tab-0}), except for the Doppler factor, which was
here chosen to be $\delta=20$. The fit to the 2009 SED is surprisinly good, if we
consider the fact that we used the same parameter set as in DPM14 with only a small modification in the value of the Doppler factor.
We note that the MW photon spectra shown in both panels have not been corrected for EBL absorption in order to demonstrate
the relative peak energy and flux of the neutrino and $\pi^0$ components. The $\gamma$-ray emission from the $\pi^0$
decay is shown in both panels as a bump in the photon spectrum, peaking at $\sim 10^{30}$~Hz. Although
the model prediction for the ratio of the $\pi^0$ $\gamma$-ray luminosity to the total neutrino luminosity is roughly two (see e.g. DMPR12),
the $\pi^0$ component shown in Fig.~\ref{mrk421} is suppressed because of internal photon-photon absorption.
To demonstrate better the effect of the latter, we over-plotted the $\pi^0$ component that is
obtained when photon-photon absorption is omitted (dash-dotted grey lines). In the figures that follow,
we will display, for clarity reasons, the $\pi^0$ component only before its attenuation by the internal synchrotron and EBL photons.
\cite{Abdo2011}, whose data we use here, have also presented a hadronic fit to the SED. Thus,
a qualitative comparison between the two models is worthwhile. Both models are similar in that they require a compact emitting region and
hard injection energy spectra of protons and primary electrons. Morevoer,
the $\gamma$-ray emission from MeV to TeV energies, in both models, has a significant
contribution from the hadronic component. However, the models differ in several aspects because of: (i) the differences
in the adopted values for the magnetic field strength and maximum proton energy; (ii) the Bethe--Heitler process, which
acts as an injection mechanism of relativistic pairs. Because of the strong magnetic field and large $\gamma_{\rm p,max}$ used in \cite{Abdo2011},
the $\gamma$-ray emission from Mrk~421 is mainly explained as proton and muon synchrotron radiation. In our model, however, these
contributions are not important, and the $\gamma$-rays in the Fermi/LAT and MAGIC energy ranges are explained
mainly by synchrotron emission from $ {\rm p}\pi$ pairs (see DPM14).
Moreover, the emission from hard X-rays up to sub-MeV energies is attributed to different
processes. In our model, it is the result of synchrotron emission from Bethe--Heitler pairs, whereas
in the hadronic model of \cite{Abdo2011}, this is explained as synchrotron emission from pion-induced cascades. It is important to note
that the Bethe--Heitler process was not included in the hadronic model of \cite{Abdo2011}.
As far as the neutrino emission is concerned,
we find that the model-predicted neutrino flux is below but close to the 1 $\sigma$ error bar of
the derived value for the associated neutrino 9 (PR14).
This proximity is noteworthy and suggests that the proposed leptohadronic model for Mrk~421
can be confirmed or disputed in the near future, as IceCube collects more data.
Even in the case of a future rejection of the proposed model for Mrk~421, there
will still be room
for leptohadronic models that operate in a different regime of the parameter space. For example, in the LHs model,
the neutrino spectrum is expected to
peak at higher energies and to be less luminous than the one derived here, because
of the higher values of $\gamma_{\rm p,max}$ and $B$ (see also in DPM14); this would also hold for the hadronic
scenario used in \cite{Abdo2011}, although no discussion of neutrinos is presented therein. For such
leptohadronic variants, $\sim$PeV neutrino observations cannot be model constraining.
\subsubsection{PG~1553+113}
\label{pg1553}
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{pg1553_v2.eps}
\caption{SED of blazar PG~1553+113 for the period 2005-2009 and the neutrino flux for the corresponding IceCube event (ID 17).
The data are obtained from \citealt{aleksic12}.
Grey diamonds correspond to the KVA minimum and maximum
fluxes and orange circles are the Swift/UVOT observations.
Light and dark red diamonds are observations from Swift/XRT in 2005 corresponding
to an intermediate and high flux level, respectively. With light green circles
(from top to bottom) are plotted the average RXTE flux and the average 15-150 keV Swift/BAT flux. Green circles represent the average Fermi/LAT spectrum
(August 2008- February 2009) and black circles the MAGIC observations over the periods 2005-2006 and 2007-2009.}
\label{pg1553+113}
\end{figure*}
The hybrid SED of blazar PG~1553+113 is shown in Fig.~\ref{pg1553+113}. The contribution of
different emission processes to the total SED is depicted with different types of lines, which
are explained in the legend above the figure. The same applies to all figures that follow. We did
not
do the same for the case of Mrk~421, since we quoted results originally presented elsewhere, where a detailed description
of the various processes contributing to the overall SED can be found. In any case, it was the prediction
of the model about the neutrino flux from Mrk~421 we wanted to emphasize and not the spectral modelling of the SED.
A few things about the hybrid SED are worth mentioning.
It is the SSC emission of primary electrons (blue dotted line) that mainly contributes to the GeV-TeV $\gamma$-ray regime, whereas
the $ {\rm p}\pi$ component is hidden below it (magenta dash-dotted line). Notice also that the other component of photohadronic origin,
namely the synchrotron emission from Bethe--Heitler pairs (magenta dashed line), has approximately the same peak flux as the $ {\rm p}\pi$ component.
The fact that the SSC emission is favoured with respect to the emission from photohadronic processes
has to do with the particular choice of parameter values (Table~\ref{tab-0}). Besides the weak
magnetic field ($B=0.05$~G), the combination of
a large source ($R =2\times10^{17}$~cm) and a relatively high Doppler factor ($\delta = 30$),
decreases the compactness of the primary synchrotron photon field, which
is the target for the photohadronic interactions, and eventually leads
to a suppression of the proton cooling due to photohadronic processes and of the respective emission signatures
(for more details, see Sect.~3.3 in PM15).
In our framework, the total neutrino luminosity is expected to be of the same order of magnitude as
the synchrotron luminosity of $ {\rm p}\pi$ pairs that is emitted as high-energy $\gamma$-rays (see also PM15).
This is clearly shown in Fig.~\ref{pg1553+113}, where the peak luminosity of the $ {\rm p}\pi$ component
is similar to the one of the neutrino component (red dotted line). Taking into account that
the $ {\rm p}\pi$ component has a small contribution to the $\gamma$-ray emission, we expect also
the peak neutrino flux to be lower than the peak $\gamma$-ray flux, which is verified numerically (see Fig.~\ref{pg1553+113}).
The previous results may be quantified through the ratio
$Y_{\nu \gamma}$, which is defined as
\begin{eqnarray}
Y_{\nu \gamma} = \frac{L_{\nu}}{L_{\gamma, \rm TeV}},
\label{Ygn}
\end{eqnarray}
where $L_{\nu}$ and $L_{\gamma, \rm TeV}$ are the total neutrino and $0.01-1$~TeV $\gamma$-ray luminosities, respectively, as measured
in the observer's frame. The ratio $Y_{\nu \gamma}$
expresses the link between the observed $\gamma$-ray luminosity and the model-predicted neutrino luminosity from a blazar, and as such
is as an important parameter of this study; this will become clear later in Sect.~\ref{ratio}.
For PG~1553+113, in particular, we find $Y_{\nu \gamma} \sim 0.1$, which
suggests that the $ {\rm p}\pi$ component is not the sole contributor to the observed
$\gamma$-ray flux.
Since the low value of $Y_{\nu \gamma}$ is a matter of parameter values, it might make the reader wonder why we adopted
the specific parameter values in the first place, and why we did not search for parameter sets that would favour
photopion production. There are two profound reasons for our particular choice: the shape of the Fermi/LAT
spectrum, and the hard X-ray observations with Swift/BAT. Although a
higher proton injection luminosity or a more compact emitting region would lead to higher neutrino fluxes and could, thus,
account for the neutrino event ID 17, it would inevitably raise the Bethe--Heitler component. Since the emission from pe
pairs falls, in general, in the hard X-ray and soft gamma-ray energy bands, the Bethe--Heitler component would not only exceed
the BAT observations but also destroy the fit to the low-energy Fermi/LAT data. One must also take into account that the
$ {\rm p}\pi$ synchrotron spectrum is broader around its peak than the SSC one, which would make it difficult to explain
the observed Fermi/LAT spectrum.
Inspection of Table~\ref{tab-0} reveals that PG~1553+113 is the second
most energetically demanding
case, with $L'_{\rm p, inj}$ being approximately one order of magnitude higher than the other values.
In terms of parameter values, such as the size $R$ and the magnetic field, it is an ``outlier'' of the sample.
This should not come as a surprise, since PG~1553+113 is the most distant BL~Lac of
our sample with an estimated redshift $z \sim 0.4$ (see Sect.~\ref{case}), and
its $\gamma$-ray flux is approximately equal to that of Mrk~421.
PG~1553+113 is, therefore, an intrinsically more luminous source than Mrk~421 and requires
an accordingly higher proton injection luminosity as measured in the comoving frame,
unless a larger value for Doppler factor ($\delta \gtrsim 30$) was adopted. For a given $R$
this would lower the required proton compactness, and thus
$L'_{\rm p, inj}$.
We attempted, however, to avoid higher Doppler factors than those listed in Table~\ref{tab-0}, since those
would shift the neutrino peak to hundreds of PeV in energy (see equation~(\ref{Ev})).
Even for the adopted parameters, the peak energy of the model-derived neutrino spectrum
is $\sim 10$~PeV. If we combine this
with the hardness of the spectrum we obtain a much lower flux than the observed one,
at the energy of the associated neutrino event;
still, this lies within the 3 $\sigma$ error bars (see also Sect.~\ref{neutrino-emission}).
In Fig.~\ref{pg1553+113} we have also included the X-ray high flux state of 2005 for illustrating reasons. An association
of a higher X-ray flux with a higher neutrino flux would be intriguing, since it would bring
the model-derived flux shown in Fig.~\ref{pg1553+113} even closer to the observed one.
However, such an association is not trivial in the absence of simultaneous observations in high-energies; neither GeV nor
TeV data were available during the 2005 X-ray high state (see Fig.~1 in \citealt{aleksic12}). Consider for example
a scenario where the increase in the X-ray flux is caused by an increase of the electron injection
compactness. This would also result in a higher $\gamma$-ray flux, since
the emission from both the SSC and $ {\rm p}\pi$ components would increase. Finally,
it would lead to a higher neutrino flux because
of the higher synchrotron photon compactness. However, if the X-ray high flux state was not accompanied by an increase
in the $\gamma$-ray emission,
the aforementioned scenario would not be viable.
More than one model parameters should be changed, and, depending on their combination, the derived
neutrino flux would also differ.
Because of the wide range of possibilities we do not, therefore,
attempt to fit the high-state observed in the X-rays, nor to calculate the respective neutrino
flux.
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{1ES1011_v2.eps}
\caption{SED of blazar 1ES 1011+496 for the period 2008-2010 and the neutrino flux for the corresponding IceCube event (ID 9).
Optical, soft X-ray and TeV data were obtained during the first MW campaign in 2008 (\citealt{reinthal12}), whereas
GeV observations are from the first (1FGL) and second (2FGL) Fermi/LAT catalogs (quasi-simultaneous).
The first TeV detection is over-plotted with grey points
for comparison reasons, as it corresponds to a high flux state; the respective data points were obtained from ASDC.
The grey arrow denotes the upper limit to the 20~keV-40~keV X-rays from the fourth INTEGRAL/IBIS catologue (\citealt{ubertinietal09}).
}
\label{1es1011}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{H2356_v2.eps}
\caption{SED of blazar H~2356-309 for the period 2004-2010 and the neutrino flux for the corresponding IceCube event (ID 10).
The data are obtained from ASI Science Data Center
using the SED builder tool. All types of symbols and lines are explained in the legend
above the plot.}
\label{h2356-309}
\end{figure*}
\subsubsection{1ES 1011+496}
\label{1es}
The hybrid SED is shown in Fig.~\ref{1es1011} and the resemblance to the model SED of
PG~1553+113 is obvious. The $\gamma$-ray emission
is mainly attributed to the SSC emission of primary electrons, since the flux of the $ {\rm p}\pi$ component is only a small fraction
of the observed $\gamma$-ray one. The fact that the model neutrino peak
flux is lower than the peak $\gamma$-ray flux should come as no surprise after the discussion in Sect.~\ref{pg1553}.
The model in Fig.~\ref{1es1011} is mainly constrained by the simultaneous optical, soft X-ray and TeV data of the 2008 MW campaign, since
there are no significant spectral variations in the Fermi/LAT observations covering the period 2008-2010.
As already noted in PM15, any information on the hard X-ray emission, e.g.
$\epsilon \gtrsim 20$~keV, would be even more constraining for the model.
This becomes also evident from the cases of PG 1553+113 and H~2356-309 shown in Figs.~\ref{pg1553+113}
and \ref{h2356-309}, respectively.
We searched, therefore, {\sl a posteriori} for possible detections or upper limits
in the hard X-ray band, even if these were not, strictly speaking, simultaneous with the 2008 data. For this, we used
the fourth INTEGRAL/IBIS catalog in 20-40 keV \citep{ubertinietal09}, that covers the period 2003-2008. The source has never been detected
within this period, which resulted in the upper limit of $\sim 6.1\times 10^{-12}$~erg cm$^{-2}$ s$^{-1}$ shown with a grey arrow in Fig.~\ref{1es1011}.
The model SED exceeds by a factor of $\sim 3$ the upper limit. This suggests that the photohadronic emission (both pe and $ {\rm p}\pi$) should
decrease, if we were to include the upper limit into our fitting procedure.
Since the most straightforward way to decrease the photohadronic emission in order to accommodate
the hard X-ray constraint is to assume a lower $\ell_{\rm p, inj}$, the derived value for $L_{\rm p}^{' \rm inj}$ listed in Table~\ref{tab-0}
should be considered only as upper limit. A lower $\ell_{\rm p, inj}$ (keeping all other parameters unaltered) would also result in an accordingly lower
neutrino flux and, thus, lower $Y_{\nu \gamma}$.
Given that the model neutrino flux at the same energy bin with neutrino event 9
is already below the observed one
by approximately two orders of magnitude (see Figs.~\ref{1es1011} and \ref{neutrinos}), we argue that
the hard X-ray constraint makes the connection of 1ES~1011+496 with
the neutrino event 9 even less plausible.
\subsubsection{H~2356-309}
H~2356-309 is one of the sources in the sample that has also been detected in TeV energies by H.E.S.S.
and is the most probable counterpart for the IceCube neutrino with ID 10.
Our model SED and neutrino flux are shown in Fig.~\ref{h2356-309}.
In contrast to PG~1553+113, the Fermi/LAT spectrum can be approximated as $\nu F_{\nu} \propto \nu^0$ within the error bars and
the $\gamma$-ray flux at $\sim 0.1$~TeV is $\sim 1$ order of magnitude below the peak flux of the low-energy hump. The properties
of the observed SED of H~2356-309 allows us, therefore, to search for parameters that favour the $ {\rm p}\pi$ emission and suppress
the contribution of the SSC component, contrary to the cases of PG~1553+113 and 1ES~1011+496.
By adopting $R =3\times 10^{15}$~cm, which is smaller by two and one orders of magnitude than
the source's radius used for PG~1553+113 and 1ES~1011+496, respectively, we increase
the photon compactness of the source. Moreover,
the choice of a stronger magnetic field ($B=5$~G) results in higher magnetic compactness ($\ell_{\rm B}$) than before, where
$\ell_{\rm B} \propto B^2 / R$ and is a measure of the synchrotron cooling rate.
Thus, the synchrotron emission by secondary pairs
from pe and $ {\rm p}\pi$ processes is enhanced compared to the previous two cases.
Figure~\ref{h2356-309} displays the above. It is, indeed, the emission from the $ {\rm p}\pi$ pairs that mainly contributes
to the observed $\gamma$-ray flux, while the SSC component is $\sim 1.5$ orders of magnitude less
luminous. Moreover, the ``Bethe--Heitler bump'' becomes a prominent feature of the SED. Both
results suggest a higher efficiency of both photohadronic processes compared to the previous two cases.
The neutrino spectral shape is similar to the previous cases and peaks at $\sim 10$~PeV.
However, the ratio $Y_{\nu \gamma}$ increases to $0.8$,
an eightfold increase from the case of PG~1553+113, which reflects the different efficiencies of photopion production in these cases.
Although $Y_{\nu \gamma}$ is close to unity,
the peak neutrino peak flux is higher than the peak $\gamma$-ray flux because
of the different spectral shapes of the two components. At the energy of the detected neutrino ($\sim 0.1$~PeV)
the model-derived flux is low, but lies within the 3~$\sigma$ error bars. This discrepancy is similar
to the one we found in previous cases, such as PG~1553+113, although the photopion production efficiency and $Y_{\nu \gamma}$
are both higher in the case of H~2356-309.
At first sight, this result seems contradictory.
However, one has to take also into consideration the differences in the
ratios of the observed $\gamma$-ray and neutrino fluxes. For example, in blazars PG~1553+113 and H~2356-309,
this ratio is close to and much less than unity, respectively. Thus, in the case of H~2356-309, the higher $Y_{\nu \gamma}$
is compensated by the lower ratio of observed $\gamma$-ray and neutrino luminosities.
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{1H1914_v2.eps}
\caption{SED of blazar H~1914-194 and the neutrino flux for the corresponding IceCube event (ID 22).
The data are obtained from ASI Science Data Center
using the SED builder tool. All types of symbols and lines are explained in the legend
above the plot.}
\label{h1914-194}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth]{1RXS_v2.eps}
\caption{SED of blazar 1 RXS~J054357.3-553206 and neutrino flux for the corresponding IceCube event (ID 19).
The data are obtained from ASI Science Data Center
using the SED builder tool. All types of symbols and lines are explained in the legend
above the plot.}
\label{rxs}
\end{figure*}
\subsubsection{1H~1914-194}
\label{h1914}
Blazar 1H~1914-194 is one of the two sources that are potential TeV emitters and
has the poorest MW coverage among the BL~Lacs of the sample (for details, see Sect.~\ref{case}).
Because of the inhomogeneity (in time domain) of the available MW observations,
we present only
an indicative fit, where we attempted to describe at best the Fermi/LAT data.
Being the least constrained case we chose parameters that ensure high
efficiency in photopion production, i.e. we present
the most optimistic, for neutrino emission, scenario.
We caution the reader that contemporaneous data in optical, X-rays and GeV $\gamma$-rays
may strongly constrain the model (see e.g. Mrk~421 and PG~1553+113)
and lead to different parameter values
from those listed in Table~\ref{tab-0}. The neutrino flux shown in Fig.~\ref{h1914-194}
should be therefore considered as an upper limit. We also
note that the non-detection of TeV $\gamma$-rays is not as important for the predicted
neutrino flux as the simultaneous coverage in the aforementioned energy bands. For example,
the peak neutrino energy is solely determined by the peak frequency
of the low energy component and the value of the Doppler factor (see equation (\ref{Ev})), while
the neutrino flux is related to the sub-TeV $\gamma$-ray flux (see Figs.~\ref{pg1553+113}, \ref{h2356-309} and Fig.~8 in PM15).
The SED we obtained is the most complex one in terms of the number of emission
components that appear in the MW spectrum.
Apart from the very low frequencies (below UV) where the primary leptonic component
contributes, the spectrum is comprised roughly of: proton synchrotron radiation (between UV and $\sim$keV),
synchrotron emission from Bethe--Heitler pairs (between $\sim$keV and $\sim$100 MeV) and synchrotron emission ($\lesssim 1$~TeV) from
secondary pairs produced through pion decay and $\gamma \gamma$ absorption of VHE $\pi^0$ $\gamma$-rays (see equation (14)
in \citealt{petromast12}).
In fact the role of $\gamma \gamma$ absorption is non-negligible in this case, since the unattenuated
$\pi^0$ component has a flux higher by a factor of
$\sim 2$ than the neutrino one (see dash-dotted grey curve). The attenuation of
VHE $\gamma$-rays initiates an EM cascade \citep{mannheim91} that
transfers energy from the PeV regime to lower energies. Our numerical approach does not
allow us to isolate multiple photon generations produced
in the electromagnetic cascade. Thus, strictly speaking, the pe and $ {\rm p}\pi$ components shown in Fig.~\ref{h1914-194}
are the result of synchrotron radiation not only from pairs produced through photohadronic interactions but
also from pairs produced in the EM cascade.
In this case, we find $Y_{\nu \gamma} \simeq 2$,
which is the highest value amongst the sample (see Table~\ref{tab-0}) and close
to the maximum value ($Y_{\nu \gamma}^{\max}=3$) that is allowed in this theoretical framework (see Sect. 3.2 in PM15).
The model-predicted neutrino flux is close to the observed flux (within the 1 $\sigma$ error bars), thus strengthening
the association of neutrino event 22 with this source. We have to note, however,
that the model-predicted neutrino flux might decrease, should more constraints on the SED be available in the future.
\subsubsection{1 RXS~J054357.3-553206}
1~RXS~J054357.3-553206 \ is the second HBL source that has not yet been detected in TeV energies and the most probable
counterpart of IceCube event 19 according to PR14.
The hybrid SED obtained with our model is shown in Fig.~\ref{rxs} and shares many features
with the model SED of PG~1553+113.
According to the discussion in Sect.~\ref{pg1553}, we can infer
that the observed $\gamma$-ray emission originates only partially
from photohadronic processes, since the model derived neutrino peak flux
is lower than the $\gamma$-ray one. In fact, we find that $Y_{\nu \gamma}=0.1$, same as in the case
of PG~1553+113. However, because the flux corresponding to neutrino event ID 19 is higher than the $\gamma$-ray one, the
model-predicted flux (at the same energy bin as event 19) is more than 3 $\sigma$ off from the observed value, and
the model prediction falls short of explaining the observed flux (see also Fig.~\ref{neutrinos}).
In contrast to PG~1553+113 the SED of blazar 1~RXS~J054357.3-553206 \ is less constrained both in (sub)MeV and TeV energies.
This, along with the large discrepancy between the model and observed neutrino fluxes, makes 1~RXS~J054357.3-553206 \ the best
case for studying the effects of a different parameter set on the neutrino flux.
We search, therefore, for parameters that increase the contribution of the $ {\rm p}\pi$ component to the observed
$\gamma$-ray emission and at the same time lead to a higher neutrino flux. We do not focus, however, on the goodness of the SED fit but we rather
try to roughly describe the observed spectrum.
The aim of what follows is (i) to demonstrate the flexibility of the leptohadronic model in describing the MW photon spectra and (ii)
to find parameters that maximize the predicted neutrino flux from the particular blazar, albeit
at the cost of a worse description of the SED.
Our results are summarized in Fig.~\ref{rxs-comp}, where we plot the
hybrid SED obtained for a different parameter set presented in Table~\ref{tab-1}.
From this point on, we will refer to it as Model~2.
For a better comparison, we over-plotted with dashed lines
the photon and neutrino spectra that were obtained for the parameters listed in Table~\ref{tab-0} (Model~1).
In Model~2 the $\gamma$-ray emission is mainly attributed to the $ {\rm p}\pi$ component and the neutrino flux is
comparable to the observed $\gamma$-ray one; in particular, we find $Y_{\nu \gamma} \simeq 3$.
In our framework, this is the most optimistic scenario and the derived neutrino flux can be considered as an upper limit on the flux
expected from this source (see also discussion in Sect.~\ref{h1914}).
However, the description of the SED is not as good as in our baseline model. In particular, the model-derived spectrum
(i) cannot explain the highest energy
Fermi/LAT data, (ii) is less hard (in $\nu F_{\nu}$ units) at GeV energies and more broad
than the observed one -- see also Sect.~\ref{pg1553}, and (iii) is marginally above the highest flux observation in X-rays.
This excess is caused by the
synchrotron emission from Bethe--Heitler pairs, which appears in
the hard X-ray/sub-MeV energy range, and it is even more luminous than the $ {\rm p}\pi$ component.
It is also noteworthy that the Bethe--Heitler emission in Model~2 plays a crucial role in
the neutrino production: if we were to neglect pair injection from the Bethe--Heitler process,
the neutrino flux in Model~2 would be lower by a factor of $\sim 10$.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{1RXS_comp.eps}
\caption{SED of blazar 1 RXS~J054357.3-553206 and neutrino flux for the corresponding IceCube event (ID 19).
The hybrid model SEDs obtained for two different parameter sets presented in Tables~\ref{tab-0} and \ref{tab-1} are respectively shown
with dashed and solid lines.
}
\label{rxs-comp}
\end{figure}
The detection of a luminous component in the 20~keV--20~MeV energy range
could, therefore, serve as an indirect probe of high-energy neutrino emission from a
blazar source within the leptohadronic scenario, as recently proposed in PM15.
\subsection{Neutrino emission}
\label{neutrino-emission}
\begin{figure*}
\centering
\includegraphics[width=0.6\textwidth, height=8cm]{neutrinos_c.eps}
\caption{Comparison of the model (lines) and the observed (circles) neutrino fluxes as defined in PR14 for
the six BL~Lacs of the sample.
The Poissonian 1 $\sigma$ error bars for each event, which are adopted by PR14, are also shown. We note that the respective 3 $\sigma$ error bars
correspond to -2.9, +0.9 dex \citep{gehrels86}.}
\label{neutrinos}
\end{figure*}
The model-derived electron and muon $\nu+\bar{\nu}$ spectra for
the six blazars of the sample are summarized in Fig.~\ref{neutrinos}. For a better comparison, we over-plotted
the fluxes that correspond to the respective neutrino events along with the Poissonian $1 \sigma$ error bars calculated for a single event
(PR14). We note that respective 3 $\sigma$ error bars correspond to -2.9, +0.9 dex \citep{gehrels86}.
We focus first on the neutrino event 9, which according to the model-independent analysis of PR14, has two plausible astrophysical counterparts;
namely
blazars Mrk~421 and 1ES~1011+496.
For Mrk~421, in particular, we show the neutrino spectrum depicted in the right panel of Fig.~\ref{mrk421}.
According to the discussion in Sect.~\ref{1es1011},
the neutrino spectrum for 1ES~1011+496 (dashed line) should be considered as an upper limit,
since inclusion of additional constraints in the SED fitting would
lower the neutrino flux by at least a factor of $\sim3$. Thus,
our results strongly favour Mrk~421 against 1ES~1011+496.
As already discussed in the respective paragraphs of Sect.~\ref{results}, the differences between the neutrino fluxes
originate from the differences in their SEDs. Thus, the case of neutrino ID 9 reveals in the best way how
detailed information from the photon emission may be used to lift possible degeneracies between multiple astrophysical counterparts.
1~RXS~J054357.3-553206 \ was selected by PR14 as the astrophysical counterpart of neutrino event 19. Figure \ref{neutrinos}
shows the respective neutrino spectrum (yellow line) we obtained for the best-fit model of its SED. The model-derived flux
at the energy of $\sim 0.2$~PeV lies below the 3 $\sigma$ error bars.
In this regard, 1~RXS~J054357.3-553206 \ may be excluded at the present time from being the astrophysical counterpart
of its associated neutrino event, namely event 19. We note, however, that a
higher neutrino flux (within the 3 $\sigma$ error bars)
can be obtained at the cost of a worse description of the SED (see Sect.~\ref{rxs}).
In all other cases, the model-derived\footnote{We refer to the neutrino emission calculated
for best-fit SED models.} neutrino flux at the energy bin of the detected neutrino
is below the 1 $\sigma$ error bars, but still within the 3 $\sigma$ error bars.
Thus, strictly speaking, the association of these sources cannot be excluded at the present time.
In particular, Mrk~421 and 1H~1914-194 are the two cases
with the smallest discrepancy between the observed and model-derived fluxes and, in this regard,
the most interesting, because their association with the respective IceCube events
can be either verified or disputed in the near future.
Besides the comparison between the model and the observed neutrino fluxes, Fig.~\ref{neutrinos} also
demonstrates the similarity of the model neutrino spectra, despite the fact that they were
obtained for BL~Lacs with different properties.
Our results suggest that the shape of the neutrino spectrum depends only weakly on
the $\gamma$-ray luminosity of the source, which varies approximately two orders
of magnitude among the members of the sample (see Table~\ref{tab-0}).
In particular, the neutrino spectrum may
be described as a power-law, i.e.
$E_{\nu}^2 dN_{\nu}/dE_{\nu} \propto E_{\nu}^{\alpha}$ with $\alpha\sim 1.1-1.3$.
Before closing this section, we must point out that
the present theoretical framework allows us to establish a
connection between the {\sl observed} $\gamma$-ray luminosity and the {\sl predicted} total neutrino luminosity.
In fact, we showed that these are simply proportional as $L_{\nu} = Y_{\nu \gamma} L_{\gamma, \rm TeV}$, with
$Y_{\nu \gamma}$ in the range $0.1-2$ (Table~\ref{tab-0}). Although
there is a maximum value predicted by the model, i.e. $Y_{\nu \gamma} \lesssim 3$ (see also PM15),
there is no restriction on the lowest possible value of this ratio.
Values of $Y_{\nu \gamma} \ll 1$ will only imply that the leptohadronic model simplifies into
a pure leptonic one, with the SSC emission of primary electrons being responsible for the observed
emission. This is illustrated best through the cases of PG~1553+113, 1ES~1011+496, and 1~RXS~J054357.3-553206.
\begin{table}
\centering
\caption{Parameter values used for the fit presented in Table~\ref{tab-0} (Model 1) and in the fiducial model described in text, which maximizes
the neutrino flux (Model~2) from blazar 1~RXS~J054357.3-553206. The symbols have the same meaning as in Table~\ref{tab-0}.}
\begin{tabular}{c c c}
\hline \hline
Parameter & Model~1 & Model~2 \\
\hline \hline
B & 0.1 & 1 \\
$R$ (cm)& $3\times 10^{16}$ & $3\times 10^{15}$ \\
$\delta$ & 31 & 50 \\
\hline
$\gamma_{\rm e, \min}$ & 1 & 1 \\
$\gamma_{\rm e, \max}$ & $1.2 \times 10^5$ & $2.5\times 10^4$ \\
$s_{\rm e}$ & 1.7 & 1.7 \\
$\ell_{\rm e, inj}$ & $2.5 \times 10^{-5}$ & $1.6 \times 10^{-5}$ \\
\hline
$\gamma_{\rm p, \min}$ & 1& 1\\
$\gamma_{\rm p, \max}$ & $1.2 \times 10^7$ & $1.2\times 10^6$\\
$s_{\rm p}$ & 2.0 & 1.9 \\
$\ell_{\rm p, inj}$ & $8 \times 10^{-4}$& $1.6 \times 10^{-2}$\\
\hline \hline
$L_{\gamma, \rm TeV}$~(erg/s)& $6.3\times 10^{45}$ & $1.6\times 10^{45}$ \\
$L_{\nu}$~(erg/s) & $6.3\times 10^{44}$ & $5\times 10^{45}$ \\
$Y_{\nu \gamma}$ & 0.1 & 3.1 \\
\hline
\end{tabular}
\label{tab-1}
\end{table}
\subsection{The ratio $Y_{\nu \gamma}$ and the energetics}
\label{ratio}
In this section we examine in more detail the
results presented in Table~\ref{tab-0}. We attempt to extract information about
the dependence of
the ratio $Y_{\nu \gamma}$, which is a crucial derivable quantity of our model,
on observables, such as the photon index in the Fermi energy range and the $\gamma$-ray luminosity $L_{\gamma, \rm TeV}$.
Moreover, using the parameter values of Table~\ref{tab-0}, we calculate the jet power for each source
and comment on the
energetics.
The results that follow, particularly those related to
trends and correlations, should be considered with caution because of the limited size of the sample. Although these cannot
be directly applied to a wider sample of BL~Lacs, they are useful in that they provide
a better understanding of the modelling results.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{Y_phindex.eps}
\includegraphics[width=0.48\textwidth]{Y_Lg.eps}
\caption{Top panel: Plot of the photon index in the Fermi/LAT energy range versus the ratio $Y_{\nu \gamma}$.
Sources with ratio values above and below the mean value $\bar{Y}_{\gamma \nu}=0.8$ are plotted with red and black symbols, respectively.
The arrow stresses the fact that the ratio derived for blazar 1ES~1011+496 is only an upper limit (see Sect.~\ref{1es1011}).
The values of the photon index are adopted from the 1FGL catalog (\citealt{abdo10}). Bottom panel: The observed
(0.01-1) TeV $\gamma$-ray luminosity plotted against the ratio $Y_{\nu \gamma}$ (logarithmic scale). The arrow and black (red) symbols
have the same meaning as
in the top panel. The dashed line is a linear fit ($\log L_{\gamma, \rm TeV} = A_1\log Y_{\nu \gamma} +A_0$) to the points
with $A_1=-1.1 \pm 0.4$, $A_0=45.2 \pm 0.2$.}
\label{Y_phindex}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{lB.eps}
\caption{Logarithmic plot of the proton (filled symbols) and electron (open symbols)
injection compactnesses as a function of the magnetic compactness $\ell_{\rm B}$ (for the definition, see text) for the
six BL~Lacs of the sample. Red (black) colored symbols correspond to ratio values $Y_{\nu \gamma}$ above (below)
the average value derived from the sample, i.e. $\bar{Y}_{\nu \gamma} =0.8$. The $\ell_{\rm p, inj}$ value
for 1ES~1011+496 is an upper limit, shown with an arrow.
The dashed arrow demonstrates the transition from
SSC to $ {\rm p}\pi$ dominated $\gamma$-ray emission. }
\label{lB}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{Lp_Le_v1.eps}
\caption{$L'_{\rm e, inj}$ versus $L'_{\rm p, inj}$ in logarithmic scale. The color coding is the same as in Fig.~\ref{lB}.
The solid and dashed lines have slopes equal to one and two, respectively, and are plotted for guiding the eye.}
\label{LpLe}
\end{figure}
In the top panel of Fig.~\ref{Y_phindex} we plot the photon index in the Fermi/LAT energy range
\citep{abdo10} versus the ratio $Y_{\nu \gamma}$.
Sources with ratio values
above and below the mean value of the sample $\bar{Y}_{\nu \gamma}=0.8$ are plotted with black and red symbols, respectively.
We remind the reader that for 1ES~1011+496 the ratio $Y_{\nu \gamma}=0.5$ was derived without including in our analysis
the upper limit to the hard X-rays. As discussed in Sect.~\ref{1es1011}, inclusion of the hard X-ray constraint would
lower the current ratio, at least by a factor of $\sim 3$; this is illustrated with an arrow in both panels of Fig.~\ref{Y_phindex}.
There is no evident trend between the ratio $Y_{\nu \gamma}$ and the
photon index in the GeV energy range, although in Sect.~\ref{pg1553} we used
the hardness of the $\gamma$-ray spectrum as an argument for choosing parameters that
favoured SSC against $ {\rm p}\pi$ radiation. These findings suggest that the Fermi photon index alone
is not a good indicator of the model-derived ratio $Y_{\nu \gamma}$. However,
we find a negative trend between
the observed $\gamma$-ray luminosity in the 0.01-1 TeV energy range (listed in Table~\ref{tab-0})
and the ratio $Y_{\nu \gamma}$ (bottom panel in Fig.~\ref{Y_phindex})\footnote{One could argue that
we were bound to find an anti-correlation, since we are plotting $L_{\gamma, \rm TeV}$ against the ratio $Y_{\nu \gamma}$
(note that $L_{\gamma, \rm TeV} = Y_{\nu \gamma}^{-1} L_{\nu}$). This would be, indeed, correct, if there was an {\sl a priori}
knowledge of the model-derived neutrino luminosity. In fact, there was no guarantee that the neutrino luminosity obtained
for all six BL~Lacs would be approximately constant (see Table~\ref{tab-0}), which is an interesting outcome of the present analysis.}.
The size of our sample is limited, yet
our results suggest that the contribution of the $ {\rm p}\pi$ component to the blazar's $\gamma$-ray emission
is smaller in sources that are more $\gamma$-ray luminous.
The dependence of $Y_{\nu \gamma}$ on the $\gamma$-ray luminosity
is particularly useful if we attempt to
calculate the BL~Lac diffuse neutrino flux by extrapolating
the findings of this sample to the wider class of BL~Lacs; this will be part of a future study.
Finally, we point out that the negative trend shown in Fig.~\ref{Y_phindex}
does not necessarily mean that less powerful BL~Lacs in $\gamma$-rays inject more power in relativistic protons, as it
will become evident at the end of this section.
The bottom panel of Fig.~\ref{Y_phindex} provides the first hints of a separation between the sources of the sample with ratios
above and below the average value. However, this becomes clear in Fig.~\ref{lB}, where we plot
the proton (filled symbols) and electron (open symbols) injection compactnesses
against the magnetic compactness defined as $\ell_{\rm B}= \sigma_{\rm T} R B^2 / 8 \pi m_{\rm e} c^2$.
Red and black colored symbols are used to denote values of $Y_{\nu \gamma}$ above and below the mean value, respectively.
Figure~\ref{lB} reveals the presence of two sub-groups within the sample. As $\ell_{\rm B}$ increases,
the synchrotron cooling of both primary and secondary electrons is enhanced. On the one hand, this
leads to an increase of the synchrotron emission from pairs produced by $ {\rm p}\pi$ interactions that emit in the $\gamma$-ray regime,
while on the other hand, it suppresses the SSC emission of primary electrons. Thus, as $\ell_{\rm B}$ increases, the contribution
of the $ {\rm p}\pi$ component to the $\gamma$-ray emission increases against the SSC one. This is finally reflected in
an increase of the ratio $Y_{\nu \gamma}$.
In Fig.~\ref{LpLe} we plot $L_{\rm e}^{'\rm inj}$ as a function of $L_{\rm p}^{'\rm inj}$.
The color coding is the same as in Fig.~\ref{lB} and the arrow stresses the fact
that the derived value for 1ES 1011+496 is an upper limit (see Sect.~\ref{1es}).
The injection luminosities of primary particles, as measured in the comoving frame, follow a more than
linear relation, i.e.
$L_{\rm e}^{' \rm inj} \propto \left(L_{\rm p}^{'\rm inj} \right)^x$
with $1 \le x \le 2$. More important, though, is that
sources with $Y_{\nu \gamma} \ge 0.8$ lie at the lower left corner of the plot, i.e. they are described
by low proton and electron luminosities (in the comoving frame) with respect to the BL~Lacs with smaller $Y_{\nu \gamma}$.
The jet power is a useful quantity when it comes to the energetics of the model.
In the one-zone leptohadronic model, the energy densities of
both relativistic electrons ($u'_{\rm e}$) and protons ($u'_{\rm p}$), as well as the energy density of the
magnetic field ($u'_{\rm B}$) and radiation ($u'_{\rm r}$), contribute to the jet power, which may be written as
\begin{eqnarray}
L_{\rm j} \simeq \pi R^2 \delta^2 \beta c \left(u'_{\rm e} + u'_{\rm p} + u'_{\rm B} + u'_{\rm r}\right),
\label{jet1}
\end{eqnarray}
where the energy densities of the various components are measured at steady state and in the comoving frame, $\delta \approx \Gamma$,
and $\beta = \sqrt{1-1/\Gamma^2} \simeq 1$. For the assumptions used to derive equation~(\ref{jet1}), see e.g. \cite{celottighisellini08}.
The jet luminosity as
well as $u'_{\rm i}$ are listed, for reference, in Table~\ref{tab-2}.
For all sources we find
high jet luminosities, and emission regions far from equipartition $u'_{\rm p}/u'_{\rm B} \gg 1$, with the proton component being
the dominant one. These results are typical of leptohadronic models (e.g. \cite{boettcherreimer13, cerrutietal14}) and often constitute a
point for criticism.
However, our analysis highlighted that sources with $Y_{\nu \gamma} \sim 1$, namely with significant contribution of the photohadronic component to
the $\gamma$-ray emission, are described by less extreme, i.e. $\lesssim 10^{49}$~erg/s, jet powers and have lower $u'_{\rm p}/u'_{\rm B}$ ratios;
this was not expected beforehand. For the rest of the sources, i.e. those with $Y_{\nu \gamma} \ll 1$, the derived jet luminosities exceed
$10^{49}$~erg/s, mainly because of the larger $R$ compared to the other sources (see Table~\ref{tab-2}).
\begin{table*}
\centering
\caption{Jet luminosity as defined in equation~(\ref{jet1}) and (comoving) energy densities
of the proton, electron, radiation and magnetic components.}
\begin{tabular}{l c c c c c}
\hline \hline
Source & $L_{j}$ (erg/s) & $u'_{\rm p}$ (erg/cm$^3$) & $u'_{\rm e}$ (erg/cm$^3$) & $u'_{\rm r}$ (erg/cm$^3$) & $u'_{\rm B}$ (erg/cm$^3$)\\
\hline \hline
Mrk~421 & $2\times 10^{48}$ & $4\times 10^3$ & $4\times 10^{-3}$& $4.9\times 10^{-2}$ & $9.9\times 10^{-1}$ \\
PG~1553+113 & $3.7 \times 10^{49}$ & $1.1 \times 10$ & $4 \times 10^{-4}$ & $2.3 \times 10^{-4}$ & $9.9 \times 10^{-5}$ \\
1ES~1011+496 & $8.3 \times 10^{49}$ & $9\times 10^2$ & $6.6 \times 10^{-3}$ & $5.3 \times 10^{-4}$ & $4 \times 10^{-4}$ \\
H 2356-309 & $4.4 \times 10^{48}$ & $5.7 \times 10^3$ & $1.4 \times 10^{-3}$ & $1.6 \times 10^{-2}$ & $9.9\times 10^{-1}$ \\
1H 1914-194 & $8 \times 10^{48}$ & $2.9 \times 10^4$ & $7.4 \times 10^{-2}$ & $8.2 \times 10^{-2}$ & $9.9\times 10^{-1}$\\
1~RXS~J054357.3-553206 & $1.5 \times 10^{49}$ & $1.8 \times 10^2$ & $3.3 \times 10^{-3}$ & $7\times 10^{-4}$ & $4 \times 10^{-4}$ \\
\hline
\end{tabular}
\label{tab-2}
\end{table*}
\section{Discussion}
\label{discussion}
In the present paper we have applied a one-zone leptohadronic model to six BL~Lacs, which
have been recently selected as the probable astrophysical counterparts (PR14) of five of the
IceCube neutrino events \citep{aartsen14}, and calculated the neutrino flux in each case.
In contrast to other studies, where the neutrino emission is calculated under the assumption
of a generic blazar SED, here we obtained the neutrino signal from BL~Lacs by fitting
their individual SEDs.
Thus, for each source, the photon and particle distributions used to
derive the neutrino emission have to be determined
individually, so as to give the observed SEDs.
For the SED modelling, we have adopted one of the possible leptohadronic model variants that allows
us to directly associate the $\gamma$-ray blazar emission with a high-energy ($2-20$~PeV)
neutrino signal. In particular, the low-energy hump of the SED is explained by synchrotron radiation of primary
relativistic electrons, whereas the observed high-energy (GeV-TeV) emission is the combined result
of synchrotron self-Compton emission (from primary electrons) and of
synchrotron radiation from secondary pairs.
These are produced in a variety of ways, such as in Bethe--Heitler and $\gamma \gamma$ pair production, and through
the decays of charged pions, which themselves are the by-product of $ {\rm p}\pi$ interactions
of co-accelerated protons with the internally produced synchrotron photons.
Secondary pairs produced in different processes have, in general, different
energy distributions. Thus, in our framework, it is mainly the synchrotron emission
of $ {\rm p}\pi$ pairs that falls in the $\gamma$-ray regime, and in combination with the SSC radiation
is used to explain the observed blazar emission in $\gamma$-rays.
BL~Lacs are known for their variability across the electromagnetic spectrum and, because of this,
we used SEDs that are comprised of nearly simultaneous MW observations. For those sources that were
targets of recent MW observing campaigns we have adopted the data from the respective publications.
In any other case,
we have compiled the SEDs using the SED builder tool of the ASI Science Data Centre (ASDC).
Only for one source in the sample (1H~1914-194), which had the poorest simultaneous MW coverage, had we to
construct an SED using non-simultaneous data.
We used the numerical code described in DMPR12 in order to calculate steady-state photon spectra
that were later applied to the MW observations.
Because of the multi-parameter nature of the problem (11-13 free parameters) and of the long numerical simulation time
required for every parameter set, we have not performed a blind search of the available parameter space.
To obtain a first estimate of the required parameter set,
we used instead the analytical expressions presented in PM15.
We then performed a series of numerical simulations
using parameter values lying close to the initial parameter set,
until a reasonably good fit to each SED was obtained.
Although the model-derived neutrino emission
from the six BL~Lacs that were selected as counterparts of the IceCube events
was the main focus of this paper, our study highlighted also
the flexibility of the leptohadronic model and its success in describing the blazar SEDs.
This is particularly important, if one takes into account the following:
(i) all sources in the sample were fitted using
reasonable parameter values; (ii) these BL~Lacs were not {\sl a priori} selected for modelling purposes;
and (iii) besides the fact that all sources in the sample were HBLs,
they differed in many aspects, e.g. in redshift, in $\gamma$-ray luminosity and spectra.
The only source that may be excluded, at the present time, from being the astrophysical counterpart of the detected neutrino (event 19)
is 1~RXS~J054357.3-553206. \ In all other cases, the model-predicted neutrino flux is
below the 1~$\sigma$ but still within the 3~$\sigma$ error bars of the respective IceCube neutrino events.
In particular, for blazars Mrk~421 and 1H~1914-194
the model-derived neutrino fluxes at the same energy with the respective neutrino events (ID 9 and 22, respectively)
were found to be very close to the low 1~$\sigma$ error bars (Fig.~\ref{neutrinos}). Although
our results cannot be considered a proof of a firm association between these BL~Lacs and the respective neutrinos,
they can be verified or disputed in the near future, as IceCube collects more data and the observed fluxes become
more constraining for the model.
Among the blazars of the sample, Mrk~421 is a special case. Its neutrino emission was calculated
in a previous study (DPM14), at a time when there was no hint of a possible association between Mrk~421 and the IceCube events.
DPM14 found that the neutrino flux from Mrk~421 was close
to the current IceCube (IC-40) sensitivity limit for this source (left panel in Fig.~\ref{mrk421}). Of particular interest are
the following:
\begin{itemize}
\item the prediction of DPM14 was found {\sl a posteriori} to be close to the observed flux related to neutrino ID 9;
\item the same parameter set, besides the Doppler factor, used in DPM14 to fit the optical/X-ray/TeV data of 2001, has been successfully
(right panel in Fig.~\ref{mrk421}) applied to the more complete dataset of the 2009 campaign \citep{abdo10}.
\end{itemize}
Apart from a direct comparison of the model and observed neutrino fluxes, we have shown that we can establish
a direct connection between the observed $\gamma$-ray emission and the putative neutrino emission
from a BL~Lac source. We have quantified this relation by introducing
the ratio of the total neutrino to the $0.01-1$ TeV luminosity ($Y_{\nu \gamma}$).
In our model, for a given observed $\gamma$-ray luminosity that
is purely explained by the synchrotron emission of $ {\rm p}\pi$ pairs, i.e.
when the SSC contribution is negligible, there is an upper limit to the neutrino
luminosity. It can be shown that the limit is $\sim 3 L_{\gamma, \rm TeV}$ (see PM15).
If, however, the SSC component is dominant in the $\gamma$-ray regime, we expect $Y_{\nu \gamma} \rightarrow \epsilon \ll 1$; in this case,
our model simplifies into a leptonic one. Thus, the spectrum of possible values is
$[\epsilon \ll 1, 3]$ and corresponds to different contributions of the $ {\rm p}\pi$ and SSC components to the $\gamma$-ray emission.
For the six blazars we have found ratios in the range $0.1 \lesssim Y_{\nu \gamma}\lesssim 2$, with a mean value
$\bar{Y}_{\nu \gamma} \simeq 0.8$. We have identified two sub-groups
in the sample: blazars Mrk~421, H~2356-309 and 1H~1914-194 with $Y_{\nu \gamma} \ge \bar{Y}_{\nu \gamma}$ and the rest
with ratios below the average value (Table~\ref{tab-0}).
Despite the small size of our sample, we have investigated the dependence of this ratio on observables of blazar emission.
No hints of correlation with the photon index in the Fermi/LAT energy range were found (Fig.~\ref{Y_phindex}, top panel).
However, we showed that there is a negative trend between $L_{\gamma, \rm TeV}$ and $Y_{\nu \gamma}$, i.e. more $\gamma$-ray luminous BL~Lacs
have a smaller photohadronic contribution to their $\gamma$-ray spectra (bottom panel in Fig.~\ref{Y_phindex}).
Although this could be interpreted as a lower injection luminosity of relativistic protons, we showed
that this is not the reason (see Fig.~\ref{LpLe} and Table~\ref{tab-2}). Instead,
the separation of sources according to $Y_{\nu \gamma}$ can be understood
in terms of the magnetic compactness. In Fig.~\ref{lB} we demonstrated the transition
from SSC to $ {\rm p}\pi$ dominated $\gamma$-ray emission as the
magnetic compactness gradually increases, which also translates to an increase of $Y_{\nu \gamma}$.
Application of the model to a larger sample of BL~Lacs is, however, necessary for
a better understanding of the aforementioned trends.
Our analysis showed that the shape of the neutrino spectrum does not
strongly depend on the $\gamma$-ray luminosity. In all cases,
the neutrino spectra are described as $dN_{\nu}/dE_{\nu} \propto E_{\nu}^{\alpha -2}$
with $\alpha \simeq 1.1-1.3$ (Fig.~\ref{neutrinos}). The peak energy of the neutrino spectrum, in the present context,
depends only on one model parameter, namely the Doppler factor, and two observables, i.e.
the peak frequency ($\nu_{\rm s}$) of the low-energy hump of the SED and the redshift of the source $z$. The dependence on the latter is not
strong.
Since all the sources of the sample are HBLs and we have obtained fits with similar values of $\delta$, we should
not expect large differences in the peak neutrino energy. We have verified this numerically, as demonstrated
in Fig.~\ref{neutrinos}. In addition to the similarity in the spectral shape, we have
connected the expected neutrino luminosity from a BL~Lac with
its $\gamma$-ray luminosity in the 0.01-1 TeV energy range. Thus, one could go one step further and
calculate the diffuse neutrino flux from BL~Lacs, in order to see how it compares
with the current IceCube detections and upper limits in the sub-EeV regime.
Such a calculation requires an extrapolation of our
findings to the wider class of BL~Lacs, including low-synchrotron peaked blazars (LBLs), and the use
of detailed simulations of the blazar population of the type done by \cite{padovanigiommi15}.
This will be the focus of a future study.
Similarly to other leptohadronic models, such as the proton synchrotron model (e.g. \citealt{boettcherreimer13}),
the derived jet power is high, ranging from $10^{48}$~erg/s up to a few $10^{49}$~erg/s, which
implies a very low radiative efficiency.
Moreover, the emitting region is far from equipartition, with the proton energy density being the dominant one --
see Tables~\ref{tab-2} in this work and
\cite{boettcherreimer13}. The demanding energetics and the large departure from
equipartition constitute the main point for criticism of our model, and of leptohadronic models in general.
However, there are some important differences
compared to the proton synchrotron model, which make this leptohadronic variant (LH$\pi$) more flexible.
On the one hand, strong magnetic fields ($\gg 10$~G) are not necessary, while, on the other hand,
proton acceleration to ultra-high energies ($\gamma_{\rm p} \sim 10^9-10^{10}$) is not required. On the contrary, the upper cutoff of
the proton distribution should be just a few times higher than $\sim 3.5 \times 10^7 (1+z)^{-1} \delta_1 \nu_{\rm s, 16}^{-1}$
(see equation~(\ref{eq2})). In terms of the available parameter space of leptohadronic models,
the LH$\pi$ variant occupies a sub-region, which corresponds, roughly speaking, to: low-to-moderate magnetic field strengths,
moderate maximum proton energies, sub-pc sizes of the emission region ($\sim 10^{15}-10^{16}$~cm) and high jet powers.
A detailed calculation of the neutrino emission from blazars following the so-called ``blazar sequence'' \citep{ghisellinietal98} has been
presented in two recent studies by \cite{murase14} and \cite{dermermurase14}. Our analysis showed, however, that
relatively high neutrino luminosities (of the order of $\sim 10^{45}$~erg/s) can be obtained from classical BL~Lacs, with
Mrk~421 being the most notable case. We note also that the $ {\rm p}\pi$ production efficiency is low in all six blazars
(see Table~\ref{tab-0}), and is in agreement
with the expected values from BL~Lacs (see equation~(20) in \citealt{murase14} and equation~(28) in PM15).
The question is then whether and how those results are compatible with each other. Figure~9 in \cite{murase14} shows
that the typical neutrino luminosity of individual BL~Lacs with no contamination by external photon fields (two lower blue curves)
are $\sim 10^{42}$~erg/s, which is approximately 3 orders of magnitude below our typical values. The reason behind
this discrepancy is the luminosity that is being injected into relativistic protons. In our analysis, we systematically
used 2 to 3 orders of magnitude higher proton injection luminosities than \cite{murase14}. This is mainly due to
the fact that we try to model the observed blazar $\gamma$-ray emission in terms of photohadronic processes,
whereas in \cite{murase14} and \cite{dermermurase14} it is not clear what the fate of the photons produced via photohadronic processes is, and why
they do not appear in the observed SED.
In any case, we compensate for the low number density of internal radiation by pushing the proton injection luminosity to higher values.
Another way to go around the problem of low photon number density in BL~Lacs, which is related to low neutrino production efficiency,
has been also recently presented in \cite{tavecchioetal14}. Their approach is different than ours, though. In their model, BL~Lacs can still be efficient neutrino
emitters if the number density of photons, which are the targets for $ {\rm p}\pi$ interactions, is enhanced.
In order to achieve this, \cite{tavecchioetal14} assume
a structured BL~Lac jet with a fast spine and slow layer
\citep{ghisellinietal05, tavecchioghisellini08}, where the target field for $ {\rm p}\pi$ interactions
is the radiation field from the layer that appears boosted in the rest frame of the spine.
The recent correlation analysis of IceCube neutrino events and $\gamma$-ray sources by PR14 led to a somewhat
unexpected result: instead of the more powerful FSRQs, it was BL~Lacs that
were among the most plausible astrophysical counterparts of 9 IceCube neutrinos.
In general, FSRQs are expected to be efficient PeV neutrino emitters, because of their higher $\gamma$-ray
luminosity, but more importantly, because of the dense external photon fields (broad line region and/or dusty torus).
The absence of FSRQs from the list of probable astrophysical counterparts may be
related to the ``energetic criterion'' applied by PR14. This was totally empirical and observationally
driven, since it assumed only a
tight photon -- neutrino connection; any other more sophisticated choice of a diagnostic would be model-dependent.
Thus, sources where a simple extrapolation of
the $\gamma$-ray flux to the PeV energies lied much below the observed neutrino flux (taking into account
the uncertainty in the flux measurement) were discarded. Relaxing this constraint would certainly
increase the number of candidates, which might include FSRQs, but these would
then depend on the specific model assumed.
In principle, the leptohadronic variant we presented here can also be applicable to
FSRQs, with the appropriate choice of parameters. Because of the higher $\gamma$-ray luminosities and lower peak energies of FSRQs with respect
to HBLs, a straightforward way to apply our model
to FSRQs would be to increase the proton injection luminosity, while lowering the maximum proton energy. Taking into account the presence of additional
photons originating from the broad line region and/or the torus, the model-predicted (PeV) neutrino luminosity would be much higher than the one
derived for the HBLs in our sample and possibly close to, or at, the limit of flux detected by IceCube.
The application of the model to the SEDs of FSRQs, however, may not be as simple as described above. Higher $L'_{\rm p, inj}$
leads not only to increased emission from $ {\rm p}\pi$ interactions but also from Bethe-Heitler interactions. Given its spectral shape
and the typical energy range that emerges (see Figs.~\ref{mrk421}-\ref{rxs-comp}), it would be difficult to reconcile its emission
with the flat X-ray spectra of FSRQs.
However, more concrete answers can be given only with a detailed fitting of these type of sources.
\section{Summary}
\label{summary}
Following the identification of eight BL Lac objects as likely
sources of IceCube neutrinos by PR14, we set out to deduce the neutrino
emission of the ones whose redshift is known by applying a leptohadronic
model, using a variant of the model first discussed in \cite{petromast12}, albeit in a
more general context. We used data
from near simultaneous MW observations for all sources (as far as possible), and fitted the
resulting SEDs in each individual case with steady-state photon spectra, which
were obtained with the DMPR12 numerical code.
Through the fitting procedure we determined the distributions of injected electrons and
protons in what we assumed were spherical volumes of certain radii and magnetic field strengths,
moving towards us with certain Lorentz factors; all the above being treated as free parameters.
Of the six BL Lacs with known redshift that we studied, one of them, Mrk~421,
had been considered for its potential as a neutrino source in the past, by DPM14;
the present study verifies that the neutrino event associated with it (ID 9) falls within
its predicted flux. A good match between modeled and observed neutrino (ID 22) flux was also found for
1H~1914-194. 1~RXS J054357.3-553206, on the other hand, was the only BL lac to be confidently excluded as a
source of its associated neutrino (ID 19), with the remaining three BL Lacs displaying some discrepancy between modeled
and observed fluxes, but not enough to rule them out as sources at this stage. Depending on the shape of their
SEDs, we found that some cases favour fits dominated by SSC while others favour fits dominated
by photohadronic interactions,
with the former having a lower ratio of neutrino to $0.01-1$ TeV $\gamma$-ray luminosities than the latter.
This ratio, $Y_{\nu\gamma}$, which could prove important for calculations of the diffuse neutrino
emission in the future, was then evaluated for its dependence on observable and model
parameters. Ultimately, the results of our model should be confirmed or disproven in the near
future, with the accumulation and data analysis of additional IceCube observations. Whatever the
outcome, it will be important for constraining the available
parameter space for leptohadronic models of blazar emission.
This illustrates the importance of multi-messenger high-energy astronomy.
\section*{Acknowledgments}
MP was supported by NASA
through Einstein Postdoctoral
Fellowship grant number PF3~140113 awarded by the Chandra X-ray
Center, which is operated by the Smithsonian Astrophysical Observatory
for NASA under contract NAS8-03060.
ER acknowledges the support of the Deutsche Forschungsgemeinschaft (DFG)
\bibliographystyle{mn2e}
|
\section{Introduction}
\label{ch:introduction}
The apparent unification of the Standard Model (SM) gauge couplings at the scale
$M_{{\rm GUT}} \simeq 2\times10^{16}$~GeV in the presence of low scale supersymmetry(SUSY)~\cite{book}
is nicely consistent with supersymmetric grand unified theories (GUTs), but it does not significantly
constrain the sparticle spectrum.
On the other hand, imposing $t\mhyphen b\mhyphen\tau$ Yukawa coupling unification (YU) condition at $M_{\rm GUT}$~\cite{yukawaUn}
can place significant constraints on the supersymmetric spectrum
in order to fit the top, bottom and tau masses~\cite{Baer:2012cp,Gogoladze:2012ii,Gogoladze:2010fu}.
Most work on $t\mhyphen b\mhyphen\tau$
YU condition
has been performed in the framework of gravity mediation SUSY breaking~\cite{Chamseddine:1982jx}.
Some well known choices for GUT-scale boundary conditions on the soft supersymmetry breaking (SSB) terms
yield, with
$t\mhyphen b\mhyphen\tau$ YU condition, quite severe constraints on the sparticle
spectrum~\cite{Baer:2012cp,Gogoladze:2012ii,Gogoladze:2010fu,Blazek:2001sb,Auto},
which is further exacerbated after the discovery of the SM-like Higgs boson with mass
$m_h \simeq 125 - 126$~GeV ~\cite{mhatlas,mhcms}.
Models with gauge mediated SUSY breaking (GMSB) provide a compelling
resolution of the SUSY flavor problem, since the SSB terms are generated by the flavor blind gauge
interactions~\cite{Dine:1993yw,Giudice:1998bp,Meade:2008wd}.
In both the minimal~\cite{Dine:1993yw} and general~\cite{Meade:2008wd} GMSB scenarios, the trilinear SSB A-terms
are relatively small at the messenger scale, even if an additional sector is added to generate
the $\mu/B\mu$ terms~\cite{Komargodski:2008ax}.
Because of the small A-terms, accommodating the light CP-even Higgs boson mass around 125~GeV requires a stop mass in the
multi-TeV range. This, in turn, pushes the remaining sparticle mass spectrum to values that are out of reach
of the 14~TeV LHC~\cite{Ajaib:2012vc,mh125spart}. There is hope, however, that the predictions can be tested at 33~TeV High-Energy LHC (HE-LHC).
This tension between the Higgs mass and the sparticle spectrum can be relaxed if we assume the
existence of low scale vector-like particles that provide significant contribution to the
CP-even Higgs boson mass~\cite{Babu:2008ge,Martin:2012dg}. We will not consider this possibility here.
In this paper we investigate YU in the framework of minimal GMSB (mGMSB) scenario.
Here, the decoupled messenger fields generate SSB terms of particular non-universal pattern. We know from previous studies
in gravity mediation SUSY breaking scenarios that non-universal SSB terms are necessary to achieve successful YU \cite{Gogoladze:2010fu}.
Moreover, below $M_{\rm GUT}$, the mGMSB model contains messenger fields that modify the evolution of the gauge and Yukawa
couplings, thereby affecting YU.
We consider mGMSB in which the messenger fields reside in $5+\bar{5}$ of $SU(5)$, which is the simplest
scenario that preserves unification of the MSSM gauge couplings.
We note that previous studies of mGMSB only considered evolution below the messenger scale, which is usually much lower
than $M_{\rm GUT}$, and therefore could not address the question of YU.
The remainder of this paper is organized as follows. In Section~\ref{sec:gmsb} we briefly outline the mGMSB
framework that we use for our analysis. We describe our scanning procedure along with the experimental constraints
we applied in Section~\ref{sec:constraints}.
In Section~\ref{sec:results} we present our results, focusing in particular on the
sparticle mass spectrum. The table in this section presents some benchmark points which summarize
the prospects for testing these predictions at the LHC. Our conclusions are presented in Section~\ref{conclusions}.
In Appendix~\ref{sec:appnd},
we briefly discuss the renormalization group equations (RGEs) from the messenger scale to $M_{{\rm GUT}}$.
\section{Minimal Gauge Mediation}
\label{sec:gmsb}
In GMSB models supersymmetry breaking is communicated from a hidden sector to the superpartners of SM particles
via some messenger fields~\cite{Giudice:1998bp}. In the minimal version, the messengers interact with the visible sector only via
SM gauge interactions and to the hidden sector through an arbitrary Yukawa coupling. The {theory is described by} the superpotential
\begin{equation}
W = W_{MSSM}+\lambda \hat{S} \Phi \bar{\Phi},
\label{superpot}
\end{equation}
where $W_{MSSM}$ is the usual superpotential of MSSM, $\hat{S}$ is the hidden sector gauge singlet chiral superfield, and $\Phi$, $\bar{\Phi}$ are the messenger fields.
In order to preserve perturbative gauge coupling unification, the messenger fields
are usually taken to form complete vector-like multiplets of $SU(5)$. Thus one could have $N_5$ (1 to 4) pairs of $5+\bar{5}$,
or a single pair of $10+\overline{10}$, or the combination $5+\bar{5}+10+\overline{10}$.
For simplicity, we only consider the case with $N_5$ (one to four) pairs of $5+\bar{5}$ vector-like multiplets.
The singlet field $\hat{S}$ develops non-zero VEVs for both its scalar and auxiliary components,
$\vev{\hat{S}}=\vev{S} +\theta^2 \vev{F}$, from the hidden sector dynamics.
This results in masses for the bosonic and fermionic components of the messenger superfields,
\begin{equation}
m_b = M_{\rm mess}\sqrt{1\pm\frac{\Lambda}{M_{\rm mess}}},\quad m_f = M_{\rm mess}\, ,
\end{equation}
where $M_{\rm mess}=\lambda \vev{S}$ is the messenger scale, and $\Lambda = \vev{F}/\vev{S}$ sets the scale of the SSB terms.
Note that $\vev{F} < \lambda\vev{S}^2$ to avoid tachyonic messengers, and in {many} realistic cases
$\vev{F} \ll \lambda\vev{S}^2$.
Below the messenger scale $M_{\rm mess}$, the fields $\Phi$ and $\bar{\Phi}$ decouple generating SSB masses for
MSSM fields.
The masses of the MSSM gauginos are generated at $M_{\rm mess}$ from 1-loop diagrams with messenger fields. In the
approximation $\vev{F} \ll \lambda\vev{S}^2$,
the gaugino masses are given by
\begin{equation}
M_i(M_{\rm mess})=\frac{\alpha_i}{4\pi} N_5 \Lambda,
\label{BCgaugino}
\end{equation}
where $i=1,2,3$ stand for $U(1)_Y$, $SU(2)_L$ and $SU(3)_c$ sectors, respectively.
Since the MSSM scalars do not couple directly to the messengers, their
SSB masses are induced at two loop level:
\begin{equation}
m^2_\phi(M_{\rm mess})=2 N_5 \Lambda^2 \sum_{i=1}^{3} C_i(\phi) \left( \frac{\alpha_i}{4\pi}\right)^2,
\label{BCscalar}
\end{equation}
where $C_i(\phi)$ is the appropriate quadratic Casimir associated with the MSSM scalar field $\phi$.
The trilinear A-terms only arise at two-loop level, and thus are vanishing at the messenger scale.
However, they are generated below the messenger scale from the RGE evolution.
{ The bilinear SSB term also vanishes at $M_{\rm mess}$, although it is often ignored. We do not impose the relation $B=0$,
anticipating that the value needed to achieve the electroweak symmetry breaking (EWSB)
can be explained by some other mechanism operating at the messenger scale~\cite{Komargodski:2008ax}.}
The mGMSB spectrum is therefore completely specified by the following set of parameters:
\begin{equation}
M_{\rm mess},\ \Lambda,\ N_5,\ \tan\beta,\ sign(\mu),\ c_{grav}\, ,
\label{para}
\end{equation}
where $\tan\beta$ is the ratio of the two MSSM Higgs doublet VEVs.
The magnitude of $\mu$ is determined by
EWSB condition at the weak
scale. The parameter $c_{grav} (\ge 1$) affects the decay rate of the sparticles into gravitino, but it does not
affect the MSSM spectrum, and we set it equal to unity.
It is worth noting that the gaugino masses in mGMSB are non-universal at the messenger scale (if $M_{\rm mess}<M_{\rm
GUT}$). However, identical relations at the same scale among the gauginos can be obtained through RGE evolution in the gravity mediated SUSY breaking framework starting with universal
gaugino masses at $M_{\rm GUT}$.
On the other hand, Eq.~(\ref{BCscalar}) implies a universal
SSB mass terms for the up ($H_{u}$) and down ($H_{d}$) Higgs doublets at the messenger scale.
Within the gravity mediated framework this relation can be obtained at the same scale if one imposes
non-universal Higgs masses $m_{H_u}^2< m_{H_d}^2$ at $M_{\rm GUT}$.
Indeed, this non-universality is a necessary condition for compatibility with $t\mhyphen b\mhyphen\tau$ YU in
the gravity mediation case with universal gaugino masses at $M_{\rm GUT}$~\cite{Blazek:2001sb,Auto}.
However, the relation $m_{H_u}^2 < m_{H_d}^2$ is introduced by hand
in gravity mediated SO(10) GUT,
while $t\mhyphen b\mhyphen\tau$ YU in mGMSB can be achieved
without such {\it ad hoc} conditions, as we will see later.
From Eq.~(\ref{BCscalar}) the sfermion masses in mGMSB are non-universal at the messenger scale,
but it is not easy to find the same pattern at $M_{\rm mess}$ within the gravity mediation scenario.
Hence, the sfermion sector could be a good place to see differences between
gravity and gauge mediation SUSY breaking scenarios that are compatible with YU condition at the GUT scale.
\section{Scanning Procedure and Experimental Constraints}
\label{sec:constraints}
For our scan over the fundamental parameter space of mGMSB, we employed ISAJET~{7.84} package~\cite{ISAJET}
that we modified to include the RGE evolution above $M_{\rm mess}$ in MSSM with $N_5$ pairs of $5+\bar{5}$ messengers.
In this package, the weak-scale values of gauge and third generation Yukawa
couplings are evolved from $M_Z$ to $M_{\rm mess}$ via the MSSM RGEs in the $\overline{DR}$ regularization scheme.
Above $M_{\rm mess}$ the messenger fields are present, which changes the beta-functions for the SM gauge and Yukawa couplings.
The modified RGEs between $M_{\rm mess}$ and $M_{\rm GUT}$ are given in Appendix~\ref{sec:appnd}.
$M_{{\rm GUT}}$ is defined as the scale at which $g_1=g_2$ and is approximately equal to $2\times 10^{16}$~GeV,
the same as in MSSM.
We do not strictly enforce the unification condition $g_3=g_1=g_2$ at $M_{\rm GUT}$, since a few percent deviation
from unification can be accounted by unknown GUT-scale threshold corrections~\cite{Hisano:1992jj}.
For simplicity, we do not include the Dirac neutrino Yukawa coupling
in the RGEs, whose contribution is expected to be small.
The SSB terms are induced at the messenger scale and we set them according to expressions (\ref{BCgaugino}) and
(\ref{BCscalar}). From $M_{\rm mess}$ the SSB parameters, along with the gauge and Yukawa couplings, are evolved
down to the weak scale $M_Z$.
In the evolution of Yukawa couplings the SUSY threshold
corrections~\cite{Pierce:1996zz} are taken into account at the
common scale $M_{\rm SUSY}= \sqrt{m_{\tilde t_L}m_{\tilde t_R}}$,
where $m_{\tilde t_L}$ and $m_{\tilde t_R}$
denote the soft masses of the left and right-handed top squarks.
The entire parameter set is iteratively run between $M_Z$ and $M_{\rm
GUT}$ using the full 2-loop RGEs until a stable solution is
obtained. To better account for leading-log corrections, one-loop
step-beta functions are adopted for the gauge and Yukawa couplings, and
the SSB parameters $m_i$ are extracted from RGEs at multiple scales:
at the scale of its own mass, $m_i=m_i(m_i)$, for the unmixed sparticles,
and at the common scale $M_{\rm SUSY}$ for the mixed ones.
The RGE-improved 1-loop effective potential is minimized at $M_{\rm SUSY}$, which effectively
accounts for the leading 2-loop corrections. Full 1-loop radiative
corrections~\cite{Pierce:1996zz} are incorporated for all sparticle masses.
We have performed random scans over the model parameters (\ref{para}) in the following range:
\begin{eqnarray}
10^{3}~{\rm GeV}\leq & \Lambda & \leq 10^{7}~{\rm GeV}, \nonumber \\
10^{5} ~{\rm GeV}\leq & M_{\rm mess} &\leq 10^{16} {\rm GeV}, \nonumber \\
40 \leq & \tan\beta & \leq 60, \\
N_{5}=1, & \mu < 0. \nonumber
\label{parameterRange}
\end{eqnarray}
Varying $N_{5}$ from 1 to 4, we find that the low scale sparticle spectrum does not change
significantly~\cite{Ajaib:2012vc}. Thus we present results only for $N_{5}=1$ case.
Regarding the MSSM parameter $\mu$, its magnitude but not its sign is determined by the radiative electroweak
symmetry breaking. In our model we set ${\rm sign}(\mu)=-1$.
A negative $\mu$-term together with a positive gaugino mass $M_2$ gives negative
contribution to the muon anomalous magnetic moment calculation~\cite{Moroi}
which would disagree with the measured value for muon $g-2$~\cite{gm2}.
This, however, is not a problem, since the sparticle spectrum is quite heavy in our scenario.
As we will see, the smuons are often heavier than 1~TeV. Hence, the SUSY contribution to muon $g-2$ is not significant,
which also is the case in $t\mhyphen b\mhyphen\tau$ YU scenarios~\cite{Baer:2001kn,Gogoladze:2010fu}.
On the other hand, a negative sign of $\mu$ with positive signs for all gauginos helps to realize YU.
The reason for this is that in order to implement YU, we require significant SUSY threshold corrections
to the bottom quark Yukawa coupling.
One of the dominant finite corrections is proportional to $\mu M_3$~\cite{Hall} and
it helps to realize Yukawa unification if this combination has negative sign~\cite{Gogoladze:2010fu}.
Finally, we employ the current central value for the top mass, $m_{t}=173.3$~GeV.
Our results are not too sensitive to one or two sigma variation
of $m_{t}$~\cite{Gogoladze:2011db}.
In scanning the parameter space, we employ the Metropolis-Hastings
algorithm as described in Ref.~\cite{Belanger:2009ti}. The data points collected all satisfy
the requirement of radiative electroweak symmetry breaking (REWSB). After collecting the data, we impose
the mass bounds on the particles~\cite{pdg} and use the IsaTools package~\cite{isatools,Baer:2001kn}
to implement the various phenomenological constraints. We successively apply mass bounds
including the Higgs~\cite{mhatlas,mhcms} and gluino masses~\cite{gluinoLHC},
and the constraints from the rare decay processes $B_s \rightarrow \mu^+ \mu^-$~\cite{BsMuMu},
$b \rightarrow s \gamma$~\cite{Amhis:2012bh} and $B_u\rightarrow\tau \nu_{\tau}$~\cite{Asner:2010qj}.
The constraints are summarized below in Table~\ref{table1}.
Notice that we used wider range for the Higgs boson mass, since there is an approximate error of around 2~GeV in the estimate of the mass
that largely arises from theoretical uncertainties in the calculation of the minimum of the
scalar potential, and to a lesser extent from experimental uncertainties in the values of $m_t$ and $\alpha_s$.
\begin{table}[h]\centering
\begin{tabular}{rlc}
$ 123\, {\rm GeV} \leq m_h \leq127$ \,{\rm GeV} &
\\
$ m_{\tilde{g}} \geq1.5$ \,{\rm TeV} & \\
$0.8\times 10^{-9} \leq{\rm BR}(B_s \rightarrow \mu^+ \mu^-)
\leq 6.2 \times10^{-9} \;(2\sigma)$ &
\\
$2.99 \times 10^{-4} \leq
{\rm BR}(b \rightarrow s \gamma)
\leq 3.87 \times 10^{-4} \; (2\sigma)$ &
\\
$0.15 \leq \frac{
{\rm BR}(B_u\rightarrow\tau \nu_{\tau})_{\rm MSSM}}
{{\rm BR}(B_u\rightarrow \tau \nu_{\tau})_{\rm SM}}
\leq 2.41 \; (3\sigma)$ &
\end{tabular}
\caption{Phenomenological constraints implemented in our study.}
\label{table1}
\end{table}
Following ref.~\cite{Belanger:2009ti} we define the parameters $R_{tb\tau}$ and $R_{b\tau}$ which quantify $t\mhyphen b\mhyphen\tau$ YU
and $b\mhyphen\tau$ YU respectively:
\begin{equation}
R_{tb\tau} = \frac{{\rm Max}(y_{t},y_{b},y_{\tau})}{{\rm Min}(y_{t},y_{b},y_{\tau})},
\hspace{0.5cm}
R_{b\tau} = \frac{{\rm Max}(y_{b},y_{\tau})}{{\rm Min}(y_{b},y_{\tau})}
\end{equation}
{Thus, $R$ is a useful indicator for Yukawa unification with $R\lesssim 1.1$,
for instance, corresponding to Yukawa unification to within 10\%, while
$R=1.0$ denotes `perfect' Yukawa unification.}
\section{Results}
\label{sec:results}
\begin{figure}[h!]
\centering
\includegraphics[scale=1]{gmsb_Rtbtau_Lambda.png}\hfill
\includegraphics[scale=1]{gmsb_Rtbtau_Mmess.png}\hfill
\includegraphics[scale=1]{gmsb_Mmess_Lambda.png}\hfill
\includegraphics[scale=1]{gmsb_munegative_muLow_Mmess.png}
\caption{Plots in $R_{tb\tau}-\Lambda$, $R_{tb\tau}-M_{\rm mess}$, $M_{\rm mess}-\Lambda$ and $\mu - M_{\rm mess}$
planes. All points are consistent with REWSB. Green points also satisfy mass bounds and B-physics constraints.
Red points are a subset of green and they are compatible with t-b-$\tau$ YU.
In the top panels regions below the horizontal line are compatible with YU such that $R_{tb\tau}\leq 1.1$}.
\label{figure1}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=1]{gmsb_munegative_Rtbtau_tanb.png}\hfill
\includegraphics[scale=1]{gmsb_munegative_Rbtau_tanb.png}
\caption{Plots in $R_{tb\tau}-\tan\beta$ and $R_{b\tau}-\tan\beta$ planes.
The color coding is the same as in Figure~\ref{figure1}.}
\label{figure2}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=1]{gmsb_munegative_mh0_tanb_tbtau.png}\hfill
\includegraphics[scale=1]{gmsb_munegative_mh0_tanb_btau.png}
\caption{Plots in $m_h -\tan\beta$ plane for $t\mhyphen b\mhyphen\tau$ YU (left panel) and $b\mhyphen\tau$ YU (right panel).
The color coding is the same as in Figure~\ref{figure1} except the Higgs mass bound is not applied in these panels.}
\label{figure3}
\end{figure}
\begin{figure}[h!]
\includegraphics[scale=1]{gmsb_munegative_stop1_stop2.png}\hfill
\includegraphics[scale=1]{gmsb_munegative_uL_glu.png}\hfill
\includegraphics[scale=1]{gmsb_munegative_stau1_stau2.png}\hfill
\includegraphics[scale=1]{gmsb_char_neut.png}
\caption{Plots in $m_{\tilde t_{1}}-m_{\tilde t_{2}}$, $m_{\tilde u_{L}}-m_{\tilde g}$,
$m_{\tilde \tau_{1}}-m_{\tilde \tau_{2}}$ and $m_{\tilde{\chi}^{\pm}_{1}}-m_{\tilde{\chi}^{0}_{1}}$ planes.
The color coding is the same as in Figure~\ref{figure1}.}
\label{figure4}
\end{figure}
\begin{figure}[h!]\hspace{0.1cm}
\includegraphics[scale=1.05]{gmsb_munegative_Rtbtau_mgrav.png}\hfill
\includegraphics[scale=1.05]{gmsb_munegative_Rbtau_mgrav.png}
\caption{Plots in $M_{\rm mess}-m_{\tilde{G}}$, $\Lambda-m_{\tilde{G}}$, $R_{tb\tau}-m_{\tilde{G}}$ and
$R_{b\tau}-m_{\tilde{G}}$ planes. The color coding is the same as in Figure~\ref{figure1}.}
\label{figure5}
\end{figure}
\begin{table}[t!]
\centering
\scalebox{0.89}{
\begin{tabular}{|c|ccccc|}
\hline
& Point 1 & Point 2 & Point 3 & Point 4 & Point 5 \\
\hline
\hline
$\Lambda$ & $0.94\times 10^{6}$ & $1.51\times 10^{6}$ & $7.6\times 10^{5}$ & $4.7\times 10^{5}$ & $1.16\times 10^{6}$ \\
$M_{\rm mess}$ & $1.5\times 10^{14}$ & $1.3\times 10^{14}$ & $1.2\times 10^{8}$ & $8.6\times 10^{13}$ & $1.3\times10^{15}$\\
$\tan\beta$ & 56.1 & 59.3 & 58.7 & 54.8 & 52.7\\
\hline
$\mu$ & -4906 & -6727 & -2995 & -4268 & -3984\\
$A_{t}$ & -3564 & -5479 & -1933 & -1834 & -4513\\
$A_{b}$ & -3600 & -5540 & -1946 & -1891 & -4868\\
$A_{\tau}$ & -268.1 & -462.9 & -123.2 & -140.4 & -535.1\\
\hline
$m_h$ & 124 & 125 & 123 & 123 & 124\\
$m_H$ & 1148 & 1640 & 929 & 1665 & 2378\\
$m_A$ & 1141 & 1629 & 923 & 1654 & 2363\\
$m_{H^{\pm}}$ & 1153 & 1643 & 934.2 & 1668 & 2380\\
\hline
$m_{\tilde{\chi}^0_{1,2}}$
& 1305, 2438 & 2098, 3868 & 1045, 1971 & 654, 1249 & 1602, 2960\\
$m_{\tilde{\chi}^0_{3,4}}$
& 4902, 4903 & 6725, 6725 & 3001, 3002 & 4190, 4190 & 3993, 3995\\
$m_{\tilde{\chi}^{\pm}_{1,2}}$
& 2440, 4902 & 3871, 6724 & 1972, 3002 & 1251, 4154 & 2961, 3996\\
$m_{\tilde g}$ & 6113 & 9350 & 4995 & 3372 & 7178\\
\hline $m_{ \tilde u_{L,R}}$
& 9545, 8590 & 13434, 12152 & 6903, 6469 & 8140, 7206 & 7818, 7255\\
$m_{\tilde t_{1,2}}$
&6325, 7784 & 9152, 11112 & 5569, 6186 & 5011, 6515 & 5752, 6745\\
\hline $m_{ \tilde d_{L,R}}$
& 9545, 8317 & 13434, 11796 & 6903, 6412 & 8140, 6937 & 7819, 7124\\
$m_{\tilde b_{1,2}}$
& 6280, 7743 & 9134, 11053 & 5617, 6148 & 5054, 6482 & 6022, 6709\\
\hline
$m_{\tilde\nu_{e,\mu}}$
& 5181 & 7122 & 2727 & 4659 & 3560\\
$m_{\tilde\nu_{\tau}}$
& 4843 & 6625 & 2659 & 4358 & 3372\\
\hline
$m_{ \tilde e_{L,R}}$
& 5186, 3698 & 7129, 5044 & 2735, 1478 & 4663, 3324 & 3567, 2378\\
$m_{\tilde \tau_{1,2}}$
& 2600, 4845 & 3397, 6628 & 1193, 2667 & 2340, 4357 & 1748, 3381\\
\hline
$R_{tb\tau}$ &1.00 & 1.09 & 1.09 & 1.06 & 1.23\\
$R_{b\tau}$ &1.00 & 1.08 & 1.06 & 1.02 & 1.07\\
\hline
\end{tabular}}
\caption{Benchmark points for exemplifying our results. All masses and scales are given in GeV units.
Point 1 depicts a solution with perfect YU. Point 2 shows a solution with 125~GeV mass for h consistent with YU. Point 3 displays
a solution with essentially the lowest stau mass consistent with YU.
Similarly, point 4 shows essentially the lightest mass for the lightest neutralino.
This point also demonstrates the lightest gluino and stop masses obtained from our scans.
Point 5 displays a solution which is compatible with $b\mhyphen\tau$ YU,
but not with $t\mhyphen b\mhyphen\tau$ YU.}
\label{benchgmsb}
\end{table}
Figure~\ref{figure1} displays
the regions in the parameter space that are compatible with $t\mhyphen b\mhyphen\tau$ YU with plots in
$R_{tb\tau}-\Lambda$, $R_{tb\tau}-M_{\rm mess}$, $M_{\rm mess}-\Lambda$ and
$\mu - M_{\rm mess}$ planes. All points are consistent with REWSB. Green points satisfy the mass bounds
and B-physics constraints. Red points are a subset of green ones and they are compatible
with $t\mhyphen b\mhyphen\tau$ YU. In the top panels the dashed line indicates
the region consistent with YU with $R_{tb\tau}\leq 1.1$. The $R_{tb\tau}-\Lambda$ plane shows that the parameter
$\Lambda$ in mGMSB is constrained by
$10^{5}\,\rm{GeV} \lesssim \Lambda \lesssim 10^{6}$~GeV for $t\mhyphen b\mhyphen\tau$ YU with $R_{tb\tau} \leq 1.1$.
As seen from the $R_{tb\tau}-M_{\rm mess}$ plane, YU imposes a far less stringent constraint on $M_{\rm mess}$,
namely $10^{8}\,\rm{GeV} \lesssim M_{\rm mess} \lesssim 10^{14}$~GeV.
We also present, for additional clarity, the results in $M_{\rm mess}-\Lambda$ plane.
The MSSM parameter $\mu$ lies in the range $-8\,\rm{TeV} \lesssim \mu \lesssim -3$~TeV for compatibility with YU.
We see that there are plenty of solutions consistent with the mass bounds and constraints
from rare B-decays. However, only the thin red stripe is compatible with YU.
The strong impact of $t\mhyphen b\mhyphen\tau$ YU on the fundamental parameter space is somewhat relaxed if only $b\mhyphen\tau$ YU is imposed.
Figure~\ref{figure2} shows the differences
between $t\mhyphen b\mhyphen\tau$ and $b\mhyphen\tau$ YU cases with plots in $R_{tb\tau}-\tan\beta$ and
$R_{b\tau}-\tan\beta$ planes. The color coding is the same as in Figure~\ref{figure1}.
One sees that, $t\mhyphen b\mhyphen\tau$ YU only allows a very narrow range
for $\tan\beta$, namely $54 \lesssim \tan\beta \lesssim 60$, and perfect
$t\mhyphen b\mhyphen\tau$ YU consistent with the experimental constraints can be realized when $\tan\beta \approx 56$. On the other hand,
the $R_{b\tau}-\tan\beta$ plane shows that the range for
$\tan\beta$ is slightly enlarged, $50 \lesssim \tan\beta \lesssim 60$,
with $b\mhyphen\tau$ YU, and perfect $b\mhyphen\tau$ YU is obtained
for $54 \lesssim \tan\beta \lesssim 57$.
A similar comparison based on the range for the SM-like Higgs boson mass
is given in Figure~\ref{figure3} with plots in $m_h -\tan\beta$ planes for $t\mhyphen b\mhyphen\tau$
(left panel) and $b\mhyphen\tau$ (right panel) YU cases.
The color coding is the same as in Figure~\ref{figure1}, except that the Higgs mass bound is not applied
in these panels. The left panel shows that demanding $10\%$ or better YU,
the lightest CP-even Higgs boson mass can be $ m_{h} \approx 119-125$~GeV for $t\mhyphen b\mhyphen\tau$ case,
while it can be as light as 117~GeV in the case of $b\mhyphen\tau$ YU, as shown in the right panel.
The panels of Figure \ref{figure3} highlight how stringent the Higgs boson mass constraint is
on the fundamental parameter space of mGMSB compatible with YU.
The left panel shows that the 125~GeV Higgs boson mass and $t\mhyphen b\mhyphen\tau$ YU more or less equivalently constrain the parameter space.
On the other hand, as seen from the right panel, $b\mhyphen\tau$ YU can be realized for $\tan\beta \gtrsim 40$,
while the mass bound on the Higgs boson excludes the region with $\tan\beta \lesssim 50$.
Figure~\ref{figure4} displays the mass spectrum of supersymmetric particles in
$m_{\tilde t_2}-m_{\tilde t_1}$, $m_{\tilde u_L}-m_{\tilde g}$,
$m_{\tilde \tau_2}-m_{\tilde \tau_1}$ and
$m_{\tilde{\chi}^{\pm}_{1}}-m_{\tilde{\chi}^{0}_{1}}$ planes,
with the color coding the same as in Figure~\ref{figure1}. As seen from the top panels,
the colored sparticles are quite heavy. Thus $t\mhyphen b\mhyphen\tau$ YU requires
$m_{\tilde{t}_1} \gtrsim 4$~TeV. Since the SSB trilinear scalar interaction term
$A_t$ is relatively small~\cite{Ajaib:2012vc}, the heavier stop is expected to be about
the same mass ($\gtrsim 4$ TeV) as the lighter stop.
Similarly, the squarks of the first two families and the gluino have masses in the
few TeV range. The $m_{\tilde q_L}-m_{\tilde g}$ plane shows that $m_{\tilde g} \gtrsim 3$~TeV
and $m_{\tilde q_L} \gtrsim 6$~TeV for consistency with $t\mhyphen b\mhyphen\tau$ YU as well as
the experimental constraints.
The bottom panels of Figure~\ref{figure4} display the mass spectra
for the lightest sparticles which are found to be stau, chargino and neutralino.
According to the $m_{\tilde \tau_2}-m_{\tilde \tau_1}$ panel, $\tilde \tau_1$ can be as
light as 1~TeV, while $m_{\tilde \tau_2} \gtrsim 2$~TeV.
Similarly, the $m_{\tilde{\chi}^{\pm}_{1}}-m_{\tilde{\chi}^{0}_{1}}$ plane displays the masses
of the lightest chargino and neutralino with the latter $\gtrsim 600$~GeV, and the former $\gtrsim 1200$~GeV.
Finally, Figure~\ref{figure5} displays the gravitino mass,
$m_{\tilde{G}}$, in $R_{tb\tau}-m_{\tilde{G}}$ and $R_{b\tau}-m_{\tilde{G}}$ planes with the color coding as in Figure~\ref{figure1}. As is well-known,
the gravitino is usually the LSP in the mGMSB framework, and its mass can be as heavy as 10~TeV.
A light gravitino is a plausible dark matter candidate and it can also manifest itself through
missing energy in colliders~\cite{Feng:2010ij}.
In the standard scenarios, the relic density bound ($\Omega h^2 \simeq 0.11$~\cite{Komatsu:2008hk})
is satisfied with a gravitino mass $\sim$ 200~eV~\cite{Feng:2010ij},
which makes it a hot dark matter candidate.
The hot component of dark matter, however, cannot be more than $15\%$ which, in turn, implies
that the gravitino mass should be less than about $30$~eV~\cite{Feng:2010ij}. This indeed is predicted
from Figure~\ref{figure5}.
The panels of Figure~\ref{figure5} represent the impact of YU on the gravitino mass.
Both $t\mhyphen b\mhyphen\tau$ and $b\mhyphen\tau$ YU do not allow a gravitino that is too heavy,
preferring instead a gravitino lighter than 100~eV or so. The $R_{tb\tau}-m_{\tilde{G}}$ plane
shows that $t\mhyphen b\mhyphen\tau$ YU is mostly realized with $m_{\tilde{G}} \lesssim 30 $~eV,
with essentially perfect YU possible for nearly massless gravitinos.
The bound on the gravitino mass is not significantly relaxed for $b\mhyphen\tau$ YU.
However, essentially perfect YU is found for $m_{\tilde{G}} \lesssim 50$~eV.
In order to have a complete dark matter scenario one could invoke axions as cold dark matter.
In Table~\ref{benchgmsb} we present five benchmark points
that exemplify our results. The points are chosen to be consistent with the
experimental constraints and YU. The masses are given in GeV. Point 1 depicts a solution with perfect YU.
Point 2 shows the heaviest mass in our scan for the lightest CP-even Higgs boson, while
point 3 displays the lightest stau mass that we found, consistent with YU.
Similarly, point 4 represents the lightest neutralino mass foundin our scan.
This point also shows the lightest gluino and stop masses that we obtained.
Point 5 displays a solution which is compatible with $b\mhyphen\tau$ YU, but not with $t\mhyphen b\mhyphen\tau$ YU.
\section{Conclusion}
\label{conclusions}
{ Models based on the gauge mediated SUSY breaking are highly motivated, as they provide an atractive solution
to the SUSY flavor problem.}
We have explored the implications of $t\mhyphen b\mhyphen\tau$ and $b\mhyphen\tau$ YU condition on the sparticle spectroscopy in the framework
of mGMSB { where the messengers reside in $5+\bar{5}$ multiplets of $SU(5)$ and only interact with visible sector via gauge interactions.}
We find that YU leads to a heavy sparticle spectrum with the lowest mass sparticles being the lightest neutralino
($m_{\tilde{\chi}_{1}^{0}}\gtrsim 600$~GeV), chargino ($m_{\tilde{\chi}_{1}^{\pm}}\gtrsim 1200$~GeV)
and stau ($m_{\tilde \tau} \gtrsim 1000$~GeV). In addition, YU prefers values of the CP-odd Higgs boson mass
$m_{A}$ to be greater than 1~TeV and all colored sparticle masses above 3~TeV.
Such heavy sparticles can escape detection at the 13-14~TeV LHC,
but should be accessible at HE-LHC 33~TeV or the proposed 100~TeV collider.
{ We find that YU requirement in the mGMSB framework leads to a spectrum of gauginos, Higgses and staus, that is very similar to the one in gravity mediated $SO(10)$ GUT, although for different reasons.
The details of the rest of sfermion spectrum are different, but they are at about the same heavy scale.
However, YU in mGMSB framework does not require any {\it ad hoc} assumptions, such as as SSB non-universalities that are necessary in
gravity mediated GUTs.}
It is also interesting to note that $t\mhyphen b\mhyphen\tau$ and $b\mhyphen\tau$ YU favor a gravitino mass less than 30~eV,
which makes it a hot dark matter candidate.
In this case, the gravitino cannot comprise more than 15\% of the dark matter density,
and one may invoke axions in order to have a complete dark matter scenario.
\section*{Acknowledgments}
We thank David Shih for useful discussions.
AM would like to thank FTPI at the University of Minnesota for hospitality during the final stages of this
project.
This work is supported in part by DOE Grants DE-FG02-91ER40626 (IG and QS) and DE-SC0010504 (AM).
This work used the Extreme Science and Engineering Discovery Environment (XSEDE),
which is supported by the National Science Foundation Grant Number OCI-1053575.
Part of the numerical calculations reported in this paper were performed at the National Academic Network and
Information Center (ULAKBIM) of Turkey Scientific and Technological Research Institution (TUBITAK),
High Performance and Grid Computing Center (TRUBA Resources).
IG acknowledges support from the Rustaveli National Science Foundation under grant No.~31/98.
|
\section{Introduction and Main Results}
In \cite{BPS}, a class of particle ensembles on the real line with many-body interactions was introduced, generalizing invariant
random matrix models, and results on the global asymptotics were obtained. G\"otze and the second author
\cite{GoetzeVenker} presented a stochastic linearization procedure which allowed to study local correlations of particles in the
special case of two-body interactions. They showed that in the limit of infinitely many particles the bulk correlations are
described by the celebrated sine kernel law. The purpose of this paper is to extend the
investigations of
\cite{GoetzeVenker} to the edge.
To be more precise, we consider particle ensembles on $\R$ with density proportional to
\begin{align}
\prod_{i<j}\phi(x_i-x_j)e^{- N\sum_{j=1}^NQ(x_j)},\label{e2}
\end{align}
where $\map{Q}{\R}{\R}$ is a continuous function of sufficient growth at infinity compared to the continuous function
$\map{\phi}{\R}{[0,\infty)}$.
Apart from some technical conditions we will assume that
\begin{align}
\phi(0)=0,\quad \phi(t)>0\ \text{for }t\not=0\quad \text{and }\ \lim_{t\to0}\frac{\phi(t)}{\lv t\rv^2}=c>0,\label{e4}
\end{align}
or, in other words, $0$ is the only zero of $\phi$ and it is of order $2$. The connection to random matrix theory is as follows: If
$\phi(t)=t^2$, then we have the unitary invariant ensembles $P_{N,Q}$ given by
\begin{align}
P_{N,Q}(x):=\frac{1}{Z_{N,Q}}\prod_{i<j}\lv x_i-x_j\rv^2 e^{- N\sum_{j=1}^NQ(x_j)}.\label{e1}
\end{align}
Here and in the following, we slightly abuse notation by using the same symbols for a measure and its density.
$P_{N,Q}$ is the joint distribution of the eigenvalues for the random $(N\times N)$ Hermitian matrix with density proportional to
\begin{align*}
e^{-N\textup{Tr}(Q(X))}
\end{align*}
w.r.t. the Lebesgue measure on the space of $(N\times N)$ Hermitian matrices. Here $\textup{Tr}$ denotes the trace and $Q(X)$ is
defined by spectral calculus. This matrix distribution is invariant under unitary conjugations. See e.g.
\cite{Deift,Handbook,PasturShcherbina,AGZ} for general references on random matrix theory. The unitary invariant ensembles
\eqref{e1} have been well-understood in terms of asymptotic behavior of the eigenvalues. A common way to describe their properties
is to use correlation functions. For a probability measure $P_N(x)dx$ on $\R^N$ and a natural number $k\leq N$ define
\begin{align*}
\rho_N^k(x_1,\dots,x_k):=\int_{\R^{N-k}}P_N(x)dx_{k+1}\dots dx_N.
\end{align*}
The $k$-th correlation function is a probability density on $\R^k$, the corresponding measure is called $k$-th correlation measure.
Note that in the Random Matrix literature the $k$-th correlation function often differs from the definition above by the
combinatorial factor $N!/(N-k)!$. The key for the detailed understanding of the local eigenvalue statistics of unitary invariant
ensembles $P_{N, Q}$ is that they are determinantal ensembles with a kernel that can be expressed in terms of orthogonal
polynomials. More precisely, denote by $K_{N,Q}$ the Christoffel-Darboux kernel of degree $N$ associated to the
measure $e^{-NQ(x)}dx$, i.e.
\begin{align*}
K_{N,Q}(t,s):=\sum_{j=0}^{N-1}p_j^{(N,Q)}(t)p_j^{(N,Q)}(s) e^{-\frac{N}{2}(Q(t)+Q(s))},
\end{align*}
where $p_j^{(N,Q)}$ is the unique polynomial of degree $j$ with positive leading coefficient such that $(p_j^{(N,Q)})_{j\in\mathbb
N}$ forms an orthonormal sequence in $L^2(e^{-NQ(x)}dx)$.
Then, for all $1\leq k\leq N$ one can express the $k$-th correlation function by
\begin{align}
\frac{N!}{(N-k)!} \rho_{N,Q}^k(x_1,\dots,x_k) = \det\lb (K_{N,Q}(x_i,x_j))_{1\leq i,j\leq k}\rb. \label{determinantal_relations}
\end{align}
To state our results, it is convenient to rewrite \eqref{e2}. Let $\map{h}{\R}{\R}$ be continuous, even and bounded below. Let
$\map{Q}{\R}{\R}$ be continuous, even and of sufficient growth at infinity. Consider the probability density on $\R^N$ given by
\begin{align}
P_{N,Q}^h(x):=\frac{1}{Z_{N,Q}^h}\prod_{i<j}\lv x_i-x_j\rv^2\exp\{- N\sum_{j=1}^NQ(x_j)-\sum_{i<j}h(x_i-x_j)\} , \label{e3}
\end{align}
where $Z_{N,Q}^h$ is the normalizing constant. Choosing $\phi(t):=t^2\exp\{-h(t)\}$, we see that $P_{N,Q}^h$ is in the form
$\eqref{e2}$. Conditions \eqref{e4} are satisfied as $h$ has no singularities.
G. Borot has pointed out to the second author that $P_{N,Q}^h$ is the eigenvalue distribution of
the unitary invariant ensemble of Hermitian matrices with density proportional to
\begin{align*}
e^{-N\textup{Tr}(Q(X))-\textup{Tr}(h(X\otimes I-I\otimes X))},
\end{align*}
where $I$ denotes the identity matrix (cf. \cite{Borot1}). Nevertheless, we will prefer to view $P_{N,Q}^h$ as particle ensemble
and
therefore
speak of particles
instead of eigenvalues. Furthermore, the ensemble does not seem to be determinantal, in contrast to $P_{N,Q}$.
Our method of proof requires the external field $Q$ to be sufficiently convex. To quantify this, we define for $Q$ being twice
differentiable and convex $\a_Q:=\inf_{t\in\R}Q''(t)$. In the following we will denote by $\rho_{N,Q}^{h,k}$
the $k$-th correlation function or measure of $P_{N,Q}^h$.
The first information we need is the global behavior of the
particles, i.e. their asymptotic location and density. To this end, we state the following theorem which was proved in
\cite{GoetzeVenker}.
\begin{thrm}[{\cite[Theorem 1.1]{GoetzeVenker}}]\label{ethrm1}
Let $h$ be a real analytic and even Schwartz function. Then there
exists a constant $\a^h\geq 0$ such that for all real analytic, strictly convex and even $Q$ with $\a_Q> \a^h$, the following
holds:\\
There exists a compactly supported probability measure $\mu_Q^h$ having a non-zero and continuous density on the interior of its
support and for
$k=1,2,\dots$, the $k$-th correlation measure of $P_{N,Q}^h$ converges weakly to the $k$-fold product of $\mu_Q^h$,
that is for any
bounded and continuous function $\map{g}{\R^k}{\R}$,
\begin{align}
\lim_{N\to \infty}\int g(t_1,\dots,t_k)\, \rho_{N,Q}^{h,k}(t_1,\dots,t_k)\, d^kt=\int g(t_1,\dots,t_k)\, d \mu_Q^h(t_1)\dots
d\mu_Q^h(t_k).
\end{align}
\end{thrm}
\begin{remark}\label{remark_thrm1}\noindent
\begin{itemize}
\item[a)] If $h$ is positive-definite (i.e. with nonnegative Fourier transform), $\a^h$ can be chosen as
$\a^h=\sup_{t\in\R}-h''(t)$ (cf. Remark \ref{remark_definiteness} and \cite[Remark after Theorem 1.1]{GoetzeVenker}).
\item[b)] It was shown in \cite[Sec. 3]{GoetzeVenker} that the measure $\mu_Q^h$ is the equilibrium measure with respect to
the external field $V(x):=Q(x)+\int h(x-t) d\mu_Q^h(t)$. This fact leads to a useful representation of $\mu_Q^h$ (see e.g.
Appendix \ref{Sec2}). We note for later reference that the evenness of $Q$ and $h$ carries over to $\mu_Q^h$ (cf.
\cite[Theorem 1.1]{GoetzeVenker}) and hence to $V$.
\item[c)] In \cite{BPS}, ensembles with many-body interactions are considered and global asymptotics
but not local correlations are derived (see also the discussion on page 326 in \cite{Totik}). In the case of pair
interactions, the classes of admissible
interactions in \cite{BPS}
and
in this paper are different. In \cite{BPS}, a convexity condition is posed, depending solely on the additional interaction
potentials where our conditions depend on both $Q$ and $h$. The characterization of the limiting measure is different, too.
Asymptotic expansions of the partition functions can be found in \cite{Borot2}. Further global asymptotics for
repulsive particle systems have been studied in \cite{Ch1,Ch2}.
\end{itemize}
\end{remark}
On the local scale, there are basically two different regimes of interest. The local regime in the bulk of the particles has
been
investigated in \cite{GoetzeVenker}, yielding asymptotics in terms of the sine kernel. These results on the bulk have been extended
in two ways. Firstly, in the recent \cite{SchubertVenker} the sine kernel asymptotics have been obtained under an unfolding of the
particles, which is a non-linear rescaling that transforms the equilibrium measure to a uniform distribution. This in particular
allows for an efficient treatment of empirical spacings between particles. Secondly, bulk universality has been shown in
\cite{Venker12} for the $\b$-variant of \eqref{e3} which is obtained by replacing the exponent 2 in \eqref{e3} by some arbitrary
$\b>0$.
Returning to $\b=2$, the
asymptotics at the edge are governed in terms of the Airy kernel. The Airy function $\map{\Ai}{\R}{\R}$ is uniquely determined by
the requirements that it is a solution to the differential equation $f''(t)=tf(t)$ and has the asymptotics $\Ai(t)\sim
(4\pi\sqrt{t})^{-1/2}e^{-\frac{2}{3}t^{3/2}}$ as $t\to\infty$. The Airy kernel $K_\Ai$ is defined on $\R^2$ and can be
represented by
\begin{align}
K_\Ai(t,s)=\int_0^\infty \Ai(t+r)\Ai(s+r)dr=\frac{\Ai(t)\Ai'(s)-\Ai'(t)\Ai(s)}{t-s},\label{def_KAi}
\end{align}
where the latter representation only holds for $t\not=s$.
The first main result of the present work is the edge universality of the correlation functions of $P_{N,Q}^h$. Here the
universal aspect
is that in the limit $N\to\infty$ the correlations do not depend on $Q$ and $h$ after an appropriate linear transformation
$t\mapsto b+t/(c^* N^{2/3})$. The quantities $b$ and $c^*$ depend on $Q$ and $h$ through the function $V$ defined in Remark
\ref{remark_thrm1} b) as follows. The number $b$ is the supremum of the support of the equilibrium measure $\mu_Q^h$ and $c^*$ is
related to the prefactor of the square root vanishing of the density of $\mu_Q^h$ at $b$,
\begin{align}\label{def_c^*}
c^*:=\dfrac{2^{-1/3}}{b}G_V(1)^{2/3},
\end{align}
where the function $G_V$ is defined in \eqref{functionG}.
\begin{thrm}\label{ethrm2}
Let $h$ and $Q$ satisfy the assumptions of Theorem \ref{ethrm1} and assume furthermore that the Fourier transform $\hat{h}$ of $h$ decays
exponentially fast. Let
$q\in \R$ be given and let $b$ and $c^*$ be as above.
Then we have for $N\to\infty$
\begin{enumerate}
\item \begin{align}
&\lb\frac{N^{1/3}}{c^*}\rb^k\rho_{N,Q}^{h,k}\lb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rb=\det\left[K_\textup{Ai}(t_i,t_j)\right]_{1\leq i,j\leq
k}+o(1)\label{Ttheorem3}
\end{align}
with the $o(1)$ term being uniform in $t\in[q,\infty)^k$.
\item If $h$ is negative-definite, i.e. $\hat{h}\leq0$, then for any $0<\s<1/3$ the $o(1)$ term in \eqref{Ttheorem3} can be replaced
by $\O(N^{-\s})$, again uniformly in $t\in[q,\infty)^k$.
\end{enumerate}
\end{thrm}
\begin{remark}\label{remark_definiteness_vorne}\noindent
\begin{enumerate}
\item Even in the situation of part b) where we have better bounds on the rate of convergence, the
Airy kernel determinant is the leading term in \eqref{Ttheorem3} only for $t\in[q,o((\log
N)^{2/3})]^k$.
\item The condition of exponential decay of the Fourier transform of $h$ allows for writing $h$ as difference of positive-definite
and real-analytic functions. This technical condition is needed and explained in Section \ref{proof_Theorem_2}. In particular, if
$h$ is positive-definite or negative-definite, the assumption on the decay of the Fourier transform is automatically satisfied
\cite{LukacsSzasz}.
\item If in the situation of part a) of the previous theorem, $h$ is positive-definite, then $\a^h$ can be chosen as
$\a^h=\sup_{t\in\R}-h''(t)$ (cf. Remark \ref{remark_thrm1}). This remark extends to all of our results.
The roles of the different conditions on the definiteness of $h$ are explained in Remark \ref{remark_definiteness} below.
\item
Edge universality has been established for a wide variety of random matrix ensembles. For $\b$-variants of \eqref{e1} with
one-cut support we refer the reader to the recent works \cite{KRV}, \cite{Bourgadeetal13} and
\cite{BFG}. It should be noted that the limit laws depend on the value of $\b$ \cite{RRV}. For results on invariant matrix ensembles and on
Wigner matrices we refer to the books and recent surveys \cite{DeiftGioev, PasturShcherbina, TaoVuSurvey, ErdosSurvey} and
references therein.
\end{enumerate}
\end{remark}
We now turn to the largest particle $x_{\max}$ and provide an extensive description of its distributional properties.
In analogy to the classical central limit theorem for independent variables, we prove a limit theorem on the fluctuations of
$x_{\max}$ around the rightmost endpoint $b$ (Theorem \ref{ethrm3}). It also implies the weak law of large numbers with respect to
a natural product measure. The strong law of large numbers is also shown (Corollary \ref{a.s.convergence}). Special
emphasis is given to the study of the upper tail of the distribution function of $x_{\max}$. Strong results on moderate and
large deviations are presented in Theorem \ref{thrm_MD} and
Corollary \ref{corTW}. We already point out, that it is only in the regime of large deviations that the differences between the
repulsive particle systems $P_{N,Q}^h$ and the determinantal ensembles become visible in the leading order behavior.
We start with the fluctuations of $x_{\max}$ around $b$. The limiting law is not Gaussian but is given by the Tracy-Widom
distribution. Its
distribution function $F_2$ can be defined as a Fredholm determinant implying in particular the representation
\begin{align}
F_2(s)=1+\sum_{k=1}^\infty\frac{(-1)^k}{k!}\int_{(s,\infty)^k}\det\left(K_{\Ai}(t_i,t_j)\right)_{i,j\leq k}dt.\label{F_2_rep}
\end{align}
It can also be expressed as
\begin{align*}
F_2(s)=\exp\{-\int_s^\infty(t-s)q(t)^2dt\},
\end{align*}
where $q$ is the solution of the differential equation $q''(t)=tq(t)+2q(t)^3$ which has the asymptotics $q(t)\sim\Ai(t)$ as
$t\to\infty$ \cite{TracyWidom94}. For later use we note that the large-$s$-asymptotics of $F_2(s)$ are known (see e.g. \cite[(1),
(25)]{BBDF}) to be
\begin{align}
1-F_2(s)=\frac{1}{16\pi}\frac{e^{-\frac43s^{3/2}}}{s^{3/2}}\left(1+\O\left(\frac{1}{s^{3/2}}\right)\right) ,
\quad \text{for} \ \ s \to \infty .\label{TW_asymptotics}
\end{align}
\begin{thrm}\label{ethrm3}
Let $x_{\max}$ denote the largest component of a vector $x$. Under the conditions and with the notation of Theorem \ref{ethrm2} we
have for $N\to\infty$
\begin{enumerate}
\item
\begin{align}
P_{N,Q}^h\lb \lb x_{\max}-b\rb c^*N^{2/3}\leq s\rb=F_2(s)+o(1)\label{TW1}
\end{align}
uniformly for $s\in\R$, where $F_2$ is the distribution function of the ($\b=2$) Tracy-Widom distribution.
\item If $h$ is negative-definite, then the $o(1)$ term in \eqref{TW1} can be replaced by $\O(N^{-\s})$, where $0<\s<1/3$.
The $\O$ term is uniform for $s\in[q,\infty)$ with arbitrary $q\in\R$.
\end{enumerate}
\end{thrm}
Note that Theorem \ref{ethrm3} implies the weak law of large numbers w.r.t.~the product measure
$\mathbb{P}:=\bigotimes_{N=1}^\infty P_{N,Q}^h$ (on an appropriate measurable space), i.e. $x_{\max}$ converges in probability to
$b$. The strong law, i.e. almost sure convergence, is the content of Corollary \ref{a.s.convergence} below.
Theorem \ref{ethrm3} can be viewed as an analogue of the central limit theorem for the distribution of the largest
particle. This point of view leads naturally to the question of what can be said about its large and moderate deviations? There one
is interested, roughly speaking, in describing the decay of the probabilities $P_{N,Q}^h\lb \lb x_{\max}-b\rb c^*N^{2/3}\leq - s\rb$
(lower tail) and $P_{N,Q}^h\lb \lb x_{\max}-b\rb c^*N^{2/3} > s\rb$ (upper tail) when $s$ is allowed to grow to infinity at some
$N$-dependent rate. In this paper we deal exclusively with the upper tail. Note that even for determinantal ensembles the lower
tail has not been studied at the level of detail required for our analysis.
The large (and moderate) deviations theory generally deals with
probabilities decaying at an exponential rate. In a general setting with probability measure $P$ and random variables $X_N$, the
deviation $s$ is often parametrized as $s=rN^\a$ for some $r, \a>0$ and one proves statements like
\begin{align}
\lim_{N\to\infty}\frac{\log P(X_N>rN^\a)}{N^\g}=-I(r).\label{logarithmic_deviations}
\end{align}
Then $N^\g$ is called the speed and $I(r)$ the rate function. In our situation $P$ corresponds to $\mathbb{P}:=\bigotimes_{N=1}^\infty P_{N,Q}^h$ and $X_N$ corresponds to $\lb x_{\max}-b\rb c^*N^{2/3}$ with $x \in \R^N$. The regime of moderate deviations extends over $0 < \a < 2/3$ and the regime of large deviations is associated with $\a = 2/3$. Note that Theorem \ref{ethrm3} does not provide any information on either the speed or on the rate function for any $\a > 0$ due to the fast decay of $1 - F_2$ described in \eqref{TW_asymptotics}.
The next theorem and its two subsequent corollaries remedy this situation. E.g.
in the case of moderate deviations Corollary \ref{corLogarithm} b) shows for all $0 < \a < 2/3$ that we have speed $N^{3\a /2}$ and rate function $I(r)=\frac{4}{3} r^{3/2}$. For large deviations $(\a = 2/3)$ we learn from Corollary \ref{corLogarithm} a) that the speed is $N$ and the rate function $I(r)=\eta_V(1 + r/(b c^*))$ with $\eta_V$ defined in \eqref{functioneta}. Observe that in the moderate regime the rate function is universal, whereas the rate function depends on $V$ and is hence non-universal for large deviations.
The situation becomes more subtle, if one studies the asymptotics of the tail probabilities and not just their logarithms.
These provide finer information about the tails and we show how the $V$-dependence of the leading order behavior becomes
stronger within the region of moderate deviations (Remark \ref{remarkUniversality}).
In recent years superlarge deviations have been studied as well, which would correspond to $\a > 2/3$. We will not
treat them here. See Remark \ref{remarkcomparisonSchueler} for a description of the difficulties that appear in that regime.
We now state Theorem \ref{thrm_MD} that summarizes our results on moderate and large deviations
for the upper tail of the largest particle. It is stated for $P_{N,Q}^h(x_{\max} > t)$ and hence $s$ in the
discussion above corresponds to $t = b + s/(c^* N^{2/3})$. In this scaling the regime of large deviations is described (more
naturally as for $s$) by those values of $t$ for which $t-b$ is of order $1$ and the regime of moderate deviations reads $N^{-2/3 }
<< t-b << 1$.
Theorem \ref{thrm_MD} provides upper and lower bounds on $P_{N,Q}^h(x_{\max} > t)$
for $t$ belonging to an arbitrary but fixed bounded subset of $(b + N^{-2/3}, \infty)$.
In the regime of moderate deviations the quotient of the upper and lower bound is seen to converge to $1$ as
$N \to \infty$ and thus the leading order behavior of $P_{N,Q}^h(x_{\max} > t)$ is determined. It is given by
\begin{align}\label{functionFNV}
\FNV(t) := \frac{b^2e^{-N\eta_V(t/b)}}{4\pi N(t^2-b^2)\eta_V'(t/b)}.
\end{align}
The $V$-dependence is contained in the function
$\map{\eta_V}{[1,\infty)}{\R}$,
\begin{align}
\eta_V(t) := \int_1^t\sqrt{s^2-1}~G_V(s)ds ,\label{functioneta}
\end{align}
where the function $G_V$ is defined in \eqref{functionG} below for the external field $V$ introduced in Remark \ref{remark_thrm1}
above.
In the regime of large deviations, our result is weaker. There we show that the quotient of $\FNV(t)$ and $P_{N,Q}^h(x_{\max} > t)$
and its inverse are both bounded.
\begin{thrm}\label{thrm_MD}
Assume that the conditions of Theorem
\ref{ethrm2} are satisfied and
fix $T > b$. Then the following holds:
\begin{enumerate}
\item There exists a constant $B > 1$ such that for all $N$ and all $t \in (b + N^{-2/3}, T)$ we have
\begin{align}\label{deviations_lower}
\frac{1}{B} \FNV(t) \leq P_{N,Q}^h(x_{\max} > t) \leq B \FNV(t).
\end{align}
\item For any sequences $(p_N)_N$, $(q_N)_N$ of reals with $p_N < q_N$, $N^{2/3}(p_N -b) \to \infty$, and $q_N -b \to 0$ for $N \to \infty$ we have
\begin{align}\label{deviations_general}
P_{N,Q}^h\lb x_{\max} > t\rb=\FNV(t) [ 1+ o(1) ]
\end{align}
uniformly in $t \in (p_N, q_N)$ as $N \to \infty$.
\item If $h$ is negative-definite, then $o(1)$ in statement b) can be replaced by the more precise
\begin{align}\label{E6b}
\O\lb
\frac{1}{N\lv t-b\rv^{3/2}}\rb +\O\lb \sqrt{t-b} \rb
\end{align}
with the $\O$ terms being uniform in $N$ and in $t \in (b + N^{-2/3}, T)$.
\end{enumerate}
\end{thrm}
Note that for negative-definite functions $h$ statement c) implies the upper bound in \eqref{deviations_lower} but not the
lower bound.
\begin{remark}\label{remarkcomparisonSchueler}
Theorem \ref{thrm_MD} should be compared with the result \cite[(1.14)]{DissSchueler} (cf. \cite{EKS}) for determinantal ensembles.
Even our best
error bounds in the case of negative-definite functions $h$ are worse than the bounds there, if $\lv t-b \rv >> N^{-1/2}$. This is
due to the $\O (\sqrt{t-b} )$ term in \eqref{E6b}. We believe that this is not an artefact of our proof. In fact, relations
\eqref{p6b5} and \eqref{p6b10} of the proof show that we can express the distribution of the largest particle by averaging
determinantal ensembles with $V$ replaced by $V-f/N$ over $f$. Tracing back the origin of the $\O (\sqrt{t-b} )$ term in
\eqref{E6b} we see that it does not stem from a lack of control on error terms but on the dependence of the upper endpoint
$b_{V-f/N}$
on $f$, that can actually be computed to first order.
In the regime of large deviations this effect is most prominent and this is the reason why we loose the convergence result for the
distribution of the largest particle. However, we conjecture that also in the regime of large deviations we have convergence
\begin{align}\nonumber
\lim_{N\to\infty} P_{N,Q}^h\lb x_{\max} > t\rb= B(t) \FNV(t)
\end{align}
for some strictly monotone increasing function $B$ with $B(t) > 1$. Observe that our previous results all show that
$P_{N,Q}^h$ has the same limiting behavior as the determinantal ensemble $P_{N,V}$. However, for $P_{N,V}$ one has $B(t)=1$, i.e.
we conjecture the breakdown of comparability between $P_{N,Q}^h$ and $P_{N,V}$ in this regime! To show this one would need to
apply Laplace's method to the
infinite dimensional integral that represents the averaging over $f$. This is a technically demanding enterprise but not without hope
due to the Gaussian character of the probability measure on $f$. We plan to come back to this problem in future work.
The situation becomes even more challenging in the superlarge regime, where the additional difficulty arises that one may not
truncate $P_{N,Q}^h$ to some finite interval $[-L, L]^N$ that is used in the proof of Theorem \ref{thrm_MD} to obtain good bounds on
the Gaussian field $f$ used for the linearization method (see e.g. discussion in the paragraph before \eqref{Gaussian_LD}). In
contrast, for
determinantal ensembles the leading order behavior of the upper tail $P_{N,V}\lb x_{\max} > t\rb$ was shown to be given
by $\FNV(t)$ in both the large and the superlarge regime for a suitable class of convex functions $V$ \cite[Theorem
1.1]{DissSchueler} (cf. \cite{EKS}).
\end{remark}
Next we discuss the connection with the Tracy-Widom law $F_2$ beyond the central limit regime already covered in Theorem
\ref{ethrm3}.
It is convenient to use the scaling of $x_{\max}$ as in the statement of Theorem \ref{ethrm3}, i.e. $t \equiv b + s / (c^* N^{2/3})$.
The positivity and real analyticity of $G_V$ (see \eqref{functionG} and thereafter), definitions \eqref{functioneta},
\eqref{functionFNV} and the definition of $c^*$ in \eqref{def_c^*} lead by a straightforward calculation to the following
representation of $\FNV$ in the regime of moderate deviations, i.e $s/N^{2/3} = o(1)$:
\begin{align}\nonumber
\FNV(t(s)) &= \frac{e^{-N \eta_V (t(s)/b)}}{16 \pi s^{3/2}} \left[ 1 + \O \lb \frac{s}{N^{2/3}}
\rb \right] \quad \text{with}\\
\label{expansion_eta}
N \eta_V (t(s)/b) &= \frac{4}{3}s^{3/2} + \sum_{j=1}^{\infty} d_{j, V} \frac{s^{j + \frac{3}{2}}}{N^{\frac{2}{3} j}} =
\frac{4}{3}s^{3/2} + \O \lb \frac{s^{5/2}}{N^{2/3}} \rb
\end{align}
for some sequence $(d_{j, V})_{j \geq 1}$ of real numbers depending on $V$. The authors would like to thank F.~G\"otze for
pointing out that the power series
\begin{align*}
\frac{4}{3}s^{3/2} + \sum_{j=1}^{\infty} d_{j, V} \frac{s^{j + \frac{3}{2}}}{N^{\frac{2}{3} j}}
\end{align*}
is an analogue to the Cram\'er series \cite{Cramer} in the deviations theory for sums of independent random
variables. It follows that
\begin{align}\nonumber
e^{-N \eta_V (t(s)/b)}=e^{-\frac{4}{3}s^{3/2}}[1 + o(1)]
\end{align}
as $N \to \infty$ for $0 \leq s= o \lb N^{4/15} \rb$, or equivalently for $0 \leq t(s)-b= o \lb N^{-2/5} \rb$. Comparing with the
asymptotics of the Tracy-Widom distribution \eqref{TW_asymptotics} and using in addition that
\begin{align}\nonumber
\sqrt{\frac{s}{N^{2/3}}} = \O\left(\frac1{s^{3/2}}\right)+\O\left(\frac{s^{5/2}}{N^{2/3}}\right) \quad \text{for} \ \ s > 0 ,
\end{align}
we obtain from statements b) and c) of Theorem
\ref{thrm_MD}:
\begin{cor}\label{corTW}
Under the condition and with the notation of Theorem
\ref{thrm_MD} the following relations hold.
\begin{enumerate}
\item For any sequences $(\hat{p}_N)_N$, $(\hat{q}_N)_N$ of reals with $\hat{p}_N < \hat{q}_N$, $\hat{p}_N \to \infty$, and
$N^{-4/15} \hat{q}_N \to 0$ for $N \to \infty$ we have:
\begin{align}\nonumber
P_{N,Q}^h\lb (x_{\max}-b)c^*N^{2/3} > s\rb=(1-F_2(s))(1+o(1)),
\end{align}
uniformly in $s \in (\hat{p}_N, \hat{q}_N)$ as $N \to \infty$.
\item If $h$ is negative-definite, then
\begin{align}\nonumber
P_{N,Q}^h\lb (x_{\max}-b)c^*N^{2/3} > s \rb=(1-F_2(s)) \left[1
+\O\left(\frac1{s^{3/2}}\right)+\O\left(\frac{s^{5/2}}{N^{2/3}}\right) \right]
\end{align}
with the $\O$ terms being uniform in $N$ and in $s \in (1 , N^{4/15})$.
\end{enumerate}
\end{cor}
\begin{remark}\label{remarkUniversality}
With Corollary \ref{corTW} we have identified the region for which the leading order of the upper tail of the distribution of the
largest particle is universal, i.e. all of its $Q$ and $h$ dependence is encoded in the two numbers $b$ and $c^*$ that define the
linear rescaling and play the same role as mean and variance in the Central Limit Theorem. Relations \eqref{expansion_eta} show
nicely how the dependence of the leading order $\FNV(t)$ on $Q$ and $h$ grows gradually within the regime of moderate deviations.
Set $\gamma_k := 2/(2k+5)$. For $0 < t -b = o \lb N^{-\g_k} \rb$, i.e. $s(t) = o\lb N^{-\g_k + 2/3} \rb$, the function $\FNV(t)$
depends to leading order on $Q$ and $h$
only through the $k+2$ numbers $b$, $c^*$, $d_{1,V}, \ldots, d_{k,V}$. This is a generic phenomenon already observed in \cite[Remark
4.12]{DissSchueler} (cf. \cite{EKS}) for determinantal ensembles.
\end{remark}
In the logarithmic form, it
is straightforward to derive the following results from Theorem \ref{thrm_MD} and relations \eqref{expansion_eta},
\eqref{functionFNV}, \eqref{functioneta} and \eqref{functionG}.
\begin{cor}\label{corLogarithm}
Under the condition and with the notation of Theorem
\ref{thrm_MD} the following statements hold.
\begin{enumerate}
\item
\begin{align}\nonumber
\frac{\log P_{N,Q}^h(x_{\max} > t)}{N} = - \eta_V \lb t/b \rb - \frac{\log \lb N (t-b)^{3/2}\rb}{N} + \O \lb \frac{1}{N} \rb,
\end{align}
where the $\O$ term is uniform in $N$ and in $t \in (b + N^{-2/3}, T)$.
\item
For any sequences $(\hat{p}_N)_N$, $(\hat{q}_N)_N$ of reals with $\hat{p}_N < \hat{q}_N$, $\hat{p}_N \to \infty$, and $N^{-2/3}
\hat{q}_N \to 0$ for $N \to \infty$ we have
\begin{align}\nonumber
\frac{\log P_{N,Q}^h\lb (x_{\max}-b)c^*N^{2/3} > s\rb}{s^{3/2}}= -\frac{4}{3} - \frac{\log (16 \pi s^{3/2})}{s^{3/2}} + o\lb \frac{1}{s^{3/2}} \rb + \O \lb \frac{s}{N^{2/3}} \rb,
\end{align}
with error bounds that are uniform in $N$ and $s \in (\hat{p}_N, \hat{q}_N)$. If $h$ is negative-definite, then the $o \lb
s^{-3/2} \rb$ term
can
be replaced by $\O \lb s^{-3} \rb $ that is also uniform in $s \in (\hat{p}_N, \hat{q}_N)$.
\end{enumerate}
\end{cor}
Finally
we present the strong law of large numbers for the largest particle that will be proved at the end of Section \ref{TSec4}.
\begin{cor}\label{a.s.convergence}
Assume the conditions of Theorem \ref{ethrm2} and set $\mathbb{P}:=\bigotimes_{N=1}^\infty P_{N,Q}^h$. Then
\begin{align*}
\mathbb{P}(\lim_{N\to\infty}x_{\max}=b)=1.
\end{align*}
\end{cor}
\begin{remark}
Corollary \ref{corLogarithm} a) implies a full large deviations principle on $\R$ for $x_{\max}$ with speed $N$ and good
rate function
$t\mapsto\eta_V(t/b)$, where we set $\eta_V(t)=\infty$ for $t<1$.
Such a large deviations principle has already been proved in a more general setting
for $\b$-variants of \eqref{e1} (see \cite{Borot2} and references therein). It can be derived via a contraction principle from a
large deviations principle for the empirical distribution of the particles.
We also refer the reader to \cite{LedouxRider} and for results on combinatorial models that are related to invariant matrix
ensembles to \cite{JohT, BDMMZ, Merkl1, Merkl2}.
As mentioned above, stronger results in non-logarithmic form as in Theorem \ref{thrm_MD} and Corollary \ref{corTW} have been shown
for determinantal ensembles in \cite{DissSchueler,EKS}.
\end{remark}
The paper is organized as follows. Section \ref{Sec3} provides a sketch of the
central ideas introduced in \cite{GoetzeVenker} that guide the proofs of all our main results. Then, in Section
\ref{proof_Theorem_2}, we derive our universality result on the correlation functions. The last section deals with our various
results on the distribution
of the largest particle.
In Appendix \ref{Sec2}, we collect all results from
\cite{KSSV} on determinantal ensembles that
are needed in this paper and bring them in a form that is suitable for our purposes.
\section{Outline of the Method}\label{Sec3}
We first mention the basic steps from \cite{GoetzeVenker} in the analysis of $P_{N,Q}^h$. The additional interaction term
$\sum_{i<j}h(x_i-x_j)$ is in general of order $N^2$, so it may influence the limiting measure. The idea is to split it into a term
contributing to the formation of the limiting measure and a perturbation term of lower order. To this end, let us introduce for a
probability measure $\mu$ on $\R$ the notation $h_\mu(t):=\int h(t-s)d\mu(s)$ and $h_{\mu\mu}:=\int\int h(t-s)d\mu(t)d\mu(s)$. We
can then write
\begin{align}
&\sum_{i<j}h(x_i-x_j)=\frac{1}{2}\sum_{i,j}h(x_i-x_j)-\frac{N}{2}h(0)\nonumber\\
=&N\sum_{j=1}^Nh_\mu(x_j)+\frac{1}{2}\sum_{i,j}\left[h(x_i-x_j)-h_\mu(x_i)-h_\mu(x_j)+h_{\mu\mu}\right]+C_N,\label{e5}
\end{align}
where $C_N:=-(N/2) h(0)-(N^2/2)h_{\mu\mu}$. The term in brackets in \eqref{e5} is the Hoeffding decomposition of the statistic
$\sum_{i<j}h(x_i-x_j)$ w.r.t. the measure $\mu$. The term $N\sum_{j=1}^Nh_\mu(x_j)$ will be added to the external field $Q$ forming
a new potential $V_\mu:=Q+h_\mu$. Setting
\begin{align}
\U_\mu(x):=-\frac{1}{2}\sum_{i,j}\left[h(x_i-x_j)-h_\mu(x_i)-h_\mu(x_j)+h_{\mu\mu}\right],\label{def_U}
\end{align}
we arrive at the representation
\begin{align}\label{relation}
P_{N,Q}^h(x)=\frac{Z_{N,V_\mu}}{Z_{N,V_\mu,\U_\mu}}P_{N,V_\mu}(x)e^{\U_\mu(x)}
\end{align}
with $Z_{N,V_\mu,\U_\mu}:=Z_{N,Q}^he^{C_N}$. Formula \eqref{relation} establishes a relation to a determinantal ensemble,
indicating that asymptotics of $P_{N,Q}^h$ may be deduced from asymptotics of $P_{N,V_\mu}$. Our aim is to find a measure $\mu$ such
that the ratio $Z_{N,V_\mu}/Z_{N,V_\mu,\U_\mu}$ and its reciprocal are bounded in $N$. It is straightforward to see that
\begin{align}
\frac{Z_{N,V_\mu,\U_\mu}}{Z_{N,V_\mu}}=\E_{N,V_\mu}e^{\U_\mu},\label{M1}
\end{align}
where $\E_{N,V_\mu}$ denotes expectation w.r.t.~$P_{N,V_\mu}$. In view of the desired boundedness of \eqref{M1}, $\U_\mu$ should
be centered at least asymptotically, which motivates the condition that $\mu$ should be the equilibrium measure to the field
$V_\mu$.
This implicit problem was solved in
\cite[Lemma 3.1]{GoetzeVenker} by
a fixed point argument, yielding existence but not uniqueness of such a $\mu$. The uniqueness followed later by proving
that any
measure $\mu$ which is the equilibrium measure to $V_\mu$, is the limiting measure for $P_{N,Q}^h$. From now on let $\mu$ denote
the unique measure with this property, write $V:=V_\mu$ and $\U:=\U_\mu$. Note that this definition of $V$ is consistent with that
in Remark \ref{remark_thrm1} b).
At this stage, it can be proved that the ratio
$Z_{N,V}/Z_{N,V,\U}$ is bounded in $N$ and bounded away from $0$ provided that $\a_Q$ is large enough. More precisely, given
$\l>0$, there is an $\a(\l)<\infty$ such that there are constants $0<C_1(\l)<C_2(\l)<\infty$ such that for $\a_Q\geq \a(\l)$
\begin{align}\label{concentration_U}
C_1(\l)\leq\E_{N,V}e^{\l\U}\leq C_2(\l)
\end{align}
for all $N$ (see \cite[Proposition 4.7]{GoetzeVenker} and \cite[Remark 4.8]{GoetzeVenker}).
One main tool to derive bound \eqref{concentration_U} is the following representation of $\U$ in terms of linear
statistics (cf. \cite[Lemma 4.6]{GoetzeVenker}) using Fourier techniques,
\begin{align}\label{Fourier}
\U(x)=-\frac{1}{2\sqrt{2\pi}}\int \lv \sum_{j=1}^Ne^{i tx_j}-N\int e^{i
ts}d\mu(s)\rv^2\F{h}(t)dt,
\end{align}
with $\F{h}(t)=(2\pi)^{-1/2}\int e^{-i ts}h(s)ds$.
The other ingredient to the proof of \eqref{concentration_U} is the
following concentration of measure inequality for linear statistics which will be needed later on, too.
\begin{prop}\label{Concentration}
Let $Q$ be a real analytic external field with $Q''\geq c>0$. Then there exists a positive constant $C$ such that
for any Lipschitz function $g$ whose third derivative is bounded on an open interval $I\subset \R$ containing the support of the
equilibrium measure
$\mu_Q$, we have for any
$\epsilon>0$ and $N$
\begin{align*}
\E_{N,Q}\exp\lee{\epsilon\lbb\sum_{j=1}^N g(x_j)-N\int g(t)d\mu_Q(t)\rbb}\ree\leq \exp\lee{\frac{\epsilon^2\Lip{g}^2}{2c}}+
\e C (\|g\|_\infty+\|g^{(3)}\|_\infty)\ree,
\end{align*}
where $\Lip{g}$ denotes the Lipschitz constant of $g$ and $\|\cdot\|_\infty$ denotes the sup norm on $I$.
\end{prop}
The proposition can be found in \cite[Corollary 4.4]{GoetzeVenker} and follows from the fact that for strongly convex $Q$,
$P_{N,Q}$ fulfills a log-Sobolev inequality which implies
concentration of the Lipschitz function $\sum_{j=1}^N g(x_j)$ around its expectation (see e.g. \cite[Proposition 4.4.26]{AGZ}). To
obtain Proposition
\ref{Concentration}, we combine this with the following estimate on the distance of the expectation to its large $N$ limit
from
\cite[Theorem 1]{Shcherbina} (see also \cite[Theorem 1]{KriecherbauerShcherbina}),
\begin{align*}
\Big\lvert \E_{N,Q}\sum_{j=1}^N g(x_j)-N\int gd\mu\Big\rvert\leq C(\|g\|_\infty+\|g^{(3)}\|_\infty),
\end{align*}
where $C$ is the constant that then appears in Proposition \ref{Concentration}.
The main idea to tackle the
local universality is linearizing
the bivariate statistics $\U$ to transform it into linear statistics. To this end, assume that $-h$ is positive-definite (i.e.
$\F{h} \leq 0$; the
general case is reduced to this case). It is well-known (see e.g. \cite{AdlerTaylor}) that due to the positive-definiteness
of $-h$, there exists a centered, stationary Gaussian process $(\tilde{f}(t))_{t\in\R}$ which has $-h$ as its covariance function. That is,
the finite-dimensional distributions of $\tilde{f}$ are Gaussian, $\E \tilde{f}(t)=0$ and $\E \tilde{f}(t)\tilde{f}(s)=-h(t-s)$, where $\E$ denotes expectation
w.r.t. the probability space underlying the Gaussian process.
The key observation now is that for fixed $x \in \R^N$ the random variable $\sum_{j=1}^N\tilde{f}(x_j)-N\int \tilde{f}d\mu$
is again a centered Gaussian with variance $2\,\U(x)$, yielding
\begin{align}
&\exp\{\U(x)\}=\E \exp\{\sum_{j=1}^N\tilde{f}(x_j)-N\int \tilde{f}d\mu\}=\E \exp\{\sum_{j=1}^Nf(x_j)\}\label{e7}\\
&\text{with}\quad f:=\tilde{f}-\int \tilde{f}d\mu.\label{def_f}
\end{align}
Combining \eqref{e7} with \eqref{relation} and \eqref{M1}, the ensemble $P_{N,Q}^h$ can now be represented as an average over determinantal ensembles $P_{N,V-f/N}$ as follows
\begin{align}
P_{N,Q}^h(x)=\frac{\E\left[Z_{N,V-f/N} P_{N,V-f/N}(x)\right]}{\E\, Z_{N,V-f/N}}.\label{Lago1}
\end{align}
Relation \eqref{Lago1} immediately extends to correlation functions. At this point the key observation for our universality proofs
is the fact that for almost all $f$ the correlation functions $\rho_{N,V-f/N}^k$ and $\rho_{N,V}^k$ coincide in the region of
interest to leading order if $N$ is sufficiently large. We will show in the appendix how to extract this information from
\cite{KSSV} (cf. Proposition
\ref{keyproposition}).
We end the brief overview of the method of proof by noting that in order to use the rich theory on Gaussian processes, in particular the sub-Gaussianity of maxima of Gaussian processes
on a compact, we first need to truncate the ensemble to some compact interval $[-L,L]^N$.
Finally, we comment on the role that positive/negative-definiteness of $h$ plays in the approach of \cite{GoetzeVenker} and how it extends to the present paper.
\begin{remark}\label{remark_definiteness}
The reader might have wondered about the different results depending on $h$ being positive-definite or negative-definite
(cf. Remark \ref{remark_definiteness_vorne}).
If $h$ is positive-definite, then $\U$ in \eqref{def_U} is negative for all configurations $x$, which can be easily seen from the
Fourier representation \eqref{Fourier}. This immediately gives the upper bound of \eqref{concentration_U}, without any condition
on $\a_Q$. The lower bound and the further analysis in the case of positive-definite $h$ only need $V$ to be strongly convex, e.g.
to apply Proposition \ref{Concentration}. With the crude bound $\a_Q>\sup_{t\in\R}-h''(t)$, given in Remark \ref{remark_thrm1},
this is guaranteed.\\
If $h$ is not negative-definite, then the linearization method as outlined in \eqref{e7} needs to be modified. The
function $h$ can be decomposed into positive-definite functions $h^\pm$, such that $h=h^+-h^-$.
If $h^+\not=0$, then $-h$ is not a covariance function, but all functions $zh^++h^-$
with $z>0$
are and can be used for the linearization of a modified version $\U_z$ of $\U$. We will then employ Vitali's Theorem from
complex analysis (a
consequence of Montel's Theorem and the Identity Principle) to transfer the convergence results for $z>0$ to $z=-1$ that
corresponds to the original $\U = \U_{-1}$. In this process, however, bounds on the rates of convergence are lost. This is the
reason why we obtian in all of our results better bounds in the case of negative-definite functions $h$.
\end{remark}
\section{Detailed proof of Theorem \ref{ethrm2}}\label{proof_Theorem_2}
\subsection*{Representation of correlation functions and truncation}\noindent
We begin with a different representation of the correlation functions. For this we need a slightly generalized invariant
ensemble: Define for a continuous $Q$ of sufficient growth, continuous $f$ of moderate growth and $M\in\mathbb{N}$ the density on
$\R^N$ by
\begin{align}
P_{N,Q,f}^M(x):=\frac{1}{Z_{N,Q,f}^M}\prod_{1\leq i<j\leq N}\lv x_i-x_j\rv^2e^{-M\sum_{j=1}^N Q(x_j)+\sum_{j=1}^N f(x_j)}.
\end{align}
We will usually have $M=N+k$ for some $k$. If $M=N$, we will abbreviate $P_{N,Q,f}:= P_{N,Q,f}^M$ and $ P_{N,Q}^M:= P_{N,Q,f}^M$ if
$f=0$. If $f=0$ and $M=N$, we have $P_{N,Q}$.
The $k$-th correlation function of $P_{N,V}$ at points $t_1,\dots,t_k$ can be written as
\begin{align}
&\rho^k_{N,V}(t_1,\dots,t_k)\nonumber\\
=&\int_{\R^{N-k}}\frac{1}{Z_{N,V}}\exp\lee-N\sum_{j=k+1}^NV(x_j)+2\sum\limits_{i<j;\ i,j>k} \log\lvb x_j-x_i\rvb
\ree\nonumber\\
&\times \exp\lee-N\sum_{j=1}^k
V(t_j)+2\sum_{i<j;\ i,j\leq
k} \log\lv t_i-t_j\rv\ree\nonumber\\
&\times\exp\lee 2\sum_{i\leq k,\ j>k}\log\lv t_i-x_j\rv\ree dx_{k+1}\dots dx_N\nonumber\\
&=F(t)\frac{Z_{N-k,V}^N}{Z_{N,V}}\E_{N-k,V}^N\exp\lee{2\sum_{i\leq k,\ j>k}\log\lv t_i-x_j\rv}\ree,\quad \text{where}\nonumber\\
&F(t):=\exp\lee{-N\sum_{j=1}^kV(t_j)+2\sum_{i<j;\ i,j\leq
k}\log\lv t_i-t_j\rv}\ree\label{e12}
\end{align}
is the factor (\ref{e12}), which depends only on the fixed particles and $\E_{N-k,V}^N$ is the expectation of $P_{N-k,V}^N$.
We label the random eigenvalues of the ensemble $P_{N-k,V}^N$ by $x_{k+1},\dots,x_N$ and abbreviate
$(t_1,\dots,t_k,x_{k+1},\dots,x_N)$ by $(t,x)$.
Setting
\begin{align}\label{def_R}
R(t,x):=R_{N-k,V}^N(t,x):=2\sum_{i\leq k,\ j>k}\log\lv
t_i-x_j\rv+\log\big[F(t)\frac{Z_{N-k,V}^N}{Z_{N,V}}\big],
\end{align}
we arrive at the shorthand
\begin{align}
&\rho^k_{N,V}(t_1,\dots,t_k)=\E_{N-k,V}^N\exp\lee{R(t,\cdot)}\ree.
\end{align}
Using \eqref{relation} and \eqref{M1} and recalling that we set $V=V_\mu$ and $\U=\U_\mu$, we see that the $k$-th correlation
function $\rho_{N,Q}^{h,k}$ of $P_{N,Q}^h$ at $t_1\dots,t_k$ can be
written as
\begin{align}
\rho_{N,Q}^{h,k}(t_1,\dots,t_k)=\frac{1}{\E_{N,V}\exp\lee{\U}\ree}\E_{N-k,V}^N\exp\lee{(\U+R)(t,\cdot)}\ree.\label{e8}
\end{align}
The representation \eqref{e8} allows to control the effects of truncation on the correlation functions (see \cite[Lemma
6.3]{GoetzeVenker}). More precisely, let $\E_{N,V;L}^M$ denote the expectation w.r.t. the ensemble $P_{N,V;L}^M$ obtained by
normalizing the ensemble $P_{N,V}^M$
restricted to $[-L,L]^N$. Furthermore, let $R_L$ be the analogue of $R$ in which the partition functions in \eqref{def_R} are
replaced by their truncated versions. Then \cite[Lemma 6.3]{GoetzeVenker} (see also \cite{Johansson98},\cite{BPS}) states that for
each $k$ there are $L,C>0$
such that
for all $N$ and
for all $t_1,\dots,t_k$
\begin{align}
\lv\rho_{N,Q}^{h,k}(t_1,\dots,t_k)-\frac{1}{\E_{N,V;L}\exp\lee{\U}\ree}\E_{N-k,V;L}^N\exp\lee{(\U+R_L)(t,\cdot)}
\ree\rv\leq
e^{-CN}.\label{truncation}
\end{align}
Moreover, we learn from the last inequality in the proof of \cite[Lemma 6.3]{GoetzeVenker} that $L$ can be chosen such that in addition there exists a constant $c>0$ with
\begin{align}
\lvert\rho_{N,Q}^{h,k}(t)\rvert\leq e^{-cN}\label{Lago2}
\end{align}
for all $t\in\R^k\setminus(-\infty,L]^k$. It can be shown (see \cite[Remark 4.5, Remark 4.8]{GoetzeVenker} and the proof of
\cite[Proposition 4.7]{GoetzeVenker}) that for such a choice of $L$ and any $\l>0$ we can again find constants $C_1(\l),C_2(\l)>0$
with
\begin{align}
0<C_1(\l)\leq\E_{N,V;L}e^{\l\U}\leq C_2(\l)\label{Lago3}
\end{align}
for all $N$, provided $\a_Q\geq\a(\l)$, where $\a(\l)$ can be chosen as in \eqref{concentration_U} and does therefore not depend on $L$.
\subsection*{Linearization and proof of Theorem \ref{ethrm2} b)}\noindent
We now give a more detailed description of the linearization procedure for negative-definite $h$. In this case
we can
indeed view $-h$
as covariance function of a centered stationary Gaussian process on $\R$ such that \eqref{e7} holds. Analyticity
of the sample paths can be deduced from the spectral representation of $\tilde{f}$ which we now explain in some detail.
Let $(B_t)_{t\geq0}$ denote a standard 1-dimensional Brownian motion, in particular $B_0=0$ a.s.. Recall that by the law of the
iterated logarithm, we know that $\lv B_t\rv$ is almost surely $\O(\sqrt{2t\log\log t})$. Thus, for
$\map{g}{[0,\infty)}{\R}$
of sufficient decay and smoothness, we can define the Wiener integral
\begin{align}
\int_0^\infty g(s)dB_s:=-\int_0^\infty B_sdg(s)=-\int_0^\infty B_sg'(s)ds.\label{gaussian_process}
\end{align}
Recall that $(B_t)_t$ is a Gaussian process, that is, its finite dimensional distributions are Gaussian which is equivalent to the
property that finite linear combinations of the family of random variables $\{B_t:t\geq0\}$ are Gaussian. By a limit argument we
see that \eqref{gaussian_process} is a Gaussian random variable and by elementary computations we find that the mean is 0 and the
variance is $\int_0^\infty g(s)^2ds$.
Now, for the representation of $\tilde{f}$, let $(B^1_t)_t, (B^2_t)_t$ denote two independent Brownian motions. Then we can define
\begin{align}
\tilde{f}(t):=\lb\frac2\pi\rb^{1/4}\int_0^\infty \cos(ts)\sqrt{-{\hat{h}(s)}}dB_s^1+\lb\frac2\pi\rb^{1/4}\int_0^\infty
\sin(ts)\sqrt{-{\hat{h}(s)}}dB_s^2.\label{process_representation}
\end{align}
To verify that $\tilde{f}$ defined in this way has the desired properties, it is enough to note that the right hand side of
\eqref{process_representation} forms a Gaussian
process on $\R$ (which can be easily checked using the characterization mentioned above) with mean $0$ and covariance function
$-h$. It is a somewhat surprising and maybe not so well-known fact \cite{LukacsSzasz} that the Fourier transform of a
positive-definite real-analytic function decays exponentially at infinity (see also the discussion preceding \eqref{e472}). By
this exponential decay of $\hat{h}$, representation \eqref{process_representation} continues to
hold for $w$ from a strip $\{x+iy:x\in\R, \lv y\rv<\d\}$ for some $\d>0$. For later use, we define the region
$D:=(-L-1,L+1)\times(-\d/3,\d/3)\subset \C$ with an appropriate choice of $L$ made precise below. We thus see that
$\tilde{f}$ is analytic on $D$ a.s.. It follows
also
from \eqref{process_representation} that the extended process $(\tilde{f}(w))_{w\in D}$ is a complex-valued centered Gaussian process and
it
is straightforward to show that the covariance function is
\begin{align}
\E (\tilde{f}(w_1)\overline{\tilde{f}(w_2)})=-h(w_1-\overline{w_2}).\label{covariance}
\end{align}
Before starting the proof we finally introduce an abbreviation for the Airy kernel determinants (see also \eqref{def_KAi})
\begin{align}
\AI_k(t):=\det\left[K_\textup{Ai}(t_i,t_j)\right]_{1\leq i,j\leq k}.\label{AI}
\end{align}
\begin{proof}[Proof of Theorem \ref{ethrm2} b)]
Let $k$ be fixed and choose $L$ such that \eqref{truncation} and \eqref{Lago2} are satisfied. We first restrict our attention to the
case $t\in [q,N^\e]$ with $0<\e<\min(2/3-2\s,2/15)$. Note that this includes the region where the Airy kernel determinant
describes the leading order behavior.
Using \eqref{truncation} and the lower bound of $\eqref{Lago3}$, it suffices to show
\begin{align}
&(c^*)^{-k}N^{
k/3 }\E_{N-k,V;L}^N\exp\lee{(\U+R_L)(b+\frac{t}{c^*N^{2/3}},\cdot)}\ree-\E_{N,V;L}\exp\lee{\U}\ree\AI_k(t)\label{e9}\\
&=\O\lb
N^{-\s}\rb\nonumber
\end{align}
uniform in $t$.
Let $\tilde{f}$ be defined as in \eqref{process_representation} and set $f:=\tilde{f}-\int
\tilde{f}d\mu$ (cf. \eqref{def_f}).
We now apply the linearization procedure and obtain from \eqref{Lago1}
\begin{align*}
\rho_{N,Q;L}^{h,k}(t)=\frac{\E\left[Z_{N,V-f/N;L} \rho_{N,V-f/N;L}^k(t)\right]}{\E\, Z_{N,V-f/N;L}}.
\end{align*}
Observe that
\begin{align}
Z_{N,V-f/N;L}=Z_{N,V;L}\,\E_{N,V;L}\exp\lbb\sum_{j=1}^Nf(x_j)\rbb\label{Lago4}
\end{align}
together with \eqref{e7} leads to
\begin{align*}
\E\,Z_{N,V-f/N;L}=Z_{N,V;L}\,\E_{N,V;L}\exp(\U).
\end{align*}
The last three relations together with a truncated version of \eqref{e8} yield
\begin{align*}
&\E_{{N}-k,V;L}^{N}\exp\lee{(\U+R_L)(b+\frac{t}{c^*{N}^{2/3}},\cdot)}\ree=\E_{N,V;L}\exp(\U)\rho_{N,Q;L}^{h,k}(b+\frac{t}{c^*{N}^{2/3}})\nonumber\\
=&\E\Big[\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f(x_j)\ree \rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb\Big],
\end{align*}
where $\rho_{N,V,f;L}^k$ is the $k$-th correlation function of the unitary
invariant ensemble $P_{N,V-f/N;L}=:P_{N,V,f;L}$ defined
on $[-L,L]^N$. We thus get that \eqref{e9} is equal to
\begin{align}
\E\Big[\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree\Big( (c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t)\Big)\Big].\label{e11}
\end{align}
By Proposition \ref{keyproposition} (observe that $c^*=\g_V$ and $a_V=-b_V$), by the determinantal relations \eqref{determinantal_relations},
and by the almost sure analyticity of $f$, we obtain for the term in the inner
parenthesis
\begin{align*}
(c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t) = \O(N^{-\s})
\end{align*}
for almost all $f$ with the $\O$-term uniform for $\|f\|_D\leq N^\kappa$ with $\k:=1/3-\s-\e/2$.
Since $\E\,\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree=\E_{N,V;L}
\exp\lee\U\ree$ is uniformly bounded in $N$ (cf. \eqref{Lago3}), we have derived
\begin{align}
\E\Big[\,&\dopp{1}_{\{\|f\|_D\leq N^\kappa\}}\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree\nonumber\\
&\times\Big( (c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t)\Big)\Big]=\O(N^{-\s}).\label{Lago42}
\end{align}
Hence the proof is finished for the case $t\in[q,N^\e]^k$ if we can show
\begin{align}
\E\Big[\,&\dopp{1}_{\{\|f\|_D>N^\kappa\}}\times\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree\nonumber\\
&\times\Big( (c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t)\Big)\Big]=\O(e^{-cN^{2\k}})\label{Gaussian_truncation}
\end{align}
for some $c>0$ uniformly in $t\in [q,N^{\e}]^k$. This bound will follow from an application of H\"older's inequality to
separate
the three $f$-dependent factors in \eqref{Gaussian_truncation}. Here we will use $L^3$ norms for convenience.
We first treat $\dopp{1}_{\{\|f\|_D>N^\kappa\}}$. It follows
readily from \eqref{process_representation} that real and imaginary parts of $\tilde{f}$ on $D$ are (real-valued) centered Gaussian
processes.
By Borell's inequality (see e.g. \cite[Theorem 2.1.1]{AdlerTaylor}) the supremum $\|X\|_\infty$
of a real-valued continuous centered Gaussian process $X_t$ over a compact $K$ is sub-Gaussian, i.e. dominated by a Gaussian random variable
with a certain expectation and variance $\s:=\sup_{t\in K}\E
X_t^2$. To apply this fact, we will without further notice always bound the supremum over the open set $D$ by the supremum
over $\overline{D}$. As the sum of sub-Gaussian random variables is also sub-Gaussian, we see using $\sup_{w\in D}\lv
\tilde{f}(w)\rv\leq
\sup_{w\in D}\lv \Re \tilde{f}(w)\rv+\sup_{w\in D}\lv \Im \tilde{f}(w)\rv$ that $\sup_{w\in D}\lv \tilde{f}(w)\rv$ has
sub-Gaussian tails. Since $\lv\int \tilde{f}d\mu\rv\leq\sup_{w\in D}\lv
\tilde{f}(w)\rv$ we conclude that $\sup_{w\in D}\lv f(w)\rv$ is sub-Gaussian. Note that $\lvert\Re\tilde{f}\rvert$ and
$\lvert\Im\tilde{f}\rvert$ are both bounded by $\lvert \tilde{f}\rvert$ and that $\E \lvert \tilde{f}(w)\rvert^2=-h(2i\Im w)$ by
\eqref{covariance}. We therefore find that $\sup_{w\in D} \lvert f(w)\rvert$ is dominated by a Gaussian with a variance $4\sup_{w\in
[-\d/3,\d/3]}-h(2iw)$. Thus there exists $c>0$ such that
\begin{align}\label{Gaussian_LD}
P\{\|f\|_D>N^\k\}=\O(e^{-3cN^{2\k}})\ \text{ and hence }\ \lbb\E\dopp{1}_{\{\|f\|_D>N^\kappa\}}^3\rbb^{1/3}=\O(e^{-cN^{2\k}}).
\end{align}
We now turn to the second factor $\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree$ in \eqref{Gaussian_truncation}.
Proposition \ref{Concentration} extends to the case of a truncated ensemble,
yielding an error of exponential order which we omit in the following (cf. \cite[Remark 4.5]{GoetzeVenker}). Thus we have
\begin{align}
\E\,\Big[ \E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f({x_j})\ree\Big]^3\leq \E\exp\lee{\frac{3\Lip{f}^2}{2\a_V}}+
3C(\|f\|_{[-L,L]}+\|f^{(3)}\|_{[-L,L]})\ree.\label{e10}
\end{align}
Observe that the
Lipschitz constant is evaluated over $[-L,L]$ instead of $\R$.
In order to estimate this $L^3$ norm, we argue in addition to the sub-Gaussianity of $\|\tilde{f}\|_{[-L,L]}$ that also the
processes $\|\tilde{f}'\|_{[-L,L]}, \|\tilde{f}^{(3)}\|_{[-L,L]}$ are sub-Gaussian. To this end note that the process $\tilde{f}'$
is a centered stationary Gaussian process with covariance function $h''$.
Borell's inequality yields again sub-Gaussianity of $\Lip{f}$ and analogous arguments prove
sub-Gaussianity of $\|f^{(3)}\|_{[-L,L]}$ as well. This shows the finiteness of the r.h.s.~of \eqref{e10} provided that $\a_Q$ and
hence $\a_V$ is sufficiently large.
It is noteworthy that the condition on $\a_Q$ is determined by the variance $h''(0)$ of $\tilde{f}'$ only. As $\tilde{f}'$ is
stationary, this $\a_Q$ is independent of $L$ and hence of $k$. In fact, this is the very reason that in Theorem \ref{ethrm2} the
convergence of \textit{all} correlation functions can be derived for a given function $h$.
Next, we will estimate $(c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb$.
From \eqref{determinantal_relations} we see that
$(\tilde{K}_{N}(t_i,t_j))_{1\leq i,j\leq k}=:A$ (where $\tilde{K}_N$ is now a shorthand for the kernel $\tilde{K}_{N,V,f;L}$ associated to $P_{N,V,f;L}$
evaluated at rescaled variables $b+\frac{t_j}{c^*N^{2/3}}$, $1\leq j\leq k$) is positive
semi-definite
and can hence be
written as $A=B^2$ for some Hermitian matrix $B$. Now using Hadamard's inequality we obtain
\begin{align}\label{THad}
&\det A=\lb \det B\rb^2\leq \prod_{j=1}^k\sum_{i=1}^k\lv B_{ij}\rv ^2=\prod_{j=1}^kA_{jj}.
\end{align}
In our case this reads
\begin{align}
&\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb\leq (N-k)!/(N!) \prod_{j=1}^k\tilde{K}_N(t_j,t_j)\nonumber\\
&\leq e^k\prod_{j=1}^k
\rho_{N,V,f;L}^{1}(b+\frac{t_j}{c^*N^{2/3}}).\label{Hadamard}
\end{align}
Now, we use the well-known (see e.g. \cite{Freud}) representation
\begin{align*}
\rho^1_{N,V,f;L}(t)= \frac{e^{-NV+f}}{N\l_N(e^{-NV+f},t)}, \l_N(e^{-NV+f},t):=\inf_{P_{N-1}(t)=1}\int_{-L}^{L} \lv
P_{N-1}(s)\rv^2e^{-NV(s)+f(s)}ds,
\end{align*}
where the infimum is taken over all polynomials $P_{N-1}$ of at most degree $N-1$ with the property that $P_{N-1}(t)=1$. From this
it is obvious that
\begin{align*}
\rho_{N,V,f;L}^{1}(b+\frac{t_j}{c^*N^{2/3}})\leq \rho_{N,V;L}^{1}(b+\frac{t_j}{c^*N^{2/3}})e^{2\|f\|_{[-L,L]}}.
\end{align*}
By Proposition \ref{keyproposition},
\eqref{determinantal_relations}, and by the uniform boundedness of the Airy kernel in the region of interest, we find that
\begin{align}
&(c^*)^{-1}N^{1/3
}\rho_{N,V,f;L}^{1}(b+\frac{t_j}{c^*N^{2/3}})\leq C e^{2\|f\|_{[-L,L]}}\ \text{ and thus}\nonumber\\
&(c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb\leq (Ce^{2\|f\|_{[-L,L]} +1})^k,\label{e15}
\end{align}
where $C$ does not depend on $f$.
This estimate together with the uniform boundedness of $\AI_k$ and the sub-Gaussianity proves the uniform boundedness of the third factor
\begin{align*}
(c^*)^{-k}N^{k/3
}\rho_{N,V,f;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t)
\end{align*}
in \eqref{Gaussian_truncation} in the $L^3$ norm. This completes the case $t\in[q,N^\e]^k$.
Assume now that at least one component of $t$, say $t_j$, is larger than $N^\e$. Then the asymptotics of the Airy kernel (see e.g.
\eqref{KAi_asymp}) imply that all entries of the $j$-th column of $(K_{\textup{Ai}}(t_i,t_l))_{1\leq i,l\leq k}$ and consequently
$\AI_k(t)$ are bounded by $\O(\exp(-N^{\e}))$ (cf. Remark \ref{remark_definiteness_vorne} a)). We are left to prove a similar bound
for the l.h.s.~of \eqref{Ttheorem3}. If one of the components $t_j$ is such that $b+\frac{t_j}{c^*N^{2/3}}>L$, this follows from
\eqref{Lago2} and for the remaining cases the estimate readily follows from Proposition \ref{prop_LD} a), \eqref{Hadamard}
and Proposition \ref{keyproposition} for those components of $t$ which lie in $[q,N^\e]$.
Of course, Proposition \ref{keyproposition} can only be applied for $\| f \| \leq N^{\k}$. One may use the same arguments as in the
case $t\in[q,N^\e]^k$ to bound the contribution of $\{ f : \| f \| > N^{\k}\}$.
\end{proof}
\subsection*{Extension to general $h$}\noindent
If $-h$ is not positive-definite, we may write it as a difference of positive-definite functions. Denoting by $g_{\pm}$
nonnegative and negative
part of a function, we first write $\F{h}=(\F{h})_+-(\F{h})_-$. Setting
$h^\pm$ as the inverse Fourier transform of $\F{h}_\pm$, we get a
decomposition
$h=h^+-h^-$ of $h$ into positive-definite, real-analytic functions. It is this step where the assumption of exponential decay
of $\hat{h}$ is needed. Exponential decay of $\hat{h}$ is equivalent to $h^\pm$ being real-analytic. Sufficiency is easily seen as
the exponential decay allows for an entension of the Fourier representation of $h$ to a strip $D$ from which analyticity of $h^\pm$
can be deduced. For the necessity we remark that any real-analytic positive-definite function has an exponentially decaying Fourier
transform \cite[Theorem 2]{LukacsSzasz} and thus with $h^\pm$ also $h$ must have this property.
Define for a complex parameter $z\in\C$
\begin{align}
\U_z(x)=&\frac{z}{2}\lbb\sum_{i,j=1}^Nh^+(x_i-x_j)-\left[h^+_\mu(x_i)+h^+_\mu(x_j)-h^+_{\mu\mu}\right]\rbb\label{e472}\\
+&\frac{1}{2}\lbb\sum_{i,j=1}^Nh^-(x_i-x_j)-\left[h^-_\mu(x_i)+h^-_\mu(x_j)-h^-_{\mu\mu}\right]\rbb\label{e482}.
\end{align}
We have $\U_{-1}=\U$.
As seen in \eqref{e9}, we have to show
\begin{align}
&(c^*)^{-k}N^{
k/3
}\E_{N-k,V;L}^N\exp\lee{(\U_{z}+R_L)(b+\frac{t}{c^*N^{2/3}},\cdot)}\ree-\E_{N,V;L}\exp\lee{\U_{z}}\ree\AI_k(t)=o(1)\label{e92}
\end{align}
for $z=-1$ uniformly in $t\in[q,\infty)^k$. Let us recall two basic facts
from analysis. As our
linearization procedure only allows for nonnegative real $z$, we will prove
\eqref{e92} for positive real $z$ and use complex analysis to deduce \eqref{e92} also for $z=-1$. Recall Vitali's Theorem for instance from
\cite[5.21]{Titchmarsh}: Let $(f_n)_n$ be a sequence of analytic functions on a domain $U\subset\C$ with $\lv f_n(z)\rv\leq M$ for
all
$n$ and all $z\in U$.
Assume that $\lim_{n\to\infty}f_n(z)$ exists for a set of $z$ having a limit point in $U$. Then $\lim_{n\to\infty}f_n(z)$ exists
for all $z\in U$ and the limit is an analytic function in $z$.
To capture the uniformity of the convergence in \eqref{e92} in a way which preserves analyticity in $z$, we use the following
obvious characterization. A sequence of complex-valued functions $(f_n)_n$ converges uniformly on the
sequence of sets $(A_n)_n$, $A_n\subset \R^l$ towards a function $f$ if and only if for all sequences
$(n_m)_m\subset \mathbb{N}$ with $\lim_{m\to\infty}n_m=\infty$ and all sequences $(t_m)_m$ with $t_m\in A_{n_m}$ we have
$\lim_{m\to\infty}f_{n_m}(t_m)-f(t_m)=0$.
For our application, we take $A:=A_N:=[q,\infty)^k$. Let $(N_m)_m\subset\N$ be a
sequence going to
infinity and $(t_{(m)})_m$ be a sequence with $t_{(m)}\in A$.
Define $\map{W_m}{\C}{\C}$ by
\begin{align}
W_m(z):=&(c^*)^{-k}N_m^{k/3
}\E_{{N_m}-k,V;L}^{N_m}\exp\lee{(\U_{z}+R_L)(b+\frac{t_{(m)}}{c^*{N_m}^{2/3}},\cdot)}\ree\nonumber\\
&-\E_{N_m,V;L}\exp\lee{\U_{z}}
\ree\AI_k(t_{(m)}).\label{W_m}
\end{align}
With these two observations, we are ready to complete the proof of Theorem \ref{ethrm2}.
\begin{proof}[Proof of Theorem \ref{ethrm2} a)]
We start by showing convergence of $W_m(z)$ uniformly for $z\in[0,1]$.
We will mostly omit the index $m$ in the following. Define $-h_z:=zh^++h^-$ such that $h_{-1}=h$, denote by $\tilde{f}_z$
the corresponding Gaussian process, and set $f_z:=\tilde{f}_z-\int \tilde{f}_zd\mu$. Now, using the same arguments given in the
proof of Theorem \ref{ethrm2} b) we find that
\begin{align*}
&\E\Big[\E_{N,V;L}
\exp\lee\sum_{j=1}^{N}f_z({x_j})\ree\Big( (c^*)^{-k}N^{k/3
}\rho_{N,V,f_z;L}^k\lbb
b+\frac{t_1}{c^*N^{2/3}},\dots,b+\frac{t_k}{c^*N^{2/3}}\rbb- \AI_k(t)\Big)\Big]\\
&=\O(N^{-\s}).
\end{align*}
The convergence is uniform in $z\in[0,1]$. This is a consequence of the boundedness of $\sup_{w\in [-\d/3,\d/3]}-h_z(2iw)$ for
$z\in[0,1]$ (cf. the arguments above \eqref{Gaussian_LD}). Note that the choice of $\a_Q$ in the hypothesis of Theorem \ref{ethrm2}
depends on this bound.
This settles the convergence of $W_m$ on $[0,1]$.
We complete the proof by demonstrating uniform boundedness of $W_m$ on the domain $G:=\{z\in\C:\Re z< 1\}$. We first observe that applying \eqref{def_U} and \eqref{Fourier} to $h^+$ instead of $h$ yields the non-negativity of $\sum_{i,j=1}^Nh^+(x_i-x_j)-\left[h^+_\mu(x_i)+h^+_\mu(x_j)-h^+_{\mu\mu}\right]$. Thus $\Re\, \U_z(x)\leq\U_1(x)$ for all $z\in G$ and all $x$ and consequently we have
\begin{align}
\sup_{N}\E_{N,V;L}\lv \exp(\U_z)\rv\leq\sup_{N}\E_{N,V;L}\exp(\U_1)<\infty,\label{U1_bound}
\end{align}
where the last inequality follows by linearization as above. Hence the second term in the definition of $W_m$ is uniformly bounded
on $G$. Furthermore, the uniform boundedness of the first term in \eqref{W_m} follows from $\Re\, \U_z\leq\U_1$ for all $z\in G$,
from the convergence and hence boundedness of $W_m(1)$ and from \eqref{U1_bound}. We are now in a position to apply Vitali's
theorem, providing the desired convergence of $W_m(-1)$ to $0$.
\end{proof}
Finally, we would like to point out a subtlety that can be explained e.g. by looking at relation \eqref{e8}. The $z$-dependence used
above is introduced by twice replacing $\U$ by $\U_z$ on the right hand side of \eqref{e8}. Observe that this is in general not the
same as replacing $h$ by $h_z$ on the left hand side. In fact, we never consider the ensembles $P_{N, Q}^{h_z}$ with $z \neq -1$,
i.e. for $h_z \neq h$. The function $V$ and the associated equilibrium measure $\mu$ do not depend on $z$ either.
\section{Proofs of Results on the Largest Particle}\label{TSec4}
\begin{proof}[Proof of Theorem \ref{ethrm3}]
We have to compute the limit of the gap probability
\begin{align*}
&P_{N,Q}^h\lbb x_j\notin(b+\frac{s}{c^*N^{2/3}},\infty),\quad j=1,\dots,N\rbb.
\end{align*}
This proof uses the same techniques and route as the proof of Theorem \ref{ethrm2}. However, in that proof the truncation threshold
$L$ depends on the order of the correlation function $k$, leading to an $\O$-term depending on $k$ in a non-obvious way, e.g.~in
the expectation of quantities like $\|f\|_D$ or $\Lip{f}$. It is therefore convenient to truncate the event before expanding in
terms of correlation functions. To this end, we recall from the last inequality in the proof of
the truncation lemma \cite[Lemma 6.3]{GoetzeVenker} that
\begin{align}
\rho_{N,Q}^{h,1}(t)\leq \exp\{{CN}-c_1N[V(t)-c_2\log(1+t^2)]\}\label{truncation_Q}
\end{align}
for some positive $C,c_1,c_2$. Hence, if $L$ is chosen large enough, we have for some $c>0$
\begin{align*}
&P_{N,Q}^h\lbb x_j\notin[-L,L]\ \text{ for some }
j\rbb\leq 2N\int_{L}^\infty \rho_{N,Q}^{h,1}(t)dt=\O(e^{-cN}).
\end{align*}
From this we also conclude that the replacement of the normalizing constant $Z_{N,Q}^h$ by $Z_{N,Q;L}^h$
is negligible, where $Z_{N,Q;L}^h$ denotes the normalizing constant for the ensemble $P_{N,Q;L}^h$ that arises from $P_{N,Q}^h$
by restricting it onto $[-L,L]^N$.
We thus have
\begin{align}
&P_{N,Q}^h\lbb x_j\notin(b+\frac{s}{c^*N^{2/3}},\infty) \text{ for all }
j\rbb=P_{N,Q;L}^h\lbb
x_j\notin(b+\frac{s}{c^*N^{2/3}},L) \text{ for all }
j\rbb+\O(e^{-cN}).\label{truncation_probability}
\end{align}
To evaluate the latter probability, we proceed as in the proofs of Theorem \ref{ethrm2} a) and b). We will write
explicitly only the case of negative-definite $h$, the general case follows then with the additional arguments given in the proof
of Theorem \ref{ethrm2} a). Note that the seemingly stronger uniformity for $s\in\R$ in a) is a simple consequence of the
continuity of $F_2$. In order to show
\begin{align*}
P_{N,Q;L}^h\lbb x_j\notin(b+\frac{s}{c^*N^{2/3}},L), j=1,...,N\rbb-F_2(s)=\O(N^{-\s})
\end{align*}
it suffices to prove the bound
\begin{align}\label{Lago5}
\E\,
\big[\E_{N,V;L}\exp\{\sum_{j=1}^Nf(x_j)\}
\lbb P_{N,V,f;L}\lbb
x_j\notin(b+\frac{s}{c^*N^{2/3}},L),j=1,...,N\rbb-F_2(s)\rbb\big]=\O(N^{-\s})
\end{align}
(cf.~\eqref{e11}). In order to see this, integrate \eqref{Lago1} over $[-L,b+\frac{s}{c^*N^{2/3}}]^N$ and use \eqref{Lago4}
together with the lower bound of \eqref{Lago3}. As in the proof of Theorem \ref{ethrm2} b) we choose $\e$ and $\k$ depending on
$\s$. Again we can neglect the contribution of those $f$ to the expectation in \eqref{Lago5} with $\|f\|_D>N^\k$.
We are therefore left to prove
\begin{align}
P_{N,V,f;L}\lbb
x_j\notin(b+\frac{s}{c^*N^{2/3}},L),j=1,...,N\rbb-F_2(s) =\O(N^{-\s})\label{boils_down}
\end{align}
uniformly for all $f \in X_D$ with $\|f\| \leq N^{\k}$ and all $s\in[q,(L-b)c^*N^{2/3}]$.
To this end, we will represent both probabilities as Fredholm determinants. It is well-known (see e.g. \cite{TracyWidom} for a
nice
derivation)
that
\begin{align*}
&P_{N,V,f;L}\lbb
x_j\notin(b+\frac{s}{c^*N^{2/3}},L),j=1,...,N)=\det(I-\hat{\mathcal{K}}_{N,V,f;L})=:\Delta(\hat{\mathcal{K}}_{N,V,f;L}),
\end{align*}
where $\hat{\mathcal{K}}_{N,V,f;L}$ denotes the integral operator on $L^2((s,(L-b)c^*N^{2/3}))$ with kernel
\begin{align*}
\hat{K}_{N,V,f;L}(t_1,t_2) := \frac{1}{N^{2/3}c^*}K_{N,V,f;L}(b+\frac{t_1}{c^*N^{2/3}},b+\frac{t_2}{c^*N^{2/3}}).
\end{align*}
Observe that $\hat{K}$ agrees with the definition in Proposition \ref{keyproposition} since $b=b_V$ and $c^* = \g_V$.
For
$F_2$ we have $F_2(s)=\Delta(\mathcal{K}_\Ai)$ with $\mathcal{K}_\Ai$ (cf. \eqref{F_2_rep}) being the integral operator w.r.t.~the
Airy kernel on
$L^2((s,\infty))$. For comparison, it will turn out convenient to consider Fredholm determinants of integral operators
that are defined on the same $L^2$-space. To this end, define $\mathcal{K}_{\Ai;L}$ as the integral operator on
$L^2((s,(L-b)c^*N^{2/3}))$ w.r.t.~the Airy kernel. Let us show that replacing $F_2(s)$ by $\Delta(\mathcal{K}_{\Ai;L})$ only
results in a negligible error. Let $\zeta$ denote the
determinantal point process on $\R$ determined by the
Airy kernel (see e.g. \cite[Proposition 4.2.30]{AGZ}). Then both $F_2(s)$ and $\Delta(\mathcal{K}_{\Ai;L})$ can be represented as
gap probabilities w.r.t.~$\zeta$. Let $P$ be the probability measure underlying $\zeta$. From the asymptotics
\eqref{TW_asymptotics} of $F_2$
we conclude
\begin{align*}
&\lv F_2(s)-\Delta(\mathcal{K}_{\Ai;L})\rv=\lv P(\zeta\cap(s,\infty)=\emptyset)-P(\zeta\cap(s,(L-b)c^*N^{2/3})=\emptyset)\rv\\
&\leq P(\zeta\cap[(L-b)c^*N^{2/3},\infty)\not=\emptyset )=1-F_2((L-b)c^*N^{2/3})=\O(e^{-cN})
\end{align*}
for some $c>0$.
Thus it suffices to estimate the difference $\Delta(\hat{\mathcal{K}}_{N,V,f;L})-\Delta(\mathcal{K}_{\Ai;L})$.
Here we use again the classic series representation of Fredholm determinants \cite[Def. 3.4.3]{AGZ} and the inequality
\cite[Lemma
3.4.5]{AGZ} on
the difference between the Fredholm determinants of two integral operators $\mathcal{S,Q}$ with bounded kernels $S,Q$ on
an interval $(c,d)$. Let $\nu$ be a finite measure on $(c,d)$ with total mass $\|\nu\|_1$. Let $\Delta(\mathcal
S,\nu):=\det(I-\mathcal S)$ denote the Fredholm determinant of $\mathcal S$ on $L^2((c,d),\nu)$. Then
\begin{align}
\lv \Delta(\mathcal S, \nu)-\Delta(\mathcal Q, \nu)\rv\leq
\left(\sum_{k=1}^\infty\frac{k^{1+k/2}\|\nu\|_1^k\max(\|S\|_\infty,\|Q\|_\infty)^{k-1}}{k!}\right)\|S-Q\|_\infty,
\label{Fredholm_inequ}
\end{align}
where $\|\cdot\|_\infty$ denotes the sup-norm on $(c,d)^2$. A natural choice would be to use the Lebesgue measure for $\nu$ on
the
intervals $(s,(L-b)c^*N^{2/3})$. However, $\|\nu\|_1$ would then be of order $N^{2/3}$ and consequently the series in
\eqref{Fredholm_inequ} could not be bounded uniformly in $N$. For our application one can circumvent this issue by transfering
the fast decay of the kernels onto the measure $\nu$.
In order to do so
we work with $L^2((s,(L-b)c^*N^{2/3}),\nu)$, $d\nu(t):=e^{-2t}dt$ and set
\begin{align*}
S_N(t_1,t_2):= \hat{K}_{N,V,f;L}(t_1, t_2)e^{t_1+t_2}
,\qquad
Q(t_1,t_2):=K_{\Ai}(t_1,t_2)e^{t_1+t_2}.
\end{align*}
Using the above mentioned representation of the Fredholm determinant one immediately obtains
$\Delta({\hat{\mathcal{K}}_{N,V,f;L}})=\Delta({\hat{\mathcal{K}}_{N,V,f;L}},dt)=\Delta(\mathcal{S}_N,\nu)$ and
$\Delta(\mathcal{K}_{\Ai;L}){=\Delta(\mathcal{K}_{\Ai;L},dt)}=\Delta(\mathcal{Q},\nu)$. Observe that $\|\nu\|_1$ is now bounded uniformly in
$N$. {The same holds for
$\|Q\|_\infty$ which follows from the asymptotic behavior} of the Airy kernel as presented e.g.~in \cite[(4.23)]{KSSV}). The uniform
boundedness of $\|S_N\|_\infty$ can be derived
as follows. Note first from the determinantal formula \eqref{determinantal_relations} for $k=2$, from the positivity of the
$2$-point correlation function and from the symmetry of the kernel that
\begin{align}
S_N(t_1,t_2)^2 \leq S_N(t_1,t_1)S_N(t_2,t_2). \label{Testimate1}
\end{align}
The boundedness of $S_N(t,t)$, $t \in (s,(L-b)c^*N^{2/3})$ on the diagonal can in turn be derived from the first two statements of
Proposition \ref{keyproposition} and from Proposition \ref{prop_LD} a), where the asymptotic behavior of the Airy kernel and a
lower bound on
$\eta_V(x)$ of the form $c' (x-1)^{3/2}$ are being used in addition. We have
now shown that the series on the right hand side of \eqref{Fredholm_inequ} is uniformly bounded in $N$ if we choose $\mathcal S_N$
for $\mathcal S$. Establishing \eqref{boils_down} uniformly for $s\in[q,(L-b)c^*N^{2/3}]$ thus reduces to proving $\|S_N-Q\|_\infty = \O(N^{-\s})$. To this end recall $\e=2(\frac{1}{3} - \s - \k) > 0$. We distinguish two cases. For $t_1$, $t_2 \in [q, N^{\e}]$ Proposition \ref{keyproposition}
yields the desired bound. In the case
that one of the variables is larger than $N^{\e}$, the crude estimate $|S_N-Q| \leq |S_N| + |Q|$ together with
\eqref{Testimate1} and Proposition \ref{prop_LD} a) provides a bound of the form $\O(e^{-cN^{3\e /2}})$ that suffices easily.
\end{proof}
We continue with the proof of our result for the regimes of moderate and large deviations for the upper tail of the distribution of
the largest
particle, which combines the analysis devised in \cite{DissSchueler,EKS} and the asymptotic results of \cite{KSSV} with the
procedure of the previous proof.
\begin{proof}[Proof of Theorem \ref{thrm_MD}]
We start by observing that it suffices to prove the statements of Theorem \ref{thrm_MD} for $N \geq N_0$ and in the cases of
statements a), c) for $t \in (b + C N^{-2/3}, T)$ for arbitrarily large but fixed constants $N_0$ and $C$. Let us first turn our
attention to statement c). We begin with the truncation of the ensemble to $[-L,L]$ as in \eqref{truncation_probability}. It
follows from \eqref{truncation_Q} that the constant $c>0$ in the error $\O(e^{-cN})$ of that truncation can be chosen arbitrarily
large by increasing $L$. Thus, we may choose $L > T+1$ so large (depending on the choice of $T$ in the statement of the
theorem) that we have
for all $t \in (b + N^{-2/3}, T)$:
\begin{align}\label{p6b5}
\frac{P_{N,Q}^h\lbb
x_{\max}>t\rbb}{\FNV (t)} = \frac{\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U(x)}\right]}{\E_{N,V;L}e^{\U(x)}\FNV(t)}
\left[ 1 +\O \lb e^{-cN} \rb
\right]
\end{align}
for some new positive constant $c$ and with $A:=\{x\in\R^N\,:\,x_{\max}>t\}$.
Now we apply the stochastic linearization to rewrite the numerator and obtain (use in particular the relation for the partition
functions provided by \eqref{Lago4})
\begin{align}\label{p6b10}
\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U(x)}\right]=\E\left[
\E_{N,V;L}e^{\sum_{j=1}^Nf(x_j)}P_{N,V,f;L}(x_{\max}>t)\right].
\end{align}
In view of \eqref{p6b5}, \eqref{p6b10} we first derive upper and lower estimates on $P_{N,V,f;L}(x_{\max}>t) / \FNV(t)$.
Since $P_{N,V,f;L}$ is determinantal we have
\begin{align}\label{p6b15}
P_{N,V,f;L}(x_{\max}>t) = 1 - \Delta(\mathcal{K}_{N, V, f; L}) ,
\end{align}
where $\Delta(\mathcal{K}_{N,V,f;L})$ denotes the Fredholm determinant related to the integral operator with kernel $K_{N,V,f;L}$
and acting on $L^2((t,L),dx)$. Next we apply Proposition \ref{prop_LD} b). The formula there holds for all $t \in (b + C_0
N^{-2/3}, T)$ (since we have ensured above that $L > T +1$) and for all $f \in X_D$ satisfying $\|f\| \equiv \|f\|_{D}\leq
c_0N(t-b)$, where $c_0$, $C_0$ denote some positive constants. In this regime a first simple consequence is
\begin{align}\nonumber
\int_t^L \!\! K_{N,V,f;L}(y,y) dy=\O \lb
\frac{1}{N\lv t-b\rv^{3/2}} \rb,
\end{align}
where one should use from the proof of Proposition \ref{prop_LD} b) that
\begin{align}\nonumber
e^{-N\eta_V(t/b_V) + \O \lb \sqrt{t-b_V} \|f\| \rb} = e^{-N\eta_{V,f}\lb\l_{V,f}^{-1}(t)\rb} \leq 1.
\end{align}
The definition of the Fredholm determinant as in \cite[Def. 3.4.3]{AGZ}, Hadamard's inequality (see
\eqref{THad}) and Proposition \ref{prop_LD} b) then yield the finer estimates
\begin{align}\label{p6b20}
1 - \Delta(\mathcal{K}_{N, V, f; L}) &= \lb \int_t^L \!\! K_{N,V,f;L}(y,y) dy\rb \left[ 1 + \O \lb
\frac{1}{N\lv t-b\rv^{3/2}} \rb
\right]\\
&= \FNV(t) e^{\O \lb \sqrt{t-b_V} \|f\| \rb}
\left[1
+ \O \! \lb \frac{\|f\|}{N\lv t-b\rv} \rb
+\O \! \lb
\frac{1}{N\lv t-b\rv^{3/2}} \! \rb
\right] .
\nonumber
\end{align}
Recalling \eqref{p6b15} we have derived for all $t \in (b + C_0 N^{-2/3}, T)$ the upper bound
\begin{align}\label{p6b25}
\frac{P_{N,V,f;L}(x_{\max}>t)}{\FNV(t)} \leq 1 + \g(\|f\|) , \quad \text{with}
\end{align}
\begin{align}\label{p6b30}
\g(r) :=
\begin{cases}
\frac{C}{N \lv t-b\rv^{3/2}}&\text{for} \quad 0 \leq r < \frac{1}{N \lv t-b\rv^2} \\
Cr\sqrt{t-b}&\text{for} \quad \frac{1}{N \lv t-b\rv^2} \leq r < \frac{1}{\sqrt{t-b}} \\
e^{Cr\sqrt{t-b}}&\text{for} \quad \frac{1}{\sqrt{t-b}} \leq r < c_0N(t-b) \\
\frac{1}{\FNV(t)}&\text{for} \quad c_0N(t-b) \leq r
\end{cases}
\end{align}
for some suitable constant $C$. In the case $r \geq c_0N(t-b)$ we have used that the numerator of the left hand side of
\eqref{p6b25} is a probability and thus bounded by $1$.
We now formulate a basic integration lemma that is convenient for estimating the right hand side of \eqref{p6b10}. To this end
we introduce a function $F$ that is related to the distribution function for $\|f\|$ induced by the Gaussian measure on $f$ and
modified by the positive density $\E_{N,V;L}e^{\sum_{j=1}^Nf(x_j)}$:
\begin{align}\label{p6b35}
F(\a) := \E \lb \dopp{1}_{\{\|f\|_D\geq \a\}} \E_{N,V;L} e^{\sum_{j=1}^N f(x_j)} \rb.
\end{align}
\begin{lemma}\label{Lp6b}
Let $0 \leq \a < \b \leq \infty$ and assume that $\g : [\a, \b) \to [0, \infty)$ is a continuously differentiable function. Then
\begin{align}\label{p6b40}
\E \lb \dopp{1}_{\{\a \leq \|f\|_D < \b\}} \E_{N,V;L} e^{\sum_{j=1}^N f(x_j)} (1 + \g(\|f\|)) \rb \leq F(\a)-F(\b) + \g(\a) F(\a) + \int_{\a}^{\b} \!\! F(r) \g'(r) dr.
\end{align}
\end{lemma}
To see this, observe first that the left hand side of \eqref{p6b40} is given by the Riemann-Stieltjes integral
\begin{align}\nonumber
-\int_{\a}^{\b} 1 + \g(r) dF(r) = - (1+\g) F \mid_{\a}^{\b} + \int_{\a}^{\b} F(r) \g'(r) dr
\end{align}
after integration by parts. Using $-\g(\b) F(\b) \leq 0$ completes the derivation of \eqref{p6b40}.
We apply Lemma \ref{Lp6b} on four intervals $[\a, \b) = [\a_{j-1}, \a_j)$ for $j =1$, $2$, $3$, $4$ according to the cases in \eqref{p6b30}, i.e.
$\a_0:=0$, $\a_1:=\frac{1}{N \lv t-b\rv^2}$, $\a_2:=\frac{1}{\sqrt{t-b}}$, $\a_3:=c_0N(t-b)$, and $\a_4:=\infty$. Combining
\eqref{p6b5}, \eqref{p6b10}, \eqref{p6b25}, \eqref{p6b30} and using $\E_{N,V;L}e^{\U(x)} = F(0)$ (see \eqref{e7}) we obtain
\begin{align}\nonumber
\frac{\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U(x)}\right]}{\E_{N,V;L}e^{\U(x)}\FNV(t)} \leq 1 &+ \g(0) +
\sum_{j=1}^{3}\g(\a_j)\frac{F(a_j) }{F(0)}
+ C\sqrt{t-b} \int_{\a_1}^{\a_2} \frac{F(r)}{F(0)} dr \\ &+ C\sqrt{t-b} \int_{\a_2}^{\a_3} e^{Cr\sqrt{t-b}}\frac{F(r)}{F(0)} dr.\nonumber
\end{align}
We are left to show that the last six summands on the right hand side are of the form $\O \left(1/(N \lv t-b\rv^{3/2})\right)$ or
of the form $\O (\sqrt{t-b})$. Since $\g(0)$ and $\g(\a_1)$ are of the first form and trivially $F(\a_1)/F(0) < 1$ the first two of
these summands are good. Since $\int_{\a_1}^{\a_2} F(r) dr \leq F(0)$ it is clear that the first integral in the sum is of the
second form. For the remaining three terms we use that there exist positive constants $d$, $K$ such that $F(r)/F(0) \leq
K e^{-dr^2}$ for all $r \geq 0$ and for all $N$. To see this, use the
sub-Gaussianity of $\|f\|_D$ analogously to the derivation of
\eqref{Gaussian_truncation} together with the lower bound on $F(0)=\E_{N,V;L}e^\U$ provided by \eqref{Lago3}.
Then it follows
\begin{align}\nonumber
\g(\a_2) F(\a_2)/F(0) = \O \lb e^{-d/\lv t-b\rv} \rb = \O (\sqrt{t-b}).
\end{align}
Moreover, setting $x:=N \lv t-b\rv^{3/2} \geq C_0 \geq 1$, $y:=x^{4/3}$ and $c_1:= d c_0^2$, we have
\begin{align}\nonumber
\g(\a_3) F(\a_3)/F(0) = \O \lb x e^{x} e^{-c_1 N^{2/3}y} \rb = \O \lb e^{(2-c_1 N^{2/3})y} \rb = \O \lb e^{-c_1 N^{2/3}/2} \rb
\end{align}
for $N$ sufficiently large (e.g. $N^{2/3} > 4/c_1$ would do) and is therefore of the first form. Finally, we consider the last integral in the sum. It can be estimated by
\begin{align}\nonumber
\O \lb \sqrt{t-b} \int_{\a_2}^{\infty} e^{-d r^2/2}dr \rb = \O \lb \sqrt{t-b} /( d \a_2) \rb = \O \lb \sqrt{t-b} \rb.
\end{align}
This completes the analysis of the upper bound. The lower bound is proved in exactly the same way. One
only needs to replace $1+\g$ by $1-\g$. In summary we obtain
\begin{align}\label{Testim5}
\frac{\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U(x)}\right]}{\E_{N,V;L}e^{\U(x)}\FNV(t)} = 1 + \O\lb
\frac{1}{N\lv t-b\rv^{3/2}}\rb +\O\lb \sqrt{t-b} \rb
\end{align}
for all $t \in (b + C_0 N^{-2/3}, T)$. In view of the remark at the very beginning of the proof and using \eqref{p6b5} we have
established statement c).
We turn our attention to the case of general functions $h$ that are not necessarily negative-definite. Observe first that
\eqref{Testim5} still holds if $U$ is replaced by $U_z$ for any positive $z$ where we do not insist on uniformity of the $\O$-terms
in $z$.
To see this one may use exactly the same proof of \eqref{Testim5} as above with $F$ being replaced $F_z$ which is defined via \eqref{p6b35} with the Gaussian process $f_z$ instead of $f$. One may again show that the tails of $F_z$ are sub-Gaussian and that $\inf_N F_z(0) > 0$. These were the crucial properties of $F$ used in the derivation of \eqref{Testim5}. Using in addition \eqref{U1_bound} we conclude
\begin{align}
\frac{\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U_z(x)}\right]}{\FNV(t)}-\E_{N,V;L}e^{\U_z} \to 0 \quad \text{for} \ \ N \to \infty \label{deviations_Vitali2}
\end{align}
for $0<z\leq1$ and for those values of $t$ that are considered in statement b) of the theorem. For later use we note that by the same reasoning, the left hand side of \eqref{deviations_Vitali2} remains bounded for $0<z\leq1$ and all $t \in (b + C_0 N^{-2/3}, T)$. In order to apply Vitali's theorem and in view of \eqref{U1_bound} it suffices to show boundedness of
\begin{align}\label{deviations_Vitali}
\frac{\E_{N,V;L}\left[\dopp{1}_A(x)e^{\U_z(x)}\right]}{\FNV(t)}
\end{align}
e.g. on $G \equiv \{ z \in \C : \Re z < 1\}$. Using the pointwise estimate $\Re\, \U_z(x)\leq\U_1(x)$ for all $z\in G$ this follows
from \eqref{deviations_Vitali2} for $z=1$ and from \eqref{U1_bound}. Vitali now yields convergence \eqref{deviations_Vitali2} also
for $z=-1$, i.e. for $U_z$ being replaced by $U$. The lower bound of \eqref{Lago3} for $\l =1$ together with \eqref{p6b5} prove
statement b).
Observe that we have already shown the upper bound of a). In fact, by \eqref{p6b5} and by the lower bound of \eqref{Lago3} for $\l
=1$ we are only required to prove an upper bound on \eqref{deviations_Vitali} for $z=-1$. This can be achieved as above where one
needs to use the boundedness of \eqref{deviations_Vitali2} on all of $(b + C_0 N^{-2/3}, T)$.
We establish the lower bound by contradiction. Assume that there is no such bound. Then, again by \eqref{p6b5}, for any fixed
(large) $C > 1$ there is a sequence $(t_N)_N$ in $(b + C N^{-2/3}, T)$
such that
\begin{align}
\lim_{N\to\infty}\frac{\E_{N,V;L}\left[\dopp{1}_{A_N}(x)e^{\U(x)}\right]}{\E_{N,V;L}e^{\U(x)}\FNV(t_N)}=0,\label{contradiction}
\end{align}
where $A_N:=\{x\in\R^N\,:\,x_{\max}>t_N\}$.
By monotonicity of $\U_z(x)$ for real values of $z$ we have for $z\in (-\infty,-1]$
\begin{align*}
\frac{\E_{N,V;L}\left[\dopp{1}_{A_N}(x)e^{\U(x)}\right]}{\FNV(t_N)}\geq
\frac{\E_{N,V;L}\left[\dopp{1}_{A_N}(x)e^{\U_z(x)}\right]}{\FNV(t_N)}
\end{align*}
and thus, by the boundedness of $\E_{N,V;L}e^\U$, both sides converge to 0. By Vitali's theorem, this
convergence extends to $z \in (0, 1)$ as well. However, we will prove that for $z=1/2$, \eqref{deviations_Vitali} is bounded
away
from 0 uniformly in $N$ and $t \in (b + C N^{-2/3}, T)$ for sufficiently large values of $C$, thus providing the desired
contradiction.
Recall the definition of $F_{1/2}$ below \eqref{Testim5} and
choose $\b$ such that $F_{1/2}(\b) \leq F_{1/2}(0)/2$.
In view of \eqref{p6b15} and \eqref{p6b20} there exists $C>0$ such that we have for all $t \in (b + C N^{-2/3}, T)$:
\begin{align}\nonumber
\tilde{d}(T) := \inf_{\|f\| \leq \b} P_{N,V,f;L}(x_{\max}>t) / \FNV(t) > 0.
\end{align}
Thus the right hand side of \eqref{p6b10} with $\U$ replaced by $\U_{1/2}$ and devided by $\FNV(t)$ is bounded below by
$\tilde{d}(T)(1-F_{1/2}(\b)) \geq \tilde{d}(T) F_{1/2}(0)/2$. Since in addition $F_{1/2}(0) \geq F(0) =\E_{N,V;L}e^\U$ is bounded
away from $0$ we have arrived at the advertised lower bound on \eqref{deviations_Vitali} for $z=1/2$, completing the proof of
Theorem \ref{thrm_MD}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{a.s.convergence}]
The corollary follows from an application of the Borel-Cantelli lemma. We thus have to show for any fixed $\e>0$
\begin{align*}
\sum_{N=1}^\infty P_{N,Q}^h(\lv x_{\max}-b\rv>\e)<\infty.
\end{align*}
The necessary estimate for the probability of $x_{\max}>b+\e$ is provided by Theorem \ref{thrm_MD} a).
It
remains to derive a bound for the lower tail of the distribution of $x_{\max}$. This can be done as follows. Given $\e>0$, let
$g$ be
a smooth, nonnegative Lipschitz function with $\int_{\R} g d\mu=1$ and with support $[b-\e+\d,b-\d]$, where $0<\d<\e/2$. Then
\begin{align*}
P_{N,Q}^h(x_{\max}<b-\e)\leq P_{N,Q}^h\lbb\sum_{j=1}^Ng(x_j)=0\rbb\leq P_{N,Q}^h\lbb \sum_{j=1}^Ng(x_j)-N \int
gd\mu\geq N\rbb.
\end{align*}
With \eqref{relation}, \eqref{M1}, and \eqref{concentration_U} an application of H\"older's inequality gives
\begin{align*}
P_{N,Q}^h\lbb\sum_{j=1}^Ng(x_j)-N\int
gd\mu\geq N\rbb\leq C\left[P_{N,V}\lbb\sum_{j=1}^Ng(x_j)-N\int
gd\mu \geq N\rbb\right]^{1/\l}
\end{align*}
for some $C>0$ and some $\l>1$.
An application of
Chebyshev's inequality to the statement of Proposition \ref{Concentration} with a good choice of $\e$ shows that this last
probability is of order
$e^{-cN^2}$ for
some $c>0$, provided that $N$ is large enough. The corollary is proved.
\end{proof}
|
\section{Introduction}
The \texttt{iopart-num} Bib\TeX{} style is intended for use in
preparing manuscripts for Institute of Physics Publishing journals,
including Journal of Physics. It provides numeric citation with
Harvard-like formatting, based upon the specification in ``How to
prepare and submit an article for publication in an IOP journal using
\LaTeXe'' by Graham Douglas (2005).
The \texttt{iopart-num} package is available on the Comprehensive
\TeX{} Archive Network (CTAN) as \texttt{/biblio/bibtex/contrib/iopart-num}.
\section{General instructions}
To use the \texttt{iopart-num} style, include the command
\verb+\bibliographystyle{iopart-num}+ in the
document preamble. The reference section is then inserted into the
document with the command \verb+
|
\section{Introduction}
\label{sec:intro}
Processes in hadron collisions where two partons from each hadron take part in separate partonic subprocesses, double parton scattering (DPS), contribute to several final states of interest at the LHC. DPS is a relevant background to precise Higgs boson coupling measurements and searches for new physics \cite{DelFabbro:1999tf,Hussein:2006xr,Bandurin:2010gn,Baer:2013yha}. The theory for DPS is still fragmentary, but major improvements have been made over the last couple of years moving towards a reliable description within perturbative QCD \cite{Manohar:2012jr,Diehl:2011tt,Blok:2010ge,Bansal:2014paa}. Despite this development there are still several important questions which have to be worked out.
For sufficiently inclusive cross sections DPS is formally a power suppressed contribution, but in certain regions of phase space double and single parton scattering contribute at the same power \cite{Diehl:2011yj}. Even for inclusive cross sections, DPS can in specific situations compete with single parton scattering - for example when the single parton scattering is suppressed by multiple small coupling constants. DPS is increasingly relevant with collider energy, and will hence be further enhanced when the LHC restarts to collide protons at larger center of mass energies. The reason is the rapid increase of the density of partons with energy and towards smaller $x$-fractions.
DPS signals have been measured at the LHC by both ATLAS \cite{Aad:2013bjm} and CMS \cite{Chatrchyan:2013xxa} in the $W$-boson plus dijet final state. Of particular interest for our present study is the LHCb measurements of double open-charm production \cite{Aaij:2012dz}, in final states such as $D^0D^0$. Among the most promising channels for a clean separation of double from single parton scattering are the production of two same sign $W$-bosons and double open charm quarks \cite{Luszczak:2011zp,Cazaroto:2013fua,Gaunt:2014rua,Gaunt:2010pi,vanHameren:2014ava,Maciula:2013kd,Golec-Biernat:2014nsa}. In fact, studies have shown that for double open-charm production in the kinematical region of the LHCb measurement, double parton scattering dominates over single parton scattering \cite{vanHameren:2014ava}.
DPS cross sections are factorized into two hard partonic subprocesses and two double parton distributions (DPDs). Only little is known about the size of the DPDs. They have been studied in a variety of quark models \cite{Rinaldi:2014ddl,Rinaldi:2013vpa,Broniowski:2013xba,Chang:2012nw}, including correlations between the two partons inside the same proton. The correlations have generally been found to be sizable. An open question in double parton scattering is the effects of quantum-number correlations. These include correlations between the spins, colors, flavors and fermion numbers of the partons \cite{Kasemets:2014yna,Manohar:2012jr,Diehl:2011yj,Mekhfi:1988kj}. Upper limits on the DPDs describing quantum-number correlations have been derived \cite{Kasemets:2014yna, Diehl:2013mla}. For polarized DPDs these limits have further been shown to hold under radiative corrections from the leading-order double DGLAP evolution up to higher scales.
In particular the spin correlations (described by polarized DPDs) have direct relations to the directions of the final state particles, and thus have the potential to change both the sizes of the DPS cross sections and the distributions of the produced particles. For example, azimuthal modulations have been found for double vector boson production \cite{Kasemets:2012pr}. The effects of the quantum correlations on DPS cross sections have been calculated \cite{Kasemets:2012pr,Manohar:2012jr} but so far no numerical results at the cross section level have been obtained. Through studies of the scale evolution of the DPDs, limits on the degree of polarization and thereby its possible effect on DPS cross sections at different scales were set in \cite{Diehl:2013mla}.
In this paper we examine the effect that polarization in DPS can have on the double $c\bar{c}$ production in kinematic regions resembling those of the LHCb $D^0D^0$ measurement \cite{Aaij:2012dz}. Several studies of this process already exist in the literature, but so far all have neglected the possibility of spin correlations. We demonstrate for the first time the quantitative impact of polarization on any DPS cross section.
The structure of the paper is as follows: In \sect{sec:gDPD} we discuss some basics of DPS with focus on polarization, introduce the different polarized and unpolarized double gluon distributions and discuss their scale evolution. In \sect{sec:xsec} we present the analytical results for the cross section calculation including all possible polarizations of two gluons in an unpolarized proton. In \sect{sec:model} we discuss the models for the DPDs which we use in order to obtain numerical results -- which we present and compare to LHCb data in \sect{sec:data}. We summarize our findings and discuss their implications in \sect{sec:concl}.
\section{Double gluon distributions}
\label{sec:gDPD}
Under the assumption of factorization, as illustrated in figure \ref{fig:DPS}, the DPS cross section can be expressed schematically as
\begin{align}\label{eq:xsec}
\frac{d\sigma}{\prod_{i=1}^2 dx_i d\bar{x}_i }\Bigg|_{DPS} & = \frac{1}{C} \hat{\sigma}_1\hat{\sigma}_2 \int d^2\vek{y}\ F(x_1,x_2,\vek{y}) \bar{F}(\bar{x}_1,\bar{x}_2,\vek{y}) \, ,
\end{align}
where $\hat{\sigma}_i$ represents hard subprocess $i$ and $C$ is a combinatorial factor equal to two (one) if the final states of the two subprocesses are (not) identical. $F$ ($\bar{F}$) labels the double parton distribution of the proton with momentum $p$ ($\bar{p}$). The DPDs depend on the longitudinal momentum fractions of the two partons $x_i$ ($\bar{x}_i$) and their transverse separation $\vek{y}$. No complete proof for factorization in DPS exists, but several important ingredients have been established \cite{Manohar:2012jr,Diehl:2011yj}. The cross section expression \eqref{eq:xsec} is schematic as the labels for the different flavors, colors, fermion numbers and spins of the four partons are implicit. The possibility of interference between the two hard interactions, and correlations between the two partons inside each proton renders this structure significantly more complicated in DPS than for the case with only one hard interaction. Of particular interest for our purposes are the correlations between the spins of two gluons, and between the spins and the transverse separation, which lead to polarized gluon DPDs.
\begin{figure}[htb]
\centering
\includegraphics[width=0.45\textwidth]{%
Plots/DoubleGluon2.eps}
\caption{\label{fig:DPS} Gluon induced double parton scattering. The green fields represents the two DPDs while the blue and red fields represent the first and second hard interaction. $x_i$ ($\bar{x}_i$) are the longitudinal momentum fractions of the partons from the proton with momentum $p$ ($\bar{p}$). $\vek{y}$ and $\mathbf{0}$ are the transverse positions of the first and second hard interaction.}
\end{figure}
For the DPDs describing two gluons in an unpolarized right-moving proton we
write \cite{Diehl:2011yj}
\begin{align}
\label{eq:dpds}
F_{a_1a_2}(x_1,x_2,\vek{y})
& = 2p^+ (x_1\mskip 1.5mu p^+)^{-1}\, (x_2\mskip 1.5mu p^+)^{-1}
\int \frac{dz^-_1}{2\pi}\, \frac{dz^-_2}{2\pi}\, dy^-\;
e^{i\mskip 1.5mu ( x_1^{} z_1^- + x_2^{} z_2^-)\mskip 1.5mu p^+}
\nonumber \\
& \quad \times \left<p|\, \mathcal{O}_{a_2}(0,z_2)\,
\mathcal{O}_{a_1}(y,z_1) \,|p\right> \,.
\end{align}
We use light-cone coordinates $v^\pm = (v^0 \pm v^3) /\sqrt{2}$, bold font to denote the
transverse component $\vek{v} = (v^1, v^2)$ of any four-vector $v$ and $v_T=|\vek{v}|$.
The operators expressed in terms of the gluon field strength tensor reads
\begin{align}
\label{eq:gluon-ops}
\mathcal{O}_{a_i}(y,z_i)
&= \Pi_{a_i}^{jj'} \, G^{+j'}\bigl( y - {\textstyle\frac{1}{2}} z_i \bigr)\,
G^{+j}\bigl( y + {\textstyle\frac{1}{2}} z_i \bigr)
\Big|_{z_i^+ = y^+_{\phantom{i}} = 0,\; \vek{z}_i^{} = \vek{0}} \, ,
\end{align}
with projections
\begin{align}
\label{eq:gluon-proj}
\Pi_g^{jj'} &= \delta^{jj'} \,,
&
\Pi_{\Delta g}^{jj'} &= i\epsilon^{jj'} \,,
&
[\Pi_{\delta g}^{kk'}]^{jj'} &= \tau^{jj'\!,kk'}
\end{align}
onto unpolarized gluons ($g$), longitudinally polarized gluons ($\Delta
g$) and linearly polarized gluons ($\delta g$). The tensor
\begin{align}
\tau^{jj'\!,kk'} = {\textstyle\frac{1}{2}} \mskip 1.5mu \bigl( \delta^{jk}\delta^{j'k'}
+ \delta^{jk'}\delta^{j'k} - \delta^{jj'}\delta^{kk'} \bigr)
\end{align}
satisfies $\tau^{jj'\!,kk'} \tau^{kk'\!,\,ll'} = \tau^{jj'\!,\,ll'}$ and
is symmetric and traceless in each of the index pairs $(jj')$ and $(kk')$.
A decomposition of the
nonzero distributions for two gluons in terms of real-valued scalar
functions has been given in \cite{Diehl:2013mla}
\begin{align}
\label{eq:def-gg}
F_{gg}(x_1,x_2,\vek{y}) & = f_{gg}(x_1,x_2,\vek{y}) \,,
\nonumber\\
F_{\Delta g \Delta g}(x_1,x_2,\vek{y}) & =
f_{\Delta g \Delta g}(x_1,x_2,\vek{y}) \,,
\nonumber\\
F_{g \mskip 1.5mu \delta g}^{jj'}(x_1,x_2,\vek{y}) & =
\tau^{jj'\!,\vek{y}\y} M^2 f_{g \mskip 1.5mu \delta g}(x_1,x_2,\vek{y}) \,,
\nonumber\\
F_{\delta g \mskip 1.5mu g}^{jj'}(x_1,x_2,\vek{y}) & =
\tau^{jj'\!,\vek{y}\y} M^2 f_{\delta g \mskip 1.5mu g}(x_1,x_2,\vek{y}) \,,
\nonumber\\
F_{\delta g \delta g}^{jj',kk'}(x_1,x_2,\vek{y}) & =
{\textstyle\frac{1}{2}}\mskip 1.5mu \tau^{jj'\!,\,kk'} f_{\delta g \delta g}(x_1,x_2,\vek{y})
+ \bigl( \tau^{jj'\!,\vek{y}\tilde{\vek{y}}} \tau^{kk'\!,\vek{y}\tilde{\vek{y}}}
- \tau^{jj'\!,\vek{y}\y} \tau^{kk'\!,\vek{y}\y} \bigr) \,
M^4 f_{\delta g \delta g}^t(x_1,x_2,\vek{y}) \, ,
\end{align}
where $M$ is the proton mass and $\tilde{\vek{y}}^j = \epsilon^{jj'} \vek{y}^{j'}$. The notation where vectors $\vek{y}$ or
$\tilde{\vek{y}}$ appear as an index of $\tau$ denote contraction, i.e.\
$\tau^{jj'\!,\vek{y}\y} = \tau^{jj'\!,kk'}\, \vek{y}^k \vek{y}^{k'}$ etc. The distributions of longitudinally polarized gluons carry open transverse indices $j,j',k,k'=\{1,2\}$ corresponding to the polarization vectors of the gluons which are contracted with the partonic cross sections.
The double parton distribution $f_{gg}$ represents the probability of finding two gluons with momentum fractions $x_1$ and $x_2$ at a transverse separation $\vek{y}$. The distribution of longitudinally polarized gluons $f_{\Delta g \Delta g}$ describe the difference in probability between finding the two gluons with their helicities aligned rather than anti-aligned, while linearly polarized gluons are described by helicity interference distributions, see for example \cite{Kasemets:2013nma,Diehl:2013mla} in the context of DPS.
\subsection{Evolution of the double gluon distributions}
The scale evolution of the DPDs is governed by a generalization of the DGLAP evolution equations. Two versions exist in the literature: one homogenous equation describing two independent branchings of the two partons, and another including the splitting of a parent parton into the two partons which subsequently undergo hard scatterings \cite{Kirschner:1979im,Shelest:1982dg,Snigirev:2003cq,Gaunt:2009re,Ceccopieri:2010kg}. Which one is the correct one for describing DPS is still under debate \cite{Gaunt:2011xd,Gaunt:2012dd,Diehl:2011tt,Diehl:2011yj,%
Ryskin:2011kk,Ryskin:2012qx,Manohar:2012pe,Blok:2011bu,Blok:2013bpa,Snigirev:2014eua,Ceccopieri:2014ufa}. The contribution from the splitting term was investigated in \cite{Gaunt:2014rua} for double $c\bar{c}$ production and was seen to give a sizable contribution to the cross section, but also that the perturbative splitting preferred to take place early on -- and evolve as two separate branches for most of the evolution range. Including such a term in our study could naturally lead to an enhancement of the effect of the polarization and we will return to this discussion in section \ref{sec:concl}. In the following we will make use of the homogeneous version, under the assumption that the physics of the single parton splitting contribution can be treated separately.
The evolution equation for the unpolarized double
gluon distribution then reads
\begin{align}
\label{eq:evol-parton-1}
\frac{d f_{gg}(x_1,x_2,\vek{y}; \mu)}{d \ln\mu^2}
& = \frac{\alpha_s}{2\pi}\sum_{a=g,q,\bar{q}} \left[ P_{ga} \otimes_1 f_{ag} + P_{ga} \otimes_2 f_{ga} \right] \,,
\end{align}
where
\begin{align}
\label{eq:otim}
P_{ab}( \, .\, ) \otimes_1
f_{bc}( \, .\, ,x_2,\vek{y};\mu)
& = \int_{x_1}^{1-x_2} \frac{d u_1}{u_1}
P_{ab}\left( \frac{x_1}{u_1} \right)
f_{bc}(u_1,x_2,\vek{y};\mu)\,,
\nonumber\\
P_{ab}( \, .\, ) \otimes_2
f_{cb}( x_1, \, .\, ,\vek{y};\mu)
& = \int_{x_2}^{1-x_1} \frac{d u_2}{u_2}
P_{ab}\left( \frac{x_2}{u_2} \right)
f_{bc}(x_1,u_2,\vek{y};\mu)
\end{align}
are convolutions in the first and second argument of the DPDs with the leading-order
splitting kernels $P_{ab}$ known from DGLAP evolution of single-parton
distributions. Polarized DPDs follow equivalent evolution equations with the splitting kernels replaced by their polarized analogues. A more thorough discussion of the evolution of the polarized DPDs and expressions for all splitting kernels are given in \cite{Diehl:2013mla}. The evolution of the unpolarized gluon distribution leads to a violent increase at low momentum fractions, in particular at low scales where the QCD coupling constant is large. This is due to the $1/x$ behavior of the unpolarized splitting kernel in the limit where $x$ tends to zero. The splitting kernel for a longitudinally polarized gluon on the other hand approaches a constant in this limit, while the one for linearly polarized gluons goes as $x$. The polarized distributions therefore do not experience this rapid increase and evolution will suppress the relevance of polarized gluons -- in particular the linearly polarized ones. The rate at which this suppression takes effect leads to the expectation that at large scales (and not too large $x$) polarized gluons can be neglected in phenomenological calculations of DPS cross sections \cite{Diehl:2014vaa}. However, for double $c\bar{c}$ production the scales are low and there is only little room for evolution. This motivates the study of the effects of polarization in this process, and could, when confronted with experimental results lead to the first measurements of, or limits on, polarization effects in DPS.
\section{Double $c\bar{c}$ cross sections}
\label{sec:xsec}
We next present the analytic results of the cross section calculation, dividing the results into contributions from the different polarizations. The non-zero results come from gluons which are unpolarized, longitudinally polarized, mixed unpolarized-linearly polarized and purely linearly polarized.
Following experimental conventions, we present our results in the center-of-mass (CM) frame of the two protons, with $\hat{z}$-axis along the proton with momentum $p$ and $\hat{x}$-axis as pointing towards the centre of the LHC ring. With this choice of $\hat{x}$-axis, without reference to any direction defined by the process itself, any azimuthal dependence must show up as differences between the azimuthal angles describing the transverse directions of the final state particles. Double $c\bar{c}$ production in the kinematic region of interest is, to good approximation, initiated by gluons (see e.g. \cite{vanHameren:2014ava}). Therefore we limit ourselves to the partonic subprocesses of figure~\ref{fig:hard}, where the $c\bar{c}$ systems are produced by $s$-channel gluons or t-channel (u-chanel) charm quarks.
\begin{figure}[htb]
\centering
\includegraphics[width=0.2\textwidth]{%
Plots/ccbarS.eps}
\includegraphics[width=0.2\textwidth]{%
Plots/ccbarT.eps}
\includegraphics[width=0.2\textwidth]{%
Plots/ccbarU.eps}
\caption{\label{fig:hard} Leading order Feynman diagrams contributing to each of the two partonic subprocesses. }
\end{figure}
The cross section contribution from unpolarized gluons reads
\begin{align}
\frac{d\sigma_{(gg)(gg)}}{dy_1dy_2d^2\vek{p}_{1}d^2\vek{p}_2} & = \frac{1}{2} \left(8\pi^2\alpha_s^2\frac{N_c}{N_c^2-1}\right)^2 \int dx_1dx_2
\prod_{i=1}^2\frac{1}{16\pi^2\hat{s}_i^2} \frac{2x_i\bar{x}_i}{2x_i-x_{Ti}e^{y_i}}
\nonumber\\
& \quad\times \frac{(1-z_1)^2+z_1^2-1/N_c^{2}}{(1-z_1)z_1}
\left[(1-z_1^2)^2+z_1^2+4z_1(1-z_1) \Big(1-\frac{m^2}{m_{T1}^2}\Big)\frac{m^2}{m_{T1}^2}\right]
\nonumber\\
& \quad\times \frac{ (1-z_2)^2+z_2^2-1/N_c^2 }{(1-z_2)z_2}
\left[(1-z_2^2)^2+z_2^2+4z_2(1-z_2) \left(1-\frac{m^2}{m_{T2}^2}\right)\frac{m^2}{m_{T2}^2}\right]
\nonumber\\
& \quad\times \int d^2\vek{y} f_{gg}(x_1,x_2,\vek{y})\bar{f}_{gg}(\bar{x}_1,\bar{x}_2,\vek{y}) \, ,
\end{align}
where $y_i$ and $\vek{p}_i$ are the rapidity and transverse momentum of the charm quark produced in interaction $i=1,2$. The variables in the cross section are given by
\begin{align}
z_i & = \frac{m^2-\hat{t}_i}{\hat{s}_i} = -\frac{x_{Ti}}{2\bar{x}_{i}}e^{-y_{i}}, &
\bar{x}_i &=\frac{x_ix_{Ti}e^{-y_i}}{2x_i-x_{Ti}e^{y_i}}, &
x_{Ti} & = \frac{2m_{Ti}}{\sqrt{s}}, &
m_{Ti} = \sqrt{\vek{p}_{i}^2 + m^2}.
\end{align}
$\hat{s}_i$ and $\hat{t}_i$ are the usual Mandelstam variables of the partonic cross section $i$. $s$ is the center of mass energy of the proton collision, $m$ is the charm mass and $y_i$ is the rapidity of the charm quark from interaction $i$.
For gluons with longitudinal polarization the cross section is
\begin{align} \label{eq:xseclong}
\frac{d\sigma_{(\Delta g \Delta g)(\Delta g \Delta g)}}{dy_1dy_2d^2\vek{p}_1d^2\vek{p}_2} & = \frac{1}{2}\left(8\pi^2\alpha_s^2\frac{N_c}{N_c^2-1}\right)^2\int dx_1dx_2
\prod_{i=1}^2\frac{1}{16\pi^2\hat{s}_i^2} \frac{2x_i\bar{x}_i}{2x_i-x_{Ti}e^{y_i}}
\nonumber\\
& \quad\times \frac{(1-z_1)^2+z_1^2-1/N_c^{2}}{(1-z_1)z_1} \left( 1 - \frac{2m^2}{m_{T1}^2} \right) \left[(1-z_1)^2+z_1^2 \right]
\nonumber\\
& \quad\times \frac{(1-z_2)^2+z_2^2-1/N_c^{2}}{(1-z_2)z_2} \left( 1 - \frac{2m^2}{m_{T2}^2} \right) \left[(1-z_2)^2+z_2^2 \right]
\nonumber\\
& \quad\times \int d^2\vek{y} f_{\Delta g \Delta g}(x_1,x_2,\vek{y})\bar{f}_{\Delta g \Delta g}(\bar{x}_1,\bar{x}_2,\vek{y}).
\end{align}
Worth noticing is that differences in the partonic cross section between longitudinally and unpolarized gluons are suppressed by $m^2/m_{Ti}^2$. For transverse momenta of the outgoing charm quarks above a few GeV, this suppression is strong and already at $p_{Ti}=3\operatorname{GeV}$ we have $m^2/m_{Ti}^2 = 0.16$. This tells us that differences in the distributions of the final state charm quarks produced at large $p_T$ between longitudinally polarized and unpolarized gluons will to good approximation originate in differences between $f_{gg}$ and $f_{\Delta g \Delta g}$.
Unpolarized mixed with linearly polarized gluons gives the cross section
\begin{align}
\frac{d\sigma_{(\delta g g)(g \delta g)}}{dy_1dy_2d^2\vek{p}_1d^2\vek{p}_2} & = \left(8\pi^2\alpha_s^2\frac{N_c}{N_c^2-1}\right)^2\int dx_1dx_2
\prod_{i=1}^2\frac{1}{16\pi^2\hat{s}_i^2} \frac{2x_i\bar{x}_i}{2x_i-x_{Ti}e^{y_i}}
\nonumber\\
& \quad\times \left ((1-z_1)^2+z_1^2-1/N_c^{2} \right) \frac{m^2}{m_{T1}^2} \left( 1 - \frac{m^2}{m_{T1}^2} \right)
\nonumber\\
& \quad\times \left( (1-z_2)^2+z_2^2-1/N_c^{2} \right) \frac{m^2}{m_{T2}^2} \left( 1 - \frac{m^2}{m_{T2}^2} \right)
\nonumber\\
& \quad\times \cos (2\Delta \phi) \int d^2\vek{y} \; \vek{y}^4M^4f_{\delta g g}(x_1,x_2,\vek{y})\bar{f}_{ g \delta g}(\bar{x}_1,\bar{x}_2,\vek{y}),
\end{align}
and the same result for the case when the linear and unpolarized gluons are interchanged, i.e. with $g \leftrightarrow \delta g$. $\Delta \phi = \phi_1-\phi_2$, where $\phi_i$ is the azimuthal angle of the outgoing $c$-quark from hard interaction $i$. The $\cos (2\Delta\phi)$ dependence is an effect of the difference in helicity between the amplitude and conjugate amplitude for the linearly polarized gluons. This term gives rise to the same kind of modulation in the azimuthal angle as observed by LHCb in the $D^0D^0$ final state \cite{Aaij:2012dz}. However, in our leading order cross section, the whole contribution for the mixed linear-unpolarized gluons is suppressed by $m^2/m_{Ti}^2$ for each of the two hard subprocesses. This suppression gives low analyzing power, and indicates already at this level that the contribution should be small. The suppression arises in the terms where there is an helicity flip in the hard cross section. For zero quark masses these terms in the partonic cross section tend to zero. The nonzero charm mass allows for a nonzero result of the mixed un- and linearly-polarized gluons, but only at the price of the suppression factor. This has previously been discussed in the context of heavy quark production with transverse momentum dependent parton distributions \cite{Pisano:2013cya}. It is however interesting to note that the suppression could be lifted if the gluons are given a transverse momentum, for example by radiating off a gluon. The next-to-leading-order (NLO) correction to the cross section is expected to be large \cite{vanHameren:2014ava}, and a large NLO contribution in combination with a lifting of the suppression have the potential to result in a large enhancement of this contribution. We will return to this point in the discussion of \sect{sec:concl}.
The cross section for gluons with linear polarization is
\begin{align}
\frac{d\sigma_{(\delta g \delta g)(\delta g \delta g)}}{dy_1dy_2d^2\vek{p}_1d^2\vek{p}_2} & = \frac{1}{4}\left(8\pi^2\alpha_s^2\frac{N_c}{N_c^2-1}\right)^2\int dx_1dx_2
\prod_{i=1}^2\frac{1}{16\pi^2\hat{s}_i^2} \frac{2x_i\bar{x}_i}{2x_i-x_{Ti}e^{y_i}}
\nonumber\\
& \quad\times \left ((1-z_1)^2+z_1^2-1/N_c^{2} \right) \left( (1-z_2)^2+z_2^2-1/N_c^{2} \right)
\nonumber\\
& \quad\times \left[ \left(1-\frac{m^2}{m_{T1}^2} \right)^2\left(1-\frac{m^2}{m_{T2}^2}\right)^2 \cos (4\Delta \phi) +\frac{m^8}{m_{T1}^4m_{T2}^4} \right]
\nonumber\\
& \quad\times \int d^2\vek{y} f_{\delta g \delta g}(x_1,x_2,\vek{y})\bar{f}_{\delta g \delta g}(\bar{x}_1,\bar{x}_2,\vek{y}).
\end{align}
The pure linearly polarized contribution to the cross section has a $\cos(4\Delta\phi)$ dependent part and a $\Delta\phi$ independent part. The $\Delta\phi$ independent term comes with a double helicity flip in the two hard cross sections and is heavily suppressed. Notice that the $f^t_{\delta g \delta g}$ term in \eqref{eq:def-gg} does not contribute. This is because it results in a dependence on the angle between the directions of the outgoing charm quarks and the direction $\vek{y}$ between the two hard subprocesses, which vanishes upon integration over $\vek{y}$.
\section{Simple model for DPDs}
\label{sec:model}
In order to obtain numerical results we need an initial ansatz for the DPDs at some low starting scale. We decompose the unpolarized DPDs into two single parton distributions and a $\vek{y}$ dependent function assumed to be universal and $x_i$ independent
\begin{align}\label{eq:dpd-factorized}
f_{gg}(x_1,x_2,\vek{y};Q_0) = f_g(x_1,Q_0)f_g(x_2;Q_0)G(\vek{y}).
\end{align}
This is an ansatz commonly used for DPS phenomenology but its validity is questionable and in some kinematic regions wrong. The easiest way to see this is in the region of large $x_i$. Momentum conservations forces $x_1+x_2 \leq 1$ on the left side of \eqref{eq:dpd-factorized}, but the right hand side does, as it stands, not respect this constraint and gives nonzero values as long as both momentum fractions individually are below 1. A way to reinstate this limit is to multiply the ansatz with the factor $(1-x_1-x_2)$ to some positive power, however for the charm production the contribution of the large $x_i$ region is negligible and we apply a strict cutoff at the kinematic limit. Despite its limitations, the ansatz provides a useful starting point for DPS studies and we use it as input for the unpolarized DPDs at some low starting scale. These input distributions will then be evolved to higher scales with the double DGLAP equations. The numerical results will only be given in terms of ratios of cross sections, in which the $\vek{y}$ dependence cancels. For unpolarized distributions, the difference between separately evolving the two parton distribution functions (PDFs) or evolving the DPD with the factorized initial ansatz is small, except in the large $x_i$ region \cite{Diehl:2014vaa}.
For polarized distributions, which describe the correlation between the spin of the two partons, it does not make sense to decompose it into polarized single parton distributions -- which describe the correlation between the spin of one parton and the spin of the proton. Instead we use the positivity bounds in \cite{Diehl:2013mla} to set upper limits on the sizes of the polarized distributions in terms of the unpolarized. We are interested in examining the maximal effects possible from the different polarizations and therefore saturate the bounds for each polarized DPD independently. This results in the relations
\begin{align}
f_{\Delta g \Delta g} (x_1,x_2,\vek{y};Q_0) & = f_{gg}(x_1,x_2,\vek{y};Q_0) ,\nonumber\\
f_{g \delta g} (x_1,x_2,\vek{y};Q_0)& = \vek{y}^2M^2 f_{gg} (x_1,x_2,\vek{y};Q_0), \nonumber\\
f_{\delta g \delta g} (x_1,x_2,\vek{y};Q_0)& = \vek{y}^4M^4 f_{gg} (x_1,x_2,\vek{y};Q_0)
\end{align}
at the initial scale $Q_0$. If the bounds are fulfilled at an initial scale they will remain valid at all larger scales, but typically be violated at lower scales. We therefore choose to saturate the bounds at some low $Q_0$, and use the double DGLAP evolution \eqref{eq:evol-parton-1} with polarized splitting kernels to obtain the polarized DPDs at higher scales.
A larger $Q_0$ gives less room for evolution and therefore less suppression of the polarized contribution to the cross section. $Q_0$ should be chosen such that one is in a regime where perturbative QCD is expected to give sensible results. For the usual PDFs, the starting scale is often chosen somewhere around 1-2 GeV, and several of the leading order distributions go negative when evolved backwards to scales below 1 GeV. Another issue is the large uncertainty even for the distribution of a single unpolarized gluon at small scales and momentum fractions. The smaller values we take for the initial scale the larger this uncertainty is, and we would like to stay at a scale where we can compare between different sets of single parton distributions. We use this as a guidance in choosing starting scales at which to saturate our polarized bounds, and conclude that a choice somewhere between 1 and 2 GeV is reasonable. We investigate the impact of this choice by varying the input scale between the two values.
For input PDFs we will use the leading-order GJR distributions \cite{Gluck:2007ck}. At these scales there are still large differences between different PDF sets, and we have investigated how they influence our results by switching to the MSTW2008lo distributions \cite{Martin:2009iq}. There are clear differences, especially with $Q_0=1\operatorname{GeV}$, but the differences are smaller than those obtained by changing from $Q_0=1\operatorname{GeV}$ to $Q_0=2\operatorname{GeV}$. As a rule of thumb, the MSTW distributions give smaller polarization than the GJR distributions -- see \cite{Diehl:2014vaa} for more details on the effects of changing between different sets of PDFs. We use the values of $\alpha_s$ and the charm mass $m$ used in the PDF sets for concistency.
In addition, we examine the effect of changing our modeling of the polarized DPDs. Instead of taking the maximal allowed polarization, we can create a model built on ratios of splitting kernels describing the branching of a parent parton into the two gluons which subsequently undergo hard scatterings. For the longitudinally polarized DPD this results in
\begin{align}
f_{\Delta g \Delta g} = \frac{T_{g\rightarrow \Delta g \Delta g}}{T_{g\rightarrow g g}} f_{gg} = \frac{zz'(2-zz')}{z^2+z'^2+z^2 z'2} f_{gg}
\end{align}
where $z=x_1/(x_1+x_2)$ and $z'=1-z$. This model has been described in more detail in \cite{Diehl:2014vaa}, which gives a complete list of expressions for the different DPDs. We will only display results obtained in this model for longitudinally polarized gluons.
\section{Numerical results and comparison with data}
\label{sec:data}
We next turn to the numerical evaluation of the cross section in the kinematic regions probed by the double open-charm measurement by the LHCb Collaboration \cite{Aaij:2012dz}. The two charm quarks are required to have a transverse momentum in the region $3 \leq p_{Ti} \leq 12 \operatorname{GeV}$ and rapidities in the range $2 \leq y_i \leq 4$, at $\sqrt{s}=7$~TeV. The phase-space of the two anticharm quarks is integrated over, without experimental cuts, since they remain undetected.
The DPD evolution equations are evaluated by the code described in \cite{Gaunt:2009re}, which has been modified to suit our purposes as described in \cite{Diehl:2014vaa}. The main modifications are the use of the homogeneous evolution equations and the incorporation of the polarized splitting kernels for the evolution of the polarized DPDs, listed in appendix A of \cite{Diehl:2013mla}. We generated gridfiles for the DPDs in the range $10^{-6} \leq x_i \leq 1$ with 240 gridpoints in each direction, and 60 points in $\ln \mu^2$ in the range $Q^2_0 < \mu^2 < 2\times10^6\operatorname{GeV}$. The phase-space integrations were performed numerically.
Care must be taken when comparing the data to the results of our calculation. While the calculation produces two pairs of charm-anticharm quarks, out of which only the two charm quarks are measured, the data is for $D^0D^0$. Simply interpreting the variables of the charm quarks as those of the final state mesons neglects the effects from hadronization/fragmentation. The assumption that the direction of the charm quark is approximately equal to that of the $D^0$ is commonly made \cite{Maciula:2013kd}. For the effect on the absolute size of the transverse momenta, the approximation is less accurate, but charm fragmentation functions typically peak around rather large $z$ values \cite{Peterson:1982ak,Braaten:1994bz}. However this approximation on the normalized cross section is not likely to change the spectrum at the level of precision we are interested in here. Normalizing the results to the total cross section cancels the effects on the absolute size, such as the branching ratio of $c\rightarrow D^0$. Our primary purpose is not to make exact predictions for the $D^0D^0$ cross section, but rather to examine the effects that polarization has on double charm production.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/pt_2mc_final.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/pt_mt_final.eps}
\caption{\label{fig:pT} Normalized cross section vs the transverse momentum of one of the charm quarks at $\mu=2m$ (left) and $\mu=m_T$ (right). Overlaid are the LHCb $D^0D^0$ data \cite{Aaij:2012dz}. The lower panels show the relative size of the polarized contribution. }
\end{figure}
Figure \ref{fig:pT} shows the dependence of the normalized cross section on the transverse momentum of one of the two charm quarks, $p_T$, as well as the ratio of the polarized over unpolarized contribution. The left panel shows the results with the scale choice of $\mu = 2m$ and the right with $\mu=m_T$, where
\begin{equation}
m_T=\frac{m_{T1}+m_{T2}}{2}
\end{equation}
is the average transverse mass of the two charm quarks. For both cases, we make visible the dependence of the result on the choice of input scale by displaying the results for $Q_0=1\operatorname{GeV}$ and $Q_0=2\operatorname{GeV}$, as discussed in section \ref{sec:model}. The two lower panels show the relative size of the polarized contribution compared to the unpolarized. The cross section result in figure \ref{fig:pT} reproduces the data reasonably well. The shape of the cross section only has a tiny dependence on the choice of $Q_0$, whilst the contribution from the polarized distributions changes with $Q_0$. Likewise, there is little difference in the shape of the cross section with the two scale choices, but the polarized contribution is larger for $\mu=2m$ than for $\mu=m_T$. This is expected since the latter choice allows for a larger evolution range and thus a stronger enhancement of the unpolarized over the polarized DPDs. With $\mu=m_T$ the suppression due to evolution also increases with the $p_T$, since increasing $p_T$ increases $m_T$, counteracting the enhancement of the polarized contribution from the partonic cross sections. The relative size of the polarized contribution does have a small dependence on $p_T$ with $\mu=2m$, but is rather flat for $\mu=m_T$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/y_2mc_final.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/y_mt_final.eps}
\caption{\label{fig:Dy} Same as in figure \ref{fig:pT}, but for the cross section differential in the rapidity difference $\Delta y$ between the two charm quarks.}
\end{figure}
Figure \ref{fig:Dy} shows the dependence of the normalized cross section on the rapidity difference between the charm quarks. The cross section results are stable under variations of the scales and nicely reproduce the shape of the data. The two input scales have a strong impact on the size of the polarized contributions. The relative polarized contribution displays no dependence on the rapidity difference. With $\mu=2m$ the ratio of polarized over unpolarized is 30\% for $Q_0=2\operatorname{GeV}$ and around $4\%$ for $Q_0=1\operatorname{GeV}$. Changing to $\mu=m_T$ decreases the ratio to about half, $15\%$ for $Q_0=2\operatorname{GeV}$ and 2\% for $Q_0=1\operatorname{GeV}$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/m_2mc_final.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/m_mt_final.eps}
\caption{\label{fig:mcc}Same as in figure \ref{fig:pT}, but for the cross section differential in the invariant mass of the two charm quarks $M_{cc}$.}
\end{figure}
The cross section dependence on the invariant mass of the two charm quarks, $M_{cc}$, is shown in figure \ref{fig:mcc}. As in the previous figures, the data is rather well reproduced by the double charm cross section calculation. The polarized contribution has some dependence on $M_{cc}$ with $\mu=2m$ and thus a small impact on the shape of the cross section, but this effect disappears for $\mu=m_T$. With the lower input scale the polarized contribution is a few percent. With the larger $Q_0$ the polarized gluon contribution is $30\%$ of the unpolarized at small $M_{cc}$ and increases up to $50\%$ at large $M_{cc}$ for $\mu=2m$, while $\mu=m_T$ gives a ratio just above $10\%$ in the entire $M_{cc}$ range.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/phi_2mc_final.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/phi_mt_final.eps}
\caption{\label{fig:Dphi} Same as in figure \ref{fig:pT}, but for the cross section differential in the azimuthal angle between the two charm quarks.}
\end{figure}
An intriguing aspect of the $D^0D^0$ results is the azimuthal correlation between the two mesons. This correlation differs from that observed between meson final states with an equal number of charm quarks and anti-quarks, such as $D^0\bar{D}^0$ and $D^\pm D^\mp$ \cite{Aaij:2012dz}, which are dominated by single parton scattering. The angular modulation in $D^0D^0$ resembles that of a $\cos2\Delta\phi$ dependence, which is naturally produced by the DPS cross section involving a mixture of unpolarized and linearly polarized gluons.
In our LO calculation, the polarized contribution can still be sizable as demonstrated in figure~\ref{fig:Dphi}, but the large contribution originates from the longitudinally polarized gluons and is thus independent of $\Delta \phi$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/y_2mc_final_splitt.eps}
\caption{\label{fig:Dy-splitt} Normalized cross section in the splitting scenario vs the difference of rapidity between the two charm quarks at $\mu=2m$ (left). Overlaid are the LHCb $D^0D^0$ data \cite{Aaij:2012dz}. The lower panel shows the relative size of the polarized contribution.}
\end{figure}
Instead of the maximal polarization model for the DPDs, we can use the splitting model - where the ratios of the perturbative splittings of an unpolarized parent parton into two (unpolarized or polarized) partons are the basis of the relations between the different DPDs as explained in \sect{sec:model}. This scenario has a smaller polarization, but different $x_i$ dependences of the unpolarized compared to polarized DPDs - which could show up as shape differences in the rapidity spectrum. Figure~\ref{fig:Dy-splitt} shows the cross section as a function of the rapidity difference, with the polarized DPDs from the splitting model for $Q_0=2\operatorname{GeV}$ and $\mu=2m$. The size of the polarized cross section is reduced in the splitting scenario to about 5-10\%, with some dependence on the rapidity difference.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/pt_2mc_final_ext.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/pt_mt_final_ext.eps}
\caption{\label{fig:DpT-ext} Normalized cross section vs the transverse momentum of one of the charm quarks in the extended region down to $p_{Ti} = 1\operatorname{GeV}$, at $\mu=2m$ (left) and $\mu=m_T$ (right). The lower panels show the relative size of the polarized contribution.}
\end{figure}
Extending the kinematic region to examine in particular the effects of going down towards lower values of $p_T$, figure~\ref{fig:DpT-ext} shows the cross section dependence on $p_T$ in the kinematic range $1\leq p_{Ti} \leq 12 \operatorname{GeV}$. This decreases the size of the polarized contribution, which is not surprising since the longitudinally polarized cross section in \eqref{eq:xseclong} has a $(1-2 m/m_{Ti})$ factor for each of the two partonic processes, which decreases when going to smaller $p_{Ti}$. Although the polarization in this region is rather small, the effect on the shape of the cross section is interesting.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/pt_2mc_final_ext2.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/pt_mt_final_ext2.eps}
\caption{\label{fig:DDpT} Normalized double differential cross section vs the transverse momentum of the two charm quarks in the extended region down to $p_{Ti} = 1\operatorname{GeV}$, at $\mu=2m$ (left) and $\mu=m_T$ (right). The lower panels show the relative size of the polarized contribution. }
\end{figure}
In figure~\ref{fig:DDpT} we show the double differential cross section, in $p_{T1}$ and $p_{T2}$. We see a strong $p_{Ti}$ dependence of the polarized contribution in combination with a large absolute size, which starts at $0\%$ for $p_{Ti}=1\operatorname{GeV}$ and goes up to $60\%$ of the unpolarized for $p_{Ti}$ approaching 12 GeV, with $\mu=2m$ and $Q_0=2\operatorname{GeV}$. The results with $\mu=m_T$ have less polarization, with a maximal ratio reduced to about 10\%. Some of the $p_{Ti}$ dependence remains but most of it is at $p_{Ti}$ values below $3\operatorname{GeV}$. A precise measurement of this double differential cross section could be able to distinguish some of the different scenarios, and either see first indications of or set first limits on the longitudinally polarized gluons in DPS. In particular if it is possible to extend the measured region down to lower transverse momenta.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/phi_2mc_final_ratio.eps}
\caption{\label{fig:Dphi-ext} Normalized cross section vs the azimuthal angle between the two charm quarks, at $\mu=2m$, with kinematic region extended down to $p_{Ti} = 1\operatorname{GeV}$. }
\end{figure}
Investigating the dependence of the relative size of the mixed unpolarized - linearly polarized gluons on the lower limit of transverse momenta we expect a rather large increase in the relative size (compared to the cross section contribution without azimuthal dependence). This is also visible in figure~\ref{fig:Dphi-ext}, where the relative size of the mixed contribution is increased by almost an order of magnitude when extending the kinematic region down to $p_{Ti}=1\operatorname{GeV}$. The amplitude is still small compared to the angular modulation in the data, but it is another indication that allowing for a non-zero transverse momentum of the initial gluons, through for example NLO correction, could lead to significant enhancements.
\subsection{Predictions at $\sqrt{s}=13$ TeV}
In this section we show predictions at a hadronic center of mass energy of $13$~TeV. The results are generally very similar to those at $\sqrt{s}=7$~TeV, and we will therefore keep the discussion rather brief. Figure \ref{fig:pT13} (first row) displays the normalized cross section as a function of the $p_T$ of one of the two quarks. The change in CM energy as compared to figure \ref{fig:pT} flattens the cross section slightly and leads to a small decrease of the polarization. The second row of figure \ref{fig:pT13} shows the cross section results as a function of the rapidity difference $\Delta y$. The change in CM energy has no visible impact on the shape of the cross section, and only leads to a small decrease of the polarized contribution.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/pt_2mc_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/pt_mt_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/y_2mc_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/y_mt_final_13.eps}
\caption{\label{fig:pT13} Collisions at $\sqrt{s}=13$~TeV. Top: normalized cross section vs the transverse momentum of one of the charm quarks with $\mu=2m$ (left) and $\mu=m_T$ (right). The lower panels show the relative size $R$ of the polarized contribution. Bottom: normalized cross section vs the rapidity difference $\Delta y$ between the two charm quarks with $\mu=2m$ (left) and $\mu=m_T$ (right) and relative size $R$ of the polarized contribution. }
\end{figure}
This small decrease of the polarization as well as the small, if any, changes to the shape of the cross section is observed also for the dependence on $M_{cc}$ and $\Delta \phi$, as demonstrated by figure \ref{fig:mcc13}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/m_2mc_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/m_mt_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/phi_2mc_final_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/phi_mt_final_13.eps}
\caption{\label{fig:mcc13} Collisions at $\sqrt{s}=13$~TeV. Top: normalized cross section vs the invariant mass of the two charm quarks with $\mu=2m$ (left) and $\mu=m_T$ (right). The lower panels show the relative size $R$ of the polarized contribution. Bottom: normalized cross section vs the azimuthal angle between the two charm quarks $\Delta \phi$ between the two charm quarks with $\mu=2m$ (left) and $\mu=m_T$ (right) and relative size $R$ of the polarized contribution.}
\end{figure}
In the extended $p_{Ti}$ region for the double differential cross section, the results have large polarization with a strong dependence on the transverse momentum at $\mu=2m$ and $Q_0=2\operatorname{GeV}$, as shown in figure \ref{fig:DDpT13}. The contribution of the polarization decreases, as does the shape dependence, when going to $\mu=m_T$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.4\textwidth]{%
Plots/pt_2mc_final_ext2_13.eps}
\includegraphics[width=0.4\textwidth]{%
Plots/pt_mt_final_ext2_13.eps}
\caption{\label{fig:DDpT13} Collisions at $\sqrt{s}=13$~TeV. Normalized double differential cross section vs the $p_{T}$ of the two charm quarks, with $\mu=2m$ (left) and $\mu=m_T$ (right), in the extended region down to $p_{Ti} = 1\operatorname{GeV}$.}
\end{figure}
\section{Conclusions}
\label{sec:concl}
We have investigated the effects of polarization in the double open-charm cross section, when the two charm quarks are produced in the kinematic region probed in the $D^0D^0$ measurement by the LHCb Collaboration \cite{Aaij:2012dz}. Polarization can give sizable effects on the magnitude of the cross section, reaching up above $50\%$ of the unpolarized contribution in certain kinematic regions. The size strongly depends on the choices made when modeling the polarized double gluon distributions and on the large uncertainties for the single gluon distributions at the relevant low scales and small momentum fractions. We have presented the results obtained with $\sqrt{s}$ equal to both $7$ and $13$ TeV. The change of energy scale only has minor impact on the shape of the DPS cross section results as well as the relative size of the polarized contributions.
The shape of the polarized contributions to the cross section are in most variables quite similar to the unpolarized results. In these cases it is difficult to disentangle the polarized contribution from other contributions in the DPS cross section, such as a single parton splitting and color interference contributions. We therefore identify variables and kinematic regions where the polarization does introduce some shape dependence. The most prominent shape dependence is found for the cross section double differential in the $p_T$ of the two charm quarks, where the polarized contribution can vary with $p_T$ from 0 up to 60\% of the unpolarized.
We compare the results of our calculation with the measurement of $D^0D^0$ mesons by the LHCb \cite{Aaij:2012dz}. For most distributions, the leading order calculation reproduces the experimental data rather well. The data cannot discriminate between the different models for the polarized DPDs as the polarization does not introduce any strong shape changes. The exception is the dependence on the azimuthal angle between the two mesons, which exhibits an approximate $\cos2\Delta\phi$ modulation. Polarized double parton scattering naturally produces such a modulation in the combination of linearly polarized and unpolarized gluons. However, the leading order DPS cross section for this term is too small to reproduce the modulation in the data. It is possible, however, that the size of this term changes drastically when including higher orders. Higher order effects for the process are expected to be large \cite{vanHameren:2014ava}. Such large NLO corrections, in combination with an expectation that the higher order corrections will lift the strong suppression of the mixed (unpolarized - linearly polarized) contribution present at tree level, can lead to a significant enhancement of the amplitude of the azimuthal modulation. Unfortunately, the theoretical formalism for the description of DPS needs to be further developed to reach a state where higher order effects can be systematically included.
In the double differential cross section, looking at the $p_{T}$ of both of the charm quarks, the longitudinal polarization can have a larger impact on both the size and shape. Measurements of this double differential cross section could therefore give first experimental indications of, or limits on, the effects of polarization in double parton scattering.
We have used the homogeneous double DGLAP evolution equations, which do not include any single part splitting term. The effect of the single parton splitting on the unpolarized DPS cross section was studied in \cite{Gaunt:2014rua}. Including it also for the polarized terms of the DPS cross section could further enhance the effects of polarization. In addition, we have employed an ansatz which splits the unpolarized gluon DPD in two single gluon PDFs and a factor depending only on the transverse distance between the two partons. This approach, common in DPS studies, is useful as a first approximation of the gluon DPD, but neglects several effects. These include correlations between kinematical variables and the color of the two gluons. The cross section ratios which we present, are likely to be more stable to such corrections than the absolute size of the cross section. However, further phenomenological as well as experimental studies are required to better constrain these effects.
\section*{Acknowledgements}
Many thanks to Daniel Boer, Markus Diehl, Robert Fleischer, Francesco Hautmann and Hannes Jung for useful discussions and Vanya Belyaev for supplying us with the LHCb data points. We are in debt to Jonathan Gaunt for providing us with his original evolution code. We acknowledge
financial support from the European Community under the ``Ideas'' program
QWORK (contract 320389). M.G.E.~is supported by the ``Stichting voor Fundamenteel Onderzoek der Materie'' (FOM), which is financially supported by the ``Nederlandse Organisatie voor Wetenschappelijk Onderzoek'' (NWO). C.P. acknowledges support by the "Fonds Wetenschappelijk Onderzoek - Vlaanderen" (FWO) through a postdoctoral Pegasus Marie Curie Fellowship. Figures were made using JaxoDraw \cite{Binosi:2003yf}.
|
\section{#1}\setcounter{equation}{0}}
\renewcommand{\theequation}{\thesection.\arabic{equation}}
\def{\hbox{ 1\kern-.8mm l}}{{\hbox{ 1\kern-.8mm l}}}
\def{\hbox{ 0\kern-1.5mm 0}}{{\hbox{ 0\kern-1.5mm 0}}}
\def\,{\rm Tr}\, {\,{\rm Tr}\, }
\def\,{\rm Tr}_M^{q,y}\, {\,{\rm Tr}_M^{q,y}\, }
\def\tau{\tau}
\def\,{\rm e}{\,{\rm e}}
\def\hat{n}{\hat{n}}
\def\tilde{g}{\tilde{g}}
\newcommand{\hs}[1]{\mbox{hs$[#1]$}}
\newcommand{\w}[1]{\mbox{$\mathcal{W}_\infty[#1]$}}
\title{Stringy Symmetries and the Higher Spin Square}
\author{Matthias R.\ Gaberdiel$^{a}$, and}
\author{Rajesh Gopakumar$^{b}$ }
\affiliation[a]{Institut f\"ur Theoretische Physik, ETH Zurich, \\
$\;$ CH-8093 Z\"urich, Switzerland }
\affiliation[b]{Harish-Chandra Research Institute, \\ Chhatnag Road, Jhusi,
Allahabad, India 211019}
\emailAdd{<EMAIL>}
\emailAdd{<EMAIL>}
\abstract{Tensionless string theory on ${\rm AdS}_3\times {\rm S}^3\times \mathbb{T}^4$,
as captured by a free symmetric product orbifold, has a large set of conserved currents which can be usefully
organised in terms of representations of a ${\cal N}=(4,4)$ supersymmetric higher spin algebra. In this paper we focus
on the single particle currents which generate the asymptotic stringy symmetry algebra on ${\rm AdS}_3$, and
whose wedge modes describe the unbroken gauge symmetries of string theory in this background. We show that this global
subalgebra contains two {\it distinct} higher spin algebras that generate the full
algebra as a `higher spin square'. The resulting unbroken stringy symmetry algebra is exponentially
larger than the two individual higher spin algebras.}
\begin{document}
\maketitle
\section{Introduction}
The uniqueness of string (or M-)theory is one of its most attractive features, and yet the reason that it is so,
is not transparent. Indeed, until the advent of string dualities it was not even clear that the handful of different
looking perturbative string theories were limits of a single theory. Whenever we encounter in physics a rigidity of
structure it is most often due to an underlying symmetry (or invariance) principle. Hence, we might try and look
for such a symmetry principle constraining string theory. However, if a powerful constraining symmetry is present in
string theory it is mostly hidden from view. Perturbative string theories expanded about flat space manifest
diffeomorphism and gauge invariance in the massless sector, but little else. But there have been persistent
hints \cite{Gross:1988ue, Witten:1988zd, Moore:1993qe, Sagnotti:2011qp}, e.g., in high energy scattering,
of the restoration of a larger broken symmetry. If this is so, one needs a more symmetric vacuum to expand
around, to manifest (at least some more of) the additional symmetries.
String theories in ${\rm AdS}$ backgrounds might be better test cases for this purpose.
Indirectly, through gauge-string duality, we are led to the conclusion that the ${\rm AdS}$ string theory
dual to a free large $N$ QFT would have a large set of unbroken gauge symmetries
\cite{Sundborg:2000wp, Witten, Mikhailov:2002bp, Sezgin:2002rt}. These are the symmetries of the
Vasiliev higher spin theory \cite{Vasiliev:2003ev}. In fact, the Vasiliev theories of interacting higher spin
gauge fields are a very elegant realisation of how invariance principles can largely constrain the form of a theory.
At least when coupled to a finite number of low spin matter fields the classical equations of motion appear to be
unique (perhaps up to some parameters).\footnote{The classical and 1-loop Vasiliev theory successfully
capture the large $N$ limit of vector-like models with no propagating gauge fields
\cite{Klebanov:2002ja, Sezgin:2003pt, Giombi:2009wh, Giombi:2010vg, Gaberdiel:2010pz, %
Aharony:2011jz, Giombi:2011kc, Giombi:2013fka, Giombi:2014iua}, see \cite{Gaberdiel:2012uj, Giombi:2012ms}
for reviews.} But string theory is not simply a Vasiliev theory --- it has an infinite set of matter fields (of
arbitrarily high spin) coupled to the Vasiliev gauge fields. Also the gauge symmetries of the Vasiliev theory are
only one infinite tower or one Regge trajectory of the string theory. So this is not quite a large enough unbroken
stringy symmetry --- a genuinely stringy symmetry should be exponentially larger in size.
Are there vacua which exhibit such a larger unbroken symmetry?
String theories on ${\rm AdS}_3$ have an even larger unbroken symmetry than a generic
${\rm AdS}_{d+1}$ background. Again this can be indirectly inferred from gauge-string duality. For free gauge
theories in $d\geq 3$, the only single trace conserved currents are built from bilinears and their derivatives,
as follows from unitarity bounds for conserved currents in a CFT. In the bulk these correspond exactly to the
Vasiliev gauge symmetries and no more. But CFTs in two dimensions which are dual to ${\rm AdS}_3$
string backgrounds are, in many cases, believed to be deformations of free symmetric product orbifolds,
see \cite{David:2002wn} for a review. These have a very large set of conserved currents at their
orbifold point --- those built from bilinears of the underlying free bosons or fermions form only a tiny subset.
As we will see below, this is true even after isolating the single particle (`single trace') generators from
amongst all the currents. Assuming that this free point corresponds to a tensionless point in the moduli
space of ${\rm AdS}_3$ string vacua, we conclude that there must be a much larger unbroken gauge symmetry
in the bulk string theory here than in the higher dimensional cases \cite{Gaberdiel:2014cha}.
We should, however, keep in mind that string theory (or any theory of gravity, for that matter) on ${\rm AdS}_3$
always exhibits an enhanced asymptotic symmetry algebra. Indeed, as first observed by Brown and Henneaux,
diffeomorphism invariance on ${\rm AdS}_3$ has its global algebra of
$\mathfrak{sl}(2, \mathbb{R})\oplus \mathfrak{sl}(2, \mathbb{R})$ enhanced to two copies of the Virasoro algebra
with a universal central charge \cite{Brown:1986nw}.
Similarly, the higher spin gauge invariances of a Vasiliev theory
have their global symmetry algebra of $\mathfrak{sl}(N) \oplus \mathfrak{sl}(N)$ (or more generally,
${\rm hs}[\lambda]\oplus {\rm hs}[\lambda]$) enhanced to two copies of a ${\cal W}_N$
(or ${\cal W}_{\infty}[\lambda]$, respectively) asymptotic symmetry algebra
\cite{Henneaux:2010xg, Campoleoni:2010zq, Gaberdiel:2011wb, Campoleoni:2011hg}.
In order to extract the global part of the asymptotic symmetry algebra of string theory
we have to restrict it to its `wedge' modes --- these consist
of the `single particle'
generators that annihilate both the in- and out-vacuum --- and take the central charge to infinity.
The resulting `wedge' algebra --- for example, for the case of the bosonic ${\cal W}_N$ algebra
this is just $\mathfrak{sl}(N)$ ---
will then tell us about the unbroken stringy symmetries in the bulk. It is also this
algebra which is gauged and which will, hopefully,
help in constraining the structure of string theory --- just as the Vasiliev theory based
on ${\rm hs}[\lambda]\times {\rm hs}[\lambda]$ was largely so, by its algebra.
\smallskip
In this paper we make a beginning towards getting an explicit handle on this unbroken stringy algebra for the
case of superstrings on ${\rm AdS}_3\times {\rm S}^3\times \mathbb{T}^4$ in its tensionless limit, as captured
by the symmetric product orbifold SCFT of $\mathbb{T}^4$. We will exploit the fact that the whole symmetric orbifold chiral
algebra can be
organised in terms of representations of the superconformal ${\cal W}_\infty^{({\cal N}=4)}$
higher spin algebra \cite{Gaberdiel:2014cha};
the relevant representations are of the form $(0;\Lambda)$, where
$\Lambda$ is a representation of ${\rm SU}(N)$ \cite{Gaberdiel:2013vva, Gaberdiel:2014cha}.
First, we will identify which of these representations correspond to {\it single particle} symmetry generators.
These are the analogues of `single trace' operators in gauge theories, and are built from a single cycle of the
symmetric group. Their wedge modes then generate the global part of the symmetry algebra which,
as mentioned above, is the unbroken gauge algebra of the bulk string theory.
For the case of the symmetric orbifold of $\mathbb{T}^4$, the single particle generators are
in one-to-one correspondence with the elements of the chiral algebra of $\mathbb{T}^4$. (Indeed, to each
element $\rho$ in the chiral algebra of $\mathbb{T}^4$, we can associate a generator of the symmetric
orbifold algebra via the symmetrised sum, $\sum_i \rho^i$.) Note that when we take the large $N$
limit of the symmetric orbifold, all these generators become independent (unlike the
generators of the chiral algebra of a single $\mathbb{T}^4$ which cannot all be independent).
The results of \cite{Gaberdiel:2014cha} imply that these resulting generators must organise themselves
in terms of representations of the wedge subalgebra ${\rm shs}_2[\lambda=0]$ of
${\cal W}_{\infty}^{({\cal N}=4)}[0]$ --- the representation of the full ${\cal W}_\infty^{({\cal N}=4)}[0]$
algebra will also contain multi-particle states, and it is only the wedge modes that map single particle
generators to themselves. We find that the relevant representations (which carry the same label
both for ${\cal W}_{\infty}^{({\cal N}=4)}$ and its wedge subalgebra), are $(0;\Lambda)$, where $\Lambda$
is described in terms of Dynkin labels as $\Lambda=[m, 0, \ldots, 0,n]$, and each representation appears with
multiplicity one. In simpler terms, these are the
completely symmetric powers of the fundamental/antifundamental representation.
Here the generators from ($m=1$, $n=0$), ($m=0$, $n=1$) and ($m=n=1$)
are those of the ${\cal W}_{\infty}^{({\cal N}=4)}$
algebra itself (the free fermions, and the invariant bilinears together with their superconformal descendants).
The others account for the additional single particle generators of the full symmetric orbifold chiral
algebra.
In order to describe the Lie algebra structure of these wedge generators, it is
convenient to go back to the full asymptotic symmetry
algebra. Furthermore, it is instructive to analyse first the structure of a simpler
bosonic toy model, the $N$'th symmetric orbifold of a single real boson. In this case the single particle
generators of the chiral algebra are in one-to-one correspondence with the chiral sector of a single boson. While this
provides a complete characterisation of the algebra, it is useful to refine this description in terms of its
higher spin subalgebras. This is what the `higher spin square' structure uncovers.
A natural ${\cal W}_\infty$ subalgebra, in this case, is spanned by the bilinear expressions
\begin{equation}\label{1.1}
\sum_{i=1}^{N}\, (\partial^{m_1} \phi^i)\, (\partial^{m_2} \phi^i) \ , \qquad m_1,m_2\geq 1 \ ,
\end{equation}
suitable linear combinations of which define the quasiprimary generators of spin \linebreak $s=m_1+m_2$. They
generate the even spin subalgebra ${\cal W}^{(\rm{e})}_\infty[1]$ of ${\cal W}_\infty[1]$. On the other hand, the
most general `single-particle' generators involve the symmetrised products of order $p\geq 1$
\begin{equation}\label{ppower}
\sum_{i=1}^{N} \, (\partial^{m_1} \phi^i) \cdots (\partial^{m_p} \phi^i ) \ , \qquad
m_1,\ldots, m_p\geq 1 \ .
\end{equation}
The (singular part of the) OPE of a ${\cal W}_\infty$ generator with an order $p$ generator (\ref{ppower}) contains,
up to possible terms of lower order, only generators of order $p$; corrected by suitable lower order terms,
the generators of order $p$ therefore form a representation of ${\cal W}_\infty$.\footnote{Strictly speaking,
they only form a representation of the wedge subalgebra of ${\cal W}_\infty$. If we want
to think of them as honest representations of ${\cal W}_\infty$ we must also include the multiparticle
states that are made up of one generator of order $p$, together with an arbitrary
number of bilinear ${\cal W}_\infty$ generators.}
In order to visualise this,
we can arrange these representations in parallel vertical columns, where each column comprises
of `descendants' with the ${\cal W}^{(\rm{e})}_\infty[1]$ algebra acting `vertically', see Figure~1. In other words, the
commutators with the generators from the higher spin algebra
${\cal W}^{(\rm{e})}_\infty[1]= {\cal W}^{({\rm vert})}_\infty$ move one up and down along each of the green columns.
Note that each column really becomes `thicker' as we move vertically downwards, corresponding to the
exponentially increasing number of descendants, i.e., of splitting up the total spin among the $p$ integers
$m_1,\ldots,m_p$. These columns generalise to the representations $(0;\Lambda)$ with
$\Lambda=[m, 0, \ldots, 0,n]$ in the ${\cal N}=4$ case.
\smallskip
The crucial observation is now that the commutators involving the `top' generators from each
column {\it also} define a higher spin symmetry algebra, the `horizontal' ${\cal W}$-algebra
${\cal W}_{\infty}^{({\rm hor})}$. In fact, the generators associated to
$\sum_i(\partial \phi^i)^l$ close, up to lower order correction terms, to form the
${\cal W}_\infty$ algebra ${\cal W}_{1+\infty}[0]$; this was first observed in \cite{Pope:1991ig},
and is basically a consequence of the fact that we may equivalently describe the real boson theory in terms
of a complex fermion, whose associated ${\cal W}_\infty$ algebra is ${\cal W}_{1+\infty}[0]$.
We can therefore equally well organise the full stringy algebra in terms of representations of this
horizontal ${\cal W}$-algebra (or rather again its wedge subalgebra), i.e., in terms of the red rows
of Figure~1, see eq.~(\ref{3.21}) below. From this perspective, the left-most terms in each row
(except for the first row) are the bilinear currents of the form (\ref{1.1}) with spin $s\geq 3$, while the
top left entry corresponds to $\sum_i \partial^l \phi^i$, i.e., to
the spin $s=1$ generator of ${\cal W}_{1+\infty}[0]$, together with its $L_{-1}$ descendants --- these
states sit in a single irreducible representation of the (wedge subalgebra of the)
vertical algebra ${\cal W}^{({\rm vert})}_\infty$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=15.5cm]{diagram1.pdf}
\end{center}
\vspace*{-2.5cm}
\caption{The Higher Spin Square.}
\end{figure}
The picture we thus arrive at is that the commutators along a given row are governed by the horizontal higher spin algebra
${\cal W}_{\infty}^{({\rm hor})}$, while commutators down a given column are governed by the vertical
higher spin algebra ${\cal W}_{\infty}^{({\rm vert})}$. Knowing both thus determines, in
principle, an arbitrary commutator between elements of the chiral algebra.
We may use, for instance, ${\cal W}_{\infty}^{({\rm vert})}$ to bring any element to the
topmost row of its column, and then employ ${\cal W}_{\infty}^{({\rm hor})}$ to evaluate commutators
with any other entry.\footnote{Or we may just as well employ ${\cal W}_{\infty}^{({\rm hor})}$ to bring any
element to the left most column, and then utilise ${\cal W}_{\infty}^{({\rm vert})}$.}
Therefore, the global horizontal and vertical higher spin symmetry algebras together effectively
generate the full stringy symmetry --- the wedge algebra of the above chiral algebra. We will thus refer to
this symmetry structure as a `higher spin square'.\footnote{The terminology must not
mislead one into viewing this as a square (or tensor product) of the two higher spin symmetries.
The rank is, in a precise sense, exponentially larger than that of either the vertical or horizontal symmetry
algebra since it involves generators built effectively from multilinear products of the individual higher spin
generators.} We believe the existence of this higher spin square is
an indication that the stringy symmetries are best understood by decomposing them in terms of the higher
spin symmetries.
\smallskip
While the analysis for this bosonic toy model is rather natural and symmetrical (and can be checked fairly
directly, see in particular Section~4.1), we have so far not succeeded in finding
an equally canonical description in the ${\cal N}=4$ case, although a number of possible constructions
exist (see Section~4.3). On the other hand, we have found a
neat finite-dimensional analogue --- associated to the Clifford
algebra --- which builds intuition for this structure, and that is studied in some detail in Section~4.2.
In that case, one has $\gamma$ matrices in $2n$ dimensions.
The bilinears in the $\gamma$ matrices are the generators of $\mathfrak{so}(2n)$ which plays the role of the
vertical algebra. The other multi-linears define various antisymmetric representations of $\mathfrak{so}(2n)$
and can be arranged in vertical columns. The $\gamma$-matrix algebra gives us a natural set of commutators
for elements across different columns --- a horizontal algebra --- and the full Clifford algebra is generated by the
horizontal and vertical algebras. In this case it is easy to see that the resulting Lie algebra --- the `Clifford
algebra square' --- is just $\mathfrak{su}(2^n)$ since the gamma matrices generate all linearly independent
matrices of size $2^n$.
This shows that there is an exponential enhancement in rank in going from the vertical algebra
$\mathfrak{so}(2n)$ to the full Clifford algebra $\mathfrak{su}(2^n)$. We find this to be a useful model to keep in mind as some kind of a finite
truncation of the higher spin square.
\medskip
The outline of this paper is as follows. Section~2 contains the identification of the single particle symmetry generators
for the symmetric orbifold of $\mathbb{T}^4$.
Section~3 discusses the decomposition of the chiral algebra for free bosons and free fermions in terms of both horizontal and vertical bosonic
${\cal W}_{\infty}$ representations. These building blocks are then put
together for the SUSY ${\cal N}=4$ case.
Section~4 uses these results and describes the higher spin square, first in the case of the free boson (Section~4.1) and then
for the ${\cal N}=4$ case (Section~4.3). We also introduce the Clifford algebra square in Section~4.2.
Finally Section~5 contains miscellaneous closing remarks.
\section{Single Particle Symmetry Generators}\label{sec:symmetric}
Let us begin by identifying the single particle currents of the symmetric orbifold of $\mathbb{T}^4$.
The basic idea is to write the chiral sector of the orbifold partition function
for ${\rm Sym}^N(\mathbb{T}^4)$ (in the limit of large $N$) in a second quantised form.
Since the operation of taking the symmetric product is essentially one of `multi-particling', it is not greatly
surprising that the generating function of the single particle currents is the same as that of four free bosons
and fermions --- the content of the supersymmetric $\mathbb{T}^4$ theory. The new observation will be the
decomposition of these generators into specific representations of the small ${\cal N}=4$
${\rm shs}_2[\lambda=0]$ algebra, the wedge subalgebra of the supersymmetric ${\cal W}_\infty$ algebra.
Let us start with the familiar formula for the generating function for the untwisted sector of a general
symmetric orbifold \cite{Dijkgraaf:1996xw}
\begin{equation}\label{generating}
\sum_{N=0}^{\infty} p^N Z_{\rm RR}^{({\rm U})} ({\rm Sym}^k(X)) = \prod_{\Delta,\bar\Delta,\ell,\bar\ell}
\frac{1}{(1 - \, p q^{\Delta} \bar{q}^{\bar\Delta} y^{\ell} \bar{y}^{\bar{\ell}} )^{
c(\Delta,\bar\Delta, \ell,\bar\ell) }} \ ,
\end{equation}
where the coefficients $c(\Delta,\bar\Delta,\ell,\bar\ell)$ are the expansion coefficients of the
R-R partition function of $X$ (with the insertion of $(-1)^{F+\tilde{F}}$),
\begin{equation}
Z_{\rm RR}(X) = \sum_{\Delta,\bar\Delta,\ell,\bar\ell} c(\Delta,\bar\Delta, \ell,\bar\ell) \,
q^{\Delta} \bar{q}^{\bar\Delta} y^{\ell} \bar{y}^{\bar{\ell}} \ .
\end{equation}
We are interested in the ${\cal W}$ algebra of the symmetric orbifold of $X=\mathbb{T}^4$.
In particular, we want to analyse only
the purely left-moving states and we want to describe them in the NS-sector. Explicitly applying the
spectral flow, this leads to
\begin{equation}\label{genNS}
\sum_{N=0}^{\infty} p^N\, Z^{(N)}_{\rm vac, NS}(\tau,y) = \prod_{n=0}^{\infty} \prod_{\ell \in \mathbb{Z}}
\frac{1}{\bigl( 1 + (-1)^\ell \, p\, q^{n+\ell/2+1/4} \, y^{\ell+1}\bigr)^{c(n,\ell)}} \ ,
\end{equation}
where the $c(n,\ell)$ are the expansion coefficients of the R-sector chiral partition function
of $\mathbb{T}^4$ (with the insertion of $(-1)^F$)
\begin{equation}\label{cdef}
(y - 2 + y^{-1}) \prod_{n=1}^{\infty} \frac{ (1 - y q^n)^2 (1-y^{-1} q^n)^2}{(1-q^n)^4} =
\sum_{n=0}^{\infty} \sum_{\ell \in\mathbb{Z}} c(n,\ell) q^n y^\ell \ .
\end{equation}
Note that the term with $n=0$ and $\ell=-1$ in eq.~(\ref{genNS}) describes the contribution of the NS vacuum
with $q^{-1/4}$ --- it follows from eq.~(\ref{cdef}) that $c(0,\pm 1)=1$ --- while all other
terms in eq.~(\ref{genNS}) lead to positive powers of $q$. Thus
the NS vaccuum character stabilises for sufficiently large $N$ to
\begin{equation}
q^{N/4} \, Z^{(N)}_{\rm vac, NS}(\tau,y) = \prod_{n=0}^{\infty} \prod_{\ell \in \mathbb{Z}} {}^\prime
\frac{1}{\bigl( 1 + (-1)^\ell \, q^{n+\ell/2+1/2} \, y^{\ell+1}\bigr)^{c(n,\ell)}} \ ,
\end{equation}
where the prime at the product means that now the term with $n=0$ and $\ell=-1$ is excluded.
This can now also be written directly in terms of the NS sector expansion coefficients
\begin{equation}\label{NSex}
q^{N/4} \, Z^{(N)}_{\rm vac, NS}(\tau,y) = \prod_{r=\frac{1}{2}}^{\infty} \prod_{\ell \in \mathbb{Z}}
\frac{1}{\bigl( 1 - (-1)^{2r} \, q^{r} \, y^{\ell}\bigr)^{d(r,\ell)}} \ ,
\end{equation}
where $r$ runs over half-integers and integers and the coefficients $d(r,\ell)$ are now the expansion
coefficients of the spectrally flowed version of (\ref{cdef})
\begin{equation}\label{NSpart}
\prod_{n=1}^{\infty} \frac{(1 - y q^{n-\frac{1}{2}})^2 (1 - y^{-1} q^{n-\frac{1}{2}})^2}{(1-q^n)^4} = 1 +
\sum_{r=\frac{1}{2}}^{\infty} \sum_{\ell\in\mathbb{Z}} d(r,\ell)\, q^r y^\ell \ .
\end{equation}
The expression in eq.~(\ref{NSex}) is now very suggestive: it is already of `multi-particle' form, where each
state appearing in the NS character of $\mathbb{T}^4$ in eq.~(\ref{NSpart}) defines a separate particle.
Note that because of the $L_{-1}$ descendants, each holomorphic current
of spin $s$ contributes not just as
$1/(1-q^s)$ (if bosonic) or $(1+q^s)$ (if fermionic), but rather as
\begin{equation}\label{wedgeout}
\prod_{n=0}^\infty \frac{1}{(1-q^{s+n})} \quad \hbox{(bosonic)} \qquad \hbox{or}\qquad
\prod_{n=0}^\infty (1+q^{s+n}) \quad \hbox{(fermionic)} \ .
\end{equation}
Thus the full chiral algebra of the symmetric orbifold is generated by $N(r,\ell)$ fields
of spin $r$ and $\mathfrak{u}(1)$ charge $\ell$ where
\begin{equation}\label{Ndef}
(1-q) \, \Bigl[
\prod_{n=1}^{\infty} \frac{(1 + y q^{n-\frac{1}{2}})^2 (1 + y^{-1} q^{n-\frac{1}{2}})^2}{(1-q^n)^4} -1 \Bigr]
= \sum_{r=\frac{1}{2}}^{\infty} \sum_{\ell \in \mathbb{Z}} N(r,\ell) \, q^r\, y^\ell \ .
\end{equation}
Equivalently, we can then write the character of the chiral algebra of the stringy extension as
\begin{equation}\label{NSex1}
q^{N/4} \, Z^{(N)}_{\rm vac, NS}(\tau,y) = \prod_{r=\frac{1}{2}}^{\infty} \prod_{\ell \in \mathbb{Z}}
\, \prod_{n=0}^{\infty} \frac{1}{\bigl( 1 - (-1)^{2r} \, q^{r+n} \, y^{\ell}\bigr)^{ (-1)^{2r}N(r,\ell)}} \ .
\end{equation}
As mentioned earlier, this result for the single particle generators {\it per se}
may not come as a big surprise since
the process of taking the symmetric product is essentially one of `multi-particling'.
Thus it is
natural that the single particle generators of the $\mathbb{T}^4$ (or, more accurately, $\mathbb{R}^4$)
theory are in one-to-one correspondence with the chiral generators of four free bosons and their four
fermionic superpartners (as captured in the generating function (\ref{Ndef})).\footnote{Note that we can
only make a correspondence and not an identification
since each independent generator is a fully symmetrised expression involving (derivatives of) all $4N$ bosons and fermions.}
What is important from our present point of view is to organise these symmetry generators in terms of
the representations $(\Lambda_+; \Lambda_-)$ of the wedge algebra of the
super ${\cal W}_{\infty}^{(4)}[0]$ algebra. There is, in fact,
a simple way to characterise the representations which contribute to the single particle generators: they
correspond to
\begin{equation}\label{wreps}
(\Lambda_+; \Lambda_-) = \bigl(0;[m,0,\ldots.0,n] \bigr) \ .
\end{equation}
This is the natural class of representations to consider since they are the only ones which contain exactly a
single cycle singlet of the symmetric group (see the discussion in appendix~C.1 of \cite{Gaberdiel:2014cha}).
This single cycle singlet is the analogue of the single trace operators when one
gauges the symmetric group.
Combining this observation with the earlier generating function
then leads to the prediction for an identity for the wedge
characters\footnote{The full characters would also include multi-particle terms, and thus only the wedge
characters appear here.}
of these super ${\cal W}_{\infty}^{(4)}[0]$ representations
\begin{eqnarray}\label{wedgegen}
\sum_{r,\, \ell} N(r, \ell) \, q^r\, y^\ell & = & (y+y^{-1}) \Bigl[ 2 q^{1/2} + 4 \sum_{r=\frac{3}{2},\frac{5}{2},\ldots} q^r \Bigr]
+ \sum_{n=1}^{\infty} \, (y^2 + 6 + y^{-2}) q^n \qquad \label{wcont} \\
& & {} + (1-q)\, \sum_{m,n\geq 0}{}^{'} \chi^{({\rm wedge}) \, {\cal N}=4[0]}_{(0;[m,0,\ldots,0,n])} (q,y) \ . \label{wprim}
\end{eqnarray}
Here the terms on the RHS in (\ref{wcont}) describe the ${\cal W}_{\infty}^{(4)}[0]$ generators, while the contributions from
the sum in (\ref{wprim}) account for the wedge modes of the single-particle representations (\ref{wreps}),
with the terms generated from $L_{-1}$ subtracted out.\footnote{The prime in the sum indicates that the cases
$(m,n)=(1,0), (0,1), (1,1)$ are excluded. In effect, these terms precisely describe the contribution of
the first line, i.e., of (\ref{wcont}).}
Note that we have to restrict to the `wedge' part of the characters, since the other contributions
correspond to `multi-particle' states involving, in addition to a generator from $\bigl(0;[m,0,\ldots.0,n] \bigr)$, an arbitrary number of ${\cal W}_{\infty}^{(4)}[0]$ fields.
We have tested this identity, using eq.~(\ref{Ndef}) as well as the wedge characters from
\cite{Gaberdiel:2014cha}, up to order $q^{6}$. In the next section we shall also see
how to understand it in terms of bosonic and fermionic building blocks.
\section{Building Blocks}
The decomposition of the modes of four free bosons and four free fermions into characters of the
wedge algebra of ${\cal W}_{\infty}^{(4)}[0]$ is a generalisation of simpler decompositions of the chiral sector
of a single free boson or fermion into characters of the wedge subalgebras of the corresponding bosonic
${\cal W}_{\infty}$ algebras, as already sketched in the introduction.
In this section we first exhibit these simpler decompositions and then explain an identity involving
two distinct ${\cal W}_{\infty}$ algebras, see eq.~(\ref{3.21}), that is at the heart of the two dual
higher spin square decompositions. Combining these simpler decompositions will also
allow us to prove eq.~(\ref{wedgegen}), see the discussion in Section~\ref{sec:together}.
\subsection{Free Boson}
To understand the identity in its simplest incarnation, consider the
chiral algebra generated by the symmetric orbifold of a single free real boson. The analogue of
eq.~(\ref{wedgegen}), in conjunction with eq.~(\ref{Ndef}), is in this case
\begin{equation}\label{bosid}
(1-q)\prod_{n=1}^{\infty}\frac{1}{(1-q^n)} =\sum_r N(r) \, q^r =
(1-q)\Bigl(1+ \sum_{n=1}^{\infty}b^{({\rm wedge}) [\lambda=1]}_{([0^{n-1},1,0,\ldots,0];0)} (q) \Bigr) \ .
\end{equation}
Here the first term is, up to the $(1-q)$ prefactor, the partition function of a single free boson
while the last is the sum of the
bosonic wedge characters
$b^{({\rm wedge})}_{(\Lambda;0)} (q)$ of the (even) ${\cal W}_{\infty}^{({\rm e})}[1]$ algebra
of \cite{Gaberdiel:2011nt}, see also \cite{CGKV}. Note that the ${\cal W}$-algebra
generators themselves are counted by the wedge character corresponding to $n=2$, see
eq.~(\ref{boswed}) below,
\begin{equation}
b^{({\rm wedge})[\lambda=1]}_{([0,1,0,\ldots,0];0)} (q) = \frac{q^2}{(1-q) (1-q^2)} \ .
\end{equation}
Indeed, taking away the $(1-q)$ factor which just accounts for derivatives, this leads to a spectrum
of a single higher spin field for each even spin (which is what the theory of real bosons gives rise to).
In this case we can write down the branching functions (and therefore, wedge characters) explicitly following the results derived in
\cite{Gaberdiel:2011zw}. The wedge characters are
given by Schur polynomials (see eq.~(2.8) and below of \cite{Gaberdiel:2011zw}), which are
essentially ${\rm U}(\infty)$ characters. Simplified expressions for these have been given in Appendix~A
of \cite{Gaberdiel:2011zw}, see eqs.~(A.13), (A.14) and (A.16). In particular,
\begin{equation}
\chi^{{\rm U}(\infty)}_{[n,0,\ldots,0]}(q) = \frac{q^{\frac{n}{2}}}{\prod_{k=1}^n(1-q^k)} \ ,
\end{equation}
using the fact that
\begin{equation}
\prod_{(ij)\in [n,0,\ldots,0]}(1-q^{h_{ij}}) = \prod_{k=1}^n(1-q^k) \ ,
\end{equation}
where $h_{ij}$ are the hook lengths of the boxes in the Young tableaux.
After multiplying by the factor of $q^{\frac{\lambda}{2} n}$ (see eq.~(2.8) of \cite{Gaberdiel:2011zw})
and using the usual transposition rule between ${\rm U}(\infty)$ and coset Young tableaux,
we have (for $\lambda=1$),
\begin{equation}\label{boswed}
b^{({\rm wedge})[\lambda=1]}_{([0^{n-1},1,0,\ldots,0];0)} (q) = \frac{q^{n}}{\prod_{k=1}^n(1-q^k)} \ .
\end{equation}
Thus the identity of eq.~(\ref{bosid}) reduces to showing
\begin{equation}\label{bosid1}
\prod_{n=1}^{\infty}\frac{1}{(1-q^n)} = 1+ \sum_{n=1}^{\infty}\frac{q^{n}}{\prod_{k=1}^n (1-q^k)} \ .
\end{equation}
This identity is in fact a particular case of the so-called $q$-binomial theorem (see for example Section~10.2 of \cite{AAR}).
Actually, in our case we can easily prove this identity using the combinatorics of a free scalar field which also sheds
light on the nature of the associated chiral algebra generators.
We organise the Fock space of the left moving excitations of a bosonic field as the union of $n$-particle excitations
\begin{equation}\label{bosmodes}
\alpha_{-m_1}\alpha_{-m_2}\cdots\alpha_{-m_n}|0 \rangle = \alpha_{-1-k_1}\alpha_{-1-k_1-k_2}\cdots
\alpha_{-1-\sum_i^nk_i} |0\rangle \ .
\end{equation}
On the right-hand-side we have written out the modes $m_i\geq 1$ in ascending order so that $k_i \geq 0$.
The different contributions of such terms to the partition function is precisely given by
$\frac{q^{n}}{\prod_{k=1}^n (1-q^k)}$. Thus summing over $n$ gives us indeed eq.~(\ref{bosid1}).
\smallskip
For the following it will also be important to consider the case of a complex boson, i.e., two
real bosons. Using the fact that the wedge characters associated to boxes and anti-boxes
simply mutliply, this can then be written as
\begin{equation}\label{complxbos}
\prod_{n=1}^{\infty} \frac{1}{(1-q^n)^2} = \sum_{n,m=0}^{\infty}
b^{({\rm wedge})[\lambda=1]}_{([0^{n-1},1,0,\ldots,0,1,0^{m-1}];0)}(q) \ ,
\end{equation}
where we have used that the term with $m=n=0$ is just the identity. In this case, the ${\cal W}$-algebra
generators themselves can be obtained from the wedge character with $m=n=1$,
\begin{equation}
b^{({\rm wedge})[\lambda=1]}_{([1,0,\ldots,0,1];0)}(q) = \frac{q^2}{(1-q)^2} \ ,
\end{equation}
accounting indeed for one field of each integer spin $s\geq 2$; this then generates ${\cal W}_\infty[1]$, without
any truncation to even spin.
\subsection{Free Fermion}\label{sec:ff}
There is also a fermionic version of the above identity, for which the relevant wedge characters are
evaluated at $\lambda=0$. It takes the form\footnote{\label{foot5} In the free fermion case, we have the equivalence
of representations $(\Lambda;0) \cong (0;\Lambda^\ast)$ \cite{Gaberdiel:2011aa}. Thus, at this stage, it is
a matter of convenience to write this expression in terms of the wedge characters of $(0;\Lambda)$, rather
than $(\Lambda;0)$.}
\begin{equation}\label{fersid}
(1-q)\prod_{r=1}^{\infty} (1+y \, q^{r-1/2}) =\sum_r \tilde{N}(r,\ell) \, q^r\, y^{\ell} =
(1-q)\Bigl(1+ \sum_{n=1}^{\infty}b^{({\rm wedge})[\lambda=0]}_{(0;[n,0,\ldots,0])} (q,y) \Bigr) \ .
\end{equation}
At $\lambda=0$ the relevant wedge characters equal
\begin{equation}\label{ferwed}
b^{({\rm wedge})[\lambda=0]}_{(0;[n,0,\ldots,0])} (q,y) = \frac{y^n q^{n^2/2}}{\prod_{k=1}^n(1- q^k)} \ ,
\end{equation}
and again the term with $n=2$ corresponds to the ${\cal W}_\infty^{(\rm e)}[0]$ generators themselves,
\begin{equation}
b^{({\rm wedge})[\lambda=0]}_{(0;[2,0,\ldots,0])} (q) = \frac{y^2 q^{2}}{(1- q)(1- q^2)} \ .
\end{equation}
The above identity is then a direct consequence of another special instance of a $q$-binomial theorem,
\begin{equation}\label{fermid}
\sum_{k=0}^{\infty} \frac{y^k q^{k^2/2}}{\prod_{l=1}^k(1-q^l)} =\prod_{r=0}^{\infty}(1+y \, q^{r+1/2}) \ .
\end{equation}
Once again, we can easily prove this identity using the combinatorics of a free fermion field.
As before, we organise the Fock space of the left moving excitations of a fermionic field as the union of
$n$-particle excitations
\begin{equation}
\psi_{-m_1-{1\over 2}}\psi_{-m_2-{1\over 2}}\cdots\psi_{-m_n-{1\over 2}}|0 \rangle = \psi_{-{1\over 2}-k_1}\psi_{-\frac{3}{2}-k_1-k_2}\cdots
\psi_{-\frac{2n-1}{2}-\sum_i^nk_i} |0\rangle \ .
\end{equation}
On the right-hand-side we have written out the modes $m_i\geq 0$ in ascending order so that $k_i \geq 0$.
The different contributions of such terms to the partition function is precisely given by
$$
\frac{y^n q^{\frac{n^2}{2}}}{\prod_{k=1}^n (1-q^k)}\ .
$$
Thus summing over $n$ gives us indeed eq.~(\ref{fermid}).
\subsection{Bosonisation}
Actually, one can combine both fermionic and bosonic versions of these identities into dual
identities for a single generating function. The basic idea is that, because of bosonisation, the
neutral sector of the theory of a complex fermion is equivalent to that of a real boson.
For a complex fermion, the relevant decomposition identity equals
\begin{equation}\label{fersid2}
\prod_{r=1}^{\infty} (1+y\, q^{r-1/2})\, (1+y^{-1} q^{r-1/2})
= \sum_{n,m=0}^{\infty}b^{({\rm wedge})[\lambda=0]}_{(0;[n,0,\ldots,m])} (q,y) \ .
\end{equation}
Here we have used that the conjugate of the wedge characters (\ref{ferwed}) take the form
\begin{equation}\label{ferwedc}
b^{({\rm wedge})[\lambda=0]}_{(0;[0,\ldots,0,n])} (q,y) = \frac{y^{-n} \,q^{n^2/2}}{\prod_{k=1}^n(1- q^k)} \ ,
\end{equation}
as well as the fact that the wedge characters associated
to boxes and anti-boxes multiply. Then noting that the $1$ in eq.~(\ref{fersid}) just corresponds to $n=0$,
the identity eq.~(\ref{fersid2})
follows directly from the square of (\ref{fersid}). Note that in this
language, the associated ${\cal W}_\infty$ algebra now corresponds to the sector $(0;[1,0,\ldots,0,1])$
for which the wedge character equals
\begin{equation}
b^{({\rm wedge})[\lambda=0]}_{(0;[1,0,\ldots,0,1])} (q) = \frac{q^{1}}{(1-q)^2} \ .
\end{equation}
This then leads to one generator for each integer spin (including $1$), i.e., to ${\cal W}_{1+\infty}[0]$.
\noindent If we restrict the complex fermion theory to the neutral sector, we have the identity
\begin{equation}
\left. \prod_{r=1}^{\infty} (1+y q^{r-1/2}) (1+y^{-1} q^{r-1/2}) \right|_{y^0} = \prod_{n=1}^{\infty} \frac{1}{(1-q^n)} \ ,
\end{equation}
i.e., the neutral sector of two complex fermions agrees precisely with the theory of a single real boson. Thus
we can rewrite the bosonic partition function also in terms of the `neutral' wedge characters of the $\lambda=0$
theory,
\begin{equation}\label{bosid2}
(1-q)\prod_{n=1}^{\infty}\frac{1}{(1-q^n)} =\sum_r N(r) \, q^r =
(1-q)\Bigl(1+ \sum_{n=1}^{\infty}b^{({\rm wedge}) [\lambda=0]}_{(0;[n,0,\ldots,0,n])} (q) \Bigr) \ .
\end{equation}
This is equivalent to the identity
\begin{equation}\label{newid}
\prod_{n=1}^{\infty} \frac{1}{(1-q^n)} = 1 + \sum_{n=1}^{\infty} \frac{q^{n^2}}{\prod_{l=1}^{n} (1-q^l)^2} \ ,
\end{equation}
which is indeed true (see Corollary 10.9.4 and below of \cite{AAR}). Thus, for the case of a single real
boson we have two alternative identities (see eq.~(\ref{bosid}))
\begin{equation}\label{3.21}
\prod_{n=1}^{\infty}\frac{1}{(1-q^n)} =
\Bigl(1+ \sum_{n=1}^{\infty}b^{({\rm wedge}) [\lambda=1]}_{([0^{n-1},1,0,\ldots,0];0)} (q) \Bigr) =
\Bigl(1+ \sum_{n=1}^{\infty}b^{({\rm wedge}) [\lambda=0]}_{(0;[n,0,\ldots,0,n])} (q) \Bigr) \ .
\end{equation}
This fact will play an important role in the following. Since fermionisation and bosonisation involves
`non-perturbative' field-redefinitions, it is natural to express the bosonic and fermionic expansions
in terms of the representations $(\Lambda;0)$ and $(0;\Lambda)$, respectively, see
\cite{Gaberdiel:2012ku}. This motivates in retrospect the choice described in footnote~\ref{foot5}.
\subsection{Putting the Blocks Together}\label{sec:together}
Given that we have identities for the partition functions of a complex boson and a complex fermion,
see eqs.~(\ref{complxbos}) and (\ref{fersid2}), we can now easily obtain the expression for the
case with ${\cal N}=4$, see eq.~(\ref{Ndef}). Indeed, taking squares of the two expressions we conclude that
\begin{eqnarray}
& & \prod_{n=1}^{\infty} \frac{(1 + y q^{n-\frac{1}{2}})^2 (1 + y^{-1} q^{n-\frac{1}{2}})^2}{(1-q^n)^4} \nonumber \\[4pt]
& & \quad = \sum_{m_1,\ldots,m_4,n_1,\ldots,n_4}
b^{({\rm wedge})[\lambda=1]}_{([0^{n_1-1},1,0,\ldots,0,1,0^{m_1-1}];0)}(q) \, \,
b^{({\rm wedge})[\lambda=1]}_{([0^{n_2-1},1,0,\ldots,0,1,0^{m_2-1}];0)}(q) \nonumber \\
& & \qquad \qquad \qquad \qquad \qquad b^{({\rm wedge})[\lambda=0]}_{(0;[n_3,0,\ldots,0,m_3])} (q,y) \, \,
b^{({\rm wedge})[\lambda=0]}_{(0;[n_4,0,\ldots,0,m_4])} (q,y) \ . \label{3.22}
\end{eqnarray}
Using the explicit expressions for these wedge characters, the product on the right-hand-side can be written as
\begin{equation}
\frac{q^{n_1+n_2+m_1+m_2+\frac{1}{2}( n_3^2+n_4^2+m_3^2+m_4^2)} \, y^{n_3+n_4-m_3-m_4}}
{\prod_{j=1}^{4} \prod_{k=1}^{n_j} (1-q^k) \, \prod_{k=1}^{m_j} (1-q^k) } \ .
\end{equation}
Eq.~(\ref{3.22}) now reproduces (\ref{wprim}) provided we identify
\begin{equation}\label{3.24}
\chi^{({\rm wedge}) [{\cal N}=4]}_{(0;[m,0,\ldots,0,0])} (q,y) =
\sum_{n_1+m_1+n_3+m_3=m} b^{({\rm wedge})[\lambda=1]}_{([0^{n_1-1},1,0,\ldots,0,1,0^{m_1-1}];0)}(q) \,\,
b^{({\rm wedge})[\lambda=0]}_{(0;[n_3,0,\ldots,0,m_3])} (q,y) \ ,
\end{equation}
and similarly for $\chi^{({\rm wedge}) [{\cal N}=4]}_{(0;[0,0,\ldots,0,n])} (q,y)$, as well as their products. It follows
from straightforward combinatorical considerations\footnote{We thank Constantin Candu for analysing this
in some detail.} that eq.~(\ref{3.24}) is indeed the wedge character of the ${\cal N}=4$ coset representation
$(0;[m,0,\ldots,0,0])$. This therefore effectively proves eq.~(\ref{wprim}).
\section{The Higher Spin Square}
In the previous section we have seen how the single particle generators of the stringy asymptotic symmetry algebra
can be organised in terms of representations of the wedge subalgebra ${\rm shs}_2[0]$
of the ${\cal N}=4$ superconformal ${\cal W}_\infty^{({\cal N}=4)}[0]$ algebra,
see eq.~(\ref{3.22}) and below. These
single particle generators should now define a Lie algebra, the wedge algebra of the full stringy ${\cal W}$
algebra (i.e., the ${\cal W}$ algebra of the symmetric orbifold). This Lie algebra is the stringy
generalisation of the ${\rm shs}_2[0]$ algebra that underlies the construction of the Vasiliev
higher spin theory, and it should be thought of as describing the symmetries of string theory
on this background.
The aim of this section is to make first steps towards characterising this global Lie algebra.
Since we are viewing this question from the point of view of the dual CFT, the discussion will
be mainly couched in terms of the full asymptotic symmetry algebra (whose wedge algebra
we are seeking to describe). We shall be able to give a fairly compelling and natural description of
the full algebra in terms of a `higher spin square' for the the simpler symmetric orbifold of a single free boson,
using the fact that we can equivalently describe the boson theory in terms of a complex fermion. We have also
found a neat finite-dimensional toy model in terms of the familiar Clifford algebra of
gamma matrices, which exhibits nicely the exponential enhancement in rank. However, while
it is easy to show that a similar structure also exists in the ${\cal N}=4$ case,
the constructions of the horizontal ${\cal W}_\infty$ algebra we have found so far are
not particularly canonical.
\subsection{Free Bosons}\label{sec:free bosons}
Let us first begin with the simple case of the symmetric orbifold of a single free real boson, i.e., the theory of
$N$ real bosons $\phi^j$ orbifolded by the permutation action $S_N$ on the colour index $j$. The symmetric
orbifold contains, in particular, the ${\rm SO}(N)$ invariant bilinears of the free boson generators, which
are known to generate ${\cal W}_{\infty}^{({\rm e})}[1]$. Furthermore, as we saw in eq.~(\ref{bosid1}),
the remaining generators can be organised in terms of the representations $([0^{n-1},1,0,\ldots,0];0)$
of ${\cal W}_{\infty}^{({\rm e})}[1]$,
with $n=2$ corresponding to ${\cal W}_{\infty}^{({\rm e})}[1]$ itself. Note that it follows from
eq.~(\ref{bosmodes}) that we can think of the generators coming from $([0^{n-1},1,0,\ldots,0];0)$
as describing the symmetric invariants of $n$ powers of the field $\partial\phi^j$
with additional derivatives sprinkled on them (corresponding to the $k_i$ in eq.~(\ref{bosmodes})).
In order to understand the resulting structure we now want to think of the
generators of the full stringy ${\cal W}$ algebra as arranged in an
infinite table or square as follows (see Figure~1).
We start with a horizontal row corresponding to the modes of $\sum_j (\partial \phi^j)^n$ with $n =1,2 \ldots$, labelling
the columns, i.e., the representations $([0^{n-1},1,0,\ldots,0];0)$. The vertical column below each entry in this row
consists of all the ${\cal W}_{\infty}^{({\rm e})}[1]$ descendants of this state, i.e., of all the states involving
$n$ bosons with an arbitrary number of derivatives sprinkled on them. As mentioned above,
the $n=2$ column corresponds
to the ${\cal W}_{\infty}^{({\rm e})}[1]$ generators themselves. We will refer to this as the vertical higher spin or
${\cal W}^{({\rm vert})}_{\infty}$ symmetry.
What is intriguing about this description is that the top horizontal row of this square {\em also}
generates a (distinct) ${\cal W}_{\infty}$ higher spin algebra --- which we will call the horizontal
higher spin algebra ${\cal W}_{\infty}^{({\rm hor})}$.
In fact, the $\sum_j(\partial\phi^j)^n$ fields (corrected by terms
involving a smaller number of powers of $(\partial\phi^j)$ and additional derivatives) are what one obtains by
bosonising the fermionic bilinears (built from a complex fermion and its conjugate) that generate a
${\cal W}_{1+\infty}$ algebra \cite{Pope:1991ig}. From this perspective, the
$\sum_j(\partial\phi^j)^n$ terms correspond to the bilinear generators of the form
\begin{equation}
\sum_l \bar{\psi}^l\, \partial^{n-1}\psi^l\ ,
\end{equation}
or rather a sum of such bilinear terms with the $n-1$ derivatives distributed between $\psi^l$ and
$\bar{\psi}^l$. The stress tensor corresponds to the $n=2$ term which is common to the vertical
${\cal W}_{\infty}^{({\rm e})}[1]$ algebra but the terms with $n\neq 2$ are not part of the vertical algebra.
Note that the correction terms mentioned above are actually necessary for the operators to be quasi-primary
with respect to the stress tensor; for example, for the first few spins one finds explicitly \cite{Pope:1991ig}
\begin{eqnarray}\label{horspin}
S^{(1)} &= {\displaystyle \sum_j\partial\phi ^j\ , }\qquad
& S^{(2)} = \tfrac{1}{2} \sum_j (\partial\phi^j)^2\ , \\
S^{(3)} &= {\displaystyle \tfrac{1}{3} \sum_j (\partial\phi^j)^3\ , } \quad
& S^{(4)} = \tfrac{1}{4} \sum_j (\partial\phi^j)^4
-\tfrac{3}{20} \sum_j (\partial^2\phi^j)^2
+ \tfrac{1}{10} \sum_j(\partial\phi^j)(\partial^3\phi^j) \ . \nonumber
\end{eqnarray}
Thus we conclude that the stringy ${\cal W}$-algebra contains actually {\em two} higher spin algebras:
the usual bilinear ${\cal W}_{\infty}^{(\rm e)}[1]$ algebra --- what we called the vertical algebra above --- and
the ${\cal W}_{1+\infty}$ algebra --- what we called the horizontal algebra above -- generated by the higher powers of $(\partial \phi)^n$, that contains
in particular ${\cal W}_{\infty}[0]$, see, e.g., \cite{Gaberdiel:2013jpa}. The two algebras
do not commute with one another, but rather generate the full algebra upon taking successive commutators,
i.e., what we have called the `higher spin square'. This structure is somewhat reminiscent of a Yangian symmetry
that is believed to be present in
${\cal N}=4$ super Yang-Mills theory at the planar level \cite{Drummond:2009fd}.
In principle, this formulation characterises all commutators of the stringy algebra recursively, and one can
also restrict this description to the wedge algebra (where instead of ${\cal W}_\infty^{({\rm e})}[1]$ and
${\cal W}_\infty[0]$ we deal with the wedge subalgebras ${\rm hs}^{({\rm e})}[1]$ and ${\rm hs}[0]$,
respectively). However, unfortunately, we have not yet managed to find any
closed form expressions for these commutators. It would be very interesting to understand the structure of this algebra in more detail.
We should emphasise that the above analysis is only correct in the large $N$ limit; for finite $N$
the relevant ${\cal W}$ algebras are truncated to finitely generated algebras (as follows, for example,
from the fact that their central charge is finite). However, unfortunately, these truncations are rather
complicated and in general not explicitly known.
\subsection{The Clifford Algebra Square -- a Toy Model}
In order to get a feeling for how this construction works explicitly, it may be useful to consider a
finite toy model that also exhibits the same kind of structure. Consider the
gamma matrix algebra in an even number of (euclidean) dimensions $2n$,
\begin{equation}
\{\gamma^i, \gamma^j \}=2\delta^{ij} \ , \qquad (i,j =1,\ldots, 2n)\ .
\end{equation}
Then, as is familiar, the terms with $k$ distinct gamma matrices $(1 \leq k \leq 2n)$ correspond to the
$k^{\rm th}$ antisymmetric representation of $\mathfrak{so}(2n)$ and are of the form
$\gamma^{[i_1}\gamma^{i_2}\cdots \gamma^{i_k]}$. Let us denote this subset of the full Clifford algebra
as $H^{(k)}$, with $H^{(0)}$ denoting the identity.
Since each $H^{(k)}$ is a representation of $\mathfrak{so}(2n)$, there is a highest weight state and a
`vertical' column of $\mathfrak{so}(2n)$ descendants.
In the context of the Clifford construction, there is a natural commutator one can define
on the level of the generators in $H^{(k)}$. This can either be done abstractly via the so-called Clifford
commutators, see, e.g., \cite{LachiezeRey:2010au} for an introduction, or more explicitly since we can think of the
$\gamma$-matrices as matrices of dimension $2^{n}$. On the level of vector spaces, the resulting commutators
have the structure
\begin{equation}\label{cliffstru}
[H^{(k)}, H^{(m)}] = \sum_{r={\rm max}(0, (\frac{k+m}{2}-n))}^{{\rm min}(k,m)}
\bigl(1- (-1)^{r+km} \bigr) \, H^{(k+m-2r)}\ .
\end{equation}
This therefore gives a very concrete realisation of the commutation relations of these additional
horizontal generators (as well as their associated columns). Note that the term with $k+m-2r=0$
never arises since in that case $r=m=k$, and $r + km$ is necessarily even.
The $\gamma$-matrices can be thought of as matrices in dimension $2^{n}$, and since
\begin{equation}
\dim \Bigl( \bigoplus_{k=0}^{2n}H^{(k)} \Bigr) = 2^{2n} \ ,
\end{equation}
we deduce that the elements of the whole Clifford algebra generate the full space of these matrices.
Excluding the term with $k=0$ that never arises in commutators, we conclude that the full Lie
algebra is $\mathfrak{su}(2^n)$.
We should mention that the resulting structure is again very
similar to what we saw in the previous section. In particular, the rank of this Lie
algebra grows exponentially with the number of species $n$, whereas the rank of the vertical
Lie algebra $\mathfrak{so}(2n)$ is linear in $n$. Again, this is nicely parallel to what one should
expect for the relation between the higher spin symmetry and the stringy symmetry.
The number of higher spin symmetry generators have a linear growth in the number
of free fields, e.g., the theory based on $N$ free fermions leads to ${\cal W}_{1+N}$
with $N+1$ generating fields,
whereas the stringy symmetry generators have an exponential (Cardy) growth as one increases
$N$ in the symmetric product orbifold by $S_N$.
While we have described the horizontal action of commutators between the various columns in
eq.~(\ref{cliffstru}), we have not quite identified a small horizontal algebra (with rank of order $n$)
analogous to ${\cal W}_{\infty}^{({\rm hor})}$. It would perhaps be interesting to do so.
\smallskip
\subsection{The ${\cal N}=4$ Higher Spin Square}\label{sec:susyhspsq}
For the case with ${\cal N}=4$ we should expect, similar to the bosonic case,
to be able to construct
a `horizontal' ${\cal W}_\infty$ algebra that will generate, together with the
vertical ${\cal W}_{\infty}^{(4)}[0]$ algebra, the full set of generators (\ref{wedgegen})
of the stringy extension. In particular, the horizontal algebra should contain at least
one generator from each of the representations $(0;[m,0,\ldots,0,n])$ with $m,n\geq 0$.
One possible choice for such a horizontal algebra is to consider the
algebra that is generated by the various powers of $(\partial \phi^1)$ and
$(\partial\bar\phi^1)$, where $(\phi^1,\bar{\phi}^1)$ is one arbitrarily chosen complex boson.
The generator $(\partial \phi^1)^m \, (\partial\bar\phi^1)^n$ will then contribute to the
representation $(0;[m,0,\ldots,0,n])$, and hence this construction will account for all the generators.
Furthermore, up to lower order correction terms, these generators define
a bosonic ${\cal W}_\infty$ algebra, as follows from the same arguments as in
Section~\ref{sec:free bosons}. It is also clear from the analysis of Section~\ref{sec:together} that
the full generating function can be expanded in terms of the corresponding characters, using
twice the identity eq.~(\ref{3.21}).
However, there are a few rather unnatural aspects about this construction. First of all,
the horizontal algebra is not really of higher spin type since the number of generators of
spin $s$ grows linearly with $s$ --- the relevant generators are schematically of the form
\begin{equation}
(\partial \phi^1)^m \, (\partial\bar\phi^1)^{s-m} \ , \qquad m=0,\ldots, s \ .
\end{equation}
One way to avoid this, is to make the vertical algebra slightly larger by
including all bilinear terms of the underlying free fields, i.e., by adding in the
representations $(0;[2,0,\ldots,0])$ and $(0;[0,\ldots,0,2])$. (It is fairly obvious
that this will also lead to a consistent ${\cal W}_\infty$ algebra whose number of
generators does not grow with the spin.) Then we can make the horizontal algebra
slightly smaller by only including the powers of a single real boson $(\partial \Phi^1)^m$.
The resulting horizontal higher spin algebra is then the same as the one analysed in
Section~\ref{sec:free bosons}, i.e., it defines the bosonic algebra ${\cal W}^{({\rm e})}_\infty[1]$.
In particular, it is therefore of higher spin type.
However, both of these constructions are somewhat unnatural since the resulting
horizontal algebras are purely bosonic, and in particular, do not
contain the ${\cal N}=4$ superconformal algebra. It is possible to overcome this problem,
but the only way we have found so far, is again rather ad hoc. Instead of the above algebras
we consider the ${\cal W}_\infty$ algebra that is generated by bilinears of the generators
\begin{equation}
(\partial \Phi^1)^{m_1} \ , \qquad \quad (\partial^{m_i} \Phi^i) \ , \ \ i = 2,3,4 \ , \qquad \qquad
(\partial^{r_j} \psi^j) \ , \ \ j=1, 2, 3, 4 \ ,
\end{equation}
where the $\Phi^j$, with $j=1,2,3,4$, denote the $4$ real bosonic fields, and the $\psi^j$ are the corresponding
fermionic partners. The resulting algebra contains the ${\cal N}=4$ superconformal
algebra, and its structure is fairly similar to that of the bilinear extension of ${\cal W}^{(4)}_\infty[0]$ --- in particular,
the number of generators does not grow with spin, and the algebra is of `higher spin' type.
Again, it is clear that the full generating function can be expanded in terms of the corresponding
characters, using as before the identity eq.~(\ref{3.21}). However, the construction singles
out one of the four real boson fields, which is not particularly natural. So while these
considerations demonstrate
that also in the ${\cal N}=4$ case, the higher spin square can be generated by two higher
spin algebras, we suspect that there exists a more natural (and symmetrical) construction.
\section{Concluding Remarks}
In this paper we have analysed the
symmetry algebra of string theory on ${\rm AdS}_3\times {\rm S}^3\times \mathbb{T}^4$ in the tensionless
limit. We have found that the unbroken symmetries in this ${\rm AdS}_3$ vacuum exhibit the structure
of a higher spin square which we have described. The lesson seems to be (at least as far as these symmetries
are concerned) that knowing the commutators of the horizontal and vertical higher spin algebras completely
determine all commutators of the exponentially larger unbroken stringy symmetry algebra. This may indicate that
higher spin symmetries play an important role in string theory, beyond just
describing the dynamics of the leading Regge trajectory.
Of course, one would hope to be able to exploit the power of this symmetry to deduce more about the
broken symmetry or higgsed phase.
For this it will be important to go beyond the in-principle determination of
commutators mentioned above to more explicit characterisations of the algebra. It would also be
interesting to understand the relation of this symmetry algebra to the Yangian that is believed to
control the planar limit of ${\cal N}=4$ super Yang Mills, as well as to the spin chain analysis of
\cite{Borsato:2014exa,Sax:2014mea}, see also \cite{Sfondrini:2014via} for a recent review.
It would be good to see whether the multi-particle algebras, studied recently by Vasiliev \cite{Vasiliev:2012tv} as a way to generalise higher spin symmetry in string theory, have a connection with the higher spin square structure.
A related question is
to understand the representations of the higher spin square, especially those relevant for the description of the
fields of string theory. In \cite{Gaberdiel:2014cha} we described how the twisted
and untwisted sectors of the symmetric orbifold theory can be organised in terms of representations of
what we would now call the vertical ${\cal W}_\infty$ algebra. Once again, after considering the
single particle excitations, these must assemble into bigger representations of the stringy symmetry
above. It would be interesting to see how many such representations appear.
One should also be able to use this information to constrain matrix elements of
the perturbation by the $\mathfrak{su}(2)_R$ singlet twist two operator
{\it a la} Wigner-Eckart. A somewhat different approach to studying such perturbations might be to see whether current algebra methods, so successful for vector like theories \cite{Maldacena:2012sf}, can be adapted to the case with a larger stringy symmetry and give useful constraints.
Another direction, which we have only touched on, concerns the finite truncation of the higher spin square. As mentioned,
the ${\cal W}$-algebras at finite $N$ will be finitely generated and there are relations between hitherto
(at $N=\infty$) independent generators. This truncation is not entirely straightforward even for a symmetric
orbifold of $N$ free bosons. Such a deformed algebra reflects a quantum modification of the symmetry algebra
and is presumably responsible for non-perturbative effects like the stringy exclusion principle.
It would be interesting to see if the Clifford algebra square arises naturally in any such truncation.
Another
modification comes into play when we consider the theory on a finite $\mathbb{T}^4$. We have been
implicitly considering the theory on $\mathbb{R}^4$ by neglecting modes which carry charge on the
$\mathbb{T}^4$ --- as usually done in the D1-D5 system. Putting the theory on $\mathbb{T}^4$ again
breaks part of the symmetry, and one may study how the corresponding perturbation is constrained by the
unbroken symmetry.
The unbroken symmetries described here are those of a particular background of string theory. But the
underlying symmetries of the theory are independent of the choice of background --- the background enters
only in manifesting which symmetries are unbroken. One may ask whether the higher spin square described
here is the largest possible unbroken symmetry. As mentioned in the introduction, in higher dimensional
${\rm AdS}$ vacua, the maximal unbroken symmetries are smaller in size (though not subalgebras of the
above stringy algebra). Are there backgrounds which manifest
even larger unbroken symmetries? Even amongst ${\rm AdS}_3$ vacua there are others (such as
the case where $\mathbb{T}^4$ is replaced by K3) which will have different unbroken stringy symmetries at,
say, the orbifold point. Finally, necessary and sufficient conditions for 2d permutation orbifold CFTs to have a
holographic dual have recently been formulated in \cite{Haehl:2014yla,Belin:2014fna} (using results of
\cite{Hartman:2014oaa}). It would be nice to extend the above results for the symmetric orbifold to the more
general permutation orbifolds with potential holographic duals.
\section*{Acknowledgements}
We thank Niklas Beisert, Constantin Candu, Avinash Dhar, Shiraz Minwalla, Suvrat Raju, Ashoke Sen, Zhenya Skvortsov,
Spenta Wadia, and Misha Vasiliev
for helpful discussions. The work of M.R.G.\ is partially supported by the Swiss National Science Foundation,
in particular through the NCCR SwissMAP, and he thanks Cambridge University and the IAS in Princeton for
hospitality during various stages of this work.
R.G., as always, expresses his gratitude to the people of India
for supporting the enterprise of theoretical physics. He also thanks the Pauli Center of
ETH Zurich and ICTS-TIFR, Bangalore for hospitality during different stages of this project.
|
\section*{}
\section{Introduction}
Herbig Ae/Be stars with masses in the range 2-15\,$M_{\odot}$
lie at the interface between low-mass and high-mass
star formation. One of the key goals is to unravel whether the magnetospheric accretion
model that is very successfully applied in low-mass ($M < 2 M_{\odot}$)
T Tauri stars may also be of relevance for
higher mass stars. Whilst {\it circular} Stokes $V$ spectropolarimetry can be used to measure magnetic fields,
{\it linear} Stokes $QU$ polarimetry may be employed to probe the gas geometry within
the innermost regions between the star and the disk -- on the scale of just a few stellar radii.
Such work is needed to unravel the complex accretion flows onto T Tauri and Herbig stars, in order to understand
issues such as the mass-accretion rate $\dot{M}_{\rm acc}$ versus stellar mass
relation (Garcia-Lopez et al. 2006; Mendigutia et al. 2012) as well as the more general questions
as to whether even the highest mass stars in the Universe form by disk accretion or
whether more exotic formation mechanisms, involving for instance stellar collisions, need to
be invoked (e.g. Bestenlehner et al. 2011).
\begin{figure}
\centerline{\includegraphics[width=8cm]{noline.ps}}
\caption{Cartoon indicating the simplest case of {\it no line effect}. On the left,
the polarization spectrum {\it triplot} -- and on the right its accompanying $QU$ diagram.
A typical emission is shown in the lower panel of the triplot,
the \%Pol is given in the middle panel, while the Position Angle (PA) is
shown in the upper panel of the triplot. See Vink et al. (2002) for further details.}
\label{noline}
\end{figure}
We will explore the alignment between disk position angles (PAs)
measured from linear $QU$ spectropolarimetry and those from Herbig binaries, showing that our data are
fully consistent with disk accretion and fragmentation in the intermediate mass 2-15\,$M_{\odot}$ (Harmanec 1988) range. Note that
this is independent of the polarizing mechanism in Herbig Ae/Be stars.
\section{The method}
Linear Stokes $QU$ polarimetry can
be used to measure flattening of the circumstellar medium.
In principle, continuum polarization would already be able to inform us about the presence
of an asymmetric disk-like structure on the sky, but in practice this
issue is complicated by the roles of intervening circumstellar (as e.g. in the UXOR phenomenon; Grinin 1994) and/or interstellar
dust, as well as instrumental polarization. This is one of the reasons why {\it spectro}polarimetry, measuring
the change in the degree of linear polarization across spectral lines is such a powerful tool, as {\it intrinsic}
information can directly be obtained from the Stokes $QU$ plane.
The second reason is the additional bonus that it may provide kinematic information
of the flows around young stars.
\begin{figure}
\centerline{\includegraphics[width=8cm]{depolariz.ps}}
\caption{Cartoon indication {\it depolarization} or {\it dilution} of the emission line.
The depolarization across the line is as broad as the Stokes $I$ emission. Depolarization
translates into Stokes $QU$ space as a linear excursion. See Vink et al. (2002).}
\label{depolariz}
\end{figure}
Figures 1-3 show spectropolarimetry
cartoons (both in terms of polarization ``triplot'' spectra and $QU$ planes)
for the case that a spatially unresolved object is (i)
spherically symmetric on the sky with ``no line effect'', (ii) asymmetric
subject to line ``depolarization'' where the emission line simply acts to ``dilute'' the polarized
continuum, or (iii) where the line effect is more subtle, involving
PA flips across intrinsically polarized emission lines.
In its simplest form, spectropolarimetry can be used to detect a difference
between an unpolarized emission line such as H$\alpha$ and a polarized continuum that results
from scattering off a circumstellar disk (Fig.~\ref{depolariz}). This tool has for instance
been employed on samples of classical Be stars in the 1970s (e.g. Poeckert \& Marlborough 1976). These
observations gave an incidence rate of order 60\%, which is fully consistent with {\it all} classical Be stars
being embedded in electron scattering disks. Given that the PA measured from linear spectropolarimetry
was consistent with those from interferometry, the technique was considered to be a particularly efficient
and accurate method for discovering disks around stars that would otherwise remain unresolved.
In more recent years, we have found incidences of intrinsic {\it line} polarization, implying
that it is the line itself that is polarized, e.g. by a rotating accretion disk around a T Tauri star
that scatters line photons from the interior regions close to the star (Fig.~\ref{narrow}), that may arise
from the magnetospheric accretion model (Vink et al. 2003).
\begin{figure}
\centerline{\includegraphics[width=8cm]{narrow.ps}}
\caption{Cartoon showing a compact source of line photons scattered off a {\it rotating} disk.
The polarization signatures are relatively narrow compared
to the Stokes $I$ emission line. The PA flip is associated with a loop
in $QU$ space. See Vink et al. (2002).}
\label{narrow}
\end{figure}
\section{Results}
\begin{figure}
\centerline{\includegraphics[width=6cm]{bd+40.ps}}
\caption{
The Herbig Be star BD$+$40 4124 showing line ``depolarization'' or ``dilution''.
The {\it triplot} shows Stokes $I$ emission in the lowest panel,
the \%Pol is given in the middle panel, while the Position Angle (PA) is
given in the upper panel.
The data are re-binned such that the 1$\sigma$ error in the polarization
corresponds to 0.05\% as calculated from photon statistics.
See Vink et al. (2002) for further details.}
\label{bd40}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=6cm]{qubd+40.ps}}
\caption{
$QU$ diagram of the Herbig Be star BD$+$40 4124 showing depolarization/dilution.
See Vink et al. (2002) for further details.}
\label{qubd40}
\end{figure}
Over the last decade, we have been active in obtaining medium resolution ($R$ $\sim$ 8000)
linear spectropolarimetry on samples of several tens of T Tauri and Herbig (single and binary) stars
on 4m class telescopes at high signal-to-noise ratios $>$ 1000 using
instruments such as the ISIS spectrograph on the William Herschel Telescope (WHT).
\subsection{A transition in mode of mass accretion between Herbig Ae and Be stars}
\begin{figure}
\centerline{\includegraphics[width=6cm]{xyper}}
\caption{
The Herbig Ae star XY Per shows intrinsic {\it line} polarization.
Note the flip in the PA, which would translate into a {\it loop} when the
data were plotted in a $QU$ diagram. See Vink et al. (2002) for further details.
}
\label{xyper}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=6cm]{quxyper}}
\caption{
$QU$ diagram of the Herbig Ae star XY Per showing a loop representative of
intrinsic {\it line} polarization.
See Vink et al. (2002) for further details.
}
\label{quxyper}
\end{figure}
Figures\,\ref{bd40} to\,\ref{quxyper} show the difference in linear H$\alpha$ spectropolarimetry
discovered by Vink et al. (2002). Herbig Ae stars show PA flips in the upper panel of the
triplot, which is caused by intrinsic line emission scattered off a rotating disk. These
PA flip ($QU$ loop) data are the same in T Tauri stars, where the magnetospheric accretion model
has been successfully applied. The Herbig Be stars show spectropolarimetric behaviour that is
notably different: here the data are consistent with disk accretion. Note that this
difference between the Ae and the Be stars is not confined to H$\alpha$, as
H$\beta$ and H$\gamma$ show exactly the same differences between the Herbig Ae and the
Herbig Be stars (Mottram et al. 2007).
\subsection{Constraining the disk inner radius}
\begin{figure}
\centerline{\includegraphics[width=9cm]{1rstar}}
\caption{
Stokes $QU$ Monte Carlo model data for the case of {\it no} inner hole, subject to
a double $QU$ loop, which could be represented by 2 PA rotations, when plotted versus wavelength, such as in Fig.\,\ref{xyper}. See Vink et al. (2005a) for more detail.
}
\label{1rstar}
\end{figure}
\begin{figure}
\centerline{\includegraphics[width=9cm]{5rstar}}
\caption{
Stokes $QU$ Monte Carlo model data for the case of a larger inner hole, subject to
a single $QU$ loop, which could also be represented by a single PA rotation, when plotted versus wavelength, such as in Fig.\,\ref{xyper}. See Vink et al. (2005a) for more detail.
}
\label{5rstar}
\end{figure}
Vink et al. (2005a) performed 3D Monte Carlo scattering experiments
using {\sc torus} (Harries 2000)
to constrain the disk inner radius in T Tauri and Herbig Ae stars.
The results were obtained for both a flat disk and a constant opening angle ``theta''
disk and the model objects were found to show loops in the $QU$ diagram data (corresponding to
PA flips as shown in the upper panel of the triplot in Fig.\,\ref{xyper}).
There is a notable difference between a disk with no inner hole
showing {\it double QU} loops, as shown in Fig.\,\ref{1rstar}, versus
a disk with a small inner hole, subject to {\it single} QU loops, such
as shown in Fig.\,\ref{5rstar}.
Using this qualitatively different
behaviour, it was shown how this methodology can be used to derive {\it quantitative} constraints on
the sizes of disk inner holes around T Tauri and Herbig stars.
Figure\,\ref{incl} summarizes our methodology. For instance, data on the bright
T~Tauri star GW~Ori show the presence of a gradual PA change
across H$\alpha$, which may be indicating the presence of a relatively small
inner hole of less than two stellar radii for an inclined disk ($i$ $\simeq$ 75 degrees), or a
more pole-on disk, but with a larger inner hole (Vink et al. 2005a; Vink et al. 2005b).
\begin{figure}
\includegraphics[width=8cm]{incl}
\caption{
Constraining the disk inner hole. Single QU loops are given by the dark shaded area, whilst
double QU loops are indicated by the light shaded areas of the disk inner radius vs. disks inclination
plane. Transitional behaviour is represented by the white shaded area of the plane. See Vink et al. (2005).
}
\label{incl}
\end{figure}
\subsection{Disk alignment in Herbig Ae/Be binaries}
Vink et al. (2005b) and Wheelwright et al. (2011) compared their spectropolarimetric disk
PAs to the disk PAs derived from alternative imaging techniques, such as interferometry.
The excellent agreement
proved that linear spectropolarimetry is an efficient tool to determine disk PAs, and
that this is independent of the polarizing mechanism, whether that involves
scattering (Vink et al. 2005; Milic \& Faurobert 2014) or optical pumping (Kuhn et al. 2007).
The disk PAs were compared to binary PAs by Wheelwright et al. (2011)
from Herbig Ae/Be binaries. Our PA data were found to be
inconsistent with a random distribution, which might perhaps have
been expected in case primary stars capture their companions in
a cluster-like competitive accretion scenario. Instead, the PA data were found to be
consistent with a disk accretion scenario, in which both the
primary and secondary object fragment from the same accretion disk.
In other words, our data suggest that stars up to 15\,$M_{\odot}$ may indeed form by disk accretion, as
modeled by e.g. Kratter \& Matzner (2006), Krumholz et al. (2009), and Kuiper et al. (2010).
\section{Current applications and limitations}
We have shown polarimetric line profiles for scattering off rotating
disks, but we have not yet given any preference to a particular type of
scattering particle.
Moreover, there is usually the issue of interstellar polarization,
implying that the observed continuum PA is generally not equal to zero and
that the level of continuum polarization is affected by a foreground contribution
due to interstellar grains.
Nonetheless, the {\it differential} effect between line and continuum is not affected by
foreground polarization, and the shapes of the loops in the $QU$ plane
remain exactly the same (they are just shifted).
Furthermore, there is the complication of unpolarized line emission.
For classical Be stars, H$\alpha$ is believed to form in the circumstellar environment, rather than at
the stellar surface, as assumed in our Monte Carlo models.
So, there are two potential polarization effects (dilution and intrinsic line polarization), which
could sometimes be at work simultaneously, but the PA rotations due to intrinsic line
polarization should be distinguishable from dilution effects because of
their contrasting characteristics in the $QU$ plane.
So far, we have assumed idealized disk geometries.
One may for instance also wish to consider more sophisticated flaring disk geometries, as
infrared spectral energy distribution modeling has indicated a preference
for this (Kenyon \& Hartmann 1987; Chiang \& Goldreich 1997.
These flaring disks may even posses puffed-up inner rims (e.g. Dullemond et al. 2001).
A flaring disk may be able to intercept light at larger distances from the stars, which may
contribute to the continuum polarization percentage. However, the differences between
a flared and constant opening angle disk are expected to be minor as far as the predicted
polarization changes across spectral line profiles are concerned.
The issue of the polarizing agent has not yet been settled.
For hot stars, such as
classical Be stars, the polarizing agent is usually attributed to electron scattering
(although hydrogen continuum opacity is thought to have an effect on broad-band
spectropolarimetry at ionization edges). Electrons are known to be
able to smear out line polarimetric profiles because of their large thermal motions
compared to the bulk motions of the stellar envelope (Wood \& Brown 1994).
Therefore, some of the Stokes $Q$ and $U$ structure across lines
would be diminished for hot stars by this thermal broadening effect. This would especially be
true for lower sensitivity measurements. To date, line polarimetry with 4m-class telescopes
is usually performed at the S/N level of 1000, and the accuracies are therefore
of the order of about 0.1\%. Demands
on ``differential'' measurements across spectral lines are less severe than absolute
ones. Already some of the published data from the WHT and the Anglo Australian Telescope (AAT)
has performed better than at the 0.1\% level (see
the data on Zeta Pup by Harries \& Howarth 1996).
Nonetheless, in the current era of 8m-class, and in the upcoming era of 40m-class telescopes,
larger photon collecting areas will make routine high precision spectropolarimetry feasible,
such that even in the presence of thermal broadening, subtle changes in the
polarization and PA may become measurable.
For cooler stars, dust may be the principal polarigenic agent.
Although the matrices for Mie and Rayleigh scattering (as used in our Monte Carlo modeling)
are different, both favour forward and
backward scattering, such that the differences between our predictions and those for Mie scattering are
expected to be only qualitative. Furthermore, dust grains are not expected to have large thermal velocities that would
result in significant line broadening.
Another potential opacity source for which smearing is not expected to be significant is that
of neutral hydrogen.
Added to the high demand on sensitivity, another key aspect is that
of spectral resolution.
Currently, the resolution that can be achieved on
common-user optical instruments is limited to
$R$ $\la$ 10 000. Vink et al. (2005a) checked whether their predicted PA changes
would be resolvable when they degraded their model spectra to $R$ $=$ 10 000. They
found that the instrumental resolution would not wipe out the predicted $QU$ profiles.
The greater limitation is therefore sensitivity. This is because in the handling of
spectropolarimetric data there is usually degeneracy between sensitivity and spectral resolution.
Currently, one generally re-bins pixels across a spectral line,
to gain the signal needed to achieve the required polarimetric accuracy.
We conclude that while there is already plenty of evidence of single PA rotations in
T Tauri and Herbig Ae stars that are entirely consistent with disrupted disks,
the sensitivity is typically too poor to allow for quantitative comparisons
between models and data. In any case, no instance of a double PA flip has yet been positively
identified. However, this must be seen as absence of evidence rather than evidence of absence
until appreciably greater sensitivity (with S/N $>>$ 1000) becomes routine.
Note that many of the PA rotation amplitudes we have predicted correspond to
changes of only 0.05-0.1\% in Stokes $U$ throughout the spectral line.
Upon occasion, this is measurable with
today's instrumentation provided the integration times are long enough.
This is well worth the effort, since line polarimetry
can uniquely obtain combined constraints on disk inclination and inner hole radius,
as shown in Fig.\,\ref{incl}.
\section{Future applications}
The future of linear line polarimetry seems may become even brighter, as there are
many exciting future applications to be considered. To name just a few:\\
\begin{enumerate}
\item Infrared spectropolarimetry (Oudmaijer et al. 2005)
\item Larger samples \& Surveys
\item Monitoring data
\item Extra-galactic stars
\end{enumerate}
With respect to (i) the (near) infrared, it might become possible not only to detect the innermost
regions of the accretion disks around the obscured most massive young massive (O-type) stars, but also
the earliest phases of low- and intermediate mass star formation in pre-T Tauri and pre-Herbig systems.
With respect to the (ii) larger samples, it is noteworthy to realize that with large surveys, such as
IPHAS (Drew et al. 2005) and VPHAS$+$ (Drew et al. 2014), thousands of young H$\alpha$ emitting stars are being discovered, allowing
us to probe the mass-accretion physics as a function of mass, age, and environment (Barentsen et al. 2011; 2013; Kalari et al. in prep.).
\begin{figure}
\includegraphics[width=8cm]{costigan}
\caption{
The amplitude in mass-accretion rate $\dot{M}_{\rm acc}$ variations
for different timescales. Note that the maximum amplitude is reached on the
(rotational) timescale (of days) and not on shorter timescales. On the other hand,
sampling on significantly longer timescales does not seem to be required. See Costigan et al. (2014).
}
\label{cost}
\end{figure}
\begin{figure*}
\includegraphics[width=16cm]{spotmontage}
\caption{
3D Monte Carlo simulation using {\sc torus} (Harries 2000) of a
rotational modulation of 2 diametrically opposed spots on the stellar surface.
The sequence consists of a full 360 degrees cycle. I.e. the eight snapshots from 0 to 315 degrees are
taken every 45 degrees, i.e. at 0, 45, 90, 135, 180, 225, 270, and 315 degrees.
Note that the upper spot is initially (at 0 degrees) directed towards the observer.
}
\label{spots}
\end{figure*}
With respect to (iii) the monitoring, it has recently become clear that it is the rotational
timescale that is dominant in the mass-accretion physics in both T Tauri and Herbig Ae stars
(Costigan et al. 2014; see Fig.\,\ref{cost}). This means that linear spectropolarimetry modeling
involving monitoring data such as predicted through Monte Carlo simulations
shown in Fig.\,\ref{spots} can be employed to map the stellar-disk mass-accretion system in 3D.
This might become feasible with the space-based ultraviolet and
optical spectropolarimeter Arago/UVMag\footnote{lesia.obspm.fr/UVMag} (PI Neiner).
Finally, with current 8m telescopes such as VLT, Herbig stars can already be discovered in the
low metallicity environment of the Large Magellanic Cloud (LMC). The mass-accretion rate ($\dot{M}_{\rm acc}$)
for the Herbig B[e] candidate VFTS\,822 (Kalari et al. 2014) seems particularly high, and in order
to find out whether such rates are realistic we need the appropriate 3D data that will only be possible
with linear spectropolarimetry.
Currently, the main limitation is still sensitivity, but we are living in
exciting times where the possibility of extremely large 40m telescopes (ELTs) is about to become reality.
If the ELTs materialize with the badly needed polarization optics, we might
be able to obtain spectropolarimetric data at a level of precision that has been feasible
with 1D Stokes $I$ data for more than a century.
It is really important to note that 3D Monte Carlo radiative transfer
is well able to perform the required modeling, but the main limitation is still the lack of accurate 3D data!
\section{Conclusions}
\begin{itemize}
\item Herbig Ae/Be stars have accretion disks on the smallest spatial scales.
\item Linear spectropolarimetry data are entirely consistent with disk accretion and fragmentation.
\item There is a {\it transition} in mass-accretion physics between Herbig Ae and Herbig Be stars.
\item The rotational timescale is key to changes in $\dot{M}_{\rm acc}$.
\item We {\it require} linear Stokes $QU$ monitoring to map the 3D geometry.
\end{itemize}
\acknowledgments
I would like to thank my long term specpol collaborators Rene Oudmaijer, Tim Harries, and Janet Drew,
as well as my recent PMS PhD students Geert Barentsen, Grainne Costigan, and Venu Kalari for there
input in mass-accretion topics in more recent times.
\nocite{*}
\bibliographystyle{spr-mp-nameyear-cnd}
|
\section{Introduction}
It is well known that the Tsallis distribution gives excellent fits to the transverse momentum distributions observed at
the Relativistic Heavy Ion Collider (RHIC)~\cite{STAR,PHENIX1,PHENIX2} and at the Large Hadron
Collider (LHC)~\cite{ALICE,CMS1,CMS2,ATLAS,ALICE2} with only three parameters, $q$, $T$ and $dN/dy$ (or, alternatively, a
volume $V$~\cite{ryb, cleymans1, azmi-cleymans}). The parameter $q$ is referred as Tsallis parameter and
discussed elsewhere~\cite{wilk4} in detail. The parameter $T$ obeys the standard thermodynamic relation,
\begin{equation*}
\left. T = \frac{\partial E}{\partial S}\right |_{N, V},
\end{equation*}
however, the entropy used in the above equation is the Tsallis entropy~\cite{tsallis1} not the standard Boltzmann-Gibbs (BG) entropy. The parameter $V$ is not necessarily related to a volume deduced from other models like HBT calculations.
It was recently shown that these fits extend to values~\cite{CMS3} of $p_T$ up to 200 GeV/c~\cite{wilk2,tsallis-poland} . This is unexpected because in this kinematic range hard scattering processes become important~\cite{wilk3}. A description of the high $p_T$ results has been discussed in \cite{Urmossy} where a model using a combination of Tsallis at low $p_T$ and QCD hard scattering at high $p_T$ was considered. The present analysis shows that the Tsallis distribution describes measurements up to the highest $p_T$ using the same Tsallis parameters as those obtained at low $p_T$.
A power law based on the Tsallis distribution~\cite{tsallis1} is used to fit the $p_T$ spectra of charged particles measured by the ATLAS and CMS collaborations. The ATLAS collaboration has reported the transverse momentum in an inclusive phase space region taking into account at least two charged particles in the kinematic range $|\eta|$ < 2.5 and $p_T$ > 100 MeV~\cite{ATLAS}.
The CMS collaboration has presented the differential transverse momentum distribution covering a $p_T$ range up to 200 GeV/c, the
largest range ever measured in a colliding beam experiment~\cite{CMS3}.
The results for the charged particle multiplicities are consistent with those of ATLAS measurements.
Moreover, the results are compared to those obtained in~\cite{wilk2,tsallis-poland,wilk3,wilk5,wilk6} where very good fits to transverse momentum distributions were presented. We confirm the quality of the fits but obtain different values of the parameters albeit using a different version of the Tsallis model.
\section{Tsallis Distribution}
Power law distributions have been widely applied in high energy physics~\cite{STAR,PHENIX1,PHENIX2,ALICE,CMS1,CMS2,ATLAS,ALICE2} to describe the transverse momentum spectra of secondary particles produced in proton - proton collisions. With the advent of the LHC, transverse
momenta of hundreds of GeV have been measured. It is of great interest to choose a distribution which describes such high $p_T$ measurements very well. A well suited form of the Tsallis distribution is discussed in this paper and our main criterium for choosing such a distribution was thermodynamic consistency which has not always been implemented fully~\cite{worku1,worku2,conroy}.
It is well established that there are numerous physical systems under which BG statistics encounters many difficulties. In particular, when analysing the $p_T$ spectra of hadrons it is found that spectra decrease far slower than predicted by BG statistics, and appear to follow some power law at high $p_T$. The Tsallis distribution was first proposed
about twenty - seven years~\cite{tsallis1} ago as a generalization of the BG distribution.
The Tsallis form of the BG distribution is given by:
\begin{align}
f(E) \equiv \exp_{q}\left(-\frac{E-\mu}{T}\right) \label{eq1},
\end{align}
where $E$ is the energy and $\mu$ is the chemical potential and
the function $\exp_{q}(x)$ is usually defined as:
\begin{equation}
\exp_{q}(x) \equiv
\begin{cases}
[1+(q-1)x]^{1/(q-1)} & \quad \text{if } x > 0\\
[1+(1-q)x]^{1/(1-q)} & \quad \text{if } x \le 0\\
\end{cases}
\end{equation}
in the limit where $q \rightarrow$ 1 it reduces to the standard exponential:
\begin{eqnarray*}
\lim_{q \to 1} \exp_q(x) \rightarrow \exp(x).
\end{eqnarray*}
The expressions of the relevant thermodynamic quantities like entropy, $S$, energy density, $\epsilon$ (= E/V), pressure, $P$, and
particle number, $N$, are given below using the $q$-logarithm function,
\begin{equation*}
{\rm ln}_q(x) \equiv \frac{x^{1-q} - 1}{1 - q},
\end{equation*}
as integrals over the function defined in eq. (\ref{eq1})~\cite{worku1,worku2,azmi-cleymans}:
\begin{gather}
S = - gV\int\frac{d^3p}{(2\pi)^3}\left[f^{q}{\rm ln}_{q}f - f\right],\\
\epsilon = g\int\frac{d^3p}{(2\pi)^3}~E~f^{q},\\
P = g\int\frac{d^3p}{(2\pi)^3}\frac{p^{2}}{3E}~f^{q},\\
N = gV\int\frac{d^3p}{(2\pi)^3} f^{q} \label{eq7},
\end{gather}
where $g$ is the degeneracy factor and $V$ is the volume.
The Tsallis distribution used in this paper is thermodynamically consistent which means that the following relations, derived from the differential form of the first and second laws of thermodynamics, are satisfied~\cite{worku2}:
\begin{align}
T &= \left.\frac{\partial \epsilon}{\partial s}\right|_n, & \mu &= \left.\frac{\partial \epsilon}{\partial n}\right|_s, \nonumber\\
n &= \left.\frac{\partial P}{\partial \mu}\right|_T, & s &= \left.\frac{\partial P}{\partial T}\right|_\mu,
\label{consistency}
\end{align}
where $s$ and $n$ are the entropy and particle number densities respectively.
The corresponding momentum distribution deduced from eq. (\ref{eq7}) is given by:
\begin{equation}
E \frac{d^{3}N}{dp^3} = gVE \frac{1}{(2\pi)^3}\left[1 + (q-1)\frac{E-\mu}{T}\right]^{-\frac{q}{q-1}},
\end{equation}
which in terms of the rapidity, $y$, and transverse mass, $m_T$, variables ($E = m_{T}\cosh y$) at mid rapidity and
for $\mu$ = 0 becomes~\cite{azmi-cleymans,worku1}:
\begin{equation}
\left.\frac{d^{2}N}{dp_{T}dy}\right|_{y = 0} = gV\frac{p_{T}m_{T}}{(2\pi)^2}\left[1 + (q - 1) \frac{m_T}{T}\right]^{-\frac{q}{q-1}} \label{eq9}.
\end{equation}
The relationship between the above parameterization and the one used by the RHIC and LHC
collaborations~\cite{STAR,PHENIX1,PHENIX2,ALICE,CMS1,CMS2,ATLAS,ALICE2} has been discussed in~\cite{cleymans1}.
Additionally, comparison with high-energy data has been done in~\cite{wilk2,tsallis-poland,wilk3,wilk5,wilk6}, using
a simplified form:
\begin{equation}
\left.\frac{d^{2}N_{ch}}{dp_Tdy}\right|_{y = 0} = A\left[1 + (q - 1)\frac{p_T}{T}\right]^{-\frac{1}{q - 1}}
\end{equation}
where A is a normalization factor and $T$ is a parameter used to fit the transverse momentum distribution and is not
related to a temperature in the thermodynamic sense as in eq. (\ref{eq9}).
Furthermore, integration of eq. (\ref{eq9}) over the transverse momentum leads to~\cite{ryb}:
\begin{gather*}
\left.\frac{dN}{dy}\right|_{y = 0} = \frac{gV}{(2\pi)^2} \int_0^\infty p_{T}dp_{T}m_{T} \left[1 + (q - 1) \frac{m_T}{T}\right]^{-\frac{q}{q-1}}
\end{gather*}
\begin{gather}
= \frac{gVT}{(2\pi)^2} \left[\frac{(2 - q)m^{2} + 2mT + 2T^{2}}{(2 - q)(3 - 2q)}\right]\left[1 + (q - 1) \frac{m}{T}\right]^{-\frac{1}{q-1}}, \label{eq11}
\end{gather}
where $m$ is the rest mass.
Using the values of parameters $q$, $T$ and $V$ obtained from Tsallis fit, using eq. (\ref{eq9}), above equation enable us to obtain the values of $dN/dy$ for pions, kaons and protons. Their sum give us the values of $dN/dy$ for charged particles.
\section{Fit Details}
The charged particle transverse momentum spectra measured by the ATLAS and CMS detectors in proton - proton
collisions at LHC energies are fitted using eq. (\ref{eq9}).
It is well established that the charged particle $p_T$ spectra is composed of contribution from pions, kaons and protons. Hence, a sum of three Tsallis distributions for pions, $\pi^{+}$'s, kaons, $K^{+}$'s and protons, $p$, has been applied to fit the charged particle $p_T$ spectra in following way:
\begin{align}
\left.\frac{d^2N_{ch}}{dp_Tdy}\right|_{y = 0} &= 2p_T\frac{V}{(2\pi)^{2}}\sum_{i = 1}^3g_{i}m_{T,i}\left[1 + (q - 1) \frac{m_{T,i}}{T}\right]^{-\frac{q}{q-1}} \label{eq12}
\end{align}
where $i = \pi^{+}, K^{+}, p$. The additional factor of 2 on the right-hand side of eq. (\ref{eq12}) takes into account the negatively
charged particles $\pi^{-}, K^{-}, \bar{p}$. The relative weights between particles were determined by the corresponding degeneracy factors and given by $g_{\pi^{+}}$ = $g_{K^{+}}$ = 1 and $g_p$ = 2. The square root of the sum of the squares of the statistical and systematic errors (quadrature) was taken into account for fitting the data.
The charged particle multiplicity as a function of the transverse momentum measured by the ATLAS collaboration for events
with $n_{ch} \ge 2$, $p_T$ > 100 MeV and $|\eta|$ < 2.5 at $\sqrt{s}$ = 0.9 and 7 TeV in proton - proton collisions are shown
in figure~\ref{atlas1}. The parameterisation given in eq. (\ref{eq12}) has been used to fit the spectra.
In order to show the quality of the fits in more detail, Figure~\ref{atlas2} shows the ratios of the data to the
fit values at $\sqrt{s}$ = 0.9 and 7 TeV for the ATLAS data. The largest deviations occur at the lower energy 0.9 TeV while the
7 TeV data show deviations from the Tsallis fits which are at most in the 10 \% range.
The charged particle differential transverse momentum yields in proton - proton collisions at $\sqrt{s}$ = 0.9 and 7 TeV have been
measured within $|\eta|$ < 2.4 by the CMS detector. The CMS measurements are described by the Tsallis fit and displayed in figure~\ref{cms1}.
The ratios of the data to the Tsallis fit at $\sqrt{s}$ = 0.9 and 7 TeV are shown in figure~\ref{cms2}.
Again, in order to sow the quality of the fits in more detail, Figure~\ref{cms2} shows the ratios of the data to the
fit values at $\sqrt{s}$ = 0.9 and 7 TeV for the CMS data. The largest deviations occur at the highest values of $p_T$
where also the error bars are largest.
\section{Results and Discussions}
A thermodynamically consistent form of the Tsallis distribution, given in eq. (\ref{eq9}), has been used to fit the transverse
momentum spectra of the charged particle measured by both ATLAS and CMS collaborations in proton - proton collisions
at $\sqrt{s}$ = 0.9 and 7 TeV. It is observed from figure~\ref{atlas1} and figure~\ref{cms1} that the Tsallis distribution fits well the measured $p_T$ spectra, albeit
that the kinematical conditions of the data of both detectors were not identical.
It is clear that the proposed form of the Tsallis distribution fits not only the low $p_T$ but covers a wide range of $p_T$ spectra upto 200 GeV/c.
The ratios of the experimental data over the fit values, depicted in figure~\ref{atlas2} and figure~\ref{cms2}, shows an
intriguing log-periodic oscillation as a function of the transverse momentum not only for high $p_t$ data sets~\cite{CMS3}
but also for low $p_T$ data sets~\cite{ATLAS}. This was first noticed in~\cite{wilk5,wilk6} for CMS data sets only.
The origin of such oscillations is a matter of investigation in order to draw some useful physics out of it.
The values of the fit parameters, $q$, $T$ and the corresponding $\chi^2/NDF$ obtained from the fits of the $p_T$ spectra
are given in table~\ref{par1}.
For the highest values of $p_T$, reaching 200 GeV/c, a value of $\chi^2/NDF\approx 0.52/24$ for the CMS data is obtained.
Instead of parameter $V$, a parameter $R \equiv [V \frac{3}{4\pi}]^{1/3}$, referred as
radius, is listed in the table. It has been found that the values of $q$ increases slowly but clearly with
increase in the center-of-mass energy.
The thermodynamically consistent version leads to a temperature $T \approx 75$ MeV at both values of $\sqrt{s}$.
It can be observed that the parameter $R$ $\approx$ 4.5 fm and 5.0 fm for $\sqrt{s}$ = 0.9 and 7 TeV respectively for both ATLAS and CMS data sets. These results are in agreement with the previous analysis done for lower $p_T$ data~\cite{azmi-cleymans,cleymans1}.
Table~\ref{par2} shows the values of $\left.\frac{dN}{dy}\right|_{y=0}$ for pions, kaons and protons and total
charged particles evaluated using eq. (\ref{eq11}).
The $\left.\frac{dN_{ch}}{d\eta}\right|_{\eta=0}$ values measured by the ATLAS detector is also present. It can be found that the production of pions are less and more kaons and protons are produced at higher energies. The charged particle multiplicity found for ATLAS using Tsallis fit are close to the measured values. An estimate of the charged particle multiplicity is presented for CMS which can be verified by the CMS collaboration.
\section{Conclusions}
We find it quite remarkable that the transverse momentum distributions measured up to 200 GeV/c in $p_T$ can be described
consistently over 14 orders of magnitude by a straightforward Tsallis distribution as presented in eq. (\ref{eq12}).
The values of $\chi^2/NDF$ are listed in table 1 and for the highest values of beam energy and $p_T$ it is found to be 0.52/24 for the CMS data.
The distribution used is part of a consistent thermodynamic description satisfying the basic consistency conditions given in eq. (\ref{consistency}).
In conclusion, we have shown that the proposed version of the Tsallis distribution only fits the transverse momentum spectra with
minimal values of $\chi^2$ values and an estimate of the particle multiplicities and particle ratio has been obtained.
\begin{table*}
\caption{Values of the $q$, $T$ and $R$ parameters and $\chi^2/NDF$ obtained using eq. (\ref{eq9}) to fit the $p_T$ spectra measured by the ATLAS~~\cite{ATLAS} and CMS~\cite{CMS3} detectors.}
\label{par1}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}llllll@{}}
\hline
{\bf Experiment} & \multicolumn{1}{c}{\bf $\sqrt{s}$ (TeV)} & \multicolumn{1}{c}{\bf $q$} & \multicolumn{1}{c}{\bf $T$ (MeV)} & \multicolumn{1}{c}{\bf $R$ (fm)} & \multicolumn{1}{c}{\bf $\chi^2/NDF$} \\
\\
\hline
ATLAS & 0.9 & 1.129 $\pm$ 0.005 & 74.21 $\pm$ 3.55 & 4.62 $\pm$ 0.29 & 0.657503/36 \\
ATLAS & 7 & 1.150 $\pm$ 0.002 & 75.00 $\pm$ 3.21 & 5.05 $\pm$ 0.07 & 4.35145/41 \\
\hline
CMS & 0.9 & 1.129 $\pm$ 0.003 & 76.00 $\pm$ 0.17 & 4.32 $\pm$ 0.29 & 0.648806/17 \\
CMS & 7 & 1.153 $\pm$ 0.002 & 73.00 $\pm$ 1.42 & 5.04 $\pm$ 0.27 & 0.521746/24 \\
\hline
\end{tabular*}
\end{table*}
\begin{table*}
\caption{Values of $\left.\frac{dN}{dy}\right|_{y=0}$ for pions, kaons, protons and total charged particles obtained
using eq. (\ref{eq11}) and $\left.\frac{dN_{ch}}{d\eta}\right|_{\eta=0}$ measured by ATLAS detector.}
\label{par2}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}lllllll@{}}
\hline
{\bf Experiment} & \multicolumn{1}{c}{\bf $\sqrt{s}$ (TeV)} & \multicolumn{1}{c}{\bf{$\left.\frac{dN^{\pi^{+}+\pi^{-}}}{dy}\right|_{y=0}$}} & \multicolumn{1}{c}{\bf$\left.\frac{dN^{K^{+}+K^{-}}}{dy}\right|_{y=0}$} & \multicolumn{1}{c}{\bf$\left.\frac{dN^{p+\bar{p}}}{dy}\right|_{y=0}$} & \multicolumn{1}{c}{\bf{$\left.\frac{dN}{dy}\right|_{y=0}$}} & \multicolumn{1}{c}{\bf$\left.\frac{dN_{ch}}{d\eta}\right|_{\eta=0}$}\\
\\
\hline
ATLAS & 0.9 & 2.84 & 0.76 & 0.32 & 3.92 & 3.85 $\pm$ 0.18 \\
ATLAS & 7 & 4.27 & 1.37 & 0.72 & 6.36 & 6.25 $\pm$ 0.30 \\
\hline
CMS & 0.9 & 2.52 & 0.70 & 0.30 & 3.52 & \\
CMS & 7 & 3.92 & 1.24 & 0.66 & 5.82 & \\
\hline
\end{tabular*}
\end{table*}
|
\section{Introduction}
The analysis of socio-economic data often involves the integration of various spatial data sources. Those data are often independently collected by a variety of offices and for different purposes. The zonal set systems used by distinct offices are rarely compatible and this leads to many difficulties. The problem of merging data bases on different spatial supports is called the areal interpolation or basis change problem (Goodchild and Lam 1980). In France, the need for official statistics at a more and more refined territorial level has been recognized by INSEE. In Europe, one of the objectives of the EU directive ``INSPIRE'', for INfrastructure for SPatial InfoRmation, is to harmonize quality geographic information to support the formulation and evaluation of public policies and activities which directly or indirectly impact the environment. Many statistical methods are proposed in the literature to handle this problem (dasymetric methods, regression methods, smoothing techniques) and the reader is referred to Do et al. (2014) for a recent review of the simplest ones. The problem of their relative accuracy is most often treated at an empirical level (see for example Reibel and Bufalino (2005), Mennis (2006), Flowerdew and Green (1992), Flowerdew et al. (1991), Reibel and Agrawal (2007), Gregory (2002)). At the theoretical level, only few articles address this problem (Sadahiro (1999, 2000)) and this is the objective of this work.
Comparing the accuracy of the different methods is difficult because the relative accuracy depends on several factors: nature of the target variable, correlation between the target and auxiliary variables, shapes of zonal sets, relative size between the two zonal sets,... In order to derive theoretical results, we need to consider simplifying restrictions. For this reason, in this document, we first of all restrict attention to data obtained from counts (see section \ref{count}): they are frequent in the literature and cover most of the cases in the socio-economic applications. We also restrict the comparison to the simplest classes of methods which are the dasymetric and the regression ones. At last, we make the assumption that target zones are nested within source zones. Indeed, this is not really a restriction since the intersections between sources and targets are always nested within sources and it is immediate to go from intersection level to target level by aggregating the predictions as we will see later. In section \ref{count}, we define what we mean by data obtained from counts and we introduce a mathematical model adapted to this case. In order to illustrate the methods and check our theoretical results, we present two sets of simulated data that we use later. In section \ref{methods}, we recall the formulas for the dasymetric and Poisson regression areal interpolation. Finally, in section \ref{acc1}, we compare the relative accuracy of areal weighting and dasymetric methods with finite distance results whereas in section \ref{acc2}, we compare the relative accuracy of dasymetric and Poisson regression methods with
asymptotic methods. In both sections, we comment the results obtained on the toy examples presented in section \ref{count}. All proofs are in the appendix.
\section{Count data and Poisson point pattern model}
\label{count}
The variable of interest $Y$ that needs to be interpolated is called the {\bf target variable} and it needs to have a meaning on any subregion of the given space. $Y_D$ will denote the value of the target variable on the subregion $D$ of the region of interest $\Omega.$
\noindent In the general area-to-area reallocation problem, the original data for the target variable is available for a set of source zones $S_s$ ($s= 1, \cdots, n_S,$) and has to be transferred to an independent set of target zones $T_t$ ($ t=1, \cdots, n_T$). The variable $Y_{S_s}$ will be denoted by $Y_s$ for simplicity and similarly for $Y_{T_t}$ by $Y_t$. The source zones and target zones are not necessarily nested and their boundaries do not usually coincide.
\noindent Overlaps between the two sets are called intersection zones and denoted by $A_{st}$ for the intersection between the source $S_s$ and the target $T_t$. For simplicity, $Y_{A_{st}}$ will be denoted by $Y_{st}.$ Many methods involve the areas of different subregions (sources, targets or other). We will denote by $\mid A \mid$ the area of any subregion $A$.
\noindent Most of economic data collected at regional level result from aggregating point data and are only released in this aggregated form. Intuitively, let us say that a point data set is a set of a random number of random points in a given region of geographical space. The collection of corresponding numbers of such points in given subdivisions of this region is a count data set.
For example with census data, a population count on a given zone is the number of inhabitants of the zone. This number is obtained from the knowledge of the addresses of these people. The collection of coordinates of such addresses is the underlying point data set. Examples of areal interpolation of population or subpopulation counts can be found for example in
Goodchild and Lam (1980), Langford (2005), Mennis and Hultgren (2006), Reibel and Agrawal (2007). Other types of counts are encountered frequently, for example number of housing units in Reibel and Bufalino (2005).
Another frequent type of count related variable is the number of points per areal unit associated to a point data set: it is a density type variable. Examples of areal interpolation of population densities can be found in Yuan et al. (1997) and Murakami (2011). An even more general type is when the variable is a ratio of counts such as number of doctors per patient. There is an easy one to one correspondence between a count variable and a density variable which allows to transform one type into the other so that
any treatment of counts can be extended to densities and reversely.
A count variable belongs to the family of extensive variables, which are variables whose value on a region is obtained by summing up its values on any partition into subregions (aggregation formula hereafter). A density variable belongs to the family of intensive variables, which are variables whose value on a region is obtained from values on any partition into subregions by a weighted sum (see Do et al. 2014 for more details). In the case of population density, the weights are given by the areas of the subregions of the partition. In the remainder of this paper, we will concentrate on pure count variables.
We introduce a model for an extensive count variable by assuming that there exists an underlying (unreleased) Poisson point pattern $Z_Y$ (in the population example, the positions of the individuals of the population) and that the target variable $Y$ on a subzone $A$ is the number of points of $Z_Y$ in $A$. For a partition $\Omega_i, i=1,2,...,k$ of the region $\Omega$, the extensive property is clearly satisfied
$$Y_\Omega = \sum_{i=1}^k Y_{\Omega_i}$$
\noindent With the proposed Poisson point pattern assumption, for any zone $A$, $Y_A= \sum_i {\mathbf 1}_A(Z_i)$ is a Poisson distributed random variable with mean $\lambda_A =\int_A \lambda_{Z_Y} (s) ds,$ where $\lambda_{Z_Y}$ is the intensity of the point process $Z_Y$.
\noindent This model implies that $Y_A$ and $Y_B$ are automatically independent for all disjoint couples of subregions $A$ and
$B$ due to the Poisson process nature. We could use point pattern models with interaction effects while retaining the extensive property but we rather devote this article to this first case, keeping the interaction case for further developments.
\noindent As we will see in the next section, some methods we want to compare (dasymetric and univariate regression) make use of an auxiliary information. For the auxiliary variable $X$ to be relevant, there must be some relationship between the target variable and the auxiliary variable. In many cases a categorical information is used such as land cover: Reibel and Agrawal (2007) and Yuan et al. (1997) use land cover type data on a 30 meters resolution grid,
Mennis and Hulgren (2006) use 5 types of land cover obtained manually from aerial photography.
Li et al. (2007) just use a binary information such as unpopulated versus populated zones.
Reibel and Bufalino (2005) interpolate the 1990 census tract counts of people and housing using length of streets as auxiliary information. Mugglin and Carlin (1998) exploit population to interpolate the number of leukemia cases.
The use of a continuous auxiliary information can also be found: Murakami (2011) utilizes distance and land price to predict population density. In the rest of the paper, we concentrate on a single extensive auxiliary variable $X$ that is also a count in order to be able to consider the accuracy of all methods simultaneously (more details at the end of section \ref{methods}). Therefore it corresponds to another underlying point process $Z_X$ with intensity $\lambda_{Z_X}$.
\noindent The auxiliary variable $X,$ has to be known at intersection level in the case of dasymetric and at the target level in the case of regression. We need to write a formal relationship between our target variable and the auxiliary information. The model we propose assumes that the following relationship holds between the two underlying point processes intensity functions
\begin{equation}\label{intens}
\lambda_{Z_Y} (s) = \alpha + \beta \lambda_{Z_X}(s),
\end{equation}
\noindent where $s$ is location.
\noindent Therefore, the following relationship holds between $Y$ given $X$:
at the level of any subset $A$ of the region, the conditional distribution of $Y_A$ given $X_A = x_A$ is given by
\begin{equation}\label{model}
Y_{A} \sim {\mathcal P}(\alpha |A|+ \beta x_A)
\end{equation}
\noindent This relationship will be used at target level $A = T$ and at source level $A=S$. This model in its general form will be called auxiliary information model (AIM). In this model, the intensity of $Z_Y$ is driven by two effects: the effect of the auxiliary variable $X$ and the effect of the area of the zone. If we look at target level, the target variable is Poisson distributed with a mean comprising two parts ${\mathbb E}(Y_T)=\alpha|T| +\beta x_T$: the first part $\alpha |T|$ reflects the impact of the area of the zone $T$, whereas $\beta x_T$ is the impact of the auxiliary variable. The linearity of the expected value of $Y$ with respect to the area and to the auxiliary information is not canonical in a Poisson regression model for counts but in our case it derives naturally from \eqref{intens}.
\noindent In sections \ref{homAIM} and \ref{phomAIM}, we introduce two sub-models of model \eqref{model} depending on the intensity function $\lambda_{Z_Y}$. We consider the case of a constant intensity (homogeneous model) and the case of a piecewise constant intensity (piecewise homogeneous model).
\section{Methods}
\label{methods}
\subsection{Prediction techniques}
Do et al.(2014) classify areal interpolation methods into three groups: smoothing, dasymetric and regression based
methods. We discard smoothing since it is concerned with continuous target variables which is not adapted to our count data model.
We therefore focus here on the remaining two groups: dasymetric and regression based methods.
Dasymetric is a class of methods using a weighting scheme to allocate the original data to the intersections and then applying an aggregation step to get to target level. The simplest method in the dasymetric class is the areal weighting interpolation which uses area as weighting scheme: the data is allocated to the targets based on the assumption that the target variable is homogeneous at source level:
$$\hat{Y}_t=\sum_{s: s \cap t \not= \emptyset} \hat{Y}_{st}=\sum_{s: s \cap t \not= \emptyset} \dfrac{|A_{st}|}{|S_s|} {Y}_{s}.$$
\noindent Note that areal interpolation does
not use any auxiliary information other than area which is usually available.
The general dasymetric method is supposed to improve upon the areal weighting interpolation method when an additional variable is known to be linked to the target variable leading to alternative weighting schemes.
Voss et al. (1999) use road segment length and the number of road nodes for allocating demographic characteristics. Population, which is collected at fine levels in general, is often used as an auxiliary information for other variables like in Gregory (2002) or Mugglin and Carlin (1998).
Instead of homogeneity, the dasymetric method with auxiliary information $X$ assumes that the target variable is proportional to the auxiliary variable at intersection level.
$$\hat{Y}_t=\sum_{s: s \cap t \not= \emptyset} \hat{Y}_{st}=\sum_{s: s \cap t \not= \emptyset} \dfrac{X_{st}}{X_s} {Y}_{s}$$
\noindent where $X_s=\sum_t X_{st}$. This entails that $X$ has to be known at intersection level, which is quite restrictive.
Concerning the regression based methods, there are several types of regression based methods also involving auxiliary information (see Do et al, 2014). Given the nature of the target variable in our model (\ref{model}), we concentrate on the Poisson regression presented in Flowerdew et al. (1991) for the purpose of predicting population (which is an extensive variable) with categorical auxiliary information. Based on model (\ref{model}), a Poisson regression with identity link is performed at source level yielding estimators $\hat \alpha, \hat \beta$ for the parameters $\alpha$ and $\beta.$
\noindent The prediction of the target variable at intersection level is then obtained by
\begin{equation}\label{RegPredictor}
\hat{Y}^{REG}_{st}=\hat{\alpha} |A_{st}|+ \hat{\beta} X_{st}
\end{equation}
\noindent and the final step aggregates intersections predictions at target levels. The regression based methods can be considered as more powerful than the dasymetric methods in the sense that they can incorporate multivariate auxiliary information and that the knowledge of auxiliary information is only needed at target level and not at intersection level. However, the purpose of this paper being to compare the accuracy of dasymetric methods and Poisson regression methods from a methodological point of view and for the case of extensive count data, we therefore concentrate on the unidimensional auxiliary count variable case.
\noindent One property often quoted concerning these methods is the pycnophylactic property. This property requires the preservation of the initial data in the following sense at some geographical level: at source level for example, it means that the predicted value for source $S_s$ obtained by aggregating the predicted values on intersections with $S_s$ should coincide with the observed value on $S_s$. The enforcement of this property will allow us to introduce an improved version of the basic Poisson regression method.
\subsection{Prediction error criteria}
The accuracy assessment necessitates the choice of a prediction error criterion and of a geographic level. In this framework, examples of criteria are root mean square error or mean square error (Sadahiro 1999, Reibel 2006,...) at regional level (that is the union of all sources), or relative absolute error at target level (Langford, 2007).
We denote by MET a generic method of prediction and let $MET$ be $DAW$ for the areal weighting method, $DAX$ for the general dasymetric method, $REG$ for the Poisson regression method and $ScR$ for the scaled regression method which will be presented later in section \ref{acc2}. We recall that we assume all target zones are nested within source zones.
\noindent In section \ref{acc1}, we use mean square error at source level to compare the areal weighting and dasymetric methods. For method MET, the source level error is then computed as follows
\begin{equation}
{\text {Er}}^{MET}_S=\sum_{t \subset S}{\text {Er}}^{MET}_t=\sum_{t \subset S} {\mathbb E}(\hat{Y}^{MET}_t-Y_t)^2
\end{equation}
\noindent and the overall regional error is
\begin{equation}\label{overall}
{\text {Er}}^{MET}=\sum_{S} \sum_{t \in S} {\mathbb E}(\hat{Y}^{MET}_t -Y_t)^2
\end{equation}
In section \ref{acc2}, we use mean square error at target level
\begin{equation}
{\text {Er}}^{MET}_t={\mathbb E}(\hat{Y}^{MET}_{t}-Y_t)^2
\end{equation}
\noindent to compare the dasymetric and Poisson regression methods.
\noindent In general, we will also use the relative error criterion defined as
\begin{equation}
{\text {Re}}^{MET}_S={\sqrt{{\text {Er}}^{MET}_S} \over {\mathbb E}(Y_S)}
\end{equation}
where ${\text {Re}}^{MET}_S$ is the relative error of method $MET$ at source level for source $S$ with method $MET$.
\section{Relative accuracy of areal weighting and dasymetric: finite distance assessment}
\label{acc1}
Let us briefly summarize the findings of the assessments found in the literature for the comparison of general dasymetric and areal weighting.
For empirical assessments, several authors report that the dasymetric method improves upon areal weighting. Depending on the context, the improvement varies: Langford (2007) reports improvements of 54\%, 57\%, and 59\% better depending on the auxiliary information used; Reibel and Bufalino (2005) reports improvements of 71.26\% and 20.08\% with street length auxiliary information for the two target variables: housing units and total population. For theoretical assessments, Sadahiro (1999, 2000) compares the areal weighting interpolation and the point-in-polygon method with a theoretical model. We did not mention yet the point-in-polygon method because it is a very elementary one consisting in allocating a source value to the target which contains its centroid. Using a stochastic model, he finds that the factors that impact the accuracy of the methods are the size and shape of target and source zones, the properties of underlying points.
\noindent In this section, we prove some theoretical properties in subsection \ref{genAIM}, with two particular cases in \ref{homAIM} and \ref{phomAIM}, and a toy example in \ref{firsttoy}.
\noindent Since targets are nested within sources, the predictors of the two methods depend only on the source that contains the concerned target zone. For that reason, we focus on studying one source zone denoted by $S$. For a target $T$ in $S$, the two predictors are as follows
\begin{equation}\label{DAW}
\hat{Y}^{DAW}_{T}=\dfrac{|T|}{|S|}Y_S
\end{equation}
and
\begin{equation}\label{DAX}
\hat{Y}^{DAX}_{T}=\dfrac{x_T}{x_S}Y_S
\end{equation}
\subsection{General auxiliary information model}
\label{genAIM}
Lemma \ref{bias} gives the expression of the prediction bias and variance in model AIM for areal weighting interpolation and dasymetric interpolation at target level.
\begin{lemma}\label{bias} In model AIM, the prediction biases and variances
of areal weighting interpolation and dasymetric methods are given by
\begin{align}
{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)&=\beta x_S (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})\label{BiasTW}\\
{\mathbb E}(\hat{Y}^{DAX}_{T}-Y_T)&= \alpha |S| (\dfrac{x_T}{x_S}-\dfrac{|T|}{|S|})\label{BiasTx}\\
Var(\hat{Y}^{DAW}_{T}-Y_T)&=\beta x_S (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})\label{VarTW}\\
Var(\hat{Y}^{DAX}_{T}-Y_T)&=\alpha |S| (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})\label{VarTx}
\end{align}
\end{lemma}
\noindent First note that the two biases have opposite signs, in other words, if the areal weighting interpolation method underestimates then the dasymetric method overestimates and vice versa. This fact can be interpreted as follows: while the true intensity comprises two effects, these methods treat only one of them which causes the contrast. Although the signs of biases are opposite, their absolute values are both proportional to $\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S}$ which measures the divergence between the share of the auxiliary information in target $T$ with respect to $S$ and the share of the area of $T$ with respect to $S$. This divergence is also proportional to $\dfrac{x_S}{|S|}-\dfrac{x_T}{|T|}$ and hence can be viewed as a distance to proportionality between area and auxiliary information. The bias of the areal interpolation method with its assumption of homogeneity is independent in the areal effect $\alpha |S|$ but is proportional to the ignored auxiliary information effect, and reversely the dasymetric method which focuses on the effect of the auxiliary information gets rid of the $\beta x_S$ in its bias but is proportional to the ignored areal effect. We will build on this to propose a new method in the next section.
\noindent
The two variances have a common part $\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})$ which we can interpret as the loss of information when transferring data from a large source zone to a smaller target zone. For the remaining part, the same explanations as for the bias stands. Both variances have a parabola shape with respect to $x_T$ (respectively to $|T|$) with a maximum at $x_T=\dfrac{1}{2} x_S,$ (resp. $|T|=\dfrac{1}{2}|S|$): we can say loosely that the variances are maximum when the target zone is around a haft of the source. They vanish when the target zone is either empty or coincide with the source which makes sense. The reallocation to a larger target intuitively decreases the difficulty of the disaggregation problem except that the error also depends on the expected number of points so we should turn attention to relative error. If one divides the variances by the square of the expected number of points in the target zone ${\mathbb E}(Y_T)$, we can see that the relative error will tend to zero as ${\mathbb E}(Y_T)$ tends to infinity.
\noindent Since the dasymetric method is pycnophylactic, the bias at source level is zero. Lemma \ref{variances} reports the expression of the prediction variances in model AIM for areal weighting interpolation and dasymetric interpolation at source level.
\begin{lemma}\label{variances} In model AIM, the variances of areal weighting and dasymetric methods at the source level are
\begin{align}
Var_S^{DAW}&=\beta x_S \sum_T(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_S(1-\sum_T \dfrac{x_T^2}{x_S^2}) +\alpha |S| (1-\sum_T \dfrac{|T|^2}{|S|^2})\label{VarSW}\\
Var_S^{DAX} &=\alpha |S| \sum_T(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_S(1-\sum_T \dfrac{x_T^2}{x_S^2}) +\alpha |S| (1-\sum_T \dfrac{|T|^2}{|S|^2}) \label{VarSx}
\end{align}
\end{lemma}
\noindent To get an insight at impact of the number $n_T$ of the target zones, we consider the special case where all targets have the same size. In this case, $\dfrac{|T|}{|S|}=\dfrac{1}{n_T}$ for any $T$, and we get
\begin{align*}
Var_S^{DAW}&=(1 - \dfrac{1}{n_T})(\alpha |S|+\beta x_S )\\
Var_S^{DAX} &=(1 - \dfrac{1}{n_T})(\alpha |S|+\beta x_S )+ \sum_T (\dfrac{x_T^2}{x_S^2} - \dfrac{1}{n_T})(\alpha |S|-\beta x_S )
\end{align*}
It is obvious that the larger the number of the target zones, the larger the variances, which agrees with our conclusion concerning the size of targets. Indeed, when the area of the target zones gets smaller, the error on each target decreases but the total error at the source level gets larger due to the effect of the number of the targets.
\noindent We are now ready to compute the mean square error difference between the two methods.
We introduce the following quantities which quantify a relative contribution of the corresponding effect to the overall mean at the geographical level of a subregion $A$: $$I_A(X)= \frac{\beta x_A}{\alpha |A| + \beta x_A}$$
$I_A(X)$ is the relative contribution of variable $X$ and similarly $I_A(|.|)= \frac{\alpha |A|}{\alpha |A| + \beta x_A}$ is the relative contribution of the areal effect.
The imbalance between the two effects is measured by the difference
$$\Delta_A = I_A(|.|) - I_A(X) = \frac{\alpha |A| - \beta x_A}{{\mathbb E}(Y_A)}.$$
This quantity ranges between $-1$ when there is a pure $X$ effect and $1$ when there is a pure areal effect with a value of zero when the two effects are of equal size.
\noindent We can derive from lemmas \ref{bias} and \ref{variances} the expression of the absolute and relative errors of the two methods at source level as a function of the relative contributions terms.
\begin{theorem}\label{reler}
\begin{align}
{\text {Er}}^{DAW}_S=&I_S(X)^2 {\mathbb E}(Y_S)^2 D+ I_S(X) {\mathbb E}(Y_S)(D+B)+I_S(|.|){\mathbb E}(Y_S)C \label{ErDAW}\\
\Rightarrow ({\text {Re}}^{DAW}_S)^2&=I_S(X)^2 D+{1 \over {\mathbb E}(Y_S)}[ I_S(X)(D+B-C) +C]\label{ReDAW}
\end{align}
\begin{align}
({\text {Re}}^{DAX}_S)^2=I_S(|.|)^2 D +{1 \over {\mathbb E}(Y_S)}[I_S(|.|)(D-B+C) +B]\label{ReDAX}
\end{align}
where $D=\sum_T(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2, B=1- \sum_T\dfrac{x_T^2}{x_S^2}, C=1- \sum_T\dfrac{|T|^2}{|S|^2}$ are positive.
\end{theorem}
\noindent Note that $B,C$ and $D$ only depend on the the geometry of the problem and the auxiliary information, whereas the relative contribution terms and ${\mathbb E}(Y_S)$ depend on the coefficients $\alpha$ and $\beta$. It is interesting to mention the symmetry between the two methods which stands clearly in these formulas when we exchange the two contributions terms. One can derive from this theorem the difference between the relative errors of the two methods
\begin{equation}\label{DiffReDAWDAX}
({\text {Re}}^{DAW}_S)^2-({\text {Re}}^{DAX}_S)^2=-D * \Delta_S (1+{1 \over {\mathbb E}(Y_S)})
\end{equation}
\noindent which turns out to be clearly proportional to the imbalance term $\Delta_S$. Similarly, one can approximate the ratio of the two relative errors when ${\mathbb E}(Y_S)$ is large on the target $A=T$ and on the source $A=S$ by
\begin{align}
{{\text {Re}}^{DAW}_A\over {\text {Re}}^{DAX}_A}&\approx {I_A(X) \over I_A(|.|)}\label{RaReDAWDAX}
\end{align}
\noindent This ratio roughly ranges from $0$ to $+\infty$ at the extreme cases of a pure $X$ or areal effect showing that one can outperform the other by a large amount.
\noindent Let us now turn attention to the difference between the two errors.
\begin{theorem} \label{TheoDiffErDaxDaw} The difference between the errors of areal weighting and dasymetric methods on a target zone $T$ is
\begin{align*}
{\text {Er}}^{DAW}_T&-{\text {Er}}^{DAX}_T=(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\Delta_S {\mathbb E}(Y_S)({\mathbb E}(Y_S)+1)
\end{align*}
\end{theorem}
\noindent The important conclusion of this result is that the sign of the difference in error agrees with the sign of $\Delta_S$, i.e. the sign of $(\alpha |S| - \beta x_S)$. Moreover, the stronger the effect of the auxiliary information $I_S(X)$ is, the better the dasymetric method and the larger the difference between the two methods.
This computation result leads to a very interesting consequence: if one of two effect dominates on a given source, the related method wins on all target zones belonging to the source. It also shows that two methods will have the same accurracy if the two effects are balanced or the auxiliary variable is homogeneous.
The normalized difference between the two effects $\Delta_S$ clearly determines which method is best.
\noindent At this point, it seems natural to look for a linear combination of these two predictors
\begin{align}\label{compositeclass}
\hat{Y}^C_T (w)&= w\hat{Y}^{DAW}_{T}+(1-w)\hat{Y}^{DAX}_{T},
\end{align}
\noindent which would combine their good properties. It turns out that in the class of linear combinations of areal weighting and dasymetric predictors, the best predictor is given by the following theorem
\begin{theorem}\label{composite}
In model AIM, the best predictor in the sense of minimizing (with respect to the weight $w$) the errors on any target zone $T$ in the class \eqref{compositeclass} is
\begin{equation}\label{Com}
\hat{Y}^C_T =\hat{Y}^C_T (w^*)=\dfrac{\alpha |T|+\beta x_T}{\alpha |S|+\beta x_S}Y_S
\end{equation}
for $w^*=\dfrac{\alpha |T|}{\alpha |S|+\beta x_S}$. Its error and relative error are respectively given by
\begin{align}
{\text {Er}}^C_T&= \dfrac{\lambda_T(\lambda_S - \lambda_T)}{\lambda_S} \label{ErComDAX}\\
({\text {Re}}^C_S)^2&=\dfrac{1}{4{\mathbb E}(Y_S)}[\Delta_S^2 D+2\Delta_S (C-B)+D+2B+2C]\label{ErComDAW}
\end{align}
Moreover, this predictor coincides with the best linear unbiased predictor in model AIM.
\end{theorem}
\noindent Because $\dfrac{\lambda_T(\lambda_S - \lambda_T)}{\lambda_S} = Var(\hat{Y}^{DAX}_T-Y_T)-\lambda_S(\dfrac{x_T}{x_S}-\dfrac{\lambda_T}{\lambda_S})^2
= Var(\hat{Y}^{DAW}_T-Y_T)-\lambda_S(\dfrac{|T|}{|S|}-\dfrac{\lambda_T}{\lambda_S})^2 ,$ the prediction error of the best predictor is smaller than the variances of the other two methods and the distance is the more important that the auxiliary information is further from homogeneity. Of course, $\hat{Y}^C_T$ is not a feasible predictor since it depends on the unknown coefficients $\alpha$ and $\beta$ of model AIM but we will use it as a benchmark tool on the one hand and we will relate it later on to the regression predictor. If we look at the error at the level of source $S$, we have that
${\text {Er}}^C_S= \lambda_S - \sum_{T}\dfrac{\lambda_T^2}{\lambda_S}\leq \lambda_S-\dfrac{\lambda_S}{n_T(S)},$ where $n_T(S)$ is the number of targets in source $S$, and hence this predictor's accuracy is worse when all targets have the same expected number of points $\dfrac{\lambda_S}{n_T(S)}$. It is interesting to note that the relative error (at source level $S$) of the best predictor tends to zero as the expected number of points in the source $S$ tends to infinity, which was not the case for the dasymetric methods. For a fixed expected number of points in a given source $S$, we can easily find the value of the imbalance $\Delta_S$ which minimizes the relative error of $\hat{Y}^C_T$ $\Delta^* = \dfrac{B-C}{D}=\dfrac{\sum_T (\dfrac{|T|}{|S|})^2-(\dfrac{x_T}{x_S})^2}{\sum_T (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2}$ and thus derive a lower bound for the relative error for a given geometry.
\noindent Because intuitively, it is natural to think that areal weighting should be outperformed by dasymetric when the underlying process is inhomogeneous, we consider the two cases of homogeneous and piecewise homogeneous submodels.
\subsection{Homogeneous model}
\label{homAIM}
Areal weighting interpolation is a simple and natural rule which is based on the assumption that the target variable is homogeneous at source level. Indeed in model AIM, it is equivalent to assume that the point process is homogeneous and its intensity is therefore constant (equal to $\alpha > 0$) leading to:
$$Y_A \sim {\mathcal P}(\alpha |A|).$$
Substituting $\beta=0$ in \eqref{BiasTW}, \eqref{VarTW}, \eqref{VarSW} we get the bias, variance and error in this case:
\begin{align*}
{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)&=0\\
{\text {Er}}^{DAW}_T&= Var(\hat{Y}^{DAW}_{T}-Y_T)=\alpha |T| (1-\dfrac{|T|}{|S|})\\
{\text {Er}}^{DAW}_S&= Var_S^{DAW}=\alpha |S| (1-\sum_T \dfrac{|T|^2}{|S|^2})
\end{align*}
Since $\dfrac{1}{n_T} \leq \sum_T \dfrac{|T|^2}{|S|^2} \leq 1$, the error at source level is maximum when all target zones have the same size, and minimal when there is a unique target which coincides with the source.
Substituting $\beta=0$ in \eqref{Com} leads to the conclusion that the best linear unbiased predictor in the homogeneous AIM model is given by the areal weighting method which is a natural result.
Let us now turn attention to a very simple non homogeneous model to illustrate the intuitive fact that the areal weighting interpolation method is not the best choice in a non homogeneous situation.
\subsection{Piecewise homogeneous model}
\label{phomAIM}
Suppose the source zone $S$ comprises two homogeneous subzones $C_1$ and $C_2$ called control zones with intensities $\alpha_1$ and $\alpha_2$ respectively. In this case, we get
$$Y_A\sim {\mathcal P}(\alpha_* |A|)$$
where $A \subset C_*$ with $*=1,2$.
For simplification reasons, we assume the target zones to be nested within the control zones. The results of lemmas \ref{bias} and \ref{variances} give in this case
\begin{align*}
&{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)_{T: T \subset C_1}=\dfrac{|T|}{|S|}(\alpha_2-\alpha_1) |C_2|\\
&{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)_{T: T \subset C_2}=\dfrac{|T|}{|S|}(\alpha_1-\alpha_2) |C_1|\\
&Var_S^{DAW}=\alpha_1|C_1| (1-\sum_{T: T\subset C_1} \dfrac{|T|^2}{|S|^2})+\alpha_2|C_2| (1-\sum_{T: T\subset C_2} \dfrac{|T|^2}{|S|^2})\\
&{\text {Er}}^{DAW}_S=Var_S^{DAW}+\sum_{T: T\subset C_1}\dfrac{|T|^2}{|S|^2}(\alpha_2-\alpha_1)^2 |C_2|^2+\sum_{T: T\subset C_2}\dfrac{|T|^2}{|S|^2}(\alpha_2-\alpha_1)^2 |C_2|^2
\end{align*}
The variance has a similar structure to the one of the homogeneous model. The bias clearly shows that the difference between the two intensities of the subzones will drive the size of the error.
\subsection{Toy example}
\label{firsttoy}
In order to illustrate our findings, we use a simulated toy example. We intentionally drop the assumption that targets are nested in sources which was made for mathematical convenience and this will allow us to test the robustness of the results with respect to that assumption. On a square grid with 25 cells, we design three sources and seven targets as unions of cells. On Figure \ref{toyex1}, we see the design of sources and targets together with the cell counts for two target variables $Y_1$ and $Y_2$ and one auxiliary variable $X$ (for one particular draw). To generate $X$, we simulate a Poisson point process with an inhomogeneous intensity. We then recover the counts at the cell level to get the auxiliary information. The two target variables are then generated according to their relationship with the auxiliary variable (model (\ref{intens})) and the source values are obtained by aggregation of cells. The true value of two target variables is also shown at target level for accuracy comparison for one particular draw. For $Y_1$, we use the set of parameters $\alpha=80$ and $\beta=1$ and for $Y_2$ we use $\alpha=0$ and $ \beta=1$ so that the area has a strong impact on $Y_1$ and that $Y_2$ is only driven by $X$. Conditionally upon one draw of $X$ (for which we
observe 1011 points), we draw 1000 repetitions of $Y_1$ and $Y_2$ and present the relative error and the error at target level on Figure \ref{resulsDAWDAX}.
\begin{figure}[h!]
\begin{center}
\subfloat[$Y_1$ at cells]{\includegraphics[width = 1.5in]{FIGS/Y.pdf}}~
\subfloat[$X$ at cells]{\includegraphics[width = 1.5in]{FIGS/x1.pdf}}~
\subfloat[$Y_2$ at cells]{\includegraphics[width = 1.5in]{FIGS/Y0.pdf}}~
\subfloat[Targets]{\includegraphics[width = 1.5in]{FIGS/targets.pdf}}~\\
\subfloat[$Y_1, Y_2$ at sources]{\includegraphics[width = 1.5in]{FIGS/sources.pdf}}~
\subfloat[$X$ at intersections]{\includegraphics[width = 1.5in]{FIGS/intersections.pdf}}~
\subfloat[$Y_1, Y_2$ at targets]{\includegraphics[width = 1.5in]{FIGS/targets1.pdf}}
\end{center}
\caption{Toy example: Data at source, target, intersection zones.}\label{toyex1}
\end{figure}
The accuracy criterion is an average of the error over all 1000 draws. We present the relative error and the error at target level on Figure \ref{resulsDAWDAX}.
\begin{figure}[h!]
\begin{center}
\subfloat[DAW for $Y_1$]{\includegraphics[width = 1.5in]{FIGS/dawy1.pdf}}
\subfloat[DAX for $Y_1$]{\includegraphics[width = 1.5in]{FIGS/daxy1.pdf}}
\subfloat[DAW for $Y_2$]{\includegraphics[width = 1.5in]{FIGS/dawy2.pdf}}
\subfloat[DAX for $Y_2$]{\includegraphics[width = 1.5in]{FIGS/daxy2.pdf}}
\caption{Toy example. Comparison of areal weighting interpolation and dasymetric method.}
\label{resulsDAWDAX}
\end{center}
\end{figure}
Figure \ref{resulsDAWDAX} is in agreement with the theoretical results: areal weighting interpolation is better than dasymetric for $Y_1$ for which the areal effect dominates and worse for $Y_2$ for which the auxiliary information effect driven by $X$ dominates. We can see that $Y_2$ is less homogeneous than $Y_1$ on Figure \ref{toyex1} (a) and (c): at the right-bottom target zone $T_5$ has very small $Y_2$ counts. The dasymetric predictor is therefore very good for $Y_2$ (Figure \ref{resulsDAWDAX} (d)). To be more precise, we compute the ratio of the two average errors at target level for the two methods and it shows that areal weighting is best for $Y_1$ with a ratio of square root of errors of ${201 \over 452} = 44\%$, whereas dasymetric is best for $Y_2$ with ratio of square root of errors of ${26 \over 197} = 13\%$. Table \ref{MSEtoyex1} reports the square root of the overall regional error (from formula (\ref{overall})) for the three methods: areal weighting, dasymetric and Poisson regression. For $Y_1$, the regression method is best, for $Y_2$ dasymetric is best because the impact of $X$ is strong (almost no areal effect).
\begin{table}
\begin{center}
\caption {\bf Square root of overall regional errors}\label{MSEtoyex1}
\begin{tabular}{|m {6cm}|c|c|}
\hline
\centering Methods&$Y_1$&$Y_2$\\
\hline
Areal weighting interpolation&201&197\\
\hline
Dasymetric&452&26\\
\hline
Regression&55&33\\
\hline
\end{tabular}
\end{center}
\end{table}
For the practitioner, an important question is to be able to guess which method will perform better in a given situation. We might believe that a good correlation between $Y$ and $X$ is a sign that dasymetric based on $X$ will perform better than areal weighting. However in our case the correlation between $Y_1$ and $X$ is $0.94$ and the correlation between $Y_2$ and $X$ is $0.98$ which shows that this is a bad idea to rely on correlation. We could look at a measure of homogeneity to predict that areal weighting is the best method: in our case, the Gini coefficient of $Y_1$ is $0.14$ and of $Y_2$ is $0.40$ which goes in that direction. As we have seen in Theorem \ref{TheoDiffErDaxDaw}, the sign of the imbalance index at source level $\Delta_S$ determines which method is best (see Table \ref{BalanceToyex1}).
To be more precise, let us examine the results of Table \ref{ErrorInterLevel} in comparison with Theorem \ref{TheoDiffErDaxDaw}. Theorem \ref{TheoDiffErDaxDaw} shows that the difference of errors at target level is influenced by three factors: the mean number of points of the source, the imbalance of the source and the inhomogeneity of the auxiliary information of the given target. Let us look at each impact. For the influence of the inhomogeneity of the auxiliary variable, we compare targets of a source $S_1$ for example. The first two impacts are constant (1164 points and 0.10 imbalance), and we see that the more homogenous the auxiliary information is (in increasing order $A_{16}, A_{13}, A_{17}, A_{12}$), the more distant the two methods are (18, 20, 1728, 2529 respectively). To examine the impact of the imbalance, we consider intersection zones $A_{17}$ and $A_{33}$ which are nested within two sources with similar number of points ($S_1$ with 1164 points and $S_3$ with 1168 points respectively). Even though their inhomogeneity are not very different (0.12 vs 0.10), the difference of the errors on $A_{33}$ is 3.7 times larger than on $A_{17}$. This fact is explained by the distance between the two imbalances of $S_1$ and $S_3$ (0.10 vs 0.51). The last but not least effect is the number of points of the source zones. The comparison between $A_{24}$ and $A_{35}$ shows its impact: they have similar inhomogeneities (0.20 vs 0.19), not too different imbalances (0.41 vs 0.51), but the difference in errors for $A_{35}$ (25707) is three times larger than the difference in errors for $A_{24}$ (7938) and this is linked to the discrepancy in the mean numbers of points (1168 vs 679). As Table \ref{ErrorInterLevel} shows, the combination of the three impacts is very complex. In other words, choosing between the areal weighting interpolation and the dasymetric is not an easy problem.
\begin{table}
\begin{center}
\caption{\bf Imbalance index at source level}\label{BalanceToyex1}
\begin{tabular}{|c|c|c|c|}
\hline
Source zones&$S_1$&$S_2$&$S_3$\\
\hline
$Y_1$& 0.10& 0.41& 0.51\\
\hline
$Y_2$& -1& -1& -1\\
\hline
\end{tabular}
\end{center}
\end{table}
{\small
\begin{table}[h]
\hspace{-2.5cm}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{1}{|c|}{Sources}&\multicolumn{4}{c|}{$S_1$}&\multicolumn{3}{c|}{$S_2$}&\multicolumn{6}{c|}{$S_3$}\\
\hline
${\mathbb E}(Y_1)$&\multicolumn{4}{c|}{1164}&\multicolumn{3}{c|}{679}&\multicolumn{6}{c|}{1168}\\
\hline
Imbalance&\multicolumn{4}{c|}{0.10}&\multicolumn{3}{c|}{0.41}&\multicolumn{6}{c|}{0.51}\\
\hline
\multicolumn{1}{|c|}{Intersections}&$A_{12}$&$A_{13}$&$A_{16}$&$A_{17}$&$A_{21}$&$A_{24}$&$A_{25}$&$A_{31}$&$A_{32}$&$A_{33}$&$A_{34}$&$A_{35}$&$A_{37}$\\
\hline
\multicolumn{1}{|c|}{$|{|A|\over |S|}-{X_A \over X_S}|$}&0.14&0.01&0.01&0.12&0.03&0.20&0.18&0.02&0.17&0.10&0.06&0.19&0.09\\
\hline
$ {\text {Er}}^{DAX}-{\text {Er}}^{DAW}$& 2529&20&18&1728&135&7938&5999&199&19858&6435&16817&25707&2240\\
\hline
${\text {Re}}^{DAX}/{\text {Re}}^{DAW}$& 1.5&1.3&1.1&1.4&1.8&5.4&5.4&3.1&8.8&7.5&8.7&8.8&6.0\\
\hline
\end{tabular}
\caption{\bf Errors at intersection level for $Y_1$}\label{ErrorInterLevel}
\end{table}
}
\section{Relative accuracy of the other methods: asymptotic assessment}
\label{acc2}
Let us now try to extend the comparison to the Poisson regression method. This cannot be done anymore by finite distance methods and so we introduce an asymptotic framework.
Model \eqref{model} yields at source level
\begin{equation}
Y_s\sim {\mathcal P}(\alpha |S_s|+\beta x_s)
\end{equation}
where $x_s=\sum_{t: t\cap s \not= \emptyset} x_{st}$. Besides the Poisson regression predictor defined by \eqref{RegPredictor}, inspired by Theorem \ref{composite}, we propose a new predictor called scaled Poisson regression predictor defined as follows
\begin{equation}\label{SclPredictor}
\hat{Y}^{ScR}_{st}=\dfrac{\hat{\alpha}|A_{st}| + \hat{\beta}x_{st}}{\hat{\alpha}|S_s| + \hat{\beta}x_{s}}Y_S,
\end{equation}
\noindent where $\hat\alpha$ and $\hat\beta$ are the estimators of $\alpha$ and $\beta$ obtained through the Poisson regression at source level. Note that if model \eqref{intens} contains only one of the two effects (that of $X$ for example), then it is easy to see that
the predictor of the scaled regression method coincides with the dasymetric method (corresponding to $X$):
\begin{align*}
\hat{Y}^{ScR}_T={\hat{\beta}x_T\over \hat{\beta}x_S}Y_S=\hat{Y}^{DAX}_T.
\end{align*}
In section \ref{est}, we establish the asymptotic properties of the estimators $\hat\alpha$ and $\hat \beta$ and these results will enable us to compare the predictors in section \ref{predict}. Section \ref{subsectionToyexample} illustrates these results on a toy example.
\subsection{Estimators of the regression coefficients}
\label{est}
In this section, we adapt proofs from Fahrmeir and Kaufmann (1985) to establish the consistency and asymptotic normality of the estimators $\hat{\alpha}, \hat{\beta}$. We first need to describe an asymptotic framework. To be realistic, we assume that the whole region $\Omega$ is fixed and that the number of source zones $n_S$ (hereafter denoted by $n$) increases to infinity. In this section, the source zones will be denoted by $S_{n,i}: i=1,2,...,n$ and $\Omega = \cup_i S_{n,i}$.
Because of the extensive property of $X$, we also assume a similar property of $X_{n,i}$: the total auxiliary information on the region $\Omega$ remains constant $x_\Omega = \sum_i x_{n,i}.$ In order to get a consistent regression however we need the amount of information at source level to increase and we thus assume that the intensity of $Y$ increases with a rate $k_n \longrightarrow \infty$ so that
$$Y_{A}\sim {\mathcal P}(\alpha \widetilde{|A|}+\beta \tilde{x}_A)$$
where $\widetilde{|A|}=k_n|A|, \tilde{x}_A=k_n x_A$.
\noindent Let $\gamma=\colvec{\alpha}{\beta}, Z_A=\colvec{|A|}{x_A}, Z_{n,i}=\colvec{|S_{n,i}|}{x_{n,i}}$. With these notations we have $\lambda_A=\gamma' Z_A, \tilde{Z}_A= \colvec{\widetilde{|A|}}{\tilde{x}_A}=k_n Z_A$ and $Y_A \sim {\mathcal P}(k_n \lambda_A)$. The true value of the parameter $\gamma$ will be denoted by $\gamma_o=\colvec{\alpha_o}{\beta_o}$.
\noindent The log likelihood function $l_n(\gamma)$, the score function $s_n (\gamma)$ and the information matrix $F_n (\gamma)$ are then given by
\begin{align*}
l_n(\gamma) = \sum_{i=1}^n y_{n,i} \ln (\gamma' \tilde{Z}_{n,i})-\gamma' \tilde{Z}_{n,i}-\ln(y_{n,i}!)
\end{align*}
\begin{align*}
s_n(\gamma)&=\dfrac{\partial l_n(\gamma)}{\partial \gamma} = \sum_{i=1}^n \dfrac{\tilde{Z}_{n,i}}{\gamma' \tilde{Z}_{n,i}} y_{n,i} - \tilde{Z}_{n,i}\\
F_n(\gamma)&= cov_\gamma (s_n(\gamma))=\sum_{i=1}^n \dfrac{\tilde{Z}_{n,i}\tilde{Z}'_{n,i}}{\gamma' \tilde{Z}_{n,i}}
\end{align*}
Differentiation of the score yields
\begin{align*}
H_n(\gamma)&=-\dfrac{\partial s_n(\gamma)}{\partial \gamma}=\sum_{i=1}^n \dfrac{\tilde{Z}_{n,i}\tilde{Z}'_{n,i}}{(\gamma' \tilde{Z}_{n,i})^2}y_{n,i}
\end{align*}
It is easy to see that ${\mathbb E}_\gamma(s_n(\gamma))=0, {\mathbb E}_\gamma(H_n(\gamma))=F_n(\gamma)$. We further simplify the notations and use $s_n, F_n, H_n, {\mathbb E}$ instead of $s_n(\gamma_o), F_n(\gamma_o), H_n(\gamma_o), {\mathbb E}_{\gamma_o}$. It is clear that the matrix $H_n$ is positive definite and therefore the log likelihood function is concave which leads to a unique minimum. In the sequel, we also need the square root $F_n^{1/2}$ of the symmetric matrix $F_n$, i.e. $F_n^{1/2}F_n^{1/2}=F_n$.
\noindent Our asymptotic framework differs from that of Fahrmeir and Kaufmann (1985) in the sense that at each step they have one new observation whereas in our case at each step all observations are new and we have one more than at the previous step. For this reason, we modify slightly their conditions and assume that
\begin{itemize}
\item[(C1)] $\{\tilde {Z}_{n,i}\} \subset {\mathscr Z} \,\forall n,i$ with ${\mathscr Z}$ is a compact set.
\item[(C2)] $\lambda_{min} (\sum_i \tilde {Z}_{n,i}' \tilde {Z}_{n,i}) \to \infty$ as $n \to \infty$ where $\lambda_{min} (W)$ denotes the minimum eigenvalue of the matrix $W$.
\end{itemize}
\noindent Condition (C1) is satisfied if there exists two positive numbers $c_1, c_2$ (note that $||\tilde{Z}_{n,i}|| \not= 0$) s.t.
\begin{equation}\label{conditionC1}
c_1 <||\tilde{Z}_{n,i}||< c_2
\end{equation}
\noindent In that case, the number of source zones increases with the rate of growth of the intensity at a similar rate and the number of points in one source zone is quite stable during the change process.
\noindent Under these conditions, we get the following asymptotic behavior for the Poisson regression coefficients.
\begin{theorem}\label{ConsitencyNormality}
Under conditions (C1) and (C2), the following statements holds for the Poisson regression estimator $\hat{\gamma}_n$ of $\gamma$
\begin{itemize}
\item[(i)]$\hat{\gamma}_n \to_p \gamma_o$ (weak consistency)\\
\item[(ii)] $F_n^{1/2} (\hat{\gamma}_n - \gamma_o) \to_d {\cal N}(0,{\mathbf I})$ (asymptotic normality)
\end{itemize}
\end{theorem}
\noindent In the next section, we use these results to study the asymptotic behavior of the predictors.
\subsection{Predictors}
\label{predict}
In this section, we consider the asymptotic properties of the following two predictors: the regression predictor \eqref{RegPredictor} and the scaled regression predictor \eqref{SclPredictor}. We prove that the scaled regression predictor is asymptotically as accurate as the unfeasible composite predictor. We also compare these two methods with areal weighting interpolation and dasymetric interpolation.
The first proposition is concerned with the pycnophylactic property, which is of interest in the areal interpolation literature. It shows that, whereas it is satisfied by the scaled Poisson regression predictor, it is not satisfied at source level by the ordinary Poisson regression predictor but only at region level.
\begin{proposition}\label{ProPycno}
The scaled Poisson regression predictor satisfies the pycnophylactic property at source level. The ordinary Poisson regression predictor is pycnophylactic at region level and asymptotically pycnophylactic at source level.
\end{proposition}
\noindent To prove proposition \ref{ProPycno}, we need the following asymptotic normality result for the target variables
\begin{equation}\label{app}
\dfrac{Y_{ni}-\gamma_o' \tilde{Z}_{n,i}}{\sqrt{\gamma_o' \tilde{Z}_{n,i}}}\to_d {\mathcal N}(0,1)
\end{equation}
\noindent We now turn attention to the asymptotic behavior of the prediction error for the ordinary Poisson regression predictor.
\begin{theorem}\label{TheNormalityREG}
The asymptotic normality of the prediction error of the Poisson regression predictor at source level is given by
$$\dfrac{\hat{Y}_{ni}^{REG}-Y_{ni}}{\sqrt{\gamma_o'\tilde{Z}_{n,i}}} \to_d {\mathcal N}(0,1)$$
\noindent If we also assume a lower bound for $\tilde{Z}_{T}$, the following similar result at the target level holds
$$\dfrac{\hat{Y}_{T}^{REG}-Y_{T}}{\sqrt{\gamma_o'\tilde{Z}_{T}}} \to_d {\mathcal N}(0,1)$$
\end{theorem}
\noindent The next result is about the quadratic prediction error and relative prediction error of the Poisson regression predictor.
\begin{theorem}\label{TheErrReg}
For any $\eta >0$, there exists a sequence of sets $\{Q_i\}_i: {\mathbb P}(Q_i) \to 1$ such that
\begin{align*}
-\eta +{\gamma}_o'\tilde{Z}_T <{\mathbb E}(\hat{Y}^{REG}_T-Y_T)^2{\mathbf 1}_{Q_i}<\eta +{\gamma}_o'\tilde{Z}_T
\end{align*}
\noindent If the number of target zones contained in one source zone $S_{n,i}$ is bounded, the relative error at source level can be approximated by
\begin{equation}\label{approRelativeErrorREG}
{\text {Re}}^{REG}_{n,i}\approx {1 \over \sqrt{{\mathbb E}(Y_{n,i})}}={1 \over \sqrt{\alpha_o \widetilde {|S_{n,i}|} + \beta_o \tilde{x}_{n,i}}}
\end{equation}
\end{theorem}
\noindent Because $Var(Y_T)={\mathbb E}(Y_T)={\gamma}_o'\tilde{Z}_T$, this theorem says that the quadratic prediction error of the regression predictor is asymptotically equivalent to the variance of the underlying process. Equation \eqref{approRelativeErrorREG} shows that the relative error of the regression predictor is going to be small when the number of points on a source zone is large. However, this number being bounded by condition (C1), this relative error cannot converge to zero in this framework.
\noindent Let us now turn attention to the difference between the relative prediction errors of the Poisson regression method and that of the areal weighting and the dasymetric methods.
If the target zones are nested within the source zones and the number of target zones contained in one source is bounded, we get the following approximation at source level for the differences between the relative errors of the methods when ${\mathbb E}(Y_{n,i})$ are large:
\begin{align}\label{diffapproRelEr}
[({\text {Re}}^{REG}_{n,i})^2-({\text {Re}}^{DAW}_{n,i})^2
&\approx -(1 + \Delta_{n,i})^2 \sum_T (\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2\\
[({\text {Re}}^{REG}_{n,i})^2-({\text {Re}}^{DAX}_{n,i})^2
&\approx -(1 - \Delta_{n,i})^2 \sum_T (\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2\\
\end{align}
\noindent This result shows that, among the three methods: areal weighting, dasymetric and Poisson regression, regression outperforms the other two methods asymptotically (negative sign). However, from the proof in the annex, we can see that if $(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})=0$ then the regression is less accurate than areal weighting and dasymetric asymptotically so that none of them is always dominant.
\noindent For areal weighting and dasymetric predictors, we have seen that if one method is better on one target, then it is also true on all targets contained on the same source zone. The difference between the accuracy of the regression method and the other two methods depends on the difference of ratios ${|T| \over x_T}-{|S| \over x_S}$: the higher this difference, the larger the difference between regression and the other two.
\noindent The fact that the regression predictor doesn't satisfy the pycnophylactic property is not surprise but the fact that it does satisfy this property on the whole region is interesting. The idea of scaling to obtain the pycnophylactic property can be found also in Yuan et. al. (1997) for ordinary linear regression without theoretical justifications; we have extended it to the Poisson regression case and provided some theoretical motivation for it.
\noindent We now turn attention to the scaled regression and prove it is better than the unscaled one and that
its accuracy can be approximated by that of the unfeasible composite predictor.
\noindent The first lemma proves an asymptotic equivalence between the scaled regression predictor and the unfeasible composite predictor.
\begin{lemma}\label{LeScRCom}
For any target $T$,
\begin{equation}
\hat{Y}^{ScR}_T-\hat{Y}^{C}_T \to_p 0
\end{equation}
\end{lemma}
\noindent The next result is about the quadratic prediction error of the scaled Poisson regression predictor.
\begin{theorem}\label{TheScR}
For any $\eta >0$, there exists a sequence of sets $\{Q_i\}_i: {\mathbb P}(Q_i) \to 1$ such that
$$-\eta +\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}<{\mathbb E}(\hat{Y}^{ScR}_T-Y_T)^2{\mathbf 1}_{Q_i}< \eta +\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}
$$
\end{theorem}
Since ${\text {Er}}^{C}_T=\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}$, this theorem shows that the quadratic prediction error of the scaled regression predictor is asymptotically equivalent to the one of the composite. Consequently, the scaled regression method is the best among the areal weighting, the dasymetric and the regression predictors.
\subsection{Accuracy: simulation assessment with a toy example}\label{subsectionToyexample}
We devise a simple simulation to illustrate these results. On a square region $\Omega$ with $16 \times 16$ cells, we build three systems of sources with respectively 14, 7 and 4 sources (see Figure \ref{cellsandsources}). We simulate two Poisson point patterns (our auxiliary information) with an expected overall number of points of $100,000$: $X_1$ is very inhomogeneous (Gini coefficient of cell counts of 0.74 with 100,247 points) and $X_2$ is very homogeneous (Gini coefficient of cell counts of 0.03 with 100,008 points).
\begin{figure}[h!]\label{cellsandsources}
\begin{center}
\subfloat[Cells]{\includegraphics[width = 2.5in]{FIGS/cells.pdf}}
\subfloat[Sources 1]{\includegraphics[width = 2.5in]{FIGS/sources1.pdf}}\\
\subfloat[Sources 2]{\includegraphics[width = 2.5in]{FIGS/sources2.pdf}}
\subfloat[Sources 3]{\includegraphics[width = 2.5in]{FIGS/sources3.pdf}}
\caption{Spatial polygons}
\end{center}
\end{figure}
\noindent Target variables are then generated following model (\ref{model}). For each of the auxiliary variables, we choose four couples of coefficients $\alpha, \beta$ to study the effects of imbalance so that we get eight different target variables. Table \ref{imb} reports the minimum, maximum and average imbalance for each case and for each system of source zones.
{\small
\begin{table}
\caption{\bf Imbalances}\label{imb}
\vskip 0.5cm
\hspace{-1cm}
\begin{tabular}{|c|c|r|r|r|}
\hline
\bf Case&\bf Sources&\bf Min& \bf Mean & \bf Max\\
\hline
$\alpha=100$&14 sources& -0.92& -0.40& 0.96
\\
\cline{2-5}
$ \beta=1$&7 sources& -0.90&-0.59&0.64
\\
\cline{2-5}
$ X_1$&4 sources& -0.87&-0.44& 0.64
\\
\hline\hline
$\alpha=600$&14 sources& -0.62& 0.11&0.99
\\
\cline{2-5}
$ \beta=1$&7 sources& -0.53&-0.05&0.93
\\
\cline{2-5}
$ X_1$&4 sources& -0.43& 0.07& 0.93
\\
\hline \hline
$\alpha=1000$&14 sources& -0.43&0.29& 1
\\
\cline{2-5}
$ \beta=1$&7 sources& -0.33&0.17&0.96
\\
\cline{2-5}
$ X_1$&4 sources& -0.21& 0.26& 0.96
\\
\hline \hline
$\alpha=1000$&14 sources& 0.6& 0.86& 1
\\
\cline{2-5}
$ \beta=0.1$&7 sources& 0.67&0.84& 1
\\
\cline{2-5}
$ X_1$&4 sources& 0.74& 0.86& 1
\\
\hline
\end{tabular}
\quad
\begin{tabular}{|c|c|r|r|r|}
\hline
\bf Case&\bf Sources&\bf Min& \bf Mean & \bf Max\\
\hline
$\alpha=100$&14 sources& -0.63& -0.6& -0.58
\\
\cline{2-5}
$ \beta=1$&7 sources& -0.61& -0.6&-0.59
\\
\cline{2-5}
$ X_2$&4 sources& -0.6&-0.6&-0.59
\\
\hline \hline
$\alpha=600$&14 sources& 0.16& 0.21& 0.22
\\
\cline{2-5}
$ \beta=1$&7 sources& 0.18& 0.2& 0.22
\\
\cline{2-5}
$ X_2$&4 sources&0.19& 0.2& 0.22
\\
\hline \hline
$\alpha=1000$&14 sources& 0.39& 0.43& 0.45
\\
\cline{2-5}
$ \beta=1$&7 sources& 0.41& 0.43& 0.44
\\
\cline{2-5}
$ X_2$&4 sources& 0.42& 0.43& 0.44
\\
\hline \hline
$\alpha=1000$&14 sources&0.92& 0.92& 0.93
\\
\cline{2-5}
$ \beta=0.1$&7 sources&0.92& 0.92& 0.93 \\
\cline{2-5}
$ X_2$&4 sources&0.92& 0.92& 0.93 \\
\hline
\end{tabular}
\end{table}
}
We then apply the four considered methods (areal weighting, dasymetric, Poisson regression and scaled Poisson regression) to transfer the data from each of the three systems of source zones to cell level. For each case, we generate the data 1000 times, and calculate prediction errors for each method and each iteration. Table \ref{errorToyex2} (respectively Table \ref{relativeErrorToyex2}) reports the average absolute square root of prediction errors (respectively the average absolute square root of relative prediction errors). The two tables also present the mean of the target variables at region level ${\mathbb E}(Y_\Omega)=\alpha |\Omega|+ \beta x_\Omega$ (because it appears in Theorem \ref{TheErrReg}) and the theoretical composite prediction error as a benchmark (see Theorem \ref{TheScR}).
\begin{center}
\begin{table}
\caption{\bf Square root of prediction errors.}\label{errorToyex2}
\vskip 0.5cm
\begin{tabular}{|>{\centering}p{2cm}|r|r|c|r|r|r|r|}
\hline
\multicolumn{1}{|c|}{\bf Methods} & {\centering{\bf $\sqrt{{\mathbb E}(Y_\Omega)}$}}& {\centering \textbf{Sources}} &{\centering \textbf{DAW }} & {\centering \textbf{DAX}} &{\centering \textbf{REG}}&{\centering \textbf{ScR}}&{\centering{\bf Composite}}\\
\hline
$\alpha=100$&\multirow{3}{*}{ 354.7
}&14 sources&6580.5 &1621.6 &353.0 &333.1 & 332.1 \\
\cline{3-8}
$ \beta=1$&&7 sources&6962.7 &2087.7 &352.1 & 341.3 &341.2
\\
\cline{3-8}
$ X_1$&&4 sources&7215.7& 2090.6 & 352.1 &347.6 &347.8\\
\hline
\hline
$\alpha=600$&\multirow{3}{*}{ 503.8
}&14 sources&6589.6 &9532.5& 500.6& 481.9& 482.1
\\
\cline{3-8}
$ \beta=1$&&7 sources& 6971.3& 12363.7 & 502.9 & 493.4 & 491.7 \\
\cline{3-8}
$ X_1$&&4 sources& 7225.3 &12374.7 & 502.5 & 498.8& 497.4 \\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 596.9
}&14 sources& 6597.3& 15878.2& 594.0 & 574.5& 574.2\\
\cline{3-8}
$ \beta=1$&&7 sources& 6982.2& 20595.2& 595.6& 586.5 & 584.6
\\
\cline{3-8}
$ X_1$&&4 sources&7229.6 &20614.5 & 594.4 & 590.8 & 590.3
\\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 515.8
}&14 sources&826.7 &15875.6 & 513.6 & 500.7 &500.9 \\
\cline{3-8}
$ \beta=0.1$&&7 sources& 861.9& 20592.8& 515.0& 509.4 & 508.3
\\
\cline{3-8}
$ X_1$&&4 sources& 883.5 &20612.4 & 514.5 & 512.3 & 511.6 \\
\hline
\hline
$\alpha=100$&\multirow{3}{*}{ 354.4
}&14 sources&458.8& 353.5& 356.6& 348.3&344.5 \\
\cline{3-8}
$ \beta=1$&&7 sources& 469.0& 358.6& 357.7& 354.2 &349.5
\\
\cline{3-8}
$ X_2$&&4 sources&474.2 &361.3 &359.6 &358.2& 351.6
\\
\hline
\hline
$\alpha=600$&\multirow{3}{*}{ 503.6
}&14 sources&573.9 &677.4& 504.9& 492.9 &489.6 \\
\cline{3-8}
$ \beta=1$&&7 sources& 587.3& 690.0& 508.1&503.1& 496.6 \\
\cline{3-8}
$ X_2$&&4 sources&591.2& 697.7 &509.1& 507.2& 499.6\\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 596.7
}&14 sources&654.8& 969.9& 599.0 &585.0 & 580.1 \\
\cline{3-8}
$ \beta=1$&&7 sources& 665.5& 992.6& 601.7& 595.7 & 588.4 \\
\cline{3-8}
$ X_2$&&4 sources& 671.8 &1002.6 &602.2 & 600.0 & 592.0 \\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 515.8
}&14 sources&502.9& 924.7 &518.6& 506.8 &501.4 \\
\cline{3-8}
$ \beta=0.1$&&7 sources& 510.3 & 947.4 & 520.4 & 515.5& 508.6 \\
\cline{3-8}
$ X_2$&&4 sources& 512.9 & 960.9& 522.2 &520.2& 511.7 \\
\hline
\end{tabular}
\end{table}
\end{center}
\begin{center}
\begin{table}
\vskip 0.5cm
\begin{tabular}{|>{\centering}p{2cm}|r|r|c|r|r|r|r|}
\hline
\multicolumn{1}{|c|}{\bf Methods} & {\centering{\bf ${\mathbb E}(Y_\Omega)$}}& {\centering \textbf{Sources}} &{\centering \textbf{DAW }} & {\centering \textbf{DAX}} &{\centering \textbf{REG}}&{\centering \textbf{ScR}}&{\centering{\bf Composite}}\\
\hline
$\alpha=100$&\multirow{3}{*}{ 125847
}&14 sources& 50.459 & 47.464 & 6.979 & 6.872& 6.873
\\
\cline{3-8}
$ \beta=1$&&7 sources& 55.664 & 54.021 & 6.970 & 6.941& 6.951
\\
\cline{3-8}
$ X_1$&&4 sources& 57.122 & 54.054 & 6.978 & 6.966& 6.969
\\
\hline
\hline
$\alpha=600$&\multirow{3}{*}{ 253847
}&14 sources& 20.458 & 59.087 & 3.492 & 3.422 & 3.421
\\
\cline{3-8}
$ \beta=1$&&7 sources& 21.865 & 73.067 & 3.495 & 3.471& 3.466
\\
\cline{3-8}
$ X_1$&&4 sources& 22.557 & 73.208& 3.490 & 3.480& 3.481
\\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 356247
}&14 sources& 14.859 & 62.210 & 2.813 & 2.754& 2.757
\\
\cline{3-8}
$ \beta=1$&&7 sources& 15.827 & 77.405 & 2.823& 2.801 & 2.794
\\
\cline{3-8}
$ X_1$&&4 sources& 16.324 & 77.589 & 2.820& 2.811& 2.808
\\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 266025
}&14 sources& 4.297 & 72.322 & 3.093& 3.018 &3.020
\\
\cline{3-8}
$ \beta=0.1$&&7 sources& 4.433 & 89.662 & 3.100 & 3.069& 3.064 \\
\cline{3-8}
$ X_1$&&4 sources& 4.508 & 90.028 & 3.098 & 3.086& 3.082 \\
\hline
\hline
$\alpha=100$&\multirow{3}{*}{ 125608
}&14 sources& 5.685 & 4.503 & 4.545& 4.438 & 4.391
\\
\cline{3-8}
$ \beta=1$&&7 sources& 5.799 & 4.569 & 4.559 & 4.514& 4.455
\\
\cline{3-8}
$ X_2$&&4 sources& 5.859 & 4.603 & 4.581 & 4.564 & 4.482
\\
\hline
\hline
$\alpha=600$&\multirow{3}{*}{ 253608
}&14 sources& 3.574 & 4.121 & 3.186 & 3.108 & 3.088 \\
\cline{3-8}
$ \beta=1$&&7 sources& 3.655 & 4.196 & 3.205 & 3.174 &3.134
\\
\cline{3-8}
$ X_2$&&4 sources& 3.679 & 4.236 & 3.211 & 3.199& 3.153 \\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 356008
}&14 sources& 2.918 & 4.084 & 2.692& 2.627& 2.606
\\
\cline{3-8}
$ \beta=1$&&7 sources& 2.966 & 4.172 & 2.704 & 2.677 &2.645
\\
\cline{3-8}
$ X_2$&&4 sources& 2.993 & 4.211 & 2.706 & 2.696& 2.661
\\
\hline
\hline
$\alpha=1000$&\multirow{3}{*}{ 266001
}&14 sources& 3.022 & 5.133 & 3.118& 3.045& 3.015
\\
\cline{3-8}
$ \beta=0.1$&&7 sources& 3.068 & 5.246 & 3.129 & 3.099& 3.059
\\
\cline{3-8}
$ X_2$&&4 sources& 3.084 & 5.320 & 3.140 & 3.128 & 3.078
\\
\hline
\end{tabular}
\caption{\bf Relative prediction errors (in percentages).}\label{relativeErrorToyex2}
\end{table}
\end{center}
Table \ref{errorToyex2} shows that when $\beta$ is fixed and $\alpha$ increases, resulting in an increase of the mean of the target variables at region level ($\alpha |\Omega|+ \beta x_\Omega$), all errors get larger. For fixed coefficients $\alpha, \beta$, the errors increase from the first set of sources to the third one, which is natural since the available information decreases from 14 observations for the first case, to 4 for the third.
In accordance with the toy example of section \ref{firsttoy}, the errors for $X_2$ are much smaller than the ones for $X_1$ for the areal weighting and the dasymetric methods due to the difference of homogeneity of the auxiliary variables. The more homogeneous the auxiliary variable is, the more accurate the areal weighting interpolation and dasymetric methods are. As we discussed earlier, if the auxiliary information is almost homogeneous, the regression might be less exact than the areal weigthing. But Table 4 shows that the regression methods are still quite good for $X_2$: in general, they are better than the two classical methods except in some particular cases. The errors of the regression and scaled regression methods are very comparable for $X_1$ and $X_2$. Indeed, the prediction errors of the regression are very similar to the mean of the target variables $Y$ and the accuracy of the scaled regression predictor is equivalent to the one of the composite predictor.
The effect of imbalance can be studied by looking at a change of $\alpha$ with fixed $\beta$. A larger $\alpha$ corresponds to a larger influence of the areal effect $\alpha |S|$ which is expected to lead to the domination of the areal weighting interpolation method (indeed we can observe this effect in the table for both auxiliary variables $X_1$ and $ X_2$). The imbalance also affects the regression and scaled regression methods: if one of the two effects $\alpha |S|$ and $\beta X_S$ is much larger than the other one, the corresponding errors seem to be further from their benchmarks (respectively the mean of $Y$ and the composite prediction error): see for example the cases $\alpha=1000, \beta=0.1$. However this effect is not very large for regression and scaled regression. One factor which influences more these two methods is the homogeneity of the auxiliary variable: comparing the results for $X_1$ and $X_2$ illustrates this. For $X_1$, the regression prediction errors are almost equal to their respective benchmarks and the amount of initial information does not seem to have a big influence (the errors are not monotonic from the first to the third set of sources). For $X_2,$ the accuracy increases with the number of source zones (the best being for the first one) and the errors of the regression method tend to the mean of $Y$. If we consider the particular case $\alpha=1000, \beta=0.1$ for $X_2$, the areal impact is much stronger than the auxiliary information impact, and we see that the areal weighting interpolation is the best method, and that the scaled regression predictors can catch up the areal weighting interpolation when there are more source zones.
\noindent Table \ref{relativeErrorToyex2} contains the corresponding relative errors. For example, in the case $\alpha = 1000, \beta = 1$ for $X_2$, we see that whatever the number of sources the relative error is around $2.6\%$ to $2.7\%$ for the scaled regression (very close to the benchmark given by the last column) whereas dasymetric is around $4\%$ and areal weighting around $3\%$. Looking at the second column, we see that when the expected number of points increases, the relative prediction error tends to decrease which was naturally not the case for the prediction error itself.
\noindent We now turn attention to the robustness of the methods with respect to the model. As previously with the same geometrical design, we generate two auxiliary information scenarios: $X_1$ is as in the previous simulation, and $X_3$ is inhomogeneous and uncorrelated with $X_1$ (correlation coefficient of $-0.16$).
A target variable $Y$ is generated from $X_3$ with the relationship $Y_A \sim {\mathcal P}(600 |A| + X_3)$. We transfer $Y$ from the first set of 14 sources to the cells (Figure \ref{cellsandsources}) by using areal weighting interpolation, dasymetric interpolation with $X_1$ and $X_3$ as auxiliary variables, the regression methods (REG and SCR) with the true model (areal effect and $X_3$), a simple model with only the areal effect, an auxiliary variable model with an irrelevant variable (with area and $X_1$), an auxiliary variable model involving an unnecessary variable (the area and both $X_1$ and $ X_3$). Table \ref{tableRobust} presents the results.
\begin{table}[h]
\begin{center}
\begin{tabular}{|l|r|}
\hline
\multicolumn{1}{|c|}{ \bf Methods} & {\centering \textbf{Relative error}} \\
\hline
\hline
DAW& 7.74\\
\hline
\hline
DAX with $X_3$&9.49 \\
\hline
REG with area and $X_3$ &2.66 \\
\hline
ScR with area and $X_3$&2.62\\
\hline
\hline
DAX with $X_1$&14.70\\
\hline
REG with area and $X_1$&10.48\\
\hline
ScR with area and $X_1$&8.26 \\
\hline
\hline
REG with area &10.62\\
\hline
ScR with area&7.74\\
\hline
\hline
REG with area, $X_1$ and $X_3$& 2.66\\
\hline
ScR with area, $X_1$ and $X_3$&2.62\\
\hline
\end{tabular}
\end{center}
\caption{\bf Robusness of methods.}\label{tableRobust}
\end{table}
The most accurate method is the scaled regression with area and $X_3$ (true model). Note that the relative error for DAW and ScR with area only is the same which was expected since we proved that in that case the two methods coincide. The regression methods for the model involving area plus $X_1$ and $X_3$ as auxiliary have the same errors (2.66\% and 2.62\%): in other words using unnecessary variables in the regression does not decrease the accuracy. On the other hand, if we use the regression with a wrong choice of auxiliary variable, it gives bad predictions (10.48\% and 8.26\% for the model with area and $X_1$, 10.62\% and 7.74\% for the model with only areal effect). The dasymetric method with $X_3$ is better than with $X_1$ (9.49\% vs 14.70\%) which makes sense because the correlation of the target variable $Y$ with $X_3$ is $0.998$ while with $X_1$ it is of $-0.159$ however we see that despite the strong correlation between $Y$ and $X_3$ the dasymetric method with $X_3$ is not so good because the areal effect is strong. The scaled regression is always better than the regression method and the scaled regression in the case of areal effect model yields the same result as the areal weighting interpolation method.
\section{Conclusion}
In this paper we have analyzed the accuracy of four areal interpolation methods: areal weighting interpolation, dasymetric interpolation, Poisson regression and scaled Poisson regression for the case of count data. We have introduced a model based on an underlying Poisson point pattern to be able to evaluate the accuracy of the different methods. We have proposed a scaled version of the Poisson regression method resulting in the enforcement of the pycnophylactic property. Areal weighting interpolation and dasymetric interpolation have been compared with a finite distance approach and the regression methods have been compared together and with the previous ones with an asymptotic approach.
We found out that one shouldn't rely on the correlation of the target variable and the auxiliary variable or on the homogeneity of the target variable to decide between areal interpolation or dasymetric but we should also take into account the relative imbalance between the areal effect and the auxiliary effect. A strong areal effect leads to the dominance of the areal weighting interpolation and a strong auxiliary effect is in favor of the dasymetric method. Moreover, the imbalance index allows to approximate the ratio of the two relative errors and their lower bounds as the number of points on the source zones gets large. We establish the formula for the best linear predictor (therefore better than the areal weighting and the dasymetric), which leads to the introduction of the scaled regression method.
For the comparison of areal weighting and dasymetric, a combination of several factors explains the complexity of the behavior: the size of sources, the auxiliary information, the number and size of target zones, \ldots The error at source level
is better when sources are divided into a smaller number of target zones. A large number of points makes the error at source level worse but improves the accuracy of the relative error. These two types of errors have the same behavior as a function of the imbalance index.
The impact of the expected number of points and of the inhomogeneity on the comparative advantage of the methods should not be forgotten: indeed when we have several sources, the sign of the imbalance index may vary from source to source and the overall effect, being an aggregate of the source level effect, will also depend on the magnitude of the source error differences which is driven by the expected number of points and by the inhomogeneity.
We proved that the accuracy of the unfeasible composite predictor is decreasing when the expected number of points are similar on all targets and this fact extends to scaled regression (due to the approximation results).
To be able to include the regression methods in the comparison, we need to resort to some asymptotic approach. We propose an asymptotic framework and prove that the Poisson regression prediction error is equivalent to the variance of the underlying process and for the scaled regression, it is approximated by the composite's prediction error. These results show the regression predictor is not automatically better than the areal weighting interpolation or the dasymetric method, but when the number of points at source level is large, it is in general better. Finally the scaled regression turns out to be the best one among the considered methods. These results are confirmed by our simulation study of the last section.
The robustness with respect to the model is also considered. The simulations show that a model with extra auxiliary variables doesn't create any loss while missing variables or unrelated variables (in place of the correct ones) decrease the accuracy of all methods.
\section{Appendix}\label{sectionAppendix}
\subsection{Proofs}
\subsubsection{Proof of Lemma \ref{bias} and lemma \ref{variances}}
From \eqref{DAW}, \eqref{DAX} and the properties of a Poisson point process we have
\begin{align*}
{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)&= {\mathbb E}(\dfrac{|T|}{|S|} Y_S-Y_T)=\dfrac{|T|}{|S|}(\alpha |S|+\beta x_S)-(\alpha |T|+\beta x_T)=\beta x_S (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})
\end{align*}
\begin{align*}
{\mathbb E}(\hat{Y}^{DAX}_{T}-Y_T)&= {\mathbb E}(\dfrac{x_T}{x_S} Y_S-Y_T)=\dfrac{x_T}{x_S}(\alpha |S|+\beta x_S)-(\alpha |T|+\beta x_T)=\alpha |S| (\dfrac{x_T}{x_S}-\dfrac{|T|}{|S|})
\end{align*}
Taking into account the independence of two disjoint target zones with the fact that the target $T$ is a portion of the source $S$ the variances of each method are given as follows
\begin{align*}
Var(\hat{Y}^{DAW}_{T}-Y_T)&=Var(\dfrac{|T|}{|S|}Y_S-Y_T)\\
&=\dfrac{|T|^2}{|S|^2} Var(Y_S)+Var(Y_T)-2\dfrac{|T|}{|S|}Cov(Y_S, Y_T)\\
&=\dfrac{|T|^2}{|S|^2} {\mathbb E}(Y_S)+{\mathbb E}(Y_T)-2\dfrac{|T|}{|S|}Cov(Y_{S\setminus T}+Y_T, Y_T)\\
&=\dfrac{|T|^2}{|S|^2} {\mathbb E}(Y_S)+{\mathbb E}(Y_T)-2\dfrac{|T|}{|S|}Var(Y_T)\\
&=\dfrac{|T|^2}{|S|^2} (\alpha |S|+\beta x_S)+(\alpha |T|+\beta x_T)-2\dfrac{|T|}{|S|}(\alpha |T|+\beta x_T)\\
&=\beta x_S (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})
\end{align*}
\begin{align*}
Var(\hat{Y}^{DAX}_{T}-Y_T)&=Var(\dfrac{x_T}{x_S}Y_S-Y_T)\\
&=\dfrac{x_T^2}{x_S^2} Var(Y_S)+Var(Y_T)-2\dfrac{x_T}{x_S}Cov(Y_S, Y_T)\\
&=\dfrac{x_T^2}{x_S^2} {\mathbb E}(Y_S)+{\mathbb E}(Y_T)-2\dfrac{x_T}{x_S}Cov(Y_{S\setminus T}+Y_T, Y_T)\\
&=\dfrac{x_T^2}{x_S^2} {\mathbb E}(Y_S)+{\mathbb E}(Y_T)-2\dfrac{x_T}{x_S}Var(Y_T)\\
&=\dfrac{x_T^2}{x_S^2} (\alpha |S|+\beta x_S)+(\alpha |T|+\beta x_T)-2\dfrac{x_T}{x_S}(\alpha |T|+\beta x_T)\\
&=\alpha |S| (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})
\end{align*}
Summing up the variances at target level with the fact that $\sum_T {|T| \over |S|}=\sum_T {x_T\over x_S}=1$, we get the variances at source level
\begin{align*}
Var_S^{DAW}&=\sum_T Var(\hat{Y}^{DAW}_{T}-Y_T)\\
&=\sum_T \beta x_S (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_T(1-\dfrac{x_T}{x_S}) +\alpha |T| (1-\dfrac{|T|}{|S|})\\
&=\beta x_S \sum_T(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_S(1-\sum_T \dfrac{x_T^2}{x_S^2}) +\alpha |S| (1-\sum_T\dfrac{|T|^2}{|S|^2})
\end{align*}
\begin{align*}
Var_S^{DAX}&=\sum_T Var(\hat{Y}^{DAX}_{T}-Y_T)\\
&=\alpha |S| \sum_T(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+\beta x_S(1-\sum_T \dfrac{x_T^2}{x_S^2}) +\alpha |S| (1-\sum_T\dfrac{|T|^2}{|S|^2})
\end{align*}
\subsubsection{Proof of Theorem \ref{reler}}
From Lemma \ref{bias} and the fact that $\alpha |S|=I_S(|.|){\mathbb E}(Y_S), \beta X_S=I_S(X){\mathbb E}(Y_S)$ we have
\begin{align*}
{\text {Er}}_T^{DAW}&= I_S(X){\mathbb E}(Y_S) (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+I_S(X){\mathbb E}(Y_S)(\dfrac{x_T}{x_S} - \dfrac{x_T^2}{x_S^2}) +I_S(|.|){\mathbb E}(Y_S) (\dfrac{|T|}{|S|} -\dfrac{|T|^2}{|S|^2})\\
&+ I_S(X)^2{\mathbb E}(Y_S)^2 (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\\
{\text {Er}}_T^{DAX}&= I_S(|.|){\mathbb E}(Y_S) (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+I_S(X){\mathbb E}(Y_S)(\dfrac{x_T}{x_S} - \dfrac{x_T^2}{x_S^2}) +I_S(|.|){\mathbb E}(Y_S) (\dfrac{|T|}{|S|} -\dfrac{|T|^2}{|S|^2})\\
&+ I_S(|.|)^2{\mathbb E}(Y_S)^2 (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2
\end{align*}
If the expectation of the number of points is sufficiently large, we can approximate the ratio of the two errors as follows
$$\dfrac{{\text {Er}}_T^{DAW}}{{\text {Er}}_T^{DAX}}\approx\dfrac{I_S(X)^2}{I_S(|.|)^2}$$
and also
$$\dfrac{{\text {Re}}_T^{DAW}}{{\text {Re}}_T^{DAX}}\approx\dfrac{I_S(X)}{I_S(|.|)}$$
At source level, we get a similar result by adding up errors on all target zones using the fact that $\sum_{T}\dfrac{|T|}{|S|}=\sum_{T}\dfrac{x_T}{x_S}=1$
\begin{align*}
{\text {Er}}_S^{DAW}&= I_S(X){\mathbb E}(Y_S)\sum_{T} (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+I_S(X){\mathbb E}(Y_S)(1 -\sum_{T} \dfrac{x_T^2}{x_S^2}) +I_S(|.|){\mathbb E}(Y_S) (1 -\sum_{T}\dfrac{|T|^2}{|S|^2})\\
&+ I_S(X)^2{\mathbb E}(Y_S)^2\sum_{T} (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\\
{\text {Er}}_S^{DAX}&= I_S(|.|){\mathbb E}(Y_S) (1-\sum_{T}\dfrac{x_T}{x_S})^2+I_S(X){\mathbb E}(Y_S)(1 - \sum_{T}\dfrac{x_T^2}{x_S^2}) +I_S(|.|){\mathbb E}(Y_S) (1 -\sum_{T}\dfrac{|T|^2}{|S|^2})\\
&+ I_S(|.|)^2{\mathbb E}(Y_S)^2 \sum_{T}(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\\
&\Rightarrow\\
{\text {Re}}_S^{DAW}&= \dfrac{1}{{\mathbb E}(Y_S)} [I_S(X)\sum_{T} (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2+I_S(X)(1 -\sum_{T} \dfrac{x_T^2}{x_S^2}) +I_S(|.|) (1 -\sum_{T}\dfrac{|T|^2}{|S|^2})]\\
&+ I_S(X)^2\sum_{T} (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\\
{\text {Re}}_S^{DAX}&= \dfrac{1}{{\mathbb E}(Y_S)} [I_S(|.|) (1-\sum_{T}\dfrac{x_T}{x_S})^2+I_S(X)(1 - \sum_{T}\dfrac{x_T^2}{x_S^2}) +I_S(|.|) (1 -\sum_{T}\dfrac{|T|^2}{|S|^2})]\\
&+ I_S(|.|)^2 \sum_{T}(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2
\end{align*}
Using the relationship $I_S(|.|)+I_S(X)=1$, the above results prove Theorem \ref{reler}.
\subsubsection{Proof of Theorem \ref{TheoDiffErDaxDaw}}
Lemma \ref{bias} yields
\begin{align*}
{\text {Er}}^{DAW}_T&-{\text {Er}}^{DAX}_T=Var(\hat{Y}^{DAW}_{T}-Y_T)+[{\mathbb E}(\hat{Y}^{DAW}_{T}-Y_T)]^2 \\
&-Var(\hat{Y}^{DAX}_{T}-Y_T)-[{\mathbb E}(\hat{Y}^{DAX}_{T}-Y_T)]^2\\
&= (\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2(\beta x_S-\alpha |S|) +(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2(\beta^2 x_S^2-\alpha^2 |S|^2)\\
&=(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2(\beta x_S-\alpha |S|)((\beta x_S+\alpha |S|+1)\\
&=(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2\dfrac{(\beta x_S-\alpha |S|)}{(\beta x_S+\alpha |S|)}((\beta x_S+\alpha |S|+1)(\beta x_S+\alpha |S|)\\
&=(\dfrac{|T|}{|S|}-\dfrac{x_T}{x_S})^2 \Delta_S ({\mathbb E}(Y_S)+1){\mathbb E}(Y_S)
\end{align*}
\subsubsection{Proof of Theorem \ref{composite}}
We calculate the error of the composite predictors then minimize with respect to $w$ to find the optimal $w^*$
\begin{align*}
\hat{Y}^C_T&= w\hat{Y}^{DAW}_{T}+(1-w)\hat{Y}^{DAX}_{T}=[w \dfrac{|T|}{|S|}+ (1-w)\dfrac{x_T}{x_S}]Y_S := u Y_S\\
Bias_T^2&=[{\mathbb E}(\hat{Y}^{DAW}_{T}-\hat{Y}^{DAX}_{T})]^2=(u\lambda_S -\lambda_T)^2\\
Var_T&=Var(u Y_S-Y_T)=u^2 \lambda_S + \lambda_T -2 u\lambda_T\\
{\text {Er}}_T&=u^2\lambda_S(\lambda_S+1) -2u\lambda_T(\lambda_S+1)+\lambda_T^2+\lambda_T\\
u^*&=argmin_u {\text {Er}}_T=\dfrac{\lambda_T}{\lambda_S}\\
\Leftrightarrow & w^* \dfrac{|T|}{|S|}+ (1-w^*)\dfrac{x_T}{x_S}=\dfrac{\alpha |T|+\beta x_T}{\alpha |S|+\beta x_S}\\
\Leftrightarrow & w^*=\dfrac{\alpha |T|}{\alpha |S|+\beta x_S}
\end{align*}
Substituting the $w^*$ in \eqref{compositeclass} we get the composite predictor \eqref{Com}.
The bias, variance and error of the above composite predictor are calculated as follows
\begin{align*}
Bias={\mathbb E}(\hat{Y}^C_T-Y_T)&=0\\
{\text {Er}}^C_T=Var(\hat{Y}^C_T-Y_T)&=Var(\dfrac{\lambda_T}{\lambda_S}Y_S -Y_T)\\
&=\dfrac{\lambda_T^2}{\lambda_S^2}Var(Y_S) + Var(Y_T)-2\dfrac{\lambda_T}{\lambda_S} Cov(Y_S,Y_T)\\
&=\dfrac{\lambda_T^2}{\lambda_S^2} \lambda_S +\lambda_T-2 \dfrac{\lambda_T}{\lambda_S} \lambda_T\\
&=\lambda_T - \dfrac{\lambda_T^2}{\lambda_S}\\
&=\dfrac{x_T^2}{x_S^2} \lambda_S+\lambda_T-2\dfrac{x_T}{x_S}\lambda_T -\dfrac{x_T^2}{x_S^2} \lambda_S+2\dfrac{x_T}{x_S}\lambda_T- \dfrac{\lambda_T^2}{\lambda_S}\\
&=\dfrac{x_T^2}{x_S^2} \lambda_S+\lambda_T-2\dfrac{x_T}{x_S}\lambda_T -\lambda_S(\dfrac{x_T}{x_S}-\dfrac{\lambda_T}{\lambda_S})^2\\
&=Var(\hat{Y}^{DAX}_T-Y_T)-\lambda_S(\dfrac{x_T}{x_S}-\dfrac{\lambda_T}{\lambda_S})^2\\
&=Var(\hat{Y}^{DAW}_T-Y_T)-\lambda_S(\dfrac{|T|}{|S|}-\dfrac{\lambda_T}{\lambda_S})^2
\end{align*}
Since
\begin{align*}
Y_T|Y_S \sim Bi(Y_S, \dfrac{{\mathbb E}(Y_T)}{{\mathbb E}(Y_S)})
\end{align*}
we have
\begin{align*}
{\mathbb E}(Y_T|Y_S)= \dfrac{{\mathbb E}(Y_T)}{{\mathbb E}(Y_S)}Y_S=\hat{Y}^C_T
\end{align*}
This shows that the composite predictor is the best linear predictor.
\subsubsection{Proof of Theorem \ref{ConsitencyNormality}}
To prove the theorem, we will prove the following lemmas
\begin{lemma}\label{lemNormedScore}
Under conditions (C1) and (C2), the normed score function $F_n^{-1/2}s_n$ is asymptotically normal
\begin{equation}
F_n^{-1/2}s_n \to_d {\mathcal N}(0,{\mathbf I})
\end{equation}
\end{lemma}
\begin{lemma}\label{normed2derivative}
Under conditions (C1) and (C2), for all $\delta >0$
\begin{equation}\label{EqNormed2Deri}
max_{\gamma \in N_n(\delta)} ||V_n(\gamma)-{\mathbf I}|| \to_p 0
\end{equation}
where $N_n(\delta)=\{\gamma: ||F_n^{1/2} (\gamma - \gamma_o)||\leq \delta\}, V_n (\gamma)=F_n^{-1/2}H_n(\gamma)F_n^{-1/2}$.
\end{lemma}
Lemma \ref{lemNormedScore} is proved by using the Lindeberg-Feller theorem.
Indeed, for $\tau$ fixed with $\tau' \tau =1$, considering the triangular array
\begin{equation}
z_{n,i}=\tau' F_n^{-1/2} \dfrac{\tilde{Z}_{n,i}}{\gamma' \tilde{Z}_{n,i}}(y_{n,i}-\gamma' \tilde{Z}_{n,i})
\end{equation}
we have
\begin{align*}
{\mathbb E}(z_{n,i})&=0\\
\sum_i Var(z_{n,i})&=1
\end{align*}
We will show that the Lindeberg condition is satisfied, i.e. for any $\varepsilon >0$
\begin{equation}\label{LinCondition}
\sum_i {\mathbb E}(z_{n,i}^2 {\mathbf 1}_{|z_{n,i}|>\varepsilon}) \to 0
\end{equation}
as $n\to \infty$.
Let $a_{n,i}=\tau' F_n^{-1/2} \dfrac{\tilde{Z}_{n,i}}{\gamma' \tilde{Z}_{n,i}}$, because $z_{n,i}^2 = a_{n,i}^2 (y_{n,i}-\gamma' \tilde{Z}_{n,i})^2$ , ${\mathbb E}(z_{n,i}^2 {\mathbf 1}_{|z_{n,i}|>\varepsilon}) = a_{n,i}^2{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>\frac{\varepsilon}{|a_{n,i}|}})$ , yields
\begin{align*}
\sum_i {\mathbb E}(z_{n,i}^2 {\mathbf 1}_{|z_{n,i}|>\varepsilon}) &=\sum_i a_{n,i}^2{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>\frac{\varepsilon}{|a_{n,i}|}})\\
&\leq (\sum_i a_{n,i}^2 )sup_{i}{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>\frac{\varepsilon}{|a_{n,i}|}})
\end{align*}
Moreover, condition (C1) yields that there is a positive number $K_1$ s.t. ${1\over \gamma' \tilde{Z}_{n,i}}<K_1, \forall (n,i)$, hence
\begin{align*}
\sum_i a_{n,i}^2 &=\tau' F_n^{-1/2}\sum_i \dfrac{\tilde{Z}_{n,i}\tilde{Z}'_{n,i}}{(\gamma' \tilde{Z}_{n,i})^2} F_n^{-1/2}\tau\\
&<K_1 \tau' F_n^{-1/2}\sum_i \dfrac{\tilde{Z}_{n,i}\tilde{Z}'_{n,i}}{\gamma' \tilde{Z}_{n,i}} F_n^{-1/2}\tau=K_1
\end{align*}
In addition, conditions (C1) (C2) lead to
$$max_i \dfrac{\varepsilon}{|a_{n,i}|} \to \infty$$
as $n \to \infty$, hence for any $M>0, \exists n_1$ s.t. $\forall n>n_1$
\begin{align*}
sup_{i}&{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>\frac{\varepsilon}{|a_{n,i}|}})\leq sup_{i}{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>M})\\
&\leq sup_{i}\sqrt{{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^4{\mathbb E}({\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>M})}\\
&\leq sup_{i}\sqrt{{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^4\dfrac{Var(y_{n,i}-\gamma' \tilde{Z}_{n,i})}{M^2}}\\
&=sup_{i}\sqrt{\gamma' \tilde{Z}_{n,i}(1+3\gamma' \tilde{Z}_{n,i})\dfrac{\gamma' \tilde{Z}_{n,i}}{M^2}}\\
&< {K_2 \over M}
\end{align*}
hence
$$sup_{i}{\mathbb E}((y_{n,i}-\gamma' \tilde{Z}_{n,i})^2{\mathbf 1}_{|y_{n,i}-\gamma' \tilde{Z}_{n,i}|>\frac{\varepsilon}{|a_{n,i}|}}) \to 0 \text{ as } n \to \infty $$
where the existence of $K_2$ is derived from condition (C1). This argument shows that the \eqref{LinCondition} holds. So does Lemma \ref{lemNormedScore}.
\noindent {\it Proof of Lemma \ref{normed2derivative}}
Using the same notation in the proof of Lemma \ref{lemNormedScore}, $\tau$ fixed s.t. $\tau'\tau=1$, let $b_{n,i}=\tau' F_n^{-1/2} \tilde{Z}_{n,i}$, the equation \eqref{EqNormed2Deri} can be rewritten as
\begin{equation}
\tau'(V_n (\gamma)-{\mathbf I})\tau=A_n+B_n+C_n
\end{equation}
where
\begin{align}
A_n&=\sum_i b_{n,i}^2 ({1 \over (\gamma' \tilde{Z}_{n,i})^2}-{1 \over (\gamma'_o \tilde{Z}_{n,i})^2})(y_{n,i}-\gamma' \tilde{Z}_{n,i})\label{An}\\
B_n&=\sum_i b_{n,i}^2 {1 \over (\gamma'_o \tilde{Z}_{n,i})^2}(y_{n,i}-\gamma' \tilde{Z}_{n,i})\label{Bn}\\
C_n&=\sum_i b_{n,i}^2 ({1 \over \gamma' \tilde{Z}_{n,i}}-{1 \over \gamma'_o \tilde{Z}_{n,i}})\label{Cn}
\end{align}
We will prove that the three terms converge in probability to 0 as $n$ tends to $\infty$.
To prove \eqref{Bn}, we first study its properties. We have
\begin{align*}
{\mathbb E}(B_n)&=0\\
Var(B_n)&=\sum_i b_{n,i}^4 {1 \over (\gamma'_o \tilde{Z}_{n,i})^4}Var(y_{n,i}-\gamma' \tilde{Z}_{n,i})\\
&=\sum_i b_{n,i}^4 {1 \over (\gamma'_o \tilde{Z}_{n,i})^4}\gamma' \tilde{Z}_{n,i}\\
&\leq \sum_i b_{n,i}^2 {1 \over \gamma'_o \tilde{Z}_{n,i}}\sup_i {b_{n,i}^2 \over (\gamma'_o \tilde{Z}_{n,i})^3}\gamma' \tilde{Z}_{n,i}\\
&=\sup_i b_{n,i}^2 {1 \over (\gamma'_o \tilde{Z}_{n,i})^3}\gamma' \tilde{Z}_{n,i}<K_3 \sup_i b_{n,i}^2
\end{align*}
Because of the boundedness of $(\gamma'_o \tilde{Z}_{n,i})^3$ and the definition of $N_n(\delta)$, $\gamma' \tilde{Z}_{n,i}$ is bounded when $n$ is large enough, moreover, $\sup_i b_{n,i}^2 \to 0$ due to the condition (C1) (C2), therefore
$$B_n \to_p 0$$
We can use similar argument to prove $A_n \to_p 0, C_n \to 0$, and this shows that the lemma \ref{normed2derivative} holds.
\subsubsection{Proof of Theorem \ref{TheNormalityREG}}
Let $$z_{nij}\sim {\mathcal P}(\gamma_o'{Z}_{n,i})-\gamma_o'{Z}_{n,i}:=\tilde{z}_{n,i},\, j=1,2,...,k_n \text{ i.i.d }$$
This yields
$$\sum_j z_{nij} = Y_{n,i}-\gamma_o'\tilde{Z}_{n,i}$$
We have
\begin{align*}
& {\mathbb E}(z_{nij})=0\\
& \sum_j Var(z_{nij})=\gamma_o'\tilde{Z}_{n,i}
\end{align*}
We will prove that this array satisfies the Lindeberg-Feller condition, i.e. $\forall \delta >0$
$$\sum_j {\mathbb E}(z_{nij}^2 {\mathbf 1}_{|z_{nij}|> \delta})\to 0, \text{ as } n \to \infty $$
Indeed,
\begin{align*}
\sum_j {\mathbb E}(z_{nij}^2 {\mathbf 1}_{|z_{nij}|> \delta})=&k_n {\mathbb E}(\tilde{z}_{n,i}^2 {\mathbf 1}_{|\tilde{z}_{n,i}|> \delta})={\mathbb E}(u_{n,i}^2 {\mathbf 1}_{|u_{n,i}|> \sqrt{k_n} \delta})
\end{align*}
where $u_{n,i}=\sqrt{k_n} \tilde{z}_{n,i}$. Because ${\mathbb E} u_{n,i}=0$, ${\mathbb E} u_{n,i}^2=Var u_{n,i}=k_n Var \tilde{z}_{n,i}=\gamma_o'\tilde{Z}_{n,i} < \infty$. Moreover $k_n \to \infty$ as $n \to \infty$, we have
$${\mathbb E}(u_{n,i}^2 {\mathbf 1}_{|u_{n,i}|> \sqrt{k_n} \delta}) \to 0$$
as $n \to \infty$
From the Lindeberg-Feller theorem we get
$${Y_{n,i} - \gamma_o'\tilde{Z}_{n,i} \over \sqrt{\gamma_o'\tilde{Z}_{n,i}}} \to_d {\mathcal N}(0, 1)$$
This proof can be applied at the target level, i.e.
$${Y_{T} - \gamma_o' \tilde{Z}_{T} \over \sqrt{\gamma_o'\tilde{Z}_{T}}} \to_d {\mathcal N}(0, 1)$$
\subsubsection{Proof of Proposition \ref{ProPycno}}
The pycnophylactic property of the scaled regression predictor is obvious.
To prove the pycnophylactic property of the regression predictor at region level, we sum up regression predictors over source zones
\begin{align*}
\hat{Y}_{\Omega}=\sum_i\sum_{T: T \subset S_{n,i}}\hat{Y}^{REG}_{T}&=\sum_i\sum_{T: T \subset S_{n,i}} \hat{\gamma}_n\tilde{Z}_{T}=\hat{\gamma}_n'\tilde{Z}_\Omega
\end{align*}
Recall that $\hat{\gamma}$ is the solution of the score equation $s_n(\gamma)=0$, i.e.
\begin{align*}
&\sum_{i=1}^n \dfrac{\tilde{Z}_{n,i}}{\hat{\gamma}' \tilde{Z}_{n,i}} y_{n,i} - \tilde{Z}_{n,i}=0\\
\Rightarrow &\sum_{i=1}^n \dfrac{\hat{\gamma}'\tilde{Z}_{n,i}}{\hat{\gamma}' \tilde{Z}_{n,i}} y_{n,i} - \hat{\gamma}'\tilde{Z}_{n,i}=0\\
\Leftrightarrow &\sum_{i=1}^n y_{n,i} - \hat{\gamma}'\tilde{Z}_{\Omega}=0\\
\Leftrightarrow &\hat{\gamma}'\tilde{Z}_{\Omega}=y_{\Omega}
\end{align*}
In other words, the regression predictor satisfies the pycnophylactic property on the region $\Omega$.
To study the pycnophylactic property of the regression predictor at source level, we consider
$$\hat{Y}_{n,i}^{REG}-Y_{n,i}$$
We have
\begin{align*}
\hat{Y}_{n,i}^{REG}-Y_{n,i}&=\hat{\gamma}_n'\tilde{Z}_{n,i}-Y_{n,i}\\
&=(\hat{\gamma}_n'-\gamma_o' )\tilde{Z}_{n,i}-(Y_{n,i}-\gamma_o' \tilde{Z}_{n,i})\\
&=F_n^{-1/2}F_n^{1/2}(\hat{\gamma}_n'-\gamma_o' )\tilde{Z}_{n,i}-(Y_{n,i}-\gamma_o'\tilde{Z}_{n,i})\\
\end{align*}
The first term converges to 0 in distribution due to the conditions (C1), (C2) and the theorem \ref{ConsitencyNormality}. The second term is different from 0, even asymptotically (Proposition \ref{ProPycno}).
Moreover, because of the boundedness of $\tilde{Z}_{n,i}$, the above argument yields
$$\dfrac{\hat{Y}_{n,i}^{REG}-Y_{n,i}}{\sqrt{\gamma_o'\tilde{Z}_{n,i}}} \to_d {\mathcal N}(0,1)$$
This completes the proof of proposition \ref{ProPycno}.
If $\tilde{Z}_{T}$ is bounded below, a similar result at target level holds
$$\dfrac{\hat{Y}_{T}^{REG}-Y_{T}}{\sqrt{\gamma_o'\tilde{Z}_{T}}} \to_d {\mathcal N}(0,1)$$
\subsubsection{Proof of Theorem \ref{TheErrReg}}
For any target $T$, the error of the regression predictor on the target is
\begin{align*}
{\mathbb E}(\hat{Y}^{REG}_T-Y_T)^2&={\mathbb E}( \hat{\gamma}_n' \tilde{Z}_T-{\gamma}_o'\tilde{Z}_T )^2+{\mathbb E}({\gamma}_o'\tilde{Z}_T -Y_T)^2-2{\mathbb E}( \hat{\gamma}_n' \tilde{Z}_T-{\gamma}_o'\tilde{Z}_T )({\gamma}_o'\tilde{Z}_T -Y_T)
\end{align*}
From Theorem \ref{ConsitencyNormality} and condition (C1), for any $\eta_1>0, \exists \varepsilon >0$ s.t.when $n$ is sufficiently large
\begin{align}
{\mathbb E}(\hat{\gamma}_n'\tilde{Z}_T -{\gamma}_o'\tilde{Z}_T )^2 {\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}&<\eta_1\label{eq5.1.1}\\
||2{\mathbb E}( \hat{\gamma}_n' \tilde{Z}_T-{\gamma}_o'\tilde{Z}_T )({\gamma}_o'\tilde{Z}_T -Y_T){\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}||&<\eta_1\label{eq5.1.2}
\end{align}
As we proved in Proposition \ref{ProPycno} $( \hat{\gamma}_n' \tilde{Z}_T-{\gamma}_o'\tilde{Z}_T ) \to_p 0$, we have
\begin{align*}
{\mathbb P}(||\hat{\gamma}_n-{\gamma}_o||< \varepsilon)&\to 1 \text{ as } n \to \infty
\end{align*}
In addition
$${\mathbb E}({\gamma}_o'\tilde{Z}_T -Y_T)^2={\gamma}_o'\tilde{Z}_T $$
Hence there is $n_1$ s.t.
$${\mathbb E}({\gamma}_o'\tilde{Z}_T -Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||\geq \varepsilon}<
{\mathbb E}({\gamma}_o'\tilde{Z}_T -Y_T)^4 {\mathbb P}(||\hat{\gamma}_n-{\gamma}_o||\geq \varepsilon)<\eta_1$$
for $n>n_1$. In other words,
$${\gamma}_o'\tilde{Z}_T -\eta_1<{\mathbb E}({\gamma}_o'\tilde{Z}_T -Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||> \varepsilon}<{\gamma}_o'\tilde{Z}_T $$
This implies $\forall \eta >0, \exists \varepsilon >0, n_1$ s.t. for $n>n_1$
\begin{align}
-\eta +{\gamma}_o'\tilde{Z}_T <{\mathbb E}(\hat{Y}^R_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}<\eta +{\gamma}_o'\tilde{Z}_T \label{eq5.1.3}
\end{align}
with a remark that ${\mathbb P}(||\hat{\gamma}_n-{\gamma}_o||< \varepsilon)\to 1 \text{ as } n \to \infty$.
Combining \eqref{eq5.1.1}, \eqref{eq5.1.2}, \eqref{eq5.1.3} we get Theorem \ref{TheErrReg}.
\subsubsection{Proof of equations (\ref{diffapproRelEr})}
We rewrite the error of the areal interpolation and dasymetric for the asymptotic model. For a target $T\subset S_{n,i}$, from \eqref{ErComDAX}, \eqref{ErComDAW}, and Lemma \ref{bias} we have
\begin{align*}
{\text {Er}}^{DAW}_T &= \beta^2 \tilde{x}_{n,i}
\gamma_o'\tilde{Z}_T-\dfrac{(\gamma_o'\tilde{Z}_T)^2}{\gamma_o'\tilde{Z}_{n,i}}+\gamma_o'\tilde{Z}_{n,i}(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\gamma_o'\tilde{Z}_T}{\gamma_o'\tilde{Z}_{n,i}})^2+\beta^2 \tilde{x}_S^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2\\
{\text {Er}}^{DAX}_T &=\gamma_o'\tilde{Z}_T-\dfrac{(\gamma_o'\tilde{Z}_T)^2}{\gamma_o'\tilde{Z}_{n,i}}+\gamma_o'\tilde{Z}_{n,i}(\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}}-\dfrac{\gamma_o'\tilde{Z}_T}{\gamma_o'\tilde{Z}_{n,i}})^2+\alpha^2 \widetilde{|S_{n,i}|}^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2
\end{align*}
A similar argument as in the proof of theorem \ref{TheErrReg} shows that, for any $\eta_1>0, \varepsilon>0$, $\exists n_1$ s.t. $\forall n>n_1$ \begin{align*}
{\text {Er}}^{DAW}_T-\eta_1&<{\mathbb E}(\hat{Y}_{T}^{DAW}-Y_{T})^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||<\varepsilon}<{\text {Er}}^{DAW}_T\\
{\text {Er}}^{DAX}_T-\eta_1&<{\mathbb E}(\hat{Y}_{T}^{DAX}-Y_{T})^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||<\varepsilon}<{\text {Er}}^{DAX}_T
\end{align*}
With $\varepsilon$ chosen as in theorem \ref{TheErrReg}, let $Q_i=\{||\hat{\gamma}_n-{\gamma}_o||<\varepsilon\}$, we have
\begin{align*}
{\gamma}_o'\tilde{Z}_T-{\text {Er}}^{DAW}_T -\eta<{\mathbb E}(\hat{Y}_{T}^{REG}-Y_{T})^2{\mathbf 1}_{Q_i}-{\mathbb E}(\hat{Y}_{T}^{DAW}-Y_{T})^2{\mathbf 1}_{Q_i} & < {\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAW}_T+\eta+\eta_1\\
{\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAX}_T-\eta<{\mathbb E}(\hat{Y}_{T}^{REG}-Y_{T})^2{\mathbf 1}_{Q_i}-{\mathbb E}(\hat{Y}_{T}^{DAX}-Y_{T})^2{\mathbf 1}_{Q_i} & < {\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAX}_T+\eta+\eta_1
\end{align*}
for all $n>n_1$. Moreover,
\begin{align*}
{\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAW}_T
&=\beta \tilde{x}_{n,i} {\tilde{x}_T^2 \over \tilde{x}_{n,i}^2} +\alpha \widetilde{|S_{n,i}|} {\widetilde{|T|}^2 \over \widetilde{|S_{n,i}|}^2}-\beta \tilde{x}_{n,i}(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2-\beta^2 \tilde{x}_{n,i}^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2\\
{\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAW}_
&=\beta \tilde{x}_{n,i} {\tilde{x}_T^2 \over \tilde{x}_{n,i}^2} +\alpha \widetilde{|S_{n,i}|} {\widetilde{|T|}^2 \over \widetilde{|S_{n,i}|}^2}-\alpha \widetilde{|S_{n,i}|}(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2-\alpha^2 \widetilde{|S_{n,i}|}^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2
\end{align*}
Taking the sum over all target zones which belong to $S_{n,i}$ then scaling the sum by ${\mathbb E}(Y_{n,i})$ and calculating the differences in terms of $\Delta_{n,i}=\Delta_{S_{n,i}}$, we have
\begin{align*}
4{\sum_T {\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAW}_T \over {\mathbb E}(Y_{n,i})^2}&=4{1 \over {\mathbb E}(Y_{n,i})}[{\beta \tilde{x}_{n,i} \over {\mathbb E}(Y_{n,i})}\sum_T {\tilde{x}_T^2 \over \tilde{x}_{n,i}^2} +{\alpha \widetilde{|S_{n,i}|}\over {\mathbb E}(Y_{n,i})}\sum_T {\widetilde{|T|}^2 \over \widetilde{|S_{n,i}|}^2}-{\beta \tilde{x}_{n,i} \over {\mathbb E}(Y_{n,i})}\sum_T(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2]\\
&-4({\beta\tilde{x}_{n,i}\over {\mathbb E}(Y_{n,i})})^2 \sum_T (\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2\\
&=2{1 \over {\mathbb E}(Y_{n,i})}[(1 + \Delta_{n,i})\sum_T {{x}_T^2 \over {x}_S^2} +(1 - \Delta_{n,i})\sum_T {|T|^2 \over |S_{n,i}|^2}-\\
&-(1 + \Delta_{n,i})\sum_T(\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2]-4(1 + \Delta_{n,i})^2 \sum_T (\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2
\end{align*}
\begin{align*}
4{\sum_T {\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAX}_T \over {\mathbb E}(Y_{n,i})^2}&=4{1 \over {\mathbb E}(Y_{n,i})}[{\beta \tilde{x}_{n,i} \over {\mathbb E}(Y_{n,i})}\sum_T {\tilde{x}_T^2 \over \tilde{x}_{n,i}^2} +{\alpha \widetilde{|S_{n,i}|}\over {\mathbb E}(Y_{n,i})}\sum_T {\widetilde{|T|}^2 \over \widetilde{|S_{n,i}|}^2}-{\alpha \widetilde{|S_{n,i}|}\over {\mathbb E}(Y_{n,i})}\sum_T(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2]\\
&-4({\alpha \widetilde{|S_{n,i}|}\over {\mathbb E}(Y_{n,i})})^2 \sum_T (\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2\\
&=2{1 \over {\mathbb E}(Y_{n,i})}[(1 + \Delta_{n,i})\sum_T {{x}_T^2 \over {x}_{n,i}^2} +(1 - \Delta_{n,i})\sum_T {|T|^2 \over |S_{n,i}|^2}-\\
&-(1 - \Delta_{n,i})\sum_T(\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2]-4(1 - \Delta_{n,i})^2 \sum_T (\dfrac{{|T|}}{{|S_{n,i}|}}-\dfrac{{x}_T}{{x}_{n,i}})^2
\end{align*}
If $(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})=0$ then the regression is less accurate than areal weighting and dasymetric asymptotically. If this difference increases, the difference between the regression and the other two methods gets smaller and then the regression method can do better than the other two methods. Indeed, for example when $\dfrac{x_T}{|T|}=\dfrac{\beta x_{S_{n,i}}-\alpha|S_{n,i}|}{2 \beta |S|}$, this yields $T, x_T$ satisfy $(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}}) \not= 0$, we have
$$\dfrac{(\gamma_o'\tilde{Z}_T)^2}{\gamma_o'\tilde{Z}_{n,i}}-{1 \over \gamma_o'\tilde{Z}_{n,i}}\beta^2 \tilde{x}_{n,i}^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2=0$$
therefore,
$${\gamma}_o'\tilde{Z}_T -{\text {Er}}^{DAW}_T=-\beta^2 \tilde{x}_{n,i}^2(\dfrac{\widetilde{|T|}}{\widetilde{|S_{n,i}|}}-\dfrac{\tilde{x}_T}{\tilde{x}_{n,i}})^2<0$$
Choosing $\eta, \eta_1$ to be sufficient small, the regression predictor is asymptotically better than the areal weighting interpolation predictor. A similar result for the case of the dasymetric predictor can be proved similarly.
We therefore proved that none of the considered three methods is always dominant.
\subsubsection{Proof of Lemma \ref{LeScRCom}}
Assume $T \in S_{n,i}$, the difference between the predictors of scaled regression and composite predictor is given by
\begin{align*}
\hat{Y}^{ScR}_T-\hat{Y}^{C}_T&=\dfrac{\hat{\gamma}_n'\tilde{Z}_T }{\hat{\gamma}_n'\tilde{Z}_{n,i} }Y_{n,i}-\dfrac{{\gamma}_o'\tilde{Z}_T }{\tilde{Z}_{n,i} {\gamma}_o}Y_{n,i}\\
&=(\hat{\gamma}_n'-{\gamma}_o')\dfrac{\tilde{\gamma}_n' \tilde{Z}_{n,i} \tilde{Z}_T -\tilde{\gamma}_n' \tilde{Z}_T \tilde{Z}_{n,i} }{(\tilde{Z}_{n,i} \tilde{\gamma}_n)^2}Y_{n,i}\\
&=(\hat{\gamma}_n'-{\gamma}_o')F_n^{-T/2}F_n^{T/2}Y_{n,i}\dfrac{\tilde{\gamma}_n' \tilde{Z}_{n,i} \tilde{Z}_T -\tilde{\gamma}_n' \tilde{Z}_T \tilde{Z}_{n,i} }{(\tilde{Z}_{n,i} \tilde{\gamma}_n)^2}
\end{align*}
where $\tilde{\gamma}_n $ belongs to the segment of $\hat{\gamma}_n$ and ${\gamma}_o$.
From Theorem \ref{ConsitencyNormality}, property \eqref{app}, conditions (C1), (C2), we have
\begin{align*}
F_n^{T/2}(\hat{\gamma}_n-{\gamma}_o) &\to_d {\mathcal N}(0,I)\\
F_n^{-T/2}\dfrac{Y_{n,i}-\tilde{Z}_{S_{n,i}} {\gamma}_o }{\sqrt{\tilde{Z}_{S_{n,i}} {\gamma}_o }}&\to_d 0\\
\dfrac{\tilde{\gamma}_n' \tilde{Z}_{n,i} \tilde{Z}_T -\tilde{\gamma}_n' \tilde{Z}_T \tilde{Z}_{n,i} }{(\tilde{Z}_{n,i} \tilde{\gamma}_n)^2}& \text{ bounded,}
\end{align*}
In other words,
$$\hat{Y}^{ScR}_T-\hat{Y}^{C}_T \to_p 0$$
\subsubsection{Proof of Theorem \ref{TheScR}}
Because of the boundedness of $\tilde{Z}_{n,i}$, upper boundedness of $\tilde{Z}_T$, there exists
$$M=sup_{ \tilde{Z}_T, \tilde{Z}_{n,i}, {\gamma} \in B(\gamma_o, 1)}||\dfrac{\tilde{Z}_{n,i}\gamma_n \tilde{Z}_T -\tilde{Z}_T \gamma_n \tilde{Z}_{n,i} }{(\tilde{Z}_{n,i} \gamma_n)^2}||{\mathbb E}(Y_{n,i}^2)$$
where $B(\gamma_o, 1)=\{\gamma: ||\gamma_o - \gamma ||< 1\}$. Since $\hat{\gamma}_n-{\gamma}_o \to_p 0$, the sequence $\hat{\gamma}_n, n=1,2,...$ is bounded, therefore for any $\varepsilon>0$, when $n$ is large enough
$$sup_{ \tilde{Z}_T, \tilde{Z}_{n,i}, \tilde{\gamma} \in segment(\gamma_o, \hat{\gamma}_n)}||\dfrac{\tilde{Z}_{n,i}\tilde{\gamma}_n \tilde{Z}_T -\tilde{Z}_T \tilde{\gamma}_n \tilde{Z}_{n,i}}{(\tilde{Z}_{n,i} \tilde{\gamma}_n)^2}||{\mathbb E}(Y_{n,i}^2){\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}<M$$
For any $\eta >0$, there is an $\varepsilon >0$ s.t.
\begin{align*}
{\mathbb E}(\hat{Y}^{ScR}_T-\hat{Y}^{C}_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}&={\mathbb E}(||\dfrac{\tilde{Z}_{n,i}\tilde{\gamma}_n \tilde{Z}_T -\tilde{Z}_T \tilde{\gamma}_n \tilde{Z}_{n,i}}{(\tilde{Z}_{n,i} \tilde{\gamma}_n)^2}||^2||Y_{n,i}||^2||(\hat{\gamma}_n-{\gamma}_o)||^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon})<M^2 \varepsilon^2 < \eta
\end{align*}
Evaluating the error on the set $\{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon\}$, we have
\begin{align*}
{\mathbb E}(\hat{Y}^{ScR}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}&={\mathbb E}(\hat{Y}^{ScR}_T-\hat{Y}^{C}_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}+{\mathbb E}(\hat{Y}^{C}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}\\
&-2{\mathbb E}(\hat{Y}^{ScR}_T-\hat{Y}^{C}_T)(\hat{Y}^{C}_T-Y_T){\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}
\end{align*}
Moreover
\begin{align*}
{\mathbb E}(\hat{Y}^{C}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}&\leq {\mathbb E}(\hat{Y}^{C}_T-Y_T)^2=Var(\hat{Y}^{C}_T-Y_T)=\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}
\end{align*}
With the same argument as in theorem \ref{TheErrReg}, when $n$ is large enough
\begin{align*}
{\mathbb E}(\hat{Y}^{C}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}&= {\mathbb E}(\hat{Y}^{C}_T-Y_T)^2-{\mathbb E}(\hat{Y}^{C}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||\geq \varepsilon}\\
&>\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o} - \eta
\end{align*}
Using a similar argument as above, we can prove $\forall \eta >0, \exists \varepsilon >0$ and $n$ large enough such that
$$||{\mathbb E}(\hat{Y}^{ScR}_T-\hat{Y}^{C}_T)(\hat{Y}^{C}_T-Y_T){\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}||<\eta $$
In other words
$$-3\eta +\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}<{\mathbb E}(\hat{Y}^{ScR}_T-Y_T)^2{\mathbf 1}_{||\hat{\gamma}_n-{\gamma}_o||< \varepsilon}< 2\eta +\tilde{Z_T}\gamma_o - {(\tilde{Z}_T \gamma_o)^2 \over \tilde{Z}_{n,i}\gamma_o}
$$
Note that ${\mathbb P}(||\hat{\gamma}_n-{\gamma}_o||< \varepsilon) \to 1$ as $n \to \infty$ and the theorem holds.
|
\section{Introduction}
\setcounter{equation}{0}
Let $\Omega$ be a bounded domain in $\mathbb{R}^N$
(with $N=2$ or $N=3$), and let $S_h$ be a finite element subspace of
$H^1_0(\Omega)$ consisting of continuous piecewise polynomials of
degree $r\geq 1$ subject to certain quasi-uniform triangulation of
the domain $\Omega$.
We consider the parabolic equation
\begin{align}\label{PDE0}
\left\{
\begin{array}{ll}
\partial_tu- \nabla \cdot (a \nabla u) =f
&\mbox{in}~~\Omega\times(0,\infty),\\
\displaystyle u=0
&\mbox{on}~~\partial\Omega\times(0,\infty),\\
u(\cdot,0)=u^0 &\mbox{in}~~\Omega,
\end{array}
\right.
\end{align}
and its finite element approximation
\begin{align}\label{FEEq0}
\left\{
\begin{array}{ll}
\big(\partial_tu_h,v_h\big)
+ (a\nabla u_h, \nabla v_h)
=(f,v_h) , ~~\forall~v_h\in S_h,
\\
u_h(0)=u^0_h ,
\end{array}
\right.
\end{align}
where $f$ is a given function,
and $a=(a_{ij}(x))_{N\times N}$ is an $N \times N$ symmetric matrix which
satisfies the ellipticity condition
\begin{align}\label{coeffcond}
\Lambda^{-1}|\xi|^2\leq
\sum_{i,j=1}^Na_{ij}(x)\xi_i\xi_j
\leq \Lambda|\xi|^2
,\quad\mbox{for}~~x\in\Omega,
\end{align}
for some positive constant $\Lambda$.
If we define the elliptic operator $A: H^1_0(\Omega)\rightarrow H^{-1}(\Omega)$
and its finite element approximation $A_h: S_h\rightarrow S_h$ by
\begin{align}
& (Aw,v):= (a \nabla w, \nabla v),
\qquad\qquad\,\, \forall~w,v\in
H^1_0(\Omega),
\label{A-e}
\\
&(A_hw_h,v_h):= (a\nabla w_h, \nabla v_h)
,\qquad \forall~w_h,v_h\in S_h ,
\label{Ah-e}
\end{align}
then the solutions of \refe{PDE0} and \refe{FEEq0} can be expressed by
\begin{align}
&u(t)=E (t)u^0 +\int_0^t E (t-s)f(s){\rm d} s ,\\
&u_h(t)=E_h(t)u^0_h+\int_0^t E_h(t-s)f(s){\rm d} s ,
\end{align}
where $\{E(t)=e^{-tA}\}_{t>0}$ and
$\{E_h(t)=e^{-tA_h}\}_{t>0}$ denote the semigroups generated by the
operators $-A$ and $-A_h$, respectively.
By the theory of parabolic equations and \cite{Ouhabaz},
it is well known that
$\{E(t)\}_{t>0}$ is an analytic semigroup on
$C_0(\overline\Omega)$ satisfying
\begin{align}
&\|E(t)v\|_{L^\infty}
+t\|\partial_tE(t)v\|_{L^\infty}\leq C\|v\|_{L^\infty}
,\quad\forall\, v\in C_0(\overline\Omega), ~ \forall\, t>0 ,
\end{align}
which is equivalent to the resolvent estimate
\begin{align*}
&\|(\lambda+A)^{-1}v\|_{L^\infty} \leq C\lambda^{-1}\|v\|_{L^\infty}
,\quad\forall\, v\in C_0(\overline\Omega) ,
\,\, \forall\, \lambda\in\Sigma_{\theta+\pi/2} ,
\end{align*}
where $\Sigma_{\theta+\pi/2}:=\{z\in \mathbb{C} : |{\rm arg}(z)|<\theta+\pi/2\}$.
The counterparts of these two inequalities above
for the discrete finite element
operator $A_h$ are the analyticity of the semigroup $\{E_h(t)\}_{t>0}$
on $L^\infty\cap S_h$:
\begin{align}
\|E_h(t)v_h\|_{L^\infty} +t\|\partial_tE_h(t)v_h\|_{L^\infty}\leq
C\|v_h\|_{L^\infty} ,\quad\forall\, v_h\in S_h,\, \forall\, t>0 ,
\label{STLEst}
\end{align}
and the resolvent estimate
\begin{align*}
\|(\lambda+A_h)^{-1}v_h\|_{L^\infty} \leq C\lambda^{-1}\|v_h\|_{L^\infty}
,\quad\forall\, v_h\in S_h , \,\, \forall\, \lambda\in\Sigma_{\varphi+\pi/2} .
\end{align*}
The estimates of the discrete semigroup
have attracted much attention in the past
several decades. With these estimates,
one may reach more precise analyses
of finite element solutions, such as
maximum-norm analysis of FEMs
\cite{Lucas,Tho, TW1,Wah}, error estimates of fully discrete FEMs
\cite{LN, Pal,Tho} and the discrete maximal $L^p$ regularity for
parabolic finite element equations \cite{Gei1,Gei2,KLL,Li,LS2}.
The proof of \refe{STLEst} dates back to Schatz et. al. \cite{STW1},
who proved \refe{STLEst} with a logarithmic factor
for the heat equation in
a two-dimensional smooth convex
domain with the linear finite element method.
The logarithmic factor was removed in the
case $r\geq 4$ for $N=1,2,3$ in \cite{NW},
and the analysis was further extended to the case
$1\leq N\leq 5$ in \cite{Chen}.
Later, a unified approach was presented in
\cite{STW98} by Schatz et. al., where they proved \refe{STLEst}
with the Neumann boundary condition for all
$r\geq 1$ and $N\geq 1$. The result was extended to
the Dirichlet boundary condition in \cite{TW2}
for the linear finite element method.
Some other maximum-norm error estimates can be found in
\cite{DLSW,DM,ELWZ,Han,Ley,Lin1}, and the resolvent estimates
can be found in \cite{Bak, BTW}.
A related topic is the discrete maximal $L^p$ regularity
(when $u^0=0$ and $1<p,q<\infty$)
\begin{align}
&\|\partial_tu_h \|_{L^p((0,T);L^q)}+\|A_hu_h \|_{L^p((0,T);L^q)}
\leq C_{p,q}\|f\|_{L^p((0,T);L^q)} ,
\label{LpqSt3}
\end{align}
which resembles the maximal $L^p$ regularity
of the continuous parabolic problem and was
proved by Geissert \cite{Gei1,Gei2}.
A straightforward
application of \refe{LpqSt3} is the $L^p$-norm error estimate
\begin{align}
\| P_h u - u_h \|_{L^p((0,T);L^q)}
\le C_{p,q}(\| P_h u^0 - u_h^0 \|_{L^q}
+ \| P_h u - R_h u \|_{L^p((0,T);L^q)} ) ,
\label{error}
\end{align}
where $R_h$ is the Ritz projection
associated with the operator $A$
and $P_h$ is the $L^2$ projection
onto the finite element space.
All these estimates
were established under the assumption
that the coefficients $a_{ij}$ and the domain
$\Omega$ are smooth enough
so that the parabolic Green's function satisfies
\begin{align}\label{g-cond000}
|\partial_t^{\gamma}\partial_x^\beta G(t, x, y)|\leq
C(t^{1/2}+|x-y|)^{-(N+2\gamma+|\beta|)} e^{-\frac{|x-y|^2}{Ct}}
,~~\forall~0\leq \gamma\leq 2,\, 0\leq |\beta|\leq 2 \, .
\end{align}
Although the condition on the coefficients was relaxed to
$a_{ij} \in C^{2+\alpha}(\overline\Omega)$
in \cite{Gei1}, this assumption
is still too strong for many physical applications.
One of the examples is an
incompressible miscible flow in porous media \cite{Dou,LS1},
where the diffusion-dispersion tensor $[a_{ij}]_{i,j=1}^N$
is only a Lipschitz continuous function of the velocity field.
In a recent work \cite{Li}, the first author proved
\refe{STLEst} in a smooth domain
under the assumption $a_{ij} \in W^{1,\infty}(\Omega)$,
together with the estimate (when $u^0=0$
and $1<p,q<\infty$)
\begin{align}
\|u_h \|_{L^p((0,T);W^{1,q})}
\leq C_{p,q}\|f\|_{L^p((0,T);W^{-1,q})} ,
\label{LpqSt2}
\end{align}
which were then applied to the incompressible
miscible flow in porous media \cite{LS2}.
Moreover, the problem in a polygon or a
polyhedron is of high interest
in practical cases, while the
inequality \refe{g-cond000}
does not hold in arbitrary convex polygons or polyhedra,
and all the analyses of
{\rm\refe{LpqSt3}}-{\rm(\ref{LpqSt2})} are limited
to smooth domains so far.
For the problem in two-dimensional
polygons with constant coefficients,
the inequality (\ref{STLEst}) with an extra logarithmic factor
was proved in \cite{CLTW,Ran,Tho}
by using the following
estimate of the discrete Green's function $\Gamma_h$:
$$
\int_{\Omega} | \Gamma_h(t,x,x_0) | dx \le C |\ln h|
\, .
$$
The corresponding results in three-dimensional polyhedra are unknown.
More interested is whether these stability estimates hold with the natural regularity $a_{ij} \in W^{1,p}(\Omega)$ for some $1<p<\infty$, since
such estimates are important for the extension of the analysis
to a general nonlinear model.
This paper focuses on \refe{STLEst}-\refe{LpqSt3} and
\refe{LpqSt2} in a convex polygon or polyhedron
with a weaker regularity of the diffusion coefficient.
Instead of estimating $\Gamma_h$ directly,
we present a more precise estimate for the error function
$F:=\Gamma_h-\Gamma$ (see Lemma 2.2) with which the
logarithmic factor can be removed (this idea was used in \cite{STW98}),
where $\Gamma$ is a regularized Green's function.
To compensate the lack of pointwise estimate of
the second-order derivatives of the Green's function,
we use local $W^{1,\infty}$ estimate
and local energy estimates of the
second-order derivatives (see Lemma 4.1).
Our main result is the following theorem.
\begin{theorem}\label{MainTHM1}
{\it Assume that $a_{ij}\in W^{1,N+\alpha}(\Omega)$ for some $\alpha>0$,
satisfying the condition {\rm (\ref{coeffcond})}, and assume that
$\Omega$ is either a
convex polygon in $\mathbb{R}^2$ or a convex polyhedron in $\mathbb{R}^3$.
Then
(1) the semigroup estimate {\rm\refe{STLEst}} holds,
(2) the solution of {\rm(\ref{FEEq0})}
satisfies {\rm(\ref{LpqSt3})}
when $f\in L^p((0,T);L^q)$ and $u^0=0$,
(3) the solution of {\rm(\ref{FEEq0})}
satisfies {\rm(\ref{LpqSt2})}
when $f\in L^p((0,T);W^{-1,q})$ and $u^0=0$.
}
\end{theorem}
Under the assumptions in Theorem \ref{MainTHM1}
and assuming that the solution of \refe{PDE0} satisfies
$u \in C(\overline\Omega\times[0,T])$,
\refe{error} follows immediately from {\rm(\ref{LpqSt3})}.
The rest of this paper is organized as follows. In section 2,
we introduce some notations and present a key lemma based on which
our main theorem can be proved. In section 3, we present superapproximation
results for smoothly truncated finite element functions and present several
estimates for the parabolic Green's functions under the assumed
regularity of the coefficients and the domain. Based on these estimates,
we prove our key lemma in section 4.
\section{Notations, assumptions and sketch of the proof}
\setcounter{equation}{0}
\subsection{Notations}
For any nonnegative integer $k $ and $1\leq p\leq\infty$,
we let $W^{k,p}(\Omega)$ be the conventional Sobolev space of functions
defined in $\Omega$, and let $W^{1,p}_0(\Omega)$ be the subspace of
$W^{1,p}(\Omega)$ consisting of functions whose traces vanish on $\partial\Omega$.
As conventions, we denote the dual space of $W^{1,p}_0(\Omega)$
by $W^{-1,p'}(\Omega)$ for $1\leq p<\infty$, and denote
$H^k(\Omega):=W^{k,2}(\Omega)$ and $L^p(\Omega):=W^{0,p}(\Omega)$ for
any integer $k$ and $1\leq p\leq \infty$.
Let $ Q_T:=\Omega\times(0,T)$.
For any Banach space $X$ and a given $T>0$,
we let $L^p((0,T);X)$ be the Bochner spaces equipped with the norm
\begin{align*}
&\|f\|_{L^p((0,T);X)} =\left\{
\begin{array}{ll}
\displaystyle\biggl(\int_0^T
\|f(t)\|_X^pdt\biggl)^\frac{1}{p},
&
1\leq p<\infty
,\\[10pt]
\displaystyle{\rm ess\,\,}\sup_{\!\!\!\!\!\!\!\!\!\! t\in(0,T)}\, \|f(t)\|_X .
& p=\infty,\end{array}
\right.
\end{align*}
To simplify notations, in the following sections, we write $L^p$, $H^k$
and $W^{k,p}$ as the abbreviations of $L^p(\Omega)$, $H^k(\Omega)$
and $W^{k,p}(\Omega)$, respectively, and denote by $(\cdot\, , \cdot\, )$ the inner product in $L^2(\Omega)$.
For any subdomain $Q\subset Q_T$, we define
\begin{align*}
&Q^t:=\{x\in\Omega:~ (x,t)\in Q\},\\[5pt]
&\|f\|_{L^{\infty,2}(Q)}:={\rm ess}\!\!\sup_{t\in(0,T)}
\|f(\cdot,t)\|_{L^2(Q^t)} ,\\
&
\|f\|_{L^p(Q)}:=\bigg(\iint_Q |f(x,t)|^p{\rm d} x{\rm d} t\bigg)^{\frac{1}{p}} ,
\quad\forall\, 1\leq p<\infty,
\end{align*}
and denote $w(t)=w(\cdot,t)$ for any function $w$ defined on $ Q_T$.
We assume that $\Omega$ is partitioned into quasi-uniform triangular
elements $\tau_l^h$, $l=1,\cdots,L$, with $h=\max_{l}\{\mbox{diam}\,\tau_l^h\}$,
and let $S_h$ be a finite element subspace of $H^1_0(\Omega)$ consisting
of continuous piecewise polynomials of degree $r\geq 1$ subject to
the triangulation.
Let $a(x)=\left(a_{ij}(x) \right)_{N\times N}$ be the coefficient matrix
and define the operators
\begin{align*}
&A:H^1_0\rightarrow H^{-1},\qquad~\, A_h:S_h\rightarrow S_h,\\
&R_h:H^1_0\rightarrow S_h, \qquad~~\, P_h:L^2\rightarrow S_h ,
\end{align*}
by
\begin{align*}
&\big(A\phi,v\big)= \big(a\nabla \phi,\nabla
v\big) &&
\mbox{for all~~$\phi ,v \in H^1_0$},\\[3pt]
&\big(A_h\phi_h,v\big)= \big(a\nabla \phi_h,\nabla
v\big) &&
\mbox{for all~~$\phi_h\in S_h$, $v\in S_h$},\\[3pt]
&\big(A_hR_hw,v\big)=\big(Aw,v\big) &&\mbox{for
all~~$w \in H^1_0$ and $v\in S_h$} ,
\\[3pt]
&\big(P_h\phi,v\big)=\big(\phi,v\big) &&
\mbox{for all~~$\phi \in L^2$
and $v \in S_h$} .
\end{align*}
Clearly, $R_h$ is the Ritz projection operator associated to the elliptic
operator $A$ and $P_h$ is the $L^2$ projection operator onto the finite element
space.
The following estimates are useful in this paper.
\begin{lemma}\label{LemMaxLp}
{\it
If $\Omega$ is a bounded convex domain
and $a_{ij}\in W^{1,N+\alpha}(\Omega)$, $N\geq 2$,
then we have
\begin{align}
&\|w\|_{H^2}\leq
C\| \nabla\cdot(a\nabla w)\|_{L^2},
&&\forall\, w\in H^1_0,
\label{H2Reg0}\\
&\|\nabla w\|_{L^\infty}
\leq C_{p}\| \nabla\cdot(a\nabla w)\|_{L^p}, \qquad
\mbox{for any given $p>N$},
&&\forall\, w\in H^1_0 ,
\label{W1inftyReg}
\end{align}
and the solution of \refe{PDE0}
with $u^0=0$ satisfies
\begin{align}
&\|\partial_tu\|_{L^p((0,T);L^q)}
+\|Au\|_{L^p((0,T);L^q)}
\leq C_{p,q}\|f \|_{L^p((0,T);L^q)} ,
\label{LpW2quf}\\
&\|\partial_tu\|_{L^p((0,T);W^{-1,q})}
+\|u\|_{L^p((0,T);W^{1,q})}
\leq C_{p,q}\|f \|_{L^p((0,T);W^{-1,q})},
\label{LpW1quf}
\end{align}
for all $1<p,q<\infty$.
}
\end{lemma}
In the Lemma above, \refe{H2Reg0} is the
standard $H^2$-regularity estimate in convex domains and
\refe{W1inftyReg} is a simple consequence of
the Green's function estimates given
in Theorem 3.3--3.4 of \cite{GW},
and \refe{LpW2quf}-\refe{LpW1quf} are consequences of the
maximal $L^p$ regularity (see Appendix for details).
\subsection{Properties of the finite element space and Green's functions}
\label{Sec2-2}
For any subdomain $D\subset\Omega$,
we denote by $S_h(D)$ the space of functions restricted to
the domain $D$, and denote by $S_h^0(D)$ the subspace of $S_h(D)$ consisting of functions which equal zero
outside $D$. For any given
subset $D\subset\Omega$, we denote
$B_d (D) =\{x\in\Omega: {\rm dist}(x,D)\leq d\}$ for $d>0$.
Then there exist positive constants
$K $ and $\kappa$ such that the
triangulation and the corresponding
finite element space $S_h$ possess the following properties
($K$ and $\kappa$ are independent of the subset $D$ and $h$).
\medskip
{\bf(P0)$\,$ Quasi-uniformity:}
For all triangles (or tetrahedron) $\tau_l^h$ in the partition,
the diameter $h_l$ of $\tau_l^h$ and the radius $\rho_l$
of its inscribed ball satisfy
$$
K^{-1}h\leq \rho_l \leq h_l\leq Kh .
$$
{\bf(P1)$\,$ Inverse inequality:}
If $D$ is a union of elements in the partition, then
\begin{align*}
\|\chi_h\|_{W^{l,p}(D)}
\leq K h^{-(l-k)-(N/q-N/p)}\|\chi_h\|_{W^{k,q}(D)} ,
\quad\forall\,\,\chi_h\in S_h,
\end{align*}
for $0\leq k\leq l\leq 1$ and
$1\leq q\leq p\leq\infty$.
{\bf(P2)$\,$ Local approximation and superapproximation:}
(1) There exists a linear operator
$I_h:H^1_0(\Omega)\rightarrow S_h$
such that if $d\geq \kappa h$, then
\begin{align*}
&\|v-I_hv\|_{L^2(D)}
\leq K\sum_{l=0}^kh^{k}d^{-l}\|v\|_{H^{{k-l}}(B_d(D))} ,
\quad \forall\,\,v\in H^k\cap H^1_0 , \,\,\,
1\leq k\leq 2 .
\end{align*}
Moreover, if supp$(v)\subset \overline D$,
then $I_hv\in S_h^0(B_{d}(D))$.
For example, the Cl\'ement interpolation operator
defined in \cite{Clement} has these properties.
Also, the Lagrange interpolation operator $\Pi_h$ satisfies
\begin{align*}
&\|v-\Pi_hv\|_{L^2(D)}
+h\|\nabla(v-\Pi_hv)\|_{L^2(D)}
\leq Kh^2\|\nabla^2v\|_{L^2(B_d(D))} ,
\quad \forall\,\,v\in H^2\cap H^1_0 .
\end{align*}
(2) If $d\geq \kappa h$,
$\omega=0$ outside $B_{2d}(D)$
and $|\partial^\beta\omega|\leq Cd^{-|\beta|}$
for all multi-index $\beta$,
then for any $\psi_h\in S_h(B_{3d}(D))$
there exists $\eta_h\in S_h^0(B_{3d}(D))$ such that
\begin{align*}
&\|\omega\psi_h-\eta_h\|_{H^k(B_{3d}(D))}
\leq K h^{1-k}d^{-1}
\|\psi_h\|_{L^2(B_{3d}(D))} ,
\quad k=0,1 .
\end{align*}
Furthermore, if $\omega\equiv 1$
on $B_{d}(D)$, then $\eta_h=\psi_h$ on $D$ and
\begin{align*}
&\|\omega\psi_h-\eta_h\|_{H^k(B_{3d}(D))}
\leq K h^{1-k}d^{-1}
\|\psi_h\|_{L^2(B_{3d}(D)\backslash D)} ,
\quad k=0,1 .
\end{align*}
For example, $\eta_h=\Pi_h(\omega\psi_h)$ has
these properties.\medskip
{\bf(P3)$\,$ Regularized Delta function:}
For any $x_0\in\overline\tau_j^h$, there exists a function
$\widetilde\delta_{x_0}\in C^3(\overline\Omega)$ with support in $\tau_j^h$
such that
\begin{align*}
&\chi_h(x_0)=\int_{\tau_j^h}\chi_h \widetilde\delta_{x_0}{\rm d} x,
\quad\forall\,\chi_h\in S_h ,\\
&\|\widetilde\delta_{x_0}\|_{W^{l,p}}
\leq K h^{-l-N(1-1/p)}
\quad\mbox{for}\,\,\,1\leq p\leq\infty,
\,\,\, l=0,1,2,3 .
\end{align*}
{\bf(P4)$\,$ Discrete Delta function}
Let $\delta_{x_0}$ denote the Dirac Delta function centered at
$x_0$, i.e. $\int_\Omega\delta_{x_0}(y)\varphi (y){\rm d} y=\varphi(x_0)$ for
any $\varphi\in C(\overline\Omega)$. The discrete Delta function
$P_h \widetilde \delta_{x_0}$ satisfies that
\begin{align*}
& P_h \widetilde\delta_{x_0}(x)
\leq Kh^{-N} e^{-\frac{|x-x_0| }{K h }} ,
\quad \forall\, x,x_0\in\Omega .
\end{align*}
The properties (P0)-(P4) hold for any quasi-uniform partition with
those standard finite element spaces and also, have been used in
many previous works such as \cite{Li,STW98,SW95,TW2}.
The proof can be found in the appendix of \cite{SW95}.\medskip
For an element $\tau_l^h$ and a point $x_0\in \overline\tau_l^h$, we
let $G(t,x,x_0)$ be the Green's function of the parabolic
equation, defined by
\begin{align}\label{GFdef}
&\partial_tG(t,\cdot,x_0)+ AG(t,\cdot,x_0)=0\quad
\mbox{for $t>0$ with $G(0,x,x_0)=\delta_{x_0}(x)$} ,
\end{align}
The regularized Green's function $\Gamma(t,x,x_0)$ is defined by
\begin{align}\label{GMFdef}
&\partial_t\Gamma(\cdot,\cdot,x_0)+A\Gamma(\cdot,\cdot,x_0)=0\quad
\mbox{for~ $t>0$~ with~
$\Gamma(0,\cdot,x_0)=\widetilde\delta_{x_0}$},
\end{align}
where $\widetilde\delta_{x_0}$ is given in (P2), and the discrete
Green's function $\Gamma_h(\cdot,\cdot,x_0)$ is defined by
\begin{align}
&\partial_t\Gamma_{h}(\cdot,\cdot,x_0)
+ A_h\Gamma_{h}(\cdot,\cdot,x_0)=0 \label{GMhFdef}
\quad\mbox{for~ $t>0$~ with~
$\Gamma_{h}(0,\cdot,x_0)=P_h\widetilde\delta_{x_0}$} .
\end{align}
The functions $G(t,x,x_0)$ and $\Gamma_h(t,x,x_0)$
are symmetric with respect
to $x$ and $x_0$.
By the fundamental estimates of parabolic equations,
there exists a positive constant $C$ such that
(\cite{FS}, Theorem 1.6; note that the Green's function
in the domain $\Omega$ is less than the Green's function in $\mathbb{R}^N$)
\begin{align}
&|G(t,x,y)|\leq
C(t^{1/2}+|x-y|)^{-N}e^{-\frac{|x-y|^2}{Ct}} .
\label{FEstP}
\end{align}
By estimating $\Gamma(t,x,x_0)=\int_\Omega G(t,x,y)\widetilde\delta_{x_0}(y){\rm d} y$, it is easy to see that (\ref{FEstP}) also holds
when $G$ is replaced by $\Gamma$ and
when $\max(t^{1/2},|x-y|)\geq 2h$.
\subsection{Decomposition of the domain $\Omega\times(0,T)$}
\label{SecGF}
Here we present some further notations
on a dyadic decomposition of the domain $\Omega\times(0,T)$, which were introduced in \cite{STW98} and also used in many other articles \cite{Gei1,Ley,Li,TW2}.
Let $R_0$ be the smallest distance between
a corner and a closed face which does not contained this corner.
For the given polygon/polyhedron $\Omega$,
there exists a positive constant $K_0\geq \max(1,R_0)$
(which depends on the interior angle of the edges/corners
of $\Omega$) such that
(1) if $z_0$ is a point in the interior of $\Omega$ and
$B_\rho(z_0)$
intersects a face of $\Omega$, then
$B_\rho(z_0)\subset B_{2\rho }(z_1)$ for
some $z_1$ which is on a face of $\Omega$;
(2) if $z_1$ is on a face of $\Omega$ and
$B_\rho(z_1)$
intersects another face, then
$B_\rho(z_1)\subset B_{\rho K_0}(z_2)$ for
some $z_2$ which is on an edge of $\Omega$;
(3) if $z_2$ is on an edge of $\Omega$ and
$B_\rho(z_2)$
intersects another face which does not contain this edge,
then $B_\rho(z_2)\subset B_{\rho K_0}(z_3)$ for
some $z_3$ which is a corner of $\Omega$.
For any integer $j\geq 1$, we define $d_j=2^{-j-3}R_0K_0^{-2}$.
For a given $x_0\in\Omega$,
we let $J_*$ be an integer satisfying
$d_{J_*}=2^{-J_*-3}R_0K_0^{-2}= C_*h$
with $C_*\geq \max(10,10\kappa,R_0K_0^{-2}/8)$
to be determined later. Thus,
$J_*=\log_2[R_0K_0^{-2}/(8C_*h)]\leq \log_2(2+1/h)$
and $J_*>1$ when $h<R_0K_0^{-2}/(16C_*) $.
Let
\begin{align*}
&Q_*(x_0)=\{(x,t)\in\Omega_T: \max (|x-x_0|,t^{1/2})\leq d_{J_*}\},
\\
&\Omega_*(x_0)=\{x\in \Omega: |x-x_0|\leq d_{J_*}\} \, , \\
&Q_j(x_0)=\{(x,t)\in \Omega_T:d_j\leq
\max (|x-x_0|,t^{1/2})\leq2d_j\},
\\
&\Omega_j(x_0)=\{x\in \Omega: d_j\leq|x-x_0|\leq2d_j\} ,
\\
&D_j(x_0)=\{x\in \Omega: |x-x_0|\leq2d_j\}
\end{align*}
for $j\geq 1$; see Figure \ref{DomainFig}.
\begin{figure}[ht]
\vspace{0.1in}
\centering
\epsfig{file=Fig1.eps,height=1.6in,width=2.1in}
\caption{Illustration of the
subdomains $Q_j$, $\Omega_j$ and $D_j$.}
\label{DomainFig}
\end{figure}
For $j=0$ we define $Q_0(x_0)=Q_T\backslash Q_1(x_0)$ and
$\Omega_0(x_0)=\Omega\backslash\Omega_1(x_0)$, and
for $j<0$ we simplify define $Q_j(x_0)=\Omega_j(x_0)=\emptyset$.
For all $j\geq 1$ we define
\begin{align*}
&\Omega_j'(x_0)=\Omega_{j-1}(x_0)\cup\Omega_{j}(x_0)\cup\Omega_{j+1}(x_0), \\
&\Omega_j''(x_0)=\Omega_{j-2}(x_0)\cup\Omega_{j}'(x_0)\cup\Omega_{j+2}(x_0),\\
&\Omega_j'''(x_0)=\Omega_{j-2}(x_0)\cup\Omega_{j}''(x_0)\cup\Omega_{j+2}(x_0),\\
& Q_j'(x_0)=Q_{j-1}(x_0)\cup Q_{j}(x_0)\cup Q_{j+1}(x_0), \\
&
Q_j''(x_0)=Q_{j-2}(x_0)\cup Q_{j}'(x_0)\cup Q_{j+2}(x_0),\\
&
Q_j'''(x_0)=Q_{j-2}(x_0)\cup Q_{j}''(x_0)\cup Q_{j+2}(x_0) ,\\
&D_j'(x_0)=D_{j-1}(x_0)\cup D_{j}(x_0) ,\\
& D_j''(x_0)=D_{j-2}(x_0)\cup D_{j}'(x_0) ,\\
& D_j'''(x_0)=D_{j-3}(x_0)\cup D_{j}''(x_0) .
\end{align*}
Then we have
\begin{align*}
&\Omega_T=\bigcup^{J_*}_{j=0}Q_j(x_0)\,\cup Q_*(x_0)
\quad\mbox{and}\quad
\Omega=\bigcup^{J_*}_{j=0}\Omega_j(x_0)\,\cup \Omega_*(x_0),
\end{align*}
We refer to $Q_*(x_0)$ as the ``innermost" set.
We shall write $\sum_{*,j}$ when the innermost set is included and
$\sum_j$ when it is not. When $x_0$ is fixed, if there is no ambiguity,
we simply write
$Q_j=Q_j(x_0)$, $Q_j'=Q_j'(x_0)$, $Q_j''=Q_j''(x_0)$,
$\Omega_j=\Omega_j(x_0)$, $\Omega_j'=\Omega_j'(x_0)$ and
$\Omega_j''=\Omega_j''(x_0)$.
In the rest of this paper, we denote by $C$ a generic positive constant, which
will be independent of $h$, $x_0$, and the undetermined constant
$C_*$ until it is determined at the end of section \ref{dka7}.
\subsection{Proof of Theorem \ref{MainTHM1}}
The keys to the proof of Theorem \ref{MainTHM1} are
several more precise estimates of the Green's functions.
Let $F(t)=\Gamma_h(t)-\Gamma(t)$.
Then for any $x_0 \in \Omega$, we have
\begin{align}
(E_h(t)v_h)(x_0) &= (F (t), v_h) + (\Gamma(t), v_h)
\nonumber\\
&=\int_0^t(\partial_tF (s), v_h){\rm d} s+(F(0), v_h) + (\Gamma(t), v_h) ,
\label{EhtFt1} \\[10pt]
(t\partial_tE_h(t)v_h)(x_0)
&= (t\partial_tF (t), v_h) + (t\partial_t\Gamma(t), v_h)
\nonumber \\ &
=\int_0^t(s\partial_{ss}F (s)+\partial_{s}F (s), v_h){\rm d} s
+ (t\partial_t\Gamma(t), v_h) ,
\label{EhtFt2}
\end{align}
with $\|F(0)\|_{L^1}=\|\widetilde\delta_{x_0}
-P_h\widetilde\delta_{x_0}\|_{L^1}\leq C $ (according to (P3) and (P4)).
Moreover, by the analyticity of the continuous parabolic semigroup
on $L^1(\Omega)$, we have
$$
\|\Gamma(t)\|_{L^1}+t\|\partial_t\Gamma(t)\|_{L^1}
\leq C \|\Gamma(0)\|_{L^1}
=C \|\widetilde\delta_{x_0}\|_{L^1}
\leq C .
$$
We present some estimates of these Green's
functions in the following lemma.
The proof of the lemma is the major work of this paper and
will be given in the next two sections.
\begin{lemma}
\label{lemma2-1}
{\it Under the assumptions of Theorem \ref{MainTHM1},
we have
\begin{align}
&\int_0^\infty\int_\Omega\big( |
\partial_tF(t,x,x_0) \big|
+\big|t\partial_{tt}F(t,x,x_0 ) |\big){\rm d} x{\rm d} t \leq C , \label{FFEst2}\\
&|\nabla\partial_tG(t,x,x_0)|\leq C\max(t^{1/2},|x-x_0|)^{-3-N}
\quad\mbox{for}\,\,\,\, (x,t)\in \Omega\times(0,1) .
\label{DxtGrEst}
\end{align}
}
\end{lemma}
The estimates in Lemma \ref{lemma2-1} were proved in \cite{STW98}
for parabolic equations with the Neumann boundary condition
and in \cite{TW2} for the Dirichlet boundary condition.
However, their proofs are only valid for smooth coefficients
and smooth domains (as clearly mentioned in
their papers).
Later, these estimates were proved in \cite{Li}
for parabolic equations in smooth domains
of arbitrary dimensions under the Neumann boundary condition with
Lipschitz continuous coefficients.
Here we are concerned with the problem
in a convex polyhedron in two or three dimensional spaces under the Dirichlet
boundary condition with
$a_{ij} \in W^{1,N+\alpha}$.
\vskip0.1in
\noindent{\it Proof of Theorem \ref{MainTHM1}: }
Firstly, from \refe{EhtFt1}-\refe{EhtFt2} we see that
(\ref{STLEst}) is a consequence of (\ref{FFEst2}).
Secondly, we can view $E_h(t)$ as an analytic semigroup on
$L^q(\Omega)$, defined by
$$
(E_h(t)v)(x_0):=\int_\Omega \Gamma_h(t,x,x_0)v(x){\rm d} x ,
$$
whose generator is $A_hP_h$.
From \cite[Theorem 4.2]{Weis1} and
\cite[Lemma 4.c]{Weis2} (with a duality argument
for the case $q\geq 2$) we know that the maximal $L^p$
regularity \refe{LpqSt3} holds if
the following maximal ergodic estimate holds:
\begin{align}\label{maxergd}
\bigg\|\sup_{t>0}\frac{1}{t}\int_0^t|E_h(s)|\, v \,{\rm d} s\bigg\|_{L^q}
\leq C\|v\|_{L^q} ,\quad \forall\, v\in L^q(\Omega) ,
\end{align}
where
$$
(|E_h(s)|v) (x_0):=\int_\Omega |\Gamma_h(t,x,x_0)| v(x){\rm d} x .
$$
Let $G_{\rm tr}(t,x,x_0)$ be a truncated Green's function
which is symmetric with
respect to $x$ and $x_0$ and satisfies
$G_{\rm tr}(t,x,x_0)=G(t,x,x_0)$ when $(x,t)$ is outside
$Q_*(x_0)$ (see \cite[Section 4.2]{Li} on its construction).
Then we have
(assuming that $\tau_0^h$ is the triangle/tetrahedron which
contains $x_0$)
\begin{align*}
&\iint_{[\Omega\times(0,\infty)]\backslash
Q_*(x_0)}|\partial_t\Gamma(t,x,x_0)
-\partial_tG_{\rm tr}(t,x,x_0)|{\rm d} x{\rm d} t \\
&= \iint_{[\Omega\times(0,1)]\backslash Q_*(x_0)}\biggl|\int_\Omega
\partial_tG(t,x,y)\widetilde\delta_{x_0}(y){\rm d}
y-\partial_tG(t,x,x_0)\biggl|{\rm d} x{\rm d} t\\
&\quad
+\iint_{\Omega\times(1,\infty)}|\partial_t\Gamma(t,x,x_0)
-\partial_tG(t,x,x_0)|{\rm d} x{\rm d} t\\
&\leq
Ch\iint_{[\Omega\times(0,1)]\backslash Q_*(x_0)}\sup_{y\in\tau_0^h}\big|
\nabla_y\partial_tG(t,x,y)\big|{\rm d} x{\rm d} t \\
&\quad
+C\int_{1}^\infty t^{-1}( \|\Gamma(t/2,\cdot,x_0)\|_{L^1}
+\|G(t/2,\cdot,x_0)\|_{L^1}){\rm d} t
\quad \mbox{[by semigroup estimate]} \\
&=Ch\sum_{j}\iint_{Q_j(x_0)}\sup_{(y,t)\in Q_j'(x)}\big|
\nabla_y\partial_tG(t,x,y)\big|{\rm d} x{\rm d} t
+C\int_{1}^\infty t^{-1-N/2}{\rm d} t
\quad \mbox{[see \refe{FEstP}] }\\
&\leq C\sum_{j}\frac{h}{d_j} +C
\quad \mbox{[see \refe{DxtGrEst}]}\\
&\leq C .
\end{align*}
By using energy estimates, it is easy to see
\begin{align*}
&\iint_{Q_*(x_0)}(|\partial_t\Gamma(t,x,x_0)|
+|\partial_tG_{\rm tr}(t,x,x_0)|){\rm d} x{\rm d} t \\
&\leq
d_{J_*}^{N/2+1}
(\|\partial_t\Gamma(\cdot,\cdot,x_0)\|_{L^2(\Omega\times(0,1))}
+\|\partial_tG_{\rm tr}(\cdot,\cdot,x_0)\|_{L^2(\Omega\times(0,1))})\leq
C_*^{N/2+1} ,
\end{align*}
where the constant $C_*$ will be determined
at the end of Section 4.
Then \refe{FFEst2} and the last two inequalities imply
\begin{align*}
&\int_0^\infty\int_{\Omega}|\partial_t\Gamma_h(t,x,x_0)
-\partial_tG_{\rm tr}(t,x,x_0)|{\rm d} x{\rm d} t \leq C .
\end{align*}
In other words, the symmetric kernel
$K(x,y):=\int_0^\infty|\partial_t\Gamma_h(t,x,y)
-\partial_tG^*_{\rm tr}(t,x,y)|{\rm d} t$
satisfies
$$
\sup_{y\in\Omega}\int_\Omega K(x,y){\rm d}
x+\sup_{x\in\Omega}\int_\Omega K(x,y){\rm d} y\leq C ,
$$
and therefore, Schur's lemma implies that the corresponding operator $M_K$,
defined by $M_K{v}(x)=\int_\Omega K(x,y){v}(y){\rm d} y$, is bounded on
$L^q(\Omega)$ for all $1\leq q\leq\infty$.
Let
$E^*_{\rm tr}(t){v}(x)=\int_\Omega G_{\rm tr}^*(t,x,y){v}(y){\rm d} y$
and note that
$E^*_{\rm tr}(t){v}(x)\leq E(t)|{v}|(x)$ (because
$G_{\rm tr}^*(t,x,y)\leq G(t,x,y)$).
We have
\begin{align*}
&{ \sup_{t>0}(|E_h(t)P_h|v)(x)}\\
&
\leq
{ \sup_{t>0}(|E_h(t)P_h| v - E^*_{\rm tr}(t) v)(x)
+\sup_{t>0}(E^*_{\rm tr}(t) v)(x) }\\
&
\leq { \sup_{t>0}\big(|E_h(t)P_h-E^*_{\rm
tr}(t)|\,|v|\big)(x)+\sup_{t>0}(E^*_{\rm tr}(t)v)(x) }\\
&
={
\sup_{t>0}\bigg|(|P_h\widetilde\delta_{x}|,|v|)+\int_0^t
\int_\Omega|
\partial_t\Gamma_h(s,x,y)
-\partial_tG^*_{\rm tr}(s,x,y)|\, |v(y)|{\rm d}
y{\rm d} s\bigg| }\\
&\quad +
{ \sup_{t>0}(E^*_{\rm tr}(t)|v|)(x) }\\
&
\leq
{ (|P_h\widetilde\delta_{x}|,|v|)
+(M_K|v|)(x)+\sup_{t>0}(E(t)|v|)(x)
}
\end{align*}
where
$$
{ \|\sup_{t>0}E(t) |v|\|_{L^q}\leq C_q\|v\|_{L^q}, }
\quad\forall~1<q<\infty ,
$$
is a simple consequence of the Gaussian estimate \refe{FEstP}
(Corollary 2.1.12 and Theorem 2.1.6 of \cite{Grafakos}).
This proves a stronger estimate than
\refe{maxergd}. The proof of \refe{LpqSt3} is completed.
Finally, \refe{PDE0}-\refe{FEEq0}
imply that the error $e_h=P_hu-u_h$ satisfies the equation
(when $u^0=u^0_h=0$)
\begin{align}
\partial_t(A_h^{-1}e_h)+A_h(A_h^{-1}e_h)=P_hu-R_hu .
\end{align}
By applying \refe{LpqSt3} to the equation above, we obtain
\begin{align}
\|e_h\|_{L^p((0,T);L^q)}
\leq C_{p,q}\|P_hu-R_hu\|_{L^p((0,T);L^q)}
\leq C_{p,q}h\|u\|_{L^p((0,T);W^{1,q})}
\end{align}
for $1<p<\infty$ and $2\leq q<\infty$,
where we have used the inequality
$\|P_hu-R_hu\|_{L^q}\leq C_{q}h\|u\|_{W^{1,q}}$,
which only holds for $2\leq q<\infty$
in convex polygons/polyhedra.
Then, by using an inverse inequality
and \refe{LpW1quf}, we have
\begin{align*}
\|e_h\|_{L^p((0,T);W^{1,q})}
&\leq Ch^{-1}\|e_h\|_{L^p((0,T);L^q)} \\
&\leq C_{p,q}\|u\|_{L^p((0,T);W^{1,q})} \\
&\leq C_{p,q}\|f\|_{L^p((0,T);W^{-1,q})} ,
\end{align*}
which implies \refe{LpqSt2} for the case
$1<p<\infty$ and $2\leq q<\infty$.
In the case $1<p<\infty$ and $1<q\leq 2$,
we define $\vec g=\nabla \Delta^{-1}P_hf$ and express
the solution of \refe{FEEq0} by
(when $u_h^0=0$)
\begin{align*}
\nabla u_h=
{\mathcal L}_h \vec g:=\int_0^t\nabla A_h^{-1/2}
A_hE_h(t-s) A_h^{-1/2}\nabla\cdot
\vec g(s) {\rm d} s .
\end{align*}
In order to prove the boundedness of the operator
${\mathcal L}_h$ on $L^p((0,T);(L^q)^N)$, we only need to
prove the boundedness of its dual operator
${\mathcal L}_h'$ on $L^{p'}((0,T);(L^{q'})^N)$.
It is easy to see that
\begin{align*}
\int_0^T({\mathcal L}_h \vec g , \vec \eta){\rm d} t
=\int_0^T\bigg(\vec g,\int_s^T
\nabla A_h^{-1/2} A_hE_h(t-s) A_h^{-1/2}\nabla\cdot
\vec \eta(t) {\rm d} t\bigg){\rm d} s ,
\end{align*}
which gives
\begin{align*}
{\mathcal L}_h' \vec \eta:=\int_s^T\nabla A_h^{-1/2}
A_hE_h(t-s) A_h^{-1/2}\nabla\cdot
\vec \eta(s) {\rm d} s .
\end{align*}
If we define the backward finite element problem
\begin{align}
\left\{
\begin{array}{ll}
-\big(\partial_tw_h,v_h\big)
+ (a\nabla w_h, \nabla v_h)
=(\nabla\cdot\vec\eta,v_h) , \,\,\,\,\forall\, v_h\in S_h,
\\
w_h(T)=0 ,
\end{array}
\right.
\end{align}
then ${\mathcal L}_h' \vec \eta=\nabla w_h$.
By a time reversal we obtain, as shown in the last paragraph,
\begin{align*}
\|\nabla w_h\|_{L^{p'}((0,T);L^{q'})}
\leq C_{p,q}\|\nabla\cdot\vec\eta\|_{L^{p'}((0,T);W^{-1,q'})}
\leq C_{p,q}\|\vec \eta\|_{L^{p'}((0,T);L^{q'})} ,
\end{align*}
for $1<p'<\infty$ and $2\leq q'<\infty$,
which implies the boundedness of
${\mathcal L}_h'$ on $L^{p'}((0,T);(L^{q'})^N)$.
By duality, we derive the boundedness of
${\mathcal L}_h$ on $L^p((0,T);(L^q)^N)$ and therefore,
\begin{align*}
\|\nabla u_h\|_{L^{p}((0,T);L^{q})}
&\leq C_{p,q}\|\vec g\|_{L^{p}((0,T);L^{q})} \\
&\leq C_{p,q}\|P_hf\|_{L^p(0,T);W^{-1,q})} \\
&\leq C_{p,q}\|f\|_{L^p(0,T);W^{-1,q})} .
\end{align*}
This proves \refe{LpqSt2} in the case $1<p<\infty$ and $1<q\leq 2$.
The proof of Theorem \ref{MainTHM1} is completed
(based on Lemma \ref{lemma2-1}).
\,\vrule height8pt width 5pt depth 0pt \medskip
\noindent{\bf Remark 2.1}$\,\,$
In the proof of \refe{LpqSt2}, we have used
an $L^q$ error estimate of the Ritz projection for $2\leq q<\infty$, which
can be proved in the same way as used in \cite[Corollary]{RS} by using
the $W^{1,q}$-stability of the Ritz projection.
This $W^{1,q}$-stability
is based on an interpolation between these two cases
$q=2$ and $q=\infty$.
The case $q=2$ is trivial and the
case $q=\infty$ was studied by several authors,
such as \cite{RS} for $r=1$ and 2D convex polygons
(which requires $H^2$ regularity of the elliptic problem),
\cite{Schatz} for $r\geq 2$ and 2D arbitrary polygons
(as a consequence of the $L^\infty$ stability proved therein,
which only requires $H^{3/2+\varepsilon}$ regularity
of the elliptic problem),
and \cite{GLRS} for $r\geq 1$ in 3D convex polyhedra
(which requires $H^2$ and $C^{1+\alpha}$ regularity of the elliptic problem).
These essential properties used by \cite{GLRS,RS,Schatz} are all
possessed by the elliptic problem when
the domain is convex polygonal/polyhedral and the coefficients
$a_{ij}$ are $W^{1,N+\alpha}$.
In the rest of this paper, we focus on the proof of Lemma
\ref{lemma2-1}.
\section{Preliminary analysis}
\setcounter{equation}{0}
In this section, we present two propositions.
\subsection{Superapproximation of smoothly
truncated finite element functions}
In this subsection, we prove the following proposition,
which is needed in proving Lemma \ref{lemma2-1}.
\begin{proposition}\label{lemsupa0}
{\it If $0\leq \omega \leq 1$ is a smooth cut-off function which equals zero in
$\Omega\backslash D$, satisfying
$|\partial^\beta\omega |\leq Cd^{-|\beta|}$
for all multi-index $\beta$ such that
$|\beta|=0,1,\cdots,r+1$ and $d\geq 10\kappa h$,
then for any $\psi_h\in S_h$ there exists $\chi_h\in S_h^0(B_d(D))$
such that
\begin{align}\label{lemsupa}
d^2\|R_h(\omega\psi_h)-\chi_h\|_{H^1}
+d\|\omega \psi_h-\chi_h\|_{L^2}\leq Ch\|\psi_h\|_{L^2(B_d(D))} .
\end{align}
}
\end{proposition}
\noindent{\it Proof.}$\,\,$
Define $0\leq \widetilde\omega\leq 1$ as
a smooth cut-off function which is zero
outside $B_{0.8d}(D)$, satisfying that $\widetilde \omega=1$
on $B_{0.7d}(D)$ and
$|\partial^\beta\widetilde\omega |\leq Cd^{-|\beta|}$
for $|\beta|=0,1,\cdots,r+1$.
First we prove the following inequality
\begin{align}\label{PInEq1}
\|\omega\psi_h-R_h(\omega\psi_h) \|_{L^2 }
\leq Chd^{-1}\|\psi_h\|_{L^2(B_{0.3d}(D))}
\end{align}
by a duality argument. We define
$\phi$ as the solution of the elliptic PDE
$$
\left\{
\begin{array}{ll}
-\nabla\cdot(a\nabla \phi)= v &\mbox{in}\,\,\,\, \Omega,\\
\phi=0 &\mbox{on}\,\,\, \partial\Omega .
\end{array}\right.
$$
We see that
\begin{align*}
\big( v ,\omega\psi_h-R_h(\omega\psi_h)\big) &=
\big(a\nabla \phi,\nabla (\omega\psi_h-R_h(\omega\psi_h))\big) \\
&=\big(a\nabla (\phi -R_h\phi),\nabla (\omega\psi_h-R_h(\omega\psi_h))\big)\\
&=\big(a\nabla (\phi -R_h\phi),\nabla (\omega\psi_h-\Pi_h(\omega\psi_h))\big ) \\
&\leq C\|\phi -R_h\phi \|_{H^1}\|\omega\psi_h-\Pi_h(\omega\psi_h)\|_{H^1} \\
&\leq Chd^{-1}\|v\|_{L^2} \|\psi_h\|_{L^2(B_{0.3d}(D))} .
\end{align*}
where we have used the superapproximation property
(P2) in section \ref{Sec2-2}
and the $H^1$ error estimate:
$$
\|\phi -R_h\phi \|_{H^1} \leq C h\|\phi\|_{H^2}
\leq Ch\|v\|_{L^2} .
$$
\refe{PInEq1} follows these inequalities.
Secondly, it is noted that the following inequality
\begin{align}\label{PInEq2}
\|R_h(\omega\psi_h) \|_{H^1(B_{d}(D)\backslash B_{0.5d}(D))}
\leq Cd^{-1}\|R_h(\omega\psi_h) \|_{L^2(B_{d}(D)\backslash B_{0.3d}(D))}
\end{align}
was proved in Lemma 4.4 of \cite{SW77}
(also see Page 1374 of \cite{STW98})
as a consequence of the discrete elliptic equation
$$
(a\nabla R_h(\omega\psi_h),\nabla\eta)=0 ,\quad\mbox{ for }\,\,\,\,
\eta\in S_h^0(B_{d}(D)\backslash D) \, .
$$
Let $\chi_h=\Pi_h[\widetilde\omega R_h(\omega\psi_h)]$
and note that the support of $\chi_h$
is contained in $B_{0.8d}(D)$.
By using the superapproximation
property (P2), we have
\begin{align}\label{PInEq3}
&d^{-1}\|R_h(\omega\psi_h)-\chi_h\|_{L^2}
+\|R_h(\omega\psi_h)-\chi_h\|_{H^1} \nonumber\\
&\leq d^{-1}\|R_h(\omega\psi_h)-\widetilde \omega R_h(\omega\psi_h) \|_{L^2}
+d^{-1}\|\widetilde \omega R_h(\omega\psi_h)
-\Pi_h[\widetilde \omega R_h(\omega\psi_h)]\|_{L^2} \nonumber \\
&\quad
+\|R_h(\omega\psi_h)-\widetilde \omega R_h(\omega\psi_h) \|_{H^1}
+\|\widetilde \omega R_h(\omega\psi_h)
-\Pi_h[\widetilde \omega R_h(\omega\psi_h)]\|_{H^1} \nonumber \\
&\leq Cd^{-1}\|R_h(\omega\psi_h)\|_{L^2(B_{d}(D)\backslash B_{0.3d}(D))} \nonumber\\
&=Cd^{-1}\|R_h(\omega\psi_h)-\omega\psi_h \|_{L^2(B_{d}(D)\backslash B_{0.3d}(D))}
\,\,\quad\mbox{(because $\omega=0$ on
$B_{d}(D)\backslash B_{0.3d}(D)$)} \nonumber\\
&\leq Chd^{-2}\|\psi_h\|_{L^2(B_{0.3d}(D))}
\qquad\qquad\qquad\qquad\quad
\,\mbox{(as a consequence of \refe{PInEq1})}
\end{align}
and from \refe{PInEq1} we see that
\begin{align}
\|\omega\psi_h -\chi_h\|_{L^2}
&\leq \|\omega\psi_h -R_h(\omega\psi_h)\|_{L^2}
+\|R_h(\omega\psi_h) -\chi_h\|_{L^2} \nonumber\\
&\leq Chd^{-1}\|\psi_h\|_{L^2(B_{0.3d}(D))} \, .
\end{align}
\refe{lemsupa} follows immediately and the proof of
the Proposition \ref{lemsupa0} is completed.
\quad \vrule height8pt width 5pt depth 0pt \medskip
\noindent{\bf Remark 3.1}$\,\,\,$
In the proof of Proposition \ref{lemsupa0}
we have assumed that $d\geq 10\kappa d$ and used
$B_{0.3d}(D)$, $B_{0.7d}(D)$, $B_{0.8d}(D)$ ...
to make sure that their radius differ from each other
by at least $\kappa h$ so that
the superapproximation property (P2) can be used.
\subsection{Local error estimate}
The following proposition is
concerned with a local energy
error estimate of parabolic equations.
\begin{proposition}\label{LocEEst}
{\it
Suppose that $\phi,\partial_t\phi\in L^2((0,T);H^1_0)$ and
$\phi_h\in H^1((0,T);S_h)$, and
$e=\phi_h-\phi$ satisfies the equation
\begin{align}
(e_t,\chi)+(a\nabla e,\nabla \chi) =0,
\quad\forall\, \chi\in S_h,~t>0 ,
\label{p32}
\end{align}
with $\phi(0)=0$ in $\Omega_j''''$.
Then for any $m>0$,
there exists a constant $C_m>0$,
independent of $h$ and $d_j$, such that
\begin{align}
&\|e_t\|_{L^2(Q_j)} + d_j^{-1}\|\nabla e\|_{L^2(Q_j)} \nonumber\\
&\leq
C_m\big(I_j(\phi_{h}(0))+X_j(\Pi_h\phi-\phi)
+H_j(e)+d_j^{-2}\|e\|_{L^2(Q_j''')}\big),
\label{3-1}
\end{align}
where
\begin{align*}
&I_j(\phi_{h}(0))=d_j^{-1}\|\phi_{h}(0)\|_{L^2(\Omega_j''')}
+\|\phi_{h}(0)\|_{H^1(\Omega_j''')} ,\\
&X_j(\Pi_h\phi-\phi)=d_j\|\nabla\partial_t(\Pi_h\phi-\phi)\|_{L^2(Q_j''')}
+\|\partial_t(\Pi_h\phi-\phi)\|_{L^2(Q_j''')} \\
&\qquad \qquad \qquad\
+d_j^{-1}\|\nabla(\Pi_h\phi-\phi)\|_{L^2(Q_j''')}+
d_j^{-2}\|\Pi_h\phi-\phi\|_{L^2(Q_j''')},\\
&H_j(e)=(h/d_j)^m\big(\|e_t\|_{L^2(Q_j''')}
+d_j^{-1}\|\nabla e\|_{L^2(Q_j''')}\big) .
\end{align*}
}
\end{proposition}
Before we prove Proposition \ref{LocEEst},
we present a local energy estimate for
finite element solutions of parabolic equations.
\begin{lemma}\label{lemlocEng0}
{\it
Suppose that $\phi_h(t)\in S_h$ satisfies
\begin{align*}
&\big(\partial_t\phi_h ,\chi\big)
+\big(a\nabla \phi_h ,\nabla\chi\big) =0,
\qquad\mbox{for}\,\,\,\,\chi\in S_h^0(\Omega_j''),\,\,\, t\in(0,d_j^2] ,\\
&\big(\partial_t \phi_h ,\chi\big)
+\big(a\nabla \phi_h ,\nabla\chi\big)
=0 ,\qquad\mbox{for}\,\,\,\,\chi\in S_h^0(D_j'') ,\,\,\, t\in(d_j^2/4,4d_j^2) .
\end{align*}
Then for any $m>0$ there exists $C_m>0$,
independent of $h$ and $d_j$, such that
\begin{align}\label{locEng0}
&\|\partial_t\phi_h\|_{L^2(Q_j)} +d_j^{-1}\|\nabla \phi_h\|_{L^2(Q_j)} \nonumber\\
&\leq C_m\left(\|\nabla \phi_h(0)\|_{L^2(\Omega_j'')}
+d_j^{-1}\|\phi_h(0)\|_{L^2(\Omega_j'')} \right) \nonumber\\
&\quad
+ C_m\left(\frac{h}{d_j}\right)^{m}\big(
\|\partial_t\phi_h\|_{L^2(Q_{j}'')}
+ d_j^{-1}\|\nabla\phi_h\|_{L^2(Q_{j}'')}\big)
+ C_md_j^{-2} \|\phi_h\|_{L^2(Q_{j}'')} .
\end{align}
}
\end{lemma}
\noindent{\it Proof.}$\,\,$
Note that $Q_j=[\Omega_j\times(0,d_j^2) ]
\cup [D_j\times (d_j^2,4d_j^2)]$.
We first present estimates in the domain $\Omega_j\times(0,d_j^2)$
and then present estimates in the domain $D_j\times (d_j^2,4d_j^2)$.
Let $\omega$ be a smooth cut-off function
which equals $1$ in $\Omega_j$
and vanishes outside $\Omega_j'$,
and let $\widetilde\omega$ be a smooth cut-off function
which equals $1$ in $\Omega_j'''$
and vanishes outside $\Omega_j''''$ with
\begin{align}
|\partial^{\beta}\omega|\leq Cd_j^{-|\beta|}
\qquad\mbox{and}\qquad
|\partial^{\beta}\widetilde\omega|\leq Cd_j^{-|\beta|}.
\label{3-2}
\end{align}
Let $v_h=\Pi_h(\widetilde\omega\phi_h)\in S_h^0(\Omega_j''''')$
so that $v_h=\phi_h$ in $\Omega_j''$, satisfying
(due to the superapproximation property (P2))
\begin{align*}
&\|v_h\|_{L^2 }
\leq C\|\phi_h\|_{L^2(\Omega_j'''')} ,\\
&\|\nabla v_h\|_{L^2 }
\leq C\|\nabla \phi_h\|_{L^2(\Omega_j'''')}
+ Cd_j^{-1}\| \phi_h\|_{L^2(\Omega_j'''')}
\end{align*}
and
$$
\big(\partial_tv_h,\chi\big)+\big(a\nabla v_h,\nabla\chi\big)
=0 ,\quad\forall\, \chi\in S_h^0(\Omega_j'') \, .
$$
It follows that
\begin{align*}
&\frac{1}{2}\frac{{\rm d}}{{\rm d} t}\|\omega v_h\|^2
+(\omega^2 a\nabla v_h,\nabla v_h)\\
&=\big[\big(\partial_tv_h,\omega^2v_h-\chi_h\big)
+\big(a\nabla v_h,\nabla(R_h(\omega^2v_h)-\chi_h)\big)\big]\\
&\quad
+\big[\big(\omega^2a\nabla v_h,\nabla v_h\big)
-\big(a\nabla v_h,\nabla(\omega^2v_h)\big)\big] \\
&\leq
\big[ C\|\partial_tv_h\|_{L^2}\|v_h\|_{L^2}hd_j^{-1}
+ C\|\nabla v_h\|_{L^2}\|v_h\|_{L^2}hd_j^{-2} \big]
+ C\big( \omega a\nabla v_h,2 v_h\nabla \omega\big),
\end{align*}
where we have used (P2) and Proposition \ref{lemsupa0},
and from \refe{3-2} we see that
\begin{align*}
\big( \omega a\nabla v_h,2 v_h\nabla \omega\big)
\leq \big( |\omega a\nabla v_h|, \, 2 |v_h| \big) d_j^{-1}
\leq C \|\omega a\nabla v_h \|_{L^2} \|v_h\|_{L^2} d_j^{-1}
\, .
\end{align*}
The last two inequalities imply
\begin{align}\label{locEng1}
&\|\phi_h\|_{L^\infty((0,d_j^2)\times L^2(\Omega_j))}
+\|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j))} \nonumber\\
&\leq C\|\phi_h(0)\|_{L^2(\Omega_j'''')}
+C\|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j''''))}h \nonumber \\
&\quad
+C\|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j''''))} h d_j^{-1}
+C\|\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j''''))} d_j^{-1} .
\end{align}
By using Proposition \ref{lemsupa0} again, we derive that
\begin{align*}
&\|\omega^2\partial_tv_h\|_{L^2}^2
+\frac{1}{2}\frac{{\rm d}}{{\rm d} t}(\omega^4 a\nabla v_h,\nabla v_h) \\
&=\big[(\partial_tv_h,\omega^4\partial_tv_h-\chi_h)
+(a\nabla v_h,\nabla[R_h(\omega^4\partial_tv_h)-\chi_h])\big]
+(4\omega^3a\nabla v_h,\partial_tv_h\nabla \omega) \\
&\leq C\|\partial_tv_h\|_{L^2}^2hd_j^{-1}
+C\|\nabla v_h\|_{L^2}\|\partial_tv_h\|_{L^2}hd_j^{-2}
+C\|\omega\nabla v_h\|_{L^2}
\|\omega^2\partial_t v_h\|_{L^2} d_j^{-1} \\
&\leq C\|\partial_tv_h\|_{L^2}^2hd_j^{-1}
+ C\|\nabla v_h\|_{L^2}^2 d_j^{-2}
+\frac{1}{2} \|\omega^2\partial_t v_h\|_{L^2}^2 ,
\end{align*}
which reduces to
\begin{align*}
&\|\omega^2\partial_tv_h\|_{L^2((0,d_j^2);L^2(\Omega))}^2 \\
&\leq C\|\nabla v_h(0)\|_{L^2(\Omega)}^2
+ C\|\partial_tv_h\|_{L^2((0,d_j^2);L^2(\Omega))}^2hd_j^{-1}
+ C\|\nabla v_h\|_{L^2((0,d_j^2);L^2(\Omega))}^2 d_j^{-2} .
\end{align*}
The inequality above further implies
\begin{align}\label{locEng2}
\|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j))}
&\leq C(\|\nabla\phi_h(0)\|_{L^2(\Omega_j'''')}
+d_j^{-1}\|\phi_h(0)\|_{L^2(\Omega_j'''')}) \nonumber\\
&\quad
+ C\|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j''''))} h^{1/2}d_j^{-1/2}
\nonumber\\
&\quad
+ Cd_j^{-1}(\|\phi_h\|_{L^\infty((0,d_j^2)\times L^2(\Omega_j''''))}
+\|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j''''))} ) .
\end{align}
With an obvious change of indices
(from $\Omega''''$ to $\Omega'''$ on the right-hand side,
and from $\Omega$ to $\Omega'$ on the left-hand side),
\refe{locEng1}-\refe{locEng2}
imply
\begin{align}\label{locEng1-2}
& \|\phi_h\|_{L^\infty((0,d_j^2); L^2(\Omega_j'))}
+\|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'))} \nonumber \\
&\leq C\|\phi_h(0)\|_{L^2(\Omega_j''')}
+C h \|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'''))} \nonumber \\
&\quad
+C h d_j^{-1} \|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'''))}
+C d_j^{-1} \|\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'''))} .
\end{align}
and
\begin{align}\label{locEng2-2}
\|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j))}
&\leq C(\|\nabla\phi_h(0)\|_{L^2(\Omega_j')}
+d_j^{-1}\|\phi_h(0)\|_{L^2(\Omega_j')}) \nonumber\\
&\quad
+ C h^{1/2}d_j^{-1/2} \|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'))}
\nonumber\\
&\quad
+ Cd_j^{-1} (\|\phi_h\|_{L^\infty((0,d_j^2); L^2(\Omega_j'))}
+\|\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j'))} ) .
\end{align}
In the same way as we derive \refe{locEng1-2}-\refe{locEng2-2},
by choosing $\overline\omega(x,t)=\omega_1(x)\omega_2(t)$
with $\omega_1=1$ in $D_j'$, $\omega_1=0$ outside $D_j''$,
$\omega_2=1$ for $t\in(d_j^2,4d_j^2)$ and
$\omega_2=0$ for $t\in(0,d_j^2/2)$,
we can derive that
\begin{align}\label{locEng3}
&\|\overline\omega\phi_h\|_{L^\infty((0,4d_j^2); L^2(\Omega))}
+\|\overline\omega\nabla \phi_h\|_{L^2((0,4d_j^2);L^2(\Omega))} \nonumber\\
&\leq C h \|\partial_t\phi_h\|_{L^2((d_j^2/4,4d_j^2);L^2(D_{j}'''))} \nonumber\\
&\quad
+C h d_j^{-1} \|\nabla \phi_h\|_{L^2((d_j^2/4,4d_j^2);L^2(D_{j}'''))} \nonumber\\
&\quad
+C d_j^{-1} \|\phi_h\|_{L^2((d_j^2/4,4d_j^2);L^2(D_{j}'''))}
\end{align}
and
\begin{align}\label{locEng4}
\|\partial_t \phi_h\|_{L^2((d_j^2,4d_j^2);L^2(D_{j}))}
&\le C h^{1/2}d_j^{-1/2}
\|\partial_t \phi_h\|_{L^2((d_j^2/4,4d_j^2);L^2(D_{j}'''))}
\nonumber\\
&\quad
+ Cd_j^{-1} (\|\overline\omega\phi_h\|_{L^\infty((0,4d_j^2); L^2(\Omega))}
+\|\overline\omega\nabla \phi_h\|_{L^2((0,4d_j^2);L^2(\Omega))} ) .
\end{align}
By noting the definition of $\omega$ and $\overline \omega$,
we have
\begin{align*}
& \|\partial_t \phi_h\|_{L^2(Q_j)} \le
C ( \|\partial_t\phi_h\|_{L^2((0,d_j^2);L^2(\Omega_j))}
+ \|\partial_t \phi_h\|_{L^2((d_j^2,4d_j^2);L^2(D_{j}))} )
\\
& \|\nabla \phi_h\|_{L^2(Q_j)} \le
C( \|\omega\nabla \phi_h\|_{L^2((0,d_j^2);L^2(\Omega))}
+ \|\overline\omega\nabla \phi_h\|_{L^2((0,4d_j^2);L^2(\Omega))} )
\, .
\end{align*}
With the last two inequalities, combining \refe{locEng1}-\refe{locEng4} gives
\begin{align*
\|\partial_t \phi_h\|_{L^2(Q_j)} +d_j^{-1}\|\nabla \phi_h\|_{L^2(Q_j)}
&\leq C(\|\nabla \phi_h(0)\|_{L^2(\Omega_j''')}
+d_j^{-1}\| \phi_h(0)\|_{L^2(\Omega_j''')} ) \\
&\quad
+ C\left(\frac{h}{d_j}\right)^{\frac{1}{2}}
\big( \|\partial_t \phi_h\|_{L^2(Q_{j}''' )}
+ d_j^{-1}\|\nabla\phi_h\|_{L^2(Q_{j}''' )}\big) \\
&\quad
+ Cd_j^{-2} \| \phi_h\|_{L^2(Q_{j}''')} .
\end{align*}
Iterating the inequality above and
changing the indices, we derive \refe{locEng0}.
\quad \vrule height8pt width 5pt depth 0pt
\vskip0.1in
Now we are ready to prove Proposition \ref{LocEEst}.
\medskip
\noindent{\it Proof of Proposition \ref{LocEEst}}$\quad$
Let $\widetilde\omega(x,t)$ be a smooth cut-off function
which equals $1$ in $Q_j''$
and vanishes outside $Q_j'''$,
and let $\widetilde \phi=\widetilde\omega \phi$.
Then $\widetilde\phi(0)=0$ and we have
\begin{align*}
&\big(\partial_t(\widetilde \phi-\phi_h),\chi\big)
+\big(a\nabla (\widetilde \phi-\phi_h),\nabla\chi\big)
=0 ,\qquad\mbox{for}\,\,\,\,\chi\in S_h^0(\Omega_j ') ,\,\,\, t\in(0,d_j^2] ,\\
& \big(\partial_t(\widetilde \phi-\phi_h),\chi\big)
+\big(a\nabla (\widetilde \phi-\phi_h),\nabla\chi\big)
=0 ,\qquad\mbox{for}\,\,\,\,\chi\in S_h^0(D_j ') ,\,\,\, t\in(d_j^2/4,4d_j^2) .
\end{align*}
Let $\widetilde \phi_h\in S_h$ be the solution of
\begin{align}\label{wdtzh}
\big(\partial_t(\widetilde\phi-\widetilde \phi_h),\chi_h\big)
+\big(a\nabla (\widetilde\phi-\widetilde \phi_h),\nabla\chi_h\big)
=0 ,\qquad\mbox{for}\,\,\,\,\chi_h\in S_h
\end{align}
with $\widetilde\phi_h(0)=\Pi_h\widetilde \phi(0)= 0$ so that
\begin{align}
&\big(\partial_t(\widetilde \phi_h-\phi_h),\chi_h\big)
+\big(a\nabla (\widetilde \phi_h-\phi_h),\nabla\chi_h\big)
=0 ,\quad\forall\,\,\chi_h\in S_h^0(\Omega_j'),\,\,\, t\in(0,d_j^2] ,
\label{wdtzh2}\\
&\big(\partial_t(\widetilde \phi_h-\phi_h),\chi_h\big)
+\big(a\nabla (\widetilde \phi_h- \phi_h),\nabla\chi_h\big)
=0 ,\quad\forall\,\,\chi_h\in S_h^0(D_j') ,\,\,\, t\in(d_j^2/4,4d_j^2) .
\label{wdtzh3}
\end{align}
Substituting $\chi_h=P_h\widetilde\phi-\widetilde\phi_h$
into (\ref{wdtzh}) we obtain
\begin{align*}
&\|\widetilde\phi-\widetilde \phi_h\|_{L^\infty((0,T);L^2)}^2
+\int_0^T\big(a\nabla (\widetilde\phi-\widetilde\phi_h),
\nabla(\widetilde\phi-\widetilde\phi_h)\big){\rm d} t \\
&=\int_0^T\big(\partial_t(\widetilde\phi-\widetilde \phi_h),
\widetilde\phi-P_h\widetilde\phi \big){\rm d} t
+\int_0^T\big(a\nabla (\widetilde\phi-\widetilde\phi_h),
\nabla(\widetilde\phi-P_h\widetilde\phi)\big){\rm d} t ,
\end{align*}
which implies
\begin{align} \label{H1phiphih}
&\|\widetilde\phi-\widetilde \phi_h\|_{L^\infty((0,T); L^2(Q_T)}^2
+\|\nabla (\widetilde \phi-\widetilde \phi_h)\|_{L^2(Q_T)}^2 \nonumber\\
&
\leq C\|\partial_t (\widetilde \phi-\widetilde \phi_h)\|_{L^2(Q_T)}
\| \widetilde \phi \|_{L^2(Q_T)}
+C\| \nabla\widetilde\phi \|_{L^2(Q_T)}^2 .
\end{align}
Substituting $\chi_h=\partial_t(P_h\widetilde\phi-\widetilde\phi_h)$
into (\ref{wdtzh}) we obtain
\begin{align*}
&\|\partial_t(\widetilde\phi-\widetilde \phi_h)\|_{L^2(Q_T)}^2
+\sup_{0\leq t\leq T}\big(a\nabla (\widetilde\phi-\widetilde\phi_h),
\nabla(\widetilde\phi-\widetilde\phi_h)\big) \\
&=\int_0^T\big(\partial_t(\widetilde\phi-\widetilde \phi_h),
\partial_t(\widetilde\phi-P_h\widetilde\phi)\big){\rm d} t
+\int_0^T\big(a\nabla (\widetilde\phi-\widetilde\phi_h),
\nabla\partial_t(\widetilde\phi-P_h\widetilde\phi)\big){\rm d} t \\
&\leq \|\partial_t(\widetilde\phi-\widetilde \phi_h)\|_{L^2(Q_T)}
\|\partial_t\widetilde\phi \|_{L^2(Q_T)}
+\|\nabla(\widetilde\phi-\widetilde \phi_h)\|_{L^2(Q_T)}
\|\nabla\partial_t\widetilde\phi\|_{L^2(Q_T)} ,
\end{align*}
which implies
\begin{align} \label{H1tphiphih}
&\|\partial_t(\widetilde \phi-\widetilde \phi_h)\|_{L^2(Q_T)}^2
\leq C\|\partial_t \widetilde \phi \|_{L^2(Q_T)}^2
+C\|\nabla (\widetilde \phi-\widetilde \phi_h)\|_{L^2(Q_T)}
\| \nabla\partial_t\widetilde \phi \|_{L^2(Q_T)} ,
\end{align}
It follows that
\begin{align*}
&\|\partial_t(\widetilde \phi-\widetilde\phi_h)\|_{L^2(Q_T)}^2
+d_j^{-2}\|\nabla (\widetilde \phi-\widetilde\phi_h)\|_{L^2(Q_T)}^2
+d_j^{-2}\|\widetilde\phi-\widetilde \phi_h\|_{L^\infty((0,T);L^2)}^2\\
&\leq \frac{1}{2}\|\partial_t (\widetilde\phi-\widetilde\phi_h)\|_{L^2(Q_T)}^2
+Cd_j^{-4}\| \widetilde \phi \|_{L^2(Q_T)}^2
+Cd_j^{-2}\| \nabla\widetilde \phi \|_{L^2(Q_T)}^2,\\
&\quad
+C\|\partial_t \widetilde \phi \|_{L^2(Q_T)}^2
+\frac{1}{2}d_j^{-2}\|\nabla (\widetilde \phi-\widetilde \phi_h)\|_{L^2(Q_T)}^2
+Cd_j^{2}\| \nabla\partial_t\widetilde\phi \|_{L^2(Q_T)}^2 .
\end{align*}
which in turn produces
\begin{align*}
&\|\partial_t(\widetilde\phi-\widetilde\phi_h)\|_{L^2(Q_T)}
+d_j^{-1}\|\nabla (\widetilde\phi-\widetilde\phi_h)\|_{L^2(Q_T)}
+d_j^{-1}\|\widetilde\phi-\widetilde \phi_h\|_{L^\infty((0,T);L^2)}
\nonumber\\
&\leq Cd_j^{-2}\| \widetilde\phi \|_{L^2(Q_T)}
+Cd_j^{-1}\| \nabla\widetilde\phi \|_{L^2(Q_T)}
+C\|\partial_t \widetilde\phi \|_{L^2(Q_T)}
+Cd_j \| \nabla\partial_t\widetilde \phi \|_{L^2(Q_T)} .
\end{align*}
By applying Lemma \ref{lemlocEng0} to \refe{wdtzh2}-\refe{wdtzh3}
and using the inequality above,
we derive that
\begin{align*
&\|\partial_t(\widetilde\phi_h-\phi_h)\|_{L^2(Q_j)}
+ d_j^{-1}\|\nabla(\widetilde \phi_h-\phi_h)\|_{L^2(Q_j)} \nonumber\\
&\leq C_m\left(\|\nabla (\widetilde \phi_h-\phi_h)(0)\|_{L^2(\Omega_j'')}
+d_j^{-1}\|(\widetilde \phi_h-\phi_h)(0)\|_{L^2(\Omega_j'')} \right) \nonumber\\
&\quad
+ C_m\left(\frac{h}{d_j}\right)^{m}\big(
\|\partial_t(\widetilde \phi_h-\phi_h)\|_{L^2(Q_{j}'')}
+ d_j^{-1}\|\nabla(\widetilde \phi_h-\phi_h)\|_{L^2(Q_{j}'')}\big)\nonumber\\
&\quad
+ C_md_j^{-2} \| \widetilde \phi_h-\phi_h\|_{L^2(Q_{j}'')} \nonumber\\
&\leq C_m\left(\|\nabla \phi_h(0)\|_{L^2(\Omega_j'')}
+d_j^{-1}\|\phi_h(0)\|_{L^2(\Omega_j'')} \right) \nonumber\\
&\quad
+ C_m\left(\frac{h}{d_j}\right)^{m}\big(
\|\partial_te\|_{L^2(Q_{j}'')}
+ d_j^{-1}\|\nabla e\|_{L^2(Q_{j}'')}\big)
+ C_md_j^{-2} \|e\|_{L^2(Q_{j}'')} \nonumber\\
&\quad
+ C_m\left(\frac{h}{d_j}\right)^{m}\big(
\|\partial_t(\widetilde\phi-\widetilde\phi_h)\|_{L^2(Q_{j}'')}
+ d_j^{-1}\|\nabla (\widetilde\phi-\widetilde\phi_h)\|_{L^2(Q_{j}'')}\big) \nonumber\\
&\quad
+ C_md_j^{-1} \| \widetilde \phi-\widetilde\phi_h\|_{L^\infty L^2(Q_j'')} \nonumber\\
&\leq C_m\left(\|\nabla \phi_h(0)\|_{L^2(\Omega_j'')}
+d_j^{-1}\|\phi_h(0)\|_{L^2(\Omega_j'')} \right) \nonumber\\
&\quad
+ C_m\left(\frac{h}{d_j}\right)^{m}\big(
\|\partial_te\|_{L^2(Q_{j}'')}
+ d_j^{-1}\|\nabla e\|_{L^2(Q_{j}'')}\big)
+ C_md_j^{-2} \|e\|_{L^2(Q_{j}'')} \nonumber\\
&\quad
+ C_md_j^{-2}\| \widetilde\phi \|_{L^2(Q_T)}
+C_md_j^{-1}\| \nabla\widetilde\phi \|_{L^2(Q_T)}
+C_m\|\partial_t \widetilde\phi \|_{L^2(Q_T)} \nonumber\\
&\quad
+C_md_j \| \nabla\partial_t\widetilde \phi \|_{L^2(Q_T)}
\end{align*}
The last two inequalities imply
\begin{align}\label{etQjdj}
\|e_t\|_{L^2(Q_j)}
+ d_j^{-1}\|\nabla e\|_{L^2(Q_j)}
\leq
C_m\big(I_j(\phi_{h}(0))+X_j(\phi)
+H_j(e)+d_j^{-2}\|e\|_{L^2(Q_j''')}\big) .
\end{align}
We have proved that
any $e=\phi_h-\phi$ satisfying \refe{p32}
also satisfies \refe{etQjdj}.
Since
$e=\phi_h-\Pi_h\phi-(\phi-\Pi_h\phi)$
and $\phi(0)-\Pi_h\phi(0)=0$ in $\Omega_j'''$,
we can replace $\phi_h$ by $\phi_h-\Pi_h\phi$
and $\phi$ by $\phi-\Pi_h\phi$
in \refe{etQjdj}. Then \refe{3-1} follows immediately.
\,\vrule height8pt width 5pt depth 0pt
\section{Proof of Lemma \ref{lemma2-1}}
\setcounter{equation}{0}
Now we turn back to the proof of Lemma \ref{lemma2-1}.
\subsection{
The proof of \refe{DxtGrEst}}\label{EGF}
In this subsection, we present several local energy estimates for the Green's
function, the regularized Green's function and the discrete Green's
function, which then are used to prove \refe{DxtGrEst}.
These energy estimates
will also be used to prove \refe{FFEst2} in the next subsection.
In this subsection we let $T=1$
and fix $x_0\in\Omega$. We write
$G$ and $\Gamma$ as abbreviations for
the functions $G(\cdot,\cdot,x_0)$ and
$\Gamma(\cdot,\cdot,x_0)$, respectively,
when there is no ambiguity.
We use the decomposition of Section \ref{SecGF}
for all $j\geq 1$ (not restricted to $j\leq J_*$)
and so we do not
require $h< R_0K_0^{-2}/(16C_*)$ in this subsection.
\begin{lemma}\label{GFEst1}
{\it
For the Green's functions $\Gamma, G$ and $\Gamma_h$ defined in
\refe{GFdef}-\refe{GMhFdef}, we have the following estimates:
\begin{align}
&\sum_{l,k=0}^2d_j^{2l+k-1+N/2}
(\|\nabla^k\partial_{t}^lG(\cdot,\cdot,x_0)\|_{L^2(Q_j(x_0))}
+\|\nabla^k\partial_{t}^l\Gamma(\cdot,\cdot,x_0)\|_{L^2(Q_j(x_0))})
\leq C, \label{GFest01}\\
&\|\nabla^2G(\cdot,\cdot,x_0) \|_{L^{\infty,2}(\cup_{k\leq
j}Q_k(x_0))}\leq Cd_j^{-N/2-2} \label{GFest03}
,\\
&\|
\nabla_{x_0}\partial_tG(\cdot,\cdot, x_0 )\|_{L^{\infty}(Q_j(x_0))}\leq Cd_j^{-N-3},
\label{DxtGL1}\\
&\|\partial_tG(\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))}
+\|t\partial_{tt}G(\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))} \nonumber\\
&\quad
+\|\partial_t\Gamma(\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))}
+\|t\partial_{tt}\Gamma(\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))}
\leq C ,\label{GFest0423}
\end{align}
}
\end{lemma}
\noindent{\it Proof.}$\,\,$
For the given $x_0$ and $j$, we define a coordinate transformation
$x-x_0=d_j\widetilde x$ and $t=d_j^2\widetilde t$, and
define $\widetilde
G(\widetilde t,\widetilde x):=G(t,x,x_0)$, $\widetilde
a(\widetilde x):=a(x)$.
Via the coordinates transformation, we assume that
the sets $Q_j$, $Q_j'$, $\Omega_j$,
$\Omega_j'$ and $\Omega$ are transformed to
$\widetilde Q_j$, $\widetilde Q_j'$, $\widetilde \Omega_j$,
$\widetilde \Omega_j'$ and $\widetilde \Omega$, respectively.
Let $0\leq \widetilde\omega_i(x,t)\leq 1$, $i=0,1,2,3$,
be smooth cut-off functions which vanishes outside
$\widetilde Q_j'$ and equals $1$ in $\widetilde Q_{j}$.
Moreover, $\widetilde\omega_{i}$ equals $1$
at the points where $\widetilde\omega_{i+1}\neq 0$,
and $|\nabla \widetilde\omega_i|\leq C$,
$|\partial_t\widetilde\omega_i|\le C$ for $i=0,1,2,3$.
Since $ \cup_{k\geq j}\widetilde\Omega_k'
\cup \widetilde\Omega_*$ is of unit size,
there exists a convex domain
$\widetilde D=B_{\rho}(z)\cap \widetilde\Omega
\supset \cup_{k\geq j}\widetilde\Omega_k'
\cup \widetilde\Omega_*$,
with $4\leq \rho\leq 8K_0^2$, which belongs to one
of the following cases
(there are only a finite
number of shapes for $\widetilde D$):
(i) $z\in \widetilde\Omega$, $\rho=4$, and
$B_4(z)$ has no intersection with the boundary of
$ \widetilde\Omega$, thus $B_\rho(z)\cap\widetilde\Omega=B_\rho(z)$,
(ii) $z$ is on a face of $ \widetilde\Omega$, $\rho=8$
and $B_8(z)$ has no intersection with other faces of $\widetilde\Omega$,
thus $B_\rho(z)\cap\Omega$ is a half ball,
(iii) $z$ is on an edge of $\widetilde\Omega$,
$\rho=8K_0$ and $B_{8K_0}(z)$
has no intersection with any closed faces of $\widetilde\Omega$
which do not contain this edge,
thus $B_\rho(z)\cap\Omega$ is the intersection of a ball with
a sector spanned by the edge,
(iv) $z$ is a corner of $ \widetilde\Omega$ and $\rho=8K_0^2<R_0$,
and $B_\rho(z)\cap \widetilde\Omega$ coincides with the intersection
of the ball $B_\rho(z)$ with the cone spanned by the corner $z$.
Note that $\widetilde D\times(0,16)$ contains $\widetilde Q_j'$,
and consider $\widetilde\omega_i\widetilde G$, $i=1,2$,
which are solutions of
\begin{align}\label{omegaG1}
\partial_{\widetilde t}(\widetilde\omega_1\widetilde G)
-\nabla_{\widetilde x} \cdot(\widetilde a
\nabla_{\widetilde x} (\widetilde\omega_1\widetilde G))
= \widetilde\omega_0\widetilde G\partial_{\widetilde t}\widetilde\omega_1
+ \widetilde\omega_0\widetilde G\nabla_{\widetilde x}\cdot\big(\widetilde
a\nabla_{\widetilde x} \widetilde\omega_1\big)
-\nabla_{\widetilde x} \cdot(2\widetilde
a \widetilde\omega_0\widetilde G\nabla_{\widetilde x} \widetilde\omega_1)
\end{align}
and
\begin{align}\label{omegaG2}
\partial_{\widetilde t}(\widetilde\omega_2\widetilde G)
-\nabla_{\widetilde x} \cdot(\widetilde a
\nabla_{\widetilde x} (\widetilde\omega_2\widetilde G))
= \widetilde\omega_1\widetilde
G\partial_{\widetilde t}\widetilde\omega_2
+ \widetilde\omega_1\widetilde
G \nabla_{\widetilde x}\cdot\big(\widetilde
a\nabla_{\widetilde x} \widetilde\omega_2\big)
-\nabla_{\widetilde x} \cdot(2\widetilde
a \widetilde\omega_1\widetilde
G\nabla_{\widetilde x} \widetilde\omega_2)
\end{align}
in the domain $\widetilde D\times(0,16)$, respectively,
both with zero boundary/initial conditions.
Since $\widetilde D$ is a convex domain,
for $p=N+\alpha$ and $p_0=Np/(N+p)$ so that
$W^{1,p}(\widetilde D)\hookrightarrow L^\infty(\widetilde D)$ and
$W^{1,p_0}(\widetilde D)\hookrightarrow L^p(\widetilde D)$,
the standard $L^p((0,16);W^{1,p}(\widetilde D))$ estimate
of \refe{omegaG1} (the inequality \refe{LpW2quf}-\refe{LpW1quf} with $p=q$,
Lemma \ref{LemMaxLp}) gives
\begin{align*}
&\|\nabla_{\widetilde x}(\widetilde\omega_1
\widetilde G)\|_{L^{p}((0,16);L^p(\widetilde D))} \\
& \leq
C\|\widetilde\omega_0\widetilde G \|_{L^{p}((0,16);L^{p_0}(\widetilde D))}
+C\|\widetilde\omega_0\widetilde G \nabla_{\widetilde x}\widetilde a \|_{L^{p}((0,16);L^{p_0}(\widetilde D))}
+C\|\widetilde\omega_0\widetilde G \|_{L^{p}((0,16);L^p(\widetilde D))} \\
&\leq
C\|\widetilde\omega_0\widetilde G \|_{L^{p}((0,16);L^p(\widetilde D))}
(C+C\|\nabla_{\widetilde x}\widetilde a \|_{L^N(\widetilde D)}) \\
&\leq
C\|\widetilde\omega_0\widetilde G \|_{L^{p}((0,16);L^p(\widetilde D))}
,
\end{align*}
and the maximal $L^p$ regularity of \refe{omegaG2} yields that
(see inequality \refe{LpW2quf},
Lemma \ref{LemMaxLp})
\begin{align*}
&\|\partial_{\widetilde t}(\widetilde\omega_2\widetilde G ))
\|_{L^{p}((0,16);L^p(\widetilde D))} +\|\nabla_{\widetilde x} \cdot(\widetilde a
\nabla_{\widetilde x} (\widetilde\omega_2\widetilde G ))
\|_{L^{p}((0,16);L^p(\widetilde D))} \\
&
\leq
C\|\widetilde \omega_1\widetilde G\|_{L^{p}((0,16);L^p(\widetilde D))}
+C\|\widetilde \omega_1\widetilde G\nabla_{\widetilde x}\widetilde a\|_{L^{p}((0,16);L^p(\widetilde D))} \\
&\quad
+C\|\nabla_{\widetilde x}(\widetilde \omega_1\widetilde G )\|_{L^{p}((0,16);L^p(\widetilde D))} \\
&\leq
C\|\widetilde \omega_1\widetilde G\|_{L^{p}((0,16);L^{\infty}(\widetilde D))}
(C+C\|\nabla_{\widetilde x}\widetilde a\|_{L^{p}(\widetilde D)} )
+C\|\nabla_{\widetilde x}(\widetilde \omega_1\widetilde G )\|_{L^{p}((0,16);L^p(\widetilde D))} \\
&\leq
C\|\nabla_{\widetilde x}(\widetilde \omega_1\widetilde G )\|_{L^{p}((0,16);L^p(\widetilde D))} .
\end{align*}
By using \refe{H2Reg0}-\refe{W1inftyReg}, we have
\begin{align*}
&\|\nabla_{\widetilde x}(\widetilde\omega_2\widetilde G )
\|_{L^{\infty}(\widetilde D)}
+\sum_{k=0}^2
\|\nabla_{\widetilde x}^k(\widetilde\omega_2\widetilde G)\|_{L^{2}(\widetilde D)}\leq
C\|\nabla_{\widetilde x} \cdot(\widetilde a
\nabla_{\widetilde x} (\widetilde\omega_2\widetilde G ))
\|_{L^{p}(\widetilde D)} .
\end{align*}
The last three inequalities imply that
\begin{align*}
&\|\nabla_{\widetilde x}(\widetilde\omega_2\widetilde G )
\|_{L^2((0,16);L^{\infty}(\widetilde D))}
+\sum_{k=0}^2\|\nabla_{\widetilde x}^k
(\widetilde\omega_2\widetilde G)\|_{L^2((0,16);L^2(\widetilde D))} \nonumber\\
&\quad
+\|\partial_{\widetilde t}(\widetilde\omega_2\widetilde G )
\|_{L^{N+\alpha}((0,16);L^{N+\alpha}(\widetilde D))} \\
&\leq
C\|\widetilde\omega_0\widetilde G \|_{L^{N+\alpha}((0,16);L^{N+\alpha}(\widetilde D))} .
\end{align*}
Similarly,
replacing $\widetilde G$ by $\partial_{\widetilde t}\widetilde G$
and $\partial_{\widetilde t}^2\widetilde G$
in the above estimates, respectively, one can derive that
\begin{align*}
&\sum_{l=0}^2\|\nabla_{\widetilde x}(\widetilde\omega_3
\partial_{\widetilde t}^l\widetilde G )
\|_{L^2((0,16);L^{\infty}(\widetilde D))}
+\sum_{l,k=0}^2\|\nabla_{\widetilde x}^k(
\widetilde\omega_3\partial_{\widetilde t}^l
\widetilde G)\|_{L^{2}((0,16);L^2(\widetilde D))} \\
&\leq
C\|\widetilde\omega_0\widetilde G \|_{L^{N+\alpha}((0,16);L^{N+\alpha}(\widetilde D))}
\leq
C\|\widetilde\omega_0\widetilde G \|_{L^{\infty}((0,16);L^\infty(\widetilde D))} .
\end{align*}
Since $\widetilde\omega_3\widetilde G\equiv 0$ at $t=0$,
it follows that
\begin{align*}
\|\nabla_{\widetilde x}\partial_{\widetilde t}(\widetilde\omega_3
\widetilde G)\|_{L^\infty((0,16);L^\infty(\widetilde D))}
&\leq C
\|\nabla_{\widetilde x}\partial_{\widetilde t}^2
(\widetilde\omega_3\widetilde G)\|_{L^2((0,16);L^\infty(\widetilde D))}\\
&
\leq C\|\widetilde\omega_0\widetilde G \|_{L^{\infty}((0,16);L^\infty(\widetilde D)))} ,
\end{align*}
and
\begin{align*}
\sum_{k=0}^2\|\nabla_{\widetilde x}^k
(\widetilde\omega_3\widetilde G)\|_{L^\infty((0,16);L^2(\widetilde D))}
&\leq C\sum_{k=0}^2\|\partial_{\widetilde t} \nabla_{\widetilde x}^k
(\widetilde\omega_3\widetilde G)\|_{L^2((0,16);L^2(\widetilde D))} \\
&\leq C\|\widetilde\omega_0\widetilde G \|_{L^{\infty}((0,16);L^\infty(\widetilde D))}
\, .
\end{align*}
Moreover, from the last three inequalities, we have
\begin{align*}
&\|\nabla_{\widetilde x} \partial_{\widetilde t}
\widetilde G \|_{L^{\infty}(\widetilde Q_j)}
+\sum_{k=0}^2\|\nabla_{\widetilde x}^k
\widetilde G\|_{L^{\infty,2}(\widetilde Q_j)}
+\sum_{l,k=0}^2\|\partial_{\widetilde t}^l
\nabla_{\widetilde x}^k\widetilde G\|_{L^2(\widetilde Q_j)}
\leq C\|\widetilde G\|_{L^{\infty}(\widetilde Q_j')} .
\end{align*}
Transforming back to the $(x,t)$ coordinates,
we see from the last two inequalities that
\begin{align*}
&d_j^{3}\|\nabla\partial_t G \|_{L^{\infty}(Q_j)}
+\sum_{k=0}^2 d_j^{k-N/2}\| \nabla^k G\|_{L^{\infty,2}(Q_j)}
+\sum_{l,k=0}^2 d_j^{2l+k-1-N/2}\|\partial_t^l\nabla^k G\|_{L^2(Q_j)} \\
&
\leq C\| G \|_{L^{\infty}(Q_j')}\leq Cd_j^{-N} ,
\end{align*}
where we have used \refe{FEstP} in the last inequality.
By the symmetry of $G$ with respect to $x$ and $x_0$
we also get
$$
d_j^{3}\|\nabla_{x_0}\partial_t G (\cdot,\cdot,x_0)\|_{L^{\infty}(Q_j)}
\leq C\| G \|_{L^{\infty}(Q_j')}\leq Cd_j^{-N} .
$$
This proves \refe{GFest01}-\refe{DxtGL1}
for $G$.
By using the expression
\begin{align}
\Gamma(x,t\, ;x_0)
=\int_\Omega G(x,t\, ;y)\widetilde\delta(y,x_0){\rm d} y .
\end{align}
one can derive the same estimates for $\Gamma$:
\begin{align*}
\|\nabla^k\partial_t^l\Gamma(\cdot,\cdot \, ; x_0)\|_{L^2(Q_j)}
&\leq\int_\Omega \|\nabla^k\partial_t^lG(\cdot,\cdot \, ;y)\|_{L^2(Q_j)}
|\widetilde\delta(y,x_0)|{\rm d} y \\
&\leq \int_\Omega Cd_j^{-2l-k+1-N/2}|\widetilde\delta(y,x_0)|{\rm d} y
\leq Cd_j^{-2l-k+1-N/2} .
\end{align*}
Finally, we note that
\begin{align} \label{Gammatt}
&\|\partial_t\Gamma (t,\cdot,x_0)\|_{L^\infty}
+t\|\partial_{tt}\Gamma (t,\cdot,x_0)\|_{L^\infty} \nonumber \\
&\leq C\|\partial_t\Gamma (t,\cdot,x_0)\|_{H^2}
+Ct\|\partial_{tt}\Gamma (t,\cdot,x_0)\|_{H^2} \nonumber \\
&\leq C\|\partial_tA\Gamma (t,\cdot,x_0)\|_{L^2}
+Ct\|\partial_{tt}A \Gamma (t,\cdot,x_0)\|_{L^2}
&&\mbox{[$H^2$ estimate, Lemma \ref{LemMaxLp}]} \nonumber \\
&= C\|\partial_{tt}\Gamma (t,\cdot,x_0)\|_{L^2}
+Ct\|\partial_{ttt}\Gamma (t,\cdot,x_0)\|_{L^2} \nonumber \\
&\leq Ct^{-2}\|\Gamma (t/2,\cdot,x_0)\|_{L^2}
&&\mbox{[semigroup estimate]} \nonumber \\
&\leq Ct^{-2-N/2} ,
\end{align}
which implies the first part of (\ref{GFest0423}) and
the second part of (\ref{GFest0423})
(the estimates of $G$) can be proved in the same way.
The proof of Lemma \ref{GFEst1} is completed.
\,\vrule height8pt width 5pt depth 0pt \medskip
When $\max(t^{1/2},|x-x_0|)<d_1$ the inequality
\refe{DxtGrEst} follows from \refe{DxtGL1}.
When $\max(t^{1/2},|x-x_0|)\geq d_1$,
the estimate
$\|\nabla_{x_0}\partial_tG(\cdot,\cdot, x_0 )\|_{L^{\infty}(Q_j(x_0))}\leq C$
can be proved directly (without using the scale transformation)
in the same way as above.
\subsection{Proof of \refe{FFEst2}}
\label{dka7}
The proof is also based on Lemma \ref{GFEst1}.
First we consider the case $h < h_0:=(R_0K_0^{-2}/(16C_*)$ and let $T=1$.
The basic energy estimates of the equations
\refe{GMFdef}-\refe{GMhFdef} yield
\begin{align*}
\|\partial_t\Gamma_h\|_{L^2(Q_T)}
+\|\partial_t\Gamma\|_{L^2(Q_T)}
&\leq \|\nabla\Gamma_h(0)\|_{L^2(\Omega)}
+\|\nabla\Gamma(0)\|_{L^2(\Omega)} \\
&
=\|\nabla P_h\widetilde\delta_{x_0}\|_{L^2(\Omega)}
+\|\nabla \widetilde\delta_{x_0}\|_{L^2(\Omega)}
\leq Ch^{-1-N/2}
\end{align*}
and
\begin{align*}
\|\partial_{tt}\Gamma_h\|_{L^2(Q_T)}
+\|\partial_{tt}\Gamma\|_{L^2(Q_T)}
&\leq \|\nabla\partial_t\Gamma_h(0)\|_{L^2(\Omega)}
+\|\nabla\partial_t\Gamma(0)\|_{L^2(\Omega)} \\
&=\|\nabla A_hP_h\widetilde\delta_{x_0}\|_{L^2(\Omega)}
+\|\nabla A\widetilde\delta_{x_0}\|_{L^2(\Omega)}
\leq Ch^{-3-N/2} ,
\end{align*}
which imply
\begin{align*}
\|\partial_tF\|_{L^2(Q_*)} +\|t\partial_{tt}F\|_{L^2(Q_*)}
&\leq Ch^{-1-N/2}+Cd_{J_*}^2h^{-3-N/2} \\
&\leq Ch^{-1-N/2}+CC_*^2h^{-1-N/2} .
\end{align*}
Hence we have
\begin{align}\label{Bd31K}
&\|\partial_tF\|_{L^1( Q_T)}+\|t\partial_{tt}F\|_{L^1( Q_T)} \nonumber\\
&\leq Cd_{J_*}^{1+N/2}\big(\|\partial_tF\|_{L^2(Q_*)} +\|t\partial_{tt}F\|_{L^2(Q_*)}\big) \nonumber\\
&\quad +\sum_{j}Cd_j^{1+N/2}
\big(\|\partial_tF\|_{L^2(Q_j)} +\|t\partial_{tt}F\|_{L^2(Q_j)}\big) \nonumber\\
&\leq CC_*^{3+N/2}
+\sum_{j}Cd_j^{1+N/2}\big(\|\partial_tF\|_{L^2(Q_j)} +\|t\partial_{tt}F\|_{L^2(Q_j)}\big) \nonumber\\
&\leq CC_*^{3+N/2}+C{\mathcal K} ,
\end{align}
where
\begin{align}\label{KdKj}
{\mathcal K}:=\sum_{j}d_j^{1+N/2}(
d_{j}^{-1}\|\nabla F\|_{L^2(Q_j)}
+\|\partial_tF\|_{L^2(Q_j)}
+d_{j}^{2}\|\partial_{tt}F\|_{L^2(Q_j)}) .
\end{align}
We proceed to estimate ${\mathcal K}$.
We set $e=F$ ($\phi_h=\Gamma_h$ and $\phi=\Gamma$)
and $e=\partial_tF$ ($\phi_h=\partial_t\Gamma_h$ and $\phi=\partial_t\Gamma$)
in \refe{p32} (Proposition \ref{LocEEst}), respectively,
and note that $\Gamma(0)=\partial_t\Gamma(0)=0$ on $\Omega_j'$.
We obtain that
\begin{align}
&d_{j}^{-1}\|\nabla F\|_{L^2(Q_j)}+\|\partial_tF\|_{L^2(Q_j)}
+d_{j}^{2}\|\partial_{tt}F\|_{L^2(Q_j)} \nonumber\\
&\leq
C(\widehat{I_j}+\widehat{X_j}+\widehat{H_j}+ d^{-2}_j\| F \|_{L^2(Q_j')})
\end{align}
where
\begin{align*}
&\widehat{I_j}=\|\nabla P_h\widetilde\delta_{x_0}\|_{L^2(\Omega_j')}
+d_j^{-1}\|P_h\widetilde\delta_{x_0}\|_{L^2(\Omega_j')} \\
&\qquad
+d_{j}^{2}\|\nabla A_hP_h\widetilde\delta_{x_0}\|_{L^2(\Omega_j')}
+d_j \|A_hP_h\widetilde\delta_{x_0}\|_{L^2(\Omega_j')},
\\
&\widehat{X_j}=
d_j\|\nabla\partial_t(\Pi_h\Gamma-\Gamma) \|_{L^2(Q_j')}
+\|\partial_t(\Pi_h\Gamma-\Gamma)\|_{L^2(Q_j')}
+d_j^{-1}\|\nabla(\Pi_h\Gamma-\Gamma)\|_{L^2(Q_j')}\\
&\qquad
+ d_j^{-2}\|\Pi_h\Gamma-\Gamma\|_{L^2(Q_j')} +
d_j^3\| \nabla\partial_{tt}(\Pi_h\Gamma-\Gamma) \|_{L^2(Q_j')}
+d_j^2\|\partial_{tt}(\Pi_h\Gamma-\Gamma)\|_{L^2(Q_j')} \\
&\widehat{H_j}=\big(h/d_j\big)^m\big( \|\partial_{t}F\|_{L^2(Q_j')}
+d_j^{-1}\|\nabla F\|_{L^2(Q_j')}
+d_j^{2}\|\partial_{tt}F \|_{L^2(Q_j')}
+d_j \|\nabla\partial_{t}F\|_{L^2(Q_j')} \big)
\, .
\end{align*}
By noting the exponential decay of
$ P_h\widetilde \delta_{x_0}(y) $ (see (P2) in section \ref{Sec2-2})
we derive
\begin{align*}
&\widehat{I_j} \leq Chd_{j}^{-2-N/2},\\
&\widehat{X_j}
\leq (d_j h+ h^2)\|\nabla^2\partial_{t}\Gamma\|_{L^2(Q_j'')}
+(d_j^{-1}h+ d_j^{-2}h^2)\|\nabla^2\Gamma\|_{L^2(Q_j'')} \\
&\quad
+( d_j^3 h+ d_j^{ 2}h^2)\|\nabla^2\partial_{tt}\Gamma\|_{L^2(Q_j'')} \\
&~
\leq C hd_j^{-2-N/2} ,
\quad
\mbox{[by using Lemma \ref{GFEst1}]} \\
&\widehat{H_j}
\leq \big(h/d_j\big)^m \big( \|\partial_{t}F\|_{L^2(Q_T)}
+d_j^{-1}\|\nabla F\|_{L^2(Q_T)}
+d_j^{2}\|\partial_{tt}F \|_{L^2(Q_T)}
+d_j \|\nabla\partial_{t}F\|_{L^2(Q_T)} \big) \\
&~
\leq C\big(h/d_j\big)^m \big(\|P_h\widetilde\delta_{x_0} \|_{H^1}
+d_j^{-1}\|P_h\widetilde\delta_{x_0} \|_{L^2}
+d_j^{2}\|A_hP_h\widetilde\delta_{x_0} \|_{H^1}
+d_j \|A_hP_h\widetilde\delta_{x_0} \|_{L^2}\big)
\\
&~
\leq C\big(h/d_j\big)^m \big(h^{-1-N/2}
+d_j^{-1}h^{-N/2}+d_j^{2}h^{-3-N/2}
+d_j h^{-2-N/2}\big)
\quad \mbox{[by (P3)-(P4)]}\\
&~
\leq Chd_j^{-2-N/2},\quad \mbox{[by choosing $m=4+N/2$]}
\end{align*}
Therefore, by \refe{KdKj},
\begin{align}\label{dlj6}
{\mathcal K}
&\leq \sum_{j} Chd_j^{-1}+\sum_{j}Cd_j^{-1+N/2}\|F\|_{L^2(Q_j')} \nonumber\\
&\leq C+C\sum_{j}d_j^{-1+N/2}\|F\|_{L^2(Q_j')} .
\end{align}
To estimate $\|F\|_{L^2(Q_j')}$, we apply a duality argument.
Let $w$ be the solution of the backward parabolic equation
\begin{align*}
\left\{\begin{array}{ll}
-\partial_tw+Aw=v &\mbox{in}\,\,\,\Omega ,\\
w=0 &\mbox{on}\,\,\,\partial\Omega,\\
w(T)=0 &\mbox{in}\,\,\,\Omega ,
\end{array}\right.
\end{align*}
where $v$ is a function which is supported on $Q_j'$ and
$\|v\|_{L^2(Q_j')}=1$. Multiplying the above equation by $F$,
with integration by parts we get
\begin{align}\label{dka6}
\iint_{ Q_T} Fv dx dt
=(F(0),w(0)) + \iint_{ Q_T} \partial_tF w dx dt
+\sum_{i,j=1}^N \iint_{ Q_T} a_{ij}\partial_j F \partial_i w dx dt ,
\end{align}
where
\begin{align*}
(F(0),w(0))&=(P_h\widetilde\delta_{x_0}-\widetilde\delta_{x_0},w(0))\\
&=(P_h\widetilde\delta_{x_0}-\widetilde\delta_{x_0},w(0)-I_hw(0))\\
&=
(P_h\widetilde\delta_{x_0},w(0)-I_hw(0))_{\Omega_j''}
+(P_h\widetilde\delta_{x_0}-\widetilde\delta_{x_0},
w(0)-I_hw(0))_{(\Omega_j'')^c}\\
&:={\mathcal I}_1+{\mathcal I}_2 .
\end{align*}
By using the exponential decay of $P_h\widetilde\delta_{x_0}$
(see (P4) of section 2) and the local approximation
property (see (P2) of section 2), we derive that
\begin{align}
&|{\mathcal I}_1|\leq
Ch\|P_h\widetilde\delta_{x_0}\|_{L^2(\Omega_j'')}
(d_j^{-1}\|w(0)\|_{L^2(\Omega)} + \|\nabla w(0)\|_{L^2(\Omega)} )
\nonumber\\
&\quad\,\,\, \leq
Ch^{-N/2+1}e^{-Cd_j/h} (d_j^{-1}\|v\|_{L^{2(N+2)/(N+4)}(Q_j')} +\|v\|_{L^2(Q_j')})
\nonumber\\
&\quad\,\,\, \leq
Ch^{-N/2+1}e^{-Cd_j/h}\|v\|_{L^2(Q_j')}
\nonumber\\
&\quad\,\,\, \leq
C(d_j/h)^{1+N/2}e^{-Cd_j/h}h^{2}d_j^{-1-N/2}
\nonumber\\
&\quad\,\,\, \leq
Ch^{2}d_j^{-1-N/2 }, \label{SF2}\\[5pt]
&|{\mathcal I}_2|\leq C\| \widetilde\delta_{x_0}\|_{L^{2} }
\|w(0)-I_hw(0)\|_{L^{2}((\Omega_j'')^c)}
\nonumber\\
&\quad\,\,\, \leq
Ch^{2-N/2}\sum_{k=0}^2 d_j^{k-2}\|\nabla^k w(0)\|_{L^{2}((\Omega_j'')^c)} .
\label{SF22}
\end{align}
To estimate $\|\nabla^kw(0)\|_{L^{2}((\Omega_j'')^c)}$,
we let $W_j$ be a set containing $(\Omega_j'')^c$ but its
distance to $\Omega_j'$ is larger than $d_j/8$. Since
$$
\nabla^k_xw(x,0)
=\int_0^{T}\int_{\Omega}
\nabla^k_xG(s,x,y)v(y,s){\rm d} y{\rm d} s ,
$$
by noting the fact
$$
|x-y|+s^{1/2}\geq d_j /8
\quad\mbox{for $x\in W_j$ and $(y,s)\in Q_j'$}
$$
and using (\ref{GFest03}), we
further derive
\begin{align}
&\sum_{k=0}^2d_j^{k-2}\|\nabla^kw(\cdot,0)\|_{L^{2}(W_j)} \nonumber\\
&\leq C\sum_{k=0}^2d_j^{k-2} \sup_{y\in\Omega}\|\nabla^k
G(\cdot,\cdot,y)\|_{L^{\infty,2}(\bigcup_{k\leq
j-3}Q_k(y))}\|v\|_{L^{1}(Q_j')}\nonumber\\
&\leq \sum_{k=0}^2d_j^{k-2} d_j^{-N -k+N/2}\|v\|_{L^{1}(Q_j')}
\nonumber\\
&\leq \sum_{k=0}^2d_j^{k-2} d_j^{-N -k+N/2}d_j^{N/2+1}\|v\|_{L^2(Q_j')}
=C d_j^{ -1 } . \label{SF3}
\end{align}
From (\ref{SF2})-(\ref{SF3}), we see that the
first term on the right-hand side of \refe{dka6} is bounded by
\begin{align}
|(F(0),w(0))| \leq Ch^{2}d_j^{-N/2 -1}+Ch^{2-N/2}d_j^{ -1}
\leq Ch^{2-N/2}d_j^{-1} ,
\label{f69}
\end{align}
and the rest terms are bounded by
\begin{align}\label{sd80}
\iint_{ Q_T} \partial_tF w dx dt
& +\sum_{i,j=1}^N \iint_{ Q_T} a_{ij}\partial_j F \partial_iw dx dt
\nonumber \\
&= \iint_{ Q_T} \partial_tF (w-\Pi_hw) dx dt
+\sum_{i,j=1}^N \iint_{ Q_T} a_{ij}\partial_j F \partial_i (w-\Pi_hw)dx dt
\nonumber\\
&\leq
\sum_{*,i} C\|\nabla^2w\|_{L^2(Q_i')} (h^2\|\partial_tF\|_{L^2(Q_i)}
+ h\|\nabla F\|_{L^2(Q_i)} )
.
\end{align}
Moreover, to estimate $\|\nabla^2w\|_{L^2(Q_i')}$
we consider the expression
$$
\nabla^2_x w(x,t)
=\int_0^{T}\int_{\Omega}
\nabla^2_x G(s-t,x,y)v(y,s)1_{s>t}\,{\rm d} y{\rm d} s .
$$
For $i\leq j-3$ (so that $d_i>d_j$), we see that $w(x,t)=0$ for
$t>16d_j^2$ (because $v$ is supported in $Q_j'$); $d_i/2\leq |x-y|\leq 4 d_i$
and $s-t<d_i^2$ for $t< 16d_j^2$, $(x,t)\in Q_i$ and
$(y,s)\in Q_j'$.
Therefore, $(x,t)\in Q_i'(y)$ and we obtain
\begin{align*}
\|\nabla^2 w\|_{L^2(Q_i')}
&\leq \sup_{y}\|\nabla^2G(\cdot,\cdot,y)\|_{L^2(Q_i'(y))}
\|v\|_{L^1(Q_j')}\\
&\leq
Cd_i^{-N/2-1}d_j^{N/2+1}\|v\|_{L^2(Q_i')}\\
&\leq
C(d_j/d_i)^{N/2+1}
\leq \frac{Cd_j}{d_i} .
\end{align*}
For $i\geq j+3$ (so that $d_i< d_j$),
$\max(|s-t|^{1/2},|x-y|)\geq d_{j+1}$ for $(x,t)\in Q_i$ and therefore,
\begin{align*}
\|\nabla^2w\|_{L^2(Q_i')}
&\leq \ \sup_{y\in \Omega}\|\nabla^2G(\cdot,\cdot,y)
1_{\bigcup_{k\leq j+1}Q_{k}(y)}\|_{L^2(Q_i')}
\|v\|_{L^1(Q_j')}\\
&\leq Cd_i \sup_{y}\|\nabla^2G(\cdot,\cdot,y)
\|_{L^{\infty,2}(\bigcup_{k\leq j+1}Q_{k}(y))}
\|v\|_{L^2(Q_j')}d_j^{N/2+1} \\
&\leq C d_i d_j^{-N -2+N/2}d_j^{N/2+1}
=\frac{ C d_i}{d_j} .
\end{align*}
Finally for $|i-j|\leq 2$, applying the standard energy estimate leads to
$$\|w\|_{L^2((0,T);H^2)}\leq C\|v\|_{L^2(Q_T)}=C.$$
Combining the three cases, we have
\begin{align}\label{sd808}
\|\nabla^2w\|_{L^2(Q_i')}\leq C \min \big(d_i/d_j,d_j/d_i\big) :=C m_{ij}.
\end{align}
Substituting \refe{f69}-\refe{sd808} into (\ref{dka6}) gives the estimate
\begin{align}
\|F\|_{L^2(Q_j')}\leq Ch^2d_j^{-N/2-1}
+C\sum_{*,i}m_{ij}
(h^2\|\partial_tF\|_{L^2(Q_i)}
+h\|\nabla F\|_{L^2(Q_i)}) ,
\end{align}
which together with \refe{dlj6} implies
\begin{align*}
{\mathcal K}&\leq C+C\sum_j \left(\frac{h}{d_j}\right)^{2-N/2} +C\sum_j d_j^{N/2-1}
\sum_{*,i}m_{ij}\big(h^2\|\partial_tF\|_{L^2(Q_i)}
+h\|\nabla F\|_{L^2(Q_i)}\big)\\
&\leq
C +C\sum_{*,i}\left(h^2\|\partial_tF\|_{L^2(Q_i)}
+h\|\nabla F\|_{L^2(Q_i)} \right)\sum_jd_j^{N/2-1}m_{ij}\\
&\leq
C +C\sum_{*,i}\left(h^2\|\partial_tF\|_{L^2(Q_i)}
+h\|\nabla F\|_{L^2(Q_i)}\right) d_i^{N/2-1}\\
&\leq C +C\left(h^2\|\partial_tF\|_{L^2(Q_*)}
+h\|\nabla F\|_{L^2(Q_*)}\right)(C_* h )^{N/2-1} \\
&\quad +C\sum_id^{1+N/2}_i
\left(\|\partial_tF\|_{L^2(Q_i)}+d_i^{-1}\|\nabla F\|_{L^2(Q_i)}\right
)\left(\frac{h}{d_i}\right) \\
&\leq
C +CC_*^{-1 +N/2 }+C\sum_id^{1+N/2}_i\left(\|\partial_tF\|_{L^2(Q_i)}+d_i^{-1}\|\nabla F\|_{L^2(Q_i)}\right
)\left(\frac{h}{d_i}\right) \\
&\leq C_2+C_2 C^{-1 +N/2 }_* +C_2C^{-1}_*{\mathcal K}
\end{align*}
for some positive constant $C_2$.
By choosing
\begin{align}\label{Cstar}
C_*= \max(10,10\kappa,R_0K_0^{-2}/8) +2C_2,
\end{align}
the above inequality shows that ${\mathcal K} \le C$.
Returning to \refe{Bd31K}, the boundedness of ${\mathcal K}$
implies
\begin{align}\label{Bd31K2}
&\|\partial_tF\|_{L^1( Q_T)}+\|t\partial_{tt}F\|_{L^1( Q_T)}\leq C.
\end{align}
From \refe{KdKj} we also see that, the boundedness
of ${\mathcal K}$ implies
\begin{align*}
\|\partial_tF\|_{L^2( Q_j)}
+\|t\partial_{tt}F\|_{L^2(Q_j)}\leq Cd_j^{-1-N/2} .
\end{align*}
Since $\Omega\times(1/4,1)\subset \cup_{d_j\geq 1/2}Q_j$,
it follows that
\begin{align*}
&\|\partial_tF\|_{L^2(\Omega\times(1/4,1))}^2
+\|t\partial_{tt}F\|_{L^2(\Omega\times(1/4,1))}^2 \\
&\leq \sum_{d_j\geq 1/2}\big(\|\partial_tF\|_{L^2(Q_j)}^2
+\|t\partial_{tt}F\|_{L^2(Q_j)}^2\big)\\
&
\leq \sum_{d_j\geq 1/2} C d_j^{-2-N} \leq C .
\end{align*}
The above inequality and (\ref{Gammatt}) imply
$$
\|\partial_t\Gamma_h\|_{L^2(\Omega\times(1/4,1))}
+\|t\partial_{tt}\Gamma_h\|_{L^2(\Omega\times(1/4,1))} \leq C .
$$
Furthermore, differentiating the equation (\ref{GMhFdef}) with respect to
$t$ and multiplying the result by $\partial_t\Gamma_h$ give
\begin{align*}
&\frac{{\rm d}}{{\rm d} t}\|\partial_t\Gamma_h(t,\cdot,x_0)\|_{L^2}^2
+c_0\|\partial_t\Gamma_h(t,\cdot,x_0)\|_{L^2}^2 \\
&\leq \frac{{\rm d}}{{\rm d} t}\|\partial_t\Gamma_h(t,\cdot,x_0)\|_{L^2}^2
+(A_h\partial_t\Gamma_h(t,\cdot,x_0),\partial_t\Gamma_h(t,\cdot,x_0))
= 0 ,\quad\mbox{for}\quad t\geq 1,
\end{align*}
which further shows that
$$ \|\partial_t\Gamma_h(t,\cdot,x_0)\|_{L^2}^2\leq
Ce^{-c_0(t-1)}\|\partial_t\Gamma_h(\cdot,\cdot,x_0)\|_{L^2((1/4,1);L^2(\Omega))}^2
\leq Ce^{-c_0t} \quad\mbox{for}\,\,\, t\geq 1. $$
In a similar way one can derive
$\|\partial_{tt}\Gamma_h(t,\cdot,x_0)\|_{L^2} \leq Ce^{-c_0t}$
for $t\geq 1$.
These inequalities together with \refe{GFest0423} imply
\begin{align}
\|\partial_tF (\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))}
+\|t\partial_{tt}F (\cdot,\cdot,x_0)\|_{L^1(\Omega\times(1,\infty))}\leq C ,
\end{align}
which together with \refe{Bd31K2} leads to
(\ref{FFEst2}) for the case
$h<h_0:=R_0K_0^{-2}/(16C_*)$ with $C_*$ being given by \refe{Cstar}.
Secondly when $h\geq h_0$, the decomposition in subsection 2.3 is not needed and the energy estimates of
\refe{GMFdef}-\refe{GMhFdef} yield
\begin{align*}
&\|\partial_t\Gamma_h\|_{L^2(Q_T)}
+\|\partial_t\Gamma\|_{L^2(Q_T)}
\leq \|\nabla\Gamma_h(0)\|_{L^2(\Omega)}
+\|\nabla\Gamma(0)\|_{L^2(\Omega)} \leq Ch_0^{-1-N/2} \\
&\|\partial_{tt}\Gamma_h\|_{L^2(Q_T)}
+\|\partial_{tt}\Gamma\|_{L^2(Q_T)}
\leq \|\nabla\partial_t\Gamma_h(0)\|_{L^2(\Omega)}
+\|\nabla\partial_t\Gamma(0)\|_{L^2(\Omega)} \leq Ch_0^{-3-N/2} ,
\end{align*}
which imply
\begin{align}
&\int_0^1\int_\Omega\big( |\partial_tF(t,x,x_0) \big|
+\big|t\partial_{tt}F(t,x,x_0 ) |\big){\rm d} x{\rm d} t \leq C .
\end{align}
Since both $\|\partial_{t}\Gamma_h(t)\|_{L^2(\Omega)}
+\|\partial_{tt}\Gamma_h(t)\|_{L^2(\Omega)}$
and $\|\partial_{t}\Gamma(t)\|_{L^2(\Omega)}
+\|\partial_{tt}\Gamma(t)\|_{L^2(\Omega)}$ decay
exponentially as $t\rightarrow \infty$,
it follows that \refe{FFEst2} still holds
when $h\geq h_0$.
The proof of Lemma \ref{lemma2-1} is completed.
~\vrule height8pt width 5pt depth 0pt
\section{Conclusion}
In this paper we have proved that the
discrete elliptic operator $-A_h$
generates a bounded analytic semigroup
and has the maximal $L^p$ regularity, uniformly with respect to $h$,
in arbitrary convex polygons and polyhedra under the
regularity assumption $a_{ij}\in W^{1,N+\alpha}$.
We have assumed the quasi-uniformity of the triangulation, and
analysis of the problem under non-quasi-uniform triangulations
remains open. As far as we know,
only the analytic semigroup estimate \refe{STLEst}
and its equivalent resolvent estimate
were studied with an extra logarithmic factor
for some special cases of non-quasi-uniform triangulations, see \cite{CT01, Thomee07}.
The discrete maximal regularity estimates
(\ref{LpqSt3}) and (\ref{LpqSt2}) have not been established
with more general triangulations even
in smooth settings.
\section*{Appendix: The proof of Lemma 2.1}
\renewcommand{\thelemma}{A.\arabic{lemma}}
\renewcommand{\thetheorem}{A.\arabic{theorem}}
\renewcommand{\theequation}{A.\arabic{equation}}
\setcounter{equation}{0}
{\it Proof of \refe{H2Reg0}:}$\,\,$
The inequality \refe{H2Reg0} is similar to
Theorem 3.1.3.1 of \cite{Grisvard},
which was proved by using the local energy inequality of
Lemma 3.1.3.2, and the lemma was proved under the assumption
$a_{ij}\in W^{1,\infty}(\Omega)$, where $\Omega$ is a convex domain.
In the following, we show that
this assumption can be relaxed to
$a_{ij}\in W^{1,N+\alpha}(\Omega)
\hookrightarrow C^{\gamma}(\overline\Omega)$,
where $\gamma=\alpha/(N+\alpha)$
\begin{lemma}\label{LemLocH2}
{\it
If $\Omega$ is convex and $a_{ij}\in W^{1,N+\alpha}(\Omega)$, then
each point $y\in \Omega$ has a neighborhood
$B_R(y)\cap\Omega$ such that
\begin{align}
\|u\|_{H^2(\Omega)}
\leq C\|Au\|_{L^2(\Omega)}
+C\|u\|_{H^1(\Omega)} .
\label{A1}
\end{align}
for all $u\in H^2\cap H^1_0$ such that
the support of $u$ is contained in $B_R(y)\cap\overline\Omega$.
The radius $R$ depends only on the semi-norms
$|a_{ij}|_{W^{1,N+\alpha}(\Omega)}$
and $|a_{ij}|_{C^\gamma(\overline\Omega)}$.
}
\end{lemma}
\noindent{\it Proof.}$\,\,\,$
Following the proof of Lemma 3.1.3.2 in \cite{Grisvard}
(see page 143, (3.1.3.4) and the equality above (3.1.3.5)),
we have (using our notations)
\begin{align*}
\|u\|_{H^2(\Omega)}
&\leq C\|Au\|_{L^2(\Omega)}
+C\sum_{i,j=1}^N\max_{x\in V_y}|a_{ij}(y)-a_{ij}(x)|\|u\|_{H^2(\Omega)} \\
&\quad
+C\sum_{i,j=1}^N\|\partial_i a_{ij}\partial_j u\|_{L^2(\Omega)}\\
&\leq C\|Au\|_{L^2(\Omega)}
+CR^\beta\|u\|_{H^2(\Omega)} \\
&\quad
+C\sum_{i,j=1}^N\|\nabla a_{ij}\|_{L^{N+\alpha}(\Omega)}
\|\nabla u\|_{L^{2(N+\alpha)/(N-2+\alpha)}(\Omega)} .
\end{align*}
When $R$ is small enough we have
\begin{align*}
\|u\|_{H^2(\Omega)}
\leq C\|Au\|_{L^2(\Omega)}
+C\|\nabla u\|_{L^{2(N+\alpha)/(N-2+\alpha)}(\Omega)} .
\end{align*}
Since $H^2$ is compactly embedded into
$W^{1,2(N+\alpha)/(N-2+\alpha)}$
which is again embedded into $H^1$,
there exists $\theta_\alpha\in (0,1)$ such that
\begin{align*}
\|\nabla u\|_{L^{2(N+\alpha)/(N-2+\alpha)}(\Omega)}
&\leq \epsilon \|u\|_{H^2(\Omega)}
+ C_\epsilon \|u\|_{H^1(\Omega)} ,\quad\forall\,\epsilon\in(0,1).
\end{align*}
Choosing $\epsilon$ small enough, \refe{A1} follows from the last
the last two inequalities
This completes the proof of Lemma \ref{LemLocH2}. \,\vrule height8pt width 5pt depth 0pt
\medskip
Then \refe{H2Reg0} can be proved by using Lemma \ref{LemLocH2}
and a perturbation procedure (as mentioned in
the proof of \cite[Theorem 3.1.3.1]{Grisvard}).\\
\noindent{\it Proof of \refe{W1inftyReg}:}$\,\,$
Theorem 3.4 of [13] states that
if $\Omega\subset\mathbb{R}^N$ ($N\geq 3$) is convex
and the coefficients $a_{ij}$ are H\"older continuous (so that
(3.1)-(3.3) of [13] hold),
the Green's function of the elliptic operator
$A$
with the Dirichlet boundary condition
satisfies
\begin{align}
| \nabla_xG_e(x,y)|+| \nabla_yG_e(x,y)|
\leq \frac{C}{|x-y|^{N-1}} ,
\label{A2}
\end{align}
where we have used the symmetry
$G_e(x,y)=G_e(y,x)$.
Therefore, any $H^1_0$ solution of the equation
$-\nabla\cdot(a\nabla u)=f$ satisfies
\begin{align*}
|\nabla u(x)|=\bigg|\int_\Omega \nabla_xG_e(x,y)f(y){\rm d} y\bigg|
\leq \bigg\|\frac{C}{|x-y|^{N-1}}\bigg\|_{L^{p'}(\Omega)}
\|f\|_{L^p(\Omega)}
\leq C\|f\|_{L^p(\Omega)}
\end{align*}
for $p>N$. As pointed out in \cite{Fromm} (page 227,
the paragraph below Proposition 1),
the inequality \refe{A2} for $N=2$ can be proved
with some minor modifications
on the proof of \cite[Theorem 3.3--3.4]{GW}
since Theorem 3.3--3.4 of \cite{GW} only requires
$a_{ij}$ being
H\"older continuous coefficients.
\vskip0.1in
\noindent{\it Proof of \refe{LpW2quf}-\refe{LpW1quf}:}
Since $W^{1,N+\alpha}(\Omega)\hookrightarrow C(\overline\Omega)$,
Theorem 1 of \cite{JLW} implies that
the solution of the elliptic equation
\begin{align}
\left\{\begin{array}{ll}
A u = f
&\mbox{in}\,\,\,\Omega, \\
u=0 &\mbox{on}\,\,\,\partial\Omega,
\end{array}\right.
\end{align}
with continuous coefficients $a_{ij}$ in a
convex domain $\Omega\subset\mathbb{R}^N$ satisfies
\begin{align}\label{W1quf}
\|u\|_{W^{1,q}(\Omega)}\leq C\|f\|_{W^{-1,q}(\Omega)} ,
\quad\forall\, 1<q<\infty ,
\end{align}
where we have noted that a continuous function
is $(\delta,R)$ vanishing, and
a convex domain is $(\delta,\sigma,R)$-quasiconvex \cite{JLW}.
Since the solution of \refe{PDE0} with $f=0$
satisfies (integrating the equation against $|u|^{q-2}u$)
$$
\|u\|_{L^q(\Omega)} \leq \|u^0\|_{L^q(\Omega)} \, ,
$$
it follows that the semigroup generated by the
elliptic operator $A$ is a
contraction semigroup on $L^q(\Omega)$.
By Theorem 1, Section 2, Chapter 3 of
\cite{Stein}, the semigroup
$\{E(t)\}_{t>0}$ generated by the elliptic operator $A$
has an analytic continuation
(analyticity of the semigroup $\{E(t)\}_{t>0}$). Moreover,
by the maximum principle
we have $u^0\geq 0$ $\implies$ $u\geq 0$
(positivity of the semigroup $\{E(t)\}_{t>0}$) and then,
by Corollary 4.d of \cite{Weis2}, the
solution of \refe{PDE0} with $u^0=0$
has the maximal $L^p$ regularity
\refe{LpW2quf}.
In other words, the map from $f$ to $Au$ given by the
formula
$$
Au=\int_0^t AE(t-s)f(\cdot,s){\rm d} s
$$
is bounded in $L^p((0,T);L^q)$, for all $1<p,q<\infty$.
Since
$$
A^{1/2}u=\int_0^t A^{1/2}E(t-s)f(\cdot,s){\rm d} s
=\int_0^t AE(t-s) A^{-1/2}f(\cdot,s){\rm d} s ,
$$
it follows that
\begin{align}
\|A^{1/2}u\|_{L^p((0,T);L^{q})}
&\leq C_{p,q}\|A^{-1/2}f\|_{L^p((0,T);L^{q})} ,\quad\forall\,
1<p,q<\infty .
\end{align}
It remains to prove the boundedness of the Riesz transform
$\nabla A^{-1/2}$:
\begin{align}\label{RiezTrf}
\|\nabla A^{-1/2}f\|_{L^{q}}
&\leq C_{q}\|f\|_{L^{q}} ,\quad\forall\,
1<q<\infty .
\end{align}
Then the last two inequalities imply
\begin{align*}
\|\nabla u\|_{L^p((0,T);L^{q})}
&=\|\nabla A^{-1/2}(A^{1/2}u)\|_{L^p((0,T);L^{q})} \\
&\leq C_q\|A^{1/2}u\|_{L^p((0,T);L^{q})} \\
&\leq C_{p,q}\|A^{-1/2}f\|_{L^p((0,T);L^{q})} \\
&\leq C_{p,q}\|f\|_{W^{-1,q}} ,
&&\forall\, 1<p,q<\infty ,
\end{align*}
where the last step of the inequality above is due to the following
duality argument ($A^{-1/2}$ is self-adjoint):
\begin{align*}
(A^{-1/2}f,g) =(f, A^{-1/2}g)
\leq C\|f\|_{W^{-1,q}}\|\nabla A^{-1/2}g\|_{L^{q'}}
\leq C\|f\|_{W^{-1,q}}\|g\|_{L^{q'}} .
\end{align*}
It has been proved in \cite[Theorem B]{Shen05}
that the Riesz transform is bounded on $L^q(\Omega)$
(i.e. the inequality (\ref{RiezTrf}) holds) if and only if the solution of
the homogeneous equation
\begin{align}\label{HomoEqu}
A u = 0
\end{align}
satisfies the local estimate
\begin{align}\label{HomoEst0}
\bigg(\frac{1}{r^N}\int_{\Omega\cap B_r(x_0)}
|\nabla u|^q{\rm d} x\bigg)^{\frac{1}{q}}
\leq C
\bigg(\frac{1}{r^N}\int_{\Omega\cap B_{\sigma_0r}(x_0)}
|\nabla u|^2{\rm d} x\bigg)^{\frac{1}{2}}
\end{align}
for all $x_0\in\Omega$ and $0<r<r_0$,
where $r_0$ and $\sigma_0\geq 2$
are any given small positive constants
such that $\Omega\cap B_{\sigma_0r_0}(x_0)$ can be given
by the intersection of $B_{\sigma_0r_0}(x_0)$ with a Lipschitz graph.
It remains to prove \refe{HomoEst0}.
Let $\omega$ be a smooth cut-off function which equals
zero outside $B_{2r}:=B_{2r}(x_0)$
and equals 1 on $B_r$.
Extend $u$ to be zero on $B_{2r}\backslash\Omega$
and denote by $u_{2r}$ the average of $u$ over $B_{2r}$.
Then (\ref{HomoEqu}) implies
\begin{align}\label{HomoEqu2}
&\sum_{i,j=1}^N\partial_i(a_{ij}\partial_j (\omega (u-u_{2r}))) \nonumber \\
&=\sum_{i,j=1}^N\partial_i(a_{ij}(u-u_{2r})\partial_j\omega )
+\sum_{i,j=1}^N a_{ij}\partial_i\omega\partial_j (u-u_{2r})
\quad\mbox{in}\,\,\,\Omega,
\end{align}
and the $W^{1,q}$ estimate (\ref{W1quf})
implies
\begin{align*
\|\omega (u-u_{2r})\|_{W^{1,q}(\Omega)}
&\leq C\|(u-u_{2r})\partial_j\omega\|_{L^q(\Omega)}
+C\|\partial_i\omega\partial_j u\|_{W^{-1,q}(\Omega)} \\
&\leq C\|(u-u_{2r})\partial_j\omega\|_{L^q(\Omega)}
+C\|\partial_i\omega\partial_j u\|_{L^s(\Omega)} \\
&= C\|(u-u_{2r})\partial_j\omega\|_{L^q(B_{2r})}
+C\|\partial_i\omega\partial_j u\|_{L^s(B_{2r})} \\
&\leq Cr^{-1}\|\nabla u\|_{L^s(B_{2r})} ,
\end{align*}
where $s=qN/(q+N)<q$ satisfies
$L^s(\Omega)\hookrightarrow W^{-1,q}(\Omega)$
and $W^{1,s}(\Omega)\hookrightarrow L^q(\Omega)$.
The last inequality implies
\begin{align}\label{HomoEst2}
\|\nabla u\|_{L^q(\Omega\cap B_{r})}
&\leq Cr^{-1}\|\nabla u\|_{L^s((\Omega\cap B_{2r})} .
\end{align}
If $s\leq 2$ then one can derive
\begin{align*
\|\nabla u\|_{L^q(\Omega\cap B_{r})}
&\leq Cr^{N/q-N/2}\|\nabla u\|_{L^2(\Omega\cap B_{2r})} .
\end{align*}
by using one more H\"older's inequality on the right-hand side.
Otherwise,
one only needs a finite number of iterations of (\ref{HomoEst2})
to reduce $s$ to be less than $2$.
This completes the proof of (\ref{HomoEst0}).
The proof of \refe{LpW2quf}-\refe{LpW1quf} is complete.
\,\vrule height8pt width 5pt depth 0pt
\bigskip
\noindent{\bf Acknowledgement}~~
We would like to thank the anonymous referees
for many valuable comments and suggestions,
which are very helpful to improve
both the quality and presentation of this paper.
|
\section{Introduction}
High-frequency quasi-periodic oscillations (HFQPOs) have been observed in some
black hole (BH) X-ray binaries, which only appear in "steep power law" state
in high luminosity ($L>0.1L_{\rm{Edd}}$) and are in range of 40 to
$450\rm{Hz}$. These frequencies are comparable to the orbital frequency of
innermost stable circular orbit (ISCO) of a stellar-mass BH. It is believed
that the mechanism behind them is closely related to the dynamics of inner
regions of BH accretion disks
\citep[see][for reviews]{Remillard06,Kato_book,Belloni12}.
The pioneering work of \cite{Kato78} studied the effect of viscosity on the
disk stability and found the sonic-point instability (its original name is
pulsational instability), which is the excitation
process of so called $p$-mode oscillations in transonic accretion flows
(this oscillation has the other name, inertial-acoustic oscillation, but
we only call it $p$-mode oscillation in this paper). The
mechanism of this instability is analogous to the $\epsilon$-mechanism in
stellar pulsations. If the viscous parameter $\alpha$ of $\alpha$-prescription
\citep{SS73} is larger than a critical value, the instability will arise from
the phase relation between the viscous heat generation and oscillation,
i.e. the viscous heat generation increase (decrease) in the compressed
(expanded) phase of disk oscillation \citep{Kato78,Kato88}.
A series of works \citep{Matsumoto88,Matsumoto89,CT95,MHL14} were dedicated to
study the role of viscosity on HFQPOs. They all found the oscillation, whose
frequency is close to the maximum of epicyclic frequency $\kappa_{\rm{max}}$.
Additionally, \cite{MHL14} also found many overtones, whose frequencies are
close to the integer multiples of $\kappa_{\rm{max}}$.
Though the famous $\alpha$-prescription has been used extensively since 1973,
it may be too simple to closely accord with actual accretion flows. The
magnetohydrodynamic (MHD) shearing box simulations of \cite{Hirose09} implied
that there is certain time-delay between the viscous stress and total pressure
instead of the instantaneity introduced by $\alpha$-prescription. The other
work of \cite{Penna13}, which is based on a few relativistic MHD global
simulations, pointed out that the parameter $\alpha$ is a function of radius
but not constant in the inner disk region. More than one decade ago,
\cite{YK96} analyzed effects of viscous time-delay on the $p$-mode oscillations
and they found that it has a little of inhibition on the growth of
oscillations but excitation still exists for high viscosity cases.
Our work presented in this letter is also dedicated to study the role of
viscosity on QPOs. The distinction of our model is the evolutionary viscous
stress equation (instead of the $\alpha$-prescription), which is constructed
to mimic effects of viscous time-delay and inconstant $\alpha$ basing on the
results of \cite{Hirose09} and \cite{Penna13}.
\section{Equations}
In this letter, we consider the axisymmetric relativistic accretion flows
around black holes. The Boyer-Lindquist coordinates $t$, $r$, $\theta$, $\phi$
are used to describe the space-time. The governing equations of accretion
flows are written as following,
\begin{eqnarray}
&&\frac{\partial\Sigma}{\partial t} = -\frac{r\Delta^{1/2}}{\gamma A^{1/2}}
\left[\Sigma\frac{\partial u^t}{\partial t}+\frac{1}{r}\frac{\partial}{\partial
r}\left(r\Sigma\frac{V}{\sqrt{1-V^2}}\frac{\Delta^{1/2}}{r}\right)\right],
\label{continuity} \\
&&\frac{\partial V}{\partial t} = \frac{\sqrt{1-V^2}\Delta}{\gamma A^{1/2}}
\left[-\frac{V}{1-V^2}\frac{\partial V}{\partial r}+\frac{{\cal A}}{r}
-\frac{1-V^2}{\rho}\frac{\partial p}{\partial r}\right], \label{radial} \\
&&\frac{\partial\mathcal{L}}{\partial t} = -\frac{\Delta}{\gamma A^{1/2}}
\frac{V}{\sqrt{1-V^2}}\frac{\partial\mathcal{L}}{\partial r}
-\frac{\Delta^{1/2}}{\gamma A^{1/2}\Sigma}\frac{\partial}{\partial r}
\left(\frac{2\Delta}{r} S_{r\phi}\right),\label{angular}\\
&&\frac{\partial H}{\partial t} = -\frac{UH}{r} -\frac{1}{\gamma}
\frac{V}{\sqrt{1-V^2}}\frac{\partial H}{\partial r},\label{thickness}\\
&&\nonumber \frac{\partial U}{\partial t} = \frac{\Delta^{1/2}r}{\gamma^2
A^{1/2}H}{\cal R}-\frac{U}{\gamma^2}\left(\frac{V}{(1-V^2)^2}
\frac{\partial V}{\partial t}+\frac{{\cal L}r^2}{A}
\frac{\partial\cal L}{\partial t}\right)\\
&&~~~~~-\frac{U}{H}\frac{\partial H}{\partial t}, \label{vertical}\\
&&\nonumber \frac{\partial T}{\partial t} = \frac{1}{\Sigma}
\frac{r\Delta^{1/2}}{\gamma A^{1/2}}\left[\frac{F^+-F^-}{c_V}
+(\Gamma_3-1)T\Sigma\left(-\frac{\partial u^t}{\partial t}
-\frac{1}{r^2}\frac\partial{\partial r}(r^2u^r)\right)\right]\\
&&~~~~~-\frac{V\Delta}{\gamma\sqrt{1-V^2}A^{1/2}}
\frac{\partial T}{\partial r}.\label{energy}
\end{eqnarray}
Where $\Sigma$, $V$, $\mathcal{L}$, $H$, $U$, $T$, and $S_{r\phi}$ are the
surface density, radial velocity (measured in the corotating frame), angular
momentum per unit mass ($\mathcal{L}\equiv u_{\phi}$), half thickness of disk,
vertical velocity of the surface, local temperature of accreted gas, and
viscous stress (only $r\phi$-component is non-vanishing), respectively.
These equations are all derived from the conservation of the stress-energy
tensor, and almost the same as those in our previous paper \cite{PaperII}
(see that paper for the detailed derivations as well as definitions of
$\gamma$, $A$, $\Delta$, and etc).
As mentioned in previous section, we adopt an additional evolutionary stress
equation to describe the viscosity instead of the $\alpha$-prescription. This
equation can be written as
\begin{equation}
n\tau^*\frac{\partial S_{r\phi}}{\partial t}=S^*_{r \phi}-S_{r \phi},
\label{stress}
\end{equation}
where the factor $n\tau^*$ is the practical viscous time-delay, which is
scaled with the typical delay $\tau^*$ by parameter $n$; $S^*_{r\phi}$ is the
expected stress by turbulence. The definitions of $\tau^*$ and $S^*_{r\phi}$
can be written as
\begin{eqnarray}
&&\tau^*=-\left(\frac{\gamma^2 A\Omega}{r^4}
\frac{\partial\ln\Omega}{\partial\ln r}\right)^{-1}, \label{def-tau} \\
&&S^*_{r\phi}=-\frac{\nu\Sigma A^{3/2}\gamma^3}{2r^3\Delta^{1/2}}
\frac{\partial\Omega}{\partial r}, \label{def-Srphi*}
\end{eqnarray}
where
\begin{eqnarray}
&&\Omega=\frac{d\phi}{dt}, \\
&&\nu=\frac{2}{3}\alpha H \sqrt{\frac{p}{\rho}}, \label{nu}\\
&&\alpha=\alpha_0\left(\frac{1-2Mr^{-1}+a^2r^{-2}}{1-3Mr^{-1}+2aM^{1/2}r^{-3/2}}
\right)^6. \label{alpha-prof}
\end{eqnarray}
If $\alpha=\rm{const}$ and $n\rightarrow 0$, equation (\ref{stress}) will
reduce to $S_{r\phi}=S^*_{r\phi}$, which is the same as the
$\alpha$-prescription in \cite{PaperII}. It means that equation (\ref{stress})
contains the $\alpha$-prescription as a trivial case.
Equation (\ref{alpha-prof}) determines the dependence of $\alpha$ on radius
$r$, BH mass $M$ and spin $a$. The radial factor, including the exponent $6$,
was suggested by \cite{Penna13}. Under this profile, $\alpha$ is almost a constant
$\alpha_0$ in outer disk region with large $r$ and increases to higher value radially
inwards. We set $\alpha_0=0.1$ and fix the mass supplying rate $\dot{M}_{\rm{out}}=0.06\dot{M}_{\rm{Edd}}$
(see table \ref{tab1} for the definition of Eddington accretion rate, $\dot{M}_{\rm{Edd}}$) at the outer boundary
in this work. In practice, these settings are sufficient to make the disk thermally unstable and
we indeed observe the limit-cycle outburst from running code.
After defining $S_{r\phi}$, the viscous heating rate $F^+$ in equation
(\ref{energy}) is redefined as
\begin{equation}
F^+=-2\frac{\gamma A^{1/2}\Delta^{1/2}}{r^3}\frac{\partial\Omega}{\partial r}
\cdot S_{r\phi}, \label{vis_heating}
\end{equation}
which would be reduced to the one in \cite{PaperII} when $n\rightarrow 0$.
\section{Instabilities}
We update our previous code established in \cite{PaperII} and run it for
eleven numerical models, whose parameters are listed in table \ref{tab1}.
\citet{Lin11} and \citet{Ciesielski12} studied the impact of viscous
time-delay on the thermal instability and found that it is not remarkable for
small enough time-delay. Indeed, we observe the expected limit-cycle
outbursts on all models with the time delay parameters $n$ in the range $0$ to
$4$. Therefore, the time-delay implemented in our code does not affect the
limit-cycle outbursts.
Among these models, S8 and S30 are two typical ones. S8 is a disk around a
non-spinning black hole as well as S30 around a fast-spinning one. In figure
\ref{Fig:LightCurves}, we show the bolometric light curves for these two
models respectively. For each point on light-curves, the disk luminosity is
made by integrating the local radiation cooling fluxes, and the sampling time
corresponds to the frequency $10^4\rm{Hz}$, which is enough to reveal any
harmonic with frequency lower than $5\times 10^3\rm{Hz}$ in the power spectral
density (PSD). Due to the difficulties in hydrodynamical calculation, we only
obtain one outburst light curve for S8 and $0.7\rm{s}$-long luminosity
ascending light curve for S30. However, they are long enough for the
calculation of PSDs. In figure \ref{Fig:PSDs}, we show the relevant PSDs for
S8 and S30. The fundamental frequency (the lowest frequency of harmonics) is
$\sim74.9\rm{Hz}$ for S8 and $\sim285.6\rm{Hz}$ for S30, which are both close
to $71.3\rm{Hz}$ and $300\rm{Hz}$, the respective maximal epicyclic frequencies,
which are also the maximum predicted frequencies of axisymmetric oscillations in
the trapped $p$-mode theory
\citep[see][and a detailed relativistic analysis is ongoing by our colleague Ji\u{r}\'{i} Hor\'{a}k]{Kato01,Kato_book}.
The spectrum of axisymmetric ($m=0$), horizontal $p$-modes
was recently computed by \cite{Giussani14}, who show that
in addition to a discrete set of lower frequency modes which
are trapped in the inner disk, there are modes of frequency
very close to the maximal epicyclic frequency
in which the oscillation is transmitted to the outer disk.
In table \ref{tab1}, we also list the fundamental
frequencies observed from the other oscillating models, which are all close to the
theoretical values though they vary in a narrow frequency range.
In figure \ref{Fig:OneCycle}, we show an oscillating cycle of model S8 for $V$
(upper two panels) and $S_{r\phi}$ (lower two panels). The oscillations arise
near the inner boundary and propagate outwards. In the figure, one can follow
the motion of individual wavelets. The negative
$V$ denotes inflow and negative gradient of $V$ corresponds to
the compression (inflow speed of inner is slower than outer) as well as positive
gradient to the expansion (inflow speed of inner is faster than outer). Thus, any wavelet on
the $V$-profiles can be divided into the compressed wave-front and expanded
wave-rear regions. For example, the left-most wavelet on initial $V$-profile (red
curve in the upper-left panel) can be divided into the wave-front (between dashed
and dash-dotted lines) and wave-rear (between solid and dashed lines) regions.
The relevant variation of $S_{r\phi}$, which is proportional to the viscous heating, is showed in
the lower-left panel. $S_{r\phi}$ (as well as the viscous heating) monotonously increases
in compressed wave-front (see the red curve between dashed and dash-dotted lines) but it unceasingly
increases after the maximal compression (takes place at the dashed line) and then decreases
in the expanded wave-rear. This non-monotonicity of $S_{r\phi}$ in wave-rear
is due to the time-delay contained by equation \ref{stress}. We deem that this is the phase
relation between the viscous heat generation and disk oscillation, which implies the arise of
sonic-point instability for $p$-mode oscillation \citep{Kato78,Kato88}. Thus, we also
deem that the oscillations actually arise near the sonic-point, which is included in
our computational domain and very close to the inner boundary.
\section{Results and Discussions}
In table \ref{tab1}, there are the other nine models with different viscous
settings but with the same BH mass, BH spin and mass supply rate (fixed
accretion rate at the outer boundary) as S8. We only observe oscillations on
the models with non-vanishing delay and large enough constant $\alpha$
($\gtrsim 0.3$) or $\alpha$-profile. In fact, the effective value on the
$\alpha$-profile increases inwards from constant $0.1$ in outer disk region
to the maximum $0.45$ at the inner boundary. Thus, the impact of
$\alpha$-profile is similar to the large constant $\alpha$ while it just
becomes large enough near the origin of instability, sonic-point. The facts
of these oscillating models imply that the large $\alpha$ (at least near
sonic-point) is a necessary condition for the arise of sonic-point
instability as well as $p$-mode oscillation, which is consistent with the
analysis of \cite{Kato88} and \cite{YK96}.
On the other view, we note the effect of delay on the appearance of
oscillations. The oscillating models with $n=1$ (S7 and S8) have the
oscillations during the whole outburst. The other oscillating models with
$n\neq 1$ (S1, S2 and S17) lose the oscillations in different luminosity
stages. The models with $n=0$ (S3, S4, and S16) have no any oscillation though
they have large enough $\alpha$, while the disappearance of the oscillations
on models S5 and S15 is due to the small $\alpha$. These facts imply that
$\tau^*$ may be a favorable delay for oscillation excitation on the disk
around a $10M_{\odot}$ Schwarzschild BH under the large $\alpha$. The last
model in table \ref{tab1}, S30 is a special case for a fast-spinning Kerr BH,
which has another favorable delay $4\tau^*$. This may imply the dependence
between the viscous time-delay and BH spin.
Focusing on the luminosity of the oscillating models, we observe oscillations
only in the limit-cycle outburst state ($L\gtrsim0.2L_{\rm{Edd}}$) when the
inner disk region has switched to slim disk mode. On the contrary, there is no
any oscillation observed in the limit-cycle quiet state
($L\thicksim 0.01L_{\rm{Edd}}$). This is consistent with the HFQPO
observations, but cannot be compared with the sonic-point instability theory
which does not discriminate between accretion rates. Recently, the shearing
box simulation of \cite{Hirose14} implied that the effective $\alpha$ is
enhanced by the vertical convection during the outburst, which is similar to
the conception of \cite{MCT94}. Thus, larger $\alpha$ required by the $p$-mode
oscillation may be caused by the outburst, explaining why HFQPOs are observed
only in high luminosity state.
Beside the fundamental harmonic, there are many overtones in both of the two
spectra in figure \ref{Fig:PSDs}. The frequency ratio of fundamental
harmonic and its overtones is a regular integer series, which is also observed
by \cite{MHL14} in their 2D-simulations (in radial and azimuthal dimensions)
when the axisymmetric $p$-mode become dominant for large $\alpha$. However,
no overtones are observed by \cite{CT95}. Perhaps, this is due
to their adopting of the viscous prescription with
$S_{r\phi}\propto p_{\rm{gas}}$ instead of $S_{r\phi}\propto p_{\rm{total}}$
in \cite{MHL14} and our models. These interesting overtones may be potentially
useful for explaining the observational QPO pairs, which are always in a specific
integer frequency ratio.
Subsequently, the virtual observation from our numerical results will be
logical and interesting. However, it would require more careful treatment. For
example, the effective time-delay of arrival caused by the gravitational
bending on the trajectories of emitted photons and the large observational
view-angle, and effective blocking caused by the gravitational red-shift and
the other shields.
As a rough evaluation for the effective time-delay of arrival, one can
consider the observation from an almost edge-on disk, on which the photons
emitted from two locations apart from distant $\Delta r=203M$ at the same
observer's time ($t$) will arrive at the observer with the rough delay
$\Delta t=0.01\rm{s}$ comparable with the period of $71.3\rm{Hz}$ (our results
presented in figures \ref{Fig:LightCurves} and \ref{Fig:PSDs} can be roughly
regarded as from the face-on disks). The distant $\Delta r=203M$ is comparable
with the radius of outward wavelets propagating area, so it is possible to
change observer's final view on the oscillation power spectrum. This also
implies that the code for ray-tracing calculation on virtual observation must
be time-dependent.
In order to roughly demonstrate the effective blocking, we calculate the
light-curves without the radiation contribution from different inner cutting
regions for the same model S8 and we show the relevant power spectra in
figure \ref{Fig:4PSDs}. It is remarkable that the fundamental harmonic (inside
the rectangle in all four panels) cannot be easily removed from PSDs because
of the outward propagation of the oscillation from sonic-point. It implies
that the measurement of BH spin with HFQPO will be very robust even in a case
when modulation of the innermost disk is not visible.
So far, we have found some features of our model fortunately coinciding with
the counterparts of HFQPO observations. All of these are only due to the
adopting of a special evolutionary stress equation in our model, which mimics
the viscous features induced from MHD simulations. However, our model still
lacks some abilities for capturing various complicated features associated
with the real accretion flows around BHs. In fact, it is almost impossible
for seeing the oscillations during the whole outburst state (like those in
figure \ref{Fig:LightCurves}) in real accretion flows. It is because the
effective $\alpha$ and viscous time-delay, determining the appearance of
oscillations, is turbulent stochastic in real accretion flows though their
expected values can be determined by certain laws. Further more MHD
simulations on the mean behaviors of effective turbulent viscosity is
necessary for improving our understanding on accretion process around BHs
\cite[e.g.][]{Hirose09,Hirose14,Penna13}. It is also impossible for capturing
the resonance between the radial and vertical oscillations suggested by
\cite{Kluzniak04}, because our model is a vertically integrated model and
there is only radial dependence reserved. The further analytic works and MHD
simulations on the roles of radial and vertical oscillations are necessary for
explaining the observational HFQPO pairs though our model has the intrinsic
multi-frequency feature. In fact, the thermal instability required by our
model is the theoretical one whose deviations from the observations have been
found a decade ago \citep{GD04}. While the existence of thermal instability on
the BH accretion disks is still an open issue at present, we only can adopt
this theoretical thermal instability to produce the high luminosity outburst
required by the oscillations in our model. We believe the plausibility of
oscillations observed on our models though it is "dancing" on a poor-quality
stage.
\acknowledgements
This work was supported by the National Natural Science Foundation of China
under grants 11233006 and 11373002, Polish NCN grant UMO-2011/01/B/ST9/05439
and 2013/08/A/ST9/00795, and Czech ASCRM100031242 CZ.1.07/2.3.00/20.0071 Synergy
(Opava) project.
|
\section{Introduction}
The direct detection of a possibly planetary mass object near the
star Fomalhaut by Kalas et al. (2008)
was widely regarded as a great success for the direct imaging detection method.
The separation between Fomalhaut A and b is some 100 au or 13 arc sec.
In addition to this published planetary mass companion candidate (called Fomalhaut b),
Fomalhaut A (the central star) is surrounded by a well resolved dust belt,
which was most recently studied with Herschel (Acke et al. 2012) and ALMA (Boley et al. 2012).
The projected position of the tentative companion was interpreted to indicate that it had cleared the gap in this belt.
The presence of the belt close to the companion candidate constrained the upper mass
limit of the companion candidate to a few Jupiter masses (Kalas et al. 2008).
The star Fomalhaut\footnote{This star is also called $\alpha$ PsA,
i.e. the brightest star in the Southern Fish, the name
{\it Fomalhaut} comes from the Arabic {\it fam al-\d{h}\={u}t al-jan\={u}b\={\i}}
meaning {\it mouth of the southern fish}, (Kunitzsch \& Smart 1986).}
(Fomalhaut A) has the following relevant properties (all for the star A):
\begin{itemize}
\item Position J2000.0: $\alpha$ = 22h 57m 39s and $\delta = -29^{\circ}~37^{\prime}~20^{\prime \prime}$ (Hipparcos, van Leeuwen 2007).
\item The distance as measured by Hipparcos is $7.70 \pm 0.03$ pc (van Leeuwen 2007).
\item Proper motion as also measured by Hipparcos is $\mu _{\alpha} = 328.95 \pm 0.50$ mas/yr
and $\mu _{\delta} = -164.67 \pm 0.35$ mas/yr (van Leeuwen 2007).
\item The spectral type is A4V as obtained by an optical spectrum (Gray et al. 2006);
given this spectral type, the color index is close to zero, e.g. B-V=0.09 mag (e.g. Ducati 2002).
\item The optical brightness is V = 1.16 mag (e.g. Ducati 2002).
\item The age was recently determined to be $440 \pm 40$ Myr by kinematic membership to
the young Castor Moving Group (Barrado y Navascues 1998, Mamajek 2012).
\end{itemize}
The companion was originally discovered in the optical bands of the Hubble Space Telescopes (HST) Advanced Camera for Survey (ACS, Ford et al. 1998).
However, several attempts to detect the object in the near and mid infrared (see e.g. Kalas et al. 2008 and Janson et al. 2012) failed (see Table 1).
This was most troublesome, given that a (few hundred Myr) young cooling Jovian-mass object should be much brighter in the infrared than in the optical.
Furthermore, the latest astrometric measurements by Kalas et al. (2013) indicate that the object would either cross the dust belt
or that it would be on a highly eccentric orbit which is not in alignment with the belt at all.
These two facts together have prompted us to seek for an alternative explanation which might explain all the observations and
finally resolve some of the apparent contradictions. In the following we will first briefly discuss the various scenarios that have been proposed so far and will then present our own considerations.
In this paper, we first review the observations of Fomalhaut b (Sect. 2) and the interpretations
as Jupiter-mass planet (Sect. 2.1), as super-Earth (Sect. 2.2), and as dust cloud (Sect. 2.3).
Then, we consider the background hypothesis as either a White Dwarf (Sect. 3.1) or neutron star (Sect. 3.2.);
we discuss the astrometry, the X-ray data, and the optical and IR photometry, to constrain the neutron star
properties (to be consistent with all observables). In Sect. 3.2.5, we also discuss the probability to
find a background object or even a neutron star close to a star like Fomalhaut. We conclude in Sect. 4.
\section{Fomalhaut~\lowercase{b} as a gravitationally bound object}
\subsection{Fomalhaut~b as a Jupiter-mass planet}
The first interpretation of the available data by Kalas et al. (2008) led to the conclusion that the object may be a giant planet.
From stability considerations of the dust belt,
Kalas et al. (2008) and Chiang et al. (2009) inferred that the mass of the object should be $\leq$ 3\,M$_{\rm J}$.
Larger masses would lead to either smaller orbits than could be inferred from the astrometry,
or higher belt eccentricities than are observed.
Kalas et al. (2008) concluded that if the flux in the optical wavelength range originates in the photosphere of a cooling planet,
then the object needs to be cooler than 400\,K. Otherwise too much flux would be produced at 1.6\,$\mu$m.
They suspected that their non-detections at 1.6\,$\mu$m and 3.8\,$\mu$m might be due to model uncertainties.
However, Marengo et al. (2009) and Janson et al. (2012) present Spitzer IRAC upper detection limits at 4.5\,$\mu$m,
which puts additional constraints on the mass of a possible giant planet.
Janson et al. (2012) conclude that the optical flux cannot stem from a planet's photosphere,
especially since the flux at 0.6\,$\mu$m is 20 to 40 times brighter than expected for an object with $\sim$3\,M$_J$ and a few
hundred Myr (Fortney et al. 2008, Burrows et al. 2003).
On the contrary, Currie et al. (2012) and Galicher et al. (2013) argue that an 0.5-1\,M$_J$ object would not have been detected at 4.5\,$\mu$m
(using models by Spiegel \& Burrows 2012 and Baraffe et al. 2003).
In addition, Galicher et al. (2013) present upper detection limits at 1.1\,$\mu$m,
which are consistent with this upper mass limit.
However, there is a general agreement in all aforementioned studies that the flux in the optical wavelength range cannot stem completely (or at all) from a planet's photosphere.
In addition to the discussed over-luminosity in the optical wavelength range, Kalas et al. (2008) and Janson et al. (2012) report significant (5-8 $\sigma$) variability
of the flux at 0.6\,$\mu$m, which could not be explained by a thermal emission from a planet's photosphere.
Kalas et al. (2008) propose that there might be a 20-40\,R$_J$ accretion disk around the assumed planet.
The disk would reflect light from the primary star, which explains the optical excess flux.
Furthermore, they argue that the variability could then be explained by accretion driven H$_\alpha$ emission.
Janson et al. (2012) strongly disagree, stating that at the high system age moons should have formed in a possible accretion disk,
and thus the reflective surface should be reduced.
Also, they think it is unlikely that accretion driven H$\alpha$ emission can explain the variability,
because the accretion rate would have to be similar to young T Tauri stars.
Currie et al. (2012) and Galicher et al. (2013) re-analyzed the same optical data and did not detect any significant variability at 0.6\,$\mu$m.
Thus, they are not excluding a dust disk around a Jupiter-mass planet.
The most recent study by Kalas et al. (2013) incorporates new astrometric measurements taken with the HST STIS (Space Telescope Imaging Spectrograph, Woodgate et al. 1998) in 2010 and 2012.
They find that the object is most likely on a ring-crossing orbit with a high semimajor axis and eccentricity.
One explanation for that, assuming Fomalhaut b is a planet, would be that it had a close encounter with a further-in massive planet
and was scattered out.
However, Kenworthy et al. (2013) performed deep coronagraphic imaging and can rule out further-in objects with 12-20\,M$_J$ at 4-10\,au.
They state that this effectively rules out scattering scenarios, which makes the orbit elements recovered by Kalas et al. (2013) somewhat peculiar.
Given that five astrometric data points are available only for four different epochs separated by a few years,
any orbit fits with periods of hundreds of years (Kalas et al. 2013) suffer from high uncertainties anyway.
Given the large separation between the two objects (star and presumable planet), even if bound,
other planet detection techniques like the radial velocity or transit technique cannot be applied.
Given the various and sometimes contradictory arguments,
the existence of a giant planet at the position of Fomalhaut b is still possible, but seems increasingly problematic.
\begin{table*}
\caption{Photometric observation epochs and analysis by various authors of Fomalhaut b.}
\label{tab: obs}
\begin{tabular}{lllllccc}
\hline
Date & Telescope & Instrument & Filter & Ref. & Det. ? & app. magnitude / & flux \\
& & & & & & upper limit [mag] & [erg/cm$^{2}$/s/\AA ] \\ \hline
2004 Sep. 26 & HST & ACS/HRC & F606W & Currie et al. 2012 & yes & $24.92 \pm 0.10$ & $3.14 \pm 0.29 \cdot 10^{-19}$ \\
2004 Oct. 25 & HST & ACS/HRC & F606W & Kalas et al. 2008 & yes & $24.43 \pm 0.09$ & $4.93 \pm 0.41 \cdot 10^{-19}$ \\
2004 Oct. 26 & HST & ACS/HRC & F606W & Kalas et al. 2008 & yes & $24.29 \pm 0.08$ & $5.61 \pm 0.41 \cdot 10^{-19}$ \\
2005 July 21 & KeckII & NIRC2 & H & Kalas et al. 2008 & no & $\ge 22.9$ & $\le 8.28 \cdot 10^{-20}$ \\
2005 Oct. 21 & KeckII & NIRC2 & CH$_4$S & Kalas et al. 2008 & no & $\ge 20.6$ & \\
2006 July 14-20 & HST & ACS/HRC & F435W & Kalas et al. 2008 & no & $\ge 24.7$ & $\le 8.36 \cdot 10^{-19}$ \\
2006 July 14-20 & HST & ACS/HRC & F435W & Currie et al. 2012 & yes & $25.22 \pm 0.18$ & $5.18 \pm 0.86 \cdot 10^{-19}$ \\
2006 July 14-20 & HST & ACS/HRC & F606W & Kalas et al. 2008 & yes & $25.13 \pm 0.09$ & $2.59 \pm 0.21 \cdot 10^{-19}$ \\
2006 July 14-20 & HST & ACS/HRC & F606W & Currie et al. 2012 & yes & $24.97 \pm 0.09$ & $3.00 \pm 0.25 \cdot 10^{-19}$ \\
2006 July 14-20 & HST & ACS/HRC & F814W & Kalas et al. 2008 & yes & $24.55 \pm 0.13$ & $1.69 \pm 0.20 \cdot 10^{-19}$ \\
2006 July 14-20 & HST & ACS/HRC & F814W & Currie et al. 2012 & yes & $24.91 \pm 0.20$ & $1.21 \pm 0.22 \cdot 10^{-19}$ \\
2008 Sep. 17-18 & Gemini North & NIRI & L' & Kalas et al. 2008 & no & $\ge 16.6$ & $\le 1.22 \cdot 10^{-18}$ \\
2009 Aug. 16 & Subaru & IRCS & J & Currie et al. 2012 & no & $\ge 22.22$ & $\le 4.01 \cdot 10^{-19}$ \\
2010 Aug. 8 - & Spitzer & IRAC & 4.5$\mu$m & Janson et al. 2012 & no & $\ge 16.7$ & $\le 5.71 \cdot 10^{-19}$ \\
2011 Jul. 23 & & & & & & & \\
\hline
\end{tabular}
\end{table*}
\subsection{Fomalhaut~b as a super-Earth}
If there is a central object associated with the source Fomalhaut b,
then Janson et al. (2012) state that its mass should be limited to $\leq$10\,M$_{Earth}$ if there is a ring-crossing orbit, which the astrometry suggests.
This is in order to prevent the object from significantly influencing the observed belt geometry (see also Kennedy \& Wyatt 2011).
For such an object to exhibit the observed fluxes in the optical wavelenth range,
the object would need to be significantly hotter than a cooling planet of that mass.
Janson et al. (2012) propose a scenario where the object had undergone an intense bombardment of planetesimals within the time scale of $\sim$10$^4$yr.
However, they recognize that the observation of such an event seems improbable due to the short time scale,
as compared to the age of the system. In addition, this scenario is not entirely compatible with the high flux at 0.6$\mu$m.
Another scenario proposed by Janson et al. (2012) is a 10\,M$_{Earth}$ object with a cloud of planetessimals which
are producing the dust that reflects the starlight.
However, this scenario would not explain the aforementioned variability at 0.6$\mu$m which was detected by two independent studies.
It is also questionable why the orbit of such an object would exhibit a high semi-major axis and
eccentricity as found by Kalas et al. (2013) if scattering scenarios can be ruled out (Kenworthy et al. 2013).
In general, while a smaller planetary mass object does not exhibit the same problems with the infrared detection limits as a more massive object,
the orbit of such a low mass companion still seems peculiar. In addition, it is still challenging to explain the optical flux in such a scenario.
\subsection{Fomalhaut~b as a dust cloud}
\label{dust-cloud}
Kalas et al. (2008) originally discussed the possibility that there might not be a central object associated with the source Fomalhaut~b,
but that it is rather a dust cloud produced by the recent collision of two planetesimals.
They reject this possibility because they think it is improbable to observe such a collision at the location of Fomalhaut~b,
due to the low density of planetessimals outside the dust belt. In addition,
they state that a dust cloud would not account for the variability at 0.6$\mu$m.
Janson et al. (2012) note that the observation of such a dust cloud might not be as improbable as Kalas et al. (2008) state.
They argue that such collisions should indeed happen more frequently inside the dust belt,
but are not observable at this location due to the speckle-like nature of such sources.
Thus the probability of observing one such collision outside the dust belt is not negligible.
Galicher et al. (2013) argue that their detection of Fomalhaut b at 0.4$\mu$m would fit well with a dust cloud younger than 500\,yr composed of water ice
or refractory carbonatious small grains, as originally proposed by Kalas et al. (2008).
Furthermore, Galicher et al. (2013) do not detect the variability at 0.6$\mu$m,
which was one of the main arguments against this scenario by Kalas et al. (2008).
However, Currie et al, (2012) contend that the observation of an unbound dust cloud should be unlikely because
Keplerian shear would spread out such a cloud. This small time frame as compared to the system age would make the observation of such a cloud implausible.
We want to note that the study by Galicher et al. (2013) finds that the object Fomalhaut~b can be fitted slightly better with an extended source (0.58\,au)
than with a point source. This is, however, on a very low significance level, and other studies have not mentioned the possible resolved nature of the source.
Overall the dust cloud scenario may appear to be a possible scenario if Fomalhaut~b is extended and gravitationally bound to Fomalhaut A.
\section{Fomalhaut~\lowercase{b} as a background object}
When evaluating the background hypothesis compared to the possibility of a gravitational bound companion, usually a non-moving background object is assumed:
One first tests the hypothesis of a bound companion with the null hypothesis that both the central star and the companion candidate have identical proper motion;
then one tests the background hypothesis with the null hypothesis that the central star has its finite (known) proper motion
and that the companion candidate proper motion is zero.\footnote{Even if both objects have common proper motion, this is not yet a proof that
they orbit each other. Even the detection of curvature in the motion of the companion may not yet be a proof that they orbit each other,
if the curvature would also be consistent with a hyperbolic orbit of a recently ejected object. The detection of curvature in a form
that is not consistent with a hyperbolic orbit would be a proof that both objects orbit each other.}
However, this method may fail if the object of interest is a moving background object with a considerable proper motion.
Therefore, we discuss the possibility that Fomalhaut~b could be a moving background object, unrelated to the primary star.
The HST magnitudes of Fomalhaut~b point to a flat SED (Kalas et al. 2008, Currie et al, 2012) that rules out any ordinary star given the faint magnitudes.
Since the position of Fomalhaut~b changes only slightly over years with respect to the primary star,
its proper motion should be roughly that of the primary star. Only a white dwarf or a neutron star are fast moving and dim enough to match these constrains.
\subsection{Fomalhaut~b as a white dwarf}
The absolute visible magnitude of white dwarfs ranges from 10 to 15~mag (Wood \& Oswald 1998), Fomalhaut b has an apparent visible magnitude of $\sim 25$ mag (Table 1).
Even taking visual absorption caused by the ISM into account,
the putative white dwarf would have a distance of at least 0.5~kpc, but
up to 5~kpc (brightest white dwarf; no absorption). From these estimates,
the object would have a projected spatial velocity (2D) of 900-9000~km/s (applying Fomalhaut's proper motion).
In addition the putative white dwarf would have an unknown radial velocity component that would raise the spatial velocity (3D) to even larger values.
These numbers are rather unrealistic, since the fastest known white dwarf moves
with 450~km/s (Oppenheimer et al. 2001, Wood \& Oswald 1998). Hence, a white dwarf as a putative background object is very unlikely.
\begin{table}
\centering
\caption{Photometry of two brightest known isolated neutron stars as measured with HST (Kaplan et al. 2011)}
\begin{tabular}{ccc} \hline
wavelength & width & flux F$_{\lambda}$ \\
$[\AA ]$ & [\AA ] & [erg/cm$^{2}$/s/\AA ] \\ \hline
\multicolumn{3}{c}{RXJ~1856.4-3754} \\ \hline
1707.4 & 515.2 & $1.50 \pm 0.13 \cdot 10^{-17}$ \\
2960.5 & 877.3 & $2.34 \pm 0.14 \cdot 10^{-18}$ \\
4444.1 & 1210.5 & $4.63 \pm 0.24 \cdot 10^{-19}$ \\
4739.1 & 1186.4 & $3.84 \pm 0.50 \cdot 10^{-19}$ \\
5734.4 & 2178.2 & $1.51 \pm 0.47 \cdot 10^{-19}$ \\ \hline
\multicolumn{3}{c}{RXJ~0720.4-3125} \\ \hline
1370 & 320 & $7.94 \pm 1.34 \cdot 10^{-18}$ \\
1480 & 280 & $5.92 \pm 0.57 \cdot 10^{-18}$ \\
2320 & 1010 & $1.07 \pm 0.13 \cdot 10^{-18}$ \\
4739.1 & 1186.4 & $1.13 \pm 1.45 \cdot 10^{-19}$ \\
5850 & 4410 & $7.73 \pm 0.68 \cdot 10^{-20}$ \\ \hline
\end{tabular}
\end{table}
\subsection{Fomalhaut~b as a neutron star}
Young (1~Myr), hot (1~MK), and close-by ($\le 500$ pc) neutron stars (NSs) have visual magnitudes
ranging from 25-27~mag (see e.g. Kaplan et al. 2011 for a compilation), i.e. as faint as Fomalhaut b, but more distant.
These objects are not necessarily radio pulsars and therefore Fomalhaut~b (if a NS) could have remained undetected.
Indeed, the radio-quiet X-ray emitting NS RXJ~1856.4-3754,
the first such object not powered by rotation, has V=25.6 mag, and it was discovered
by coincidence by Walter et al. (1996) as they actually searched for T Tauri stars.
Due to the supernova kick, the spatial velocities of NSs are on average much larger than those of
e.g. white dwarfs and peak at $\sim 400$~km/s, the fastest known NS moves with 1500~km/s (Hobbs et al. 2005).
\subsubsection{Astrometry: Proper motion and parallax}
Since the companion candidate to Fomalhaut (i.e. object b) was found to be (at least nearly) co-moving to Fomalhaut A,
both Fomalhaut A and Fomalhaut b have very similar proper motions. We can therefore estimate the motion of Fomalhaut b by using the proper
motion of Fomalhaut A: $\mu_{\alpha} = 328.95 \pm 0.50$ mas/yr and $\mu_{\delta} = -164.67 \pm 0.35$ mas/yr (Hipparcos).
The proper motion of both Fomalhaut~A and b are then $\sim 368$~mas/yr.
This proper motion of Fomalhaut b would then be equivalent to a tangential (2D) velocity of $\sim 30$~km/s for 11~pc distance
(i.e. background to Fomalhaut A), or 170~km/s for 100~pc distance. Such velocities are fully consistent with NS velocities.
Since Fomalhaut A is a very nearby star (7.7 pc), its parallactic motion (wobble) is large (a parallax of 130 mas).
Since very precise astrometry for both Fomalhaut A and b are available (which were used to show their common proper motion),
one can check whether some differential parallactic motion between the star and the companion candidate are detectable:
If the presumable companion candidate would be in the distant background, one would not detect any significant parallactic
motion for the companion, but of course still large parallactic motion for the star Fomalhaut A.
To investigate whether the relative astrometry between Fomalhaut A and b would be consistent with Fomalhaut b
being a background object, we tried to fit the data points with a differential proper motion and differential parallax.
Results are shown in Fig. 1. In principle, our best fit in terms of reduced $\chi^2$ (0.16) yields a
differential parallax of $39.1$\,mas and a differential proper motion of $-50.4$\,mas/yr in RA and $107.8$\,mas/yr in Dec.
This corresponds to Fomalhaut b being in the background behind Fomalhaut A at a total distance of 11\,pc.
However, as indicated by the small reduced $\chi^2$, this result is not significant due to the low number of data
points and their uncertainties. We also fit differential motion to the data points without any differential parallax.
The resulting linear fit has only a marginally worse reduced $\chi^2$ (0.33) as compared to the best
fit and is also fully consistent with all measurements. Finally, we repeated the same fitting procedure
but with the maximum possible differential parallax of $129.8$\,mas. The resulting fit is consistent with all
measurements but the one taken in 2010.
However, this measurement has the highest uncertainty of all astrometric data points and
lies still within two $\sigma$ of the fit and thus also for this scenario a reasonable reduced $\chi ^2$ of 1.41 was calculated.
This analysis shows that the astrometry is in principle compatible with Fomalhaut b being a background object at any distance (behind Fom A),
although a distance of 11\,pc is slightly favored. However, as pointed out, we can not exclude that Fomalhaut b is at the same distance as Fomalhaut A.
Fomalhaut is located at a galactic latitude of $-64.9^{\circ}$, i.e. south of the Galactic plane.
The proper motion of both the star Fomalhaut A and its companion candidate (or the nearby neutron star)
are moving (in both equatorial and galactic coordinates) towards the south-south-east, i.e. away from the Galactic plane.
This would be consistent with a young neutron star which was recently born in the Galactic plane.
Since the Sun is currently $26 \pm 3$ pc north of the Galactic plane (Majaess et al. 2009),
and since Fomalhaut is $7.70 \pm 0.03$ pc away from the Sun (mostly towards the galactic south),
Fomalhaut (and its companion candidate) are currently $33.7 \pm 3$ pc south of the Galactic plane.
For the largest one-dimensional velocity known for a neutron star (1285 km/s for PSR B2011+38, Hobbs et al. 2005),
our object would have needed (at least) $2.6 \pm 0.2$ kyr to travel from the Galactic plane to its current position;
for the mean one-dimensional neutron star velocity ($133 \pm 8$ km/s, Hobbs et al. 2005),
it would have needed $248 \pm 37$ kyr.
Of course, it could have formed outside of the Galactic plane, or it may have oscillated around the plane
one or several times (and/or have orbited the Galactic center one or more times).
Its current position south of the Galactic plane together with its motion away from the plane
would be consistent with a young neutron star.
\begin{figure*}
\centering
\includegraphics[width=0.96\textwidth]{differential-parallax-new-fit-final.pdf}
\caption{{\bf Change in position angle (left) and separation (right) with time:}
The five astrometric data points for the object called Fomalhaut b are shown (epochs as listed in Table 1)
in order to try to measure its own proper motion and parallax.
Differential proper motion and differential parallactic fits to the relative astrometric measurements of Fomalhaut A and b.
The solid (black) line shows our best fit of a differential proper motion of $-50.4$\,mas/yr in RA and $107.8$\,mas/yr in Dec
as well as a {\em differential} parallax of $39.1$\,mas, yielding a total distance of 11 pc for Fomalhaut b.
For comparison, we also show a fits with no differential parallax (dashed red line) and with the maximum possible differential
parallax of $129.8$\,mas (dotted blue line),
corresponding to a distance estimate of 8 pc in the first case and any (larger) distance in the second case.
A differential parallax between Fom A and b that is equal to the measured absolute parallax of Fom A implies
no measureable parallax of Fom b. Then, Fom b could be located at any (larger) distance.
In the first (more likely) case the differential motion in RA and declination is $-42.8$\,mas/year
and $110.9$\,mas/year, respectively. In the second case it would be $-68.1$\,mas/year and $100.5$\,mas/year, respectively..
The relative change in position of Fomalhaut A and b can be due to different distances.
The best fit shown results in $\sim 11$ pc as distance for the companion candidate, but it is not significant.
Within $\ge 2~\sigma$ error bars, any other larger distance is also possible,
}
\end{figure*}
\subsubsection{X-ray data}
Young and nearby NSs are detectable as bright X-ray sources (e.g., Walter et al, 1996, Haberl et al. 1997, and Haberl 2007 for a review).
Therefore, we checked the X-ray archives whether there is a source located at the position of Fomalhaut~b.
Only one 1.5~ks {\it EINSTEIN} IPC (Miller et al. 1978) pointing from the year 1979 and a 6.2~ks PSPC exposure with
ROSAT (Tr\"umper 1983) from January 1996 are available in the archive (in addition to a 170~sec exposure from the ROSAT All-Sky-Survey),
see Figs. 2 and 3. The {\it EINSTEIN} IPC observation shows many artifacts that mimic sources, but there is no evidence of X-ray emission
at the current or past position of Fomalhaut~b (Fig. 2). Two potential X-ray sources (denoted as ``source 1" and ``source 2", respectively)
in Fig. 2 are too distant from Fomalhaut~b's past position, considering its proper motion, to be identified with Fomalhaut~b.
Also the ROSAT PSPC data give no evidence of X-ray emission at the position of Fomalhaut~b.
Furthermore, ``source 1" (out of view) and ``source 2" detected in the {\it EINSTEIN} IPC pointing are also not visible,
suggesting that the latter is an artifact or a variable source (Fig. 3).
\begin{figure*}
\centering
\includegraphics[width=0.96\textwidth, viewport=8 88 695 520]{EINSTEIN.pdf}
\caption{{\bf X-ray observation with Einstein:} A 1.5~ks {\it EINSTEIN} IPC pointing exhibits numerous artifacts and two potential X-ray sources.
However, there is no evidence for X-ray emission at Fomalhaut's former position (1979) in the center of this image.}
\label{e}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.75\textwidth]{ROSAT-2.pdf}
\caption{{\bf X-ray observation with ROSAT:}
Data from {\it ROSAT} PSPC (6283~sec, 0.11-2.4~keV, 1996 Jan 19) do not show X-ray emission near Fomalhaut ($\alpha$ PsA).
Fomalhaut b is 13 arc sec NW of A, there is neither a source nor more background photons.
``Source~1" from \autoref{e} is not in the field of view and ``source~2" is not detected even though the exposure time is larger.}
\label{r}
\end{figure*}
Based on these non-detections of Fomalhaut~b in the X-ray images, one can put rough constraints on the properties of the putative NS.
In the 6.2~ks ROSAT exposure obtained with the Boron filter (which blocks about $90~\%$ of the soft flux below 0.3 keV),
an upper limit count rate of $\le 0.00066$ cts/s was determined (Schmitt 1997);
while this upper limit was determined for the star Formalhaut A, it should also apply to Fomalhaut b given the small separation
(see also Fig. 3, where no source is detected).
The isolated NS RXJ~1856.4-3754 appears as bright source with pure blackbody emission (therefore often serves as calibration target, see Mereghetti et al. 2012)
in the X-ray energies with a ROSAT PSPC count rate of 3.67 cts/sec in the 0.11-2.4~keV band.
Hence, RXJ~1856.4-3754 is at least 556 times brighter than Fomalhaut~b in the ROSAT PSPC energy band (having taken into account that
RXJ~1856.4-3754 was observed without Boron filter, while Fomalhaut was observed with Boron filter).
According to the most recent parallax measurements by Walter et al. (2010), RXJ~1856.4-3754 has a distance of $\sim 123$ pc,
yielding an emitting area of $\sim 4.4$ km as origin of the X-ray radiation.
Assuming that Fomalhaut~b emits as blackbody, too,
the temperature $\mathrm{T_{\infty}}$ of its X-ray emitting area must be
below 380,000~K, if its radiation would have the same
normalization\footnote{Due to high gravity and curved space around a NS, an observer at infinity
measures temperature $\mathrm{T_{\infty}=T\sqrt{1-r_s/R}}$ and radius $\mathrm{R_{\infty}=R/\sqrt{1-r_s/R}}$,
where $\mathrm{r_s}$ is the Schwarzschild radius.} of $\mathrm{N=(4.4~km/123~pc)^2}$ as RXJ~1856.4-3754,
or its luminosity must scale with $\mathrm{f<(4.4~km/123~pc)^2\times(10^{6}~K)^4}$
(Tr\"umper 2003, Tr\"umper et al. 2004, Walter et al. 2010) to obey the upper limit derived from the {\it ROSAT} data.
The non-detection of Fomalhaut b as NS is consistent with a NS with at least a few Myr
age - even if only slightly background to Fomalhaut A, see below.
This also applies, if it would be a NS of a different kind, i.e. other than RXJ~1856.4-3754 and RXJ~0720.4-3125.
The latter two have kinematic ages of $0.5$ to 1 Myr (Table 4), while the X-ray non-detection
constraint applies to all NSs older than $\sim 10^{5.5}$ yr (Sect. 3.2.4).
The fact that Fomalhaut\,A (spectral type A4) is not detected in X-rays is not surprising,
since A stars do not have strong hot winds nor a corona.
However, some A4V stars exhibit X-ray luminosities that would correspond to about 50 times the detection limit
of Fomalhaut\,A (Schr\"oder \& Schmitt 2007). Many
of those (if not all) X-ray detected A-type stars are considered to host a very close stellar companion;
this is not the case for Fomalhaut\,A; there are, however, Fomalhaut B as a wide K4-type stellar companion (0.3 pc away, Mamajek 2012)
and Fomalhaut C (LP 876-10) as 2nd wide M4-type stellar companion (0.77 pc away, Mamajek et al. 2013).
\subsubsection{Photometry and SED}
Absorption/extinction caused by the ISM must be taken into account (note that according to L\"ohne et al. 2012a,b extinction caused by the disk around Fomalhaut is negligible).
We use the average Galactic extinction curve provided by Fitzpatrick \& Massa (2007). They fitted this extinction curve with a spline function based on several
anchor points that were calculated by comparing model spectra with measured data from individual reference stars.
The fit errors are in the order of several mmag, whereas we stress that the extinction curves for individual stars
(hence, individual directions) can deviate significantly (Fitzpatrick \& Massa 2007).
Taking blackbody normalization, temperature (of the optically emitting area), and A$_{\rm V}$ as free input parameters,
the resulting magnitudes have to fit those measured by Currie et al. (2012) and Kalas et al. (2008).\footnote{We have tested
and verified the fit procedure with the optical data of RXJ~1856.4-3754, its optical magnitudes, the distance 123 pc,
a radius of 17 km, and negligible extinction.}
As an estimate for the fit quality, we introduce the factor\footnote{The low number of data points prevent a fit quality estimate in terms
of reduced $\mathrm{\chi^2}$.} $\mathrm{q^2=1/k^2\sum_{i}^{k}(m_i - \bar{m}_i)^2}$, where $\mathrm{m_i}$ are the
measured magnitudes (Currie et al. 2012, Kalas et al. 2008) and $\mathrm{\bar{m}_i}$ the fitted magnitudes in the filter $i$.
Since Fomalhaut~b is only detected in three filters, the total number of different magnitudes is $k=3$.
Distance, temperature (of the optically emitting area), and extinction are free fit parameters, the radius was assumed to be 17 km.
Currie et al. (2012) and Kalas et al. (2008) list magnitude errors in the order of 0.1~mag to 0.2~mag for the HST photometry (see Table 1).
However, the magnitudes of the same filter differ by 0.5~mag for different measurements and by different authors,
suggesting that the systematic errors are much larger than the statistical errors.
We calibrated the modelled blackbody flux to the Vega magnitude system (our results were checked with the average Vega flux densities in the different
HST filters as listed in the HST handbook), since Currie et al. (2012) and Kalas et al. (2008) give Fomalhaut~b's magnitudes in the Vega system,
and corrected for non-infinite aperture (see HST handbook), see Table 1.
Furthermore, we calculated the HST magnitudes of the two optically detected isolated NSs with known distance, see Table 2,
and compared our results to those given by Kaplan et al. (2011) in ST the magnitude system as an additional check.
We can find blackbody spectral energy distribution which fit the observed data,
they are in the parameter range of $\mathrm{T_{\infty}=6,810-126,500~K}$, $\mathrm{D=1.6-33~pc}$ and $\mathrm{A_V<2.5~mag}$
(but not all combinations in the parameter ranges are possible).
In Table 3, we show the ten allowed combinations as examples.
The resulting
effective temperatures (of the optical emitting area)
yield ages of at least
$\sim 10^{5.5}$ yr
for the putative NS
according to cooling curves in Aguilera et al. (2008) and Page et al. (2009);
however, since NSs cool very rapidly for
effective temperatures (of the optical and/or X-ray emitting area)
of $\le 100,000$ K and ages
above at around that age,
a precise age estimate from the temperature is not well possible.
X-ray non-detection would not be surprising for Fomalhaut b being a relatively old NS.
We would like to point out that we assume just one (effective) temperature for both the
optical and X-ray emitting area, i.e. the whole neutron star surface.
This temperature fits the optical magnitudes known and agrees with the X-ray upper limit as observed.
It is of course possible that the polar caps are hotter than the remaining surface area,
so that the object as neutron star would show pulses, even if not yet detected.
A more realistic model of the emission of a young NS is a two component model:
a cool ($\mathrm{T_{\infty}\approx300,000-400,000~K}$) blackbody emitting in the optical (representing the major part of the NS surface) and
a hot ($\mathrm{T_{\infty}\approx10^5-10^6~K}$) blackbody, visible at X-ray energies, caused by
the hot spot(s) at the magnetic poles (Kaplan et al. 2011).
However, this would mean fitting three magnitudes with five parameters ($A_V$,$D$ or $\mathrm{R_{1\infty}}$,$\mathrm{R_{2\infty}}$,$\mathrm{T_{1\infty}}$,$\mathrm{T_{2\infty}}$)
and -- as expected -- did not lead to useful results:
E.g. any high effective temperature value
(of the optical and X-ray emitting area, i.e. the whole surface)
can be compensated by large $\mathrm{A_V}$ values flattening the SED.
To put constraints on the two blackbody model it is necessary to derive more data points for the SED, in particular at UV energies.
\begin{table}
\centering
\caption[]{Ten allowed parameter combinations consistent with the photometry of Fomalhaut b.
We list magnitudes, effective temperature ($\mathrm{T_{\infty}}$)
(of the optical emitting area)
as seen from an observer at infinity,
predicted distance to the Sun ($D$), and interstellar extinction $\mathrm{A_V}$.}
\label{best_ten}
\begin{tabular}{ccccccc}
\hline
F435W & F606W & F814W & $\mathrm{T_{\infty}}$ & $D$ & $\mathrm{A_V}$ & $q$ \\
{[mag]} & [mag] & [mag] & [K] & [pc]& [mag] & \\
\hline
25.13 & 25.07 & 24.87 & 19360 & 6.2 & 0.53 & 0.045\\
25.15 & 25.09 & 24.87 & 37890 & 9.0 & 0.85 & 0.048\\
25.12 & 25.02 & 24.81 & 14320 & 4.9 & 0.33 & 0.050\\
25.12 & 25.08 & 24.87 & 45710 & 10.1 & 0.86 & 0.051\\
25.22 & 25.11 & 24.84 & 54190 & 10.5 & 1.04 & 0.052\\
25.24 & 25.12 & 24.87 & 23840 & 6.8 & 0.77 & 0.053\\
25.07 & 25.02 & 24.86 & 13960 & 5.1 & 0.22 & 0.054\\
25.20 & 25.13 & 24.89 & 55520 & 11.1 & 0.97 & 0.054\\
25.09 & 25.09 & 24.92 & 62010 & 12.4 & 0.85 & 0.058\\
25.11 & 25.03 & 24.79 & 53960 & 10.4 & 0.98 & 0.059\\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\includegraphics[angle=270,width=0.96\textwidth]{new-fig-sed.jpg}
\caption{{\bf Optical photometry for the object called Fomalhaut b compared to two neutron stars:}
(a, top) in black (Kalas et al. 2008) and
(b, bottom) in red (Currie et al. 2012)
with IR upper limits as arrows (band width indicated), data from Table 1.
The optical emission from RXJ~1856.4-3754 can well be fitted with a 380,000 K blackbody, the emission from RXJ~0720.4-3125
is consistent with 112,000 K (both assumed to be unabsorbed); we show their spectral energy distribution
(Planck functions for 112,000 K and 125,500 K) as pink and blue line, respectively. See also Fig. 5.}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[angle=270,width=0.96\textwidth]{plot_sed_rne.jpg}
\caption{{\bf Fomalhaut b photometry compared to typical hot neutron stars and a typical cold planet.}
The available optical photometry for the object called Fomalhaut b in red (Currie et al. 2012) and black (Kalas et al. 2008)
with IR upper limits as arrows (band width indicated), data from Table 1, as also plotted in Fig. 4.
We also show the optical data points for the two well-known isolated middle-aged neutron stars
RXJ~1856.4-3754 and RXJ~0720.4-3125 in pink and blue, respectively (data from Table 2).
The optical emission from RXJ~1856.4-3754 can well be fitted with a 380,000 K blackbody, the emission from RXJ~0720.4-3125
is consistent with 112,000 K (both assumed to be unabsorbed).
The model atmosphere for 400 K in green (here from AMES COND for log g = 4.0 (Chabrier et al. 2000, Allard et al. 2001) for a few hundred Myr planet,
from phoenix.ens-lyon.fr/Grids/AMES-Cond/SPECTRA, as an example (similar for other models),
scaled in y-axis as far down as possible to still fit the optical data points)
does not fit the Fomalhaut b data as known before -- it is inconsistent with the IR non-detection.
We show all data in the upper panel and then again (for clarity) in the lower panels the Fomalhaut b data
with Planck functions for 380,000 K and 112,000 K (bottom left) and with Planck functions for 112,000 K and 125,500 K (bottom right),
which do fit
(fits with 17 km radius as RXJ~1856.4-3754).
The y-axis (flux) ranges are identical, the x-axis (wavelength) ranges differ slightly.}
\end{figure*}
\subsubsection{Constraining the neutron star properties}
Given the observed astrometry and photometry, we can now try to constrain the properties of Fomalhaut b,
if it would be a NS, in particular its distance and age range.
We assume for most part of this section that Fomalhaut b as NS would have the typical radius of NSs, $\sim 10$ to 17 km.
First, we assumed $\sim 17$ km radius for the optically emitting area and $\sim 4.4$ km as radius of the
X-ray emitting area in our spectral fits; these are values similar as for RXJ~1856.4-3754 (and maybe RXJ~0720.4-3125).
Afterwards, we compare the object called Fomalhaut b with other neutron stars,
which have smaller and/or hotter polar caps than RXJ~1856.4-3754 and RXJ~0720.4-3125.
At the very end of this subsection, we also consider even smaller emitting areas (and smaller radii)
like in strange (quark) stars.
The non-detection of a supernova remnant places a lower limit to the age of a NS to roughly $10^{5}$ yr,
a supernova remnant has diffused and faded away. In such a case, one would regard this NS (Fomalhaut b) as middle-aged,
isolated NS ({\em isolated} means that there is neither a companion nor a supernova remnant).
Given the small distance of Fomalhaut b (even as NS), $\sim 11$ pc being our best fit, see above, the supernova remnant
would have a large extend on sky: The Vela remnant at a distance of $\sim 290$ pc (Caraveo et al. 2001, Dodson et al. 2003)
and an age of $\sim 11$ kyr (Dodson et al. 2002), both measured for the Vela pulsar, has an apparent size of $255^{\prime}$ (Green 2009).
A supernova remnant some $\sim 26$ times closer would have a size of $\sim 111^{\circ}$.
Even at such a large size, it might have been noticed in the ROSAT All-Sky Survey, but no such (large) remnant was detected
(in particular not at that position), see e.g. Busser (1998) and Schaudel et al. (2002).
If such a large remnant would not have been detectable, Fomalhaut b as NS would still not be a young NS,
because it would then be bright in X-rays (if young), which is not the case.
Hence, if a NS, Fomalhaut b is most likely middle-aged or old.
If Fomalhaut b as NS would be related to the Local Bubble ($\sim 50$ Myr old), a volume around the Sun with very low interstellar
medium density, then it might be possible that its SN did not form a detectable remnant; in this case, the NS might be younger than 50 Myr
(but still older than $\sim 10^{5.5}$ yr due to X-ray non-detection).
The non-detection of radio pulsations
may simply be due to the fact that the pulses
are not beamed towards Earth; otherwise, they place a lower limit to the age of a NS to roughly $10^{8}$ yr,
the so-called death-line or graveyard of radio pulsars:
Most known pulsating NSs are younger than $10^{8}$ yr given their (characteristic) spin-down age (Gyr old milli-second pulsars recycled
by mass transfer from their (former) companion are exceptions).
In such a case, one would regard this NS (Fomalhaut b) as middle-aged to old.
The proper motion is definitely in the possible range for NSs.
The astrometry (parallactic motion) would be well consistent with
$\sim 11$ pc (best fit),
but also much larger distances are not excluded (Fig. 1).
In principle, the color of Fomalhaut b could also be compared to the NSs RXJ~1856.4-3754 and RXJ~0720.4-3125;
however, Kalas et al. (2008) and Currie et al. (2012) do not agree well on the magnitudes
and the differences and error bars in their values are on the order or larger than the colors,
and also, while Fomalhaut b was detected by HST with F606W, F435W, and F814W, neither RXJ~1856.4-3754 nor RXJ~0720.4-3125
were observed with F435W nor F814W (but detected in F606W).
We have re-reduced the HST photometry and arrived at values close to those of Kalas et al. (2008) and Currie et al. (2012),
but also our photometry error bars are comparable to the error bars and absolute differences between the results in
Kalas et al. (2008) and Currie et al. (2012).
The optical detections and the upper limits in the infrared and X-rays allow good fits for
a range in effective temperature
(of the optical and X-ray emitting area)
of up to roughly $\sim 100,000$ K with small to negligible extinction (see Table 3 for a few examples);
this is compatible with the more conservative X-ray temperature upper limit (up to 380,000 K)
For a NS, this would then yield a distance of up to 33 pc and an age of
at least $\sim 10^{5.5}$ yr
(according to the cooling
curves in Aguilera et al. 2008 and Page et al. 2009), NSs start to cool very rapidly for temperatures $\le 100,000$ K
(of the optical and X-ray emitting area)
at ages somewhere between $10^{5.5}$ to $10^{6.5}$ yr, so that a precise age estimate from the temperature
(of the optical and X-ray emitting area)
is hardly possible in this regime.
The non-detection in X-rays (Figs. 2 and 3) can place limits on temperature
(of the X-ray emitting area)
and distance by
comparison with other middle-aged isolated NSs (all without supernova remnant),
namely the NSs RXJ~1856.4-3754 and RXJ~0720.4-3125, which are middle-aged, isolated, and which have a known distance;
in Table 4, we list their ROSAT PSPC count rates, distances, ages, and V-band magnitudes -
to be used to scale to Fomalhaut b (we assume negligible extinction for Fomalhaut b and the NSs here).
Given that RXJ~1856.4-3754 and Fomalhaut b have very similar optical photometric magnitudes (Tables 1 and 2),
we can relate distances $d$ and temperatures $T$ (since flux scales with $T^{4}$ and $d^{-2}$).
If the temperature ratio between RXJ~1856.4-3754 (380,000 K) and Fomalhaut b (say 100,000 K to 150,000 K) is the same for the warm surface
responsible for the optical emission as for the hot polar spots responsible for X-ray emission, then we can scale
from the temperature ratio and the distance ratio (123 pc for RXJ~1856.4-3754 and, say, 11 pc for Fomalhaut b as NS)
as well as the X-ray count rate of RXJ~1856.4-3754 (3.67 cts/sec, Table 4)
also to the expected X-ray count rate of Fomalhaut: With the PIMMS software,
we obtain $\sim 0.00066$ cts/sec for ROSAT PSPC with Boron filter for 112,000 to 126,500 K at $\sim 11$ pc;
this is exactly the upper limit count rate obtained for Fomalhaut (A and b): 0.00066 cts/sec (Schmitt 1997).
Hence, for a distance range of 11 pc (best fit obtained from the astrometry),
a NS would need to have a temperature
(of the X-ray emitting area)
of 112,000 K to 126,500 K to obey the X-ray upper limit.
The X-ray non-detection of Fomalhaut b is then consistent with being a NS.
Indeed, the temperature of 380,000 K is both the upper limit on the temperature of the X-ray emitting region
(based on the comparison with RXJ~1856.4-3754, above), and it is also close to the temperature of the
optical emitting region of RXJ~1856.4-3754, see e.g. Kaplan et al. (2011).
In Figs. 4 \& 5, we show the available photometry and upper limits of the object known as Fomalhaut b,
compared to the NSs RXJ~1856.4-3754 (380,000 K) and RXJ~0720.4-3125 (112,000 K) as well as compared to several blackbodies
with temperatures of 112,000 K to 126,500 K
(for the X-ray emitting area),
as well as a typical model atmosphere for 400 K (as should be expected for a planet, Kalas et al. 2008).
The blackbodies of 380,000 K to 112,000 K do fit the Fomalhaut b data.
For the comparison with RXJ~0720.4-3125, one should keep in mind
that Kaplan et al. (2011) showed that a Rayleigh-Jeans tail with a temperature of 112,000 K
would not fit the spectrum without an additional power law component.
For RXJ~1856.4-3754, however, there is no evidence for a deviation from a blackbody.
The constraints from optical data and X-ray non-detection are also consistent with an age above
$\sim 10^{5}$ yr as derived from the non-detection of a supernova remnant
or even
$\sim 10^{8}$ yr as derived from the non-detection of radio pulsations (if beaming towards us).
By comparison with RXJ~1856.4-3754 and RXJ~0720.4-3125 (i.e. same temperature and area of the emitting polar caps),
Fomalhaut b as NS would be $\sim 10^{5.5}$ yr (or older) as derived from the non-detection of X-rays, but see below.
Let us now also compare the object called Fomalhaut b with NSs other than RXJ~1856.4-3754 and RXJ~0720.4-3125,
namely with NSs with smaller and/or hotter emitting areas (polar caps).
E.g., the radio pulsars PSR J0108-1431, PSR B1929+10 and PSR B0950+08 are detected in X-rays at large distance;
the existence of radio-silent NS with hot and/or small emitting regions is possible.
According to recent X-ray observations with Chandra and XMM, the relevant parameters are known.
PSR J0108-1431 can be fitted with a blackbody with $k \cdot T = 0.28$ keV and an X-ray emitting area of $53^{+32}_{-21}$ m$^{2}$,
or a power law with $\gamma = 2$ (Pavlov et al. 2009), it has an age of $\sim 160$ Myr and a distance of $\sim 210$ pc (Taylor \& Cordes 1993).
Then, using the PIMMS software, we expect $0.045$ count per second (0.027 to 0.072 for full $1~\sigma$ error range) with ROSAT PSPC
with boron filter (same setup as used in the observation of Fomalhaut), if such a NS would be at $\sim 11$ oc distance only.
This is more than the ROSAT PSPC Fomalhaut upper limit being 0.00066 cts/sec (Schmitt 1997).
According to Posselt et al. (2012), this NS has an energy of $k \cdot T = 0.11$ keV with the radius of the X-ray emitting area being 43 m.
Then, we would obtain with PIMMS a ROSAT PSPC coutn rate of 0.0047 cts/sec with boron filter, again at 11 pc, again larger than the upper limit.
A NS like PSR J0108-1431 would have been detected up to 0.00066 cts/sec up to a distance of 90 to 120 pc,
but is not detected at the position of what is called Fomalhaut b.
PSR B1929+10 is $\sim 3$ Myr old at $\sim 361$ pc distance with hot polar caps ($\sim 0.3$ keV or $\sim 3.5 \cdot 10^{6}$ K, with
a projected emitting area of $\sim 3000$ m for the two X-ray emitting polar caps together,
i.e. a radius of of the X-ray emitting area of $\sim 21.5$ m each for two circular polar caps),
resulting in an X-ray luminosity (or the polar caps) of $\sim 1.7 \cdot 10^{30}$ erg/s in the 0.3-10 keV band,
see e.g. Misanovic et al. (2008) or Slowikowska et al. (2005) as well as references therein.
Again using PIMMS, we would expect a ROSAT PSPC count rate of 3.47 cts/sec with the boron filter at 11 pc,
so that a NS like PSR B1929+10 would be detectable until $\sim 800$ pc at 0.00666 cts/sec (but is not detected at the Fom b position).
PSR B0950+08 is $\sim 17$ Myr old at $\sim 262$ pc distance with hot polar caps ($k \cdot T = 0.086$ keV or $\sim 10^{6}$ K
with $\sim 250$ m radius of the X-ray emitting caps),
resulting in an X-ray luminosity of $\sim 3 \cdot 10^{29}$ erg/s from the polar caps,
see e.g. Zavlin \& Pavlov (2004) or Becker et al. (2004) as well as references therein.
Again using PIMMS, we would expect a ROSAT PSPC count rate of 1.097 cts/sec with the boron filter at 11 pc,
so that a NS like PSR B1929+10 would be detectable until $\sim 450$ pc at 0.00666 cts/sec (but is not detected at the Fom b position).
By comparison with RXJ~1856.4-3754 (X-ray luminosity of $3.8 \cdot 10^{31}$ erg/s at 120 pc, Burwitz et al. 2003),
PSR B1929+10 would be detectable at up to five times larger distances than RXJ~1856.4-3754,
and PSR B0950+08 would be detectable at up to eleven times larger distances than RXJ~1856.4-3754.
PSR J0108-1431, PSR B1929+10, and PSR B0950+08 would have been detectable in X-rays at a distance of only $\sim 11$ pc
(best fit for Fomalhaut b) -- even given the short exposure time of the ROSAT pointing.
The astrometry of Fomalhaut b does not exclude larger distances, where those two NSs would also have been detectable.
Hence, if Fomalhaut b as NS would be similar to PSR J0108-1431, PSR B1929+10, or PSR B0950+08
in both temperature and radius of both the X-ray emitting area
as well as age, than it would be detectable at 11 pc. Or, putting it another way around,
Fomalhaut b as NS would need to be older than the previously given lower limit of $\sim 10^{5.5}$ yr.
On the other hand, the available optical photometry of Fomalhaut b allows good fits only for blackbody
temperatures (of the optically emitting areas) from 112,000 to 126,500 K, which may be too small for PSR B1929+10 and PSR B0950+08.
If Fomalhaut A is a member of the several hundred Myr ($440 \pm 40$ Myr) old Castor Moving Group (Barrado y Navascues 1998),
it might well be possible that Fomalhaut b as NS (and/or its progenitor) also belongs to this Moving Group -
given that Fomalhaut b (even as NS) and Fomalhaut A have a similar proper motion (as most members of
Moving Group have a similar proper motion). \\
The Castor Moving Group has 26 known members (plus the two stellar companions
to Fomalhaut A, being Fomalhaut B and C with K4 and M6) with at least eight A-type stars,
four stars with A0-2 (Barrado y Navascues 1998), so that it is not impossible that there was originally also one
early B-type star in this group (the progenitor of Fomalhaut b as NS):
We have converted the known spectral types of the Castor Moving Group members (Barrado y Navascues 1998, Mamajek 2012,
Mamajek et al. 2013) to main-sequence masses to investigate the present mass function; extrapolating
from the bin with the largest masses (a bin with stars above 1~M$_{\odot}$) by using the exponent $N \simeq 2.7 \pm 0.7$
from the Kroupa et al. (1993) initial mass function to even larger masses, we can estimate the expectation number
for core-collapse supernova progenitors (with at least 8~M$_{\odot}$) in the Castor Moving Group to be $0.2 ^{+0.6} _{-0.4}$
within $2\sigma$ error bars. Hence, it may appear unlikely, but possible. \\
If the progenitor of Fomalhaut b as NS would indeed have been a member of the Castor Moving Group, then Fomalhaut b as NS would now
have an age only slightly below the age of the Castor Moving Group (given the short life-time of its
progenitor), so that it would have an age of a few hundred Myr. Such an age is fully consistent with the cooling curves
of NSs given its brightness in the optical and its non-detection in the X-rays.
(The proper motion of the two presumable stellar companions Fomalhaut B and C being co-moving with Fomalhaut A
can be interpreted either that they form a triple stellar system (with undetected orbital motion around the
common center-of-mass) or that all three stars are (independant) members of the same Moving Group.)
The proper motion of Fomalhaut A and b would be equivalent to a tangential (2D) velocity of $\sim 30$~km/s
for $\sim 11$ pc distance,
which is a low, but a possible velocity for a NS. The unknown radial velocity has to be added.
Only if Fomalhaut b (as NS) got a very small kick in its supernova, then its velocity can now still be similar
to the velocity of the progenitor star, i.e. the typical Castor velocity; a small velocity would be consistent
with a small SN kick. \\
We would like to stress that, for our interpretation of Fomalhaut b as NS, it is not essential that it would
be a member of the Castor Moving Group. \\
At an age of $440 \pm 40$ Myr (or, say, hundreds of Myr), detections of radioisotopes on Earth
due to the very nearby SN explosion is also very difficult:
At the proper motion and current distance of the Fomalhaut companion candidate,
it would have moved $\sim 16$ kpc in 440 Myr
(the space velocity used considers only the known two-dimensional motion, not its unknown radial velocity);
it would have orbited the Galactic Center almost twice,
so that it is not possible to constrain well the location of the SN; hence, it is also completely unknown,
whether the SN took place within a few tens or hundreds of pc around Earth. Also, Firestone (2014) had to restrict their study
to SNe (in radioisotopes) within the last 300 kyr given also the half lifes and measurement precision of relevant radionucleids.
Could Fomalhaut b as NS be just $\sim 2$ Myr young~?
It is at least $\sim 10$ times closer than RXJ~0720.4-3125 (at $\sim 280$ pc, Eisenbeiss 2010),
so that it would be expected to be at least $\sim 100$ times brighter, if at the same age ($\sim 1$ Myr, Table 2)
and with the same temperature
of the emitting area
(Figs. 4 \& 5), and radius ($\sim 17$ km, Eisenbeiss 2010).
In fact, it is only $\sim 3$ times brighter in the optical than RXJ~0720.4-3125 ($\sim 1$ Myr, Tables 1, 2, and 4).
If Fomalhaut b as NS would be 2 to 3 times smaller than RXJ~0720.4-3125, it would be 4 to 9 times fainter.
The remaining factor could easily be obtained by a slightly different temperature
of the emitting area
(luminosity scaling with
the forth power of the temperature), even if at the same radius; also, cooling tracks at that age are
highly uncertain. Hence, from these considerations, Fomalhaut b as NS could be as young as $\sim 2$ Myr. \\
If Fomalhaut b as NS would be as young as $\sim 2$ Myr, it would have moved $\sim 61$ pc since 2 Myr
(the space velocity used considers only the known two-dimensional motion, not its unknown radial velocity).
Given its current distance of $\sim 11$ pc (best fit),
it had a distance of somewhere with $\sim 72$ pc at birth (i.e. still inside the Local Bubble).
A supernova that close might have left some effect on Earth, e.g. a $^{60}$Fe signal,
such as the one detected at an age of 2 Myr under the Earth ocean crust (Knie et al. 1999, Fields 2004, Bishop et al. 2013).
We conclude that, for certain parameter combinations,
Fomalhaut b as NS could be 2 Myr young and could be the NS that was born
in the nearby supernova that left $^{60}$Fe on Earth. However, this is speculative.
The true age could be constrained with, e.g., an X-ray detection.
If Fomalhaut b as compact, non-planetary object would not have the typical NS radius, nor the same radius as
the NSs RXJ~1856.4-3754 or RXJ~0720.4-3125, to which we compared it, but a different radius, also a different
distance estimate would apply. Theoretical considerations show that a NS cannot be much larger than
RXJ~1856.4-3754 ($17 \pm 3$ km, Tr\"umper et al. 2004, Walter et al. 2010). However, some equations-of-state
predict smaller radii such as half the radius, namely for strange (quark) stars.
Our considerations above hold in a similar way for both normal NSs and strange stars, with one exception:
If strange stars are half a large as normal NSs, then the allowed distance range for Fomalhaut b as strange star
would need to be accordingly smaller by a factor of $\sqrt 2$.
For similar temperatures
(of both the X-ray and optical emitting areas),
the radius of the emitting area cannot be smaller by more than a factor of 5 than in RXJ~1856.4-3754:
A somewhat larger distance of, say, 20 pc -- devided by the square root of 5 -- gives 8 pc, the lowest allowed distance
(at or just behing Fomalhaut A).
If Fomalhaut b would be a NS or a strange star, we would not expect photometric variability (except maybe pulsations),
and we would not expect the object to be resolved/extended.
\begin{table*}
\caption{Properties of two isolated neutron stars. We list both the kinematic age from tracing back the motion of the object to its presumable birth place
inside an OB association as well as its characteristic (spin-down) age $\tau _{ch}$, which is to be considered an
upper limit to the true age; the kinematic ages fit better with cooling curves than spin-down ages (Tetzlaff et al. 2010). \\
References: Wal96: Walter et al. 1996, Wal10: Walter et al. 2010, vKK08: van Kerkwijk \& Kaplan 2008, vKK01;
Tet10: Tetzlaff et al. 2010; van Kerkwijk \& Kulkarni 2001, Hab97: Haberl et al. 1997, Eis10: Eisenbeiss 2010,
Kap05: Kaplan \& van Kerkwijk 2005; Tet11: Tetzlaff et al. 2011.}
\begin{tabular}{l|cc|cc|cccc|cc} \\ \hline
NS & X-ray & Ref & distance & Ref & \multicolumn{4}{c}{age [Myr]} & F606W & Ref \\
name & cts/sec & & [pc] & & $\tau _{ch}$ & Ref & kin. & Ref & [mag] & \\ \hline
RXJ1856 & 3.67 & Wal96 & $123 ^{+15} _{-11}$ & Wal10 & 3.8 & vKK08 & $\sim 0.5$ & Tet10 & 25.6 & vKK01 \\
RXJ0720 & 1.65 & Hab97 & $280 ^{+210} _{-85}$ & Eis10 & 1.9 & Kap05 & 0.7-1.0 & Tet11 & 26.8 & Eis10 \\ \hline
\end{tabular}
\end{table*}
\subsubsection{Probability considerations}
From the age of the Galaxy and the core-collapse supernova rate of a few events per century (Tammann et al. 1994),
as well as from several other considerations (metallicity of the ISM, pulsar birth rates, $^{26}$Al content, etc.),
$\sim 10^{8}$ NSs should have been produced in our Galaxy so far.
Of course, only the hottest (i.e. youngest) ones can be detected in the optical:
For the bright isolated NS RXJ~1856.4-3754, the optical emission and size indicates a surface temperature
(of the optical emitting area)
of a few $10^{5}$ K
(Tr\"umper 2003, Tr\"umper et al. 2004, Walter et al. 2010).
According to several sets of NS cooling curves, and depending on assumption about their interior and their mass
(see, e.g., cooling curves in Aguilera et al. 2008 and Page et al. 2009),
NSs with that temperature
(of the optical emitting area)
might be sufficiently hot and, hence, detectable in the optical (like RXJ~1856.4-3754), until an age of $\sim 10^{6}$ yr
(or a few times $\sim 10^{6}$ yr).
For a constant NS formation rate since $\sim 1.4 \cdot 10^{9}$ yr (the age of the Galaxy),
there should then be some $\sim 72,000$ detectable (young, hot) NS.
Given the velovity distribution of NSs, many of them can leave the Galaxy,
but the young ones that we consider here could not leave it, yet.
Given their high velocities, NSs are not restricted to the Galactic plane;
if they would be distributed uniformly on the sky ($4\pi$~sr = 41253 square degree),
then we would expect $\sim 1.75$ detectable NS per square degree.
To get a rough estimate of the area that was covered by instruments capable of detecting Fomalhaut\,b,
we used the HST archived exposure catalog (provided by STScI in 2007).
This is a sensible approach, since Fomalhaut\,b has so far only been detected with the HST, only in the optical,
and the majority of the HST measurements of Fomalhaut\,b including the discovery epoch were taken before 2007.
From the catalog, we extracted all imaging exposures which were taken in bands with a central wavelength shorter than 1\,$\mu$m
and with longer exposure times than 1000\,s, i.e. exposures in which objects as faint as Fomalhaut\,b should have been detected.
After removal of duplicate exposures (i.e. exposures with a separation of less than 1\,arcmin), we found 862 exposures with ACS/HRC,
2216 exposures with ACS/WFC, and 2900 exposures with WFPC2 matching our criteria.
Given the fields of view of these instruments, these exposures cover a total area of 11.65 square degrees on the sky.
Thus, $\sim 20$ NSs could be contained in this area.
A few well-known middle-aged NSs are indeed optically detected by HST (like RXJ~1856.4-3754
and the other the so-called Magnificent Seven NSs, see review in Haberl 2007), namely in deep exposures,
some of them at high galactic latitude.
If we restrict this estimate to the Galactic plane ($|b| \le 20^{\circ}$), where almost all stars (and most NSs) are located (we deal here with a
potential NS in the background to a star with small apparent separation to that star), then we are left with only 1.64 square degrees,
so that we would expect $\sim 3$ detectable NSs.
Therefore, it is not unlikely to discover a previously unknown NS in one or a few of these exposures.
If we would restrict the estimate to a small area on the sky around a bright nearby star (like Formalhaut A), we can arrive at a different estimate:
Fomalhaut b is some 13 arc sec off Fomalhaut A, so that a circle with 13 arc sec radius is relevant here for a background probability estimate.
The probablity to find any one of $\sim 72,000$ detectable NSs (as given above) within 13 arc sec around one particular star
(on the whole sky) is then only $\sim 7 \times 10^{-3}$.
However, several thousand different young and/or nearby stars were surveyed with deep exposures for planetary companions by direct imaging,
e.g. the directly imaged planet candidate (or brown dwarf) companion near GQ Lup was first detected in 1999 by the HST in an optical imaging snapshot program
(Neuh\"auser et al. 2005).
Then, the expectation number of background (or foreground) NSs within 13 arc sec around any one of several thousand stars (say, 5000 stars)
is $\sim 0.35$. This estimate is not significantly different from one object found.
Hence, probability estimates show that it is not very unlikely to find one unrelated object (or even a NS)
with deep imaging in the background (or foreground) of one of the surveyed stars.
Above, we have estimated the probability for young, self-luminous NSs only. It is also possible that older NSs get
re-heated due to Bondi-Hoyle accretion of interstellar material (e.g. Madau \& Blaes 1993).
However, it may well be that re-heating of old NSs is not important, because otherwise many more such
re-heated NSs should have been detected by, e.g., the ROSAT All-Sky Survey and other X-ray missions
(see e.g. Neuh\"auser \& Tr\"umper 1999) -- such old accreting NSs were not detected (in particular not in large numbers)
probably due to both higher space velocities $v$ (Bondi-Hoyle accretion scales with $v^{-3}$)
and larger magnetic fields (propeller effect) than assumed in Madau \& Blaes (1993).
On the other hand, with PSR J0108-1431 at $\sim 160$ Myr, there also exists an example of a
relatively old NS which is still quite hot -- probably due to internal re-heating
(Tauris et al. 1994, Pavlov et al. 2009, Posselt et al. 2012).
If we would add those NSs in our probability estimate, the probability for finding one NS would increase.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth, viewport=90 265 480 560]{pmF1.pdf}
\caption{{\bf Proper motion} (scaled to arbitrary units) of Fomalhaut's neighbouring stars according
to the {\it SIMBAD} data base. Fomalhaut is marked by the black solid star in the centre.}
\label{pmF1}
\end{figure}
Formalhaut b is not only located close to a bright star (Formalhaut A), but it also moves with a similar proper motion.
We checked the proper motions of all stars (projected) near Fomalhaut, where this quantity is measured, see Fig. 6.
Many stars move from north-east to south-west (including Fomalhaut),
suggesting a preferred direction of motion of stars in this field.
Therefore, an apparently co-moving object located a few parsec behind Fomalhaut (in the same Galactic spiral arm) is well possible.
\subsubsection{Gamma-ray detection~?}
Before we conclude let us also check whether the object might have been detected by some $\gamma$-ray detector.
We have cross-correlated the position of Fomalhaut b with all sources from BATSE and Fermi.
We found two positive possible correlations: \\
Fomalhaut b is located in the positional error ellipse of
a BATSE $\gamma$-ray burst (GRB) source, namely $0.34^{\circ}$ off the BATSE source 11591a at 2000-02-17T06:54:21,
which has a large positional error of $9.4^{\circ}$ and a very low flux of 0.106 photons cm$^{-2}$ s${-1}$
(Stern et al. 2001). \\
Fomalhaut b is also located $0.96^{\circ}$ off the BATSE source named 951022.99 inside its
large positional error ellipse ($11.3^{\circ}$); this source did not even lead to a follow-up trigger
(Kommers et al. 2001). \\
Given the large error bars of these two BATSE $\gamma$-sources, it is unlikely that Fomalhaut b is related to any of them.
Then, we have retrieved the FERMI LAT data ourselves to search for a source near Fomalhaut.
We used the data up to $15^{\circ}$ around Fomalhaut with a resolution of $0.1^{\circ}$ per pixel in
the energy range from 100 to 300,000 MeV. We detected only the two known sources 2FGL J2258.9-2759
(separation $97.8^{\prime}$) und 2FGL J2250.8-2808 (separation $126.5^{\prime}$) near Fomalhaut.
Then, we have subtracted those sources from the data, to search again for a faint remaining source,
but we could not detect anything near Fomalhaut.
There is no Fermi source or even Fermi pulsar anywhere near Fomalhaut b.
\section{Conclusions}
The faint object near Formalhaut A (called Fomalhaut~b) remains the subject of intense discussion.
If one assumes that Fomalhaut~b is gravitationally bound to Fomalhaut~A,
then the most likely hypothesis seems to be an expanding dust cloud without a central source of $\geq$10\,M$_{Earth}$.
We show that the body of observations (optical photometry, proper motion, X-ray non-detection) can in principal be fit with a NS.
In particular, we can explain the SED in the optical wavelength range and the non-detections in the near- and mid-infrared:
The available photometry allows good fits for a blackbody with temperature range from 112,000 to 126,500 K
(of the optical emitting area)
-- for a neutron star,
this temperature range would yield a distance of $\sim 11$ pc to remain undetected in X-rays as observed. While this may appear to be a fine-tuned
parameter range, one should also keep in mind that such parameters are consistent with {\em all} observables,
while the planetary interpretation has problems with two observational issues (ring-crossing orbit and non-detection in IR).
We also show that it is not unlikely to find one such faint, but unrelated object (or even one of $\sim 10^{8}$ Galactic NSs)
near one of the many bright stars surveyed with deep imaging for planets.
If Fomalhaut b is a neutron star rather than a planet, then the eccentricity of the dust ring around Fomalhaut A might
be explained by either one or more lower-mass, as yet undetected planet(s) or possibly by an eccentric stellar companion.
Given our rough distance estimate for Fomalhaut b as neutron star,
it might also be possible that Fomalhaut b is both a neutron star and a companion in the Fomalhaut system ($\sim 8$ pc),
i.e. that they orbit around each other - Fomalhaut b as neutron star currently located some tens to hundreds of au behind
(or before) Fomalhaut A. It should be less problematic for a neutron star component to orbit on an eccentric and/or inclined
orbit (than for a planetary companion).
In such a case, one might expect Bondi-Hoyle accretion of circumstellar material onto the NS and, hence, variable brightness,
but such a variable brightness would be hard to detect for an object as faint as $\sim 25$ mag with large error bars
(see Table 1 and Figs. 4 \& 5).
If the companion candidate to Fomalhaut is indeed a NS at some 11 pc distance, it would be the clostest known NS.
With a total of some $10^{8}$ NSs in the Galaxy ($\sim 30$ kpc diameter, $\sim 0.6$ kpc thickness),
we would expect $\sim 2.2$ NSs between 8 and 14 pc (i.e. $\pm 3$ pc around the best fit, the lower value being the lowest
allowed distance). We may have noticed one of them.
If we would restrict this estimate to young (self-luminous) NSs, the expected number would be smaller.
If the companion candidate to Fomalhaut is indeed a NS, then it may either be an exception
(as self-luminous very nearby NS) -- or it may be an old NS re-heated by accretion from the interstellar material.
For the latter, this NS would need to travel with small velocity $v$ through the interstellar material
(Bondi-Hoyle accretion scales with $v^{-3}$), which is indeed the case: $\sim 30$~km/s only as two-dimensional
velocity from its proper motion at $\sim 11$ pc, respectively, which is slow for NSs.
Our NS hypothesis could be tested by observations in the UV (e.g. with the ACS Solar Blind Camera):
If the object is detected and shows a similar flat SED as in the optical,
than a planet or dust cloud could be ruled out and a background neutron star becomes the most likely explanation.
If radio, X-ray, or gamma-ray pulsations (in the typical range as for NSs, i.e. few milli-seconds to $\sim 20$ seconds)
could be detected in the Fomalhaut companion candidate in very deep observations,
it would certainly be a neutron star (we would expect pulsations around $\sim 1$ to $\sim 10$ seconds,
rather than below one second, because it would need to be a middle-aged to old neutron star).
An X-ray detection would also yield a more stringent constraint on the age.
A detection of gravitational waves due to rotation of a non-spherical object would confirm its compactness.
\section*{Acknowledgments}
RN, MMH and JGS acknowledges support by the Deutsche Forschungsgemeinschaft (DFG) through SFB/TR 7 ``Gravitationswellenastronomie''.
We would like to thank Alexander Krivov and Markus Mugrauer for valueable comments on an earlier version of this manuscript.
CG wishes to acknowledge Deutsche Forschungsgemeinschaft (DFG) for grant MU 2695/ 13-1.
We obtained the AMES models from phoenix.ens-lyon.fr/Grids/AMES-Cond/SPECTRA.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
This research has made use of NASA's Astrophysics Data System Bibliographic Services.
We express special thanks to Donna Keeley for language editing of the paper.
We also would like to thank an anonymous referee for very good comments.
|
\section{Introduction}
Optimal control theory is concerned with the analysis of controlled dynamical systems, where one aims at steering such a system from a given configuration to some desired target by minimizing some criterion. The Pontryagin maximum principle (in short, PMP), established at the end of the 50's for general finite-dimensional nonlinear continuous-time dynamics (see \cite{pont}, and see \cite{gamk} for the history of this discovery), is certainly the milestone of the classical optimal control theory. It provides a first-order necessary condition for optimality, by asserting that any optimal trajectory must be the projection of an extremal. The PMP then reduces the search of optimal trajectories to a boundary value problem posed on extremals. Optimal control theory, and in particular the PMP, has an immense field of applications in various domains, and it is not our aim here to list them.
\medskip
We speak of a {\em purely continuous-time optimal control problem}, when both the state $q$ and the control $u$ evolve continuously in time, and the control system under consideration has the form
$$
\dot q(t) = f(t,q(t),u(t)), \qquad \text{for a.e. } t \in \mathbb{R}^+,
$$
where $q(t) \in \mathbb{R}^n$ and $u(t) \in \Omega \subset \mathbb{R}^m$. Such models assume that the control is permanent, that is, the value of $u(t)$ can be chosen at each time $t \in \mathbb{R}^+$. We refer the reader to textbooks on continuous optimal control theory such as \cite{agrach,Bon-Chy03,trel2,bres,brys,BulloLewis,hest,Jurdjevic,lee,pont,Schattler,seth,trel} for many examples of theoretical or practical applications.
\medskip
We speak of a {\em purely discrete-time optimal control problem}, when both the state $q$ and the control $u$ evolve in a discrete way in time, and the control system under consideration has the form
$$
q_{k+1}-q_k = f(k,q_k,u_k), \qquad k \in \mathbb{N},
$$
where $q_k \in \mathbb{R}^n$ and $u_k \in \Omega \subset \mathbb{R}^m$. As in the continuous case, such models assume that the control is permanent, that is, the value of $u_k$ can be chosen at each time $k \in \mathbb{N}$. A version of the PMP for such discrete-time control systems has been established in \cite{Halkin,holt2,holt} under appropriate convexity assumptions. The considerable development of the discrete-time control theory was in particular motivated by the need of considering digital systems or discrete approximations in numerical simulations of differential control systems (see the textbooks \cite{bolt,cano,mord,seth}).
It can be noted that some early works devoted to the discrete-time PMP (like \cite{fan}) are mathematically incorrect. Some counterexamples were provided in \cite{bolt} (see also \cite{mord}), showing that, as is now well known, the exact analogous of the continuous-time PMP does not hold at the discrete level. More precisely, the maximization condition of the continuous-time PMP cannot be expected to hold in general in the discrete-time case. Nevertheless, a weaker condition can be derived, in terms of nonpositive gradient condition (see \cite[Theorem 42.1]{bolt}).
\medskip
We speak of an {\em optimal sampled-data control problem}, when the state $q$ evolves continuously in time, whereas the control $u$ evolves in a discrete way in time. This hybrid situation is often considered in practice for problems in which the evolution of the state is very quick (and thus can be considered continuous) with respect to that of the control. We often speak, in that case, of {\em digital control}. This refers to a situation where, due for instance to hardware limitations or to technical difficulties, the value $u(t)$ of the control can be chosen only at times $t=kT$, where $T>0$ is fixed and $k \in \mathbb{N}$. This means that, once the value $u(kT)$ is fixed, $u(t)$ remains constant over the time interval $[kT,(k+1)T)$. Hence the trajectory $q$ evolves according to
$$
\dot q(t) = f(t,q(t),u(kT)), \qquad \text{for a.e. } t \in [kT,(k+1)T),\quad k\in\mathbb{N}.
$$
In other words, this {\em sample-and-hold} procedure consists of ``freezing" the value of $u$ at each \textit{controlling time} $t=kT$ on the corresponding \textit{sampling time interval} $[kT,(k+1)T)$, where $T$ is called the \textit{sampling period}. In this situation, the control of the system is clearly nonpermanent.
\medskip
To the best of our knowledge, the classical optimal control theory does not treat general nonlinear optimal sampled-data control problems, but concerns either purely continuous-time, or purely discrete-time optimal control problems.
It is one of our objectives to derive, in this paper, a PMP which can be applied to general nonlinear optimal sampled-data control problems.
Actually, we will be able to establish a PMP in the much more general framework of {\em time scales}, which unifies and extends continuous-time and discrete-time issues.
But, before coming to that point, we feel that it is of interest to enunciate a PMP in the particular case of sampled-data systems and where the set $\Omega$ of pointwise constraints on the controls is convex.
\paragraph{PMP for optimal sampled-data control problems and $\Omega$ convex.}
Let $n$, $m$ and $j$ be nonzero integers. Let $T>0$ be an arbitrary sampling period. In what follows, for any real number $t$, we denote by $E(t)$ the integer part of $t$, defined as the unique integer such that $E(t)\leq t< E(t)+1$. Note that $k = E(t/T)$ whenever $kT\leq t<(k+1)T$.
We consider the general nonlinear optimal sampled-data control problem
\begin{equation*}
{\bf (OSDCP)}^{\R^+}_{\N T} \qquad \left\{\begin{split}
& \min \; \int_0^{t_f} f^0 ( \tau, q(\tau), u( k(\tau) T ) ) \, d\tau , \quad \text{with } k(\tau) = E(\tau /T), \\
& \dot q(t) = f ( t,q(t), u( k(t) T ) ) , \quad \text{with } k(t) = E(t /T), \\
& u(kT) \in \Omega , \\
& g(q(0),q(t_f)) \in \mathrm{S} .
\end{split}\right.
\end{equation*}
Here, $f: \mathbb{R}\times \mathbb{R}^{n} \times \mathbb{R}^{m} \rightarrow \mathbb{R}^{n}$ and $f^0: \mathbb{R}\times \mathbb{R}^{n} \times \mathbb{R}^{m} \rightarrow \mathbb{R}$ are continuous, and of class $\mathrm{C}^1$ in $(q,u)$,
$g: \mathbb{R}^n \times \mathbb{R}^n \rightarrow \mathbb{R}^j$ is of class $\mathrm{C}^1$,
and $\Omega$ (resp., $\mathrm{S}$) is a non-empty closed convex subset of $\mathbb{R}^m$ (resp., of $\mathbb{R}^j$).
The final time $t_f \geq 0$ can be fixed or not.
Note that, under appropriate (usual) compactness and convexity assumptions, the optimal control problem ${\bf (OSDCP)}^{\R^+}_{\N T}$ has at least one solution (see Theorem~\ref{propexistence} in Section~\ref{subsectionOSCP}).
Recall that $g$ is said to be submersive at a point $(q_1,q_2) \in \mathbb{R}^n \times \mathbb{R}^n$ if the differential of $g$ at this point is surjective. We define the Hamiltonian $H:\mathbb{R}\times\mathbb{R}^n \times \mathbb{R}^n \times \mathbb{R} \times \mathbb{R}^m\rightarrow \mathbb{R}$, as usual, by
$H(t,q,p,p^0,u) = \langle p, f(t,q,u) \rangle_{\mathbb{R}^n} + p^0 f^0(t,q,u)$.
\begin{theorem}[Pontryagin maximum principle for ${\bf (OSDCP)}^{\R^+}_{\N T}$]\label{thmmainintro}
If a trajectory $q^*$, defined on $[0,t^*_f]$ and associated with a sampled-data control $u^*$, is an optimal solution of ${\bf (OSDCP)}^{\R^+}_{\N T}$, then there exists a nontrivial couple $(p,p^0)$, where $p : [0,t^*_f] \rightarrow \mathbb{R}^n$ is an absolutely continuous mapping (called adjoint vector) and $p^0 \leq 0$, such that the following conditions hold:
\begin{itemize}
\item Extremal equations:
\begin{equation*}
\dot q^*(t) = \frac{\partial H}{\partial p} (t,q^*(t),p(t),p^0,u^*(k(t)T)), \qquad
\dot p(t) = - \frac{\partial H}{\partial q} (t,q^*(t),p(t),p^0,u^*(k(t)T)) ,
\end{equation*}
for almost every $t\in[0,t^*_f)$, with $k(t)=E(t/T)$.
\item Maximization condition:\\
For every controlling time $kT \in [0,t^*_f)$ such that $(k+1)T \leq t^*_f$, we have
\begin{equation}\label{secondconditionintro}
\left\langle \int_{kT}^{(k+1)T} \frac{\partial H}{\partial u} (\tau,q^*(\tau),p(\tau),p^0,u^*(kT)) \; d\tau \; , \; y-u^*(kT) \right\rangle_{\mathbb{R}^m} \leq 0,
\end{equation}
for every $y \in \Omega$. In the case where $kT \in [0,t^*_f)$ with $(k+1)T > t^*_f$, the above maximization condition is still valid provided $(k+1)T$ is replaced with $t^*_f$.
\item Transversality conditions on the adjoint vector:\\
If $g$ is submersive at $(q^*(0),q^*(t^*_f))$, then the nontrivial couple $(p,p^0)$ can be selected to satisfy
\begin{equation*}
p(0) = - \left( \frac{\partial g}{\partial q_1} (q^*(0),q^*(t^*_f)) \right)^\top \psi,\qquad
p(t^*_f) = \left( \frac{\partial g}{\partial q_2} (q^*(0),q^*(t^*_f)) \right)^\top \psi,
\end{equation*}
where $-\psi$ belongs to the orthogonal of $\mathrm{S}$ at the point $g (q^*(0),q^*(t^*_f)) \in \mathrm{S}$.
\item Transversality condition on the final time:\\
If the final time is left free in the optimal control problem ${\bf (OSDCP)}^{\R^+}_{\N T}$, if $t^*_f>0$ and if $f$ and $f^0$ are of class $\mathrm{C}^1$ with respect to $t$ in a neighborhood of $t^*_f$, then the nontrivial couple $(p,p^0)$ can be moreover selected to satisfy
\begin{equation*}
H(t^*_f, q^*(t^*_f), p(t^*_f),p^0,u^*(k^*_f T) ) = 0,
\end{equation*}
where $k^*_f=E(t^*_f/T)$ whenever $t^*_f \notin \mathbb{N} T$, and $k^*_f=E(t^*_f/T)-1$ whenever $t^*_f \in \mathbb{N} T$.
\end{itemize}
\end{theorem}
Note that the only difference with the usual statement of the PMP for purely continuous-time optimal control problems is in the maximization condition. Here, for sampled-data control systems, the usual pointwise maximization condition of the Hamiltonian is replaced with the inequality~\eqref{secondconditionintro}.
This is not a surprise, because already in the purely discrete case, as mentioned earlier, the pointwise maximization condition fails to be true in general, and must be replaced with a weaker condition.
The condition~\eqref{secondconditionintro}, which is satisfied for every $y \in \Omega$, gives a necessary condition allowing to compute $u^*(kT)$ in general, and this, for all controlling times $kT \in [0,t^*_f)$. We will provide in Section~\ref{sectionexemple} a simple optimal consumption problem with sampled-data control, and show how these computations can be done in a simple way.
Note that the optimal sampled-data control problem ${\bf (OSDCP)}^{\R^+}_{\N T}$ can of course be seen as a finite-dimensional optimization problem where the unknowns are $u^*(kT)$, with $k \in \mathbb{N}$ such that $kT \in [0,t^*_f)$. The same remark holds, by the way, for purely discrete-time optimal control problems. One could then apply classical Lagrange multiplier (or KKT) rules to such optimization problems with constraints (numerically, this leads to direct methods). The Pontryagin maximum principle is a far-reaching version of the Lagrange multiplier rule, yielding more precise information and reducing the initial optimal control problem to a shooting problem (see, \textit{e.g.}, \cite{Trelat_JOTA2012} for such a discussion).
\paragraph{Extension to the time scale framework.}
In this paper, we actually establish a version of Theorem~\ref{thmmainintro} in a much more general framework, allowing for example to study sampled-data control systems where the control can be permanent on a first time interval, then sampled on a finite set, then permanent again, etc.
More precisely, Theorem~\ref{thmmainintro} can be extended to a general framework in which the set of controlling times is not $\mathbb{N} T$ but some arbitrary non-empty closed subset of $\mathbb{R}$ (\textit{i.e.}, a \textit{time scale}), and also the state may evolve on another time scale.
We will state a PMP for such general systems in Section~\ref{secPMP} (see Theorem~\ref{thmmain}). Since such systems, where we have two time scales (one for the state and one for the control), can be viewed as a generalization of sampled-data control systems, we will refer to them as {\em sampled-data control systems on time scales}.
Let us first recall and motivate the notion of time scale. The time scale theory was introduced in \cite{hilg} in order to unify discrete and continuous analysis. By definition, a time scale $\mathbb{T}$ is an arbitrary non-empty closed subset of $\mathbb{R}$, and a dynamical system is said to be posed on the time scale $\mathbb{T}$ whenever the time variable evolves along this set $\mathbb{T}$. The time scale theory aims at closing the gap between continuous and discrete cases, and allows one to treat general processes involving both continuous-time and discrete-time variables. The purely continuous-time case corresponds to $\mathbb{T}=\mathbb{R}^+$ and the purely discrete-time case corresponds to $\mathbb{T}=\mathbb{N}$. But a time scale can be much more general (see, \textit{e.g.}, \cite{gama,may} for a study of a seasonally breeding population whose generations do not overlap, and see \cite{ati} for applications to economics), and can even be a Cantor set.
Many notions of standard calculus have been extended to the time scale framework, and we refer the reader to \cite{agar2,agar3,bohn,bohn3} for details on that theory.
The theory of the calculus of variations on time scales, initiated in \cite{bohn2}, has been well studied in the existing literature (see, \textit{e.g.}, \cite{torr8,bohn4,bour12,torr9,hils,hils3}). In \cite{hils1,hils2}, the authors establish a weak version of the PMP (with a nonpositive gradient condition) for control systems defined on general time scales. In \cite{bour11}, we derived a strong version of the PMP, in a very general time scale setting, encompassing both the purely continuous-time PMP (with a maximization condition) and the purely discrete-time PMP (with a nonpositive gradient condition).
All these works are concerned with control systems defined on general time scales with permanent control. The main objective of the present paper is to handle control systems defined on general time scales with nonpermanent control, that we refer to as {\em sampled-data control systems on time scales}, and for which we assume that the state and the control are allowed to evolve on different time scales (the time scale of the control being a subset of the time scale of the state). This framework is the natural extension of the classical sampled-data setting, and allows to treat simultaneously many sampling-data control situations.
Our main result is a {\em PMP for general finite-dimensional nonlinear optimal sampled-data control problems on time scales}. Note that our result will be derived without any convexity assumption on the set $\Omega$ of pointwise constraints on the controls. Our proof is based on the construction of appropriate needle-like variations and on the Ekeland variational principle. In the case of a permanent control, our statement encompasses the time scale version of the PMP obtained in \cite{bour11}, and \textit{a fortiori} it also encompasses the classical continuous and discrete versions of the PMP.
\paragraph{Organization of the paper.}
In Section~\ref{section1}, after having recalled several basic facts in time scale calculus, we define a general nonlinear optimal sampled-data control problem defined on time scales, and we state a Pontryagin maximum principle (Theorem~\ref{thmmain}) for such problems.
Section~\ref{sectionapplication} is devoted to some applications of Theorem~\ref{thmmain} and further comments. Section~\ref{annexe} is devoted to the proof of Theorem~\ref{thmmain}.
\section{Main result}\label{section1}
Let $\mathbb{T}$ be a time scale, that is, an arbitrary non-empty closed subset of $\mathbb{R}$. Without loss of generality, we assume that $\mathbb{T}$ is bounded below, denoting by $a = \min \mathbb{T} $, and unbounded above.\footnote{In this paper we only work on a bounded subinterval of type $[a,b] \cap \mathbb{T}$ with $a$, $b \in \mathbb{T}$. It is not restrictive to assume that $a = \min \mathbb{T} $ and that $\mathbb{T}$ is unbounded above. On the other hand, these two assumptions widely simplify the notations introduced in Section~\ref{secprelimtimescale} (otherwise we would have, in all further statements, to distinguish between points of $\mathbb{T} \backslash \{ \max\mathbb{T} \}$ and $\max\mathbb{T}$).}
Throughout the paper, $\mathbb{T}$ will be the time scale on which the state of the control system evolves.
We start the section by recalling some useful notations and basic results of time scale calculus, in particular the notion of Lebesgue $\Delta$-measure and of absolutely continuous function within the time scale setting. The reader already acquainted with time scale calculus may jump directly to Section~\ref{subsectionOSCP}.
\subsection{Preliminaries on time scale calculus}\label{secprelimtimescale}
The forward jump operator $\sigma : \mathbb{T} \rightarrow \mathbb{T}$ is defined by $\sigma (t) = \inf \{ s \in \mathbb{T}\ \vert\ s > t \}$ for every $t \in \mathbb{T}$. A point $t \in \mathbb{T}$ is said to be \textit{right-scattered} whenever $\sigma (t) > t$. A point $t \in \mathbb{T}$ is said to be \textit{right-dense} whenever $\sigma (t) = t$. We denote by $\mathrm{RS}$ the set of all right-scattered points of $\mathbb{T}$, and by $\mathrm{RD}$ the set of all right-dense points of $\mathbb{T}$.
Note that $\mathrm{RS}$ is at most countable (see \cite[Lemma 3.1]{caba}) and that $\mathrm{RD}$ is the complement of $\mathrm{RS}$ in $\mathbb{T}$. The graininess function $\mu : \mathbb{T} \rightarrow \mathbb{R}^+$ is defined by $\mu(t) = \sigma (t) -t$ for every $t \in \mathbb{T}$.
For every subset $A$ of $\mathbb{R}$, we denote by $A_\mathbb{T} = A \cap \mathbb{T}$. An interval of $\mathbb{T}$ is defined by $I_\mathbb{T}$ where $I$ is an interval of $\mathbb{R}$. For every $b \in \mathbb{T} \backslash \{ a \}$ and every $s \in [a,b)_\mathbb{T} \cap \mathrm{RD}$, we set
\begin{equation}\label{defVbs}
\mathcal{V}^{s,b} = \{ \beta \geq 0 \mid s+\beta \in [s,b]_\mathbb{T} \}.
\end{equation}
Note that $0$ is not isolated in $\mathcal{V}^{s,b}$.
\paragraph{$\Delta$-differentiability.}
Let $n \in \mathbb{N}^*$. The notations $\Vert \cdot \Vert_{\mathbb{R}^n}$ and $\langle \cdot, \cdot \rangle_{\mathbb{R}^n}$ respectively stand for the usual Euclidean norm and scalar product of $\mathbb{R}^n$. A function $q : \mathbb{T} \rightarrow \mathbb{R}^n$ is said to be $\Delta$-differentiable at $t \in \mathbb{T}$ if the limit
$$q^\Delta (t) = \lim\limits_{\substack{s \to t \\ s \in \mathbb{T}}} \frac{q^\sigma (t) -q(s)}{\sigma (t) -s} $$
exists in $\mathbb{R}^n$, where $q^\sigma = q \circ \sigma $. Recall that, if $t \in \mathrm{RD}$, then $q$ is $\Delta$-differentiable at $t$ if and only if the limit of $\frac{q(t)-q(s)}{t-s}$ as $s \to t$, $s \in \mathbb{T}$, exists; in that case it is equal to $q^\Delta (t)$.
If $t \in \mathrm{RS}$ and if $q$ is continuous at $t$, then $q$ is $\Delta$-differentiable at $t$, and $q^\Delta (t) = \frac{q^\sigma(t) - q(t)}{\mu (t)}$ (see \cite{bohn}).
If $q$, $q': \mathbb{T} \rightarrow \mathbb{R}^n$ are both $\Delta$-differentiable at $t \in \mathbb{T}$, then the scalar product $\langle q , q' \rangle_{\mathbb{R}^n}$ is $\Delta$-differentiable at $t$ and
\begin{equation}\label{eqleibniz}
\langle q, q' \rangle_{\mathbb{R}^n}^\Delta (t) = \langle q^\Delta (t), q'^\sigma(t) \rangle_{\mathbb{R}^n} + \langle q (t), q'^\Delta(t) \rangle_{\mathbb{R}^n} = \langle q^\Delta (t), q'(t) \rangle_{\mathbb{R}^n} + \langle q^\sigma (t), q'^\Delta(t) \rangle_{\mathbb{R}^n}.
\end{equation}
These equalities are usually called Leibniz formulas (see \cite[Theorem 1.20]{bohn}).
\paragraph{Lebesgue $\Delta$-measure and Lebesgue $\Delta$-integrability.}
Let $\mu_\Delta$ be the Lebesgue $\Delta$-measure on $\mathbb{T}$ defined in terms of Carath\'eodory extension in \cite[Chapter 5]{bohn3}. We also refer the reader to \cite{agar,aulb,caba,guse} for more details on the $\mu_\Delta$-measure theory. For all $(c,d)\in\mathbb{T}^2$ such that $c \leq d $, one has $\mu_\Delta ([c,d)_\mathbb{T}) = d-c$. Recall that $A \subset \mathbb{T}$ is a $\mu_\Delta$-measurable set of $\mathbb{T}$ if and only if $A$ is an usual $\mu_L$-measurable set of $\mathbb{R}$, where $\mu_L$ denotes the usual Lebesgue measure (see \cite[Proposition 3.1]{caba}). Moreover, if $A \subset \mathbb{T} $, then
$$\mu_\Delta ( A ) = \mu_L (A) + \sum_{r \in A \cap \mathrm{RS}} \mu (r).$$
Let $A \subset \mathbb{T}$. A property is said to hold $\Delta$-almost everywhere (in short, $\Delta$-a.e.) on $A$ if it holds for every $t \in A \backslash A'$, where $A' \subset A$ is some $\mu_\Delta$-measurable subset of $\mathbb{T}$ satisfying $\mu_\Delta (A') = 0$. In particular, since $\mu_\Delta (\{ r \}) = \mu (r) > 0$ for every $r \in \mathrm{RS}$, we conclude that, if a property holds $\Delta$-a.e. on $A$, then it holds for every $r \in A \cap \mathrm{RS}$.
Similarly, if $A \subset \mathbb{T}$ is such that $\mu_\Delta (A) = 0$, then $A \subset \mathrm{RD}$.
Let $n \in \mathbb{N}^*$ and let $A \subset \mathbb{T}$ be a $\mu_\Delta$-measurable subset of $\mathbb{T}$. Consider a function $q$ defined $\Delta$-a.e.\ on $A$ with values in $\mathbb{R}^n$. Let $\tilde{A}=A \cup (r,\sigma(r))_{r \in A \cap \mathrm{RS}}$, and let $\tilde{q}$ be the extension of $q$ defined $\mu_L$-a.e.\ on $\tilde{A}$ by $\tilde{q} (t)=q(t)$ whenever $t \in A$, and by $\tilde{q}(t)=q(r)$ whenever $t \in (r,\sigma(r))$, for every $r \in A \cap \mathrm{RS}$.
Recall that $q$ is $\mu_\Delta$-measurable on $A$ if and only if $\tilde{q}$ is $\mu_L$-measurable on $\tilde{A}$ (see \cite[Proposition 4.1]{caba}).
Let $n \in \mathbb{N}^*$ and let $A \subset \mathbb{T}$ be a $\mu_\Delta$-measurable subset of $\mathbb{T}$. The functional space $\mathrm{L}^\infty_\mathbb{T} (A,\mathbb{R}^n)$ is the set of all functions $q$ defined $\Delta$-a.e.\ on $A$, with values in $\mathbb{R}^n$, that are $\mu_\Delta$-measurable on $A$ and bounded $\Delta$-almost everywhere. Endowed with the norm $\Vert q \Vert_{\mathrm{L}^\infty_\mathbb{T} (A,\mathbb{R}^n)} = \supess_{\tau \in A} \Vert q(\tau) \Vert_{\mathbb{R}^n}$, it is a Banach space (see \cite[Theorem 2.5]{agar}). The functional space $\mathrm{L}^1_\mathbb{T} (A,\mathbb{R}^n)$ is the set of all functions $q$ defined $\Delta$-a.e.\ on $A$, with values in $\mathbb{R}^n$, that are $\mu_\Delta$-measurable on $A$ and such that
$\int_{A} \Vert q(\tau) \Vert_{\mathbb{R}^n} \, \Delta \tau < + \infty$.
Endowed with the norm $\Vert q \Vert_{\mathrm{L}^1_\mathbb{T} (A,\mathbb{R}^n)} = \int_{A} \Vert q(\tau) \Vert_{\mathbb{R}^n} \, \Delta \tau$, it is a Banach space (see \cite[Theorem 2.5]{agar}).
We recall here that if $q \in \mathrm{L}^1_\mathbb{T} (A,\mathbb{R}^n)$ then
$$ \int_{A} q(\tau) \,\Delta \tau = \int_{\tilde{A}} \tilde{q}(\tau) \,d\tau = \int_{A} q(\tau) \,d\tau + \sum_{r \in A \cap \mathrm{RS}} \mu (r) q(r), $$
see \cite[Theorems 5.1 and 5.2]{caba}. Note that if $A$ is bounded then $\mathrm{L}^\infty_\mathbb{T} (A,\mathbb{R}^n) \subset \mathrm{L}^1_\mathbb{T} (A,\mathbb{R}^n)$.
\paragraph{Absolutely continuous functions.}
Let $n \in \mathbb{N}^*$ and let $(c,d) \in \mathbb{T}^2$ such that $c < d$. Let $\mathrm{C} ([c,d]_\mathbb{T},\mathbb{R}^n)$ denote the space of continuous functions defined on $[c,d]_\mathbb{T}$ with values in $\mathbb{R}^n$. Endowed with its usual uniform norm $\Vert \cdot \Vert_\infty$, it is a Banach space. Let $\mathrm{AC}([c,d]_\mathbb{T},\mathbb{R}^n)$ denote the subspace of absolutely continuous functions.
Let $t_0 \in [c,d]_\mathbb{T}$ and $q:[c,d]_\mathbb{T} \rightarrow \mathbb{R}^n$.
It is easily derived from \cite[Theorem 4.1]{caba2} that $q \in \mathrm{AC}([c,d]_\mathbb{T},\mathbb{R}^n)$ if and only if $q$ is $\Delta$-differentiable $\Delta$-a.e.\ on $[c,d)_\mathbb{T}$ and satisfies $q^\Delta \in \mathrm{L}^1_\mathbb{T} ([c,d)_\mathbb{T},\mathbb{R}^n)$, and for every $t \in [c,d]_\mathbb{T}$ one has
$ q(t) = q(t_0) + \int_{[t_0,t)_\mathbb{T}} q^\Delta (\tau) \, \Delta \tau$
whenever $t \geq t_0$, and
$ q(t) = q(t_0) - \int_{[t,t_0)_\mathbb{T}} q^\Delta (\tau) \, \Delta \tau $
whenever $ t \leq t_0$.
Assume that $q \in \mathrm{L}^1_\mathbb{T} ([c,d)_\mathbb{T},\mathbb{R}^n)$, and let $Q$ be the function defined on $[c,d]_\mathbb{T}$ by $Q(t) = \int_{[t_0,t)_\mathbb{T}} q (\tau) \,\Delta \tau$ whenever $ t \geq t_0$, and by $ Q(t) = - \int_{[t,t_0)_\mathbb{T}} q (\tau) \,\Delta \tau $ whenever $t \leq t_0$.
Then $Q \in \mathrm{AC}([c,d]_\mathbb{T})$ and $Q^\Delta = q$ $\Delta$-a.e.\ on $[c,d)_\mathbb{T}$.
Note that, if $q \in \mathrm{AC}([c,d]_\mathbb{T}, \mathbb{R}^n )$ is such that $q^\Delta = 0$ $\Delta$-a.e.\ on $[c,d)_\mathbb{T}$, then $q$ is constant on $[c,d]_\mathbb{T}$, and that, if $q$, $q' \in \mathrm{AC}([c,d]_\mathbb{T}, \mathbb{R}^n )$, then $\langle q,q' \rangle_{\mathbb{R}^n} \in \mathrm{AC} ([c,d]_\mathbb{T}, \mathbb{R} )$ and the Leibniz formula~\eqref{eqleibniz} is available $\Delta$-a.e.\ on $[c,d)_\mathbb{T}$.
For every $q \in \mathrm{L}^1_\mathbb{T} ([c,d)_\mathbb{T},\mathbb{R}^n)$, we denote by $\mathscr{L}_{[c,d)_\mathbb{T}} (q)$ the set of points $t \in [c,d)_\mathbb{T}$ that are $\Delta$-Lebesgue points of $q$. It holds $\mu_\Delta (\mathscr{L}_{[c,d)_\mathbb{T}}(q)) = \mu_\Delta ([c,d)_\mathbb{T}) = d-c$, and
\begin{equation*}
\lim\limits_{\substack{\beta \to 0 \\ \beta \in \mathcal{V}^{s,d}}} \frac{1}{\beta} \int_{[s,s+\beta)_\mathbb{T}} q(\tau) \, \Delta \tau = q(s),
\end{equation*}
for every $s \in \mathscr{L}_{[c,d)_\mathbb{T}}(q) \cap \mathrm{RD}$, where $\mathcal{V}^{s,d}$ is defined by~\eqref{defVbs}.
\subsection{Optimal sampled-data control problems on time scales}\label{subsectionOSCP}
Let ${\mathbb{T}_1}$ be another time scale. Throughout the paper, ${\mathbb{T}_1}$ will be the time scale on which the control evolves. We assume that ${\mathbb{T}_1} \subset \mathbb{T}$.\footnote{Indeed, it is not natural to consider controlling times $t \in {\mathbb{T}_1}$ at which the dynamics does not evolve, that is, at which $t \notin \mathbb{T}$. The value of the control at such times $t \in {\mathbb{T}_1} \backslash \mathbb{T}$ would not influence the dynamics, or, maybe, only on $[t^* ,+\infty[_\mathbb{T}$ where $t^*= \inf \{ s \in \mathbb{T} \,| \,s \geq t \}$. In this last case, note that $t^* \in \mathbb{T}$ and we can replace ${\mathbb{T}_1}$ by $({\mathbb{T}_1} \cup \{ t^* \} ) \backslash \{ t \}$ without loss of generality.}
Similarly to $\mathbb{T}$, we assume that $\min {\mathbb{T}_1} = a$ and that ${\mathbb{T}_1}$ is unbounded above. As in the previous paragraph, we introduce the notations $\sigma_1$, $\mathrm{RS}_1$, $\mathrm{RD}_1$, $\mathcal{V}^{s,b}_1$, $\Delta_1$, etc., associated with the time scale ${\mathbb{T}_1}$. Since ${\mathbb{T}_1} \subset \mathbb{T}$, note that $\mathrm{RS} \subset \mathrm{RS}_1$ and $\mathrm{RD}_1 \subset \mathrm{RD}$. We define the map
$$ \fonction{\Phi}{\mathbb{T}}{{\mathbb{T}_1}}{t}{\Phi(t) = \sup \,\{ s \in {\mathbb{T}_1} \,| \,s \leq t \} .} $$
For every $t \in {\mathbb{T}_1}$, we have $\Phi(t)=t$. For every $t \in \mathbb{T} \backslash {\mathbb{T}_1}$, we have $\Phi(t) \in \mathrm{RS}_1$ and $\Phi(t) < t < \sigma_1(\Phi(t))$. Note that, if $t \in \mathbb{T}$ is such that $\Phi(t) \in \mathrm{RD}_1$, then $t \in {\mathbb{T}_1}$.
In what follows, given a function $u:{\mathbb{T}_1}\rightarrow\mathbb{R}$, we denote by $u^\Phi$ the composition $u \circ \Phi : \mathbb{T} \rightarrow \mathbb{R}$. Of course, when dealing with functions having multiple components, this composition is applied to each component. Let us mention, at this step, that if $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\mathbb{R})$ then $u^\Phi \in \mathrm{L}^\infty_{\mathbb{T}} (\mathbb{T},\mathbb{R})$ (see Proposition~\ref{propnumber1} and more properties in Section~\ref{annexe1uetuphi}).
\medskip
Let $n$, $m$ and $j$ be nonzero integers. We consider the general nonlinear optimal sampled-data control problem on time scales
\begin{align}[left={\bf (OSDCP)}^{\T}_{\TU}\qquad\empheqlbrace]
& \min \int_{[a,b)_\mathbb{T}} f^0 (\tau,q (\tau), u^\Phi(\tau) ) \, \Delta \tau , \notag \\
& q^\Delta(t) = f (t,q(t),u^\Phi(t)), \label{DD-CS} \\
& u\in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\Omega) , \notag \\
& g(q(a),q(b)) \in \mathrm{S} . \notag
\end{align}
Here, the trajectory of the system is $q : \mathbb{T} \rightarrow \mathbb{R}^n$, the mappings $f:\mathbb{T}\times\mathbb{R}^{n} \times \mathbb{R}^{m} \rightarrow \mathbb{R}^{n}$ and $f^0: \mathbb{T}\times \mathbb{R}^{n} \times \mathbb{R}^{m} \rightarrow \mathbb{R}$ are continuous, of class $\mathrm{C}^1$ in $(q,u)$, the mapping $g: \mathbb{R}^n \times \mathbb{R}^n \rightarrow \mathbb{R}^j$ is of class $\mathrm{C}^1$, $\Omega$ is a non-empty closed subset of $\mathbb{R}^m$, and $\mathrm{S}$ is a non-empty closed convex subset of $\mathbb{R}^j$.
The final time $b \in \mathbb{T}$ can be fixed or not.
\begin{remark}
We recall that, given $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\mathbb{R}^m)$, we say that $q$ is a solution of~\eqref{DD-CS} on $I_\mathbb{T}$ if:
\begin{enumerate}
\item $I_\mathbb{T}$ is an interval of $\mathbb{T}$ satisfying $a \in I_\mathbb{T}$ and $I_\mathbb{T} \backslash \{ a \} \neq \emptyset$;
\item For every $c \in I_\mathbb{T} \backslash \{ a \}$, $q \in \mathrm{AC} ([a,c]_\mathbb{T},\mathbb{R}^n)$ and~\eqref{DD-CS} holds for $\Delta$-a.e. $t \in [a,c)_\mathbb{T}$.
\end{enumerate}
Existence and uniqueness of solutions (Cauchy-Lipschitz theorem on time scales) have been established in \cite{bour10}, and useful results are recalled in Section~\ref{recallCL}.
\end{remark}
\begin{remark}
The time scale ${\mathbb{T}_1}$ stands for the set of controlling times of the control system~\eqref{DD-CS}. If $\mathbb{T} = \mathbb{T}_1$, then the control is permanent. The case $\mathbb{T} = \mathbb{T}_1 = \mathbb{R}^+$ corresponds to the classical continuous case, whereas $\mathbb{T} = \mathbb{T}_1 = \mathbb{N}$ coincides with the classical discrete case. If $\mathbb{T}_1 \varsubsetneq \mathbb{T}$, the control is nonpermanent and sampled. In that case, the sampling times are given by $t \in \mathrm{RS}_1$ such that $\sigma(t) < \sigma_1 (t)$ and the corresponding sampling time intervals are given by $[t,\sigma_1(t))_\mathbb{T}$. The classical optimal sampled-data control problem ${\bf (OSDCP)}^{\R^+}_{\N T}$ investigated in Theorem~\ref{thmmainintro} corresponds to $\mathbb{T}=\mathbb{R}^+$ and ${\mathbb{T}_1}=\mathbb{N} T$, with $T >0$.
\end{remark}
\begin{remark}\label{remarkinclusiondeuxT1}
Let us consider two optimal control problems ${\bf (OSDCP)}^{\T}_{\TU}$ and ${\bf (OSDCP)}^{\T}_{\T_2}$, posed on the same general time scale $\mathbb{T}$ for the trajectories, but with two different sets of controlling times ${\mathbb{T}_1}$ and $\mathbb{T}_2$, and let us assume that $\mathbb{T}_2 \subset {\mathbb{T}_1}$. We denote by $\Phi_1$ and $\Phi_2$ the corresponding mappings from $\mathbb{T}$ to ${\mathbb{T}_1}$ and $\mathbb{T}_2$ respectively. If $u_1 \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\Omega)$ is an optimal control for ${\bf (OSDCP)}^{\T}_{\TU}$ and if there exists $u_2 \in \mathrm{L}^\infty_{\mathbb{T}_2} (\mathbb{T}_2,\Omega)$ such that $u_2^{\Phi_2} (t) = u_1^{\Phi_1} (t)$ for $\Delta$-a.e. $t \in \mathbb{T}$, it is clear that $u_2$ is an optimal control for ${\bf (OSDCP)}^{\T}_{\T_2}$. We refer to Section~\ref{sectionexemple} for examples.
\end{remark}
\begin{remark}\label{remarkparam}
The framework of ${\bf (OSDCP)}^{\T}_{\TU}$ encompasses optimal parameter problems. Indeed, let us consider the parametrized dynamical system
\begin{equation}\label{eqparam}
q^\Delta(t) = f (t,q(t),\lambda), \quad \Delta\text{-a.e. } t \in \mathbb{T},
\end{equation}
with $\lambda \in \Omega$. Then, considering $\mathbb{T}_1 = \{ a \} \cup [b,+\infty[_{\mathbb{T}}$, \eqref{DD-CS} coincides with \eqref{eqparam} where $u(a)$ plays the role of $\lambda$. In this situation, Theorem~\ref{thmmain} (stated in Section~\ref{secPMP}) provides necessary conditions for optimal parameters $\lambda$. We refer to Section~\ref{sectionexemple} for examples.
\end{remark}
\begin{remark}\label{remmultiscale}
A possible extension is to study dynamical systems on time scales with several sampled-data controls but with different sets of controlling times:
$$ q^\Delta(t) = f ( t ,q(t), u_1^{\Phi_1}(t),u_2^{\Phi_2} (t) ), \quad \Delta\text{-a.e. } t \in \mathbb{T}, $$
where $\mathbb{T}_1$ and $\mathbb{T}_2$ are general time scales contained in $\mathbb{T}$, and $\Phi_1$ and $\Phi_2$ are the corresponding mappings from $\mathbb{T}$ to ${\mathbb{T}_1}$ and $\mathbb{T}_2$. Our main result (Theorem~\ref{thmmain}) can be easily extended to this framework. Actually, this \textit{multiscale version} will be useful in order to derive the transversality condition on the final time (see Remark~\ref{remfindepreuvefreefinaltime}).
\end{remark}
\begin{remark}\label{rem8}
Another possible extension is to study dynamical systems on time scales with sampled-data control where the state $q$ and the constraint function $f$ are also sampled:
$$ q^\Delta(t) = f ( \Phi_1(t) ,q^{\Phi_2} (t), u^{\Phi_3}(t) ), \quad \Delta\text{-a.e. } t \in \mathbb{T}, $$
where $\mathbb{T}_1$, $\mathbb{T}_2$ and $\mathbb{T}_3$ are general time scales contained in $\mathbb{T}$, and $\Phi_1$, $\Phi_2$ and $\Phi_3$ are the corresponding mappings from $\mathbb{T}$ to ${\mathbb{T}_1}$, $\mathbb{T}_2$ and $\mathbb{T}_3$ respectively. In particular, the setting of \cite{bour11} corresponds to the above framework with $\mathbb{T} = \mathbb{R}^+$ and $\mathbb{T}_1 = \mathbb{T}_2 = \mathbb{T}_3$ a general time scale.
\end{remark}
Although this is not the main objective of our paper, we provide hereafter a result stating the existence of optimal solutions for ${\bf (OSDCP)}^{\T}_{\TU}$, under some appropriate compactness and convexity assumptions. Actually, if the existence of solutions is stated, the necessary conditions provided in Theorem~\ref{thmmain}, allowing to compute explicitly optimal sampled-data controls, may prove the uniqueness of the optimal solution. We refer to Section~\ref{sectionexemple} for examples.
Let $\mathcal{M}$ stand for the set of trajectories $q$, associated with $b \in \mathbb{T}$ and with a sampled-data control $u\in \mathrm{L}^\infty_{{\mathbb{T}_1}}({\mathbb{T}_1},\Omega)$, satisfying \eqref{DD-CS} $\Delta$-a.e. on $[a,b)_\mathbb{T}$ and $g(q(a),q(b)) \in \mathrm{S}$. We define the set of extended velocities $ \mathcal{W}(t,q) = \{ ( f(t,q,u) , f^0(t,q,u) )^\top \mid u \in \Omega \} $ for every $(t,q) \in \mathbb{T} \times \mathbb{R}^n$.
\begin{theorem}\label{propexistence}
If $\Omega$ is compact, $\mathcal{M}$ is non-empty, $\Vert q \Vert_{\infty}+b \leq M$ for every $q \in \mathcal{M}$ and for some $M \geq 0$, and if $ \mathcal{W}(t,q)$ is convex for every $(t,q) \in \mathbb{T} \times \mathbb{R}^n$, then ${\bf (OSDCP)}^{\T}_{\TU}$ has at least one optimal solution.
\end{theorem}
The proof of Theorem~\ref{propexistence} is done in Section~\ref{proofexistence}. Note that, in this theorem, it suffices to assume that $g$ is continuous. Besides, the assumption on the boundedness of trajectories can be weakened, by assuming, for instance, that the extended dynamics have a sublinear growth at infinity (see, \textit{e.g.}, \cite{Cesari}; many other easy and standard extensions are possible).
\subsection{Pontryagin maximum principle for ${\bf (OSDCP)}^{\T}_{\TU}$}\label{secPMP}
\subsubsection{Preliminaries on convexity and stable $\Omega$-dense directions}
The orthogonal of the closed convex set $\mathrm{S}$ at a point $x \in \mathrm{S}$ is defined by
$$ \mathcal{O}_\mathrm{S} [x] = \{ x' \in \mathbb{R}^j \mid \forall x'' \in \mathrm{S}, \ \langle x',x''-x \rangle_{\mathbb{R}^j} \leq 0 \}. $$
It is a closed convex cone containing $0$.
We denote by $d_\mathrm{S}$ the distance function to $\mathrm{S}$ defined by $d_\mathrm{S} (x) = \inf_{x' \in \mathrm{S}} \Vert x-x' \Vert_{\mathbb{R}^j}$, for every $x\in\mathbb{R}^j$.
Recall that, for every $x \in \mathbb{R}^j$, there exists a unique element $\mathrm{P}_\mathrm{S} (x)\in\mathrm{S}$ (projection of $x$ onto $\mathrm{S}$) such that $d_\mathrm{S} (x) = \Vert x - \mathrm{P}_\mathrm{S}(x) \Vert_{\mathbb{R}^j}$. It is characterized by the property
$\langle x-\mathrm{P}_\mathrm{S} (x) , x'-\mathrm{P}_\mathrm{S} (x) \rangle_{\mathbb{R}^j} \leq 0$ for every $x'\in \mathrm{S}$. In particular, $x-\mathrm{P}_\mathrm{S} (x) \in \mathcal{O}_\mathrm{S} [\mathrm{P}_\mathrm{S} (x)]$. The function $\mathrm{P}_\mathrm{S}$ is $1$-Lipschitz continuous. We recall the following obvious lemmas.
\begin{lemma}\label{lemconvex}
Let $(x_k)_{k \in \mathbb{N}}$ be a sequence of points of $\mathbb{R}^j$ and $(\zeta_k)_{k \in \mathbb{N}}$ be a sequence of nonnegative real numbers such that $x_k \to x \in \mathrm{S}$ and $\zeta_k (x_k - \mathrm{P}_\mathrm{S} (x_k) ) \to x' \in \mathbb{R}^j$ as $k \to +\infty$.
Then $x' \in \mathcal{O}_\mathrm{S} [x]$.
\end{lemma}
\begin{lemma}
The function $d_\mathrm{S}^2 : x \mapsto d_\mathrm{S} (x)^2$ is differentiable on $\mathbb{R}^j$, with $\mathrm{d} d_\mathrm{S}^2 (x) ( x')= 2\langle x-\mathrm{P}_\mathrm{S} (x), x' \rangle_{\mathbb{R}^j}$.
\end{lemma}
Hereafter we recall the notion of stable $\Omega$-dense directions and we state an obvious lemma. We refer to \cite[Section~2.2]{bour11} for more details.
\begin{definition}\label{defstable}
Let $v \in \Omega$. A direction $y \in \Omega$ is said to be a stable $\Omega$-dense direction from $v$ if there exists $\varepsilon>0$ such that $0$ is not isolated in $\{ \alpha \in [0,1], \; v' + \alpha (y-v') \in \Omega \}$ for every $v' \in \overline{\mathrm{B}}_{\mathbb{R}^m} (v,\varepsilon) \cap \Omega$. The set of all stable $\Omega$-dense directions from $v$ is denoted by $\mathrm{D}^{\Omega}_{\mathrm{stab}} (v)$.
\end{definition}
\begin{lemma}\label{lemstable}
If $\Omega$ is convex, then $\mathrm{D}^{\Omega}_{\mathrm{stab}} (v) = \Omega$ for every $v \in \Omega$.
\end{lemma}
\subsubsection{Main result}
Recall that $g$ is said to be submersive at a point $(q_1,q_2) \in \mathbb{R}^n \times \mathbb{R}^n$ if the differential of $g$ at this point is surjective. We define the Hamiltonian $H:\mathbb{T}\times \mathbb{R}^n \times \mathbb{R}^n \times \mathbb{R} \times \mathbb{R}^m \rightarrow \mathbb{R}$ of ${\bf (OSDCP)}^{\T}_{\TU}$ by $H(t,q,p,p^0,u) = \langle p, f(t,q,u) \rangle_{\mathbb{R}^n} + p^0 f^0(t,q,u)$.
\begin{theorem}[Pontryagin maximum principle for ${\bf (OSDCP)}^{\T}_{\TU}$]\label{thmmain}
If a trajectory $q^*$, defined on $[a,b^*]_\mathbb{T}$ and associated with a sampled-data control $u^* \in \mathrm{L}^\infty_{{\mathbb{T}_1}}({\mathbb{T}_1},\Omega)$, is an optimal solution of ${\bf (OSDCP)}^{\T}_{\TU}$,
then there exists a nontrivial couple $(p,p^0)$, where $p \in \mathrm{AC} ([a,b^*]_\mathbb{T},\mathbb{R}^n)$ (called adjoint vector) and $p^0 \leq 0$, such that the following conditions hold:
\begin{itemize}
\item Extremal equations:
\begin{equation}\label{extremal_equations}
q^{*\Delta}(t) = \frac{\partial H}{\partial p} (t,q^*(t),p^\sigma(t),p^0,u^{*\Phi}(t)), \qquad
p^\Delta(t) = - \frac{\partial H}{\partial q} (t,q^*(t),p^\sigma(t),p^0,u^{*\Phi}(t)) ,
\end{equation}
for $\Delta$-a.e. $t\in[a,b^*)_\mathbb{T}$.
\item Maximization condition:
\begin{itemize}
\item For $\Delta_1$-a.e. $s \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RD}_1$, we have
\begin{equation*}\label{maxcondition}
u^*(s) \in \argmax_{z \in \Omega} H(s,q^*(s),p(s),p^0,z) .
\end{equation*}
\item For every $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ such that $\sigma_1(r) \leq b^*$, we have
\begin{equation}\label{secondcondition}
\left\langle \int_{[r,\sigma_1(r))_\mathbb{T}} \frac{\partial H}{\partial u} (\tau,q^*(\tau),p^\sigma(\tau),p^0,u^*(r)) \; \Delta \tau \; , \; y-u^*(r) \right\rangle_{\mathbb{R}^m} \leq 0,
\end{equation}
for every $y \in \mathrm{D}^\Omega_{\mathrm{stab}} (u^*(r))$. In the case where $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ with $\sigma_1(r) > b^*$, the above maximization condition is still valid provided $\sigma_1(r)$ is replaced with $b^*$.
\end{itemize}
\item Transversality conditions on the adjoint vector: \\
If $g$ is submersive at $(q^*(a),q^*(b^*))$, then the nontrivial couple $(p,p^0)$ can be selected to satisfy
\begin{equation}\label{transv_cond}
p(a) = - \left( \frac{\partial g}{\partial q_1} (q^*(a),q^*(b^*)) \right)^\top \psi,\qquad
p(b^*) = \left( \frac{\partial g}{\partial q_2} (q^*(a),q^*(b^*)) \right)^\top \psi,
\end{equation}
where $-\psi \in \mathcal{O}_\mathrm{S} [g (q^*(a),q^*(b^*))]$.
\item Transversality condition on the final time: \\
If the final time is left free in the optimal control problem ${\bf (OSDCP)}^{\T}_{\TU}$, if $b^*$ belongs to the interior of $\mathbb{T}$ (for the topology of $\mathbb{R}$), and if $f$ and $f^0$ are of class $\mathrm{C}^1$ with respect to $t$ in a neighborhood of $b^*$, then the nontrivial couple $(p,p^0)$ can be moreover selected such that the Hamiltonian function $t \mapsto H(t,q^*(t),p(t),p^0,u^{*\Phi} (t)) $ coincides almost everywhere, in some neighborhood of $b^*$, with a continuous function vanishing at $t=b^*$.
In particular, if $u^{*\Phi} (t)$ has a left-limit at $t=b^*$ (denoted by $u^{*\Phi} (b^*_-)$), then the transversality condition can be written as
$$ H(b^*,q^*(b^*),p(b^*),p^0,u^{*\Phi} (b^*_-)) = 0 .$$
\end{itemize}
\end{theorem}
Theorem~\ref{thmmain} is proved in Section~\ref{annexe}. Several remarks are in order.
\begin{remark}\label{remarknormaliser}
As is well known, the nontrivial couple $(p,p^0)$ of Theorem~\ref{thmmain}, which is a Lagrange multiplier, is defined up to a multiplicative scalar. Defining as usual an \textit{extremal} as a quadruple $(q,p,p^0,u)$ solution of the extremal equations~\eqref{extremal_equations}, an extremal is said to be \textit{normal} whenever $p^0\neq 0$ and \textit{abnormal} whenever $p^0=0$. In the normal case $p^0\neq 0$, it is usual to normalize the Lagrange multiplier so that $p^0=-1$.
\end{remark}
\begin{remark}\label{remencompass}
Theorem~\ref{thmmain} encompasses the time scale version of the PMP derived in \cite{bour11} when the control is permanent, that is, when ${\mathbb{T}_1} = \mathbb{T}$. Indeed, in that case, for every $r \in [a,b^*)_\mathbb{T} \cap \mathrm{RS}$, $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ and $ \sigma_1(r) = \sigma (r) \leq b^*$. Then the condition~\eqref{secondcondition} can be written as the nonpositive gradient condition
\begin{equation*}
\left\langle \frac{\partial H}{\partial u} (r,q^*(r),p(\sigma(r)),p^0,u^*(r)) \; , \; y-u^*(r) \right\rangle_{\mathbb{R}^m} \leq 0,
\end{equation*}
for every $y \in \mathrm{D}^\Omega_{\mathrm{stab}} (u^*(r))$. Moreover, in the case of a free final time, under the assumptions made in the fourth item of Theorem~\ref{thmmain}, $b^*$ also belongs to the interior of ${\mathbb{T}_1} = \mathbb{T}$, and then in that case we recover the classical condition
$$ \max_{z \in \Omega} H(b^*,q^*(b^*),p(b^*),p^0,z) = 0. $$
\textit{A fortiori}, Theorem~\ref{thmmain} encompasses both the classical continuous-time and discrete-time versions of the PMP, that is respectively, when $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$ and $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$.
\end{remark}
\begin{remark}
If the final time is free, under the assumptions made in the fourth item of Theorem~\ref{thmmain}, and if moreover $b^* \notin {\mathbb{T}_1}$ (or if $b^*$ is a left-scattered point of ${\mathbb{T}_1}$), $u^{*\Phi}$ is constant on $[\rho_1(b^*),b^*)_\mathbb{T}$ where $\rho_1 (b^*) = \max \{ s \in {\mathbb{T}_1} \mid s < b^* \} < b^*$ (and thus, in particular, $u^{*\Phi}(t)$ has a left-limit at $t=b^*$), and therefore $ H(b^*,q^*(b^*),p(b^*),p^0,u^* ( \rho_1(b^*) )) = 0 $. This is similar to the situation of Theorem~\ref{thmmainintro}.
\end{remark}
\begin{remark}\label{remarkconditionsterminales}
Let us describe some typical situations of terminal conditions $g(q(a),q(b)) \in \mathrm{S}$ in ${\bf (OSDCP)}^{\T}_{\TU}$, and of the corresponding transversality conditions on the adjoint vector.
\begin{itemize}
\item If the initial and final points are fixed in ${\bf (OSDCP)}^{\T}_{\TU}$, that is, if we impose $q(a) = q_a$ and $q(b) = q_b$, then $j=2n$, $g(q_1,q_2) = (q_1,q_2)$ and $\mathrm{S} = \{ q_a \} \times \{ q_b \}$. In that case, the transversality conditions on the adjoint vector give no additional information.
\item If the initial point is fixed, that is, if we impose $q(a) = q_a$, and if the final point is left free in ${\bf (OSDCP)}^{\T}_{\TU}$, then $j=n$, $g(q_1,q_2) = q_1$ and $\mathrm{S} = \{ q_a \} $. In that case, the transversality conditions on the adjoint vector imply that $p(b^*) = 0$. Moreover, we have $p^0 \neq 0$\footnote{Indeed, if $p^0 =0$, then the adjoint vector $p$ is trivial from the extremal equation and from the final condition $p(b^*)=0$. This leads to a contradiction since the couple $(p,p^0)$ has to be nontrivial.} and we can normalize the Lagrange multiplier so that $p^0=-1$ (see Remark~\ref{remarknormaliser}).
\item If the initial point is fixed, that is, if we impose $q(a) = q_a$, and if the final point is subject to the constraint $G(q(b)) \in (\mathbb{R}^+)^k$ in ${\bf (OSDCP)}^{\T}_{\TU}$, where $G = (G^1,\ldots,G^k) : \mathbb{R}^n \rightarrow \mathbb{R}^k$ is of class $\mathrm{C}^1$ and is submersive at any point $q_2 \in G^{-1} ((\mathbb{R}^+)^k)$, then $j=n+k$, $g(q_1,q_2) = (q_1,G(q_2))$ and $\mathrm{S} = \{ q_a \} \times (\mathbb{R}^+)^k $. The transversality conditions on the adjoint vector imply that
$$ p(b^*) = \displaystyle \sum_{i=1}^k \lambda_i \partial_{q_2} G^i (q^*(b^*)), $$
with $\lambda_i \geq 0$, $i=1,\ldots,k$.
\item If the periodic condition $q(a)=q(b)$ is imposed in ${\bf (OSDCP)}^{\T}_{\TU}$, then $j=n$, $g(q_1,q_2) = q_1- q_2$ and $\mathrm{S} = \{ 0 \}$. In that case, the transversality conditions on the adjoint vector yield that $p(a) = p(b^*)$.
\end{itemize}
We stress that, in all examples above, the function $g$ is indeed a submersion.
\end{remark}
\begin{remark}\label{remarknonsubmersive}
In the case where $g$ is not submersive at $(q^*(a),q^*(b^*))$, to obtain transversality conditions on the adjoint vector, Theorem~\ref{thmmain} can be reformulated as follows:
\begin{quote}
\textit{If a trajectory $q^*$, defined on $[a,b^*]_\mathbb{T}$ and associated with a sampled-data control $u^* \in \mathrm{L}^\infty_{{\mathbb{T}_1}}({\mathbb{T}_1},\Omega)$, is an optimal solution of ${\bf (OSDCP)}^{\T}_{\TU}$,
then there exists a nontrivial couple $(\psi,p^0) \in \mathbb{R}^j \times \mathbb{R}$, with $-\psi \in \mathcal{O}_\mathrm{S} [g (q^*(a),q^*(b^*))]$ and $p^0 \leq 0$, and there exists $p \in \mathrm{AC} ([a,b^*]_\mathbb{T},\mathbb{R}^n)$ such that the extremal equations, the maximization conditions and the transversality conditions are satisfied.}
\end{quote}
However, with this formulation, the couple $(p,p^0)$ may be trivial\footnote{Indeed, if $p^0 =0$ and $\psi$ belongs to the kernel of $(\partial_{q_2} g (q^*(a),q^*(b^*)) )^\top$, then the couple $(p,p^0)$ is trivial. This situation leads to a contradiction if $g$ is submersive at $(q^*(a),q^*(b^*))$.} and, as a consequence, the result may not provide any information. We refer to Sections~\ref{section46} and \ref{section47} for more details.
\end{remark}
\begin{remark}
In this paper, the closedness of $\Omega$ is used in a crucial way in our proof of the PMP. Indeed, the closure of $\Omega$ allows us to define the Ekeland functional on a complete metric space (see Section~\ref{section41}). However, if the initial point is fixed, that is, if we impose $q(a) = q_a$, and if the final point is left free in ${\bf (OSDCP)}^{\T}_{\TU}$, then Theorem~\ref{thmmain} can be proved with a simple calculus of variations, without using the Ekeland variational principle. In this particular case, the closedness assumption can be removed.
\end{remark}
\begin{remark}\label{remarkbolza}
If the cost functional to be minimized in ${\bf (OSDCP)}^{\T}_{\TU}$ is $ \int_{[a,b)_\mathbb{T}} f^0 (\tau,q (\tau), u^\Phi(\tau) ) \, \Delta \tau + \ell (b,q(a),q(b))$, where $\ell : \mathbb{T} \times \mathbb{R}^n \times \mathbb{R}^n \rightarrow \mathbb{R}$ is continuous, of class $\mathrm{C}^1$ in $(q_1,q_2)$, then the transversality conditions on the adjoint vector become
\begin{multline*}
p(a) = - \left( \partial_{q_1} g (q^*(a),q^*(b^*)) \right)^\top \psi - p^0 \partial_{q_1} \ell (b^*,q^*(a),q^*(b^*)), \\ \text{and } p(b^*) = \left( \partial_{q_2} g (q^*(a),q^*(b^*)) \right)^\top \psi + p^0 \partial_{q_2} \ell (b^*,q^*(a),q^*(b^*)) . $$
\end{multline*}
Moreover, in the fourth item of Theorem~\ref{thmmain}, if $\ell$ is of class $\mathrm{C}^1$ in a neighborhood of $b^*$, the transversality condition on the final time must be replaced by:
\begin{quote}
\textit{The nontrivial couple $(p,p^0)$ can be selected such that the Hamiltonian function $ t \mapsto H(t,q^*(t),p(t),p^0,u^{*\Phi} (t)) $ coincides almost everywhere, in some neighborhood of $b^*$, with a continuous function that is equal to $-p^0 \partial_b \ell (b^*,q^*(a),q^*(b^*))$ at $t=b^*$.}
\end{quote}
To prove this claim, it suffices to modify accordingly the Ekeland functional in the proof of Theorem~\ref{thmmain} (see Section~\ref{section41}).
\end{remark}
\section{Applications and further comments}\label{sectionapplication}
In this section, we first give, in Section~\ref{sectionexemple}, a very simple example of an optimal control problem on time scales with sampled-data control, that we treat in details and on which all computations are explicit. The interest is that this example provides as well a simple situation where it is evident that some of the properties that are valid in the classical continous-time PMP do not hold anymore in the time-scale context. We gather these remarks in Section~\ref{nonextension}.
\subsection{A model for optimal consumption with sampled-data control}\label{sectionexemple}
Throughout this subsection, $\mathbb{T}$ and ${\mathbb{T}_1}$ are two time scales, unbounded above, satisfying ${\mathbb{T}_1} \subset \mathbb{T}$, $\min \mathbb{T} = \min {\mathbb{T}_1} = 0$ and $12 \in \mathbb{T}$. In the sequel, we study the following one-dimensional dynamical system with sampled-data control on time scales:
\begin{equation}\label{exequationq}
q^\Delta (t) = u^\Phi(t)q(t), \qquad \Delta\text{-a.e. } t \in [0,12)_\mathbb{T},
\end{equation}
with the initial condition $q(0)=1$, and subject to the constraint $u(t) \in [0,1]$ for $\Delta_1$-a.e. $t \in [0,12)_{\mathbb{T}_1}$. Since the final time $b=12$ is fixed, we can assume that $12 \in {\mathbb{T}_1}$ without loss of generality.
The above model is a classical model for the evolution of a controlled output of a factory during the time interval $[0,12]_\mathbb{T}$ (corresponding to the twelve months of a year). Precisely, $q(t) \in \mathbb{R}$ stands for the output at time $t \in \mathbb{T}$ and $u(t) \in [0,1]$ stands for the fraction of the output reinvested at each controlling time $t \in {\mathbb{T}_1}$. We assume that this fraction is sampled at each sampling time $t \in {\mathbb{T}_1}$ such that $t \in \mathrm{RS}_1$ and $\sigma (t) < \sigma_1(t)$, on the corresponding sampling interval $[t,\sigma_1(t))_\mathbb{T}$. In the sequel, our goal is to maximize the total consumption
\begin{equation}\label{exC}
C(u) = \int_{[0,12)_\mathbb{T}} (1-u^\Phi(\tau))q(\tau) \,\Delta \tau.
\end{equation}
In other words, our aim is to maximize the quantity of the output that we do not reinvest.
\begin{remark}
In the continuous case and with a permanent control (that is, with $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$), the above optimal control problem is a very well-known application of the classical Pontryagin maximum principle. We refer for example to~\cite[Exercice~2.3.3. p.82]{seie} or \cite[p.92]{JBHU}. In this section, our aim is to solve this optimal control problem in cases where the control is nonpermanent and sampled. We first treat some examples in the continuous-time setting $\mathbb{T} = \mathbb{R}^+$ and then in the discrete-time setting $\mathbb{T} = \mathbb{N}$.
\end{remark}
The above optimal sampled-data control problem corresponds to ${\bf (OSDCP)}^{\T}_{\TU}$ with $n=m=j=1$, $a=0$, $b=12$ (fixed final time), $\Omega = [0,1]$ (convex), $ g(q_1,q_2)=q_1$ and $\mathrm{S} = \{ 1 \}$ (fixed initial value $q(0)=1$ and free final value $q(12)$), $f(t,q,u) = uq$ and $f^0(t,q,u) = (u-1)q$ (our aim is to minimize $-C(u)$).
Since $f$ and $f^0$ are affine in $u$ and since $\Omega$ is compact, Theorem~\ref{propexistence} asserts that ${\bf (OSDCP)}^{\T}_{\TU}$ admits an optimal solution $q^*$, defined on $[0,12]_\mathbb{T}$ and associated with a sampled-data control $u^* \in \mathrm{L}^\infty_{\mathbb{T}_1}( {\mathbb{T}_1},[0,1])$. We now apply Theorem~\ref{thmmain} in order to compute explicitly the values of $u^*$ at each controlling time $t \in [0,12)_{\mathbb{T}_1}$. The nontrivial couple $(p,p^0)$ satisfies $p^0 = -1$ and $p(12)=0$ (see Remark~\ref{remarkconditionsterminales}). The adjoint vector $p \in \mathrm{AC} ([0,12]_\mathbb{T},\mathbb{R})$ is a solution of
\begin{equation}\label{exequationp}
p^\Delta(t) = - u^{*\Phi}(t) p^\sigma(t) + u^{*\Phi}(t) -1 , \qquad \Delta\text{-a.e. } t \in [0,12)_\mathbb{T}.
\end{equation}
Moreover, one has the following maximization conditions:
\begin{enumerate}
\item for $\Delta_1$-a.e. $s \in [0,12)_{\mathbb{T}_1} \cap \mathrm{RD}_1$,
\begin{equation}\label{exequationmax1}
u^*(s) \in \argmax_{z \in [0,1]} \,( z (p(s)-1)+1 ) q^*(s) .
\end{equation}
\item for every $r \in [0,12)_{\mathbb{T}_1} \cap \mathrm{RS}_1$,
\begin{equation}\label{exequationmax2}
\int_{[r,\sigma_1(r))_\mathbb{T}} q^*(\tau) (p^\sigma (\tau) - 1) (y-u^*(r)) \; \Delta \tau \leq 0,
\end{equation}
for every $y \in \mathrm{D}^\Omega_{\mathrm{stab}}(u^*(r)) = \Omega = [0,1]$ (since $\Omega$ is convex, see Lemma~\ref{lemstable}).
\end{enumerate}
Since $q^*$ is a solution of~\eqref{exequationq} and satisfies $q^*(0)=1$, one can easily see that $q^*$ is monotonically increasing on $[0,12]_\mathbb{T}$ and then $q^*$ has positive values. From~\eqref{exequationp} and since $p(12)=0$, one can easily obtain that $p$ is monotonically decreasing on $[0,12]_\mathbb{T}$ and then $p$ has nonnegative values.
\subsubsection{Continuous-time setting $\mathbb{T} = \mathbb{R}^+$}\label{sectionexcontinu}
\paragraph*{Case ${\mathbb{T}_1} = \mathbb{R}^+$ (permanent control).} Since $q^*$ has positive values, one can conclude from~\eqref{exequationp} and~\eqref{exequationmax1} that:
$$ u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ for a.e. } t \text{ such that } p(t)-1 > 0, \\
0 & \text{ for a.e. } t \text{ such that } p(t)-1 < 0,
\end{array} \right. \;
\dot{p}(t) = \left\lbrace \begin{array}{ll}
-p(t) & \text{ for a.e. } t \text{ such that } u^*(t)=1, \\
-1 & \text{ for a.e. } t \text{ such that } u^*(t)=0.
\end{array} \right. $$
Since $p(12)=0$, one can easily prove that the optimal (permanent) control $u$ is unique and given by $ u^*(t) = 1 $ for a.e. $t \in [0,11)$ and $ u^*(t) = 0$ for a.e. $t \in [11,12)$. The associated optimal consumption is $ C(u^*) = e^{11} \simeq 59874.142 $. We refer to~\cite[Exercice~2.3.3. p.82]{seie} or \cite[p.92]{JBHU} for more details.
\paragraph*{Case ${\mathbb{T}_1} $ discrete (sampled-data control).} Solving the differential equations~\eqref{exequationq} and \eqref{exequationp} leads to $ q^*(t) = q^*(\sigma_1(r)) e^{u^*(r)(t-\sigma_1(r))}$ and
$$ p^\sigma (t) = p(t) = \left\lbrace \begin{array}{ll}
p(\sigma_1(r)) e^{-u^*(r)(t-\sigma_1(r))} + \frac{u^*(r)-1}{u^*(r)} ( 1 - e^{-u^*(r)(t-\sigma_1(r))} ) & \text{if } u^*(r) \neq 0, \\
p(\sigma_1(r)) + \sigma_1 (r) - t & \text{if } u^*(r) = 0,
\end{array} \right. $$
for every $r \in [0,12)_{{\mathbb{T}_1}}$ and every $t \in [r,\sigma_1(r)]$. Then, \eqref{exequationmax2} can be written as $ q^*( \sigma_1(r) ) (y-u^*(r)) \Gamma_r (u^*(r)) \leq 0$, where $\Gamma_r : [0,1] \rightarrow \mathbb{R}$ is the continuous function given by:
$$ \Gamma_r (x) = \left\lbrace \begin{array}{ll}
\dfrac{1}{x^2} \Big[e^{-\mu_1(r)x} - \Big( 1+\mu_1(r)x(x(1-p(\sigma_1(r)))-1) \Big) \Big] & \text{if } x \neq 0, \\
\mu_1(r) \left( p(\sigma_1(r)) + \dfrac{\mu_1(r)}{2} -1 \right) & \text{if } x = 0.
\end{array} \right. $$
Since~\eqref{exequationmax2} holds true for every $y \in [0,1]$ and since $q^*$ has positive values, the following properties are satisfied for every $r \in [0,12)_{{\mathbb{T}_1}}$:
\begin{itemize}
\item if $\Gamma_r$ has negative values on $(0,1]$, then $u^*(r) =0$;
\item if $\Gamma_r$ has positive values on $[0,1)$, then $u^*(r) =1$;
\item if $\Gamma_r (0) > 0$ and $\Gamma_r (1) < 0 $, then $u^*(r) \in (0,1)$ is a solution of the nonlinear equation $\Gamma_r (x) = 0$.
\end{itemize}
Note that $\Gamma_r$ depends only on $\mu_1(r)$ and $p(\sigma_1(r))$. As a consequence, the knowledge of the value $p(12)=0$ and the above properties allow to compute $u^*(r_0)$ where $r_0$ is the element of $[0,12)_{{\mathbb{T}_1}}$ such that $\sigma_1 (r_0) =12$ (and $\mu_1 (r_0) = 12-r_0$). Then, the knowledge of $u^*(r_0)$ allows to compute $p(r_0)$ from~\eqref{exequationp}. Then, the knowledge of $p(r_0)$ and the above properties allow to compute $u^*(r_1)$ where $r_1$ is the element of $[0,12)_{{\mathbb{T}_1}}$ such that $\sigma_1 (r_1) =r_0$ (and $\mu_1 (r_1) = r_0-r_1$), etc. Actually, this recursive procedure allows to compute $u^*(r)$ for every $r \in [0,12)_{{\mathbb{T}_1}}$. Numerically, we obtain the following results:
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
${\mathbb{T}_1} = \mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,1,\ldots, 10 \} \\
0 & \text{ if } t =11
\end{array} \right.$ & $C(u^*)=e^{11} \simeq 59874.142$ \\ \hline
${\mathbb{T}_1} = 2\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,2,4,6,8 \} \\
0 & \text{ if } t =10
\end{array} \right.$ & $C(u^*)= 2 e^{10} \simeq 44052.932 $ \\ \hline
${\mathbb{T}_1} = 3\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,3,6 \} \\
0.4536 & \text{ if } t =9
\end{array} \right.$ & $C(u^*) \simeq 28299.767 $ \\ \hline
${\mathbb{T}_1} = 4\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,4 \} \\
0.6392 & \text{ if } t =8
\end{array} \right.$ & $C(u^*) \simeq 20013.885$ \\ \hline
${\mathbb{T}_1} = 9\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t = 0 \\
0.4536 & \text{ if } t =9
\end{array} \right.$ & $C(u^*) \simeq 28299.767$ \\ \hline
${\mathbb{T}_1} = 12\mathbb{N}$ & $u^*(0) \simeq 0.9083$ & $C(u^*) \simeq 5467.24$ \\ \hline
${\mathbb{T}_1} = 12\mathbb{N} \cup \{ 10 , 11.5 \} $ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t = 0 \\
0.9072 & \text{ if } t = 10 \\
0 & \text{ if } t =11.5
\end{array} \right.$ & $C(u^*) \simeq 49476.604 $ \\ \hline
\end{tabular}
\end{center}
\begin{remark}
In this example, Theorem~\ref{propexistence} states the existence of an optimal solution. In all cases above studied, the Pontryagin maximum principle proves that the optimal solution is moreover unique.
\end{remark}
\begin{remark}
The case ${\mathbb{T}_1} = \mathbb{N}$ can be easily deduced from the permanent case ${\mathbb{T}_1} = \mathbb{R}^+$ (see Remark~\ref{remarkinclusiondeuxT1}). Similarly, the case ${\mathbb{T}_1} = 9\mathbb{N}$ can be deduced from the case ${\mathbb{T}_1} = 3\mathbb{N}$.
\end{remark}
\begin{remark}
The case ${\mathbb{T}_1} = 12\mathbb{N}$ corresponds to an optimal parameter problem (see Remark~\ref{remarkparam}).
\end{remark}
\begin{remark}\label{remarkcontinu2N}
For the needs of Section~\ref{nonextension}, let us give some details on the case ${\mathbb{T}_1} = 2\mathbb{N}$. In that case, $q^*(t)=e^t$ on $[0,10]$ and $q^*(t)=e^{10}$ on $[10,12]$. Moreover, $p(t) = 2e^{10-t}$ on $[0,10]$ and $p(t)=12-t$ on $[10,12]$. Hence, for every $t \in [10,11)$, $\argmax_{z \in \Omega} H(t,q^*(t),p(t),p^0,z) = \{ 1 \}$. The Hamiltonian $ t \mapsto H(t,q^*(t),p(t),p^0,u^{*\Phi}(t)) $ is equal to $t \mapsto 2e^{10}$ a.e. on $[0,10)$ and is equal to $t \mapsto e^{10}$ a.e. on $[10,12)$. Finally, the maximized Hamiltonian $ t \mapsto \max_{z \in \Omega} H(t,q^*(t),p^\sigma(t),p^0,z)$ is equal to $t \mapsto 2e^{10}$ on $[0,10)$, to $t \mapsto (12-t) e^{10}$ on $[10,11)$ and to $t \mapsto e^{10}$ on $[11,12)$.
\end{remark}
\paragraph*{Case ${\mathbb{T}_1}$ hybrid (sampled-data control).} In this paragraph, we study the hybrid case ${\mathbb{T}_1} = [0,6] \cup \{ 10 \} \cup [11.5,+\infty)$. Similarly to the permanent case ${\mathbb{T}_1} = \mathbb{R}^+$, one can easily conclude from~\eqref{exequationmax1} that:
$$ u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ for a.e. } t \in [0,6) \cup [11.5,12) \text{ such that } p(t)-1 > 0, \\
0 & \text{ for a.e. } t \in [0,6) \cup [11.5,12) \text{ such that } p(t)-1 < 0.
\end{array} \right. $$
Since $p(12)=0$, one can easily prove that $u^*(t) = 0$ for a.e. $t \in [11.5,12)$, and then $p(11.5)=0.5$. From the knowledge of $p(11.5)=0.5$ and using similar arguments than in the previous paragraph, one can compute $u^*(10) \simeq 0.9072$. From~\eqref{exequationp}, it gives $p(10) \simeq 2.2462$. From the knowledge of $p(10) \simeq 2.2462 $, one can compute $u^*(6)=1$. From~\eqref{exequationp}, it gives $p(6) \simeq 122.6402$. Since $p$ is monotonically decreasing, we conclude that $p(t)-1 > 0$ for every $t \in [0,6)$. We finally conclude that the optimal sampled-data control $u^*$ is unique and is given by $u^*(t) = 1$ for a.e. $t \in [0,6)$, $u^*(6)=1$, $u^*(10) \simeq 0.9072$ and $u^*(t) = 0$ for a.e. $t \in [11.5,12)$. The associated optimal consumption is $ C(u^*) \simeq 49476.604 $.
\begin{remark}
The discrete case ${\mathbb{T}_1} = 12\mathbb{N} \cup \{ 10 , 11.5 \}$ can now be seen as a consequence of the hybrid case ${\mathbb{T}_1} = [0,6] \cup \{ 10 \} \cup [11.5,+\infty)$ (see Remark~\ref{remarkinclusiondeuxT1}).
\end{remark}
\subsubsection{Discrete-time setting $\mathbb{T} = \mathbb{N}$}\label{sectionexdiscret}
In the discrete-time setting $\mathbb{T} = \mathbb{N}$, similarly to the continuous-time one, we can prove that \eqref{exequationmax2} can be written as $ q^*(r)(y-u^*(r)) \Lambda_r (u^*(r)) \leq 0$, where $\Lambda_r : [0,1] \rightarrow \mathbb{R}$ is the continuous function given by:
$$ \Lambda_r (x) = \left\lbrace \begin{array}{ll}
\dfrac{1}{x^2} \Big[ 1- (1+x)^{\mu_1(r)-1} \Big( 1+\mu_1(r)x(x(1-p(\sigma_1(r)))-1)+x \Big) \Big] & \text{if } x \neq 0, \\
\mu_1(r) \left( p(\sigma_1(r)) + \dfrac{\mu_1(r)-3}{2} \right) & \text{if } x = 0.
\end{array} \right. $$
The same recursive procedure (than in the previous section) allows to compute $u^*(r)$ for every $r \in [0,12)_{{\mathbb{T}_1}}$. Numerically, we obtain the following results:
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
${\mathbb{T}_1} = 2\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,2,4,6,8 \} \\
0 & \text{ if } t =10
\end{array} \right.$ & $C(u^*)= 2^{11} = 2048 $ \\ \hline
${\mathbb{T}_1} = 3\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,3,6 \} \\
0 & \text{ if } t =9
\end{array} \right.$ & $C(u^*) = 3 \cdot 2^9 = 1536 $ \\ \hline
${\mathbb{T}_1} = 4\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t \in \{ 0,4 \} \\
0.2886 & \text{ if } t =8
\end{array} \right.$ & $C(u^*) \simeq 1108.882$ \\ \hline
${\mathbb{T}_1} = 6\mathbb{N}$ & $u^*(t) = \left\lbrace \begin{array}{ll}
1 & \text{ if } t =0 \\
0.5725 & \text{ if } t =6
\end{array} \right.$ & $C(u^*) \simeq 674.787 $ \\ \hline
${\mathbb{T}_1} = 12\mathbb{N}$ & $u^*(0) \simeq 0.8145$ & $C(u^*) \simeq 159.647$ \\ \hline
\end{tabular}
\end{center}
\begin{remark}\label{remarknonunique}
In the permanent discrete case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$, the optimal control is not unique. Indeed, every control $u^*$ satisfying $u^*(t)=1$ for every $t \in \{ 0,1,\ldots,9 \}$, $u^*(10) \in [0,1]$ and $u^*(11)=0$ is optimal. Let us give some details on the proof and let us note that the PMP does not provide any constraint on the value $u^*(10)$ in that case.
With $r=11$ and $p(12)=0$, $\Lambda_r$ has negative values on $(0,1]$ (constantly equal to $-1$), then $u^*(11) = 0$ from \eqref{exequationmax2}. From~\eqref{exequationp}, we compute $p(11)=1$. With $r=10$ and $p(11)=1$, $\Lambda_r$ is constantly equal to $0$ and then \eqref{exequationmax2} does not provide any constraint on the value $u^*(10)$. Actually, it does not matter since $p(10)$ can still be computed from~\eqref{exequationp} ($p(10)=2$) and the recursive procedure can be pursued. We obtain $u^*(t)=1$ for every $t \in \{ 0,1,\ldots,9 \}$. From~\eqref{exC}, we have
$$ C(u^*) = (1-u^*(10)) q^*(10) + q^*(11) = (1-u^*(10)) \cdot 2^{10} + (1+u^*(10)) q^*(10) = 2 \cdot 2^{10} = 2^{11}. $$
Finally, if $u^*(11)=0$ and $u^*(t)=1$ for every $t \in \{ 0,1,\ldots,9 \}$, the value of $u^*(10) \in [0,1]$ does not influence $C(u^*)$. This concludes the proof and the remark.
\end{remark}
\begin{remark}
The case ${\mathbb{T}_1} = 2\mathbb{N}$ can be seen as a consequence of the permanent case ${\mathbb{T}_1} = \mathbb{N}$ (see Remarks~\ref{remarkinclusiondeuxT1} and \ref{remarknonunique}).
\end{remark}
\begin{remark}
The case ${\mathbb{T}_1} = 12\mathbb{N}$ corresponds to an optimal parameter problem (see Remark~\ref{remarkparam}).
\end{remark}
\subsection{Non-extension of several classical properties}\label{nonextension}
In this section, we recall some basic properties that occur in classical optimal control theory in the continuous-time setting and with a permament control, that is, with $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$. Our aim is to discuss their extension (or their failure) to the general time scale setting and to the nonpermament control case. We will provide several counterexamples in the discrete-time setting with a permanent control ($\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$) and in the continuous-time setting with a nonpermanent control (${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{R}^+$). In the following paragraphs, except the last one, the final time $b \in \mathbb{T}$ can be fixed or not.
\paragraph*{Pointwise maximization condition of the Hamiltonian.}
In the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$, it is well known that an optimal (permanent) control $u^*$ satisfies the maximization condition $ u^*(t) \in \argmax_{z \in \Omega} H(t,q^*(t),p(t),p^0,z)$ for a.e. $t \in [0,b^*)$. We refer to \cite[Example~7]{bour11} for a counterexample showing the failure of this maximization condition in the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$. We refer to Remark~\ref{remarkcontinu2N} for a counterexample in the case ${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{R}^+$.
\paragraph*{Continuity of the Hamiltonian.}
In the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$, it is well known that the Hamiltonian function $ t \mapsto H(t,q^*(t),p(t),p^0,u^*(t)) $ coincides almost everywhere on $[0,b^*]$ with the continuous function $ t \mapsto \max_{z \in \Omega} H(t,q^*(t),p(t),p^0,z)$. Remark~\ref{remarkcontinu2N} provides a counterexample showing the failure of this regularity property in the case ${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{R}^+$.
\begin{remark}
Nevertheless, in the case of a free final time, under the assumptions of the fourth item of Theorem~\ref{thmmain}, the Hamiltonian function $ t \mapsto H(t,q^*(t),p(t),p^0,u^{*\Phi}(t)) $ coincides almost everywhere, in some neighborhood of $b^*$, with a continuous function.
\end{remark}
\paragraph*{The autonomous case.}
In the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$, if the Hamiltonian $H$ is autonomous (that is, does not depend on $t$), it is well known that the function $t \mapsto H(q^*(t),p(t),p^0,u^*(t))$ is almost everywhere constant on $[0,b^*]$, this constant being equal to the maximized Hamiltonian. We refer to \cite[Example~8]{bour11} for a counterexample showing the failure of this constantness property in the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$, and we refer to Remark~\ref{remarkcontinu2N} for a counterexample in the case ${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{R}^+$ (and there, even the maximized autonomous Hamiltonian is not constant).
\paragraph*{Saturated constraint set $\Omega$ for Hamiltonian affine in $u$.}
In this paragraph, we assume that $\Omega$ is convex. In the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$, if the Hamiltonian is \textit{affine in $u$}, that is, if it can be written as
$$ H(t,q,p,p^0,u) = \langle H_1 (t,q,p,p^0) , u \rangle_{\mathbb{R}^m} + H_2 (t,q,p,p^0), $$
one can easily prove that $H_1 (t,q^*(t),p(t),p^0) \in \mathcal{O}_\Omega [u^*(t)]$ for almost every $t \in [0,b^*)$. It follows that an optimal (permament) control $u^*$ must take its values at the boundary of $\Omega$ (saturation of the constraints) for almost every $t \in [0,b^*)$ such that $H_1 (t,q^*(t),p(t),p^0) \neq 0_{\mathbb{R}^m}$.
\begin{remark}
This classical property can be extended to the case $\mathbb{T} = {\mathbb{T}_1} =\mathbb{N}$. Indeed, in that case, the nonpositive gradient condition is given by
$$ \left\langle \frac{\partial H}{\partial u} (t,q^*(t),p(t+1),p^0,u^*(t)) , y -u^*(t) \right\rangle_{\mathbb{R}^m} = \langle H_1 (t,q^*(t),p(t+1),p^0) , y - u^*(t) \rangle_{\mathbb{R}^m} \leq 0, $$
for every $y \in \Omega$, that is, $H_1 (t,q^*(t),p(t+1),p^0) \in \mathcal{O}_\Omega [u^*(t)]$.
\end{remark}
\begin{remark}
Remark~\ref{remarknonunique} provides an interesting example in the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$. Indeed, in that case, the control defined by $u^*(t) =0$ for every $t \in \{ 0, \ldots ,9 \}$, $u^*(10) = 1/2$ and $u^*(11)=1$ is an optimal (permanent) control. However, it does not saturate the constraint set $\Omega$ at $t=10$. It is not a surprise since, in that case, $H_1(t,q^*(t),p(t+1),p^0) = 0$ at $t=10$.
Note that, in the case ${\mathbb{T}_1} = 4\mathbb{N}$, Section~\ref{sectionexdiscret} provides a counterexample showing the failure of this classical property in the case ${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{N}$. Similarly, Section~\ref{sectionexcontinu} in the case ${\mathbb{T}_1} = 3\mathbb{N}$ provides a counterexample in the case ${\mathbb{T}_1} \varsubsetneq \mathbb{T} = \mathbb{R}^+$.
\end{remark}
\begin{remark}
Figure \ref{reffig} represents the values of the optimal sampled-data control $u^*$ of Section~\ref{sectionexcontinu}, in the case ${\mathbb{T}_1} = 12\mathbb{N} \cup \{ \lambda \}$, where $\lambda$ is a parameter evolving in $(0,12)$. In that case, $u^*(0)$ (resp., $u^*(\lambda)$) saturates the constraint set $\Omega = [0,1]$ approximately for $\lambda \in (0,11.9245)$ (resp., for $\lambda \in (9.9866,12)$).
\end{remark}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=0.25]{figure2.png}
\end{center}
\caption{Optimal sampled-data control $u^*$.}\label{reffig}
\end{figure}
\paragraph*{Vanishing of the maximized Hamiltonian at the final time.}
In this paragraph, we assume that the final time is left free.
In the case $\mathbb{T} = {\mathbb{T}_1} = \mathbb{R}^+$, under the assumptions of the fourth item of Theorem~\ref{thmmain}, it is well known that the maximized Hamiltonian vanishes at $t=b^*$ (see Remark~\ref{remencompass}). In Theorem~\ref{thmmain}, we have established that this property is still valid in the time scale setting under some appropriate conditions, the main one being that $b^*$ must belong to the interior of $\mathbb{T}$. In the discrete case $\mathbb{T} = \mathbb{N}$, the interior of $\mathbb{T}$ is empty and then the latter assumption is never satisfied. The Hamiltonian at the final time may then not vanish, and we refer to \cite[Example~8]{bour11} for a counterexample with $\mathbb{T} = {\mathbb{T}_1} = \mathbb{N}$.
\section{Proofs}\label{annexe}
The section is structured as follows. Subsections~\ref{annexe1}, \ref{sectionvariations} and \ref{annexe3} are devoted to the proof of Theorem~\ref{thmmain}. In Subsection~\ref{annexe1}, we recall some known Cauchy-Lipschitz results on time scales and we establish some preliminary results on the relations between $u$ and $u^\Phi$. In Subsection~\ref{sectionvariations}, we introduce appropriate needle-like variations of the control. Finally, in Subsection~\ref{annexe3}, we apply the Ekeland variational principle to an adequate functional in an appropriate complete metric space, and then we prove the PMP. In Subsection~\ref{proofexistence} (that the reader can read independently of the rest of Section~\ref{annexe}), we detail the proof of Theorem~\ref{propexistence}.
\subsection{Preliminaries}\label{annexe1}
\subsubsection{Relations between $u$ and $u^\Phi$}\label{annexe1uetuphi}
We start with a lemma whose arguments of proof will be used several times.
\begin{lemma}\label{lemnumber1}
Let $m \in \mathbb{N}^*$ and let $c < d$ be two elements of $\mathbb{T}$ with $c \in {\mathbb{T}_1}$. Let $u$, $v : [c,d)_{{\mathbb{T}_1}} \rightarrow \mathbb{R}^m$ be two functions. Then, $u = v$ $\Delta_1$-a.e. on $[c,d)_{{\mathbb{T}_1}}$ if and only if $u^\Phi = v^\Phi$ $\Delta$-a.e. on $[c,d)_{\mathbb{T}}$.
\end{lemma}
\begin{proof}
Without loss of generality, we can assume that $v$ is constant equal to $0_{\mathbb{R}^m}$. Let us define $ A = \{ t \in [c,d)_{{\mathbb{T}_1}} \,| \,u(t) \neq 0_{\mathbb{R}^m} \} $ and $ B = \{ t \in [c,d)_{\mathbb{T}} \,| \,u^\Phi(t) \neq 0_{\mathbb{R}^m} \} $. Since $\Phi (t) = t$ for every $t \in {\mathbb{T}_1}$, the inclusion $A \subset B$ holds. Firstly, let us assume that $\mu_{\Delta_1}(A) = 0$. Hence, $A \subset \mathrm{RD}_1$ and $\mu_{\Delta_1}(A) = \mu_L (A) = 0$. Since $A \subset \mathrm{RD}_1 \subset \mathrm{RD}$, we deduce that $\mu_\Delta (A) = \mu_L (A) = 0$. On the other hand, for every $t \in B$, $\Phi(t) \in A \subset \mathrm{RD}_1$, then $t \in {\mathbb{T}_1}$ and $t = \Phi(t) \in A$. We conclude that $A = B$ and $\mu_\Delta (B) = 0$. Secondly, let us assume that $\mu_\Delta (B) = 0$ and $\mu_{\Delta_1}(A) > 0$. Since $A \subset B$, we deduce that $\mu_\Delta (A) = 0$, $A \subset \mathrm{RD}$ and $\mu_\Delta (A) = \mu_L (A) = 0$. Since $\mu_L (A) = 0$ and $\mu_{\Delta_1}(A) > 0$, we conclude that there exists $t_0 \in \mathrm{RS}_1 \cap A \subset \mathrm{RS}_1 \cap \mathrm{RD}$. Consequently, $u^\Phi$ is constant (different of $0_{\mathbb{R}^m}$) on $[t_0, \min ( \sigma_1(t_0),d))_\mathbb{T}$. As a consequence, $\mu_\Delta (B) \geq \min ( \sigma_1(t_0),d) - t_0 > 0$. This raises a contradiction.
\end{proof}
\begin{proposition}\label{propnumber1}
Let $m \in \mathbb{N}^*$ and let $c < d$ be two elements of $\mathbb{T}$ with $c \in {\mathbb{T}_1}$.
\begin{enumerate}
\item For every $u \in \mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$, we have $u^\Phi \in \mathrm{L}^1_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)$ and
\begin{equation}\label{inequalitynotequality}
\Vert u^\Phi \Vert_{\mathrm{L}^1_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)} \leq \Vert u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)}.
\end{equation}
\item For every $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$, we have $u^\Phi \in \mathrm{L}^\infty_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)$ and
$$ \Vert u^\Phi \Vert_{\mathrm{L}^\infty_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)} = \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)}. $$
\end{enumerate}
\end{proposition}
\begin{proof}
Let $u \in \mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$. We first treat the $\mu_{\Delta}$-measurability of $u^\Phi$. From Lemma~\ref{lemnumber1}, we can consider that $u$ is defined everywhere on $[c,d)_{{\mathbb{T}_1}}$ and is $\mu_{\Delta_1}$-measurable on $[c,d)_{{\mathbb{T}_1}}$. Let us prove that $u^\Phi$ is $\mu_{\Delta}$-measurable on $[c,d)_{\mathbb{T}}$. We introduce $d' = \inf \{ s \in {\mathbb{T}_1} \,| \,s \geq d \} \in {\mathbb{T}_1}$. Note that $d' \geq d$, with equality if and only if $d \in {\mathbb{T}_1}$. From~\cite{caba}, $u$ is $\mu_{\Delta_1}$-measurable on $[c,d)_{{\mathbb{T}_1}}$ if and only if the extension $\tilde{u}$ defined on $[c,d')$ by
$$ \tilde{u}(t) = \left\lbrace \begin{array}{lcl}
u(t) & \text{if} & t \in {\mathbb{T}_1}, \\
u(r) & \text{if} & t \in (r,\sigma_1(r)), \,\text{where} \,r \in \mathrm{RS}_1,
\end{array} \right. $$
is $\mu_L$-measurable on $[c,d')$. By hypothesis, $\tilde{u}$ is $\mu_L$-measurable on $[c,d')$ and consequently, the restriction $\tilde{u}_{|_{[c,d)}}$ is $\mu_L$-measurable on $[c,d)$. Still from~\cite{caba}, $u^\Phi$ is $\mu_{\Delta}$-measurable on $[c,d)_{\mathbb{T}}$ if and only if the extension $\widetilde{u^\Phi}$ defined on $[c,d)$ by
$$ \widetilde{u^\Phi}(t) = \left\lbrace \begin{array}{lcl}
u^\Phi(t) & \text{if} & t \in \mathbb{T}, \\
u^\Phi(r) & \text{if} & t \in (r,\sigma(r)), \,\text{where} \,r \in \mathrm{RS},
\end{array} \right. $$
is $\mu_L$-measurable on $[c,d)$. Hence, it is sufficient to see that $\widetilde{u^\Phi} = \tilde{u}_{|_{[c,d)}}$. This is true since, for any $t \in [c,d)$, we have:
\begin{itemize}
\item either $t \in {\mathbb{T}_1}$, and then $t \in \mathbb{T}$ and thus $\widetilde{u^\Phi}(t) = u^\Phi(t) = u(\Phi(t)) = u(t) = \tilde{u}(t)$.
\item either $t \in \mathbb{T} \backslash {\mathbb{T}_1}$, and then $t \in (\Phi(t),\sigma_1(\Phi(t)))$ where $\Phi(t) \in \mathrm{RS}_1$ and thus $\widetilde{u^\Phi}(t) = u^\Phi(t) = u(\Phi(t)) = \tilde{u}(t)$.
\item either $t \notin \mathbb{T}$, and then $t \in (r,\sigma(r))$ where $r \in \mathrm{RS}$ and then $t \in (\Phi(r),\sigma_1(\Phi(r)))$ where $\Phi(r) \in \mathrm{RS}_1$ and thus $\widetilde{u^\Phi}(t) = u^\Phi(r) = u(\Phi(r)) = \tilde{u}(t)$.
\end{itemize}
This establishes the $\mu_{\Delta}$-measurability of $u^\Phi$ on $[c,d)_{\mathbb{T}}$.
Considering $\Vert u \Vert_{\mathbb{R}^m}$ instead of $u$, we can consider that $m = 1$ and $u \in \mathrm{L}^1_{{\mathbb{T}_1}}([c,d)_{{\mathbb{T}_1}},\mathbb{R}^+)$. From~\cite{caba}, we have
\begin{multline*}
\displaystyle \int_{[c,d)_\mathbb{T}} u^\Phi (\tau) \,\Delta \tau = \int_{[c,d)} \widetilde{u^\Phi}(\tau) \,d\tau
= \int_{[c,d)} \tilde{u}_{|_{[c,d)}}(\tau) \,d\tau \\
= \int_{[c,d')} \tilde{u}(\tau) \,d\tau - \int_{[d,d')} \tilde{u}(\tau) \,d\tau
= \int_{[c,d)_{{\mathbb{T}_1}}} u(\tau) \,\Delta_1 \tau - (d'-d)u(\Phi(d)).
\end{multline*}
Noting that $(d'-d)u(\Phi(d)) \geq 0$ concludes the proof of the first point.
Let us prove the second point. Let $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$. The $\mu_{\Delta}$-measurability of $u^\Phi$ is already proved since $\mathrm{L}^\infty_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m) \subset \mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$. Let $M \geq 0$ be a constant. With the same arguments as in the proof of Lemma~\ref{lemnumber1}, we can prove that $ \Vert u(t) \Vert_{\mathbb{R}^m} \leq M$ for $\Delta_1$-a.e. $t \in [c,d)_{{\mathbb{T}_1}}$ if and only if $ \Vert u^\Phi(t) \Vert_{\mathbb{R}^m} \leq M$ for $\Delta$-a.e. $t \in [c,d)_{\mathbb{T}}$. As a consequence, we get $u^\Phi \in \mathrm{L}^\infty_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)$ and $ \Vert u^\Phi \Vert_{\mathrm{L}^\infty_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)} = \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)}$.
\end{proof}
\begin{remark}
From the proof above, we see that the inequality~\eqref{inequalitynotequality} is an equality if and only if $d \in {\mathbb{T}_1}$ or $u(\Phi(d)) = 0$. Then, considering $\mathbb{T} = \mathbb{N}$, ${\mathbb{T}_1} = 2\mathbb{N}$, $m=1$, $c=0$, $d=1$ and $u$ the constant function equal to $1$ provides a counterexample. Indeed, in that case, we have $u \in \mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)$ and $u^\Phi \in \mathrm{L}^1_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)$ with
$ \Vert u^\Phi \Vert_{\mathrm{L}^1_{\mathbb{T}}( [c,d)_{\mathbb{T}},\mathbb{R}^m)} = 1 < 2 = \Vert u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([c,d)_{{\mathbb{T}_1}},\mathbb{R}^m)}. $
\end{remark}
\subsubsection{Recalls on $\Delta$-Cauchy-Lipschitz results}\label{recallCL}
According to \cite[Theorem 1]{bour10}, for every control $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\mathbb{R}^m)$ and every initial condition $q_a \in \mathbb{R}^n$, there exists a unique maximal solution of~\eqref{DD-CS} such that $q(a)=q_a$, denoted by $q(\cdot,u,q_a)$, and defined on a maximal interval, denoted by $I_\mathbb{T}(u,q_a)$. The word \textit{maximal} means that $q(\cdot,u,q_a)$ is an extension of any other solution.
Moreover, we recall that (see \cite[Lemma~1]{bour10})
$$ \forall t \in I_\mathbb{T} (u,q_a), \quad q(t,u,q_a) = q_a + \displaystyle \int_{[a,t)_\mathbb{T}} f(\tau,q(\tau,u,q_a),u^\Phi(\tau)) \,\Delta \tau .$$
Finally, either $I_\mathbb{T} (u,q_a) = \mathbb{T}$, that is, $q(\cdot,u,q_a)$ is a \textit{global} solution of~\eqref{DD-CS}, or $I_\mathbb{T} (u,q_a) = [a,c)_\mathbb{T}$ where $c$ is a left-dense point of $\mathbb{T}$, and in that case, $q(\cdot,u,q_a)$ is unbounded on $I_\mathbb{T} (u,q_a)$ (see \cite[Theorem~2]{bour10}).
\begin{definition}\label{defadm}
For a given $b \in \mathbb{T}$, a couple $(u,q_a) \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\mathbb{R}^m) \times \mathbb{R}^n$ is said to be \textit{admissible} on $[a,b]_\mathbb{T}$ whenever $b \in I_\mathbb{T} (u,q_a)$.
\end{definition}
For a given $b \in \mathbb{T}$, we denote by $\mathcal{UQ}^{b}$ the set of all admissible couples $(u,q_a)$ on $[a,b]_\mathbb{T}$. It is endowed with the norm
$ \Vert (u,q_a) \Vert_{\mathcal{UQ}^{b}} = \Vert u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^{m})} + \Vert q_a \Vert_{\mathbb{R}^n} . $
\subsection{Needle-like variations of the control, and variation of the initial condition}\label{sectionvariations}
Throughout this section, we consider $b \in \mathbb{T}$ and $(u,q_a) \in \mathcal{UQ}^{b}$.
We are going, in particular, to define appropriate needle-like variations. As in \cite{bour11}, we have to distinguish between right-dense and right-scattered times, along the time scale ${\mathbb{T}_1}$. We will also define appropriate variation of the initial condition.
In the sequel, the notation $\Vert \cdot \Vert$ stands for the usual induced norm of matrices in $\mathbb{R}^{n,n}$, $\mathbb{R}^{n,m}$ and $\mathbb{R}^{m,m}$.
\subsubsection{General variation of $(u,q_a)$}\label{section31prel}
In the first lemma below, we prove that $\mathcal{UQ}^{b}$ is open. Actually we prove a stronger result, by showing that $\mathcal{UQ}^{b}$ contains a neighborhood of any of its point in $\mathrm{L}^1_{{\mathbb{T}_1}} \times \mathbb{R}^n$ topology, which will be useful in order to define needle-like variations.
\begin{lemma}\label{prop30-1}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$.
There exist $\nu_R > 0$ and $\eta_R > 0$ such that the set
\begin{equation*}
\mathrm{E}_R(u,q_a) = \Big( \overline{\mathrm{B}}_{\mathrm{L}^\infty_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} (0,R) \cap \overline{\mathrm{B}}_{\mathrm{L}^1_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} (u,\nu_R) \Big) \times \overline{\mathrm{B}}_{\mathbb{R}^n} (q_a,\eta_R)
\end{equation*}
is contained in $\mathcal{UQ}^{b}$.
\end{lemma}
Before proving this lemma, let us recall a time scale version of Gronwall's Lemma (see \cite[Chapter 6.1]{bohn}).
The generalized exponential function is defined by
$e_L (t,c) = \exp ( \int_{[c,t)_\mathbb{T}} \xi_{\mu (\tau)} (L) \ \Delta \tau )$,
for every $L \geq 0$, every $c \in \mathbb{T}$ and every $t \in [c,+\infty)_\mathbb{T}$,
where
$\xi_{\mu (\tau)} (L) = \log ( 1+ L \mu (\tau)) / \mu (\tau)$ whenever $\mu (\tau) > 0$, and $\xi_{\mu (\tau)} (L) = L$ whenever $\mu (\tau) =0$
(see \cite[Chapter 2.2]{bohn}). Note that, for every $L \geq 0$ and every $c \in \mathbb{T}$, the function $e_L (\cdot,c)$ is positive and increasing on $[c,+\infty)_\mathbb{T}$.
\begin{lemma}[\cite{bohn}]\label{lemgronwall}
Let $c < d$ be two elements of $\mathbb{T}$, let $L_1$ and $L_2$ be two nonnegative real numbers, and let $q \in \mathrm{C} ([c,d]_\mathbb{T}, \mathbb{R} )$ satisfying
$0 \leq q(t) \leq L_1 + L_2 \int_{[c,t)_\mathbb{T}} q(\tau) \;\Delta \tau,$
for every $t \in [c,d]_\mathbb{T}$. Then
$0 \leq q(t) \leq L_1 e_{L_2} (t,c),$
for every $t \in [c,d]_\mathbb{T}$.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{prop30-1}]
By continuity of $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$, the set
$$
K_R = \{ (t,x,v) \in [a,b]_\mathbb{T}\times\mathbb{R}^{n} \times \overline{\mathrm{B}}_{\mathbb{R}^m}(0,R) \mid \Vert x-q(t,u,q_a) \Vert_{\mathbb{R}^n} \leq 1 \}
$$
is a compact subset of $\mathbb{T} \times \mathbb{R}^n \times \mathbb{R}^m $. Therefore $ \Vert \partial f / \partial q \Vert $ and $ \Vert \partial f / \partial u \Vert $ are bounded on $K_R$ by some $L_R \geq 0$ and then, from convexity and from the mean value inequality, it holds:
\begin{equation}\label{eqlipK}
\Vert f(t,x_2,v_2) - f(t,x_1,v_1) \Vert_{\mathbb{R}^n} \leq L_R ( \Vert x_2 -x_1 \Vert_{\mathbb{R}^n} + \Vert v_2 -v_1 \Vert_{\mathbb{R}^m} ),
\end{equation}
for all $(t,x_1,v_1),(t,x_2,v_2) \in K_R$.
Let $\nu_R >0$ and $0 < \eta_R <1$ such that $ (\eta_R + \nu_R L_R) e_{L_R} (b,a) < 1$.
Let $(u',q'_a) \in \mathrm{E}_R(u,q_a)$. Our aim is to prove that $b \in I_\mathbb{T}(u',q'_a)$.
By contradiction, assume that the set
$$ A = \{ t \in I_\mathbb{T}(u',q'_a) \cap [a,b]_\mathbb{T}\ \vert \ \Vert q(t,u',q'_a) - q(t,u,q_a) \Vert_{\mathbb{R}^n} > 1 \}$$
is not empty and let $t_0 = \inf A$. Since $\mathbb{T}$ is closed, $t_0 \in I_\mathbb{T}(u',q'_a) \cap [a,b]_\mathbb{T}$ and $[a,t_0]_\mathbb{T} \subset I_\mathbb{T}(u',q'_a) \cap [a,b]_\mathbb{T}$. If $t_0$ is a minimum then $ \Vert q(t_0,u',q'_a) - q(t_0,u,q_a) \Vert_{\mathbb{R}^n} > 1$. If $t_0$ is not a minimum then $t_0 \in \mathrm{RD}$ and by continuity we have $ \Vert q(t_0,u',q'_a) - q(t_0,u,q_a) \Vert_{\mathbb{R}^n} \geq 1$. Moreover one has $t_0 > a$ since $\Vert q(a,u',q'_a) - q(a,u,q_a) \Vert_{\mathbb{R}^n} = \Vert q'_a - q_a \Vert_{\mathbb{R}^n} \leq \eta_R < 1$. Hence $\Vert q(\tau,u',q'_a) - q(\tau,u,q_a) \Vert_{\mathbb{R}^n} \leq 1$ for every $\tau \in [a,t_0)_\mathbb{T}$. Therefore $(\tau,q(\tau,u',q'_a),u'^{\Phi}(\tau))$ and $(\tau,q(\tau,u,q_a),u^{\Phi}(\tau))$ are elements of $K_R$ for $\Delta$-a.e. $\tau \in [a,t_0)_\mathbb{T}$.
Since one has
$$
q(t,u',q'_a) - q(t,u,q_a) = q'_a - q_a + \int_{[a,t)_\mathbb{T}} \left( f(\tau,q(\tau,u',q'_a),u'^{\Phi}(\tau)) - f(\tau,q(\tau,u,q_a),u^{\Phi}(\tau)) \right) \Delta \tau ,
$$
for every $t \in I_\mathbb{T}(u',q'_a) \cap [a,b]_\mathbb{T}$, it follows from~\eqref{eqlipK} that, for every $t \in [a,t_0]_\mathbb{T}$,
\begin{multline*}
\Vert q(t,u',q'_a) - q(t,u,q_a) \Vert_{\mathbb{R}^n} \leq \Vert q'_a - q_a \Vert_{\mathbb{R}^n} + L_R \int_{[a,t)_\mathbb{T}} \Vert u'^{\Phi}(\tau) - u^{\Phi}(\tau) \Vert_{\mathbb{R}^m} \, \Delta \tau \\
+ L_R \int_{[a,t)_\mathbb{T}} \Vert q(\tau,u',q'_a) - q(\tau,u,q_a) \Vert_{\mathbb{R}^n} \, \Delta \tau,
\end{multline*}
which implies from Lemma~\ref{lemgronwall} that, for every $t \in [a,t_0]_\mathbb{T}$,
\begin{equation*}
\Vert q(t,u',q'_a) - q(t,u,q_a) \Vert_{\mathbb{R}^n} \leq (\Vert q'_a - q_a \Vert_{\mathbb{R}^n} + L_R \Vert u'^{\Phi}- u^{\Phi} \Vert_{\mathrm{L}^1_\mathbb{T}([a,b)_\mathbb{T},\mathbb{R}^m)}) e_{L_R} (b,a),
\end{equation*}
which finally implies from Proposition~\ref{propnumber1} that, for every $t \in [a,t_0]_\mathbb{T}$,
\begin{equation*}
\begin{split}
\Vert q(t,u',q'_a) - q(t,u,q_a) \Vert_{\mathbb{R}^n} & \leq (\Vert q'_a - q_a \Vert_{\mathbb{R}^n} + L_R \Vert u'- u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}) e_{L_R} (b,a) \\
& \leq (\eta_R + \nu_R L_R) e_{L_R} (b,a) < 1.
\end{split}
\end{equation*}
This raises a contradiction at $t=t_0$. Therefore $A$ is empty and thus $q(\cdot,u',q'_a)$ is bounded on $I_\mathbb{T}(u',q'_a) \cap [a,b]_\mathbb{T}$. It follows from \cite[Theorem 2]{bour10} that $b \in I_\mathbb{T}(u',q'_a)$, that is, $(u',q'_a) \in \mathcal{UQ}^{b}$.
\end{proof}
\begin{remark}\label{rmK}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$ and $(u',q'_a) \in \mathrm{E}_R(u,q_a)$. With the notations of the above proof, since $[a,b]_\mathbb{T} \subset I_\mathbb{T}(u',q'_a)$ and $A$ is empty, we infer that $\Vert q(t,u',q'_a) - q(t,u,q_a) \Vert_{\mathbb{R}^n} \leq 1$, for every $t \in [a,b]_\mathbb{T}$.
Therefore $(\tau,q(\tau,u',q'_a),u'^{\Phi}(\tau)) \in K_R$ for every $(u',q'_a) \in \mathrm{E}_R(u,q_a)$ and for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$.
\end{remark}
\begin{lemma}\label{prop30-1-1}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$. The mapping
\begin{equation*}
\fonction{F_R(u,q_a)}{(\mathrm{E}_R(u,q_a),\Vert \cdot \Vert_{\mathcal{UQ}^{b}})}{(\mathrm{C}([a,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{(u',q'_a)}{q(\cdot,u',q'_a)}
\end{equation*}
is Lipschitzian.
In particular, for every $(u',q'_a) \in \mathrm{E}_R(u,q_a)$, $q(\cdot,u',q'_a)$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ when $u'$ tends to $u$ in $\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)$ and $q'_a $ tends to $q_a$ in $\mathbb{R}^n$.
\end{lemma}
\begin{proof}
Let $(u',q'_a)$ and $(u'',q''_a)$ be two elements of $\mathrm{E}_R(u,q_a) \subset \mathcal{UQ}^{b}$. It follows from Remark~\ref{rmK} that $(\tau,q(\tau,u'',q''_a),u''^{\Phi}(\tau))$ and $(\tau,q(\tau,u',q'_a),u'^{\Phi}(\tau))$ are elements of $K_R$ for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$. Following the same arguments as in the previous proof, it follows from~\eqref{eqlipK}, from Lemma~\ref{lemgronwall} and from Proposition~\ref{propnumber1} that, for every $t \in [a,b]_\mathbb{T}$,
\begin{equation*}
\Vert q(t,u'',q''_a) - q(t,u',q'_a) \Vert_{\mathbb{R}^n} \leq (\Vert q''_a - q'_a \Vert_{\mathbb{R}^n} + L_R \Vert u'' - u' \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}}([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}) e_{L_R} (b,a).
\end{equation*}
The lemma follows.
\end{proof}
\subsubsection{Needle-like variation of $u$ at a point $r \in \mathrm{RS}_1$}\label{section31RS}
Let $r \in [a,b)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ and let $y \in \mathbb{R}^m$. We define the needle-like variation $\Pi = (r,y)$ of $u$ at $r$ by
\begin{equation*}
u_\Pi (t,\alpha) = \left\{ \begin{array}{lcl}
u(r) + \alpha (y-u(r)) & \textrm{if} & t=r, \\
u(t) & \textrm{if} & t \neq r,
\end{array} \right.
\end{equation*}
for $\Delta_1$-a.e. $t \in [a,b)_{{\mathbb{T}_1}}$ and for every $\alpha \in [0,1]$. In the sequel, let us denote by $\sigma^*_1(r) = \min(\sigma_1(r),b)$.
\begin{lemma}\label{lem32-1}
There exists $0 < \alpha_0 \leq 1$ such that $(u_\Pi (\cdot,\alpha),q_a) \in \mathcal{UQ}^{b}$ for every $\alpha \in [0,\alpha_0]$.
\end{lemma}
\begin{proof}
Let $R = \max ( \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)},\Vert u(r) \Vert_{\mathbb{R}^m} + \Vert y \Vert_{\mathbb{R}^m} ) +1 > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$. We use the notations $K_R$, $L_R$, $\nu_R$ and $\eta_R$, defined in Lemma~\ref{prop30-1} and in its proof. One has $\Vert u_\Pi (\cdot,\alpha) \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} \leq R$ for every $\alpha \in [0,1]$, and
$$
\Vert u_\Pi (\cdot,\alpha) - u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} = \mu_1 (r) \Vert u_\Pi (r,\alpha) - u(r) \Vert_{\mathbb{R}^m} = \alpha \mu_1 (r) \Vert y - u(r) \Vert_{\mathbb{R}^m} .
$$
Hence, there exists $0 < \alpha_0 \leq 1$ such that $\Vert u_\Pi (\cdot,\alpha) - u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} \leq \nu_R$ for every $\alpha \in [0,\alpha_0]$, and hence $(u_\Pi (\cdot,\alpha),q_a) \in \mathrm{E}_R(u,q_a)$. The claim follows then from Lemma~\ref{prop30-1}.
\end{proof}
\begin{lemma}\label{lem32-1-1}
The mapping
\begin{equation*}
\fonction{F_\Pi(u,q_a)}{([0,\alpha_0],\vert \cdot \vert)}{(\mathrm{C}([a,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\alpha}{q (\cdot,u_\Pi (\cdot,\alpha),q_a)}
\end{equation*}
is Lipschitzian.
In particular, for every $\alpha \in [0,\alpha_0]$, $q (\cdot,u_\Pi (\cdot,\alpha),q_a)$ converges uniformly to $q (\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $\alpha$ tends to $0$.
\end{lemma}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem32-1}. It follows from Lemma~\ref{prop30-1-1} that there exists $C \geq 0$ (the Lipschitz constant of $F_R(u,q_a)$) such that
\begin{equation*}
\begin{split}
\Vert q (\cdot,u_\Pi (\cdot,\alpha_2),q_a) - q (\cdot,u_\Pi (\cdot,\alpha_1),q_a) \Vert_\infty
& \leq C \Vert (u_\Pi (\cdot,\alpha_2),q_a)- (u_\Pi (\cdot,\alpha_1),q_a) \Vert_{\mathcal{UQ}^{b}} \\
& = C \vert \alpha_2 - \alpha_1 \vert \mu_1 (r) \Vert y - u(r) \Vert_{\mathbb{R}^m} ,
\end{split}
\end{equation*}
for all $\alpha_1$ and $\alpha_2$ in $ [0,\alpha_0]$.
The lemma follows.
\end{proof}
We define the so-called \textit{first variation vector} $h_\Pi (\cdot,u,q_a)$ associated with the needle-like variation $\Pi = (r,y)$ as the unique solution on $[r,\sigma^*_1(r)]_\mathbb{T}$ of the linear $\Delta$-Cauchy problem
\begin{equation}\label{varvect_scatteredh}
h^\Delta(t) = \frac{\partial f}{\partial q} (t,q (t,u,q_a),u(r)) \, h(t) + \frac{\partial f}{\partial u} (t,q (t,u,q_a),u(r)) \, (y-u(r)), \quad
h(r) = 0.
\end{equation}
The existence and uniqueness of $h_\Pi (\cdot,u,q_a)$ are ensured by \cite[Theorem 3]{bour10}.
\begin{proposition}\label{prop32-10}
The mapping
\begin{equation*}
\fonction{F_\Pi(u,q_a)}{ ([0,\alpha_0],\vert \cdot \vert)}{(\mathrm{C} ([r,\sigma^*_1(r)]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\alpha}{q (\cdot,u_\Pi (\cdot,\alpha),q_a)}
\end{equation*}
is differentiable at $0$, and one has $DF_\Pi(u,q_a)(0) = h_\Pi (\cdot,u,q_a)$.
\end{proposition}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem32-1}. Recall that $(\tau,q(\tau,u_\Pi(\cdot,\alpha),q_a),u^\Phi_\Pi (\tau,\alpha)) \in K_R$ for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$ and for every $\alpha \in [0,\alpha_0] $ (see Remark~\ref{rmK}). For every $\alpha \in (0,\alpha_0]$ and every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$, we define
$$
\varepsilon_\Pi (t,\alpha) = \frac{q(t,u_\Pi(\cdot,\alpha),q_a) - q (t,u,q_a)}{\alpha }- h_{\Pi} (t,u,q_a).
$$
It suffices to prove that $\varepsilon_\Pi (\cdot,\alpha)$ converges uniformly to $0$ on $[r,\sigma^*_1(r)]_\mathbb{T}$ as $\alpha$ tends to $0$.
For every $\alpha \in (0,\alpha_0]$, since the function $\varepsilon_\Pi (\cdot,\alpha)$ vanishes at $t=r$ and is absolutely continuous on $[r,\sigma^*_1(r)]_\mathbb{T}$, $\varepsilon_\Pi (t,\alpha) = \int_{[r,t)_\mathbb{T}} \varepsilon^\Delta_\Pi (\tau,\alpha) \,\Delta \tau$
for every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$, where
\begin{equation*}
\begin{split}
\varepsilon_\Pi^\Delta (\tau,\alpha) = & \frac{f(\tau,q(\tau,u_\Pi(\cdot,\alpha),q_a),u(r)+\alpha (y-u(r)))-f(\tau,q (\tau,u,q_a),u(r))}{\alpha} \\ & - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u(r)) \, h_{\Pi} (\tau,u,q_a) - \frac{\partial f}{\partial u} (\tau,q(\tau,u,q_a),u(r)) \, (y-u(r)) ,
\end{split}
\end{equation*}
for $\Delta$-a.e. $\tau \in [r,\sigma^*_1(r))_\mathbb{T}$.
Using the Taylor formula with integral remainder, we get
\begin{equation*}\begin{split}
\varepsilon_\Pi^\Delta (\tau,\alpha) = & \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta \ \varepsilon_\Pi (\tau,\alpha)\\ & + \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u(r)) \right) h_{\Pi} (\tau,u,q_a) \\ & + \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial u} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial u} (\tau,q(\tau,u,q_a),u(r)) \right) (y-u(r)),
\end{split}\end{equation*}
where
$$ \star_{\theta \alpha \tau} = ( \tau, q (\tau,u,q_a) + \theta (q(\tau,u_\Pi(\cdot,\alpha),q_a) - q (\tau,u,q_a)),u(r)+\theta \alpha (y-u(r)) ) \in K_R.$$
It follows that
$\Vert \varepsilon_\Pi^\Delta (\tau,\alpha) \Vert_{\mathbb{R}^n} \leq \chi_\Pi (\tau,\alpha) + L_R \Vert \varepsilon_\Pi (\tau,\alpha) \Vert_{\mathbb{R}^n}$, where
\begin{equation*}\begin{split}
\chi_\Pi (\tau,\alpha) = & \left\Vert \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u(r)) \right) \, h_{\Pi} (\tau,u,q_a) \right\Vert_{\mathbb{R}^n} \\ & + \left\Vert \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial u} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial u} (\tau,q(\tau,u,q_a),u(r)) \right) \, (y-u(r)) \right\Vert_{\mathbb{R}^n}.
\end{split}\end{equation*}
Therefore, one has
\begin{equation*}
\Vert \varepsilon_\Pi (t,\alpha) \Vert_{\mathbb{R}^n} \leq \int_{[r,\sigma^*_1(r))_\mathbb{T}} \chi_\Pi (\tau,\alpha) \,\Delta \tau +L_R \int_{[r,t)_\mathbb{T}} \Vert \varepsilon_\Pi (\tau,\alpha) \Vert_{\mathbb{R}^n} \, \Delta \tau,
\end{equation*}
for every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$. From Lemma~\ref{lemgronwall},
$\Vert \varepsilon_\Pi (t,\alpha) \Vert_{\mathbb{R}^n} \leq \Upsilon_\Pi (\alpha) e_{L_R} (\sigma^*_1(r),r)$,
for every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$, where
$\Upsilon_\Pi (\alpha) = \int_{[r,\sigma^*_1(r))_\mathbb{T}} \chi_\Pi (\tau,\alpha) \,\Delta \tau $.
To conclude, it remains to prove that $\Upsilon_\Pi (\alpha)$ converges to $0$ as $\alpha$ tends to $0$. Since $q(\tau,u_\Pi(\cdot,\alpha),q_a)$ converges uniformly to $q(\tau,u,q_a)$ on $[r,\sigma^*_1(r)]_\mathbb{T}$ as $\alpha$ tends to $0$ (see Lemma~\ref{lem32-1-1}) and since $\partial f / \partial q$ and $\partial f / \partial u$ are uniformly continuous on $K_R$, the conclusion follows.
\end{proof}
Then, we define the so-called \textit{second variation vector} $w_\Pi (\cdot,u,q_a)$ associated with the needle-like variation $\Pi = (r,y)$ as the unique solution on $[\sigma^*_1(r),b]_\mathbb{T}$ of the linear $\Delta$-Cauchy problem
\begin{equation}\label{varvect_scatteredw}
w^\Delta(t) = \frac{\partial f}{\partial q} (t,q (t,u,q_a),u^\Phi(t)) \, w(t) , \quad
w(\sigma^*_1(r)) = h_\Pi(\sigma^*_1(r),u,q_a).
\end{equation}
The existence and uniqueness of $w_\Pi (\cdot,u,q_a)$ are ensured by \cite[Theorem 3]{bour10}.
\begin{proposition}\label{prop32-1}
The mapping
\begin{equation*}
\fonction{F_\Pi(u,q_a)}{ ([0,\alpha_0],\vert \cdot \vert)}{(\mathrm{C} ([\sigma^*_1(r),b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\alpha}{q (\cdot,u_\Pi (\cdot,\alpha),q_a)}
\end{equation*}
is differentiable at $0$, and one has $DF_\Pi(u,q_a)(0) = w_\Pi (\cdot,u,q_a)$.
\end{proposition}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem32-1}. From Proposition~\ref{prop32-10}, the case $\sigma^*_1(r) = b$ is already proved. As a consequence, we only focus here on the case $\sigma^*_1(r) = \sigma_1(r) < b$.
Recall that $(\tau,q(\tau,u_\Pi(\cdot,\alpha),q_a),u^\Phi_\Pi (\tau,\alpha)) \in K_R$ for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$ and for every $\alpha \in [0,\alpha_0] $ (see Remark~\ref{rmK}). For every $\alpha \in (0,\alpha_0]$ and every $t \in [\sigma_1(r),b]_\mathbb{T}$, we define
$$
\varepsilon_\Pi (t,\alpha) = \frac{q(t,u_\Pi(\cdot,\alpha),q_a) - q (t,u,q_a)}{\alpha }- w_{\Pi} (t,u,q_a).
$$
It suffices to prove that $\varepsilon_\Pi (\cdot,\alpha)$ converges uniformly to $0$ on $[\sigma_1(r),b]_\mathbb{T}$ as $\alpha$ tends to $0$.
For every $\alpha \in (0,\alpha_0]$, since the function $\varepsilon_\Pi (\cdot,\alpha)$ is absolutely continuous on $[\sigma_1(r),b]_\mathbb{T}$, we have
$\varepsilon_\Pi (t,\alpha) = \varepsilon_\Pi (\sigma_1(r),\alpha) + \int_{[\sigma_1(r),t)_\mathbb{T}} \varepsilon^\Delta_\Pi (\tau,\alpha) \, \Delta \tau$
for every $t \in [\sigma_1(r),b]_\mathbb{T}$, where
\begin{multline*}
\varepsilon_\Pi^\Delta (\tau,\alpha) = \frac{f(\tau,q(\tau,u_\Pi(\cdot,\alpha),q_a),u^\Phi(\tau))-f(\tau,q (\tau,u,q_a),u^\Phi(\tau))}{\alpha} \\
- \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \, w_{\Pi} (\tau,u,q_a) ,
\end{multline*}
for $\Delta$-a.e. $\tau \in [\sigma_1(r),b)_\mathbb{T}$.
Using the Taylor formula with integral remainder, we get
\begin{multline*}
\varepsilon_\Pi^\Delta (\tau,\alpha) = \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta \, \varepsilon_\Pi (\tau,\alpha)\\ + \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{\Pi} (\tau,u,q_a),
\end{multline*}
where:
$$ \star_{\theta \alpha \tau} = ( \tau, q (\tau,u,q_a) + \theta (q(\tau,u_\Pi(\cdot,\alpha),q_a) - q (\tau,u,q_a)),u^\Phi(\tau) ) \in K_R.$$
It follows that
$\Vert \varepsilon_\Pi^\Delta (\tau,\alpha) \Vert_{\mathbb{R}^n} \leq \chi_\Pi (\tau,\alpha) + L_R \Vert \varepsilon_\Pi (\tau,\alpha) \Vert_{\mathbb{R}^n}$, where
\begin{equation*}
\chi_\Pi (\tau,\alpha) = \left\Vert \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \alpha \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau, q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{\Pi} (\tau,u,q_a) \right\Vert_{\mathbb{R}^n}.
\end{equation*}
Therefore, one has
\begin{equation*}
\Vert \varepsilon_\Pi (t,\alpha) \Vert_{\mathbb{R}^n} \leq \Vert \varepsilon_\Pi (\sigma_1(r),\alpha) \Vert_{\mathbb{R}^n} + \int_{[\sigma_1(r),b)_\mathbb{T}} \chi_\Pi (\tau,\alpha) \,\Delta \tau +L_R \int_{[\sigma_1(r),t)_\mathbb{T}} \Vert \varepsilon_\Pi (\tau,\alpha) \Vert_{\mathbb{R}^n} \, \Delta \tau,
\end{equation*}
for every $t \in [\sigma_1(r),b]_\mathbb{T}$. It follows from Lemma~\ref{lemgronwall} that
$\Vert \varepsilon_\Pi (t,\alpha) \Vert_{\mathbb{R}^n} \leq \Upsilon_\Pi (\alpha) e_{L_R} (b,\sigma_1(r))$,
for every $t \in [\sigma_1(r),b]_\mathbb{T}$, where
$\Upsilon_\Pi (\alpha) = \Vert \varepsilon_\Pi (\sigma_1(r),\alpha) \Vert_{\mathbb{R}^n} + \int_{[\sigma_1(r),b)_\mathbb{T}} \chi_\Pi (\tau,\alpha) \, \Delta \tau$.
To conclude, it remains to prove that $\Upsilon_\Pi (\alpha)$ converges to $0$ as $\alpha$ tends to $0$.
First, from Proposition~\ref{prop32-10}, it is easy to see that $\Vert \varepsilon_\Pi (\sigma_1(r),\alpha) \Vert_{\mathbb{R}^n} $ converges to $0$ as $\alpha$ tends to $0$.
Second, since $q(\tau,u_\Pi(\cdot,\alpha),q_a)$ converges uniformly to $q(\tau,u,q_a)$ on $[\sigma_1(r),b]_\mathbb{T}$ as $\alpha$ tends to $0$ (see Lemma~\ref{lem32-1-1}) and since $\partial f / \partial q$ is uniformly continuous on $K_R$, we infer that $\int_{[\sigma_1(r),b)_\mathbb{T}} \chi_\Pi (\tau,\alpha) \, \Delta \tau$ converges to $0$ as $\alpha$ tends to $0$. The conclusion follows.
\end{proof}
\begin{lemma}\label{lem32-2h}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b)_{\mathbb{T}_1},\mathbb{R}^m)}$ and let $(u_k,q_{a,k})_{k \in \mathbb{N}}$ be a sequence of elements of $\mathrm{E}_R(u,q_a)$. If $u_k$ converges to $u$ $\Delta_1$-a.e. on $[a,b)_{\mathbb{T}_1}$ and $q_{a,k}$ converges to $q_a$ in $\mathbb{R}^n$ as $k$ tends to $+\infty$, then
$h_{\Pi}(\cdot,u_k,q_{a,k})$ converges uniformly to $h_\Pi(\cdot,u,q_a)$ on $[r,\sigma^*_1(r)]_\mathbb{T}$ as $k$ tends to $+\infty$.
\end{lemma}
\begin{proof}
We use the notations $K_R$, $L_R$, $\nu_R$ and $\eta_R$, defined in Lemma~\ref{prop30-1} and in its proof.
Let us consider the absolutely continuous function defined by $\varepsilon_k (\cdot) = h_{\Pi}(\cdot,u_k,q_{a,k}) - h_{\Pi}(\cdot,u,q_a)$ on $[r,\sigma^*_1(r)]_\mathbb{T}$. Let us prove that $\varepsilon_k$ converges uniformly to $0$ on $[r,\sigma^*_1(r)]_\mathbb{T}$ as $k$ tends to $+\infty$. Since $\varepsilon_k (r) = 0$, one has
\begin{equation*}\begin{split}
\varepsilon_k (t) = & \int_{[r,t)_\mathbb{T}} \frac{\partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) \, \varepsilon_k (\tau) \,\Delta \tau \\
& + \int_{[r,t)_\mathbb{T}} \left(\frac{ \partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u (r)) \right) \, h_{\Pi} (\tau,u,q_a) \,\Delta \tau \\
& + \int_{[r,t)_\mathbb{T}} \left( \frac{ \partial f}{\partial u} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) - \frac{\partial f}{\partial u} (\tau,q(\tau,u,q_a),u (r)) \right) \, (y-u(r)) \,\Delta \tau \\
& + \int_{[r,t)_\mathbb{T}} \frac{ \partial f}{\partial u} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) \, (u(r)-u_k(r)) \,\Delta \tau
\end{split}\end{equation*}
for every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$ and every $k \in \mathbb{N}$. Since $(u_k,q_{a,k}) \in \mathrm{E}_R(u,q_a)$ for every $k \in \mathbb{N}$, it follows from Remark~\ref{rmK} that $(\tau,q(\tau,u_k,q_{a,k}),u^\Phi_k(\tau)) \in K_R$ for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$. Hence it follows from Lemma~\ref{lemgronwall} that
$ \Vert \varepsilon_k (t) \Vert_{\mathbb{R}^n} \leq \Upsilon_k e_{L_R} (b,r),$
for every $t \in [r,\sigma^*_1(r)]_\mathbb{T}$, where $\Upsilon_k$ is given by
\begin{equation*}\begin{split}
\Upsilon_k & = L_R \mu_1(r) \Vert u(r)-u_k(r) \Vert_{\mathbb{R}^m} \\
& + \int_{[r,\sigma^*_1(r))_\mathbb{T}} \left\Vert \frac{ \partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u (r)) \right\Vert \Vert h_{\Pi} (\tau,u,q_a) \Vert_{\mathbb{R}^n} \,\Delta \tau \\
& + \int_{[r,\sigma^*_1(r))_\mathbb{T}} \left\Vert \frac{ \partial f}{\partial u} (\tau,q(\tau,u_k,q_{a,k}),u_k (r)) - \frac{\partial f}{\partial u} (\tau,q(\tau,u,q_a),u (r)) \right\Vert \Vert y-u(r) \Vert_{\mathbb{R}^m} \,\Delta \tau .
\end{split}\end{equation*}
Since $\mu_{\Delta_1} (\{ r\} ) = \mu_1 (r) > 0$, $u_k(r)$ converges to $u(r)$ as $k$ tends to $+\infty$. Moreover, from the Lebesgue dominated convergence theorem, $(u_k,q_{a,k})$ converges to $(u,q_a)$ in $(\mathrm{E}_R(u,q_a),\Vert \cdot \Vert_{\mathcal{UQ}^{b}} )$ and, from Lemma~\ref{prop30-1-1}, $q(\cdot,u_k,q_{a,k})$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $k$ tends to $+\infty$. Since $\partial f / \partial q$ and $\partial f / \partial u$ are uniformly continuous on $K_R$, we conclude that $\Upsilon_k$ converges to $0$ as $k$ tends to $+\infty$. The lemma follows.
\end{proof}
\begin{lemma}\label{lem32-2w}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b)_{\mathbb{T}_1},\mathbb{R}^m)}$ and let $(u_k,q_{a,k})_{k \in \mathbb{N}}$ be a sequence of elements of $\mathrm{E}_R(u,q_a)$. If $u_k$ converges to $u$ $\Delta_1$-a.e. on $[a,b)_{\mathbb{T}_1}$ and $q_{a,k}$ converges to $q_a$ in $\mathbb{R}^n$ as $k$ tends to $+\infty$, then
$w_{\Pi}(\cdot,u_k,q_{a,k})$ converges uniformly to $w_\Pi(\cdot,u,q_a)$ on $[\sigma^*_1(r),b]_\mathbb{T}$ as $k$ tends to $+\infty$.
\end{lemma}
\begin{remark}
With the same arguments of the proof of Lemma~\ref{lemnumber1}, the hypotheses of Lemma~\ref{lem32-2w} give the convergences of $u^\Phi_k$ to $u^\Phi$ $\Delta$-a.e. on $[a,b)_\mathbb{T}$ and, from the Lebesgue dominated convergence theorem, of $(u_k,q_{a,k})$ to $(u,q_a)$ in $(\mathrm{E}_R(u,q_a),\Vert \cdot \Vert_{\mathcal{UQ}^{b}} )$.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{lem32-2w}]
We use the notations $K_R$, $L_R$, $\nu_R$ and $\eta_R$, defined in Lemma~\ref{prop30-1} and in its proof. From Lemma~\ref{lem32-2h}, the case $\sigma^*_1(r) = b$ is already proved. As a consequence, we only focus here on the case $\sigma^*_1(r) = \sigma_1(r) < b$.
Let us consider the absolutely continuous function defined by $\varepsilon_k (\cdot) = w_{\Pi}(\cdot,u_k,q_{a,k}) - w_{\Pi}(\cdot,u,q_a)$ on $[\sigma_1(r),b]_\mathbb{T}$. Let us prove that $\varepsilon_k$ converges uniformly to $0$ on $[\sigma_1(r),b]_\mathbb{T}$ as $k$ tends to $+\infty$. One has
\begin{multline*}
\varepsilon_k (t) = \varepsilon_k (\sigma_1(r)) + \int_{[\sigma_1(r),t)_\mathbb{T}} \frac{\partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u^\Phi_k (\tau)) \, \varepsilon_k (\tau) \,\Delta \tau \\
+ \int_{[\sigma_1(r),t)_\mathbb{T}} \left(\frac{ \partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u^\Phi_k (\tau)) - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{\Pi} (\tau,u,q_a) \,\Delta \tau ,
\end{multline*}
for every $t \in [\sigma_1(r),b]_\mathbb{T}$ and every $k \in \mathbb{N}$. Since $(u_k,q_{a,k}) \in \mathrm{E}_R(u,q_a)$ for every $k \in \mathbb{N}$, it follows from Remark~\ref{rmK} that $(\tau,q(\tau,u_k,q_{a,k}),u^\Phi_k(\tau)) \in K_R$ for $\Delta$-a.e. $\tau \in [\sigma_1(r),b)_\mathbb{T}$. Hence it follows from Lemma~\ref{lemgronwall} that
$$ \Vert \varepsilon_k (t) \Vert_{\mathbb{R}^n} \leq (\varepsilon_k (\sigma_1(r)) + \Upsilon_k ) e_{L_R} (b,\sigma_1(r)),$$
for every $t \in [\sigma_1(r),b]_\mathbb{T}$, where $\Upsilon_k$ is given by
\begin{equation*}
\Upsilon_k = \int_{[\sigma_1(r),b)_\mathbb{T}} \left\Vert \frac{ \partial f}{\partial q} (\tau,q(\tau,u_k,q_{a,k}),u^\Phi_k (\tau)) - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right\Vert \Vert w_{\Pi} (\tau,u,q_a) \Vert_{\mathbb{R}^n} \,\Delta \tau.
\end{equation*}
Since $(u_k,q_{a,k})$ converges to $(u,q_a)$ in $(\mathrm{E}_R(u,q_a),\Vert \cdot \Vert_{\mathcal{UQ}^{b}})$ and from Lemma~\ref{prop30-1-1}, $q(\cdot,u_k,q_{a,k})$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $k$ tends to $+\infty$. Since $\partial f / \partial q$ is continuous and bounded on $K_R$ and since $u^\Phi_k$ converges to $u^\Phi$ $\Delta$-a.e. on $[a,b)_\mathbb{T}$, the Lebesgue dominated convergence theorem concludes that $\Upsilon_k$ converges to $0$ as $k$ tends to $+\infty$. Finally, from Lemma~\ref{lem32-2h}, one has $\varepsilon_k (\sigma_1(r)) = w_{\Pi}(\sigma_1(r),u_k,q_{a,k}) - w_{\Pi}(\sigma_1(r),u,q_a) = h_{\Pi}(\sigma_1(r),u_k,q_{a,k}) - h_{\Pi}(\sigma_1(r),u,q_a) $
converges to $0$ as $k$ tends to $+\infty$. The lemma follows.
\end{proof}
\subsubsection{Needle-like variation of $u$ at a point $s \in \mathrm{RD}_1$}\label{section31RD}
Let $s \in \mathscr{L}_{[a,b)_\mathbb{T}} (f(\cdot,q(\cdot,u,q_a),u^\Phi)) \cap \mathrm{RD}_1$ and $z \in \mathbb{R}^m$. Note that $s \in {\mathbb{T}_1}$ and then $\Phi(s)=s$. We define the needle-like variation $\amalg = (s,z)$ of $u$ at $s$ by
\begin{equation*}
u_\amalg (t,\beta) = \left\{ \begin{array}{lcl}
z & \textrm{if} & t \in [s,s+\beta)_{{\mathbb{T}_1}} , \\
u(t) & \textrm{if} & t \notin [s,s+\beta)_{{\mathbb{T}_1}}.
\end{array} \right.
\end{equation*}
for $\Delta_1$-a.e. $t \in [a,b)_{{\mathbb{T}_1}}$ and for every $\beta \in \mathcal{V}^{s,b}_1 \subset \mathcal{V}^{s,b}$.
\begin{lemma}\label{lem33-1}
There exists $\beta_0 > 0$ such that $(u_\amalg (\cdot,\beta),q_a) \in \mathcal{UQ}^{b}$ for every $\beta \in \mathcal{V}^{s,b}_1 \cap [0,\beta_0]$.
\end{lemma}
\begin{proof}
Let $R = \max ( \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)},\Vert z \Vert_{\mathbb{R}^m}) +1 > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$. We use the notations $K_R$, $L_R$, $\nu_R$ and $\eta_R$, defined in the proof of Lemma~\ref{prop30-1}.
For every $\beta \in \mathcal{V}^{s,b}_1$, we have $\Vert u_\amalg (\cdot,\beta) \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} \leq R$ and
$$
\Vert u_\amalg (\cdot,\beta) - u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} = \int_{[s,s+\beta)_{{\mathbb{T}_1}}} \Vert z - u(\tau) \Vert_{\mathbb{R}^m} \,\Delta_1 \tau \leq 2 R \beta.
$$
Hence, there exists $\beta_0 > 0$ such that for every $\beta \in \mathcal{V}^{s,b}_1 \cap [0,\beta_0]$, $\Vert u_\amalg (\cdot,\beta) - u \Vert_{\mathrm{L}^1_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)} \leq \nu_R$ and thus $(u_\amalg (\cdot,\beta),q_a) \in \mathrm{E}_R(u,q_a)$. The conclusion then follows from Lemma~\ref{prop30-1}.
\end{proof}
\begin{lemma}\label{lem33-1-1}
The mapping
\begin{equation*}
\fonction{F_\amalg(u,q_a)}{(\mathcal{V}^{s,b}_1 \cap [0,\beta_0],\vert \cdot \vert)}{(\mathrm{C}([a,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\beta}{q (\cdot,u_\amalg (\cdot,\beta),q_a)}
\end{equation*}
is Lipschitzian.
In particular, for every $\beta \in \mathcal{V}^{s,b}_1 \cap [0,\beta_0]$, $q (\cdot,u_\amalg (\cdot,\beta),q_a)$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $\beta$ tends to $0$.
\end{lemma}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem33-1}. From Lemma~\ref{prop30-1-1}, there exists $C \geq 0$ (Lipschitz constant of $F_R(u,q_a)$) such that
\begin{equation*}\begin{split}
\Vert q (\cdot,u_\amalg (\cdot,\beta_2),q_a) - q (\cdot,u_\amalg (\cdot,\beta_1),q_a) \Vert_\infty
& \leq C \Vert (u_\amalg (\cdot,\beta_2),q_a) - (u_\amalg (\cdot,\beta_1),q_a) \Vert_{\mathcal{UQ}^{b}} \\
& \leq 2 C R \vert \beta_2-\beta_1 \vert,
\end{split}\end{equation*}
for all $\beta_1$ and $\beta_2$ in $\mathcal{V}^{s,b}_1 \cap [0,\beta_0]$.
The lemma follows.
\end{proof}
According to \cite[Theorem 3]{bour10}, we define the \textit{variation vector} $w_{\amalg}(\cdot,u,q_a)$ associated with the needle-like variation $\amalg = (s,z)$ as the unique solution on $[s,b]_\mathbb{T}$ of the linear $\Delta$-Cauchy problem
\begin{equation}\label{varvect_densew}
w^\Delta(t) = \frac{\partial f}{\partial q} (t,q(t,u,q_a),u^\Phi(t)) \, w(t) , \quad w(s) = f(s,q(s,u,q_a),z) - f(s,q(s,u,q_a),u(s)).
\end{equation}
\begin{proposition}\label{prop33-1}
For every $\delta \in \mathcal{V}^{s,b} \backslash \{ 0 \}$, the mapping
\begin{equation*}
\fonction{F_\amalg(u,q_a)}{(\mathcal{V}^{s,b}_1 \cap [0,\beta_0],\vert \cdot \vert)}{(\mathrm{C} ([s+\delta,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\beta}{q (\cdot,u_\amalg (\cdot,\beta),q_a)}
\end{equation*}
is differentiable at $0$, and one has $ DF_\amalg(u,q_a)(0) = w_\amalg (\cdot,u,q_a)$.
\end{proposition}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem33-1}. Recall that $(\tau,q (\tau,u_\amalg (\cdot,\beta),q_a),u^\Phi_\amalg (\tau,\beta))$ and $(\tau,q (\tau,u_\amalg (\cdot,\beta),q_a),z)$ belong to $K_R$ for every $\beta \in \mathcal{V}^{s,b}_1 \cap [0,\beta_0]$ and for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$ (see Remark~\ref{rmK}). For every $\beta \in \mathcal{V}^{s,b}_1 \cap (0,\beta_0]$ and every $t \in [s+\beta,b]_\mathbb{T}$, we define
$$ \varepsilon_\amalg (t,\beta) = \frac{q (t,u_\amalg (\cdot,\beta),q_a) - q(t,u,q_a)}{ \beta } - w_{\amalg} (t,u,q_a). $$
It suffices to prove that $\varepsilon_\amalg (\cdot,\beta)$ converges uniformly to $0$ on $[s+\beta,b]_\mathbb{T}$ as $\beta$ tends to $0$. Note that, for every $\delta \in \mathcal{V}^{s,b} \backslash \{ 0 \}$, it suffices to consider $\beta \leq \delta$.
For every $\beta \in \mathcal{V}^{s,b}_1 \cap (0,\beta_0]$, the function $\varepsilon_\amalg (\cdot,\beta)$ is absolutely continuous on $[s+\beta,b]_\mathbb{T}$ and
$\varepsilon_\amalg (t,\beta) = \varepsilon_\amalg (s+\beta,\beta) + \int_{[s+\beta,t)_\mathbb{T}} \varepsilon^\Delta_\amalg (\tau,\beta) \, \Delta \tau$,
for every $t \in [s+\beta,b]_\mathbb{T}$,
where
\begin{multline*}
\varepsilon_\amalg^\Delta (\tau,\beta) = \frac{f(\tau,q (\tau,u_\amalg (\cdot,\beta),q_a),u^\Phi(\tau))-f(\tau,q (\tau,u,q_a),u^\Phi(\tau))}{\beta} \\
- \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \, w_{\amalg} (\tau,u,q_a),
\end{multline*}
for $\Delta$-a.e. $\tau \in [s+\beta,b)_\mathbb{T}$.
Using the Taylor formula with integral remainder, we get
\begin{multline*}
\varepsilon_\amalg^\Delta (\tau,\beta) = \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \beta \tau}) \,d\theta \, \varepsilon_\amalg (\tau,\beta)\\
+ \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \beta \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{\amalg} (\tau,u,q_a),
\end{multline*}
where:
$$ \star_{\theta \beta \tau} = ( \tau, q (\tau,u,q_a) + \theta (q(\tau,u_\amalg(\cdot,\beta),q_a) - q (\tau,u,q_a)),u^\Phi(\tau) ) \in K_R.$$
It follows that
$\Vert \varepsilon_\amalg^\Delta (\tau,\beta) \Vert_{\mathbb{R}^n} \leq \chi_\amalg (\tau,\beta) + L_R \Vert \varepsilon_\amalg (\tau,\beta) \Vert_{\mathbb{R}^n}$, where
\begin{equation*}\begin{split}
\chi_\amalg (\tau,\beta) = & \left\Vert \left( \displaystyle \int_0^1 \dfrac{\partial f}{\partial q} (\star_{\theta \beta \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{\amalg} (\tau,u,q_a) \right\Vert_{\mathbb{R}^n}.
\end{split}\end{equation*}
Therefore, one has
\begin{equation*}
\Vert \varepsilon_\amalg (t,\beta) \Vert_{\mathbb{R}^n} \leq \Vert \varepsilon_\amalg (s+\beta,\beta) \Vert_{\mathbb{R}^n} + \int_{[s+\beta,b)_\mathbb{T}} \chi_\amalg (\tau,\beta) \,\Delta \tau +L_R \int_{[s+\beta,t)_\mathbb{T}} \Vert \varepsilon_\amalg (\tau,\beta) \Vert_{\mathbb{R}^n} \, \Delta \tau ,
\end{equation*}
for every $t \in [s+\beta,b]_\mathbb{T}$, and it follows from Lemma~\ref{lemgronwall} that
$\Vert \varepsilon_\amalg (t,\beta) \Vert_{\mathbb{R}^n} \leq \Upsilon_\amalg (\beta) e_{L_R} (b,s)$,
for every $t \in [s+\beta,b]_\mathbb{T}$, where
$ \Upsilon_\amalg (\beta) = \Vert \varepsilon_\amalg (s+\beta,\beta) \Vert_{\mathbb{R}^n} + \int_{[s+\beta,b)_\mathbb{T}} \chi_\amalg (\tau,\beta) \, \Delta \tau$.
To conclude, it remains to prove that $\Upsilon_\amalg (\beta)$ converges to $0$ as $\beta$ tends to $0$. Since $q (\cdot,u_\amalg (\cdot,\beta),q_a)$ converges uniformly to $q(\cdot,u,q_a)$ on $[s+\beta,b]_\mathbb{T}$ as $\beta$ tends to $0$ (see Lemma~\ref{lem33-1-1}) and since $\partial f / \partial q$ is uniformly continuous on $K_R$, we first infer that $\int_{[s+\beta,b)_\mathbb{T}} \chi_\amalg (\tau,\beta) \, \Delta \tau$ converges to $0$ as $\beta$ tends to $0$.
Secondly, let us prove that $\Vert \varepsilon_\amalg (s+\beta,\beta) \Vert_{\mathbb{R}^n} $ converges to $0$ as $\beta$ tends to $0$. By continuity, $w_\amalg (s+\beta,u,q_a)$ converges to $w_\amalg (s,u,q_a)$ as $\beta$ to $0$. Moreover, since $q (\cdot,u_\amalg (\cdot,\beta),q_a)$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $\beta$ tends to $0$ and since $f$ is uniformly continuous on $K_R$, it follows that $f(\cdot, q (\cdot,u_\amalg (\cdot,\beta),q_a),z)$ converges uniformly to $f(\cdot, q(\cdot,u,q_a),z)$ on $[a,b]_\mathbb{T}$ as $\beta$ tends to $0$. Therefore, it suffices to note that
$$\frac{1}{ \beta } \int_{[s,s+\beta)_\mathbb{T}} \left( f(\tau,q (\tau,u,q_a),z) - f(\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \Delta \tau$$
converges to $w_\amalg (s,u,q_a) = f(s,q(s,u,q_a),z) - f(s,q(s,u,q_a),u(s))$ as $\beta$ tends to $0$ since $s$ is a $\Delta$-Lebesgue point of $f(\cdot,q(\cdot,u,q_a),z)$ by continuity and of $f(\cdot,q(\cdot,u,q_a),u^\Phi(\cdot))$ by hypothesis.
Then $\Vert \varepsilon_\amalg (s+\beta,\beta) \Vert_{\mathbb{R}^n}$ converges to $0$ as $\beta$ tends to $0$, and hence $\Upsilon_\amalg (\beta)$ converges to $0$ as well.
\end{proof}
\begin{lemma}\label{lem33-2}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b)_{\mathbb{T}_1},\mathbb{R}^m)}$ and let $(u_k,q_{a,k})_{k \in \mathbb{N}}$ be a sequence of elements of $\mathrm{E}_R(u,q_a)$. If $u_k$ converges to $u$ $\Delta_1$-a.e. on $[a,b)_{\mathbb{T}_1}$, $u_k(s)$ converges to $u(s)$ and $q_{a,k}$ converges to $q_a$ as $k$ tends to $+\infty$, then $w_{\amalg}(\cdot,u_k,q_{a,k})$ converges uniformly to $w_\amalg(\cdot,u,q_a)$ on $[s,b]_\mathbb{T}$ as $k$ tends to $+\infty$.
\end{lemma}
\begin{proof}
The proof is similar to the one of Lemma~\ref{lem32-2w}, replacing $\sigma_1(r)$ with $s$.
\end{proof}
\subsubsection{Variation of the initial condition $q_a$}\label{section32}
Let $q'_a \in \mathbb{R}^n$.
\begin{lemma}\label{lem34-1}
There exists $\gamma_0>0$ such that $(u,q_a+\gamma q'_a) \in \mathcal{UQ}^{b}$ for every $\gamma \in [0,\gamma_0]$.
\end{lemma}
\begin{proof}
Let $R = \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}+1 > \Vert u \Vert_{\mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\mathbb{R}^m)}$. We use the notations $K_R$, $L_R$, $\nu_R$ and $\eta_R$, defined in the proof of Lemma~\ref{prop30-1}.
There exists $\gamma_0>0$ such that $\Vert q_a+\gamma q'_a - q_a \Vert_{\mathbb{R}^n} = \gamma \Vert q'_a \Vert_{\mathbb{R}^n} \leq \eta_R$ for every $\gamma \in [0,\gamma_0]$, and hence $(u,q_a+\gamma q'_a) \in \mathrm{E}_R(u,q_a)$. Then the claim follows from Lemma~\ref{prop30-1}.
\end{proof}
\begin{lemma}\label{lem34-1-1}
The mapping
\begin{equation*}
\fonction{F_{q'_a}(u,q_a)}{([0,\gamma_0],\vert \cdot \vert)}{(\mathrm{C}([a,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\gamma}{q (\cdot,u,q_a+\gamma q'_a)}
\end{equation*}
is Lipschitzian.
In particular, for every $\gamma \in [0,\gamma_0]$, $q (\cdot,u,q_a+\gamma q'_a)$ converges uniformly to $q (\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $\gamma$ tends to $0$.
\end{lemma}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem34-1}. From Lemma~\ref{prop30-1-1}, there exists $C \geq 0$ (Lipschitz constant of $F_R(u,q_a)$) such that
\begin{equation*}\begin{split}
\Vert q (\cdot,u,q_a + \gamma_2 q'_a ) - q (\cdot,u,q_a + \gamma_1 q'_a ) \Vert_\infty
& \leq C \Vert (u,q_a+\gamma_2 q'_a)-(u,q_a+\gamma_1 q'_a) \Vert_{\mathcal{UQ}^{b}} \\
& = C \vert \gamma_2 - \gamma_1 \vert \Vert q'_a \Vert_{\mathbb{R}^n} .
\end{split}\end{equation*}
for all $\gamma_1$ and $\gamma_2$ in $[0,\gamma_0]$.
\end{proof}
According to \cite[Theorem 3]{bour10}, we define the \textit{variation vector} $w_{q'_a}(\cdot,u,q_a)$ associated with the perturbation $q_a'$ as the unique solution on $[a,b]_\mathbb{T}$ of the linear $\Delta$-Cauchy problem
\begin{equation}\label{varvect_pointinit}
w^\Delta(t) = \frac{\partial f}{\partial q} (t,q (t,u,q_a),u^\Phi(t)) \, w(t), \quad w(a) = q'_a.
\end{equation}
\begin{proposition}\label{prop34-1}
The mapping
\begin{equation*}
\fonction{F_{q'_a}(u,q_a)}{([0,\gamma_0],\vert \cdot \vert)}{(\mathrm{C} ([a,b]_\mathbb{T},\mathbb{R}^{n}),\Vert \cdot \Vert_\infty)}{\gamma}{q (\cdot,u,q_a+\gamma q'_a )}
\end{equation*}
is differentiable at $0$, and one has $DF_{q'_a}(u,q_a)(0) = w_{q'_a} (\cdot,u,q_a)$.
\end{proposition}
\begin{proof}
We use the notations of the proof of Lemma~\ref{lem34-1}. Note that, from Remark~\ref{rmK}, $(\tau,q (\tau,u,q_a+\gamma q'_a),u^\Phi(\tau)) \in K_R$ for every $\gamma \in [0,\gamma_0]$ and for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$.
For every $\gamma \in (0,\gamma_0]$ and every $t \in [a,b]_\mathbb{T}$, we define
\begin{equation*}
\varepsilon_{q'_a} (t,\gamma) = \frac{q (t,u,q_a+\gamma q'_a) - q(t,u,q_a)}{ \gamma } - w_{q'_a} (t,u,q_a).
\end{equation*}
It suffices to prove that $\varepsilon_{q'_a} (\cdot,\gamma)$ converges uniformly to $0$ on $[a,b]_\mathbb{T}$ as $\gamma$ tends to $0$.
For every $\gamma \in (0,\gamma_0]$, since the function $\varepsilon_{q'_a} (\cdot,\gamma)$ vanishes at $t=a$ and is absolutely continuous on $[a,b]_\mathbb{T}$, $\varepsilon_{q'_a} (t,\gamma) = \int_{[a,t)_\mathbb{T}} \varepsilon^\Delta_{q'_a} (\tau,\gamma) \, \Delta \tau$, for every $t \in [a,b]_\mathbb{T}$, where
\begin{multline*}
\varepsilon_{q'_a}^\Delta (\tau,\gamma) = \frac{f(\tau,q (\tau,u,q_a+\gamma q'_a),u^\Phi(\tau))-f(\tau,q (\tau,u,q_a),u^\Phi(\tau))}{\gamma} \\
- \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau),\tau) \, w_{q'_a} (\tau,u,q_a) ,
\end{multline*}
for $\Delta$-a.e. $\tau \in [a,b)_\mathbb{T}$.
Using the Taylor formula with integral remainder, we get
\begin{multline*}
\varepsilon_{q'_a}^\Delta (\tau,\gamma) = \displaystyle \int_0^1 \frac{\partial f}{\partial q} (\star_{\theta \gamma \tau}) \,d\theta \cdot \varepsilon_{q'_a} (\tau,\gamma) \\
+ \left( \displaystyle \int_0^1 \frac{\partial f}{\partial q} (\star_{\theta \gamma \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{q'_a} (\tau,u,q_a),
\end{multline*}
where
$$ \star_{\theta \gamma \tau} = ( \tau, q(\tau,u,q_a) + \theta (q(\tau,u,q_a+\gamma q'_a) - q(\tau,u,q_a)) , u^\Phi(\tau)) \in K_R. $$
It follows that $\Vert \varepsilon_{q'_a}^\Delta (\tau,\gamma) \Vert_{\mathbb{R}^n} \leq \chi_{q'_a} (\tau,\gamma) + L_R \Vert \varepsilon_{q'_a} (\tau,\gamma) \Vert_{\mathbb{R}^n}$, where
\begin{equation*}
\chi_{q'_a} (\tau,\gamma) = \left\Vert \left( \displaystyle \int_0^1 \frac{\partial f}{\partial q} (\star_{\theta \gamma \tau}) \,d\theta - \frac{\partial f}{\partial q} (\tau,q(\tau,u,q_a),u^\Phi(\tau)) \right) \, w_{q'_a} (\tau,u,q_a) \right\Vert_{\mathbb{R}^n} .
\end{equation*}
Hence
\begin{equation*}
\Vert \varepsilon_{q'_a} (t,\gamma) \Vert_{\mathbb{R}^n} \leq \int_{[a,b)_\mathbb{T}} \chi_{q'_a} (\tau,\gamma) \, \Delta \tau +L_R \int_{[a,t)_\mathbb{T}} \Vert \varepsilon_{q'_a} (\tau,\gamma) \Vert_{\mathbb{R}^n} \, \Delta \tau,
\end{equation*}
for every $t \in [a,b]_\mathbb{T}$, and it follows from Lemma~\ref{lemgronwall} that
$\Vert \varepsilon_{q'_a} (t,\gamma) \Vert_{\mathbb{R}^n} \leq \Upsilon_{q'_a} (\gamma) e_{L_R} (b,a)$,
for every $t \in [a,b]_\mathbb{T}$, where
$
\Upsilon_{q'_a} (\gamma) = \int_{[a,b)_\mathbb{T}} \chi_{q'_a} (\tau,\gamma) \, \Delta \tau.
$
To conclude, it remains to prove that $\Upsilon_{q'_a} (\gamma)$ converges to $0$ as $\gamma$ tends to $0$. Since $q(\cdot,u,q_a+\gamma q'_a)$ converges uniformly to $q(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $\gamma$ tends to $0$ (see Lemma~\ref{lem34-1-1}) and since $\partial f / \partial q$ is uniformly continuous on $K_R$, the conclusion follows.
\end{proof}
\begin{lemma}\label{lem34-2}
Let $R > \Vert u \Vert_{\mathrm{L}^\infty_\mathbb{T}([a,b)_\mathbb{T},\mathbb{R}^m)}$ and let $(u_k,q_{a,k})_{k \in \mathbb{N}}$ be a sequence of elements of $\mathrm{E}_R(u,q_a)$. If $u_k$ converges to $u$ $\Delta_1$-a.e. on $[a,b)_{\mathbb{T}_1}$ and $q_{a,k}$ converges to $q_a$ in $\mathbb{R}^n$ as $k$ tends to $+\infty$, then $w_{q'_a}(\cdot,u_k,q_{a,k})$ converges uniformly to $w_{q'_a}(\cdot,u,q_a)$ on $[a,b]_\mathbb{T}$ as $k$ tends to $+\infty$.
\end{lemma}
\begin{proof}
The proof is similar to the one of Lemma~\ref{lem32-2w}, replacing $\sigma_1(r)$ with $a$.
\end{proof}
\subsection{Proof of Theorem~\ref{thmmain}}\label{annexe3}
We are now in a position to prove the PMP. In the sequel, we consider $q^*$ an optimal trajectory, associated with an optimal sampled-data control $u^*$ and with $b^* \in \mathbb{T}$, with $b^* = b$ if the final time is fixed. We set $q^*_a = q^*(a)$.
\subsubsection{The augmented system
As in \cite{lee,pont}, we consider the \textit{augmented system} in $\mathbb{R}^{n+1}$
\begin{equation} \label{DD-CS_augm}
\bar q^\Delta(t) = \bar f (t,\bar q(t),u^\Phi(t)),
\end{equation}
with $\bar q=(q,q^0)^\top$ the augmented state with values in $\mathbb{R}^n\times\mathbb{R}$, and $\bar f:\mathbb{T}\times\mathbb{R}^{n+1}\times\mathbb{R}^m \rightarrow\mathbb{R}^{n+1}$, the augmented dynamics, defined by
$\bar f(t,\bar q,u)=(f(t,q,u),f^0(t,q,u))^\top$.
Note that $\bar f$ does not depend on $q^0$. We will always impose as an initial condition $q^0(a)=0$, so that $q^0(b)= \int_{[a,b)_\mathbb{T}} f^0 ( \tau , q (\tau), u^\Phi(\tau) ) \, \Delta \tau$. Hence, the additional coordinate $q^0$ stands for the cost.
Denoting $\bar q^*_a = (q^*_a,0)^\top$, we have $(u^*,\bar q_a^*) \in \mathcal{U\bar{Q}}^{b^*}$. We set $ \bar q^* = \bar q(\cdot,u^*,\bar q^*_a) = (q^*,q^{0*})^\top$. Hence, $\bar q^*$ is a solution of~\eqref{DD-CS_augm} on $[a,b^*]_\mathbb{T}$ associated with the control $u^*$, satisfying $q^{0*}(a) = 0$ and $ g( q^*(a), q^*(b^*)) \in \mathrm{S}$ and minimizing $ q^0(b) $
over all possible trajectories $\bar q$ solutions of~\eqref{DD-CS_augm} on $[a,b]_\mathbb{T}$, where $b \in \mathbb{T}$ and associated with a control $u\in \mathrm{L}^\infty_{{\mathbb{T}_1}} ([a,b)_{{\mathbb{T}_1}},\Omega)$, satisfying $q^0(a) = 0$ and $ g( q(a), q(b)) \in \mathrm{S}$.
\subsubsection{Application of the Ekeland variational principle}\label{section41}
For the completeness, we recall a simplified (but sufficient) version of the Ekeland variational principle.
\begin{theorem}[\cite{ekel}
Let $(\mathrm{E},d_\mathrm{E})$ be a complete metric space and let $J : \mathrm{E} \rightarrow \mathbb{R}^+$, $\lambda \mapsto J(\lambda)$ be a continuous nonnegative mapping. Let $\varepsilon > 0$ and $\lambda^* \in \mathrm{E}$ such that $J(\lambda^*) \leq \varepsilon$.
Then, there exists $\lambda_\varepsilon \in \mathrm{E}$ such that $d_\mathrm{E} (\lambda_\varepsilon,\lambda^*) \leq \sqrt{\varepsilon}$ and, $-\sqrt{\varepsilon} \, d_\mathrm{E} (\lambda,\lambda_\varepsilon) \leq J(\lambda) - J(\lambda_\varepsilon)$, for every $\lambda \in \mathrm{E}$.
\end{theorem}
Let $R > \Vert u^* \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1}([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)}$. Recall that $\mathrm{E}_R(u^*,\bar{q}^*_a)\subset\mathcal{U\bar{Q}}^{b^*}$ (see Lemma~\ref{prop30-1}). We set
$$
\mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a) = \{ (u,\bar{q}_a) \in \mathrm{E}_R(u^*,\bar{q}^*_a) \ \vert \ u\in \mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\Omega), \, \bar{q}_a=(q_a,0) \}.
$$
Note that $(u^*,\bar{q}^*_a) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$. Since $\Omega$ is closed, it follows from the (partial) converse of the Lebesgue dominated convergence theorem that $(\mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a),\Vert \cdot \Vert_{\mathcal{U\bar{Q}}^{b^*}})$ is a closed subset of $\mathrm{L}^1_{\mathbb{T}_1}([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m) \times \mathbb{R}^{n+1}$ and then is a complete metric space.
For every $\varepsilon > 0$, we define the functional $J^R_\varepsilon:(\mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a),\Vert \cdot \Vert_{\mathcal{U\bar{Q}}^{b^*}}) \rightarrow \mathbb{R}^+$ by
$$
J^R_\varepsilon(u,\bar{q}_a)
=
\left( \left( \left( q^0(b^*,u,\bar q_a)-q^{0*}(b^*) + \varepsilon \right)^+ \right)^2 + d^2_\mathrm{S} \left( g \left( q_a,q(b^*,u,\bar q_a) \right) \right) \right)^{1/2} .
$$
Since $g$ and $d^2_{\mathrm{S}}$ are continuous and so is $F_R(u^*,\bar{q}^*_a)$ (see Lemma~\ref{prop30-1-1}), it follows that $J^R_\varepsilon$ is continuous on $(\mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a),\Vert \cdot \Vert_{\mathcal{U\bar{Q}}^{b^*}})$. Moreover, one has $J^R_\varepsilon (u^*,\bar{q}^*_a) = \varepsilon$ and, from optimality of $q^{0*}(b^*)$, $J^R_\varepsilon (u,\bar{q}_a) > 0$ for every $(u,\bar{q}_a) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$.
It follows from the Ekeland variational principle that, for every $\varepsilon > 0$, there exists $(u^R_\varepsilon,\bar{q}^R_{a,\varepsilon}) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$ such that $\Vert (u^R_\varepsilon,\bar{q}^R_{a,\varepsilon})-(u^*,\bar{q}^*_a) \Vert_{\mathcal{U\bar{Q}}^{b^*}} \leq \sqrt{\varepsilon}$ and
\begin{equation}\label{eqconsequenceekeland}
-\sqrt{\varepsilon} \, \Vert (u,\bar{q}_a) - (u^R_\varepsilon,\bar{q}^R_{a,\varepsilon}) \Vert_{\mathcal{U\bar{Q}}^{b^*}} \leq J^R_\varepsilon (u,\bar{q}_a) - J^R_\varepsilon (u^R_\varepsilon,\bar{q}^R_{a,\varepsilon}),
\end{equation}
for every $(u,\bar{q}_a) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$.
In particular, $u^R_\varepsilon$ converges to $u^*$ in $\mathrm{L}^1_{\mathbb{T}_1}([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)$ and $\bar{q}^R_{a,\varepsilon}$ converges to $\bar{q}^*_a$ as $\varepsilon$ tends to $0$. Besides, setting
\begin{equation}\label{defp0R}
\psi^{0R}_\varepsilon = \frac{-1}{J^R_\varepsilon (u^R_\varepsilon,\bar{q}^R_{a,\varepsilon})} \big(q^0 (b^*,u^R_\varepsilon,\bar{q}^R_{a,\varepsilon}) - q^{0*}(b^*) + \varepsilon \big)^+ \leq 0
\end{equation}
and
\begin{equation}\label{defpsiepsR}
\psi^R_\varepsilon = \frac{-1}{J^R_\varepsilon (u^R_\varepsilon,\bar{q}^R_{a,\varepsilon})} \left( g (q^R_{a,\varepsilon},q(b^*,u^R_\varepsilon,\bar{q}^R_{a,\varepsilon})) - \mathrm{P}_{\mathrm{S}} \big( g (q^R_{a,\varepsilon},q(b^*,u^R_\varepsilon,\bar q^R_{a,\varepsilon})) \big) \right) \in \mathbb{R}^{j} ,
\end{equation}
note that $\vert \psi^{0R}_\varepsilon \vert^2 + \Vert \psi^R_\varepsilon \Vert_{\mathbb{R}^j}^2 = 1 $ and $-\psi^R_\varepsilon \in \mathcal{O}_\mathrm{S} [\mathrm{P}_{\mathrm{S}} (g (q^R_{a,\varepsilon},q(b^*,u^R_\varepsilon,\bar{q}^R_{a,\varepsilon})))]$.
Using a compactness argument, the continuity of $F_R(u^*,\bar{q}^*_a)$ (see Lemma~\ref{prop30-1-1}), the $\mathrm{C}^1$-regularity of $g$ and the (partial) converse of the Lebesgue dominated convergence theorem, we infer that there exists a sequence $(\varepsilon_k)_{k \in \mathbb{N}}$ of positive real numbers converging to $0$ such that $u^R_{\varepsilon_k}$ converges to $u^*$ $\Delta_1$-a.e. on $[a,b^*)_{\mathbb{T}_1}$,
$\bar{q}^R_{a,\varepsilon_k}$ converges to $\bar{q}^*_a$,
$g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}))$ converges to $g(q^*_a,q^*(b^*)) \in \mathrm{S}$,
$\mathrm{d}g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}))$ converges to $\mathrm{d}g(q^*_a,q^*(b^*))$,
$\psi^{0R}_{\varepsilon_k}$ converges to some $\psi^{0R} \leq 0$,
and $\psi^R_{\varepsilon_k}$ converges to some $\psi^R \in \mathbb{R}^j$ as $k$ tends to $+\infty$,
with $\vert \psi^{0R} \vert^2 + \Vert \psi^R \Vert_{\mathbb{R}^j}^2 = 1 $ and $-\psi^R \in \mathcal{O}_{\mathrm{S}} [g (q^*_a,q^*(b^*))]$ (see Lemma~\ref{lemconvex}).
In the next lemmas, we use the inequality~\eqref{eqconsequenceekeland} respectively with needle-like variations of $u^R_{\varepsilon_k}$ at right-scattered points of ${\mathbb{T}_1}$ and at right-dense points of ${\mathbb{T}_1}$, and then variations of $\bar{q}^R_{a,\varepsilon_k}$. Hence, we infer some important inequalities by taking the limit in $k$.
Note that these variations were defined in Section~\ref{sectionvariations} for any dynamics $f$, and that we apply them here to the augmented system~\eqref{DD-CS_augm}, associated with the augmented dynamics $\bar f$.
\begin{lemma
For every $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ and every $y \in \mathrm{D}^\Omega_\mathrm{stab}(u^*(r))$, considering the needle-like variation $\Pi = (r,y)$ as defined in Section~\ref{section31RS}, one has
\begin{equation}\label{inegfondamscattered}
\psi^{0R} w^{0}_\Pi (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi^R , w_\Pi (b^*,u^*,\bar{q}^*_{a}) \right\rangle_{\mathbb{R}^n} \leq 0 ,
\end{equation}
where the variation vector $\bar w_\Pi=(w_\Pi,w^0_\Pi)^\top$ is defined by~\eqref{varvect_scatteredw} (replacing $f$ with $\bar f$).
\end{lemma}
\begin{proof}
Since $u^R_{\varepsilon_k}$ converges to $u^*$ $\Delta_1$-a.e. on $[a,b^*)_{\mathbb{T}_1}$, it follows that $u^R_{\varepsilon_k} (r)$ converges to $u^*(r)$ as $k$ tends to $+\infty$, where $\Vert u^*(r) \Vert_{\mathbb{R}^m} < R$. It follows that $ \Vert u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m} < R$ and $0$ is not isolated in $\Xi = \{ \alpha \in [0,1], \; u^R_{\varepsilon_k} (r)+\alpha (y-u^R_{\varepsilon_k} (r)) \in \Omega \}$ for any sufficently large $k$, see Definition~\ref{defstable}. Fixing such a large $k$, one has $u^R_{\varepsilon_k,\Pi} (\cdot,\alpha) \in \mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\Omega)$ and
\begin{equation*}\begin{split}
\Vert u^R_{\varepsilon_k,\Pi} (\cdot,\alpha) \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)}
& \leq \max (\Vert u^R_{\varepsilon_k} \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)}, \Vert u^R_{\varepsilon_k,\Pi} (r,\alpha) \Vert_{\mathbb{R}^m} ) \\
& \leq \max (R, \Vert u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m} + \alpha \Vert y - u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m} ) ,
\end{split}\end{equation*}
for every $\alpha \in \Xi$. Moreover, one has
\begin{equation*}\begin{split}
\Vert u^R_{\varepsilon_k,\Pi} (\cdot,\alpha) - u^* \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)}
& \leq \Vert u^R_{\varepsilon_k,\Pi} (\cdot,\alpha) - u^R_{\varepsilon_k} \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} + \Vert u^R_{\varepsilon_k}-u^* \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} \\
& \leq \alpha \mu_1 (r) \Vert y-u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m} + \sqrt{\varepsilon_k}.
\end{split}\end{equation*}
Therefore $(u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k}) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$ for every $\alpha \in \Xi$ sufficiently small and every $k$ sufficiently large.
It then follows from~\eqref{eqconsequenceekeland} that
\begin{equation*}
-\sqrt{\varepsilon_k} \, \Vert u^R_{\varepsilon_k,\Pi} (\cdot,\alpha) - u^R_{\varepsilon_k} \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} \leq J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k}) - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}),
\end{equation*}
and thus
\begin{equation*}
-\sqrt{\varepsilon_k} \, \mu_1 (r) \Vert y-u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m}
\leq \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k})^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\alpha ( J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k}) + J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) ) } .
\end{equation*}
Using Proposition~\ref{prop32-1}, we infer that
\begin{equation*}\begin{split}
& \lim_{\alpha \to 0} \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k})^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\alpha}
\\
& = 2 \big( q^0 (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) - q^{0*}(b^*) + \varepsilon_k \big)^+ w^{0}_\Pi (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \\
& \qquad + 2 \Big\langle g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) - \mathrm{P}_{\mathrm{S}} \big( g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \big), \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \frac{\partial g}{\partial q_2} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \, w_\Pi (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \Big\rangle_{\mathbb{R}^j} .
\end{split}\end{equation*}
Since $J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\Pi} (\cdot,\alpha),\bar{q}^R_{a,\varepsilon_k})$ converges to $J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})$ as $\alpha$ tends to $0$, using~\eqref{defp0R} and~\eqref{defpsiepsR} it follows that
\begin{multline*}
-\sqrt{\varepsilon_k} \, \mu_1 (r) \Vert y-u^R_{\varepsilon_k} (r) \Vert_{\mathbb{R}^m} \leq -\psi^{0R}_{\varepsilon_k} w^{0}_\Pi (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \\
- \left\langle \Big( \frac{\partial g}{\partial q_2} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \Big)^\top \psi^R_{\varepsilon_k} , w_\Pi (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \right\rangle_{\mathbb{R}^n} .
\end{multline*}
By letting $k$ tend to $+\infty$ and using Lemma~\ref{lem32-2w}, the lemma follows.
\end{proof}
We define the sets
\begin{equation*}
\begin{split}
A &= \left\lbrace t \in [a,b^*)_{{\mathbb{T}_1}} \mid u^{R}_{\varepsilon_k}(t) \,\text{does not converge to} \,u^*(t) \,\text{when} \,k \,\text{tends to} \,+\infty \right\rbrace ,\\
A_k &= \left\lbrace t \in [a,b^*)_{{\mathbb{T}_1}} \mid t \notin \mathscr{L}_{[a,b^*)_\mathbb{T}} (\bar f( \cdot, \bar q(\cdot,u^R_{\varepsilon_k},\bar q^R_{a,k}),u^{R\Phi}_{\varepsilon_k} )) \right\rbrace,\quad k \in \mathbb{N}.
\end{split}
\end{equation*}
\begin{lemma
We have $ \mu_{\Delta_1} \left( A \cup \bigcup_{k \in \mathbb{N}} A_k \right) = 0. $
\end{lemma}
\begin{proof}
Since $\mu_{\Delta_1} (A)=0$, it suffices to prove that $\mu_{\Delta_1} ( A_k ) = 0$ for every $k \in \mathbb{N}$. Let $k \in \mathbb{N}$. We set
$$ B_k = \left\lbrace t \in [a,b^*)_{\mathbb{T}} \mid t \notin \mathscr{L}_{[a,b^*)_\mathbb{T}} (\bar f(\cdot , \bar q(\cdot,u^R_{\varepsilon_k},\bar q^R_{a,k}),u^{R\Phi}_{\varepsilon_k})) \right\rbrace. $$
We know that $\mu_{\Delta} (B_k)=0$. Hence, $B_k \subset \mathrm{RD}$ and consequently, $\mu_\Delta (B_k) = \mu_L(B_k) = 0$. Since $A_k \subset B_k$, $\mu_L(A_k) = 0$. To conclude, it suffices to prove that $\mu_{\Delta_1} ( A_k ) = \mu_L (A_k)$. To see this, let us prove that $A_k \subset \mathrm{RD}_1$. By contradiction, let us assume that there exists $t \in A_k \cap \mathrm{RS}_1$. Since $t \in A_k \subset \mathrm{RD}$, we conclude that $u^{R\Phi}_{\varepsilon_k}$ is constant on $[t,\min(\sigma_1(t),b^*))_{\mathbb{T}}$, where $(t,\min(\sigma_1(t),b^*))_\mathbb{T} \neq \emptyset$. As a consequence, $\bar f( \cdot , \bar q(\cdot,u^R_{\varepsilon_k},\bar q^R_{a,k}),u^{R\Phi}_{\varepsilon_k})$ is continuous on $[t,\min(\sigma_1(t),b^*))_{\mathbb{T}}$ and consequently $t \in \mathscr{L}_{[a,b^*)_\mathbb{T}} (\bar f(\cdot, \bar q(\cdot,u^R_{\varepsilon_k},\bar q^R_{a,k}),u^{R\Phi}_{\varepsilon_k}))$. This leads to a contradiction.
\end{proof}
We define the set of Lebesgue times $t \in [a,b^*)_{{\mathbb{T}_1}}$ by
$$ \mathscr{L}^{R,1}_{[a,b^*)_\mathbb{T}} = [a,b^*)_{{\mathbb{T}_1}} \backslash \left( A \cup \bigcup_{k \in \mathbb{N}} A_k \right). $$
Note that $\mu_{\Delta_1} (\mathscr{L}^{R,1}_{[a,b^*)_\mathbb{T}}) = \mu_{\Delta_1} ([a,b^*)_{{\mathbb{T}_1}}) $ and then $ \mu_{\Delta_1} (\mathscr{L}^{R,1}_{[a,b^*)_\mathbb{T}} \cap \mathrm{RD}_1) = \mu_{\Delta_1} ([a,b^*)_{{\mathbb{T}_1}} \cap \mathrm{RD}_1)$.
\begin{lemma
For every $s \in \mathscr{L}^{R,1}_{[a,b^*)_\mathbb{T}} \cap \mathrm{RD}_1$ and for every $z \in \Omega \cap \overline{\mathrm{B}}_{\mathbb{R}^m}(0,R)$,
considering the needle-like variation $\amalg = (s,z)$ as defined in Section~\ref{section31RD}, one has
\begin{equation}\label{inegfondamdense}
\psi^{0R} w^{0}_\amalg (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi^R , w_\amalg (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n} \leq 0 ,
\end{equation}
where the variation vector $\bar w_\amalg=(w_\amalg,w^0_\amalg)^\top$ is defined by~\eqref{varvect_densew} (replacing $f$ with $\bar f$).
\end{lemma}
\begin{proof}
For every $k \in \mathbb{N}$ and any $\beta \in \mathcal{V}^{s,b^*}_1$, we recall that $u^R_{\varepsilon_k,\amalg} (\cdot,\beta) \in \mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\Omega) $ and
\begin{equation*}
\Vert u^R_{\varepsilon_k,\amalg} (\cdot,\beta) \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} \leq \max (\Vert u^R_{\varepsilon_k} \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} , \Vert z \Vert_{\mathbb{R}^m} ) \leq R ,
\end{equation*}
and
\begin{equation*}\begin{split}
\Vert u^R_{\varepsilon_k,\amalg} (\cdot,\beta) - u^* \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)}
& \leq \Vert u^R_{\varepsilon_k,\amalg} (\cdot,\beta) - u^R_{\varepsilon_k} \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} + \Vert u^R_{\varepsilon_k}-u^* \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} \\
& \leq 2R \beta + \sqrt{\varepsilon_k}.
\end{split}\end{equation*}
Therefore $(u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k}) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$ for every $\beta$ sufficiently small and every $k$ sufficiently large.
It then follows from~\eqref{eqconsequenceekeland} that
\begin{equation*}
-\sqrt{\varepsilon_k} \, \Vert u^R_{\varepsilon_k,\amalg} (\cdot,\beta) - u^R_{\varepsilon_k} \Vert_{\mathrm{L}^1_{\mathbb{T}_1} ([a,b^*)_{\mathbb{T}_1},\mathbb{R}^m)} \leq J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k}) - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) ,
\end{equation*}
and thus
\begin{equation*}
-2 R \sqrt{\varepsilon_k}
\leq \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k})^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\beta ( J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k}) + J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) ) } .
\end{equation*}
Using Proposition~\ref{prop33-1}, we infer that
\begin{equation*}\begin{split}
& \lim_{\beta \to 0} \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k})^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\beta} \\
& = 2 \big( q^0 (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) - q^{0*}(b^*) + \varepsilon_k \big)^+ w^{0}_\amalg (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \\
& \qquad + 2 \Big\langle g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) - \mathrm{P}_{\mathrm{S}} \big( g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \big), \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \frac{\partial g}{\partial q_2}(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \, w_\amalg (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \Big\rangle_{\mathbb{R}^j}.
\end{split}\end{equation*}
Since $J^R_{\varepsilon_k} (u^R_{\varepsilon_k,\amalg} (\cdot,\beta),\bar{q}^R_{a,\varepsilon_k})$ converges to $J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})$ as $\beta$ tends to $0$, using~\eqref{defp0R} and~\eqref{defpsiepsR} it follows that
\begin{equation*}
-2R \sqrt{\varepsilon_k} \leq -\psi^{0R}_{\varepsilon_k} w^{0}_\amalg (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})
- \left\langle \Big( \frac{\partial g}{\partial q_2} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \Big)^\top \psi^R_{\varepsilon_k} , w_\amalg (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \right\rangle_{\mathbb{R}^n}.
\end{equation*}
By letting $k$ tend to $+\infty$, and using Lemma~\ref{lem33-2}, the lemma follows.
\end{proof}
\begin{lemma
For every $\bar{q}_a \in \mathbb{R}^n \times \{ 0 \}$, considering the variation of initial point as defined in Section~\ref{section32}, one has
\begin{multline}\label{inegfondampoint}
\psi^{0R} w^0_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a)+ \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi^R , w_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n}
\\
\leq - \left\langle \Big( \frac{\partial g}{\partial q_1} (q^*_a,q^*(b^*)) \Big)^\top \psi^R , q_a \right\rangle_{\mathbb{R}^n} ,
\end{multline}
where the variation vector $\bar w_{\bar{q}_a}=(w_{\bar{q}_a},w^0_{\bar{q}_a})^\top$ is defined by~\eqref{varvect_pointinit} (replacing $f$ with $\bar f$).
\end{lemma}
\begin{proof}
For every $k \in \mathbb{N}$ and every $\gamma\geq 0$, one has
$$
\Vert \bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a - \bar{q}^*_a \Vert_{\mathbb{R}^{n+1}} \leq \gamma \Vert \bar{q}_a \Vert_{\mathbb{R}^{n+1}} + \Vert \bar{q}^R_{a,\varepsilon_k} - \bar{q}^*_a \Vert_{\mathbb{R}^{n+1}} \leq \gamma \Vert \bar{q}_a \Vert_{\mathbb{R}^{n+1}} + \sqrt{\varepsilon_k}.
$$
Therefore $(u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a) \in \mathrm{E}^{\Omega,0}_R (u^*,\bar{q}^*_a)$ for every $\gamma$ sufficiently small and every $k$ sufficiently large.
It then follows from~\eqref{eqconsequenceekeland} that
\begin{equation*}
-\sqrt{\varepsilon_k} \, \Vert \bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a - \bar{q}^R_{a,\varepsilon_k} \Vert_{\mathbb{R}^{n+1}} \leq J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a) - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}),
\end{equation*}
and thus
\begin{equation*}
-\sqrt{\varepsilon_k} \, \Vert \bar{q}_a \Vert_{\mathbb{R}^{n+1}}
\leq \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a)^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\gamma ( J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a) + J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) ) } .
\end{equation*}
Using Proposition~\ref{prop34-1}, we infer that
\begin{equation*}\begin{split}
& \lim_{\gamma \to 0} \frac{J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a)^2 - J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})^2}{\gamma} \\
& = 2 \big( q^0 (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) - q^{0*}(b^*) + \varepsilon_k \big)^+ w^0_{\bar{q}_a} (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \\
& \quad + 2 \Big\langle g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) - \mathrm{P}_{\mathrm{S}} \big( g(q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \big), \\
& \qquad\qquad \frac{\partial g}{\partial q_1} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \, q_a + \frac{\partial g}{\partial q_2} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \, w_{\bar{q}_a} (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \Big\rangle_{\mathbb{R}^j}.
\end{split}\end{equation*}
Since $J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}+\gamma \bar{q}_a)$ converges to $J^R_{\varepsilon_k} (u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})$ as $\gamma$ tends to $0$, using~\eqref{defp0R} and~\eqref{defpsiepsR} it follows that
\begin{multline*}
-\sqrt{\varepsilon_k} \, \Vert \bar{q}_a \Vert_{\mathbb{R}^{n+1}} \leq -\psi^{0R}_{\varepsilon_k} w^{0}_{\bar{q}_a} (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) - \left\langle \Big( \frac{\partial g}{\partial q_1} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \Big)^\top \psi^R_{\varepsilon_k} , q_a \right\rangle_{\mathbb{R}^n} \\
- \left\langle \Big( \frac{\partial g}{\partial q_2} (q^R_{a,\varepsilon_k},q(b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k})) \Big)^\top \psi^R_{\varepsilon_k} , w_{\bar{q}_a} (b^*,u^R_{\varepsilon_k},\bar{q}^R_{a,\varepsilon_k}) \right\rangle_{\mathbb{R}^n}.
\end{multline*}
By letting $k$ tend to $+\infty$, and using Lemma~\ref{lem34-2}, the lemma follows.
\end{proof}
At this step, we have obtained in the previous lemmas the three inequalities~\eqref{inegfondamscattered},~\eqref{inegfondamdense} and~\eqref{inegfondampoint}, valuable for any $R> \Vert u^* \Vert_{\mathrm{L}^\infty_{\mathbb{T}_1} ([a,b)_{\mathbb{T}_1},\mathbb{R}^m)}$. Recall that $\vert \psi^{0R} \vert^2 + \Vert \psi^R \Vert_{\mathbb{R}^j}^2 = 1 $ and that $-\psi^R \in \mathcal{O}_{\mathrm{S}} [g (q^*_a,q^*(b^*))]$.
Then, considering a sequence of real numbers $R_\ell$ converging to $+\infty$ as $\ell$ tends to $+\infty$, we infer that there exist $\psi^0 \leq 0$ and $\psi \in \mathbb{R}^j$ such that $\psi^{0R_\ell}$ converges to $\psi^0$ and $\psi^{R_\ell}$ converges to $\psi$ as $\ell$ tends to $+\infty$, and moreover $\vert \psi^0 \vert^2 + \Vert \psi \Vert_{\mathbb{R}^j}^2 = 1 $ and $-\psi \in \mathcal{O}_{\mathrm{S}} [g (q^*_a,q^*(b^*))] $.
We set $\mathscr{L}^{1}_{[a,b^*)_\mathbb{T}} = \bigcap_{\ell \in \mathbb{N}} \mathscr{L}^{R_\ell,1}_{[a,b^*)_\mathbb{T}}$. Note that $\mu_{\Delta_1} (\mathscr{L}^{1}_{[a,b^*)_\mathbb{T}}) = \mu_{\Delta_1} ([a,b^*)_{{\mathbb{T}_1}}) $ and consequently $ \mu_{\Delta_1} (\mathscr{L}^{1}_{[a,b^*)_\mathbb{T}} \cap \mathrm{RD}_1) = \mu_{\Delta_1} ([a,b^*)_{{\mathbb{T}_1}} \cap \mathrm{RD}_1)$. Taking the limit in $\ell$ in~\eqref{inegfondamscattered},~\eqref{inegfondamdense} and~\eqref{inegfondampoint}, we get the following lemma.
\begin{lemma}
For every $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$, and every $y \in \mathrm{D}^\Omega_\mathrm{stab}(u^*(r))$, one has
\begin{equation}\label{varineqscattered}
\psi^{0} w^{0}_\Pi (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi , w_\Pi (b^*,u^*,\bar{q}^*_{a}) \right\rangle_{\mathbb{R}^n} \leq 0 ,
\end{equation}
where the variation vector $\bar w_\Pi=(w_\Pi,w^0_\Pi)^\top$ associated with the needle-like variation $\Pi = (r,y)$ of $u$ is defined by~\eqref{varvect_scatteredw} (replacing $f$ with $\bar f$).
For every $s \in \mathscr{L}^{1}_{[a,b^*)_\mathbb{T}} \cap \mathrm{RD}_1$ and every $z \in \Omega$, one has
\begin{equation}\label{varineqdense}
\psi^{0} w^{0}_\amalg (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi , w_\amalg (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n} \leq 0 ,
\end{equation}
where the variation vector $\bar w_\amalg=(w_\amalg,w^0_\amalg)^\top$ associated with the needle-like variation $\amalg = (s,z)$ of $u$ is defined by~\eqref{varvect_densew} (replacing $f$ with $\bar f$);
For every $\bar{q}_a \in \mathbb{R}^n \times \{ 0 \}$, one has
\begin{multline}\label{varineqpointinit}
\psi^{0} w^0_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a)+ \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi , w_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n}
\\
\leq - \left\langle \Big( \frac{\partial g}{\partial q_1} (q^*_a,q^*(b^*)) \Big)^\top \psi , q_a \right\rangle_{\mathbb{R}^n} ,
\end{multline}
where the variation vector $\bar w_{\bar{q}_a}=(w_{\bar{q}_a},w^0_{\bar{q}_a})^\top$ associated with the variation $\bar{q}_a$ of the initial point $\bar{q}_a^*$ is defined by~\eqref{varvect_pointinit} (replacing $f$ with $\bar f$).
\end{lemma}
This result concludes the application of the Ekeland variational principle.
The last step of the proof consists of deriving the PMP from these inequalities.
\subsubsection{Proof of Remark~\ref{remarknonsubmersive}}\label{section46}
In this subsection, we prove the formulation of the PMP mentioned in Remark~\ref{remarknonsubmersive}. Note that we do not prove, at this step, the transversality condition on the final time.
We define $\bar{p} =(p ,p^0 )^\top$ as the unique solution on $[a,b^*]_\mathbb{T}$ of the backward shifted linear $\Delta$-Cauchy problem
\begin{equation}\label{equationpourpdanspreuve}
\bar{p}^\Delta(t) = - \left( \frac{\partial \bar{f}}{\partial \bar{q}} (t, \bar{q}^*(t),u^{*\Phi}(t)) \right)^\top \bar{p}^\sigma(t), \quad
\bar{p}(b^*) = \left( \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi,\psi^0 \right)^\top.
\end{equation}
The existence and uniqueness of $\bar{p}$ are ensured by \cite[Theorem 6]{bour10}. Since $\bar f$ does not depend on $q^0$, it is clear that $p^0$ is constant with $p^0=\psi^0 \leq 0$.
\paragraph{Right-dense points.} Let $s \in \mathscr{L}^1_{[a,b^*)_\mathbb{T}} \cap \mathrm{RD}_1$ and $z \in \Omega$. Since the function $\langle \bar{p} , \bar{w}_\amalg(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}} $ is absolutely continuous and satisfies $\langle \bar{p} , \bar{w}_\amalg(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}^\Delta = 0$ $\Delta$-a.e. on $[s,b^*)_\mathbb{T}$ from the Leibniz formula~\eqref{eqleibniz}, this function is constant on $[s,b^*]_\mathbb{T}$. It thus follows from~\eqref{varineqdense} that
\begin{equation*}
\begin{split}
\langle \bar{p}(s) , \bar{w}_\amalg(s,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}
& = \langle \bar{p}(b^*) , \bar{w}_\amalg(b^*,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}} \\
& = \psi^{0} w^{0}_\amalg (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi , w_\amalg (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n} \leq 0,
\end{split}
\end{equation*}
and since $\bar{w}_\amalg(s,u^*,\bar{q}^*_a) = \bar{f} (s,\bar{q}^*(s),z) - \bar{f} (s,\bar{q}^*(s),u^*(s))$, we finally get
$$
\langle \bar{p}(s) , \bar{f} (s,\bar{q}^*(s),z) \rangle_{\mathbb{R}^{n+1}} \leq \langle \bar{p}(s) , \bar{f} (s,\bar{q}^*(s),u^*(s)) \rangle_{\mathbb{R}^{n+1}}.
$$
Since this inequality holds for every $z \in \Omega$, we have obtained the maximization condition
$$ H(s,q^*(s),p(s),p^0,u^*(s)) = \max_{z \in \Omega} H(s,q^*(s),p(s),p^0,z) . $$
\paragraph{Right-scattered points.} Let $r \in [a,b^*)_{\mathbb{T}_1} \cap \mathrm{RS}_1$ and $y \in \mathrm{D}^\Omega_\mathrm{stab}(u^*(r))$. Since $\langle \bar{p} , \bar{w}_\Pi(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}$ is absolutely continuous and satisfies $\langle \bar{p} , \bar{w}_\Pi(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}^\Delta = 0 $ $\Delta$-a.e. on $[\sigma^*_1(r),b^*)_\mathbb{T}$ from the Leibniz formula~\eqref{eqleibniz}, this function is constant on $[\sigma^*_1(r),b^*]_\mathbb{T}$. It thus follows from~\eqref{varineqscattered} that
\begin{multline*}
\langle \bar{p}(\sigma^*_1(r)) , \bar{w}_\Pi(\sigma^*_1(r),u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}
= \langle \bar{p}(b^*) , \bar{w}_\Pi(b^*,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}} \\
= \psi^{0} w^{0}_\Pi (b^*,u^*,\bar{q}^*_a) + \left\langle \Big( \frac{\partial g}{\partial q_2} (q^*_a,q^*(b^*)) \Big)^\top \psi , w_\Pi (b^*,u^*,\bar{q}^*_{a}) \right\rangle_{\mathbb{R}^n} \leq 0 .
\end{multline*}
We recall that $\bar{w}_\Pi(\sigma^*_1(r),u^*,\bar{q}^*_a) = \bar{h}_\Pi(\sigma^*_1(r),u^*,\bar{q}^*_a)$ where the variation vector $\bar h_\Pi=(h_\Pi,h^0_\Pi)^\top$ associated with the needle-like variation $\Pi = (r,y)$ of $u$ is defined by~\eqref{varvect_scatteredh} (replacing $f$ by $\bar{f}$).
Since $\bar{h}_\Pi(r,u^*,\bar{q}^*_a) = 0$, it follows that
$$ \int_{[r,\sigma^*_1(r))_\mathbb{T}} \langle \bar{p} , \bar{h}_\Pi(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}^\Delta (\tau) \; \Delta \tau = \langle \bar{p}(\sigma^*_1(r)) , \bar{h}_\Pi(\sigma^*_1(r),u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}} \leq 0. $$
Using the Leibniz formula~\eqref{eqleibniz} in the integrand and using~\eqref{varvect_scatteredh} (replacing $f$ by $\bar{f}$) and \eqref{equationpourpdanspreuve}, we finally get
\begin{equation*}
\left\langle \int_{[r,\sigma^*_1(r))_\mathbb{T}} \frac{\partial H}{\partial u} (\tau,q^*(\tau),p^\sigma(\tau),p^0,u^*(r)) \; \Delta \tau \; , \; y-u^*(r) \right\rangle_{\mathbb{R}^m} \leq 0.
\end{equation*}
\paragraph{Transversality conditions.}
The transversality condition on the adjoint vector $p$ at the final time $b^*$ has been obtained by definition (note that $-\psi \in \mathcal{O}_\mathrm{S} [g(q^*_a,q^*(b^*))] $ as mentioned previously). Let us now establish the transversality condition on the adjoint vector $p$ at the initial time $a$.
Let $\bar{q}_a \in \mathbb{R}^n \times \{ 0 \}$. With the same arguments as before, we prove that the function $\langle \bar{p}, \bar{w}_{\bar{q}_a}(\cdot,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}$ is constant on $[a,b^*]_\mathbb{T}$. It thus follows from~\eqref{varineqpointinit} that
\begin{equation*}
\begin{split}
\langle \bar{p}(a) , \bar{w}_{\bar{q}_a}(a,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}}
& = \langle \bar{p}(b^*) , \bar{w}_{\bar{q}_a}(b^*,u^*,\bar{q}^*_a) \rangle_{\mathbb{R}^{n+1}} \\
& = \psi^{0} w^0_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a)+ \left\langle \Big( \frac{\partial g}{\partial q_2} (\bar{q}^*_a,\bar{q}^*(b^*)) \Big)^\top \psi , w_{\bar{q}_a} (b^*,u^*,\bar{q}^*_a) \right\rangle_{\mathbb{R}^n} \\
& \leq - \left\langle \Big( \frac{\partial g}{\partial q_1} (q^*_a,q^*(b^*)) \Big)^\top \psi , q_a \right\rangle_{\mathbb{R}^n} ,
\end{split}
\end{equation*}
and since $\bar{w}_{\bar{q}_a}(a,u^*,\bar{q}^*_a)=\bar{q}_a=(q_a,0)$, we finally get
\begin{equation*}
\left\langle p(a)+ \Big( \frac{\partial g}{\partial q_1} (q^*_a,q^*(b^*)) \Big)^\top \psi , q_a \right\rangle_{\mathbb{R}^n} \leq 0.
\end{equation*}
Since this inequality holds for every $\bar{q}_a \in \mathbb{R}^n \times \{ 0 \}$, the left-hand equality of~\eqref{transv_cond} follows.
\subsubsection{End of the proof}\label{section47}
In this subsection, we conclude the proof of Theorem~\ref{thmmain}. Note that we prove the fourth item of Theorem~\ref{thmmain} in Remark~\ref{remfindepreuvefreefinaltime} below. To conclude the proof of Theorem~\ref{thmmain}, we will use the result claimed in Remark~\ref{remarknonsubmersive}. Let us separate two cases.
Firstly, let us assume that $g$ is submersive at $(q^*_a,q^*(b^*))$. In that case, we have just to prove that the couple $(p,p^0)$ is not trivial. Let us assume that $p$ is trivial. Then $p(a)=p(b^*)=0$ and, from the transversality conditions on the adjoint vector, $\psi$ belongs to the kernels of $( \partial_{q_1} g (q^*_a,q^*(b^*)) )^\top $ and $( \partial_{q_2} g (q^*_a,q^*(b^*)) )^\top $. It follows that $\psi$ belongs to the orthogonal of the image of the differential of $g$ at $(q^*_a,q^*(b^*))$. Since $g$ is submersive at $(q^*_a,q^*(b^*))$, it implies that $\psi = 0$. Since the couple $(\psi,p^0)$ is not trivial, we conclude that $p^0 \neq 0$ and then $(p,p^0)$ is not trivial. This concludes the proof of Theorem~\ref{thmmain} in this first case.
Secondly, let us assume that $g$ is not necessarily submersive at $(q^*_a,q^*(b^*))$. In that case, one has to note that if $q^*$ is an optimal solution of ${\bf (OSDCP)}^{\T}_{\TU}$ associated with the function $g$ and with the closed convex $\mathrm{S}$, then $q^*$ is also an optimal solution of ${\bf (OSDCP)}^{\T}_{\TU}$ associated with the function $\tilde{g}$ defined by $\tilde{g}(q_1,q_2) = (q_1,q_2)$ (which is submersive at any point) and with the closed convex $\tilde{\mathrm{S}} = \{ q^*_a \} \times \{ q^*(b^*) \}$. Then, we get back to the above first case, but with a different function $g$ and a different closed convex $\mathrm{S}$ that would impact only the third item of Theorem~\ref{thmmain}.
\begin{remark}\label{remfindepreuvefreefinaltime}
Throughout this remark, we assume that all asumptions of the fourth item of Theorem~\ref{thmmain} are satisfied. As mentioned in Remark~\ref{remmultiscale}, Theorem~\ref{thmmain} can be easily extended (without the fourth item, for now) to the framework of dynamical systems on time scales with several sampled-data controls with different sets of controlling times. To prove the transversality condition on the final time, we use the \textit{multiscale version} (without the fourth item) of Theorem~\ref{thmmain}.
Let $\delta>0$ be such that $[b^* - \delta , b^* + \delta] \subset \mathbb{T}$ and such that $(f,f^0)$ is of class $\mathrm{C}^1$ on $[b^* - \delta , b^* + \delta] \times \mathbb{R}^n \times \mathbb{R}^m$. Obviously, the trajectory $q^*$, associated with $b^*$ and $u^*$, is an optimal solution of the optimal sampled-data control problem, with free final time $b \in (b^* - \delta, b^* + \delta)$, defined by
\begin{align}[left={\bf (O)}_{[b^* - \delta, b^* + \delta]} \qquad\empheqlbrace]
& \min \int_{b^* - \delta}^b f^0 (\tau,q (\tau), u^\Phi(\tau) ) \, d \tau , \notag \\
& \dot{q}(t) = f (t,q(t),u^\Phi(t)), \quad \text{for a.e. $t \in [b^* - \delta,b)$, } \notag \\
& u\in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\Omega) , \notag \\
& (q(b^* - \delta),g(q^*(a),q(b))) \in \{ q^*(b^* - \delta) \} \times \mathrm{S}. \notag
\end{align}
We set $\rho^* (s) = \delta (s-1) + b^*$, $x^*(s) = q^* \circ \rho^* (s)$ and $w^* (s) = \delta$, for every $s \in [0,1]$. With the change of variable $x = q \circ \rho$ and with $\dot{\rho} = w$, it is clear that the augmented trajectory $(\rho^*,x^*)$, associated with the augmented control $(w^*,u^*)$, is an optimal solution of the optimal sampled-data control problem, with fixed final time $s=1$, defined by
\begin{align}[left={\bf (O)}_{[0,1]} \qquad\empheqlbrace]
& \min \int_{0}^1 w(\tau) f^0 (\rho(\tau), x (\tau), u^{\Phi \circ \rho }(\tau) ) \, d \tau , \notag \\
& (\dot{\rho} (t),\dot{x}(t)) = (w(t),w(t) f ( \rho (t) , x(t) , u^{\Phi \circ \rho }(t)) ), \quad \text{for a.e. $t \in [0,1)$, } \notag \\
& (w,u) \in \mathrm{L}^\infty ( [0,1], [ \delta/2, + \infty ) ) \times \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\Omega) , \notag \\
& (\rho(0),x(0)) = (b^* - \delta,q^*(b^* - \delta)), \notag \\
& (\rho (1) , g(q^*(a),x(1)) ) \in [b^* - \delta , b^* + \delta ] \times \mathrm{S} . \notag
\end{align}
In this new optimal sampled-data control problem, the sampled-data controls $w$ and $u$ are defined on different time scales. Its Hamiltonian $\tilde{H}$ is
$ \tilde{H}(\rho,x,p_\rho,p_x,p^0,w,u) = p_\rho w + \langle p_x , w f(\rho,x,u) \rangle_{\mathbb{R}^n} + p^0 w f^0 (\rho,x,u).$
Applying the \textit{multiscale version} (without the fourth item) of Theorem~\ref{thmmain}, since $w^*$ takes its values in the interior of the constraint set $[\delta/2,+\infty)$, it follows in particular that
$$ \dfrac{\partial \tilde{H}}{\partial w} (\rho^* (s) , x^*(s) , p_{\rho} (s) , p_x (s), p^0 , w^*(s), u^{*,\Phi \circ \rho^* }(s) ) = 0, $$
for almost every $s \in [0,1)$. Moreover, since $\rho^* (1) = b^*$ belongs to the interior of $(b^* - \delta , b^* + \delta )$, we can select $(p_\rho,p_x)$ such that $p_\rho (1) = 0$ (usual transversality condition on the adjoint vector). It follows that
$$ \langle p_x (s) , f(\rho^*(s),x^*(s),u^{*,\Phi \circ \rho^* }(s) \rangle_{\mathbb{R}^n} + p^0 f^0 (\rho^*(s),x^*(s),u^{*,\Phi \circ \rho^* }(s)) = - p_\rho (s) ,$$
for almost every $s \in [0,1]$.
We have thus proved that the function $t \mapsto H(t,q^*(t),p(t),p^0,u^{*\Phi} (t) )$ coincides, almost everywhere on $[b^*-\delta,b^*)$, with a continuous function vanishing at $t=b^*$.
\end{remark}
\begin{remark}
Another method in order to prove the fourth item of Theorem~\ref{thmmain} consists of considering variations of the final time $b$ in a neighborhood of $b^*$ and to modify accordingly the functional of Section~\ref{section41} to which Ekeland's variational principle is applied.
\end{remark}
\subsection{Proof of Theorem~\ref{propexistence}}\label{proofexistence}
The following proof can be read independently of the rest of Section~\ref{annexe}. It is inspired from \cite[Chap. 6]{trel}.
We only treat the case where the final time $b \in \mathbb{T}$ is fixed. The proof can be easily adapted to the case of a free final time.
Let us consider a sequence $(q_k)_{k \in \mathbb{N}}$ of $\mathcal{M}$, associated with sampled-data controls $u_k \in \mathrm{L}^\infty_{{\mathbb{T}_1}}({\mathbb{T}_1},\Omega)$, minimizing the cost considered in ${\bf (OSDCP)}^{\T}_{\TU}$. It follows from the assumptions that the sequence $( f (\cdot,q_k,u^\Phi_k ),f^0 (\cdot,q_k,u^\Phi_k ) )_{k \in \mathbb{N}}$ is bounded in $\mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$. Hence a subsequence (that we do not relabel) converges in the weak topology of $\mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$ to a function $(F,F^0)$. Moreover, a subsequence of $( q_k (a) )_{k \in \mathbb{N}}$ (that we do not relabel) converges in $\mathbb{R}^n$ and we denote the limit by $x(a)$. We define the absolutely continuous functions $ x(t) = x(a) + \int_{[a,t)_\mathbb{T}} F(\tau) \Delta \tau$ and $x^0 (t) = \int_{[a,t)_\mathbb{T}} F^0(\tau) \Delta \tau $. In particular, note the pointwise convergence of $( q_k )_{k \in \mathbb{N}}$ to $x$ on $[a,b]_\mathbb{T}$. Since $g$ is continuous and since $\mathrm{S}$ is closed, we have $g(x(a),x(b)) \in \mathrm{S}$. Note that $x^0 (b)$ is equal to the infimum of admissible costs. To conclude the proof, we have to prove the existence of $u \in \mathrm{L}^\infty_{{\mathbb{T}_1}} ({\mathbb{T}_1},\Omega)$ such that $ x(t) = x(a) + \int_{[a,t)_\mathbb{T}} f(\tau,x(\tau),u^\Phi(\tau)) \Delta \tau$ and $x^0 (t) = \int_{[a,t)_\mathbb{T}} f^0(\tau,x(\tau),u^\Phi(\tau)) \Delta \tau $.
The sequence $ ( f (\cdot,x,u^\Phi_k ),f^0 (\cdot,x,u^\Phi_k ) )_{k \in \mathbb{N}}$ is bounded in $\mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$. Hence, a subsequence (that we do not relabel) converges in the weak topology of $\mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$ to some $(\tilde{F},\tilde{F}^0)$. It follows from the Lebesgue dominated convergence theorem and from the global Lipschitz continuity of $f$ in $q$ on $[a,b]_\mathbb{T} \times \overline{\mathrm{B}}_{\mathbb{R}^n}(0,M) \times \Omega$ that $ \int_{[a,b)_\mathbb{T}} \varphi(\tau) F(\tau) \Delta \tau = \int_{[a,b)_\mathbb{T}} \varphi(\tau) \tilde{F}(\tau) \Delta \tau $ for every $\varphi \in \mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R})$. Hence $F=\tilde{F}$ (and similarly $F^0=\tilde{F}^0$) $\Delta$-a.e. on $[a,b)_\mathbb{T}$.
We define the set $ \mathcal{W}$ of functions $\Theta \in \mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$ such that $\Theta (t) \in \mathcal{W}(t,x(t))$ for $\Delta$-a.e. $t \in [a,b)_\mathbb{T}$. It follows from the assumptions that $\mathcal{W}$ is a closed (convex) subset of $\mathrm{L}^2_{\mathbb{T}} ([a,b)_\mathbb{T},\mathbb{R}^{n+1})$ with its usual topology, and thus with its weak topology as well (see \cite{brez}). We infer that $(F,F^0)=(\tilde{F},\tilde{F}^0) \in \mathcal{W}$. Hence, for $\Delta$-a.e. $t \in [a,b)_\mathbb{T}$, there exists $v(t) \in \Omega$ such that $ (F(t),F^0(t)) = (f (t,x(t),v(t) ) , f^0 (t,x(t),v(t) ) ) $. Note that $v$ can be selected $\mu_\Delta$-measurable on $[a,b)_\mathbb{T}$.\footnote{Indeed, similarly to the proof of Proposition~\ref{propnumber1}, one has just to consider the extension function $\tilde{v}$ defined on $[a,b)$ by $\tilde{v} (t) = v(t)$ if $t \in \mathbb{T}$ and by $\tilde{v}(t) = v(r)$ if $t \in (r,\sigma(r))$, where $r \in \mathrm{RS}$. Then, the function $\tilde{v}$ can be selected $\mu_L$-measurable on $[a,b)$ from \cite[Lemma~3A p. 161]{lee}, which implies the $\mu_\Delta$-measurability of $v$ on $[a,b)_\mathbb{T}$ from~\cite{caba}.}
Since $\mathrm{RS}_1 \cap [a,b)_\mathbb{T}$ is at most countable, using a diagonal argument, there exists a subsequence of $(u_k)_{k \in \mathbb{N}}$ (that we do not relabel) such that $u_k (r)$ converges to some $u_r \in \Omega$ for every $r \in \mathrm{RS}_1 \cap [a,b)_\mathbb{T}$. It follows from the Lebesgue dominated convergence theorem that $ x(t) = x(a) + \int_{[a,t)_\mathbb{T}} f(\tau,x(\tau),u^\Phi(\tau)) \Delta \tau$ and $x^0 (t) = \int_{[a,t)_\mathbb{T}} f^0(\tau,x(\tau),u^\Phi(\tau)) \Delta \tau $, where $u$ is defined by $u(t)=u_r$ if $t=r \in \mathrm{RS}_1 \cap [a,b)_\mathbb{T}$ and $u(t)=v(t)$ if $t \in \mathrm{RD}_1 \cap [a,b)_\mathbb{T}$. It is easy to prove that $u \in \mathrm{L}^\infty_{\mathbb{T}_1} ([a,b)_{\mathbb{T}_1},\Omega)$ (the $\mu_{\Delta_1}$-measurability of $u$ on $[a,b)_{\mathbb{T}_1}$ is established as in the proof of Proposition~\ref{propnumber1}). This concludes the proof of Theorem~\ref{propexistence}.
\bigskip
\noindent{\bf Acknowledgment.}
The second author was partially supported by the Grant FA9550-14-1-0214 of the EOARD-AFOSR.
\bibliographystyle{plain}
|
\section{Introduction}
The wetting of a substrate by a fluid is an important physical process and understanding such behaviour is crucial in a variety of fields from industrial processes such as lubrication and painting to biological applications such as tear films in the eyes or mucus linings in the lungs. The wetting behaviour of a fluid\cite{degennes85,hansen13} is determined by the manner in which the atoms or molecules within the fluid interact with each other and with those forming the substrate. Determining the macroscopic fluid properties, wetting behaviour and thermodynamics, starting from an understanding of the (microscopic) molecular interactions is one of the cornerstone problems in liquid state science.\cite{hansen13}
On the macroscopic scale, the wetting behaviour of a fluid in contact with a solid substrate can be characterised by the contact angle that a liquid drop makes with that substrate. Three regimes of wetting behaviour can be identified: complete wetting, partial wetting and non wetting, these three states are defined by contact angles of $\theta=0^\circ$, $0^\circ < \theta < 180^\circ$ and $\theta = 180^\circ$ respectively.\cite{degennes85} Surface tension forces arise from interfaces in the fluid and these can be related to the contact angle by Young's equation\cite{hansen13}
\begin{equation}\label{eq:youngs}
\cos\theta=\frac{\gamma_{wg} - \gamma_{wl}}{\gamma_{lg}},
\end{equation}
where $\gamma_{lg}$, $\gamma_{wl}$ and $\gamma_{wg}$ are the liquid-gas, wall-liquid and wall-gas surface tensions respectively.
The effective {interface} Hamiltonian (IH) model,\cite{dietrich88, schick90, macdowell13b, macdowell14} also referred to as the {interface} free energy model, describes the height profile of a mesoscopic liquid film on a substrate. One can find the equilibrium shape of a droplet by minimising the free energy functional
\begin{equation}\label{eq:fullcurveenergy}
F[h] = \int \left[ g(h) + \gamma_{lg} \sqrt{1+(\nabla h)^2} \right] \text{d}\ul{x},
\end{equation}
where $h({\bf x})$ is the liquid film thickness at some point $\ul{x}$ on the substrate and $g(h)$ is the binding {potential, which is also referred to as the effective interface potential.\cite{dietrich88, schick90, macdowell02, macdowell13b, macdowell14, henderson05} $g(h)$ is a restricted free energy, i.e.~the free energy subject to the constraint that the thickness of the liquid layer adsorbed on the surfaces is $h$. A good discussion on the subject of restricted free energies can be found in Ref.~\onlinecite{Doi_book}.} The binding potential describes the interaction between two interfaces and is related to the disjoining pressure $\Pi = -\partial g/ \partial h$. In the IH model, the binding potential is often approximated by simple expressions that give the qualitatively correct behaviour. A common example of such an approximation would be an asymptotic expansion which is valid only for larger film thicknesses (cf.\ Sec.\ \ref{sec:binding}). This paper sets out a method to directly calculate the binding potential from a microscopic basis, namely via density functional theory (DFT).\cite{evans79,evans92,hansen13,Lutsko10,Tarazona08,evans10,lowen10} As input, the method takes the interactions between particles in the fluid and also the forces on the fluid particles due to any external fields, for example that due to the wall of a container or a surface on which the liquid is deposited. This calculation yields an expression for the binding potential that is valid for all film thicknesses. The term ``particle'' is used here generically to refer to the atoms/molecules/colloids in the fluid, depending on the exact system under study. Note that it is also possible to calculate $g(h)$ from computer simulations -- see Refs.\,\onlinecite{macdowell05, macdowell06, herring10, macdowell11, rane11, gregorio12, tretyakov13, benet14}.
The IH model is particularly useful due to its application in dynamical studies. By making the assumption of small surface gradients and contact angles, Eq.\,\eqref{eq:fullcurveenergy} reduces to
\begin{equation}\label{eq:thinEnergy}
F[h] = \int \left[g(h) + \frac{\gamma_{lg}}{2} (\nabla h)^2 \right] \text{d}\ul{x}
\end{equation}
where we have omitted a constant contribution. Equation~\eqref{eq:thinEnergy} can then be employed in the thin film (or long-wave) evolution equation,\cite{oron97,thiele07} which describes the time evolution of a thin film of liquid on a flat solid substrate. In gradient dynamics form it is written\cite{mitlin93,thiele10}
\begin{equation}\label{eq:thinFilm}
\frac{\partial h}{\partial t} = \nabla \cdot \left[ Q(h) \nabla \frac{\delta F[h]}{\delta h} \right],
\end{equation}
where $Q(h)$ is a mobility factor that depends on the film thickness $h(\ul{x},t)$. Equation~\eqref{eq:thinFilm} may be derived by making a long wave approximation in the governing Navier-Stokes equation.\cite{safran94, thiele07} There are many applications of this equation to model different situations. Steady state solutions, where $\partial h/\partial t=0$, such as drop profiles, are found by minimising the free energy, Eq.\,\eqref{eq:thinEnergy}, with respect to the film height profile, subject to a volume constraint. More specifically, it amounts to solving
\begin{equation}
\frac{\delta F}{\delta h} = \alpha,
\end{equation}
where $\alpha$ is a Lagrange multiplier stemming from the constraint
\begin{equation}
\int \text{d} \ul{x} \ h(\ul{x}) = V_0,
\end{equation}
where $V_0$ is a specified drop volume. A typical drop profile resulting from such a calculation can be seen in Fig.\,\ref{fig:profs}(a). Note the very thin non-zero height `precursor' film that is present to the left and right of the drop.
\begin{figure}
\vspace{-2em}
\includegraphics[width=1.05\columnwidth]{intro_height}
\vspace{-3em}
\includegraphics[width=1.05\columnwidth]{intro_density}
\vspace{-2em}
\caption{Two different descriptions of a liquid drop on a surface: (a) a height profile calculated via Eq.\ \eqref{eq:filmHeight} from a mesoscopic free energy (cf.\ Eq~\eqref{eq:fullcurveenergy}) and (b) a density profile, which gives the fluid number density at a distance $z$ above the surface. These are both for a fluid with $\beta \epsilon = 0.9$ and $\beta \epsilon_w = 0.6$ (see Sec.\ \ref{sec:lattice} for further details).}
\label{fig:profs}
\end{figure}
A fully microscopic description has statistical mechanics as its basis. Statistical mechanics calculates an average over an ensemble of all possible states of the system, i.e.\ it averages over all possible configurations of the particles. This average leads to determining the fluid one body density profile $\rho(\ul{r})$, which represents the likelihood of finding a particle at a given point $\ul{r}$ in the system.\cite{hansen13} This statistical mechanical point of view is the basis for DFT. From DFT, the grand potential, $\Omega$, of the system is calculated and the equilibrium density profile, $\rho(\ul{r})$, which minimises $\Omega$, can be found. A typical equilibrium density profile of a liquid drop on a solid substrate is displayed in Fig.\,\ref{fig:profs}(b). Other examples of drop profiles calculated using DFT can be found in Refs.\,\onlinecite{berim08, ruckenstein10, pereira11, nold11, nold14}. Note that by identifying the surface of the liquid as the surface where the density equals a specified value, $\rho_{int}$, where $\rho_l>\rho_{int}>\rho_g$ and where $\rho_l$ and $\rho_g$ are the coexisting liquid and gas densities of the fluid, the description can be further reduced to obtain a film height profile very similar to that displayed in Fig.\,\ref{fig:profs}(a). One possible choice is to choose $\rho_{int}=(\rho_l+\rho_g)/2$. Alternatively, by integrating in the $z$ direction over such a density profile we can obtain a drop height profile. Here, we define the height of the liquid film on a substrate as the adsorption divided by the liquid-gas bulk density difference
\begin{equation}\label{eq:filmHeight}
h(x,y)=\frac{\int_0^\infty \left[\rho(x,y,z)-\rho_g \right] dz}{\rho_l - \rho_g}.
\end{equation}
It is worth noting that using DFT it is just as easy to compute the profile for a liquid droplet that makes a contact angle with the substrate that is greater than $90^\circ$, than one with $\theta < 90^\circ$. It is significantly more difficult to find drop profiles for $\theta > 90^\circ$ in the mesoscopic approach because for these contact angles, Eqs.\ \eqref{eq:fullcurveenergy} and \eqref{eq:thinEnergy} cannot be employed.
The remainder of this paper is set out as follows: in Sec.~\ref{sec:binding}, the binding potential and the procedure for calculating it are discussed. A simple DFT, for the lattice-gas model, is presented in Sec.\ \ref{sec:lattice} that is used to demonstrate the procedure. The dependence of the fluid behaviour on the particle interactions is discussed in Sec.~\ref{sec:ranges}. In particular, it is shown that truncating the range of the dispersion interactions between the particles has a profound effect on the binding potential and interfacial phase behaviour. The method of fitting a function to the calculated data is given in Sec.~\ref{sec:fitting}, followed by the results of passing the binding potential from the lattice-gas model to the thin film IH model in Sec.~\ref{sec:results}. Finally, conclusions are drawn in Sec.~\ref{sec:conclusion}.
\section{The Binding Potential}\label{sec:binding}
For any fluid, coexistence between the liquid and gas phases occurs when the temperature $T$, chemical potential $\mu$ and pressure $p$ of the two phases are equal. It then follows, for a given volume $V$, that the bulk grand free energy, $\Omega=-pV$, of either phase occupying the same volume is also equal. For a system where the liquid and gas phases of a fluid exist together at the point of liquid-gas coexistence, any excess, over bulk, contributions to the free energy of the system must stem solely from the interface that forms between the two phases. This excess grand potential per unit area of the interface defines the surface tension between those two phases. In this case, it is the liquid-gas surface tension $\gamma_{lg}$. Now consider a system with chemical potential $\mu=\mu_{coex}$, the value at coexistence, where a film of liquid separates the bulk gas from a solid surface (cf.\ Fig.\,\ref{fig:system}). The excess free energy now consists of the sum of the two interfacial tensions and also the interaction between the two interfaces. The excess grand potential in such a system is given by $\Omega_{ex}(h)\equiv\Omega+pV=\omega_{ex}(h)A$, where $A$ is the area of the interface and
\begin{equation}\label{eq:freeEnergy}
\omega_{ex}(h)=\gamma_{wl}+\gamma_{lg}+g(h).
\end{equation}
$h$ is the liquid film thickness, $\gamma_{wl}$ and $\gamma_{lg}$ are the wall-liquid and liquid-gas surface tensions respectively.\cite{marchand11} The final term $g(h)$ is the binding potential, which gives the contribution to the free energy from the interaction between the two interfaces. This has the property that as $h \to \infty$, $g(h) \to 0$. The absolute minimum of the grand potential defines the equilibrium film thickness $h$.
In the case of liquid droplets surrounded by their vapour on a solid substrate, the absolute minimum of the binding potential is directly related to the equilibrium contact angle of the drop,\cite{churaev95a,rauscher08}
\begin{equation}\label{eq:bindCont}
\theta=\cos^{-1}\left(1+\frac{g(h_0)}{\gamma_{lg}}\right),
\end{equation}
where $g(h_0)$ is the value at the minimum of the binding potential. This corresponds directly to Young's equation, Eq.\,\eqref{eq:youngs}; the absolute minimum of the binding potential corresponds to the equilibrium state of the system and gives the equilibrium excess grand potential, i.e.\ the wall-gas surface tension:
\begin{equation}
\omega(h_0) = \gamma_{wl}+ \gamma_{lg} + g(h_0) = \gamma_{wg}.
\end{equation}
Replacing $\gamma_{wg}$ in Eq.\,\eqref{eq:youngs} with this expression leads to Eq.\,\eqref{eq:bindCont}.
Equation~\eqref{eq:freeEnergy} is given above as a function of the film height $h$. From a microscopic viewpoint it is often more convenient to use the adsorption of the fluid as the order parameter characterising the fluid at the interface, instead of the film thickness. The total adsorption is readily calculated from the fluid density profile as
\begin{equation}\label{eq:adsorption}
\Gamma=\frac{1}{A} \int \text{d}\ul{r} (\rho(\ul{r}) - \rho_b).
\end{equation}
$\rho_b$ is the bulk fluid density, which in the cases considered here is the gas density $\rho_g$. We may also define a local adsorption [c.f.~Eq.~\eqref{eq:filmHeight}] as follows:
\begin{equation}\label{eq:filmAds}
\Gamma(x,y)=\int_0^\infty \left[\rho(x,y,z)-\rho_g \right] dz,
\end{equation}
so that when the definition in Eq.~\eqref{eq:filmHeight} for the film height $h$ is used, we have $\Gamma = h(\rho_l- \rho_g)$. For other definitions of the film height $h$, then
\begin{equation}\label{eq:adsorption_approx}
\Gamma \approx h(\rho_l- \rho_g).
\end{equation}
The dependence on how precisely $h$ is defined becomes negligible, for large $h$. Thus, Eq.~\eqref{eq:freeEnergy} and also the binding potential $g$ may be given as a function of the adsorption. Note, however, that negative adsorptions are possible, e.g.\ at a purely repulsive wall, and in such a situation describing the liquid film at the wall via a film thickness $h$ becomes meaningless since the quantity $h$ defined in Eq.~(\ref{eq:filmHeight}) then becomes negative.
For a system with chemical potential $\mu=\mu_{coex}+\delta\mu$, i.e.\ off-coexistence, then there is an additional contribution $\Gamma\delta\mu$, that must be added to the right hand side of Eq.~\eqref{eq:freeEnergy}.\cite{evans89, schick90} Together with Eq.\ \eqref{eq:adsorption_approx}, this gives
\begin{equation}\label{eq:freeEnergy_with_dmu}
\omega_{ex}(h)\approx\gamma_{wl}+\gamma_{lg}+(\rho_l-\rho_g)h\delta\mu+g(h).
\end{equation}
\begin{figure}
\begin{center}
\includegraphics{schematic}
\caption{A schematic of the system: a film of liquid of thickness $h$ separating a semi-infinite volume of gas from a solid surface.}
\label{fig:system}
\end{center}
\end{figure}
Often, only asymptotic forms of binding potentials are used which are strictly valid only in the limit of a large film thickness.\cite{schick90} There are two main asymptotic forms of the binding potential that are considered, the choice of which depends on the assumed particle interactions and the range of those interactions. For van der Waals' (dispersion) interactions, the following asymptotic form is appropriate:\cite{dietrich88,ibagon13,ibagon14,stewart05b}
\begin{equation}\label{eq:longAsym}
g(h)=\frac{a}{h^2} + \frac{b}{h^3} + \cdots,
\end{equation}
The equivalent disjoining pressure is\cite{thiele07,oron97}
\begin{equation}
\Pi(h)= \frac{2a}{h^3}+ \frac{3b}{h^4}+ \cdots.
\end{equation}
Under certain approximations the asymptotic behaviour shown in Eq.\,\eqref{eq:longAsym} can be calculated analytically including the values of the coefficients $a$ and $b$, and how they depend on the temperature.
With only short ranged interactions between particles, the binding potential can be expressed asymptotically as\cite{fernandez11,henderson05,brezin83,macdowell02,dietrich88}
\begin{equation}\label{eq:shortAsym}
g(h)=a \exp(-h/ \xi) + b \exp(-2h/\xi) + \cdots.
\end{equation}
The length $\xi$ is the bulk correlation length in the liquid phase at the interface. These asymptotic forms, truncated after a few terms, are used frequently throughout the literature (independently or as combinations of the two forms\cite{degennes85, thiele07, evans09, frastia12}) even though they are only strictly valid for thick liquid films and can not describe the binding potential as $h\to 0$. To describe the small $h$ behaviour, the full form of $g(h)$ is required. A fully microscopic theory, that describes the fluid structure at the wall, is required to obtain this.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{binding}
\caption{A series of binding potentials for a fluid with fixed inverse temperature $\beta \epsilon=0.9$ against a solid substrate with varying attraction strength $\beta \epsilon_w$. A change in wetting behaviour, from non-wetting to wetting occurs, as indicated by the change in the position of the minimum in $g(\Gamma)$ from a low finite adsorption to a large $(\Gamma \to \infty)$ adsorption, as $\beta \epsilon_w$ increases.}
\label{fig:bindingPot}
\end{center}
\end{figure}
Fig.~\ref{fig:bindingPot} shows binding potentials for various values of a parameter $\epsilon_w$, which determines whether or not the fluid wets the wall. These binding potentials are calculated using the microscopic DFT-based approach that is introduced below (cf. Sec.\ \ref{sec:lattice}). The parameter $\epsilon_w$ characterises the interaction strength between the wall and the fluid particles. The full form of the long range (algebraic) dispersion interactions between the particles are included and so for large $\Gamma$, $g(\Gamma)$ has the asymptotic form given in Eq.\,\eqref{eq:longAsym}. In Fig.~\ref{fig:bindingPot} we see that for small values of $\epsilon_w$ (weakly attracting, solvophobic wall) the global minimum in $g$ is at a small value of $\Gamma$, i.e.\ the liquid does not wet the wall. As $\epsilon_w$ is increased, there is a first order wetting transition when $\beta\epsilon_w\approx0.74$ and for $\beta\epsilon_w>0.74$ the global minimum in $g(\Gamma)$ is at $\Gamma\to\infty$, i.e.\ there is a macroscopically thick film of liquid on the wall at coexistence $\mu=\mu_{coex}$.
DFT gives a route by which the free energy may be calculated, taking into account the microscopic structure of the fluid at the wall and the interactions of the particles within it. The equilibrium state of the system is found by minimising the grand potential functional:\cite{evans79,evans92,hansen13,Lutsko10,Tarazona08,evans10,lowen10}
\begin{equation}
\Omega[\rho({\bf r})] = {\cal F}[\rho({\bf r})] + \int \text{d}{\bf r} V_{ext}({\bf r})\rho({\bf r}) - \mu \int \text{d}{\bf r} \rho({\bf r}),
\end{equation}
where ${\cal F}$ is the intrinsic Helmholtz free energy functional, $\mu$ is the chemical potential and $V_{ext}$ is the external potential. The equilibrium density profile, $\rho(\ul{r})$, is that which satisfies the Euler-Lagrange equation
\begin{equation}
\frac{\delta \Omega[\rho]}{\delta \rho({\bf r})}=0.
\label{eq:EL_eq}
\end{equation}
Once an approximation for ${\cal F}$ is specified, this equation is solved numerically via a scheme in which an initial density profile is supplied and then iterated until a specified convergence criterion is reached.\cite{hughes13,roth10}
In order to calculate the binding potential as a function of the adsorption using DFT, it is required to evaluate the free energy for any specified $\Gamma$, i.e.\ in addition to the equilibrium profile we require other non-equilibrium profiles for a range of values of the adsorption $\Gamma$. By using the procedure developed and justified in Ref.~\onlinecite{archer11} (see also Ref.~\onlinecite{archer13}), the excess density $(\rho({\bf r}) - \rho_b)$ is normalised at each iterative step, which is equivalent to including a (fictitious) additional effective external potential that stabilises a wetting film of the desired thickness with the specified value of $\Gamma$. Using this normalisation procedure, the excess grand potential can be obtained for a range of values of $\Gamma$. The binding potential is then given via Eq.~\eqref{eq:freeEnergy} by subtracting the values of the solid-liquid and liquid-gas surface tensions.
\section{A Microscopic Model}\label{sec:lattice}
The method outlined above for calculating the binding potential is valid for any DFT model. To illustrate the procedure, a simple approximate DFT for a lattice-gas, is used. The model is only briefly described here: a full description and derivation can be found in Ref.~\onlinecite{hughes13}. The system is discretised by a cubic lattice and the fluid particles, assumed to all be identical and spherical, occupy only one cell each on the lattice. The diameter of each particle, and the width of each cell, is $\sigma$ and there are $M=M_i M_j M_k$ lattice cells and $N<M$ fluid particles in the system. A point in the lattice is denoted ${\bf i}=(i,j,k)$ and $M_i$, $M_j$, and $M_k$ are the number of cells in the $i$, $j$ and $k$ directions, respectively. Note $i$, $j$ and $k$ are integers. Any configuration of particles in the system can then be described by the set of occupation numbers, $\{ n_{\bf i}\}$, where $n_{\bf i}=1$ if there is a particle in cell ${\bf i}$ and $n_{\bf i}=0$ otherwise. The average of an ensemble of all such systems may be taken and then the average occupation number of each cell is described by the density $\rho_{\bf i}=\langle n_{\bf i} \rangle$. It may be assumed that the equilibrium density profile only varies in one direction, or, as for the calculations below, it is assumed that the density profile is invariant in the third dimension, so only a two-dimensional (2D) slice of the system needs to be studied. Here, the $z$-axis direction, indexed by $k$, is assumed to be invariant. The energy of the full three-dimensional system is given by the Hamiltonian
\begin{equation}\label{eq:hamiltonian}
E= - \frac12 \sum_{{\bf i}=1}^M \sum_{{\bf j} \neq {\bf i}} \tilde{\epsilon}_{{\bf i},{\bf j}} n_{\bf i} n_{\bf j} + \sum_{{\bf i}=1}^M \tilde{V}_{\bf i} n_{\bf i},
\end{equation}
where $\tilde{V}_{\bf i}$ is the external potential and $\tilde{\epsilon}_{{\bf i},{\bf j}}$ is a Lennard-Jones-like pair potential between particles at lattice sites ${\bf i}$ and ${\bf j}$:
\begin{equation}
\tilde{\epsilon}_{{\bf i},{\bf j}} = v(r_{{\bf i},{\bf j}}) =
\begin{cases}
- \epsilon / {r_{{\bf i},{\bf j}}}^6 & \text{ for } r_{{\bf i},{\bf j}} \geq \sigma, \\
\infty & \text{ for } r_{{\bf i},{\bf j}} < \sigma,
\end{cases}\label{eq:pair_pot}
\end{equation}
where $r_{{\bf i},{\bf j}}$ is the distance between a pair of particles located at lattice sites ${\bf i}$ and ${\bf j}$. If one assumes that the system is invariant along the direction indexed by $k$, then a mean-field approximation for the internal energy $U$ is the following 2D sum:
\begin{equation}
U = \langle E\rangle = -\frac12 \sum_{{\bf i}_0=1}^{M_{2d}} \sum_{{\bf j}_0 \neq {\bf i}_0} \epsilon_{{\bf i}_0,{\bf j}_0} \rho_{{\bf i}_0} \rho_{{\bf j}_0} + \sum_{{\bf i}_0=1}^{M_{2d}} V_{{\bf i}_0} \rho_{{\bf i}_0},
\end{equation}
where $\sum_{{\bf i}_0=1}^{M_{2d}}$ sums over the $M_{2d}=M_iM_j$ sites in the 2D lattice plane where $k=0$ and ${\bf i}_0=(i,j,0)$. The third (invariant) dimension is now accounted for in the interaction weights $\epsilon_{{\bf i},{\bf j}}$ and $V_{\bf i}$. The sums are written explicitly in Eqs.\,\eqref{eq:external} to \eqref{eq:truncWeights} {in Appendix \ref{app:appB}, these equations also give the weights when the range of particle interactions is truncated.} The grand potential of this system can be approximated as follows:\cite{hughes13}
\begin{align}\label{eq:latGas}
\Omega(\{\rho_{{\bf i}_0}\}) & = U-TS-\mu N, \nonumber \\
& = k_BT \sum_{{\bf i}_0=1}^{M_{2d}} \left[ \rho_{{\bf i}_0} \ln(\rho_{{\bf i}_0})+(1- \rho_{{\bf i}_0})\ln(1- \rho_{{\bf i}_0})\right] \nonumber \\
& \qquad- \frac12 \sum_{{\bf i}_0=1}^{M_{2d}} \sum_{{{\bf j}_0} \neq {{\bf i}_0}} \epsilon_{{{\bf i}_0},{{\bf j}_0}} \rho_{{\bf i}_0} \rho_{{\bf j}_0} + \sum_{{\bf i}_0=1}^{M_{2d}} \rho_{{\bf i}_0}(V_{{\bf i}_0} - \mu),
\end{align}
where $k_B$ is Boltzmann's constant and $S$ is the entropy\footnote{Note that the contribution to the lattice gas free energy density $\beta f=\rho\ln\rho-(1-\rho)\ln(1-\rho)$ in Eq.~\eqref{eq:latGas} contains both the ideal gas $\beta f_{id}=\rho\ln\rho-\rho$ and an excess contribution $\beta f_{ex}=\rho+(1-\rho)\ln(1-\rho)$; i.e.\ $f=f_{id}+f_{ex}$. The origin of this form arises from the hole-particle $n_{\bf i}\leftrightarrow1-n_{\bf i}$ symmetry in the lattice model (see e.g.\ Ref.\,\onlinecite{hughes13}). The second term, $f_{ex}$, stems from the particles having a hard-core interaction, preventing multiple occupancy of a cell. It is instructive to compare this with the free energy of a binary mixture. The ideal-gas contribution to the free energy of a mixture is $\beta f=\rho_1\ln\rho_1-\rho_1+\rho_2\ln\rho_2-\rho_2$, where $\rho_1$ is the density of species 1 and $\rho_2$ the density of species 2. Re-writing this in terms of the total density $\rho=\rho_1+\rho_2$ and concentration $x=\rho_1/\rho$, we obtain $\beta f=\rho\ln\rho-\rho+\rho[x\ln x+(1-x)\ln(1-x)]$. The final term is the ideal-gas contribution to the entropy of mixing and it is interesting to note the similarity in form to the lattice-gas free energy. For an incompressible binary mixture $\rho$ is a constant and so the first two terms are often neglected. As a consequence of this, the free energy of an incompressible binary fluid can have a very similar mathematical form to the lattice-gas free energy considered here.}. Included implicitly in the above equation is the particle diameter $\sigma=1$.
The set of lattice densities that describes the system at equilibrium is found by solving [c.f.\ Eq.\ \eqref{eq:EL_eq}]:
\begin{equation}\label{eq:eqlbrmCon}
\frac{\partial \Omega}{\partial \rho_{\bf i}}=0,
\end{equation}
for every lattice site ${\bf i}$. From Eqs.\,\eqref{eq:latGas} and \eqref{eq:eqlbrmCon} we obtain
\begin{equation}\label{eq:difLatGas}
k_BT\ \ln\left(\frac{\rho_{\bf i}}{1-\rho_{\bf i}}\right) - \sum_{{\bf j}=1}^{M_iM_j} \epsilon_{{\bf i},{\bf j}} \rho_{\bf j} + V_{\bf i}-\mu = 0.
\end{equation}
Solving this coupled set of equations \eqref{eq:difLatGas} gives the equilibrium fluid profile $\{\rho_{\bf i}\}$. However, in order to obtain the non-equilibrium profile with specified adsorption $\Gamma_d$, the iterative method developed in Ref.\,\onlinecite{archer11} must be used. This is equivalent to solving the set of coupled equations [c.f.\ Eq.\,\eqref{eq:difLatGas}]
\begin{equation}
k_BT\ \ln\left(\frac{\rho_{\bf i}}{1-\rho_{\bf i}}\right) - \sum_{{\bf j}=1}^{M_iM_j} \epsilon_{{\bf i},{\bf j}} \rho_{\bf j} + V_{\bf i} + V_{\bf i}^\text{eff} -\mu = 0,
\end{equation}
where $V_{\bf i}^\text{eff}$ is the fictitious additional potential mentioned at the end of Sec.\ \ref{sec:binding}. It has the property that $V_{\bf i}^\text{eff} \to 0$ as $i \to \infty$, far from the wall. This additional potential stabilises a film of liquid with the specified adsorption against the wall. Note that $V_{\bf i}^\text{eff}$ is a-priori unknown, it is calculated on-the-fly self consistently as part of the minimisation algorithm.\cite{archer11} This is done by re-normalising the density profile during every iteration of the algorithm by replacing the value of the density with
\begin{equation}\label{eq:renorm}
\rho_{\bf i}^\text{new} = \frac{\Gamma_d}{\Gamma_\text{old}} (\rho_{\bf i}^\text{old} - \rho_b) + \rho_b,
\end{equation}
where $\Gamma_d$ is the desired adsorption and $\Gamma_\text{old}$ is the adsorption corresponding to the density profile $\{\rho_{\bf i}^\text{old}\}$, obtained from iterating Eq.\,\eqref{eq:difLatGas}. More details about this algorithm and its properties can be found in Ref.\ \onlinecite{archer11}.
{Note that this procedure does not require the bulk phase to be at coexistence, although in all the results presented here it is at coexistence. As the bulk fluid state point $(\mu,T)$ is varied, the form of the restricted free energy $g(\Gamma)$ also changes as well as the bulk densities and, in consequence, the interface tensions. Note also that one can vary the adsorption $\Gamma$ by the standard method of varying the value of the chemical potential $\mu$. Calculations (not presented here) show that the main difference between results from our method for varying the film thickness via a fictitious external potential, with results from varying it by changing $\mu$ are to be seen when the adsorption is small. This is particularly so when the fluid wets the wall, because in this case, to obtain a small adsorbed film height by varying the chemical potential requires a large shift in the value away from that at coexistence.}
The bulk fluid phase diagram is displayed in Fig.\,\ref{fig:phase}. For details about how this is calculated see e.g.\ Refs.\, \onlinecite{hughes13,robbins11}. The binodal and spinodal are both displayed. The binodal gives the densities of the coexisting gas and liquid states. Within the spinodal curve, the uniform fluid is unstable and spontaneous demixing occurs. The bulk critical point is at $\rho\sigma^3=0.5$ and temperature $k_BT/\epsilon=1.5$.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{phaseDiagram}
\caption{The phase diagram for the lattice fluid in the temperature $k_BT/\epsilon$ versus density $\rho\sigma^3$ plane. The solid (red) line is the binodal curve and the dashed (blue) line is the spinodal curve.}
\label{fig:phase}
\end{center}
\end{figure}
DFT is a statistical mechanical theory - i.e.\ in principle it should give the ensemble average density profile of the fluid. A statistical description of a fluid confined in an external potential should yield an (ensemble average) density profile with the same symmetry as that potential.\cite{evans79,malijevsky13} Thus, for a planar wall, the equilibrium density profiles only vary with the distance from the wall. Fig.\,\ref{fig:prof1d} shows typical examples of such profiles for the lattice-gas model and of the corresponding points on the binding potential that they represent. This series of density profiles range from small (including negative) values of the adsorption to large values, where the profiles indicate that there is a thick film of liquid at the wall.
The corresponding fictitious potentials are also displayed in Fig.\,\ref{fig:prof1d}. At the minimum in $g(\Gamma)$, the fictitious potential $V_{\bf i}^\text{eff}$ is, of course, zero, because this corresponds to the equilibrium state. Moving away to either side of this minimum, $V_{\bf i}^\text{eff}$ increases rapidly and the largest magnitude potentials are observed for low adsorptions $\Gamma$, for values of $\Gamma$ where the gradient in $g(\Gamma)$ is largest. In Fig.\,\ref{fig:prof1d} the density profile (and associated potential $V_{\bf i}^\text{eff}$) corresponding to $\Gamma \sigma^2 = 0.304$ is very close to at equilibrium and so $V_{\bf i}^\text{eff}$ is very small everywhere. In contrast, the fictitious potential for $\Gamma \sigma^2=0.004$ is much larger, as the gradient of $g(\Gamma)$ at this point is also large. For larger adsorption values, in the tail of the binding potential, $V_{\bf i}^\text{eff}$ can be either weakly attractive or weakly repulsive. This is due to the oscillations in $g(\Gamma)$ stemming from the fact that we are dealing with a lattice model. In a continuum DFT model these oscillations decrease in amplitude as $\Gamma$ increases or are entirely absent, depending on the fluid state point. In the subsection below, we discuss this issue further.
\begin{figure*}
\includegraphics[width=0.66\columnwidth]{mrkdBinding_alt}
\includegraphics[width=0.66\columnwidth]{oneDProfs_alt}
\includegraphics[width=0.66\columnwidth]{effPots_alt}
\includegraphics[width=0.66\columnwidth]{bind_bigAds}
\includegraphics[width=0.66\columnwidth]{bigEffs}
\caption{In (a) we display the binding potential $g(\Gamma)$ for a fluid with bulk gas density $\rho \sigma^3=0.107$ and temperature $\beta \epsilon=0.9$ against a planar wall with attraction strength $\beta \epsilon_w=0.6$. In (d) we display a magnification of tail of $g(\Gamma)$, for larger values of the adsorption $\Gamma$. Marked on $g(\Gamma)$ are points that correspond to the density profiles displayed in (b), and the corresponding fictitious potentials $V_{\bf i}^\text{eff}$, which are displayed in (c) and (e). The marked points have adsorption values: $\Gamma \sigma^2 = 0.004$, 0.304, 0.6, 1.6, 3.6, 4.6, 5.6, 6.6, 7.6, 8.6, 9.6 (note that the density profiles for the final four values are not displayed in (a) and (b), for clarity). The fictitious potential is that which must be applied to stabilise a film of liquid with the given adsorption in an open (grand-canonical) system. We observe a clear relation between the gradient of $g(\Gamma)$ and the magnitude and sign of the fictitious external potential.}
\label{fig:prof1d}
\end{figure*}
Note that even when the typical microstates of the system consist of liquid drops on the surface, after performing a statistical average over all states, the resultant density profile should be invariant in the direction parallel to the substrate.\cite{malijevsky13} In order to study liquid drops or liquid ridges (invariant in one direction along the surface) one must break the translational symmetry and impose that the centre of mass be located at a particular point or line on the surface. By constraining the centre of mass, 2D drop profiles (i.e.\ liquid ridges) can also be found for the same system even though the external potential only varies in one direction.
Fig.~\ref{fig:prof2d} displays such 2D density profiles for external wall potentials of varying attraction strengths. These are calculated by solving Eq.\,\eqref{eq:difLatGas}. However, the value of $\mu$ is not at the outset imposed, instead the total adsorption \eqref{eq:adsorption} is specified. This is done by renormalising the density profile at each iteration of the algorithm via Eq.\,\eqref{eq:renorm}. For more details about this algorithm, see Ref.\ \onlinecite{hughes13}.
Fig.~\ref{fig:prof2d} shows that for weakly attracting walls (small $\epsilon_w$), it is energetically favourable for the liquid not to be in contact with the wall and the contact angle is large. As $\epsilon_w$ is increased, the contact angle decreases as the fluid seeks to have greater contact with the wall. As discussed below (see also Fig.~\ref{fig:bindingPot}), there is a wetting transition at $\beta \epsilon_w\approx 0.74$ and for values of $\beta \epsilon_w$ greater than this, the liquid spreads over the surface ($\theta=0$). {Note that the lattice-gas DFT, with various different choices for $\epsilon_{{\bf i},{\bf j}}$, has been used extensively to study wetting and also to calculate profiles for liquid drops on surfaces -- see e.g.\ Refs.\,\onlinecite{detcherverry03, porcheron06, monson08} for some recent example. As far as we are aware, in all previous studies the adsorption on the surface is determined by the choice of $\mu$, either explicitly, or by fixing the total number of particles in the system, $N$. Since the fictitious potential that we use in our method to stabilise the drops is almost always rather small, the density profiles we obtain are actually rather similar to those found previously when the interaction and wall potentials are the same.}
\begin{figure}
\includegraphics[width=\columnwidth]{profa}
\includegraphics[width=\columnwidth]{profb}
\includegraphics[width=\columnwidth]{profc}
\includegraphics[width=\columnwidth]{profd}
\includegraphics[width=\columnwidth]{profe}
\caption{A series of 2D density profiles for a drop of the lattice-gas fluid with $\beta \epsilon=0.9$, deposited on various solid substrates with attraction strength parameters $\beta \epsilon_w = 0.2$, 0.5, 0.6, 0.72 and 0.74. The fluid-fluid particle interaction range is truncated to $L=5$.}
\label{fig:prof2d}
\end{figure}
\subsection{The Binding Potential via the Sharp-Kink Approximation}\label{sec:SK}
Some of the interfacial and wetting behaviour of the lattice-gas model can be elucidated in a particularly simple manner by making the so-called sharp kink (SK) approximation. \cite{schick90,dietrich88} Specifically, the lattice gas density profile is taken to be
\begin{equation}
\rho_i =
\begin{cases}
0 &\mbox{for } i \leq 0, \\
\rho_l &\mbox{for } 0 < i \leq h, \\
\rho_g &\mbox{for } i > h,
\end{cases}
\end{equation}
where $h$ is the position of the liquid-gas interface and the surface of the solid wall is at the lattice site $i=0$. No minimisation is necessary under this approximation and the binding potential for any given film thickness $h$ can be calculated directly. For a system with truncated fluid-fluid interactions, the binding potential can be calculated as
\begin{equation}\label{eq:fullAsym}
g(h) = (\rho_l- \rho_g){\sigma^3}\sum_{i=h+1}^\infty V_i,
\end{equation}
when $h>L$, the range of the fluid-fluid interactions. An asymptotic expansion of this sum gives (see also Ref.\,\onlinecite{macdowell14}):
\begin{equation}\label{eq:sk_asym}
g \sim \epsilon_w \pi (\rho_l- \rho_g){\sigma^3}\left(\frac{{\sigma^2}}{12h^2} + \frac{{\sigma^3}}{12h^3} + \cdots \right),
\end{equation}
Truncating this series gives an approximation that is valid for large $h$. The coefficients given in Eq.\,\eqref{eq:longAsym} can therefore be calculated explicitly for the lattice-gas model under this approximation. Figure \ref{fig:sklg_comp} shows the binding potential for $\beta\epsilon=0.9$ and $\beta\epsilon_w=0.55$ on a log-log plot, calculated using the full lattice-gas model and comparing with results from the SK approximation. The analytically calculated asymptotic limit in Eq.\,\eqref{eq:sk_asym} is plotted to $O(h^{-2})$, which agrees very well with a numerical evaluation of the full SK approximation. The numerical and analytic SK results begin to slightly drift apart for larger adsorptions. This is because in the numerical calculation the external potential is truncated to a range of $100\sigma$ whereas the analytic calculation assumes an infinite interaction range. This truncation results in the numerical result diverging from the $\sim h^{-2}$ decay in Eq.\,\eqref{eq:sk_asym} for large $h$.
The density profile obtained by minimising the DFT, in contrast to the SK approximation, incorporates a better approximation for the true shape of the liquid-gas interface and also the effect of this interface being close to the wall. Thus, for small values of $\Gamma$ (i.e.\ small $h$), this approximation is far more reliable. Also, it includes the correct $\Gamma^{-2}$ asymptotic decay for large $\Gamma$. Note that in Fig.\,\ref{fig:sklg_comp} the oscillations in the tail of the binding potential are present as a result of the system being discretised on a lattice. The free energy is lower when the liquid-gas interface is between two lattice sites, rather than on a lattice site. Thus, as the specified adsorption is increased, the liquid-gas interface moves continuously away from the wall which leads to oscillations in $g(\Gamma)$. These oscillations lie on top of the correct $\sim \Gamma^{-2}$ asymptotic decay; i.e.\ for large $h$ the binding potential is of the form $g \approx \epsilon_w \pi {\sigma^5}(\rho_l- \rho_g)/(12h^2)-B\cos(2\pi h/\sigma)$, where the amplitude of the oscillations $B$ is a small number that depends on the state point and manner in which the interactions are treated. As the range of the fluid-fluid interactions is increased, the amplitude of these oscillations $B$ decreases (c.f.\ Fig.\,\ref{fig:aPlot}). Thus, apart from these oscillations, the binding potential calculated from the full minimisation matches up well for large $\Gamma$ with the results from the SK approximation. {We attempted to extract the coefficients for the higher terms in the expansion in Eq.~\eqref{eq:longAsym} from our numerical results, since analytic expressions for these exist.\cite{dietrich91} However, the oscillations induced by the lattice in $g(h)$ make this problematic.} Note too that the free energy contribution from the liquid-gas free interface $\gamma_{lg}$ is a little different in the SK approximation compared to that obtained from the full minimisation. The respective values are subtracted when calculating $g(\Gamma)$ in the two methods.
For small values of the adsorption, large differences are seen between the SK results for $g(\Gamma)$ and those from the full minimisation. This demonstrates that the SK approximation should only be used for thicker films and a microscopic theory for the fluid structure (DFT) is required to accurately calculate a binding potential that is valid for thin ($h\lesssim 3\sigma)$ liquid films.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{sklg_comp}
\caption{The binding potential for $\beta \epsilon=0.9$ and $\beta \epsilon_w=0.55$ calculated for the lattice gas model and compared with results from the sharp kink approximation, displayed on a log-log plot. The green dashed line corresponds to the leading order ($\sim h^{-2}$) term in Eq.\,\eqref{eq:sk_asym} and the dotted blue line to a numerical evaluation of the sum in Eq.\,\eqref{eq:fullAsym}. The red solid line is the result from a constrained minimisation of the full functional.}
\label{fig:sklg_comp}
\end{center}
\end{figure}
\section{The Influence of the Range of Particle Interactions}\label{sec:ranges}
In Sec.\ \ref{sec:binding} the manner in which the particle interactions influence the (asymptotic) form of the binding potentials was discussed; i.e.\ for long range dispersion interactions the algebraic decay form in Eq.\,\eqref{eq:longAsym} is appropriate, whereas for short-range forces, the exponential form in Eq.\,\eqref{eq:shortAsym} is the correct asymptotic form. In this section, the cross-over from one regime to the other is explored as the interaction potentials in the lattice-gas model are truncated at different ranges.
When comparing systems with different interaction ranges, it is important that the truncation does not change the bulk fluid phase diagram. Otherwise, even if comparing systems at the same temperature $T$, the value of $(T-T_c)$, where $T_c$ is the bulk critical temperature, would not be the same for systems with a different truncation range $L$. For the simple mean-field approximation to the free energy in Eq.\,\eqref{eq:latGas}, the bulk uniform fluid free energy is determined by the integrated interaction strength of the pair potential $\sum \epsilon_{{\bf i},{\bf j}}$, i.e.\ the total potential that arises from the interaction of a single particle with all others within the interaction range. As the interaction range is adjusted, one must vary the value of $\epsilon$, the parameter governing the overall strength of the pair interactions, to ensure that the integrated interaction strength remains constant so that the bulk fluid phase diagram remains unchanged. All values of $\beta \epsilon$ quoted here are the strength of the interaction when the interaction range is truncated to only the nearest neighbour lattice sites $(L=1)$.
Truncating the range over which particles in the system interact changes the overall shape of the resulting binding potential. In particular, if all interaction potentials (both fluid-fluid and wall-fluid) are truncated, then the tail of the binding potential decays exponentially to zero. If there are any long ranged (not truncated) interactions then, in three dimensions, the binding potential tail decays algebraically $\sim h^{-2}$. By varying the range of the particle interactions in the lattice-gas model, both of these regimes are seen and also there is a crossover from one to the other as the truncation range $L$ is varied. These results are displayed in Figs.~\ref{fig:aPlot} and \ref{fig:bPlot}.
Fig.~\ref{fig:aPlot} shows the binding potential for $\beta\epsilon=0.8$, $\beta\epsilon_w=0.5$ and at different truncation ranges, $L$, of the fluid-fluid interactions. Note, however, that the interactions between wall and fluid particles extends over the entire domain in all cases (the domain size is $M_k=100\sigma$). As the truncation range $L$ is increased, it becomes energetically favourable for the fluid not to wet the substrate and so a minimum in $g(\Gamma)$ at a finite value of the adsorption $\Gamma$ appears; i.e.\ the interfacial phase behaviour changes purely as a result of how the fluid-fluid particle interactions are modelled. Note that since the value of $g(\Gamma)$ at the minimum $g(\Gamma_0)\equiv g(h_0)$ decreases as $L$ increases, from Eq.\,\eqref{eq:bindCont}, this indicates that increasing the interaction range $L$ makes the fluid less wetting and increases the contact angle. This shows that care should always be taken when modelling the interaction between two fluid particles: truncating the pair potentials at too small a distance may result in significant errors in predictions for the interfacial phase behaviour. Fig.~\ref{fig:aPlot} also shows the $\sim \Gamma^{-2}$ decay of $g(\Gamma)$ as $\Gamma\to\infty$ that is present in all cases, due to the presence of the long ranged wall-fluid interactions.
The effect of truncating the range of all interactions (including the wall-fluid potential) on the binding potential is shown in Fig.\,\ref{fig:bPlot} for the case when $\beta \epsilon=0.8$ and $\beta \epsilon_w=0.7$. The binding potential for $L=80$ is the longest ranged, having the slowest decay to zero as $\Gamma$ increases. As $L$ is decreased, the range of $g(\Gamma)$ decreases. The inset of the figure shows the same data on a log-log scale which enables one to clearly see the form of the asymptotic decay of $g(\Gamma)$ as $\Gamma \to \infty$. As expected from the discussion above, as the range of the interactions $L$ is increased, an increasingly large portion of algebraic $\sim\Gamma^{-2}$ decay is present. However, it should be pointed out that formally speaking, it is only when $L\to\infty$ that the ultimate asymptotic decay of $g(\Gamma)$ as $\Gamma \to\infty$ changes from exponential to algebraic.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{rangeTrunc1}
\caption{The binding potential for the fluid with inverse temperature $\beta \epsilon=0.8$ at a wall with attraction strength $\beta \epsilon_w=0.5$. The various different curves correspond to truncating the fluid-fluid pair interactions at different values of the truncation length $L$. The wall-fluid interactions remain fixed at a truncation of $100\sigma$. Note that we vary $L$ in a manner that does not change the bulk fluid phase diagram. We see that the fluid is predicted to be more wetting as $L$ becomes shorter; in fact, the interfacial phase behaviour changes from non-wetting to wetting, purely as a result of truncating the interaction range. The inset shows the same curves plotted on a log-log scale which shows that the asymptotic decay form for large $\Gamma$ is the same in all cases because of the long-ranged wall-fluid interactions.}
\label{fig:aPlot}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{rangeTrunc2}
\caption{The binding potential for the fluid with inverse temperature $\beta \epsilon=0.9$ at a wall with attraction strength $\beta \epsilon_w=0.7$. For these values, the fluid wets the wall. In contrasts to the case in Fig.~\ref{fig:aPlot}, here we truncate both the wall-fluid and fluid-fluid interactions at the same range, $L$. The different curves correspond to varying $L$ from between 1 to 80 particle diameters. As $L$ is decreased, the range of $g(\Gamma)$ decreases. In the inset, the black dashed line is $0.001\Gamma^{-2}$. As $L$ is increased, the binding potential approaches this large $\Gamma$ asymptotic decay form.}
\label{fig:bPlot}
\end{center}
\end{figure}
\section{A Fitting Function for the Binding Potential}\label{sec:fitting}
In order to take the binding potentials calculated via DFT in the previous section and use them with the mesoscale IH model, a suitable fit function is required. The fit function should have the same form as that observed in the DFT results so that it can accurately fit the data. Several aspects of the binding potential curves are particularly important for the fit function to be correct: At small values of the adsorption $\Gamma$, the binding potential exhibits a minimum and a maximum when the fluid is non-wetting or near to the wetting transition. These need to be fitted well; in particular the value at the minimum $g(h_0)$ needs to be the correct value [c.f. Eq.\,\eqref{eq:bindCont}]. Also, the true binding potential is finite for $\Gamma=0$ (unlike the approximation in Eq.\,\eqref{eq:longAsym} and other such power-series) and so the fit function should not diverge at $\Gamma=0$. Finally, when dispersion interactions are present, the fit function should exhibit the correct $\Gamma^{-2}$ decay as $\Gamma\to\infty$. Therefore, the following form for the fit function is suggested:
\begin{equation}\label{eq:expFit}
g(\Gamma) = A \frac{\exp[- P(\Gamma)] -1}{\Gamma^2},
\end{equation}
with
\begin{equation}
P(\Gamma) = a_0 \Gamma^2 e^{-a_1 \Gamma} +a_2 \Gamma^2 + a_3 \Gamma^3 + a_4\Gamma^4 + a_5\Gamma^5.
\end{equation}
The rationale for this choice is as follows: For small $x$, $\exp(x) \approx 1+x$, and so at low adsorptions this form gives $g(\Gamma) \approx A(a_0 e^{-a_1 \Gamma} +a_2 + a_3 \Gamma + \dots)$. For high adsorptions this form gives $g(\Gamma) \sim -A \Gamma^{-2}$, which is the correct form for the asymptotic $\Gamma\to\infty$ decay. The coefficient $a_5$ of the highest order term must be positive. The second exponential within the outer exponential function [i.e.\ the exponential term with coefficient $a_0$ in $P(\Gamma)$] helps to correctly fit the minimum at low adsorptions, which is often asymmetric, being much steeper on the low adsorption side of the minimum, compared to the other side, as can seen in Fig.\,\ref{fig:sk_comp}. The constant $a_1$ in Eq.\,\eqref{eq:expFit} is usually quite large and positive so that the inner exponential term has almost no effect on the form of $g(\Gamma)$ on the large adsorption side of the minimum.
An example of the fitting function is displayed in Fig.\,\ref{fig:sk_comp}. The function is plotted with the original data and the SK results as a comparison. The SK results do not describe the behaviour at small $\Gamma$ whereas the fit function gives a very good approximation to the data over the whole range. The leading order coefficient of the decay of the binding potential, $A$ in Eq.\,\eqref{eq:expFit}, can either be calculated directly using the SK approximation or can be found by fitting the DFT data, both methods give similar results. The value for $A$ from the SK approximation is used here. In {Appendix A} we give the fit parameter values in Eq.~\eqref{eq:expFit} for a range of values of $\beta\epsilon$, $\beta\epsilon_w$ and $L$.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{sk_comp}
\caption{The binding potential for a fluid with $\beta \epsilon=0.9$ and $\beta \epsilon_w=0.55$ with truncated interaction ranges of $5\sigma$ and $100\sigma$, for the fluid-fluid and wall-fluid interactions, respectively. The data calculated from the DFT model is shown with the fitting function Eq.\,\eqref{eq:expFit} (parameter values are given in {Appendix A}) and the sharp-kink binding potential for the same parameters. The inset shows the same data on a log-log scale.}
\label{fig:sk_comp}
\end{center}
\end{figure}
\section{Droplet Profiles Obtained Using the Binding Potential}\label{sec:results}
Using the binding potential calculated from the DFT, a drop thickness profile can be calculated from the IH model, by minimising the free energy in Eqs.\,\eqref{eq:fullcurveenergy} or \eqref{eq:thinEnergy}. This drop profile, despite being the result of a mesoscale calculation, contains information about the nature of the microscopic interactions between particles in the system via the binding potential. Of course, a drop profile can also be calculated directly using DFT; the result of such a calculation is shown in Fig.\,\ref{fig:prof2d}. However, it is computationally easier to treat larger systems using the IH model. Also, non-equilibrium situations are much more easily modelled via Eq.\,\eqref{eq:thinFilm} than with a dynamical DFT model that includes all the hydrodynamics.\cite{archer06, archer09} However, since the two approaches are both based upon the same microscopic interactions, the resulting drop profiles from each method should be the same at the mesoscopic scale. This section shows how the two approaches compare. {We find that overall, the drop shape profiles from the DFT and from the IH model are in good agreement, a result which {\em a-priori} is not obvious, if bearing in mind the degree of coarse-graining in going from a density profile to a film height profile. The good agreement between the two is thus more than a mere consistency test.}
Performing the minimisation described in Sec.\ \ref{sec:lattice} for the lattice-gas DFT{, constrained via Eq.\,\eqref{eq:renorm},} on a 2D domain yields density profiles such as those displayed in Fig.\,\ref{fig:prof2d}. These 2D profiles correspond to the density profile of a cross section through a liquid ridge on a surface with centre of mass along the line $x=0$. The location of the liquid-gas interface can be calculated using Eq.\,\eqref{eq:filmHeight}. The alternative procedure, which yields a very similar result, is to just plot the contour where the density $\rho_{\bf i}=(\rho_g+\rho_l)/2=0.5$. Note that this treats the (strictly) discrete density profile as a continuous function. In Fig.\,\ref{fig:drops} are displayed drop profiles obtained from DFT and using Eq.\, \eqref{eq:filmHeight}, compared with results obtained from minimising Eq.\,\eqref{eq:fullcurveenergy} together with the binding potential obtained from DFT for various values of $\beta \epsilon_w$. These results show that drop profiles obtained via the two methods coincide well with each other, {as was also observed in Ref.~\onlinecite{nold14}}. In Fig.\,\ref{fig:full/longwave} are comparisons of a droplet profile found via the DFT route with droplets calculated from a minimisation of both the full curvature free energy Eq.\,\eqref{eq:fullcurveenergy} and the long wavelength approximation free energy, Eq.\,\eqref{eq:thinEnergy}. This shows that the full curvature free energy \eqref{eq:fullcurveenergy} gives closer agreement with the DFT results than the approximation in Eq.\,\eqref{eq:thinEnergy}, as one would expect. Note too that when the contact angle is very small, then the drop shape is very sensitive to variations in the parameters.
\begin{figure}
\vspace{-3em}
\includegraphics[width=0.48\textwidth]{dropComparison}
\vspace{-3em}
\caption{Comparison of drop profiles calculated from DFT (points) and from the IH model, Eq.\,\eqref{eq:fullcurveenergy} (dashed lines). We compare drops of equal maximum height for a range of values of the wall attraction strength $\epsilon_w$. The two methods coincide well across the range of contact angles. These are for the case when the particle interactions are truncated to a range of $L=5$ with a strength of $\beta \epsilon=0.9$. The parameters used in the fitting functions for $g(h)$, used to obtain the IH results, can be found {in Appendix A}.}
\label{fig:drops}
\end{figure}
\begin{figure}
\includegraphics[width=0.48\textwidth]{longwave001}
\includegraphics[width=0.48\textwidth]{longwave005}
\caption{Drop profiles calculated from the DFT density profiles together with those from the IH model using both the full curvature free energy \eqref{eq:fullcurveenergy}, and the long wavelength approximation, Eq.\,\eqref{eq:thinEnergy}. These results are for $\beta\epsilon=0.9$ and $\beta\epsilon_w=0.5$ (top figure) and $\beta\epsilon_w=0.7$ (bottom figure). In the inset we show a magnification of the contact line region. Note the `precursor' film height $\ll \sigma$ indicating that this is sub-monolayer.}
\label{fig:full/longwave}
\end{figure}
Determining the contact angle from drop profiles obtained using the IH model is not straight-forward because the profiles have a precursor film. The DFT calculations show (see e.g.\ Figs.\,\ref{fig:profs}, \ref{fig:prof1d} and \ref{fig:prof2d}) positive adsorption on the surface outside of a droplet. However, the fluid density in the first ($i=1$) layer of lattice sites is substantially less than one -- see also the inset in Fig.\,\ref{fig:full/longwave}. This indicates that the density on the surface outside of the drops is not even a complete monolayer and so calling it a ``precursor film'' is arguably misleading.
Returning to the issue of calculating the contact angle in a IH pre-cursor film model: here, this is calculated from the curvature at the top of the drop, at $x=0$. Liquid droplets of a suitable size have the shape of a spherical cap. This is the case when the volume of the drop is small enough that gravity does not play a significant role but also large enough that the shape at the maximum is not distorted by the binding potential. It is droplets of this size that are studied here. In these cases where the drops have this spherical cap shape, the contact angle for the drop is calculated by fitting a circle to the highest point of the drop profile. The contact angle that this circle makes with the substrate is taken to be the contact angle of the drop. This procedure is straight-forward: The curvature at a stationary point, i.e.\ the maximum point of the liquid drop, is simply the second derivative of the height profile at that point and so the radius of curvature is given as
\begin{equation}
r_c= \frac{1}{h''(x_\text{max})}
\end{equation}
where $x_\text{max}$ is the point where the height of the droplet $h$ is at it's maximum value. Defining $a$ to be the distance from the point that the circle meets the substrate to the centre of the base of the drop, then the contact angle is given by
\begin{equation}
\theta_c = \frac{\pi}{2}- \cos\left(\frac{a}{r_c}\right).
\end{equation}
Comparisons made in this way can only be used to relate the two theories for systems where a liquid drop exists, i.e.\ for a partially wetting system. The focus here is also on drops with an acute contact angle as these profiles can be calculated by minimising Eq.\,\eqref{eq:thinEnergy}, but droplets with obtuse contact angles cannot. For non-equilibrium problems one makes the assumption of small contact angles (long-wavelength) and utilises the thin film equation, Eq.~\eqref{eq:thinFilm}, with the free energy functional given by Eq.\,\eqref{eq:thinEnergy}. We should also mention that making the long-wave approximation, i.e.\ going from Eq.~\eqref{eq:fullcurveenergy} to Eq.~\eqref{eq:thinEnergy}, also results in the drop shape away from the surface ceasing to be the arc of a circle and instead being a parabola.\cite{safran94} Thus, for the long-wave theory \eqref{eq:thinEnergy}, we fit the drops with a parabola, in order to make a fair comparison. Fig.~\ref{fig:cntAngle} shows the comparison of contact angles calculated for different values of $\beta \epsilon_w$, from the DFT model and from the IH model, for both the full curvature case and the long-wavelength approximation.
An alternate way to obtain the contact angle is by calculating the largest gradient of the drop profile. The inverse tangent of this gives the contact angle. These `maximum gradient' contact angles are also shown in Fig.\,\ref{fig:cntAngle}. This figure shows that results based on the full curvature free energy agree well with those calculated from the DFT model. The IH model in the long-wave approximation only agrees with the DFT results for very small contact angles, as expected. For the full curvature model, the contact angle found by fitting a circle to the profile matches the DFT results better than those from the maximum gradient method. The droplet shape calculated under the long-wavelength approximation is no longer that of a spherical cap, it is instead parabolic, and so a similar procedure to the circle fitting method is employed where a parabola is fitted to the droplet profile. Results for the contact angle obtained via these fits are shown in Fig.\,\ref{fig:cntAngle}. Clearly, these results depend heavily on the method used to find the contact angle of the droplets and although the circle/parabola fitting method gives better results it is only applicable to specific drop shapes -- see Ref.~\onlinecite{tretyakov13} for a further discussion on extracting a contact angle from a drop profile. Note also that the contact angles in Fig.\,\ref{fig:cntAngle} calculated via DFT are the only results to extend beyond $90^\circ$; droplet profiles can not be found in this range using the IH model.
The calculated contact angles {are weakly dependent on the size of the drop}. All contact angles {extracted from the IH model are calculated} from droplets that have a height of $35\sigma$. Calculating contact angles from droplets of a specific height, rather than a specific volume, seems to give more consistent results. {The range at which particle interactions are truncated does not significantly affect how well the results from the two models (DFT an IH) agree with each other. Over the range of parameters studied, up to a truncation range of $L=40$, the discrepancy in the contact angle between the two models is typically a few percent.}
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{angleComp}
\caption{The contact angle for a range of different values of $\beta \epsilon_w$, for $\beta\epsilon=0.9$, calculated from both the DFT and IH models. Contact angles are calculated for both the full curvature and long-wave approximation versions of the IH model, using both the circle fitting and maximum gradient methods. The circle fitting method does not well represent the actual contact angle of the droplet from the long-wave IH model.}
\label{fig:cntAngle}
\end{center}
\end{figure}
\section{Conclusions}\label{sec:conclusion}
In this paper we have developed a microscopic DFT based method for calculating the binding potential $g(\Gamma)$ at various different state points and interaction potentials of various different strengths and ranges between the surface and the liquid. These were subsequently used as input to a mesoscopic {interface} free energy model, which was then used to calculate the height profile of liquid drops on surfaces. The liquid height profiles are very similar in shape to the profiles obtained directly from the DFT, indicating that the coarse-graining procedure used to obtain $g(\Gamma)$ is valid.
The binding potential is calculated as a function of the adsorption $\Gamma$, which can be related to the film height via Eq.\ \eqref{eq:adsorption_approx}. It is calculated using the method developed in Ref.\ \onlinecite{archer11} for studying nucleation of the liquid from the gas phase. This method constrains the fluid density profile to have a specified value of $\Gamma$, which is equivalent to imposing an additional fictitious external potential that stabilises the adsorbed film with the given adsorption. Note that this fictitious potential is not known a-priori and is calculated on-the-fly as part of the minimisation to obtain the fluid density profile.
The method presented for calculating the binding potential is general and it should be possible to use if for any DFT. Here it has been implemented using a simple DFT for a lattice-gas, that in Eq.\ \eqref{eq:latGas}. A better description of the free energy of a lattice-gas is possible, we refer the interested reader to e.g.\ Ref.\ \onlinecite{edison14b} for a more sophisticated and accurate approximation for the lattice-gas free energy. The lattice-gas approximation was used in order to more easily test the results for a range of different drop sizes. Note that to calculate the 2D density profiles for drops of the size in Fig.\ \ref{fig:prof2d} (or even larger) with a more sophisticated continuum DFT can be computationally time consuming -- e.g.\ using the commonly used approximation
${\cal F}[\rho({\bf r})]={\cal F}_{FMT}[\rho({\bf r})]+{\cal F}_{att}[\rho({\bf r})]$, where ${\cal F}_{FMT}[\rho({\bf r})]$ is the fundamental-measure theory approximation of the free energy of a fluid of hard-spheres and ${\cal F}_{att}[\rho({\bf r})]$ is a simple mean-field approximation for the contribution to the free energy due to the attractive interactions between the fluid particles.\cite{evans92,hansen13,Lutsko10,Tarazona08,evans10,lowen10} However, one of the draw-backs of the lattice-gas model is that the discretization onto a lattice leads to unrealistic small amplitude oscillations in $g(\Gamma)$, particularly for larger values of $\beta \epsilon$ (i.e.\ for lower temperatures) -- see e.g.\ Fig.\ \ref{fig:sklg_comp}. A more realistic continuum DFT model will not exhibit such oscillations that do not decay in amplitude with distance from the wall. Note, however, that oscillations may be present in the true binding potential. These, if present, stem from the packing of the particles, rather than from the discretization onto a lattice and are to be expected whenever the fluid exhibits layering transitions.
{Another drawback of the lattice-gas is that the mapping of the continuum fluid onto the lattice constitutes a significant approximation. This prevents a quantitative comparison between our results with existing simulation results for continuum models, such as those in Refs.\ \onlinecite{macdowell05, macdowell06, herring10, macdowell11, rane11, gregorio12, tretyakov13, benet14}. However, since we do find good qualitative agreement, we are currently applying the method using continuum (FMT) DFT. Results from this follow-on study will be published elsewhere, including comparison with simulation results.}
It should also be emphasised that it remains straight-forward to calculate droplet profiles using DFT rather than using the mesoscale model over the full range of contact angles, i.e.\ including when the contact angle $>90^\circ$. However, the IH model is particularly advantageous when considering the non-equilibrium situation, which can be described using Eq.\,\eqref{eq:thinFilm}, albeit this equation is derived under the assumption of small contact angles and only accounts for convective transport with no slip at the substrate. It is shown in Fig.\,\ref{fig:cntAngle} how the accuracy rapidly breaks down for larger contact angles. Non-equilibrium phenomena can also be studied using dynamical DFT, see e.g.\ Refs.\ \onlinecite{marconi99, marconi00, archer04a, archer04b} for more on this approach.
One of the striking results of the present work is the observation of the degree to which the binding potential changes when the range at which particle interactions are truncated is changed -- see in particular Figs.\ \ref{fig:aPlot} and \ref{fig:bPlot}. Crucially, the value of $g(\Gamma)$ at the minimum that can be present at low values of the adsorption depends very sensitively on the truncation range. This minimum can change from being just a local minimum (a metastable equilibrium point) to become the global energetic minimum which represents a shift in the phase behaviour from wetting to non-wetting. In fact, for very short truncation ranges, the minimum can be completely removed. The conclusion from this part of our study is that to determine accurately the location of a wetting transition in theory or simulations one must ensure that the truncation is not so severe as to induce such errors. The tails of the potentials matter!
{This is important in the context of (coarse-grained) Molecular Dynamics simulations where dispersion forces are often cut off at distances of a few particle diameters. Ref.~\onlinecite{tretyakov13}, for instance, employs a cut-off length for the Lenard Jones interactions of two times the equilibrium bead distance and finds, in consequence, that the extracted binding potential is well fitted by a sum of exponentials as in Eq.~(\ref{eq:shortAsym}), similarly to our results when only short-range interactions are considered.}
Note also that for weakly attracting (solvophobic) walls the minimum in $g(\Gamma)$ is at very small, or potentially even negative, values of $\Gamma$. In this regime describing the sub-monolayer adsorption via a film-thickness $h$ is a somewhat misleading concept. In this regime, the minimum in the binding potential is very asymmetric, rising very sharply on the small $\Gamma$ side of the minimum, but rising far less steeply as $\Gamma$ is increased from the value at the minimum. To incorporate this behaviour in the fit-function for $g(\Gamma)$, we had to use the exponential in the function $P(\Gamma)$ in Eq.\ \eqref{eq:expFit}. This also means that the minimum in $g(\Gamma)$ is not well-approximated by a quadratic function when the minimum is at a small value of $\Gamma$ and so, of course, capillary-wave theory also does not apply in this regime, since capillary-wave theory assumes Gaussian fluctuations in a harmonic potential. Also, when the adsorption at the surface is sub-monolayer, then a non-equilibrium situation can no longer be described by Eq.\ \eqref{eq:thinFilm}; it can instead be described via a gradient dynamics model with a diffusive dynamics.
Finally, it should also be pointed out that although in the present work we have developed a method for calculating a better approximation for the binding potential, our approach is still a mean-field theory and therefore does not include all fluctuation effects, because our theory is based on Eq.\ \eqref{eq:fullcurveenergy}. Some effects of fluctuations at wetting transitions can be included by replacing $\gamma_{lg}$ in Eq.\ \eqref{eq:fullcurveenergy} with a function $\gamma(h)$. However, as Parry and co-workers have showed, the {effective interface} Hamiltonian is in fact non-local.\cite{parry06, parry07, bernardino09} Non-locality and fluctuation effects are particularly important at wetting transitions.
\section*{Acknowledgements}
A.P.H. acknowledges support through a Loughborough University Graduate School Studentship. All authors thank the Center of Nonlinear Science (CeNoS) of the University of M\"unster for recent support of the authors collaboration.
\begin{appendix}
\section{Fit function parameters}\label{app:param}
The following table contains the parameters to the fitted binding potentials that have been used throughout this paper {and also for a selection of other state points}. For ease of reference, the fitting function is repeated here:
\begin{equation}
g(\Gamma) = A \frac{\exp[- P(\Gamma)] -1}{\Gamma^2},
\end{equation}
with
\begin{equation}
P(\Gamma) = a_0 \Gamma^2 e^{-a_1 \Gamma} +a_2 \Gamma^2 + a_3 \Gamma^3 + a_4\Gamma^4 + a_5\Gamma^5. \nonumber
\end{equation}
\begin{table*}
\begin{tabular}{c>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$}>{$}c<{$} }
Figure & \beta \epsilon & \beta \epsilon_w & L & A & a_0 & a_1 & a_2 & a_3 & a_4 & a_5 & a_6\\[1em] \hline
\ref{fig:drops} \& \ref{fig:full/longwave}& 0.9 & 0.5 & 5 & -0.073 & 1.142 & 8 & -3.078 & 3.283 & -1.272 & 0.173 & 0.000 \\
\ref{fig:drops} \& \ref{fig:sk_comp} & 0.9 & 0.55 & 5 & -0.081 & 1.080 & 8 & -2.153 & 2.074 & -0.705 & 0.084 & 0.000 \\
\ref{fig:drops} & 0.9 & 0.6 & 5 & -0.088 & 0.964 & 8 & -1.325 & 0.930 & -0.036 & -0.097 & 0.019 \\
\ref{fig:drops} & 0.9 & 0.65 & 5 & -0.096 & 0.770 & 8 & -0.534 & -0.245 & 0.762 & -0.358 & 0.051 \\
\ref{fig:full/longwave} & 0.9 & 0.7 & 5 & -0.103 & 0.609 & 8 & 0.138 & -1.037 & 1.158 & -0.451 & 0.060 \\
- & 0.8 & 0.5 & 2 & -0.062 & 0.431 & 8 & 0.431 & -1.448 & 1.562 & -0.649 & 0.097 \\
- & 0.8 & 0.5 & 3 & -0.062 & 0.562 & 8 & -0.203 & -0.900 & 1.175 & -0.468 & 0.062 \\
- & 0.8 & 0.5 & 4 & -0.062 & 0.669 & 8 & -0.473 & -0.548 & 0.903 & -0.363 & 0.047 \\
- & 0.8 & 0.5 & 5 & -0.062 & 0.776 & 8 & -0.679 & -0.222 & 0.634 & -0.262 & 0.033 \\
- & 0.8 & 0.5 & 10 & -0.062 & 0.998 & 8 & -1.035 & 0.415 & 0.107 & -0.074 & 0.009 \\
- & 0.8 & 0.5 & 20 & -0.062 & 1.125 & 8 & -1.194 & 0.752 & -0.162 & 0.011 & 0.0001 \\
- & 0.8 & 0.5 & 40 & -0.062 & 1.104 & 8 & -1.179 & 0.678 & -0.100 & -0.007 & 0.002 \\
- & 0.9 & 0.6 & 2 & -0.081 & 0.620 & 8 & -0.108 & -0.902 & 1.878 & -1.093 & 0.228 \\
- & 0.9 & 0.6 & 3 & -0.088 & 0.786 & 8 & -0.851 & 0.177 & 0.670 & -0.405 & 0.068 \\
- & 0.9 & 0.6 & 4 & -0.096 & 0.824 & 8 & -1.049 & 0.573 & 0.211 & -0.182 & 0.030 \\
- & 0.9 & 0.6 & 5 & -0.102 & 0.823 & 8 & -1.133 & 0.766 & -0.019 & -0.078 & 0.014 \\
- & 0.9 & 0.6 & 10 & -0.110 & 0.851 & 8 & -1.273 & 1.023 & -0.280 & 0.026 & 0.000 \\
- & 0.9 & 0.6 & 20 & -0.118 & 0.797 & 8 & -1.234 & 0.965 & -0.258 & 0.023 & 0.000 \\
- & 0.9 & 0.6 & 40 & -0.125 & 0.748 & 8 & -1.170 & 0.898 & -0.236 & 0.021 & 0.000 \\
- & 1.0 & 0.65 & 2 & -0.104 & 0.830 & 8 & -1.186 & 1.224 & 0.172 & -0.420 & 0.112 \\
- & 1.0 & 0.65 & 3 & -0.104 & 1.000 & 8 & -1.961 & 2.217 & -0.720 & 0.008 & 0.026 \\
- & 1.0 & 0.65 & 4 & -0.104 & 1.060 & 8 & -2.232 & 2.584 & -1.036 & 0.148 & 0.000 \\
- & 1.0 & 0.65 & 5 & -0.104 & 1.072 & 8 & -2.385 & 2.718 & -1.088 & 0.153 & 0.000 \\
- & 1.0 & 0.65 & 10 & -0.104 & 1.057 & 8 & -2.594 & 2.863 & -1.120 & 0.150 & 0.000 \\
- & 1.0 & 0.65 & 20 & -0.104 & 1.037 & 8 & -2.646 & 2.882 & -1.111 & 0.146 & 0.000 \\
- & 1.0 & 0.65 & 40 & -0.104 & 1.026 & 8 & -2.658 & 2.880 & -1.105 & 0.144 & 0.000
\end{tabular}
\caption{This table lists the parameters used in the fitting function to generate the results shown in the figures within the main text. The parameters are identified by the figure in which the binding potential is used, the title of that curve in the figure and some additional identification where applicable. Values are rounded to three decimal places and the value of $a_1$ is enforced.}
\end{table*}
\section{Particle interaction potentials}\label{app:appB}
The net interaction potential between a fluid particle and a planar wall made of particles interacting with the fluid particle via a Lennard-Jones-like potential, with the same form as in Eq.~\eqref{eq:pair_pot} is
\begin{equation}\label{eq:external}
V_{\bf i} = \epsilon_{wf} \sum_{i'=-\infty}^0 \sum_{j'=-\infty}^\infty \sum_{k'=-\infty}^\infty ((i-i')^2 + j'^2 + k'^2)^{-3},
\end{equation}
where we assume the surface of the wall is located in the plane $i=0$. The parameter $\epsilon_{wf}$ characterises the strength of the interaction between a single wall particle and a fluid particle. The triple sum above spans all of the particles in the wall; i.e.\ the wall is modelled as being discretised on a lattice just as the fluid is and that a wall particle is present on every lattice site where $i \leq 0$. The sums in Eq.\ \eqref{eq:external} can be simplified to obtain
\begin{equation}\label{eq:netInteraction}
V_{\bf i} =
\begin{cases}
-\epsilon_w/i^3 & \mbox{ for } i \geq 1 \\
\infty & \mbox{ for } i < 1
\end{cases}
\end{equation}
where the parameter $\epsilon_w$ determines the net strength of attraction to the wall.
The (2D) effective fluid-fluid particle interactions are governed in a similar manner by the potential
\begin{equation}\label{eq:weights}
\epsilon_{{\bf i},{\bf j}} = \epsilon \left[ (i'^2 + j'^2)^{-3} + 2 \sum_{k'=1}^\infty (i'^2+j'^2+k'^2)^{-3} \right].
\end{equation}
The parameter $\epsilon$ is the strength of a single Lennard-Jones pair interaction between two fluid particles and $\epsilon_{{\bf i},{\bf j}}$ is the weighted interaction between two particles taking into account all of the interactions in the invariant $k$ dimension. In the right hand side of Eq.\,\eqref{eq:weights}, $i'$, $j'$, and $k'$ are the distances between a pair of lattice sites in the $i$, $j$ and $k$ directions respectively.
Due to the fact that calculating long ranged particle interactions can be time consuming, making computer calculations rather slow, it is often necessary to truncate the interactions to some interaction range $L \sigma$. When particle interactions are truncated at a range of $L \sigma$ the external potential Eq.\,\eqref{eq:external} becomes
\begin{widetext}
\begin{equation}\label{eq:truncExternal}
V_{\bf i} = \epsilon_{wf} \sum_{i'=-L}^0 \sum_{j'=-L}^L \sum_{k'=-L}^L
\begin{cases}
((i-i')^2 + j'^2 + k'^2)^{-3} & \text{ for } ((i-i')^2 + j'^2 + k'^2)\leq L^2 \\
0 & \text{ otherwise}
\end{cases}
\end{equation}
and the fluid-fluid particle interaction weights \eqref{eq:weights} are then given as
\begin{equation}\label{eq:truncWeights}
\epsilon_{{\bf i},{\bf j}} =
\begin{cases}
\epsilon \left[ (i'^2 + j'^2)^{-3} + 2 \sum\limits_{k'=1}^{k' \leq \sqrt{L^2-i'^2-j'^2}} (i'^2+j'^2+k'^2)^{-3} \right] & \text{ for } |{\bf i}-{\bf j}|\leq L \\
0 & \text{ otherwise. }
\end{cases}
\end{equation}
\end{widetext}
\end{appendix}
|
\section{Introduction}
The \textit{maximum genus} $\gamma_M(G)$ of a graph $G$ is the
maximum integer $g$ such that $G$ has a cellular embedding in
the orientable surface of genus $g$. A result of Duke
\cite{duke} implies that a graph $G$ has a cellular embedding
in the orientable surface of genus $g$ if and only if
$\gamma(G) \le g \le \gamma_M(G)$ where $\gamma(G)$ denotes the
(minimum) genus of $G$. The problem of determining the set of
genera of orientable surfaces upon which $G$ can be embedded
thus reduces to calculation of $\gamma(G)$ and $\gamma_M(G)$.
Computing the minimum genus of a graph is a notoriously
difficult problem, which is known to be NP-complete even for
cubic graphs (see \cite{Th,T2}). Nevertheless, the minimum
genus can be calculated in linear time for graphs with bounded
genus or bounded treewidth by \cite{KMR:2008}. Moreover, for
graphs with fixed treewidth and bounded maximum degree
\cite{gross:2014} provides a polynomial-time algorithm
obtaining the complete genus distribution $\{g_i\}$ of
the graph $G$, where $g_i$ denotes the number of cellular
embeddings of $G$ into the orientable surface of genus $i$. For
graphs of bounded maximum degree \cite{CS:2013} has recently
proposed a polynomial-time algorithm constructing an embedding
with genus at most $O(\gamma(G)^{c_1}\log^{c_2} n)$ where $c_1$
and $c_2$ are constants. On the other hand, for every
$\epsilon>0$ and every function $f(n) = O(n^{1-\epsilon})$
there is no polynomial-time algorithm that constructs an
embedding of any graph $G$ with $n$ vertices into the surface
of genus at most $\gamma(G) + f(n)$ unless P $=$ NP (see
\cite{CKK,CKK2}).
For maximum genus, the situation is quite different, as maximum
genus admits a good (min-max) characterisation by Xuong's and
Nebesk\'y's theorems, see \cite{KOK,X} and \cite{KG,N},
respectively. From among these results the best known is
Xuong's theorem stating that $\gamma_M(G) = (\beta(G) - \min_T
\mathrm{odd}(G-E(T)))/2$, where $\beta(G)$ is the cycle rank of $G$,
$\mathrm{odd}(G-E(T))$ is the number of components of $G-E(T)$ with an
odd number of edges, and the minimum is taken over all spanning
trees $T$ of $G$. Building on these results, Furst
et~al.~\cite{FGM} and Glukhov \cite{G} independently devised
polynomial-time algorithms for determining the maximum genus of
an arbitrary graph. The algorithm of \cite{FGM} uses Xuong's
characterisation of maximum genus and exploits a reduction to
the linear matroid parity on an auxiliary graph; its running
time is bounded by $O(mn\Delta\log^6m)$, where $n,m$, and
$\Delta$ are the number of vertices, edges, and the maximum
degree of the graph, respectively. A matroidal structure is
also in the backgroung of the algorithm derived in~\cite{G},
albeit in a different way. Starting with any spanning tree $T$
of $G$, the algorithm greedily finds a sequence of graphs
$F_i$ such that $T = F_0 \subseteq \cdots \subseteq F_n
\subseteq G$, $|E(F_{i+1}) - E(F_i)| = 2$, and $\gamma_M(F_i) =
i$ for all $i$, and $\gamma_M(F_n) = \gamma_M(G)$. The running
time of this algorithm is bounded by $O(m^6)$.
Although two polynomial-time algorithms for the maximum genus
problem are known, both are relatively complicated. It is
therefore desirable to have a simpler way to determine the
maximum genus, at least approximately. A greedy approximation
algorithm for the maximum genus of a graph was proposed by Chen
\cite{chen}. The algorithm has two main phases. First, it
modifies a given graph $G$ into a $3$-regular graph $H$ by
vertex splitting, chooses an arbitrary spanning tree $T$ of
$H$, and finds a set $P$ of disjoint pairs of adjacent edges
in $H-E(T)$ with the maximum possible size. Second, it
constructs a single-face embedding of $T\cup P$ and then
inserts the remaining edges into the embedding while trying to
raise the genus as much as possible. A high-genus embedding of
$G$ in the same surface is then constructed by contracting the
edges created by vertex splitting. The algorithm constructs an
embedding of $G$ with genus at least $\gamma_M(G)/4$ and its
running time $O(m\log n)$ is dominated by the second phase,
that is, by operations on an embedded spanning subgraph of $H$.
In this paper we show that there is a much simpler way to
approximate maximum genus. Our algorithm repeatedly removes
arbitrary pairs of adjacent edges from $G$ while keeping the
graph connected. We prove that this simple idea leads to at
least $\gamma_M(G)/2$ pairs removed, providing an algorithm
that returns an integer $k$ such that $\gamma_M(G)/2 \le k \le
\gamma_M(G)$. This process can be implemented with running time
$O(m^2\log^2n/(n\log\log n))$. The algorithms developed in
\cite{chen} can then be used to efficiently construct an
embedding with genus $k$. Our result provides the first
method to approximate maximum genus that can be easily
implemented and improves the previous more complicated
algorithm of Chen~\cite{chen}, which can guarantee
embedding with genus only $\gamma_M(G)/4$. Structurally,
our approach yields a natural $2$-approximate counterpart of
Xuong's theorem.
\section{Background}
In this section we present definitions and results that provide
the background for our algorithm.
Our terminology is standard and consistent with \cite{MT}.
By a graph we mean a finite undirected graph with loops and
parallel edges permitted. Throughout, all embeddings into
surfaces are cellular, forcing our graphs to be connected,
and the surfaces are orientable. For more details and the
necessary background we refer the reader to \cite{GT} or
\cite{MT}; a recent survey of maximum genus can be found in
\cite[Chapter~2]{topics}.
One of the earliest results on embeddings of graphs is the
following observation, which is sometimes called Ringeisen's
edge-addition lemma. Although it is implicit in \cite{NSW},
Ringeisen \cite{ringeisen:1972} was perhaps the first to draw
an explicit attention to it.
\begin{lemma}\label{lemma:edge-addition-technique}
Let $\Pi$ be an embedding of a connected graph $G$ and let $e$
be an edge not contained in $G$, but incident with vertices in
$G$
\begin{itemize}
\item[{\rm (i)}] If both ends of $e$ are inserted into the
same face of $\Pi$, then this face splits into two
faces of the extended embedding of $G+e$ and the genus
does not change.
\item[{\rm (ii)}] If the ends of $e$ are inserted into two
distinct faces of $\Pi$, then in the extended embedding
of $G+e$ these faces are merged into one and the genus
raises by one.
\end{itemize}
\end{lemma}
The next lemma, independently obtained in \cite{KOK},
\cite{jungerman:1978}, and \cite{X}, constitutes the
cornerstone of proofs of Xuong's theorem. It follows easily
from Lemma~\ref{lemma:edge-addition-technique}.
\begin{lemma}\label{lemma:adding-pairs-single-face}
Let $G$ be a connected graph and $\{e,f\}$ a pair of adjacent
edges not contained in $G$, but incident with vertices in $G$.
If $G$ has an embedding with a single face, then so does
$G\cup\{e,f\}$.
\end{lemma}
Recall that by Xuong's theorem $\gamma_M(G) = (\beta(G) -
\min_T \mathrm{odd}(G-E(T)))/2$, where $\mathrm{odd}(G-E(T))$ is the number of
components of the cotree $G-E(T)$ with an odd number of edges.
It is not difficult to see that every cotree component with an
even number of edges can be partitioned into pairs of adjacent
edges, and that every cotree component with an odd number of
edges can be partitioned into pairs of adjacent edges and one
unpaired edge. Therefore, any spanning tree $S$ minimising
$\mathrm{odd}(G-E(T))$ maximises the number of pairs in the above
partition of the cotree. The proof strategy of Xuong's theorem
can now be summarised as follows. First, embed $S$ in the
$2$-sphere arbitrarily. Then repeatedly apply
Lemma~\ref{lemma:adding-pairs-single-face} to pairs obtained
from the partition of the components of $G-E(S)$, each time
rising the genus by one. Finally, add the remaining edges.
Lemma~\ref{lemma:edge-addition-technique} guarantees that the
addition cannot lower the genus. The result of this process is
an embedding of $G$ with genus at least $(\beta(G) - \min_T
\mathrm{odd}(G-E(T)))/2$.
The fact that a spanning tree minimising $\mathrm{odd}(G-E(T))$
maximises the number of pairs of adjacent edges in the cotree
suggests a slightly different combinatorial characterisation of
maximum genus. It is due to Khomenko et al.~\cite{KOK} and in
fact is older than Xuong's theorem itself.
\begin{theorem}\label{thm:KOK}
The maximum genus of a connected graph equals the maximum
number of disjoint pairs of adjacent edges whose removal leaves
a connected graph.
\end{theorem}
The following useful lemma, found for example in
\cite{CKK}, is an extension of
Lemma~\ref{lemma:adding-pairs-single-face} to embeddings with
more than one face. It can either be proved directly by using
Ringeisen's edge-adding technique or can be derived from
Xuongs's theorem. We may note in passing that this lemma was
used in \cite{CKK} to devise an algorithm that constructs an
embedding of genus $\gamma_M(G)-1$ whenever such an embedding
exists.
\begin{lemma}\label{lemma:adding-pairs}
Let $G$ be a connected graph and $\{e,f\}$ a pair of adjacent
edges not contained in $G$, but incident with vertices in $G$.
Then $\gamma_M(G\cup\{e,f\})\ge \gamma_M(G) + 1$.
\end{lemma}
Our main observation is that
Lemma~\ref{lemma:adding-pairs} can be applied to sets of pairs
of adjacent edges which do not necessarily have the
maximum possible size. Indeed, if we find any $k$ pairs
of adjacent edges $(e_i,f_i)_{i=1}^{k}$ in a graph $G$ such
that $G-\bigcup_{i=1}^{k} \{e_i,f_i\}$ is connected, then by
Lemma~\ref{lemma:adding-pairs} we can assert that the maximum
genus of $G$ is at least $k$. This suggests that identifying a
large number of pairs of adjacent edges whose removal leaves a
connected subgraph can be utilised to obtain a
simple approximation algorithm for the maximum genus.
Indeed, in the following section we show that choosing
the pairs of adjacent edges arbitrarily yields an
effective approximation of maximum genus.
\section{Algorithm}
In this section we present a greedy algorithm for finding at
least $\gamma_M(G)/2$ pairs of adjacent edges while the rest of
the graph remains connected. The idea is simple: if the removal
of a pair of adjacent edges does not disconnect the graph, then
we remove it.
To prove that the set output by Greedy-Max-Genus Algorithm
always contains at least $\gamma_M(G)/2$ pairs of adjacent
edges we employ the following lemma, which can be easily proved
either using Xuong's theorem or directly from
Lemma~\ref{lemma:edge-addition-technique}.
\begin{lemma}\label{lemma:edge-removal}
Let $G$ be a connected graph and let $e$ be an arbitrary edge
of $G$ such that $G-e$ is connected. Then
$$\gamma_M(G)-1\le\gamma_M(G-e)\le \gamma_M(G).$$
\end{lemma}
\vbox{
\hrule
\vspace*{1mm}
Greedy-Max-Genus Algorithm
\hrule
\vspace*{1mm}
\noindent
Input: Connected graph $G$ \\
Output: Set $P$ of paiwise disjoint pairs of adjacent edges of $G$ such that $G-P$ is \\
\hspace*{8mm} a connected spanning subgraph of $G$\\
\hspace*{4mm} 1: $H \leftarrow G$, $P \leftarrow \emptyset$\\
\hspace*{4mm} 2: {\bf repeat }\\
\hspace*{4mm} 3: \qquad choose adjacent edges $e, f$ from $H$ \\
\hspace*{4mm} 4: \qquad {\bf if} $H - \{e,f\}$ is connected \\
\hspace*{4mm} 5: \qquad \qquad $H \leftarrow H - \{e,f\}$\\
\hspace*{4mm} 6: \qquad \qquad $P \leftarrow P\cup (e,f)$ \\
\hspace*{4mm} 7: {\bf until} all pairs of adjacent edges of $H$ have been tested\\
\hspace*{4mm} 8: return $P$
\vspace*{1mm}
\hrule
}
The final ingredient for our Greedy-Max-Genus Algorithm is the
following characterisation of graphs with maximum genus $0$.
\begin{theorem}\label{thm:mgZero}
The following statements are equivalent for every
connected graph $G$.
\begin{itemize}
\item[{\rm (i)}] $\gamma_M(G)=0$
\item[{\rm (ii)}] No two cycles of $G$ have a vertex in
common.
\item[{\rm (iii)}] $G$ contains no pair of adjacent edges
whose removal leaves a connected graph.
\end{itemize}
\end{theorem}
The equivalence (i) $\Leftrightarrow$ (ii) in
Theorem~\ref{thm:mgZero} was first proved by Nordhaus
et~al.~in \cite{NRSW}. The equivalence (ii) $\Leftrightarrow$
(iii) is easy to see, nevertheless it is its appropriate
combination with Lemma~\ref{lemma:edge-removal} which yields
the desired performance guarantee for Greedy-Max-Genus
algorithm, as shown in the following theorem.
\begin{theorem}\label{thm:alg}
For every connected graph $G$, the set of pairs output
by Greedy-Max-Genus Algorithm run on $G$ contains at least
$\gamma_M(G)/2$ pairs of adjacent edges.
\end{theorem}
\begin{proof}
Assume that the algorithm stops after the removal of $k$
disjoint pairs of adjacent edges from~$G$. For
$i\in\{0,1,\dots,k\}$ let $H_i$ denote the graph obtained from
$G$ by the removal of the the first $i$ pairs of edges.
By Lemma~\ref{lemma:edge-removal}, the removal of a single edge
from a graph can lower its maximum genus by at most one.
Therefore, the removal of two edges can lower the maximum genus
by at most two. It follows that $\gamma_M(H_i) \ge \gamma_M(G)
-2i$ for each $i$; in particular, $\gamma_M(H_k) \ge
\gamma_M(G) -2k$. From Theorem \ref{thm:mgZero} we get that
$\gamma_M(H_k) = 0$. By combining these expressions we get $2k
\ge \gamma_M(G)$, which yields $k\ge \gamma_M(G)/2$, as
desired.
\end{proof}
\begin{remark}
Let $n$ and $m$ denote the number of vertices and edges
of $G$, respectively, and let $k$ be the number of pairs of
adjacent edges produced by Greedy-Max-Genus Algorithm run on
$G$. An embedding of $G$ with genus at least $k$ can be
constructed from the set of pairs of adjacent edges in time
$O(n+k\log n)$, see the proof of Theorem 4.5 in \cite{chen}
for details.
\end{remark}
Observe that any maximal set of pairs of adjacent edges
of $G$ whose removal from $G$ yields a connected graph can be
output by Greedy-Max-Genus Algorithm run on $G$. Hence, as a
corollary of Theorem~\ref{thm:alg} we obtain the following
2-approximate counterpart of Xuong's theorem.
\begin{theorem}\label{thm:approx}
Let $G$ be a connected graph and let $P$ be any inclusion-wise
maximal set of disjoint pairs of adjacent edges of $G$ whose
removal leaves a connected subgraph. Then
$|P|\ge\gamma_M(G)/2$.
\end{theorem}
The following example shows that the bound of Theorem
\ref{thm:approx} is tight and Greedy-Max-Genus Algorithm can
output the value $\gamma_M(G)/2$ for infinitely many graphs
$G$.
\begin{example}
\label{ex:1}
Take the star $K_{1,2n}$ where $n$ is an arbitrary
positive integer, replace every edge with a~pair of parallel
edges, and add a loop to every vertex of degree $2$. Denote the
resulting graph by~$G_n$. Using Theorem~\ref{thm:KOK} it is
easy to see that $\gamma_M(G_n)= 2n$. Indeed, take a set
$P$ of $2n$ disjoint pairs of adjacent edges, each
consisting of a loop and one of its adjacent edges. Since the
edges of $G_n$ not in $P$ form a spanning tree, $P$ has maximum
size with respect to the property that $G-P$ is
connected. Thus $\gamma_M(G_n) = 2n$ by Theorem~\ref{thm:KOK}.
On the other hand, consider a set $P'$ of $n$ disjoint
pairs of adjacent edges that include only edges incident with
the central vertex. The removal of these pairs from $G_n$
leaves a spanning tree of $G_n$ with a loop attached to every
pendant vertex, so $P'$ is a maximal set of pairs for which
$G_n-P'$ is connected. Since $|P'|=n=\gamma_M(G_n)/2$, our
example confirms that the bound in Theorems~\ref{thm:alg}
and~\ref{thm:approx} is best possible.
\end{example}
Example \ref{ex:1} also implies that processing vertices in
the decreasing order with respect to their degrees does not
necessarily lead to a better performance of the algorithm.
\section{Implementation}
To implement the algorithm it is clearly sufficient to consider
all pairs of edges $\{e,f\}$ with a common end-vertex and test
whether removing the pair does not disconnect the graph.
The running time is thus
$O((\tau+\rho)\sum_{i=1}^n d_i^2)$, where $\tau$ is the time required to
test the connectivity, $\rho$ is the time required to update the
underlying data structure, and $d_i$ is the degree of the
$i$-th vertex.
If the input graph is simple, then $\sum_{i=1}^n d_i^2
= O(m^2/n)$ by \cite{caen:1998}. If the input graph is not
simple, we can preprocess it as follows. Let $P=\emptyset$.
From every set $F$ of pairwise parallel edges we repeatedly
remove pairs of edges and add them into $P$ until $F$
contains at most two parallel edges. Simirarly, from every set
$F$ of loops incident with a single vertex we remove pairs of
loops and add them into $P$ until $F$ contains at most one
loop.
Let $G'$ denote the resulting graph.
Finally, the set $P$ of pairs of adjacent edges is added
to the set $P'$ produced by Greedy-Max-Genus Algorithm on $G'$.
It can be easily seen that in $G'$ the sum
of squares of the degrees is again in $O(m^2/n)$ and that the
preprocessing phase can be done in $O(m)$, where $m$ and $n$ is the number of edges and vertices of the input graph. Regarding the
testing of connectivity, we have $\tau=O(m)$ for instance by using
DFS, in which case there is no need for additional updates of data
structure and thus $\rho=O(m)$. Therefore, we obtain an
implementation which reduces essentially to a series of
connectivity tests and has running time $O(m^3/n)$. Using the
dynamic graph algorithm for connectivity from \cite{WN:2013} it
is possible to support updates in $\rho = O(\log^2 n/\log\log n)$
amortized time and queries in $\tau = O(\log n/\log\log n)$
worst-case time. This yields the total running time
$O(m^2\log^2n/(n\log\log n))$.
\section{Discussion}
We have presented an approximation algorithm for the maximum
genus problem that for any connected graph $G$ outputs an
integer $k$ such that $\gamma_M(G)/2 \le k \le \gamma_M(G)$.
This result shows that the classical ideas dating back to
Norhaus et al.~\cite{NRSW} and Ringeisen \cite{ringeisen:1972},
and to characterisations of maximum genus by Xuong \cite{X} and
Khomenko et al.~\cite{KOK}, can be used also for efficient
approximation of maximum genus. Our algorithm is much simpler
than both the $4$-approximation algorithm of Chen
\cite{chen} and the precise polynomial-time algorithms for the
maximum genus problem of Furst et al. \cite{FGM} and
Glukhov~\cite{G}, outperforms the existing
$4$-approximation algorithm \cite{chen}, and provides the first
approximation of maximum genus that can be easily implemented.
On the structural side, we have obtained a natural
$2$-approimation counterpart of Xuong's theorem.
\section*{Acknowledgement}
The first author was partially supported by Ministry of
Education, Youth, and Sport of Czech Republic, Project
No.~CZ.1.07/\-2.3.00/\-30.0009. The second author was
partially supported by VEGA grant 1/0474/15. The authors would
like to thank Rastislav Kr\'alovi\v{c} and Jana
Vi\v{s}\v{n}ovsk\'a for reading preliminary versions of this
paper and making useful suggestions.
\scriptsize
\bibliographystyle{plain}
|
\section{Introduction}
\label{sec:introduction}
In 1881, G.~Darboux determined all one-parametric spatial motions with only planar curves as point paths
\cite{darboux81}. These properties determine a unique motion which is nowadays named after Darboux.
It is the composition of a planar elliptic motion with a suitably oscillating translation
perpendicular to that plane. If the elliptic motion degenerates to a rotation, the Darboux motion
is called \emph{vertical,} if this is not the case, we call the Darboux motion \emph{non-vertical}
or \emph{general.} In both cases, the generic point paths are ellipses.
The Darboux motion is a special Schoenfließ motion and we call the direction perpendicular to the planar elliptic motion component the \emph{axis direction.}
These and other properties are well known in literature, see for example \cite{blaschke58} or
\cite[Chapter~9, \S3]{bottema90}.
In this article, we construct 7R linkages with one degree of freedom such that one
link during (one component of) the motion performs a general Darboux motion.
These results complement recently presented linkages that generate vertical Darboux
motions \cite{lee12}. The beginnings of this research can be found in
\cite{li14:_spatial_straight_line_linkages} where we investigated spatial
linkages with a straight line trajectory. The linkages we present in this article are novel.
Unlike the linkages of \cite{lee12}, we generate a general (non-vertical) Darboux motion
and our linkages are free of prismatic or cylindrical joints. One of our linkages has already be
found in \cite{li14:_spatial_straight_line_linkages}. Our description here is more
general and detailed.
Our linkage construction is based on the factorization of rational motions
which has been first introduced in \cite{hegedus13:_factorization2}.
In the dual quaternion model of rigid body displacements, the Darboux motion
is parameterized by a cubic motion polynomial. In general, a cubic motion can
be generated by 6R linkages \cite{hegedus14:_four_pose_synthesis,li13:_anglesymmetric,li14:_ck}
but specialties of the Darboux motion prevent application of the
factorization algorithm for generic motion polynomials. Therefore,
we resort to new factorization techniques for non-generic motion polynomials
\cite{gallet15,li15:_factorization,li15:_minimal_degree}, i.e., we
multiply the motion polynomial with a real
polynomial which
divides the primal part of the motion polynomial. The new motion polynomial can be factored in different ways which allows us to obtain new closed 7R linkages
by combining different factorizations. A unified algorithm for factorization of
generic and non-generic motion polynomials
is introduced in \cite{li15:_factorization}. Here, we just mention the basic idea of this algorithm: By factoring out linear motion polynomial factors (from the left and from the right) we create a generic motion polynomial that can be factored by the original algorithm of \cite{hegedus13:_factorization2}.
We continue this text with a quick introduction to the dual quaternion model
of $\SE$, the group of rigid body displacements, and to motion polynomials
and their factorization (Section~\ref{sec:preliminaries}).
In Section~\ref{sec:factorizations} we describe the Darboux motion in the dual quaternion model,
present different factorizations, and combine them to form 7R linkages.
\section{Preliminaries}
\label{sec:preliminaries}
The investigations in this article are based on the dual quaternion model of the
group $\SE$ of rigid body displacements. We assume familiarity with dual
quaternions and their relation to kinematics but we provide a quick introduction
of concepts that cannot be found in standard textbooks. For more information on
basics of dual quaternions and kinematics we refer to
\cite{mccarthy90:_introduction_theoretical_kinematics, selig05:_geometric_fundamentals_robotics, husty12:_kinematics_algebraic_geometry}. A short presentation of the topic
with close relations to this article can also be found in \cite{li15:_survey}.
\subsection{Dual quaternions}
A dual quaternion is an object of the shape
\begin{equation}
\label{eq:dualquat}
h = h_0 + h_1{\mathbf{i}} + h_2{\mathbf{j}} + h_3{\mathbf{k}} + \varepsilon(h_4 + h_5{\mathbf{i}} + h_6{\mathbf{j}} + h_7{\mathbf{k}})
\end{equation}
with real numbers $h_0,\ldots,h_7$. The associative but non-commutative
multiplication of dual quaternions is defined via the relations
\begin{gather*}
{\mathbf{i}}^2 = {\mathbf{j}}^2 = {\mathbf{k}}^2 = {\mathbf{i}}{\mathbf{j}}{\mathbf{k}} = -1,
\quad
\varepsilon^2 = 0,\\
{\mathbf{i}}\varepsilon = \varepsilon{\mathbf{i}},\quad
{\mathbf{j}}\varepsilon = \varepsilon{\mathbf{j}},\quad
{\mathbf{k}}\varepsilon = \varepsilon{\mathbf{k}}.
\end{gather*}
For example
\begin{equation*}
\begin{aligned}
(1 - \varepsilon{\mathbf{i}})({\mathbf{j}} + \varepsilon{\mathbf{k}}) &= {\mathbf{j}} + \varepsilon{\mathbf{k}} - \varepsilon{\mathbf{i}}{\mathbf{j}} - \varepsilon^2{\mathbf{i}}{\mathbf{k}} \\
&= {\mathbf{j}} + \varepsilon{\mathbf{k}} - \varepsilon{\mathbf{k}} - 0 = {\mathbf{j}}.
\end{aligned}
\end{equation*}
Conjugate dual quaternion and dual quaternion norm are
\begin{gather*}
\cj{h} = h_0 - h_1{\mathbf{i}} - h_2{\mathbf{j}} - h_3{\mathbf{k}} + \varepsilon(h_4 - h_5{\mathbf{i}} - h_6{\mathbf{j}} - h_7{\mathbf{k}}),\\
\Norm{h} = h\cj{h}.
\end{gather*}
The primal part of $h$ is $p = h_0 + h_1{\mathbf{i}} + h_2{\mathbf{j}} + h_3{\mathbf{k}}$ and the
dual part is $d = h_4 + h_5{\mathbf{i}} + h_6{\mathbf{j}} + h_7{\mathbf{k}}$. Primal and dual
part are (ordinary) quaternions. The set of quaternions is denoted by
$\H$, the set of dual quaternions by $\DH:=\mathbb{D}\otimes_{\mathbb{R}}\H$. Writing $h = p + \varepsilon d$
with $p, d \in \H$, we have $\Norm{h} = p\cj{p} + \varepsilon(p\cj{q}+q\cj{p})$.
Note that both $p\cj{p}$ and $p\cj{q}+q\cj{p}$ are real numbers and the
norm itself is a \emph{dual number.}
If the dual quaternion $h = p + \varepsilon d$ has
a real norm
($p\cj{q}+q\cj{p} = 0$), its action on a point $[x_0,x_1,x_2,x_3]$ of
real projective three-space is defined via
\begin{equation}
\label{eq:action}
x \mapsto y = \frac{px\cj{p}+p\cj{q}-q\cj{p}}{p\cj{p}}
\end{equation}
where $x = x_0 + x_1{\mathbf{i}} + x_2{\mathbf{j}} + x_3{\mathbf{k}}$,
$y = y_0 + y_1{\mathbf{i}} + y_2{\mathbf{j}} + y_3{\mathbf{k}}$ and the image point has projective
coordinates $[y_0,y_1,y_2,y_3]$. The map \eqref{eq:action} describes
a rigid body displacement and the composition of such displacements corresponds
to dual quaternion multiplication. Note that $h$ and $\lambda h$ with
$\lambda \in {\mathbb{R}} \setminus \{0\}$ describe the same map.
The dual quaternion $h$ as in \eqref{eq:dualquat} describes a rotation
or translation, if and only if $h_4 = 0$. Among these, translations
are characterized by $h_1 = h_2 = h_3 = 0$. We speak
accordingly of \emph{rotation} or \emph{translation quaternions.}
In the former case, the revolute axis has Plücker coordinates $[h_1,h_2,h_3,-h_5,-h_6,-h_7]$.
In particular, the axis direction is $(h_1,h_2,h_3)$.
\subsection{Motion polynomials}
Now we study polynomials
\begin{equation}
\label{eq:polynomial}
C = c_nt^n + \cdots + c_1t + c_0
\end{equation}
in the indeterminate $t$ and with dual quaternion coefficients
$c_n,\ldots,c_0 \in \DH$. The set of these quaternions is denoted by $\DH[t]$.
Multiplication of polynomials in $\DH[t]$ is defined by the convention that
indeterminate $t$ and coefficients commute. The conjugate polynomial $\cj{C}$ is
the polynomial with conjugate coefficients $\cj{c_n},\ldots,\cj{c_0}$.
A polynomial $C \in \DH[t]$ is called a \emph{motion polynomial,} if its
leading coefficient $c_n$ is invertible (has non-zero primal part) and the
so-called norm polynomial $C\cj{C}$ has real coefficients. In this case,
it parameterizes a rational motion via the map \eqref{eq:action}. Note that
$C$ and $QC$ parameterize the same rational motion if $Q$ is a real polynomial
without zeros. This observation will be important in the next section.
The linear polynomial $C = t - h$ with $h$ as in \eqref{eq:dualquat} is
a motion polynomial if and only if $h_4 = h_1h_5 + h_2h_6 + h_3h_7 = 0$.
Assuming $(h_1,h_2,h_3) \neq (0,0,0)$, the displacements described by $C$ for
varying $t$ are rotations about a fixed axis but with different revolute angle.
The axis' Plücker coordinates are $[h_1,h_2,h_3,-h_5,-h_6,-h_7]$ and $(h_1,h_2,h_3)$
is a direction vector of the axis.
A generic motion polynomial $C$ of degree $n$ admits $n!$ factorizations of the shape
\begin{equation}
\label{eq:factorization}
C = (t - h_1) \cdots (t - h_n)
\end{equation}
with rotation quaternions $h_1,\ldots,h_n$. The precise statement and a
factorization algorithm can be found in \cite{hegedus13:_factorization2}.
Here is suffices to say that a Darboux motion is not generic and the general
theory of \cite{hegedus13:_factorization2} gives no information about existence
or non-existence of factorizations.
Equation~\eqref{eq:factorization} admits an important kinematic interpretation:
Each factorization corresponds to a one-parametric motion of an open chain of $n$
revolute joints whose end-effector follows the motion parameterized by~$C$. The
revolute axes are determined by the rotation quaternions $h_1,\ldots,h_n$. Different
factorizations may be combined to form closed loop linkages with this coupler motion.
In the remainder of this text we need auxiliary results related to the
factorization of motion polynomials. Their proofs can be found in
\cite{hegedus13:_factorization2}:
\begin{lemma}[Polynomial division]
\label{lem:division}
If $C$ and $D$ are polynomials in $\DH[t]$ and the leading
coefficient of $D$ is $1$, there exist polynomials $Q,R \in \DH[t]$
such that $C = QD + R$ and $\deg R < \deg D$.
\end{lemma}
Note that the computation of quotient $Q$ and remainder $R$ is possible by
straightforward polynomial long division in a non-commutative setting.
\begin{lemma}[Zeros and right factors]
\label{lem:zero}
The dual quaternion $h$ is a zero of the polynomial $C \in \DH[t]$ if and only
if there exists a polynomial $Q \in \DH[t]$ such that $C = Q(t-h)$.
\end{lemma}
The statement of Lemma~\ref{lem:zero} is less trivial than the corresponding
statement for real polynomials. In particular, with $C$ as in \eqref{eq:polynomial},
the value of $C(h)$ (and hence also the term ``zero of $C$'') is \emph{defined} as $c_nh^n + \cdots + c_1h + c_0$. It is not
admissible to simply plug $h$ in a factorized representation of the shape \eqref{eq:factorization}.
With $C=(t-k)(t-h)$ and $h,k \in \DH[t]$ we have, for example, $C = t^2 - (h + k)t + kh$ and thus
\begin{equation*}
C(h) = -kh + kh = 0
\quad\text{but}\quad
C(k) = hk + kh \neq 0.
\end{equation*}
Lemma~\ref{lem:zero} asserts that the ``usual'' relation between zeros linear
factors holds at least for \emph{right factors.}
\section{Factorizations of the Darboux motion}
\label{sec:factorizations}
Following \cite[Equation~(3.4)]{bottema90}, the general Darboux motion is given
by the parametric equations
\begin{equation*}
\begin{aligned}
X &= x \cos\varphi - y \sin\varphi \\
Y &= x \sin\varphi + y \cos\varphi + a \sin\varphi \\
Z &= z + b \sin\varphi + c(1 - \cos\varphi).
\end{aligned}
\end{equation*}
Here, $\varphi \in [-\pi,\pi)$ is the motion parameter, $a$, $b$, $c$ are real
constants and $x$, $y$, $z$ resp. $X$, $Y$, $Z$ are coordinates in the moving and
the fixed frame. For $a = 0$, the motion is a vertical Darboux motion. In this paper
we exclude this case and assume $a \neq 0$.
Using the conversion formulas between matrix and dual quaternion representation of
rigid body displacements (see for example \cite{husty12:_kinematics_algebraic_geometry}),
we find the polynomial parameterization
\begin{multline}
C_0 = ({\mathbf{k}} + c\varepsilon) t^3+(1 - \varepsilon(a{\mathbf{i}} - c{\mathbf{k}} + b)) t^2\\
+ ({\mathbf{k}} - \varepsilon(a{\mathbf{j}} + b{\mathbf{k}}))t + 1
\end{multline}
for the Darboux motion. Left-dividing this motion polynomial by the leading
coefficient ${\mathbf{k}} + c\varepsilon$ (this amounts to a coordinate change in the fixed frame), we arrive at
\begin{equation}\label{dbmotion1}
C = t^3 - ({\mathbf{k}} - \varepsilon(a{\mathbf{j}} - b{\mathbf{k}}))t^2 + (1 - \varepsilon(b + a{\mathbf{i}} - c{\mathbf{k}}))t - {\mathbf{k}} +c\varepsilon.
\end{equation}
Our further investigations will be based on this parametric representation of the Darboux motion.
Note that the primal part of $C$ is $(t^2+1)(t - {\mathbf{k}})$. The presence of a real
factor violates the assumptions of \cite[Theorem~1]{hegedus13:_factorization2}
(existence of factorizations) and prevents application of Algorithm~1
(computation of factorizations) of that paper. Nonetheless, we will see below
that $C$ admits
infinitely many factorizations with three linear motion polynomials.
Moreover, suitable polynomial multiples of $C$ admit factorizations as well and parameterize
the same Darboux motion. This we will use for our construction of 7R Darboux linkages.
In \cite{li14:_spatial_straight_line_linkages}, the authors gave two types of
factorizations for the motion $C$ in Equation~\eqref{dbmotion1}. The first type is with three linear
motion polynomials (FI), the second type (FII) is with five linear motion polynomials
whose product does not equal $C$ but $C$ times a real quadratic polynomial $P$. This
makes an algebraic difference but does not change the kinematics: Both $C$ and $PC$
parameterize the same motion. Moreover, two successive linear factors in FII are identical.
Hence, it gives rise to an open 4R chain which can generate the Darboux motion. In combination with
the 3R chain obtained from FI we obtain a 7R linkage.
In this section, we recall the construction of \cite{li14:_spatial_straight_line_linkages}
and we give two further factorizations (FIII and FIV) of \eqref{dbmotion1}. Similar to FII,
these factorizations also use five linear motion polynomials but give only four axes because
two successive linear motion polynomials are same. Combining it with the already known 3R chain,
we obtain further 7R Darboux linkages.
\subsection{First factorization FI}
We want to find linear motion polynomials $Q_1$, $Q_2$, $Q_3$ such that
$C = Q_1Q_2Q_3$. By Lemma~\ref{lem:zero}, a necessary and sufficient condition
for $Q_3 = t - q_3$, $q_3 \in \DH$, to be a right factor of $C$ is $C(q_3) = 0$.
A straightforward calculation shows that $q_3$ is necessarily of the shape
\begin{equation*}
q_3 = {\mathbf{k}} + \varepsilon(v{\mathbf{i}} + w{\mathbf{j}}), \quad v, w \in {\mathbb{R}}.
\end{equation*}
Next, we consider the quotient polynomial $Q = P + \varepsilon D$, defined by $C = QQ_3$.
Its primal part is $P = t^2+1$. Thus, the motion parameterized by $Q$ is a curvilinear
translation. Since $Q$ is quadratic and $P$ has no real zero, the trajectories are congruent
ellipses. This motion admits a factorization $Q = Q_1Q_2$ of the required shape only
if it is a \emph{circular translation} (see \cite{li15:_survey} for a detailed proof of this statement). This
puts
a condition on $q_3$, namely
\begin{equation}
\label{eq:circular-translation}
v = -a^{-1}bc,
\quad
w = - (2a)^{-1}(a^2+b^2-c^2).
\end{equation}
If \eqref{eq:circular-translation} is satisfied, the remainder $Q$ is a circular
translation and can be factored in infinitely many ways as $Q = Q_1Q_2$ with linear
rotation polynomials $Q_1$, $Q_2$. Each pair of these factorizations correspond to a
parallelogram linkage that generates the circular translation. One example of a
factorization of $C$ is
\begin{equation}\label{FI}
\begin{aligned}
Q_1 & = t + E - \frac{bc}{a} {\mathbf{i}} \varepsilon + \frac{a^2+c^2-b^2}{2a} {\mathbf{j}} \varepsilon - b {\mathbf{k}} \varepsilon , \\
Q_2 & = t - E, \\
Q_3 & = t - {\mathbf{k}} +\frac{bc}{a} {\mathbf{i}}\varepsilon + \frac{a^2+b^2-c^2}{2a} {\mathbf{j}}\varepsilon
\end{aligned}
\end{equation}
where
\begin{equation}
\label{eq:E}
E = \frac{2ac}{a^2+b^2+c^2} {\mathbf{i}} + \frac{2ab}{a^2+b^2+c^2} {\mathbf{j}} + \frac{a^2-b^2-c^2}{a^2+b^2+c^2} {\mathbf{k}}.
\end{equation}
In particular, we have shown
\begin{proposition}
The general Darboux motion parameterized by \eqref{dbmotion1} can be decomposed
into a rotation about an axis parallel to the Darboux motion's axis direction and a circular
translation in a plane orthogonal to the vector quaternion $D$ of \eqref{eq:E}.
\end{proposition}
Combining two of the infinitely many factorizations with three linear factors gives
a parallelogram linkage to whose coupler a dangling link (corresponding to the right
factor $Q_3$ which is the same for all factorizations) is attached. This 6R linkage
has two trivial degrees of freedom and is not particularly interesting.
\subsection{Second factorization FII}
Now we construct a further factorization FII of the general Darboux motion. We do,
however, alter the parameterization \eqref{dbmotion1} and multiply it with the quadratic
polynomial $P := t^2+1$. This changes the motion polynomial but not the parameterized
motion and gives additional freedom to find factorizations. In contrast to FI, we have more
than three linear factors.
We begin by setting the right factor to $Q_4 := t - {\mathbf{k}}$ and then use polynomial division
(Lemma~\ref{lem:division}) to compute
\begin{equation*}
C_1 = t^2 + \varepsilon(a{\mathbf{j}} - b{\mathbf{k}})t + 1 + c\varepsilon{\mathbf{k}}.
\end{equation*}
such that $C = C_1Q_4$. The polynomial $PC_1$ can be written as the product of four
linear motion polynomials. Again, there exist infinitely many factorizations. In order
to keep the number of joints small, we choose one with identical middle factors, i.e.,
$PC_1=Q_7 Q_6^2 Q_5,$ where
\begin{equation*}
\begin{aligned}
Q_7 & = t + {\mathbf{i}} + \frac{a+c}{2} \varepsilon{\mathbf{j}} - \frac{b}{2} \varepsilon{\mathbf{k}}, \\
Q_6 & = t - {\mathbf{i}}, \\
Q_5 & = t + {\mathbf{i}} + \frac{a-c}{2} \varepsilon{\mathbf{j}} - \frac{b}{2} \varepsilon{\mathbf{k}}.
\end{aligned}
\end{equation*}
The multiplicity of the middle factor $Q_6$ allows us to make a 3R chain for the
motion parameterized by $PC_1$. Together with $C = Q_1Q_2Q_3$ and $Q_4$, we can get
a 7R linkage. It can be seen that the axes of $Q_1$, $Q_2$ are parallel, as are the
axes of $Q_3,\ Q_4$ and $Q_5$, $Q_6$, $Q_7$.
For a concrete example, a Gröbner basis computation reveals that the configuration
set of this 7R linkage is one dimensional \cite{myweb7R}. Note that its configuration
curve consists of different components and only one corresponds to the Darboux motion
parameterized by~$C$.
We also want to mention that the construction of the second factorization FII
leaves many degrees of freedom. We just presented one factorization of this type.
\subsection{Third factorization FIII}
In order to find a third factorization FIII that is suitable for linkage construction,
we proceed in a similar fashion to case FII,
multiplying the motion polynomial $C$ of
Equation~\eqref{dbmotion1} with $P = t^2+1$. But we do so \emph{prior to splitting
off a right factor.} The polynomial $PC$ has infinitely many factorizations into
products of \emph{five} linear motion polynomials, one of them being of the shape
$CP=Q'_7 Q'_6Q'_5Q'^2_4.$
We start by finding a suitable factor $Q'_4$. Because $Q'_4\cj{Q'_4}$ is a factor of
$(CP)(\cj{CP}) = P^5$, we have the necessary constraint $Q'_4\cj{Q'_4} = P$ but
still have infinitely many choices. We take
\begin{equation*}
Q'_4 = t - {\mathbf{k}} - x {\mathbf{i}}\varepsilon - y {\mathbf{j}}\varepsilon
\end{equation*}
with $x, y \in {\mathbb{R}}$. After division of $PC$ by $(Q'_4)^2$ (Lemma~\ref{lem:division}),
we are left with the cubic motion polynomial $C_2$ defined by $CP = C_2Q'^2_4$, i.e.,
\begin{multline*}
C_2 = t^3 + ({\mathbf{k}} + \varepsilon(a{\mathbf{j}} - b{\mathbf{k}} + 2x{\mathbf{i}} + 2y{\mathbf{j}})) t^2 \\
+ (1 + \varepsilon(a{\mathbf{i}} + c{\mathbf{k}} + 2y{\mathbf{i}} - 2x{\mathbf{j}} + b))t + {\mathbf{k}} - c\varepsilon.
\end{multline*}
This new cubic motion polynomial $C_2$ parameterizes again a Darboux motion but the
parameterization is not in the standard form \eqref{dbmotion1}. Nonetheless, there
exist again infinitely many factorizations similar to \eqref{FI}. For example, we
have $C_2=Q'_7 Q'_6Q'_5$ where
\begin{equation*}
\begin{aligned}
Q'_5 &= t + {\mathbf{k}} + \Bigl(\frac{b^2x-abc-2bcy-c^2x}{T^2+4x^2}+x\Bigr){\mathbf{i}}\varepsilon \\
& \quad + \Bigl( \frac{ab^2-ac^2+2b^2y+4bcx-2c^2y}{2(T^2+4x^2)}+\frac{a}{2}+y \Bigr){\mathbf{j}}\varepsilon,\\
Q'_6 &= t - \frac{2(ac-2bx+2cy)}{b^2+c^2+T^2+4x^2}{\mathbf{i}} + \frac{2(ab+2by+2cx)}{b^2+c^2+T^2+4x^2}{\mathbf{j}}\\
& \quad + \frac{T^2-b^2-c^2+4x^2}{b^2+c^2+T^2+4x^2}{\mathbf{k}},\\
Q'_7 &=\cj{Q'_6} \phantom{+} \frac{Sx+bcT+4x(ay+x^2+y^2)}
{T^2+4x^2}{\mathbf{i}}\varepsilon \\
& \quad +
\frac{(S+4x^2)aT+4(ay(a(3y+a)+2y^2)-abcx)}
{2a(T^2+4x^2)}{\mathbf{j}}\varepsilon \\
& \quad - b{\mathbf{k}}\varepsilon,\\
\end{aligned}
\end{equation*}
and we abbreviated $S = a^2-b^2+c^2$ and $T = a + 2y$.
Because $C = Q_1Q_2Q_3$ and $PC =C_2Q'^2_4 = Q'_7Q'_6Q'_5Q'^2_4$ parameterize
the same motion, we can combine these two factorizations to form a 7R linkage
where each rotation is defined by $Q'_7$, $Q'_6$, $Q'_5$, $Q'_4$, $Q_3$, $Q_2$, $Q_1$.
It can be seen that the axes of $Q_1$, $Q_2$ are parallel, as are the axes of $Q_3$, $Q'_4$,
$Q_5$ and $Q'_7$, $Q'_6$.
As revealed by a Gröbner basis computation, the configuration space of a concrete numeric
example of this 7R linkage is really a curve \cite{myweb7R}. Thus, we have indeed constructed
another 7R linkage whose coupler motion is a non-vertical Darboux motion. Note that the
configuration curve contains several components, not all of them rational. One component
corresponds to the rational curve parameterized by~$C$.
In Figure~\ref{fig:darboux1}, we present nine configurations (the first one and the last
one are from same configuration) of this linkage in an orthographic projection parallel to~${\mathbf{k}}$.
\begin{figure*}
\centering
\begin{overpic}[width=0.49\textwidth,trim=0 32 0 45,clip]{example4/example4+0}
\put(70,20){$Q_1$}
\put(57,37){$Q_2$}
\put(26,42){$Q_3$}
\put(68,4){$Q_4$}
\put(50,4){$Q_5$}
\put(38,11){$Q_6$}
\put(20,35){$Q_7$}
\end{overpic} \includegraphics[width=0.49\textwidth,trim=0 32 0 45,clip]{example4/example4+1}\\
\includegraphics[width=0.49\textwidth,trim=0 28 0 28,clip]{example4/example4+2} \includegraphics[width=0.49\textwidth,trim=0 28 0 28,clip]{example4/example4+3}\\
\includegraphics[width=0.49\textwidth,trim=0 0 0 0,clip]{example4/example4+4} \includegraphics[width=0.49\textwidth,trim=0 0 0 0,clip]{example4/example4+5}\\
\includegraphics[width=0.49\textwidth,trim=0 17 0 18,clip]{example4/example4+6} \includegraphics[width=0.49\textwidth,trim=0 17 0 18,clip]{example4/example4+7}\\
\includegraphics[width=0.49\textwidth,trim=0 40 0 43,clip]{example4/example4+8} \includegraphics[width=0.49\textwidth,trim=0 40 0 43,clip]{example4/example4+0}
\caption{The first 7R linkage generates a non-vertical Darboux motion.}
\label{fig:darboux1}
\end{figure*}
\subsection{Fourth factorization FIV}
Finally, we present an exceptional case of the 7R linkage obtained
from the third factorization FIII. Among the infinitely many choices
for $a$, $b$, $c$, $x$, $y$ we choose the special case $a=1$, $b=2$,
$c=0$, $x=0$, $y=0$. We then have
\begin{equation*}
\begin{aligned}
Q''_1 & = t + \frac{4}{5} {\mathbf{j}} - \frac{3}{5} {\mathbf{k}}
- \frac{3}{2} {\mathbf{j}}\varepsilon -2 {\mathbf{k}}\varepsilon, \\
Q''_2 & = t - \frac{4}{5} {\mathbf{j}} + \frac{3}{5} {\mathbf{k}}, \\
Q''_3 & = t - {\mathbf{k}} + \frac{5}{2}{\mathbf{j}}\varepsilon,\\
Q''_4 & = t - {\mathbf{k}}, \\
Q''_5 & = t + {\mathbf{k}} + \frac{5}{2}{\mathbf{j}}\varepsilon,\\
Q''_6 & = t + \frac{4}{5} {\mathbf{j}} - \frac{3}{5} {\mathbf{k}}, \\
Q''_7 & = t - \frac{4}{5} {\mathbf{j}} + \frac{3}{5} {\mathbf{k}}
- \frac{3}{2} {\mathbf{j}}\varepsilon -2 {\mathbf{k}}\varepsilon.
\end{aligned}
\end{equation*}
We can combine this factorization to form a 7R linkage where each rotation is
defined by $Q''_7$, $Q''_6$, $Q''_5$, $Q''_4$, $Q''_3$, $Q''_2$, $Q''_1$. It can be
seen that the axes of $Q''_1$, $Q''_2$, $Q''_7$, $Q''_6$ are parallel, as are the axes
of $Q''_3$, $Q_4''$, $Q''_5$. Since four neighboring axes are parallel, this linkages
contains a planar 4-bar linkage. Furthermore, there are two pairs of parallel 3R sub chains
which give us two Sarrus linkages if we fix the remaining joint. Hence, this 7R linkage has
at least two degrees of freedom.
A Gröbner basis computation confirms that the configuration space is indeed two-dimensional
\cite{myweb7R}. Moreover, a decomposition of the configuration variety shows that the
linkage is not kinematotropic, which means that there are components of different dimension
\cite{Galletti01:_kinematotropic}:
It consists of three irreducible 2-dimensional components. Two components are rational.
We do not know whether the other two-dimensional
component which contains the rational curve parameterized by~$C$ is rational or not.
In Figure~\ref{fig:darboux3}, we present eleven configurations
(the first one and the last one are the same) of this linkage in
an orthographic projection parallel to the direction of ~${\mathbf{k}}$.
Because the axes of $Q''_1$, $Q''_2$, $Q''_7$, and $Q''_6$ are parallel,
the remaining axes are depicted as points throughout the motion. We can
observe the parallelity of axes and possible singularities (intersections).
The two configurations where the axes of $Q''_3$ and $Q''_5$ coincide give us
two rational configuration sets of dimension two: A planar 4R loop
(the coupler motion is a circular translation) and a single rotation
rational component. None of these two contains the Darboux motion.
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{example5/example5+0} \includegraphics[width=0.49\textwidth]{example5/example5+1}\\
\includegraphics[width=0.49\textwidth]{example5/example5+2} \begin{overpic}[width=0.49\textwidth]{example5/example5+3}
\put(52,28){$Q''_3 = Q''_5$}
\end{overpic}\\
\includegraphics[width=0.49\textwidth]{example5/example5+4} \includegraphics[width=0.49\textwidth]{example5/example5+5}\\
\includegraphics[width=0.49\textwidth]{example5/example5+6} \includegraphics[width=0.49\textwidth]{example5/example5+7}\\
\begin{overpic}[width=0.49\textwidth]{example5/example5+8}
\put(52,8){$Q''_3 = Q''_5$}
\end{overpic} \includegraphics[width=0.49\textwidth]{example5/example5+9}\\
\includegraphics[width=0.49\textwidth]{example5/example5+10} \includegraphics[width=0.49\textwidth]{example5/example5+0}
\caption{The second 7R linkage generates a non-vertical Darboux motion.}
\label{fig:darboux3}
\end{figure*}
\section{Conclusions}
\label{sec:conclusions}
Using factorization of motion polynomials in non-generic cases, we constructed several examples of
new 7R Darboux linkages. The main idea is to find an irreducible real polynomial $P$ and
factor $PC$ instead of $C$. The obtained 7R linkages generate the general (non-vertical)
Darboux motion and exhibit some interesting specialities like parallel axes or
circular translations as relative motions between certain links. One can replace two R-joints by other
R-joints which generate
same circular translations.
Verifying the validity of the presented factorizations and corresponding linkages is
trivial but, admittedly, we did not explain in detail how to actually find the
factorizations. The factorization theory for non-generic motion polynomials,
including theoretical results, algorithms and upper bounds for the number of factors,
is currently being worked out. Its presentation is left to future publications.
\section*{Acknowledgements}
This research was supported by the Austrian Science Fund (FWF):
P\;26607 (Algebraic Methods in Kinematics: Motion Factorisation and
Bond Theory).
\bibliographystyle{iftomm}
|
\section{Introduction}
The recent observation of two pulsars with approximately two solar
masses \cite{Demorest:2010bx,Antoniadis:2013pzd} presents a severe
challenge to the theoretical description of cold high-density matter
in $\beta$-equilibrium.
The equation of state (EoS) has to be sufficiently stiff in order to support
such high masses of compact stars. Many models that are solely based on
nucleonic (neutrons and protons) and leptonic (electrons and muons)
degrees of freedom are able to reproduce maximum neutron star masses
above two solar masses if the effective interaction between the nucleons becomes
strongly repulsive at high baryon densities. However,
additional hadronic particle species can appear at densities
above two or three times the nuclear saturation density $n_{\rm sat}
\approx 0.16$~fm$^{-3}$. In most cases, these additional degrees of
freedom lead to a substantial softening of the EoS resulting in a reduced
maximum mass of the compact star below the observed values.
This feature is well-known for
models with hyperons -- the so-called ''hyperon puzzle'', see, e.g.,
\cite{Lonardoni:2014bwa,Fortin:2014mya} and references therein --
but was also observed in approaches
that take excited states of the nucleons such as $\Delta(1232)$ resonances into
account, see, e.g., \cite{Drago:2014oja,Cai:2015hya} and references therein.
Usually, only specifically designed interactions can avoid the problem
of too low maximum masses.
Successful models of the baryonic contribution to the stellar EoS
should be scrutinized whether they comply with other experimental
constraints, e.g.\ with respect to the employed interactions.
In the center of compact stars very high baryon densities are reached
exceeding several times $n_{\rm sat}$ and the corresponding Fermi
momenta of the particles are much larger than those at saturation.
This is particularly significant for models with only nucleonic degrees of freedom.
Hence, not only the density dependence of the effective in-medium
interaction between nucleons but also their momentum dependence
becomes relevant. For densities near $n_{\rm sat}$ this information is
contained in the optical potential of nucleons that can be extracted
from the systematics of elastic proton scattering on nuclei, see,
e.g., \cite{Hama:1990vr,Cooper:1993nx}. A saturation of the real part
of the optical potential is observed at high kinetic energies
approaching $1$~GeV. The momentum dependence of the in-medium
interaction is also crucial in simulations of heavy-ion collisions
\cite{Zhang:1994hpa}.
Typical approaches for
the baryonic contribution to the stellar EoS are
energy density functionals that originate from
nonrelativistic or relativistic mean-field models of nuclear matter.
The most prominent cases among the former class are Skyrme energy
density functionals, see reference \cite{Dutra:2012mb} for an overview of different
parametrizations. They are derived originally from the zero-range Skyrme interaction
with a two-body contribution, which is an expansion up to second order
in the particle momenta, and a density dependent three-body
contribution, which is included in order to reproduce saturation
properties of nuclear matter. Obviously, an extrapolation of the model
to high momenta is questionable given the limited form of the momentum
dependence. Examples of the latter class, frequently denominated covariant
density functionals, can be inferred from relativistic Lagrangian
densities. In conventional models, see reference \cite{Dutra:2014qga}
for a wide collection of different parametrizations, a nucleon optical
potential in the medium can be derived from the relativistic scalar and
vector self-energies, see section \ref{sec:res}. It exhibits a linear increase with
energy, which is in contradiction with the expectation from
experiment. In general, one would expect that the nucleon self-energies
itself depend explicitly on the particle momentum or energy as, e.g.,
in Dirac-Brueckner calculations of nuclear matter \cite{Fuchs:2003zn}. However, this is
not realized in standard RMF models.
There are particular extensions of relativistic mean-field
(RMF) models that contain
nucleon self-energies with an explicit energy or momentum dependence.
This dependence cannot be introduced in a relativistic model in a
simple parametric form because it affects, e.g., the definition of the
conserved currents.
In extended systematic approaches new derivative couplings between the nucleon and
meson fields are introduced that allow to reproduce the energy
dependence of the optical potential as extracted from experiments.
One of the earliest RMF models with scalar derivative couplings was
presented in reference \cite{Zimanyi:1990np}. A rescaling of the
nucleon fields removed the explicit momentum dependence of the
self-energies but lead to a considerable softening of the EoS.
More general couplings of the mesons to linear derivatives of the
nucleon fields were considered in reference \cite{Typel:2002ck} with an
application to uniform nuclear matter. With appropriately chosen
coupling constants a reduction of the optical potential was found
as compared to the strong linear energy dependence in conventional RMF
models. The model was further extended in reference
\cite{Typel:2005ba} assuming a density dependence of the couplings. It
was successfully applied to the description of finite nuclei. Notable
new features were the increase of the effective nucleon masses
(usually rather small in order to explain the strong spin-orbit
interaction in nuclei) and correspondingly higher level densities
close to the Fermi energy in nuclei
in better concordance with expectations from
experiments. Though, using couplings only linear in the
derivatives leads to a quadratic dependence of the optical potential
on the kinetic energy with a decrease for energies
exceeding $1$~GeV. Couplings to all orders in the derivative of the
nucleons were introduced in the so-called nonlinear derivative (NLD)
model \cite{Gaitanos:2009nt,Gaitanos:2011yb} assuming a particular exponential
dependence on the derivatives but no density dependence of the
nucleon-meson couplings or self-couplings of the mesons.
The general formalism was developed and applied to infinite
isospin symmetric and asymmetric nuclear matter with a particular choice of the
non-linear derivative terms that lead to an energy dependence of the self-energies.
In reference \cite{Gaitanos:2012hg} the approach was extended with
generalized non-linear derivative couplings of any functional form
in the field-theoretical formalism
allowing for a momentum or energy dependence.
Nonlinear self-couplings of the $\sigma$ meson field were added
in order to improve the
description of characteristic nuclear matter parameters at saturation.
The application of this version of the NLD model to stellar
matter yielded a maximum neutron star mass of $2.03~M_{\rm sol}$ barely satisfying
the observational constraints but the dependence of the result on the
model parameters was not explored in detail.
The NLD model was also applied to the description of bulk properties of nuclear matter
in reference \cite{Chen:2012rn}. A softening of the EoS and
maximum masses of neutron stars substantially below $2~M_{\rm sol}$
were found with different parametrizations
assuming an energy dependence of the couplings but
nonlinear self-couplings of the mesons or density dependent
meson-nucleon couplings were not considered.
Properties of finite nuclei were studied in reference \cite{Chen:2014xha}
after adding meson self-interactions in the Lagrangian.
A qualitative description similar to conventional RMF models was achieved but neutron
star properties were not examined in this extended model.
In this work we introduce a more flexible extension of the
nonlinear derivative model assuming density dependent
meson-nucleon couplings in addition. Instead of using derivative operators that generate
an explicit momentum dependence of the self-energies, we will use a functional
form that leads to an energy dependence. This approach will also be more
suitable for a future applications of the DD-NLD approach to nuclei
since the relevant equations and their numerical implementation
are simplified. Here, the equations of state
of symmetric and asymmetric nuclear matter will be calculated for
different choices of the derivative coupling operators that lead to a
saturation of the optical potential at high energies as derived from experiments.
They are compared to
the results of a standard RMF model with density dependent couplings
that is consistent with essentially all modern constraints for the
characteristic nuclear matter parameters at saturation.
The parameters of the DD-NLD models are chosen such that these
saturation properties are reproduced. The effect of the optical
potential constraint on the mass-radius relations of
neutron stars will be studied.
The paper is organized as follows: In section \ref{sec:model} the
Lagrangian density of the DD-NLD approach is presented. The field
equations in mean-field approximation and the
energy-momentum tensor will be derived. The relevant equations for the
case of infinite nuclear matter will be considered in more detail in
section \ref{sec:inm}. The parametrization of the
density dependent couplings and the functional form of the derivative
coupling functions is discussed in section \ref{sec:para}. Results for
the energy dependence of the optical potential, the EoS of nuclear matter and the
mass-radius relation are presented in section \ref{sec:res} for
various versions of the model. Conclusions are given in section
\ref{sec:con}. Detailed expression for various densities are collected
in \ref{sec:app_a}.
\section{Lagrangian density and field equations of the DD-NLD model}
\label{sec:model}
In most RMF models the effective interaction between nucleons is
described by an exchange of mesons. Usually, $\sigma$ and $\omega$
mesons are introduced to consider the attractive and
repulsive contributions to the nucleon-nucleon potential, respectively. They are
represented by isoscalar Lorentz scalar and Lorentz vector fields
$\sigma$ and $\omega_{\mu}$. In order to model the isospin dependence
of the interaction, the exchange of $\rho$ mesons is included. It
is denoted by the isovector Lorentz vector field $\bm{\rho}_{\mu}$ in
the following. The Lagrangian density in the DD-NLD
approach\footnote{Natural units with $\hbar = c = 1$ are used in the following.}
\begin{equation}
\mathcal{L} = \mathcal{L}_{\rm nuc} + \mathcal{L}_{\rm mes} +
\mathcal{L}_{\rm int}
\end{equation}
contains contributions of the free nucleons $\Psi = (\Psi_{p},\Psi_{n})$ with mass $m$
\begin{equation}
\label{eq:L_nuc}
\mathcal{L}_{\rm nuc} = \frac{1}{2} \left(
\overline{\Psi}\gamma_{\mu}i\overrightarrow{\partial^{\mu}}\Psi -
\overline{\Psi} i \overleftarrow{\partial^{\mu}}\gamma_{\mu}\Psi
\right) - m \overline{\Psi}\Psi
\end{equation}
in a symmetrized form and of free mesons
\begin{eqnarray}
\mathcal{L}_{\rm mes} & = & \frac{1}{2}
\left( \partial_{\mu}\sigma\partial^{\mu}\sigma
- m^{2}_{\sigma}\sigma^{2}
- \frac{1}{2} F^{(\omega)}_{\mu\nu}F^{(\omega)\mu\nu}
+ m^{2}_{\omega}\omega_{\mu}\omega^{\mu}
\right. \\ \nonumber & & \left.
- \frac{1}{2} \bm{F}^{(\rho)}_{\mu\nu} \bm{F}^{(\rho)\mu\nu}
+ m^{2}_{\rho} \bm{\rho}_{\mu} \bm{\rho}^{\mu}
\right)
\end{eqnarray}
with the field tensors
\begin{equation}
F^{(\omega)}_{\mu\nu} = \partial_{\mu} \omega_{\nu} - \partial_{\nu}
\omega_{\mu}
\quad \mbox{and} \quad
\bm{F}^{(\rho)}_{\mu\nu} = \partial_{\mu} \bm{\rho}_{\nu} - \partial_{\nu}
\bm{\rho}_{\mu}
\end{equation}
of the isoscalar $\omega$ meson and the isovector $\rho$ meson,
respectively. The arrows in equation (\ref{eq:L_nuc}) denote the
direction of differentiation.
Standard RMF models assume a minimal coupling of the nucleons to the
meson fields leading to
\begin{equation}
\label{eq:L_int_st}
\mathcal{L}_{\rm int} =
\Gamma_{\sigma} \sigma \overline{\Psi} \Psi
- \Gamma_{\omega} \omega_{\mu}\overline{\Psi}\gamma^{\mu}\Psi
- \Gamma_{\rho}
\bm{\rho}_{\mu}\overline{\Psi}\bm{\tau}\gamma^{\mu}\Psi
\end{equation}
for the interaction contribution to the total Lagrangian density
$\mathcal{L}$ with meson-nucleon couplings
$\Gamma_{i}$ ($i=\sigma,\omega,\rho$). We assume that they depend on
the vector density $n_{v}$, see equation (\ref{eq:n_v}) for the
explicit definition.
In the derivative coupling model the nucleon
field $\Psi$ ($\overline{\Psi}$) is replaced in $\mathcal{L}_{\rm int}$ by $\mathcal{D}_{m}\Psi$
($\overline{\mathcal{D}_{m}\Psi}$) with operator functions
$\mathcal{D}_{m}$, which can be different for the various mesons
$m=\sigma,\omega,\rho$. They can be expanded in a series
\begin{equation}
\mathcal{D}_{m}(x) = \sum_{n=0}^{\infty} \frac{d_{n}^{(m)}}{n!} x^{n}
\end{equation}
with numerical coefficients $d_{n}^{(m)}$.
The argument $x$ contains derivatives $i\partial_{\beta}$ that
act on the nucleon field. More specifically we write
\begin{equation}
\label{eq:def_x}
x = v^{\beta}i\partial_{\beta} - sm
\end{equation}
as a hermitian Lorentz scalar operator
with an auxiliary Lorentz vector $v^{\beta}=(v_{0},\vec{v})$ and
a scalar factor $s$.
Hence the interaction contribution in the NLD model is written as
\begin{eqnarray}
\mathcal{L}_{\rm int} & = &
\frac{1}{2} \Gamma_{\sigma} \sigma \left(
\overline{\Psi}\overleftarrow{\mathcal{D}}_{\sigma}\Psi
+ \overline{\Psi} \overrightarrow{\mathcal{D}}_{\sigma} \Psi \right)
\\ \nonumber & &
- \frac{1}{2} \Gamma_{\omega} \omega_{\mu}\left(
\overline{\Psi}\overleftarrow{\mathcal{D}}_{\omega}\gamma^{\mu}\Psi
+ \overline{\Psi}\gamma^{\mu}\overrightarrow{\mathcal{D}}_{\omega}\Psi \right)
\\ \nonumber & &
- \frac{1}{2} \Gamma_{\rho} \bm{\rho}_{\mu} \left(
\overline{\Psi}\overleftarrow{\mathcal{D}}_{\rho}\gamma^{\mu}\bm{\tau}\Psi
+ \overline{\Psi}\bm{\tau}\gamma^{\mu}\overrightarrow{\mathcal{D}}_{\rho}\Psi
\right)
\end{eqnarray}
in a symmetrized form with respect to the derivative operators $\mathcal{D}_{m}$,
i.e.,
\begin{eqnarray}
\overrightarrow{\mathcal{D}}_{m} &=& \sum^{\infty}_{k=0} C_{k}^{(m)}
(v^{\beta }i \overrightarrow{\partial}_{\beta})^{k} \\
\overleftarrow{\mathcal{D}}_{m} &=& \sum^{\infty}_{k=0} C_{k}^{(m)}
(-v^{\beta }i \overleftarrow{\partial}_{\beta})^{k}
\end{eqnarray}
with coefficients
\begin{equation}
C_{k}^{(m)} = \sum^{k}_{n=0} \frac{d_{n}^{(m)}}{n!} \binom{n}{k} (-sm)^{n-k} \: .
\end{equation}
Obviously, no derivatives appear for the choice $\mathcal{D}_{m}=1$ (corresponding to
$d_{n}^{(m)} = \delta_{n0}$) and the standard form (\ref{eq:L_int_st}) is
recovered.
When spatially inhomogeneous systems with Coulomb interaction are
considered, $\mathcal{L}_{\rm mes}$ and $\mathcal{L}_{\rm int}$ can be
complemented with the appropriate contributions. Since only
uniform matter is considered in the following, we do not give them
here explicitly.
The field equations of nucleons are derived from the
generalized Euler-Lagrange equation
\begin{equation}
\label{eq:EuLa}
\frac{\partial\mathcal{L}}{\partial \varphi_r}
+ \sum^{\infty}_{i=1}(-)^i \partial_{\alpha_1,\dots,\alpha_i}
\frac{\partial\mathcal{L}}{\partial(\partial_{\alpha_1,\dots,\alpha_i}\varphi_r)} = 0
\end{equation}
for $\varphi_{r}=\Psi, \overline{\Psi}$. For the meson fields $\varphi_{r}= \sigma, \omega_{\mu},
\bm{\rho}_{\mu}$ the standard Euler-Lagrange equation applies, i.e.\
only the $i=1$ term in equation (\ref{eq:EuLa}) is relevant since higher-order
derivatives of the meson fields do not appear in the Lagrangian density $\mathcal{L}$.
Details can found in references
\cite{Gaitanos:2009nt,Gaitanos:2012hg}.
The Dirac equation
\begin{equation}
\label{eq:Dirac}
\left[\gamma_{\mu} \left(i\partial^{\mu} - \Sigma^{\mu}\right)
- \left( m - \Sigma\right)\right] \Psi = 0
\end{equation}
for the nucleons looks formally the same as in standard RMF approaches
but the scalar ($\Sigma$) and vector ($\Sigma^{\mu}$) self-energy operators now
contain the derivative operators $\mathcal{D}_{m}$. They are given by
\begin{equation}\
\label{eq:Sigma_s}
\Sigma = \Gamma_{\sigma} \sigma \overrightarrow{\mathcal{D}}_{\sigma}
\end{equation}
and
\begin{equation}
\label{eq:Sigma_v}
\Sigma^{\mu}= \Gamma_{\omega}\omega^{\mu}\overrightarrow{\mathcal{D}}_{\omega}
+ \Gamma_{\rho}\bm{\tau} \cdot \bm{\rho}^{\mu}
\overrightarrow{\mathcal{D}}_{\rho} + \Sigma^{\mu}_{R}
\end{equation}
with the 'rearrangement' contribution
\begin{eqnarray}
\Sigma^{\mu}_{R} & = &
\frac{j^{\mu}}{n_{v}} \left[
\Gamma_{\omega}^{\prime}
\omega^{\nu}
\frac{1}{2} \left( \overline{\Psi} \overleftarrow{\mathcal{D}}_{\omega} \gamma_{\nu} \Psi
+ \overline{\Psi} \gamma_{\nu} \overrightarrow{\mathcal{D}}_{\omega} \Psi \right)
\right. \\ \nonumber & & \left.
+ \Gamma_{\rho}^{\prime}
\bm{\rho}^{\nu} \frac{1}{2} \left(
\overline{\Psi} \overleftarrow{\mathcal{D}}_{\rho} \gamma_{\nu} \bm{\tau} \Psi
+ \overline{\Psi} \gamma_{\nu} \bm{\tau} \overrightarrow{\mathcal{D}}_{\rho}
\Psi \right)
\right. \\ \nonumber & & \left.
- \Gamma_{\sigma}^{\prime} \sigma \frac{1}{2}
\left( \overline{\Psi} \overleftarrow{\mathcal{D}}_{\sigma} \Psi
+ \overline{\Psi} \overrightarrow{\mathcal{D}}_{\sigma} \Psi \right) \right]
\end{eqnarray}
containing derivatives
\begin{equation}
\Gamma_{i}^{\prime} = \frac{d\Gamma_{i}}{dn_{v}}
\end{equation}
of the coupling functions.
In the case of inhomogeneous systems and a non-vanishing three-vector
component $\vec{v}$ of the auxiliary vector $v^{\beta}$, additional
contributions in (\ref{eq:Sigma_s}) and (\ref{eq:Sigma_v}) will
appear. In the present application of the DD-NLD model, however, we
do not consider this case.
The field equations of the mesons are found as
\begin{eqnarray}
\partial_{\mu}\partial^{\mu} \sigma + m^{2}_{\sigma} \sigma
& = & \frac{1}{2} \Gamma_{\sigma} \left(
\overline{\Psi} \overleftarrow{\mathcal{D}}_{\sigma} \Psi
+ \overline{\Psi}\overrightarrow{\mathcal{D}}_{\sigma} \Psi \right)
\\
\partial_{\mu} F^{(\omega)\mu\nu} + m^{2}_{\omega} \omega^{\nu}
& = & \frac{1}{2} \Gamma_{\omega} \left(
\overline{\Psi} \overleftarrow{\mathcal{D}}_{\omega} \gamma^{\nu} \Psi
+ \overline{\Psi}\gamma^{\nu} \overrightarrow{\mathcal{D}}_{\omega} \Psi \right)
\\
\partial_{\mu} \bm{F}^{(\rho)\mu\nu} + m^{2}_{\rho} \bm{\rho}^{\nu}
& = & \frac{1}{2} \Gamma_{\rho} \left(
\overline{\Psi} \overleftarrow{\mathcal{D}}_{\rho} \gamma^{\nu} \bm{\tau}\Psi
+ \overline{\Psi}\bm{\tau}\gamma^{\nu}
\overrightarrow{\mathcal{D}}_{\rho} \Psi \right)
\end{eqnarray}
with source terms containing derivative operators.
The conserved baryon current in the DD-NLD model is given by
\begin{equation}
\label{eq:J}
J^{\mu} = \sum_{i=p,n} \langle\overline{\Psi}_{i}N^{\mu}\Psi_{i}\rangle
\end{equation}
with the norm operator
\begin{equation}
\label{eq:norm}
N^{\mu} = \gamma^{\mu} + \Gamma_{\sigma} \sigma
\left(\partial^{\mu}_{p} \mathcal{D}_{\sigma} \right)
- \Gamma_{\omega}\omega_{\alpha} \gamma^{\alpha}
\left( \partial^{\mu}_{p} \mathcal{D}_{\omega}\right)
- \Gamma_{\rho} \bm{\rho} _{\alpha}
\gamma^{\alpha} \bm{\tau} \left( \partial^{\mu}_{p}
\mathcal{D}_{\rho}\right)
\end{equation}
where $\partial^{\mu}_{p}\mathcal{D}_{m}$ is the derivative of $\mathcal{D}_{m}$
operator with respect to the momentum $p_{\mu}=i\partial_{\mu}$, i.e.\
\begin{equation}
\partial^{\mu}_{p}\mathcal{D}_{m}
= v^{\mu}\sum^{\infty}_{k=1} kC_{k}^{(m)} (v^{\beta }i\partial_{\beta})^{k-1}
\: ,
\end{equation}
and $\langle \dots \rangle$ denotes the summation over all occupied
states.
The current (\ref{eq:J}) is not identical to the vector current
\begin{equation}
\label{eq:J_v}
J_{v}^{\mu} = \sum_{i=p,n}
\langle\overline{\Psi}_{i}\gamma^{\mu}\Psi_{i}\rangle \: ,
\end{equation}
which is used to define the vector density
\begin{equation}
\label{eq:n_v}
n_{v} = \sqrt{J_{v}^{\mu}J_{v\mu}}
\end{equation}
appearing as the argument of the coupling functions $\Gamma_{i}$.
The energy-momentum tensor assumes the form
\begin{equation}
T^{\mu \nu} = \sum_{i=p,n}
\langle\overline{\Psi}_{i}N^{\mu} p^{\nu} \Psi_{i}\rangle
- g^{\mu\nu}\langle \mathcal{L} \rangle \: .
\end{equation}
Then the energy density $\varepsilon$ and pressure $p$ are found from
$\varepsilon = T^{00}$ and $p = \sum_{i=1}^{3}T^{ii}/3$, respectively.
\section{DD-NLD model for nuclear matter}
\label{sec:inm}
In the case of stationary nuclear matter, the equations simplify considerably
since the system is homogeneous and the meson fields, which are
treated as classical fields, are constant in space and
time. Positive-energy solutions of the
Dirac equation (\ref{eq:Dirac}) are plane waves $\Psi_{i} = u_{i}
\exp \left( - i p_{i}^{\mu} x_{\mu}\right)$ for protons
and neutrons with Dirac spinors $u_{i}$, which are normalized according to
\begin{equation}
\overline{\Psi}_{i} N^{0}
\Psi_{i} = \bar{u}_{i} N^{0} u_{i} = 1
\end{equation}
with the time component of the norm operator (\ref{eq:norm}).
They depend on the effective mass
\begin{equation}
m_{i}^{\ast} = m_{i} - \Sigma_{i}
\end{equation}
and effective momentum
\begin{equation}
p_{i}^{\ast\mu} = p_{i}^{\mu} - \Sigma_{i}^{\mu}
\end{equation}
related by the dispersion relation
\begin{equation}
p_{i}^{\ast\mu} p_{i\mu}^{\ast} = \left(m_{i}^{\ast}\right)^{2} \: .
\end{equation}
The derivative $i\partial^{\beta}$ in the $\mathcal{D}_{m}$ operators
can be replaced by the corresponding four-momentum
$p_{i}^{\beta}=(E_{i},\vec{p}_{i})$
resulting in a simple function $D_{m}$ depending on the energy $E_{i}$ and
the momentum $\vec{p}_{i}$ of the nucleon.
Using the identity
\begin{equation}
N^{\mu} \Psi_{i} = \left[ \gamma^{\mu} + \partial_{p}^{\mu}
\Sigma_{i} - \gamma_{\alpha} \partial_{p}^{\mu} \left(
\Sigma_{i}^{\alpha} - \Sigma_{R}^{\alpha} \right)
\right] \Psi_{i}
\end{equation}
the conserved current and the energy-momentum tensor can be written as
\begin{equation}
J^{\mu} = \sum_{i=p,n} \kappa_{i}
\int \frac{d^{3} p}{(2\pi)^{3}} \: \frac{\Pi^{\mu}_{i}}{\Pi^{0}_{i}}
\end{equation}
and
\begin{equation}
T^{\mu\nu}= \sum_{i=p,n} \kappa_{i} \int \frac{d^{3}p}{(2\pi)^{3}} \:
\frac{\Pi^{\mu}_{i} p^{\nu}}{\Pi^{0}_{i}} -
g^{\mu\nu}\langle\mathcal{L}\rangle \: ,
\end{equation}
respectively, with the four-momentum
\begin{equation}
\Pi^{\mu}_{i} = p^{\ast\mu}_{i} + m^{\ast}_{i} \left(\partial^{\mu}_{p}\Sigma_{i}\right)
- p^{\ast}_{i\beta} \left[\partial^{\mu}_{p} \left( \Sigma^{\beta}_{i}-
\Sigma^{\beta}_{R}\right)\right]
\end{equation}
and spin degeneracy factors $\kappa_{i}=2$.
The integration runs over all momenta $p$ with modulus lower than the
Fermi momenta $p_{Fi}$ in the no-sea approximation. They are defined through the individual
nucleon densities
\begin{equation}
\label{eq:n_i}
n_{i} = \frac{\kappa_{i}}{6\pi^{2}} p_{Fi}^{3} \: .
\end{equation}
Without the preference for a particular direction in infinite nuclear
matter, the spatial components of
the Lorentz vector meson fields vanish and the auxiliary vector in
equation (\ref{eq:def_x}) is set to $v^{\beta}=\delta_{\beta 0}$ such
that the $D_{m}$ functions only depend on the nucleon energy $E_{i}$.
Without isospin changing processes, only the third
component of the isovector $\rho$ field has to be considered in the
field equations for the mesons. Using the abbreviations
$\omega = \omega^{0}$ and $\rho = \bm{\rho}^{0}_{3}$ the meson fields
are immediately obtained from
\begin{eqnarray}
\label{eq:feq_sigma}
\sigma & = & \frac{\Gamma_{\sigma}}{m_{\sigma}^{2}} n_{\sigma}
= \frac{\Gamma_{\sigma}}{m_{\sigma}^{2}} \sum_{i=p,n} \langle \overline{\Psi}_{i} D_{\sigma}
\Psi_{i} \rangle
\\
\label{eq:feq_omega}
\omega & = & \frac{\Gamma_{\omega}}{m_{\omega}^{2}} n_{\omega}
= \frac{\Gamma_{\omega}}{m_{\omega}^{2}} \sum_{i=p,n} \langle \overline{\Psi}_{i} \gamma^{0} D_{\omega}
\Psi_{i} \rangle
\\
\label{eq:feq_rho}
\rho & = & \frac{\Gamma_{\rho}}{m_{\rho}^{2}} n_{\rho}
= \frac{\Gamma_{\rho}}{m_{\rho}^{2}} \sum_{i=p,n} \langle \overline{\Psi}_{i} \gamma^{0}
\tau_{3} D_{\rho} \Psi_{i} \rangle
\end{eqnarray}
with source densities $n_{\sigma}$, $n_{\omega}$, and $n_{\rho}$.
The self-energies simplify to
\begin{eqnarray}
\Sigma_{i} & = & \Gamma_{\sigma} \sigma D_{\sigma}
\\
\Sigma_{i}^{0} & = & \Gamma_{\omega}\omega D_{\omega}
+ \Gamma_{\rho} \tau_{3,i} \rho D_{\rho}
+ \Sigma^{0}_{R}
\\
\vec{\Sigma}_{i} & = & 0
\end{eqnarray}
with $\tau_{3,i} = 1$ ($-1$) for protons (neutrons) and
the 'rearrangement' contribution
\begin{equation}
\Sigma^{0}_{R} =
\Gamma_{\omega}^{\prime} \omega n_{\omega}
+ \Gamma_{\rho}^{\prime} \rho n_{\rho}
- \Gamma_{\sigma}^{\prime} \sigma n_{\sigma}
\end{equation}
is independent of the nucleon energy.
The dispersion relation reads
\begin{equation}
\label{eq:disp}
E_{i} = \sqrt{p^{2}+\left( m_{i} - S_{i} \right)^{2}} + V_{i}
\end{equation}
if we introduce the energy-dependent scalar potentials $S_{i}(E) = \Sigma_{i}$ and
vector potentials $V_{i}(E) = \Sigma_{i}^{0}$.
Explicit expressions for the various densities and
thermodynamic quantities of the DD-NLD model are given in \ref{sec:app_a}.
\section{Parametrization of the DD-NLD model}
\label{sec:para}
For the application of the DD-NLD model to nuclear matter the parameters need to be
specified. Besides the usual parameters of a RMF model with density
dependent couplings the form of the $D_{m}$ functions has to be given.
We assume identical functions for all mesons, i.e.\ $D = D_{\sigma} =
D_{\omega} = D_{\rho}$, and consider three functional dependencies:
\begin{itemize}
\item[D1] a constant $D = 1$, which corresponds to a usual RMF model with density
dependent couplings,
\item[D2] a Lorentzian form $D = 1/(1+x^{2})$,
\item[D3] an exponential dependence $D=\exp\left( -x \right)$
\end{itemize}
with $x=(E_{i}-m_{i})/\Lambda$ because we set
$v^{\beta}=\delta_{\beta 0}$ and $s=1$ in equation
(\ref{eq:def_x}). The parameter $\Lambda$ regulates the strength of
the energy dependence.
For the proton and neutron masses the experimental values
of $m_{p} = 938.272046$~MeV/c${}^{2}$ and $m_{n} = 939.565379$~MeV/c${}^{2}$,
respectively, are used. The meson masses are set to
$m_{\sigma}=550$~MeV/c${}^{2}$, $m_{\omega}=783$~MeV/c${}^{2}$, and $m_{\rho}=763$~MeV/c${}^{2}$.
The density dependence of the meson nucleon couplings has the same
form as introduced in reference \cite{Typel:1999yq}.
We assume a dependence on the vector density $n_{v}$ as defined
in equation (\ref{eq:n_v}), which is not identical to the zero-component of the
conserved baryon current (\ref{eq:J}). See \ref{sec:app_a} for explicit expressions.
This choice simplifies the rearrangement terms considerably.
For the isoscalar
mesons $m=\sigma,\omega$ the coupling is written as
\begin{equation}
\Gamma_{m}(n_{v}) = \Gamma_{m}(n_{\rm ref}) f_{m}(x)
\end{equation}
with functions
\begin{equation}
f_{m}(x) = a_{m} \frac{1 + b_{m} (x + d_{m})^{2}}{1 + c_{m} (x +d_{m})^{2}}
\end{equation}
that depend on the argument $x=n_{v}/n_{\rm ref}$ and contain
coefficients $a_{m}$, $b_{m}$, $c_{m}$, and $d_{m}$. For the
$\rho$-meson coupling we set
\begin{equation}
\Gamma_{\rho}(n_v) = \Gamma_{\rho}(n_{\rm ref}) \exp\left[-a_{\rho}
\left(x-1\right)\right] \: .
\end{equation}
In order to reduce the number of independent parameters we demand that
the conditions $f_{\sigma}(1)=f_{\omega}(1)=1$ and $f_{\sigma}^{\prime\prime}(0) =
f_{\omega}^{\prime\prime}(0)=0$ hold. Hence, there are only two
independent coefficients in the functions $f_{m}$ for each of the isoscalar mesons.
The overall magnitude of the couplings is given by
the couplings $\Gamma_{m}(n_{\rm ref})$ at a reference density
$n_{\rm ref}$.
We require that the characteristic saturation properties for the three
choices of the $D$ function are identical and close to current values
extracted from experiments.
In particular, we set the saturation density to $n_{\rm sat}=
0.15$~fm${}^{-3}$, the binding energy per nucleon at saturation to
$B=16$~MeV, the compressibility to $K = 240$~MeV, the symmetry energy
to $J = 32$~MeV and the symmetry energy slope coefficient to $L=60$~MeV.
Furthermore, we set the effective nucleon mass at saturation to
$m_{\rm eff} = 0.5625~m_{\rm nuc}$ (related to the strength of the
spin-orbit potential in nuclei) and fix the ratios
$f_{\omega}^{\prime}(1)/f_{\omega}(1)=-0.15$ and
$f_{\omega}^{\prime\prime}(1)/f_{\omega}^{\prime}(1)=-1.0$ in order
to determine the coefficients in the functions $f_{m}$ uniquely.
These values are close to those of the parametrization DD2 \cite{Typel:2009sy}
that was fitted to properties of nuclei and predicts a neutron star maximum mass of
$2.4~M_{\rm sol}$.
\begin{table}[t]
\caption{\label{tab:1
Parameters of the meson coupling functions for three choices of the $D$
functions and different values of cut-off parameter
$\Lambda$.}
\centering
\begin{tabular}{clccccc}
\toprule
meson & parameter & \multicolumn{1}{c}{D1}
&\multicolumn{2}{c}{D2} & \multicolumn{2}{c}{D3}
\\
\midrule
& $\Lambda$ [MeV] & $-$ & $400$ & $500$ & $600$ & $700$ \cr
\midrul
$\sigma$ & $\Gamma_{\sigma}(n_{\rm ref})$ & 10.72913 & 10.93466 & 10.86315 & 9.74679& 9.89158 \cr
& $a_{\sigma}$ & 1.36402 & 1.35816 & 1.36015 & 1.38410 & 1.38064 \cr
& $b_{\sigma}$ & 0.53404 & 0.51914 & 0.52433 & 0.61515 & 0.60127 \cr
& $c_{\sigma}$ & 0.86714 & 0.83989 & 0.84931 & 1.00615 & 0.98211 \cr
& $d_{\sigma}$ & 0.62000 & 0.62998 & 0.62648 & 0.57558 & 0.58258 \cr
\midrul
$\omega$ & $\Gamma_{\omega}(n_{\rm ref})$ &13.29858 & 13.56462 & 13.47215 & 12.0503 & 12.23457 \cr
& $a_{\omega}$ & 1.3822 & 1.3822 & 1.3822 & 1.3822 & 1.3822 \cr
& $b_{\omega}$ & 0.42253 & 0.42253 & 0.42253 & 0.42253 & 0.42253 \cr
& $c_{\omega}$ & 0.71932 & 0.71932 & 0.71932 & 0.71932 & 0.71932 \cr
& $d_{\omega}$ & 0.60473 & 0.68073 & 0.68073 & 0.68073 & 0.68073 \cr
\midrul
$\rho$ & $\Gamma_{\rho}(n_{\rm ref})$ & 3.59367 & 3.67852& 3.64957 & 3.18819 & 3.25233 \cr
& $a_{\rho}$ & 0.48762 & 0.48954 & 0.48872 & 0.34777 & 0.36279\cr
\midrul
& $n_{\rm ref}$ [fm$^{-3}$] & 0.15000 & 0.14618 & 0.147485 & 0.16515 & 0.16268\cr
\bottomrul
\end{tabular}
\end{table}
Explicit values of the model parameters are given
in table \ref{tab:1} with two choices of the cut-off parameter
$\Lambda$ for the cases of Lorentzian and exponential functions $D$.
Note that the coefficients of the function $f_{\omega}$ are identical for all five
parametrizations due to the constraints. The reference density
$n_{\rm ref}$ is not necessarily identical to the saturation density
$n_{\rm sat}$ because in the case of explicit derivative couplings the
vector density $n_{v}$ is different from the conserved baryon density
$n_{B} = J^{0}=n_{p}+n_{n}$.
\section{Results}
\label{sec:res}
\begin{figure}[t]
\centering
\includegraphics[width=0.80\textwidth]{Uopt_D2.eps}
\caption{\label{fig:1
The optical potential $U_{\rm opt}$ as a function of the kinetic
energy $E_{\rm kin}=E-m_{\rm nuc}$ of a nucleon in symmetric nuclear matter at
saturation density in RMF models with parametrizations D1 and
D2 compared to two fits from Dirac phenomenology. See text for
details.}
\end{figure}
The nonlinear derivative couplings are introduced in the
RMF model in order to improve the energy dependence of
the optical potential $U_{\rm opt}$. The elastic proton
scattering on nuclei of different mass number $A$ can be well described
in Dirac phenomenology with scalar ($S$) and vector ($V$) potentials, which smoothly
vary with $A$ and the energy of the projectile
\cite{Hama:1990vr,Cooper:1993nx}. From these global fits
the optical potential in symmetric nuclear
matter at saturation density is
obtained as a function of the kinetic energy $E_{\rm kin} = E - m_{\rm
nuc}$ in the limit $A\to \infty$. There are different definitions of the
nonrelativistic optical potential when it is derived from relativistic
scalar and vector self-energies. Here we use the form
\begin{equation}
\label{eq:U_opt}
U_{\rm opt}(E) = \frac{E}{m_{\rm nuc}}V -S + \frac{S^{2} - V^{2}}{2m_{\rm
nuc}}
\end{equation}
with $S=\Sigma_{p}$ and $V = \Sigma_{p}^{0}$ as in references
\cite{Typel:2002ck,Typel:2005ba,Gaitanos:2009nt,Gaitanos:2011yb,Gaitanos:2012hg}.
In conventional RMF models without derivative couplings, the scalar
and vector potentials are constant in energy and the optical potential
(\ref{eq:U_opt}) is just a linear function in energy. This is clearly
seen in figures \ref{fig:1} and \ref{fig:2} as a full black line for
the calculation with the parametrization D1. In contrast, the optical potentials
derived from the scalar and vector potentials in Dirac phenomenology
from two different fits \cite{Hama:1990vr} are much smaller at high energies and exhibit a
saturation for $E_{\rm kin}$ approaching $1$~GeV. At low energies, the optical
potentials from experiment behave more similar as that of the
theoretical model concerning the absolute strength and
the energy dependence. In figure \ref{fig:1} (\ref{fig:2}) the result for
the DD-NLD parametrization D2 (D3) is depicted for two values of the
cut-off parameter $\Lambda$. Here, a reasonable description of the
experimental optical potential is achieved due to the energy
dependence of the nucleon self-energies. The difference between
parametrizations D2 and D3 is not very significant. The dependence of
$U_{\rm opt}$ on $\Lambda$ is stronger for D2. The deflection of the
DD-NLD curve for that of the standard RMF model D1 appears at lower
kinetic energies for the D3 parametrization as compared to the D2
case. In the DD-NLD model, the optical potential can be calculated
easily for other baryon densities and arbitrary neutron-proton
asymmetries. Because there are no experimental data available for
these general cases, we refrain from presenting the results here.
But see reference \cite{Typel:2002ck} for the systematics with density in
the linear derivative coupling model.
\begin{figure}[t]
\centering
\includegraphics[width=0.80\textwidth]{Uopt_D3.eps}
\caption{\label{fig:2
The optical potential $U_{\rm opt}$ as a function of the kinetic
energy $E_{\rm kin}$ of a nucleon in symmetric nuclear matter at
saturation density in RMF models with parametrizations D1 and
D3 compared to two fits from Dirac phenomenology. See text for
details.}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{SM.eps}
\caption{\label{fig:3
Energy per nucleon $E/A$ as a function of the baryon
density $n_{B}$ for symmetric nuclear matter (a) and pure neutron matter (b).
Results are given for three different choices of the $D$ function
and different values of the cut-off parameter $\Lambda$.}
\end{figure}
The reduction of the optical potential at high kinetic energies, which
originates from the energy dependence of the self-energies, is also
reflected in the equation of state. In figure \ref{fig:3} the
energy per nucleon $E/A$ (without the rest mass contribution)
is depicted as a function of the baryon
density $n_{B}$ in symmetric nuclear matter (left panel) and neutron
matter (right panel). In both cases, a substantial softening of the
EoS is found as compared to the standard RMF calculation with
parametrisation D1. The effect is stronger for an exponential energy
dependence of the self-energies (D3) than for the case of a Lorentzian
dependence (D2). By construction, all EoS are identical at the
saturation density $n_{\rm sat}$.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\textwidth]{MR.eps}
\caption{\label{fig:4
Mass-radius relation of neutron stars for different
choices of the $D$ functions and cut-off parameters $\Lambda$ in the
DD-NLD model. The two shaded bands refer to astrophysical mass measurements of the pulsars
PSR~$J1614-2230$ \cite{Demorest:2010bx} and PSR~$J0348+0432$ \cite{Antoniadis:2013pzd}.}
\label{MR}
\end{figure}
The DD-NLD model can be used to predict the properties of
neutron stars. Here, the EoS of stellar matter is required. It is
obtained by adding the contribution of electrons to the
energy density and pressure of the baryons. The conditions of charge
neutrality and $\beta$ equilibrium fix the lepton density and
proton-neutron asymmetry uniquely. Since the present model calculations treat
only homogeneous matter, a suitable EoS for the crust of neutron stars
has to be added at low densities. We use the standard
Baym-Pethick-Sutherland (BPS) crust EoS \cite{Baym:1971pw}.
The mass-radius relation of
neutron stars is found finally by solving the
Tolman-Oppenheimer-Volkoff equations
\cite{Oppenheimer:1939ne,Tolman:1939jz}.
It is shown for the five models of this work in figure \ref{fig:4}
together with the masses of the two most massive pulsars observed so
far. The model without an energy dependence (black full line) can explain without any
difficulties the large neutron star masses from astrophysical
observations \cite{Demorest:2010bx,Antoniadis:2013pzd} because the EoS
is rather stiff at high densities. In contrast, for the EoS of neutron
star matter in the DD-NLD models that are consistent with the optical
potential constraint, a serious reduction of the maximum neutron star
mass is seen.
These models have even problems to reach typical masses of about
$1.4~M_{\rm sol}$ of ordinary neutron stars. Deviations from the
predictions of the standard model D1 start to appear already at masses
below $0.7~M_{\rm sol}$. There is also an effect on the neutron star
radius, which is found to be smaller in the D2 and in
the D3 model. Explicit values for the maximum mass as well as the
radius and central density at this extreme conditions are given in
table \ref{tab:2}. For the DD-NLD models D2 and, in particular, D3 the
central densities is a star of maximum mass are considerably higher
than those of the standard RMF model without an energy dependence of the couplings.
It is worthwhile to compare our results for the mass-radius relation of neutron stars
with those of previous versions of the NLD model. In the approach of
reference \cite{Gaitanos:2012hg}
it was possible to reach a maximum neutron star mass of about $2~M_{\rm sol}$ with their choice
of parameters. In contrast, the results of reference \cite{Chen:2012rn} using different
functional forms of the couplings indicate a reduction of the maximum mass below the observed
values in line with our calculations. All three versions of the NLD model are adjusted
to similar values of the nuclear matter parameters at saturation, such as saturation density,
binding energy, compressibility or symmetry energy consistent with experimental constraints.
Nevertheless, the predictions for matter properties at supra-saturation densities are rather
different due to the various choices in the models to represent the effective
in-medium interaction.
In our paper, an explicit density dependence of the couplings is considered whereas in
\cite{Gaitanos:2012hg} nonlinear self-couplings of the meson fields were
assumened. In the approach of \cite{Chen:2012rn} to nuclear matter neither meson self-couplings
nor a density dependence of the couplings were used.
A reasonable description of nuclear matter near saturation does not
determine the high-density behaviour of the EoS uniquely, in particular due to the
additional freedom with the
explicit momentum/energy dependence in the NLD approach as compared to
standard RMF models. Of course, the maximum mass constraint could be used directly
in the determination of the model parameters. But also
additional constraints at high densities, e.g.\ from heavy-ion collisions,
might help to reduce the uncertainties in the extrapolation from low to high
densities in the future.
\begin{table}[t]
\caption{\label{tab:2} Maximum mass, corresponding radius and central
density of neutrons stars in the DD-NLD models with different parametrizations.}
\vspace{2mm}
\centering
\begin{tabular}{lcccc}
\toprul
model & $M_{\rm max}$ [$M_{\rm sol}$] & $R(M_{\rm max})$ [km] &
$n_{\rm central}(M_{\rm max})$ [fm${}^{-3}$] \\
\midrul
D1 & 2.37 & 11.57 & 0.88 \cr
D2, $\Lambda = 500$~MeV & 1.48 & 11.75 & 0.95 \\
D2, $\Lambda = 400$~MeV & 1.36 & 11.63 & 0.97 \\
D3, $\Lambda = 700$~MeV & 1.26 & 10.13 & 1.41 \\
D3, $\Lambda = 600$~MeV & 1.19 & 9.99 & 1.46 \\
\bottomrul
\end{tabular}
\end{table}
\section{Conclusions}
\label{sec:con}
There are several aspects that have to be taken into account in
constraining models of dense matter for the application to neutron
stars. In phenomenological models, the characteristic saturation
properties of nuclear matter and the density dependence of the effective
interaction are usually addressed. However, less attention is paid to
its energy or momentum dependence.
Introducing non-linear derivative
couplings into RMF models, it is possible to generate an energy
dependence of the nucleon self-energies such that the optical
potential in nuclear matter, which is extracted in Dirac phenomenology
from elastic proton-nucleus scattering experiments, can be well described
up to energies of $1$~GeV.
Considering density-dependent nucleon-meson couplings at the same time,
a very flexible model is obtained. Its parameters can be fitted to the
usual nuclear matter constraints even for different functional forms
of the energy dependent couplings.
In the current version of the DD-NLD model, the
energy dependence of the self-energies causes a softening of the
EoS at high densities, both for symmetric nuclear matter and pure
neutron matter. This effect is independent of the appearance
of additional degrees of freedom, such as hyperons
or deltas. As a result, it becomes more difficult to obtain
very massive neutron stars consistent with the observational
constraints. In the present work, the density dependence of the
$\omega$ meson coupling was kept fixed in the parametrizations and only
a few choices for the energy dependence were tested. Only constraints near
the nuclear saturation density were used to determine the model parameters.
There remains
substantial freedom in the model to be explored in order to find
a suitable parametrization that is consistent with the optical potential and
maximum neutron mass constraint.
Nevertheless, the results of the our study indicate that the optical
potential constraint has to be taken seriously into account in the
development of realistic phenomenological models for dense matter.
RMF models with self-energies that explicitly depend on the
nucleon momentum or energy can be applied to the simulation of
heavy-ion collisions using relativistic transport approaches. Here the
equation of state can be tested at supra-saturation densities. It is well known
that an energy/momentum dependence of the effective in-medium interaction is mandatory
for a proper description and analysis of experimental data.
Such an approach can help to constrain the parameters of the present model at densities
that are not accessible in the description of finite nuclei.
In our work, an explicit energy dependence of the
nucleon-meson couplings was favored. It allows to apply the DD-NLD
approach to the description of nuclei without major difficulties.
Such an additional investigation of the model will permit a better
control on the parameters. In particular, the interplay between the
choice of the functional form of the energy dependence, the cutoff parameters and the
density dependence of the couplings can be studied with
its impact for the prediction of maximum neutron star masses.
With a larger number of observables as constraints
it can be expected that firmer conclusions can be drawn on the compatibility of
nuclear matter and finite nuclei descriptions in the DD-NLD model.
Work in this direction is in progress.
\section*{Acknowledgements}
This work was supported by the Helmholtz Association (HGF) through the
Nuclear Astrophysics Virtual Institute (NAVI, VH-VI-417). S.A. acknowledges
support from the Helmholtz Graduate School for Hadron and Ion Research
(HGS-HIRe for FAIR).
|
\section{Introduction}\label{sec:intro}
Recall that if $L$ is a locally compact topological group acting continuously on a locally Hausdorff topological space $M$ then this action is called {\it \textbf{proper}} if for every compact subset $C \subset M$ the set
$$L(C):=\{ g\in L \ | \ g\cdot C \cap C \neq \emptyset \}$$
is compact. If $L$ is discrete and acts properly on $M$ then we say that L acts {\it \textbf{properly discontinuously}} on $M.$ We will restrict our attention to the case where $M=G/H$ is a homogeneous space of reductive type and $L \subset G.$ In more detail we assume that $G$ is a linear connected reductive real Lie group with the Lie algebra $\mathfrak{g}$ and $H\subset G$ is a closed subgroup of $G$ with finitely many connected components. Also let $\mathfrak{h}$ be the Lie algebra of $H.$
\begin{definition}
The subgroup $H$ is reductive in $G$ if $\mathfrak{h}$ is reductive in $\mathfrak{g},$ that is, there exists a Cartan involution $\theta $ of $\mathfrak{g}$ for which $\theta (\mathfrak{h}) = \mathfrak{h}.$
The space $G/H$ is called the homogeneous space of reductive type.
\label{def1}
\end{definition}
The problem of proper actions was studied for example in \cite{kul}, \cite{kob2}, \cite{kob1}, \cite{ben}, \cite{ok}, \cite{kas} and \cite{bt}. We also refer the reader to the excellent survey \cite{ko}.
In the present paper we would like to examine the connection between the following three conditions
\begin{itemize}
\item C1 := the space $G/H$ admits a properly discontinuous action of an infinite subgroup of $G,$
\item C2 := the space $G/H$ admits a properly discontinuous action of a non-virtually abelian infinite subgroup of $G,$
\item C3 := the space $G/H$ admits a proper action of a subgroup $L\subset G$ locally isomorphic to $SL(2,\mathbb{R}),$
\end{itemize}
(a discrete group is non-virtually abelian if it does not contain an abelian subgroup of finite index). Notice that if $H$ is compact then every closed subgroup of $G$ acts discontinuously on $G/H.$ For example the mentioned three conditions are equivalent:
$$\text{\rm C1} \Leftrightarrow \text{\rm C2} \Leftrightarrow \text{\rm C3}.$$
However this is not the case with non-compact $H.$ Indeed let
$$\mathfrak{g}=\mathfrak{k}+\mathfrak{p}$$
be a Cartan decomposition of $\mathfrak{g}$ induced by a Cartan involution $\theta$ and choose a maximal abelian subspace $\mathfrak{a}$ of $\mathfrak{p}.$ One can show that all maximal abelian subspaces of $\mathfrak{p}$ are conjugate and therefore we can define the \textbf{\textit{real rank}} of $\mathfrak{g}$ by
$$\text{\rm rank}_{\mathbb{R}}\mathfrak{g} := \text{\rm dim} (\mathfrak{a}).$$
In this setting we can state the famous Calabi-Markus phenomenon proved by T. Kobayashi (c.f. Corollary 4.4 in \cite{kob2})
\begin{center}
$G/H$ fulfills the condition C1 if and only if $\text{\rm rank}_{\mathbb{R}}\mathfrak{g} > \text{\rm rank}_{\mathbb{R}}\mathfrak{h}.$
\end{center}
Take $K\subset G$ the maximal compact subgroup corresponding to $\mathfrak{k}.$ The group $K$ acts on $\mathfrak{p}$ by the adjoint representation. Define the Weyl group $W_{g}:=N_{K}(\mathfrak{a})/Z_{K}(\mathfrak{a})$ where $N_{K}(\mathfrak{a})$ (resp. $Z_{K}(\mathfrak{a})$) denotes the normalizer (resp. centralizer) of $\mathfrak{a}$ in $K.$ Then $W_{g}$ acts on $\mathfrak{a}$ by orthogonal transformations and $W_{g}$ is isomorphic to the finite reflection group generated by the restricted root system of $\mathfrak{g}$ (see Sections \ref{sec:resroot} and \ref{sec:ref}). Let $w_{0} \in W_{g}$ be the longest element. Put
\begin{equation}
\mathfrak{b}:= \{ X\in \mathfrak{a} \ | \ -w_{0}(X)=X \}.
\label{eq1}
\end{equation}
Also let $\mathfrak{h}=\mathfrak{k}_{h}+\mathfrak{p}_{h}$ be a Cartan decomposition of $\mathfrak{h}$ with respect to $\theta |_{\mathfrak{h}}.$ Notice that we can choose a maximal abelian subspace $\mathfrak{a}_{h}$ of $\mathfrak{p}$ so that $\mathfrak{a}_{h} \subset \mathfrak{a}.$
We have the following.
\begin{theorem}[Theorem 1 in \cite{ben}]
The group $G$ contains a discrete and non-virtually abelian subgroup which acts properly discontinuously on $G/H$ if and only if for every $w$ in $W_{g}$, $w \cdot \mathfrak{a}_{h}$ does not contain $\mathfrak{b}.$
\label{ben}
\end{theorem}
For example the space $SL(3,\mathbb{R})/SL(2,\mathbb{R})$ fulfills the condition C1 but does not fulfill the condition C2 and therefore the condition C3 (see Example 1 in \cite{ben}). Furthermore, consider the following example.
\begin{example}
Take $SO(4,4)/U$ where $U$ has a Lie algebra denoted by $\mathfrak{u}$ and $\mathfrak{u}\cong \mathbb{R}^{2}+\mathfrak{sl}(2,\mathbb{R}).$ The embedding $U\hookrightarrow G$ is induced by $\mathfrak{u} \hookrightarrow \mathfrak{so}(4,4)$ and is described in the Appendix.
\label{ex1}
\end{example}
This space fulfills the condition C2 but not C3. Thus for a space of reductive type we only have
$$C1 \Leftarrow C2 \Leftarrow C3.$$
The aim of this paper is to show that one can impose a condition on ``regularity'' of the embedding $H \hookrightarrow G$ so that conditions C2 and C3 become equivalent and close to C1. Let $\mathfrak{t}$ be a maximal abelian subspace of $\mathfrak{z}_{k}(\mathfrak{a}):= \{ Y\in \mathfrak{k} \ | \ \forall_{X\in \mathfrak{a}} \ [Y,X]=0 \}.$ Then
$$\mathfrak{j}:=\mathfrak{t}+\mathfrak{a}$$
is a Cartan subalgebra of $\mathfrak{g}$ called the \textbf{\textit{split Cartan subalgebra}} (for an arbitrary $\theta$ stable Cartan subalgebra $\mathfrak{w}$ the dimension $\text{\rm dim}(\mathfrak{a}\cap \mathfrak{w})$ is called the \textit{non-compact dimension} of the Cartan subalgebra, therefore a Cartan subalgebra is split if it is of maximal non-compact dimension). Analogously define a split Cartan subalgebra $\mathfrak{j}_{[h,h]}$ of $[\mathfrak{h},\mathfrak{h}].$ We propose the following definition.
\begin{definition}
The space $G/H$ is strongly regular if $\mathfrak{j}_{[h,h]}\subset \mathfrak{j} \subset \mathfrak{n}_{g}([\mathfrak{h} ,\mathfrak{h}])$ where $\mathfrak{n}_{g}([\mathfrak{h} ,\mathfrak{h} ]):=\{ Y\in \mathfrak{g} \ | \ \forall_{X\in [\mathfrak{h} ,\mathfrak{h} ]} \ [Y,X]\in [\mathfrak{h}, \mathfrak{h} ] \}$ is a normalizer of $[\mathfrak{h} ,\mathfrak{h} ]$ in $\mathfrak{g}.$
\end{definition}
\begin{remark}
The space $G/H$ is strongly regular if and only if $\mathfrak{j}_{[h,h]}\subset \mathfrak{j}$ (up to conjugation) and $[\mathfrak{j},[\mathfrak{h},\mathfrak{h}]]\subset \mathfrak{h}.$
\end{remark}
Examples of strongly regular spaces include (but are not limited to) spaces of parabolic type (i.e. spaces where $H$ is a semisimple part of the Levi factor of any parabolic subgroup of $G$), spaces induced by regular complex spaces (i.e. spaces where $G$ and $H$ are split real forms of $G^{\mathbb{C}}$ and its regular subgroup $H^{\mathbb{C}},$ respectively), and some k-symmetric spaces (spaces where $H$ is an isotropy subgroup induced by an automorphism of $G$ of order $k$). For more details see Section \ref{sec:strreg}.
\begin{example}
Take $a,b,c \in \mathbb{N}$ so that $0<c<a<b.$ The following spaces are strongly regular (notation is close to \cite{ov2})
$$SL(b,\mathbb{R})/SL(a,\mathbb{R}), \ Sp(b,\mathbb{R})/Sp(a,\mathbb{R}), \ SO(b,b)/SO(a,a),$$
$$ SO(b,b+1)/SO(a,a+1), \ SO(b,b)/SL(a,\mathbb{R}), \ SO(b,b+1)/SL(a,\mathbb{R}), $$
$$SO(a,b)/SL(c,\mathbb{R}), \ SO(2a,2b)/SO(2c,2b), \ SO(2a,2b+1)/SO(2c,2b+1), $$
$$ SU(a,b)/SU(c,b), \ SU(a,b)/SL(c,\mathbb{C}), \ SU^{\ast}(2b+2)/SU^{\ast}(2a+2), $$
$$Sp(a,b)/SU^{\ast}(2c+2), \ E_{8}^{VIII}/SO(6,6), \ E_{8}^{IX}/E_{7}^{VII}, $$
$$ E_{6}^{III}/SU(1,6), \ F_{4}^{I}/Sp(3,\mathbb{R}), \ F_{4}^{I}/SL(3,\mathbb{R}).$$
\end{example}
Notice that the notion of strong regularity is a real analogue of the definition of regular subgroup in the complex case.
Define the \textbf{\textit{a-hyperbolic rank}} of a real reductive Lie algebra as
$$\text{\rm rank}_{a-hyp}\mathfrak{g}:=\text{\rm dim}\mathfrak{b},$$
where $\mathfrak{b}$ is given by (\ref{eq1}). The a-hyperbolic rank can be easily calculated using Table \ref{tab1} (see Section \ref{sec:ahyp}).
\begin{definition}
We say that $G$ is of inner type if $\text{\rm rank}_{a-hyp}\mathfrak{g}=\text{\rm rank}_{\mathbb{R}}\mathfrak{g}.$
\end{definition}
\begin{remark}
The group $G$ is of inner type if and only if $\text{\rm rank}\mathfrak{k}=\text{\rm rank}\mathfrak{g}.$
\end{remark}
The main result of this paper is stated in the following theorem.
\begin{theorem}
Let $G/H$ be a strongly regular space such that $H$ is of inner type. Then
$$\text{\rm C2} \Longleftrightarrow \text{\rm C3} \Longleftrightarrow \text{\rm rank}_{a-hyp}\mathfrak{g}> \text{\rm rank}_{a-hyp}\mathfrak{h}.$$
\label{twg}
\end{theorem}
\begin{corollary}
If $G/H$ is a strongly regular homogeneous space such that $G$ and $H$ are of inner type then
$$\text{\rm C1} \Longleftrightarrow \text{\rm C2} \Longleftrightarrow \text{\rm C3} \Longleftrightarrow \text{\rm rank}_{\mathbb{R}}\mathfrak{g}> \text{\rm rank}_{\mathbb{R}}\mathfrak{h}.$$
\end{corollary}
Taking into account the Calabi-Markus phenomenon we obtain the following characterization of strongly regular spaces $G/H$ for which $H$ is of inner type
\begin{itemize}
\item $\text{\rm C1} \Longleftrightarrow \text{\rm rank}_{\mathbb{R}}\mathfrak{g}> \text{\rm rank}_{\mathbb{R}}\mathfrak{h},$
\item $\text{\rm C2} \Longleftrightarrow \text{\rm rank}_{a-hyp}\mathfrak{g}> \text{\rm rank}_{\mathbb{R}}\mathfrak{h},$
\item $\text{\rm C3} \Longleftrightarrow \text{\rm rank}_{a-hyp}\mathfrak{g}> \text{\rm rank}_{\mathbb{R}}\mathfrak{h}.$
\end{itemize}
\noindent
Notice that the assumption of $H$ being of inner type in Theorem \ref{twg} is important. Consider the space from Example \ref{ex1}. The space $SO(4,4)/U$ fulfills the condition of strong regularity: $\mathfrak{j}_{[u,u]}\subset \mathfrak{j}_{so(4,4)} \subset \mathfrak{n}_{so(4,4)}([\mathfrak{u} ,\mathfrak{u}]).$ But
$$1=\text{\rm rank}_{a-hyp}\mathfrak{u}< \text{\rm rank}_{\mathbb{R}}\mathfrak{u}=3.$$
\section{Preliminaries}
In this section we introduce a notation and properties used in the proof of Theorem \ref{twg}. We retain the notation from the previous section.
\subsection{Restricted roots and the Weyl group}\label{sec:resroot}
For more information on this subject please refer to Section 4 in Chapter 4 of \cite{ov}.
Let $\mathfrak{g}^{\mathbb{C}}$ and $\mathfrak{j}^{\mathbb{C}}$ be complexifications of the reductive Lie algebra $\mathfrak{g}$ and the split Cartan subalgebra $\mathfrak{j}=\mathfrak{t}+\mathfrak{a} \subset \mathfrak{g},$ respectively. Denote by $\Delta_{gc}$ the system of roots for $\mathfrak{g}^{\mathbb{C}}$ with respect to $\mathfrak{j}^{\mathbb{C}}.$ Then
$$\mathfrak{g}^{\mathbb{C}}=\mathfrak{j}^{\mathbb{C}}+\sum_{\beta \in \Delta_{gc}}\mathfrak{g}^{c}_{\beta}.$$
Consider a subspace
$$\mathfrak{j}^{\mathbb{C}}(\mathbb{R}):=i\mathfrak{t}+ \mathfrak{a} \subset \mathfrak{j}^{\mathbb{C}}$$
and let $\mathfrak{j}^{\mathbb{C}}(\mathbb{R})^{\ast}$ be the dual space with respect to the Killing form of $\mathfrak{g}^{\mathbb{C}}$. Denote by
\begin{equation}
r:\mathfrak{j}^{\mathbb{C}}(\mathbb{R})^{\ast} \rightarrow \mathfrak{a}^{\ast}
\label{eq2}
\end{equation}
the restriction map and define
$$\Sigma_{g} \cup \{ 0 \} := r(\Delta_{gc} \cup \{ 0 \} ).$$
Then
\begin{equation}
\mathfrak{g}=\mathfrak{j}+\sum_{\alpha \in \Sigma_{g}}\mathfrak{g}_{\alpha}, \ \mathfrak{g}_{\alpha}=\mathfrak{g} \cap \sum\limits_{\substack{\beta \in \Delta_{gc} \\ r(\beta)=\alpha}}\mathfrak{g}_{\beta}^{c}
\label{eq35}
\end{equation}
is a root space decomposition for $\mathfrak{g}.$ It follows that for $\mathfrak{g}_{0} := \mathfrak{j}$ and $\alpha_{1},\alpha_{2} \in \Sigma_{g} \cup \{ 0 \}$
\begin{equation}
[\mathfrak{g}_{\alpha_{1}},\mathfrak{g}_{\alpha_{2}}] \subset \mathfrak{g}_{\alpha_{1} +\alpha_{2}}, \ \ \text{\rm where} \ \ \mathfrak{g}_{\alpha_{1} +\alpha_{2}} \neq \{ 0 \} \ \ \text{\rm iff} \ \ \alpha_{1}+\alpha_{2} \in \Sigma_{g} \cup \{ 0 \}.
\label{eq3}
\end{equation}
The Weyl group $W_{g}$ of $\mathfrak{g}$ is the finite group of orthogonal transformations of $\mathfrak{a}$ generated by reflections by hyperplanes $C_{\alpha}:=\{ X\in \mathfrak{a} \ | \ \alpha (X)=0 \}$ for $\alpha \in \Sigma_{g}.$ One can prove the following.
\begin{proposition}[Proposition 4.2, Ch. 4 in \cite{ov}]
The group $W_{g}$ coincides with the group of transformations induced by automorphisms $Ad_{k}$ $(k\in N_{K}(\mathfrak{a}))$ and also with the group of transformations induced by automorphisms $Ad_{g}$ $(g\in N_{G}(\mathfrak{a})).$ Therefore
$$W_{g}\cong N_{K}(\mathfrak{a})/Z_{K}(\mathfrak{a})\cong N_{G}(\mathfrak{a})/Z_{K}(\mathfrak{a}).$$
\end{proposition}
\subsection{Strongly regular homogeneous spaces}\label{sec:strreg}
In this subsection we would like to introduce the notion of a \textit{\textbf{strongly regular}} homogeneous space $G/H.$ For the simplicity we will assume (for this subsection only) that $\mathfrak{h}$ is semisimple.
Recall that the complex semisimple Lie subalgebra $\mathfrak{d}$ of complex semisimple Lie algebra $\mathfrak{e}$ is called \textit{\textbf{regular}} with respect to a Cartan subalgebra $\mathfrak{s}$ of $\mathfrak{e}$ if $\mathfrak{s} \subset \mathfrak{n}_{e}(\mathfrak{d}).$ In this setting if
$$\mathfrak{e}=\mathfrak{s}+ \sum_{\beta \in \Delta_{e}}\mathfrak{e}_{\beta}$$
then
$$\mathfrak{d}=\mathfrak{s}_{d}+\sum_{\beta \in \Delta_{d}}\mathfrak{e}_{\beta},$$
where $\Delta_{d}$ is a closed and symmetric subsystem of $\Delta_{e}$ and
$$\mathfrak{s}_{d}:=\text{\rm Span}(\{ [\mathfrak{e}_{\beta},\mathfrak{e}_{-\beta}] \ | \ \beta \in \Delta_{d} \})$$
is a Cartan subalgebra of $\mathfrak{d}.$ (see Proposition 1.1, Ch. 6 in \cite{ov}). Recall that
\begin{definition}
A subsystem $\Delta_{d}$ of $\Delta_{e}$ is closed if for every $\beta_{1} , \beta_{2} \in \Delta_{d}$
$$\beta_{1}+ \beta_{2} \in \Delta_{e} \ \Longrightarrow \ \beta_{1}+\beta_{2} \in \Delta_{d}.$$
It is symmetric if for every $\beta \in \Delta_{d}$ we have $-\beta \in \Delta_{d} .$
\end{definition}
Let $\mathfrak{c}_{d}$ be the orthogonal complement to $\mathfrak{s}_{d}$ in $\mathfrak{s}$ with respect to the Killing form. Notice that
$$\forall_{\beta \in \Delta_{d}} \ \forall_{C\in \mathfrak{c}_{d}, X\in \mathfrak{g}_{\beta}} \ [C,X]=\beta (C) X=0,$$
as $C\in C_{\beta}.$ Therefore
\begin{equation}
[\mathfrak{c}_{d},\mathfrak{d}]=0.
\label{eq77}
\end{equation}
In the real case we cannot just assume that \textit{some} Cartan subalgebra of $\mathfrak{g}$ is contained in $\mathfrak{n}_{g}(\mathfrak{h})$ because not all Cartan subalgebras in $\mathfrak{g}$ are conjugate. Therefore we modify our requirement to
$$\mathfrak{j}_{h}\subset \mathfrak{j} \subset \mathfrak{n}_{g}(\mathfrak{h})$$
(recall that $\mathfrak{j}_{h}=\mathfrak{t}_{h}+\mathfrak{a}_{h}$ and $\mathfrak{j}=\mathfrak{t}+\mathfrak{a}$ are split-Cartan subalgebras of $\mathfrak{h}$ and $\mathfrak{g},$ respectively). Also let
$$\mathfrak{g}=\mathfrak{j}+\sum_{\alpha \in \Sigma_{g}}\mathfrak{g}_{\alpha}$$
be the root space decomposition for $\mathfrak{g}.$ In this setting we obtain the following result.
\begin{lemma}
Let $G/H$ be a strongly regular homogeneous space. Then
\begin{equation}
\mathfrak{h}=\mathfrak{j}_{h}+\sum_{\alpha \in \Sigma_{h}}\mathfrak{h}_{\alpha}
\label{eq4}
\end{equation}
is a root space decomposition for $\mathfrak{h}$ such that $\Sigma_{h} \subset \Sigma_{g}$ is a subsystem and so by (\ref{eq35}) we have $\mathfrak{h}_{\alpha}\subset \mathfrak{g}_{\alpha}$ for $\alpha \in \Sigma_{h}.$ Also
$$\mathfrak{a}_{h}=\text{\rm Span}(\{ [\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}] \ | \ \alpha \in \Sigma_{h} \}).$$
\label{lg}
\end{lemma}
\begin{proof}
First notice that $\mathfrak{h}^{\mathbb{C}}$ is a regular subalgebra of $\mathfrak{g}^{\mathbb{C}}$ with respect to $\mathfrak{j}^{\mathbb{C}}.$ Therefore
$$\mathfrak{h}^{\mathbb{C}}=\hat{\mathfrak{j}}^{\mathbb{C}}_{h}+\sum_{\beta \in \Delta_{hc}}\mathfrak{g}_{\beta}^{c},$$
where $\Delta_{hc}$ is a subsystem of $\Delta_{gc}$ and $\hat{\mathfrak{j}}_{h}^{\mathbb{C}} \subset \mathfrak{j}^{\mathbb{C}}$ is some Cartan subalgebra of $\mathfrak{h}^{\mathbb{C}}.$ It follows from the assumption of strong regularity that $\mathfrak{j}_{h}^{\mathbb{C}} \subset \mathfrak{j}^{\mathbb{C}}.$ Also $\hat{\mathfrak{j}}_{h}^{\mathbb{C}}$ and $\mathfrak{j}_{h}^{\mathbb{C}}$ as Cartan subalgebras of the complex semisimple lie algebra $\mathfrak{h}^{\mathbb{C}}$ are conjugate by an element $w \in H^{\mathbb{C}} \subset G^{\mathbb{C}}.$
Assume for the moment that $w\in W_{gc}.$ It follows that for $w\hat{\mathfrak{j}}_{h}^{\mathbb{C}}=\mathfrak{j}_{h}^{\mathbb{C}}$ we have
$$\mathfrak{h}^{\mathbb{C}} = w\mathfrak{h}^{\mathbb{C}}=w\hat{\mathfrak{j}}_{h}^{\mathbb{C}}+w\sum_{\beta \in \Delta_{hc}}\mathfrak{g}_{\beta}^{c}=\mathfrak{j}_{h}^{\mathbb{C}}+\sum_{\beta \in w(\Delta_{hc})}\mathfrak{g}_{\beta}^{c}.$$
Thus we may assume that
$$\hat{\mathfrak{j}_{h}^{\mathbb{C}}}=\mathfrak{j}_{h}^{\mathbb{C}}$$
and therefore
$$\mathfrak{h}^{\mathbb{C}}=\mathfrak{j}_{h}^{\mathbb{C}}+\sum_{\beta \in \Delta_{hc}}\mathfrak{g}_{\beta}^{c}.$$
Now the lemma follows from the definition of the restricted root system and the restriction map (defined by (\ref{eq2})).
We only need to prove that $w\in W_{gc}.$ Let $\mathfrak{c}$ be the orthogonal complement of $\hat{\mathfrak{j}}_{h}^{\mathbb{C}}$ in $\mathfrak{j}^{\mathbb{C}}.$ We have
$$\mathfrak{j}^{\mathbb{C}}=\hat{\mathfrak{j}}_{h}^{\mathbb{C}} + \mathfrak{c}$$
and thus by (\ref{eq77})
$$w\mathfrak{j}^{\mathbb{C}}=w\hat{\mathfrak{j}}_{h}^{\mathbb{C}} + w\mathfrak{c}=\mathfrak{j}_{h}^{\mathbb{C}} + \mathfrak{c} \subset \mathfrak{j}^{\mathbb{C}}.$$
as $w\in H^{\mathbb{C}}.$ Therefore $\mathfrak{j}^{\mathbb{C}}_{h}$ and $\hat{\mathfrak{j}}_{h}^{\mathbb{C}}$ are conjugate by an element of $N_{G^{\mathbb{C}}}(\mathfrak{j}^{\mathbb{C}}).$ Without loss of generality we may assume that this element belongs to $W_{gc}$ because
$$W_{gc}\cong N_{G^{\mathbb{C}}}(\mathfrak{j}^{\mathbb{C}})/Z_{G^{\mathbb{C}}}(\mathfrak{j}^{\mathfrak{j}^{\mathbb{C}}}).$$
\end{proof}
It follows from (\ref{eq35}) and (\ref{eq3}) that if $\mathfrak{h}$ is given by (\ref{eq4}) then it fulfills the condition of strong regularity. Therefore Lemma \ref{lg} can be used to characterize strongly regular homogeneous spaces with semisimple $\mathfrak{h}.$
It also follows from Lemma \ref{lg} how one can construct some easy examples of strongly regular homogeneous spaces. For instance
\begin{enumerate}
\item If $\mathfrak{g}$ and $\mathfrak{h}$ are split real forms of $\mathfrak{g}^{\mathbb{C}}$ and its regular subalgebra $\mathfrak{h}^{\mathbb{C}},$ respectively. Then the root system of $\mathfrak{h}^{\mathbb{C}}$ can be constructed from the extended Dynkin diagram of $\mathfrak{g}^{\mathbb{C}}.$ Since for the split real form we can identify $\Delta_{gc}$ with $\Sigma_{g}$ this yields a class of examples of strongly regular homogeneous spaces (see \cite{ov}, Chapter 6, Section 1.1).
\item Also if $\mathfrak{o}$ is a centralizer in $\mathfrak{g}$ of any subspace in $\mathfrak{a}$ then $\mathfrak{h}:=[\mathfrak{o},\mathfrak{o}]$ induces a strongly regular homogeneous space. In this case the Satake diagram of $\mathfrak{h}$ can be obtain from the Satake diagram of $\mathfrak{g}$ by deleting any subset (different subsets correspond to different centralizers) of white vertexes and all connections and arrows which lead to this subset of white vertexes (see Theorem 1.6 and comments after, Ch. 6 in \cite{ov}).
\end{enumerate}
A more nontrivial example (as we may obtain $\mathfrak{h}_{\alpha} \subsetneq \mathfrak{g}_{\alpha}$) can be constructed as follows. Take a connected semisimple real Lie group $M$ of inner type with the Lie algebra $\mathfrak{m}$. Then there exists a closed subgroup $\tilde{U} \subset M$ locally isomorphic to $SL(2,\mathbb{R}) \times ... \times SL(2,\mathbb{R})=[SL(2,\mathbb{R})]^{\text{\rm rank}\mathfrak{m}}.$ The Lie algebra of $\tilde{U}$ is induced by a set of $\text{\rm rank}\mathfrak{m}$ orthogonal restricted roots generating commuting $\mathfrak{sl}(2,\mathbb{R})$ subalgebras (such set of roots exists since $M$ is of inner type). The resulting homogeneous space $M/\tilde{U}$ is strongly regular.
Also some \textit{k-symmetric spaces} are strongly regular. For instance it is easy to see that a 3-symmetric space $SO(n,n+3)/U(1)\times SO(n,n+1),$ $n\geq 1$ is strongly regular (see \cite{g1} for a classification of 3-symmetric spaces).
\subsection{Finite groups of reflections}\label{sec:ref}
For a more detailed treatment of this subject please refer to Chapter I.1 in \cite{hum}. Recall that we denote by $W_{g}$ the finite reflection group generated by roots in $\Sigma_{g}.$ In more detail if $\alpha \in \Sigma_{g}$ then let $s_{\alpha}$ be a reflection in $\mathfrak{a}$ through a hyperplane
\begin{equation}
C_{\alpha}= \{ X\in \mathfrak{a} \ | \ \alpha (X)=0 \}
\label{eq5}
\end{equation}
with respect to the scalar product given by the restriction of the Killing form of $\mathfrak{g}$ to $\mathfrak{a}$. Let $\Pi_{g} \subset \Sigma_{g}$ be a subset of simple roots. Then $W_{g}$ is generated by \textit{simple reflections} $s_{\alpha}$ for $\alpha \in \Pi_{g}.$ We also have $W_{g}(\Sigma_{g})=\Sigma_{g}.$
For an arbitrary $w\in W_{g}$ let $r$ be the smallest number such that $w=s_{\alpha_{1}}...s_{\alpha_{r}}$ is a product of simple reflections.
\begin{definition}
The number $r$ is called the length of $w$ and is denoted by $l(w):=r.$ By definition $l(e)=0.$
\end{definition}
Using the above definition we can describe a unique element of the Weyl group called the longest element of $W_{g}.$
\begin{definition}[see section 1.8 in \cite{hum}]
The longest element of the Weyl group is the element $w_{0}\in W_{g}$ such that $l(w_{0}w)=l(w_{0})-l(w)$ for any $w\in W_{g}.$
\end{definition}
As a consequence we obtain the following property.
\begin{fact}
The element $w_{0}$ is an involution (that is $w_{0}^{2}=id$). Also $w_{0}(\Pi_{g})=-\Pi_{g}.$
\end{fact}
Define \textbf{\textit{the fundamental Weyl chamber}} $\mathfrak{a}^{+}$ as
$$\mathfrak{a}^{+}=\{ X\in \mathfrak{a} \ | \ \forall_{\alpha \in \Pi_{g}} \ \alpha (X) \geq 0 \}.$$
Then $w_{0}(\mathfrak{a}^{+})=-\mathfrak{a}^{+}$ and
\begin{theorem}[c.f. Theorem 1.12 in \cite{hum}]
The subset $\mathfrak{a}^{+}$ is a fundamental domain of the action of $W_{g}$ on $\mathfrak{a}.$ Also
\begin{enumerate}
\item if $wX_{1}=X_{2}$ for some $X_{1},X_{2}\in\mathfrak{a}^{+}$ and $w\in W_{g}$ then $X_{1}=X_{2}.$ Moreover, if $X_{3}\in \text{\rm Int}(\mathfrak{a}^{+}):=\{ X\in \mathfrak{a} \ | \ \forall_{\alpha \in \Pi_{g}} \ \alpha (X) > 0 \}$ then the isotropy group of $X_{3}$ is trivial.
\item if $U\subset V$ is a subset then the subgroup $W_{U}\subset W_{g}$ fixing $U$ pointwise is generated by those reflections $s_{\alpha}$ which are contained in $W_{U}.$
\end{enumerate}
\label{weyl}
\end{theorem}
We will also need the following lemma.
\begin{lemma}
Let $e\neq w\in W_{g}$ and $S\subset \mathfrak{a}$ be a set that $w$ fixes pointwise. Then there exists $\alpha \in \Sigma_{g}$ such that $S\subset C_{\alpha}$ (where $C_{\alpha}$ is defined by (\ref{eq5})).
\label{ref}
\end{lemma}
\begin{proof}
Let $W_{S}\subset W_{g}$ be the subgroup fixing $S$ pointwise. Then $w\in W_{S}$ so $W_{S}$ is nontrivial. It follows from Theorem \ref{weyl} that there exists a reflection in $W_{S}.$ Take $\alpha$ to be a root generating this reflection.
\end{proof}
\subsection{The a-hyperbolic rank}\label{sec:ahyp}
A-hyperbolic dimensions of simple real Lie algebras can be calculated using data in Table \ref{tab1} (for a detailed description how to calculate the a-hyperbolic rank of a simple Lie algebra please refer to \cite{bt}).
\begin{center}
\begin{table}[h]
\centering
{\footnotesize
\begin{tabular}{| c | c | c |}
\hline
\multicolumn{3}{|c|}{ \textbf{\textit{a-hyperbolic ranks of simple Lie algebras}}} \\
\hline
$\mathfrak{g}$ & $\text{rank}_{a-hyp} (\mathfrak{g})$ & $\text{rank}_{\mathbb{R}} (\mathfrak{g})$ \\
\hline
$\mathfrak{sl}(2k,\mathbb{R})$ & k & 2k-1 \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{sl}(2k+1,\mathbb{R})$ & k & 2k \\
{\scriptsize $k\geq 1$} & & \\
\hline
$\mathfrak{su}^{\ast}(4k)$ & k & 2k-1 \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{su}^{\ast}(4k+2)$ & k & 2k \\
{\scriptsize $k\geq 1$} & & \\
\hline
$\mathfrak{so}(2k+1,2k+1)$ & 2k & 2k+1 \\
{\scriptsize $k\geq 2$} & & \\
\hline
$\mathfrak{e}_{6}^{\text{I}}$ & 4 & 6 \\
\hline
$\mathfrak{e}_{6}^{\text{IV}}$ & 1 & 2 \\
\hline
\end{tabular}
}
\captionsetup{justification=centering}
\caption{
The table contains all simple real Lie algebras $\mathfrak{g},$ for which $\text{rank}_{\mathbb{R}}(\mathfrak{g}) \neq \text{rank}_{a-hyp}(\mathfrak{g})$ (notation is close to Table 9, page 312 of \cite{ov2}).}
\label{tab1}
\end{table}
\end{center}
Table \ref{tab1} can also be used to calculate the a-hyperbolic rank of a reductive real Lie algebra.
\begin{enumerate}
\item The a-hyperbolic rank of a semisimple Lie algebra equals the sum of a-hyperbolic ranks of all its simple parts.
\item The a-hyperbolic rank of a reductive Lie algebra equals the a-hyperbolic rank of its derived subalgebra.
\end{enumerate}
There is a close relation between $\mathfrak{b}$ and the set of antipodal hyperbolic orbits in $\mathfrak{g}.$ Recall that an element $X \in \mathfrak{g}$ is called {\it \textbf{hyperbolic}} if $X$ is semisimple (that is, $ad_{X}$ is diagonalizable) and all eigenvalues of $ad_{X}$ are real.
\begin{definition}
An adjoint orbit $GX:=Ad(G)(X)=\{ gX:=Ad_{g}X \ | \ g\in G \}$ is said to be hyperbolic if $X$ (and therefore every element of $GX$) is hyperbolic. An orbit $GY$ is antipodal if $-Y\in GY$ (and therefore for every $Z\in GY,$ $-Z\in GY$).
\end{definition}
We have the following lemma.
\begin{lemma}
For every hyperbolic orbit $GX$ in $\mathfrak{g}$ the set $GX \cap \mathfrak{a}$ is a single $W_{g}$ orbit in $\mathfrak{a}$.
Furthermore hyperbolic orbit $GY$ in $\mathfrak{g}$ is antipodal if and only if it meets $\mathfrak{b},$ that is
$$\mathfrak{b} \cap GY$$
is nonempty.
\label{lma}
\end{lemma}
\begin{proof}
It is standard to show that $GX \cap \mathfrak{a}$ is a single $W_{g}$ orbit (see for example Proposition 2.4 in \cite{kos}). If $X \in \mathfrak{b}$ then the longest element $w_{0}\in W_{g}$ takes $X$ to $-X$ so $GX$ is antipodal. Conversely if $GX$ is antipodal then it follows from Theorem \ref{weyl} that there exists $Y\in GX$ such that $Y\in \mathfrak{a}^{+}.$ Also $-Y\in -\mathfrak{a}^{+}$ and so $w_{0}(-Y)\in \mathfrak{a}^{+}.$ Again it follows from Theorem \ref{weyl} that $w_{0}(-Y)=Y.$ Thus
$$w_{0}Y=-Y$$
and so $Y\in \mathfrak{b}.$
\end{proof}
\subsection{Criterion of proper actions}\label{sec:prop}
Let $L\subset G$ be a subgroup of reductive type in $G.$ After conjugation by an element of $G$ we may assume that the Lie algebra $\mathfrak{l}$ of $L$ has a split Cartan subalgebra
$$\mathfrak{j}_{l}=\mathfrak{t}_{l}+\mathfrak{a}_{l}$$
such that $\mathfrak{a}_{l}\subset \mathfrak{a}.$ Then the following criterion of proper actions holds.
\begin{theorem}[Theorem 4.1 in \cite{kob2}] The following conditions are equivalent
\\ (i) $H$ acts on $G/L$ properly,
\\ (ii) $L$ acts on $G/H$ properly,
\\ (iii) $\mathfrak{a}_{h} \cap W_{\mathfrak{g}} \mathfrak{a}_{l} = \{ 0 \}.$
\label{tkob}
\end{theorem}
Now assume that $L$ is locally isomorphic to $SL(2,\mathbb{R}).$ Then $\mathfrak{a}_{l}$ is 1-dimensional. On the other hand $\mathfrak{l}\cong \mathfrak{sl}(2,\mathbb{R})$ and so $\mathfrak{l}$ is generated by three vectors $(H,E,F)$ of $\mathfrak{g}$ (called \textbf{\textit{sl(2,R)-triple}}) such that $H$ is semisimple and
$$[H,E]=2E, \ \ [H,F]=-2F \ \text{\rm and} \ [E,F]=H.$$
Thus $\mathfrak{a}_{l}=\mathbb{R}H$ and we can reformulate the above criterion using Lemma \ref{lma}.
\begin{lemma}
$L$ acts properly on $G/H$ if and only if $W_{g}H \cap \mathfrak{a}_{h} = \emptyset .$
\label{proper}
\end{lemma}
\section{Proof of Theorem \ref{twg}}\label{sec:proof}
First notice the following.
\begin{lemma}
Let $B,V_{1},...,V_{n}$ be subspaces of a finite dimensional Euclidean space $V$ and let
$$B\subset \bigcup_{k=1}^{n} V_{k}.$$
Then there exists $j, \ 1\leq j \leq n$ such that $A\subset V_{j}.$
\label{lemlem}
\end{lemma}
\begin{proof}
We will use induction on $n.$ If $n=1$ then $B \subset V_{1}.$ So assume that the lemma is valid for the sum of $n-1$ linear subspaces of $V.$ Let
$$B\subset \bigcup_{k=1}^{n} V_{k}.$$
If there exists $i, \ 1\leq i \leq n$ such that
\begin{equation}
B \subset \bigcup\limits_{k=1,k\neq i}^{n} V_{k}
\label{eqq1}
\end{equation}
then we can apply induction to (\ref{eqq1}). If it is not the case then we can find nonzero vectors $X,Y \in B$ such that
$$X\in \bigcup\limits_{k=1,k\neq i}^{n} V_{k}, \ X\notin V_{i}, \ Y\notin \bigcup\limits_{k=1,k\neq i}^{n} V_{k}, \ Y\in V_{i}.$$
But then for $a\neq 0$ we have $aX+bY \notin V_{i}$ and $aX+bY \in B \subset \bigcup\limits_{k=1}^{n} V_{k}.$ Therefore
$$\{ aX+bY \ | \ a\neq 0 \} \subset \bigcup\limits_{k=1,k\neq i}^{n} V_{k}.$$
After taking closures of this sets we obtain
$$\text{\rm Span}(X,Y) \subset \bigcup\limits_{k=1,k\neq i}^{n} V_{k} \ \Rightarrow \ Y\in \bigcup\limits_{k=1,k\neq i}^{n} V_{k}.$$
A contradiction.
\end{proof}
Let $w_{0}$ and $w_{0}^{h}$ be the longest elements for $W_{g}$ and $W_{h},$ respectively. Define $\mathfrak{b}$ and $\mathfrak{b}_{h}$ as in (\ref{eq1}). We can assume that
\begin{equation}
\mathfrak{b}_{h}\subset \mathfrak{b}.
\label{eq8}
\end{equation}
Indeed if $HX$ is an antipodal hyperbolic orbit in $\mathfrak{h},$ $X \in \mathfrak{b}_{h}\subset \mathfrak{a}_{h} \subset \mathfrak{a},$ then $GX$ is an antipodal hyperbolic orbit in $\mathfrak{g}.$ Therefore there exists $g\in G$ such that
$$gX\in \mathfrak{b} \subset \mathfrak{a}.$$
It follows from Lemma \ref{lma} that there exists $w\in W_{g}$ for which
$$wX \in \mathfrak{b} \ \Rightarrow X\in w^{-1}\mathfrak{b}.$$
We obtain
$$\mathfrak{b}_{h} \subset \bigcup_{w\in W_{g}} w\mathfrak{b}.$$
Since $W_{g}$ is finite it follows from Lemma \ref{lemlem} that there exists $w_{1}\in W_{g}$ with the following property: $\mathfrak{b}_{h} \subset w_{1}\mathfrak{b}.$ After conjugating $\mathfrak{h}$ by $w_{1}$ we obtain (\ref{eq8}).
\textbf{Step 1}: $C2\Rightarrow C3$
Since $\mathfrak{b}_{h}\subset \mathfrak{b}$ it follows from Theorem \ref{ben} that $\text{\rm dim}\mathfrak{b}_{h}< \text{\rm dim}\mathfrak{b}.$ Also, since $G/H$ is a strongly regular space we can assume that $\Sigma_{h} \subset \Sigma_{g}$ (see Lemma \ref{lg}). Thus $W_{h}\subset W_{g}$ and $w_{0}^{h}\neq w_{0}.$
Since $H$ is of inner type we have $\mathfrak{a}_{h}=\mathfrak{b}_{h}.$ Therefore
$$\forall_{X\in \mathfrak{a}_{h}} \ w_{0}w_{0}^{h}(X)=w_{0}(-X)=X$$
and so $\mathfrak{a}_{h}$ is a set that $w_{0}w_{0}^{h}\in W_{g}$ fixes pointwise. It follows from Lemma \ref{ref} that we can choose $\alpha \in \Sigma_{g}$ so that
\begin{equation}
\alpha (\mathfrak{a}_{h})\equiv 0.
\label{eq10}
\end{equation}
For every root $\alpha \in \Pi_{g}$ construct a sl(2,R)-triple $( H_{\alpha}, E_{\alpha}, F_{-\alpha} )$ so that $E_{\alpha} \in \mathfrak{g}_{\alpha},$ $F_{-\alpha} \in \mathfrak{g}_{-\alpha}$ and $H_{\alpha}\in [\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}].$ Define $H \in \mathfrak{a}$ by: $\alpha (H)=2$ for every $\alpha \in \Pi_{g}$ and $H\in [\mathfrak{g},\mathfrak{g}].$ Since $\{ H_{\alpha} \ | \ \alpha \in \Pi_{g} \}$ spans $\mathfrak{a} \cap [\mathfrak{g},\mathfrak{g}]$ we can find $\{ a_{\alpha} \ | \ \alpha \in \Pi_{g} \} \subset \mathbb{R}$ so that
$$H= \sum_{\alpha \in \Pi_{g}} a_{\alpha}H_{\alpha}.$$
Set
$$E:=\sum_{\alpha \in \Pi_{g}}E_{\alpha}, \ \ F:=\sum_{\alpha \in \Pi_{g}} a_{\alpha}F_{-\alpha}.$$
Then $$[H,E] = \sum\limits_{\alpha \in \Pi_{g}} \alpha (H) E_{\alpha}=2E,$$
$$[H,F] = \sum\limits_{\alpha \in \Pi_{g}} a_{\alpha}(-\alpha (H)) F_{-\alpha}=-2F,$$
\begin{equation}
[E,F] = \sum\limits_{\alpha ,\beta \in \Pi_{g}} a_{\beta}[E_{\alpha},F_{-\beta}]=\sum\limits_{\alpha \in \Pi_{g}} a_{\alpha}[E_{\alpha},F_{-\alpha}]=\sum\limits_{\alpha \in \Pi_{g}} a_{\alpha} H_{\alpha}=H,
\label{eq11}
\end{equation}
and (\ref{eq11}) follows from (\ref{eq3}) since the difference of two simple roots is never a root. Therefore $( H,E,F )$ defines a sl(2,R)-triple in $\mathfrak{g}.$
We will show that the subgroup $L\subset G$ induced by $\mathfrak{l}=\text{\rm Span}(H,E,F)\cong \mathfrak{sl}(2,\mathbb{R})$ acts properly on $G/H.$ By Lemma \ref{proper} we only need to show that
$$W_{g}H \cap \mathfrak{a}_{h} = \emptyset .$$
Assume that there exists $w_{2}\in W_{g}$ such that
$$w_{2}H \in \mathfrak{a}_{h} \ \Rightarrow \ H\in w_{2}^{-1}\mathfrak{a}_{h}.$$
By (\ref{eq10}) $(w_{2}^{-1}\alpha) (H)=0$ and so $s_{w_{2}^{-1}\alpha}H=H$. By construction of $(H,E,F)$ we have $H \in \text{\rm Int}(\mathfrak{a}^{+}).$ But Theorem \ref{weyl} implies that the isotropy group of $H$ is trivial - a contradiction.
\textbf{Step 2}: $\text{\rm rank}_{a-hyp}\mathfrak{g}\leq \text{\rm rank}_{a-hyp}\mathfrak{h} \ \Rightarrow \ \neg C3$
It follows from (\ref{eq8}) and the equality $\mathfrak{a}_{h}=\mathfrak{b}_{h}$ that in this case $\mathfrak{b}=\mathfrak{a}_{h}.$ By Theorem $\ref{ben}$ the space $G/H$ does not fulfill the condition $C2.$ Therefore it does not fulfill the condition $C3.$
\textbf{Step 3}: $\text{\rm rank}_{a-hyp}\mathfrak{g} > \text{\rm rank}_{a-hyp}\mathfrak{h} \ \Rightarrow \ C2$
It follows that $\text{\rm dim}\mathfrak{b} > \text{\rm dim}\mathfrak{a}_{h}.$ Thus for any $w\in W_{g},$ $w\mathfrak{a}_{h}$ does not contain $\mathfrak{b}.$ Our claim follows from Theorem \ref{ben}.
\section{Appendix}\label{sec:appen}
Please recall that the split Cartan subalgebra of $\mathfrak{so}(4,4)$ can be identified with the Euclidean space $\mathbb{R}^{4}.$ Let $\{ e_{1}, e_{2}, e_{3}, e_{4} \}$ be a standard orthonormal basis of $\mathbb{R}^{4}.$ Denote by $(\mathbb{R}^{4})^{\ast}$ the dual space of $\mathbb{R}^{4}.$ Then
$$\Sigma_{so(4,4)}=\{ \pm e_{i} \pm e_{j} \ | \ 1\leq i < j \leq 4 \} \subset (\mathbb{R}^{4})^{\ast}.$$
The Weyl group $W_{so(4,4)}$ acts on $\mathbb{R}^{4}$ by permuting coordinates of a vector and changing the sign of an even number of coordinates of a vector.
As is shown in Example 5.3.7 in \cite{cmg} semisimple elements of all (up to conjugation) sl(2,R)-triples in $\mathfrak{so}(4,4)$ are given in the first column of Table \ref{tab2}. The second column of this table shows a transformation of a given semisimple element (denoted by $w(H)$) by some element of the Weyl group $W_{so(4,4)}.$ Take $a=(3,1,0,2),$ $b=(2,0,0,1)$ and $c=(0,0,1,0).$ The third column of Table \ref{tab2} presents $w(H)$ as a linear combination of vectors $a,b,c.$
\begin{center}
\begin{table}[h]
\centering
{\footnotesize
\begin{tabular}{| c | c | c |}
\hline
Semisimple elements $W$ & $w(H)$ & Linear combination of $a,b,c$ \\
\hline
$(6,4,2,0)$ & $(6,2,0,4)$ & $2a$ \\
\hline
$(4,2,2,0)$ & $(4,0,2,2)$ & $2(b+c)$ \\
\hline
$(3,3,1,1)$ & $(-3,1,3,-1)$ & $a-3b+3c$ \\
\hline
$(3,3,1,-1)$ & $(-3,1,-3,-1)$ & $a-3b-3c$ \\
\hline
$(4,2,0,0)$ & $(4,0,0,2)$ & $2b$ \\
\hline
$(2,1,1,0)$ & $(2,0,1,1)$ & $b+c$ \\
\hline
$(1,1,1,1)$ & $(1,1,1,1)$ & $a-b+c$ \\
\hline
$(1,1,1,-1)$ & $(1,1,-1,1)$ & $a-b-c$ \\
\hline
$(2,0,0,0)$ & $(0,0,2,0)$ & $2c$ \\
\hline
$(1,1,0,0)$ & $(-1,1,-0,0)$ & $a-2b$ \\
\hline
$(0,0,0,0)$ & $(0,0,0,0)$ & $0$ \\
\hline
\end{tabular}
}
\captionsetup{justification=centering}
\caption{
Semisimple elements of nilpotent orbits in $\mathfrak{so}(4,4).$
}
\label{tab2}
\end{table}
\end{center}
Therefore any $W_{g}$ orbit of a semisimple element of any sl(2,R) triple in $\mathfrak{so}(4,4)$ meets $\mathfrak{u}_{1}:=\text{\rm Span(a,b,c)}.$ Take the root $\alpha:=-e_{1}+e_{2}$ and let $\mathfrak{u}^{+},$ $\mathfrak{u}^{-}$ be root spaces corresponding to $\alpha$ and $-\alpha,$ respectively. We have $[\mathfrak{u}^{+},\mathfrak{u}^{-}]=\text{\rm Span}(-1,1,0,0)\subset \mathfrak{u}_{1}.$ Put
$$\mathfrak{u}:=\mathfrak{u}_{1}+\mathfrak{u}^{+}+\mathfrak{u}^{-}.$$
One easily sees that $\mathfrak{u}$ is a reductive subalgebra of $\mathfrak{so}(4,4)$ isomorphic to $\mathbb{R}^{2}+\mathfrak{sl}(2,\mathbb{R}).$ Moreover
$$\text{\rm rank}_{\mathbb{R}}\mathfrak{u}=3=\text{\rm dim}\mathfrak{u}_{1} \ \text{\rm and} \ \text{\rm rank}_{a-hyp}\mathfrak{u}=1,$$
as $\mathfrak{u}_{1}$ is a split Cartan subalgebra of $\mathfrak{u}.$ Let $U\subset SO(4,4)$ be the corresponding closed subgroup. It follows from Lemma \ref{proper} and the construction of $\mathfrak{u}$ that $G/U$ does not admit proper action of $SL(2,\mathbb{R}).$ Also
$$\text{\rm rank}_{a-hyp}\mathfrak{so}(4,4)=4=\text{\rm dim} \mathfrak{b}_{so(4,4)}> \text{\rm dim}\mathfrak{u}_{1}$$
therefore it follows from Theorem \ref{ben} that $G/U$ admits a properly discontinuous action of an infinite non-virtually abelian subgroup of $G.$
|
\section{Introduction\label{s:intro}}
The identification of high-energy (Regge) dynamics in QCD processes has been
long since theoretically investigated. With the advent of high-energy colliders
like HERA, Tevatron and LHC such studies have become possible at experimental
level too. The processes which are expected to be more sensitive to this
peculiar dynamical regime, where the center-of-mass energy $\sqrt{s}$ is much
larger than all hard scales involved in the scattering, are the so-called {\em
Mueller-Navelet} (MN) {\em jets} at hadron-hadron colliders~\cite{MuNa87} and
{\em forward jets} at electron-hadron colliders~\cite{Mue91}. Both of them are
defined by the presence of QCD jets at large rapidity, accompanied by any
hadronic activity which is inclusively collected in the central region.
Such processes can be theoretically described by factorization formulae which
involve several ingredients: the partonic distribution functions (PDFs) of the
incoming hadron(s), the gluon Green's function (GGF) describing the high-energy
dynamics of emitted and exchanged partons --- mostly (reggeized) gluons --- and
finally the so-called jet vertices, describing the production of a forward jet
from the interaction of one incoming parton and a reggeized gluon. In the case
of an incoming electron, an additional quantity, the photon impact factor, has
to be considered too.
At present, all ingredients are known at next-to-leading level in the respective
parameters: the PDFs which resum logarithms of collinear type, the GGF resumming
logarithms of the energy, and the jet vertices (and impact-factors) which are
computed at finite perturbative order.
Focusing for definiteness on MN jets in hadron-hadron collision, the
process-dependent part of the cross section is represented by the jet
vertex~\cite{BaCoVa02}, which depends on the jet variables and the actual jet
algorithm. By following the work of Ivanov and Papa~\cite{IP}, in this paper we
reconsider the computation of the jet vertex in the small-cone approximation
(SCA), namely for jets whose extension in the rapidity-azimuthal angle
$(y,\phi)$-plane is small. In particular we apply their method to determine the
analytic expressions of the jet vertex for two particular choices of jet
algorithms: the {\em kt algorithm}%
\footnote{Here {\em kt algorithm} denotes the whole class of clustering
algorithms based on ref.~\cite{CDSW93}, which may differ in the details of the
recombination scheme and of the resolution variable, like the {\em anti-kt}
and the {\em Cambrigde/Aachen} versions.}
\cite{CDSW93} and the {\em cone algorithm}~\cite{EKS89}. These are the mostly
used algorithms in modern jet phenomenology, in particular the kt one. On the
contrary, the algorithm used in~\cite{IP} can be traced back to the one
considered by Furman~\cite{Fur82} in early studies of QCD radiation, but not
used for practical purposes anymore, being infra-red unsafe.
The aim of our work is twofold: on one hand we want to give a precise estimate
of the error introduced by the small-cone approximation in the description of
QCD observables at high energies, i.e., at large rapidities, so as to possibly
justify its use in phenomenological analyses. On the other hand, we want to
estimate the differences occurring by choosing different jet algorithms for the
same process.
The jet algorithm dependence has already been studied in the past, and a
detailed analysis of the cone and kt algorithms in the SCA was presented in
ref.~\cite{MuVo12} in the context of collinear factorization. Here we carry out
a similar analysis in the framework of high-energy (kt-dependent) factorization,
the basic tool for the description of the Regge regime in perturbative QCD.
One should keep in mind that the small-cone expressions are fully analytic
(before their convolution with the PDFs) and compact, and allow a simple
implementation in numerical codes that run much faster than those with the exact
jet vertices. For this reason the SCA jet vertices have already been
used\cite{CIMP13,CMSS13} in quantitative comparison with available
data\cite{cmsMNjet}. However, the experimental results were extracted by
clustering jets with the kt algorithm, while the SCA jet vertices used for the
theoretical calculation were those obtained with the Furman algorithm by Ivanov
and Papa (FIP). Also the small-cone analysis performed in ref.~\cite{DSW13}
compared calculations with the exact jet vertices in the kt algorithm versus the
small-cone ones in the FIP algorithm.
Starting from these premises, and after reviewing in sec.~\ref{s:ts} the
theoretical setup for the description of Mueller-Navelet jets in terms of the
collinear and high-energy factorization formulae, in sec.~\ref{s:jetalg} we
discuss in detail the differences among the three (kt, cone and FIP) jet
algorithms in the relevant case of two near particles and we determine the
kinematical configurations in the limit of small jet ``radius'' $R$.
We then derive the small-cone jet vertices for the kt and cone algorithms in
sec.~\ref{s:scjv}, by computing their differences with respect to the FIP
algorithm induced by the different kinematical conditions.
In sec.~\ref{s:nc} we perform a numerical study in order to asses the
quantitative difference between the exact and small-cone jet vertex in the kt
algorithm, by comparing the vertices themselves as well as a typical
differential cross section for MN jets and some angular coefficients measuring
the azimuthal decorrelation between the jets. In addition we also determine the
discrepancies of the same quantities induced by a different choice of the jet
algorithm.
We discuss the results in sec.~\ref{s:conc}, where we conclude that the wrong
choice of algorithm causes sizeable errors on the predictions, while the SCA
within the same algorithm provides a good approximation to the exact quantities
and can therefore be used as a valuable tool for a quantitative description of
MN and forward jets.
\section{Theoretical setup\label{s:ts}}
\subsection{Factorization\label{s:fact}}
The process we are considering was suggested long ago by Mueller and
Navelet~\cite{MuNa87} in order to study the high-energy behaviour of QCD. It is
generated by the collision of two hadrons $H_{A,B}$ --- typically (anti)protons --- and is
characterized by the detection in the final state of two hard jets $J_{1,2}$ with
large rapidity separation:
\begin{equation}\label{proc}
H_A + H_B \to J_1 + J_2 + X
\end{equation}
where $X$ represents any additional emission. Each jet
$J_i$ represents a cluster of particles grouped together according to some given jet
algorithm and is described by 3 variables: the rapidity $y_i$, the {\em transverse}
energy $E_i\equiv|\boldsymbol{k}_{J,i}|$ and the azimuthal angle $\phi_i\equiv \arg(\boldsymbol{k}_{J,i})$,
$\boldsymbol{k}_{J,i}$ being the $i$-th jet transverse momentum.
The kinematical region where one expects the high-energy QCD dynamics to play an
important role is given by
\begin{equation}\label{hekin}
s \equiv (p_A+p_B)^2 \gg E_1^2 \sim E_2^2 \gg \Lqcd^2 \;, \qquad
|Y| \equiv |y_1 - y_2| \gg 1 \;,
\end{equation}
where the condition of hard jets ($E_i^2\gg\Lqcd^2$) is imposed for the
applicability of perturbation theory.
In the leading twist approximation, i.e., up to power suppressed correction in
the hard scale parameter $\Lqcd^2/E^2\ll1$, the hadronic cross section $\sigma$
can be factorized in the (longitudinal momentum fraction) convolution of two
partonic distribution functions (PDFs) $f_{a/H}(x)$ and a partonic cross section
$\hat\sigma
\begin{equation}\label{collfact}
\frac{\dif\sigma_{AB}(s)}{\dif J_1 \dif J_2} = \sum_{a,b}\int_0^1 \dif x_1 \dif x_2\;
f_{a/A}(x_1) f_{b/B}(x_2) \frac{\dif\hat\sigma_{ab}(x_1 x_2 s)}{\dif J_1 \, \dif J_2}
\;, \qquad \dif J_i \equiv \dif y_i \,\dif E_i \,\dif\phi_i \;,
\end{equation}
where $a,b\in\{q,g\}$ denote the parton flavours (quark or gluon) and $x_i$ the partonic
momentum fractions w.r.t.\ their parent hadrons.
In turn, in the high-energy Regge regime we are considering, the partonic cross
section for jet production can be factorized in a (transverse momentum)
convolution of (process dependent) jet vertices $V$ and a universal factor $G$
called gluon Green's function (GGF
\begin{equation}\label{hefact}
\frac{\dif\hat\sigma_{ab}(x_1 x_2 s)}{\dif J_1 \, \dif J_2} =
\int\dif^2\boldsymbol{k}_1\,\dif^2\boldsymbol{k}_2 \; V_a(x_1,\boldsymbol{k}_1;J_1) G(x_1 x_2 s, \boldsymbol{k}_1, \boldsymbol{k}_2)
V_b(x_2,\boldsymbol{k}_2;J_2) \;,
\end{equation}
where $\boldsymbol{k}_i$ denotes the (reggeized) gluon transverse momentum flowing out from
the GGF and entering the jet vertex, while $\hat{s}\equiv x_1 x_2 s$ is the
center-of-mass energy squared of the partonic subsystem. The overall
factorization structure is depicted in fig.~\ref{f:fact}.
\begin{figure}[t]
\centering
\includegraphics[width=0.4\linewidth]{factFormula.eps}
\caption{\it Diagrammatic representation of the collinear and high-energy
factorization formula for Mueller-Navelet jet production: $f_{a,b}$
represents the parton densities, $V_{a,b}$ the jet vertices and $G$ the
gluon Green's function.}
\label{f:fact}
\end{figure}
The PDFs are non-perturbative objects that depend also on the renormalization
and factorization scales $\mu_R$, $\mu_F$. While $\mu_R$ is introduced in the
renormalization of UV divergencies, $\mu_F$ enters in the treatment of the IR
collinear divergencies which are absorbed by the PDFs. The $\mu_F$ dependence of
the PDFs is governed by the DGLAP equations~\cite{DGLAP}, and their evolution
kernels (splitting functions) are known at next-to-next-to-leading order (NNLO).
By means of global fits they have been determined in a wide range of the
$(x,\mu_F^2)$ plane.
The GGF is the central object in high-energy QCD, in that it resums the
$\log(s)$ to all orders in perturbation theory. It obeys the BFKL
equation~\cite{BFKL}
\begin{align}\label{bfkl}
\omega G_{\omega}(\boldsymbol{k}_1,\boldsymbol{k}_2) &= \delta^2(\boldsymbol{k}_1+\boldsymbol{k}_2) + \int\dif^2\boldsymbol{k}\;
K(\boldsymbol{k}_1,\boldsymbol{k}) G_{\omega}(\boldsymbol{k},\boldsymbol{k}_2) \\
G(\hat{s},\boldsymbol{k}_1,\boldsymbol{k}_2) &=
\int_{-\ui\infty}^{\ui\infty}\frac{\dif\omega}{2\pi\ui} \;
\left(\frac{\hat{s}}{s_0}\right)^\omega G_{\omega}(\boldsymbol{k}_1,\boldsymbol{k}_2) \;,
\label{Gomega}
\end{align}
where $G_\omega$ is the Mellin transform of $G(\hat{s})$, defined in terms of an
arbitrary energy scale $s_0$. The BFKL kernel $K=\as K^{(0)} + \as^2 K^{(1)}$ is
known in next-to-leading-logarithmic (NLL) approximation. The coefficient
$K^{(1)}$ depends on the choice of both $\mu_R$ and $s_0$.
Finally, the jet vertices are perturbative finite objects without energy
($\sqrt{s}$) dependence, and are known in NLO approximation:
$V=\as V^{(0)}+\as^2 V^{(1)}$. They depend on the jet variables $y,E,\phi$ and
also on the arbitrary scales $\mu_R,\mu_F,s_0$, in such a way that the hadronic
cross section be independent of those scales up to NLL terms, i.e., the scale
dependence is present only in the terms of relative order $\as^2(\as\log(s))^n$.
It is apparent that the determination of the Mueller-Navelet (MN) jet cross
section involves quite a number of integrals, both in the factorization
formulae~(\ref{collfact},\ref{hefact}), and in the determination of $G$ and
$V$, as can be checked from their explicit expressions~\cite{BaCoVa02}.
Furthermore, when comparing the theoretical predictions with experiments, one
needs integrated cross sections in some of the jet variables, in order to comply
with the experimental binning. Such integrations are mostly done numerically,
and this would require a large amount of computing resources or time for reaching
a precision at the level of 1\%.
In order to cope with such a problem, two techniques can be exploited so as to
reduce the computing time and improve convergence:
\begin{itemize}
\item to project the GGF and jet vertices on a complete set of functions which
respects the symmetries of the process (e.g., azimuthal invariance);
\item to use an approximated and simpler version of the jet vertices.
\end{itemize}
These methods are often used in BFKL phenomenological analyses, and we shall
illustrate them in the following subsections: the former in order to set up the
theoretical framework and the main notations; the latter in order to introduce
the main subject of this paper.
\subsection{Representation in Mellin space\label{s:rms}}
The first method to reduce computing time does not involve any approximation, at
least at the NLL level of accuracy we are working with. It is better illustrated
in the LL approximation, where the strong coupling $\as$ is fixed and thus the
BFKL kernel and jet vertices, in addition of being invariant under azimuthal
rotations, are also scale invariant.%
\footnote{Strictly speaking, in our notations they are homogeneous functions of
the transverse momenta $\boldsymbol{k}$, $\boldsymbol{k}_J$.} In this case, a Fourier-Mellin
transform diagonalizes the transverse integrations and the ensuing expressions
are considerably simpler to evaluate.
One can proceed in this way: first of all, let's exploit the fact that, in the
LL approximation, each partonic momentum fraction coincides with the
corresponding jet's longitudinal momentum fraction (because of a $\delta(x-x_J)$
in $V^{(0)}$):
\begin{equation}\label{xJ}
x = x_J \equiv E\esp{\pm y}/\sqrt{s} \qquad
(+,- \text{ for jet } 1,2) \;,
\end{equation}
so that $\hat{s} = x_1 x_2 s = \esp{Y} E_1 E_2$, where $Y\equiv|y_1-y_2|$ is the
rapidity distance between the jets. If we adopt the convenient and natural
choice of $s_0=E_1 E_2$ as energy scale, the Green's function~(\ref{Gomega}) is
now independent of the partonic momentum fractions:
\begin{equation}\label{GY}
G(\hat{s},\boldsymbol{k}_1,\boldsymbol{k}_2) =
\int_{-\ui\infty}^{\ui\infty}\frac{\dif\omega}{2\pi\ui} \;
\esp{\omega Y} G_{\omega}(\boldsymbol{k}_1,\boldsymbol{k}_2) \equiv G(Y,\boldsymbol{k}_1,\boldsymbol{k}_2)
\;,\qquad (s_0 = E_1 E_2)
\end{equation}
Secondly, we introduce the {\em impact factor} by integrating a jet vertex with
the corresponding parton density:
\begin{equation}\label{impf}
\Phi_A(\boldsymbol{k};J) \equiv \sum_a\int_0^1\dif x\; f_{a/A}(x) V_a(x,\boldsymbol{k};J) \;.
\end{equation}
so that we can rewrite the factorization formula in the form
\begin{equation}\label{sigmaFact}
\frac{\dif\sigma_{AB}}{\dif J_1\,\dif J_2} =
\int\dif^2\boldsymbol{k}_1\,\dif^2\boldsymbol{k}_2\; \Phi_A(\boldsymbol{k}_1) G(Y,\boldsymbol{k}_1,\boldsymbol{k}_2) \Phi_B(\boldsymbol{k}_2) \;.
\end{equation}
At this point we project kernel and vertices onto the eigenfunctions of the LL
BFKL kernel
\begin{equation}\label{En}
\afn{n\nu}(\boldsymbol{k}) \equiv \frac1{\sqrt{2}\pi} |\boldsymbol{k}^2|^{\ui\nu-\frac12}
\esp{\ui n\phi} \;, \qquad(n\in\Z,\;\nu\in\R) \;,
\end{equation}
satisfying the completeness relation
\begin{equation}\label{comp}
\sum_{n\in\Z}\int_{-\infty}^{\infty} \dif\nu\;
\afn{n\nu}(\boldsymbol{k}) \afn{n\nu}^*(\boldsymbol{k}') = \delta^2(\boldsymbol{k}-\boldsymbol{k}')
\end{equation}
and providing the LL eigenvalue function $\chi^{(0)}_{n\nu}$
\begin{align}
[K^{(0)} \afn{n\nu}](\boldsymbol{k}) &\equiv \int\dif^2\boldsymbol{k}'\; K^{(0)}(\boldsymbol{k},\boldsymbol{k}')
\afn{n\nu}(\boldsymbol{k}') = \chi^{(0)}_{n\nu} \afn{n\nu}(\boldsymbol{k}) \label{K0E}\\
\chi^{(0)}_{n\nu} &= 2\psi(1)-\psi\left(\frac{1+n}{2}+\ui\nu\right)
-\psi\left(\frac{1+n}{2}-\ui\nu\right) \;.
\label{chi0}
\end{align}
Finally, by inserting a completeness~(\ref{comp}) between each pair of factors
in~(\ref{sigmaFact}), we arrive at the convenient expression for the
differential cross section
\begin{equation}\label{sigmaMel}
\frac{\dif\sigma_{AB}}{\dif J_1\,\dif J_2} =
\sum_n (-1)^n \int\dif\nu\;\Phi_{A\,n\nu} \; G_{n\nu}(Y) \; \Phi_{B\,n\nu}^*\;,
\end{equation}
where in the last equality we have used the Fourier-Mellin transforms
\begin{align}\label{Phin}
\int\dif^2\boldsymbol{k}\;\Phi(\boldsymbol{k};J) \afn{n\nu}(\boldsymbol{k}) &\equiv \Phi_{n\nu}(J) \\
\int\dif^2\boldsymbol{k}\,\dif^2\boldsymbol{k}'\; \afn{n\nu}^*(\boldsymbol{k})G(Y,\boldsymbol{k},\boldsymbol{k}') \afn{n'\nu'}(\boldsymbol{k}')
&\equiv G_{n\nu}(Y) (-1)^n \delta_{nn'}\delta(\nu-\nu') \;. \label{Gn}
\end{align}
The delta functions on the r.h.s.\ of eq.~(\ref{Gn}) are just a consequence of
the azimuthal- and scale-invariance of the kernel, and allow us to trade two
bidimensional integrals for a sum and a simple integral.
The azimuthal correlation of the MN jets is usually measured by means of the
Fourier coefficients ($m\in\Z$)
\begin{equation}\label{defCm}
C_m(E_1,y_1;E_2,y_2) \equiv
\int_0^{2\pi}\dif\phi_1\dif\phi_2 \;\cos\big(m(\phi_1-\phi_2-\pi)\big)
\frac{\dif\sigma}{\dif J_1\,\dif J_2} \;.
\end{equation}
Because of the azimuthal- and scale-invariance of the vertices, it is easy to
show that
\begin{equation}\label{phiStruct}
\Phi_{n\nu}(y,E,\phi) = \esp{\ui n \phi} (E^2)^{\ui\nu-1}
\Psi_{n\nu}(x_J) \;, \qquad \Psi_{n\nu} = \Psi_{-n\nu}
\end{equation}
where $\Psi$ is dimensionless,%
\footnote{We shall sometimes refer to the reduced impact factor $\Psi$ as jet
vertex.}
thus obtaining a factorization formula with just one integration:
\begin{equation}\label{Cm1}
C_m = \left(\frac{2\pi}{E_1 E_2}\right)^2 \int\dif\nu
\left(\frac{E_1^2}{E_2^2}\right)^{\ui\nu} \Psi_{A\,m\nu} \; G_{m\nu}(Y) \;
\Psi_{B\,m\nu}^* \;.
\end{equation}
Such a structure is preserved in the NLL approximation too, provided the impact
factors and GGF are suitably modified in order to take into account the loss of
scale-invariance due to the renormalization procedure and to the factorization
of collinear singularities, which translates in a dependence on $\log(E/\mu_R)$,
$\log(E/\mu_F)$ and $\log(E/\sqrt{s_0})$ of the GGF and impact factors. The
expression of the GGF in Mellin-space is very simply expressed in terms of the
eigenvalue $\chi_{n\nu}$ of the BFKL kernel:
\begin{equation}\label{Gnnu}
G_{n\nu}(Y) = \esp{Y\chi_{n\nu}} \;, \qquad
\chi_{n\nu} = \abs \chi_{n\nu}^{(0)} + \abs^2 \chi_{n\nu}^{(1)}
\end{equation}
However, the computation of $\Psi_{n\nu}$ at NLL level involves several
integrations. The main advantage of this procedure is that such integrations can
be done once and for all for each set of one-jet variables. Nevertheless, such
computations can still be rather lengthy, and the use of an approximate
expression of the impact factors --- to be described in the next subsection ---
turns out to be very convenient.
Let us conclude this section by noticing that the factorization
formula~(\ref{Cm1}) is very useful also in the case of cross-section integrated
in jet energies. In fact, a double integral in $E_1$ and $E_2$ factorizes into
the product of simple integrals of the impact factors (provided the integration
domain $D=I_1\times I_2$ can be factorized into the cartesian product of two
one-dimensional sets)
\begin{align}\label{intCm}
\intC_m(y_1,y_2) &\equiv \int_{D=I_1\times I_2}\dif E_1 \dif E_2 \; C_m
\\ \nonumber
&= (2\pi)^2 \int\dif\nu \; G_{m\nu}(Y)
\left[\int_{I_1} \dif E_1 \; (E_1^2)^{\ui\nu-1} \Psi_{A\,m\nu} \right] \;
\left[\int_{I_2} \dif E_2 \; (E_2^2)^{\ui\nu-1} \Psi_{B\,m\nu} \right]^*
\\ \nonumber
&\equiv (2\pi)^2 \int\dif\nu \; G_{m\nu}(Y) \; \intI_{A\,m\nu}(y_1) \;
\intI_{B\,m\nu}^*(y_2) \;.
\end{align}
In this case, the integrated impact factors $\intI$ can be computed
independently and stored in suitable grids, thus reducing a lot the
computational effort of the phase-space integration. In the general case
$D\neq I_1\times I_2$, however, the expression~(\ref{Cm1}) has to be numerically
integrated in energy (and possibly in rapidity), and a suitable approximation
(like the SCA) could be a very valuable tool to diminish the computing demand.
\subsection{Small-cone approximation\label{s:sca}}
In order to study the behaviour of the jet vertex for small values of the
``radius'' $R$ --- to be precisely defined later on --- and possibly to speed up
the computation of the jet impact factor $\Psi_{n\nu}$ and of its
energy-integrated version $\intI_{n\nu}$, one can use the small-cone
approximation (SCA), as suggested and derived in~\cite{IP}. The dependence of
the impact factor on the jet radius $R$ has the form~\cite{MuVo12} ($\mu$ is a
shorthand for $\mu_R$, $\mu_R$ and $\sqrt{s_0}$)
\begin{equation}\label{Rdep}
\Psi_{n\nu}\Big(x_J,\log\frac{E}{\mu};R\Big) = A_{\nu}(x_J) \log(R)
+ B_{n\nu}\Big(x_J,\log\frac{E}{\mu}\Big) + \ord{R^2}
\end{equation}
and the analytic expressions for the coefficients $A,B$ were explicitly
computed~\cite{IP} for a particular jet algorithm (FIP).
However, whereas the coefficient $A$ of $\log(R)$ depends only on the incoming
hadron and is given in terms of the usual splitting functions as
\begin{align}\label{A}
A_{\nu} = -\as^2\frac{\sqrt{N_c^2-1}}{\sqrt{2}\pi^2 N_c} x_J
\int_{x_J}^1\dif\zeta\; \zeta^{-2\ui\nu}
& \bigg\{ \big[P_{qq}(\zeta)+P_{gq}(\zeta)\big]
\sum_{a\in\{q,\bar{q}\}} f_{a}\Big(\frac{x_J}{\zeta}\Big) \\
& + \big[P_{gg}(\zeta)+2 n_f P_{qg}(\zeta)\big]
f_{g}\Big(\frac{x_J}{\zeta}\Big) \bigg\} \;,
\end{align}
the constant term $B$ depends also on the details of the jet algorithm. Ivanov
and Papa~\cite{IP} computed such coefficient for an algorithm which was used in
pioneering work on QCD jets by Furman~\cite{Fur82} --- we shall refer to it as
FIP algorithm --- which, however, is no more used in present day phenomenology.
The main purpose of our paper is to derive such coefficient for the two mostly
used algorithms of QCD analysis, namely the cone-algorithm and the kt-algorithm.
The computations can be repeated by following the procedure of~\cite{IP}.
The expressions for the jet vertices at LL and NLL level are extracted from the
perturbative calculation of processes with two incoming partons producing 2-jet
at LO and NLO respectively. At LO the amplitudes have just two partons in the
final states, each of which is identified with a jet. The jets are emitted
back-to-back in the azimuthal direction and have no substructure. Therefore no
dependence on the algorithm is found at LO.
The same is true at NLO as far as the virtual corrections are concerned. On the
other hand, the NLO real corrections involve 3 partons in the final state,
therefore a jet can be constituted by either one or two of them. In the case of
1-parton (simple) jet, all algorithms are designed in such a way that no further
emission is found within a region of radius $R$ in the $(y,\phi)$ plane around
the position of that parton.
Therefore, the differences among the algorithms are to be found in the 2-parton
(composite) jet configurations. In the following section we will carefully
compare the definitions the {\it cone}, {\it kt}, and {\it FIP} jet
algorithms and, from their differences, we shall compute the small-cone impact
factors for the cone and kt jets.
\section{Jet algorithms\label{s:jetalg}}
\subsection{The cone algorithm\label{s:cone}}
According to ref.~\cite{EKS89}, when two partons $p_1$ and $p_2$ are combined
into one jet of radius $R$, the resulting jet variables $(y,E,\phi)$ are defined
to be
\begin{equation}\label{conejet}
E = E_1 + E_2 \;, \qquad y = \frac{y_1 E_1 + y_2 E_2}{E} \;, \qquad
\phi = \frac{\phi_1 E_1 + \phi_2 E_2}{E} \;.
\end{equation}
To determine whether the two partons are to be combined, we see if they fit in a
cone of radius $R$ about the jet axis in the $(y,\phi)$ plane. In practice, by
denoting with $\Omega_{ij}$ the $(y,\phi)$-distance
\begin{equation}\label{dij}
\Omega_{ij}^2 \equiv (y_i-y_j)^2 + (\phi_i-\phi_j)^2
\end{equation}
one requires $\Omega_{1J}<R$ and $\Omega_{2J}<R$ for the composite jet, which
amounts the condition
\begin{equation}\label{coneComp}
\Omega_{12} < R \frac{E_1+E_2}{\max(E_1,E_2)} \;.
\end{equation}
A simple jet can then be defined only if the two partons cannot be combined,
i.e., provided
\begin{equation}\label{coneSimple}
\Omega_{12} > R \frac{E_1+E_2}{\max(E_1,E_2)} \;.
\end{equation}
\subsection{The kt algorithm\label{s:kta}}
According to ref.~\cite{CDSW93}, the kt clustering algoritm consists in an
iterative procedure which is based on comparing a set of resolution variables of
single-particle $d_{iB}$ (B for beam) and pairs $d_{ij}$
\begin{equation}\label{resVar}
d_{iB}\equiv E_i^2 \;, \qquad
d_{ij} \equiv \min(E_i^2,E_j^2)\, \frac{\Omega_{ij}^2}{R^2} \;.
\end{equation}
One then considers the smallest one: if it is a $d_{iB}$, particle $i$ is thrown
in the beam basket and removed from the list; if it is a $d_{ij}$ then particles
$i$ and $j$ are merged into a pseudoparticle $\{ij\}$ --- e.g., by using the
recombination scheme~(\ref{conejet}). The procedure is repeated from the
beginning, until all resolution variables are greater than some (hard) stopping
parameter $d_{\cut} \gg \Lqcd^2$.
In the case of three partons in the final state there are these possibilities:
\begin{itemize}
\item One of the $d_{iB}$, say $d_{1B}$ is the smallest resolution variable.
\begin{itemize}
\item If $d_{1B} < d_{\cut}$ then particle 1 belongs to the beam and we are
left with two partons which, if not in the beam, form two simple jets;
\item If $d_{1B} > d_{\cut}$ then the clustering stops and all three particles
form simple jets;
\end{itemize}
\item One of the $d_{ij}$, say $d_{12}$ is the smallest resolution variable. In
this case 1 and 2 are merged into a pseudoparticle \{12\} and one considers
three resolution variables: $d_{\{12\}B}$, $d_{3B}$ and $d_{\{12\}3}$.
\begin{itemize}
\item If the smallest is larger than $d_{\cut}$ clustering stops and we have one
simple jet \{3\} and one composite jet \{12\};
\item If the smallest is less than $d_{\cut}$, a (pseudo)particle
belongs to the beam and we cannot have two jets in the final state;
\end{itemize}
\end{itemize}
To summarize, we have a composite jet, say \{12\}, only if
\begin{equation}\label{ktComp}
d_{12} < d_{iB} \;\forall i \qquad\Longrightarrow\qquad \Omega_{12} < R \;.
\end{equation}
In the other case
\begin{equation}\label{ktSimple}
\Omega_{12} > R
\end{equation}
only simple jets are present.
\subsection{The Furman algorithm\label{s:ipa}}
In~\cite{IP} Ivanov and Papa define a jet as a set of particles within a
cone of radius $R$. To be more precise, they require that:
\begin{itemize}
\item the jet's momentum is the sum of the particles' momenta;
\item all and only the particles of the jet belong to a circle of radius $R$ in the
$(y,\phi)$ plane and centered at the jet's momentum.
\end{itemize}
This is just the jet definition of Furman~\cite{Fur82} used in the '80s for
early phenomenology of NLO QCD jets --- we denote it ``FIP algorithm''.
In the case under study, where at most two particles (1 and 2) can form a jet,
the condition for a composite jet is nothing but the prescription adopted in the
cone algorithm: two particles such that
\begin{equation}\label{ipComp}
\Omega_{12} < R \frac{E_1+E_2}{\max(E_1,E_2)}
\end{equation}
can be considered as a composite jet. On the other hand, particle 1 can form a
simple jet if no other particle is found within a distance $R$ from it, namely
\begin{equation}\label{ipSimple}
\Omega_{12} > R \;.
\end{equation}
This definition is somewhat pathological, because it may happen that a given
configuration can give rise to both a composite jet and two simple jets. This
fact can be easily understood in the case of two particles with the same
transverse energy and whose distance satisfy $R < \Omega_{12} < 2R$. They can
form simple jets because a cone of radius $R$ centered on either particle does
not contain the other; on the other hand, a cone of radius $R$ centered halfway
the two particles contains both of them. For this reason, it is not possible to
extend such jet definition into an IR-safe algorithm to all orders, i.e., with
an arbitrary number of particles, hence it has been abandoned in favour of
better algorithms like the cone and especially the kt one.
In any respect, this definition is different from both the cone algorithm (in
the case of simple jets) and the kt algorithm (in the case of composite jets),
and yields different results when used to defined any observable. It is the main
purpose of this paper to derive the correct expressions for the jet vertices in
the SCA for the kt algorithm and also for the cone one.
\section{Small cone jet vertices\label{s:scjv}}
In this section we shall compute the SCA jet vertex in the cone and kt algoritm.
The idea is to identify and calculate the contributions differing from those of
the original paper~\cite{IP} where the FIP algorithm was adopted.
As explained in the previous section, for dijet production at NLO, the
differences among jet algorithms occur only in the way that two partons can be
combined to form a jet. In the BFKL approach, two partons (1 and 2) in the
fragmentation region of an incoming hadron are produced by the interaction of a
parton stemming from the hadron and a reggeized gluon, as depicted in
fig.~\ref{f:fragKinem}.
\begin{figure}[t]
\centering
\includegraphics[width=0.22\textwidth]{fragKinem.eps}
\caption{\it Kinematics of the fragmentation region of hadron A: the collision
of the incoming parton (blue) and of the Regge gluon (red) produces a pair
of outgoing partons (black). In parentheses the longitudinal momentum
fraction and the transverse momentum of each particle.}
\label{f:fragKinem}
\end{figure}
In the high-energy kinematics, the incoming parton has just a fraction $x$ of
the pure longitudinal momentum of the hadron, while the reggeized gluon's
momentum $q$ is essentially transverse. Following Ivanov-Papa (IP)~\cite{IP} we
work in dimensional regularization ($D=4+2\epsilon$) and we indicate with $k_1$ ($k_2$)
the ($2+2\epsilon$)-dimensional transverse momentum of the outgoing parton 1 (2), and
with $\zeta$ ($\bar\zeta\equiv 1-\zeta$) its longitudinal momentum fraction with
respect to the incoming parton. In terms of these variables, the relative
rapidity and azimuthal angle between the two outgoing partons are
\begin{equation}\label{deltas}
\Delta y = \frac12\log\frac{\zeta^2 k_2^2}{\bar\zeta^2 k_1^2} \;, \qquad
\Delta\phi = \arccos\frac{k_1\cdot k_2}{|k_1|\,|k_2|} \;, \qquad \bar\zeta
\equiv 1-\zeta \;.
\end{equation}
We shall present our results by using the notations of ref.~\cite{IP}, namely in
term of their ``$I$'' quantities, which are related to our definition of jet
impact factor~(\ref{Phin}) by
\begin{equation}\label{notation}
\Phi_{n\nu}(y,E,\phi)
= \as \frac{\sqrt{N_c^2-1}}{\sqrt{2}\pi N_c} \frac{x_J}{E} \, I(n,\nu;y,E,\phi)
\;, \qquad I = \sum_{i,f} (I^R_{i;f} + I^V_{i;f}) \;,
\end{equation}
where $i$ ($f$) are labels for the initial (final) state of the sub-processes
contributing to the cross section, while the superscripts $R$ and $V$ denote
real and virtual parts respectively.
\subsection{Vertex for the cone algorithm\label{s:cav}}
The condition for composite jet in the cone algorithm~(\ref{coneComp}) coincides
with the one adopted by FIP~(\ref{ipComp}), thus no difference is expected for
the corresponding contributions.
In contrast, the conditions for simple jets~(\ref{coneSimple}) and
(\ref{ipSimple}) are different, causing different contributions to the
jet vertices. By using the notations of ref.~\cite{IP}, $k=k_1$ is the
transverse momentum of the parton forming the simple jet, $q-k=k_2$ is the
transverse momentum of the parton outside the jet (also called ``spectator''),
and one introduces the transverse vector $\Del$
\begin{equation}\label{delta}
q = \frac{k}{\zeta} + \Del
\end{equation}
which vanishes when the two partons are collinear. In fact, for small $\Del$,
we have
\begin{equation}\label{omegasc}
\Omega_{12}^2 \equiv \Delta\phi_{12}^2 + \Delta y_{12}^2
\simeq \frac{\zeta^2}{\bar\zeta^2}\frac{\Del^2}{k^2}
\end{equation}
and also
\begin{equation}\label{encomb}
\frac{E_1+E_2}{\max(E_1,E_2)} = \frac{|k_1|+|k_2|}{\max(|k_1|,|k_2|)} \simeq
\frac{|k|\left(1+\frac{\bar\zeta}{\zeta}\right)}{|k|\max\left(1,\frac{\bar\zeta}{\zeta}\right)}
= \frac1{\max(\zeta,\bar\zeta)}
\end{equation}
Therefore, the simple-jet condition~(\ref{coneComp}) becomes
\begin{equation}\label{coneDelta}
|\Del| > \frac{\bar\zeta}{\zeta} |k|\frac{R}{\max(\zeta,\bar\zeta)} \;,
\end{equation}
at variance with the FIP condition~(\ref{ipComp}) which, expressed in terms of
$\Del$, reads like eq.~(\ref{coneDelta}) but without the denominator
$\max(\zeta,\bar\zeta)$~\cite{IP}.
In order to identify the contributions of the FIP jet vertex that have to be
modified in the cone algorithm, let us recall that the simple jet configurations
were computed in two steps: {\it 1)} by allowing the spectator parton to span
the whole phase space and then {\it 2)} by subtracting the contribution stemming
from the spectator parton inside the jet cone. Therefore, we have to modify only
the subtractions by replacing $R\to R/\max(\zeta,\bar\zeta)$.
The {\bf quark initiated} jet vertex (quark+Regge-gluon $\to$ quark+gluon, see
fig.~\ref{f:fragKinem}) has two such subtractions, one for the quark-jet and one
for the gluon-jet, whose results are reported in eqs.~(5.36) and (5.38) of
ref.~\cite{IP} respectively. In the gluon-jet term the substitution $R\to
R/\max(\zeta,\bar\zeta)$ is straightforward:
\begin{align}
I_{q;g,-q}^R &= (5.36)[1] \to
-\frac{\as}{2\pi}\frac{\Gamma(1-\epsilon)}{\epsilon(4\pi)^e}\frac{\Gamma^2(1+\epsilon)}{\Gamma(1+2\epsilon)}
(k^2)^{\gamma+\epsilon-\frac{n}{2}} (k\cdot l)^n
\int_{x_J}^1\frac{\dif\zeta}{\zeta}\;\zeta^{-2\gamma} \sum_{a=q,\bar{q}}
f_a\left(\frac{x_J}{\zeta}\right) \nonumber \\
&\qquad\qquad\times\left(\frac{R}{\max(\zeta,\bar\zeta)}\right)^{2\epsilon}
\left[P_{gq}(\zeta)
\left(1+2\epsilon\log\frac{\bar\zeta}{\zeta}\right)+\epsilon C_F\zeta\right] \nonumber \\
&= I_{q;g,-q}^R - \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\int_{x_J}^1\frac{\dif\zeta}{\zeta}\;\zeta^{-2\gamma} \sum_{a=q,\bar{q}}
f_a\left(\frac{x_J}{\zeta}\right)
P_{gq}(\zeta) 2\log\big(\max(\zeta,\bar\zeta)\big) \;,
\label{diffIqgq}
\end{align}
where $\gamma\equiv\ui\nu-\frac12$ and $l\equiv e_1+\ui e_2$ is a complex vector
lying only in the first two of the $2+2\epsilon$ transverse dimensions.
In the quark-jet one has to proceed more carefully, because of the presence of a
double pole in $\epsilon$ multiplying $(R/\max(\zeta,\bar\zeta))^{\epsilon}$. This would
generate, among other things, modified simple poles and also finite double logs.
However, since the double pole is multiplied by a $\delta(1-\zeta)$, for these
terms $\max(\zeta,\bar\zeta)=1$ and the outcome is identical to the FIP
algorithm. The only difference comes from the simple pole in front of the
$P_{qq}$ splitting function, and we obtain
\begin{align}
I_{q;q,-g}^R &= (5.38)[1] \nonumber \\
&\to I_{q;q,-g}^R - \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\int_{x_J}^1\frac{\dif\zeta}{\zeta}\;\zeta^{-2\gamma} \sum_{a=q,\bar{q}}
f_a\left(\frac{x_J}{\zeta}\right)
P_{qq}(\zeta) 2\log\big(\max(\zeta,\bar\zeta)\big) \;.
\label{diffIqqg}
\end{align}
The {\bf gluon initiated} jet vertex has two subtractions too, one for the
$q\bar{q}$ final state and one for the $gg$ final state, whose results are
reported in eqs.~(5.49)%
\footnote{We note a misprint in ref.~\cite{IP}: in eqs.~(5.48-49) the first
subscript of $I^R$ should be $g$ instead of $q$. Also the last subscript in
eq.~(5.38) should be $-g$ instead of $-q$.}
and (5.56) of ref.~\cite{IP} respectively. The situation
is very similar to that of the quark initiated vertex: the subtraction in the
(anti)quark jet contains a simple pole times the $P_{qg}(\zeta)$ splitting
function in front of the $R^{2\epsilon}$ factor, while the gluon jet contains both
single and double poles, the former with the $P_{gg}(\zeta)$ splitting function
and the latter with the $\delta(1-\zeta)$ distribution.
For both types of jets the substitution $R\to R/\max(\zeta,\bar\zeta)$ simply
amounts to finite contributions proportional to
$\log\big(\max(\zeta,\bar\zeta)\big)$:
\begin{align}
I_{g;q,-\bar{q}}^R &+ I_{g;\bar{q},-q}^R + I_{g;g,-g}^R
= (5.49)[1] + \{ q \leftrightarrow \bar{q} \} + (5.56)[1] \nonumber \\
&\to I_{g;q,-\bar{q}}^R + I_{g;\bar{q},-q}^R + I_{g;g,-g}^R \label{diffIg} \\
& - \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\int_{x_J}^1\frac{\dif\zeta}{\zeta}\;\zeta^{-2\gamma} \sum_{a=q,\bar{q}}
f_a\left(\frac{x_J}{\zeta}\right)
\left[2 n_f P_{qg}(\zeta) + P_{gg}(\zeta) \right]
2\log\max(\zeta,\bar\zeta) \;. \nonumber
\end{align}
To sum up, the jet vertex for the cone algorithm in the small-cone approximation
is obtained by replacing $R\to R/\max(\zeta,\bar\zeta)$ in the final formulae
(5.39) and (5.57) of~\cite{IP}. The complete expressions are written in
eqs.~(\ref{finalQ},\ref{finalG}).
\subsection{Vertex for the kt algorithm\label{s:kav}}
The condition for simple jet in the kt algorithm~(\ref{ktSimple}) coincides with
the one adopted by FIP~(\ref{ipSimple}), while the conditions for composite
jet~(\ref{ktComp}) and (\ref{ipComp}) are different. In the composite jet
configuration the jet's transverse momentum is $k=k_1+k_2=q$ and it is
convenient to define the auxiliary transverse vector $\Del$ as
\begin{equation}\label{Del2}
k_1 = \zeta k + \Del
\end{equation}
thus obtaining
\begin{equation}\label{omegacc}
\Omega_{12}^2 = \frac{\Del^2}{k^2 \zeta^2 \bar{\zeta}^2}
\end{equation}
The composite jet condition~(\ref{ktComp}) becomes
\begin{equation}\label{ktDelta}
|\Del| < \zeta\bar\zeta |k| R \;,
\end{equation}
at variance with the FIP condition~(\ref{ipComp}) which reads~\cite{IP}
$|\Del|<\min(\zeta,\bar\zeta) |k| R$.
The corresponding contributions of~\cite{IP}, that have to be modified in order
to recover the jet vertex in the kt algorithm, are found in sec.~5.1.1.c for the
quark initiated vertex and secs.~5.2.1.b and 5.2.2.b for the gluon initiated
one. In all such cases the modification amounts to replace
\begin{equation}\label{deltaMaxRep}
|\Delta_{\max}| = \min(\zeta,\bar\zeta) |k| R \to \zeta\bar\zeta |k| R \;,
\end{equation}
and finally to substitute $\min(\zeta,\bar\zeta)\to\zeta\bar\zeta$ in the
$\zeta$-integral with the relevant splitting function.
Explicitly, in the {\bf quark-initiated} case, such integrals for the FIP and kt
algoritms reads respectively
\begin{align}\label{IqqgFIP}
I_{q;q+g}^{R\;(\mathrm{FIP})} &\propto \int_0^1\dif\zeta\;
[\min(\zeta,\bar\zeta)]^{2\epsilon} \frac{1+\bar\zeta^2+\epsilon\zeta^2}{\zeta}
= \frac1{\epsilon} -\frac32 +\left(\frac72 -\frac{\pi^2}{3}+3\log 2\right)\epsilon \\
I_{q;q+g}^{R\;(\mathrm{kt})} &\propto \int_0^1\dif\zeta\; \quad\;\;
(\zeta\bar\zeta)^{2\epsilon} \quad \frac{1+\bar\zeta^2+\epsilon\zeta^2}{\zeta}
= \frac1{\epsilon} -\frac32 +\left(\frac{13}{2} -\frac{2\pi^2}{3}\right)\epsilon \;.
\label{IqqgKT}
\end{align}
Their difference $(3-\pi^2/3-3\log2)\epsilon$, when multiplied by the overall $1/\epsilon$ pole,
provides a finite contribution that has to be added to the IP result in order to
obtain the proper expression for the kt algorithm:
\begin{align}
I_{q;q+g}^R = (5.33)[1]
\to I_{q;q+g}^R + \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\sum_{a=q,\bar{q}} f_a(x_J) C_F \left(3-\frac{\pi^2}{3}-3\log2 \right) \;.
\label{diffIqg}
\end{align}
In the {\bf gluon-initiated} case the procedure is identical; here we have two
contributions: one from the $q\bar{q}$ jet
\begin{align}
I_{g;q+\bar{q}}^R = (5.47)[1]
\to I_{g;q+\bar{q}}^R + \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\frac{C_A}{C_F} f_g(x_J) \, 2 n_f T_R \left(\frac{2}{3}\log2 -\frac{23}{36} \right) \;,
\label{diffIqqq}
\end{align}
and one from the $gg$ jet
\begin{align}
I_{g;g+g}^R = (5.54)[1]
\to I_{g;g+g}^R + \frac{\as}{2\pi} (k^2)^{\gamma-\frac{n}{2}} (k\cdot l)^n
\frac{C_A}{C_F} f_g(x_J) C_A \left(
\frac{131}{36} -\frac{\pi^2}{3} -\frac{11}{3}\log2 \right) \;.
\label{diffIggg}
\end{align}
The integrals~(\ref{IqqgFIP},\ref{IqqgKT}) and the analogous ones for the
gluon-initiated contributions, yielding
eqs.~(\ref{diffIqg}-\ref{diffIggg}), were already considered and
computed~\cite{MuVo12} in the first study of the relation between the kt and the
cone algorithm --- the latter sharing with FIP the condition of composite jet.
\subsection{Final expressions of the jet vertices\label{s:final}}
The result of the jet vertex for the cone and kt algoritms is reported
below, by adding to the original expressions of IP~\cite{IP} the modifications
computed in the previous subsections and here highlighted in boldface:
$\boldsymbol{\langle\cdots\rangle_C}$ for the cone and
$\boldsymbol{\langle\cdots\rangle_K}$ for the kt. The quark part is
\begin{align}
I_q &= \frac{\as}{2\pi} (k^2)^{\gamma} \esp{\ui n\phi} \int_{x_J}^1
\frac{\dif\zeta}{\zeta} \sum_{a=q,\bar{q}}
f_a\left(\frac{x_J}{\zeta}\right)
\Bigg\{
\left[P_{qq}(\zeta)+\frac{C_A}{C_F}P_{gq}(\zeta)\right]\log\frac{k^2}{\mu_F^2} +
\nonumber \\
&\quad- 2 \zeta^{-2\gamma}[P_{qq}(\zeta)+P_{gq}(\zeta)]\log
\frac{R}{\boldsymbol{\big\langle\max(\zeta,\bar\zeta)\big\rangle_C}}-\frac{\beta_0}{2}\log\frac{k^2}{\mu_R^2}\delta(1-\zeta)
\nonumber\\
&\quad+C_A\delta(1-\zeta)\left\{\chi^{(0)}_{n\nu}\log\frac{s_0}{k^2}+\frac{85}{18}+\frac{\pi^2}{2}
+\frac12\left[\psi'\left(1+\gamma+\frac{n}{2}\right)-\psi'\left(\frac{n}{2}-\gamma\right)-\chi^{(0)\,2}_{n\nu}
\right]\right\} \nonumber\\
&\quad+(1+\zeta^2)\bigg\{C_A\left[\frac{(1+\zeta^{-2\gamma})\chi^{(0)}_{n\nu}}{2(1-\zeta)_+}
-\zeta^{-2\gamma}\left(\frac{\log(1-\zeta)}{1-\zeta}\right)_+\right] \nonumber \\
&\quad+\left(C_F-\frac{C_A}{2}\right)
\left[\frac{\bar\zeta}{\zeta^2}I_2-\frac{2\log \zeta}{\bar\zeta}+2\left(\frac{\log(1-\zeta)}{1-\zeta}\right)_+
\right]\bigg\} \nonumber \\
&\quad + \delta(1-\zeta)\left[ C_F\left(3\log 2-\frac{\pi^2}{3}-\frac{9}{2}
+\boldsymbol{\Big\langle 3-\frac{\pi^2}{3}-3\log 2 \Big\rangle_K}\right)
-\frac{10}{9} n_f T_R\right] \nonumber \\
&\quad+C_A \zeta+C_F \bar\zeta +\frac{1+\bar\zeta^2}{\zeta}\left[C_A\frac{\bar\zeta}{\zeta}I_1+2C_A\log\frac{\bar\zeta}{\zeta}
+C_F \zeta^{-2\gamma}(\chi^{(0)}_{n\nu}-2\log\bar\zeta)\right]\Bigg\} \;, \label{finalQ}
\end{align}
where $\gamma\equiv\ui\nu-1/2$, $\beta_0\equiv(11 C_A-4 n_f T_R)/3$, $\psi$ is
the digamma function, while the splitting functions $P_{ab}(\zeta)$ and the
special functions $I_j(n,\gamma,\zeta)$ are reported in app.~\ref{a:def}.
The gluon part is
\begin{align}
I_g &= \frac{\as}{2\pi} (k^2)^{\gamma} \esp{\ui n\phi} \int_{x_J}^1
\frac{\dif\zeta}{\zeta} f_g\left(\frac{x_J}{\zeta}\right) \frac{C_A}{C_F}
\Bigg\{
\left[P_{gg}(\zeta)+\frac{C_A}{C_F} 2 n_F P_{qg}(\zeta)\right]\log\frac{k^2}{\mu_F^2} +
\nonumber \\
&\quad- 2 \zeta^{-2\gamma}[P_{gg}(\zeta)+2 n_f P_{qg}(\zeta)]\log
\frac{R}{\boldsymbol{\big\langle\max(\zeta,\bar\zeta)\big\rangle_C}}-\frac{\beta_0}{2}\log\frac{k^2}{4\mu_R^2}\delta(1-\zeta)
\nonumber\\
&\quad+C_A\delta(1-\zeta)\left\{\chi^{(0)}_{n\nu}\log\frac{s_0}{k^2}
+\frac12\left[\psi'\left(1+\gamma+\frac{n}{2}\right)-\psi'\left(\frac{n}{2}-\gamma\right)-\chi^{(0)\,2}_{n\nu}
\right] \right. \nonumber \\
&\qquad\qquad\qquad\qquad \left. +\frac1{12}+\frac{\pi^2}{6} +\boldsymbol{\Big\langle
\frac{131}{36} -\frac{\pi^2}{3} -\frac{11}{3}\log 2 \Big\rangle_K} \right\} \nonumber\\
&\quad+2C_A(1-\zeta^{-2\gamma})\left[\left(\frac1{\zeta}-2+\zeta \bar\zeta\right)\log \bar\zeta
+\frac{\log(1-\zeta)}{1-\zeta}\right] \nonumber \\
&\quad+C_A\left[\frac1{\zeta}+\frac1{(1-\zeta)_+}-2+\zeta \bar\zeta\right]
\left[(1+\zeta^{-2\gamma})\chi^{(0)}_{n\nu}-2\log \zeta+\frac{\bar\zeta^2}{\zeta^2} I_2\right] \nonumber \\
&\quad+ 2 n_f T_R \left[2\frac{C_F}{C_A}\zeta \bar\zeta+(\zeta^2+\bar\zeta^2)\left(\frac{C_F}{C_A}\chi^{(0)}_{n\nu}
+\frac{\bar\zeta}{\zeta}I_3\right) \right.\nonumber \\
&\qquad\qquad\qquad \left. +\delta(1-\zeta)\left( -\frac1{12}
+\boldsymbol{\Big\langle \frac{2}{3}\log 2-\frac{23}{36} \Big\rangle_K}
\right) \right]\Bigg\} \;. \label{finalG}
\end{align}
It is apparent from the above expression that the $\log(R)$ coefficient $A$ of the
jet vertex is independent of the jet algorithm, while the constant coefficient
$B$ depends on it.
\section{Numerical study\label{s:nc}}
In this section we assess the quantitative difference among the jet vertices in
the three algorithms (cone, kt, FIP) that we considered, and also the
corresponding accuracy of their small-cone approximations (SCA). We shall use
the term {\em exact} in the sense of ``without SCA''.
\subsection{Jet vertices versus $\boldsymbol{R}$}
We start by evaluating the ``exact'' jet vertices in the three algorithms --- we
employ the numerical code used in ref.~\cite{CSSW10} --- for various values of
the jet radius $R$, and we compare them with their SCA. We expect the SCA to be
better, the smaller the values of $R$, and increasing discrepancy with
increasing $R$. On the other hand, the differences among different algorithms
shouldn't vanish with $R$, according to our analysis.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\linewidth]{scaVSR_n0.eps}
\caption{\it Dependence on the jet radius $R$ of the NLO jet vertices $\Psi$
in the three jet algorithms discussed in the text: FIP (dashed red), cone
(dotted green) and kt (solid blue). The big dots correspond to the exact
evaluation (the small error bars showing MonteCarlo integration
uncertainties), while the straight lines denote the small-cone
approximation. Here $n=0$, $\nu=0$, $y=3.6$, $E=35\gev$.}
\label{f:scaVSR}
\end{figure}
This is actually the case, as can be seen in fig.~\ref{f:scaVSR}, where we plot
the exact (points) and small-cone (lines) NLO part of the jet vertex versus $R$,
in the three algorithms mentioned before (for a given choice of the parameters
$n$, $\nu$, $E$, $y$). The common slope of the lines represents the coefficient
$A$ of $\log(R)$, which is the same for the three algorithms, while the
intercepts at $R=1$ give the constant coefficients $B$, which clearly depend on
the algorithm. A detailed study, carried out with several values of the
parameters, shows that the small-cone approximation works very well up to $R$ of
few tenths, with an error below 1\% for $R\lesssim0.2$ which increases up to
3-4\% when $R=0.5$.
\subsection{SCA versus algorithm choice}
Next, we specialize our analysis to realistic values of the jet radius and
energy. Since the typical phenomenological studies on MN jets use $R\simeq 0.5$
and jet transverse energies $E\gtrsim 35\gev$, in the following all quantities
will be evaluated at $R=0.5$ and $E=35\gev$. In addition, nowadays the mostly
used jet algorithm is the kt. Therefore we adopt ``exact'' quantities, computed
in the kt algorithm, as reference quantities, and we estimate the deviations to
them introduced by the SCA.
To some extent, adopting the SCA at fixed $R$ is a sort of choosing a jet
algorithm. The natural question then arises: how does the discrepancy introduced by
the small-cone approximation
\begin{equation}\label{disc1}
(\text{exact-kt}) - (\text{SCA-kt})
\end{equation}
compare with the discrepancy caused by different choices of jet algorithm, i.e.,
\begin{equation}\label{disc2}
(\text{SCA-kt}) - (\text{SCA-FIP}) \;?
\end{equation}
In order to answer this question, we compute the exact NLO part%
\footnote{We recall that the LO part is independent of the jet algorithm.}
of jet vertex $\Psi_{n\nu}^{\nlo}$ in the kt algorithm as function of $\nu$,
and compare it with the SCA in the same algorithm and also with the SCA in the
FIP algorithm.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\linewidth]{scaVSnu.eps}
\caption{\it Comparison of the NLO exact jet vertex in the kt algorithm
(points) with its SCA in the same algorithm (solid blue) and in the FIP
algorithm (dashed red). Here $R=0.5$, $n=1$, $y=3.6$, $E=35\gev$.}
\label{f:scaVSnu}
\end{figure}
Fig.~\ref{f:scaVSnu} shows such a comparison for $n=1$. We can see that the
exact result in the kt algorithm is much better approximated by the SCA in the
same algorithm rather than by the FIP choice, in particular at small values of
$\nu$, which are the most important in the $\nu$ integrals of
eqs.~(\ref{Cm1},\ref{intCm}) --- since the GGF is peaked around $\nu=0$.
This conclusion is further supported by analysing the whole (LO+NLO)
$\nu$-integrand of eq.~(\ref{intCm}), again by comparing exact-kt, SCA-kt and
SCA-FIP, as in fig.~\ref{f:integrandVSnu}. The discrepancy introduced by the SCA
in the wrong FIP algorithm is about three times larger than that
introduced by the SCA in the proper kt algorithm, the latter being of the order
of 5\%.
Actually, the relative error due to the SCA is slightly larger for the full
(LO+NLO) quantities than for the pure NLO ones. This is due to the fact that the
NLO corrections usually have sign opposite to the common LO terms, giving
rise to cancellations in their sum that amplify the relative differences.
\begin{figure}[ht!]
\centering
\hspace{0.036\linewidth}\includegraphics[width=0.411\linewidth]{integrand_n0.eps}%
\hspace{0.093\linewidth}\includegraphics[width=0.428\linewidth]{integrand_n1.eps}%
\hfill\null\\
\hfill%
\includegraphics[width=0.44\linewidth]{integrand_n0_ratio.eps}%
\hspace{0.08\linewidth}%
\includegraphics[width=0.44\linewidth]{integrand_n1_ratio.eps}%
\hfill\null
\caption{\it Comparison of the exact $\nu$-integrand for kt algorithm (points)
with the SCA in the same algorithm (solid blue) and in the FIP algorithm
(dashed red). On the left: $n=0$; on the right: $n=1$. The upper plots display
absolute values in arbitrary units, and below them we show the ratios
w.r.t.\ exact integrand. The parameters are $R=0.5$, $E_1=E_2=35\gev$,
$y_1=3.6$, $y_2=2.8$.}
\label{f:integrandVSnu}
\end{figure}
\subsection{Cross section and angular coefficients}
Finally, we present the results of the differential cross section
$\dif\sigma/\dif Y$ and of few angular coefficients $C_m/C_n$.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.61\linewidth]{dsigma_dY.eps}\\
\hspace{9pt}\includegraphics[width=0.6\linewidth]{dsigma_dY_ratio.eps}
\caption{\it Comparison of the exact differential cross section
$\dif\sigma/\dif Y$ ($n=0$) for kt algorithm (solid blue) with the SCA in
the same algorithm (dash-dotted blue) and in the FIP algorithm (dashed red).
Top: absolute values in linear scale; bottom: ratios of the SCAs w.r.t.\ the
exact one. Here $R=0.5$, $E_1=E_2=35\gev$.}
\label{f:dsigmadY}
\end{figure}
In fig.~\ref{f:dsigmadY} we plot the differential cross section
$\dif\sigma/\dif Y = C_0(Y)$ by comparing again the exact kt calculation with
the small-cone approximations in the KT and FIP algorithms.
It is evident that the wrong choice of the algorithm yields a large error,
especially at lower values of $Y$, while the sole SCA with the proper algorithm
introduces an error of 4-8\%.
The shape of the $C_0$ curves, which are not monothonically decreasing
in $Y$ as one could naively expect, is due to an additional cut in
rapidity $|y_i|>y_{\min}=3$ that we have imposed just for computational
convenience, as will be shortly explained. Due to this cut, the
minimum value of $Y$ that we allow is $Y_{\min} = 2 y_{\min} = 6$ and in
this limit the cross-section vanishes. It then quickly rises for $Y>6$
before eventually decreasing at larger $Y\gtrsim 7$.
The reason for imposing the $|y_i|>y_{\min}$ cut is due to the fact that
monochromatic (fixed $E$) impact factors oscillate at large $\nu$, causing the
$\nu$ integration in eq.~(\ref{Cm1}) to be slowly convergent. Actually, the
convergence is provided by the GGF, whose modulus decreases at large $\nu$, the
decrease being faster at larger values of $Y$. On the other hand, at low values
of $Y$, and in particular with very asymmetric rapidity configurations
($\big||y_1|-|y_2|\big|\gg 1$), the numeric $\nu$-integration has to be pushed
to large $\nu$-values, thus demanding a large computational effort. Since this
problem is absent for realistic phenomenological studies (where the impact
factors are integrated in the $E$ variable and decrease themselves with
$\nu$), and because the goal of this analysis is just to compare the main
features of the different algorithms, we solve the convergence issue by imposing
the mentioned cut in rapidity $|y_i|>3$.
From the plots of figs.~\ref{f:integrandVSnu},\ref{f:dsigmadY} one can infer
that the main effect of the SCA is mostly an overall normalization change, and
that this is also true even for the wrong choice of algorithm (kt versus FIP),
though with a larger factor. If this were the case, by computing {\em ratios} of
observables such effects should cancel out and reproduce more faithfully the
exact quantities. This is partially true, as can be seen in fig.~\ref{f:CmCn},
where we plot some ratios of angular coefficients
$C_m(Y)/C_n(Y)=\langle\cos(m\Delta\phi_J)\rangle/\langle\cos(n\Delta\phi_J)\rangle$
which are usually adopted in order to measure the azimuthal decorrelation of the
MN jets.
It is nevertheless evident that the cancellation of the systematic effects is
more effective if the SCA is made with the proper algorithm, leading to a
discrepancy of about 2\% or less for all the ratios considered. A different
choice of algorithm yields definitely larger discrepancies, and therefore should
be avoided.
\begin{figure}[th!]
{\centering
\hspace{0.03\linewidth}\includegraphics[width=0.45\linewidth]{C1_C0.eps}\hfill
\includegraphics[width=0.46\linewidth]{C2_C0.eps}\\}
\hspace{0.02\linewidth}\includegraphics[width=0.46\linewidth]{C2_C1.eps}
\caption{\it Comparison of some ratios $C_m/C_n$ determined with the exact jet
vertices in the kt algorithm (solid blue) with those obtained in the SCA
with the same algorithm (dotted blue) and with the FIP algorithm (dashed
red). Here $R=0.5$, $E_1=E_2=35\gev$.}
\label{f:CmCn}
\end{figure}
\section{Conclusions\label{s:conc}}
In this paper we have reconsidered the high-energy factorization formula for
Mueller-Navelet jet production, and we have computed the NLO jet vertices in the
small-cone approximation for two different jet algorithms: the {\em kt algorithm}
and the {\em cone algorithm}.
The small-cone approximation amounts to evaluate the generic jet vertex
$\Psi$ at small jet radius $R\to 0$ according to the expansion
$\Psi=A\log(R)+B+\ord{R^2}$, where the coefficient $A$ is universal, depending
only on the infra-red properties of QCD, while the constant term $B$ depends on
the jet algorithm. It should be noted that the neglected term in the expansion
is quadratic in $R$, and that the linear term is missing. It turns out that the
small-cone approximation is quite accurate for realistic values of $R\lesssim
0.5$ and, because of its analytic and computational simplicity, it has been
and will be used with benefit.
The calculation of the Mellin-Fourier components of the jet vertices (cfr.\
sec.~\ref{s:rms}) was originally performed~\cite{IP} by adopting an algorithm
which is equivalent to Furman's one~\cite{Fur82}, but which is not used anymore
in modern phenomenological studies, being infra-red unsafe. In order to use such
vertices for the analysis of data from high-energy colliders, we have computed
them in the two most popular algorithms: the kt and the cone ones. The ensuing
analytical expressions have been taken out in sec.~\ref{s:scjv}, and the
differences among the three algorithms have been highlighted.
The quantitative difference between the small-cone approximations in the kt
algorithm and in Furman's one have been estimated in sec.~\ref{s:nc}, by
plotting several quantities of interest: the jet vertices, the integrand of the
high-energy factorization formula, the differential cross section w.r.t.\ the
rapidity distance $Y\equiv|y_1-y_2|$ of the two MN jets, and some angular
coefficients indicating the azimuthal decorrelations of the jets
themselves. Also the corresponding quantities with the exact jet vertices in the
kt algorithm have been plotted.
It turns out that the difference between the kt and Furman's algorithms is
sizeable, of the order of 20\% at the level of cross section, and about 5\% for
the ratios $C_m/C_0=\langle\cos(m\Delta\phi_J)\rangle$ and $C_m/C_n$, the
angular coefficients of azimuthal decorrelation.
On the other hand, the discrepancies between the exact results and the
small-cone approximated ones are much smaller, of the order of 5\% at the level
of cross section, and less than 2\% for the angular ratios. We therefore
conclude that the small-cone expansion, computed with the proper jet algorithm,
provides a good approximation to the Mueller-Navelet jet vertices; it can then
be used as a very convenient tool to perform phenomenological studies, in that
it requires much less computational resources than the exact computation.
This aspect could be essential when analysing observables obtained by
integrating the jet energies in a non-factorized domain, e.g., by requiring
$E_1+E_2>2 E_{\cut}$. Such condition is often used in dijet analysis since, at
fixed NLO, it yields more stable results than the condition $E_i>E_{\min}$,
while retaining the $1\leftrightarrow 2$ symmetry. On the other hand, a
non-factorizable domain of integration in energy prevents the use of
eq.~(\ref{intCm}) with independently integrated impact factors $\intI$; a
numerical integration in energy (and possibly in rapidity) of the
integrand~(\ref{Cm1}) is thus necessary, and the small-cone approximation
represents a valuable tool to reduce the computational effort of such
calculation.
\section*{Acknowledgements}
We wish to thank the {\em Galileo Galilei Institute for Theoretical Physics} where
part of this work was performed.
|
\section{Dynamical systems in number theory}
In the last decades dynamical systems became very important for the
development of modern number theory. The present paper focuses on
Furstenberg's refinements of Poincaré's recurrence theorem and
applications of these ideas to Diophantine problems.
A (measure-theoretic) dynamical system is formally given as a
quadruple $(X,\mathfrak{B}, \mu, T)$, where $(X,\mathfrak{B},\mu)$
is a probability space with $\sigma$-algebra $\mathfrak{B}$ of
measurable sets and $\mu$ a probability measure; $T\colon
X\rightarrow X$ is a measure-preserving transformation on this
space, \textit{i.e.} $\mu(T^{-1}A)= \mu(A)$ for all measurable
sets $A\in\mathfrak{B}$. In the theory of dynamical systems,
properties of the iterations of the transformation $T$ are of
particular interest. For this purpose we only consider invertible
transformations and call such dynamical systems invertible.
The first property, we consider, originates from Poncaré's famous
recurrence theorem (see Theorem 1.4 of
\cite{walters1982:introduction_to_ergodic} or Theorem 2.11 of
\cite{einsiedler_ward2011:ergodic_theory_with}) saying that
starting from a set $A$ of positive measure $\mu(A)>0$ and
iterating $T$ yields infinitely many returns to $A$. More
generally, we call a subset $\mathcal{R}\subset\mathbb{N}$ of the
positive integers a set of recurrence if for all invertible
dynamical systems and all measurable sets $A$ of positive measure
$\mu (A)>0$ there exists $n\in\mathcal{R}$ such that $\mu(A\cap
T^{-n}A)>0$. Then Poincaré's recurrence theorem means that $\N$
is a set of recurrence.
A second important theorem for dynamical systems is Birkhoff's ergodic
theorem (see Theorem 1.14 of
\cite{walters1982:introduction_to_ergodic} or Theorem 2.30 of
\cite{einsiedler_ward2011:ergodic_theory_with}). We call $T$ ergodic
if the only invariant sets under $T$ are sets of measure $0$ or of
measure $1$, \textit{i.e.} $T^{-1}A=A$ implies $\mu(A)=0$ or
$\mu(A)=1$. Then Birkhoff's ergodic theorem connects average in time
with average in space, \textit{i.e.}
\[\lim_{N\rightarrow\infty}\frac{1}{N}\sum^{N-1}_{n=0}f \circ
T^{n}(x)= \int_{X}f(x)d\mu(x)\]
for all $f\in L^{1}(X,\mu)$ and $\mu$-almost all $x \in X$.
Let us explain an important application of this theorem to number
theory. For $q\geq2$ a positive integer, consider
$T\colon[0,1)\rightarrow [0,1)$ defined by $T(x)=\{qx\},$ where
$\{t\}=t-\lfloor t\rfloor$ denotes the fractional part of $t$.
If $x\in\R$ is given by its $q$-ary digit expansion
$x=\lfloor x\rfloor+ \sum^{\infty}_{j=1}a_{j}(x)q^{-j}$, then the
digits $a_{j}(x)$ can be computed by iterating this transformation
$T$: $a_{j}(x)=i$ if $T^{j-1}x\in\left[\tfrac iq,\frac{i+1}q\right)$
with $i\in\{0,1,\ldots,q-1\}$. Moreover, since $a_j(Tx)=a_{j+1}(x)$
for $j\geq1$ the transformation $T$ can be seen as a left shift of the
expansion.
Now we call a real number $x$ simply normal in base $q$ if
\[\lim_{\mathbb{N}\rightarrow\infty}\frac{1}{N}\#\{j\leq N\colon a_{j}=d\}= \frac{1}{q}\]
for all $d=0, \ldots, q-1$, \textit{i.e.} all digits $d$ appear
asymptotically with equal frequencies $1/q.$ A number $x$ is
called $q$-normal if it is simply normal with respect to all bases
$q, q^{2},q^{3},\ldots$. This is equivalent to the fact that the
sequence $(\{q^{n}x\})_{n\in\mathbb{N}}$ is uniformly distributed
modulo $1$ (for short: u.d.~mod $1$), which also means that all
blocks $d_{1},d_{2},\ldots, d_{L}$ of subsequent digits appear in
the expansion of $x$ asymptotically with the same frequency
$q^{-L}$ (\textit{cf.} \cite{bugeaud2012:distribution_modulo_one,
drmota_tichy1997:sequences_discrepancies_and,
kuipers_niederreiter1974:uniform_distribution_sequences}). For
completeness, let us give here one possible definition of u.d.
sequences $(x_{n})$: a sequence of real numbers $x_{n}$ is called
u.d.~mod $1$ if for all continuous functions $f: [0,1]\rightarrow
\mathbb{R}$
\begin{equation}\label{ud}
\lim_{N\rightarrow\infty}\frac{1}{N}\sum^{N}_{n=1}f(x_{n})=
\int^{1}_{0}f(x)dx.
\end{equation}
Note, that by Weyl's criterion the class of continuous functions
can be replaced by trigonometric functions $e(hx)=e^{2\pi ihx}$,
$h\in \mathbb{N}$ or by characteristic functions $1_{I}(x)$ of
intervals $I=[a,b)$. Applying Birkhoff's ergodic theorem, shows
that Lebesgue almost all real numbers are $q$-normal in any base
$q\geq 2$. Defining a real number to be absolutely normal if it is
$q$-normal for all bases $q\geq 2$, this immediately yields that
almost all real numbers are absolutely normal.
In particular, this shows the existence of absolutely normal
numbers. However, it is a different story to find constructions of
(absolutely) normal numbers. It is a well-known difficult open problem
to show that important numbers like $\sqrt{2}$, $\ln 2$, $e$, $\pi$
etc. are simply normal with respect to some given base $q\geq 2$. A
much easier task is to give constructions of $q$-normal numbers for
fixed base $q$. Champernowne
\cite{champernowne1933:construction_decimals_normal} proved that
$$0.1\,2\,3\,4\,5\,6\,7\,8\,9\,10\,11\,12\ldots$$
is normal to base $10$ and later this type of constructions was
analysed in detail. So, for instance, for arbitrary base $q\geq 2$
\[0.\langle\lfloor g(1)\rfloor\rangle_{q}\; \langle\lfloor g(2)\rfloor\rangle_{q}\ldots\]
is $q$-normal, where $g(x)$ is a non-constant polynomial with real
coefficients and the $q-$normal number is constructed by concatenating
the $q-$ary digit expansions $\langle\lfloor g(n)\rfloor\rangle_{q}$
of the integer parts of the values $g(n)$ for $n=1,2,\ldots$. These
constructions were extended to more general classes of functions $g$
(replacing the polynomials) (see
\cite{nakai_shiokawa1992:discrepancy_estimates_class,
nakai_shiokawa1990:class_normal_numbers,
madritsch_thuswaldner_tichy2008:normality_numbers_generated,
madritsch_tichy2013:construction_normal_numbers,
davenport_erdoes1952:note_on_normal,
schiffer1986:discrepancy_normal_numbers}) and the concatenation of
$\langle[g(p)]\rangle_{q}$ along prime numbers instead of the positive
integers (see \cite{nakai_shiokawa1997:normality_numbers_generated,
madritsch2014:construction_normal_numbers,
copeland_erdoes1946:note_on_normal,
madritsch_tichy2013:construction_normal_numbers}).
All such constructions depend on the choice of the base number
$q\geq 2$, and thus they are not suitable for constructing
absolutely normal numbers. A first attempt to construct absolutely
normal numbers is due to Sierpinski
\cite{sierpinski1917:demonstration_elementaire_du}. However,
Turing \cite{turing1992:note_on_normal} observed that Sierpinski's
``construction'' does not yield a computable number, thus it is
not based on a recursive algorithm. Furthermore, Turing gave an
algorithm for a construction of an absolutely normal number. This
algorithm is very slow and, in particular, not polynomially in
time. It is very remarkable that Becher \text{et al.}
\cite{becher_heiber_slaman2013:polynomial_time_algorithm}
established a polynomial time algorithm for the construction of
absolutely normal numbers. However, there remain various questions
concerning the analysis of these algorithms. The discrepancy of
the corresponding sequences is not studied and the order of
convergence of the expansion is very slow and should be
investigated in detail. Furthermore, digital expansions with
respect to linear recurring base sequences seam appropriate to be
included in the study of absolute normality from a computational
point of view.
Let us now return to Poincaré's recurrence theorem which shortly states
that the set $\mathbb{N}$ of positive integers is a recurrence set. In
the 1960s various stronger concepts were introduced:
\begin{enumerate}[(i)]
\item $\mathcal{R}\subseteq \mathbb{N}$ is called a nice recurrence
set if for all invertible dynamical systems and all measurable sets
$A$ of positive measure $\mu(A)>0$ and all $\varepsilon > 0,$ there
exist infinitely many $n\in \mathcal{R}$ such that $$\mu(A\cup
T^{-n}A)>\mu(A)^{2}- \varepsilon.$$
\item $\mathcal{H}\subseteq\mathbb{N}$ is called a van der Corput set
(for short: vdC set) if for all $h\in\mathcal{H}$ the following
implications holds:
\[(x_{n+h}-x_{n})_{n\in\mathbb{N}}\; \text{is u.d.~mod $1$}\Longrightarrow
(x_{n})_{n\in\mathbb{N}}\; \text{is u.d.~mod $1$.}\]
\end{enumerate}
Clearly, any nice recurrence set is a recurrence set. By van der
Corput's difference theorem (see
\cite{kuipers_niederreiter1974:uniform_distribution_sequences,
drmota_tichy1997:sequences_discrepancies_and}) the set
$\mathcal{H}= \N$ of positive integers is a vdC set. Kamae and
Mend\`es-France \cite{kamae_mendes1978:van_der_corputs} proved that
any vdC set is a nice recurrence set. Ruzsa
\cite{ruzsa1984:connections_between_uniform} conjectured that any
recurrence set is also vdC. An important tool in the analysis of
recurrence sets is their equivalence with intersective (or difference)
sets established by Bertrand-Mathis
\cite{bertrand-mathis1986:ensembles_intersectifs_et}. We call a set
$\mathcal{I}$ intersective if for each subset $E\subseteq \mathbb{N}$
of positive (upper) density, there exists $n\in\mathcal{I}$ such that
$n=x-y$ for some $x,y\in E$. Here the upper density of $E$ is defined
as usual by
\[\overline{d}(E)=\limsup_{N\to\infty}\frac{\#(E\cap[1,N])}{N}.\]
Bourgain \cite{bourgain1987:ruzsas_problem_on} gave an example of an
intersective set which is not a vdC set, hence contradicting the above
mentioned conjecture of Ruzsa.
Furstenberg \cite{furstenberg1977:ergodic_behavior_diagonal}
proved that the values $g(n)$ of a polynomial $g\in \mathbb{Z}[x]$
with $g(0)=0$ form an intersective set and later it was shown by
Kamae and Mend\`es-France \cite{kamae_mendes1978:van_der_corputs}
that this is a vdC set, too. It is also known, that for fixed
$h\in \mathbb{Z}$ the set of shifted primes $\{p\pm h\colon
p\text{ prime}\}$ is a vdC set if and only if $h=\pm 1.$
(\cite[Corollary 10]{montgomery1994:ten_lectures_on}). This leads
to interesting applications to additive number theory, for
instance to new proofs and variants of theorems of S\'ark\H{o}zy
\cite{sarkozy1978:difference_sets_sequences1,
sarkoezy1978:difference_sets_sequences3,
sarkoezy1978:difference_sets_sequences2}. A general result
concerning intersective sets related to polynomials along primes is
due to Nair \cite{nair1992:certain_solutions_diophantine}.
In the present paper we want to extend the concept of recurrence sets,
nice recurrence sets and vdC sets to subsets of $\Z^{k},$ following
the program of Bergelson and Lesigne
\cite{bergelson_lesigne2008:van_der_corput} and our earlier
paper~\cite{bergelson_kolesnik_madritsch+2014:uniform_distribution_prime}. In
section 2 we summarize basic facts concerning these concepts,
including general relations between them and counter examples. Section
3 is devoted to a sufficient condition for establishing the vdC
property. In the final section 4 we collect various examples and give
some new applications.
\section{Van der Corput sets}
In this section we provide various equivalent definitions of van der
Corput sets in $\Z^k$. In particular, we give four different
definitions, which are $k$-dimensional variants of the one dimensional
definitions, whose equivalence is due to Ruzsa
\cite{ruzsa1984:connections_between_uniform}. These generalizations
were established by Bergelson and Lesigne
\cite{bergelson_lesigne2008:van_der_corput}. Then we present a set,
which is not a vdC set in order to give some insight into the
structure of vdC sets. Finally, we define the
higher-dimensional variant of nice recurrence sets.
\subsection{Characterization via uniform distribution}
Similarly to above we first define a van der Corput set (vdC set for
short) in $\Z^k$ via
uniform distribution.
\begin{defi}
A subset $\mathcal{H}\subset\Z^k\setminus\{0\}$ is a vdC set if any
family $(x_{\mathbf{n}})_{\mathbf{n}\in\N^k}$ of real numbers is u.d.~mod $1$
provided that it has
the property that for all $\mathbf{h}\in \mathcal{H}$ the family
$(x_{\mathbf{n}+\mathbf{h}}-x_{\mathbf{n}})_{\mathbf{n}\in\N^k}$ is
u.d.~mod $1$.
Here the property of u.d.~mod $1$ for the multi-indexed family
$(x_{\mathbf{n}})_{\mathbf{n}\in\N^k}$ is defined via a natural
extension of \ref{ud}:
\end{defi}
\begin{equation}\label{ud1}
\lim_{N_1,N_2,\ldots,N_k\to+\infty}\frac1{N_1N_2\cdots N_k}
\sum_{0\leq\mathbf{n}<(N_1,N_2,\ldots,N_k)}f(x_{\mathbf n})=
\int^{1}_{0}f(x)dx
\end{equation}
for all continuous functions $f:[0,1]\rightarrow\mathbb{R}.$ Here
in the limit $N_1, N_2,\ldots, N_k$ are tending to infinity
independently and $<$ is defined componentwise.
Using the $k$-dimensional variant of van der Corput's inequality
we could equivalently define a vdC set as follows:
\begin{defi}
A subset $\mathcal{H}\subset\Z^k\setminus\{0\}$ is a van der Corput
set if for any family $(u_\mathbf n)_{n\in\Z^k}$ of complex numbers of
modulus $1$ such that \[\forall\mathbf{h}\in \mathcal{H},\quad
\lim_{N_1,N_2,\ldots,N_k\to+\infty}\frac1{N_1N_2\cdots N_k}
\sum_{0\leq\mathbf{n}<(N_1,N_2,\ldots,N_k)}u_{\mathbf{n}+\mathbf{h}}\overline{u_{\mathbf{n}}}=0\]
the relation
\[\lim_{N_1,N_2,\ldots,N_k\to+\infty}\frac1{N_1N_2\cdots
N_k}
\sum_{0\leq\mathbf{n}<(N_1,N_2,\ldots,N_k)}u_{\mathbf{n}}=0\]
holds.
\end{defi}
\subsection{Trigonometric polynomials and spectral characterization}
The first two definitions are not very useful for proving or
disproving that a set $\mathcal{H}$ is a vdC set. Similar to the one
dimensional case the following spectral characterization involving
trigonometric polynomials is a better tool.
\begin{thm}[{\cite[Proposition
1.18]{bergelson_lesigne2008:van_der_corput}}]
A subset $\mathcal{H}\subset\Z^k\setminus\{0\}$ is a van der Corput set if and
only if for all $\varepsilon>0$, there exists a real trigonometric
polynomial $P$ on the $k$-torus $\T^k$ whose spectrum is contained
in $\mathcal{H}$ and which satisfies $P(0)=1$, $P\geq-\varepsilon$.
\end{thm}
The set of polynomials fulfilling the last theorem for a given
$\varepsilon$ forms a convex set. Moreover the conditions may be
interpreted as some infimum. Therefore we might expect some dual
problem, which is actually provided by the following theorem. For
details see Bergelson and Lesigne
\cite{bergelson_lesigne2008:van_der_corput} or Montgomery
\cite{montgomery1994:ten_lectures_on}.
\begin{thm}[{\cite[Theorem
1.8]{bergelson_lesigne2008:van_der_corput}}]
Let $\mathcal{H}\subset\Z^k\setminus\{0\}$. Then $\mathcal{H}$ is a
van der Corput set if and only if for any positive measure $\sigma$
on the $k$-torus $\T^k$ such that, for all $\mathbf{h}\in \mathcal{H}$,
$\widehat{\sigma}(\mathbf{h})=0$, this implies
$\sigma(\{(0,0,\ldots,0)\})=0$.
\end{thm}
\subsection{Examples}
The structure of vdC sets is better understood by first giving a
counter example. The following lemma shows to be very useful in
the construction of counter examples.
\begin{lem}\label{mt:infinite_intersection}
Let $\mathcal{H}\subset\N$. If there exists $q\in\N$ such that the
set $\mathcal{H}\cap q\N$ is finite, then the set $\mathcal{H}$ is
not a vdC set.
\end{lem}
\begin{proof}
The proof is a combination of the following two observations of Ruzsa
\cite{ruzsa1984:connections_between_uniform} (see Theorem 2
and Corollary 3 of \cite{montgomery1994:ten_lectures_on}):
\begin{enumerate}
\item Let $m\in\N$. The sets $\{1,\ldots,m\}$ and $\{n\in\N\colon
m\nmid n\}$ are both not vdC sets.
\item Let $\mathcal{H}=\mathcal{H}_1\cup\mathcal{H}_2\subset\N$. If
$\mathcal{H}$ is a vdC set, then $\mathcal{H}_1$ or $\mathcal{H}_2$
also has to be a vdC set.
\end{enumerate}
Suppose there exists a $q\in\N$ such that $\mathcal{H}\cap q\N$ is
finite. Then we may split $\mathcal{H}$ into the sets
$\mathcal{H}\cap q\N$ and $\mathcal{H}\setminus q\N$. The first
one is finite and the second one contains no multiples of $q$.
Therefore both are not vdC sets and hence $\mathcal{H}$ is not a
vdC set.
\end{proof}
The first counter example deals with arithmetic progressions.
\begin{lem}\label{lem:arithmetic_prog}
Let $a,b\in\N$. If the set $\{an+b\colon n\in\N\}$ is a vdC set,
then $a\mid b$.
\end{lem}
\begin{proof}
Let $b\in\N$ and $\mathcal{H}=\{an+b\colon n\in\N\}$ be a
vdC set. Then by Lemma \ref{mt:infinite_intersection} we must have
\[an+b\equiv b\equiv 0\bmod a\quad\text{infinitely often.}\]
This implies that $a\mid b$.
\end{proof}
The sufficiency (and also the necessity) of the requirement $a\mid b$
follows from the following result of Kamae and Mend\`es-France
\cite{kamae_mendes1978:van_der_corputs} (\textit{cf.} Corollary 9 of
\cite{montgomery1994:ten_lectures_on}).
\begin{lem}
Let $P(z)\in\Z[z]$ and suppose that $P(z)\to+\infty$ as
$z\to+\infty$. Then $\mathcal{H}=\{P(n)>0\colon n\in\N\}$ is a vdC set
if and only if for every positive integer $q$ the congruence
$P(z)\equiv 0\pmod q$ has a root.
\end{lem}
Now we want to establish a similar result for sets of the form
$\{ap+b\colon p\text{ prime}\}$. In this case the following result is
due to Bergelson and Lesigne
\cite{bergelson_lesigne2008:van_der_corput} which is a generalization of
the case $f(x)=x$ due to Kamae and Mend\`es-France
\cite{kamae_mendes1978:van_der_corputs}.
\begin{lem}[{\cite[Proposition
1.22]{bergelson_lesigne2008:van_der_corput}}]\label{bl:prop1.22}
Let $f$ be a (non zero) polynomial with integer coefficients and
zero constant term. Then the sets $\{f(p-1)\colon
p\in\mathbb{P}\}$ and $\{f(p+1)\colon p\in\mathbb{P}\}$ are vdC
sets in $\Z$.
\end{lem}
We show the converse direction.
\begin{lem}
Let $a$ and $b$ be non-zero integers. Then the set $\{ap+b\colon
p\in\mathbb{P}\}$ is a vdC set if and only if $\lvert a\rvert=\lvert
b\rvert$, \textit{i.e.} $ap+b=a(p\pm 1)$.
\end{lem}
\begin{proof}
It is clear from Lemma \ref{bl:prop1.22} that $\{ap+b\colon
p\in\mathbb{P}\}$ is a vdC set if $\lvert a\rvert=\lvert b\rvert$.
On the contrary a combination of Lemma \ref{mt:infinite_intersection}
and Lemma \ref{lem:arithmetic_prog} yields that $a\mid b$. Now we
consider the sequence modulo $b$. Then by Lemma
\ref{mt:infinite_intersection} we get that
\[ap+b\equiv ap\equiv 0\bmod b\quad\text{infinitely often.}\] Since $(p,b)>1$ only
holds for finitely many primes $p$ we must have $b\mid a$. Combining
these two requirements yields $\lvert a\rvert=\lvert b\rvert$.
\end{proof}
\section{A sufficient condition}
In this section we want to formulate a general sufficient
condition which provides us with a tool to show for plenty of
different examples that they generate a vdC set. This is a
generalization of the conditions of Kamae and Mend\`es-France
\cite{kamae_mendes1978:van_der_corputs} and Bergelson and Lesigne
\cite{bergelson_lesigne2008:van_der_corput}. Before stating the
condition we need an auxiliary lemma.
\begin{lem}[{\cite[Corollary
1.15]{bergelson_lesigne2008:van_der_corput}}]
\label{bl:lem_linear_algebra}
Let $d$ and $e$ be positive integers, and let $L$ be a linear
transformation from $\Z^d$ into $\Z^e$ (represented by an $e\times
d$ matrix with integer entries). Then the following assertions hold:
\begin{enumerate}
\item If $D$ is a vdC set in $\Z^d$ and if $0\not\in L(D)$, then
$L(D)$ is a vdC set in $\Z^e$.
\item Let $D\in\Z^d$. If the linear map $L$ is one-to-one, and if
$L(D)$ is a vdC set in $\Z^e$, then $D$ is a vdC set in $\Z^d$.
\end{enumerate}
\end{lem}
Our main tool is the following general result. Applications
are given in the next section.
\begin{prop}\label{mt:sufficient_condition}
Let $g_1,\ldots,g_k\colon\N\to\Z$ be arithmetic functions. Suppose
that $g_{i_1},\ldots,g_{i_m}$ is a basis of the $\Q$-vector space
$\mathrm{span}(g_1,\ldots,g_k)$. For each $q\in\N$, we introduce
\[D_q:=\left\{(g_{i_1}(n),\ldots,g_{i_m}(n)\colon n\in\N\text{ and $q!\mid
g_{i_j}(n)$ for all $j=1,\ldots,m$}\right\}.\]
Suppose further that, for every $q$, there exists a sequence
$(h^{(q)}_n)_{n\in\N}$ in $D_q$ such that, for all
$\mathbf{x}=(x_1,\ldots,x_m)\in\R^m\setminus\Q^m$, the sequence
$(h^{(q)}_n\cdot\mathbf{x})_{n\in\N}$ is uniformly distributed mod
$1$. Then
\[\widetilde{D}:=\{(g_1(n),\ldots,g_k(n))\colon n\in\N\}\in\Z^k\]
is a vdC set.
\end{prop}
\begin{proof}
We first show that the set
\[D:=\{(g_{i_1}(n),\ldots,g_{i_m}(n))\colon n\in\N\}\]
is a vdC set in $\Z^m$. For $q,N\in\N$ we define a family of trigonometric
polynomials
\[P_{q,N}:=\frac1N\sum_{n=1}^Ne\left(h_n^{(q)}\cdot\mathbf{x}\right).\]
By hypothesis, $\lim_{N\to\infty}P_{q,N}(x)=0$ for
$x\not\in\Q^m$. For fixed $q$ there exists a subsequence $(P_{q,N'})$
which converges pointwise to a function $g_q$. Since $g_q(x)=1$
(for $x\in\Q^m$ and $q$ sufficiently large) and $g_q(x)=0$ (for
$x\not\in\Q^m$), the sequence $(g_q)$ is pointwise convergent to the indicator
function of $\Q^m$. For a positive measure $\sigma$ on the
$m$-dimensional torus with vanishing Fourier transform
$\widehat{\sigma}$ on $D$, we have $\int P_{q,N}\mathrm{d}\sigma=0$
for all $q,N$. Thus $\sigma(\Q^m)=0$ follows from the dominating
convergence theorem, obviously $\sigma(\{0,0,\ldots,0\})=0$, and thus $D$ is a vdC
set.
In order to prove that $\widetilde{D}$ is a vdC set we apply Lemma
\ref{bl:lem_linear_algebra} twice. Since $g_{i_1},\ldots,g_{i_m}$ is a
base of $\mathrm{span}(g_1,\ldots,g_k)$, we can write each $g_j$ as a
linear combination (with rational coefficients) of
$g_{i_1},\ldots,g_{i_m}$. Multiplying with the common denominator of
the coefficients yields
\[a_jg_j=b_{j,1}g_{i_1}+\cdots+b_{j,m}g_{i_m}\]
for $j=1,\ldots,k$ and certain $a_j,b_{j,\ell}\in\Z$. Considering the
transformation $L\colon\Z^m\to\Z^k$ given by the matrix $(b_{j,\ell})$
and applying part (1) of Lemma \ref{bl:lem_linear_algebra} shows that
\[\{(a_1g_1(n),\ldots,a_kg_k(n))\colon n\in\N\}\]
is a vdC set for certain integers $a_1,\ldots,a_k$.
Now consider the transformation $\widetilde{L}\colon\Z^k\to\Z^k$ given by
the $k\times k$ diagonal matrix with entries $a_1,\ldots,a_k$ in the
diagonal. Then by part (2) of Lemma \ref{bl:lem_linear_algebra} also
$\widetilde{D}$ is a vdC set and the proposition is proved.
\end{proof}
\section{Various examples and applications to additive problems}
In this section we consider multidimensional variants of prime
powers, entire functions and $x^\alpha\log^\beta x$ sequences.
\subsection{Prime powers}
In a recent paper the authors together with Bergelson, Kolesnik and
Son
\cite{bergelson_kolesnik_madritsch+2014:uniform_distribution_prime}
consider sets of the form
\[\{(\alpha_1(p_n\pm1)^{\theta_1},\ldots,\alpha_k(p_n\pm1)^{\theta_k})\colon
n\in\N\},\] where $\alpha_i,\beta_i\in\R$ and $p_n\in\mathcal{P} $
runs over all prime numbers. These sets are vdC, however, we
missed the treatment of a special case in the proof. In
particular, if for some $i\neq j$ the exponents satisfy
$\theta_i=\theta_j=:\theta$, then the vector
$(p_n^{\theta},p_n^{\theta})$ is not uniformly distributed mod 1.
Here we close this gap.
\begin{thm}
If $\alpha_i$ are positive integers and $\beta_i$ are positive and non-integers, then
$$ D_1 = \{ \left( (p-1)^{\alpha_1}, \cdots , (p-1)^{\alpha_k}, [(p-1)^{\beta_1}], \cdots , [(p-1)^{\beta_\ell}] \right) | \, p \in\mathcal{P} \},$$
and
$$ D_2 = \{ \left( (p+1)^{\alpha_1}, \cdots , (p+1)^{\alpha_k}, [(p+1)^{\beta_1}], \cdots , [(p+1)^{\beta_\ell}] \right) | \, p \in \mathcal{P} \}$$
are vdC sets in $\Z^{k+\ell}$.
\end{thm}
\begin{proof}
Since $x^{\theta_1}$ and $x^{\theta_2}$ are $\Q$-linear dependent
for all $x\in\Z$ if and only if $\theta_1=\theta_2$, an
application of Proposition \ref{mt:sufficient_condition} yields
that it suffices to consider the case where all exponents are
different. However, this follows by the same arguments as in the
proof of Theorem 4.1 in
\cite{bergelson_kolesnik_madritsch+2014:uniform_distribution_prime}.
\end{proof}
\subsection{Entire functions}
In this section we consider entire functions of bounded logarithmic
order. We fix a transcendental entire function $f$ and denote by
$S(r):=\max_{\lvert z\rvert\leq r}\lvert f(z)\rvert$. Then we call
$\lambda$ the logarithmic order of $f$ if
\[\limsup_{r\to\infty}\frac{\log S(r)}{\log r}=\lambda.\]
The central tool is the following result of Baker
\cite{baker1984:entire_functions_and}.
\begin{thm}[{\cite[Theorem 2]{baker1984:entire_functions_and}}]\label{baker:uniform_distribution_of_entire_functions}
Let $f$ be a transcendental entire function of logarithmic order
$1<\lambda<\frac43$. Then the sequence
\[\left(f(p_n)\right)_{n\geq1}\]
is uniformly distributed mod $1$.
\end{thm}
Our second example of a class of vdC sets is the following.
\begin{thm}\label{thm:entire_functions_vdC}
Let $f_1,\ldots,f_k$ be entire functions with distinct logarithmic orders
$1<\lambda_1,\lambda_2,\ldots,\lambda_k<\frac43$, respectively.
Then the set
\[D:=\{(\lfloor f_1(p_n)\rfloor,\ldots,\lfloor f_k(p_n)\rfloor)\colon
n\in\N\}\]
is a vdC set.
\end{thm}
\begin{proof}
We enumerate $D=(\mathbf{d}_n)_{n\geq1}$, where
\[\mathbf{d}_n:=\left(\lfloor f_1(p_n)\rfloor,\ldots,\lfloor f_k(p_n)\rfloor\right).\]
First we show, that for every $q\in\N$ the set
\[D^{(q)}:=\{(d_1,\ldots,d_k)\in D\colon q\mid d_i\}\] has positive
relative density in $D$. We note that if
$0\leq\left\{\frac{f_i(p_n)}{q}\right\}<\frac1q$ for $1\leq i\leq k$,
then $\mathbf{d}_n\in D^{(q)}$. By Theorem
\ref{baker:uniform_distribution_of_entire_functions} the
sequence \[\left(\left(\frac{f_1(p_n)}{q},\ldots,\frac{f_k(p_n)}q\right)\right)_{n\geq1}\]
is uniformly distributed and thus $D^{(q)}$ has positive density in
$D$.
For each $q\in\N$ we enumerate the elements of
$D^{(q!)}=(\mathbf{d}^{(q!)}_n)_{n\geq1}$, such that
$\lvert\mathbf{d}^{(q!)}_n\rvert$ is increasing. Since the
logarithmic orders are distinct we immediately get that the
functions $f_i$ are $\Q$-linearly independent. Thus by Proposition
\ref{mt:sufficient_condition} it is sufficient to show that for all
$q\in\N$ and all $\mathbf{x}=(x_1,\ldots,x_k)\in\R^k\setminus\Q^k$
the sequence $(\mathbf{d}^{(q!)}_n\cdot\mathbf{x})_{n\geq1}$ is
u.d. mod 1.
Using the orthogonality relations for additive characters we get for
any non-zero integer $h$, that
\begin{multline*}
\frac1{\lvert\{n\leq N\colon\mathbf{d}_n\in D^{(q!)}\}\rvert}
\sum_{n\leq N}e\left(h\left(d^{(q!)}_n\cdot\mathbf{x}\right)\right)\\
=\frac1{\lvert\{n\leq N\colon\mathbf{d}_n\in D^{(q!)}\}\rvert}
\frac1{(q!)^k}\sum_{j_1=1}^{q!}\cdots\sum_{j_k=1}^{q!}
\frac1N\sum_{n\leq
N}e\left(d_n\cdot\left(h\mathbf{x}+\left(\frac{j_1}{q!},\ldots,\frac{j_k}{q!}\right)\right)\right).
\end{multline*}
The innermost sum is of the form
\[\sum_{n\leq N}e(g(p_n)),\]
with $g(x)=\sum_{i=1}^k\alpha_i\lfloor f_i(x)\rfloor$ for a certain
$(\alpha_1,\ldots,\alpha_k)\in\R^k\setminus\Q^k$.
By relabeling the terms we may suppose that there exists an $\ell$
such that $\alpha_1,\ldots,\alpha_\ell\not\in\Q$ and
$\alpha_{\ell+1},\ldots,\alpha_k\in\Q$. Furthermore we may write
$\alpha_j=\frac{a_j}q$ for $\ell+1\leq j\leq m$. Then
\[e(g(p_n))=e\left(\sum_{i=1}^k\alpha_k\lfloor
f_i(p_n)\rfloor\right)= \prod_{j=1}^\ell
s_j(\alpha_jf_j(p_n),f_j(p_n))\prod_{j=\ell+1}^kt_j(\lfloor
f_j(p_n)\rfloor),\] where $s_j(x,y)=e(x-\{y\}\alpha_j)$ ($1\leq
j\leq\ell$) and $t_j(z)=e\left(a_j\frac zq\right)$ ($\ell+1\leq
j\leq k$).
Since $s_j(x,y)$ is Riemann-integrable on $\T^2$ for
$j=1,\ldots,\ell$ \; and $t_j(z)$ is continuous on $\Z_q=\Z/q\Z$, the
function $\prod_{j=1}^\ell s_j\prod_{j=\ell+1}^kt_j$ is
Riemann-integrable on $\T^{2\ell}\times\Z_q^{k-\ell}$.
Now an application of Theorem
\ref{baker:uniform_distribution_of_entire_functions} yields that for
any $u\in\N$ the sequence
\[\left(\alpha_1f_1(p_n),f_1(p_n),\ldots,\alpha_\ell
f_\ell(p_n),f_\ell(p_n),\frac{f_{\ell+1}(p_n)}u,\ldots,\frac{f_k(p_n)}u\right)_{n\geq1}\]
is u.d. in $\T^{2\ell}\times\T^{k-\ell}$. Since $\lfloor
x\rfloor\equiv a\pmod q$ is equivalent to $\frac xq\in[\frac
aq,\frac{a+1}q]$, we deduce that
\[\left(\alpha_1f_1(p_n),f_1(p_n),\ldots,\alpha_\ell
f_\ell(p_n),f_\ell(p_n),\lfloor
f_{\ell+1}(p_n)\rfloor,\ldots,\lfloor
f_k(p_n)\rfloor\right)_{n\geq1}\]
is u.d. in $\T^{2\ell}\times\Z_q^{k-\ell}$, and Weyl's criterion
implies that
\[\lim_{N\to\infty}\frac1N\sum_{n\leq
N}e\left(\sum_{i=1}^k\alpha_i\lfloor f_i(p_n)\rfloor\right)=0,\]
proving the theorem.
\end{proof}
\subsection{Functions of the form $x^\alpha\log^\beta x$}
In the one-dimensional case Boshernitzan \textit{et al.}
\cite{boshernitzan_kolesnik_quas+2005:ergodic_averaging_sequences}
showed, among other things, that these sets are vdC sets. Our aim
is to show an extended result for the $k$-dimensional case.
Therefore we use the following general criterion, which is a
combination of Fejer's theorem and van der Corput's difference
theorem.
\begin{thm}[{\cite[Theorem 3.5]{kuipers_niederreiter1974:uniform_distribution_sequences}}]
Let $f(x)$ be a function defined for $x > 1$ that is $k$-times
differentiable for $x > x_0$. If $f^{(k)}(x)$ tends monotonically to
$0$ as $x\to\infty$ and if \; $\lim_{x\to\infty}x\lvert
f^{(k)}(x)\rvert=\infty$, then the sequence $(f(n))_{n\geq1}$ is
u.d. mod 1.
\end{thm}
Applying this theorem we get the following
\begin{cor}\label{cor:uniform_distribution_of_n_log_powers}
Let $\alpha\neq0$ and
\begin{itemize}
\item either $\sigma>0$ not an integer and $\tau\in\R$ arbitrary
\item or $\sigma>0$ an integer and $\tau\in\R\setminus[0,1]$.
\end{itemize}
Then the sequence $(\alpha n^\sigma\log^\tau n)_{n\geq2}$ is
u.d. mod 1.
\end{cor}
Our third example is the following class of vdC sets.
\begin{thm}
Let $\alpha_1,\ldots,\alpha_k>0$ and $\beta_1,\ldots,\beta_k\in\R$,
such that $\beta_i\not\in[0,1]$ whenever $\alpha_i\in\Z$ for
$i=1,\ldots,k$. Then the set
\[D:=\{(\lfloor n^{\alpha_1}\log^{\beta_1}n\rfloor,\ldots,\lfloor
n^{\alpha_k}\log^{\beta_k}n\rfloor)\colon n\in\N\}\]
is a vdC set.
\end{thm}
\begin{proof}
Following the same arguments as is the proof of Theorem
\ref{thm:entire_functions_vdC} and replacing the uniform distribution
result for entire functions (Theorem \ref{thm:entire_functions_vdC})
by the corresponding result for $n^\alpha\log^\beta n$ sequences (Corollary
\ref{cor:uniform_distribution_of_n_log_powers}) yields the proof.
\end{proof}
\section*{Acknowledgment}
This research work was done when the first author was a visiting
lecturer at the Department of Analysis and Computational Number Theory
at Graz University of Technology. The author thanks the institution
for its hospitality.
The second author acknowledges support of the project F 5510-N26
within the special research area ``Quasi Monte-Carlo Methods and
applications'' founded by the Austrian Science Fund.
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
\para{Associahedra}
The $n$-dimensional \defn{associahedron} is a simple polytope whose $\frac{1}{n+2}\binom{2n+2}{n+1}$ vertices correspond to Catalan objects (triangulations of an $(n+3)$-gon, binary trees on $n+1$ nodes, ...) and whose edges correspond to mutations between them (diagonal flips, edge rotations, ...). Its combinatorial structure appeared in early works of D.~Tamari~\cite{Tamari} and \mbox{J.~Stasheff~\cite{Stasheff}}, and was first realized as a convex polytope by M.~Haiman~\cite{Haiman} and C.~Lee~\cite{Lee}. Since then, the associahedron has motivated a flourishing research trend with rich connections to combinatorics, geometry and algebra: polytopal constructions~\cite{Loday, HohlwegLange, CeballosSantosZiegler, LangePilaud}, Tamari and Cambrian lattices~\cite{TamariFestschrift, Reading-latticeCongruences, Reading-CambrianLattices}, diameter and Hamiltonicity~\cite{SleatorTarjanThurston, Dehornoy, Pournin, HurtadoNoy}, geometric properties~\cite{BergeronHohlwegLangeThomas, HohlwegLortieRaymond, PilaudStump-barycenter}, combinatorial Hopf algebras \cite{LodayRonco, HivertNovelliThibon-algebraBinarySearchTrees, Chapoton, ChatelPilaud}, to cite a few. The associahedron was also generalized in several directions, in particular to secondary and fiber polytopes~\cite{GelfandKapranovZelevinsky, BilleraFillimanSturmfels}, graph associahedra and nestohedra~\cite{CarrDevadoss, Devadoss, Postnikov, FeichtnerSturmfels, Zelevinsky, Pilaud}, pseudotriangulation polytopes~\cite{RoteSantosStreinu-polytope}, cluster complexes and generalized associahedra~\cite{FominZelevinsky-YSystems, ChapotonFominZelevinsky, HohlwegLangeThomas, Stella, Hohlweg}, and brick polytopes~\cite{PilaudSantos-brickPolytope, PilaudStump-brickPolytope}.
\para{Graph associahedra}
This paper deals with graph associahedra, which were defined by M.~Carr and S.~Devadoss~\cite{CarrDevadoss} in connection to C.~De Concini and C.~Procesi's wonderful arrangements~\cite{DeConciniProcesi}. Given a simple graph~$\graphG$ with~$\connectedComponents$ connected components and~$n+\connectedComponents$ vertices, the \defn{$\graphG$-associahedron}~$\Asso(\graphG)$ is an $n$-dimensional simple polytope whose combinatorial structure encodes the connected subgraphs of~$\graphG$ and their nested structure. More precisely, the face lattice of the polar of the \mbox{$\graphG$-associahedron} is isomorphic to the \defn{nested complex}~$\Nest(\graphG)$ on~$\graphG$, defined as the simplicial complex of all collections of tubes (connected induced subgraphs) of~$\graphG$ which are pairwise compatible (either nested, or disjoint and non-adjacent). As illustrated in Figures~\ref{fig:associahedron}, \ref{fig:cyclohedron} and~\ref{fig:permutahedron}, the graph associahedra of certain special families of graphs happen to coincide with well-known families of polytopes: classical associahedra are path associahedra, cyclohedra are cycle associahedra, and permutahedra are complete graph associahedra. The graph associahedra were extended to the \defn{nestohedra}, which are simple polytopes realizing the nested complex of arbitrary building sets~\cite{Postnikov, FeichtnerSturmfels}. Graph associahedra and nestohedra have been geometrically realized in different ways: by successive truncations of faces of the standard simplex~\cite{CarrDevadoss}, as Minkowski sums of faces of the standard simplex~\cite{Postnikov, FeichtnerSturmfels}, or from their normal fans by exhibiting explicit inequality descriptions~\cite{Devadoss, Zelevinsky}. For a given graph~$\graphG$, the resulting polytopes all have the same normal fan which coarsens the type~$A$ Coxeter arrangement: its rays are the characteristic vectors of the tubes, and its cones are generated by characteristic vectors of compatible tubes. Alternative realizations of graph associahedra with different normal fans are obtained by successive truncations of faces of the cube in~\cite{Volodin, DevadossForceyReisdorfShowers}. The objective of this paper is to provide a new unrelated family of complete simplicial fans realizing the graphical nested complex~$\Nest(\graphG)$ for any graph~$\graphG$.
\begin{figure}[p]
\capstart
\centerline{\includegraphics[scale=1.13]{correspondenceAssociahedron}}
\caption{The classical associahedron (left) is the path associahedron (right).}
\label{fig:associahedron}
\end{figure}
\begin{figure}[p]
\capstart
\centerline{\includegraphics[scale=1.13]{correspondenceCyclohedron}}
\caption{The cyclohedron (left) is the cycle associahedron (right).}
\label{fig:cyclohedron}
\end{figure}
\begin{figure}[p]
\capstart
\centerline{\includegraphics[scale=1.13]{correspondencePermutahedron}}
\caption{The permutahedron (left) is the complete graph associahedron (right).}
\label{fig:permutahedron}
\end{figure}
\para{Cluster algebras and cluster fans}
Our construction is directly inspired from combinatorial and geometric properties of finite type cluster algebras and generalized associahedra introduced by S.~Fomin and A.~Zelevinsky in~\cite{FominZelevinsky-ClusterAlgebrasI, FominZelevinsky-ClusterAlgebrasII, FominZelevinsky-YSystems}. Note that A.~Zelevinsky~\cite{Zelevinsky} already underlined the closed connection between nested complexes, nested fans, and nestohedra on one hand and cluster complexes, cluster fans and generalized associahedra on the other hand. More recently, T.~Lam and P.~Pylyavskyy established an even deeper connection with linear Laurent Phenomenon algebras~\cite{LamPylyavskyy-LaurentPhenomenonAlgebras, LamPylyavskyy-LinearLaurentPhenomenonAlgebras}, see the discussion in Section~\ref{subsec:LPA}. In this paper, we use ideas from cluster algebras to obtain results on graphical nested complexes, which in turn translate to relevant properties of the geometry of finite type cluster algebras.
\defn{Cluster algebras} are commutative rings generated by a set of \defn{cluster variables} grouped into overlapping \defn{clusters}. The clusters are obtained from an initial cluster~$X^\circ$ by a mutation process. Each mutation exchanges a single variable in a cluster according to a formula controlled by a combinatorial object. We refer to~\cite{FominZelevinsky-ClusterAlgebrasI} since precise details on this mutation are not needed here. Two cluster variables are \defn{compatible} if they belong to a same cluster and \defn{exchangeable} if they are not compatible and belong to two clusters connected by a mutation. The \defn{cluster complex} of a cluster algebra~$\cA$ is the pure simplicial complex whose vertices are cluster variables of~$\cA$ and whose facets are the clusters of~$\cA$. A cluster algebra is of \defn{finite type} if it has finitely many cluster variables, and thus if its cluster complex is finite. Finite type cluster algebras were classified in~\cite{FominZelevinsky-ClusterAlgebrasII}: up to isomorphism, there is one finite type cluster algebra for each finite crystallographic root system.
The Laurent Phenomenon~\cite{FominZelevinsky-ClusterAlgebrasI} asserts that each cluster variable~$x$ can be expressed as a Laurent polynomial in terms of the cluster variables~$x_1^\circ, \dots, x_n^\circ$ of the initial cluster~$X^\circ$. The \defn{$\b{d}$-vector} of~$x$ with respect to~$X^\circ$ is the vector~$\b{d}(X^\circ, x)$ whose $i$th coordinate is the exponent of the initial variable~$x_i^\circ$ in the denominator of~$x$. In finite type, this exponent was also interpreted in~\cite{FominZelevinsky-YSystems, CeballosPilaud} as the \defn{compatibility degree}~$\compatibilityDegree{x_i^\circ}{x}$ between the cluster variables~$x_i^\circ$ and~$x$. The compatiblity degree~$\compatibilityDegree{\cdot}{\cdot}$ has the following properties: for any distinct cluster variables~$x$ and~$x'$, we have~$\compatibilityDegree{x}{x'} \ge 0$ with equality if and only if~$x$ and~$x'$ are compatible, and~$\compatibilityDegree{x}{x'} = 1 = \compatibilityDegree{x'}{x}$ if and only if $x$ and~$x'$ are exchangeable. The $\b{d}$-vectors can be used to construct a simplicial fan realization of the cluster complex, called \defn{$\b{d}$-vector fan}: it is known for certain initial clusters in finite type cluster algebras, that the cones generated by the $\b{d}$-vectors of all collections of compatible cluster variables form a complete simplicial fan realizing the cluster complex. This is proved by S.~Fomin and A.~Zelevinsky~\cite{FominZelevinsky-ClusterAlgebrasII} for the bipartite initial cluster, by S.~Stella~\cite{Stella} for all acyclic initial clusters, and by F.~Santos~\cite[Section~5]{CeballosSantosZiegler} for any initial cluster in type~$A$. In fact, we expect this property to hold for any initial cluster, acyclic or not, of any finite type cluster algebra: this paper proves it for types~$A$, $B$, and~$C$ (our general proof provides as a particular case a new proof in type~$A$, similar to that of~\cite[Section~5]{CeballosSantosZiegler}), and the other finite types are investigated in a current project of the authors in collaboration with C.~Ceballos.
There is another complete simplicial fan realizing the cluster complex, whose rays are now given by the \defn{$\b{g}$-vectors} of the cluster variables, defined as the multi-degrees of the cluster variables expressed in the cluster algebra with principal coefficients~\cite{FominZelevinsky-ClusterAlgebrasIV}. The fact that the cones generated by the $\b{g}$-vectors of all collections of compatible cluster variables form a complete simplicial fan is a consequence of~\cite{Reading-UniversalClusterAlgebra}. When the initial cluster is acyclic, the resulting $\b{g}$-vector fan is the \defn{Cambrian fan} of N.~Reading and D.~Speyer~\cite{ReadingSpeyer} and it coarsens the Coxeter fan.
\para{Compatibility fans for graphical nested complexes}
\enlargethispage{.1cm}
Motivated by the combinatorial and geometric richness of the~$\b{d}$- and $\b{g}$-vector fans described above for finite type cluster algebras, we want to construct various simplicial fan realizations of the nested complex. As it coarsens the Coxeter fan, the normal fan of the graph-associahedra and nestohedra of~\cite{CarrDevadoss, Devadoss, Postnikov, FeichtnerSturmfels, Zelevinsky} should be considered as an analogue of the $\b{g}$-vector fan. A tentative approach to construct alternative $\b{g}$-vector fans for tree associahedra can be found in~\cite{Pilaud}.
In this paper, we construct an analogue of the $\b{d}$-vector fan for graphical nested complexes. For two tubes~$\tube, \tube'$ of a graph~$\graphG$, we define the \defn{compatibility degree}~$\compatibilityDegree{\tube}{\tube'}$ of~$\tube$ with~$\tube'$ to be~$\compatibilityDegree{\tube}{\tube'} = -1$ if~$\tube = \tube'$, $\compatibilityDegree{\tube}{\tube'} = 0$ if~$\tube \ne \tube'$ are compatible, and~${\compatibilityDegree{\tube}{\tube'} = |\{\text{neighbors of $\tube$ in $\tube' \ssm \tube$}\}|}$ otherwise. Similar to the compatibility degree for cluster algebras, it satisfies~$\compatibilityDegree{\tube}{\tube'} \ge 0$ for any distinct tubes~$\tube, \tube'$ of~$\graphG$, with equality if and only if $\tube$ and~$\tube'$ are compatible, and~$\compatibilityDegree{\tube}{\tube'} = 1 = \compatibilityDegree{\tube'}{\tube}$ if and only if $\tube$ and~$\tube'$ are exchangeable. We define the \defn{compatibility vector}~$\compatibilityVector{\tubing^\circ}{\tube} \eqdef \left[ \compatibilityDegree{\tube_1^\circ}{\tube}, \dots, \compatibilityDegree{\tube_n^\circ}{\tube} \right]$ of a tube~$\tube$ with respect to an initial maximal tubing~$\tubing^\circ \eqdef \{\tube_1^\circ, \dots, \tube_n^\circ\}$, and the \defn{compatibility matrix}~$\compatibilityMatrix{\tubing^\circ}{\tubing} \eqdef [\compatibilityDegree{\tube_i^\circ}{\tube_j}]_{i \in [n], j \in [m]}$ of a tubing~$\tubing \eqdef \{\tube_1, \dots, \tube_m\}$ of~$\graphG$ with respect to the initial tubing~$\tubing^\circ$. We write~$\R_{\ge 0} \, \compatibilityMatrix{\tubing^\circ}{\tubing}$ to denote the polyhedral cone generated by the compatibility vectors of the tubes of~$\tubing$ with respect to~$\tubing^\circ$.
Notice that our degree is asymmetric, which yields a natural notion of duality. We define the \defn{dual compatibility vector}~$\dualCompatibilityVector{\tube}{\tubing^\circ} \eqdef \left[ \compatibilityDegree{\tube}{\tube_1^\circ}, \dots, \compatibilityDegree{\tube}{\tube_n^\circ} \right]$ of~$\tube$ with respect to~$\tubing^\circ$ and the \defn{dual compatibility matrix}~$\dualCompatibilityMatrix{\tubing}{\tubing^\circ} \eqdef [\compatibilityDegree{\tube_j}{\tube_i^\circ}]_{i\in [n], j \in [m]}$. To make the distinction clear from the dual compatibility vector, we often call~$\compatibilityVector{\tubing^\circ}{\tube}$ the \defn{primal} compatibility vector.
Although there is no denominators involved anymore, we still use the letter~$\b{d}$ to stand for compatibility \textbf{d}egree vector, and to match with the cluster algebra notations. Indeed, our compatibility degrees for paths and cycles coincide with the compatibility degree of~\cite{FominZelevinsky-YSystems} in types~$A$, $B$, and~$C$. Compatibility degrees on type~$A$ cluster variables correspond to compatibility (and dual compatibility) degrees on tubes of paths while compatibility degrees in type~$C$ (resp.~$B$) cluster variables correspond to compatibility (resp.~dual compatibility) degrees on tubes of cycles.
Our main result is the following analogue of the compatibility fan for path associahedra constructed by F.~Santos in~\cite[Section~5]{CeballosSantosZiegler}.
\begin{theorem}
For any graph~$\graphG$, the compatibility vectors (resp.~dual compatibility vectors) of all tubes of~$\graphG$ with respect to any initial maximal tubing~$\tubing^\circ$ on~$\graphG$ support a complete simplicial fan realizing the nested complex on~$\graphG$. More precisely, both collections of cones
\[
\compatibilityFan{\graphG}{\tubing^\circ} \eqdef \bigset{\R_{\ge 0} \, \compatibilityVector{\tubing^\circ}{\tubing}}{\tubing \text{ tubing on } \graphG}
\;\;\,\text{and}\;\;\;\;
\dualCompatibilityFan{\graphG}{\tubing^\circ} \eqdef \bigset{\R_{\ge 0} \, \dualCompatibilityMatrix{\tubing}{\tubing^\circ}}{\tubing \text{ tubing on } \graphG}
\]
are complete simplicial fans.
\end{theorem}
We then study the number of distinct compatibility fans. As in~\cite{CeballosSantosZiegler}, we consider that two compatibility fans~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\compatibilityFan{\graphG'}{\tubing'^\circ}$ are equivalent if they differ by a linear isomorphism. Such a linear isomorphism induces an isomorphism between the nested complexes~$\nestedComplex(\graphG)$ and~$\nestedComplex(\graphG')$. Besides those induced by graph isomorphisms between~$\graphG$ and~$\graphG'$, there are non-trivial nested complex isomorphisms: for example, the complementation~$\tube \to \ground \ssm \tube$ on the complete graph, or the map on tubes of the path corresponding to the rotation of diagonals in the polygon. Extending these two examples, we exhibit a non-trivial nested complex isomorphism on any spider (a set of paths attached by one endpoint to a clique). Our next statement shows that these are essentially the only non-trivial nested complex isomorphisms.
\begin{theorem}
All nested complex isomorphisms~$\nestedComplex(\graphG) \to \nestedComplex(\graphG')$ are induced by graph isomorphisms~$\graphG \to \graphG'$, except if one of the connected components of~$\graphG$ is a spider.
\end{theorem}
\begin{corollary}
Except if one of the connected components of~$\graphG$ is a path, the number of linear isomorphism classes of compatibility fans of~$\graphG$ is the number of orbits of maximal tubings on~$\graphG$ under graph automorphisms of~$\graphG$.
\end{corollary}
The next step would be to realize all these complete simplicial fans as normal fans of convex polytopes. This question remains open, except for some particular graphs: besides all graphs with at most~$4$ vertices, we settle the case of paths and cycles following a similar proof as~\cite{CeballosSantosZiegler}.
\begin{theorem}
All compatibility and dual compatibility fans of paths and cycles are polytopal.
\end{theorem}
\para{Overview}
The paper is organized as follows. We first recall in Section~\ref{sec:preliminaries} definitions and basic notions on polyhedral fans and graphical nested complexes. In particular, we state in Proposition~\ref{prop:characterizationFan} a crucial sufficient condition for a set of vectors indexed by a set~$X$ to support a simplicial fan realization of a simplicial complex on~$X$. We also briefly survey the classical constructions of graph associahedra of~\cite{CarrDevadoss, Postnikov, Zelevinsky}.
In Section~\ref{sec:compatibilityFan}, we define the compatibility degree between two tubes of a graph, review its combinatorial properties, and state our geometric results on compatibility and dual compatibility~fans.
We study various examples in Section~\ref{sec:specificGraphs}. After an exhaustive description of the compatibility fans of all graphs with at most~$4$ vertices, we study four families of graphs: paths, cycles, complete graphs, and stars. The first two families connect our construction to S.~Fomin and A.~Zelevinsky's $\b{d}$-vector fans for type~$A$, $B$, and $C$ cluster complexes.
Section~\ref{sec:furtherTopics} discusses various further topics. We first study the behavior of the compatibility fans with respect to products and links. We then describe all nested complex isomorphisms in order to show that most compatibility fans are not linearly isomorphic. We also discuss the question of the realization of our compatibility fans as normal fans of convex polytopes. We extend our construction to design nested complexes~\cite{DevadossHeathVipismakul}. Finally, we discuss the connection of this paper to Laurent Phenomenon algebras~\cite{LamPylyavskyy-LaurentPhenomenonAlgebras, LamPylyavskyy-LinearLaurentPhenomenonAlgebras}.
Finally, we have chosen to gather all proofs of our results in Section~\ref{sec:proofs} with the hope that the properties and examples of compatibility fans treated in Sections~\ref{sec:specificGraphs} and~\ref{sec:furtherTopics} help the reader's~intuition.
\section{Preliminaries}
\label{sec:preliminaries}
In this section, we briefly review classical material to recall basic definitions and fix notations. The reader familiar with polyhedral geometry and graph associahedra can skip these preliminaries and proceed directly to Section~\ref{sec:compatibilityFan}.
\subsection{Polyhedral geometry and fans}
We first recall classical definitions from polyhedral geometry. More details can be found in the textbooks~\cite[Lecture~1]{Ziegler} and~\cite[Section~2.1.1]{DeLoeraRambauSantos}.
A \defn{closed polyhedral cone} is a subset of~$\R^n$ defined equivalently as the positive span of finitely many vectors or as the intersection of finitely many closed linear halfspaces. The \defn{dimension} of a cone is the dimension of its linear span. The \defn{faces} of a cone~$C$ are the intersections of~$C$ with the supporting hyperplanes of~$C$. Faces of polyhedral cones are polyhedral cones. The faces of dimension~$1$ (resp.~codimension~$1$) are called \defn{rays} (resp.~\defn{facets}). We will only consider \defn{pointed} cones, which contain no entire line of~$\R^n$. Therefore, $C$ is the positive span of its rays and the intersection of the halfspaces defined by its facets and containing it. We say that~$C$ is \defn{simplicial} if its rays form a linear basis of its linear span.
A \defn{polyhedral fan} is a collection~$\cF$ of polyhedral cones of~$\R^n$ closed under faces and which intersect properly, \ie
\begin{enumerate}
\item if~$C \in \cF$ and~$F$ is a face of~$C$, then~$F \in \cF$;
\item the intersection of any two cones of~$\cF$ is a face of both.
\end{enumerate}
A polyhedral fan is \defn{simplicial} if all its cones are, and \defn{complete} if the union of its cones covers the entire space~$\R^n$. We will use the following characterization of complete simplicial fans, whose formal proof can be found \eg in~\cite[Corollary 4.5.20]{DeLoeraRambauSantos}.
\begin{proposition}
\label{prop:characterizationFan}
For a simplicial sphere~$\Delta$ with vertex set~$X$ and a set of vectors~$\b{V} \eqdef (\b{v}_x)_{x \in X}$ of~$\R^n$, the collection of cones~$\bigset{\R_{\ge 0}\b{V}_\triangle}{\triangle \in \Delta}$, where~$\R_{\ge 0}\b{V}_\triangle$ denotes the positive span of~$\b{V}_\triangle \eqdef \set{\b{v}_x}{x \in \triangle}$, forms a complete simplicial fan if and~only~if
\begin{enumerate}
\item there exists a facet~$\triangle$ of~$\Delta$ such that~$\b{V}_\triangle$ is a basis of~$\R^n$ and such that the open cones~$\R_{> 0}\b{V}_\triangle$ and~$\R_{> 0}\b{V}_{\triangle'}$ are disjoint for any facet~$\triangle'$ of~$\Delta$ distinct from~$\triangle$;
\item for two adjacent facets~$\triangle, \triangle'$ of~$\Delta$ with~$\triangle \ssm \{x\} = \triangle' \ssm \{x'\}$, there is a linear dependence
\[
\alpha \, \b{v}_x + \alpha' \, \b{v}_{x'} + \sum_{y \in \triangle \cap \triangle'} \beta_y \, \b{v}_y = 0
\]
on~$\b{V}_{\triangle \cup \triangle'}$ in which the coefficients~$\alpha$ and~$\alpha'$ have the same sign (different from~$0$). Note that when these conditions hold, this linear dependence is unique up to rescaling.
\end{enumerate}
\end{proposition}
A \defn{polytope} is a subset~$P$ in~$\R^n$ defined equivalently as the convex hull of finitely many points in~$\R^n$ or as a bounded intersection of finitely many closed half-spaces of~$\R^n$. The faces of~$P$ are the intersections of~$P$ with its supporting hyperplanes. The (outer) \defn{normal cone} of a face~$F$ of~$P$ is the cone generated by the outer normal vectors of the facets (codimension~$1$ faces) of~$P$ containing~$F$. Finally, the (outer) \defn{normal fan} of~$P$ is the collection of the (outer) normal cones of all its faces.
\subsection{Graphical nested complexes}
\label{subsec:nestedComplex}
We review the definitions and basic properties of graphical nested complexes. We refer to~\cite{CarrDevadoss, Devadoss} for the original construction of graph associahedra. This construction extends to nested complexes on arbitrary building sets, see~\cite{Postnikov, FeichtnerSturmfels, Zelevinsky}. Although we remain in the graphical situation, our presentation borrows results from these~papers.
Fix a graph~$\graphG$ with vertex set~$\ground$. Let~$\connectedComponents(\graphG)$ denote the set of connected components of~$\graphG$ and define~$n \eqdef |\ground|-|\connectedComponents(\graphG)|$. A \defn{tube} of~$\graphG$ is a non-empty subset~$\tube$ of vertices of~$\graphG$ inducing a connected subgraph~$\graphG{}[\tube]$ of~$\graphG$. The inclusion maximal tubes of~$\graphG$ are its connected components~$\connectedComponents(\graphG)$; all other tubes are called \defn{proper}. Two tubes~$\tube, \tube'$ of~$\graphG$ are \defn{compatible} if they are either nested (\ie $\tube \subseteq \tube'$ or~$\tube' \subseteq \tube$), or disjoint and non-adjacent (\ie $\tube \cup \tube'$ is not a tube of~$\graphG$). A \defn{tubing} on~$\graphG$ is a set~$\tubing$ of pairwise compatible proper tubes of~$\graphG$. The collection of all tubings on~$\graphG$ is a simplicial complex, called \defn{nested complex} of~$\graphG$ and denoted by~$\nestedComplex(\graphG)$.
\begin{example}
\label{exm:exmTubes}
To illustrate the content of the paper, we will follow a toy example, presented in \fref{fig:exmTubes}. We have represented a graph~$\graphG\ex$ on the left with a tube~$\tube^\circ\ex \eqdef \{a,b,d,f,g,h,i,k,l\}$, and a maximal tubing~$\tubing^\circ\ex$ on~$\graphG\ex$ on the right.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{exmTubes}}
\caption{A tube~$\tube^\circ\ex$ of~$\graphG\ex$ (left) and a maximal tubing~$\tubing^\circ\ex$ on~$\graphG\ex$ (right).}
\label{fig:exmTubes}
\end{figure}
\end{example}
For a tubing~$\tubing$ on~$\graphG$ and a tube~$\tube$ of~$\tubing \cup \connectedComponents(\graphG)$, we define~$\lambda(\tube, \tubing) \eqdef \tube \ssm \bigcup_{\tube' \in \tubing, \tube' \subsetneq \tube} \tube'$. The sets~$\lambda(\tube, \tubing)$ for~$\tube \in \tubing \cup \connectedComponents(\graphG)$ form a partition of the vertex set of~$\graphG$. When~$\tubing$ is a maximal tubing, each set~$\lambda(\tube, \tubing)$ contains a unique vertex of~$\graphG$ that we call the \defn{root} of~$\tube$ in~$\tubing$.
The nested complex~$\nestedComplex(\graphG)$ is an $(n-1)$-dimensional simplicial sphere. The dual graph of this complex is the graph whose vertices are maximal tubings on~$\graphG$ and whose edges are flips between them. A \defn{flip} is a pair of distinct maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ for some tubes~$\tube \in \tubing$ and~$\tube' \in \tubing'$. Since~$\nestedComplex(\graphG)$ is a sphere, any tube of any maximal tubing can be flipped, and the resulting tube is described in the following proposition, whose proof is left to the reader. Remember that we denote by~$\graphG{}[U]$ the subgraph of~$\graphG$ induced by a subset~$U \subseteq \ground$.
\begin{proposition}
\label{prop:flip}
Let~$\tube$ be a tube in a maximal tubing~$\tubing$ on~$\graphG$, and let~$\overline{\tube}$ be the inclusion minimal tube of~$\tubing \cup \connectedComponents(\graphG)$ which strictly contains~$\tube$. Then the unique tube~$\tube'$ such that~$\tubing' = \tubing \symdif \{\tube, \tube'\}$ is again a maximal tubing on~$\graphG$ is the connected component of~$\graphG{}[\overline{\tube} \ssm \lambda(\tube, \tubing)]$ containing~$\lambda(\overline{\tube}, \tubing)$.
\end{proposition}
We say that two distinct tubes~$\tube, \tube'$ of~$\graphG$ are \defn{exchangeable} if there exists two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. Note that several such pairs~$\{\tubing, \tubing'\}$ are possible, but they all contain certain tubes. We call \defn{forced tubes} of the exchangeable pair~$\{\tube, \tube'\}$ any tube which belongs to any adjacent maximal tubings~$\tubing, \tubing'$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. These tubes are easy to describe: they are precisely the tube~$\overline{\tube} \eqdef \tube \cup \tube'$ and the connected components of~$\overline{\tube} \ssm (\lambda(\tube,\tubing) \cup \lambda(\tube',\tubing'))$.
\begin{example}
\label{exm:exmFlip}
\fref{fig:exmFlip} illustrates the flip between two maximal tubings~$\tubing\ex$ and~$\tubing'\ex$ on~$\graphG\ex$. The flipped tubes~$\tube\ex = \{a,b,c,d,f,g,h,k,l,m\}$ (with root~$g$) and~$\tube'\ex = \{c,d,e,h,i,m\}$ (with root~$i$) are dashed green, while the forced tubes of the exchangeable pair~$\{\tube\ex, \tube'\ex\}$ are red.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{exmFlip}}
\caption{The flip between two maximal tubings~$\tubing\ex$ and~$\tubing'\ex$ on~$\graphG\ex$.}
\label{fig:exmFlip}
\end{figure}
\end{example}
\subsection{Graphical nested fans and graph associahedra}
\label{subsec:graphAssociahedra}
Although we will not use them in the remaining of this paper, we briefly survey some constructions of the nested fan and graph associahedron to illustrate the previous definitions. The nested fan is constructed implicitly in~\cite{CarrDevadoss}, and explicitly in~\cite{Postnikov, FeichtnerSturmfels, Zelevinsky} in the more general situation of nested complexes on arbitrary building sets.
Let~$(\b{e}_v)_{v \in \ground}$ be the canonical basis of~$\R^\ground$, let~${\HH \eqdef \bigset{\b{x} \in \R^\ground}{\sum_{w \in W} x_w = 0 \text{ for all } W \in \connectedComponents(\graphG)}}$ and~$\pi : \R^\ground \to \HH$ denote the orthogonal projection on~$\HH$. Let~$\b{g}(\tube) \eqdef \pi \big( \sum_{v \in \tube} \b{e}_v \big)$ denote the projection of the characteristic vector of a tube~$\tube$ of~$\graphG$, and define~$\b{g}(\tubing) \eqdef \set{\b{g}(\tube)}{\tube \in \tubing}$ for a tubing~$\tubing$ on~$\graphG$. These vectors support a complete simplicial fan realization of the nested complex:
\begin{theorem}[\cite{CarrDevadoss, Postnikov, FeichtnerSturmfels, Zelevinsky}]
\label{theo:gFan}
For any graph~$\graphG$, the collection of cones
\[\cG(\graphG) \eqdef \set{\R_{\ge 0} \, \b{g}(\tubing)}{\tubing \text{ tubing on } \graphG}\]
is a complete simplicial fan of~$\HH$, called \defn{nested fan} of~$\graphG$, which realizes the nested complex~$\nestedComplex(\graphG)$.
\end{theorem}
\begin{remark}
The cones of this fan can be encoded as follows. Fix a tubing~$\tubing$ on~$\graphG$. The \defn{spine} of~$\tubing$ is the forest~$\spine$ given by the Hasse diagram of the inclusion poset on~$\tubing \cup \connectedComponents(\graphG)$ where the vertex corresponding to a tube~$\tube$ is labeled by~$\lambda(\tube, \tubing)$. Then the cone~$\R_{\ge 0} \, \b{g}(\tubing)$ is the \defn{braid cone} of~$\spine$ and is polar to the \defn{incidence cone} of~$\spine$:
\[
\R_{\ge 0} \, \b{g}(\tubing) = \set{\b{x} \in \HH}{x_v \le x_w \text{ for all } v \to w \in \spine} = \big(\R_{\ge 0} \set{\b{e}_v-\b{e}_w}{v \to w \in \spine} \big)\polar
\]
where $v \to w \in \spine$ means that there is a directed path in~$\spine$ from the node containing~$v$ to the node of~$\spine$ containing~$w$. Spines are called $B$-trees in~\cite{Postnikov}.
\end{remark}
It is proved in~\cite{CarrDevadoss, Devadoss, Postnikov, FeichtnerSturmfels, Zelevinsky} that the nested fan comes from a polytope.
\begin{theorem}[\cite{CarrDevadoss, Devadoss, Postnikov, FeichtnerSturmfels, Zelevinsky}]
\label{theo:graphAssociahedron}
For any graph~$\graphG$, the nested fan~$\cG(\graphG)$ is the normal fan of the graph associahedron~$\Asso(\graphG)$.
\end{theorem}
In fact, these different papers define different constructions and realizations for the graph associahedron~$\Asso(\graphG)$. Originally, M.~Carr and S.~Devadoss constructed~$\Asso(\graphG)$ by iterative truncations of faces of the standard simplex. S.~Devadoss then gave explicit integer coordinates for the facets in~\cite{Devadoss}. In a different context, A.~Postnikov~\cite{Postnikov} and independently E.~M.~Feichtner and B.~Sturmfels~\cite{FeichtnerSturmfels} constructed graph associahedra (more generally nestohedra) by Minkowski sums of faces of the standard simplex. Finally, A.~Zelevinsky~\cite{Zelevinsky} discussed which polytopes can realize the nested fan using a characterization of all possible facet inequality descriptions.
\section{Compatibility degrees, vectors, and fans}
\label{sec:compatibilityFan}
In this section, we define our compatibility degree on tubes and state our main results. To go straight to results and examples, the proofs are delayed to Section~\ref{sec:proofs}. We refer to Section~\ref{sec:specificGraphs} for examples of compatibility fans of relevant graphs in connection to the geometry of type~$A$, $B$ and~$C$ cluster complexes, and to Section~\ref{sec:furtherTopics} for further properties of compatibility fans.
\subsection{Compatibility degree}
Motivated by the compatibility degrees in finite type cluster algebras, we introduce an analogous notion on tubes of graphical nested complexes.
\begin{definition}
\label{def:compatibilityDegree}
For two tubes~$\tube, \tube'$ of~$\graphG$, the \defn{compatibility degree} of~$\tube$ with~$\tube'$ is
\[
\compatibilityDegree{\tube}{\tube'} =
\begin{cases}
-1 & \text{if } \tube = \tube', \\
|\{\text{neighbors of $\tube$ in } \tube' \ssm \tube\}| & \text{if } \tube \not\subseteq \tube', \\
0 & \text{otherwise.}
\end{cases}
\]
\end{definition}
\begin{example}
\label{exm:exmCompatibilityDegree}
On the graph~$\graphG\ex$ of Examples~\ref{exm:exmTubes} and~\ref{exm:exmFlip}, the compatibility degrees of the green tubes~$\tube\ex$, $\tube'\ex$ and of the red tube~$\overline{\tube}\ex = \tube\ex \cup \tube'\ex$ of \fref{fig:exmFlip} with the tube~$\tube^\circ\ex$ of \fref{fig:exmTubes}\,(left) are given by
\[
\begin{array}{l@{\qquad}l@{\qquad}l}
\compatibilityDegree{\tube\ex}{\tube^\circ\ex} = |\{i\}| = 1, & \compatibilityDegree{\tube'\ex}{\tube^\circ\ex} = |\{g\}| = 1, & \compatibilityDegree{\overline{\tube}\ex}{\tube^\circ\ex} = 0, \\
\compatibilityDegree{\tube^\circ\ex}{\tube\ex} = |\{c,m\}| = 2, & \compatibilityDegree{\tube^\circ\ex}{\tube'\ex} = |\{c,e,m\}| = 3, & \compatibilityDegree{\tube^\circ\ex}{\overline{\tube}\ex} = 0.
\end{array}
\]
Note that the last~$0$ is forced by the last line of the definition since~$\tube^\circ\ex \subsetneq \overline{\tube}\ex$.
\end{example}
We will see in Sections~\ref{subsec:paths} and~\ref{subsec:cycles} that our compatibility degree on tubes of paths (resp.~cycles) corresponds to the compatibility degree on cluster variables in type~$A$ (resp.~$B/C$) cluster algebras defined in~\cite{FominZelevinsky-YSystems}. The compatibility degree in cluster algebras encodes compatibility and exchangeability between cluster variables. The analogous result for the graphical compatibility degree is given by the following proposition, proved in Section~\ref{subsec:proofCompatibilityDegree}.
\begin{proposition}
\label{prop:compatibilityDegree}
For any two tubes~$\tube, \tube'$ of~$\graphG$,
\begin{itemize}
\item $\compatibilityDegree{\tube}{\tube'} < 0 \iff \compatibilityDegree{\tube'}{\tube} < 0 \iff \tube = \tube'$, \label{item:identical}
\item $\compatibilityDegree{\tube}{\tube'} = 0 \iff \compatibilityDegree{\tube'}{\tube} = 0 \iff \tube$ and~$\tube'$ are compatible, and \label{item:compatible}
\item $\compatibilityDegree{\tube}{\tube'} = 1 = \compatibilityDegree{\tube'}{\tube} \iff \tube$ and~$\tube'$ are exchangeable. \label{item:exchangeable}
\end{itemize}
\end{proposition}
\begin{remark}
It can happen that~$\compatibilityDegree{\tube}{\tube'} = 1$ while~$\compatibilityDegree{\tube'}{\tube} \ne 1$, in which case~$\tube$ and~$\tube'$ are not exchangeable. This situation appears as soon as~$\graphG$ contains a cycle or a trivalent vertex. See \eg Example~\ref{exm:exmCompatibilityDegree}.
\end{remark}
Proposition~\ref{prop:compatibilityDegree} should be understood informally as follows: the compatibility degree between two tubes measures how much they are incompatible. It is natural to use this measure to construct fan realizations of the nested complex: intuitively, pairs of tubes with low compatibility degrees should correspond to rays closed to each other. We make this idea precise in the next section.
\subsection{Compatibility fans}
Similar to the $\b{d}$-vector fan defined by S.~Fomin and A.~Zelevinsky~\cite{FominZelevinsky-ClusterAlgebrasII}, we consider the compatibility vectors with respect to an arbitrary initial maximal tubing~$\tubing^\circ$. Remember that any maximal tubing on~$\graphG$ has precisely~$n \eqdef |\ground| - |\connectedComponents(\graphG)|$ tubes.
\begin{definition}
Let~$\tubing^\circ \eqdef \{\tube_1^\circ, \dots, \tube_n^\circ\}$ be an arbitrary initial maximal tubing on~$\graphG$. The \defn{compatibility vector} of a tube~$\tube$ of~$\graphG$ with respect to~$\tubing^\circ$ is the integer vector $\compatibilityVector{\tubing^\circ}{\tube} \eqdef [\compatibilityDegree{\tube_1^\circ}{\tube}, \dots, \compatibilityDegree{\tube_n^\circ}{\tube}]$. The \defn{compatibility matrix} of a tubing~$\tubing \eqdef \{\tube_1, \dots, \tube_m\}$ on~$\graphG$ with respect to~$\tubing^\circ$ is the matrix $\compatibilityVector{\tubing^\circ}{\tubing} \eqdef [\compatibilityDegree{\tube_i^\circ}{\tube_j}]_{i \in [n], j \in [m]}$.
\end{definition}
Remember that we denote by~$\R_{\ge 0} \, \b{M}$ the polyhedral cone generated by the column vectors of a matrix~$\b{M}$. Note that the compatibility vectors of the initial tubes are given by the negative of the basis vectors, while all other compatibility vectors lie in the positive orthant: $\compatibilityVector{\tubing^\circ}{\tubing^\circ} = -\b{I}_n$ and~$\compatibilityVector{\tubing^\circ}{\tube} \in \R_{\ge 0} \, \b{I}_n$ for~$\tube \notin \tubing^\circ$. Our main result asserts that these compatibility vectors support a complete simplicial fan realization of the graphical nested complex.
\begin{theorem}
\label{theo:compatibilityFan}
For any graph~$\graphG$ and any maximal tubing~$\tubing^\circ$ on~$\graphG$, the collection of cones
\[
\compatibilityFan{\graphG}{\tubing^\circ} \eqdef \set{\R_{\ge 0} \, \compatibilityVector{\tubing^\circ}{\tubing}}{\tubing \text{ tubing on } \graphG}
\]
is a complete simplicial fan which realizes the nested complex~$\nestedComplex(\graphG)$. We call it the \defn{compatibility fan} of~$\graphG$ with respect to~$\tubing^\circ$.
\end{theorem}
We prove this statement in Section~\ref{subsec:proofCompatibilityFan}. The proof relies on the characterization of complete simplicial fans presented in Proposition~\ref{prop:characterizationFan}. Unfortunately, we are not able to compute the linear dependence between the compatibility vectors involved in an arbitrary flip. To illustrate the difficulty, we show in the following example that these linear dependences may be complicated. In particular, they do not always involve only the forced tubes of the flip.
\begin{example}
\label{exm:exmLinearDependence}
Consider the initial maximal tubing~$\tubing^\circ\ex$ of the graph~$\graphG\ex$ of \fref{fig:exmTubes}\,(right) and the flip~$\tubing\ex \ssm \{\tube\ex\} = \tubing'\ex \ssm \{\tube'\ex\}$ illustrated in \fref{fig:exmFlip}.
The linear dependence between the compatibility vectors of the tubes of~$\tubing\ex \cup \tubing'\ex$ with respect to~$\tubing^\circ\ex$ is
\begin{gather*}
2 \, \compatibilityVector{\tubing^\circ\ex}{\tube\ex} + \compatibilityVector{\tubing^\circ\ex}{\tube'\ex} - \compatibilityVector{\tubing^\circ\ex}{\{d\}} - \compatibilityVector{\tubing^\circ\ex}{\{e\}} \qquad \qquad \\ \qquad \qquad - 3 \, \compatibilityVector{\tubing^\circ\ex}{\{m\}} + 4 \, \compatibilityVector{\tubing^\circ\ex}{\{k,l\}} - 3 \, \compatibilityVector{\tubing^\circ\ex}{\{c,d,h\}} = 0.
\end{gather*}
Observe that the tube~$\{d\}$ is involved in this linear dependence although it is not a forced tube of the exchangeable pair~$\{\tube\ex,\tube'\ex\}$.
\end{example}
Theorem~\ref{theo:compatibilityFan} has the following consequences, whose direct proof would require some work.
\begin{corollary}
\label{coro:fullRank}
For any initial tubing~$\tubing^\circ$ on~$\graphG$,
\begin{itemize}
\item the compatibility vector map~$\tube \longmapsto \compatibilityVector{\tubing^\circ}{\tube}$ is injective: $\compatibilityVector{\tubing^\circ}{\tube} = \compatibilityVector{\tubing^\circ}{\tube'} \; \Longrightarrow \; \tube = \tube'$.
\item the compatibility matrix~$\compatibilityVector{\tubing^\circ}{\tubing}$ of any maximal tubing~$\tubing$ on~$\graphG$ has full rank.
\end{itemize}
\end{corollary}
\subsection{Dual compatibility fan}
It is also interesting to consider the following dual notion of compatibility vectors, where the roles of~$\tube$ and~$\tube_1^\circ, \dots, \tube_n^\circ$ are reversed. The results are similar, and the motivation for this dual definition will become clear in Section~\ref{subsec:cycles}.
\begin{definition}
Let~$\tubing^\circ \eqdef \{\tube_1^\circ, \dots, \tube_n^\circ\}$ be an arbitrary initial maximal tubing on~$\graphG$. The \defn{dual compatibility vector} of a tube~$\tube$ of~$\graphG$ with respect to~$\tubing^\circ$ is the vector $\dualCompatibilityVector{\tube}{\tubing^\circ} \eqdef [\compatibilityDegree{\tube}{\tube_1^\circ}, \dots, \compatibilityDegree{\tube}{\tube_n^\circ}]$. The \defn{dual compatibility matrix} of a tubing~$\tubing \eqdef \{\tube_1, \dots, \tube_m\}$ on~$\graphG$ with respect to~$\tubing^\circ$ is the matrix~$\dualCompatibilityVector{\tubing}{\tubing^\circ} \eqdef [\compatibilityDegree{\tube_j}{\tube_i^\circ}]_{i \in [n], j \in [m]}$.
\end{definition}
The following statement is the analogue of Theorem~\ref{theo:compatibilityFan}.
\begin{theorem}
\label{theo:dualCompatibilityFan}
For any graph~$\graphG$ and any maximal tubing~$\tubing^\circ$ on~$\graphG$, the collection of cones
\[
\dualCompatibilityFan{\graphG}{\tubing^\circ} \eqdef \set{\R_{\ge 0} \, \dualCompatibilityVector{\tubing}{\tubing^\circ}}{\tubing \text{ tubing on } \graphG}\]
is a complete simplicial fan which realizes the nested complex~$\nestedComplex(\graphG)$. We call it the \defn{dual compatibility fan} of~$\graphG$ with respect to~$\tubing^\circ$.
\end{theorem}
The proof of this statement appears in Section~\ref{subsec:proofDualCompatibilityFan}. It is a direct application of Theorem~\ref{theo:compatibilityFan} using duality between compatibility and dual compatibility matrices.
\begin{example}
\label{exm:exmLinearDependenceDual}
Consider the initial maximal tubing~$\tubing^\circ\ex$ on the graph~$\graphG\ex$ of \fref{fig:exmTubes}\,(right) and the flip~$\tubing\ex \ssm \{\tube\ex\} = \tubing'\ex \ssm \{\tube'\ex\}$ illustrated in \fref{fig:exmFlip}. The linear dependence between the dual compatibility vectors of the tubes of~$\tubing\ex \cup \tubing'\ex$ with respect to~$\tubing^\circ\ex$ is
\[
2 \, \dualCompatibilityVector{\tube\ex}{\tubing^\circ\ex} + \dualCompatibilityVector{\tube'\ex}{\tubing^\circ\ex} - \dualCompatibilityVector{\{e\}}{\tubing^\circ\ex} - \dualCompatibilityVector{\{c,d,h\}}{\tubing^\circ\ex} = 0.
\]
\end{example}
\section{Examples for specific graphs}
\label{sec:specificGraphs}
In this section, we provide examples of compatibility fans for particular families of graphs. We start with graphs with few vertices to illustrate the variety of compatibility fans. We then describe compatibility fans for paths, cycles, complete graphs and stars using alternative combinatorial models (triangulations, lattice paths, ...). For paths and cycles, we give an explicit connection to the compatibility degree in cluster algebras of types~$A$, $B$, and~$C$. The examples of this section shall help the intuition for further properties studied in Section~\ref{sec:furtherTopics} and for the proofs gathered in Section~\ref{sec:proofs}.
\subsection{Graphs with few vertices}
\label{subsec:fewVertices}
\enlargethispage{.3cm}
In view of Proposition~\ref{prop:product} below, we restrict to connected graphs. The only connected graphs with $3$ vertices are the $3$-path and the triangle, whose compatibility fans are represented in \fref{fig:3vertices}. The other possible choices for the initial tubing in these pictures would produce the same fans: it is clear for the triangle as all maximal tubings are obtained from one another by graph isomorphisms; for the path, it is an illustration of the non-trivial isomorphisms between compatibility fans studied in Section~\ref{subsec:many}.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=1.05]{3vertices}}
\caption{Compatibility fans of the $3$-path (left) and of the triangle (right).}
\label{fig:3vertices}
\end{figure}
The first interesting compatibility fans appear in dimension~$3$ for connected graphs on $4$ vertices. All possibilities up to linear transformations are represented in \fref{fig:4vertices}. Instead of representing cones in the $3$-dimensional space, we intersect the compatibility vectors with the unit sphere, make a stereographic projection of the resulting points on the sphere (the pole of the projection is the point of the sphere in direction~$-\b{e}_1-\b{e}_2-\b{e}_3$), and draw the cones on the resulting planar points. Under this projection, the three external vertices correspond to the tubes of the initial tubing, and the external face corresponds to the initial tubing. For the sake of readability, we do not label the remaining vertices of the projection. Their labels can be reconstructed from the initial tubes by flips. For example, the tubes corresponding to the vertices of the top pictures of \fref{fig:4vertices} are given in \fref{fig:allLabels}.
The pictures become more complicated in dimension~$4$. To illustrate them, we have represented in \fref{fig:5vertices} the stereographic projection of the compatibility fan for an arbitrary initial maximal tubing on the path, cycle, complete graph, and star on $5$ vertices.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[width=.37\textwidth]{path1AllLabels} \quad \includegraphics[width=.37\textwidth]{path2AllLabels} \quad \includegraphics[width=.37\textwidth]{path3AllLabels}}
\vspace{-.2cm}
\caption{All tubes in the top pictures of \fref{fig:4vertices}.}
\label{fig:allLabels}
\end{figure}
\begin{figure}[p]
\capstart
\renewcommand{\arraystretch}{1.3}
\setlength{\extrarowheight}{3.4cm}
\centerline{
\begin{tabular}{@{}|c|ccc|}
\hline
\rotatebox{90}{\textbf{Path} (associahedron)} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/path1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/path2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/path3} \\
\hline
\end{tabular}
}
\centerline{
\begin{tabular}{@{}|c|ccc|}
\hline
\rotatebox{90}{\textbf{Cycle} (cyclohedron)} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle3} \\
\hline
\end{tabular}
}
\setlength{\extrarowheight}{0cm}
\centerline{
\begin{tabular}[t]{@{}|@{\hspace{.13cm}}c@{\hspace{.13cm}}|}
\hline
\textbf{Complete graph} \\[-.15cm] (permutahedron) \\
\hline
\\[-.4cm]
\includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/complete} \\
\hline
\end{tabular}
\!\!\!\!\!
\setlength{\extrarowheight}{3.4cm}
\begin{tabular}[t]{@{}|c|cc|}
\hline
& \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/tripod1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/tripod2} \\
\rotatebox{90}{\textbf{Star} (stellohedron)} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/tripod3} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/tripod4} \\
\hline
\end{tabular}
}
\caption{All possible compatibility fans up to linear isomorphism, for all connected graphs on $4$ vertices (see also the end of the picture on page~\pageref{fig:4verticesEnd} for the two remaining graphs). Instead of representing the cones in the $3$-dimensional space, we intersect the compatibility vectors with the unit sphere, make a stereographic projection of the resulting points on the sphere (the pole of the projection is the point of the sphere in direction~$-\b{e}_1-\b{e}_2-\b{e}_3$), and draw the cones on the resulting planar points.}
\label{fig:4vertices}
\end{figure}
\begin{figure}[p]
\renewcommand{\arraystretch}{1.3}
\setlength{\extrarowheight}{3.4cm}
\centerline{
\begin{tabular}{@{}|c|ccc|}
\hline
& \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge3} \\
\rotatebox{90}{\textbf{Complete graph minus one edge} \hspace{-2cm}} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge4} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge5} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusOneEdge6} \\
\hline
\end{tabular}
}
\centerline{
\begin{tabular}{@{}|c|ccc|}
\hline
& \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges3} \\
& \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges4} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges5} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges6} \\
\rotatebox{90}{\textbf{Complete graph minus two incident edges} \hspace{-4cm}} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges7} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges8} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/completeMinusTwoEdges9}
\\
\hline
\end{tabular}
}
\label{fig:4verticesEnd}
\end{figure}
\begin{figure}
\capstart
\centerline{\includegraphics[width=.55\textwidth]{4dimensionalCompatibilityFansLabeled/path} \quad \includegraphics[width=.55\textwidth]{4dimensionalCompatibilityFansLabeled/cycle}}
\vspace*{.4cm}
\centerline{\includegraphics[width=.55\textwidth]{4dimensionalCompatibilityFansLabeled/complete} \quad \includegraphics[width=.55\textwidth]{4dimensionalCompatibilityFansLabeled/star}}
\caption{Stereographic projection of the compatibility fan for particular initial maximal tubings on the path, cycle, complete graph, and star on $5$ vertices.}
\label{fig:5vertices}
\end{figure}
\subsection{Paths}
\label{subsec:paths}
We now consider the nested complex~$\nestedComplex(\pathG_{n+1})$ and the compatibility fan~$\compatibilityFan{\pathG_{n+1}}{\tubing^\circ}$ for the path~$\pathG_{n+1}$ on~$n+1$ vertices. As already mentioned in the introduction, the nested complex~$\nestedComplex(\pathG_{n+1})$ is isomorphic to the $n$-dimensional \defn{simplicial associahedron}, \ie the simplicial complex of sets of pairwise non-crossing diagonals of an $(n+3)$-gon. It is convenient to present the correspondence as follows. Consider an $(n+3)$-gon~$Q_{n+3}$ with vertices labeled from left to right by~$0, 1, \dots, n+2$ and such that all vertices~$1, \dots, n+1$ are located strictly below the boundary edge~$[0,n+2]$. We can therefore identify the path~$\pathG_{n+1}$ with the path~$1, \dots, n+1$ on the boundary of~$Q_{n+3}$. We then associate to a diagonal~$\delta$ of~$Q_{n+3}$ the tube~$\tube_\delta$ of~$\pathG_{n+1}$ whose vertices are located strictly below~$\delta$, see Figures~\ref{fig:path} and~\ref{fig:associahedron}. Finally, we associate to a set~$\Delta$ of pairwise non-crossing internal diagonals of~$Q_{n+3}$ the set of tubes~$\tubing_\Delta \eqdef \set{\tube_\delta}{\delta \in \Delta}$, see \fref{fig:associahedron}. The reader can check that the map~$\Delta \mapsto \tubing_\Delta$ defines an isomorphism between the simplicial associahedron and the nested complex~$\nestedComplex(\pathG_{n+1})$: two diagonals~$\delta, \delta'$ of~$Q_{n+3}$ are non-crossing if and only if the corresponding tubes~$\tube_\delta, \tube_{\delta'}$ of~$\pathG_{n+1}$ are compatible.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.6]{path}}
\caption{Isomorphism between the simplicial associahedron and the nested complex of a path: diagonals are sent to tubes (left), preserving the compatibility (middle) and incompatibility (right). See also \fref{fig:associahedron}.}
\label{fig:path}
\end{figure}
It follows by classical results on the associahedron that the path~$\pathG_{n+1}$ has:
\begin{itemize}
\item $\displaystyle{\frac{n(n+3)}{2}}$ proper tubes~\href{https://oeis.org/A000096}{\cite[A000096]{OEIS}} (internal diagonals of the $(n+3)$-gon),
\item $\displaystyle{\frac{1}{n+2} \binom{2n+2}{n+1}}$ maximal tubings~\href{https://oeis.org/A000108}{\cite[A000108]{OEIS}} (triangulations of the $(n+3)$-gon),
\item $\displaystyle{\frac{1}{k+1}\binom{n}{k}\binom{n+k+2}{k}}$ tubings with~$k$ tubes~\href{https://oeis.org/A033282}{\cite[A033282]{OEIS}} (dissections of the $(n+3)$-gon into~$k$ parts).
\end{itemize}
The following statement, whose proof is left to the reader, describes the behavior of the map~${\delta \mapsto \tube_\delta}$ with respect to compatibility degrees.
\begin{proposition}
\label{prop:typeA}
For any two diagonals~$\delta, \delta'$ of~$Q_{n+3}$, the compatibility degree of the corresponding tubes~$\tube_\delta$ and~$\tube_{\delta'}$ of~$\pathG_{n+1}$ is given by
\[
\compatibilityDegree{\tube_\delta}{\tube_{\delta'}} =
\begin{cases}
-1 & \text{if $\delta = \delta'$,} \\
0 & \text{if $\delta \ne \delta'$ do not cross,} \\
1 & \text{if $\delta \ne \delta'$ cross.}
\end{cases}
\]
\end{proposition}
In other words, our compatibility degree between tubes of~$\pathG_{n+1}$ coincides with the compatibility degree between type~$A$ cluster variables defined by S.~Fomin and A.~Zelevinsky in~\cite{FominZelevinsky-YSystems}, and our graphical compatibility fan coincides with the type~$A$ compatibility fan defined for an acyclic initial cluster in~\cite{FominZelevinsky-ClusterAlgebrasII} and for any initial cluster in~\cite[Section~5]{CeballosSantosZiegler}. We thus obtain an alternative proof of F.~Santos' result~\cite[Section~5]{CeballosSantosZiegler}.
\begin{corollary}
For type~$A$ cluster algebras, the denominator vectors (or compatibility vectors) of all cluster variables with~respect to any initial cluster support a complete simplicial fan which realizes the cluster complex.
\end{corollary}
\begin{remark}[Dual compatibility fan]
\label{rem:primalDualPath}
The compatibility fan~$\compatibilityFan{\pathG_{n+1}}{\tubing^\circ}$ and the dual compatibility fan~$\dualCompatibilityFan{\pathG_{n+1}}{\tubing^\circ}$ coincide since the compatibility degree is symmetric for tubes~of~$\pathG_{n+1}$.
\end{remark}
\begin{remark}[Linear dependences]
In the case of the path~$\pathG_{n+1}$, the linear dependences are explicity described in~\cite{CeballosSantosZiegler}. They are derived from the case of the octagon by edge contraction in the interpretation in terms of triangulations. They can only involve the two flipped tubes and the forced tubes, and the coefficients are either~$1$ or~$2$ for the flipped tubes and~$-1$ or~$0$ for the forced tubes. See Section~\ref{subsec:proofPolytopality} for more details.
\end{remark}
\begin{remark}
\label{rem:naive}
\enlargethispage{1.2cm}
The compatibility degree for tubes of a path takes values in~$\{-1, 0, 1\}$. It is tempting to construct compatibility fans for graphical nestohedra using the naive compatibility degree defined by~$\compatibilityDegree{\tube}{\tube'} = -1$ if~$\tube = \tube'$, $\compatibilityDegree{\tube}{\tube'} = 0$ if $\tube \ne \tube'$ are compatible, and~$\compatibilityDegree{\tube}{\tube'} = 1$ if $\tube \ne \tube'$ are incompatible. This naive approach works for the paths but fails for any other connected graph since two distinct tubes would get the same compatibility vectors. See \fref{fig:ctrexmNaive} for examples on the triangle and on the tripod.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[width=\textwidth]{ctrexmNaive}}
\caption{Counter-examples to the naive definition of compatibility degrees: both on the triangle and on the tripod, all tubes of the initial maximal tubing on the left are incompatible with the two distinct tubes on the right.}
\label{fig:ctrexmNaive}
\end{figure}
\end{remark}
\subsection{Cycles}
\label{subsec:cycles}
We now consider the nested complex~$\nestedComplex(\cycleG_{n+1})$ and the compatibility fan~$\compatibilityFan{\cycleG_{n+1}}{\tubing^\circ}$ for the cycle~$\cycleG_{n+1}$ on~$n+1$ vertices. As already mentioned in the introduction, the nested complex~$\nestedComplex(\cycleG_{n+1})$ is isomorphic to the $n$-dimensional \defn{simplicial cyclohedron}, \ie the simplicial complex of sets of pairwise non-crossings pairs of centrally symmetric internal diagonals (including duplicated long diagonals) of a regular $(2n+2)$-gon~$R_{2n+2}$. The explicit correspondence works as follows. We label the vertices of~$R_{2n+2}$ cyclically with two copies of~$[n+1]$. We then associate
\begin{itemize}
\item to a duplicated long diagonal~$\delta$ with vertices labeled by~$i$ the tube~$\tube_\delta \eqdef [n+1] \ssm \{i\}$ of~$\cycleG_{n+1}$,
\item to a pair of centrally symmetric diagonals~$\{\delta, \bar\delta\}$ the tube~$\tube_\delta$ of~$\cycleG_{n+1}$ which consists of the labels of the vertices of~$R_{2n+2}$ separated from the center of~$R_{2n+2}$ by~$\delta$ and~$\bar\delta$.
\end{itemize}
Finally, we associate to a set~$\Delta$ of pairwise non-crossing pairs of centrally symmetric internal diagonals of~$R_{2n+2}$ the set of tubes~$\tubing_\Delta \eqdef \set{\tube_\delta}{\delta \in \Delta}$. See Figures~\ref{fig:cycle} and~\ref{fig:cyclohedron}. The reader can check that the map~$\Delta \mapsto \tubing_\Delta$ defines an isomorphism between the simplicial cyclohedron and the nested complex~$\nestedComplex(\cycleG_{n+1})$: two pairs of centrally symmetric diagonals (or duplicated long diagonals)~$\{\delta, \bar\delta\}$ and~$\{\delta', \bar\delta'\}$ of~$R_{2n+2}$ are non-crossing if and only if the corresponding tubes~$\tube_\delta$ and~$\tube_{\delta'}$ of~$\cycleG_{n+1}$ are compatible.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{cycle}}
\caption{Isomorphism between the simplicial cyclohedron and the nested complex of a cycle: centrally symmetric pairs of diagonals are sent to tubes, preserving the compatibility and incompatibility. See also \fref{fig:cyclohedron}.}
\label{fig:cycle}
\end{figure}
It follows by classical results on the cyclohedron that the cycle~$\cycleG_{n+1}$ has:
\begin{itemize}
\item $n(n+1)$ proper tubes~\href{https://oeis.org/A002378}{\cite[A002378]{OEIS}} (centrally symmetric pairs of diagonals),
\item $\displaystyle{\binom{2n}{n}}$ maximal tubings~\href{https://oeis.org/A000984}{\cite[A000984]{OEIS}} (centrally symmetric triangulations),
\item $\displaystyle{\binom{n}{k}\binom{n+k}{k}}$ tubings with~$k$ tubes~\href{https://oeis.org/A063007}{\cite[A063007]{OEIS}} (centrally symmetric dissections).
\end{itemize}
The following statement, whose proof is left to the reader, describes the behavior of the map~$\{\delta,\bar\delta\} \mapsto \tube_\delta$ with respect to compatibility degrees.
\begin{proposition}
\label{prop:typeBC}
For any two pairs of centrally symmetric diagonals (or duplicated long diagonals)~$\{\delta, \bar\delta\}$ and~$\{\delta', \bar\delta'\}$ of~$R_{2n+2}$, the compatibility degree~$\compatibilityDegree{\tube_\delta}{\tube_{\delta'}}$ of the corresponding tubes~$\tube_\delta$ and~$\tube_{\delta'}$ of~$\cycleG_{n+1}$ is the number of crossings between the two diagonals~$\delta$ and~$\bar\delta$ and the diagonal~$\delta'$.
\end{proposition}
In other words, our compatibility degree (resp.~dual compatibility degree) between tubes of~$\cycleG_{n+1}$ coincides with the compatibility degree between type~$C$ (resp.~type~$B$) cluster variables defined by S.~Fomin and A.~Zelevinsky in~\cite{FominZelevinsky-YSystems}. Moreover, our graphical compatibility fan (resp.~dual compatibility fan) coincides with the type~$C$ (resp.~type~$B$) compatibility fan defined for an acyclic initial cluster in~\cite{FominZelevinsky-ClusterAlgebrasII}. This extends for any arbitrary initial cluster to the following corollary.
\begin{corollary}
For type~$B$ and~$C$ cluster algebras, the denominator vectors (or compatibility vectors) of all cluster variables with~respect to any initial cluster support a complete simplicial fan which realizes the cluster complex.
\end{corollary}
\begin{remark}[Dual compatibility fan]
Since the compatibility degree is not symmetric for tubes of~$\cycleG_{n+1}$, the compatibility fan~$\compatibilityFan{\cycleG_{n+1}}{\tubing^\circ}$ and the dual compatibility fan~$\dualCompatibilityFan{\cycleG_{n+1}}{\tubing^\circ}$ do not coincide. Figures~\ref{fig:dualCompatibilityFanCycle3} and~\ref{fig:dualCompatibilityFanCycle4} show both fans for different initial tubings on the cycles~$\cycleG_3$ and~$\cycleG_4$.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1.05]{3verticesDual}}
\caption{Compatibility (left) and dual compatibility (right) fans for the triangle.}
\label{fig:dualCompatibilityFanCycle3}
\end{figure}
\begin{figure}[h]
\capstart
\renewcommand{\arraystretch}{1.3}
\setlength{\extrarowheight}{3.4cm}
\centerline{
\begin{tabular}{@{}|c|ccc|}
\hline
\rotatebox{90}{\textbf{Compatibility fan}} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycle3} \\
\hline
\hline
\rotatebox{90}{\textbf{Dual compatibility fan}} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycleOpp1} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycleOpp2} & \includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/cycleOpp3} \\
\hline
\end{tabular}
}
\caption{Compatibility (top) and dual compatibility (bottom) fans for the cycle on $4$ vertices with respect to different initial tubings.}
\vspace{-1.7cm}
\label{fig:dualCompatibilityFanCycle4}
\end{figure}
\end{remark}
\newpage
\begin{remark}[Linear dependences]
As for paths, only finitely many linear dependences occur for all cycles~$\cycleG_{n+1}$, both on compatibility vectors as on dual compatibility vectors. Indeed, with the interpretation of the maximal tubings in terms of centrally symmetric triangulations, the same kind of arguments as in~\cite{CeballosSantosZiegler} ensure that all these dependences can be inferred by checking the cycle~$\cycleG_{8}$ on~$8$ vertices. As for the path, these linear dependences only involve flipped and forced tubes, and the coefficients of the flipped tubes may only be~$1$ or~$2$ and these of the forces tubes may only be~$0,-1$ or~$-2$. See Section~\ref{subsec:proofPolytopality} for more details.
It is easy to find an example of a maximal tubing on the tripod such that one of the linear dependences obtained with respect to this maximal tubing does not only involve forced tubes. It implies in particular that the paths and cycles are the only graphs that have this property. It is then tempting to ask whether it is a coincidence that these graphs also are the only ones whose corresponding associahedra also are generalized associahedra.
\end{remark}
\subsection{Complete graphs}
We now consider the nested complex~$\nestedComplex(\completeG_{n+1})$ and the compatibility fan~$\compatibilityFan{\completeG_{n+1}}{\tubing^\circ}$ for the complete graph~$\completeG_{n+1}$ on~$n+1$ vertices. As already mentioned in the introduction, the nested complex~$\nestedComplex(\completeG_{n+1})$ is isomorphic to the $n$-dimensional \defn{simplicial permutahedron}, \ie the simplicial complex of collections of pairwise nested subsets of~$[n+1]$. See \fref{fig:permutahedron}.
It follows by classical results on the permutahedron that the complete graph~$\completeG_{n+1}$ has:
\begin{itemize}
\item $2^n-2$ proper tubes~\href{https://oeis.org/A000918}{\cite[A000918]{OEIS}} (proper subsets of~$[n]$),
\item $n!$ maximal tubings~\href{https://oeis.org/A000142}{\cite[A000142]{OEIS}} (permutations of~$[n]$),
\item $ k! \, S(n,k)$ tubings with~$k$ tubes, where~$S(n,k)$ is the Stirling number of second kind (\ie the number of ways to partition a set of $n$ elements into $k$ non-empty subsets)~\mbox{\href{https://oeis.org/A008277}{\cite[A008277]{OEIS}}}.
\end{itemize}
For two tubes~$\tube, \tube'$ of~$\completeG_{n+1}$, the compatibility degree of~$\tube$ with~$\tube'$ is $\compatibilityDegree{\tube}{\tube'} = -1$ if~$\tube = \tube'$, $\compatibilityDegree{\tube}{\tube'} = 0$ if~$\tube$ and~$\tube'$ are distinct and nested, and~$\compatibilityDegree{\tube}{\tube'} = |\tube' \ssm \tube|$ otherwise.
This connects the compatibility vector~$\compatibilityVector{\tubing^\circ}{\tube}$ to an alternative combinatorial model for the permutahedron in terms of lattice paths. Since all maximal tubings are equivalent, we can assume that~$\tubing^\circ = \set{[i]}{i \in [n]}$. For any tube~$\tube$ of~$\completeG_{n+1}$, we consider the lattice paths~$\phi(\tube)$ and~$\psi(\tube)$ whose horizontal steps above abscissa~$[i,i+1]$ lie at height~$|\tube \ssm [i]|$ and~$\compatibilityDegree{[i]}{\tube}$ respectively. These lattice paths are illustrated in \fref{fig:latticePaths}, where~$\phi(\tube)$ is the plain path while~$\psi(\tube)$ is dotted until it meets~$\phi(\tube)$. The proof of the following statement is left to the reader.
\begin{proposition}
\begin{enumerate}[(i)]
\item For any tube~$\tube$ of~$\completeG_{n+1}$, the lattice path~$\phi(\tube)$ is decreasing from~$(0,|\tube|)$ to~$(n+1,0)$ with vertical steps of height~$0$ or~$1$.
\item $\phi$ is surjective on the decreasing paths ending at~$(n+1,0)$ with vertical steps of height~$0$ or~$1$.
\item For any tubes~$\tube, \tube'$ of~$\completeG_{n+1}$, we have~$\tube \subseteq \tube'$ if and only if~$\phi(\tube')$ decreases when~$\phi(\tube)$ decreases. In particular, the paths~$\phi(\tube)$ and~$\phi(\tube')$ are then non-crossing.
\item For a tubing~$\tubing$ on~$\completeG_{n+1}$, the map~$\sigma(\tubing) : i \longmapsto |\set{\tube \in \tubing}{\phi(\tube) \text{ has a descent at abscissa } i}|+1$ is a surjection from~$[n+1]$ to~$[|\tubing|+1]$, and therefore~$\pi(\tubing) \eqdef \bigsqcup_{j \in [|\tubing|+1]} \sigma^{-1}(j)$ is an ordered partition of~$[n+1]$ into~$|\tubing|+1$ parts. The map~$\tubing \mapsto \pi(\tubing)$ defines an isomorphism form the nested complex~$\nestedComplex(\completeG_{n+1})$ to the refinement poset of ordered partitions.
\item For a tube~$\tube$ of~$\completeG_{n+1}$ not in~$\tubing^\circ$, the path~$\psi(\tube)$ is obtained from the path~$\phi(\tube)$ by replacing the initial down stairs by an horizontal path at height~$0$. See \fref{fig:latticePaths}, where~$\psi(\tube)$ is dotted until it meets~$\phi(\tube)$.
\end{enumerate}
\end{proposition}
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{latticePaths}}
\caption{The tubing~$\{146, 12468, 123468\}$ corresponds to three non-crossing decreasing lattice paths, and to the ordered partition~$57|3|28|146$.}
\label{fig:latticePaths}
\end{figure}
\begin{remark}[Dual compatibility fan]
\label{rem:dualCompatibilityFanComplete}
As discussed in Section~\ref{subsec:many} below, the complementation~${\tube \mapsto \ground \ssm \tube}$ defines an automorphism of the nested complex~$\nestedComplex(\completeG_{n+1})$, which dualizes the compatibility degree: $\compatibilityDegree{\tube}{\tube'} = \compatibilityDegree{\ground \ssm \tube'}{\ground \ssm \tube}$ for any tubes~$\tube, \tube'$ of~$\completeG_{n+1}$. Therefore, the dual compatibility fans are compatibility fans: for any tubing~$\tubing^\circ$~on~$\completeG_{n+1}$,
\[
\dualCompatibilityFan{\completeG_{n+1}}{\tubing^\circ} = \compatibilityFan{\completeG_{n+1}}{\set{\ground \ssm \tube^\circ}{\tube^\circ \in \tubing^\circ}}.
\]
\end{remark}
\begin{remark}[Linear dependences]
For the complete graph, the linear dependences between compatibility vectors of tubes involved in a flip can already be complicated. However, the coefficients~$(\alpha,\alpha')$ of the flipped tubes in these dependences can only take the following values:
\[
(k,k) \; \text{ with } k > 0, \qquad \text{ or } \qquad (k, kp) \; \text{ with } k,p > 0, \qquad \text{ or } \qquad (kp+p, kp) \; \text{ with } k,p > 0.
\]
\end{remark}
\subsection{Stars}
We finally consider the nested complex~$\nestedComplex(\starG_{n+1})$ and the compatibility fan~$\compatibilityFan{\starG_{n+1}}{\tubing^\circ}$ for the star~$\starG_{n+1}$ with~$n+1$ vertices, \ie the tree with~$n$ leaves~$\ell_1, \dots, \ell_n$ all connected to a central vertex denoted~$*$. The graph associahedron~$\Asso(\starG_{n+1})$ is called \defn{stellohedron}. We have represented in \fref{fig:stellohedron3} two realizations of the $3$-dimensional stellohedron.
\begin{figure}[b]
\capstart
\centerline{\includegraphics[scale=1.13]{stellohedron3}}
\caption{Two polytopal realizations of the $3$-dimensional stellohedron: their normal fans are the nested fan (left) and a compatibility fan (right).}
\label{fig:stellohedron3}
\end{figure}
\enlargethispage{.8cm}
One easily checks that the star~$\starG_{n+1}$ has:
\begin{itemize}
\item $2^n + n - 1$ proper tubes~\href{https://oeis.org/A052944}{\cite[A052944]{OEIS}} (distinguish tubes containing~$*$~or~not),
\item $\displaystyle{n! \sum_{i=0}^{n} \frac{1}{i!}}$ maximal tubings~\href{https://oeis.org/A000522}{\cite[A000522]{OEIS}} (consider the minimal tube containing~$*$),
\item $\displaystyle{\sum_{i \in [k]} \binom{n}{k-i} \, (i-1)! \big( i \, S(n-k+i,i) + S(n-k+i,i-1) \big)}$ tubings with $k$ tubes, where $S(m,p)$ denotes the Stirling number of second kind (\ie the number of ways to partition a set of $m$ elements into $p$ non-empty subsets) \href{https://oeis.org/A008277}{\cite[A008277]{OEIS}} (to see it, sum over the number~$i$ of tubes containing~$*$), and
\item $\displaystyle{4 \, n! \sum_{\sum n_i = n} \frac{1}{\prod n_i} - 1 = \sum_{i \ge 1} (i+1)^n/2^i}$ tubings in total (including the empty tubing). This is the number of chains in the boolean lattice on an $n$-element set~\href{https://oeis.org/A007047}{\cite[A007047]{OEIS}} (an immediate bijection is given by the spines of the tubings).
\end{itemize}
We consider the initial maximal tubing~$\tubing^\circ \eqdef \big\{ \{\ell_1\}, \dots, \{\ell_n\} \big\}$ whose tubes are the~$n$ leaves of~$\starG_{n+1}$. The other~$2^n-1$ tubes of~$\starG_{n+1}$ are the tubes containing the central vertex~$*$ and some leaves (but not all). The compatibility degree of such a tube~$\tube$ containing~$*$ with a tube~$\{\ell_i\}$ is~$0$ if~$\ell_i \in \tube$ and~$1$ if~$\ell_i \notin \tube$. The compatibility vector~$\compatibilityVector{\tubing^\circ}{\tube}$ of~$\tube$ with respect to~$\tubing^\circ$ is thus given by the characteristic vector of the leaves of~$\starG_{n+1}$ not contained in~$\tube$. Moreover, two tubes~$\tube, \tube' \notin \tubing^\circ$ are compatible if and only if they are nested (since they both contain the vertex~$*$). Therefore, the compatibility fan~$\compatibilityFan{\starG_{n+1}}{\tubing^\circ}$ is obtained from the coordinate hyperplan fan by a barycentric subdivision of the positive orthant. Examples in dimension~$2$, $3$ and~$4$ are gathered in \fref{fig:star}.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{starSymmetric}\includegraphics[width=.37\textwidth]{3dimensionalCompatibilityFansLabeled/tripod1}\includegraphics[scale=.35]{4dimensionalCompatibilityFansLabeled/star}}
\caption{The compatibility fans~$\compatibilityFan{\starG_{n+1}}{\tubing^\circ}$ for the star~$\starG_{n+1}$ and the initial tubing~$\tubing^\circ = \big\{ \{\ell_1\}, \dots, \{\ell_n\} \big\}$ formed by its leaves (for~$n \in \{2,3,4\}$).}
\label{fig:star}
\end{figure}
\begin{remark}[Dual compatibility fan]
\label{rem:primalDualStar}
Observe that the compatibility degree of any tube of~$\starG_{n+1}$ with any leaf of~$\starG_{n+1}$ belongs to~$\{-1,0,1\}$. Therefore, $\compatibilityVector{\tubing^\circ}{\tube} = \dualCompatibilityVector{\tube}{\tubing^\circ}$ for any tube~$\tube$, so that the compatibility and dual compatibility fans with respect to the initial tubing~$\tubing^\circ$ coincide. This does not hold for arbitrary initial tubings on~$\starG_{n+1}$, see Remark~\ref{rem:naive} and \fref{fig:ctrexmNaive}\,(right).
\end{remark}
\begin{remark}[Linear dependences]
In the special case discussed in this section, all linear dependences between compatibility vectors of tubes involved in a flip are inclusion-exclusion dependences as in the beginning of the proof of Theorem~\ref{theo:compatibilityFan} in Section~\ref{subsec:proofCompatibilityFan}. The coefficients of the flipped tubes thus always equal~$1$ while those of the forced tubes (not in the initial tubing~$\tubing^\circ$) always equal~$-1$.
\end{remark}
\begin{remark}
As pointed out by F.~Santos, the stellohedron~$\Asso(\starG_{n+1})$ coincides with the secondary polytope of two concentric copies of an $(n-1)$-dimensional simplex. See \fref{fig:secondaryPolytopeStellohedron}.
\end{remark}
\begin{figure}[b]
\capstart
\centerline{\includegraphics[scale=1.13]{secondaryPolytopeStellohedron}}
\caption{The stellohedron (left) is a secondary polytope (right).}
\label{fig:secondaryPolytopeStellohedron}
\vspace*{-.8cm}
\end{figure}
\section{Further topics}
\label{sec:furtherTopics}
\enlargethispage{.3cm}
In this section, we discuss several further topics in connection to compatibility fans. Note that we only state the results for compatibility fans, but similar statements hold for dual compatibility fans. Section~\ref{subsec:product} studies the behavior of the compatibility fans with respect to products and links. In Section~\ref{subsec:many}, we show that most compatibility fans are not linearly isomorphic, which requires a description of all nested complex isomorphisms. Section~\ref{subsec:polytopality} discusses the question of the realization of our compatibility fans as normal fans of convex polytopes. In Section~\ref{subsec:designNestedComplex}, we extend our construction to design nested complexes~\cite{DevadossHeathVipismakul}. Finally, we discuss in Section~\ref{subsec:LPA} the connection of this paper to Laurent Phenomenon algebras~\cite{LamPylyavskyy-LaurentPhenomenonAlgebras, LamPylyavskyy-LinearLaurentPhenomenonAlgebras}.
\subsection{Products and restrictions}
\label{subsec:product}
In all examples that we discussed earlier, we only considered connected graphs. Compatibility fans for disconnected graphs can be reconstructed from those for connected graphs by the following statement, whose proof is left to the reader.
\begin{proposition}
\label{prop:product}
If~$\graphG$ has connected components~$\graphG_1, \dots, \graphG_k$, then the nested complex~$\nestedComplex(\graphG)$ is the join of the nested complexes~$\nestedComplex(\graphG_1), \dots, \nestedComplex(\graphG_k)$. Moreover, for any maximal tubings~$\tubing_1^\circ, \dots, \tubing_k^\circ$ on~$\graphG_1, \dots, \graphG_k$ respectively, the compatibility fan~$\compatibilityFan{\graphG}{\tubing^\circ}$ with respect to the maximal tubing ${\tubing^\circ \eqdef \tubing_1^\circ \cup \dots \cup \tubing_k^\circ}$ on~$\graphG$ is the product of the compatibility fans~$\compatibilityFan{\graphG_1}{\tubing_1^\circ}, \dots, \compatibilityFan{\graphG_k}{\tubing_k^\circ}$:
\[
\compatibilityFan{\graphG}{\tubing^\circ} = \compatibilityFan{\graphG_1}{\tubing_1^\circ} \times \dots \times \compatibilityFan{\graphG_k}{\tubing_k^\circ} = \set{C_1 \times \dots \times C_k}{C_i \in \compatibilityFan{\graphG_i}{\tubing_i^\circ} \text{ for all } i \in [k]}.
\]
\end{proposition}
\fref{fig:productFans}\,(right) illustrates Proposition~\ref{prop:product} with the compatibility fan of a graph formed by two paths. Compatibility fans of paths are discussed in Section~\ref{subsec:paths}. Besides all compatibility vectors, the cones of three different tubings are represented in \fref{fig:productFans}\,(right).
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1.05]{productFans}}
\caption{The compatibility fan of a disconnected graph is the product (right) of the compatibility fans of its connected components (left and~middle).}
\label{fig:productFans}
\end{figure}
As observed in~\cite{CarrDevadoss}, all links of graphical nested complexes are joins of graphical nested complexes. The following statement asserts that the compatibility fans reflect this property on coordinate hyperplanes. To be more precise, for a tube~$\tube^\circ$ of~$\graphG$, we denote by~$\graphG{}[\tube^\circ]$ the restriction of~$\graphG$ to~$\tube^\circ$ and by~$\graphG{}^\star\tube^\circ$ the~\defn{reconnected complement} of~$\tube^\circ$ in~$\graphG$, \ie the graph with vertex set~$\ground \ssm \tube^\circ$ and edge set~$\bigset{e \in \binom{\ground \ssm \tube^\circ}{2}}{\text{$e$ or~$e \cup \tube^\circ$ is connected in~$\graphG$}}$. A maximal tubing~$\tubing^\circ$ on~$\graphG$ containing~$\tube^\circ$ induces maximal tubings~$\tubing^\circ[\tube^\circ] \eqdef \set{\tube}{\tube \in \tubing^\circ, \tube \subsetneq \tube^\circ}$ on the restriction~$\graphG{}[\tube^\circ]$ and~${\tubing^\circ}^\star\tube^\circ \eqdef \set{\tube \ssm \tube^\circ}{\tube \in \tubing^\circ, \tube \not\subset \tube^\circ}$ on the reconnected complement~$\graphG^\star\tube^\circ$. See \fref{fig:exmReconnectedComplement}.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{exmReconnectedComplement}}
\caption{The red tube~$\tube^\circ\ex = \{a,b,d,f,g,h,i,k,l\}$ in the maximal tubing~$\tubing^\circ$~(left) yields a maximal tubing~$\tubing^\circ\ex[\tube^\circ\ex]$ on the restriction~$\graphG\ex[\tube^\circ\ex]$ (middle) and a maximal tubing~${\tubing^\circ\ex}^\star\tube^\circ\ex$ on the reconnected complement~${\graphG\ex}^\star\tube^\circ\ex$ (right).}
\label{fig:exmReconnectedComplement}
\end{figure}
\begin{proposition}
\label{prop:restriction}
The link of a tube~$\tube^\circ$ in the nested complex~$\nestedComplex(\graphG)$ is isomorphic to the join of the nested complexes~$\nestedComplex(\graphG{}[\tube^\circ])$ and~$\nestedComplex(\graphG^\star\tube^\circ)$. Moreover, for an initial maximal tubing~$\tubing^\circ$ containing~$\tube^\circ$, the intersection of the compatibility fan~$\compatibilityFan{\graphG}{\tubing^\circ}$ with the coordinate hyperplane orthogonal to~$\b{e}_{\tube^\circ}$ is the product of the compatibility fans~$\compatibilityFan{\graphG{}[\tube^\circ]}{\tubing^\circ[\tube^\circ]}$ and~$\compatibilityFan{\graphG^\star\tube^\circ}{{\tubing^\circ}^\star\tube^\circ}$.
\end{proposition}
This statement follows from Lemmas~\ref{lem:restriction} and~\ref{lem:restrictionCompatibility} and Theorem~\ref{theo:compatibilityFan}, proved in Section~\ref{sec:proofs}.
\subsection{Many compatibility fans}
\label{subsec:many}
In this section, we show that we obtained many distinct compatibility fans. Following~\cite{CeballosSantosZiegler}, we classify compatibility fans up to linear isomorphisms: two fans~$\cF, \cF'$ of~$\R^n$ are linearly isomorphic if there exists an invertible linear map which sends the cones of~$\cF$ to the cones of~$\cF'$. Observe already that if two compatibility fans~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\compatibilityFan{\graphG'}{\tubing'^\circ}$ are linearly isomorphic, then the two nested complexes~$\nestedComplex(\graphG)$ and~$\nestedComplex(\graphG')$ are (combinatorially) isomorphic, meaning that there is a bijection~$\Phi$ from the tubes of~$\graphG$ to the tubes of~$\graphG'$ which preserves the compatibility. The converse does not always hold: a nested complex isomorphism can preserve compatibility without preserving the compatibility degree. However, we prove below that the nested complex isomorphisms are so constrained that they all either preserve the compatibility degree and thus induce linear isomorphisms between compatibility fans, or exchange the compatibility and dual compatibility degrees and thus induce linear isomorphisms between compatibility and dual compatibility fans.
In the sequel, we describe all nested complex isomorphisms. Observe first that an isomorphism~$\phi$ between two graphs~$\graphG$ and~$\graphG'$ automatically induces an isomorphism~$\Phi$ between the nested complexes~$\nestedComplex(\graphG)$ and~$\nestedComplex(\graphG')$ defined by~$\Phi(\tube) \eqdef \set{\phi(v)}{v \in \tube}$ for all tubes~$\tube$ on~$\graphG$. We say that such a nested complex isomorphism~$\Phi$ is \defn{trivial}. Trivial isomorphisms clearly preserve compatibility degrees: $\compatibilityDegree{\Phi(\tube)}{\Phi(\tube')} = \compatibilityDegree{\tube}{\tube'}$ for any tubes~$\tube, \tube'$ on~$\graphG$. We are interested in non-trivial nested complex isomorphisms.
We first want to underline two relevant examples.
\begin{example}
\label{exm:nestedComplexIsomorphisms}
The reader can check that:
\begin{enumerate}[(i)]
\item The complementation~$\tube \mapsto \ground \ssm \tube$ is a non-trivial automorphism of the nested complex~$\nestedComplex(\completeG_{n+1})$ of the complete graph~$\completeG_{n+1}$. It dualizes the compatibility degree: $\compatibilityDegree{\ground \ssm \tube}{\ground \ssm \tube'} = \compatibilityDegree{\tube'}{\tube}$ for any tubes~$\tube, \tube'$ of~$\completeG_{n+1}$.
\item The map~$\rot$ defined for~$1 \le j \le k \le n+1$ by
\[
\rot[j,k] \eqdef
\begin{cases}
[k+1, n+1] & \text{if } j = 1, \\
[j-1, k-1] & \text{if } j > 1,
\end{cases}
\]
is a non-trivial automorphism of the nested complex~$\nestedComplex(\pathG_{n+1})$ of the path~$\pathG_{n+1}$. Indeed, up to conjugation by the bijection~$\delta \mapsto \tube_\delta$ of Section~\ref{subsec:paths}, the map~$\rot$ coincides with the (combinatorial) $1$-vertex clockwise rotation of the $(n+3)$-gon~$Q_{n+3}$. See \fref{fig:rotation}.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.6]{rotation}}
\vspace{.5cm}
\centerline{
$\begin{array}{c|c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c@{\;\,}c}
\tube & [1] & [2] & [3] & [4] & [5] & [6] & [1,2] & [2,3] & [3,4] & [4,5] & [5,6] & [1,3] & [2,4] & [3,5] & [4,6] & [1,4] & [2,5] & [3,6] & [1,5] & [2,6] \\
\rot\tube & [2,6] & [1] & [2] & [3] & [4] & [5] & [3,6] & [1,2] & [2,3] & [3,4] & [4,5] & [4,6] & [1,3] & [2,4] & [3,5] & [5,6] & [1,4] & [2,5] & [6] & [1,5]
\end{array}$
}
\vspace{-.1cm}
\caption{The $1$-vertex clockwise rotation of the~$(n+3)$-gon induces a non-trivial automorphism~$\rot$ of the nested complex~$\nestedComplex(\pathG_{n+1})$.}
\label{fig:rotation}
\end{figure}
Therefore, $\rot$ has order~$n+3$, and its iterated powers are explicitly described by
\[
\rot^p[j,k] \eqdef
\begin{cases}
[j-p, k-p] & \text{if } 0 \le p < j, \\
[k-p+2, n+j-p+1] & \text{if } j \le p < k+2, \\
[n+p+j-2k-1, n+p-k-1] & \text{if } k+2 \le p < n+3.
\end{cases}
\]
In fact, it is shown in~\cite[Lemma~2.2]{CeballosSantosZiegler} that the automorphism group of the nested complex~$\nestedComplex(\pathG_{n+1})$ is the dihedral group generated by the non-trivial automorphism~$\rot$ (the rotation of the \mbox{$(n+3)$-gon~$Q_{n+3}$}) and the automorphism~$\rev : [j,k] \mapsto [n+2-k, n+2-j]$ induced by the graph automorphism~${j \mapsto n+2-j}$ of~$\pathG_{n+1}$ (the vertical reflection of the $(n+3)$-gon~$Q_{n+3}$). Note that since the compatibility degree on~$\pathG_{n+1}$ is in~$\{-1,0,1\}$, any nested complex automorphism preserves the compatibility degree (and also dualizes it since it is symmetric).
\end{enumerate}
\end{example}
We now describe a generalization of both cases of Example~\ref{exm:nestedComplexIsomorphisms}. For ${\underline{n} \eqdef \{n_1, \dots, n_\ell\}} \in \N^\ell$ with~$n + 1 = \sum_{i \in [\ell]} (n_i + 1)$, define the \defn{spider}~$\spiderG_{\underline{n}}$ as the graph with vertices~$\set{v^i_j}{i \in [\ell], 0 \le j \le n_i}$ and edges~${\bigset{\big\{v^i_{j-1}, v^i_j\big\}}{i \in [\ell], j \in [n_i]} \cup \bigset{\big\{v^i_0, v^{i'}_0\big\}}{i \ne i' \in [\ell]}}$. For~$0 \le j \le k \le n_i$, we denote by~$\big[ v^i_j, v^i_k \big]$ the path between~$v^i_j$ and~$v^i_k$ in~$\spiderG_{\underline{n}}$. Informally, the spider~$\spiderG_{\underline{n}}$ consists in~$\ell$ paths~$[v^i_1, v^i_{n_i}]$ called \defn{legs} of the spider, each attached to a vertex~$v^i_0$ of a clique called \defn{body} of the spider. See \fref{fig:isomorphismSpiders}. Note that spiders are sometimes called sunlike graphs in the literature.
We now define a non-trivial automorphism~$\Omega$ of the nested complex~$\nestedComplex(\spiderG_{\underline{n}})$ of the spider~$\spiderG_{\underline{n}}$. We distinguish two kinds of tubes of~$\nestedComplex(\spiderG_{\underline{n}})$:
\begin{description}
\item[Leg tubes] A tube~$\tube$ disjoint from the body is included in a leg. The map~$\Omega$ sends~$\tube$ into its image by the transformation which cuts the leg containing~$\tube$ and glues it back to the body by its other endpoint. Formally, for~$i \in [\ell]$ and~$1 \le j \le k \le n_i$,
\[
\Omega\big( \big[ v^i_j, v^i_k \big] \big) \eqdef \big[ v^i_{n_i+1-k}, v^i_{n_i+1-j} \big].
\]
Note that~$\Omega$ sends a leg tube~$\tube$ to a leg tube~$\Omega(\tube)$ with~$|\Omega(\tube)| = |\tube|$.
\item[Body tubes] A tube~$\tube$ intersecting the body is the union of initial segments~$\big[ v^i_0, v^i_{k_i} \big]$, with~$-1 \le k_i \le n_i$ (with the convention that~$[v^i_0, v^i_{-1}] = \varnothing$). We then define
\[
\Omega\big( \bigcup_{i \in [\ell]} \big[ v^i_0, v^i_{k_i} \big] \big) \eqdef \bigcup_{i \in [\ell]} \big[ v^i_0, v^i_{n_i-1-k_i} \big].
\]
Note that~$\Omega$ sends a body tube~$\tube$ to a body tube~$\Omega(\tube)$ with~$|\Omega(\tube)| = |\ground| - |\tube|$.
\end{description}
\fref{fig:isomorphismSpiders} illustrates the map~$\Omega$ on different tubes of the spider~$\spiderG_{\{0,3,2,3,0,3,2,3\}}$.
Observe that~$\Omega$ indeed generalizes both non-trivial nested complex automorphisms of Example~\ref{exm:nestedComplexIsomorphisms}:
\begin{enumerate}[(i)]
\item The complete graph~$\completeG_{n+1}$ is the spider~$\spiderG_{\{0\}^{n+1}}$ whose legs are all empty. The automorphism~$\Omega$ of~$\nestedComplex(\spiderG_{\{0\}^{n+1}})$ specializes to the complementation~$\tube \mapsto \ground \ssm \tube$ on~$\nestedComplex(\completeG_{n+1})$.
\item The path~$\pathG_{n+1}$ is a degenerate spider whose body can be chosen at different places. Indeed, the path~$\pathG_{n+1}$ coincides with the spider~$\spiderG^1 \eqdef \spiderG_{\{n\}}$ with body~$\{1\}$ and the single leg~$[2,n+1]$, and the automorphism~$\Omega$ of~$\nestedComplex(\spiderG^1)$ is the composition of the rotation automorphism~$\rot$ with the vertical reflection automorphism~$\rev$. Similarly, for any~$2 \le p \le n+1$, the path~$\pathG_{n+1}$ coincides with the spider~$\spiderG^p \eqdef \spiderG_{\{p-2, n+1-p\}}$ with body~$\{p-1,p\}$ and legs~${[1,p-2]}$ and~${[p+1,n+1]}$, and the automorphism~$\Omega$ of~$\nestedComplex(\spiderG^p)$ is the composition of~$\rot^p$ with~$\rev$. Finally, the path~$\pathG_{n+1}$ coincides with the spider~$\spiderG^{n+2} \eqdef \spiderG_{\{n\}}$ with body~$\{n+1\}$ and the single leg~$[n]$, and the automorphism~$\Omega$ of~$\nestedComplex(\spiderG^{n+2})$ is the composition of~$\rot^{n+2}$ with~$\rev$.
\end{enumerate}
This actually suggests an alternative description of~$\Omega$ on arbitrary spiders~$\spiderG_{\underline{n}}$. Namely, $\Omega$ is equivalently described by the following steps: shift all leg tubes towards the body, complement all body tubes, delete all edges~$\bigset{\big\{v^i_0, v^{i'}_0\big\}}{i \ne i' \in [\ell]}$ of the body, replace them by the clique~${\bigset{\big\{v^i_{n_i}, v^{i'}_{n_{i'}}\big\}}{i \ne i' \in [\ell]}}$ on the feets of the spider, and finally apply the trivial isomorphism from the resulting spider back to the initial spider. Our original presentation of~$\Omega$ will nevertheless be easier to handle in the proofs.
The following statement is proved in Section~\ref{subsec:proofIsomorphisms}.
\begin{proposition}
\label{prop:exmAutomorphism}
The map~$\Omega$ is a non-trivial involutive automorphism of the nested complex~$\nestedComplex(\spiderG_{\underline{n}})$ of the spider~$\spiderG_{\underline{n}}$ which dualizes the compatibility degree: $\compatibilityDegree{\Omega(\tube)}{\Omega(\tube')} = \compatibilityDegree{\tube'}{\tube}$.
\end{proposition}
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{isomorphismSpider}}
\caption{The spider~$\spiderG_{\{0,3,2,3,0,3,2,3\}}$ and examples of the action of the non-trivial nested complex isomorphism~$\Omega$: the tubing~$\tubing$ on the left is sent to the tubing~$\Omega(\tubing)$ on the right.}
\label{fig:isomorphismSpiders}
\end{figure}
\begin{remark}
It follows from Proposition~\ref{prop:exmAutomorphism} that all dual compatibility fans of a spider~$\spiderG_{\underline{n}}$ are also compatibility fans of~$\spiderG_{\underline{n}}$: we have~$\dualCompatibilityFan{\spiderG_{\underline{n}}}{\tubing^\circ} = \compatibilityFan{\spiderG_{\underline{n}}}{\Omega(\tubing^\circ)}$ for any tubing~$\tubing^\circ$ on~$\spiderG_{\underline{n}}$. Note that we already used this observation for complete graphs in Remark~\ref{rem:dualCompatibilityFanComplete}.
\end{remark}
In fact, these non-trivial automorphisms of the nested complexes of the spiders are essentially the only non-trivial nested complex isomorphisms. The following statements are proved in Section~\ref{subsec:proofIsomorphisms}.
\begin{proposition}
\label{prop:nestedComplexIsomorphismDisconnected}
A nested complex isomorphism~$\Phi : \nestedComplex(\graphG) \to \nestedComplex(\graphG')$ restricts to nested complex isomorphisms~$\nestedComplex(\graphG[H]) \to \nestedComplex(\graphG[H]')$ between maximal connected subgraphs~$\graphG[H]$ of~$\graphG$ and~$\graphG[H]'$ of~$\graphG'$.
\end{proposition}
\begin{theorem}
\label{theo:nestedComplexIsomorphisms}
Let~$\graphG$ and~$\graphG'$ be two connected graphs and~$\Phi : \nestedComplex(\graphG) \to \nestedComplex(\graphG')$ be a non-trivial nested complex isomorphism. Then~$\graphG$ and~$\graphG'$ are spiders and there exists a graph isomorphism~$\psi : \graphG \to \graphG'$ which induces a nested complex isomorphism~$\Psi : \nestedComplex(\graphG) \to \nestedComplex(\graphG')$ (defined by~$\Psi(\tube) \eqdef \set{\psi(v)}{v \in \tube}$) such that the composition~$\Psi^{-1} \circ \Phi$ coincides with the non-trivial nested complex automorphism~$\Omega$ on~$\nestedComplex(\graphG)$.
\end{theorem}
\begin{corollary}
For connected graphs~$\graphG$, $\graphG'$, any nested complex isomorphism~$\Phi : \nestedComplex(\graphG) \to \nestedComplex(\graphG')$ either preserves or dualizes the compatibility degree: either~$\compatibilityDegree{\Phi(\tube)}{\Phi(\tube')} = \compatibilityDegree{\tube}{\tube'}$ for all~$\tube, \tube'$ of~$\graphG$, or~$\compatibilityDegree{\Phi(\tube)}{\Phi(\tube')} = \compatibilityDegree{\tube'}{\tube}$ for all~$\tube, \tube'$ of~$\graphG$.
\end{corollary}
To finish our classification of primal and dual compatibility fans up to linear isomorphisms, it remains to understand when the primal and the dual compatibility fans of~$\graphG$ with respect to the same initial maximal tubing are linearly isomorphic. For example, we already observed that
\begin{enumerate}[(i)]
\item $\compatibilityFan{\pathG_{n+1}}{\tubing^\circ} = \dualCompatibilityFan{\pathG_{n+1}}{\tubing^\circ}$ for any initial tubing~$\tubing^\circ$ on a path~$\pathG_{n+1}$, see Remark~\ref{rem:primalDualPath},
\item $\compatibilityFan{\starG_{n+1}}{\tubing^\circ} = \dualCompatibilityFan{\starG_{n+1}}{\tubing^\circ}$ for the initial tubing~$\tubing^\circ \eqdef \big\{ \{\ell_1\}, \dots, \{\ell_n\} \big\}$ of the star~$\starG_{n+1}$ whose tubes are the~$n$ leaves, see Remark~\ref{rem:primalDualStar}.
\end{enumerate}
These examples extend to all subdivisions of stars. Namely, for~$\underline{n} \eqdef \{n_1, \dots, n_\ell\} \in \N^\ell$ with ${n = \sum_{i \in [\ell]} (n_i + 1)}$, define the \defn{octopus}~$\octopusG_{\underline{n}}$ as the graph with vertices~$\big\{ * \big\} \cup \set{v^i_j}{i \in [\ell], 0 \le j \le n_i}$ and edges~${\bigset{\big\{v^i_{j-1}, v^i_j\big\}}{i \in [\ell], j \in [n_i]} \cup \bigset{\big\{*, v^i_0 \big\}}{i \in [\ell]}}$. Informally, the octopus~$\octopusG_{\underline{n}}$ consists in~$\ell$ paths~$[v^i_0, v^i_{n_i}]$ called \defn{legs}, each attached to a~\defn{head}~$*$. These graphs are often called starlike graphs in the literature, we use the term octopus to stay in the wildlife lexical field. Note that the path~$\pathG_{n+1}$ is a degenerate octopus where the head can be chosen at any vertex. As for stars, a tube of an octopus~$\octopusG_{\underline{n}}$ that does not contain its head~$*$ is either compatible or exchangeable with any other tube of~$\octopusG_{\underline{n}}$. Therefore, if~$\tubing^\circ$ is a maximal tubing on~$\octopusG_{\underline{n}}$ whose tubes do not contain the head~$*$, then the compatibility fan~$\compatibilityFan{\octopusG_{\underline{n}}}{\tubing^\circ}$ and the dual compatibility fan~$\dualCompatibilityFan{\octopusG_{\underline{n}}}{\tubing^\circ}$ coincide. The following lemma states that this only happens in this situation.
\begin{lemma}
\label{lem:comparisonPrimalDual}
Let~$\graphG$ be a connected graph and~$\tubing^\circ$ be an initial maximal tubing on~$\graphG$. If~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\dualCompatibilityFan{\graphG}{\tubing^\circ}$ are linearly isomorphic, then~$\graphG$ is an octopus whose head is contained in no~tube~of~$\tubing^\circ$.
\end{lemma}
We finally obtain our classification of primal and dual compatibility fans. We first focus on primal compatibility fans and then conclude with both primal and dual compatibility fans together.
\begin{corollary}
\label{coro:many}
The number of linear isomorphism classes of compatibility fans of a connected graph~$\graphG$ is:
\begin{enumerate}[(i)]
\item the number of triangulations of the regular $(n+3)$-gon up to the action of the dihedral group if $\graphG = \pathG_{n+1}$ is a path,
\item the number of orbits of maximal tubings on~$\graphG$ under graph automorphisms of~$\graphG$ otherwise.
\end{enumerate}
\end{corollary}
\begin{corollary}
The number of linear isomorphism classes of primal and dual compatibility fans of a connected graph~$\graphG$ is:
\begin{enumerate}[(i)]
\item the number of triangulations of the regular $(n+3)$-gon up to the action of the dihedral group if $\graphG = \pathG_{n+1}$ is a path,
\item the number of $\graphG$-automorphism orbits of maximal tubings on~$\graphG$ if~$\graphG$ is a spider but not~a~path,
\item the number of $\graphG$-automorphism orbits of maximal tubings on~$\graphG$, counted twice if not rooted at~$*$, if~$\graphG$ is an octopus with head~$*$ but not a path,
\item twice the number of $\graphG$-automorphism orbits of maximal tubings on~$\graphG$ otherwise.
\end{enumerate}
\end{corollary}
\begin{remark}
Not only we obtain many non-isomorphic complete simplicial fan realizations for graphical nested complexes (as stated in Corollary~\ref{coro:many}), but these realizations cannot be derived from the existing geometric constructions for graph associahedra. Indeed, all previous polytopal realizations of graph associahedra can be obtained by successive face truncations of a simplex~\cite{CarrDevadoss} or of a cube~\cite{Volodin, DevadossForceyReisdorfShowers}. Not all compatibility fans can be constructed in this way. For example, the leftmost compatibility fan of \fref{fig:allLabels} is not linearly isomorphic to the normal fan of a polytope obtained by face truncations of the $3$-dimensional simplex or cube.
\end{remark}
\subsection{Polytopality}
\label{subsec:polytopality}
In this section, we briefly discuss the polytopality of our compatibility fans for graphical nested complexes. A complete polyhedral fan is said to be \defn{polytopal} (or \defn{regular}) if it is the normal fan of a polytope. It is well known that not all complete polyhedral fans (even simplicial) are polytopal. Examples are easily constructed from non-regular triangulations, see \eg the discussion in~\cite[Chapter~2]{DeLoeraRambauSantos}.
\para{Polytopality of cluster fans}
The polytopality of cluster fans has been studied since the foundations of finite type cluster algebras. For compatibility fans, polytopality was shown for particular initial clusters by F.~Chapoton, S.~Fomin and A.~Zelevinsky~\cite{ChapotonFominZelevinsky} and in type~$A$ by F.~Santos~\cite[Section~5]{CeballosSantosZiegler}.
\begin{theorem}[\cite{ChapotonFominZelevinsky, CeballosSantosZiegler}]
\label{theo:polytopalityCompatibilityFanCluster}
The $\b{d}$-vector fan (or compatibility fan) is polytopal for
\begin{itemize}
\item any initial cluster in any type~$A$ cluster algebra~\cite[Section~5]{CeballosSantosZiegler}, and
\item the bipartite initial cluster in any finite type cluster algebra~\cite{ChapotonFominZelevinsky}.
\end{itemize}
\end{theorem}
The polytopality of the $\b{g}$-vector fan was studied by C.~Hohlweg, C.~Lange and H.~Thomas~\cite{HohlwegLangeThomas}. See also recent alternative proofs by S.~Stella~\cite{Stella} and V.~Pilaud and C.~Stump~\cite{PilaudStump-brickPolytope}.
\begin{theorem}[\cite{HohlwegLangeThomas, Stella, PilaudStump-brickPolytope}]
The $\b{g}$-vector fan (or Cambrian fan) is polytopal for any acyclic initial cluster in any finite type cluster algebra.
\end{theorem}
To our knowledge, the polytopality of the $\b{d}$- and $\b{g}$-vector fans remains open in all other cases. All these results rely on the following characterization of polytopality for complete simplicial fans, in connection to regular triangulations of vector configurations and to the theory of secondary polytopes~\cite{GelfandKapranovZelevinsky}, see also~\cite{DeLoeraRambauSantos}. Equivalent formulations of this characterization appear \eg in~\cite[Lemma~2.1]{ChapotonFominZelevinsky}, \cite[Proposition~6.3]{Zelevinsky}, \cite[Theorem~4.1]{HohlwegLangeThomas}, or~\cite[Lemma~5.4]{CeballosSantosZiegler}. Here, we follow the presentation of the first two which fits our previous notations.
\begin{proposition}
\label{prop:polytopalityFan}
Let~$\cF$ be a complete simplicial fan in~$\R^n$ and let~$\b{R}$ denote a set of vectors generating its rays ($1$-dimensional cones). Then the following are equivalent:
\begin{enumerate}
\item $\cF$ is the normal fan of a simple polytope in~$(\R^n)^*$;
\item There exists a map~$\omega: \b{R} \to \R_{> 0}$ such that for any two maximal adjacent cones~$\R_{\ge 0} \b{S}$ and~$\R_{\ge 0} \b{S}'$ of~$\cF$ with $\b{S}, \b{S}' \subseteq \b{R}$ and~$\b{S} \ssm \{\b{s}\} = \b{S}' \ssm \{\b{s'}\}$, we have
\[
\alpha \, \omega(\b{s}) + \alpha' \, \omega(\b{s}') + \sum_{\b{r} \in \b{S} \cap \b{S}'} \beta_{\b{r}} \, \omega(\b{r}) > 0,
\]
where
\[
\alpha \, \b{s} + \alpha' \, \b{s}' + \sum_{\b{r} \in \b{S} \cap \b{S}'} \beta_{\b{r}} \, \b{r} = 0
\]
is the unique (up to rescaling) linear dependence with~$\alpha, \alpha' > 0$ between the rays of~$\b{S} \cup \b{S}'$.
\end{enumerate}
Under these conditions, $\cF$ is the normal fan of the polytope defined by
\[
\set{\phi \in (\R^n)^*}{\dotprod{\phi}{\b{r}} \le \omega(\b{r}) \text{ for all } \b{r} \in \b{R}}.
\]
\end{proposition}
\para{Polytopality of compatibility fans}
We have seen in Section~\ref{subsec:graphAssociahedra} that the nested fan is the normal fan of the graph associahedron of~\cite{CarrDevadoss, Devadoss, Postnikov, Zelevinsky}. For the compatibility fan, the question of the polytopality remains open:
\begin{conjecture}
\label{conj:polytopality}
All primal and dual compatibility fans of graphical nested complexes are polytopal.
\end{conjecture}
To settle this conjecture, the hope would be to apply the characterization of polytopality for complete simplicial fans presented in Proposition~\ref{prop:polytopalityFan}. Besides finding an explicit function~$\omega$ on the compatibility vectors of the tubes of a graph, our main issue is that we do not control the details of the linear dependence between the compatibility vectors of the tubes involved in a flip. See the proof of Theorem~\ref{theo:compatibilityFan} in Section~\ref{subsec:proofCompatibilityFan}.
To support Conjecture~\ref{conj:polytopality}, we have studied the polytopality of the compatibility fans of the specific families of graphs discussed in Section~\ref{sec:specificGraphs}. We show in Section~\ref{subsec:proofPolytopality} that Conjecture~\ref{conj:polytopality} holds for paths and cycles.
\begin{theorem}
\label{theo:polytopalityPathsCycles}
All compatibility and dual compatibility fans of paths and cycles are polytopal.
\end{theorem}
Note that the case of paths is covered by the results of~\cite[Section~5]{CeballosSantosZiegler} presented in Theorem~\ref{theo:polytopalityCompatibilityFanCluster}. For cycles, the result was unknown except for the bipartite initial tubing by the results of~\cite{ChapotonFominZelevinsky} on type~$B$ and~$C$ cluster algebras. Via the correspondences given in Propositions~\ref{prop:typeA} and~\ref{prop:typeBC}, Theorem~\ref{theo:polytopalityPathsCycles} translates to the following relevant property of $\b{d}$-vector fans.
\begin{corollary}
In types~$A$, $B$ and~$C$ cluster algebras, the $\b{d}$-vector fan with respect to any initial cluster (acyclic or not) is polytopal.
\end{corollary}
We were not able to settle Conjecture~\ref{conj:polytopality} for arbitrary graphs. We believe that this question is worth investigating. As already mentioned, it requires a better understanding of all linear dependences between the compatibility vectors of the tubes involved in a flip.
\medskip
In another direction, we checked empirically that all $3$-dimensional compatibility and dual compatibility fans of Section~\ref{subsec:fewVertices} and \fref{fig:4vertices} are polytopal. Using the characterization given in Proposition~\ref{prop:polytopalityFan}, it boils down to check the feasibility of (many) linear programs.
\medskip
Finally, as a curiosity and to conclude this polytopality section on a recreative note, we provide a polytopal realization of the compatibility fan for the star~$\starG_{n+1}$ with respect to the initial tubing~$\tubing^\circ \eqdef \big\{ \{\ell_1\}, \dots, \{\ell_n\} \big\}$ whose tubes are the~$n$ leaves~of~$\starG_{n+1}$. We first observe that this fan is linearly isomophic to the fan~$\cG(\graphG)$ of Theorem~\ref{theo:gFan}. Therefore, it can be realized by an affine transformation of the graph associahedra constructed in Theorem~\ref{theo:graphAssociahedron}.
Here, we prefer to give a direct construction with integer coordinates. We provide both the vertex and the facet descriptions of this realization. On the one hand, for each maximal tubing~$\tubing$ on~$\starG_{n+1}$ we define a point~$\b{x}(\tubing) \in \R^n$ whose~$i$th coordinate is the cardinality of the inclusion minimal tube of~$\tubing \cup \{\ground\}$ containing the leaf~$\ell_i$ minus~$1$. The set~$\set{\b{x}(\tubing)}{\tubing \text{ maximal tubing on } \starG_{n+1}}$ is the orbit under permutation coordinates of the set~$\set{\sum_{i > k} i \, \b{e}_i}{0 \le k \le n}$. On the other hand, for a tube~$\tube$ of~$\starG_{n+1}$ containing the central vertex~$*$, we observed earlier that the compatibility vector of~$\tube$ with respect to~$\tubing^\circ$ is the characteristic vector of the leaves of~$\starG_{n+1}$ not contained in~$\tube$. Let~${f(k) \eqdef \sum_{j = k}^{n} j = \frac{1}{2}(n+k)(n+1-k)}$ and define a half-space~$\HS(\tube)$ of~$\R^n$ by
\[
\HS(\tube) \eqdef \biggset{\b{x} \in \R^n}{\dotprod{\compatibilityVector{\tubing^\circ}{\tube}}{\b{x}} \le f(|\tube|)} = \biggset{\b{x} \in \R^n}{\sum_{\substack{i \in [n] \\ \ell_i \in \tube}} x_i \le f(|\tube|)}.
\]
Finally, for the tubes of the initial tubing~$\tubing^\circ$, we define
\[
\HS(\{\ell_i\}) \eqdef \set{\b{x} \in \R^n}{\dotprod{\compatibilityVector{\tubing^\circ}{\{\ell_i\}}}{\b{x}} \le 0} = \set{\b{x} \in \R^n}{x_i \ge 0}.
\]
\begin{proposition}
\label{prop:polytopeStar}
The compatibility fan~$\compatibilityFan{\starG_{n+1}}{\tubing^\circ}$ for the initial tubing~$\tubing^\circ \eqdef \big\{ \{\ell_1\}, \dots, \{\ell_n\} \big\}$ whose tubes are the $n$ leaves~of~$\starG_{n+1}$ is the normal fan of the $n$-dimensional simple polytope defined equivalently as
\begin{itemize}
\item the convex hull of the points~$\b{x}(\tubing)$ for all maximal tubings~$\tubing$ on~$\starG_{n+1}$, or
\item the intersection of the half-spaces~$\HS(\tube)$ for all tubes~$\tube$ of~$\starG_{n+1}$.
\end{itemize}
\end{proposition}
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.36]{stellohedron4bisLabeled}\quad\includegraphics[scale=.36]{stellohedron4terLabeled}}
\caption{Two Schlegel diagrams for the $4$-dimensional stellohedron defined in Proposition~\ref{prop:polytopeStar}. The red facet corresponds to all tubings containing the tube~$\{*\}$ while the blue facet corresponds to all tubings containing the tube~$\{\ell\}$, where~$\ell$ is the bottom left leaf of~$\starG_5$.}
\label{fig:stellohedron4}
\end{figure}
The proof of this statement is given in Section~\ref{subsec:proofPolytopality}. As an illustration, the $3$-dimensional stellohedron defined in Proposition~\ref{prop:polytopeStar} is represented in \fref{fig:stellohedron3}\,(right). \fref{fig:stellohedron4} represents two Schlegel diagrams (see~\cite[Lecture~5]{Ziegler} for definition) for the $4$-dimensional stellohedron defined in Proposition~\ref{prop:polytopeStar}. In both pictures, we have distinguished two particular facets:
\begin{itemize}
\item The red facet corresponds to all tubings containing the tube~$\{*\}$. Since the reconnected complement of~$\{*\}$ in~$\starG_5$ is the complete graph~$\completeG_4$, it has the combinatorics of the permutahedron. In fact, by definition of our polytopal realization, this facet is the classical permutahedron, obtained as the convex hull of the orbit of~$\sum_{i \in [4]} i \, \b{e}_i$ under permutation of the coordinates.
\item The blue facet corresponds to all tubings containing the tube~$\{\ell\}$, where~$\ell$ is the bottom left leaf of~$\starG_5$. This facet contains the initial tubing~$\tubing^\circ$ at the back. Note that there are $4$ isometric facets to this blue facet, corresponding to the four leaves of~$\starG_5$. This is visible in \fref{fig:stellohedron4}\,(right).
\end{itemize}
The blue (resp.~red) facet is the projection facet on the left (resp.~right) picture.
\begin{remark}
To conclude, observe that we could have replaced the function~$f$ in the definition of the half-spaces~$\HS(\tube)$ by any concave function. This follows from Proposition~\ref{prop:polytopalityFan} since the linear dependence between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$ is given for any adjacent maximal tubings~$\tubing, \tubing'$ on~$\starG_{n+1}$ distinct from~$\tubing^\circ$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ by
\[
\compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} = \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \compatibilityVector{\tubing^\circ}{\overline{\tube}}
\]
where~$\overline{\tube} \eqdef \tube \cap \tube'$ and~$\underline{\tube} \eqdef \tube \cap \tube'$ (which are tubes of~$\starG_{n+1}$). Details are left to the reader.
\end{remark}
\subsection{Design nested complex}
\label{subsec:designNestedComplex}
Generalizing graphical nested complexes, S.~Devadoss, T.~Heath and C.~Vipismakul introduced design nested complexes in~\cite[Section~5]{DevadossHeathVipismakul}. To define these complexes, one considers \defn{design tubes} of~$\graphG$, which are of two types:
\begin{itemize}
\item the \defn{round tubes} are usual tubes of~$\graphG$ (including the connecting components of~$\graphG$),
\item the \defn{square tubes} are just single nodes of~$\graphG$.
\end{itemize}
We denote by~$\squareTube{v}$ the square tube containing~$v$, and still denote round tubes as sets. Two design tubes are \defn{compatible} if
\begin{itemize}
\item they are both round tubes and they are either nested, or disjoint and non-adjacent,
\item or at least one of them is a square tube and they are not nested.
\end{itemize}
The \defn{design nested complex} of~$\graphG$ is the simplicial complex~$\designNestedComplex(\graphG)$ of sets of pairwise compatible design tubes of~$\graphG$. Examples are given in \fref{fig:designGraphAssociahedra}. By definition, the nested complex~$\nestedComplex(\graphG)$ is (isomorphic to) the subcomplex of the design nested complex~$\designNestedComplex(\graphG)$ involving none of the square tubes, or equivalently containing all improper round tubes.
For a design tube~$\tube$ of~$\graphG$, set
\[
\b{g}\design(\tube) \eqdef
\begin{cases}
\sum_{v \in \tube} \b{e}_v & \text{if } \tube \text{ is a round tube,} \\
-\b{e}_v & \text{if } \tube \text{ is the square tube } \{v\}.
\end{cases}
\]
For a tubing~$\tubing$ on~$\graphG$, define~$\b{g}\design(\tubing) \eqdef \set{\b{g}\design(\tube)}{\tube \in \tubing}$. These vectors again support a complete simplicial fan realization of the design nested complex~$\designNestedComplex(\graphG)$.
\begin{theorem}[\cite{DevadossHeathVipismakul}]
For any graph~$\graphG$, the collection of cones
\[\cG\design(\graphG) \eqdef \set{\R_{\ge 0} \, \b{g}\design(\tubing)}{\tubing \text{ tubing on } \graphG}\]
is a complete simplicial fan of~$\R^\ground$, called \defn{design nested fan} of~$\graphG$, which realizes~$\designNestedComplex(\graphG)$.
\end{theorem}
By definition, the $\b{g}\design$ vector of a round tube~$\tube$ is just the characteristic vector of~$\tube$. Therefore, the non-negative part of the design nested fan~$\cG\design(\graphG)$ projects to the nested fan~$\cG(\graphG)$ defined in Section~\ref{subsec:graphAssociahedra}. Using ideas similar to the construction of the graph associahedra~$\Asso(\graphG)$ realizing~$\cG(\graphG)$, S.~Devadoss, T.~Heath and C.~Vipismakul prove that~$\cG\design(\graphG)$ is as well polytopal.
\begin{theorem}[\cite{DevadossHeathVipismakul}]
The design nested fan is the normal fan of a polytope obtained from the cube by iterated face truncations.
\end{theorem}
\begin{figure}
\capstart
\centerline{\includegraphics[scale=1.3]{designGraphAssociahedra}}
\caption{Two design graph associahedra: the design $\pathG_3$-associahedron (left) and the design $\completeG_3$-associahedron (right). For readability, round tubes are colored blue while square tubes are colored red.}
\label{fig:designGraphAssociahedra}
\end{figure}
We denote this polytope by~$\Asso\design(\graphG)$ and call it \defn{design graph associahedron} (although it is called \emph{graph cubeahedron} in~\cite{DevadossHeathVipismakul}). Observe that the face of~$\Asso\design(\graphG)$ corresponding to the tubing formed by the connected components of~$\graphG$ coincides with the graph associahedron~$\Asso(\graphG)$. \fref{fig:designGraphAssociahedra} illustrates the design $\pathG_3$-associahedron and the design $\completeG_3$-associahedron. In this figure, the attentive reader should recognize different standard graph associahedra: on the one hand, the front $2$-dimensional faces of~$\Asso\design(\pathG_3)$ and~$\Asso\design(\completeG_3)$ are respectively~$\Asso(\pathG_3)$ and~$\Asso(\completeG_3)$, and on the other hand, the design graph associahedra~$\Asso\design(\pathG_3)$ and~$\Asso\design(\completeG_3)$ themselves turn out to coincide respectively with~$\Asso(\pathG_4)$ and~$\Asso(\starG_4)$ (see Example~\ref{exm:isomorphismDesignNormal}). Starting from dimension~$4$, most design graph associahedra are not standard graph associahedra (see Proposition~\ref{prop:isomorphismDesignNormal}). Figures~\ref{fig:designCycleAssociahedron} and~\ref{fig:designStarAssociahedron} represent two Schlegel diagrams (see~\cite[Lecture~5]{Ziegler} for definition) for the $4$-dimensional design cycle associahedron~$\Asso\design(\cycleG_4)$ and design star associahedron~$\Asso\design(\starG_4)$.
\begin{figure}[p]
\capstart
\centerline{\includegraphics[scale=.36]{designCycleAssociahedron}\quad\raisebox{.3cm}{\includegraphics[scale=.36]{designCycleAssociahedronbis}}}
\caption{Two Schlegel diagrams for the $4$-dimensional design cycle associahedron. The blue facet corresponds to all round tubings while the red facet corresponds to all tubings containing the bottom right square tube of~$\cycleG_4$.}
\label{fig:designCycleAssociahedron}
\vspace{.5cm}
\end{figure}
\begin{figure}[p]
\capstart
\centerline{\raisebox{.4cm}{\includegraphics[scale=.36]{designStarAssociahedron}}\quad\includegraphics[scale=.36]{designStarAssociahedronbis}}
\caption{Two Schlegel diagrams for the $4$-dimensional design star associahedron. The blue facet corresponds to all round tubings while the red facet corresponds to all tubings containing the right square tube of~$\starG_4$.}
\label{fig:designStarAssociahedron}
\end{figure}
This section aims at showing that our compatibility fan construction extends to the design nested complex. That is, we produce a complete simplicial fan realizing the design nested complex from any initial maximal design tubing. Interestingly, we will see in Remark~\ref{rem:connectionConstructions} that the compatibility fan associated to the specific maximal design tubing consisting of all square tubes coincides with the design nested fan. This provides a relevant connection between our construction and the classical constructions~\cite{CarrDevadoss, FeichtnerSturmfels, Devadoss, Zelevinsky} for graph associahedra.
Observe that a square tube is compatible with all design tubes not containing it and exchangeable with all design tubes containing it. Definition~\ref{def:compatibilityDegree} of compatibility degree thus naturally extends on all pairs of design tubes as follows.
\begin{definition}
For two design tubes~$\tube, \tube'$ of~$\graphG$, the \defn{compatibility degree} of~$\tube$ with~$\tube'$ is
\[
\compatibilityDegree{\tube}{\tube'} =
\begin{cases}
-1 & \text{if } \tube = \tube', \\
1 & \text{if } \tube \text{ and } \tube' \text{ are nested and exactly one of them is square,} \\
|\{\text{neighbors of $\tube$ in } \tube' \ssm \tube\}| & \text{if } \tube \text{ and } \tube' \text{ are round and } \tube \not\subseteq \tube', \\
0 & \text{otherwise.}
\end{cases}
\]
\end{definition}
By construction, this compatibility degree still satisfies the conclusions of Proposition~\ref{prop:compatibilityDegree} and thus measures the incompatibility between design tubes. We then define as usual the \defn{compatibility vector} of a design tube~$\tube$ of~$\graphG$ with respect to an initial maximal design tubing~$\tubing^\circ \eqdef \{\tube_1^\circ, \dots, \tube_n^\circ\}$ as the integer vector $\compatibilityVector{\tubing^\circ}{\tube} \eqdef [\compatibilityDegree{\tube_1^\circ}{\tube}, \dots, \compatibilityDegree{\tube_n^\circ}{\tube}]$ and the \defn{compatibility matrix} of a design tubing~$\tubing \eqdef \{\tube_1, \dots, \tube_m\}$ on~$\graphG$ with respect to~$\tubing^\circ$ as the matrix $\compatibilityVector{\tubing^\circ}{\tubing} \eqdef [\compatibilityDegree{\tube_i^\circ}{\tube_j}]_{i \in [n], j \in [m]}$. We extend Theorem~\ref{theo:compatibilityFan} in the following statement, whose proof is sketched in Section~\ref{subsec:proofDesignCompatibilityFan}.
\begin{theorem}
\label{theo:compatibilityFanDesign}
For any graph~$\graphG$ and any maximal design tubing~$\tubing^\circ$ on~$\graphG$, the collection of cones
\[
\designCompatibilityFan{\graphG}{\tubing^\circ} \eqdef \set{\R_{\ge 0} \, \compatibilityVector{\tubing^\circ}{\tubing}}{\tubing \text{ design tubing on } \graphG}
\]
is a complete simplicial fan which realizes the design nested complex~$\designNestedComplex(\graphG)$. We call it the \defn{design compatibility fan} of~$\graphG$ with respect to~$\tubing^\circ$.
\end{theorem}
\enlargethispage{-.5cm}
Using the same duality trick as in the proof of Theorem~\ref{theo:dualCompatibilityFan}, the reader can obtain as well dual design compatibility fans.
\bigskip
Concerning isomorphisms, we have similar results as in Section~\ref{subsec:many}. Notice first that the conclusions of Proposition~\ref{prop:nestedComplexIsomorphismDisconnected} still hold for design nested complexes. We can thus restrict our discussion to design nested complexes of connected graphs. We first compare design and standard nested complexes, starting with the following examples.
\begin{example}
\label{exm:isomorphismDesignNormal}
The reader can check that:
\begin{enumerate}[(i)]
\item The design nested complex~$\designNestedComplex(\completeG_n)$ is isomorphic to the standard nested complex~$\nestedComplex(\starG_{n+1})$. A natural isomorphism sends a square tube~$\squareTube{v}$ of~$\completeG_n$ to the tube~$\{v\}$ of~$\starG_{n+1}$, and a round tube~$\tube$ of~$\completeG_n$ to the tube~$\{*\} \cup ([n] \ssm \tube)$ of~$\starG_{n+1}$ (where~$*$ is the central vertex). See \fref{fig:isomorphismDesignCompleteStar}.
\begin{figure}
\capstart
\centerline{\includegraphics[width=\textwidth]{isomorphismDesignCompleteStar}}
\caption{An isomorphism from the design nested complex of a complete graph with~$5$ vertices (left) to the nested complex of a star with~$6$ vertices~(right).}
\label{fig:isomorphismDesignCompleteStar}
\end{figure}
\item The design nested complex~$\designNestedComplex(\pathG_n)$ is isomorphic to the standard nested complex~$\nestedComplex(\pathG_{n+1})$. A natural isomorphism sends a square tube~$\squareTube{v}$ of~$\pathG_n$ to the tube~$\{v+1, \dots, n+1\}$ of~$\pathG_{n+1}$, and a round tube~$\tube$ of~$\pathG_n$ to the tube~$\tube$ of~$\pathG_{n+1}$. See \fref{fig:isomorphismDesignPath}. We denote this isomorphism by~$\Pi : \designNestedComplex(\pathG_n) \to \nestedComplex(\pathG_{n+1})$.
\begin{figure}
\capstart
\centerline{\includegraphics[width=\textwidth]{isomorphismDesignPath}}
\caption{An isomorphism from the design nested complex of a path with~$5$ vertices (top) to the nested complex of a path with~$6$ vertices (bottom).}
\label{fig:isomorphismDesignPath}
\end{figure}
\end{enumerate}
\end{example}
We now describe a generalization of both cases of Example~\ref{exm:isomorphismDesignNormal}. We consider the spiders and octopuses defined in Section~\ref{subsec:many}: for~${\underline{n} \eqdef \{n_1, \dots, n_\ell\}} \in \N^\ell$ with~$n = \sum_{i \in [\ell]} (n_i + 1)$,
\begin{itemize}
\item the spider~$\spiderG_{\underline{n}}$ has~$\ell$ legs~$[v_1^i, v_{n_i}^i]$ attached to its body formed by a complete graph~on~$\{v_0^i\}_{i \in [\ell]}$,
\item the octopus~$\octopusG_{\underline{n}}$ has~$\ell$ legs~$[v_0^i, v_{n_i}^i]$ attached to its head formed by a single vertex~$*$.
\end{itemize}
We now define an isomorphism~$\bar\Omega$ from the design nested complex~$\designNestedComplex(\spiderG_{\underline{n}})$ of the spider~$\spiderG_{\underline{n}}$ to the nested complex~$\nestedComplex(\octopusG_{\underline{n}})$ of the octopus~$\octopusG_{\underline{n}}$. We distinguish three kinds of tubes of~$\designNestedComplex(\spiderG_{\underline{n}})$:
\begin{description}
\item[Square tubes] for~$ i \in [\ell]$ and~$j \in[0,n_i]$, \;\;\;\, $\bar\Omega(\squareTube{v^i_j}) \eqdef \big[ v^i_0, v^i_{n_i-j} \big]$,
\item[Leg tubes] for~$i \in [\ell]$ and~$1 \le j \le k \le n_i$, \; $\bar\Omega\big( \big[ v^i_j, v^i_k \big] \big) \eqdef \big[ v^i_{n_i+1-k}, v^i_{n_i+1-j} \big]$,
\item[Body tubes] for~$-1 \le k_i \le n_i$ ($i \in [\ell]$), \quad\; $\bar\Omega\big( \bigcup\limits_{i \in [\ell]} \big[ v^i_0, v^i_{k_i} \big] \big) \eqdef \{\ast\} \cup \bigcup\limits_{i \in [\ell]} \big[ v^i_0, v^i_{n_i-1-k_i} \big]$.
\end{description}
\fref{fig:isomorphismDesignSpiderOctopus} illustrates the map~$\bar\Omega$ on different tubes of the spider~$\spiderG_{\{0,3,2,3,0,3,2,3\}}$.
Observe that~$\bar\Omega$ indeed generalizes both isomorphisms of Example~\ref{exm:isomorphismDesignNormal}:
\begin{enumerate}[(i)]
\item We have~$\completeG_n = \spiderG_{\{0\}^n}$ while~$\starG_{n+1} = \octopusG_{\{0\}^n}$, and the isomorphism~$\bar\Omega : \designNestedComplex(\spiderG_{\{0\}^n}) \to \nestedComplex(\octopusG_{\{0\}^n})$ coincides with the isomorphism of Example~\ref{exm:isomorphismDesignNormal}\,(i).
\item We have~$\pathG_n = \spiderG_{\{n\}}$ while~$\pathG_{n+1} = \octopusG_{\{n\}}$, and the isomorphism~$\bar\Omega : \designNestedComplex(\spiderG_{\{n\}}) \to \nestedComplex(\octopusG_{\{n\}})$ coincides with the isomorphism~$\Pi$ of Example~\ref{exm:isomorphismDesignNormal}\,(ii) up to the automorphisms~$\rot$ and~$\rev$ of the nested complex~$\nestedComplex(\octopusG_{\{n\}})$ described in Example~\ref{exm:nestedComplexIsomorphisms}\,(ii). More precisely if we consider the leftmost vertex of~$\pathG_n$ as the body of~$\spiderG_{\{n\}}$ and the leftmost vertex of~$\pathG_{n+1}$ as the head of~$\octopusG_{\{n\}}$, then one can check that~${\rot} \circ {\bar\Omega} = {\rev} \circ {\Pi}$.
\end{enumerate}
\begin{figure}
\capstart
\centerline{\includegraphics[width=\textwidth]{isomorphismDesignSpiderOctopus}}
\caption{An isomorphism from the design nested complex of the spider $\spiderG_{\{0,3,2,3,0,3,2,3\}}$ (left) to the nested complex of the octopus~$\octopusG_{\{0,3,2,3,0,3,2,3\}}$~(right).}
\label{fig:isomorphismDesignSpiderOctopus}
\end{figure}
The following statement is left to the reader. The proof is similar to that of Proposition~\ref{prop:exmAutomorphism}.
\begin{proposition}
The map~$\bar\Omega$ is an isomorphism from the design nested complex~$\nestedComplex(\spiderG_{\underline{n}})$ of the spider~$\spiderG_{\underline{n}}$ to the nested complex~$\nestedComplex(\octopusG_{\underline{n}})$ of the octopus~$\octopusG_{\underline{n}}$ which dualizes the compatibility degree: $\compatibilityDegree{\bar\Omega(\tube)}{\bar\Omega(\tube')} = \compatibilityDegree{\tube'}{\tube}$.
\end{proposition}
The following proposition, proved in Section~\ref{subsec:proofDesignIsomorphisms}, states that~$\bar\Omega$ is essentially the only isomorphism between design and standard nested complexes.
\begin{proposition}
\label{prop:isomorphismDesignNormal}
Let~$\bar\graphG$ and~$\graphG$ be two connected graphs and~$\Phi : \designNestedComplex(\bar\graphG) \to \nestedComplex(\graphG)$ be a simplicial complex isomorphism. Then~$\bar\graphG$ is a spider~$\spiderG_{\underline{n}}$ while~$\graphG$ is an octopus~$\octopusG_{\underline{n}}$ and~$\Phi$ coincides with~$\bar\Omega$ up to composition with a nested complex automorphism of~$\nestedComplex(\graphG)$ (described in Theorem~\ref{theo:nestedComplexIsomorphisms}).
\end{proposition}
We now classify combinatorial isomorphisms of design nested complexes. As for nested complexes, a graph isomorphism~$\phi : \graphG \to \graphG'$ induces a \defn{trivial} design nested complex isomorphism ${\Phi : \designNestedComplex(\graphG) \to \designNestedComplex(\graphG')}$, defined by~$\Psi(\tube) \eqdef \set{\psi(v)}{v \in \tube}$ and~$\Psi(\squareTube{v}) \eqdef \squareTube{\psi(v)}$, which preserves compatibility degrees. We again focus on non-trivial design nested complex isomorphisms. We first underline two examples.
\begin{example}
\label{exm:isomorphismDesignNestedComplexes}
The reader can check that:
\begin{enumerate}[(i)]
\item For the star~$\starG_n$ with central vertex~$*$, the map which preserves the square tube~$\squareTube{*}$, exchanges the square tube~$\squareTube{v}$ with the round tube~$\{v\}$, and exchanges any other round tube~$\tube$ with the round tube~$([n] \ssm \tube) \cup \{*\}$, is a non-trivial design nested complex isomorphism of~$\designNestedComplex(\starG_n)$.
\item For~$p \in [n+3]$, the conjugation~${\Pi \, \circ \! \rot^p \! \circ \, \Pi^{-1}}$ of the automorphism ${\rot^p \; : \nestedComplex(\pathG_{n+1}) \to \nestedComplex(\pathG_{n+1})}$ of Example~\ref{exm:nestedComplexIsomorphisms}\,(ii) by the isomorphism~$\Pi : \designNestedComplex(\pathG_n) \to \nestedComplex(\pathG_{n+1})$ of Example~\ref{exm:isomorphismDesignNormal}\,(ii) is a non-trivial design nested complex isomorphism of~$\designNestedComplex(\pathG_n)$.
\end{enumerate}
Note that these two isomorphisms send some square tubes to round tubes and thus are non-trivial.
\end{example}
We now describe a generalization of both cases of Example~\ref{exm:isomorphismDesignNestedComplexes}. Namely, we define an automorphism~$\Omega\design$ of the design nested complex~$\designNestedComplex(\octopusG_{\underline{n}})$ of the octopus~$\octopusG_{\underline{n}}$ for any~$\underline{n} \eqdef \{n1, \dots, n_\ell\} \in \N^\ell$ with~$n = \sum_{i \in [\ell]} (n_i + 1)$ as follows:
\begin{description}
\item[Square tubes] for~$i \in [\ell]$ and~$j \in [0, n_i]$, \qquad $\Omega\design(\squareTube{v^i_j}) \eqdef \big[ v^i_0, v^i_{n_i-j} \big]$ \quad and \quad $\Omega\design(\squareTube{\ast}) \eqdef \squareTube{\ast}$,
\item[Leg tubes]
\begin{tabular}[t]{@{}ll}
for~$i \in [\ell]$ and~$0 \le k \le n_i$, & \; $\Omega\design \big( \big[ v^i_0, v^i_k \big] \big) \eqdef \squareTube{v^i_{n_i-k}}$, \\
for~$i \in [\ell]$ and~$1 \le j \le k \le n_i$, & \; $\Omega\design \big( \big[ v^i_j, v^i_k \big] \big) \eqdef \big[ v^i_{n_i+1-k}, v^i_{n_i+1-j} \big]$,
\end{tabular}
\item[Head tubes] for~$-1 \le k_i \le n_i$ ($i \in [\ell]$), \quad $\Omega\design\big( \{\ast\} \cup \bigcup\limits_{i \in [\ell]} \big[ v^i_0, v^i_{k_i} \big] \big) \eqdef \{\ast\} \cup \bigcup\limits_{i \in [\ell]} \big[ v^i_0, v^i_{n_i-1-k_i} \big]$.
\end{description}
\fref{fig:isomorphismDesignOctopus} illustrates the map~$\Omega\design$ on different tubes of the octopus~$\octopusG_{\{0,3,2,3,0,3,2,3\}}$.
\begin{figure}[t]
\capstart
\centerline{\includegraphics[scale=.9]{isomorphismDesignOctopus}}
\caption{The octopus~$\spiderG_{\{0,3,2,3,0,3,2,3\}}$ and examples of the action of the non-trivial design nested complex isomorphism~$\Omega\design$: the design tubing~$\tubing$ on the left is sent to the design tubing~$\Omega\design(\tubing)$ on the right.}
\label{fig:isomorphismDesignOctopus}
\end{figure}
The reader is invited to check that~$\Omega\design$ indeed generalizes the non-trivial automorphisms of Example~\ref{exm:isomorphismDesignNestedComplexes}. The following statement is also left to the reader. The proof is similar to that of Proposition~\ref{prop:exmAutomorphism}.
\begin{proposition}
\label{prop:exmAutomorphismDesign}
The map~$\Omega\design$ is a non-trivial involutive automorphism of the design nested complex~$\nestedComplex(\octopusG_{\underline{n}})$ of the octopus~$\octopusG_{\underline{n}}$ which dualizes the compatibility degree: $\compatibilityDegree{\Omega\design(\tube)}{\Omega\design(\tube')} = \compatibilityDegree{\tube'}{\tube}$.
\end{proposition}
\enlargethispage{.5cm}
The following theorem, proved in Section~\ref{subsec:proofDesignIsomorphisms}, states that~$\Omega\design$ is essentially the only non-trivial design nested complex isomorphism.
\begin{theorem}
\label{theo:designNestedComplexIsomorphisms}
Let~$\graphG$ and~$\graphG'$ be two connected graphs and~$\Phi : \designNestedComplex(\graphG) \to \designNestedComplex(\graphG')$ be a non-trivial design nested complex isomorphism. Then~$\graphG$ and~$\graphG'$ are octopuses and there exists a graph isomorphism~$\psi : \graphG \to \graphG'$ which induces a design nested complex isomorphism~${\Psi : \designNestedComplex(\graphG) \to \designNestedComplex(\graphG')}$ (defined by~$\Psi(\tube) \eqdef \set{\psi(v)}{v \in \tube}$ and~$\Psi(\squareTube{v}) \eqdef \squareTube{\psi(v)}$) such that the composition~$\Psi^{-1} \circ \Phi$ coincides with the non-trivial design nested complex automorphism~$\Omega\design$ on~$\designNestedComplex(\graphG)$.
\end{theorem}
We conclude that we get many design compatibility fans (up to linear isomorphism).
\begin{corollary}
\label{coro:classificationDesignPrimal}
If a connected graph~$\graphG$ is not an octopus, then the number of linear isomorphism classes of primal (resp.~dual) design compatibility fans of~$\graphG$ is the number of orbits of maximal design tubings on~$\graphG$ under graph automorphisms of~$\graphG$.
\end{corollary}
Finally, as for the compatibility fan, we do not know in general whether the design compatibility fan is or not polytopal. Nevertheless, we know it for the following initial tubing.
\begin{remark}
\label{rem:connectionConstructions}
Consider the maximal tubing~$\tubing\design$ of a graph~$\graphG$ consisting of all square tubes of~$\graphG$. By definition of the compatibility degree with square tubes, the compatibility vector of a square tube~$\{v\}$ is just~$-\b{e}_v$ while the compatibility vector of a round tube~$\tube$ with respect to~$\tubing\design$ is just the characteristic vector of~$\tube$. Therefore, the design compatibility fan~$\designCompatibilityFan{\graphG}{\tubing\design}$ coincides with the design nested fan~$\cG\design(\graphG)$. It follows that~$\designCompatibilityFan{\graphG}{\tubing\design}$ is the normal fan of the design graph associahedron~$\Asso\design(\graphG)$ for any graph~$\graphG$. In fact, as the linear dependencies among the compatibility vectors with respect to the initial tubing~$\tubing\design$ are simply described, a direct computation shows that the map~$\omega$ defined by~$\omega(\tube) = 3^{|\ground|}|\tube| - 3^{|\tube|}$ for a round tube and~$\omega(\squareTube{v}) = C$ for a sufficiently large constant~$C$ provides a suitable weight function for Condition~(2) of Proposition~\ref{prop:polytopalityFan}.
\end{remark}
\subsection{Laurent Phenomenon algebras}
\label{subsec:LPA}
To conclude, we discuss our construction with respect to the framework of \defn{Laurent Phenomenon algebras} developed by T.~Lam and P.~Pylyavskyy in~\cite{LamPylyavskyy-LaurentPhenomenonAlgebras, LamPylyavskyy-LinearLaurentPhenomenonAlgebras}. The definition of these algebras is similar to that of cluster algebras but allows more flexibility for the variable mutation. Namely, the cluster variables of these algebras are still obtained by a mutation process from an initial cluster~$X^\circ$, but the mutated variable depends upon some mutation polynomials which are themselves updated along the mutations. We refer to~\cite{LamPylyavskyy-LaurentPhenomenonAlgebras} for the precise rules directing these variable and polynomial mutations.
Since the construction is more flexible, it yields more general algebras and cluster complexes. To our knowledge, there is no classification of the finite type Laurent Phenomenon algebras, in contrast to the case of classical cluster algebras. However, an important subclass is the class of linear Laurent Phenomenon algebras discussed in~\cite{LamPylyavskyy-LinearLaurentPhenomenonAlgebras}, where the mutation polynomials of a given cluster are all linear. A striking fact is that all graphical nested complexes appear as cluster complexes for some linear Laurent Phenomenon algebras. Namely, T.~Lam and P.~Pylyavskyy associate to each graph~$\graphG$ a Laurent Phenomenon algebra~$\cA(\graphG)$ whose cluster complex is isomorphic to the design nested complex~$\designNestedComplex(\graphG)$ discussed in Section~\ref{subsec:designNestedComplex}. The nested complex~$\nestedComplex(\graphG)$ is thus isomorphic to the cluster complex of the linear Laurent Phenomenon algebra obtained by freezing in~$\cA(\graphG)$ the variables corresponding to square tubes.
As the name suggests, Laurent Phenomenon algebras still exhibits the Laurent Phenomenon. Each cluster variable~$x$ can be expressed as a Laurent polynomial in terms of the cluster variables~$x_1^\circ, \dots, x_n^\circ$ of the initial cluster~$X^\circ$. Define the \defn{$\b{d}$-vector} of~$x$ with respect to~$X^\circ$ as the vector~$\b{d}(X^\circ, x)$ whose $i$th coordinate is the exponent of the initial variable~$x_i^\circ$ in the denominator of~$x$. It is then tempting to extend the construction of the $\b{d}$-vector fan of S.~Fomin and A.~Zelevinsky~\cite{FominZelevinsky-ClusterAlgebrasII} to all Laurent Phenomenon algebras:
\begin{question}
\label{qu:LPA}
Given a Laurent Phenomenon algebra~$\cA$ and an initial cluster seed~$X^\circ$ of~$\cA$, do the cones generated by the $\b{d}$-vectors with respect to~$X^\circ$ of all collections of compatible cluster variables in~$\cA$ always form a complete simplicial fan realizing the cluster complex of~$\cA$?
\end{question}
When studying this question, we encountered several difficulties that we want to underline here:
\begin{itemize}
\item Although similar to the problem addressed in this paper, we have not been able to prove or disprove Question~\ref{qu:LPA} even for the specific case of linear Laurent Phenomenon algebras arising from nested complexes of graphs. In contrast to our combinatorial definition of compatibility degree among tubes, it is difficult to track the denominator vectors along the mutation process. In particular, expressing explicitly the linear dependence among the denominator vectors of the cluster variables involved in a mutation seems out of reach at the moment.
\item One of the main difficulties is that, even in some finite cases, an entry of the denominator vector of a variable can depend on the whole initial cluster, not only on the variable corresponding to that entry. Namely, in certain linear Laurent Phenomenon algebras, one can find cluster variables~$x,x^\circ$ and initial clusters~$X^\circ, X'^\circ$ both containing~$x^\circ$ such that the exponent of~$x^\circ$ in the denominator of~$x$ is different when computed with respect to~$X^\circ$ or with respect to~$X'^\circ$. In this sense, the denominator vectors are not anymore compatibility vectors arising from a compatibility degree between cluster variables.
\item Even worst, given an initial cluster~$X^\circ$, an initial variable~$x^\circ$ and a cluster variable~$x$ not in~$X^\circ$ but compatible with~$x^\circ$, the exponent of~$x^\circ$ in the denominator~$x$, when expanded in the variables of~$X^\circ$ can be nonzero.
\item In particular, the denominator vectors in the linear Laurent Phenomenon algebra~$\cA(\graphG)$ do not always coincide with our compatibility vectors on tubes of~$\graphG$.
\end{itemize}
Despite all these difficulties, we hope that the study of the linear Laurent Phenomenon algebras arising from (design) nested complexes of graphs can help to answer Question~\ref{qu:LPA}. Since non-isomorphic Laurent Phenomenon algebra may have isomorphic cluster complexes (the cyclohedron is an example), another simpler question would be to look for other Laurent Phenomenon algebras whose denominators interpret our compatibility degrees.
\begin{question}
For which graph~$\graphG$ does there exist a Laurent Phenomenon algebra~$\cA(\graphG)$ whose cluster complex is isomorphic to the nested complex~$\Asso(\graphG)$ (resp.~to the design nested complex~$\Asso\design(\graphG)$) and whose denominators are given by the (primal or dual) compatibility degrees defined in this paper?
\end{question}
\section{Proofs}
\label{sec:proofs}
This section contains all proofs of our results. Some of them require additional technical steps, which motivated us to separate them from the rest of the paper. We also hope that the many examples treated in Section~\ref{sec:specificGraphs} help the reader's intuition throughout these proofs.
\subsection{Compatibility degree (Proposition~\ref{prop:compatibilityDegree})}
\label{subsec:proofCompatibilityDegree}
We start with the proof that our graphical compatibility degree encodes compatibility and exchangeability between tubes. We show the three points of Proposition~\ref{prop:compatibilityDegree}:
\medskip
\noindent$\bullet$ $\compatibilityDegree{\tube}{\tube'} < 0 \iff \compatibilityDegree{\tube'}{\tube} < 0 \iff \tube = \tube'$. \\
This is immediate from the definition since the compatibility degree between two distinct tubes is either a cardinal or~$0$.
\medskip
\noindent$\bullet$ $\compatibilityDegree{\tube}{\tube'} = 0 \iff \compatibilityDegree{\tube'}{\tube} = 0 \iff \tube$ and~$\tube'$ are compatible. \\
Consider two distinct tubes~$\tube, \tube'$ of~$\graphG$. If they are compatible, then either~$\tube \subseteq \tube'$, or~$\tube' \subseteq \tube$, or $\tube$ and~$\tube'$ are non-adjacent. In the first case,~$\compatibilityDegree{\tube}{\tube'} = 0$ by the last line of Definition~\ref{def:compatibilityDegree} of the compatibility degree. In the last two cases, $\compatibilityDegree{\tube}{\tube'} = |\{\text{neighbors of $\tube$ in } \tube' \ssm \tube\}| = |\varnothing| = 0$. Conversely, if~$\compatibilityDegree{\tube}{\tube'} = 0$, then either~$\tube$ has no neighbor in~$\tube'$, or~$\tube' \subseteq \tube$, or~$\tube \subseteq \tube'$, so that the two tubes are compatible.
\medskip
\noindent$\bullet$ $\compatibilityDegree{\tube}{\tube'} = 1 = \compatibilityDegree{\tube'}{\tube} \iff \tube$ and~$\tube'$ are exchangeable. \\
The $\Leftarrow$ part follows from the explicit flip description in Proposition~\ref{prop:flip}. Indeed, assume that~$\tube$ and~$\tube'$ are exchangeable, let~$\overline{\tube} \eqdef \tube \cup \tube'$, and let~$\tubing, \tubing'$ be two adjacent maximal tubings on~$\graphG$ such that~${\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}}$. Since $\tube'$ is the connected component of~$\graphG{}[\overline{\tube} \ssm \lab(\tube, \tubing)]$ containing~$\lab(\overline{\tube}, \tubing)$, the root~$\lab(\overline{\tube}, \tubing)$ is the unique neighbor of~$\tube$ in~$\tube' \ssm \tube$. Therefore, $\compatibilityDegree{\tube}{\tube'} = 1$, and~$\compatibilityDegree{\tube'}{\tube} = 1$ by symmetry.
Assume conversely that~$\compatibilityDegree{\tube}{\tube'} = 1 = \compatibilityDegree{\tube'}{\tube}$. Since~$\compatibilityDegree{\tube'}{\tube} = 1$, there exists a unique neighbor~$r$ of~$\tube'$ in~$\tube \ssm \tube'$. Similarly, there exists a unique neighbor~$r'$ of~$\tube$ in~$\tube' \ssm \tube$. We want to find two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. We start with the forced tubes (see the end of Section~\ref{subsec:nestedComplex}): we define~$\overline{\tube} \eqdef \tube \cup \tube'$ and we let~$\tube[s]_1, \dots, \tube[s]_\ell$ be the connected components of~$\overline{\tube} \ssm \{r,r'\}$. We choose an arbitrary maximal tubing~$\tubing[S]_i$ on~$\graphG{}[\tube[s]_i]$ for each~$i$, and an arbitrary maximal tubing~$\tubing[S]$ on~$\graphG$ containing~$\overline{\tube}$. The set of tubes
\[
\tubing[R] \eqdef \{\overline{\tube}, \tube[s]_1, \dots, \tube[s]_\ell\} \; \sqcup \; \tubing[S]_1 \! \sqcup \dots \sqcup \! \tubing[S]_\ell \; \sqcup \; \set{\tube[s]}{\tube[s] \in \tubing[S], \tube[s] \not\subseteq \overline{\tube}}.
\]
is clearly a tubing, and is compatible with both~$\tube$ and~$\tube'$. We now compute the cardinality of~$\tubing[R]$. Observe first that $|\tubing[S]| = |\ground|-|\connectedComponents(\graphG)|$ so that~$|\set{\tube[s]}{\tube[s] \in \tubing[S], \tube[s] \not\subseteq \overline{\tube}}| = |\ground|-|\connectedComponents(\graphG)|-|\overline{\tube}|$ since~$\overline{\tube}$ is a tube of~$\graphG$. Moreover, $|\tubing[S]_i| = |\tube[s]_i| - 1$ since~$\tube[s]_i$ is a tube, and~$\sum_i |\tube[s]_i| = |\bigcup_i \tube[s]_i| = |\overline{\tube} \ssm \{r,r'\}| = |\overline{\tube}| - 2$. We conclude that
\[
|\tubing[R]| = (1 + \ell) + (|\overline{\tube}| - 2 - \ell) + (|\ground|-|\connectedComponents(\graphG)|-|\overline{\tube}|) = |\ground| - |\connectedComponents(\graphG)| - 1
\]
Therefore,~$\tubing[R]$ is a ridge of the nested complex~$\nestedComplex(\graphG)$, so that~$\tubing \eqdef \tubing[R] \cup \{\tube\}$ and~$\tubing' \eqdef \tubing[R] \cup \{\tube'\}$ are maximal tubings related by the flip of~$\tube$ into~$\tube'$. \qed
\subsection{Restriction on coordinate hyperplanes (Proposition~\ref{prop:restriction})}
\label{subsec:proofLemRestriction}
We now state two lemmas needed in the proofs of Theorem~\ref{theo:compatibilityFan}. They have essentially the same content as Proposition~\ref{prop:restriction}, except that they focus on compatibility vectors (rays) and not on the other cones of the compatibility fan since Theorem~\ref{theo:compatibilityFan} is not proved yet. In particular, they will imply Proposition~\ref{prop:restriction} once Theorem~\ref{theo:compatibilityFan} will be established.
Remember that for a tube~$\tube^\circ$ of~$\graphG$, we denote by~$\graphG{}[\tube^\circ]$ the restriction of~$\graphG$ to~$\tube^\circ$ and by~$\graphG{}^\star\tube^\circ$ the reconnected complement of~$\tube^\circ$ in~$\graphG$, defined as the graph with vertex set~$\ground \ssm \tube^\circ$ and edge set ${\bigset{e \in \binom{\ground \ssm \tube^\circ}{2}}{\text{$e$ or~$e \cup \tube^\circ$ is connected in~$\graphG$}}}$.
\begin{lemma}[\cite{CarrDevadoss}]
\label{lem:restriction}
For a tube~$\tube^\circ$ of~$\graphG$, the map
\[
\tube[s] \longmapsto \widetilde{\tube[s]} \eqdef
\begin{cases}
\tube[s] & \text{if $\tube[s] \subsetneq \tube^\circ$} \\
\tube[s] \ssm \tube^\circ & \text{if $\tube[s] \supsetneq \tube^\circ$ or~$\tube[s] \cap \tube^\circ = \varnothing$}
\end{cases}
\]
between the tubes of~$\graphG$ compatible with~$\tube^\circ$ and the tubes of~$\widetilde\graphG \eqdef \graphG{}[\tube^\circ] \sqcup \graphG^\star\tube^\circ$ defines an isomorphism between the link of~$\tube^\circ$ in the nested complex~$\nestedComplex(\graphG)$ and the nested complex~$\nestedComplex(\widetilde\graphG)$.
\end{lemma}
We denote by~$\widetilde\tubing \eqdef \set{\widetilde\tube}{\tube \in \tubing}$ the image of a tubing~$\tubing$ on~$\graphG$. This map actually preserves the compatibility degrees between tubes and the compatibility vectors.
\begin{lemma}
\label{lem:restrictionCompatibility}
Let~$\tube^\circ$ be a tube of~$\graphG$. The map~$\tube[s] \mapsto \widetilde{\tube[s]}$ between the link of~$\tube^\circ$ in the nested complex~$\nestedComplex(\graphG)$ and the nested complex~$\nestedComplex(\widetilde\graphG)$ of the graph~$\widetilde\graphG \eqdef \graphG{}[\tube^\circ] \sqcup \graphG^\star\tube^\circ$ defined in Lemma~\ref{lem:restriction} preserves the compatibility degree: $\compatibilityDegree{\tube}{\tube'} = \compatibilityDegree{\,\widetilde\tube}{\widetilde\tube'}$ for any tubes~$\tube, \tube'$ of~$\graphG$ compatible with~$\tube^\circ$. Therefore, for any maximal tubing~$\tubing^\circ$ on~$\graphG$ containing~$\tube^\circ$ and any tube~$\tube$ of~$\graphG$ compatible with~$\tube^\circ$, the compatibility vector~$\compatibilityVector{\widetilde{\tubing}^\circ}{\widetilde\tube}$ is obtained from the compatibility vector~$\compatibilityVector{\tubing^\circ}{\tube}$ by deletion of its vanishing $\tube^\circ$-coordinate.
\end{lemma}
\begin{proof}
If~$\tube$ and~$\tube'$ are compatible, so are~$\widetilde\tube$ and~$\widetilde\tube'$ by Lemma~\ref{lem:restriction}, thus the result follows from Proposition~\ref{prop:compatibilityDegree}. We can therefore assume that~$\tube$ and~$\tube'$ are incompatible, so as~$\widetilde\tube$ and~$\widetilde\tube'$. Therefore, the compatibility degrees~$\compatibilityDegree{\tube}{\tube'}$ and~$\compatibilityDegree{\,\widetilde\tube}{\widetilde\tube'}$ actually count neighbors. However, it follows immediately from the definitions of the graph~$\widetilde\graphG$ and of the map~$\tube[s] \mapsto \widetilde{\tube[s]}$ that the neighbors of~$\tube$ in~$\tube' \ssm \tube$ are precisely the neighbors of~$\widetilde\tube$ in~$\widetilde\tube' \ssm \widetilde\tube$. This proves the equality between the compatibility degrees. The equality between the compatibility vectors follows coordinate by coordinate.
\end{proof}
\subsection{Compatibility fan (Theorem~\ref{theo:compatibilityFan})}
\label{subsec:proofCompatibilityFan}
In order to show that the cones of the compatibility matrices of all tubings on~$\graphG$ form a complete simplicial fan, we need the following refinement.
\begin{theorem}
\label{theo:compatibilityFanRefined}
For any graph~$\graphG$ and any maximal tubing~$\tubing^\circ$ on~$\graphG$, the compatibility vectors with respect to~$\tubing^\circ$ have the following properties.
\begin{description}
\item[Span Property] For any tube~$\tube[u]$ of~$\graphG$, the span of~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing, \tube[s] \subseteq \tube[u]}$, for a maximal tubing~$\tubing$ on~$\graphG$ containing~$\tube[u]$, is independent of~$\tubing$. \\[-.3cm]
\item[Flip Property] For any two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ with~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$, there exists a linear dependence
\[
\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'} + \sum_{\tube[s] \in \tubing \cap \tubing'} \beta_{\tube[s]} \, \compatibilityVector{\tubing^\circ}{\tube[s]} = 0
\]
between the compatibility vectors of~$\tubing \cup \tubing'$ with respect to~$\tubing^\circ$ which is:
\begin{description}
\item[Separating] the hyperplane spanned by~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing'}$ separates~$\compatibilityVector{\tubing^\circ}{\tube}$ and $\compatibilityVector{\tubing^\circ}{\tube'}$, \ie the coefficients~$\alpha$ and~$\alpha'$ have the same sign different from~$0$.
\item[Local] the dependence is supported by tubes included in~$\overline{\tube}$, \ie $\beta_{\tube[s]} = 0$ for all~$\tube[s] \not\subseteq \overline{\tube}$.
\end{description}
\end{description}
\end{theorem}
Theorem~\ref{theo:compatibilityFan} follows from the \textbf{Separating Flip Property} and the characterization of complete simplicial fans in Proposition~\ref{prop:characterizationFan}. The \textbf{Span Property} and \textbf{Local Flip Property} are not required to get Theorem~\ref{theo:compatibilityFan} but we use them to obtain the proof of the \textbf{Separating Flip Property}. Observe also that we do not need to prove that the linear dependence between the compatibility vectors of~$\tubing \cup \tubing'$ is unique: it is a consequence of Proposition~\ref{prop:characterizationFan} once we know the \textbf{Separating Flip Property}.
\medskip
Before entering details, let us sketch the general idea of the proof of Theorem~\ref{theo:compatibilityFanRefined}. We seek for a linear dependence between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$ with respect to the initial tubing~$\tubing^\circ$, that is, for a linear relation satisfied by their compatibility degrees with any tube of~$\tubing^\circ$. There are simple combinatorial relations between the compatibility degrees of the tubes of~$\tubing \cup \tubing'$ with all tubes of~$\tubing^\circ$ not contained in~$\overline{\tube}$. Our strategy is to start from such a relation and adapt it iteratively such that it holds for the other tubes of~$\tubing^\circ$ as well. This transformation is done in two steps:
\begin{itemize}
\item We first deal with the tubes of~$\tubing^\circ$ contained in~$\overline{\tube}$ and maximal for this property. They determine the coefficients of the linear dependence on the forced tubes
\item For the remaining tubes of~$\tubing^\circ$, we need to make successive corrections to the linear dependence. We first get an explicit linear dependence assuming that~$\tubing \cap \tubing'$ contains certain suitable tubes included in~$\overline{\tube}$. We then use inductively the \textbf{Span Property} and the \textbf{Local Flip Property} to get an implicit linear dependence in general.
\end{itemize}
The key of the proof is that our transformation increases the set of tubes of~$\tubing^\circ$ for which the relation between the compatibility degrees of the tubes of~$\tubing \cup \tubing'$ is valid.
\medskip
We now start the formal proof. We proceed by induction on the dimension of the nested complex~$\nestedComplex(\graphG)$. It is immediate when this dimension is~$0$. We now consider an arbitrary graph~$\graphG$ and assume that we have shown Theorem~\ref{theo:compatibilityFanRefined} and thus Corollary~\ref{coro:fullRank} for any graph~$\graphG[H]$ such that~$\dim(\nestedComplex(\graphG[H])) < \dim(\nestedComplex(\graphG))$.
Given a exchangeable pair of tubes~$\tube, \tube'$ of~$\graphG$, our first objective is to exhibit \textbf{separating} and \textbf{local} linear dependences for some adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. We will show later the \textbf{Span Property} and use it to prove that the linear dependence is \textbf{separating} and \textbf{local} for all adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$.
\begin{lemma}
\label{lem:flipProperty}
For any two exchangeable tubes~$\tube, \tube'$ of~$\graphG$, there exists adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ and a linear dependence between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$ which is both \textbf{separating} and \textbf{local}.
\end{lemma}
\begin{proof}
We fix some notations for the forced tubes of the exchangeable pair~$\{\tube,\tube'\}$. First recall that since~$\tube$ and~$\tube'$ are exchangeable, the tube~$\tube$ (resp.~$\tube'$) has a single neighbor in~$\tube'\ssm\tube$ (resp. in~$\tube\ssm\tube'$) that we denote by~$r$ (resp.~$r'$). We set~${\overline{\tube} \eqdef \tube \cup \tube'}$, and we denote by~$\underline{\tube}_1, \dots, \underline{\tube}_k$ the connected components of~$\graphG{}[\tube \cap \tube']$, by~$\tube[a]_1, \dots, \tube[a]_\ell$ the connected components of~$\graphG{}[\tube \ssm (\tube' \cup \{r\})]$, and by~$\tube[a]'_1, \dots, \tube[a]'_{\ell'}$ the connected components of~$\graphG{}[\tube' \ssm (\tube \cup \{r'\})]$. Although it is not a tube, we set~$\underline{\tube} \eqdef \bigsqcup_{i \in [k]} \underline{\tube}_i$ and we abuse notation to write~$\compatibilityDegree{\tube^\circ}{\underline{\tube}}$ for~$\sum_{i \in [k]} \compatibilityDegree{\tube^\circ}{\underline{\tube}_i}$ and similarly~$\compatibilityVector{\tubing^\circ}{\underline{\tube}}$ for~$\sum_{i \in [k]} \compatibilityVector{\tubing^\circ}{\underline{\tube}_i}$. We will use in the same way the notations~$\tube[a]$ and~$\tube[a]'$.
We need to distinguish different cases, for which the linear dependences are slightly different, while the proofs are essentially identical. To simplify the discussion, we assume in Cases~(A), (B) and~(C) below that~$\tube,\tube' \notin \tubing^\circ$ and that no tube of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$. At the end of the proof, Cases~(D) and~(E) show how to restrict to these hypotheses.
\para{(A) A first relation}
Consider a tube~$\tube^\circ$ of~$\tubing^\circ$ not contained in~$\overline{\tube}$. We claim that
\[
\compatibilityDegree{\tube^\circ}{\tube} + \compatibilityDegree{\tube^\circ}{\tube'} = \compatibilityDegree{\tube^\circ}{\overline{\tube}} + \compatibilityDegree{\tube^\circ}{\underline{\tube}}.
\]
Indeed, since~$\tube^\circ \not\subseteq \overline{\tube}$, these four compatibility degrees actually count neighbors of~$\tube^\circ$. The formula thus follows from inclusion-exclusion principle since~$\overline{\tube} = \tube \cup \tube'$ and~$\underline{\tube} = \tube \cap \tube'$.
There are other tubes of~$\tubing^\circ$ satisfying this relation. Indeed, consider a tube~$\tube^\circ$ in~$\tubing^\circ$ included in~$\overline{\tube}$ which contains~$r$ but does not contain nor is adjacent to~$r'$. Then~$\tube^\circ \subsetneq \tube \subsetneq \overline{\tube}$ so that~$\compatibilityDegree{\tube^\circ}{\tube} = \compatibilityDegree{\tube^\circ}{\overline{\tube}} = 0$. Moreover, $\tube^\circ$ is incompatible with~$\tube'$ and all~$\underline{\tube}_1, \dots, \underline{\tube}_k$. Since $r'$ is not adjacent to~$\tube^\circ$, all neighbors of~$\tube^\circ$ in~$\tube' \ssm \tube^\circ$ are in~$\underline{\tube} \ssm \tube^\circ$. Therefore~$\compatibilityDegree{\tube^\circ}{\tube'} = \compatibilityDegree{\tube^\circ}{\underline{\tube}}$. The relation follows. Similarly, the relation follows if~$\tube^\circ$ contains~$r'$ and does not contain nor is adjacent to~$r$.
If all tubes of~$\tubing^\circ$ included in~$\overline{\tube}$ satisfy the previous conditions, we have obtained a \textbf{separating} and \textbf{local} linear dependence:
\begin{equation}
\label{eq:linearDependence1}
\compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} = \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \compatibilityVector{\tubing^\circ}{\overline{\tube}}.
\end{equation}
This linear dependence would be valid for any adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that ${\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}}$ since the tube~$\overline{\tube}$ and the tubes~$\underline{\tube}$ are forced in any such pair. Unfortunately, these conditions do not always hold for all tubes of~$\tubing^\circ$. In this case, we will therefore adapt the linear dependence~\eqref{eq:linearDependence1} to cover all possible configurations for the tubes of~$\tubing^\circ$.
\para{(B) No tube of~$\tubing^\circ$ contained in~$\overline{\tube}$ contains both~$r$ and~$r'$}
Except if the linear dependence~\eqref{eq:linearDependence1} is valid, there must exist w.l.o.g.~a tube~$\tube^\circ \in \tubing^\circ$ contained in~$\overline{\tube}$, containing~$r$ and adjacent to~$r'$. Choose~$\tube^\circ$ maximal for these properties. Since we have assumed that no tube of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$, all tubes of~$\tubing^\circ$ included in~$\tube^\circ$ contain~$r$. These tubes thus form a nested chain~$\tube^\circ = \tube^\circ_0 \supsetneq \tube^\circ_1 \supsetneq \dots \supsetneq \tube^\circ_p = \{r\}$. For~$i \in [p]$, define~$\tube^\star_i$ to be the connected component of~$\graphG{}[\tube^\circ_{i-1} \ssm \{r\}]$ containing the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$.
Set
\[
\alpha \eqdef \compatibilityDegree{\tube^\circ}{\tube[a]} + \compatibilityDegree{\tube^\circ}{\tube'}
\qquad\text{and}\qquad
\alpha' \eqdef \compatibilityDegree{\tube^\circ}{\underline{\tube}} + \compatibilityDegree{\tube^\circ}{\tube[a]},
\]
and define inductively~$\beta_1, \dots, \beta_p$ by
\[
\beta_i = \alpha' \, \compatibilityDegree{\tube^\circ_i}{\tube'} - \alpha \, \compatibilityDegree{\tube^\circ_i}{\underline{\tube}} - (\alpha - \alpha')\compatibilityDegree{\tube^\circ_i}{\tube[a]} - \sum_{j \in [i-1]} \beta_j \, \compatibilityDegree{\tube^\circ_i}{\tube^\star_j}.
\]
We claim that
\begin{equation}
\label{eq:linearDependence2}
\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'} = \alpha \, \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \alpha' \, \compatibilityVector{\tubing^\circ}{\overline{\tube}} + (\alpha - \alpha') \, \compatibilityVector{\tubing^\circ}{\tube[a]} + \sum_{i \in [p]} \beta_i \, \compatibilityVector{\tubing^\circ}{\tube^\star_i}.
\end{equation}
To prove it, we check this linear dependence coordinate by coordinate.
Observe first that~$\compatibilityDegree{\tube^\circ_i}{\tube} = \compatibilityDegree{\tube^\circ_i}{\overline{\tube}} = 0$ for all~$0 \le i \le p$ since~$\tube^\circ_i \subsetneq \tube \subseteq \overline{\tube}$. Moreover, for all~$i < j$, we have by definition~$\tube^\star_j \subsetneq \tube^\circ_i$, so that~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_j} = 0$. Finally, we have~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_i} = 1$ since $\tube^\circ_i$ and~$\tube^\star_i$ are incompatible and the only neighbor of~$\tube^\circ_i$ in~$\tube^\star_i \ssm \tube^\circ_i$ is the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$. Therefore, Relation~\eqref{eq:linearDependence2} holds for~$\tube^\circ$ by definition of~$\alpha$ and~$\alpha'$ and for~$\tube^\circ_i$ by definition of~$\beta_i$.
Consider now a tube~$\tube[s]^\circ$ of~$\tubing^\circ$ not included in~$\tube^\circ$. Suppose that~$\tube[s]^\circ$ is included in~$\overline{\tube}$. Then~$\tube[s]^\circ$ contains precisely one of~$r$ and~$r'$ (it cannot contain both by assumption~(B), and it cannot avoid both as it would be compatible with both~$\tube$ and~$\tube'$). If $\tube[s]^\circ$ contains~$r$, it contains~$\tube^\circ$ and therefore equals~$\tube^\circ$ by maximality of the latter. Otherwise, $\tube[s]^\circ$ contains~$r'$, thus is adjacent to~$\tube^\circ$, thus contains it (by compatibility), and thus contains~$r$, a contradiction. We therefore obtained that~$\tube[s]^\circ$ is not included in~$\overline{\tube}$, so that all compatibility degrees~$\compatibilityDegree{\tube[s]^\circ}{\tube}$, $\compatibilityDegree{\tube[s]^\circ}{\overline{\tube}}$, $\compatibilityDegree{\tube[s]^\circ}{\underline{\tube}}$ and~$\compatibilityDegree{\tube[s]^\circ}{\tube[a]}$ actually count neighbors of~$\tube[s]^\circ$.
Observe now that~$r$ cannot be adjacent to~$\tube[s]^\circ$ (except if it belongs to~$\tube[s]^\circ$), since $r$ belongs to~$\tube^\circ$ which is compatible with~$\tube[s]^\circ$. Therefore, we~have
\[
\compatibilityDegree{\tube[s]^\circ}{\tube} = \compatibilityDegree{\tube[s]^\circ}{\underline{\tube}} + \compatibilityDegree{\tube[s]^\circ}{\tube[a]}
\qquad\text{and}\qquad
\compatibilityDegree{\tube[s]^\circ}{\tube'} = \compatibilityDegree{\tube[s]^\circ}{\overline{\tube}} - \compatibilityDegree{\tube[s]^\circ}{\tube[a]},
\]
since~$\tube = \underline{\tube} \sqcup \tube[a] \sqcup \{r\}$ and~$\tube' = \overline{\tube} \ssm \tube[a] \ssm \{r\}$. Finally, since any~$\tube^\star_i$ is contained in~$\tube^\circ$, it is compatible with~$\tube[s]^\circ$, so that~$\compatibilityDegree{\tube[s]^\circ}{\tube^\star_i} = 0$ by Proposition~\ref{prop:compatibilityDegree}. Combining these equalities, we obtain that Relation~\eqref{eq:linearDependence2} holds for any~$\tube[s]^\circ \in \tubing^\circ$.
We found a linear dependence between the compatibility vectors of~${\{\tube, \tube', \overline{\tube}\} \cup \underline{\tube} \cup \tube[a] \cup \set{\tube^\star_i}{i \in [p]}}$ with respect to the initial tubing~$\tubing^\circ$. Any two of these tubes except~$\tube, \tube'$ are compatible:
\begin{itemize}
\item the forced tubes are pairwise compatible;
\item each~$\tube^\star_i$ is a connected component of~$\tube^\circ_i \ssm \{r\}$, thus is contained in~$\overline{\tube} \ssm \{r,r'\}$, and thus is compatible with the connected components of~$\overline{\tube} \ssm \{r,r'\}$ (\ie all forced tubes);
\item for~$\tube^\circ_i \supseteq \tube^\circ_j$, any connected component of~$\tube^\circ_j \ssm \{r\}$ is contained in a connected component of~$\tube^\circ_i \ssm \{r\}$ and thus is compatible with all connected components of~$\tube^\circ_i \ssm \{r\}$. In particular~$\tube^\star_i$ and~$\tube^\star_j$ are compatible.
\end{itemize}
Therefore, there exists adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ and $\tubing \cup \tubing' \supseteq \{\tube, \tube', \overline{\tube}\} \cup \underline{\tube} \cup \tube[a] \cup \set{\tube^\star_i}{i \in [p]}$. For this choice of~$\tubing, \tubing'$, we thus obtained a \textbf{separating} and \textbf{local} linear dependence~\eqref{eq:linearDependence2} between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$.
\para{(C) A tube of~$\tubing^\circ$ contained in~$\overline{\tube}$ contains both~$r$ and~$r'$}
We distinguish again two cases.
\para{\quad (C.1) No tube of $\tubing^\circ$ contained in~$\overline{\tube}$ contains~$r$ and is adjacent to~$r'$ or conversely}
There must exist a tube~$\tube^\circ \in \tubing^\circ$ included in~$\tube$ and containing~$r$. Choose~$\tube^\circ$ maximal for these properties. With the same arguments as in Case~(B), the tubes of~$\tubing^\circ$ included in~$\tube^\circ$ form a nested chain ${\tube^\circ = \tube^\circ_0 \supsetneq \tube^\circ_1 \supsetneq \dots \supsetneq \tube^\circ_p = \{r\}}$. For~$i \in [p]$, define~$\tube^\star_i$ to be the connected component of~${\graphG{}[\tube^\circ_{i-1} \ssm \{r\}]}$ containing the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$. We define the tube~$\tube'^\circ$, the chain~${\tube'^\circ = \tube'^\circ_0 \supsetneq \tube'^\circ_1 \supsetneq \dots \supsetneq \tube'^\circ_{p'} = \{r'\}}$ and the tubes~$\tube'^\star_i$ for~$i \in [p']$ similarly.
Consider now the inclusion minimal tube~$\overline{\tube}^\circ$ of~$\tubing^\circ$ contained in~$\overline{\tube}$ and containing both~$r$ and~$r'$. Since we assumed that no tube of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$, we have~$\overline{\tube}^\circ = \tube^\circ \sqcup \tube'^\circ \sqcup \{\overline{r}\}$ where~$\overline{r} \in \underline{\tube}$. Let~$\overline{\tube}^\star$ be the connected component of~$\graphG{}[\overline{\tube}^\circ \ssm \{r, r'\}]$ containing~$\overline{r}$.
Set
\[
\alpha \eqdef \compatibilityDegree{\tube'^\circ}{\underline{\tube}} + \compatibilityDegree{\tube'^\circ}{\tube[a]'}
\qquad\text{and}\qquad
\alpha' \eqdef \compatibilityDegree{\tube^\circ}{\underline{\tube}} + \compatibilityDegree{\tube^\circ}{\tube[a]},
\]
and define inductively~$\beta_1, \dots, \beta_p$ and~$\beta'_1, \dots, \beta'_{p'}$ by
\begin{align*}
\beta_i & = \alpha' \, \compatibilityDegree{\tube^\circ_i}{\tube'} - (\alpha + \alpha') \, \compatibilityDegree{\tube^\circ_i}{\underline{\tube}} - \alpha \, \compatibilityDegree{\tube^\circ_i}{\tube[a]} + \alpha\alpha' \, \compatibilityDegree{\tube^\circ_i}{\overline{\tube}^\star} - \sum_{j \in [i-1]} \beta_j \, \compatibilityDegree{\tube^\circ_i}{\tube^\star_j}, \\
\beta'_i & = \alpha \, \compatibilityDegree{\tube'^\circ_i}{\tube} - (\alpha + \alpha') \, \compatibilityDegree{\tube'^\circ_i}{\underline{\tube}} - \alpha' \, \compatibilityDegree{\tube'^\circ_i}{\tube[a]'} + \alpha\alpha' \, \compatibilityDegree{\tube'^\circ_i}{\overline{\tube}^\star} - \sum_{j \in [i-1]} \beta'_j \, \compatibilityDegree{\tube'^\circ_i}{\tube'^\star_j}.
\end{align*}
We claim that
\begin{align}
\label{eq:linearDependence3}
\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'} = & \; (\alpha + \alpha') \, \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \alpha \, \compatibilityVector{\tubing^\circ}{\tube[a]} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube[a]'} - \alpha\alpha' \, \compatibilityVector{\tubing^\circ}{\overline{\tube}^\star} \\
\nonumber & + \sum_{i \in [p]} \beta_i \, \compatibilityVector{\tubing^\circ}{\tube^\star_i} + \sum_{i \in [p']} \beta'_i \, \compatibilityVector{\tubing^\circ}{\tube'^\star_i}.
\end{align}
To prove it, we check this linear dependence coordinate by coordinate.
We start with~$\tube^\circ$ and~$\tube'^\circ$. We have~$\compatibilityDegree{\tube^\circ}{\tube} = 0$ since $\tube^\circ \subsetneq \tube$. Moreover, $\compatibilityDegree{\tube^\circ}{\overline{\tube}^\star} = |\{\overline{r}\}| = 1$. Finally, we have by definition~$\tube^\star_j \subsetneq \tube^\circ$ so that~$\compatibilityDegree{\tube^\circ}{\tube^\star_j} = 0$ for all~$j \in [p]$. Combining these equalities, Relation~\eqref{eq:linearDependence3} follows for~$\tube^\circ$ from the definition of~$\alpha$ and~$\alpha'$. The argument is identical~for~$\tube'^\circ$.
We now consider the tubes~$\tube^\circ_i$ and~$\tube'^\circ_i$. Observe first that~$\compatibilityDegree{\tube^\circ_i}{\tube} = 0$ for all~$0 \le i \le p$ since~$\tube^\circ_i \subsetneq \tube$. Moreover, for all~$i < j$, we have by definition~$\tube^\star_j \subsetneq \tube^\circ_i$, so that~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_j} = 0$. In addition, we have~$\compatibilityDegree{\tube^\circ_i}{\tube'^\star_j} = 0$ for all~$i \in [p]$ and~$j \in [p']$ since~$\tube^\circ_i$ and~$\tube'^\star_j$ are compatible. Finally, we have~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_i} = 1$ since $\tube^\circ_i$ and~$\tube^\star_i$ are incompatible and the only neighbor of~$\tube^\circ_i$ in~$\tube^\star_i \ssm \tube^\circ_i$ is the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$. Therefore, Relation~\eqref{eq:linearDependence3} holds for~$\tube^\circ_i$ by definition of~$\beta_i$. The argument is identical for~$\tube'^\circ_i$.
Finally, we consider a tube~$\tube[s]^\circ$ of~$\tubing^\circ$ not strictly contained in~$\overline{\tube}^\circ$. With similar arguments as in Case~(B), the compatibility degrees~$\compatibilityDegree{\tube[s]^\circ}{\tube}$, $\compatibilityDegree{\tube[s]^\circ}{\tube'}$, $\compatibilityDegree{\tube[s]^\circ}{\underline{\tube}}$, $\compatibilityDegree{\tube[s]^\circ}{\tube[a]}$ and~$\compatibilityDegree{\tube[s]^\circ}{\tube[a]'}$ actually count neighbors of~$\tube[s]^\circ$, and (by assumption~$C$) satisfy
\[
\compatibilityDegree{\tube[s]^\circ}{\tube} = \compatibilityDegree{\tube[s]^\circ}{\underline{\tube}} + \compatibilityDegree{\tube[s]^\circ}{\tube[a]}
\qquad\text{and}\qquad
\compatibilityDegree{\tube[s]^\circ}{\tube'} = \compatibilityDegree{\tube[s]^\circ}{\underline{\tube}} + \compatibilityDegree{\tube[s]^\circ}{\tube[a]'}.
\]
Moreover, since all~$\tube^\star_i$, $\tube'^\star_i$ and~$\overline{\tube}^\star$ are contained in~$\overline{\tube}^\circ$, they are all compatible with~$\tube[s]^\circ$, so that $\compatibilityDegree{\tube[s]^\circ}{\tube^\star_i} = \compatibilityDegree{\tube[s]^\circ}{\tube'^\star_i} = \compatibilityDegree{\tube[s]^\circ}{\overline{\tube}^\star} = 0$ by Proposition~\ref{prop:compatibilityDegree}. Combining these equalities, we obtain that Relation~\eqref{eq:linearDependence3} holds for any~$\tube[s]^\circ \in \tubing^\circ$.
With the same arguments as in Case~(B), there exists adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ and~$\tubing \cup \tubing' \supseteq \{\tube, \tube', \overline{\tube}^\star\} \cup \underline{\tube} \cup \tube[a] \cup \tube[a]' \cup \set{\tube^\star_i}{i \in [p]} \cup \set{\tube'^\star_i}{i \in [p]}$. For this choice of~$\tubing, \tubing'$, we thus obtained a \textbf{separating} and \textbf{local} linear dependence~\eqref{eq:linearDependence3} between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$.
\para{\quad (C.2) A tube of $\tubing^\circ$ contained in~$\overline{\tube}$ contains~$r$ and is adjacent to~$r'$ or conversely}
W.l.o.g., consider an inclusion maximal tube~$\tube^\circ \in \tubing^\circ$ contained in~$\overline{\tube}$, containing~$r$ and adjacent to~$r'$. Since we have assumed that no tube of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$, all tubes of~$\tubing^\circ$ included in~$\tube^\circ$ contain~$r$. These tubes thus form a nested chain~$\tube^\circ = \tube^\circ_0 \supsetneq \tube^\circ_1 \supsetneq \dots \supsetneq \tube^\circ_p = \{r\}$. For~$i \in [p]$, define~$\tube^\star_i$ to be the connected component of~$\graphG{}[\tube^\circ_{i-1} \ssm \{r\}]$ containing the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$.
Set
\[
\alpha \eqdef \compatibilityDegree{\tube^\circ}{\tube'} - \compatibilityDegree{\tube^\circ}{\underline{\tube}} = |\{r'\}| = 1
\qquad\text{and}\qquad
\alpha' \eqdef \compatibilityDegree{\tube^\circ}{\underline{\tube}} + \compatibilityDegree{\tube^\circ}{\tube[a]},
\]
and define inductively~$\beta_1, \dots, \beta_p$ by
\[
\beta_i = \alpha' \, \compatibilityDegree{\tube^\circ_i}{\tube'} - (1 + \alpha') \, \compatibilityDegree{\tube^\circ_i}{\underline{\tube}} - \compatibilityDegree{\tube^\circ_i}{\tube[a]} - \sum_{j \in [i-1]} \beta_j \, \compatibilityDegree{\tube^\circ_i}{\tube^\star_j}.
\]
We claim that
\begin{equation}
\label{eq:linearDependence4}
\compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'} = (1 + \alpha') \, \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \compatibilityVector{\tubing^\circ}{\tube[a]} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube[a]'} + \sum_{i \in [p]} \beta_i \, \compatibilityVector{\tubing^\circ}{\tube^\star_i}.
\end{equation}
To prove it, we check this linear dependence coordinate by coordinate.
Observe first that~$\compatibilityDegree{\tube^\circ_i}{\tube} = 0$ for all~$0 \le i \le p$ since~$\tube^\circ_i \subsetneq \tube$. Moreover, for all~$i < j$, we have by definition~$\tube^\star_j \subsetneq \tube^\circ_i$, so that~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_j} = 0$. Finally, we have~$\compatibilityDegree{\tube^\circ_i}{\tube^\star_i} = 1$ since $\tube^\circ_i$ and~$\tube^\star_i$ are incompatible and the only neighbor of~$\tube^\circ_i$ in~$\tube^\star_i \ssm \tube^\circ_i$ is the singleton~$\tube^\circ_{i-1} \ssm \tube^\circ_{i}$. Therefore, Relation~\eqref{eq:linearDependence2} holds for~$\tube^\circ$ by definition of~$\alpha$ and~$\alpha'$ and for~$\tube^\circ_i$ by definition of~$\beta_i$.
Consider now a tube~$\tube[s]^\circ$ of~$\tubing^\circ$ not strictly contained in~$\tube^\circ$. With similar arguments as in Case~(B), the compatibility degrees~$\compatibilityDegree{\tube[s]^\circ}{\tube}$, $\compatibilityDegree{\tube[s]^\circ}{\tube'}$, $\compatibilityDegree{\tube[s]^\circ}{\underline{\tube}}$, $\compatibilityDegree{\tube[s]^\circ}{\tube[a]}$ and~$\compatibilityDegree{\tube[s]^\circ}{\tube[a]'}$ actually count neighbors of~$\tube[s]^\circ$, and (by assumption~$C$) satisfy
\[
\compatibilityDegree{\tube[s]^\circ}{\tube} = \compatibilityDegree{\tube[s]^\circ}{\underline{\tube}} + \compatibilityDegree{\tube[s]^\circ}{\tube[a]}
\qquad\text{and}\qquad
\compatibilityDegree{\tube[s]^\circ}{\tube'} = \compatibilityDegree{\tube[s]^\circ}{\underline{\tube}} + \compatibilityDegree{\tube[s]^\circ}{\tube[a]'}.
\]
Moreover, since all~$\tube^\star_i$ are contained in~$\tube^\circ$, they are all compatible with~$\tube[s]^\circ$, so that $\compatibilityDegree{\tube[s]^\circ}{\tube^\star_i} = 0$ by Proposition~\ref{prop:compatibilityDegree}. Combining these equalities, we obtain that Relation~\eqref{eq:linearDependence3} holds for any~$\tube[s]^\circ \in \tubing^\circ$.
With the same arguments as in Case~(B), there exists adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$ and~$\tubing \cup \tubing' \supseteq \{\tube, \tube'\} \cup \underline{\tube} \cup \tube[a] \cup \tube[a]' \cup \set{\tube^\star_i}{i \in [p]}$. For this choice of~$\tubing, \tubing'$, we thus obtained a \textbf{separating} and \textbf{local} linear dependence~\eqref{eq:linearDependence4} between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$.
\medskip
\enlargethispage{.2cm}
We assumed in Cases~(A), (B) and~(C) above that~$\tube,\tube' \notin \tubing^\circ$ and that no tube of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$. The remaining two cases show how to force this assumption.
\para{(D) A tube~$\tube^\circ$ of~$\tubing^\circ$ is compatible with both~$\tube$ and~$\tube'$}
We treat this case by induction on the number of tubes of~$\tubing^\circ$ compatible with both~$\tube$ and~$\tube'$. By Lemma~\ref{lem:restrictionCompatibility}, the compatibility vectors with respect to~$\tubing^\circ$ of the tubes of~$\graphG$ compatible with~$\tube^\circ$ correspond to the compatibility vectors of the tubes of~$\widetilde\graphG \eqdef \graphG{}[\tube^\circ] \sqcup \graphG^\star\tube^\circ$ with respect to the maximal tubing~$\widetilde\tubing^\circ$. Since there are strictly less tubes of~$\widetilde\tubing^\circ$ compatible with both~$\widetilde\tube$ and~$\widetilde\tube'$ than tubes of~$\tubing^\circ$ compatible with both~$\tube$ and~$\tube'$, the induction hypothesis ensures that there exists adjacent maximal tubings~$\widetilde\tubing, \widetilde\tubing'$ on~$\widetilde\graphG$ such that~${\widetilde\tubing \ssm \{\widetilde\tube\} = \widetilde\tubing' \ssm \{\widetilde\tube'\}}$ and a \textbf{separating} and \textbf{local} linear dependence between the compatibility vectors of~${\widetilde\tubing \cup \widetilde\tubing'}$ with respect to~$\widetilde\tubing^\circ$. By Lemma~\ref{lem:restriction}, the sets~${\tubing \eqdef \{\tube^\circ\} \cup \bigset{\tube[s]}{\widetilde{\tube[s]} \in \widetilde\tubing}}$ and ${\tubing' \eqdef \{\tube^\circ\} \cup \bigset{\tube[s]'}{\widetilde{\tube[s]}' \in \widetilde\tubing'}}$ are tubings on~$\graphG$ (since~$\tube^\circ$ is compatible with all preimages of tubes of~$\widetilde\graphG$) and they are maximal by cardinality. Moreover, Lemma~\ref{lem:restrictionCompatibility} ensures that the linear dependence between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$ with respect to~$\tubing^\circ$ coincides with the linear dependence between the compatibility vectors of the tubes of~$\widetilde\tubing \cup \widetilde\tubing'$ with respect to~$\widetilde\tubing^\circ$. The linear dependence clearly remains \textbf{separating} and \textbf{local}, which concludes when there is a tube~$\tube^\circ \in \tubing^\circ$ compatible with both~$\tube$ and~$\tube'$.
\para{(E) $\tube$ or~$\tube'$ belongs to~$\tubing^\circ$}
We can assume w.l.o.g.~that~$\tube = \tube^\circ$ belongs to~$\tubing^\circ$. Consider any two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ such that~${\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}}$. Since any tube~$\tube[s]$ in~$\tubing \cap \tubing'$ is compatible with~$\tube = \tube^\circ$, the $\tube^\circ$-coordinate of the compatibility vector~$\compatibilityVector{\tubing^\circ}{\tube[s]}$ vanishes. The same happens for the vector~${\b{v} \eqdef \compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'}}$ since $\compatibilityDegree{\tube^\circ}{\tube} = -1$ (as~$\tube = \tube^\circ$) while $\compatibilityDegree{\tube^\circ}{\tube'} = 1$ (by Proposition~\ref{prop:compatibilityDegree}). The set of vectors~$\{\b{v}\} \cup \set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing'}$ has cardinality~$|\tubing^\circ|$ but is contained in the hyperplane of~$\R^{\tubing^\circ}$ orthogonal to~$\b{e}_{\tube^\circ}$. Therefore, there is a linear dependence between these vectors, which translates into a linear dependence between the compatibility vectors of the tubes of~$\tubing \cup \tubing'$ with the same coefficient on~$\compatibilityVector{\tubing^\circ}{\tube}$ and~$\compatibilityVector{\tubing^\circ}{\tube'}$. This coefficient cannot vanish: otherwise, we would have a linear dependence on the compatibility vectors of~$\widetilde\tubing$ with respect to~$\widetilde\tubing^\circ$. Since~$\widetilde\tubing$ is a maximal tubing on~$\widetilde\graphG$, and~$\dim(\nestedComplex(\widetilde\graphG)) = \dim(\nestedComplex(\graphG)) - 1$, this would contradict Corollary~\ref{coro:fullRank} and thus the induction hypothesis for~$\widetilde\graphG$. This linear dependence is therefore \textbf{separating}. It is automatically \textbf{local} if~$\overline{\tube} = \ground$ since all tubes are then subsets of~$\overline{\tube}$. Otherwise, we prove that it is \textbf{local} by restriction. We distinguish two cases.
\para{\quad (E.1) A tube~$\overline{\tube}^\circ$ of~$\tubing^\circ$ contains~$\tube \cup \{r'\}$ and is contained in~$\overline{\tube}$}
Consider the restricted graph~${\hat\graphG = \graphG{}[\tube \cup \{r'\}]}$. Any tube~$\tube[s]$ of~$\graphG$ included in~$\tube$ is also a tube of~$\hat\graphG$. Therefore, the set~$\hat\tubing \eqdef \set{\tube[s]}{\tube[s] \in \tubing, \tube[s] \subseteq \tube}$ is a tubing on~$\hat\graphG$ for any tubing~$\tubing$ containing~$\tube$. Define also the tube~$\hat\tube' \eqdef \tube' \ssm \tube[a]'$ of~$\hat\graphG$. The existence of~$\overline{\tube}^\circ$ implies that $\compatibilityDegree{\tube[s]^\circ}{\tube'} = \compatibilityDegree{\tube[s]^\circ}{\tube[a]'}$ for any~$\tube[s]^\circ \in \tubing^\circ$ not included in~$\tube$. It follows that the compatibility vector~$\compatibilityVector{\hat\tubing^\circ}{\hat\tube'}$ is the restriction of~$\compatibilityVector{\tubing^\circ}{\tube'}-\compatibilityVector{\tubing^\circ}{\tube[a]'}$ to the coordinates indexed by~$\hat\tubing^\circ$. Similarly, for any tube~$\tube[s] \in \tubing$ contained in~$\tube$, the compatibility vector~$\compatibilityVector{\hat\tubing^\circ}{\tube[s]}$ is the restriction of~$\compatibilityVector{\tubing^\circ}{\tube[s]}$ to the coordinates indexed by~$\hat\tubing^\circ$. This shows that the linear dependence on~$\bigset{\compatibilityVector{\hat\tubing^\circ}{\tube[s]}}{\tube[s] \in \hat\tubing \cup \hat\tubing'}$ provides a linear dependence on~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cup \tubing', \tube[s] \subseteq \tube} \cup \{\compatibilityVector{\tubing^\circ}{\tube'}-\compatibilityVector{\tubing^\circ}{\tube[a]'}\}$. The resulting linear dependence on~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cup \tubing'}$ is~\textbf{local}.
\para{\quad (E.2) No tube of~$\tubing^\circ$ contains~$\tube \cup \{r'\}$ and is contained in~$\overline{\tube}$}
The proof is identical to Case~(E.1), replacing~$\compatibilityVector{\tubing^\circ}{\tube'}-\compatibilityVector{\tubing^\circ}{\tube[a]'}$ by~$\compatibilityVector{\tubing^\circ}{\tube'}-\compatibilityVector{\tubing^\circ}{\overline{\tube}}$. Details are left to the reader.
\end{proof}
We can now prove the \textbf{Span Property}.
\begin{lemma}
\label{lem:spanProperty}
For any tube~$\tube[u]$ of~$\graphG$, the span of~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing, \tube[s] \subseteq \tube[u]}$, for a maximal tubing~$\tubing$ on~$\graphG$ containing~$\tube[u]$, is independent of~$\tubing$.
\end{lemma}
\begin{proof}
We proceed by induction on the size of~$\tube[u]$. The result is immediate if~$\tube[u]$ is a singleton. Consider now two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ containing~$\tube[u]$ such that~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. Assume first that~$\tube$ and~$\tube'$ are contained in~$\tube[u]$. By Lemma~\ref{lem:flipProperty}, there exists adjacent maximal tubings~$\tubing[S], \tubing[S]'$ on~$\graphG$ containing~$\tube[u]$ such that~$\tubing[S] \ssm \{\tube\} = \tubing[S]' \ssm \{\tube'\}$ and a linear dependence between the compatibility vectors of the tubes of~$\tubing[S] \cup \tubing[S]'$ with respect to~$\tubing^\circ$ which is both \textbf{separating} and \textbf{local}. By definition, this implies that there exists~$\alpha > 0$ and~$\alpha' > 0$ such that the vector~${\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'}}$ belongs to~$\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S] \cap \tubing[S]', \tube[s] \subseteq \overline{\tube}})$. However,
\[
\set{\tube[s] \in \tubing[S] \cap \tubing[S]'}{\tube[s] \subseteq \overline{\tube}} = \{\overline{\tube}\} \, \cup \, \set{\tube[s] \in \tubing[S] \cap \tubing[S]'}{\tube[s] \subseteq \underline{\tube}} \, \cup \, \set{\tube[s] \in \tubing[S] \cap \tubing[S]'}{\tube[s] \subseteq \tube[a]} \, \cup \, \set{\tube[s] \in \tubing[S] \cap \tubing[S]'}{\tube[s] \subseteq \tube[a]'}.
\]
By induction hypothesis applied to each tube of~$\underline{\tube}$, we have
\[
\begin{array}{@{}c@{\;}c@{\;}l@{\quad}r}
\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S] \cap \tubing[S]', \tube[s] \subseteq \underline{\tube}}) & = &\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S], \tube[s] \subseteq \underline{\tube}}) & \text{(as~$\tubing[S] \ssm (\tubing[S] \cap \tubing[S]') = \tube \not\subseteq \underline{\tube}$)} \\
& = & \vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing, \tube[s] \subseteq \underline{\tube}}) & \text{(induction hypothesis)} \\
& = & \vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing', \tube[s] \subseteq \underline{\tube}}) & \text{(as~$\tubing \ssm (\tubing \cap \tubing') = \tube \not\subseteq \underline{\tube}$)}
\end{array}
\]
and similarly replacing~$\underline{\tube}$ by~$\tube[a]$ or~$\tube[a]'$. It follows that the vector~${\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'}}$ also belongs to~$\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing', \tube[s] \subseteq \tube[u]})$. Since~$\alpha \ne 0$, this implies that~$\compatibilityVector{\tubing^\circ}{\tube}$ belongs to~$\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing', \tube[s] \subseteq \tube[u]})$. Similarly, $\compatibilityVector{\tubing^\circ}{\tube'}$ belongs to~$\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing, \tube[s] \subseteq \tube[u]})$. We therefore obtained that
\[
\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing, \tube[s] \subseteq \tube[u]}) = \vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing', \tube[s] \subseteq \tube[u]}).
\]
This also clearly holds when~$\tube$ and~$\tube'$ are not contained in~$\tube[u]$. This concludes the proof since the graph of flips on the maximal tubings on~$\graphG$ containing~$\tube[u]$ is connected.
\end{proof}
We can finally conclude the proof of Theorem~\ref{theo:compatibilityFanRefined} using both Lemmas~\ref{lem:flipProperty} and~\ref{lem:spanProperty}.
\begin{proof}[Proof of Theorem~\ref{theo:compatibilityFanRefined}]
The \textbf{Span Property} is proved in Lemma~\ref{lem:spanProperty}. It only remains to show the \textbf{Flip Property} for arbitrary adjacent maximal tubings. Consider two adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$ with~$\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}$. By Lemma~\ref{lem:flipProperty}, there exists adjacent maximal tubings~$\tubing[S], \tubing[S]'$ on~$\graphG$ such that~$\tubing[S] \ssm \{\tube\} = \tubing[S]' \ssm \{\tube'\}$ and a linear dependence between the compatibility vectors of the tubes of~$\tubing[S] \cup \tubing[S]'$ with respect to~$\tubing^\circ$ which is both \textbf{separating} and \textbf{local}. By definition, this implies that there exists~$\alpha > 0$ and~$\alpha' > 0$ such that the vector~${\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'}}$ belongs to~$\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S] \cap \tubing[S]', \tube[s] \subseteq \overline{\tube}})$. Lemma~\ref{lem:spanProperty} applied to~$\underline{\tube}$ ensures that
\[
\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing', \tube[s] \subseteq \underline{\tube}}) = \vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S] \cap \tubing[S]', \tube[s] \subseteq \underline{\tube}}).
\]
and similarly replacing~$\underline{\tube}$ by~$\tube[a]$ or~$\tube[a]'$. We thus conclude that
\[
\vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cap \tubing', \tube[s] \subseteq \overline{\tube}}) = \vect(\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing[S] \cap \tubing[S]', \tube[s] \subseteq \overline{\tube}})
\]
contains the vector~${\alpha \, \compatibilityVector{\tubing^\circ}{\tube} + \alpha' \, \compatibilityVector{\tubing^\circ}{\tube'}}$ with~$\alpha > 0$ and~$\alpha' > 0$. In other words, we obtained a \textbf{separating} and \textbf{local} linear dependence on~$\set{\compatibilityVector{\tubing^\circ}{\tube[s]}}{\tube[s] \in \tubing \cup \tubing'}$.
\end{proof}
\subsection{Dual compatibility fan (Theorem~\ref{theo:dualCompatibilityFan})}
\label{subsec:proofDualCompatibilityFan}
The fact that dual compatibility vectors support a complete simplicial fan realizing the nested complex is a direct consequence of Theorem~\ref{theo:compatibilityFan}, using the following duality trick. Observe first that given $n+1$ vectors~$\{u, v, w_1, \dots, w_{n-1}\}$ in~$\R^n$, the hyperplane spanned by~$\{w_1, \dots, w_{n-1}\}$ separates~$u$ and~$v$ if and only if
\[
\det([u | w_1 | \dots | w_{n-1}]) \cdot \det([v | w_1 | \dots | w_{n-1}]) < 0.
\]
Theorem~\ref{theo:compatibilityFan} shows this condition for primal compatibility matrices and we need to show it for dual compatibility matrices. For this, we notice that the primal and dual compatibility matrices are related by
\begin{equation}
\label{eq:dualityTrick}
\dualCompatibilityMatrix{\tubing}{\tubing^\circ} \eqdef [\compatibilityDegree{\tube_j}{\tube_i^\circ}]_{i, j \in [n]} = \transpose{[\compatibilityDegree{\tube_i}{\tube_j^\circ}]_{i,j \in [n]}} \defeq \transpose{\compatibilityMatrix{\tubing}{\tubing^\circ}},
\end{equation}
where~$\transpose{M}$ denotes the transpose of the matrix~$M$.
Consider now two pairs of adjacent maximal tubings~$\tubing^\circ, {\tubing^\circ}'$ and~$\tubing, \tubing'$ on~$\graphG$. From the \textbf{Separation Flip Property} of Theorem~\ref{theo:compatibilityFan} applied to the initial maximal tubing~$\tubing$, we obtain that ${\det(\compatibilityMatrix{\tubing}{\tubing^\circ}) \cdot \det(\compatibilityMatrix{\tubing}{{\tubing^\circ}'}) < 0}$. Similarly, for the initial maximal tubing~$\tubing'$, we obtain that ${\det(\compatibilityMatrix{\tubing'}{\tubing^\circ}) \cdot \det(\compatibilityMatrix{\tubing'}{{\tubing^\circ}'}) < 0}$. Multiplying these two inequalities, we get
\[
\bigg( \det(\compatibilityMatrix{\tubing}{\tubing^\circ}) \cdot \det(\compatibilityMatrix{\tubing'}{\tubing^\circ}) \bigg) \cdot \bigg( \det(\compatibilityMatrix{\tubing}{{\tubing^\circ}'}) \cdot \det(\compatibilityMatrix{\tubing'}{{\tubing^\circ}'}) \bigg) > 0.
\]
Since the transposition preserves the determinant, we obtain by Equation~\ref{eq:dualityTrick} that
\[
\bigg( \det(\dualCompatibilityMatrix{\tubing}{\tubing^\circ}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{\tubing^\circ}) \bigg) \cdot \bigg( \det(\dualCompatibilityMatrix{\tubing}{{\tubing^\circ}'}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{{\tubing^\circ}'}) \bigg) > 0.
\]
This implies that the products~$\det(\dualCompatibilityMatrix{\tubing}{\tubing^\circ}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{\tubing^\circ})$ and~$\det(\dualCompatibilityMatrix{\tubing}{{\tubing^\circ}'}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{{\tubing^\circ}'})$ have the same sign. Since we know that~$\det(\dualCompatibilityMatrix{\tubing}{{\tubing^\circ}}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{{\tubing^\circ}}) < 0$ for the tubing~$\tubing^\circ = \tubing$, we obtain by repeated flips in~$\tubing^\circ$ that~$\det(\dualCompatibilityMatrix{\tubing}{{\tubing^\circ}}) \cdot \det(\dualCompatibilityMatrix{\tubing'}{{\tubing^\circ}}) < 0$ for any initial tubing~$\tubing^\circ$ and pair of adjacent maximal tubings~$\tubing, \tubing'$ on~$\graphG$. This shows the \textbf{Separation Flip Property} for dual compatibility vectors. We conclude again by Proposition~\ref{prop:characterizationFan}. \qed
\begin{remark}
Observe that this proof does not provide us with any explicit linear dependence between dual compatibility vectors. In particular, we do not know whether the following analogue properties of Theorem~\ref{theo:compatibilityFanRefined} hold:
\begin{description}
\item[Dual Span Property] For any tube~$\tube[u]$ of~$\graphG$, the span of~$\set{\dualCompatibilityVector{\tube}{\tubing^\circ}}{\tube \in \tubing, \tube \not\subset \tube[u]}$, for a maximal tubing~$\tubing$ on~$\graphG{}$ containing~$\tube[u]$, is independent of~$\tubing$. \\[-.3cm]
\item[Dual Local Flip Property] For any two maximal tubings~$\tubing, \tubing'$ with~${\tubing \ssm \{\tube\} = \tubing' \ssm \{\tube'\}}$, the unique linear dependence between the dual compatibility vectors of~$\tubing \cup \tubing'$ with respect to~$\tubing^\circ$ is supported by tubes not strictly included in a connected component of~$\underline{\tube} = \tube \cap \tube'$.
\end{description}
\end{remark}
\subsection{Nested complex isomorphisms (Proposition~\ref{prop:exmAutomorphism}, Proposition~\ref{prop:nestedComplexIsomorphismDisconnected} and Theorem~\ref{theo:nestedComplexIsomorphisms})}
\label{subsec:proofIsomorphisms}
We now prove various results on nested complex isomorphisms presented in Section~\ref{subsec:many}. We first show that the map~$\Omega$ on the tubes of a spider~$\spiderG_{\underline{n}}$ defines a nested complex automorphism that dualizes the compatibility degree.
\begin{proof}[Proof of Proposition~\ref{prop:exmAutomorphism}]
First, $\Omega$ clearly sends tubes of~$\spiderG_{\underline{n}}$ to tubes of~$\spiderG_{\underline{n}}$. We just have to show that~$\compatibilityDegree{\Omega(\tube)}{\Omega(\tube')} = \compatibilityDegree{\tube'}{\tube}$ for any two tubes~$\tube, \tube'$ of~$\spiderG_{\underline{n}}$, and Proposition~\ref{prop:compatibilityDegree} will imply that~$\Omega$ is a nested complex automorphism. This follows from the definition of~$\Omega$ and the fact that
\begin{gather*}
\biggCompatibilityDegree{\big[ v^i_j, v^i_k \big]}{\big[ v^{i'}_{j'}, v^{i'}_{k'} \big]} = \delta_{i = i'} \cdot \big( \delta_{j < j' \le k+1 < k'+1} + \delta_{j' < j \le k'+1 < k+1} \big), \\
\biggCompatibilityDegree{\big[ v^i_j, v^i_k \big]}{\bigcup_{h \in \ell} \big[ v^h_0, v^h_{k_h} \big]} = \delta_{j \le k_i + 1 \le k}, \\
\text{and} \qquad \biggCompatibilityDegree{\bigcup_{i \in \ell} \big[ v^i_0, v^i_{k_i} \big]}{\bigcup_{i \in \ell} \big[ v^i_0, v^i_{k'_i} \big]} = |\set{i \in [\ell]}{k_i < k'_i}| \cdot \delta_{\exists i \in [\ell], \, k_i > k'_i}. \qedhere
\end{gather*}
\end{proof}
Our objective is to show that these maps~$\Omega$ on spiders are essentially the only non-trivial nested complex isomorphisms. We fix an isomorphism~$\Phi$ between two nested complexes~$\nestedComplex(\graphG)$ and~$\nestedComplex(\graphG')$. We first show that~$\Phi$ preserves connected components in the following sense.
\begin{lemma}
\label{lem:isomorphismsPreserveConnectedComponents}
Two tubes~$\tube$ and~$\tube'$ of~$\graphG$ belong to the same connected component of~$\graphG$ if and only if their images~$\Phi(\tube)$ and~$\Phi(\tube')$ belong to the same connected component of~$\graphG'$.
\end{lemma}
\begin{proof}
Observe first that two tubes from distinct connected components are automatically compatible. Assume now that~$\tube$ and~$\tube'$ are in the same connected component of~$\graphG$. If~$\tube$ and~$\tube'$ are incompatible, then~$\Phi(\tube)$ and~$\Phi(\tube')$ are also incompatible and therefore in the same connected component of~$\graphG'$. If~$\tube$ and~$\tube'$ are compatible, then there exists a tube~$\tube''$ incompatible with both~$\tube$~and~$\tube'$:
\begin{itemize}
\item if~$\tube \cap \tube' = \varnothing$, consider a path connecting a neighbor of~$\tube$ to a neighbor of~$\tube'$ in~$\graphG{}[\ground \ssm (\tube \cup \tube')]$;
\item if~$\tube \subseteq \tube'$, consider a path connecting a neighbor of~$\tube$ to a neighbor of~$\tube'$ in~$\graphG{}[\ground \ssm \tube]$.
\end{itemize}
We obtain that~$\Phi(\tube'')$ is incompatible with both~$\Phi(\tube)$ and~$\Phi(\tube')$, so that they all belong to the same connected component of~$\graphG'$. This proves one direction. For the other direction, consider~$\Phi^{-1}$.
\end{proof}
Consequently, the nested complex~$\nestedComplex(\graphG)$ records the sizes of the connected components of the graph~$\graphG$. We call \defn{connected size partition} of~$\graphG$ the partition~$\lambda(\graphG) \eqdef |\ground_1|, |\ground_2|, \dots, |\ground_{\connectedComponents}|$ of~$|\ground|$, where~$\ground_1, \dots, \ground_\connectedComponents$ are the connected components of~$\graphG$ ordered such that~$|\ground_i| \ge |\ground_{i+1}|$.
\begin{corollary}
\label{coro:isomorphismsPreserveConnectedSizePartition}
Two graphs whose nested complexes are isomorphic have the same connected size partitions: $\nestedComplex(\graphG) \simeq \nestedComplex(\graphG') \Longrightarrow \lambda(\graphG) = \lambda(\graphG')$. In particular, they have the same number of vertices.
\end{corollary}
\begin{proof}
Consider a maximal tubing~$\tubing$ on~$\graphG$ and decompose it into subtubings~$\tubing_1, \dots, \tubing_\connectedComponents$ on the connected components~$\ground_1, \dots, \ground_\connectedComponents$ of~$\graphG$. Their images~$\Phi(\tubing_1), \dots, \Phi(\tubing_\connectedComponents)$ decompose the maximal tubing~$\Phi(\tubing)$. Moreover, Lemma~\ref{lem:isomorphismsPreserveConnectedComponents} ensures that two tubes~$\Phi(\tube) \in \Phi(\tubing_i)$ and~$\Phi(\tube') \in \Phi(\tubing_{i'})$ belong to the same connected component of~$\graphG'$ if and only if~$i = i'$. We therefore obtain that ${\lambda(\graphG) = \{|\ground_1|, \dots, |\ground_{\connectedComponents}|\} = \{|\tubing_1| + 1, \dots, |\tubing_{\connectedComponents}| + 1\} = \{|\Phi(\tubing_1)| + 1, \dots, |\Phi(\tubing_{\connectedComponents})| + 1\} = \lambda(\graphG')}$.
\end{proof}
Proposition~\ref{prop:nestedComplexIsomorphismDisconnected} is another immediate consequence of Lemma~\ref{lem:isomorphismsPreserveConnectedComponents}: since it sends all tubes in a connected component~$\graphG[H]$ of~$\graphG$ to tubes in the same connected component~$\graphG[H]'$ of~$\graphG'$, the map~$\Phi$ induces a nested complex isomorphism between~$\nestedComplex(\graphG[H])$ and~$\nestedComplex(\graphG[H]')$. From now on, we assume without loss of generality that~$\graphG$ is connected. Our next step is a crucial structural property of~$\Phi$.
\begin{lemma}
For any nested complex isomorphism~$\Phi:\nestedComplex(\graphG) \to \nestedComplex(\graphG')$ and any tube~$\tube$ of~$\graphG$, either~$|\Phi(\tube)| = |\tube|$ or~$|\Phi(\tube)| = |\ground| - |\tube|$.
\end{lemma}
\begin{proof}
By Lemma~\ref{lem:restriction}, the join of a tube~$\tube$ in~$\nestedComplex(\graphG)$ is isomorphic to the nested complex~${\nestedComplex(\graphG{}[\tube] \sqcup \graphG^\star\tube)}$ of the union of the restriction~$\graphG{}[\tube]$ with the reconnected complement~$\graphG^\star\tube$. The former has~$|\tube|$ vertices while the latter has~$|\ground| - |\tube|$ vertices. Since~$\Phi$ induces an isomorphism from the join of~$\tube$ in~$\nestedComplex(\graphG)$ to the join of~$\Phi(\tube)$ in~$\nestedComplex(\graphG')$, the result follows from Corollary~\ref{coro:isomorphismsPreserveConnectedSizePartition}.
\end{proof}
We say that~$\Phi$ \defn{maintains} the tube~$\tube$ if~$|\Phi(\tube)| = |\tube|$ and that~$\Phi$ \defn{swaps} the tube~$\tube$ if~${|\Phi(\tube)| = |\ground| - |\tube|}$.
\begin{proposition}
\label{prop:nonReversing}
If it maintains all tubes of~$\graphG$, then~$\Phi$ is the trivial nested complex isomorphism induced by the graph isomorphism~$\psi : \graphG \to \graphG'$ defined by~$\Phi(\{v\}) = \{\psi(v)\}$.
\end{proposition}
\begin{proof}
Two vertices~$v$ and~$w$ of~$\graphG$ are adjacent if and only if the two tubes~$\{v\}$ and~$\{w\}$ are incompatible. Since~$\Phi$ preserves the compatibility relation, this shows that~$v$ and~$w$ are adjacent if and only if~$\psi(v)$ and~$\psi(w)$ are, \ie that~$\psi$ defines a graph isomorphism. Let~$\Psi$ denote the nested complex isomorphism induced by~$\psi$, \ie defined by~$\Psi(\tube) \eqdef \set{\psi(v)}{v \in \tube}$.
We prove by induction on~$|\tube|$ that~$\Phi(\tube) = \Psi(\tube)$ for any tube~$\tube$ on~$\graphG$. It holds for singletons. For the induction step, consider an arbitrary tube~$\tube$ of~$\graphG$. Let~$v \in \ground \ssm \tube$ be a neighbor of~$\tube$. Since~$\{v\}$ and~$\tube$ are incompatible, so are~$\Phi(\{v\}) = \{\psi(v)\}$ and~$\Phi(\tube)$, and thus~$\psi(v)$ is a neighbor of~$\Phi(\tube)$. Let~$w \in \ground$ be such that~$\psi(w)$ is a neighbor of~$\psi(v)$ in~$\Phi(\tube)$. If~$w \notin \tube$ then it is incompatible with~$\tube \cup \{v\}$, and thus~$\Phi(\{w\}) = \{\psi(w)\}$ is incompatible with~$\Phi(\tube \cup \{v\})$. Therefore, $\Phi(\tube \cup \{v\})$ is adjacent to and does not contain~$\psi(w)$ which is in~$\Phi(\tube)$. Since~${|\Phi(\tube \cup \{v\})| = |\tube| + 1 = |\Phi(\tube)| + 1}$, this implies that~$\Phi(\tube \cup \{v\})$ is incompatible with~$\Phi(\tube)$, a contradiction. Therefore, we know that~$w \in \tube$. Let~$\tube_1, \dots, \tube_k$ denote the connected components of~$\graphG{}[\tube \ssm \{w\}]$. By induction hypothesis, ${\Phi(\tube_i) = \Psi(\tube_i)}$ for all~${i \in [k]}$. Moreover, since~$\Phi(\tube_i)$ is compatible with~$\Phi(\tube)$ and adjacent to~${\psi(w) \in \Phi(\tube)}$, it is included in~$\Phi(\tube)$. We thus obtain~${\Psi(\tube) = \{\psi(w)\} \cup \Psi(\tube_1) \cup \dots \cup \Psi(\tube_k) \subseteq \Phi(\tube)}$ and thus~$\Phi(\tube) = \Psi(\tube)$ since~${|\Phi(\tube)| = |\tube| = |\Psi(\tube)|}$.
\end{proof}
\begin{lemma}
\label{lem:structureMaintenedSwappedTubes}
If~$\Phi$ does not maintain a tube~$\tube$ of~$\graphG$ (\ie~$|\Phi(\tube)| \ne |\tube|$), then
\begin{enumerate}[(i)]
\item $\Phi$ swaps any tube of~$\graphG$ containing~$\tube$,
\item $\Phi$ maintains any tube of~$\graphG$ disjoint from and non-adjacent to~$\tube$, and
\item $\Phi$ swaps at least one singleton included in~$\tube$.
\end{enumerate}
\end{lemma}
\begin{proof}
Consider a tube~$\tube[s]$ of~$\graphG$ strictly containing~$\tube$. The link of~$\{\tube[s], \tube\}$ in~$\nestedComplex(\graphG)$ is isomorphic to the nested complex of the union of the graphs~$\graphG{}[\tube]$, $(\graphG^\star\tube)[\tube[s] \ssm \tube]$, and~$\graphG^\star\tube[s]$ with~$|\tube|$, $|\tube[s]| - |\tube|$ and~${|\ground| - |\tube[s]|}$ vertices respectively. Therefore, Corollary~\ref{coro:isomorphismsPreserveConnectedSizePartition} ensures that the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ is isomorphic to the nested complex of a graph with three connected components with~$|\tube|$, $|\tube[s]| - |\tube|$ and~$|\ground| - |\tube[s]|$ vertices respectively.
If~$\Phi(\tube[s])$ is not contained in~$\Phi(\tube)$, then the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ would be isomorphic to the nested complex of a graph with one connected component~$\graphG'[\Phi(\tube)]$ having~$|\Phi(\tube)|$ vertices. We reach a contradiction as~$|\Phi(\tube)| = |\ground| - |\tube|$ is neither~$|\tube|$ (by assumption on~$\tube$), nor~$|\ground| - |\tube[s]|$ (since~$|\tube[s]| > |\tube|$), nor~$|\tube[s]| - |\tube|$ (since~$|\tube[s]| < |\ground|$).
Therefore, $\Phi(\tube[s])$ is contained in~$\Phi(\tube)$ and the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ is isomorphic to the union of the graphs~$\graphG{}[\Phi(\tube[s])]$, $(\graphG^\star\Phi(\tube[s]))[\Phi(\tube) \ssm \Phi(\tube[s])]$, and~$\graphG^\star\Phi(\tube)$ with~$|\Phi(\tube[s])|$, $|\Phi(\tube)| - |\Phi(\tube[s])|$ and~${|\ground| - |\Phi(\tube)|}$ vertices respectively. If~$|\Phi(\tube[s])| \ne |\ground| - |\tube[s]|$, then it forces~$|\Phi(\tube[s])| = |\tube[s]| - |\tube| = |\tube[s]|$, a contradiction.
This proves~(i).
Consider now a tube~$\tube[s]$ of~$\graphG$ disjoint from and non-adjacent to~$\tube$. Note that~$|\tube[s]| + |\tube| < |\ground|$ as there is at least a vertex separing them. The link of~$\{\tube[s], \tube\}$ in~$\nestedComplex(\graphG)$ is isomorphic to the nested complex of the union of the graphs~$\graphG{}[\tube[s]]$, $\graphG{}[\tube]$, and~$(\graphG^\star\tube[s])^\star\tube$ with~$|\tube[s]|$, $|\tube|$ and~${|\ground| - |\tube[s]| - |\tube|}$ vertices respectively. Again, Corollary~\ref{coro:isomorphismsPreserveConnectedSizePartition} ensures that the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ is isomorphic to the nested complex of a graph with three connected components with~$|\tube[s]|$, $|\tube|$ and~${|\ground| - |\tube[s]| - |\tube|}$ vertices~respectively.
If~$\Phi(\tube[s])$ is not contained in~$\Phi(\tube)$, then the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ would be isomorphic to the nested complex of a graph with one connected component~$\graphG'[\Phi(\tube)]$ having~$|\Phi(\tube)|$ vertices. We reach a contradiction as~$|\Phi(\tube)| = |\ground| - |\tube|$ is neither~$|\tube|$ (by assumption on~$\tube$), nor~$|\tube[s]|$ (since~$|\tube[s]| + |\tube| < |\ground|$), nor~$|\ground| - |\tube[s]| - |\tube|$ (since~$|\tube[s]| > 0$).
Therefore, $\Phi(\tube[s])$ is contained in~$\Phi(\tube)$ and the link of~$\{\Phi(\tube[s]), \Phi(\tube)\}$ in~$\nestedComplex(\graphG')$ is isomorphic to the union of the graphs~$\graphG{}[\Phi(\tube[s])]$, $(\graphG^\star\Phi(\tube[s]))[\Phi(\tube) \ssm \Phi(\tube[s])]$, and~$\graphG^\star\Phi(\tube)$ with~$|\Phi(\tube[s])|$, $|\Phi(\tube)| - |\Phi(\tube[s])|$ and~${|\ground| - |\Phi(\tube)|}$ vertices respectively. If~$|\Phi(\tube[s])| \ne |\tube[s]|$, then it forces~$|\Phi(\tube[s])| = |\ground| - |\tube[s]| - |\tube| = |\ground| - |\tube[s]|$, a contradiction.
This proves~(ii).
Finally, to prove~(iii) we can assume that~$\tube$ is not a singleton. Thus~$\Phi(\tube)$ is not an inclusion maximal tube. Let~$\tube'$ be a tube of~$\graphG$ such that~$\Phi(\tube')$ is a maximal tube of~$\graphG'$ containing~$\Phi(\tube)$. Since~$\Phi^{-1}$ swaps~$\Phi(\tube)$ and~$\Phi(\tube')$ contains~$\Phi(\tube)$, $\Phi^{-1}$ also swaps~$\Phi(\tube')$ by~(i). Thus, $\tube'$ is a singleton swapped by~$\Phi$ and contained in~$\tube$.
\end{proof}
\begin{lemma}
\label{lem:structureMaintenedSwappedVertices}
Denote by~$M \eqdef \set{v \in V}{|\Phi(\{v\})| = 1}$ the set of vertices maintained by~$\Phi$ and by~$S \eqdef \set{v \in V}{|\Phi(\{v\}|) = |\ground| - 1}$ the set of vertices swapped by~$\Phi$. Then
\begin{enumerate}[(i)]
\item $S$ forms a clique of~$\graphG$,
\item any vertex in~$M$ has at most one neighbor in~$S$, and
\item any vertex in~$S$ has at most one neighbor in~$M$.
\end{enumerate}
\end{lemma}
\begin{proof}
Let~$s,s' \in S$. Since~$|\Phi(\{s\})| = |\Phi(\{s'\})| = |\ground| - 1$, the tubes~$\Phi(\{s\})$ and~$\Phi(\{s'\})$ are incompatible. Therefore~$\{s\}$ and~$\{s'\}$ are incompatible, so that~$s$ and~$s'$ are neighbors. The set~$S$ thus forms a clique.
To prove~(ii) and~(iii), assume that some vertices~$m \in M$ and~$s \in S$ are neighbors. The tubes~$\{m\}$ and~$\{s\}$ are thus incompatible, so that~$\Phi(\{m\})$ and~$\Phi(\{s\})$ are also incompatible. Since ${|\Phi(\{m\})| = 1}$ while~$|\Phi(\{s\})| = |\ground| - 1$, this implies~$\Phi(\{s\}) = \ground' \ssm \Phi(\{m\})$. It follows that~$m$ cannot have another neighbor swapped by~$\Phi$ and~$s$ cannot have another neighbor maintained~by~$\Phi$.
\end{proof}
We are now ready to prove that any non-trivial nested complex isomorphism coincides, up to composition with a trivial nested complex isomorphism, with the isomorphism~$\Omega$ on a spider.
\begin{proof}[Proof of Theorem~\ref{theo:nestedComplexIsomorphisms}]
The proof works by induction on the number~$|\ground|$ of vertices of~$\graphG$. It is clear when~$|\ground| \le 2$. For the induction step, assume that the result holds for all graphs on less than~$|\ground|$ vertices and consider a non-trivial nested complex isomorphism~$\Phi : \nestedComplex(\graphG) \to \nestedComplex(\graphG')$. Then~$\Phi$ does not maintain all tubes of~$\graphG$ by Proposition~\ref{prop:nonReversing}, and thus swaps at least one singleton~$\{s\}$ by Lemma~\ref{lem:structureMaintenedSwappedTubes}\,(iii). Let~$s'$ denote the vertex of~$\graphG'$ such that~$\Phi(\{s\}) = \ground' \ssm \{s'\}$.
The map~$\Phi$ induces a nested complex isomorphism between the link of~$\{s\}$ in~$\nestedComplex(\graphG)$ and the link of~$\Phi(\{s\})$ in~$\nestedComplex(\graphG')$. The former is isomorphic to the nested complex of the reconnected complement~$\widetilde\graphG \eqdef \graphG^\star\{s\}$ while the latter is isomorphic to the nested complex of the restriction~$\widetilde\graphG' \eqdef \graphG'[\Phi(\{s\})]$. Let~$\widetilde\Phi : \widetilde\tube \mapsto \Phi(\tube)$ denote the resulting nested complex isomorphism between~$\nestedComplex(\widetilde\graphG)$ and~$\nestedComplex(\widetilde\graphG')$. This isomorphism~$\widetilde\Phi$ is non-trivial: by Lemma~\ref{lem:structureMaintenedSwappedTubes}, $\Phi$ swaps any tube~$\tube$ containing~$\{s\}$, so that $\widetilde\Phi$ swaps the tube~${\tube \ssm \{s\}}$. It follows by induction hypothesis that~$\widetilde\graphG$ and~$\widetilde\graphG'$ are spiders and that there exists a graph isomorphism~${\widetilde\psi : \widetilde\graphG \to \widetilde\graphG'}$ inducing a trivial nested complex isomorphism~$\widetilde\Psi : \nestedComplex(\widetilde\graphG) \to \nestedComplex(\widetilde\graphG')$ such that~$\widetilde\Psi^{-1} \circ \widetilde\Phi \defeq \widetilde\Omega$ is the automorphism of~$\nestedComplex(\widetilde\graphG)$ described in Section~\ref{subsec:many}. In other words, we can label by~$\widetilde v^i_j$ the vertices of the spider~$\widetilde\graphG$ and by~$\widetilde v'^i_j$ the vertices of the spider~$\widetilde\graphG'$, with~$i \in [\,\widetilde\ell\,]$ and~$0 \le j \le \widetilde n_i$, such that
\begin{equation}
\label{eq:barPhi}
\widetilde\Phi\big( \big[ \widetilde v^i_j, \widetilde v^i_k \big] \big) = \big[ \widetilde v'^i_{\widetilde n_i+1-k}, \widetilde v'^i_{\widetilde n_i+1-j} \big]
\quad\text{and}\quad
\widetilde\Phi\big( \bigcup_{i \in [\,\widetilde\ell\,]} \big[ \widetilde v^i_0, \widetilde v^i_{k_i} \big] \big) = \bigcup_{i \in [\,\widetilde\ell\,]} \big[ \widetilde v'^i_0, \widetilde v'^i_{\widetilde n_i-1-k_i} \big].
\end{equation}
We now claim that~$\graphG$ and~$\graphG'$ are both spiders. To prove it, we distinguish two cases:
\begin{description}
\item[Body case] all neighbors of~$s$ in~$\graphG$ are swapped by~$\Phi$. Then they form a clique in~$\graphG$ (by Lemma~\ref{lem:structureMaintenedSwappedVertices}\,(i)), so that~$\graphG$ is the spider~$\widetilde\graphG$ where we add one more body vertex~$s$ with no attached leg. Moreover,~$s'$ is necessarily swapped by~$\Phi^{-1}$ (otherwise~$\Phi^{-1}(\{s'\})$ would be a neighbor of~$s$ maintained by~$\Phi$). We conclude by symmetry that~$\graphG'$ is the spider~$\widetilde\graphG'$ where we add one more body vertex~$s'$ with no attached leg.
\item[Leg case] $s$ has a neighbor~$m$ maintained by~$\Phi$. It is unique by Lemma~\ref{lem:structureMaintenedSwappedVertices}\,(iii) and not connected to any other vertex swapped by~$\Phi$ by Lemma~\ref{lem:structureMaintenedSwappedVertices}\,(ii). Therefore,~$\graphG$ is the spider~$\widetilde\graphG$ where we replace the edges connecting~$m$ to all other body vertices of~$\widetilde\graphG$ by a new body vertex~$s$ with an edge to~$m$. Moreover,~$s'$ is necessarily maintained by~$\Phi^{-1}$ (otherwise~$\Phi^{-1}(\{s'\})$ should be~$\ground \ssm \{s\}$ which is not connected). We conclude that~$\graphG'$ is the spider~$\widetilde\graphG'$ where we add one additional leg vertex~$s'$ to the free endpoint of a leg.
\end{description}
We now label by~$v^i_j$ the vertices of~$\graphG$ according to the labels~$\widetilde v^i_j$ of~$\widetilde\graphG$ and by~$v'^i_j$ the vertices of~$\graphG'$ according to the labels~$\widetilde v'^i_j$ of~$\widetilde\graphG'$. We follow the two cases above:
\begin{description}
\item[Body case] We set~$\ell \eqdef \widetilde\ell + 1$, $n_i \eqdef \widetilde n_i$ for~$i \in [\,\widetilde\ell\,]$ and~$n_\ell = 0$. For any~$i \in [\,\widetilde\ell\,]$ and~$0 \le j \le n_i$, we label by~$v^i_j$ the vertex of~$\graphG$ corresponding to the vertex labeled by~$\widetilde v^i_j$ in~$\widetilde\graphG$, and similarly we label by~$v'^i_j$ the vertex of~$\graphG'$ corresponding to the vertex labeled by~$\widetilde v'^i_j$ in~$\widetilde\graphG'$. Finally, we label~$s$ by~$v^\ell_0$ and~$s'$ by~$v'^\ell_0$.
\item[Leg case] Assume that the neighbor of~$s$ maintained by~$\Phi$ corresponds to the vertex labeled by~$\widetilde v^a_0$ in~$\widetilde\graphG$. Then the neighbor of~$s'$ corresponds to the vertex labeled by~$\widetilde v'^a_{\widetilde n_a}$ in~$\widetilde\graphG'$. We set~$\ell \eqdef \widetilde\ell$, $n_i \eqdef \widetilde n_i$ for~$i \in [\ell] \ssm \{a\}$ and~$n_a \eqdef \widetilde n_a + 1$. For any~$i \in [\ell]$ and~$0 \le j \le \widetilde n_i$, we label by~$v^i_j$ if~$i \ne a$ and~$v^i_{j+1}$ if~$i = a$ the vertex of~$\graphG$ corresponding to the vertex labeled by~$\widetilde v^i_j$ in~$\widetilde\graphG$ and by~$v'^i_j$ the vertex of~$\graphG'$ corresponding to the vertex labeled by~$\widetilde v'^i_j$ in~$\widetilde\graphG'$. Finally, we label~$s$ by~$v^a_0$ and~$s'$ by~$v'^a_{n_a}$.
\end{description}
By our previous description of the graphs~$\graphG$ and~$\graphG'$, these labelings are indeed valid labelings of spiders, meaning that the edges of~$\graphG$ are indeed given by~$\bigset{\big\{v^i_{j-1}, v^i_j\big\}}{i \in [\ell], j \in [n_i]} \cup \bigset{\big\{v^i_0, v^{i'}_0\big\}}{i \ne i' \in [\ell]}$, and similarly for~$\graphG'$. We moreover claim that~$\Phi$ is given by
\begin{equation}
\label{eq:Phi}
\Phi\big( \big[ v^i_j, v^i_k \big] \big) = \big[ v^i_{n_i+1-k}, v^i_{n_i+1-j} \big]
\qquad\text{and}\qquad
\Phi\big( \bigcup_{i \in [\ell]} \big[ v^i_0, v^i_{k_i} \big] \big) = \bigcup_{i \in [\ell]} \big[ v^i_0, v^i_{n_i-1-k_i} \big].
\end{equation}
It is immediate for all tubes compatible with~$\{s\} = \{v^a_0\}$ as it is easily transported from~\eqref{eq:barPhi}. Therefore, we only have to check it for the tubes of~$\graphG$ adjacent to~$s = v^a_0$ and not containing~it. Observe first that~$\Phi\big( \big[ v^a_1, v^a_k \big] \big)$ is a tube with~$k$ vertices (by Lemma~\ref{lem:structureMaintenedSwappedTubes}\,(iii)), and that it contains~$s' = v'^a_{n_a}$ since it has to be incompatible with~$\Phi(\{s\}) = \ground' \ssm \{s'\}$. Therefore, $\Phi\big( \big[ v^a_1, v^a_k \big] \big) = \big[ v^a_{n_a+1-k}, v^a_{n_a} \big]$. Consider now a tube~$\tube = \bigcup_{i \in [\ell]} \big[ v^i_0, v^i_{k_i} \big]$ not containing~$s = v^a_0$ (\ie with~$k_a = -1$). Since the nested tubes~$\tube$ and~$\tube \cup \{s\}$ are both swapped, we have~$\Phi(\tube) = \Phi(\tube \cup \{s\}) \cup \{s'\}$. Since~$\Phi(\tube \cup \{s\})$ is given by Equation~\eqref{eq:Phi}, so is~$\Phi(\tube)$. This concludes the proof that~$\Phi$ is given by Equation~\eqref{eq:Phi}, so that it coincides with~$\Omega$ up to the graph automorphism defined by~$v^i_j \mapsto v'^i_j$.
\end{proof}
We now prove that if the primal and dual compatibility fans of~$\graphG$ with respect to the same initial maximal tubing~$\tubing^\circ$ are linearly isomorphic, then~$\graphG$ is an octopus whose head is contained in no tube of~$\tubing^\circ$.
\begin{proof}[Proof of Lemma~\ref{lem:comparisonPrimalDual}]
Consider a graph~$\graphG$ and an initial tubing~$\tubing^\circ$ such that the fans~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\dualCompatibilityFan{\graphG}{\tubing^\circ}$ are linearly isomorphic. The fans~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\dualCompatibilityFan{\graphG}{\tubing^\circ}$ both contains precisely $n$ pairs of opposite rays, given by the vectors~$\b{e}_i$ of the canonical basis and their opposites~$-\b{e}_i$. Therefore, the fans~$\compatibilityFan{\graphG}{\tubing^\circ}$ and~$\dualCompatibilityFan{\graphG}{\tubing^\circ}$ have the same rays, which implies that the compatibility vector~$\compatibilityVector{\tubing^\circ}{\tube}$ and dual compatibility vector~$\dualCompatibilityVector{\tube}{\tubing^\circ}$ are collinear for any tube~$\tube$ of~$\graphG$. In other words, we have~$\compatibilityDegree{\tube^\circ_1}{\tube} \compatibilityDegree{\tube}{\tube^\circ_2} = \compatibilityDegree{\tube}{\tube^\circ_1} \compatibilityDegree{\tube^\circ_2}{\tube}$ for all tubes~$\tube$ of~$\graphG$ and~$\tube^\circ_1, \tube^\circ_2 \in \tubing^\circ$.
We now prove by induction that this condition implies that~$\graphG$ is an octopus whose head is contained in no tube of~$\tubing^\circ$. The result is clear when~$|\ground| \le 3$. Consider thus a connected graph~$\graphG$ on more than~$4$ vertices and a maximal tubing~$\tubing^\circ$ on~$\graphG$ with root~$u$ (\ie $u$ is the only vertex of~$\ground$ contained in no proper tube of~$\tubing^\circ$). The graph~$\graphG{}[\ground\ssm\{u\}]$ has connected components~$\graphG_1,\dots,\graphG_k$ and~$\tubing^\circ$ induces a maximal tubing~$\tubing^\circ_i$ on each component~$\graphG_i$. We know that~$\graphG_1$ is an octopus whose head~$v$ is contained in no tube of~$\tubing^\circ$. Otherwise, by induction hypothesis, we could find three tubes~$\tube, \tube^\circ_1, \tube^\circ_2$ of~$\graphG_1$ such that~$\compatibilityDegree{\tube^\circ_1}{\tube} \compatibilityDegree{\tube}{\tube^\circ_2} \ne \compatibilityDegree{\tube}{\tube^\circ_1} \compatibilityDegree{\tube^\circ_2}{\tube}$, which would contradict our assumption on~$\tubing^\circ$ since~$\tube, \tube^\circ_1, \tube^\circ_2$ are also tubes of~$\graphG$. Assume now that~$\graphG_1$ is not a path. We distinguish five~cases:
\begin{itemize}
\item Suppose that~$v$ is the unique vertex of~$\graphG_1$ adjacent to~$u$. Let~${\tube = \{u,v\}}$, ${\tube^\circ_1 = \graphG_1}$ and $\tube^\circ_2$ be a leg of~$\graphG_1$. Then~$\compatibilityDegree{\tube}{\tube_1^\circ} \geq 2$ and~$\compatibilityDegree{\tube_1^\circ}{\tube} = \compatibilityDegree{\tube}{\tube_2^\circ} = \compatibilityDegree{\tube_2^\circ}{\tube} = 1$.
\item Suppose that~$u$ is adjacent to~$v$ and at least another vertex~$w$ of~$\graphG_1$. Let~${\tube = \{u\}}$, ${\tube^\circ_1 = \graphG_1}$ and $\tube^\circ_2$ be the leg of~$\graphG_1$ containing~$w$. Then~$\compatibilityDegree{\tube_1^\circ}{\tube} = \compatibilityDegree{\tube_2^\circ}{\tube} = 1$ and~${\compatibilityDegree{\tube}{\tube_2^\circ} < \compatibilityDegree{\tube}{\tube_1^\circ}}$.
\item Suppose that~$u$ is adjacent to at least two legs of~$\graphG_1$. Let~${\tube = \{u\}}$, ${\tube^\circ_1 = \graphG_1}$ and $\tube^\circ_2$ be a leg of~$\graphG_1$ adjacent to~$u$. Then~${\compatibilityDegree{\tube_1^\circ}{\tube} = \compatibilityDegree{\tube_2^\circ}{\tube} = 1}$ and~${\compatibilityDegree{\tube}{\tube_2^\circ} < \compatibilityDegree{\tube}{\tube_1^\circ}}$.
\item Suppose that~$u$ is adjacent to a single leg~$\graphG[H]$ of~$\graphG_1 \ssm \{v\}$ but not to~$v$. Let~$\tube = \{u,v\} \cup \graphG[H]$, $\tube^\circ_1 = \graphG_1$ and $\tube^\circ_2$ be a leg of~$\graphG_1$ distinct from~$\graphG[H]$. Then~$\compatibilityDegree{\tube}{\tube_1^\circ} \geq 2$ and~$\compatibilityDegree{\tube_1^\circ}{\tube} = \compatibilityDegree{\tube}{\tube_2^\circ} = \compatibilityDegree{\tube_2^\circ}{\tube} = 1$.
\end{itemize}
In all cases, we have~$\compatibilityDegree{\tube^\circ_1}{\tube} \compatibilityDegree{\tube}{\tube^\circ_2} \ne \compatibilityDegree{\tube}{\tube^\circ_1} \compatibilityDegree{\tube^\circ_2}{\tube}$, contradicting our assumption on~$\tubing^\circ$. Thus,~$\graphG_1$ is a path. A similar case analysis shows that~$\graphG_1$ is attached to~$u$ only by one of its endpoints. By symmetry, all components~$\graphG_1, \dots, \graphG_k$ of~$\graphG \ssm \{u\}$ are paths attached to~$u$ only by an endpoint, so that~$\graphG$ is an octopus whose head is contained in no tube~of~$\tubing^\circ$.
\end{proof}
\subsection{Polytopality of compatibility fans (Theorem~\ref{theo:polytopalityPathsCycles} and Proposition~\ref{prop:polytopeStar})}
\label{subsec:proofPolytopality}
This section provides the proof of the polytopality results presented in Section~\ref{subsec:polytopality}. Using a similar method as~\cite[Section~5]{CeballosSantosZiegler} based on Proposition~\ref{prop:polytopalityFan}, we first prove that all compatibility and dual compatibility fans of paths and cycles are polytopal.
\begin{proof}[Proof of Theorem~\ref{theo:polytopalityPathsCycles}]
We use the characterization of polytopality of complete simplicial fans given in Proposition~\ref{prop:polytopalityFan}. For this, we need to understand better the linear dependences on compatibility vectors for paths and cycles.
\medskip
Consider first the case of the path. When~$(\tubing \cup \tubing') \cap \tubing^\circ = \varnothing$, the linear dependences can only be of the form
\begin{align*}
\compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} & = \compatibilityVector{\tubing^\circ}{\overline{\tube}} + \compatibilityVector{\tubing^\circ}{\underline{\tube}},
\\
\compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} & = \compatibilityVector{\tubing^\circ}{\tube[a]} + \compatibilityVector{\tubing^\circ}{\tube[a]'},
\\
2 \, \compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} & = \compatibilityVector{\tubing^\circ}{\overline{\tube}} + \compatibilityVector{\tubing^\circ}{\tube[a]},
\\
2 \, \compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} & = \compatibilityVector{\tubing^\circ}{\underline{\tube}} + \compatibilityVector{\tubing^\circ}{\tube[a]'},
\end{align*}
up to exchanging simultaneously~$\tube$ with~$\tube'$ and~$\tube[a]$ with~$\tube[a]'$. If~$\tubing \cap \tubing'$ contains a tube~$\tube^\circ \in \tubing^\circ$, then the compatibility degree of all tubes of~$(\tubing \cup \tubing') \ssm \{\tube^\circ\}$ with~$\tube^\circ$ vanishes, so that the tube~$\tube^\circ$ cannot appear in the linear dependence. When~$\tube, \tube' \notin \tubing^\circ$ but~$(\tubing \cap \tubing') \cap \tubing^\circ \ne \varnothing$, the relations are thus obtained from the ones above by deleting terms in their right hand sides. The dependences when~$\tube$ or~$\tube'$ belong to~$\tubing^\circ$ will be treated separately.
We now define a height function~$\omega$ on tubes on~$\pathG_{n+1}$ by
\[
\omega(\tube) =
\begin{cases}
f(|\tube|) & \text{if } \tube \notin \tubing^\circ, \\
\Omega & \text{otherwise,}
\end{cases}
\]
where~$f: \R \to \R_{>0}$ is any strictly concave increasing positive function and~$\Omega \in \R$ is a large enough constant. When~$(\tubing \cup \tubing') \cap \tubing^\circ = \varnothing$, we obtain by definition of~$\overline{\tube}$, $\underline{\tube}$, $\tube[a]$ and~$\tube[a]'$, and using that~$f$ is concave and increasing, that
\begin{align*}
\omega(\tube) + \omega(\tube') & > \omega(\overline{\tube}) + \omega(\underline{\tube}),
\\
\omega(\tube) + \omega(\tube') & > \omega(\tube[a]) + \omega(\tube[a]'),
\\
2 \, \omega(\tube) + \omega(\tube') & > \omega(\overline{\tube}) + \omega(\tube[a]),
\\
2 \, \omega(\tube) + \omega(\tube') & > \omega(\underline{\tube}) + \omega(\tube[a]').
\end{align*}
Moreover, the inequalities still hold when we delete terms in their right hand sides since~$\omega$ is positive. Therefore, $\omega$ satisfies Condition~(2) of Proposition~\ref{prop:polytopalityFan} when we do not flip an initial tube. Finally, initial tubes only appear in the left hand sides of linear dependences, so choosing~$\omega(\tube^\circ) = \Omega$ large enough ensures that $\omega$ satisfies Condition~(2) of Proposition~\ref{prop:polytopalityFan} for any flip. Observe that this is essentially the same proof as in~\cite[Section~5]{CeballosSantosZiegler}.
\medskip
We now adapt this proof for the cycle~$\cycleG_{n+1}$. Clearly, the dependences described above for the path also appear for the cycle (as cycles contain paths). Beside those, when~$(\tubing \cup \tubing') \cap \tubing^\circ = \varnothing$, a straightforward case analysis shows that the linear dependences can only be of the form
\[
\compatibilityVector{\tubing^\circ}{\tube} + \compatibilityVector{\tubing^\circ}{\tube'} = 2 \, \compatibilityVector{\tubing^\circ}{\underline{\tube}_1}, \qquad\text{where } \underline{\tube}_1 \in \underline{\tube}.
\]
Again, no tube of~$\tubing^\circ$ can appear in their right hand sides of the linear dependences. Therefore, when~$\tube, \tube' \notin \tubing^\circ$ but $(\tubing \cap \tubing') \cap \tubing^\circ \ne \varnothing$, the linear dependences are obtained from the generic ones above by deleting terms in their right hand sides. The dependences when~$\tube$ or~$\tube'$ belong to~$\tubing^\circ$ will again be treated separately.
We choose the same height function~$\omega$ as before. For the same reasons, the linear dependences for the path are again transformed to strict inequalities on~$\omega$. Moreover, as~$\tube_1 \subseteq \tube \cap \tube'$ and~$f$ is increasing, we have
\[
\omega(\tube) + \omega(\tube') > 2 \, \omega(\underline{\tube}_1).
\]
We conclude as before by choosing~$\Omega$ large enough that~$\omega$ satisfies the Condition~(2) of Proposition~\ref{prop:polytopalityFan} for any flip.
\medskip
Finally, for dual compatibility vectors, a straightforward case analysis shows that the linear dependences are all of the form
\begin{align*}
\dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = \dualCompatibilityVector{\overline{\tube}}{\tubing^\circ} + \dualCompatibilityVector{\underline{\tube}}{\tubing^\circ},
\\
\dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = \dualCompatibilityVector{\tube[a]}{\tubing^\circ} + \dualCompatibilityVector{\tube[a]'}{\tubing^\circ},
\\
2 \, \dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = \dualCompatibilityVector{\overline{\tube}}{\tubing^\circ} + \dualCompatibilityVector{\tube[a]}{\tubing^\circ},
\\
2 \, \dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = \dualCompatibilityVector{\underline{\tube}}{\tubing^\circ} + \dualCompatibilityVector{\tube[a]'}{\tubing^\circ},
\end{align*}
when none of~$\tube$, $\tube'$ and~$\overline{\tube}$ have~$n$ vertices. We can also have the linear dependences
\begin{align*}
\dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = 2 \, \dualCompatibilityVector{\overline{\tube}}{\tubing^\circ} + \dualCompatibilityVector{\underline{\tube}}{\tubing^\circ},
\\
2 \, \dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} & = 2 \, \dualCompatibilityVector{\overline{\tube}}{\tubing^\circ} + \dualCompatibilityVector{\tube[a]}{\tubing^\circ},
\end{align*}
when~$|\overline{\tube}| = n$ and
\[
\dualCompatibilityVector{\tube}{\tubing^\circ} + \dualCompatibilityVector{\tube'}{\tubing^\circ} = \dualCompatibilityVector{\underline{\tube}_1}{\tubing^\circ}, \qquad\text{where } \underline{\tube}_1 \in \underline{\tube}.
\]
when~$|\tube| = |\tube'| = n$.
Again, no tube of~$\tubing^\circ$ can appear in the right hand sides of the linear dependences. Therefore, when~$\tube, \tube' \notin \tubing^\circ$ but $(\tubing \cap \tubing') \cap \tubing^\circ \ne \varnothing$, the linear dependences are obtained from the generic ones above by deleting terms in their right hand sides.
We now define a height function~$\omega$ on tubes on~$\cycleG_{n+1}$ by
\[
\omega(\tube) =
\begin{cases}
f(|\tube|) & \text{if } \tube \notin \tubing^\circ \text{ and } |\tube| \ne n, \\
\displaystyle{\frac{f(|\tube|)}{2}} & \text{if } \tube \notin \tubing^\circ \text{ and } |\tube| = n , \\
\Omega & \text{otherwise,}
\end{cases}
\]
where~$f: \R \to \R_{>0}$ is any strictly concave increasing positive function and~$\Omega \in \R$ is a large enough constant. By definition of~$\overline{\tube}$, $\underline{\tube}$, $\tube[a]$ and~$\tube[a]'$, and using that~$f$ is concave and increasing, we obtain that~$\omega$ satisfies a strict inequality for each linear dependence above. We conclude as before by choosing~$\Omega$ large enough that~$\omega$ satisfies the Condition~(2) of Proposition~\ref{prop:polytopalityFan} for any flip.
\end{proof}
Our last proof concerns the polytopality of the compatibility fan for the star, for which we have presented a candidate in Section~\ref{subsec:polytopality}.
\begin{proof}[Proof of Proposition~\ref{prop:polytopeStar}]
We just have to show that for any tube~$\tube$ and any maximal tubing~$\tubing$ on~$\starG_{n+1}$, the point~$\b{x}(\tubing)$ belongs to the half-space~$\HS(\tube)$ and to the boundary of this half-space if and only if~$\tube \in \tubing$.
Consider first a tube~$\tube$ not in~$\tubing^\circ$. Let~$\underline{\tube}$ denote the inclusion minimal tube of~$\tubing \cup \ground$ containing the central vertex~$*$. Then the other tubes of~$\tubing$ are all leaves of~$\starG_{n+1}$ contained in~$\underline{\tube}$ and a nested chain of tubes~$\underline{\tube} = \tube_{|\underline{\tube}|} \subsetneq \tube_{|\underline{\tube}|+1} \subsetneq \dots \subsetneq \tube_{n+1} = \ground$ of~$\starG_{n+1}$. Therefore, we have~$\b{x}(\tubing)_i = 0$ if $\{\ell_i\} \subseteq \underline{\tube}$ and~$\b{x}(\tubing)_i = j-1$ if~$\{\ell_i\} = \tube_j \ssm \tube_{j-1}$. We conclude that
\[
\dotprod{\compatibilityVector{\tubing^\circ}{\tube}}{\b{x}(\tubing)} = \sum_{\substack{i \in [n] \\ \ell_i \in \tube}} \b{x}(\tube)_i = \sum_{\substack{|\underline{\tube}| \le j \le n+1 \\ \tube_j \ssm \tube_{j-1} \not\subseteq \tube}} (j-1) \le \sum_{j = |\tube|}^{n} j = f(|\tube|),
\]
with equality if and only if~$\tube_j \ssm \tube_{j-1} \not\subseteq \tube$ for all~$|\tube| \le j \le n+1$, \ie if and only if~$\tube = \tube_{|\tube|}$.
Finally, for any~$i \in [n]$, we have~$\dotprod{\compatibilityVector{\tubing^\circ}{\{\ell_i\}}}{\b{x}(\tubing)} = -\b{x}(\tubing)_i \le 0$, with equality if and only if the inclusion minimal tube of~$\tubing \cup \ground$ containing~$i$ is~$\{\ell_i\}$, \ie if and only if~$\{\ell_i\} \in \tubing$.
\end{proof}
\subsection{Design compatibility fan (Theorem~\ref{theo:compatibilityFanDesign})}
\label{subsec:proofDesignCompatibilityFan}
The proof of Theorem~\ref{theo:compatibilityFanDesign} still relies on Proposition~\ref{prop:characterizationFan}, that is on the understanding of the linear dependences of the compatibility vectors of the tubes involved in a flip. We now need to distinguish two kinds of flips: we call \defn{round flips} those exchanging two round tubes, and \defn{square flips} those exchanging a square to a round tube.
We claim that Theorem~\ref{theo:compatibilityFanRefined} still holds for round flips. Indeed, a coordinatewise verification shows that the linear dependences exhibited in Section~\ref{subsec:proofCompatibilityFan} still hold for a initial maximal design tubing~$\tubing^\circ$: the arguments are identical for coordinates corresponding to round tubes and straightforward for coordinates corresponding to square tubes.
It thus remains to show the \textbf{Separating Flip Property} for square flips. It turns out that the proof for square flips is much easier as the linear dependences only involve compatibility vectors of forced tubes. Using the duality trick presented in Section~\ref{subsec:proofDualCompatibilityFan} and the fact that any two maximal design tubings are connected by a sequence of square flips, it is equivalent to prove the \textbf{Separating Flip Property} for the compatibility fan or for the dual compatibility fan. In the sequel, we prefer to work with the dual compatibility vectors.
Fix an initial maximal design tubing~$\tubing^\circ$ on a graph~$\graphG$. Consider a round tube~$\tube$ exchangeable with a square tube~$\squareTube{v}$, that is the vertex~$v$ is contained in~$\tube$. The forced tubes of this square flip (\ie the tubes contained in all flips exchanging~$\tube$ and~$\squareTube{v}$) are the following:
\begin{itemize}
\item the square tubes~$\squareTube{w_1}, \dots \squareTube{w_k}$ for the neighbors~$w_1, \dots, w_k$ of~$\tube$ in~$\graphG$,
\item the round tubes~$\tube[a]_1,\dots,\tube[a]_{\ell}$ given by the connected components of~$\graphG{}[\tube\ssm \{v\}]$.
\end{itemize}
Let~$p$ be number of~$w_i$'s which are roots of initial round tubes in~$\tubing^\circ$ and let~$q$ be the number of~$\tube[a]_i$'s containing an initial square tube of~$\tubing^\circ$. Suppose that
\begin{itemize}
\item the~$w_i$'s are ordered such that~$w_1, \dots, w_p$ are roots of tubes~$\tube_1^\circ, \dots, \tube_p^\circ$ in~$\tubing^\circ$ with~$\tube_i^\circ \not\supseteq \tube_j^\circ$ for~$1 \le i < j \le p$, while~$\squareTube{w_{p+1}}, \dots, \squareTube{w_k}$ are square tubes of~$\tubing^\circ$,
\item the~$\tube[a]_i$'s are ordered such that~$\tube[a]_1,\dots,\tube[a]_q$ contain an initial square tube of~$\tubing^\circ$ while~$\tube[a]_{q+1},\dots,\tube[a]_{\ell}$ do not.
\end{itemize}
With these notations, the reader can check that if~$p = 0$ and~$q > 0$, then the dual compatibility vectors of these forced tubes satisfy the dependence
\begin{equation}
\label{eq:designTrivialDependence}
\dualCompatibilityVector{\tube}{\tubing^\circ} = \sum_{j \in [q]} \big(\dualCompatibilityVector{\tube[a]_j}{\tubing^\circ} - \dualCompatibilityVector{\squareTube{v}}{\tubing^\circ}\big)
\end{equation}
We now adapt this linear dependence to cover the case when~$p \ge 0$ and~$q > 0$. Namely, we claim that
\[
\dualCompatibilityVector{\tube}{\tubing^\circ} = \sum_{j \in [q]} \big(\dualCompatibilityVector{\tube[a]_j}{\tubing^\circ} - \dualCompatibilityVector{\squareTube{v}}{\tubing^\circ}\big) + \sum_{i \in [p]}\beta_i \, \dualCompatibilityVector{\squareTube{w_i}}{\tubing^\circ}
\]
where the coefficients~$\beta_1, \dots, \beta_p$ are recursively defined by
\begin{equation}
\label{eq:designFullDependence}
\beta_i = \compatibilityDegree{\tube}{\tube_i^\circ} - \bigg(\sum_{j \in [q]} \big(\compatibilityDegree{\tube[a]_{j}}{\tube_i^\circ} -\compatibilityDegree{\squareTube{v}}{\tube_i^\circ}\big) + \sum_{r \in [i-1]} \beta_r \, \compatibilityDegree{\squareTube{w_r}}{\tube_i^\circ} \bigg).
\end{equation}
To prove it, we check this linear dependence coordinate by coordinate. It boils down to Equation~\eqref{eq:designTrivialDependence} for initial tubes contained in~$\tube$. It is also clear for the initial square tubes not contained in~$\tube$ as all dual compatibility degrees involved in~\eqref{eq:designFullDependence} vanish. Moreover, the coefficients~$\beta_1, \dots, \beta_p$ are defined in order to compensate for the default of the initial tubes~$\tube_1^\circ, \dots, \tube_p^\circ$ in Equation~\eqref{eq:designTrivialDependence}. Finally, we proceed by induction for the remaining initial tubes of~$\tubing^\circ$, that is the initial round tubes not contained in~$\tube$ and whose root is not one of the~$w_i$'s. Namely, since~$q > 0$ and thanks to our special ordering on the $w_i$'s, the coordinate corresponding to such a tube in all terms of Equality~\eqref{eq:designFullDependence} is a linear combination of the coordinates corresponding to its predecessors in the spine of~$\tubing^\circ$ (\ie the inclusion poset on round tubes of~$\tubing^\circ$).
Finally, we still have to check the case where~$q = 0$, meaning where~$\tube$ is contained in an initial round tube of~$\tubing^\circ$. We denote by~$\tube^\circ$ the inclusion minimal initial round tube of~$\tubing^\circ$ containing~$\tube$ and by~$z^\circ$ its root in~$\tubing^\circ$. We then have to distinguish whether or not~$z^\circ=v$:
\begin{itemize}
\item If~$z^\circ = v$, call~$w_1,\dots,w_r$ ($r \ge 1$) the neighbors of~$\tube$ that also belong to~$\tube^\circ$. We then claim that the following linear dependence holds:
\[
\dualCompatibilityVector{\tube}{\tubing^\circ} + r \, \dualCompatibilityVector{\squareTube{v}}{\tubing^\circ}= \sum_{i \in [r]} \dualCompatibilityVector{\squareTube{w_i}}{\tubing^\circ}.
\]
Indeed an initial tube~$\tube[s]^\circ$ of~$\tubing^\circ$ either contains~$\tube$ and thus all the vertices~$v, w_1, \dots, w_r$, or does not contain~$v$ so that the equality holds by counting the vertices~$w_1, \dots, w_r$ in~$\tube[s]^\circ$.
\item If~$z^\circ \ne v$, then~$z^\circ$ belongs to one of the tubes~$\tube[a]_1, \dots, \tube[a]_\ell$, say~$\tube[a]_1$. Let~$w_1,\dots,w_r$ be the neighbors of~$\tube$ contained in~$\tube^\circ$ that are also neighbors of~$\tube[a]_1$, and~$w_{r+1},\dots,w_s$ be the other neighbors of~$\tube$ contained in~$\tube^\circ$. Observe that~$s \ge 1$ since~$\tube \subset \tube^\circ$. One can then check that the linear dependence we look for is
\[
(s-r+1) \, \dualCompatibilityVector{\tube}{\tubing^\circ} = s \, \bigg( \dualCompatibilityVector{\tube[a]_1}{\tubing^\circ}-\dualCompatibilityVector{\squareTube{v}}{\tubing^\circ}-\sum_{i = r +1}^s \dualCompatibilityVector{\squareTube{w_i}}{\tubing^\circ} \bigg) + (s-r+1)\sum_{i = 1}^s \dualCompatibilityVector{\squareTube{w_i}}{\tubing^\circ}.
\]
It clearly holds for tubes not contained in~$\tube^\circ$ and for~$\tube^\circ$ itself. Consider thus an initial round tube~$\tube[s]^\circ$ of~$\tubing^\circ$ contained in~$\tube^\circ$. Since~$z^\circ \in \tube[a]_1$, $\tube[s]^\circ$ cannot contain~$\tube[a]_1$. Thus the first term of the right hand side vanishes for~$\tube[s]^\circ$, while~$\compatibilityDegree{\tube}{\tube[s]^\circ} = \sum_{i \in [s]} \compatibilityDegree{\squareTube{w_i}}{\tube[s]^\circ}$, concluding the proof. \qed
\end{itemize}
\subsection{Design nested complex isomorphisms (Proposition~\ref{prop:isomorphismDesignNormal} and Theorem~\ref{theo:designNestedComplexIsomorphisms})}
\label{subsec:proofDesignIsomorphisms}
We now prove our characterization of the design nested complex isomorphisms announced in Section~\ref{subsec:designNestedComplex}. In both proofs, we will use the following description of links in design nested complexes, similar to that of~\cite{CarrDevadoss} for links in nested complexes. We leave this proof to the reader.
\begin{lemma}
\label{lem:linksDesign}
The link of a tube~$\tube$ in the design nested complex~$\designNestedComplex(\graphG)$ is isomorphic,
\begin{itemize}
\item for a square tube~$\tube = \squareTube{v}$, to the join~$\designNestedComplex(\graphG_1) \ast \dots \ast \, \designNestedComplex(\graphG_\ell)$ of the design nested complexes of the connected components~$\graphG_1, \dots, \graphG_\ell$ of~$\graphG{}[\ground \ssm \{v\}]$.
\item for a round tube~$\tube$, to the join of the nested complex~$\nestedComplex(\graphG{}[\tube])$ of the restriction of~$\graphG$ to~$\tube$ with the design nested complex~$\designNestedComplex(\graphG^\star\tube)$ of the reconnected complement of~$\tube$ in~$\graphG$.
\end{itemize}
\end{lemma}
Before proving Proposition~\ref{prop:isomorphismDesignNormal} and Theorem~\ref{theo:designNestedComplexIsomorphisms}, we need a technical result.
\begin{lemma}
\label{lem:groundImage}
Let~$\bar\graphG$ and~$\graphG$ be connected graphs with vertex sets~$\bar\ground$ and~$\ground$ respectively. Let~$\Phi$ be an isomorphism from the design nested complex~$\designNestedComplex(\bar\graphG)$ of~$\bar\graphG$ to the design (resp.~standard) nested complex~$\designNestedComplex(\graphG)$ (resp.~$\nestedComplex(\graphG)$) of~$\graphG$. If there exists a vertex~$v \in \ground$ such that all tubes of~$\graphG$ incompatible with~$\Phi(\bar\ground)$ are round tubes containing~$v$, then~$\bar\graphG$ and~$\graphG$ are paths.
\end{lemma}
\begin{proof}
Since~$\Phi$ is an isomorphism, it induces a bijection between the design tubes of~$\bar\graphG$ incompatible with~$\bar\ground$ and the tubes of~$\graphG$ incompatible with~$\Phi(\bar\ground)$. The former are precisely the square tubes of~$\bar\graphG$ (by definition of the compatibility of design tubes) while the later are some round tubes of~$\graphG$ containing~$v$ (by assumption). Since all the square tubes of~$\bar\graphG$ are compatible, it follows that their images by~$\Phi$ are nested in~$\graphG$. Let~$\bar{u}, \bar{v} \in \bar\ground$ be the only vertices such that~$|\Phi(\squareTube{\bar{u}})| = 1$ and~$|\Phi(\squareTube{\bar{v}})| = |\ground|-1$. For any other vertex~$\bar{w} \in \bar\ground$, the link of the tube~$\Phi(\squareTube{\bar{w}})$ is the join of two non-trivial design or standard nested complexes by Lemma~\ref{lem:linksDesign}. Since~$\Phi$ is an isomorphism, so is the link of the square tube~$\squareTube{\bar{w}}$, so that~$\bar{w}$ disconnects~$\bar\graphG$ again by Lemma~\ref{lem:linksDesign}. We conclude that all but two vertices of~$\bar\graphG$ disconnect~$\bar\graphG$, which implies that~$\bar\graphG$ is a path. In order to show that~$\graphG$ is also a path, we now distinguish two situations:
\begin{enumerate}[(i)]
\item Suppose first that~$\Phi$ is an isomorphism from the design nested complex~$\designNestedComplex(\bar\graphG)$ of~$\bar\graphG$ to the standard nested complex~$\nestedComplex(\graphG)$ of~$\graphG$. Then the composition of~$\Phi^{-1}$ with the isomorphism~$\Pi$ of Example~\ref{exm:isomorphismDesignNormal}\,(ii) is an isomorphism between the standard nested complex~$\nestedComplex(\graphG)$ of~$\graphG$ and the standard nested complex of a path, which implies by Theorem~\ref{theo:nestedComplexIsomorphisms} that~$\graphG$ itself is a path.
\item Suppose now that~$\Phi$ is an isomorphism from the design nested complex~$\designNestedComplex(\bar\graphG)$ of~$\bar\graphG$ to the design nested complex~$\designNestedComplex(\graphG)$ of~$\graphG$. Using Lemma~\ref{lem:linksDesign} and a similar argument as in the proof of Lemma~\ref{lem:isomorphismsPreserveConnectedComponents}, the image~$\Phi(\bar\ground)$ of~$\bar\ground$ is either a square tube~$\squareTube{v}$ of~$\graphG$, or a singleton round tube~$\{v\}$ of~$\graphG$, or the round tube~$\ground$ of~$\graphG$. The last two cases are discarded by our assumption since the square tube~$\squareTube{v}$ is incompatible with~$\{v\}$ and with~$\ground$. Therefore, we obtain that~$\Phi(\bar\ground) = \squareTube{v}$ and the tubes of~$\graphG$ incompatible with~$\Phi(\bar\ground) = \squareTube{v}$ are exactly the round tubes containing~$v$. So the square tubes of~$\bar\graphG$ are in bijection with the tubes of~$\graphG$ containing~$v$. Since~$|\bar\ground|=|\ground|$ (dimensions of isomorphic simplicial complexes), an immediate induction shows that~$\graphG$ is a path. \qedhere
\end{enumerate}
\end{proof}
We are now ready to prove our classification of design nested complex isomorphisms.
\begin{proof}[Proof of Proposition~\ref{prop:isomorphismDesignNormal}]
We show the result by induction on~$|\bar\ground|$, the cases~$|\bar\ground| \le 2$ being trivial. Assume that~$|\bar\ground| \ge 3$ and consider an isomorphism~$\Phi:\designNestedComplex(\bar\graphG)\to\nestedComplex(\graphG)$.
We first consider the image~$\Phi(\bar\ground)$ of the round tube~$\bar\ground$ of~$\bar\graphG$. It is either a singleton, or the complement of a singleton. Otherwise, its link would be the join of two non-trivial complexes by Lemma~\ref{lem:linksDesign}, which yields a contradiction using a similar argument as in the proof of Lemma~\ref{lem:isomorphismsPreserveConnectedComponents}. We claim that we can assume that~$\Phi(\bar\ground)$ is a singleton adjacent to at least two vertices in~$\graphG$. Indeed,
\begin{itemize}
\item if~$\Phi(\bar\ground) = \ground \ssm \{v\}$, then all tubes incompatible with~$\Phi(\bar\ground) = \ground \ssm \{v\}$ contain~$v$,
\item if~$\Phi(\bar\ground) = \{v\}$ where~$v$ has a unique neighbor~$w$ in~$\graphG$, then all tubes incompatible with~$\Phi(\bar\ground) = \{v\}$ contain~$w$,
\end{itemize}
In both cases, Lemma~\ref{lem:groundImage} ensures that~$\bar\graphG$ and~$\graphG$ are paths, and we can compose~$\Phi$ with a rotation~$\rot^p$ to ensure that~$\Phi(\bar\ground)$ is a singleton adjacent to at least two vertices in~$\graphG$.
We now consider a vertex~$\bar{w} \in \bar\ground$ which does not disconnect~$\bar\graphG$ (such a vertex always exists). By Lemma~\ref{lem:linksDesign}, the link of the square tube~$\squareTube{\bar{w}}$ is not the join of two non-trivial nested complexes. Since~$\Phi$ is an isomorphism, so is the link of~$\Phi(\squareTube{\bar{w}})$, so that~$\Phi(\squareTube{\bar{w}})$ is either a singleton or the complement of a singleton again by Lemma~\ref{lem:linksDesign}. Moreover, since~$\squareTube{\bar{w}}$ is incompatible with~$\bar\ground$ and~$\Phi$ is an isomorphism, $\Phi(\squareTube{\bar{w}})$ is incompatible with~$\Phi(\bar\ground) = \{v\}$. Therefore, $\Phi(\squareTube{\bar{w}})$ is either a singleton~$\{w\}$ adjacent to~$\{v\}$, or the complement~$\ground \ssm \{v\}$ of the singleton~$\{v\}$. Now, since~$\bar\ground$ contains at least two vertices~$\bar{w}, \bar{w}'$ which do not disconnect~$\graphG$ and since~$\Phi(\squareTube{\bar{w}}) \ne \Phi(\squareTube{\bar{w}'})$, we can assume that~$\Phi(\squareTube{\bar{w}})$ is a singleton~$\{w\}$ adjacent to~$\{v\}$.
It implies that~$\Phi$ induces an isomorphism from the design nested complex~$\designNestedComplex(\bar\graphG{}[\bar\ground\ssm\{\bar{w}\}])$ to the nested complex~$\nestedComplex(\graphG^\star\{w\})$ of the reconnected complement of the tube~$\{w\}$ in~$\graphG$. The induction hypothesis implies that
\begin{itemize}
\item the graph~$\bar\graphG{}[\bar\ground\ssm\{\bar{w}\}]$ is isomorphic to the spider~$\spiderG_{\underline{n}}$ and the graph~$\graphG^\star\{w\}$ is isomorphic to the octopus~$\octopusG_{\underline{n}}$ (with head denoted~$v$), for a certain~$\underline{n} = \{n_1, \dots, n_{\ell}\} \in \N^\ell$.
\item the image~$\Phi(\bar\ground\ssm\{\bar{w}\})$ of~$\bar\ground\ssm\{\bar{w}\}$ is the pair~$\{v,w\}$ containing the central vertex of~$\graphG^\star\{w\}$,
\item the description of the images of the tubes of~$\bar\graphG$ not containing~$\bar{w}$ is given by~$\bar\Omega$, in particular~$\Phi(\squareTube{\bar{w}'})$ is a singleton adjacent to~$v$ in~$\graphG^\star\{w\}$. Since~$\Phi(\squareTube{\bar{w}'})$ cannot be an edge in~$\graphG$, it then has to be also a singleton~$\{w'\}$ in~$\graphG$, non-adjacent to~$w$ for compatibility.
\end{itemize}
Since~$\graphG^\star\{w\}$ is an octopus with head~$v$, it follows by definition of the reconnected complement that~$\graphG$ is either an octopus with head~$v$ or an octopus with head~$v$ with an additional edge of the form~$\{v,v_i^1\}$ for a certain~$i\in[\ell]$. We can now apply the same reasonning to~$\bar{w}'$ and since~$\Phi(\squareTube{\bar{w}})$ and~$\Phi(\squareTube{\bar{w}'})$ are non-adjacent singletons in~$\graphG$, we conclude that~$\graphG$ is an octopus. Since~$\Phi(\bar\ground)=\{v\}$, the graph~$\bar\graphG$ is a spider and the reader can check that the restriction of~$\Phi$ to the link of~$\bar\ground$ in the design nested complex~$\designNestedComplex(\bar\graphG)$ is the non-trivial isomorphism~$\Omega:\nestedComplex(\bar\graphG)\to\nestedComplex(\graphG^\star\{v\})$ defined in Section~\ref{subsec:many}. It follows that~$\Phi$ coincides with~$\bar\Omega$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{theo:designNestedComplexIsomorphisms}]
If~$\graphG$ is the path~$\pathG_n$, then its design nested complex~$\designNestedComplex(\graphG)$ is isomorphic to the nested complex~$\nestedComplex(\pathG_{n+1})$ by Example~\ref{exm:isomorphismDesignNormal}\,(ii). Therefore, the design nested complex~$\designNestedComplex(\graphG')$ is also isomorphic to the nested complex~$\nestedComplex(\pathG_{n+1})$, which implies that~$G'$ is the path~$\pathG_n$ by Proposition~\ref{prop:isomorphismDesignNormal}. We can therefore assume that~$\graphG$ and~$\graphG'$ are not paths.
Let~$\ground$ and~$\ground'$ denote the vertex sets of~$\graphG$ and~$\graphG'$. By Lemma~\ref{lem:linksDesign}, the link of~$\ground$ in~$\designNestedComplex(\graphG)$ is the nested complex~$\nestedComplex(\graphG)$. With a similar argument as in the proof of Lemma~\ref{lem:isomorphismsPreserveConnectedComponents}, it follows that its image~$\Phi(\ground)$ is either a square tube of~$\graphG'$, or a singleton round tube of~$\graphG'$, or the round tube~$\ground'$ of~$\graphG'$. We treat these situations separately.
\begin{description}
\item[if~$\Phi(\ground) = \squareTube{v'}$]
All tubes of~$\graphG'$ incompatible with~$\Phi(\ground) = \squareTube{v'}$ are round tubes containing~$v'$. Thus by Lemma~\ref{lem:groundImage}, the graphs~$\graphG$ and~$\graphG'$ are paths, which we already excluded.
\item[if~$\Phi(\ground)=\ground'$]
By Lemma~\ref{lem:linksDesign}, $\Phi$ induces a non-trivial nested complex isomorphism~$\Psi$ from~$\nestedComplex(\graphG)$ to~$\nestedComplex(\graphG')$. It follows from Theorem~\ref{theo:nestedComplexIsomorphisms} that~$\graphG$ and~$\graphG'$ are isomorphic spiders and~$\Psi$ coincides with the non-trivial nested complex isomorphism~$\Omega$ described in Section~\ref{subsec:many}. It thus suffices to show that the nested complex isomorphism~$\Omega$ cannot be extended to design nested complexes. For this observe first that such an extension would send square tubes to square tubes. Now consider a singleton~$\{v\}$ of~$\graphG$ swapped by~$\Omega$. This singleton~$\{v\}$ is incompatible with the square tube~$\squareTube{v}$ and compatible with all other square tubes of~$\graphG$. Yet its image is incompatible with more than one square tube, a contradiction.
\item[if~$\Phi(\ground)=\{v'\}$]
The link of~$\{v'\}$ in the design nested complex~$\designNestedComplex(\graphG')$ is a design nested complex isomorphic to the nested complex~$\nestedComplex(\graphG)$. By Proposition~\ref{prop:isomorphismDesignNormal}, we obtain that~$\graphG$ is an octopus and~$\Phi^{-1}(\ground')$ is the singleton containing its central vertex. Since~$\Phi^{-1}(\ground')\neq\ground$, the same argument applies to~$\Phi^{-1}$ and shows that~$\graphG'$ is also an octopus with the same legs as~$\graphG$. Moreover Proposition~\ref{prop:isomorphismDesignNormal} describes the images of round tubes of~$\graphG$ by~$\Phi$ and of round tubes of~$\graphG'$ by~$\Phi^{-1}$, which together forces~$\Phi$ to coincide with the map~$\Omega\design$. \qedhere
\end{description}
\end{proof}
\section*{Acknoledgements}
\enlargethispage{.4cm}
We thank Francisco Santos for comments and suggestions on a former version of this paper. We also thank an FPSAC referee for pointing out a misleading over-simplification in a previous version of the proof of Theorem~\ref{theo:dualCompatibilityFan}.
\bibliographystyle{alpha}
|
\section{Introduction}
The cosmological recombination epoch marks an extended period over which electrons recombine with protons and Helium nuclei as cosmological expansion and cooling cause the Universe to transition from a fully ionized primordial plasma to a gas of almost completely neutral Hydrogen and Helium atoms. In the cosmological concordance model, the cosmological recombination epoch spans redshifts $500 \lesssim z \lesssim 1600$ for Hydrogen and for Helium recombination the corresponding redshifts are $1600 \lesssim z \lesssim 3500$ for HeII $\rightarrow$ HeI and $5000\lesssim z \lesssim 8000$ for HeIII $\rightarrow$ HeII. At these redshifts, electrons and photons are tightly coupled through energy exchange via Compton scattering and the thermodynamic temperature of electrons and photons are almost exactly equal (at $T\approx 3815\{(1+z)/1400\}$~K). During Hydrogen recombination, once even a small neutral fraction builds up, photons emitted in any free-bound transition to the ground state would almost immediately be re-absorbed by a nearby neutral hydrogen atom, thereby effectively compensating for the electron captured \citep{Zeldovich68, Peebles68, Chluba2007b}. Hydrogen recombination therefore relies on free-bound transitions to excited states followed by a trickle down to the ground state with the subsequent emission of numerous recombination photons. During this process, atoms are frequently photo-ionized by the huge number of CMB photons and after multiple dissociations and recaptures the atoms finally reach the ground state, emitting photons in the Lyman-$\alpha$ resonance or the $2s$--$1s$ two-photon continuum \citep{Zeldovich68, Peebles68} and, at low frequencies, in transitions among excited states \citep{Dubrovich1975, Rybicki93, Dubrovich1997}.
Recombination is expected to be stalled as the ambient radiation temperature in the Lyman-$\alpha$ transition wavelength rises and is balanced in a quasi-static equilibrium with populations in the 1$s$ and 2$p$ states. Recombination thus depends on the depletion rate of $n=2$ atoms and removal of the excess brightness in the ambient Lyman-$\alpha$ line by Hubble expansion and two photon decay from the 2$s$ state. These processes occur at a rate that is $\simeq 10^7-10^8$ times slower than the spontaneous transition probability of the Lyman-$\alpha$ and play a vital role in controlling the dynamics of recombination.
About $57\%$ of all hydrogen atoms become neutral through the $2s$--$1s$ two-photon channel \citep{Chluba2006b}, which has a vacuum decay rate of only $A_{2s1s}\simeq 8.22\,{\rm s^{-1}}$ \citep{Goeppert1931, Breit1940, Spitzer1951, Goldman1989, Labzowsky2005}. Consequently, hydrogen recombination is expected to be substantially delayed compared to what might be expected assuming equilibrium Saha recombination at the average densities and temperatures typical for our Universe.
As the number of photons per baryon is roughly $1.6 \times 10^9$, radiative processes including stimulated recombination, induced emission and absorption of photons dominate the populating of atomic levels rather than collisions \citep[e.g.,][]{Chluba2007, Chluba2010}. As the Universe gradually expands and recombination proceeds by uncompensated bound-bound and free-bound transitions, the level populations of atomic species slowly fall out of equilibrium with the radiation field, which causes departures of the CMB spectrum from an ideal Planck form. The spectral lines corresponding to these transitions are predicted to appear as redshifted additive deviations to the cosmic microwave background (CMB) spectrum, with most of the Hydrogen lines originating from redshift $z\simeq 1300-1400$ \citep{Jose2006,Chluba2006b}; the observable intensity of these spectral lines is furthermore expected to be uniform on the sky and unpolarized. The lines are also substantially broadened owing to the extended recombination time (and electron scattering in the case of HeII recombination). At low frequencies, adjacent lines overlap substantially and hence the cosmological recombination radiation is expected to manifest itself as spectral `ripples' riding on a smooth continuum, which are together an additive component of the extragalactic background light extending over radio, microwaves and near IR wavelengths.
Improvements in our understanding of the recombination history within the framework provided by the concordance cosmology, the role of $2s$--$1s$ two-photon decay and other factors governing the depletion of the $n=2$ level population, the importance of the contribution of Helium to the recombination process, progress in related atomic physics and, last but not the least important, substantial improvements in computing power have all contributed to much improved and more realistic modeling of the epoch in question \citep[see][for overview of recombination physics]{Fendt2009, Jose2010}. Detailed calculations as described in \cite{Sunyaev2009} and references therein provide a fairly precise estimate of the spectral ripples expected due to the recombination of Hydrogen and Helium.
In Fig.~\ref{fig:template}, we show the predictions from 0.1~GHz up to 3000~GHz. At high frequencies we can distinguish features caused by the Lyman-$\alpha$ line, Balmer-continuum and-$\alpha$ line, Paschen-$\alpha$ and Brackett-$\alpha$ lines \citep{Jose2006, Chluba2006, Chluba2007}. Helium contributes locally at a typical level\footnote{Enhancements of features caused by additional feedback effects \citep{Chluba2009c} were neglected for now.} of $\simeq 10\% - 30\%$ \citep{Jose2008, Sunyaev2009}. Owing to the substantial line broadening resulting from the extended period of recombination, at low frequencies the recombination radiation appears as (i) a smooth continuum to the radio, microwave and near IR backgrounds, plus (ii) more easily distinguishable spectral ripples that has a quasi-periodic frequency dependence. It is the detection of these ripply features that we focus on in this work because of its distinctive signature that distinguishes it from any other known spectral features in the CMB.
\begin{figure}[h]
\centering
\includegraphics[width=15.0cm]{predicted_recombination_line_theory.png}
\caption{The additive spectral structure expected in the intensity of the uniform extragalactic background light owing to cosmological recombination of Hydrogen and Helium \citep{Jose2006, Chluba2006b, Chluba2007, Jose2008}. The wide frequency range is successively covered in four panels. The additive spectrum is in units of [mJy\ sr$^{-1}$] in the panels in the top row and [Jy\ sr$^{-1}$] in the bottom row panels.}
\label{fig:template}
\end{figure}
\subsection{The importance of an experimental detection of the recombination radiation}
Direct observations of the predicted ripples -- the unique {\it fingerprint} of the recombination era -- would be a confirmation of the recombination theory, validating our understanding of the associated atomic physics as well as the physical processes in these early times, which in itself is a major motivating factor for experimentally detecting them. A detailed discussion on the dependence of the spectral signatures from the epoch of recombination on various contributing factors that affect their strength, shape and position is presented in \citet{Chluba2008T0} and \cite{Sunyaev2009}. We list here some of the major motivations for an experimental detection.
\begin{enumerate}
\item {\it Determining key cosmological parameters.} Observing the spectral distortions from the epochs of Hydrogen and Helium recombination would, in principle, provide an additional way to determine some of the key parameters of the Universe. For instance, the dependence of the predicted recombination spectrum on $\Omega_{\rm b} h^2$ and $T_0$ is shown in Fig.~3 and 4 of \citet{Chluba2008T0}. It may be noted here that most of the recombination of Hydrogen and Helium and thus the spectral signatures that arise from these processes occur over redshifts {\it before} the formation of the CMB anisotropies that occurred near the peak of the Thomson visibility function: spectral lines from transitions associated with the Hydrogen atoms arise from a redshift range of 1300--1400. Hence detections of the recombination lines with increasing accuracy provides a way of constraining the thermal evolution of the universe beyond the last scattering surface (see below). This also provides a novel method to measure the {\it pre-stellar abundance} of Helium and to break parameter degeneracies by combining with CMB anisotropy measurements. An illustration of the dependence of the predicted final recombination spectrum on contribution from HeIII $\rightarrow$ HeII and HeII $\rightarrow$ HeI is given in Fig.~11 in \cite{Sunyaev2009}.
\item {\it Probing energy release in the pre-recombination era.} Detection of cosmological recombination lines allows us to gain a better understanding of the thermal history of our Universe on the basis of the different redshifts of Hydrogen and Helium recombination. For instance, it is possible to compute the recombination spectrum assuming that the ambient radiation field is a distorted blackbody, where the distortion is of $y$-type \citep{Chluba2008c}. The $y$-distortion \citep{Zeldovich1969} could arise from the blue-shifting of photons by Comptonization in the early universe, due to energy-release at $z\lesssim 50000$, when the redistribution of photons in energy by Compton scattering becomes inefficient \citep{Burigana1991, Hu1993, Chluba2011therm}. This distortion is described by the Compton-$y$ parameter, $y = \int \frac{k (T_{\rm e}-T_\gamma)}{m_{\rm e} c^2}\sigma N_{\rm e} c \,{\rm d}t$. COBE/FIRAS observations place an upper limit of $y< 1.5\times 10^{-5}$ \citep[95\% c.l.;][]{Fixsen1996, Fixsen2009}. The contributions from Hydrogen and Helium to the total recombination spectrum depends both on the value of $y$ parameter and when the distortion was created \citep{Chluba2008c}. Thus the cosmological recombination radiation may allow distinguishing between Compton $y$-distortions that were caused by energy release before or after the epoch of recombination.
\item {\it Testing recombination physics}. Todays most advanced recombination codes \citep{Chluba2010b, Yacine2010c} include many subtle atomic physics and radiative transfer effects \citep[e.g.,][]{Dubrovich2005, Chluba2006, Kholupenko2007, Switzer2007II, Chluba2008a, Hirata2008, Grin2009}. These calculations ensure that the science return from {\it Planck} and upcoming CMB experiments is not compromised by inaccuracies in the recombination model \citep{Jose2010, Shaw2011}. However, the detailed dynamics of the recombination process is also reflected in the shape and position of the recombination features. Any departures of the recombination spectrum from the theoretical predictions will indicate presence of some non-standard process \citep[e.g.,][]{Chluba2008c, Chluba2010a}. Thus, by observing the recombination radiation we can directly confront our understanding of recombination physics with experimental evidence.
\end{enumerate}
\noindent Detecting nK amplitude fluctuations in the extragalactic background brightness, which are indeed a tiny perturbation to the orders of magnitude larger sky brightness temperature that arises from the CMB, extragalactic sources and Galactic emission, is a very challenging problem. However, we opine that receiver technology and the ability to produce detector arrays have progressed to the point where it is meaningful to look at the practical issues. In this first paper, we study the feasibility of such a detection by modeling the sky spectrum as recorded by an ideal instrument and examine whether it is at all possible to recover the weak signal embedded in the substantially brighter Galactic and extragalactic foregrounds. We discuss methods for fitting to data for the recovery or detection of the faint recombination line spectrum when observed embedded in the substantially brighter cosmic radio background. We also compute optimal observing frequencies for the detection and associated signal-to-noise ratio for detection with realistic receivers and a purpose-built array of spectral radiometers with due consideration to contribution from atmospheric emission further guided by radio frequency interference (RFI) over the band. Some of the challenges are very similar to those for the detection of the global 21cm signal \citep[see][]{Patra2013}; however, the recombination ripples benefit from their unique frequency dependence, which is hard to mimic by other sources and instrumental effects.
\section{The signal and noise level for a detection of the recombination radiation}\label{sec:snr}
In this section, we discuss the frequency dependence of the spectral signal and that of the background brightness and additive instrument noise to arrive at estimates for signal-to-noise ratio versus frequency for detection of the line structure. The successive peaks of the recombination spectral features correspond to the spectral lines arising from bound-bound $n \rightarrow (n-1)$ electron transitions between adjacent principal quantum states of Hydrogen, visible at frequency $\nu \simeq 4.7\,{\rm GHz}\, [n/10]^{-3}$ (for emission redshift $z\simeq 1400$). If we subtract a low order baseline component from the recombination spectrum in the 1.5--7.0 GHz band, the resulting spectral structure is as shown in Fig.~\ref{fig:template_ideal}. As a method of removing the foreground and receiver response in an observation with an ideal instrument, if a smooth baseline were to be subtracted from a recorded sky spectrum, Fig.~\ref{fig:template_ideal} is what would be expected as a residual.
\begin{figure}[ht]
\centering
\includegraphics[trim= 30mm 0mm 20mm 10mm, clip,scale = 0.4]{baselin_sub_template_nK_yaxis_fig2.png}
\caption{ The ripply recombination line signal expected to be detected in the 1.5--7.0 GHz band.}
\label{fig:template_ideal}
\end{figure}
We now address the problem of identifying the most suitable frequency range for the experimental detection of recombination lines from the Epoch of Recombination by first considering the relative amplitude of the spectral ripples arising from recombination to the noise in the detection. The total system noise is the additive sum of the spatially-varying sky background, consisting of contributions from discrete and diffuse Galactic and extragalactic sources in the sky, the cosmic microwave background and the cosmic infra-red background, as well as effects of atmospheric emission, plus receiver noise that is at lowest quantum noise associated with zero-point fluctuations. This sets a fundamental limit on the achievable detection sensitivity. Later in this section we also consider detection using receivers that have more realistic noise temperatures consistent with current technology. We also discuss constraints arising from the requirement that within the observing band we need to cover at least a distinctive segment of the quasi-periodic cosmological signal, which might provide a unique template specific to cosmological recombination that would be difficult to mimic by other astrophysical sources, atmospheric emission, RFI or instrumental effects.
\begin{figure}[Ht]
\centering
\includegraphics[ trim= 75mm 0mm 50mm 10mm, clip,scale = 0.5]{sky_background_Chajnantor_final.png}
\caption{The expected intensity of the recombination signal is shown in black. Our model for the intensity equivalent to the minimal total system noise is shown as a red dotted line where the receiver noise is assumed to have ideal quantum noise limited performance. Also shown as a blue dashed line is the intensity corresponding to system noise when observing with a state of the art cryogenically cooled receiver that has a noise temperature of 1 K and in green is the sky intensity when observing with an uncooled receiver assuming a noise temperature of 14 K. All intensities are in units of Jy~sr$^{-1}$ (surface brightness units) with the x-axis in $\log(\nu/{\rm GHz})$.}
\label{fig:lines_vs_sky_tcmb_trec}
\end{figure}
As a representation of the telescope system noise contribution from additive sky radiation, we construct a model spectrum of the background sky brightness all the way from 50~MHz up to 4~THz, extending to just beyond the redshifted Lyman-$\alpha$ line that is essentially the highest frequency at which we may expect spectral features arising from cosmological recombination. We adopt the model presented in \cite{Subrahmanyan2013} for the Galactic and extragalactic contributions to sky brightness and derive sky temperatures towards the Galactic pole at 150, 408 and 1420~MHz. A fit of a power-law form to these brightness estimates yields a temperature spectral index of $-2.5$ and a normalization corresponding to sky brightness of 438~K at 100~MHz; this model spectrum towards the Galactic pole is adopted to represent the contribution to system noise from Galactic emission plus extragalactic discrete sources. We also include a component that is an estimate of the far and mid-infrared background; this model is derived from data in Table~47 of \citet{Leinert1998}. For the receiver noise, we compute the system temperature for two cases: (i) assuming a state of the art cryogenically cooled receiver with noise temperature of 1~K above quantum noise \citep{Schleeh2012} and (ii) that for an uncooled receiver with a more realistic noise temperature of 14~K plus quantum noise \citep{Belostotski2007,Witvers2010}. We use the {\it am} atmospheric model \citep{Paine2004} with typical conditions at the Chajnantor\footnote{https://www.cfa.harvard.edu/\~spaine/am/cookbook/unix/other\_examples/Chajnantor.amc} plateau site to derive the atmospheric opacity versus frequency. The system temperature is finally translated to above atmosphere using this opacity so that it may be combined with the signal intensity derived above for calculating the signal-to-noise ratio. Our adopted model spectrum for the above-atmosphere system temperature is shown in Fig.~\ref{fig:lines_vs_sky_tcmb_trec} along with the spectrum of the expected recombination signal.
\begin{figure}[h]
\centering
\includegraphics[trim=10mm 0mm 10mm 0mm,clip,scale = 0.5]{delf_byf_vs_f_shaded.png}
\caption{Ratio of spacing between peaks of $\alpha$ transitions from $n$ and $n+2$ quantum shells of Hydrogen to the nominal frequency versus nominal observing frequency.}
\label{fig:bw_vs_nu}
\end{figure}
As a criterion for the minimum bandwidth required at any observing frequency we choose a spectral window that includes at least two adjacent broad recombination spectral lines that arise from $n \rightarrow (n-1)$ ($\alpha$) transitions. This minimum bandwidth would correspond to the spacing between peaks of $\alpha$ transitions from $n$ and $n+2$ shells. Such a spectral window would be expected to include a sufficiently high order of the variation in the signal so as to give it a distinctive signature. Fig.~\ref{fig:bw_vs_nu} shows the ratio of this minimum bandwidth to the nominal center frequency versus the center frequency. The detection of Recombination Epoch spectral lines clearly requires a larger fractional bandwidth at higher frequencies. A ratio less than unity implies that an octave bandwidth at the observing frequency would contain more than two adjacent recombination lines, satisfying our criterion. The ratio falls below unity for observing frequencies below about 18~GHz, which implies that detecting the cosmological line spectrum would require bandwidths exceeding an octave if the center frequency were to exceed 18~GHz. Receivers with bandwidths wider than an octave are susceptible to self-generated radio frequency interference from harmonics of system clocks and local oscillator frequencies; therefore, although sensitivity for a detection would undoubtedly improve with wider observing bands, for the ultra-sensitive detection experiment being considered herein it would be unwise to attempt detection at frequencies above this value.
\begin{figure}[ht]
\centering
\includegraphics[trim=30mm 0mm 30mm 0mm,clip,scale = 0.40]{log_signal_strength_vs_log_freq_GHz.png}
\caption{Variation in the logarithm of amplitude of the ripples representing the signal from recombination versus frequency in log(GHz) units.}
\label{fig:sig_cont}
\end{figure}
As an estimate of the detectable signal from Hydrogen recombination, we require a measure of the peak-to-peak spectral variations versus observing frequency. We first fit a spline through the points in the recombination line intensity template that are at frequencies corresponding to the peaks of the $n \rightarrow (n-1)$ Hydrogen transitions. We use a redshift that is a mean value, which we determine by comparing the locations of the peaks in the predicted recombination line spectrum with the corresponding locations given by the Rydberg formula. Subtracting the spline fit gives a residual recombination line spectrum that represents the detectable signal in an observation from which a mean spectral baseline has been subtracted. In Fig.~\ref{fig:sig_cont} we plot the strength of the expected recombination line signal defined as half the peak-to-peak amplitude of the baseline subtracted ripples expected in the observed band, versus observing frequency. With frequency, the expected ripple amplitude increases by orders of magnitude across the frequency range we consider here. It may be noted that although in terms of spectral intensity this ripple signal amplitude increases with frequency, as shown above the number of spectral ripples within any octave bandwidth decreases with increasing frequency. Thus, a frequency regime that maximizes both aspects is sought. (See Sect.~\ref{sec:freq_choice})
\subsection{Signal-to-noise ratio: a matched filter approach to signal detection}
The detection of spectral lines from cosmological recombination involves first subtracting a baseline from the observed spectrum to remove the relatively smoother foregrounds and CMB, after which the residual spectral segment may be examined for the expected spectral ripple using a matched filter. We evaluate here the signal-to-noise ratio for a detection method where an estimate of the amplitude of any spectral ripple that matches a theoretical template is derived from the measurement. The expected signal, following baseline subtraction, is similar to a sinusoidal ripple that has a period that varies systematically across the spectral segment. The noise or uncertainty in the estimate of the amplitude of the ripple is derived from the measurement errors in the measured intensities in the spectral channels, and from the propagation of errors from these measurements to the derived estimate of the amplitude of the signal present in the data. The measurement errors in channel data depend on the system noise, which originates from the sky brightness and receiver noise, and on the channel bandwidth and integration time.
We consider here the ideal scenario where the measurement data---the bandpass calibrated spectrum of the cosmic radio background from which a baseline has been subtracted---contains a spectral ripple that is the same as the expected theoretical template without any other residual contamination. In this case where the signal amplitude is estimated using a matched filter derived from the template, we obtain a signal-to-noise ratio that is limited purely by the noise present in the spectral channels. This case study admittedly represents an optimistic estimate of the signal-to-noise ratio attainable.
The residual spectrum after baseline subtraction is assumed to have $N$ frequency channels, each of bandwidth $\delta b$. The sampled residual sinusoidal ripple, which is also the template of the expected ripply recombination lines following baseline subtraction, is given by $a_k$ and has an amplitude of $A_0$. From this template of the expected signal we define a weighting function $w_k$, which is defined such that the weighted sum of $a_k$ is $A_0$. The matched filter is thus the weighted summation of the observed spectrum and the matched filtering yields an estimate of the amplitude $S$ of the ripple from recombination:
\begin{equation}\label{eq:sum_aw}
S=\sum_{k=0}^{N}a_kw_k
\end{equation}
The template representing the expectation and the corresponding weighting function may be scaled appropriately to search the data for signatures of the ripple from recombination assuming a range of effective redshifts for the recombination lines.
The weighting function is hence similar to the template in that it has the same ripple form but with an amplitude $w_0$. If the signal matches the template, the requirement that $S=\sum_{k=0}^{N}a_kw_k$ ought to equal $A_0$ leads to the condition that $w_0 = 2/N$.
The uncertainty in the channel data in the residual spectrum $a_k$ is
\begin{equation}
\delta a_k = \frac{T}{\sqrt{\delta b \cdot t}},
\end{equation}
where $T$ is the system temperature and $t$ is the total integration time. Propagating this error in the channel data to the error $\delta S$ in the estimate of the amplitude of the ripple from recombination:
\begin{eqnarray}
\delta S & = & \sqrt{\mathlarger{\sum}_{k=0}^{N-1} (\delta a_k)^2 \abs{w_k}^2} \nonumber \\
& = & \frac{T}{\sqrt{\delta b \cdot t}} \cdot \sqrt{\mathlarger{\sum}_{k=0}^{N-1} \abs{w_k}^2} \nonumber \\
& = & T w_0 \sqrt{\frac{N}{2 \delta b \cdot t} }. \nonumber
\end{eqnarray}
The last step above follows from the result that the average value of the square of a sinusoidal wave is half the amplitude. Substituting the earlier result that $w_0 = 2/N$ yields:
\begin{equation}
\delta S = T \sqrt{\frac{2}{N \delta b \cdot t}}.
\end{equation}
Thus the signal-to-noise ratio in the measurement is given by:
\begin{equation}\label{eq:snr}
SNR = \frac{S}{\delta S}= \frac{A_0}{T}\sqrt\frac{N \delta b \cdot t}{2} = \frac{A_0}{T}\sqrt\frac{B \cdot t}{2},
\end{equation}
where $B = N \delta b$ is the total bandwidth of the observed spectrum. For this estimate it is assumed that noise in the different channels is uncorrelated.
\subsection{Choice of Frequency for detecting ripples from cosmological recombination}\label{sec:freq_choice}
\begin{figure}[ht]
\centering
\includegraphics[trim = 30mm 0mm 0mm 0mm, clip,scale = 0.45]{SNR_chajnantor_logGHz.png}
\caption{Signal-to-noise ratio versus nominal observing frequency (in log(GHz) scale) for the detection of spectral-line signatures from recombination epoch with an octave bandwidth. The green line shows the ratio assuming ideal quantum noise limited receivers; the blue line shows the ratio for the more realistic case of receivers with 14~K receiver noise. The red line shows this ratio for a receiver which is cryogenically cooled to 1~K. The observing time is assumed to be 1~s.}
\label{fig:snr_cmb_rec}
\end{figure}
In Fig.~\ref{fig:snr_cmb_rec} we show the estimated signal-to-noise ratio versus observing frequency to guide the choice of observing frequency. In creating this plot the signal is assumed to be half the peak-to-peak magnitude of the ripple, which is the amplitude shown in Fig.~\ref{fig:sig_cont}. The noise is assumed to be the system temperature (from Fig.~\ref{fig:lines_vs_sky_tcmb_trec}) scaled by a factor $\sqrt{2/B}$, where $B$ is the octave bandwidth about a nominal central frequency. The integration time is assumed to be 1~s. The figure shows the signal-to-noise ratio for the case where the system noise corresponds to quantum-noise limited receivers (in green); we also show in red the signal-to-noise ratio for a cryogenically cooled receiver with 1~K noise temperature and in blue this ratio for the case of uncooled 14~K receivers. All three traces are assuming an ideal case where the spectral ripple in the measurement data matches the predicted template exactly.
Detection with maximum signal-to-noise ratio suggests observing in the 1--6~GHz band. If we avoid the lower end of this band where Galactic HI and terrestrial and satellite down-link related radio frequency interference is substantial, the 2--6~GHz band is suggested. Observing in octave bandwidths within this range also satisfies the criterion that at least two cycles of ripples of the recombination line spectrum ought to be contained in the observed spectral segment (see Fig.~\ref{fig:template_ideal}).
\section{A modeling of observations of the cosmological recombination spectrum}
The recombination lines from cosmological recombination are detected by a receiver along with foregrounds, which are averaged by the telescope beam over its response pattern on the sky. The averaging over a multitude of sources with different emission spectra, over the sky and along line of sight, results in a detected spectrum that would have an unknown form: even if the emissivity of the gas is a power-law form at every location, the spectral index does vary across the sky and along line of sight and, therefore, the averaging by the telescope beam would result in an observed spectrum that deviates from a single power law. A similar effect, related to the superposition of blackbodies inside the beam of an experiment, causes an inevitable $y$-type distortion of the CMB spectrum \citep{Chluba2004}; the superposition of blackbodies simply is not a blackbody anymore \citep{Zeldovich1972, Chluba2004, Stebbins2007}.
It is necessary to detect the presence of the recombination line spectrum in the observed spectrum although the line spectrum has amplitude that is {\it almost nine orders of magnitude smaller} and the additive foreground is a complex spectrum of unknown form! By simulating the sky spectrum as would be observed by an ideal system, this section addresses the question of whether such a detection is at all possible.
We choose an octave band from 3 to 6~GHz as the observing band for the simulations presented here. This band also meets our criterion of having at least two recombination spectral lines within the band. In this section we describe a code we have developed, which simulates an ideal receiver system observing the sky over the identified frequency range, and generates the temperature spectrum of the sky as would be produced by such an instrument. A calibration method is included in the pipeline.
We use all-sky maps at 408~MHz \citep{Haslam1982}, 1420~MHz \citep{Reich1982, Reich1986, Gorski2005} and at 23~GHz (WMAP science data product\footnote{WMAP Science Team}) as input maps to estimate the combined galactic and extragalactic brightness contribution to the sky spectrum between 3 and 6~GHz. The sky brightness in the Rayleigh-Jeans limit, at every sky pixel and over frequency channels spaced 10 MHz apart is derived from these three input temperature maps using a linear interpolation in log-log space.
We derive a 408-MHz all sky map from the corresponding raw data product available in the LAMBDA\footnote{http://lambda.gsfc.nasa.gov/} website. The non-destriped 408-MHz map made with a $0\fdg85$ beam that is available in the HEALPix \citep{Gorski2005} R9 nested ordering scheme is first smoothed with a Gaussian beam of FWHM $31\farcm60$ to obtain a map with $1^{\circ}$ resolution. Following this we degrade the map to an R8 nested ordering representation. This final 408-MHz map used in the simulation is a nested R8 HEALPix map in Galactic coordinates with a resolution corresponding to beam FWHM $1^{\circ}$ and with temperature in mK brightness units.
The 1420-MHz all sky map was derived from the data available in the OCEANCOLOR\footnote{http://oceancolor.gsfc.nasa.gov/AQUARIUS/DinnatEtAl2010/} website. This data is in Kelvin units with a beam size of $0\fdg6$ and a pixel size $0\fdg25$ and is in J2000 celestial coordinates. We perform gridding convolution of the data with a Gaussian of appropriate FWHM to yield an image with final resolution of $1^{\circ}$. We then transform the map to Galactic coordinates using the celestial HEALPix coordinate pixel listing for the R8 nested ordering scheme. Thus the final all-sky map we use at 1420 MHz has a beam FWHM $1^{\circ}$ and is in Galactic coordinates ordered in the nested R8 HEALPix scheme.
The deconvolved form of the WMAP 23~GHz all sky map from the LAMBDA website was smoothed to $1^{\circ}$ resolution and formatted in Galactic coordinates in the HEALPix R8 nested ordering scheme. The image units were transformed from thermodynamic to brightness temperature. The image intensities correspond to a differential measurement; therefore, the map is uncertain in its zero point, which was arbitrarily set. We assume a plane-parallel slab model for the galaxy and use the method adopted in \citet{Kogut2011} to estimate this uniform component at 23~GHz using all sky images of the absolute brightness of the sky at 150, 408 and 1420~MHz. The CMB monopole, dipole and WMAP ILC images were subtracted from all these images and the pixels with substantial contamination from discrete sources were blanked using the same blanking image that was used in the WMAP analysis. Pixels in each of the images were binned in cosecant Galactic latitude (cosecant($b$)) over the range 1.0--4.0. The run of median pixel intensity versus mean cosecant($b$) were examined in log-log coordinates. A straight line fit to this plot yields an estimate of the mean extragalactic background at each frequency. It may be noted here that we have restricted this analysis to the southern Galactic hemisphere for consistency with the WMAP analysis. For the case of the 23~GHz image, the fit yields an intercept of $0$K with an accuracy of 1 part in $10^9$ as expected for an image has been constructed to have its intercept arbitrarily set to zero. The intercepts estimated for the lower frequency images were extrapolated using a polynomial fit to get the `missing' uniform background in the 23~GHz image. The resulting value of $493~\mu$K was added as a constant to the 23~GHz image.
The three maps representing the Galactic and extragalactic radio emission (minus CMB) at 408~MHz, 1420~MHz and 23~GHz were interpolated at every image pixel separately to estimate sky temperatures at frequencies between 3 and 6~GHz. We interpolate using a first order polynomial in log-brightness-temperature versus log-frequency space, considering the spectrum at every pixel to be a smooth power-law in linear space. The total sky brightness toward any sky pixel and frequency is estimated as the sum of the galactic and extragalactic sky brightness derived from this polynomial interpolation, plus a uniform cosmic microwave background brightness computed using the Planck formula, plus a uniform component corresponding to the weak cosmological recombination line spectrum, as shown in Fig.~\ref{fig:template}.
\begin{figure}[Ht]
\centering
\includegraphics[scale = 0.55]{sample_spec_24nov13_00_1min.png}
\caption{ Sample synthetic spectrum recorded by an ideal receiver system discussed in the text. The receiver is assumed to be observing with an antenna with a cos$^2(ZA)$ beam pattern pointed at the zenith.
}
\label{fig:sample_spectrum}
\end{figure}
\begin{figure}[Ht]
\centering
\includegraphics[trim=10mm 0mm 0mm 0mm,clip,scale = 0.5]{waterfall_24h_python.png}
\caption{ A time-frequency plot showing the synthetic sky spectra as observed over 24~hours. One can see the galaxy rise and set as the sky drifts over the instrument with each spectrum recorded one minute apart.}
\label{fig:spectrum_24hr}
\end{figure}
The observing system is assumed to have a single antenna element with a frequency independent radiation pattern that has a cos$^2(ZA)$ form, where $ZA$ denotes zenith angle. The antenna bore sight (the optical axis the antenna) is assumed to be static on the ground at the observing site and directed towards zenith at all times. We neglect polarization in the sky intensity as well as the telescope response. Separately, the spectral radiometer is assumed to be a correlation spectrometer and hence does not respond to receiver noise \citep[see][]{Patra2013}. The spectral response of the observing system at any instant is then simply the weighted sum of spectra towards every pixel in the sky where the weighting is by the beam power pattern towards that sky direction. As stated above, the spectrum towards any sky pixel is not expected to be a simple power law because it is an average along the line of sight of regions that might have different spectral indices. The weighted average spectrum is an average over the sky and along the line of sight and hence even if each emitting region is of power law form the average observed spectrum may have a complex form that is unknown.
Our simulation code has the geographic location of the observing site and LST as free parameters. Our code can be easily modified to change the antenna beam pattern as well.
The generation of spectra takes into account effects of atmospheric refraction and precession; astronomical aberration is negligible even in its severest form for the problem considered here and is dropped from our calculations. Calibration is assumed to be done by recording spectra with a hot (373.0~K) and separately a cold (273.0~K) load on the antenna and dividing spectra recorded on the sky by the difference between spectra recorded with the hot and cold loads. Bandpass calibration of sufficiently high precision is assumed. Our code generates mock calibrated spectra over time whose mean temperature varies as the sky and Galactic Plane drift across the telescope beam. A sample spectrum of the sky is shown in Fig.~\ref{fig:sample_spectrum}.
Figure~\ref{fig:spectrum_24hr} shows a waterfall plot to illustrate the variation over 24~hours in the recorded spectra observed by this ideal instrument as the sky drifts across the antenna.
The synthetic sky spectra, whose amplitude is of the order of a few K, contains the recombination line spectrum as a small additive component and is representative of mock observations made with a correlation spectral radiometer that does not respond to receiver noise. The challenge is to distinguish the weak signal corresponding to the epoch of cosmological recombination, which is buried in the observed spectrum as a broadband quasi-periodic sinusoid with peak-to-peak amplitude of order $\sim10$~nK. {\it Not only is the signal a tiny fraction of the total sky spectrum, the recombination line spectrum is an additive component of a foreground spectrum whose functional form is complex and unknown.} In the next section we discuss methods to detect the cosmological recombination spectrum and challenges therein.
\section{The detection of signatures of cosmological recombination}
\label{sec:fitting}
In the previous section we have described the generation of a synthetic spectrum of the radio sky between 3 and 6~GHz, as would be observed by an ideal instrument. The signal from cosmological recombination is expected to appear as quasi-periodic ripples, some $9$ orders of magnitude smaller than the galactic and extragalactic foreground spectrum in which it is additively concealed. The discernment of the recombination signal in such a total spectrum, even under ideal conditions, is pivotal in answering the question of whether an experimental detection is indeed possible. Below we discuss the challenges involved in and describe a possible method for the detection of the cosmological recombination signal in synthetic sky spectra.
The brightness of the sky as measured by a total power radiometer is the sum of the brightness contributions from all the discrete and diffuse Galactic and extragalactic sources that lie within its beam, along the line of sight and across the sky, including the cosmic microwave background and other cosmological emissions. The precision with which the foreground needs to be modeled so that a subtraction of the model might reveal the cosmological recombination spectrum is extreme. Although the discrete sources lying in the beam might have well-measured spectra and the diffuse sky radiation might have all-sky maps at multiple frequencies, the functional form of the final cumulative spectrum of the foreground is not known {\it a priori} to the required accuracy. The foreground has necessarily to be fit to the measurement, which implies that the functional form used for the fit to the foreground needs to be of a form that accurately fits to the foreground without also fitting to the recombination spectral features. Only then may a fitting of a foreground model to the measurement set and its subsequent subtraction reveal the recombination line features in the residual.
The thermal and non-thermal processes (e.g., synchrotron and free-free emission) that contribute to the foregrounds have spectra that are smooth because they arise from impulsive emissions in time domain. The electron energy spectra in astrophysical objects and diffuse matter are also believed to be smooth over decades in energy space. Hence the cumulative spectra in thermal and non-thermal emissions are expected to be featureless over octave bandwidths and, therefore, ought to be distinguishable in principle from the cosmological recombination radiation.
The problem of recovering broad and weak spectral deviations that are buried in a total sky spectrum that is several order of magnitudes brighter is not unique to cosmological hydrogen recombination lines. All-sky (also referred to as global) spectral deviations predicted to arise from the Epoch of Reionization (EoR) also face a similar challenge in that broad spectral features of about 10--100~mK brightness are buried in foregrounds of several hundred K, the foreground being $\sim5$ orders of magnitude brighter.Thus the importance of a method to effectively fit a smooth functional form to the combined Galactic and extragalactic foreground over 1-2 octaves of bandwidth, in which the signal of interest is itself not lost in the process, cannot be understated.
\begin{figure}[Ht]
\centering
\includegraphics[scale = 0.5]{unconstrained_residual_nK.png}
\caption{ The residual obtained on subtracting a sample sky spectrum by an eight order polynomial fitting function in log-frequency and log-temperature space. While one would expect to see signatures of the spectral lines arising from cosmological recombination in the residual at the nK level, the residual is dominated only by thermal noise with the lines themselves having been absorbed by the fitting function.}
\label{fig:unconstrained_residual}
\end{figure}
In the context of detecting global EoR signal in the all-sky radio background spectrum, a polynomial functional form of high order has been adopted in the literature as the analytic function to fit to the foreground \citep{Bowman2010}. However, broad spectral lines of interest that are present in observed spectra as additive components may also be absorbed in the polynomial fit, particularly if the polynomial is of high order, such that the residual would have little trace of the cosmological signal. As an illustration, on fitting a sample spectrum generated by our simulation with an eighth-order polynomial in log-frequency versus log-temperature space without any constraints on the nature of the polynomial, and subtracting the fit spectrum from the original spectrum, we obtain a residual as shown in Fig.~\ref{fig:unconstrained_residual}. While one might hope that the polynomial fit to the foreground would leave the recombination line signal untouched, comparison with Fig.~\ref{fig:template_ideal} shows clearly that this is not the case. The ripples from cosmological recombination, the very signal that we are looking to detect, has been eliminated by the fitting process.
What is needed is a careful choice of the functional form that would leave the cosmological signals of interest almost wholly in the residue. Also needed is a method of successive approximation that would model the foreground to the required accuracy so that the faint cosmological signal might dominate the residue.
\subsection{A complete monotone approach to foreground modeling}
The CMB component has a well defined functional form. The fitting function $f_{\rm fg}(\nu)$ for the foreground may be defined to be an analytic function over the frequency range of interest, so that the model does not admit discontinuities and its derivatives are defined. The functional form describing the foreground model is expected to be `smooth' in the sense that it {\it must not be able to} fit to additive signals that have the sinusoidal structure expected of the recombination spectrum.
A potentially useful functional form is that of a {\it completely monotonic} function. Mathematically, a function $f(x)$ is said to be a complete monotone if for all values of $x$ in the interval $0 \le x < \infty$, $(-1)^n\times\frac{\textrm{d}^nf(x)}{\textrm{d}x^n} \ge 0$ for every integer $n \ge 0$. An example of a complete monotone function is $f(x)=1/(a+b x)^c$, where $a \ge 0$, $b \ge 0$ and $c \ge 0$. This implies that power-law form spectra with negative spectral indices are completely monotonic functions. It is also known that if $f(x)$ and $g(x)$ are completely monotonic, then $a f(x) + b g(x)$ where $a$ and $b$ are non-negative constants is also completely monotonic, which implies that the sum of power law spectra with negative indices would also be completely monotonic. A function $f(x)$ may also be completely monotonic over finite range $a < x < b$ where $a \ge 0$ and $b \ge 0$: this would be the case if the condition $(-1)^n\times\frac{\textrm{d}^nf(x)}{\textrm{d}x^n} \ge 0$ for every integer $n \ge 0$ holds over the range $a < x < b$. The mathematical definition suggests that if $f_{\rm fg}(\nu)$, where $\nu$ is frequency, is the foreground model over a certain bandwidth, then we may adopt a functional form for $f_{\rm fg}(\nu)$ that is a complete monotone within the bandwidth of interest.
Successive approximation of an analytic functional form to data may be done using a Taylor approximation. The Taylor series of an analytic function always converges about every point in its domain. We may represent $f_{\rm fg}(\nu)$ as a polynomial whose coefficients are constrained so that $f_{\rm fg}(\nu)$ is completely monotonic and expand the function as a Taylor polynomial by successively estimating the coefficients of the polynomial, stopping when the degree of the polynomial is such that it is a sufficiently good fit to the data ({\it i.e.} the observed sky spectrum) and the residual is dominated by the embedded cosmological recombination lines. It may be noted here that increasing further the order of the polynomial, whose coefficients are constrained such that the polynomial is a complete monotone, will only yield vanishingly smaller coefficients for the higher order terms. Since the constraint on the polynomial forces it to remain smooth, once the recombination lines dominate the residual the residual is not minimized further by the introduction of higher order terms. Thus we converge to a smooth completely monotonic approximation to the foreground that leaves the recombination line spectral structure as a residual to the fit, which no longer changes significantly with increasing the order of the completely monotonic polynomial.
$f_{\rm fg}(\nu)$ may be modeled as a completely monotonic (CM) polynomial in brightness temperature versus frequency space. More useful is a modeling in temperature versus log-frequency space, where the number of terms in the Taylor approximation would be reduced. The Taylor expansion may be conveniently computed using the lowest frequency in the range of interest as the reference value.
A CM polynomial of arbitrary order $n$, whose constant term $a_0$ is left unconstrained so as to allow arbitrary vertical translations, may be written in the form
\begin{equation}\label{eq:CM_function}
f(x) = a_0 + \sum_{i=1}^{n} (-1)^i (x - x_0)^i \{ \sum_{j=0}^{n-i} a_{i+j} C^{i+j}_{j} (x_m - x_0)^{j} \},
\end{equation}
where $C^n_k$ denotes the binomial coefficient $n!/\{k! (n-k)!\}$. An example illustrating the algorithm to construct such a CM function and thereby arriving at the functional form presented in equation \ref{eq:CM_function} is described in Appendix \ref{sec:appendix_a}.
A disadvantage of adopting a CM functional form is that the data is fitted in brightness temperature versus log-frequency space, where the foreground is expected to be CM. A functional form that is smooth in log-temperature versus log-frequency space is preferred, since a lower order polynomial would be sufficient to describe the foreground. Lower numbers of parameters makes for more robust fitting with less likelihood that optimization algorithms get trapped in local minima.
\subsection{Modeling the foreground as a {\it Maximally Smooth} function: a smooth polynomial that has no zero-crossings in derivatives}
We now consider polynomial functional forms to describe the smooth foreground in log-temperature versus log-frequency space. The first approximation in this parametrization is a straight line representing the mean spectral index of the sky region. Curvature in the mean spectrum may be represented, to lowest order, by adopting a parabolic form for the spectrum in this log space. This form has a constant second order derivative.
If we improve upon the modeling of the spectral curvature by adding a cubic term to the polynomial, we may constrain the model to be smooth and without inflections by requiring that the second derivative has no zero crossings in the domain. If we considering a polynomial model of the form
\begin{equation}
f(x) = p_0 + p_1 (x -x_0) + p_2 (x - x_0)^2 + p_3 (x - x_0)^3,
\end{equation}
we would require that second derivative
\begin{equation}
\frac{\text{d}^2f(x)}{\text{d}x^2} = (2!/0!) p_2 + (3!/1!)p_3 (x - x_0)
\end{equation}
has no zero crossings in the domain.
If we wish to improve, further, the modeling of the spectral curvature we may represent the foreground spectrum by a fourth order polynomial of the form:
\begin{equation}
f(x) = p_0 + p_1 (x -x_0) + p_2 (x - x_0)^2 + p_3 (x - x_0)^3 + p_4 (x - x_0)^4.
\end{equation}
To constrain this polynomial to be smooth and without inflection points embedded, we now require that the second derivative
\begin{equation}
\frac{\text{d}^2f(x)}{\text{d}x^2} = (2!/0!) p_2 + (3!/1!)p_3 (x - x_0) + (4!/2!)p_4(x-x_0)^2
\end{equation}
has no zero crossings. Additionally, we also require that the third derivative
\begin{equation}
\frac{\text{d}^3f(x)}{\text{d}x^3} = (3!/0!) p_3 + (4!/1!)p_4 (x - x_0)
\end{equation}
has no zero crossings.
In general, we may model the foreground by an $n^{\rm th}$-order polynomial and require that all derivatives of order 2 and higher have no zero crossings within the domain of interest. This is implemented by computing the functions
\begin{multline}
\frac{\text{d}^mf(x)}{\text{d}x^m} = \{m!/0!\} p_m + \{(m+1)!/1!\} p_{m+1} (x - x_0) + \{(m+2)!/2!\} p_{m+2} (x - x_0)^2 + \\
\{(m+3)!/3!\} p_{m+3} (x - x_0)^3 +\ .....\ + \{ n! / (n-m)! \} p_n (x - x_0)^{n-m}
\end{multline}
or
\begin{equation}
\frac{\text{d}^mf(x)}{\text{d}x^m} = \sum_{i=0}^{n-m} \{(m+i)! / i! \} p_{m+i} (x - x_0)^i
\end{equation}
for all $m$ in the range 2, 3, 4, ...., $(n-1)$ and constraining the polynomial coefficients $p_j$ so that there are no zero crossings within the domain for any of these functions. We call the polynomial functions that satisfies these constraints {\it Maximally Smooth} functions. They will not have ripples embedded.
We model the foreground in log-temperature versus log-frequency space using a Taylor series expansion about the lowest frequency log$_{10}(\nu_0)$ in our band. The polynomial is written in terms of powers of $\log_{10}(\nu/\nu_0)$. If $y(x)$ is the polynomial describing the foreground in log-space, then $10^{y(x)}$ is added to a term that describes the CMB spectrum and another term that models the recombination line spectrum to get a model for the total spectrum.
We model the recombination line spectrum component in the mock data using a single scaling parameter: the recombination line component is considered to be this scale factor times a template of the spectral ripple that is nominally expected to be present in the observation. First, the scale factor is in itself a quantitative measure of the recombination template present in the total spectrum. A scale factor close to unity indicates that the predicted template is present in the spectrum where as a small value indicates the absence of such a spectral signature. The distribution allowed for this scaling parameter by the goodness of fit yields the confidence in the detection of the predicted recombination template. Secondly, by modeling the recombination lines using a scaled template allows the residuals of the optimization to approach measurement noise if the theoretical framework that led to the template is correct.
The analytic fitting function in its final form, which includes a Maximally Smooth form for the foreground modeling, is given by
\begin{equation}\label{eq:fit_func}
T(\nu) = \left( {\frac{h\nu}{k}} \right) / \left( {e^{\frac{h\nu}{kp_{0}}}-1} \right) + p_1T_{\rm rec}(\nu)
+ 10^{ \sum_{i=0}^{n} \left[\log_{10}(\nu/\nu_0)\right]^i \,p_{i+2}} ,
\end{equation}
where $p_0$ corresponds to the CMB temperature, $p_1$ is the scale factor that multiplies the recombination line template $T_{\rm rec}(\nu)$ and $p_2$ through $p_{2+n}$ are the coefficients of the terms in the $n^{\rm th}$-order Maximally Smooth polynomial that models the foreground. Fig.~\ref{fig:template_ideal} shows the predicted recombination line signature that is expected to be present in the synthetic spectrum as an additive component; as stated earlier, this template has been derived from the predicted recombination line spectrum by subtracting a baseline of low order and is what would be expected as a residual if a smooth baseline were to be subtracted from an observation as a method of removing the foreground. Segments of this template are what form the recombination line template $T_{\rm rec}(\nu)$.
We use the downhill simplex \citep{Nelder1965} optimization algorithm to iteratively fit this model to the synthetic spectrum adopting a successive approximation strategy. We start by fitting to a function that has four coefficients, one for the CMB temperature, one for the amplitude of the recombination line spectrum and two that describe a first order model for the foreground. We successively include more terms one by one giving an initial guess of zero for each new coefficient.
Adopting the method discussed in Section~3, we have generated synthetic spectra over three octave bands --- 2.0--4.0, 2.5--5.0 and 3.0--6.0~GHz --- spaced uniformly over 24 hr in LST to test the robustness of this modeling algorithm. In this test, the noise in the spectrum was kept small so that the residual spectrum might reveal the recombination line structure obviously, if recovered successfully. As expected, the final fit parameters vary for different realizations of the sky spectrum corresponding to different observation times. Nevertheless, the algorithm did indeed converge for all LSTs and yielded a best fit scaling factor close to unity at all times. An 8$^{\rm th}$ order Maximally Smooth polynomial was used to model the foreground. The residuals were consistent with the measurement noise that had been added to the synthetic spectra and the foregrounds were indeed successfully modeled as 8$^{\rm th}$-order smooth polynomials, without also fitting to the embedded recombination line spectrum.
\begin{figure}[ht]
\centering
\includegraphics[scale = 0.5]{foreground_plus_lines.png}
\caption{ Foreground model fit to a synthetic spectrum that represents a mock observation is shown in blue. The residual, scaled by a fraction of the inverse of the chi-square of fit and and added back to the foreground model is shown in red. }
\label{fig:foreground_plus_lines}
\end{figure}
In Fig.~\ref{fig:foreground_plus_lines} we show a sample fit to a mock observation in the 3--6~GHz band. The model-fit CMB plus Maximally Smooth foreground is shown along with a second trace that is constructed by computing the residuals to this fit and adding the residuals to the model with a large rescaling. The plot shows that the model does indeed not fit to the recombination line ripple, which is present in the synthetic spectrum, and does leave the ripple as a residual. The model fitting algorithm based on the Maximally Smooth polynomials we have constructed is capable of separation of the foreground from embedded ripple, which is nine orders of magnitude smaller. The residual alone is shown in Fig.~\ref{fig:residual_constrained}, which can be compared with Fig.~\ref{fig:unconstrained_residual} where an 8$^{\rm th}$-order polynomial was used without any constraints to fit to the synthetic spectrum.
\begin{figure}[ht]
\centering
\includegraphics[trim= 20mm 10mm 10mm 10mm, clip,scale = 0.4]{constrained_residual_nK.png}
\caption{The residual on subtracting the foreground plus CMB model-fit from the synthetic spectrum.}
\label{fig:residual_constrained}
\end{figure}
\begin{figure}[h!]
\begin{centering}
\begin{minipage}[b]{0.32\linewidth}
\includegraphics[trim= 5mm 5mm 15mm 0mm,clip,width=\linewidth]{residuals_2_to_4_diff_method.png}
\label{fig:2_to_4}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\includegraphics[trim= 5mm 5mm 15mm 0mm,clip,width=\linewidth]{residuals_2pt5_to_5_diff_method.png}
\label{fig:2pt5_to_5}
\end{minipage}
\begin{minipage}[b]{0.32\linewidth}
\includegraphics[trim= 5mm 5mm 15mm 0mm,clip,width=\linewidth]{residuals_3_to_6_diff_meth.png}
\label{fig:3_to_6}
\end{minipage}
\caption{Residuals in the 2.0--4.0, 2.5--5.0 and 3--6~GHz bands following fitting and subtracting Maximally Smooth foreground models to synthetic sky spectra. The mock observations are over a 24 hour period, with spectra spaced 2 hours apart, and represent observations made by an ideal instrument as discussed in the text. Shown as a red dotted line is the recombination line template from the theoretical predictions; the overlaid colored dashed lines represent the recovered residuals from different realisations of synthetic sky spectra.}
\label{fig:residual_lst}
\end{centering}
\end{figure}
In Fig.~\ref{fig:residual_lst} we show the residuals to the fit for a smooth foreground plus CMB to synthetic spectra for mock observations in different frequency ranges and distributed over LST. The algorithm does recover the recombination ripple in all cases, giving confidence in the robustness of the method.
These residuals represent a detection of the cosmological recombination lines. The recovery demonstrates that the smooth functional form proposed here does not also fit to the nanoscale embedded ripple representing the recombination line signature, despite the functional form being allowed to be of arbitrarily high order. The recovery demonstrates that it is indeed possible to model fit a smooth functional form--- e.g., the Maximally Smooth functional---to observations and model the completely monotonic foreground---of unknown functional form---to an accuracy better than 10~nK!
\section{Confidence in Detection}
In the previous section, we presented an algorithm by which we recover cosmological recombination lines of amplitude $\sim$10~nK that are a tiny additive part of a sky spectrum which is nine orders of magnitude brighter. In this section, we first examine the probability distribution for the detected amplitude of the spectral ripple. We then examine the confidence in detection of the expected ripple and the confidence with which we may reject false positives, using a Bayes factor approach. These lead to inference of the observing time needed for detections with different confidence levels.
\subsection{Detection based on the fit amplitude of the recombination spectral ripple}
\label{sec:mcmc}
An approach to determining whether or not a given sky spectrum contains cosmological Hydrogen recombination lines is to jointly fit to the spectrum a model composed of a Planck spectral component for the CMB, a Maximally Smooth polynomial accounting for radiation from point and extended sources in the foreground, and the recombination line template scaled by a factor, which is a variable for the fitting. The functional form for such a model is given by Equation~\ref{eq:fit_func}.
We expect that the fit value of $p_1$, which represents the normalized amplitude of the recombination line ripple, would take on a value of unity if the theory is valid, and a value zero if no lines exist. We choose 0.5 as the threshold between a null detection and a positive detection, with values smaller than 0.5 assumed to indicate a null detection and greater than 0.5 deemed to be indicating a positive detection.
To determine the confidence in detection using such an approach, as well as the probability of a false positive, we synthesize two types of mock observations, one in which the theoretically expected cosmological recombination lines are added and another without. Forty spectra of each type were generated with independent thermal noise. Below we adopt the notation that data and associated terms for spectra with recombination lines present in them would be referred to as data set `$a$' and those without recombination lines as `$b$'. We fit each of the synthetic sky spectra separately with the mathematical model given by Equation~\ref{eq:fit_func} using the successive approximation approach, optimizing all the parameters using the \cite{Nelder1965} algorithm to minimize the Chi-squared difference between the model and mock data, each time increasing the degree of the Maximally Smooth polynomial in Equation~\ref{eq:fit_func} by unity until the root mean square of the residual saturates.
To sample the distribution in scaling factor $p_1$ and thus derive the confidence for detection using a threshold of 0.5 for the scaling factor, we adopt a Markov Chain Monte Carlo (MCMC) analysis. Specifically, we adopt the EMCEE package \citep{Foreman-Mackey2013}, which is a Python-scripting-language implementation of an affine invariant ensemble sampler as proposed by \cite{Goodman2010}. The {\tt emcee} package wins over traditional implementation of MCMC samplers in its high computational efficiency in generating statistically independent samples from the posterior probability distribution function (PDF). Another major advantage of {\tt emcee} is that it requires tuning of only a couple of parameters by hand in an $N$-dimensional space as opposed to $N^2$ values in other traditional methods. While data driven algorithms do exist to arrive at optimum guesses for initial parameters, for conventional methods this comes at the computational cost of lengthy burn-in chains.
Chi-squared minimization of the model $T(\nu)$ leads to optimum values for the parameters $p_0$ representing the CMB temperature, $p_1$ representing the scaling factor, and the polynomial coefficients $p_2$, $p_3$, $p_4$, $p_5$ and $p_6$ (for a $4^{th}$ order Maximally Smooth polynomial). We then invoke the EMCEE sampler with 30 walkers each of which generates a Markov sampling chain for every parameter. The different chains are initialized at random locations close to the optimum location in the $N=7$ dimensional parameter space.
The EMCEE ensemble sampler operates on the log posterior probability distribution function that we define to be:
\begin{equation}
\ln ({\rm Probability}) = \frac{1}{2}\mathlarger{\mathlarger{\sum}}\bigg(\frac{(T(\nu) - y_{a|b})^2}{\sigma^2} + {\rm ln}(2\pi\sigma^2)\bigg),
\label{eq:lnprob}
\end{equation}
where $y_{a|b}$ is the synthetic sky spectrum with or without recombination lines.
We first run a short burn-in of 1000 steps, discard these values and run a second burn-in of 6000 steps. We ensure that length of the second burn-in is at least ten times the correlation length, then discard these as well and run a final useful 500 steps. With 30 walkers exploring 500 steps each we generate a distribution of 15000 samples per parameter per synthetic sky spectrum. Marginalizing over the nuisance parameters, which in our case are all parameters with the exception of the scaling factor, we have for 40 sky spectra a distribution of 600,000 scaling factors in all. On running the EMCEE ensemble sampler on data sets `$a$' and `$b$' as described above we have two distributions of scaling factors with 600,000 samples each.
We generate multiple pairs of data sets `$a$' and `$b$' with different noise integration times; as a guide we have adopted times corresponding to different signal-to-noise ratios as defined by Equation~\ref{eq:snr}. For each integration time, the fraction of samples of the scaling factor from the Markov chain corresponding to data set `$a$' that have values exceeding the threshold of 0.5 gives an estimate of the probability of detection. Similarly the fraction of samples corresponding to data set `$b$' below 0.5 is an estimate of the probability of rejecting a false positive. These are depicted in Fig.~\ref{fig:cl_from_sf}.
\begin{figure}[ht]
\centering
\includegraphics[trim=15mm 0mm 20mm 0mm,clip,scale = 0.45]{MCMC_confidence_and_fp_Chajnantor_1K_array_percentage.png}
\caption{The confidence with which scaling factor (SF) above threshold 0.5 signifies a detection of cosmological recombination lines (shown as a blue solid line with scale on the left of the plot), and percentage likelihood of false positives appearing as a detection for this threshold (shown as a red dotted line with scale on the right side of the plot). The confidence values and percentage likelihoods of false positives are shown versus integration time assuming cryogenic cooled state-of-the-art 1~K receivers and a 128-element array of precision total-power spectrometers.}
\label{fig:cl_from_sf}
\end{figure}
Detection of the recombination ripple with 68\% confidence using uncooled receivers requires observing with about $210 \times 10^{8}$~antenna secs, or equivalently about $240000$ antenna days. An array of 128 singly-polarized total-power spectrometer elements with uncooled receivers would require about $\sim 5$ years for detection with this confidence and based on $SF$ threshold. However, using cryogenically cooled state-of-the art receivers, the observing time for detection with 68\% confidence is $\sim10 \times 10^{8}$~ antenna secs or, equivalently, 92 observing days with an array of 128 spectral radiometers deploying such cooled receivers. A 90\% confidence detection requires observing with 430 days with such an array, and 95\% confidence detection requires increasing the observing time for such an array to about 625 days.
\subsection{Detection based on Bayes factor (BF)}
\label{sec:bayes_factor}
Consider some data D and a choice between model M$_1$ characterised by a set of parameters $\theta_1$ and M$_2$ characterised by a set of parameters $\theta_2$, to describe the data. The Bayes factor is given by the ratio of the probability of the data D given the model M$_1$ to the probability of D given M$_2$:
\begin{equation}\label{eq:BF_def}
BF = \frac{P(D|M_1)}{P(D|M_2)} = \frac{\int P(\theta_1|M_1) P(D|\theta_1,M_1) d\theta_1}{\int P(\theta_2|M_2) P(D|\theta_2,M_2) d\theta_2}.
\end{equation}
An alternative expression of the Bayes factor is in terms of posterior and prior odds. Rewriting the probabilities associated with the two models as odds, we obtain:
\begin{equation}\label{eq:BF_def2}
\begin{split}
{\rm Posterior\ odds} = BF \times {\rm Prior\ odds}, \\
{\rm or\ that:~~}\frac{P(M_1|D)}{P(M_2|D)} = BF \times \frac{P(M_1)}{P(M_2)}, \\
{\rm which\ leads\ to:~~}BF = \frac{P(M_1|D)}{P(M_2|D)}\times \frac{P(M_2)}{P(M_1)}.
\end{split}
\end{equation}
When the prior odds are 1:1 the Bayes factor reduces to the likelihood ratio.
We once again adopt the notation used in Section~\ref{sec:mcmc} so that symbols representing data and associated terms for mock observations with the recombination lines present in them would be given subscript `$_a$' and those without recombination lines `$_b$'. Terms representing the null hypothesis would have subscript `$_0$' and alternative hypothesis `$_2$'. In our problem of estimating the confidence in the detection of recombination line ripples in observed spectra, we consider two hypotheses. The first hypothesis is that the spectrum does not contain recombination lines of any detectable amplitude; {\it i.e.}, within errors the mock spectrum can be adequately modeled by the total of contributions from the cosmic microwave background and a Maximally Smooth function representing the cumulative spectra from diffuse and point sources in the foreground. We refer to this as the null hypothesis $H_0$. The second hypothesis, which we refer to as the alternative hypothesis $H_2$, is that the observed sky spectrum includes the cosmological Hydrogen recombination lines as an additive component along with contributions from the CMB and foregrounds. Hypothesis $H_2$ is that the spectrum in question does contain the recombination lines exactly as given by the theoretical expectations. The Bayes factor for a comparison between these alternate hypotheses is given by:
\begin{equation}\label{eq:BF}
BF = \frac{P(D|H_2)}{P(D|H_0)}\times\frac{P(H_2)}{P(H_0)},
\end{equation}
where $P(D|H_2)$, $P(D|H_0)$ are the respective likelihoods of the alternative and null hypotheses and $P(H_2)$, $P(H_0)$ are the prior probabilities of the alternative and null hypotheses. Since in this comparison we have no {\it a priori} preference or bias towards the presence or absence of the recombination ripple in data, we may assume equal prior probabilities and hence the Bayes factor reduces to the likelihood ratio given by:
\begin{equation}\label{eq:BF_2}
BF = \frac{P(D|H_2)}{P(D|H_0)}.
\end{equation}
We now proceed to compute the likelihood functions for the above two hypotheses. We generate data set `a', a set of 100 mock observations with independent thermal noise and with each containing the recombination lines as an additive component. We then fit each of these spectra independently with two different models. The first model is the null hypothesis $H_0$ with a blackbody function to model the CMB component and a Maximally Smooth polynomial to model the foreground component of the sky spectrum. $H_0$ expects that these two components completely describe the spectrum and that the residual on subtracting this model-fit from the data must be Gaussian random with zero mean and standard deviation given by the measurement noise. We compute the likelihood of obtaining the data given $H_0$ by:
\begin{equation}\label{eq:likeli_1}
P(D_a|H_0) = \prod_{i=1}^{N}\frac{e^{-\frac{y_{res0}[i]^2}{2\sigma_0^2}}}{\sqrt{2\pi\sigma_0^2}},
\end{equation}
where $N$ is the number of independent points across the spectrum and $y_{res0}[i]$ is the residual spectrum following subtraction of the model corresponding to the null hypothesis.
The variance of the measurement noise, $\sigma_0^2$, is estimated from the data itself. This variance is assumed to be half the Allan variance (AV) of the residual, which is given by:
\begin{equation}\label{eq:av}
AV = \sum_{i=1}^{N-1}(y_{res}[i+1]-y_{res}[i])^2,
\end{equation}
where $y_{res}$ is the residual on subtracting the model from the data. Since in our cases the signal-to-noise is small in the channel data, this approach makes the estimate for measurement error robust and independent of errors in the model fit and any low-order residuals that are not represented in the model.
\begin{figure}[ht]
\centering
\includegraphics[trim = 0mm 0mm 10mm 5mm,clip,scale = 0.5]{logBF_vs_inttime_Chajnantor_1K_monopole_array.png}
\caption{Bayes factors computed for data sets `a' and `b' that are mock observations generated with and without the recombination line component respectively. log$_{10}(BF)$ is plotted versus observing time assuming cryogenic cooled state-of-the-art 1~K receiver noise temperatures and a 128-element array of precision total-power spectrometers. Filled circles represent the median values of the $BF$'s computed separately for the two data sets. The horizontally-striped region shaded in red represents the range in Bayes factors obtained for the 100 mock observations in data set `a' and the vertically-striped region shaded in green represents the range of Bayes factors obtained for data set `b'.}
\label{fig:bf_time}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[trim = 5mm 5mm 10mm 5mm,clip,scale = 0.55]{BF_confidence_Chajnantor_1K_monopole_array.png}
\caption{The confidence with which Bayes factor above unity might signify a detection, and simultaneously reject the detection to be a false positive, versus integration time. We assume cryogenic-cooled state-of-the-art 1~K receiver noise temperatures and a 128-element array of precision total-power spectrometers.}
\label{fig:mean_cl}
\end{figure}
The second model corresponds to the alternative hypothesis $H_2$. This model contains the blackbody CMB term, the Maximally Smooth polynomial representing the foreground and, in addition, a template of the recombination lines that are expected to be present in the spectra for this hypothesis. As before, we proceed by fitting the data with the model, subtracting the best fit model from the data and computing the likelihood from the residual. We again make the reasonable assumption that if the data is completely represented by the model then the residual should be Gaussian random noise corresponding to measurement noise. The likelihood function P($D_a|H_2$) is given by:
\begin{equation}\label{eq:likeli_2}
P(D_a|H_2) = \prod_{i=1}^{N}\frac{e^{-\frac{y_{res2}[i]^2}{2\sigma_2^2}}}{\sqrt{2\pi\sigma_2^2}},
\end{equation}
where $y_{res2}[i]$ is the residual on subtracting the alternative hypothesis model from the data, and variance $\sigma_2^2$ is half the Allan variance of this residual. Since the Allan variance gives the measurement noise independent of the model used to fit to the data, we expect $\sigma_0^2$ and $\sigma_2^2$ to be the same.
Having computed the likelihoods for both the models we arrive at the Bayes factor by computing the likelihood ratio:
\begin{equation}\label{eq:BF_w_line}
BF_a = \frac{P(D_a|H_2)}{P(D_a|H_0)}.
\end{equation}
We repeat this exercise with a separate data set `b', which is a set of 100 spectra corresponding to mock observations with the same measurement noise variance as in data set `a', but with no recombination lines in the synthetic spectra. We once again fit these spectra with the two models as before. We compute the likelihood functions for the null and alternative hypotheses given by $P(D_b|H_0$) and $P(D_b|H_2$) respectively, under the reasonable assumption that when a model that completely represents the data, except for random noise, is subtracted from the data the residual must be the measurement noise. Thus the Bayes factor for the 100 mock observations that are generated assuming that recombination line component is absent is given by:
\begin{equation}\label{eq:BF_n_line}
BF_b = \frac{P(D_b|H_2)}{P(D_b|H_0)}.
\end{equation}
Fig.~\ref{fig:bf_time} shows a plot of log$_{10}$($BF_a$) and log$_{10}$($BF_b$) versus integration time. With increasing integration time the measurement noise reduces and the two Bayes factors diverge, as expected.
For any data set, the Bayes factor indicates the relative preference of the data for the two hypotheses. The median $BF_a$ rises above unity with increasing integration time, increasingly preferring the hypothesis $H_2$ over $H_0$. In contrast the median $BF_b$ drops below unity with increasing integration time, increasingly preferring the hypothesis that the recombination lines are absent in data set `b'. We may set a threshold at unity and examine the confidence with which this threshold might discriminate between observations that contain a recombination ripple to those in which this additive component is absent. At any integration time, the distribution of Bayes factors indicates the probability that $BF_a$ is above unity, which gives the confidence in a detection. For the same integration time, the distribution of $BF_b$ yields the probability of obtaining a value above unity, which gives the likelihood of a false positive. Using the multiple data sets `a' and `b' we have computed the run of the confidence in detection (and simultaneous rejection of false positives) versus integration time: this is depicted in Fig.~\ref{fig:mean_cl}.
If we adopt the Bayes factor as a detection statistic, detection of the recombination ripple with 68\% confidence using uncooled receivers requires observing with about $52 \times 10^{8}$~antenna secs, or equivalently about $60,000$ antenna days. An array of 128 singly-polarized total-power spectrometer elements with uncooled receivers would require about $1.3$ years for detection with this confidence. Using cryogenically cooled state-of-the art receivers, the observing time for such detection with 68\% confidence is $2.56 \times 10^{8}$~ antenna secs or, equivalently, $23$ observing days with an array of 128 spectral radiometers deploying such cooled receivers. 90\% confidence detection requires observing with 255 days with such an array, and 95\% confidence detection requires about 440 days.
Fig.~\ref{fig:mcmc_vs_bf} shows the confidence in the detection of cosmological recombination lines in synthetic sky spectra as a function of integration time, estimated using both the Bayes factor method as well as using the fit to the amplitude of the recombination ripple (Section~\ref{sec:mcmc}). The Bayes factor approach to detection is simplest in that it only asks whether in an observed spectrum the lines as predicted by theory are present or totally absent. The method that derives an amplitude for the ripple, as estimated by a best fit to a scaling factor for a template of the expected ripple, attempts to ask a more informative question in that it attempts to evaluate the amplitude of the recombination line ripple. Unsurprisingly, the Bayes factor method has greater confidence in its answer for any integration time.
\begin{figure}[ht]
\centering
\includegraphics[trim = 15mm 5mm 10mm 5mm,clip,scale = 0.5]{BF_and_MCMC_confidence_Chajnantor_1K_monopole_array.png}
\caption{A comparison of the confidence in detecting cosmological recombination lines in sky spectra using Bayes Factors and that given by a comparison of the best-fit ripple amplitude to a threshold. The confidence values are shown versus integration time assuming cryogenic cooled state-of-the-art 1~K receiver noise temperatures and a 128-element array of precision total-power spectrometers.}
\label{fig:mcmc_vs_bf}
\end{figure}
\section{Summary}
We demonstrated that it is in principle feasible with current day technology to experimentally detect the cosmological Hydrogen and Helium recombination lines at low frequencies, although the lines are embedded in a foreground that is about nine orders of magnitude brighter with a priori unknown precise spectral shape. The recombination radiation has a smooth component that is difficult to distinguish from the smooth foregrounds; however, the recombination radiation also has a unique ripply component that may be distinguished from the foregrounds and instrumental effect. We estimate the amplitude of the ripply signal, the instrument noise arising from receivers and foregrounds, and estimate the signal-to-noise in detections using ground based spectrometers. Detection may ideally be attempted using an octave band in the 2--6~GHz window; an octave band would have a spectral segment of the recombination ripple with sufficient structural complexity so as to be distinguishable from the relatively smoother foreground. We have developed an algorithm to detect the recombination line ripple by foreground modeling that models the foreground as a {\it Maximally Smooth} polynomial, which we define enforcing no zero crossings in derivatives of order $n\geq 2$. A similar approach may be applicable to modeling and detecting the global 21-cm signal at lower frequencies.
We then evaluate the confidence with which these cosmological recombination lines may be isolated using the aforementioned algorithm and estimate the integration times required for detection with varying degrees of confidence. In its simplest form, a detection is testing an observed spectral segment addressing the question of whether or not the theoretically motivated spectral ripple is present in the sky spectrum or absent. To answer this question with 90\% confidence requires an integration time of 32,640 antenna-days with a total-power spectral radiometer with cryogenically cooled state-of-the art receivers; however, with today's best uncooled receivers the integration time for such a detection increases to as much as 660,000 antenna days. This makes a compelling case for the use of cryogenically cooled receivers. Moreover, since the antenna element for such all-sky global signals does not gain from antenna directivity, very small antenna elements of centimeter dimensions may be deployed, perhaps as a compact cluster of antenna elements housed in a moderate-size dewar. Assuming a 128 element array followed by independent spectral radiometers, a 90\% confidence detection may be achieved in 255 days integration time.
APSERa\footnote{http://www.rri.res.in/apsera/}---Array of Precision Spectrometers for the Epoch of RecombinAtion---is an experimental venture at the Raman Research Institute\footnote{http://www.rri.res.in/}, India, with the science goal of detecting spectral ripples from the epochs of Hydrogen and Helium recombination. On completion, it is to be an array of 128 miniature radio telescopes operating in the octave band of 3--6~GHz with antenna elements and receivers that are custom designed for all-sky spectral measurements with the required sensitivity to experimentally detect these cosmological recombination lines. APSERa will be deployed at a radio quiet site closer to the geographic poles to avoid radio frequency interference in these bands from downlinks of geostationary satellites.
\section{Future Investigations}
Ripple like spectral signatures arising from the epochs of Hydrogen and Helium recombination are predicted to appear as additive distortions to the sky spectrum at a level $\sim$9 orders of magnitude weaker than the total brightness of the radio sky. We have demonstrated that it is in principle possible to detect these signals. Due to their small amplitude, any small departure in the spectral radiometer from the ideal behavior assumed herein could critically affect the detection likelihood. Studies of the effects of potential instrumental non-idealities, and confusion arising from additive astrophysical contaminants, which have not been considered here, will form the next step in our future work. The key investigations are listed below.
\begin{itemize}
\item The spectral radiometer radio telescope used to detect the weak cosmological recombination lines must be capable of achieving the required spurious-free dynamic range so as to recover the embedded cosmological signal in the measurement. This implies that the receiver configuration and calibration techniques must be specifically designed to minimize systematics and that any residual systematics are either well characterized or emendable to accurate modeling so as to enable unambiguous detection of the weak signal. A future work is the design of a suitable receiver configuration, observing strategies, gain stability and calibration methods to achieve the required accuracy in the measurement.
\item In our simulations, we have generated mock sky spectra as observed by an ideal frequency independent antenna beam. The spectral response of such an antenna is {\it maximally smooth} if all sources in the beam have spectra that are complete monotones. However, designing and fabricating an antenna that is frequency independent over an octave bandwidth is non-trivial. Frequency dependence in the beam pattern and its sidelobes may result in a response that is no longer smooth and may confuse the detection of cosmological recombination. A study of this mode coupling of sky structure into spectral structure in a spectral radiometer, which depends on the type of frequency dependence in the beam pattern, would lead to design tolerances on the antenna element for this detection experiment.
\item The all-sky `ripply' signal arising from the epochs of Hydrogen and Helium recombination is inherently unpolarized and this spectral feature may be detected with an antenna that responds to any single polarization mode, either linear or circular. However, the foreground synchrotron emission from extragalactic sources as well as Galactic emission is linearly polarized and Faraday rotation during the line-of-sight propagation results in the received polarization position angle varying with observing frequency. This causes linearly polarized sources to appear with a `ripply' spectral structure in the response of linearly polarized antennas. One possible way of eliminating beam asymmetries and the effect of beam rotation across the sky is to rotate the array itself. A study of this mode coupling of polarized sky emission to spectral structure is a future study that will lead to design tolerances on the polarization properties of the antenna element and on the polarization calibration. The potential of using the unpolarised nature of the cosmological signal to discriminate it from polarised foregrounds is another aspect to be explored.
\item The spectral template of the additive ripple-like feature from cosmological Hydrogen and Helium recombination has a fairly accurate theoretical prediction. The near quasi-sinusoidal nature of this signal acts as a fingerprint providing a means to distinguish it from other weak cosmological signals. Two such cosmological signals are the $\mu$ and $y$ distortions of the CMB \citep{Zeldovich1969, Sunyaev1970mu}. Even within the standard cosmological model, these distortions are created at amplitudes that are typically larger than the cosmological recombination radiation \citep[e.g.,][]{Chluba2011therm, Sunyaev2013, Chluba2013fore, PRISM2013WPII}. A large average Compton $y$ distortions, at the level of $y\simeq 10^{-7}-10^{-6}$, is expected from reionization and structure formation \citep{Sunyaev1972b, Hu1994pert, Cen1999, Oh2003} and unresolved clusters and filaments \citep{Refregier2000, Miniati2000, Zhang2004}. At low frequencies, this type of distortion only leads to a weak frequency dependent signal, $T(\nu)\simeq -2 y (1-x^2/12)$ with $x\approx 0.018 \, [\nu/{\rm GHz}]$, making it is less problematic. For a $\mu$ distortion, the low frequency spectrum shows much richer structure, especially when varying time-dependence of the energy release mechanism \citep[Fig.~14,][]{Chluba2011therm}. Each of these introduce spectral features in the CMB with the $\mu$ distortion peaking at about 1~GHz. In an experiment optimized to detect ripples originating from the epochs of Hydrogen and Helium recombination it would be interesting to study whether $\mu$ and $y$ distortions of the CMB are a possible source of confusion or would themselves be detected as a positive by-product. Similarly, the anomalous microwave emission from spinning dust \citep[see][]{Draine1998,Haimoud2009,Planck2014} will add another spectral dependences to the low-frequency regime, which at this point we have not investigated.
\end{itemize}
\section{Acknowledgements}
The authors would like to thank Franklin H. Briggs at the Research School for Astronomy \& Astrophysics, Australian National University, for his valuable inputs throughout the preparation of this paper. JC is supported by NSF Grant No. 0244990. We acknowledge the ARC Centre of Excellence for All-sky Astrophysics (CAASTRO) and the Australian National University for support under the distinguished visitor program.
|
\section{Introduction} \label{s1}
Variational sparsity regularization based on $\ell^1$-norms became of significant interest in the past ten years with respect to inverse problems applications, e.g.~in imaging (cf., e.g.,~\cite{Scherzetal09}), but also with respect to the progress in
regularization theory for the treatment of ill-posed operator equations in infinite dimensional Hilbert and Banach spaces (cf., e.g.,~\cite{BeBu11,BreLor09,Grasm10,Grasmei11,Lorenz08,RamRes10}). Moreover, with focus on sparsity, the use of $\ell^1$-regularization can be motivated for specific classes of well-posed problems, too (cf., e.g.,~\cite{CanRomTao06}). Based on the powerful tool of variational inequalities (also called variational source conditions), in \cite{BurFleHof13} convergence rates results on $\mathbf{\ell^1}$-regularization for linear ill-posed operator equations have been formulated in infinite dimensional spaces under the condition that the sparsity assumption slightly fails, but the solution is still in $\ell^1$. In the present paper, we improve those results and illustrate the improvement level with respect to the associated convergence rates for the Ces\'aro operator equation in $\ell^2$ and for specific denoising problems. Since the variational inequality approach requires injectivity of
the forward operator (cf.~\cite[Proposition~5.6]{BurFleHof13}), we restrict all considerations in this paper to {\sl injective} linear forward operators which also ensure uniquely determined solutions for the corresponding linear operator equations. The focus on injectivity is also motivated
by the fact that the ill-posedness concept of Section~\ref{s4} suggested by M.~Z.~Nashed (cf.~\cite{Nashed86}) would require substantial technical refinements in a general Banach space setting if non-injective operators were included.
Let $\widetilde A: X \to Y$ be an {\sl injective} and {\sl bounded linear operator} mapping between an {\sl infinite dimensional separable Hilbert space} $X$ and an infinite dimensional {\sl Banach space} $Y$ with norms $\|\cdot\|_X$ and $\|\cdot\|_Y$, respectively. We are searching for the uniquely determined solution $\widetilde{x}^\dagger \in X$ of the linear operator equation
\begin{equation}
\label{eq:tildeopeq}
\widetilde A\, \widetilde x\,=\,y, \qquad \widetilde x \in X,\quad y \in \mathcal{R}(\widetilde A),
\end{equation}
where $\mathcal{R}(\widetilde A)$ denotes the range of $\widetilde A$. Typically, instead of $y$ only noisy data $y^\delta \in Y$ are available. In this context, we consider the deterministic noise model
\begin{equation} \label{eq:noise}
\|y^\delta-y\|_Y \le \delta
\end{equation}
with given noise level $\delta>0$.
If the operator $\widetilde A$ is {\sl normally solvable}, i.e.\ its range is a closed subset in $Y$, then solving
the equation (\ref{eq:tildeopeq}) is a {\sl well-posed problem}. Consequently, for injective $\widetilde A$ the inverse ${\widetilde A}^{-1}: \mathcal{R}(\widetilde A) \subset Y \to X$ exists and is also a bounded linear operator.
If, on the other hand, the range of $\widetilde A$ is not closed, the inverse ${\widetilde A}^{-1}$
is an unbounded linear operator and solving the equation (\ref{eq:tildeopeq}) is an {\sl ill-posed problem}. This means that small perturbations in the right-hand side may lead to arbitrarily large error in the solution. Then regularization methods are required for obtaining stable approximate solutions to equation (\ref{eq:tildeopeq}).
As usual we consider in the sequel the Banach spaces $\ell^q$, $1 \le q < \infty$, and $\ell^\infty$ of infinite sequences of real numbers with finite norms
$$\|x\|_{\ell^q}:=\left( \sum \limits_{k=1}^\infty |x_k|^q\right)^{1/q}\quad\text{and}\quad \|x\|_{\ell^\infty}:=\sup_{k \in \mathbb{N}} |x_k|.$$
The Banach space $c_{\scriptscriptstyle 0}$ consists of
the real sequences $(x_k)_{k \in \mathbb{N}}$ with $\lim \limits_{k \to \infty} |x_k|=0$ and is also equipped with the norm $\|x\|_{c_0}:=\sup_{k \in \mathbb{N}} |x_k|$. Moreover, by $\ell^0$ we denote the
set of sparse sequences, where a sequence is \emph{sparse} if only a finite number of components is not zero.
For such sequences the number of nonzero components is given by
$$\|x\|_{\ell^0}:=\sum_{k=1}^\infty|{\rm sgn}(x_k)|\quad\text{with}\quad
{\rm sgn}(t):=\begin{cases}1,&t>0,\\0,&t=0,\\-1,&t<0.\end{cases}$$
Throughout this paper we fix an orthonormal basis $\{u^{(k)}\}_{k \in \mathbb{N}}$ in the Hilbert space $X$.
By $x=(x_k)_{k\in\mathbb{N}}\in\ell^2$ we denote the infinite sequence of corresponding Fourier coefficients of $\widetilde{x}$, i.e.\
$$\widetilde x=\sum \limits_{k=1}^\infty x_k u^{(k)}.$$
The synthesis operator $L: \ell^1 \to X$ defined as $Lx:=\widetilde x$ is an injective bounded linear operator.
Note that this operator can be extended to $\ell^2$, but in our setting we define it only on $\ell^1$.
The focus of our studies is on \emph{almost sparse} solutions $\widetilde{x}^\dagger$ to equation (\ref{eq:tildeopeq}). This means that only a finite number of coefficients $x_k^\dagger$ from the infinite sequence $x^\dagger=(x^\dagger_k)_{k \in \mathbb{N}}$ is relevant.
Here we do not require strict sparsity, $x^\dagger \in \ell^0$, but we allow an infinite number of nonzero coefficients if they decay fast enough.
Precisely, we assume $x^\dagger \in \ell^1$ throughout this paper.
Introducing the operator $A:=\widetilde A \circ L: \ell^1 \to Y$ our goal is to recover the solution $x^\dagger\in\ell^1$ of
\begin{equation}
\label{eq:opeq}
Ax\,=\,y, \qquad x \in \ell^1,\quad y \in \mathcal{R}(A),
\end{equation}
from noisy data $y^\delta \in Y$ satisfying (\ref{eq:noise}).
The following proposition shows that solving this equation is always an ill-posed problem, even if
the original equation (\ref{eq:tildeopeq}) is well-posed.
For the proof and for further reference we note that the synthesis operator $L$ is a composition $L=U \circ \mathcal{E}_2$ of the embedding operator
$\mathcal{E}_2:\ell^1\to\ell^2$ and the Riesz isomorphism $U:\ell^2\to X$. Thus, the operator $A$ can be written as a composition
\begin{equation} \label{eq:three}
A=\widetilde A \circ U \circ \mathcal{E}_2
\end{equation}
of three injective bounded linear operators.
\begin{proposition} \label{pro:alwaysill}
The range of $A$ is not closed.
\end{proposition}
\begin{proof}
Assume that $\mathcal{R}(A)$ is closed. The full preimage of $\mathcal{R}(A)$ with respect to $\widetilde A$ is $\mathcal{R}(L)$.
Thus, $L$ has closed range, too. Looking at the composition (\ref{eq:three}), the full preimage of $\mathcal{R}(L)$ with
respect to $U$ is $\mathcal{R}(\mathcal{E}_2)=\ell^1$. Consequently, $\ell^1$ would be a closed subspace of $\ell^2$. Since $\ell^1$ is
dense in $\ell^2$ this yields the contradiction $\ell^1=\ell^2$.
\end{proof}
The proposition shows that equation (\ref{eq:opeq}) requires regularization in order to obtain stable approximate solutions.
For this purpose we use a variant of variational regularization, called $\ell^1$-regularization, where regularized solutions, denoted by ${x_\alpha^\delta}$, are minimizers of the extremal problem
\begin{equation}
\label{eq:TikBanach}
\frac{1}{p}\|Ax-y^\delta\|_Y^p + \alpha \,\|x\|_{\ell^1}
\to \min, \quad \mbox{subject to} \quad x \in \ell^1.
\end{equation}
Here, $1 < p < \infty$ is some exponent and $\alpha>0$ is a regularization parameter. This regularization parameter is chosen in an appropriate manner, a priori as $\alpha=\alpha(\delta)$ depending on the noise level $\delta$, or a posteriori as $\alpha=\alpha(\delta,y^\delta)$ depending also on the present regularized solution $y^\delta$ (for details see Sections~\ref{s2} and \ref{s3} below).
We are interested in error estimates
\begin{equation} \label{eq:genrates-1}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1} \le C_{x^\dagger}\,\varphi(\delta)\qquad \mbox{for all} \qquad 0 < \delta \le \overline \delta,
\end{equation}
where the positive constant $C_{x^\dagger}$ may depend on the solution $x^\dagger$ but not on the noise level $\delta>0$. The estimates (\ref{eq:genrates-1})
can be interpreted as convergence rates
\begin{equation} \label{eq:genrates1}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O(\varphi(\delta)) \qquad \mbox{as} \qquad \delta \to 0
\end{equation}
with rate functions $\varphi$ which are concave index functions.
Following \cite{MatPer03} and \cite{HofMat07} we call $\varphi: (0,\infty) \to (0,\infty)$ an index function if it is a continuous and strictly
increasing function with $\lim \limits_{t \to +0} \varphi(t)=0$.
The article is organized as follows: in the next section we briefly summarize results on existence, stability and convergence
of $\ell^1$-regularized solutions. Section~\ref{s3} contains the main theorem of this paper which improves the result of Theorem~5.2 from \cite{BurFleHof13} and can lead to better convergence rates.
Moreover, Section~\ref{s4} provides some insight into the interplay between Nashed's ill-posedness concept and mapping properties of the forward operator like compactness and strict singularity.
In the final Section~\ref{sec:denoise} we apply our findings to a problem of denoising type.
\section{Existence, stability and convergence of $\ell^1$-regularized solutions} \label{s2}
From the general theory of Tikhonov regularization (cf.,~e.g.,~\cite[Section~3]{HKPS07}, \cite[Sections~4.1.1 and 4.1.2]{Schusterbuch12} and \cite[Section~3.1]{ItoJin14}) one can
infer the existence and stability of
$\ell^1$-regularized solutions ${x_\alpha^\delta}$ as well as its convergence for $\delta \to 0$ to the uniquely determined solution $x^\dagger$ of equation (\ref{eq:opeq}) for appropriate choices of the regularization parameter $\alpha>0$.
For this purpose we summarize the results of Proposition~2.8 and Remark~2.9 from \cite{BurFleHof13} in the following proposition taking into account that the separable Banach space $c_0$ is a predual space of $\ell^1$,
that the operator $A: \ell^1 \to Y$ is sequentially weak*-to-weak continuous, and that $\ell^1$ satisfies the weak* Kadec-Klee property (for a proof see,~e.g.,~\cite[Lemma~2.2]{BoHo13}).
The weak*-to-weak continuity of $A$ can be shown along the lines of the proof of Lemma~2.7 in \cite{BurFleHof13} when $\mathcal{R}(A^*) \subset c_0$ is valid for the adjoint operator $A^*: Y^* \to \ell^\infty$ to $A$.
In our setting, $\mathcal{R}(A^*) \subset c_0$ is a consequence of the weak convergence $\widetilde A \,u^{(k)} \rightharpoonup 0$ in $Y$ for the prescribed orthonormal system $\{u^{(k)}\}_{k \in \mathbb{N}}$ in the Hilbert space $X$, which implies the weak convergence of formula
(\ref{eq:weakA}) in Remark~\ref{rem:distinguish} below such that the corresponding part of the proof of \cite[Proposition~2.4]{BurFleHof13} applies even if (\ref{eq:strongA}) is violated.
\begin{proposition} \label{pro:exist}
For all $1<p<\infty$, $\alpha>0$ and $y^\delta \in Y$ there exist uniquely determined minimizers ${x_\alpha^\delta} \in \ell^1$ of the extremal problem (\ref{eq:TikBanach}). These $\ell^1$-regularized solutions are always sparse, i.e.~they satisfy
$${x_\alpha^\delta} \in \ell^0.$$
Furthermore, they are always stable with respect to the data, i.e., small perturbations in $y^\delta$ in the norm topology of $Y$ lead only to small changes in ${x_\alpha^\delta}$ with respect to the $\ell^1$-norm.
\par
If $\delta_n \to 0$ and if the regularization parameters $\alpha_n=\alpha(\delta_n,y^{\delta_n})$ are chosen such that $$\alpha_n \to 0 \qquad \mbox{and} \qquad \frac{\delta_n^p}{\alpha_n}\to 0 \qquad \mbox{as} \qquad n \to \infty,$$
then
\begin{equation} \label{eq:normconvergence}
\lim \limits_{n \to \infty} \|x_{\alpha_n}^{\delta_n}-x^\dagger\|_{\ell^1}\,=\,0\,.
\end{equation}
\end{proposition}
\section{Improved convergence rates} \label{s3}
The norm convergence (\ref{eq:normconvergence}) can be arbitrarily slow.
In order to obtain convergence rates, also for the $\ell^1$-regularization with regularized solutions ${x_\alpha^\delta}$ defined as minimizers to problem (\ref{eq:TikBanach}),
a link condition between the smoothness of the solution $x^\dagger$ to (\ref{eq:opeq}) and the forward operator $A$ is required. From the studies and results of the recent paper \cite{BurFleHof13} we immediately derive the following theorem, where this link condition is a range condition imposed on all unit sequences $e^{(k)}=(0,\ldots,0,1,0,\ldots)$, $k \in \mathbb{N}$, with respect to the adjoint operator $A^*$.
We mention here that $\{e^{(k)}\}_{k \in \mathbb{N}}$ represents a Schauder basis in the Banach spaces $\ell^q$ for all $1 \le q <\infty$.
\begin{theorem} \label{thm:BFH}
Let the operator $A\colon \ell^1 \to Y$ from equation (\ref{eq:opeq})
be such that there exist elements $f^{(k)} \in Y^*$, $k \in \mathbb{N}$, satisfying the range conditions
\begin{equation} \label{eq:allrange}
e^{(k)}\,=A^*f^{(k)}.
\end{equation}
Then a variational inequality
\begin{equation} \label{eq:vi}
\|x-x^\dagger\|_{\ell^1}\le \|x\|_{\ell^1}-\|x^\dagger\|_{\ell^1}+ \varphi_1(\|Ax-Ax^\dagger\|_Y) \qquad \mbox{for all} \quad x \in \ell^1
\end{equation}
is valid for the concave index function
\begin{equation} \label{eq:phiBFH}
\varphi_1(t)=2\inf_{n\in\mathbb{N}}\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert
+t\sum_{k=1}^n\|f^{(k)}\|_{Y^*}\right).
\end{equation}
This yields the convergence rate
\begin{equation} \label{eq:genrates}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O(\varphi_1(\delta)) \qquad \mbox{as} \qquad \delta \to 0
\end{equation}
for $\ell^1$-regularized solutions ${x_\alpha^\delta}$ and for the uniquely determined solution $x^\dagger \in \ell^1$ of equation (\ref{eq:opeq}) provided that the regularization parameter
$\alpha=\alpha(\delta,y^\delta)$ is chosen
appropriately, e.g.\ according to the discrepancy principle
\begin{equation} \label{eq:discr}
\tau_1 \delta \le
\|Ax^\delta_{\alpha(\delta,y^\delta)}-y^\delta\|_Y \le \tau_2 \delta
\end{equation}
for prescribed values $1 < \tau_1 \le \tau_2<\infty$.
\end{theorem}
\begin{remark} \label{rem:curious} {\rm
For more details concerning the consequences of variational inequalities and the role of the choice of the regularization parameter for obtaining convergence rates in regularization we refer, for example, to
\cite{HKPS07} and \cite{AnzHofMat14,BoHo10,BoHo13,Flemmingbuch12,Grasm10,HofMat12}.
Making use of Gelfand triples it was shown in \cite{AnzHofRam13} that, for a wide range of applied inverse problems, the forward operators $A$ are such that link conditions of the form (\ref{eq:allrange}) apply for all $e^{(k)},\;k \in \mathbb{N}$. On the other hand, the paper \cite{FleHeg14} gives counterexamples where (\ref{eq:allrange}) fails for specific operators $A$, but alternative link conditions presented there can compensate this deficit.
}\end{remark}
We improve the convergence rate obtained in the theorem above as follows:
\begin{theorem}\label{thm:improved}
Theorem \ref{thm:BFH} remains true if $\varphi_1$ is replaced by $\varphi_2$ with
\begin{equation} \label{eq:phiimproved}
\varphi_2(t)=2\inf_{n\in\mathbb{N}}\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert
+t\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y^*}\right).
\end{equation}
\end{theorem}
\begin{proof}
From \cite[Lemma~5.1]{BurFleHof13} we know that
\begin{equation}
\label{eq:Pythagoras}
\|x-x^\dagger\|_{\ell^1}-\|x\|_{\ell^1}+\|x^\dagger\|_{\ell^1} \le 2\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert+\sum \limits_{k=1}^n |x_k-x^\dagger_k| \right).
\end{equation}
Observing
\begin{align*}
|x_k-x^\dagger_k|
&=\bigl({\rm sgn}(x_k-x^\dagger_k)\bigr)(x_k-x^\dagger_k)\\
&=\bigl({\rm sgn}(x_k-x^\dagger_k)\bigr)\langle e^{(k)},x-x^\dagger\rangle_{\ell^\infty \times \ell^1}\\
&=\bigl({\rm sgn}(x_k-x^\dagger_k)\bigr)\langle f^{(k)},Ax-Ax^\dagger\rangle_{Y^\ast \times Y}
\end{align*}
the second sum on the right-hand side can be estimated above by
\begin{equation}
\label{eq:restricted-injectivity}
\begin{aligned}
\sum_{k=1}^n |x_k-x^\dagger_k|
&=\sum_{k=1}^n \bigl({\rm sgn}(x_k-x^\dagger_k)\bigr)\langle f^{(k)},Ax-Ax^\dagger\rangle_{Y^\ast \times Y}\\
&=\left\langle\sum_{k=1}^n \bigl({\rm sgn}(x_k-x^\dagger_k)\bigr) f^{(k)},Ax-Ax^\dagger\right\rangle_{Y^\ast \times Y}\\
&\leq\left\Vert\sum_{k=1}^n \bigl({\rm sgn}(x_k-x^\dagger_k)\bigr) f^{(k)}\right\Vert_{Y^\ast}\|Ax-Ax^\dagger\|_Y\\
&\leq\|Ax-Ax^\dagger\|_Y\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y^*}
\end{aligned}
\end{equation}
and taking the infimum over all $n\in\mathbb{N}$ in the resulting inequality yields the variational inequality (\ref{eq:vi}).
\end{proof}
To understand the difference between $\varphi_1$ and $\varphi_2$ it may be helpful to note that $\varphi_1=\varphi_2$ if all $f^{(k)}$ are pairwise collinear. On the other hand, in the particular case that $Y^*=Y=\ell^2$
and $f^{(k)} =e^{(k)}$, we have
$\sum_{k=1}^n\|f^{(k)}\|_{Y^*}=n$ while
$\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y^*}=\sqrt n$.
In Example~\ref{ex:cesaro} and in Section~\ref{sec:denoise} we will show that the improved index function $\varphi_2$ yields better convergence rates for
some equations than the original function $\varphi_1$. The convergence rate result obtained in
\cite{FleHeg14} can be improved in a similar way.
\begin{example}[H\"older rates]\label{ex:hoelderrates} {\rm
If
\begin{equation}
\label{eq:Holeder-rates}
\sum_{k=n+1}^\infty\vert x^\dagger_k\vert\leq K_1\,n^{-\mu}\quad\text{and}\quad
\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y^*}\leq K_2\,n^\nu
\end{equation}
for $K_1,K_2\geq 0$ and $\mu,\nu>0$, then the convergence rate obtained in Theorem~\ref{thm:improved} is
\begin{equation} \label{firstrate}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O\bigl(\delta^{\frac{\mu}{\mu+\nu}}\bigr) \qquad \mbox{as} \qquad \delta \to 0
\end{equation}
(cf.\ \cite[Example~5.3]{BurFleHof13}).
}
\end{example}
\begin{example}[exponential decay of solution components]\label{ex:exponential} {\rm
If
$$\sum_{k=n+1}^\infty\vert x^\dagger_k\vert\leq K_1\,\exp(-n^\gamma)\quad\text{and}\quad
\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y^*}\leq K_2\,n^\nu$$
for $K_1,K_2\geq 0$ and $\gamma,\nu>0$, then the convergence rate obtained in Theorem~\ref{thm:improved} is
$$\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O\left(\delta\bigl(\log(1/\delta)\bigr)^{\frac{\nu}{\gamma}}\right) \qquad \mbox{as} \qquad \delta \to 0$$
(cf.\ \cite[Example~3.5]{BoHo13}).
}\end{example}
\begin{remark} \label{rem:distinguish} {\rm
Note that we always have weak convergence
\begin{equation} \label{eq:weakA}
Ae^{(k)}\rightharpoonup 0 \qquad \mbox{in}\;\;\;Y \quad \mbox{as} \qquad k\to\infty,
\end{equation}
because $\{e^{(k)}\}_{k\in\mathbb{N}}$ converges
weakly in $\ell^2$ and $\widetilde A$ is weak-to-weak continuous since it is norm-to-norm continuous.
In \cite[Remark~2.5]{BurFleHof13} it was shown that the slightly stronger condition
\begin{equation} \label{eq:strongA}
\lim \limits_{k \to \infty} \|Ae^{(k)}\|_Y=0.
\end{equation}
enforces $\|f^{(k)}\|_{Y^\ast}\to\infty$ in Theorems~\ref{thm:BFH} and \ref{thm:improved}.
Obviously, condition (\ref{eq:strongA}) is satisfied if the underlying operator $\widetilde{A}$ is compact since compact
operators map weakly convergent sequences to norm convergent ones (note that this property is equivalent to compactness of $\widetilde{A}$ if
$X$ is a Hilbert or at least a reflexive Banach space, cf.\ \cite[Thm.~3.4.37]{Megginson98}).
On the other hand, one easily finds examples for noncompact operators which do not satisfy (\ref{eq:strongA}).
Choose, e.g., $X=Y=\ell^2$ and let $\widetilde{A}$ be the identity. Then $\|Ae^{(k)}\|_Y=1$ for all $k \in \mathbb{N}$.
The question arises whether (\ref{eq:strongA}) is equivalent to compactness of $\widetilde A$.
The answer is `no' as the following example demonstrates.}
\end{remark}
\begin{example}[Ces\`aro operator]\label{ex:cesaro}{\rm
Let $X=Y=\ell^2$ and define $\widetilde A:\ell^2\to\ell^2$ by
\begin{equation}\label{eq:cesaro}
[\widetilde Ax]_n=\frac{1}{n}\sum_{k=1}^n x_k.
\end{equation}
This operator is injective and noncompact with nonclosed range (see \cite[Solution 177]{Halmos82} or \cite{BroHalShi65}), but we have
$$\|Ae^{(k)}\|_{\ell^2}^2=\sum_{n=k}^\infty\frac{1}{n^2}\to 0\quad\text{if}\quad k\to\infty.$$
Since with $f^{(1)}:=e^{(1)}$ and $f^{(k)}:=ke^{(k)}-(k-1)e^{(k-1)}$ for $k\geq 2$ assumption (\ref{eq:allrange}) of Theorems~\ref{thm:BFH}
and \ref{thm:improved} is satisfied, both convergence rates results apply to the specified operator.
\par
In the index function $\varphi_1$ in Theorem~\ref{thm:BFH} the second sum is
$$\sum_{k=1}^n\|f^{(k)}\|_{\ell^2}=\sum_{k=1}^n\sqrt{(k-1)^2+k^2} \le \frac{n(n+1)}{\sqrt{2}}\,.$$
From
$$\sqrt{(k-1)^2+k^2}\geq\frac{1}{\sqrt{2}}(k-1+k)=\sqrt{2}k-\frac{1}{\sqrt{2}}$$
we even obtain a lower bound of the same order
$$\sum_{k=1}^n\|f^{(k)}\|_{\ell^2}\geq\frac{\sqrt{2}n(n+1)}{2}-\frac{n}{\sqrt{2}}=\frac{n^2}{\sqrt{2}}.$$
On the other hand we now show that the supremum in the definition of $\varphi_2$ can be estimated above by
$$\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{\ell^2}
\leq\frac{2}{\sqrt{3}}\,n^{3/2}.$$
At first we calculate
\begin{align*}
\sum_{k=1}^n a_k f^{(k)}
&=a_1e^{(1)}+\sum_{k=2}^nk\,a_ke^{(k)}-\sum_{k=1}^{n-1}k\,a_{k+1}e^{(k)}\\
&=n\,a_ne^{(n)}+\sum_{k=1}^{n-1}k\,(a_k-a_{k+1})e^{(k)}
\end{align*}
for arbitrary $a_1,\ldots,a_n$. Thus,
$$\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{\ell^2}=\sqrt{n^2a_n^2+\sum_{k=1}^{n-1}k^2\,(a_k-a_{k+1})^2},$$
which attains its maximum over $(a_1,\ldots,a_n)\in\{-1,0,1\}^n$ for $a_k=(-1)^k$. Then
$$\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{\ell^2}
=\sqrt{n^2+4\sum_{k=1}^{n-1}k^2}
=\sqrt{\frac{4}{3}n^3-n^2+\frac{2}{3}n}
\leq \frac{2}{\sqrt{3}}\,n^{3/2}.$$
Here, Examples~\ref{ex:hoelderrates} and \ref{ex:exponential} apply and from these examples we see that the behaviour of the estimated sum in $\varphi_1$
and of the supremum in $\varphi_2$ directly carries over to the convergence rate. Thus, the slower growth of the supremum in
comparison to the faster growth of the sum yields a better rate for $\nu=\frac{3}{2}$ based on Theorem~\ref{thm:improved} than for $\nu=2$ based on Theorem~\ref{thm:BFH}.
}\end{example}
\begin{example}[diagonal operator]\label{ex:diagonal}{\rm
For a comparison we briefly recall Example~2.6 from \cite{BurFleHof13}, where $\widetilde A:X \to Y$ is a
{\sl compact diagonal operator}
between the separable Hilbert spaces $X$ and $Y$ with the singular system
$\{\sigma_k,u^{(k)},v^{(k)}\}_{k \in \mathbb{N}}$
and $\widetilde A u^{(k)}=\sigma_k\,v^{(k)},\;k \in \mathbb{N}$. Then the decay rate of the singular values $\sigma_k \to 0$ for $k \to \infty$ characterizes the degree of ill-posedness
of the equation (\ref{eq:tildeopeq}). For $\sigma_k \sim k^{-\zeta}, \;\zeta>0$, we have $\|Ae^{(k)}\|_Y=\|\widetilde A u^{(k)}\|_Y=\sigma_k \sim k^{-\zeta}$. The link condition (\ref{eq:allrange}) is satisfied
with $f^{(k)}=\frac{1}{\sigma_k}v^{(k)}$ and $\|f^{(k)}\|_Y \sim k^\zeta \to \infty$ as $k \to \infty$. Moreover, we have with some constant $C>0$
$$\sup_{\substack{a_k\in\{-1,0,1\}\\k=1,\ldots,n}}\left\Vert\sum_{k=1}^n a_k f^{(k)}\right\Vert_{Y} \,= \,\sqrt{\sum \limits_{k=1}^n \frac{1}{\sigma_k^2}} \,\le \, \sum \limits_{k=1}^n \frac{1}{\sigma_k}\, \le \, C\,n^{\zeta+1}, $$
hence $\nu=\zeta+1>1$ in Examples~\ref{ex:hoelderrates} and \ref{ex:exponential} based on Theorem~\ref{thm:improved}. Note that the values $\zeta>\frac{1}{2}$ and consequently $\nu>\frac{3}{2}$ correspond with the case
of Hilbert-Schmidt operators $\widetilde A$ and $\nu=\frac{3}{2}$ occurring in Example~\ref{ex:cesaro} is just a borderline case with respect to that fact.
}
\end{example}
\section{Ill-posedness of type I and II} \label{s4}
As suggested by M.~Z.~Nashed in \cite{Nashed86} we distinguish two types of ill-posedness for linear operator equations in a Banach space setting. Again, our focus is on injective operators.
\begin{definition} \label{def:type}
Let $B: Z_1 \to Z_2$ be an injective and bounded linear operator mapping between the infinite dimensional Banach spaces $Z_1$ and $Z_2$. Then the operator equation
\begin{equation} \label{eq:B}
B x=y
\end{equation}
is called {\sl well-posed} if the range $\mathcal{R}(B)$ is a closed subset of $Z_2$, consequently {\sl ill-posed} if the range is not closed, i.e.~ $\mathcal{R}(B) \not= \overline{\mathcal{R}(B)}^{Z_2}$.
\par
In the ill-posed case, the equation (\ref{eq:B}) is called {\sl ill-posed of type I} if the range $\mathcal{R}(B)$ contains an {\sl infinite dimensional closed subspace}, and it is called {\sl ill-posed of type~II} otherwise.
\end{definition}
The two types of ill-posedness differ in the behavior of corresponding regularizers (cf.~\cite{Nashed86}) and with respect to smoothing properties of the linear operators $B$. If $B:=B_1: Z_1 \to Z_2$ is such that equation (\ref{eq:B})
proves to be ill-posed of type~I and $B:=B_2: Z_1 \to Z_2$ is such that equation (\ref{eq:B}) proves to be ill-posed of type~II, then $B_2$ tends to be `more smoothing' than $B_1$. Namely, a range inclusion $\mathcal{R}(B_2) \subset \mathcal{R}(B_1)$ may occur, but $\mathcal{R}(B_1) \subset \mathcal{R}(B_2)$ cannot apply. We refer to \cite{BHTY06} for consequences of range inclusions and in particular to Example~10.2 ibidem for the interplay
of operators which characterize different types of ill-posedness.
The following proposition shows that at least for operators $B$ between Hilbert spaces $Z_1$ and $Z_2$ the type of
ill-posedness is determined by compactness properties of $B$.
\begin{proposition}\label{pro:Nashed}
If the operator equation (\ref{eq:B}) is well-posed or ill-posed of type~I, then the operator $B$ is non-compact.
Consequently, compactness of B implies ill-posedness of type~II.
\par
If $Z_1$ and $Z_2$ in equation (\ref{eq:B}) are Hilbert spaces and the equation is ill-posed,
then the equation is
ill-posed of type II if and only if $B$ is compact.
\end{proposition}
\begin{proof}
If the operator equation (\ref{eq:B}) is well-posed or ill-posed of type~I, there is an infinite dimensional Banach space $\hat Z_2$ included in the subspace $\mathcal{R}(B)$ of $Z_2$ with the same norm as in $Z_2$.
The preimage $\hat Z_1:=B^{-1}(\hat Z_2)$ is a Banach space included in $Z_1$ with the same norm as in $Z_1$. For a compact operator $B$, also its restriction $B|_{\hat Z_1}: \hat Z_1 \to \hat Z_2$ would be compact and moreover surjective. This contradicts the fact that a compact operator has only a closed range if it has a finite dimensional range, being a consequence of the non-compactness of the unit ball in a infinite dimensional
Banach space.
\par
For Hilbert spaces $Z_1$ and $Z_2$ and ill-posed equations (\ref{eq:B}), the equivalence of ill-posedness of type~I and the non-compactness of $B$ is well-known (cf.~\cite[Thm.~4.6]{Nashed86} and \cite[Lemma~5.8 and Theorem~5.9]{Doug98}).
\end{proof}
If compactness of $B$ is replaced by strict singularity, the characterization of ill-posedness types can be
made more precise for injective operators $B$.
\begin{definition}\label{def:strictsingularity}
A bounded linear operator $B$ between Banach spaces $Z_1$ and $Z_2$ is \emph{strictly singular} if its restriction to an infinite dimensional
subspace is never an isomorphism.
\end{definition}
\begin{proposition}\label{pro:Nashedsingular}
Let $B$ be an injective bounded linear operator between Banach spaces $Z_1$ and $Z_2$.
Then equation~(\ref{eq:B}) is ill-posed of type II if and only if $B$ is strictly singular.
\end{proposition}
\begin{proof}
We show that there exists an isomorphic restriction of $B$ to an infinite dimensional subspace of $Z_1$ if
and only if $\mathcal{R}(B)$ contains a closed infinite dimensional subspace.
\par
Obviously, if $Z_3$ is a closed infinite dimensional subspace of $\mathcal{R}(B)$, then the corresponding preimage
is of infinite dimension and the restriction of $B$ to this preimage is an isomorphism.
On the other hand, if there is an isomorphic restriction to an infinite dimensional subspace of $Z_1$, then
its image is also of infinite dimension and closed.
\end{proof}
Now we are going to apply Definition~\ref{def:type} to the equations (\ref{eq:tildeopeq}) and (\ref{eq:opeq}) and to interpret the different cases. First we distinguish in the subsequent remark the possible cases arising in the context of equation (\ref{eq:tildeopeq}).
\begin{remark} \label{rem:opeq} {\rm
\begin{itemize}
\item[]
\item[(a)]
\textbf{Well-posed case}: The equation (\ref{eq:tildeopeq}) can be well-posed, which takes place if $\widetilde A$ is normally solvable. Linear Volterra integral {\sl equations of the second kind} as well as more generally
linear Fredholm integral equations of the second kind with appropriate kernels represent typical
examples of this case, where $X=Y=L^2(\Omega)$ with some bounded and sufficiently regular domain $\Omega$ in $\mathbb{R}^l,\;l=1,2,...\,,$ and the operator $\widetilde A$ is of the form $\widetilde A = I-K$ with the identity operator $I$ and a compact operator $K$ such that zero does not belong to the spectrum of the operator $\widetilde A$.
Normal solvability also occurs if $X=Y$ and $\widetilde A=I$. Then solving
(\ref{eq:tildeopeq}), for given noisy data $y^\delta \in Y$, is the simplest case of a {\sl denoising} problem (cf.~Section~\ref{sec:denoise}).
\item[(b)]
\textbf{Ill-posed case}: The equation (\ref{eq:tildeopeq}) is ill-posed if the operator $\widetilde A$ fails to be normally solvable. This is just the case if the inverse $\widetilde A^{-1}: \mathcal{R}(\widetilde A) \subset Y \to X$ is unbounded.
\begin{itemize}
\item[(b1)\,]
\textbf{Type I}:
Operators $\widetilde A$ for equations (\ref{eq:tildeopeq}) which prove to be ill-posed of type I are non-compact and even not strictly singular. If $Y$ is also a Hilbert space, all ill-posed equations (\ref{eq:tildeopeq}) with non-compact operator $\widetilde A$ are of this type. Multiplication operators in $X=Y=L^2(a,b)$ with $L^\infty(a,b)$-multiplier functions possessing essential zeros and linear convolution
operators in $X=Y=L^2(\mathbb{R}^l)$ with square-integrable kernels represent typical examples for this case. Furthermore, the Hausdorff moment problem with $\widetilde A:L^2(0,1) \to \ell^2$ (cf.~\cite[Example~3.2]{Hof99}) is of this type.
\item[(b2)\,]
\textbf{Type II, non-compact}:
If $Y$ is not a Hilbert space then there exist strictly singular operators $\widetilde A$ with nonclosed range which are not compact, for example the
embedding operators from $\ell^2$ to $\ell^q$ with $2<q<\infty$ (cf.~\cite[Theorem (a)]{GoldThorp63}). This also leads to ill-posed equations (\ref{eq:tildeopeq}) of type~II. Such operators $\widetilde A$ can have a
non-separable range $\mathcal{R}(\widetilde A)$ (cf.~\cite[Remark on p.~335]{GoldThorp63}).
\item[(b3)\,]
\textbf{Type II, compact}:
The equation (\ref{eq:tildeopeq}) is ill-posed of type II if $\widetilde A: X \to Y$ is a {\sl compact} operator. Then $\widetilde A$ is strictly singular and the range $\mathcal{R}(\widetilde A)$ is a separable space. Typical examples with compact operators $\widetilde A$ are linear Fredholm and Volterra integral {\sl equations of the first kind} with square-integrable kernels in $L^2$-spaces $X$ and $Y$ over bounded and sufficiently regular domains in $\mathbb{R}^l$. Moreover, all bounded linear operators $\widetilde A: \ell^2 \to \ell^q$ are compact for $1 \le q <2$ as was mentioned in \cite{Tar72}. In particular, if also $Y$ is a Hilbert space, then vice versa ill-posedness of type~II requires compactness of $\widetilde A$.
\end{itemize}
\end{itemize}
}\end{remark}
The diagram in Figure~1 illustrates the different cases in Remark~\ref{rem:opeq} and the results of this section concerning the relations between compactness, strict singularity and type of ill-posedness
for injective forward operators. By the way, we should mention that there exist non-injective strictly singular operators possessing a range which contains an
infinite dimensional closed subspace (cf.~\cite[first example]{GoldThorp63}).
\begin{figure}[h]
\begin{center}
\begin{picture}(0,0)%
\includegraphics{figure.pdf}%
\end{picture}%
\setlength{\unitlength}{2072sp}%
\begingroup\makeatletter\ifx\SetFigFont\undefined%
\gdef\SetFigFont#1#2#3#4#5{%
\reset@font\fontsize{#1}{#2pt}%
\fontfamily{#3}\fontseries{#4}\fontshape{#5}%
\selectfont}%
\fi\endgroup%
\begin{picture}(7986,3626)(2491,-5360)
\put(8191,-4246){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\familydefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}operator compact}%
}}}}
\put(4481,-3493){\rotatebox{13.0}{\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\familydefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}operator not}%
}}}}}
\put(4574,-3785){\rotatebox{13.0}{\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\familydefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}strictly singular}%
}}}}}
\put(7839,-2998){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\familydefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}operator strictly}%
}}}}
\put(7852,-3273){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\familydefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}singular}%
}}}}
\put(4590,-2632){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\bfdefault}{\updefault}{\color[rgb]{0,0,0}well-posed}%
}}}}
\put(8438,-2450){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\bfdefault}{\updefault}{\color[rgb]{0,0,0}ill-posed of type II}%
}}}}
\put(4653,-4638){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\bfdefault}{\updefault}{\color[rgb]{0,0,0}ill-posed of type I}%
}}}}
\end{picture}%
\end{center}
\caption{Relations between strict singularity, compactness and type of ill-posedness for equations~(\ref{eq:B}) with injective bounded linear operator.}
\end{figure}
A scenario completely different from Remark~\ref{rem:opeq} occurs for equation (\ref{eq:opeq}) due to the composition structure (\ref{eq:three}) of the operator $A$. The fact that the non-compact embedding operator $\mathcal{E}_2$ is strictly
singular (cf.~\cite[Theorem]{GoldThorp63}) prevents the occurrence of well-posedness and ill-posedness of type~I in the context of this equation.
\begin{proposition} \label{pro:onlyII}
Under the assumptions stated above equation (\ref{eq:opeq}) is \linebreak always ill-posed of type~II.
\end{proposition}
\begin{proof}
Taking into account Proposition~\ref{pro:Nashedsingular} we only have to show that $A$ is always strictly singular.
But this follows immediately from the composition structure (\ref{eq:three}) and the two facts that
$\mathcal{E}_2$ is strictly singular (see~\cite{GoldThorp63}) and that the composition of a strictly singular operator with a bounded linear
operator is again strictly singular.
\end{proof}
\section{The special case of denoising} \label{sec:denoise}
Finally, we apply our results to a typical denoising problem. Given a noisy signal one wants to remove the noise.
For this purpose one decomposes the signal with respect to a wavelet basis (or any other orthonormal system) and tries to
find a sparse approximation with respect to this basis. Thus, in our setting we choose $\widetilde A$ to be the identity
on $\ell^2$. Since in some applications it might be reasonable to measure the noise in a weaker norm we extend $Y$ to $\ell^q$
with $2\leq q\leq\infty$. Then $A:=\mathcal{E}_q$ is the embedding of $\ell^1$ into $\ell^q$. In the sequel we only look at $A$ and therefore extend
the feasible values for $q$ to $1\leq q\leq\infty$.
The minimization problem (\ref{eq:TikBanach}) now reads as
\begin{equation} \label{eq:Tiklq}
\frac{1}{p}\|\mathcal{E}_q x-y^\delta\|_{\ell^q}^p + \alpha \,\|x\|_{\ell^1}
\to \min, \quad \mbox{subject to} \quad x \in \ell^1,
\end{equation}
where the exact signal $x^\dagger$ is assumed to be nearly sparse, i.e.\ $x^\dagger\in\ell^1$.
From the computational point of view it seems to be helpful to apply $p:=q$ for the exponent in the misfit term of (\ref{eq:Tiklq}) whenever $1<q<\infty$.
If we measure the error after denosing in the $\ell^1$-norm as $\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}$, then the chances of having small errors improve with decreasing values $q$, since the strength of the norm in $Y$ grows if $q$ decreases.
\begin{proposition} \label{pro:lq}
For the embedding operator $A=\mathcal{E}_q$ from $\ell^1$ to $\ell^q$ with $1<q \le \infty$ equation (\ref{eq:opeq}) is ill-posed of type~II.
We have weak convergence $Ae^{(k)} \rightharpoonup 0$ if $k\to\infty$ for $1<q \le \infty$, but no convergence in norm.
For all $1\leq q \le \infty$ and all $k \in \mathbb{N}$ the link condition (\ref{eq:allrange}) is satisfied with $f^{(k)}=e^{(k)}$.
Thus, Theorems~\ref{thm:BFH} and \ref{thm:improved} apply and the corresponding index functions are
\begin{equation}
\varphi_1(t)=2\inf_{n\in\mathbb{N}}\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert
+t\,n\right)\qquad \quad \mbox{if} \quad 1\leq q\leq\infty
\end{equation}
and
\begin{equation} \label{eq:phiimp3}
\varphi_2(t)=2\inf_{n\in\mathbb{N}}\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert+t\,n^\theta\right)\qquad
\end{equation}
with
$$\theta=\begin{cases}
1-\frac{1}{q},&\text{if}\quad 1\leq q<\infty,\\
1,&\text{if}\quad q=\infty.
\end{cases}$$
\end{proposition}
\begin{proof}
The ill-posedness of type~II is a consequence of Proposition~\ref{pro:onlyII} whenever $2 \le q \le \infty$, because the injective and bounded embedding operator from $\ell^2$ to $\ell^q$ plays here the role of $\widetilde A$. On the other hand, the non-existence of an infinite dimensional closed subspace in $\ell^q$ for $1<q<2$ included in $\ell^1$ also follows from the theorem in \cite{GoldThorp63}, since the corresponding embedding operators are strictly singular.
\par
It is evident that $e^{(k)}$ does not converge to zero in the norm of $\ell^q$, but we have $e^{(k)} \rightharpoonup 0$ in $\ell^q$ for all $1<q \le \infty$. Also the validity of (\ref{eq:allrange}) is evident.
\par
The index functions $\varphi_1$ and $\varphi_2$ in the convergence rates theorems can be computed easily for the special case
under consideration.
\end{proof}
\begin{example}[H\"older rates] {\rm If the decay rate $x^\dagger_k \to 0$ as $k \to \infty$ of the remaining solution coefficients is of power type
\begin{equation} \label{eq:lqdecay}
|x^\dagger_k| \le C_{x^\dagger}\,k^{\mu-1},\; k \in \mathbb{N}, \quad\text{or equivalently}\quad\sum_{k=n+1}^\infty\vert x^\dagger_k\vert\leq K_{x^\dagger}\,n^{-\mu},\; n \in \mathbb{N},
\end{equation}
with constants $\mu>0$ and $C_{x^\dagger},K_{x^\dagger}>0$, we immediately derive from Proposition~\ref{pro:lq} and Example~\ref{ex:hoelderrates}
the H\"older convergence rates for the denoising problem with forward operator $A=\mathcal{E}_q$ as
\begin{equation} \label{eq:lqrate}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O\left(\delta^{\frac{\mu}{\mu+1-\frac{1}{q}}}\right) \quad \mbox{as} \quad \delta \to 0 \qquad \mbox{if} \quad 1\leq q<\infty
\end{equation}
and
\begin{equation} \label{eq:lqrateinf}
\|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O\left(\delta^{\frac{\mu}{\mu+1}}\right) \quad \mbox{as} \quad \delta \to 0 \qquad \mbox{if} \qquad q=+\infty.
\end{equation}
As expected the rate grows if $q$ decreases, i.e., if the noise is measured in a stronger norm.
On the other hand, the borderline case $q=1$ leads to a well-posed equation (\ref{eq:opeq}).
In this case the index function $\varphi_2$ attains the form
$$\varphi_2(t)=2\inf_{n\in\mathbb{N}}\left(\sum_{k=n+1}^\infty\vert x^\dagger_k\vert
+t \right)=2t\qquad \quad \mbox{if} \quad q =1$$
and the corresponding rate is
$$ \|{x_\alpha^\delta}-x^\dagger\|_{\ell^1}= O\left(\delta\right) \quad \mbox{as} \quad \delta \to 0 \qquad \mbox{if} \qquad q=1,$$
which is typical for well-posed situations.
}\end{example}
The example also shows that the improved index function $\varphi_2$ from Theorem~\ref{thm:improved} indeed provides a better convergence rate for
$1\leq q<\infty$ than the original index function $\varphi_1$ from Theorem~\ref{thm:BFH}.
Note that $\varphi_1$ in Proposition~\ref{pro:lq} is for all $q$ the same function as $\varphi_2$ with $q=\infty$.
At the end we should mention that the rate results (\ref{eq:lqrate}) and (\ref{eq:lqrateinf}) yield the values $0<\nu \le 1$ in formula (\ref{firstrate}) from Example~\ref{ex:hoelderrates}. Consequently,
the H\"older rates in Examples~\ref{ex:cesaro} and \ref{ex:diagonal} with $\nu>1$ are always lower than the observed rates for the denoising case, which indicates a lower degree of ill-posedness for the denoising problem.
\subsection*{Acknowledgement}
The authors very appreciate the fruitful discussion with Radu I.~Bo\c{t} (University of Vienna) and are particularly grateful that he brought the paper \cite{GoldThorp63} to our attention. We also express our thanks to Peter Stollmann and Thomas Kalmes (TU Chemnitz) for valuable hints.
Jens Flemming and Bernd Hofmann were supported by the German Research Foundation (DFG) under grants FL~832/1-1 and HO~1454/8-2, respectively. Ivan Veseli\'c was supported by the PPP-grant 56266051 of the DAAD and the Ministry of Science of the Republic of Croatia.
|
\section{INTRODUCTION}
Sketch is an important artistic drawing style and may be the simplest form since it is only composed of lines. An interesting application is searching image databases using free-hand sketch queries \cite{smith1997visualseek, ren2014sketch}.
However, drawing a vivid sketch portrait is time consuming even for a skilled artist. Automatic face sketch generation has been studied for a long time and it has many useful applications for digital entertainment \cite{liu2006mapping}.
Another important application based on face sketch is to assist law enforcement. Assumed that we need to automatically retrieval a photo from the database for a query image, which help the police to narrow down the suspect quickly. Unfortunately, the photo of the suspect maybe unavailable in most cases. To deal with such a problem, the best substitute available is an artist drawing based on the recollection of an eyewitness.
In this situation, we only focus on sketches without exaggeration, so that the sketch can realistically reflect the real person. Figure \ref{fig:00_samples_on_CUFS} shows two samples of photo-sketch pairs in CUHK Face Sketch Database \cite{wang2009face}.
\begin{figure}[!h]
\centering
\ifeps
\includegraphics[width=0.48\textwidth]{00_samples_on_CUFS.pdf}
\else
\includegraphics[width=0.48\textwidth]{figure/00_samples_on_CUFS.jpg}
\fi
\caption{Samples of photo-sketch pairs: (a) photos; (b) sketches drawn by artist}
\label{fig:00_samples_on_CUFS}
\end{figure}
Figure \ref{fig:00_samples_on_CUFS} indicates the great difference between photos and sketches. Thus, photo based face verification methods cannot be directly applied in this problem. The key to sketch based face verification is to reduce the modality difference between photos and sketches.
One intuitive idea is to recover the photo image from a sketch. However, it's an ill-pose problem because a sketch may lose many informations during the drawing procedure. An alternative way is to generate a pseudo-sketch from a photo, which has been discussed in \cite{tang2004face, liu2005nonlinear, wang2009face}.
The original generation scheme is trying to find a transformation from photos to sketches. Intuitively, it should be a complex nonlinear mapping. Thus, former works like \cite{tang2004face, liu2005nonlinear, wang2009face, xiao2009new, zhang2011svr, wang2012semi, song_eccv14_sketch} simplify the generation problem to \emph{synthesis} form. \iffalse It's under the assumption that if two photo images (or patches) are similar, that their corresponding sketch images (or patches) should also be similar.\fi The underlying
assumption is that, if two photo images (or patches) are similar, their corresponding sketch images (or patches) should also be similar. Thus, if we find out a way to use other photo images (or patches) to synthesize a photo image (or patch), then we can use the corresponding sketch images (or patches) to synthesize its pseudo-sketch.
However, this simplification may has some limitation, especially the scalability. The time cost in synthesis based method grows linearly with the amounts of training data, because it need to use the samples in training set to synthesize a new one.
On the contrary, our method try to \emph{directly} solve the generation problem with an end-to-end model called fully convolutional network (FCN), which can be regarded as a special kind of convolutional neural networks (CNNs). More precisely, FCN is stacked by \emph{only} convolutional layers. It can solve complex nonlinear problem while producing pixel-wise outputs, which is very suitable for the photo-sketch generation problem. Please refer to Figure \ref{fig:06_network_architecture} to get an intuitive idea.
The main contributions of this work are three-folds. First, an end-to-end photo-sketch generation model is studied with a novel architecture of fully convolutional networks, which is original to the best of our knowledge. Second, we present a joint generative-discriminative formulation driving the network optimization, such that the generated face sketches can characterize the person detail as well as the discriminability against other individuals. Third, our system demonstrates superior performances compared with other state-of-the-art approaches on several benchmarks.
Rest of the paper is organized as follows: Section 2 discusses the relative work in photo-sketch synthesis and convolutional neural networks. In Section 3, we will interpret our model and its implementation. In Section 4, extensive experiments suggest that our approach outperforms other state-of-the-art methods on several benchmarks. In Section 5, we draw the conclusion of our work and list some ideas we may adopt in the future.
\section{RELATIVE WORK}
\subsection{Photo-Sketch Synthesis and Recognition}
In recent ten years, there has been several works on the topic of photo-sketch synthesis and recognition.
\citet{tang2004face} were the first to address the problem with a significant amount of database. They proposed a synthesis method based on eigen transformation (ET). It was under the assumption that the photo-sketch mapping can be approximated as linear, yet this assumption may be too strong especially when the hair region was included. Then, it used reconstruction error based distance for recognition.
\citet{liu2005nonlinear} proposed a nonlinear face sketch synthesis method called locally linear embedding (LLE), which can be considered as an improved version of \cite{tang2004face}. Instead of modelling the whole image, it applied ET on local patches with overlapping. It adopted a kernel-based nonlinear LDA discriminative classifier for sketch recognition.
Another local-based strategy proposed by \citet{xiao2009new} was based on Embedded Hidden Markov model (E-HMM). They transformed the sketches to pseudo-photos and applied eigenface algorithm for recognition.
In \cite{wang2009face}, \citeauthor{wang2009face} proposed a multiscale Markov Random Fields (MRF) for face photo-sketch synthesis, which can be both applied to photo-to-sketch synthesis and sketch-to-photo synthesis. They evaluated a series of classifiers on the pseudo-images. Random Sample LDA (RS-LDA) performed best among them. \citet{zhang2010lighting} then improved the MRF framework by adding shape priors and descriptors robust to lighting variations.
Most of methods above may lose some vital details, which influenced the visual quality and face recognition performance. Thus, \citet{zhang2011svr} added a refinement step on existing approaches. They applied a support vector regression (SVR) based model to synthesize the high-frequency information. Similarly, \citet{gao2012face} proposed a new method called SNS-SRE with two steps, \emph{i.e}.\ sparse neighbor selection (SNS) to get an initial estimation and sparse-representation-based enhancement (SRE) for further improvement.
\subsection{Pixel-Wise Predections via CNNs}
Convolutional neural network (CNN) was widely used in computer vision. Its typical structure contains a series of convolutional and pooling layers as feature extractors, and several full connected layers as classifiers to give prediction. It has achieved great success in large scale object classification, localization and detection \cite{krizhevsky2012imagenet, he2014spatial, sermanet-iclr-14, erhan2013scalable, szegedy2013deep, girshick2013rich}.
One important application of CNN was to produce dense or even pixel-wise predictions. \citet{sermanet-iclr-14} developed a CNN-based framework called OverFeat, which integrated recognition, localization and detection. The key component was a pooling layer with offsets, which can imitate sliding window technique and produce dense outputs in the final layer. A similar idea called spatial pyramid pooling (SPP) was proposed by \citet{he2014spatial}, which was also modified the last pooling layer. \citet{wang2014deep} proposed a joint architecture for generic object extraction, and \citet{luo2013pedestrian} applied a deep decompositional network for pedestrian parsing. Both of them shared an idea, which is adopting a full connected layer on the top of the network to produce a dense prediction. Moreover, patch-by-patch scanning technique on the original image has been studied by \citet{ciresan2012deep} in neuronal membrane segmentation and \citet{farabet2013learning} for scene labeling.
Recently, \citet{dong2014learning} applied CNN for image super-resolution and it can produce a pixel-wise output.
They discussed its relationship to sparse-coding-based methods, and concluded that they can use three convolutional layers to simulate the representation-mapping-reconstruction procedure in sparse coding.
Our work was mostly inspired by \cite{dong2014learning}. We conducted our early experiments via their network and then further improved it. We called it fully convolutional network (FCN, a similar idea was also proposed in \cite{long_shelhamer_fcn}), for it only contains convolutional layers and the corresponding activation function, but without any other layers like pooling, full connected, local response normalization (LRN), \emph{etc}.\ We would further discuss it in Section 3.2.
\section{PHOTO-SKETCH GENERATION}
\subsection{Formulation}
In our model, we use two constraints for sketch generation. The first one is the generated sketches should be as close to the ones drawn by artists as possible. This encourages the deep network to learn the sketching skills from the artists. The second one is the generated sketches should be able to facilitate the law enforcement. That is, given a sketch drawn by the artist, it should be able to identify the subject in the photo database. These two constraints are introduced as generative loss and discriminative regularizer in our objective function, which is similar to \citep{ding2015deep}.
Suppose there are $N$ subjects in the training set with each subject containing one photo $P_i$ and one sketch $S_i$. We use $f(\mathbf W,P_i)$ to denote the sketch generated by a fully convolutional network parameterized by $\mathbf W$. We define our loss function as follows where $L_{gen}$ and $L_{disc rim}$ represent the generative loss and the discriminative regularizer respectively. Here we use $\alpha$ to control the weight of the discriminative regularizer to the overall objective.
\begin{equation}
L(P,S,\mathbf W)=L_{gen}(P,S,\mathbf W)+\alpha L_{discrim}(P,S,\mathbf W)
\end{equation}
For the generative loss, we use a straight function which is defined as the pixel-wise difference between the ground-truth sketch and the generated one.
\begin{equation}
L_{gen}\left(P, S, \mathbf W \right) = \frac{1}{N}\sum_{i=1}^{N}\left ( S_{i} - f(\mathbf W,P_{i}) \right ) ^ 2
\label{equation:generative-loss}
\end{equation}
For the discriminative regularizer, we encourage the drawn sketch of one particular person should be different from the generated sketch of another person as defined by the following function.
\begin{equation}
\begin{split}
&L_{discrim}\left(P, S, \mathbf W \right) = \\
&\frac{1}{N\left(N-1 \right)}\sum_{i = 1}^{N}\sum_{j=1,j\not= i}^{N}\log\left(1 + \mathrm{e} ^{ - \frac{\left(S_{i} - f(W,P_{j})\right)^2}{ \lambda}} \right)
\end{split}
\label{equation:discriminative_loss}
\end{equation}
$\lambda$ is a parameter used to avoid numeric overflow.
\textbf{Parameter Optimization}
Using this model, the learning process is to minimize the loss function $L(P,S,\mathbf W)$. This is achieved by the standard network propagation algorithm in the batch training manner. More precisely, in each iteration, we randomly select a set of photo-sketch pairs and construct the generative loss items and discriminative regularizer items respectively. In order to derive the gradient with respect to the network parameter $\mathbf W$, the key step is to calculate the partial derivative of the loss function with respect to the output (generated sketch) of each photo. As one photo may be involved into several cost items, we need to go through each items to accumulate the derivative with respect to the output. Algorithm $\ref{alg:batchTraining}$ gives the details of the learning process.
\begin{small}
\begin{algorithm}[htb]
\caption{Parameter Optimization}
\label{alg:batchTraining}
\begin{algorithmic}[1]
\REQUIRE ~~\\
Training photos $\{P_i\}$ and sketches $\{S_i\}$;
\ENSURE ~~\\
Network Parameters $\mathbf{W}$
\WHILE {$t<T$}
\STATE $t\leftarrow t+1$;
\STATE Randomly select a subset of photos and sketches $\{P'_i\},\{S'_j\}$ from the training set;
\FORALL {$P'_i$}
\STATE Do forward propagation to get $f(\mathbf W, P'_i)$
\ENDFOR
\STATE $\Delta \mathbf W=0$
\FORALL {$P'_i$}
\STATE Calculate the partial derivative with respect to the output: $\frac {\partial L}{\partial f(\mathbf W, P'_i)}$
\STATE Run backward propagation to obtain the gradient with respect to the network parameter: $\Delta \mathbf W_i$
\STATE Accumulate the gradient: $\mathbf W+=\Delta \mathbf W_i$
\ENDFOR
\STATE $\mathbf{W}^t=\mathbf{W}^{t-1}-\lambda_t \Delta \mathbf{W}$
\ENDWHILE
\end{algorithmic}
\end{algorithm}
\end{small}
\begin{figure*}[!t]
\centering
\ifeps
\includegraphics[width=1.0\textwidth]{06_network_architecture.pdf}
\else
\includegraphics[width=1.0\textwidth]{figure/06_network_architecture.jpg}
\fi
\caption{The overview of our model. It takes a full-size photo image as input and directly generates a full-size pseudo-sketch as output. The middle part is the architecture of our fully convolutional network. It contains six convolutional layers, with rectified linear units as activation functions (omitted in the figure).}
\label{fig:06_network_architecture}
\end{figure*}
\subsection{Fully Convolutional Representation}
In this subsection, we will discuss the property of convolutional layers and how they can be cascaded as a pixel-wise fully convolutional representation of an input image.
\textbf{Composition of Convolutions}
A typical convolutional layer has $K$ kernels with activation function $f$ can be formulated as
\begin{equation}
y^{k}_{ij} = f((\mathbf{W}_{k} * x)_{ij}+b_{k})
\label{equation:convolutional layer}
\end{equation}
$x$ denotes the input feature maps. $k \in \{1,2,...,K\}$ denotes the index of a filter, and $\mathbf{W}_{k}$ and $b_{k}$ are the $k$-th filter weights and bias. $y^{k}_{ij}$ denotes the element at the coordinate $(i,j)$ on the $k$-th output feature map.
Equation (\ref{equation:convolutional layer}) indicates that the convolutional operation preserves the spatial relationship. What's more, composition of convolutions will not change this property, \emph{i.e}.\ we can use a stack of convolutional layers to represent a complex nonliear mapping, which can be adopted on an input of arbitrary size, and then produce a corresponding spatial output.
\textbf{Relationship to the Patch-Wise Representation}
In Equation (\ref{equation:convolutional layer}), a pixel on the output feature map only will be influenced by a \emph{patch} on the input feature map, \emph{i.e}.\ its receptive field. This property is also maintained under composition of convolutions. Thus, fully convolutional representation on the whole image can be regarded as a \emph{batch} of patch-wise representation on patches of the image. Nevertheless, fully convolutional representation is much more efficient, for the patches overlap significantly (we set stride to 1 in all layers).
\textbf{Border Effect Trade-Off}
Convolutional operations will shrink the size of the feature map. For example, if we adopt a $(3 \times 3)$ convolution operator on a $(3 \times 3)$ input, the output size will be shrank to $(1 \times 1)$.
As more convolutional layers stacked up, more shrinks will be accumulated in the outputs. A possible solution is to add proper padding before the convolution operation. But it will bring on \emph{border effects}, which has been claimed in \cite{dong2014learning}.
Thus, in the trade-off, we don't use padding in the convolutional layers during both the training and testing time. Each $(155\times 200)$ photo image will be shrank to $(143 \times 188)$ in the representation (See Figure \ref{fig:06_network_architecture}). \iffalse \emph{i.e}\. there will be some pixels on the border of input image could not find their corresponding representation on the output feature map.\fi
\subsection{Implementation}
\textbf{Pre-processing}
As it is described in \cite{wang2009face}, all the photos and sketches are translated, rotated, and scaled such that the two eye centers of all the face images are at fixed position in the pre-processing step. This simple geometric normalization step makes the same face components in different images in roughly alignment.
Another pre-processing step is inspired by the results\footnote{http://www.ee.cuhk.edu.hk/\~{}xgwang/sketch\_multiscale.html} of \cite{wang2009face}. The transformation mentioned above produces a $(200\times250)$ image, but it may have some black regions on the border areas. We crop the $(155\times200)$ center part of the image in order to exclude this negative influence.
Since we choose to avoid border effects, the sketch images need to be cropped to $(143 \times 188)$ to fit in the network output dimension.
\textbf{Spatial Patch-wise Learning with Overlapping}
From the previous works on photo-sketch synthesis\cite{liu2005nonlinear, wang2009face}, patch-wise learning with overlapping is very important to handle the non-linearity between photos and sketches. Intuitively, patches in different positions are diverse from others (\emph{e.g}.\ eye patches, nose patches and mouth patches). Therefore, learning different patch representations in different spatial positions respectively is a very straightforward idea.
We handle it by adding additional XY channels in the input, \emph{i.e}.\ the input image data contain five channels, three RGB channels of the photo image, and two channels of the corresponding coordinate $(i, j)$. This tiny modification significantly improves the result, which will be discussed in Section 4.3.
\textbf{Network Architecture}
We apply the network proposed in \cite{dong2014learning} in our early experiments, then we modify its architecture for further enhancement. We borrow the idea in \cite{simonyan2014very}, which has a great success in ILSVRC-2014. Their main contribution is to deeper the CNN under computational constraints via very
small $(3\times3)$ convolutional filters in all layers.
Similarly, we modify the 3 layers network in \cite{dong2014learning} into a 6 layers network, which is shown in Figure \ref{fig:06_network_architecture}. For the $(9\times9)$, $(1\times1)$, $(5\times5)$ convolutional layer in \cite{dong2014learning}, we replace them with two $(5\times5)$, $(1\times1)$ and $(3\times3)$ convolutional layers respectively. Moveover, we double the filter amounts in every layer to improve the network's capacity.
We call it \emph{medium network} due to its size. We have two similar architectures called small network and large network, which will be further discussed in Section 4.3.
\textbf{Training Details}
Our results are based on our medium network (See Figure \ref{fig:06_network_architecture}).
As mentioned above, we first pre-process the photo images into $(155\times200)$ and the sketch images into $(143\times188)$. The optimization objective has been discussed in Section 3.1.
We set $\lambda$ as $10^9$ and set $\alpha$ as $10^4$.
We use Caffe \cite{jia2014caffe} for implementation.
The filter weights are initialized by drawing randomly from a Gaussian distribution with zero mean and standard deviation 0.01, and the bias are initialized by zero. We set the learning rate as $10^{-11}$, then it takes several hours to converge on a NVIDIA Tesla K40 GPU.
\begin{figure*}[!t]
\centering
\ifeps
\includegraphics[width=1.0\textwidth]{01_photo_to_sketch_fig.pdf}
\else
\includegraphics[width=1.0\textwidth]{figure/01_photo_to_sketch_fig.jpg}
\fi
\caption{Compared to synthesis based method \cite{wang2009face}, our generative approach could retain more details in the original photos, \emph{e.g}.\ the two locks of hair in forehead of the man in column (e) of row 1; the two big eyes of the woman in column (a) of row 2. (a)\&(e) photos; (b)\&(f) our generative pseudo-sketches; (c)\&(g) synthesized pseudo-sketches in \cite{wang2009face}; (d)\&(h) sketches drawn by the artist. }
\label{fig:01_photo_to_sketch_fig}
\end{figure*}
\section{EXPERIMENTS}
We evaluate our model on CHUK student dataset \cite{wang2009face}. It includes 188 faces\if false and can be downloaded from this website (http:/mmlab.ie.cuhk.edu.hk/facesketch.html)\fi. 88 faces are selected for training and remaining 100 faces are selected for testing. For each face, there is a sketch drawn by the artist and a photo taken in a frontal pose, under normal lighting condition, and with a neural expression.
We mainly do three aspects of experiments, which will be shown in the next three subsections. Section 4.1 will discuss our photo-sketch generation results and comparison with the synthesis based method in \cite{wang2009face}. Section 4.2 will show that our generated pseudo-sketches significantly reduce the modality between photos and sketches. Thus they can be adopted to a sketch-based face verification system. Section 4.3 will involve empirical study about our model. Several factors such as network depth and filter numbers will be discussed on the evaluation protocol both qualitatively and quantitatively.
\subsection{Face-Sketch Generation}
In Figure \ref{fig:01_photo_to_sketch_fig}, we show some examples of our generated results and compare them to \cite{wang2009face}. For fair comparison, we list all 14 pseudo-sketches which can be found on their website.\iffalse For full result, please go to our web site. (vision.sysu.edu.cn/xxxxx)\fi
Figure \ref{fig:01_photo_to_sketch_fig} indicates that our results contain more vital details.
For example, the man in column (e) of the first row has two little locks of hair on his forehead. But this detail is missing in the pseudo-sketch in \cite{wang2009face} (in column (g)). On the contrary, our generated result (in column(f)) retains this detail information which may be very important to distinguish this man from others.
Another example is the woman in column (a) of the second row. She has a distinctive feature comparing to other persons in the database, for she has two big eyes. \iffalse The synthesis method fail on it while\fi Our method can also capture this detail in the pseudo-sketch (in column(b)).
It suggests that synthesis based methods may fail in some cases. The main reason is that they are under an assumption that the synthesized sketch image (or patch) can be reconstructed by the sketch images (or patches) in the training set. However, this assumption may be too strong, for some persons have their facial distinctions in fact
Our generative method overcomes this difficulty.
We try to \emph{directly} learn the transformation from photo space to sketch space. Even though it's a complex nonlinear mapping, FCN has the capacity to handle this challenge.
Compared to traditional synthesized method, our novel generative approach has two folds of advantages. Firstly, it could retain more detail information from the photos. Secondly, our inference time is independent on the amount of training data. The time cost in synthesis based method grows linearly with the data amounts, while the runtime of our approach is only affected by the size of the input photo.
\subsection{Sketch-Based Face Verification}
In this subsection, we will show that our generated pseudo-sketch significantly reduce the modality between photos and sketches. We follow the same testing procedure in \cite{tang2004face, liu2005nonlinear, wang2009face}, which can be concluded in two steps: (a) convert the photos in testing set into corresponding pseudo-sketches; (b) define a feature or transformation to measure the distance between the query sketch and the pseudo-sketches.
In our implementation, we use our model to generate the pseudo-sketches in procedure (a), and use our generative loss (see Equation (\ref{equation:generative-loss})) as the distance measurement. Since the sizes of our pseudo-sketches are $(188\times143)$, we also crop the query sketch into $(188\times143)$.
Following the same protocol described in \cite{tang2004face}, we compare our approach with previous methods on cumulative match score ($CMS$) in Table \ref{table:CMS-comparison-with-others}. $CMS$ measures the percentage of ``the correct answer is in the top $n$ matches'', where $n$ is called the rank.
As shown in Table \ref{table:CMS-comparison-with-others}, we form a baseline experiment, which is to convert the photo into gray scale as somehow ``pseudo-sketch'' to clarify the modality difference between photo space and sketch space. In this baseline, the accuracy for the first rank only equals to 41\%, which is far away from satisfying. In early method like ET described in \cite{tang2004face}, has got 71\% for the top one match. Latter trials in \cite{wang2009face}, \cite{zhang2010lighting}, \cite{zhang2011svr} have greatly improved the verification accuracy on top one candidate to 96\%, 99\% and even 100\%.
From the last row in Table \ref{table:CMS-comparison-with-others}, our approach also gets 100\% correct answer on its first guess. It's worth to notice that our approach is quite different from the former methods, for it applies generation instead of synthesis.
\begin{table}[!h]
\center
\begin{tabular}{|>{\hfil}p{50pt}<{\hfil}|>{\hfil}p{32pt}<{\hfil}|>{\hfil}p{32pt}<{\hfil}|>{\hfil}p{32pt}<{\hfil}|>{\hfil}p{35pt}<{\hfil}|}
\hline
& Rank 1 & Rank 3 & Rank 5 & Rank 10\\
\hline
Baseline & 41 & 56 & 59 & 70\\
\hline
ET \cite{tang2004face} & 71 & 81 & 88 & 96 \\
\hline
MRF \cite{wang2009face} & 96 & - & - & 100 \\
\hline
MRF+ \cite{zhang2010lighting} & 99 & - & - & 100 \\
\hline
SVR \cite{zhang2011svr} & 100 & - & - & - \\
\hline
Ours \iffalse Method (medium networks)\fi & \textbf{100} & \textbf{100} & \textbf{100} & \textbf{100} \\
\hline
\end{tabular}
\caption{Cumulative Match Scores ($CMS$) comparison on full training set (88 photo-sketch pairs). Our method achieves 100\% accuracy on the first guess.}
\label{table:CMS-comparison-with-others}
\end{table}
Moreover, our generative method is very robust and has excellent generalization ability. In Figure \ref{fig:07_icmr_discriminative_loss}, it suggests that our method can using a portion of training data to get a 100\% accuracy for the first rank. For example, we only need 5 training samples to get to 95\% accuracy, and 27 training samples is enough to get a 100\% score for top one candidate.
We also design a verification on our optimization objective on Figure \ref{fig:07_icmr_discriminative_loss}. It suggests that the model trained with the discriminative regularizer consistently outperforms the model trained without discriminative regularizer.
\begin{figure}[!h]
\ifeps
\includegraphics[width=0.48\textwidth]{07_icmr_discriminative_loss.pdf}
\else
\includegraphics[width=0.48\textwidth]{figure/07_icmr_discriminative_loss.jpg}
\fi
\caption{The model trained with discriminative regularizer consistently outperforms the model trained without discriminative regularizer. Both model gets 100\% accuracy on rank 1 candidate while involves less than half training samples. \emph{i.e}.\ 27 samples in 88 photo-sketch pairs.}
\label{fig:07_icmr_discriminative_loss}
\end{figure}
\subsection{Empirical Study}
In this subsection, further analyses will be conducted on the factors which affect our photo-sketch generation performance.
\textbf{Different Network Architectures}
Our first concern is about the depth of the network and the numbers of filters. We use the network proposed in \cite{dong2014learning} (we call it SR network) in our early experiments. However, the result is not quite satisfying, for this network seems too small to learn the complex nonlinear mapping from photos to sketches. Thus, we improve the network by: (a) adding more layers to the network; and (b) adding more numbers of filters to each layer.
We conduct our experiments on three different network architectures. Due to their scales, we call them small, medium and large network respectively.
The medium network architecture can be found in Figure \ref{fig:06_network_architecture}. The difference between it and SR network please refer to Section 3.3. Moreover, the only difference between our small, medium and large network is their filter numbers. The medium network has two times of filters compared to the small one. And the large network doubles the filter numbers of the medium network.
For comparison, we define a measurement call $MPRL$, which is short for \emph{multiscale pixel-wise reconstruction loss}. It evaluates the pixel-wise accuracy on different scales. To explain $MPRL$, we firstly introduce \emph{pixel-wise reconstruction loss} ($PRL$) on one photo-sketch pair.
For a photo-sketch pair, we denote the sketch as the ground truth ($GT$) and generated pseudo-sketch as the prediction ($P$). Both of their sizes are $(W \times H)$. Thus, $PRL$ can be formulated as
\begin{equation}
PRL = \frac{1}{{W \times H}} \sqrt{ \sum\limits_{x = 1}^W {\sum\limits_{y = 1}^H { \left(GT(x,y) - P(x,y) \right)^2 } } }
\end{equation}
$MPRL$ is the multiscale version of $PRL$. In practice, for each photo-sketch pair, we rescale both the $GT$ and the $P$ to three scales \{0.5, 1, 2\} and evaluate their $PRL$ to form the $MPRL$. Then, for the whole test set, we evaluate all the pairs and average their $MPRL$ as the final measurement.
Figure \ref{fig:02_small_medium_big} summarizes the results both in qualitative and quantitative. The $MPRL$ of SR network is \{34.5, 36.8, 36.2\}. Our small network reduces it to \{32.6, 35.0, 34.4\}. Then, our medium and large network get better performance as \{32.1, 34.6, 34.0\} and \{30.0, 32.5, 31.9\}.
\begin{figure}[!t]
\centering
\ifeps
\includegraphics[width=0.49\textwidth]{02_small_medium_big.pdf}
\else
\includegraphics[width=0.49\textwidth]{figure/02_small_medium_big.jpg}
\fi
\caption{Comparison on the $MPRL$ measurement and generative pseudo-sketches via different network architectures. The model achieves better performance while to deeper the network or to add more filters. (a-d) pseudo-sketches generated via network in \cite{dong2014learning}, our small network, our medium network and our large network.}
\label{fig:02_small_medium_big}
\end{figure}
The pseudo-sketch examples generated by each network are on the bottom part of Figure \ref{fig:02_small_medium_big}. It suggests that the results of SR network can capture the overall structures, but they are far away from satisfying in details. Then, Our small network's generation is quite better and has clearer contours. Moreover, the pseudo-sketches generated by medium network are more vivid than the small ones.
The results of large network are as good as medium ones, but the large network is much more time consuming.
Table \ref{table:run-time-on-different-networks} shows that our small network is almost as fast as the SR network. The medium network's cost is about 2 times of the small one, and the large network's runtime is about three times of the medium one.
To balance the effectiveness and efficiency, we prefer our medium network as the final choice.
\begin{table}[!h] \normalsize
\center
\begin{tabular}{|>{\hfil}p{22pt}<{\hfil}|>{\hfil}p{30pt}<{\hfil}|>{\hfil}p{40pt}<{\hfil}|>{\hfil}p{52pt}<{\hfil}|>{\hfil}p{41pt}<{\hfil}|
\hline
& SR Net & Small Net & Medium Net & Large Net\\
\hline
Time & 8.5ms & 8.6ms & 17.1ms & 50.8ms \\
\hline
\end{tabular}
\caption{Runtime of single $(155\times 200)$ image on NVIDIA Tesla K40 GPU}
\label{table:run-time-on-different-networks}
\end{table}
\textbf{Difference between trained with and without XY channels}
Now we consider another factor that influences the model's performance. As mentioned above, we add XY channels to the RGB color channels. We assume these additional spatial messages could help the network to distinguish different spatial patches and learning different representation on them.
Similarly, we use $MPRL$ for quantitative measurement and use pseudo-sketch comparison for qualitative analysis. In Figure \ref{fig:03_xychannel}, the left part is for results generated by our medium network but trained without XY channels, and the right part is for results generated by our medium network trained normally. Figure \ref{fig:03_xychannel} suggests that latter one outperforms former one both in numbers and perception.
\begin{figure}[!hb]
\centering
\ifeps
\includegraphics[width=0.49\textwidth]{03_xychannel.pdf}
\else
\includegraphics[width=0.49\textwidth]{figure/03_xychannel.jpg}
\fi
\caption{Comparison on the $MPRL$ measurement and generative pseudo-sketches via network trained without and with XY channels. Adding XY channels can make the model more robust. (a) pseudo-sketches generated via network trained without XY channels; (b) pseudo-sketches generated via network trained with XY channels.}
\label{fig:03_xychannel}
\end{figure}
\section{CONCLUSION AND FUTURE WORK}
In this paper, we propose an end-to-end fully convolutional network in order to directly model the complex nonlinear mapping between face photos and sketches. From the experiments, we find out that fully convolutional network is a powerful tool which can handle this difficult problem while providing a pixel-wise prediction both effectively and efficiently.
However, this solution still has some limitations. Firstly, the synthesized pseudo-sketches in \cite{wang2009face} have shaper edges and clearer contours than our generated ones. Secondly, the database involved in our experiments only contains Asiatic face images, which may limit the generalization ability of our model to other racial groups.
In future work, we will further improve our loss function and try various databases in our experiments, and we may explore about the relation between our work and those involved with non-photorealistic rendering \cite{lin2013video}.
\textbf{Acknowledgements}
We gratefully acknowledge NVIDIA for GPU donations.
\setlength{\bibsep}{0.5ex}
\bibliographystyle{abbrvnat}
|
\section{Introduction}
Star formation is a fundamental parameter of galaxies that describes how galaxies evolve, when used in conjunction with mass. As the process of star formation depletes a galaxy of its gas, it must be continuously replenished by infall from the intergalactic medium to be supported for an extended time. When massive stars die, they enrich the surrounding interstellar medium with heavy metals, thus altering a galaxy's chemical composition. Therefore, accurately tracing star formation though cosmic time gives key constraints on how galaxies are able to form and evolve \citep[e.g.,][and references therein]{tinsley68,somerville12,madau&dickinson14}.
For these reasons, great efforts have been made to calibrate a wide range of the electromagnetic spectrum that can be linked to processes involved with recent star formation \citep[see review by][]{kennicutt12}. In particular, infrared (IR) wavelength calibrations are proving to be critical to understanding galaxies in the early Universe. Deep IR surveys with the \textit{Spitzer} and \textit{Herschel} Space Telescopes have revealed that the majority of star formation that occurs at redshift $z\sim1-3$ is enshrouded by dust \citep[e.g.,][]{murphy11a, elbaz11}, making it very difficult to measure accurate SFRs at optical wavelengths. In addition, IR-bright galaxies ($L\gtrsim\times10^{11} L_\odot$) are much more prevalent during that time than today \citep[e.g.,][and references therein]{chary&elbaz01,lefloch05,magnelli09,murphy11a, elbaz11, lutz14}. Furthermore, observations suggest that $\sim$$85\%$ of today's stars were formed at redshift $0<z<2.5$ \citep{marchesini09,muzzin13,tomczak14}. Together these results have renewed interest in monochromatic (i.e., single-band) mid-IR (MIR) star formation rate (SFR) indicators, as distant galaxies can easily be observed in the MIR.
Dust emission in the MIR is more closely related to star formation than longer IR wavelengths, where heating by low mass (i.e., long-living) stars becomes important, which has allowed several wavelength bands in the MIR to be well calibrated locally as SFR diagnostics \citep{zhu08,rieke09,calzetti10}. However, difficulties arise in utilizing local calibrations because the regions of rest-frame wavelengths probed by a given band will vary with redshift. As a reference, the \textit{Spitzer} $24~\micron$ and the \textit{Herschel} $70~\micron$ bands target the rest-frame $8~\micron$ and $23~\micron$ emission, respectively, for a galaxy at redshift $z=2$. Correcting for this effect is most commonly achieved through $k$-corrections which depend heavily on the assumed galaxy spectral energy distribution (SED) template. Such templates \citep[e.g.,][]{chary&elbaz01,dale&helou02,polletta07,rieke09,brown14} typically require many photometric bands for accurate matching. Therefore, in order to fully utilize current and future deep IR imaging surveys for a greater understanding the formation and evolution of galaxies without a reliance on extensive multi-band imaging, continuous single-band SFR indicators will be imperative.
In the near future, the Mid-Infrared Instrument (MIRI; 5--28~$\mu$m) on board the James Webb Space Telescope (JWST) will expand our ability to probe galaxies in the MIR, detecting down to the regime of normal star-forming disk galaxies ($L\lesssim3\times10^{11} L_\odot$) out to $z=3$ and representing an order of magnitude improvement in sensitivity over \textit{Spitzer} bands of similar wavelength\footnote{\url{http://www.stsci.edu/jwst/instruments/miri/instrumentdesign/filters/}}. Thus, current and future cosmological surveys are highlighting the need for continuous monochromatic SFR indicators that cover, without breaks, the MIR wavelength range of $6-70~\micron$. This will provide a flexible tool that can be applied to any galaxy up to redshift $z\approx3$. With the release of the Wide-field Infrared Survey Explorer (WISE) All-Sky Survey \citep{wright10}, times are ripe for consolidating all these data into a coherent picture. In this study, we use \textit{GALEX}, Sloan Digital Sky Survey (SDSS), WISE, and \textit{Spitzer} data of a large sample of local galaxies to perform the calibration of SFR($\lambda$) in the $6-70~\micron$ range.
Throughout this work we adopt the WMAP five-year cosmological parameters, $H_0=70.5$~km/s/Mpc, $\Omega_M=0.27$, $\Omega_{\Lambda}=0.73$ \citep{komatsu09}. We assume a \citet{kroupa01} initial mass function (IMF) for all SFR calibrations. The infrared luminosity, $L_{\rm{IR}}$, of a galaxy refers to the integrated luminosity over the region from 8 to $1000~\micron$ ($L_{\rm{IR}}=\int_{8\mu\rm{m}}^{1000\mu\rm{m}} L_\nu d\nu$).
\section{Data}
\subsection{The SSGSS Sample}
The \textit{Spitzer}--SDSS--\textit{GALEX} Spectroscopic Survey (SSGSS) is a sample of 101 galaxies located within the \textit{Spitzer} Wide-Area Infrared Extragalactic (SWIRE) Survey/Lockman Hole area at $0.03 < z < 0.22$ \citep{treyer10,odowd11}. These galaxies represent a subset of the 912 galaxies within the \citet{johnson06} sample, which has extensive multi-wavelength coverage from \textit{Spitzer}, SDSS, and \textit{GALEX}, and for which \textit{Spitzer} infrared spectrograph (IRS) measurements have also been obtained. The UV data is from pipeline-processed \textit{GALEX} observations of this regions with average exposures of $\sim$1.5~ks. The optical photometry comes from the seventh data release of the SDSS main galaxy sample \citep[DR7;][]{abazajian09}. The optical spectroscopic measurements are from the Max Planck Institute for Astrophysics and Johns Hopkins University (MPA/JHU) group\footnote{\url{http://www.mpa-garching.mpg.de/SDSS/DR7/}}, which is based on the method presented in \citet{tremonti04}. The infrared photometry comes from the SWIRE survey observations \citep{lonsdale03}.
Each galaxy in this sample has been observed with the \textit{Spitzer} IRAC and MIPS bands, in addition to observations using the blue filter of the IRS peak-up facility. These blue filter peak-ups have spectral coverage from $13.3-18.7~\micron$ and give an additional photometric point at $16~\micron$, between the IRAC $8~\micron$ and MIPS $24~\micron$ bands. Aperture photometry was performed in 7\arcsec\ and 12\arcsec\ radius apertures for the $3.6-8~\micron$ IRAC and $24~\micron$ MIPS, respectively, and then aperture-corrected to 12.2\arcsec\ and $>35$\arcsec, respectively. For the MIPS 70 and $160~\micron$ bands, nearly all the galaxies can be treated as point sources, and aperture corrections were taken from the MIPS handbook. A full description of these aperture corrections is described in \citet{johnson07}. For a more detailed description of the SSGSS dataset, we refer the reader to \citet{odowd11}.
The IRS spectroscopy for the SSGSS sample is obtained through the NASA/IPAC Infrared Science Archive (IRSA) website\footnote{\url{http://irsa.ipac.caltech.edu/data/SPITZER/SSGSS/}} and is from the work of \citet{odowd11}. This study utilizes the lower resolution Short-Low (SL) and Long-Low (LL) IRS modules, as these have been obtained for the entire SSGSS sample. The SL module spans $5.2-14.5~\micron$ with resolving power $R = 60-125$ and has a slit width of $3.6-3.7$\arcsec. The LL module spans $14-38~\micron$ with resolving power $R = 57-126$ and has a slit width of $10.5-10.7$\arcsec. The spectra from the two modules were combined by weighted mean and a detailed description of the method can be found in \citet{odowd11}. At $z\sim0.1$ these galaxies are sufficiently distant such that the IRS slit encompasses a significant fraction of each galaxy ($r$-band Petrosian diameters are $\sim$$10$\arcsec), providing some of the best MIR SEDs for a continuous SFR($\lambda$) determination.
In order to accurately determine a diagnostic for star formation across the MIR, it is necessary to only consider cases where the majority of light is being contributed from stars (i.e., star-forming galaxies; SFGs) and not from an active galactic nucleus (AGN). The galaxy type is traditionally determined according to their location on the Baldwin-Phillips-Terlevich (BPT) diagram \citep{baldwin81,kewley01,kauffmann03}. By adopting the DR7 values of emission line measurements, we find that 64 of the 101 SSGSS galaxies are classified as SFGs. We note that the initial classification of SSGSS galaxies by \citet{odowd11} utilized SDSS DR4 measurements, which results in a few BPT designations to differ between these works. We further exclude 6 of the SSGSS galaxies classified as SFGs from our analysis for the following reasons: SSGSS~18 appears to be a merger, SSGSS~19, 22, and 96 have significant breaks in their IRS spectra due to low signal-to-noise (S/N), and SSGSS~35 and 51 suffer from problems with IRS confusion. This leaves 58 galaxies to be used in our calibration of a monochromatic MIR SFR indicator. Table~\ref{Tab:IRS_correct} shows the SSGSS IDs of the galaxies used for this study along with some of their properties (additional parameters in the table are introduced in later sections).
Our sample of 58 galaxies span a redshift range of $0.03\le z \le0.22$ with a median redshift of 0.075. The range of infrared luminosity is $9.53\le \log(L_{\rm{IR}}/L_{\odot})\le 11.37$, with a median of $10.55$. All measurements of $L_{\rm{IR}}$\ for these galaxies are taken from the original SSGSS dataset \citep[presented in][]{treyer10}. The selection criteria for the SSGSS sample was based on $5.8~\micron$ surface brightness and $24~\micron$ flux density, and this restricts our sample of 58 galaxies to relatively high stellar masses ($1.6\times10^9\le M/M_{\odot}\le 1.7\times10^{11}$) and metallicities ($8.7\le 12+\log(\rm{O/H})\le 9.2$). These stellar mass and metallicity estimates are updated from the SSGSS dataset values (based on DR4) to the MPA-JHU DR7 estimates.
\begin{table*}
\begin{center}
\caption{Summary of Galaxy Properties and IRS Correction Terms \label{Tab:IRS_correct}}
\begin{tabular}{ccccccccc}
\hline\hline
SSGSS& R.A. & Decl. & $z$ & log($L_{\rm{IR}}$)& $\langle SFR \rangle$ &$c_{\rm{phot}}$ & $k$ & \multirow{2}{*}{$\frac{(16+k)}{(8+k)}$} \\
ID & (J2000) & (J2000) & & $(L_{\odot})$ &$(M_{\odot}/\rm{yr})$& & & \\ \hline
1 & 160.34398 & 58.89201 & 0.066 & 10.55 & 4.22 & 0.963 & 14.10 & 1.362 \\
2 & 159.86748 & 58.79165 & 0.045 & 9.83 & 0.65 & 1.062 & 23.64 & 1.253 \\
3 & 162.41000 & 59.58426 & 0.117 & 10.59 & 3.41 & 1.091 & 54.09 & 1.129 \\
4 & 162.54131 & 59.50806 & 0.066 & 10.22 & 1.29 & 0.871 &--4.466 & 3.264 \\
5 & 162.36443 & 59.54812 & 0.217 & 11.37 & 19.55 & 1.088 & 24.48 & 1.246 \\
6 & 162.52991 & 59.54828 & 0.115 & 10.99 & 7.41 & 0.851 & 1.310 & 1.859 \\
8 & 161.48123 & 59.15443 & 0.044 & 9.97 & 0.79 & 0.843 &--5.224 & 3.882 \\
14 & 161.92709 & 56.31395 & 0.153 & 11.06 & 9.60 & 0.918 & 10.69 & 1.428 \\
16 & 162.04231 & 56.38041 & 0.072 & 10.42 & 2.15 & 0.943 & 15.73 & 1.337 \\
17 & 161.76901 & 56.34029 & 0.047 & 10.83 & 5.58 & 0.922 & 21.94 & 1.267 \\
24 & 163.53931 & 56.82104 & 0.046 & 10.59 & 3.43 & 1.077 & 0.240 & 1.971 \\
25 & 158.22482 & 58.10917 & 0.073 & 10.30 & 1.89 & 1.061 & 2.363 & 1.772 \\
27 & 159.34668 & 57.52069 & 0.072 & 11.01 & 7.70 & 0.950 &--0.933 & 2.132 \\
30 & 159.73558 & 57.26361 & 0.046 & 10.11 & 1.06 & 1.040 & 63.01 & 1.113 \\
32 & 161.48724 & 57.45520 & 0.117 & 10.82 & 6.30 & 1.121 & 278.9 & 1.028 \\
34 & 160.30701 & 57.08246 & 0.046 & 9.96 & 0.79 & 0.987 & 46.47 & 1.147 \\
36 & 159.98523 & 57.40522 & 0.072 & 10.39 & 2.01 & 0.996 &--1.629 & 2.256 \\
38 & 160.20963 & 57.39475 & 0.118 & 10.92 & 6.43 & 0.905 & 1.989 & 1.801 \\
39 & 159.38356 & 57.38491 & 0.074 & 10.14 & 1.28 & 0.684 & 1.615 & 1.832 \\
41 & 158.99098 & 57.41671 & 0.102 & 10.34 & 1.68 & 0.818 & 2.159 & 1.787 \\
42 & 158.97563 & 58.31007 & 0.155 & 11.03 & 9.32 & 1.170 & 14.84 & 1.350 \\
46 & 159.02698 & 57.78402 & 0.044 & 10.02 & 0.99 & 0.932 & 3.183 & 1.715 \\
47 & 159.22287 & 57.91185 & 0.102 & 10.68 & 4.34 & 1.331 & 61.86 & 1.115 \\
48 & 159.98817 & 58.65948 & 0.200 & 11.24 & 15.53 & 0.869 & 11.59 & 1.408 \\
49 & 159.51942 & 58.04882 & 0.091 & 10.49 & 3.13 & 1.107 & 0.495 & 1.942 \\
52 & 160.54201 & 58.66098 & 0.031 & 9.53 & 0.29 & 1.129 &--2.795 & 2.537 \\
54 & 160.41264 & 58.58743 & 0.115 & 11.20 & 11.53 & 0.901 & 6.117 & 1.567 \\
55 & 160.29353 & 58.25641 & 0.121 & 10.54 & 2.95 & 0.865 & 3.155 & 1.717 \\
56 & 160.41617 & 58.31722 & 0.072 & 10.01 & 0.85 & 0.886 & 4.645 & 1.633 \\
57 & 160.12233 & 58.16783 & 0.073 & 9.92 & 0.75 & 0.689 & 15.28 & 1.344 \\
59 & 159.89861 & 57.98557 & 0.075 & 10.36 & 2.00 & 0.928 & 15.09 & 1.346 \\
60 & 160.51027 & 57.89706 & 0.116 & 10.48 & 2.89 & 0.910 & 6.530 & 1.551 \\
62 & 160.91280 & 58.04736 & 0.133 & 11.08 & 9.85 & 0.978 & 0.688 & 1.921 \\
64 & 161.00317 & 58.76030 & 0.073 & 10.88 & 5.44 & 0.913 & 3.276 & 1.709 \\
65 & 161.37666 & 58.20886 & 0.118 & 11.18 & 11.85 & 0.965 & 12.55 & 1.389 \\
66 & 161.25533 & 57.77575 & 0.113 & 10.86 & 6.75 & 0.937 & 13.86 & 1.366 \\
67 & 161.18829 & 58.45495 & 0.031 & 10.09 & 1.21 & 1.581 & 8.790 & 1.476 \\
68 & 163.63458 & 57.15902 & 0.068 & 10.54 & 3.37 & 0.975 & 25.56 & 1.238 \\
70 & 163.17673 & 57.32074 & 0.090 & 10.49 & 2.81 & 1.037 & 255.8 & 1.030 \\
71 & 163.21991 & 57.13160 & 0.163 & 10.98 & 8.32 & 1.271 & 285.0 & 1.027 \\
72 & 163.25565 & 57.09528 & 0.080 & 10.79 & 5.01 & 0.951 & 14.00 & 1.364 \\
74 & 161.95050 & 57.57723 & 0.118 & 10.80 & 5.07 & 0.878 &--2.430 & 2.436 \\
76 & 162.02142 & 57.81512 & 0.074 & 10.55 & 2.62 & 0.938 & 6.128 & 1.566 \\
77 & 162.10524 & 57.66665 & 0.044 & 9.69 & 0.62 & 1.130 & 66.79 & 1.107 \\
78 & 162.12204 & 57.89890 & 0.074 & 10.56 & 3.17 & 1.027 &--4.810 & 3.508 \\
79 & 161.25693 & 57.66116 & 0.045 & 9.82 & 0.81 & 1.023 &--0.238 & 2.031 \\
80 & 162.07401 & 57.40280 & 0.075 & 10.35 & 2.14 & 0.992 &--1.949 & 2.322 \\
81 & 162.04674 & 57.40856 & 0.075 & 10.24 & 1.69 & 0.909 & 1.709 & 1.824 \\
82 & 161.03609 & 57.86136 & 0.121 & 10.77 & 4.98 & 0.854 & 2.444 & 1.766 \\
83 & 160.77402 & 58.69774 & 0.119 & 10.92 & 6.56 & 0.919 &--3.474 & 2.768 \\
88 & 161.38522 & 58.50156 & 0.116 & 10.51 & 2.54 & 1.106 & 42.43 & 1.159 \\
90 & 162.64168 & 59.37266 & 0.153 & 11.02 & 8.66 & 0.796 & 8.033 & 1.499 \\
91 & 162.53705 & 58.92866 & 0.117 & 10.78 & 5.31 & 0.782 &--3.869 & 2.937 \\
92 & 162.65512 & 59.09582 & 0.032 & 10.18 & 1.42 & 0.966 & 6.675 & 1.545 \\
94 & 161.80573 & 58.17759 & 0.061 & 10.13 & 0.90 & 0.919 & 9.281 & 1.463 \\
95 & 163.71245 & 58.39082 & 0.115 & 10.82 & 5.74 & 0.956 & 22.68 & 1.261 \\
98 & 164.14571 & 58.79676 & 0.050 & 10.59 & 3.70 & 0.982 & 5.797 & 1.580 \\
99 & 164.33247 & 57.95170 & 0.077 & 10.82 & 5.58 & 0.903 &--1.944 & 2.321 \\ \hline
\end{tabular}
\end{center}
\textbf{Notes.} Columns list the (1) galaxy ID number, (2) redshift, (3) integrated infrared luminosity from $8-1000~\micron$, (4) average SFR from the diagnostics in Table~\ref{Tab:calibrations}, (5) offset between IRS spectra and global photometry above $16~\micron$ (6) correction parameter for wavelength-dependent aperture loss of IRS spectrum below $16~\micron$, (7) correction factor of the spectrum at $8~\micron$ due to wavelength-dependent aperture loss.
\end{table*}
\subsection{WISE Data}
The WISE All-Sky Survey provides photometry at 3.4, 4.6, 12, and $22~\micron$ \citep{wright10} which complements the wealth of IR data available for the SSGSS sample. Most importantly for this study, the WISE $12~\micron$ band provides a crucial photometric point that bridges the gap between the \textit{Spitzer} $8~\micron$ and $24~\micron$ bands, a section of the MIR SED that experiences a transition from polycyclic aromatic hydrocarbon (PAH) emission features and dust continuum. The WISE photometry for the SSGSS sample is obtained through the NASA/IPAC Infrared Science Archive (IRSA) website\footnote{\url{http://irsa.ipac.caltech.edu/Missions/wise.html}}. Using an approach similar to that of \citet{johnson07} for the \textit{Spitzer} bands, we utilize the 13.75\arcsec\ radius aperture measurements for $12~\micron$ and then apply an aperture correction of 1.20. This correction term was found using sources in our sample with no obvious contamination from neighbors and measuring the flux density out to 24.75\arcsec\ to determine their total flux density. The WISE photometry at $22~\micron$ is less accurate than the \textit{Spitzer} $24~\micron$, owing to it having two orders of magnitude lower sensitivity \citep{dole04,wright10}, and is not used for our analysis. In addition, the $22~\micron$ observations suffer from a effective wavelength error, which systematically brightens the photometry of star forming galaxies \citep{wright10,brown14}.
\section{Analysis}
\subsection{Anchoring the IRS Spectra to Global Photometry} \label{IRS_anchor}
For this study, we focus on utilizing the \textit{Spitzer} $5.8-24~\micron$ and WISE $12~\micron$ photometry to anchor the \textit{Spitzer} IRS $5-40~\micron$ spectroscopy available for the entire sample. The reasoning for this approach is to have spectroscopy that is representative of the global flux density of each galaxy, which is required to create a calibrated continuous, monochromatic SFR($\lambda$) indicator. Offsets between the IRS spectrum and the photometry can occur from differences in data reduction methods, or from the width of the IRS slit being smaller than the size of the galaxy. The former effect results in a uniform offset across the entire spectra and can be corrected with a normalization factor. The behavior of the latter effect will be dependent on whether the galaxy observed is an unresolved point-like source. The default \textit{Spitzer} IRS custom extraction (SPICE) does include a correction for light lost from the slit due to the changing angular resolution as a function of wavelength but assumes the object to be a point source. To correct for both of these effects, we utilize photometry from the \textit{Spitzer} 8, 16, and $24~\micron$ and WISE $12~\micron$ bands as a reference. The end-of-channel transmission drop of the SL module below $\sim$5.8$~\micron$, combined with a typical redshift of $z\sim0.1$, makes the $5.8~\micron$ band region unreliable for use in most cases, and so it is not used as an anchor. However, we do make use of the $5.8~\micron$ band to inspect our photometric matching in the lowest redshift galaxies (see below). Here we outline our approach to correct for offset effects so that these spectra are well representative of global photometric measurements.
\citet{odowd11} found that IRS measurements of SSGSS galaxies from the LL module did not show evidence for significant aperture loss when compared to the \textit{Spitzer} 16 and $24~\micron$ photometry. This is attributed to the fact that the lower resolution (larger PSF) of sources in this longer wavelength module, coupled with the larger slit width of $10.6$\arcsec\, allows for the standard SPICE algorithm to accurately recover the total flux density, since objects are close to point sources \citep[see][]{odowd11}. This would suggest that any offsets between the photometry and spectroscopy beyond $16~\micron$ should be uniform across the module (i.e., a global loss in flux density). Therefore, each spectrum is first fit to match the $16~\micron$ and $24~\micron$ photometric points using a constant offset, $c_{\rm{phot}}$, found using chi-squared minimization,
\begin{equation}
c_{\mathrm{phot}}=\frac{\sum_i (S_{\mathrm{phot},i}\,S_{\mathrm{IRS},i})/\sigma(S_{\mathrm{phot},i})^2}{\sum_i (S_{\mathrm{phot},i}/\sigma(S_{\mathrm{phot},i}))^2}
\label{cphot}
\end{equation}
\begin{equation}
\chi^2=\sum_i \left(\frac{S_{\mathrm{phot},i}-c_{\mathrm{phot}} S_{\mathrm{IRS},i}}{\sigma(S_{\mathrm{phot},i})}\right)^2\,,
\label{chisq}
\end{equation}
where $S_{\mathrm{phot},i}$ is the \textit{Spitzer} photometric flux density of band $i$, $\sigma(S_{\mathrm{phot},i})$ is the uncertainty of the \textit{Spitzer} photometric flux density, and $S_{\mathrm{IRS},i}$ is the effective IRS photometric flux density found using the transmission curve for each band, $T_i(\lambda)$,
\begin{equation}
S_{\mathrm{IRS},i}= \frac{\int S_{\rm{IRS}}(\lambda)T_i(\lambda)\,d\lambda}{\int T_i(\lambda)\,d\lambda}\,.
\label{convolve}
\end{equation}
This method ignores the method of calibration that was utilized for each specific bandpass (i.e., a conversion of number of electrons measured by the detector into a flux density in Jy requires knowing the shape of the incoming flux of an object, which varies as a function of wavelength). However, discrepancies between the adopted method and correcting for calibration effects amount to $\sim$$1\%$, and is not important for this study. The offset required to match photometric values is typically small, with values of $c_{\rm{phot}}$ being between $0.7-1.6$.
In contrast to the LL module, \citet{odowd11} found that IRS measurements short-ward of $16~\micron$ from the SL module did show evidence for aperture loss when compared to the \textit{Spitzer} $8~\micron$ photometry. In this case, the increasing resolution of the SL module at shorter wavelengths results in many of the galaxies in this sample being resolved in this module. Also taking into account that the SL slit is 3.6\arcsec, which is smaller than the average extent of $\sim$$10$\arcsec\ ($r$-band Petrosian diameter) for SSGSS galaxies, implies that flux density loss in this wavelength region is more pronounced for more extended objects. For this reason, an additional correction term must be introduced below $16~\micron$ which has a $1/\lambda$ dependency to reflect the additional losses as resolution increases at shorter wavelengths (i.e., the PSF is decreasing at shorter wavelengths, resulting in less correction of light outside the slit). The correction terms adopted are summarized in the following equations,
\begin{equation}
S_{\rm{IRS,corr}}(\lambda) = \left\{
\begin{array}{lr}
S_{\rm{IRS}}(\lambda)/c_{\rm{phot}}\times \left( \frac{16+k}{\lambda+k} \right) & : \lambda <16~\mu m \\
S_{\rm{IRS}}(\lambda)/c_{\rm{phot}} & : \lambda \ge 16~\mu m
\end{array}
\right.
\end{equation}
where $k$ is a constant found by performing a Levenberg-Marquardt least-squares fit of this function, using the IDL code \texttt{MPFITFUN} \citep{markwardt09}, such that the IRS spectrum matches the 8 and $12~\micron$ photometric flux density. Smaller values of $k$ correspond to larger correction factors in the spectrum. Examples of normalizing the IRS spectroscopy to the photometry are shown in Figure~\ref{fig:SSGSS_spectra}. A list of the normalization parameters is shown in Table~\ref{Tab:IRS_correct}.
As a consistency check on the $1/\lambda$ dependency, a more accurate check is made on the few cases where spectra have high S/N and low redshifts, such that the $5.8~\micron$ band region of the spectrum is reliable to use for convolution. In nearly all these cases the correction using a $1/\lambda$ dependency matches the observed photometric point, as demonstrated by SSGSS 46 in Figure~\ref{fig:SSGSS_spectra}. In addition, the agreement of the \textit{Spitzer} $8~\micron$ and WISE $12~\micron$ data with this approach suggests these normalized spectra are well representative of photometric values.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{SSGSS_1_spectra.eps} &
\includegraphics[width=3.5in]{SSGSS_14_spectra.eps} \\
\includegraphics[width=3.5in]{SSGSS_46_spectra.eps} &
\includegraphics[width=3.5in]{SSGSS_59_spectra.eps}
\end{array}$
\end{center}
\caption{IRS spectroscopy (black line), \textit{Spitzer} photometry (green squares), and WISE photometry (cyan circle) for some SSGSS galaxies. The IRS spectrum is normalized to the \textit{Spitzer} $8~\micron$, $16~\micron$, and $24~\micron$ and WISE $12~\micron$ photometric flux densities according to the method described in \S~\ref{IRS_anchor} (red line). The effective IRS photometry, found using the transmission curve for each bandpass filter, are shown as triangles. The normalized transmission curves for the \textit{Spitzer} and WISE bands in this region are shown as green and cyan lines, respectively. \label{fig:SSGSS_spectra}}
\end{figure*}
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{DL07_FIT_SSGSS_1.eps} &
\includegraphics[width=3.5in]{DL07_FIT_SSGSS_14.eps} \\
\includegraphics[width=3.5in]{DL07_FIT_SSGSS_46.eps} &
\includegraphics[width=3.5in]{DL07_FIT_SSGSS_59.eps}
\end{array}$
\end{center}
\caption{A fit to the spectra of some SSGSS galaxies using the dust models of \citet{draine&li07}, shown as the solid red line. The IRAC 3.6, 4.5, and $5.8~\micron$ along with the MIPS 70 and $160~\micron$ photometry is used to fit the dust continuum in the absence of the IRS spectrum. The bands within the IRS region are shown for comparison and not used directly for the fit. The regions of the IRS spectrum associated with transmission drops in the instrument are not shown for clarity. We attribute the relatively poor match in the $6-20~\micron$ region to using a simple three component model. However, the purpose of these fits is only to determine the shape of emission in the $34-70~\micron$ region, for which the data is found to be in good agreement with the models. \label{fig:DLFIT_SSGSS}}
\end{figure*}
\subsection{Extending out to 70~\texorpdfstring{$\micron$}{um}} \label{DL07_method}
The IRS spectrum extends out to $40~\micron$, however, the combination of end-of-channel transmission drop around $\lambda_{\rm{obs}}\sim37~\micron$ and redshift effects makes these spectra unreliable for $\lambda_{\rm{rest}}\gtrsim37/(1+z)~\micron$. For the mean redshift of this sample ($z\sim0.1)$, this corresponds to roughly $\lambda_{\rm{rest}}\sim34~\micron$. In order to utilize the wavelength region between $34$ and $70~\micron$, which contains poor quality or no spectral data, we interpolate with dust models. This interpolation is expected to be robust since there are no sharp emission features in this wavelength range. The shape of the emission in this region is dependent on the temperature distribution of of the dust, the grain size distribution, as well as the relative importance of stochastic versus thermal equilibrium heating (the former gives an almost-constant continuum and the latter is responsible to the Wien-like rise of the spectrum).
To extend the wavelength region of our study out to $70~\micron$, we fit the dust models of \citet{draine&li07}, combined with an additional stellar continuum component, to our IR photometry and IRS spectroscopy. For these models, the emission spectrum is given by \citet{draine07} as,
\begin{multline}
S_{\nu,\rm{model}} = \Omega_{*}B_\nu(T_*) + \frac{M_{\rm{dust}}}{4\pi D_{\rm{lum}}^{2}}[(1-\gamma)p_\nu^{(0)}(j_M,U_{\rm{min}}) \\ +\gamma p_\nu(j_M,U_{\rm{min}},U_{\rm{max}},\alpha),
\end{multline}
where $\Omega_{*}$ is the solid angle subtended by stars, $T_*$ is the effective temperature of the stellar contribution, $M_{\rm{dust}}$ is the total dust mass, $D_{\rm{lum}}$ is the distance to the galaxy, $p_\nu$ is the specific power per unit dust mass, $U_{\rm{min}}$ ($U_{\rm{max}}$) is the minimum (maximum) interstellar radiation, $\gamma$ is the fraction of the dust mass exposed to radiation with intensity $U>U_{\rm{min}}$, $j_M$ corresponds to the dust model \citep[i.e., the PAH abundance relative to dust, $q_{\rm{PAH}}$; shown in table 3 of][]{draine&li07}, and $\alpha$ is the power-law factor for the starlight intensity. In summary, this emission spectrum is a linear combination of three components: (1) a stellar continuum with effective temperature $T_*$ which dominates at $\lambda\lesssim5~\micron$; (2) a diffuse ISM component with an intensity factor $U=U_{\rm{min}}$; and (3) a component arising from photo-dissociation regions (PDRs). Typically, component (2) comprises a much larger amount of the total dust mass and, as such, is dominant over component (3) in the emission spectrum \citep{draine07, draine14}.
To fit this model we follow the approach outlined in \citet{draine07}, which found that the SEDs of galaxies in the \textit{Spitzer} Infrared Nearby Galaxies Survey (SINGS) were well reproduced with fixed values of $\alpha=2$, $U_{\rm{max}}=10^6$, and $T_*=5000$~K. Holding these parameters fixed, $q_{\rm{PAH}}$, $U_{\rm{min}}$, $\gamma$, $M_{\rm{dust}}$, and $\Omega_{*}$ are varied to find the dust model that comes closest to reproducing the photometry and spectroscopy. For this work, a grid of $\gamma$ values is constructed for all $q_{\rm{PAH}}$ (MW, LMC, SMC) and $U_{\rm{min}}$ values. The value of $M_{\rm{dust}}$ for each grid point is determined by minimizing the $\chi^2$ parameter in a similar manner to eq. (\ref{cphot}) and (\ref{chisq}), as $M_{\rm{dust}}$ represents a constant offset value. The goodness-of-fit for each case is assessed using the $\chi^2$ parameter,
\begin{equation}
\chi^2\equiv\sum_i\frac{S_{\mathrm{obs},i}-S_{\mathrm{model},i}}{\sigma_{\mathrm{obs},i}^2+\sigma_{\mathrm{model},i}^2}\,,
\end{equation}
where the sum is over observed bands and spectroscopic channels, $S_{\mathrm{model},i}$ is the model spectrum (for band comparison, the model spectrum is convolved with the response function of that band), $\sigma_{\mathrm{obs},i}$ is the observational uncertainty in the observed flux density $S_{\mathrm{obs},i}$, and $\sigma_{\mathrm{model},i}=0.1S_{\mathrm{model},i}$ as adopted by \citet{draine07}. The observed flux densities used in determining the best fit is comprised of the IRS spectroscopy in addition to the IRAC 3.6, 4.5, and $5.8~\micron$ and MIPS 70, and $160~\micron$ photometry. The model which minimizes the value of $\chi^2$ is adopted for use in representing the region $\lambda_{\rm{rest}}\gtrsim37/(1+z)~\micron$. Examples of the best fitting model for SSGSS galaxies are shown in Figure~\ref{fig:DLFIT_SSGSS}. The bands within the IRS region are shown only for comparison and are not used directly for the fit.
Our choice to adopt the model which minimizes the value of $\chi^2$ is not necessarily the most accurate representation of the spectra, as the degeneracy of the model parameters can allow for multiple fits to have similar $\chi^2$ values while having different FIR SED shapes. However, we do not consider this to be of great significance for this study for two reasons. First, the flux density variation due to changes in the SED shape for cases with $(\bar\chi^2-\bar\chi^2_{\rm{min}})<1$, where $\bar\chi^2_{\rm{min}}$ is the minimum value of the reduced $\chi^2$, is typically less than $25\%$ over the $30-70~\micron$ region, which is lower than the scatter among individual galaxy templates. As we will be utilizing an average of our galaxy templates for our diagnostic, the uncertainty from model SED variations will not be the dominant source of uncertainty. Second, the parameters of these fits are not used to determine the properties of these galaxies, which are more sensitive to these degeneracy effects than the total flux density.
All cases are best fit by Milky Way dust models with $q_{\rm{PAH}}\geq2.50\%$. There is a systematic trend among most fits to underestimate the flux density around the $8~\micron$ PAH feature and overestimate the $10-20~\micron$ region. This is most likely due to the limitations of fitting only three components to the data. However, since the main focus of these fits is to provide a description of the region from $\sim$$34-70~\micron$, these deviations are not considered to be significant, as they should have little effect on matching the shape of the emission beyond $\sim$$34~\micron$. As will be discussed in \S~\ref{template_variation}, these fits are consistent with other $z=0$ SFG templates found in the literature that make use of \citet{draine&li07} models, suggesting that these deviations could be a common problem. Investigation into the cause of these discrepancies warrants additional study, as it will improve our understanding of dust properties in galaxies. We do not make use of these fits to determine $L_{\rm{IR}}$\ values, instead using the values provided in the SSGSS catalog, which have been extensively checked \citep{treyer10} across a variety of diagnostic methods.
\subsection{Determining Rest-Frame Luminosities}
As these galaxies span a redshift range of $0.03\le z\le 0.22$, the IRS spectrum and photometry of each galaxy span slightly different regions in rest-frame wavelength. This offset causes observed photometric values to vary by up to 30\% from the rest-frame values. This would affect our determination of SFR if not accounted for and introduce additional scatter. Since previous MIR calibrations have been performed for local samples of galaxies ($z\sim0$) to accuracies around 30\% \citep{rieke09,kennicutt09,calzetti10,hao11}, this is a non-negligible effect.
To correct for redshift effects, photometric values for each band are determined by convolving the spectrum at the rest-frame filter postions for each band according to eq (\ref{convolve}), only now using the corrected spectrum, $S_{\rm{IRS,corr}}(\lambda)$, instead of the original IRS spectrum. This is performed for the \textit{Spitzer} 8, 24, and $70~\micron$ and WISE 12 and $22~\micron$ bands. These corrected flux densities are used to calculate the rest-frame luminosity (erg~s$^{-1}$) of each band,
\begin{equation} \label{Lrest}
L_{\mathrm{rest}}= (\nu L_{\nu})_{\mathrm{rest}}=(\nu S_{\mathrm{IRS,corr}})_{\mathrm{obs}}\,4\pi D_{\mathrm{lum}}^2 \,,
\end{equation}
where $\nu_{\mathrm{rest}}$ and $\nu_{\mathrm{obs}}$ are the effective rest-frame and observer-frame frequency of each band, respectively, and $D_{\mathrm{lum}}$ is the luminosity distance for the galaxy, calculated from its redshift. These rest-frame luminosities are used to determine SFRs for each of the galaxies in our sample. In a similar manner, each IRS spectrum is expressed as a wavelength dependent rest-frame luminosity, $L(\lambda)_{\mathrm{rest}}$, using the continuous spectrum, $S_{\mathrm{IRS,corr}}(\lambda)$. This is used later for calibrating our wavelength continuous SFR-luminosity conversion factors, $C(\lambda)$.
To correct the \textit{Spitzer} 3.6 and $4.5~\micron$ bands, which lie outside of the IRS spectral coverage, a correction is applied assuming these bands encompass the Rayleigh-Jeans tail of the stellar continuum emission,
\begin{equation}
S'_{\mathrm{obs}}= S_{\mathrm{obs}}\left(\frac{\lambda_{\mathrm{rest}}}{\lambda_{\mathrm{obs}}}\right)^{-2}=S_{\mathrm{obs}}\times(1+z)^{-2}\,.
\end{equation}
The luminosity is then found following eq (\ref{Lrest}) using $S'_{\mathrm{obs}}$. The rest-frame \textit{Spitzer} 3.6 and $4.5~\micron$ luminosities will only be used to examine the origins of scatter within our conversion factors in \S~\ref{scatter}, and has no influence on our estimates for the MIR conversion factors. The fitting results in \S~\ref{DL07_method} suggest that this simple approach is reasonable for our sample of SFGs.
\begin{table*}
\begin{center}
\caption{Reference Star Formation Rate Calibrations \label{Tab:calibrations}}
\begin{tabular}{ccccc}
\hline\hline
Band(s) & $L_x$ Range & $SFR$ & $\log C_x$ & Reference \\
& (erg s$^{-1}$)& & & \\ \hline
FUV+TIR & \nodata & $[L(\mathrm{FUV})_{\mathrm{obs}}+0.46\,L(\mathrm{TIR})]/C_x$ & 43.35 & \citet{hao11} \\
FUV+$24~\micron$ & \nodata & $[L(\mathrm{FUV})_{\mathrm{obs}}+3.89\,L(24)]/C_x$ & 43.35 & \citet{hao11} \\
NUV+TIR & \nodata & $[L(\mathrm{NUV})_{\mathrm{obs}}+0.27\,L(\mathrm{TIR})]/C_x$ & 43.17 & \citet{hao11} \\
NUV+$24~\micron$ & \nodata & $[L(\mathrm{NUV})_{\mathrm{obs}}+2.26\,L(24)]/C_x$ & 43.17 & \citet{hao11} \\
H$\alpha$+$8~\micron$ & \nodata & $[L(\mathrm{H}\alpha)_{\mathrm{obs}}+0.011\,L(8)]/C_x$ & 41.27 & \citet{kennicutt09,hao11} \\
H$\alpha$+$24~\micron$ & \nodata & $[L(\mathrm{H}\alpha)_{\mathrm{obs}}+0.020\,L(24)]/C_x$ & 41.27 & \citet{kennicutt09,hao11} \\
H$\alpha$+$24~\micron$ & $L(24)< 4\times10^{42}$ & $[L(\mathrm{H}\alpha)_{\mathrm{obs}}+0.020\,L(24)]/C_x$ & 41.26 & \citet{calzetti10} \\
& $4\times10^{42}\le L(24)< 5\times10^{43}$ & $[L(\mathrm{H}\alpha)_{\mathrm{obs}}+0.031\,L(24)]/C_x$ & 41.26 & \citet{calzetti10} \\
& $ L(24)\ge 5\times10^{43}$ & $L(24)\times[2.03\times10^{-44}\,L(24)]^{0.048}/C_x$ & 42.77 & \citet{calzetti10} \\
H$\alpha$+TIR & \nodata & $[L(\mathrm{H}\alpha)_{\mathrm{obs}}+0.0024\,L(\mathrm{TIR})]/C_x$ & 41.27 & \citet{kennicutt09,hao11} \\
$24~\micron$ & $2.3\times10^{42}\le L(24)\le 5\times10^{43}$ & $L(24)/C_x$ & 42.69 & \citet{rieke09} \\
& $L(24)>5\times10^{43}$ & $L(24)\times(2.03\times10^{-44}L(24))^{0.048}/C_x$ & 42.69 & \citet{rieke09} \\
$70~\micron$ & $L(70)\gtrsim 1.4\times10^{42}$ & $L(70)/C_x$ & 43.23 & \citet{calzetti10} \\ \hline
\end{tabular}
\end{center}
\textbf{Notes.} Columns list the (1) bands used in the calibration, (2) luminosity range over which the calibration can be used; empty fields denote an unspecified range, (3) $SFR$ conversion formula, (4) conversion constant, and (5) reference for calibration.
\end{table*}
\subsection{Reference Monochromatic SFR Indicators}
In order to perform any calibration of luminosity as a SFR indicator, it is necessary to rely on previous, well-calibrated SFR indicators. In this work, we utilize the calibrations of \citet{kennicutt09}, \citet{rieke09}, \citet{calzetti10}, and \citet{hao11}, which incorporate the full suite of data available for this sample. This list is shown in Table~\ref{Tab:calibrations}. The reference SFR, $\langle SFR \rangle$, for each galaxy is taken to be the average of the SFRs from these calibrations. By utilizing the average of a large number of diagnostics, we limit the risk of potential biases that any single diagnostic can be subject to. Several of these reference diagnostics make use of the total infrared luminosity, $L_{\rm{TIR}}$, which refers to the integrated luminosity over the region from 3 to $1100~\micron$. All measurements of $L_{\rm{TIR}}$ for these galaxies have been obtained from the original SSGSS dataset \citep{treyer10}. For the SSGSS sample, \citet{treyer10} find that $L_{\rm{TIR}}$ is larger than $L_{\rm{IR}}$\ by $\sim$$0.04$~dex.
To utilize SDSS measurements of H$\alpha$ for a SFR estimation, it is necessary to apply an aperture correction. The diameter of the SDSS spectroscopic fiber spans 3\arcsec, which is a factor of $\sim$$3$ smaller than the typical size of the SSGSS galaxies and results in only a fraction of the light being measured. We correct for this aperture effect using the prescription from \citet{hopkins03}, which uses the difference between the $r$-band Petrosian magnitude and the $r$-band fiber magnitude \citep[see also][]{treyer10}. The values of these corrections range from $1.9-8.3$ for our sample, with the exception of SSGSS~67 with a correction of $21.4$ due to its much larger size.
As a consistency check, the SFRs inferred from each indicator for our galaxies is compared in Figure~\ref{fig:SFR_compare}. It is seen from the distribution that the majority of these values agree within the $\sim$30\% uncertainty associated with individual calibrations \citep{rieke09,kennicutt09,calzetti10,hao11}. A formal fit of this distribution to a Gaussian profile gives values of $\mu=-0.02$ and $\sigma=0.17$. Our choice in using the average of the diagnostics, $\langle SFR \rangle$, for each galaxy instead of the median value appears to cause no significant differences, with typical offsets of only a few percent between the two, which are symmetric. A formal Gaussian fit of fractional difference between the mean and median gives $\mu=-0.01$ and $\sigma=0.04$. We illustrate the relative offsets for the individual calibrations in Figure~\ref{fig:SFR_compare_hist}. We note that the MPA/JHU group provides independent estimates for SFRs based on the technique discussed in \citet{brinchmann04} using extrapolated H$\alpha$ measurements. However, we find that the spread in values for these estimates relative to $\langle SFR \rangle$ are significantly larger than our other diagnostics ($1\sigma=0.48$), which we attribute to their larger SFR uncertainties ($\sim$$50\%$), and as such were excluded from our analysis (including Figures~\ref{fig:SFR_compare} and \ref{fig:SFR_compare_hist}).
\begin{figure*}
\plotone{compare_SFR_calibrations.eps}
\caption{{\it Left:} Comparison of SFRs determined from the calibrations listed in Table~\ref{Tab:calibrations} for the SFGs in the SSGSS. Each vertical strip of values shows the SFR values for each method on a single galaxy. The reference SFR, $\langle SFR \rangle$, for these galaxies is taken to be the average of the SFR values from these calibrations. {\it Right:} Histogram showing the distribution of SFR offsets relative to the reference value. This distribution is well fit by a Gaussian with $\mu=-0.02$ and $\sigma=0.17$, suggesting that the majority of these values agree within the uncertainties associated with the individual calibrations. \label{fig:SFR_compare}}
\end{figure*}
\begin{figure*}
\plotone{compare_SFR_calibrations_hist.eps}
\caption{Histograms showing the distribution of the individual SFR estimates relative to the reference value, $\langle SFR \rangle$. The calibrations being considered are listed in Table~\ref{Tab:calibrations}. The median offset and $1\sigma$ dispersion are shown in each panel. \label{fig:SFR_compare_hist}}
\end{figure*}
\subsection{\texorpdfstring{$L_{\rm{IR}}$}{LIR} as a SFR Indicator} \label{lir}
A commonly utilized method to determine SFRs for galaxies relies on measuring the integrated luminosity over most of the IR wavelength range, $L_{\rm{IR}}$ ($8-1000~\micron$). However, physically understanding the conversion of $L_{\rm{IR}}$\ to a SFR is non-trivial and sensitive to many assumptions, such as the timescale of star formation, $\tau$, the star formation history (SFH), the metallicity, and the initial mass function \citep[IMF; see][]{murphy11b,calzetti13}. For example, a galaxy with a constant SFH, a fixed metallicity, and a fixed IMF will have the calibration constant for $L_{\rm{IR}}$\ change by a factor of 1.75 between assuming $\tau=100$~Myr and $\tau=10$~Gyr \citep{calzetti13}.
We chose to avoid the use of SFRs based solely on $L_{\rm{IR}}$\ for reference because of the sensitivity to these assumptions. However, in order to compare the accuracy of our calibration on higher redshift samples (in \S~\ref{Cx_test}) for which $L_{\rm{IR}}$\ is the only technique available to estimate SFRs, we use a SFR-$L_{\rm{IR}}$\ conversion which reproduces the values of $\langle SFR \rangle$ seen for the SSGSS sample. This occurs for a conversion factor of $\log[C(L_{\mathrm{IR}})]=43.64$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$). Utilizing Starburst99 \citep{leitherer99}, with a constant SFH, solar metallicity, a Kroupa IMF over $0.1-100~M_\odot$, and assuming all of the stellar light (UV+visible) is reradiated by dust, this corresponds to a timescale of $\tau\sim500$~Myr \citep[e.g.,][]{calzetti13}.
This adopted conversion factor differs slightly from other commonly adopted values. In the case of \citet{murphy11b}, $\log[C(L_{\mathrm{IR}})]=43.41$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$), our calibration is larger by 70\%. This large difference is due to two reasons: (1) \citet{murphy11b} assume that only UV light is being reradiated by the dust and does not account for the optical light that would also be reradited ($\sim$40\% of the discrepancy), and (2) they assume a 100~Myr constant star-forming population ($\sim$20\% of the discrepancy). In the case of \citet{kennicutt98} after converting from a \citet{salpeter55} IMF to a \citet{kroupa01} IMF, $\log[C(L_{\mathrm{IR}})]=43.53$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$), our calibration is larger by 30\%. Most of this difference is due to them assuming a $100$~Myr constant star-forming population.
\section{A Calibrated Continuous, Monochromatic SFR(\texorpdfstring{$\lambda$}{lambda})}
\subsection{Composite IRS Spectrum}
The SFR of a galaxy, using a calibrated single-band luminosity, can be written as
\begin{equation}
\mathrm{SFR} (M_\odot \mathrm{yr}^{-1})= L_x/C_x\,,
\label{sfr}
\end{equation}
where $L_x$ is the monochromatic luminosity, in units of erg~s$^{-1}$, and $C_x$ is the conversion factor between SFR and luminosity for filter $x$ \citep[following convention of][]{kennicutt12}. In this respect, the appropriate conversion factor at a given band is found by normalizing the luminosity by the SFR determined independently from a reference calibration.
This same approach is taken to calibrate our continuous wavelength conversion factors,
\begin{equation}
C(\lambda) (\mathrm{erg~s}^{-1}/(M_\odot \mathrm{yr}^{-1}))= L(\lambda)_{\mathrm{rest}}/\langle \mathrm{SFR}\rangle\,,
\end{equation}
where $L(\lambda)_{\mathrm{rest}}$ is the wavelength dependent IRS luminosity and $\langle SFR \rangle$ is the reference SFR. To achieve our calibration of $C(\lambda)$, the SFR-normalized IRS spectra are averaged together to create a composite spectrum for the group. As a result of shifting the spectra to the rest-frame, the wavelengths associated with each spectral channel no longer match exactly. Therefore, to perform this average, the channel wavelengths in the spectrum of the first galaxy in our group is taken to be the reference grid. Next, the normalized luminosity values of the other galaxies are re-gridded to this (i.e., each channel is associated to the nearest neighboring reference channel). In using this approach, the smoothing of sharp features that result from direct interpolation is avoided. The uncertainties associated with this re-griding to determine a composite spectrum are small relative to the channel flux density uncertainly. Furthermore, these uncertainties are much smaller than the scatter between spectra, which drives the uncertainty of our template, and can be considered negligible for the purposes of this study.
The result of an average for the entire sample of SFGs in the SSGSS sample is shown in Figure~\ref{fig:group_all}. The uncertainty of each channel in the composite spectrum is taken to be the standard deviation of the the group value for that channel. The sample standard deviation is the dominant source of uncertainty (typically between $20-30\%$ of the normalized luminosity value) and is larger than the flux density uncertainties of individual IRS channels (typically $\sim$2\%) by roughly an order of magnitude. This template can be used to determine the appropriate conversion factor for any luminosity within our wavelength coverage.
\subsection{Filter Smoothed Composite Spectrum} \label{filter_smooth}
In practice, observations of a galaxy are made using specific bandpass filters that encompass a portion of their SED. Therefore, it is more practical to utilize a composite spectrum that corresponds to photometric luminosities observed by various bands as functions of redshift. To accomplish this, the normalized IRS spectrum of each SFG in our sample is convolved with the filter response of specific bands as functions of redshift (i.e., the effective wavelength blue-shifts and the bandpass narrows, both by a factor of ($1+z$), as one goes to higher redshifts). Performing this convolution is similar to smoothing by the bandpass filter, only with the filter width changing with redshift. Throughout the rest of this paper, the term ``smoothed'' is used interchangeably to mean this convolution process. The redshift limit imposed for each band occurs at the shortest usable rest-frame wavelengths of the IRS spectrum for that band. The composite IRS template and the filter smoothed composites presented in this section are publicly available for download from the IRSA\footnote{\url{http://irsa.ipac.caltech.edu/data/SPITZER/MIR\_SFR}}.
Each normalized IRS spectra is smoothed using the \textit{Spitzer}\footnote{\url{http://irsa.ipac.caltech.edu/data/SPITZER/docs/irac/}}$^,$\footnote{\url{http://irsa.ipac.caltech.edu/data/SPITZER/docs/mips/}}, WISE\footnote{\url{http://www.astro.ucla.edu/~wright/WISE/passbands.html}}, and JWST/MIRI filters. The properties of these filters is listed in Table~\ref{Tab:filter_properties}. We note that the \textit{Herschel} PACS $70~\micron$ is close enough to \textit{Spitzer} $70~\micron$ that these can be interchanged for use with $C_{S70}(\lambda)$. We emphasize to the reader that care should be taken when considering the $22~\micron$ band, as it has been shown to suffer from an effective wavelength error \citep[see][]{wright10,brown14}. For the MIRI filters, we use the response functions of \citet{glasse15}. Since these curves do not take into account the wavelength dependent quantum efficiency, the instrument transmission, or the responsivity of the detector, they should be updated once better curves become available.
The composite spectra for each of these bands is created by averaging the smoothed spectra together, in the same manner as for the group average. The result of averaging the convolved spectra is shown in Figures~\ref{fig:band_calibration_all_SFG} \& \ref{fig:JWST_calibration_all_SFG}. The associated uncertainty with each channel (sample standard deviation) is slightly lower than the native composite spectrum owing to the smoothing from the convolution and is typically between $15-20\%$ of the normalized luminosity value (except for $70~\micron$ case, which is still around 30\%), making them comparable to accuracies achieved in many previous calibrations. Previously determined MIR conversion factors (from $z\sim0$ samples) are also shown and appear in good agreement.
For the filter bands considered here, the smoothed IRS spectra show a very large increase in scatter below $\sim$$6~\micron$, which is due to a combination of the end-of-channel uncertainties being very high and also from variations in the old stellar populations of these galaxies. For these reasons, we only consider the regions for which the $1\sigma$ uncertainty is less than $30\%$ suitable for calibration. In the case of the WISE $12~\micron$ band, this region occurs below $\sim$$7~\micron$ because of the significantly wider filter bandwidth. The ranges chosen for the calibration of each band is shown in Table~\ref{Tab:calibration_parameters}.
\begin{table}
\begin{center}
\caption{Filter Properties \label{Tab:filter_properties}}
\begin{tabular}{cccc}
\hline\hline
Instrument & Band & $\lambda_{\rm{eff,0}}$ & FWHM \\
& & ($\micron$) & ($\micron$) \\ \hline
IRAC & $8~\micron$ & 7.87 & 2.8 \\
MIPS & $24~\micron$ & 23.68 & 5.3 \\
MIPS & $70~\micron$ & 71.42 & 19.0 \\
WISE & $12~\micron$ & 12.08 & 8.7 \\
WISE & $22~\micron$ & 22.19$^a$ & 3.5 \\
MIRI & F1000W & 10.00 & 2.0 \\
MIRI & F1280W & 12.80 & 2.4 \\
MIRI & F1500W & 15.00 & 3.0 \\
MIRI & F1800W & 18.00 & 3.0 \\
MIRI & F2100W & 21.00 & 5.0 \\
MIRI & F2550W & 25.50 & 4.0 \\ \hline
\end{tabular}
\end{center}
\textbf{Notes.} Columns list the (1) instrument, (2) passband name, (3) rest-frame effective wavelength, and (4) full width at half maximum. $^a$The $22~\micron$ observations suffer from a effective wavelength error \citep[see][]{wright10,brown14}.
\end{table}
\begin{figure*}
\plotone{all_SFG_linterp_spectra.eps}
\caption{Normalized IRS luminosity, $L(\lambda)_{\mathrm{rest}}/\langle SFR \rangle$, for all SFG galaxies (gray solid lines). The composite spectrum of this group is shown (thick black line) along with the standard deviation from this average (black dotted lines), which at most wavelengths is between $20-30\%$. The fits to the dust continuum for each galaxy (gray dashed lines; described in \S~\ref{DL07_method}) along with the average (black dashed line) are also shown. The low dispersion among normalized spectra suggests that the $6-70~\mu$m region can be utilized for SFR diagnostics. \label{fig:group_all}}
\end{figure*}
\subsection{Fits to the Composite Spectra}
To simplify the application of our results as SFR indicators, each of the filter smoothed composite spectra is fit using a continuous function, $fit_{x}(\lambda)$. We perform Levenberg-Marquardt least-squares fits of a polynomial (up to 1st order) and Drude profiles (up to 5), $I_r(\lambda)$, to the smoothed composite spectra, using the IDL code \texttt{MPFITFUN},
\begin{equation}
fit_{x}(\lambda) = \sum_{i=1}^2 p_i\lambda^{(i-1)} + \sum_{r=1}^{5} I_r(\lambda) \,,
\end{equation}
where $x$ corresponds to the filter being considered, $p_i$ are constants, and $I_r(\lambda)$ are Drude profiles. Drude profiles, which are typically employed to characterize dust features, have the form
\begin{equation}
I_r(\lambda) = \frac{b_r\gamma_r^2}{(\lambda/\lambda_r-\lambda_r/\lambda)^2+\gamma_r^2} \,,
\end{equation}
where $\lambda_r$ is the central wavelength of the feature, $\gamma_r$ is the fractional FWHM, and $b_r$ is the central intensity, which is required to be non-negative. We emphasize that because these are smoothed spectra, the parameters of these fits are not of physical significance and are simply being employed for ease of application.
For the Drude profiles, the central wavelengths, $\lambda_r$, are fixed to wavelengths that roughly correspond to the peaks in the smoothed spectrum, while $\gamma_r$ and $b_r$ are left as free parameters. Therefore, there are up to 12 free parameters in total, two from the polynomial and ten from the Drude profiles. The values of $\lambda_r$ for each smoothed composite fit and all the other fit parameters are listed in Table~\ref{Tab:calibration_parameters}. The fits are shown for the individual bands in Figure~\ref{fig:band_calibration_all_SFG}. The fitting functions are typically accurate to within $\pm5\%$ (0.02 dex) of the true values and can be used in place of the templates [$C_x(\lambda) = fit_x(\lambda)$].
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.0in]{SSGSS_S7pt9_calibration.eps} &
\includegraphics[width=3.0in]{SSGSS_S24_calibration.eps} \\
\includegraphics[width=3.0in]{SSGSS_S70_calibration.eps} &
\includegraphics[width=3.0in]{SSGSS_W12_calibration.eps} \\
\multicolumn{2}{c}{\includegraphics[width=3.0in]{SSGSS_W22_calibration.eps}}
\end{array}$
\end{center}
\caption{The top of each panel shows the conversion factor for a \textit{Spitzer} or WISE band using all SFG galaxies (solid red lines), along with their uncertainty (dotted red lines), which for most cases is between $15-20\%$. The solid black line is a fit to the smoothed spectrum, $fit_{x}(\lambda)$. Local conversion factors from the literature are also shown for comparison (colored symbols). The region below $\sim$$6~\micron$ is excluded due to significantly increased uncertainty in the composite spectrum (see \S~\ref{filter_smooth}). The bottom of each panel shows the residuals between the conversion factor and a fit to the curve ($\log[C_{x}(\lambda)/fit_{x}(\lambda)]$). \label{fig:band_calibration_all_SFG}}
\end{figure*}
\subsection{Comparison to WISE SFR Calibrations}
The WISE All-Sky Survey \citep{wright10} provided photometry for over 563 million objects, and as such has great potential for future application of our calibrations. Recently, calibrations of the WISE bands as SFR indicators have emerged \citep{donoso12,shi12,jarrett13,lee13,cluver14}, some of which can be easily compared to our results. In particular, we focus on the results of \citet{jarrett13} and \citet{lee13} as these have linear calibrations of the WISE 12 and $22~\micron$ bands and can be directly compared to our values. The difference between the calibration values found in these works is rather large, corresponding to 0.34 dex ($\sim$120\%) and 0.15 dex ($\sim$40\%), for the WISE 12 and $22~\micron$ band, respectively. These large discrepancies are likely the result of the different approaches of the two works. \citet{jarrett13} rely of the previous calibrations of \citet{rieke09} at $24~\micron$, whereas \citet{lee13} attempt to determine SFRs from extinction-corrected H$\alpha$ emission.
The composite of our sample of SFG spectra smoothed by the WISE 12 and $22~\micron$ filters is compared to these calibrations in Figure~\ref{fig:band_calibration_all_SFG}. We find that the results lie in-between the values found by \citet{jarrett13} and \citet{lee13}. Since the WISE $22~\micron$ band is so similar in shape and location to the \textit{Spitzer} $24~\micron$ band, the calibrations of \citet{zhu08} and \citet{rieke09} are also presented and show close agreement to our work.
\begin{sidewaystable}
\begin{center}
\caption{Continuous Star Formation Rate Calibration $fit_x(\lambda)$ Parameters \label{Tab:calibration_parameters}}
\begin{tabular}{cccccccccccccccccccc}
\hline\hline
Band & $\lambda_{\rm{eff}}$ Range & $z$ limit & $p_1$ & $p_2$ & $\lambda_1$ & $b_1$ & $\gamma_1$ & $\lambda_2$ & $b_2$ & $\gamma_2$ & $\lambda_3$ & $b_3$ & $\gamma_3$ & $\lambda_4$ & $b_4$ & $\gamma_4$ & $\lambda_5$ & $b_5$ & $\gamma_5$\\
& ($\micron$) & & & & ($\micron$) & & & ($\micron$) & & & ($\micron$) & & & ($\micron$) & & & ($\micron$) & & \\ \hline
$S8$ & $6.20-7.87$ & $0.27$ & -1.229e42 & \nodata & 7.1 & 1.075e43 & 0.349 & 8.2 & 3.102e42 & 0.224 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
$S24$ & $6.20-23.68$ & $2.82$ & -2.803e42 & \nodata & 7.1 & 1.107e43 & 0.413 & 8.2 & 4.496e42 & 0.165 & 12.0 & 5.300e42 & 0.341 & 17.9 & 4.154e42 & 0.576 & 25.0 & 5.093e42 & 0.366 \\
$S70$ & $17.85-71.42$& $3.00$ & -3.213e42 & 3.125e41 & 17.9 & 1.208e42 & 0.272 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
$S70_{\rm{corr}}$ & $17.85-71.42$& $3.00$ & -1.929e42 & 2.825e41 & 35.0 & 4.065e42 & 0.461 & 47.0 & 2.029e42 & 0.732 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
$W12$ & $7.00-12.08$ & $0.73$ & -1.598e43 & 1.112e42 & 6.3 & 1.528e43 & 0.402 & 9.5 & 9.050e42 & 0.685 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
$W22$ & $6.28-22.19$ & $2.53$ & -1.791e43 & \nodata & 6.8 & 2.253e43 & 0.789 & 8.0 & 6.473e42 & 0.198 & 12.0 & 9.557e42 & 0.392 & 17.9 & 1.428e43 & 0.743 & 25.0 & 1.202e43 & 0.401 \\
$F1000W$ & $6.48-10.00$ & $0.54$ & -2.463e42 & \nodata & 6.7 & 6.078e42 & 0.569 & 7.9 & 1.074e43 & 0.247 & 10.6 & 4.042e42 & 0.065 & \nodata & \nodata & \nodata & \nodata & \nodata & \nodata \\
$F1280W$ & $6.43-12.80$ & $0.99$ & -5.267e43 & 3.409e42 & 6.7 & 3.544e43 & 0.801 & 8.0 & 7.603e42 & 0.224 & 10.6 & 2.468e42 & 0.097 & 11.9 & 6.918e42 & 0.213 & \nodata & \nodata & \nodata \\
$F1500W$ & $6.37-15.00$ & $1.35$ & -2.105e43 & \nodata & 6.7 & 2.212e43 & 0.756 & 7.9 & 1.038e43 & 0.285 & 10.5 & 1.359e42 & 0.069 & 11.9 & 1.174e43 & 0.367 & 17.7 & 1.976e43 & 0.684 \\
$F1800W$ & $6.48-18.00$ & $1.78$ & -1.339e43 & 3.510e41 & 6.4 & 8.470e42 & 0.160 & 7.8 & 2.092e43 & 0.317 & 10.7 & 1.581e42 & 0.082 & 12.0 & 1.008e43 & 0.331 & 17.7 & 8.963e42 & 0.539 \\
$F2100W$ & $6.43-21.00$ & $2.27$ & -1.493e43 & 6.926e41 & 7.0 & 1.531e43 & 0.483 & 8.1 & 7.937e42 & 0.247 & 10.6 & 8.052e41 & 0.048 & 12.0 & 8.089e42 & 0.357 & 17.3 & 4.146e42 & 0.533 \\
$F2550W$ & $6.40-25.50$ & $2.99$ & -6.947e42 & 4.408e41 & 6.5 & 7.568e42 & 0.186 & 7.9 & 1.468e43 & 0.237 & 11.9 & 6.429e42 & 0.273 & 13.5 & 8.379e41 & 0.083 & 17.3 & 2.112e42 & 0.290 \\\hline
\end{tabular}
\end{center}
\textbf{Notes.} The second fit for $S70$ has a correction term that takes into account the FIR variation of SFG SEDs with redshift (see \S~\ref{C70z}). The functional form of these fits is $f_{\rm{IRS}}(\lambda) = \sum_{i=1}^2 p_i\lambda^{(i-1)} + \sum_{r=1}^{5} I_r(\lambda)$, where $I_r(\lambda) = \frac{b_r\gamma_r^2}{(\lambda/\lambda_r-\lambda_r/\lambda)^2+\gamma_r^2}$
\end{sidewaystable}
\section{Application to Higher Redshift Galaxies}
\subsection{Demonstration}
Here we demonstrate how to apply our calibrations to a SFG with a known redshift. Let us consider using the observed $24~\micron$ flux density for the galaxy SSGSS~1 to estimate the SFR of this galaxy. At $z=0.066$ for this particular galaxy, the $24~\micron$ band has an effective wavelength of $\lambda_{\rm{eff}}(24)=22.51~\micron$ and an observed flux density of $S_{obs}(24)=9.72\times10^{-26}$~erg~s$^{-1}$~cm$^{-2}$~Hz$^{-1}$, which corresponds to a luminosity of $\log[L_{\rm{obs}}(24)]=43.11$~erg~s$^{-1}$. Knowing the effective wavelength, we next want to use the composite $24~\micron$ band smoothed spectrum to determine the appropriate conversion factor, which is found to be $\log[C_{24}(22.51~\micron)]=42.62$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$) using the smoothed composite template or the fitting function (see Figure~\ref{fig:band_calibration_all_SFG}). Finally making use of eq. (\ref{sfr}), we get that the SFR is simply the observed luminosity in this band divided by the conversion factor, $\mathrm{SFR}=10^{43.11}/10^{42.62}=3.09~M_\odot \rm{yr}^{-1}$. This value differs from the actual value of $\langle SFR \rangle=4.22~M_\odot \rm{yr}^{-1}$ by about $30\%$.
Next we can consider the slightly more distant case of SSGSS 14, at $z=0.153$, and determine the SFR from its observed $24~\micron$ flux density of $S_{obs}(24)=4.44\times10^{-26}$~erg~s$^{-1}$~cm$^{-2}$~Hz$^{-1}$, which corresponds to a luminosity of $\log[L_{\rm{obs}}(24)]=43.55$~erg~s$^{-1}$. The $24~\micron$ band has an effective wavelength of $\lambda_{\rm{eff}}(24)=20.53~\micron$, which corresponds to $\log[C_{24}(20.53~\micron)]=42.57$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$). Taking the ratio of these numbers gives $\mathrm{SFR}=9.63~M_\odot \rm{yr}^{-1}$. This value differs from the actual value of $\langle SFR \rangle=9.60~M_\odot \rm{yr}^{-1}$ by $<1\%$.
In the same manner, each calibration can be applied to any redshift that spans the $\lambda_{\rm{eff}}$ range covered by IRS. These examples highlight the importance of using large sample sizes in the application of these diagnostics, as a single case can have SED variations relative to the mean of our sample, which can give rise to slight inaccuracies in SFR estimates. It is important to emphasize that the accuracy of such an application is dependent on the shape of the SED of SFGs as a function of redshift. The extent to which this condition holds is examined in detail in \S~\ref{template_variation}.
\subsection{Limitations of this Sample} \label{limitations}
It is important to acknowledge the potential differences of this sample with respect to high-$z$ galaxies as well as the limitations for its use. As was mentioned, the selection criteria for the SSGSS sample limits it to relatively high metallicities, which may not be a well representative sample as one goes to high-$z$. In addition, if the dust content of high-$z$ galaxies is different, it is possible that the amount of UV light reprocessed by dust could change. For example, if high-$z$ galaxies had more dust, then our templates would overestimate the SFR, as it would be implicitly adding back in unobscured UV flux present in the SSGSS sample but that may not be there for the high-$z$ galaxies. Variations in the typical dust temperature of galaxies with redshift would also pose a problem, as this would result in variations in their FIR SED. The relative importance of some of these effects will be tested when we compare our SED to those at higher redshift (\S~\ref{template_variation}).
Another area for concern is in the range of $L_{\rm{IR}}$\ values spanned by the SSGSS sample. Our template is made utilizing galaxies over a range of $9.53\le \log(L_{\rm{IR}}/L_{\odot})\le 11.37$, which is lower than the range that is currently accessible at high-$z$. However, the results of \citet{elbaz11} suggest that luminous infrared galaxies (LIRGs; $L_{\rm{IR}}>10^{11}L_{\odot}$) and ultra luminous infrared galaxies (ULIRGs; $L_{\rm{IR}}>10^{12}L_{\odot}$) identified in the GOODS-\textit{Herschel} sample at high-$z$ have similar SEDs to normal SFGs, in contrast to their starburst-like counterparts found locally \citep[e.g.,][]{rieke09}. They find that the entire population of IR-bright galaxies has a distribution with a median of $\mathrm{IR8}=L_{\rm{IR}}/L_{\rm{rest}}(8\mu\mathrm{m})=4.9~[-2.2,+2.9]$, where the term in brackets is the $1\sigma$ dispersion. Elbaz et al. suggest that this population can be separated into two groups: a main-sequence (MS) of normal SFGs for which $\mathrm{IR8}=4\pm2$ consisting of $\sim$80\% of the population, and starburst (SB) galaxies which occupy the region with $\mathrm{IR8}>8$ and represent about $\sim$20\% of the population. For reference, the IR8 value of our template is 4.8, which agrees with the median of the GOODS-\textit{Herschel} sample. The uniformity of IR8 values in MS galaxies, over the range $10^9<L_{\rm{IR}}/L_{\odot}<10^{13}$, suggests that the SED of normal SFGs do not change drastically with luminosity. This also indicates that our limited range in $L_{\rm{IR}}$\ coverage for the SSGSS sample should not drastically affect its utility towards higher luminosity MS galaxies.
Perhaps the biggest factor limiting the large scale application of this technique is in the ability to identify galaxy types at higher redshifts. The calibrations presented in this work are applicable to normal SFGs, typically referred to as being on the main-sequence of star formation, and not to cases undergoing starburst activity (different SED) or with AGN (significant IR emission not associated with star formation). This topic should be thoroughly addressed before widespread applications of these calibrations can be made to specific surveys.
There are a several techniques that have been demonstrated to isolate out AGN and starburst galaxies, however, some of them rely on observations made outside the MIR. One of the most reliable techniques to identify AGN is thorough X-ray observations \citep[e.g.,][]{alexander03}, however these can miss obscured AGN and could be biased \citep{brandt&hasinger05}. In order to avoid obscuration effects, AGN selection techniques using the MIR and FIR have also been developed. These include \textit{Spitzer}+\textit{Herschel} color-cuts \citep{kirkpatrick12}, \textit{Spitzer}/IRAC color-cuts \citep{lacy04,stern05,donley12,kirkpatrick12}, and WISE color-cuts \citep{stern12,mateos12,assef13}. Emission line diagnostics, such as the BPT diagram \citep{kewley13} and the Mass-Excitation diagram \citep{juneau14}, are also effective techniques. It has been suggested that starburst galaxies can be identified as sources with $\mathrm{IR8}>8$ by \citet{elbaz11}.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.0in]{SSGSS_F1000W_calibration.eps} &
\includegraphics[width=3.0in]{SSGSS_F1280W_calibration.eps} \\
\includegraphics[width=3.0in]{SSGSS_F1500W_calibration.eps} &
\includegraphics[width=3.0in]{SSGSS_F1800W_calibration.eps} \\
\includegraphics[width=3.0in]{SSGSS_F2100W_calibration.eps} &
\includegraphics[width=3.0in]{SSGSS_F2550W_calibration.eps}
\end{array}$
\end{center}
\caption{The top of each panel shows the conversion factor for select JWST/MIRI bands using all SFG galaxies (solid red lines), along with their uncertainty (dotted red lines), which for most cases is between $15-20\%$. The region below $\sim$$6~\micron$ for each band is excluded due to significantly increased uncertainty in the composite spectrum (see \S~\ref{filter_smooth}). The bottom of each panel shows the residuals between the conversion factor and a fit to the curve ($\log[C_{x}(\lambda)/fit_{x}(\lambda)]$). \label{fig:JWST_calibration_all_SFG}}
\end{figure*}
\section{Discussion}
\subsection{Origins of the Scatter in SFR(\texorpdfstring{$\lambda$}{lambda})} \label{scatter}
In addition to grouping all of the SFGs together, we also examine grouping our galaxies based on their luminosity at rest-frame 3.6, 4.5, 8, 24, and $70~\micron$, as well as their $L_{\rm{IR}}$, $L_{\rm{IR}}$\ surface brightness, and $L($H$\alpha$)/$L(24~\micron$) ratios, in order to identify possible origins to the scatter within the SFR calibration of the entire group. These bands are chosen because $3.6~\micron$ and $4.5~\micron$ correlate with the underlying stellar population \citep[i.e., stellar mass;][]{meidt12,meidt14}, and the other bands correlate strongly with star formation \citep{zhu08,rieke09,calzetti10,hao11}. For each of these cases, the sample is divided into 6 bins with $9-10$ galaxies in each.
Looking at each of the calibrations, weak trends are found suggesting larger conversion factors, at almost all MIR wavelengths, for galaxies with higher luminosities when arranged by any of the luminosities mentioned before. A few examples are shown in Figure~\ref{fig:group_calibrations}. These trends are very weak because the separation between the groups is comparable to the scatter within each of the groups, which is $10-25\%$ (the largest scatter at lowest luminosity galaxies), and similar to the uncertainty of the entire group average values. If real, these trends could suggest that (1) galaxies with a larger old stellar population require slightly larger conversion factors at all MIR wavelengths, as might be expected if light unassociated to star formation is contaminating the MIR; and/or (2) a slightly super-linear relationship exists between luminosity and SFRs in the MIR.
Our attempts to account these effects by introducing additional terms into the conversion factors do not appear to significantly reduce the overall scatter of the conversion factors. Such additional terms would also make application of this method to higher redshift galaxies more difficult, as more information would be needed (e.g., determination of rest-frame luminosities). Given our limited range in galaxy properties to determine the validity of any trends, we adopt the simplest approach and use the entire group average for our analysis. We refrain from using higher luminosity local galaxies, such as the (U)LIRGs in the GOALS sample \citep{armus09}, as an additional test of such claims because of the significant FIR SED evolution that occurs for $L_{\rm{IR}}\gtrsim10^{11}L_{\odot}$, which is absent from galaxies of these luminosities at high-$z$ \citep[see \S~\ref{limitations};][]{elbaz11}. In addition, a large fraction of these systems are likely to host AGN \citep{u12}, which would be excluded by our selection process. For reference to the reader, we illustrate the local SED evolution by comparing the GOALS photometry \citep{u12} and a few of the \citet{rieke09} templates to our own template, normalized by $L_{\rm{rest}}(8\micron)$, in Figure~\ref{fig:GOALS_compare}.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{calibration_S24_3pt6ordered.eps} &
\includegraphics[width=3.5in]{calibration_S24_IRordered.eps} \\
\includegraphics[width=3.5in]{calibration_S70_3pt6ordered.eps} &
\includegraphics[width=3.5in]{calibration_S70_IRordered.eps}
\end{array}$
\end{center}
\caption{Top: $C_{24}(\lambda)$ conversion factor for galaxies when arranged into groups ($\sim$10 galaxies) according to $L(3.6\mu\rm{m})$, a proxy for stellar mass, and $L_{\rm{IR}}$, a proxy for the SFR. Bottom: $C_{70}(\lambda)$ conversion factor for galaxies when arranged according to $L(4.5\mu\rm{m})$ and $L_{\rm{IR}}$. Local conversion factors are also shown for comparison. The dispersion in each of the groups ($10-25\%$; see \S~\ref{scatter}) is not shown for clarity, but is comparable to the separation among the groups. Weak trends appear which would suggest larger conversion factors are needed for the higher luminosity galaxies. \label{fig:group_calibrations}}
\end{figure*}
\begin{figure}
\plotone{GOALS_template_compare.eps}
\caption{Comparison of our SFG composite template to GOALS photometry \citep{u12} and \citet{rieke09} templates. The scatter of the SSGSS template is shown as the filled gray region. The values of $\mathrm{IR8}=L_{\rm{IR}}/L_{\rm{rest}}(8\mu\mathrm{m})$ for the \citet{rieke09} templates range from $\mathrm{IR8}=4.8$ at $L_{\rm{TIR}}=10^{10}L_{\odot}$ to $\mathrm{IR8}=59.6$ at $L_{\rm{TIR}}=10^{13}L_{\odot}$, whereas high-$z$ galaxies over this luminosity range have $\mathrm{IR8}=4.9~[-2.2,+2.9]$ \citep{elbaz11}. Therefore, there is significant FIR SED evolution that occurs for local (U)LIRGS that is absent at high-$z$. \label{fig:GOALS_compare}}
\end{figure}
\subsection{Variation in SFG SEDs with Redshift} \label{template_variation}
Many studies have sought to characterize the SED of different galaxy types (e.g., SFG, AGN) as functions of redshift. In this section, the templates of \citet{elbaz11,kirkpatrick12,magdis12} and \citet{ciesla14} are compared to our own to thoroughly examine the extent to which the SED of SFGs change with redshift. Similar to the approach outlined in this work, these studies use large surveys to construct IR templates for different populations of galaxies at different redshifts. In general, the templates created in these studies suggest that the mean dust temperature of galaxies increases as one looks to higher redshifts. In addition to the change in dust temperature that is evident, \citet{magdis12} suggest that the value of IR8 increases mildly from $\mathrm{IR8}\sim4$ to $\mathrm{IR8}\sim6$ at $z>2$ for MS galaxies.
If one considers the notion that both $L_{\rm{rest}}(8\micron$) and $L_{\rm{IR}}$\ are typically used for SFR indicators, such a change in IR8 would suggest that there is a change in the SFR converstion factor of one (or both) of these luminosities with redshift and this is important to keep in mind when comparing the templates. Changes in $L_{\rm{rest}}(8\micron$) could result from variations in PAH abundances relative to the total dust content, which has been found to correlate with metallicity \citep{engelbracht05,engelbracht08,marble10}, and also to the hardness of the radiation field \citep{madden06,gordon08,engelbracht08}. Given the sensitivity of $L_{\rm{IR}}$\ to the contributions from older stellar populations \citep{calzetti10}, it is also likely that the value of the conversion factor for SFR-$L_{\rm{IR}}$\ could also vary with redshift.
The comparison between the templates from the literature to our own is shown in Figures~\ref{fig:template_compare} and \ref{fig:template_compare_model}. We have chosen to normalize the templates in two ways, both of which correlate with star formation. Normalizing by a close proxy for star formation is crucial to compare how viable our continuous calibrations are at higher redshifts. The first method is to normalize by $L_{\rm{rest}}(8\micron$), which is chosen over use of the $24~\micron$ region because it is not available for the \citet{kirkpatrick12} templates. With this choice of normalization, it is also easier to directly compare the shape of the MIR SEDs. The second method is to normalize by $L_{\rm{IR}}$, as is traditionally done in many template comparisons. We reemphasize that the observed trend of IR8 increasing from 4 to 6 implies that these choices of normalization for the templates will give different results.
First, the templates of \citet{kirkpatrick12} are examined as these provide the best sample for comparison because they are based on direct spectral measurements of higher redshift galaxies. In addition, access to spectral data allowed them to accurately identify galaxies with significant AGN contribution and create separate templates for AGN and SFGs, the latter of which is considered here. The gap in spectral coverage of their templates, shown as the vertical dotted lines in Figure~\ref{fig:template_compare}, correspond to regions lacking spectral or photometric values with which to constrain the SED, and is ignored for our comparison. Looking at the templates normalized by $L_{\rm{rest}}(8\micron$), it is seen that the templates show remarkable agreement in SED shape for $\lambda<24~\micron$ and lie almost entirely within the scatter in our local SED template. In contrast, there is clear disagreement in SED shape at $\lambda>24~\micron$ which becomes more drastic at higher redshift. This is mostly due to the larger IR8 values of these templates, which exceed the IR8 values observed in photometric samples at these redshifts \citep{magdis12}. For reference, the \citet{kirkpatrick12} $z\sim1$ template has IR$8=6.5$ and the $z\sim2$ template has IR$8=8.0$. We associate this difference to the selection criteria of this sample, which required bright sources at $24~\micron$ ($S_{24}>100~\mu$Jy) to obtain IRS spectroscopy and which corresponds to more $L_{\rm{IR}}$\ luminous galaxies at higher redshifts (see \S\ref{Cx_test} for more details). When instead normalized by $L_{\rm{IR}}$, slight offsets appear between the templates for $\lambda<24~\micron$ as a result of the larger $L_{\rm{IR}}$\ with increasing redshift. The fact that the shape remains fixed, regardless of possible offsets, gives credibility to this technique being applicable to up to $z\sim2$ for all bands at effective wavelengths below $24~\micron$. In contrast, even when normalized by $L_{\rm{IR}}$\ there is clear disagreement in SED shape at $\lambda>24~\micron$ which becomes more drastic at higher redshift. This effect, also observed by \citet{magdis12}, is argued to be due to the mean dust temperature of galaxies increasing with redshifts.
There are multiple physical mechanisms that give rise to increased dust temperatures in galaxies. Locally, similar trends are seen in galaxies with increasing values of $L_{\rm{IR}}$\ \citep[e.g.,][]{rieke09}. However, for local LIRGs and ULIRGs there is also an associated decrease in the relative strength of the $8~\micron$ PAH feature relative to the FIR, corresponding to IR8 values more similar to SB galaxies, which deviates significantly from our template. Instead, \citet{magdis12} suggest that this could be the result of a hardening of the radiation field, $\langle U \rangle$, in MS galaxies with increasing redshift ($\langle U \rangle\propto(1+z)^{1.15}$). Adopting \citet{draine&li07} models to fit their galaxy SEDs, for which $\langle U \rangle\propto L_{\rm{IR}}/M_{\rm{dust}}$, they argue that this is explained by the redshift evolution of the $M_*-Z$ and $\mathrm{SFR}-M_*$ relations. Another physical mechanism that gives rise to this effect is compactness. More compact star formation in galaxies can give rise to elevated dust temperatures and appears to occur more frequently in MS galaxies at higher redshifts \citep{elbaz11,schreiber14}. Regardless of the origin of this effect, these results suggest that the $70~\micron$ band requires additional correction to be utilized as a SFR diagnostic as a function of redshift. We perform this analysis in the next section.
Next, the templates of \citet{elbaz11} are considered. These templates make use of redshifted photometry of galaxies from $0<z<2$ to act as spectroscopic analog. The combination of all galaxies over this redshift range of results in an artificially broad FIR bump, due to the shifting of the FIR bump with $z$, and makes direct comparison of these templates tricky. In general there is good agreement in SED shape with their MS template and our own if this FIR broadening is taken into account.
Lastly, the templates based on \citet{draine&li07} model fitting of photometric data are considered, shown in Figure~\ref{fig:template_compare_model}. These include the templates of \citet{magdis12}, and \citet{ciesla14}. Considering first the $z\sim0$ cases normalized by $L_{\rm{rest}}(8\micron$), we note that all of these model-based templates show the same excess in the $10-25~\micron$ region compared to the spectral data that was seen in our own fits using \citet{draine&li07} models (cyan line in the Figure~\ref{fig:template_compare_model}). This suggests that these model-based templates may not be accurately representing the intrinsic SED over this region. The region beyond $25~\micron$ is likely to be more representative, as it is usually well fit a simple two component dust model. As with the spectral-based templates, the FIR bump peaks at shorter wavelengths with increasing redshift and also shows an increase in IR8.
Taken together, it would appear that there is no strong evidence to suggest that the shape of the SED for SFGs varies significantly over the wavelength region of $6-30~\micron$. However, vertical offsets, corresponding to a constant factor offset, cannot be ruled out without direct comparison of SFR estimates for higher redshift galaxies.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{template_compare.eps} &
\includegraphics[width=3.5in]{template_compare_LFIR.eps}
\end{array}$
\end{center}
\caption{The top of each panel shows the comparison of the composite SED of SFGs in our sample to those at higher redshifts for which spectroscopic information is available. The scatter of the SSGSS template is shown as the filled gray region. The SEDs have been normalized by $L_{\rm{rest}}(8\micron$) and $L_{\rm{IR}}$. The sections between the vertical dotted green and red lines, corresponding to the $z\sim1$ and $z\sim2$ templates from \citet{kirkpatrick12}, lack spectral data. The template of \citet{elbaz11} uses redshifted photometry to act as a spectroscopic analog. The bottom of each panel shows the residuals between our template and the other templates. The shape of the SED remains unchanged with redshift for $\lambda\lesssim20~\micron$, with only constant offsets occurring depending on the normalization. For $\lambda\gtrsim20~\micron$, significant SED evolution is present with increasing redshift. \label{fig:template_compare}}
\end{figure*}
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{template_compare_model.eps} &
\includegraphics[width=3.5in]{template_compare_LFIR_model.eps}
\end{array}$
\end{center}
\caption{The top of each panel shows the comparison of the composite SED of SFGs in the SSGSS sample to those at higher redshifts for which \citet{draine&li07} models have been used to fit the available photometry. The scatter of the SSGSS template is shown as the filled gray region. The SEDs have been normalized by $L_{\rm{rest}}(8\micron$) and $L_{\rm{IR}}$. Note that when normalized by $L_{\rm{rest}}(8\micron$) these model-based template have an excess in the $10-25~\micron$ region compared to the spectral data. This trend is seen in our own fits of \citet{draine&li07} models (cyan line), and indicates a limitation in the simple 3-component model typically adopted (see \S~\ref{DL07_method}). The bottom of each panel shows the residuals between our template and the other templates. \label{fig:template_compare_model}}
\end{figure*}
\subsection{Accounting for Dust Temperature Variation} \label{C70z}
The significant change in shape of the FIR bump makes the calibrations at the longer wavelengths (i.e., $C_{70}(\lambda)$) more difficult. Comparing the SED of the higher-$z$ galaxies at these wavelengths to the SSGSS SED, there is a significant difference (up to $0.5$~dex), which exceeds the scatter of SEDs for local SFGs. For this reason, a correction to $C_{70}(\lambda)$ does appear necessary if it is to be applied at higher redshifts.
We correct for the dust temperature variation using the the SED template grids of \citet{bethermin12}, which are built from the results of \citet{magdis12}. These templates have been normalized by $L_{\rm{IR}}$. By making use of this grid, the observed \textit{Spitzer} $70~\micron$ luminosity as function of redshift is estimated while accounting for the changing SEDs. A demonstration of how the observed luminosity changes is shown in Figure~\ref{fig:C70_z}. For example, at $z=0.5$ and $z=1$ the $70~\micron$ band measures rest-frame $46.7~\micron$ and $35.7~\micron$, respectively, and the observed band luminosity is derived from the $z=0.5$ (purple line) and $z=1$ (blue line) templates at those wavelengths. We perform a fit to this new conversion factor and present it in Table~\ref{Tab:calibration_parameters}. The accuracy of this correction will be tested in the following section.
\begin{figure}
\plotone{C70_z_dependence_LFIR_B12.eps}
\caption{The SED of SFGs changes with redshift, as demonstrated by the templates of \citet{bethermin12}. Taking this into account, the value of $C_{70}(\lambda)$ would change significantly from those derived from a $z\sim0$ SED (dashed black line). This change is demonstrated as the red dashed line, which shows the 70~$\micron$ filter convolution when accounting for the SED variations with redshift. We stress that the red dashed line is a z-dependent interpretation of the expected 70~\micron\ emission and cannot be considered a spectrum for an individual galaxy. \label{fig:C70_z}}
\end{figure}
\subsection{Testing the Calibrations} \label{Cx_test}
To test the utility of our calibrations, we compare SFRs of galaxies from other surveys to those found using our continuous, monochromatic values developed in this work. This requires a survey which has photometry available in one of the calibrated bands, as well as an independent technique to measure star formation from the those used to calibrate our conversion factors. We choose to use SFRs based on $L_{\rm{IR}}$\ measurements as these are the most readily available diagnostic for deep IR surveys. For consistency with our local SSGSS sample, we adopt a conversion factor of $\log[C(L_{\mathrm{IR}})]=43.64$~erg~s$^{-1}/(M_\odot \rm{yr}^{-1}$), which corresponded to a $\tau\sim500$~Myr constant star formation (see \S~\ref{lir}). Furthermore, sources with significant AGN components need to be to identified and removed. It is worth noting that adopting different $L_{\rm{IR}}$\ conversion factors for this analysis will only lead a constant offset between these two SFRs at all redshifts and that we are most interested in assessing where breaks from a constant relation develop.
First we use the sample of 70 sources identified as SFGs from \citet{kirkpatrick12}, corresponding to AGN contribution of less than $20\%$. These galaxies cover a redshift range of $0.3<z<2.5$ and have full \textit{Spitzer} and \textit{Herschel} photometry. We use the $L_{\rm{IR}}$\ measurements of these galaxies from \citet{kirkpatrick12} (private communication), determined from IRS measurements for the MIR and by fitting two modified blackbodies for the FIR. A comparison of SFR($C_{24}(\lambda)$) to SFR($L_{\rm{IR}}$) is shown in Figure~\ref{fig:C24LIR_compare_K12}, both as a function of redshift and $L_{\rm{IR}}$. There is general agreement between the values up to redshifts of about $z\sim1$, which corresponds to galaxies with log[$L_{\rm{IR}}$/$L_\odot]<12$. Given that the sources of \citet{kirkpatrick12} were required to be very bright in the IR to obtain IRS spectral measurements at these redshifts, it is likely that their sources at $z>1$ are slightly biased to larger $L_{\rm{IR}}$\ luminosities (demonstrated by their larger IR8 values). These values deviate significantly from deeper photometric surveys of SFGs at these redshifts, which is shown in Figure~\ref{fig:C24LIR_compare_K12}, by the templates of \citet{magdis12} (dashed cyan line). We remind the reader that the \citet{magdis12} templates are based on \citet{draine&li07} models, which was found to show significant offsets compared to the observed spectra of the SSGSS galaxies, and is only shown for reference. We also examine the comparison of SFR($C_{70}(\lambda)$) to SFR($L_{\rm{IR}}$), shown in Figure~\ref{fig:C70LIR_compare_K12}. The redshift-dependent correction of $C_{70}(\lambda)$ seems to work well for galaxies of $z\lesssim1.2$, for which data is available for this band.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{SFRC24_specz_K12.eps} &
\includegraphics[width=3.5in]{SFRC24_SFRlir_K12.eps}
\end{array}$
\end{center}
\caption{Comparison of SFRs estimated from $C_{24}(\lambda)$ and $L_{\rm{IR}}$\ for the GOODS-\textit{Herschel} sample from \citet{kirkpatrick12}. Left: Comparison as a function of redshift. The values show agreement for $z\lesssim1$, beyond which the dataset is biased towards galaxies with log[$L_{\rm{IR}}$/$L_{\odot}]\gtrsim12$. The distribution of the \citet{kirkpatrick12} sources, along with the parameters of a best-fit Gaussian to this distribution, is also shown. Right: Comparison as a function of $L_{\rm{IR}}$. The values show agreement for cases with log[$L_{\rm{IR}}$/$L_{\odot}]\lesssim12$. For cases with log[$L_{\rm{IR}}$/$L_{\odot}]\gtrsim12$, which dominate $z\gtrsim1$ for this sample, the monochromatic SFR is lower than the $L_{\rm{IR}}$\ SFR. \label{fig:C24LIR_compare_K12}}
\end{figure*}
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{SFRC70_specz_K12.eps} &
\includegraphics[width=3.5in]{SFRC70_SFRlir_K12.eps}
\end{array}$
\end{center}
\caption{Comparison of SFRs estimated from $C_{70}(\lambda)$ and $L_{\rm{IR}}$\ for the GOODS-\textit{Herschel} sample from \citet{kirkpatrick12}. Left: Comparison as a function of redshift when using our calibration based on a $z=0$ template (filled triangles) and the $z$-dependent template (open squares). The calibration derived from the $z$-dependent template appears to work better than the $z=0$ template and shows agreement for $z\lesssim1.2$, beyond which data is lacking. The distributions of the \citet{kirkpatrick12} sources when using the $z=0$ template (filled gray) and $z$-dependent template (open red), along with the parameters of a best-fit Gaussian to these distributions, are also shown. Right: Comparison as a function of $L_{\rm{IR}}$. \label{fig:C70LIR_compare_K12}}
\end{figure*}
As a second test we use the sample from \citet{elbaz11}. This sample covers a redshift range of $0.03<z<2.85$ and has \textit{Spitzer} and \textit{Herschel} photometry. The $L_{\rm{IR}}$\ values for these galaxies have been determined by \citet{schreiber14}, and are estimated from the \citet{chary&elbaz01} template that provides the best fit to the Herschel data. For our analysis we only consider sources for which at least one photometric band covers wavelengths greater than 30$~\micron$, as these cases achieve better accuracy of the FIR region. The photometric redshifts of these sources are obtained from \citet{pannella14} \citep[using the EAZY code;][]{brammer08}, and we require the sources to have suitable quality flags. These photometric redshifts achieve a relative accuracy ($\Delta z=(z_{\mathrm{phot}}-z_{\mathrm{spec}})/(1+z_{\mathrm{spec}})$) of 3\%, with less than 3\% of cases suffering from catastrophic failures \citep[$\Delta z>0.2;$][]{pannella14}. These sources lack spectroscopic measurements to identify AGN or starburst sources and we rely on the color-cut outlined by \citet{kirkpatrick12} for AGN and also remove sources with IR8$>$8 from our sample, which are believed to be predominately starburst galaxies \citep{elbaz11}. We follow the method of \citet{elbaz11} to determine rest-frame $8~\micron$ from $k$-correcting the $8~\micron$ band ($z<0.5$), $16~\micron$ band ($0.5<z<1.5$), and $24~\micron$ band ($1.5\le z\le 2.5$) assuming these galaxies follow the IR SED of M82. These constraints leaves us with a SFG sample of $825$ sources with $24~\micron$ observations and $66$ with $70~\micron$ observations.
Using the measured $L_{\rm{IR}}$\ values from these sources, the same comparison is made as before and is shown in Figure~\ref{fig:C24LIR_compare}. In this case there appears to be more agreement among the diagnostics out to redshifts of $z\lesssim2$, with a $1\sigma$ dispersion of 0.16 dex (45\%). Small changes appear to develop beyond $z>2$, which is consistent with the observed trend of IR8 going 4 to 6 by $z=3$ (a difference of $\sim$$0.2$ dex). This trend is apparent in the ratios of observed $24~\micron$ luminosity to $L_{\rm{IR}}$\ in the templates of \citet{magdis12} (dashed cyan line). Similar to the \citet{kirkpatrick12} sample, the largest discrepancies occur in galaxies with log[$L_{\rm{IR}}$/$L_\odot]>12$. Next, we examine the comparison of SFR($C_{70}(\lambda)$ to SFR($L_{\rm{IR}}$) shown in Figure~\ref{fig:C70LIR_compare}. As before, the redshift-dependent correction of $C_{70}(\lambda)$ seems to work well for galaxies of $z\lesssim1$, with a $1\sigma$ dispersion of 0.18 dex (50\%). For $z\gtrsim 1$, significant differences appear but this is likely due to the poor sensitivity of the 70$~\micron$ band detecting only the most luminous galaxies in these bands at high redshifts. With the limited number of sources available at $z>1$, the reliability of our corrections cannot be determined for this range.
Common to the $C_{24}(\lambda)-$$L_{\rm{IR}}$\ comparisons for the two samples considered is the trend that as one goes to $z\gtrsim2$ and/or $L_{\rm{IR}}\gtrsim10^{12}L_{\odot}$ the SFRs predict from $L_{\rm{IR}}$\ will be larger than those inferred from the MIR. This may demonstrate that $C_{24}(\lambda)$ is unsuitable when considering galaxies with log[$L_{\rm{IR}}$/$L_\odot]>12$ or $z\gtrsim2$, but it could also indicate a change occurs in $L_{\rm{IR}}$ conversion factor at these luminosities/redshifts. Unfortunately, it is unclear whether what we are observing is a redshift effect or a luminosity effect, as there is a degeneracy between these variables that cannot be resolved with our current data. In other words, we may simply be seeing a selection effect. Given the higher sensitivity of JWST, would including sources with $L_{\rm{IR}}\lesssim12$ at $z\gtrsim2$ follow the same trend of increasing IR8 when considering the entire population? Prior to the JWST mission, a technique that does not rely on $L_{\rm{IR}}$\ to determine SFRs for higher redshift galaxies will be necessary to state confidently what effect is occurring.
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{SFRC24_zphot.eps} &
\includegraphics[width=3.5in]{SFRC24_SFRlir.eps}
\end{array}$
\end{center}
\caption{Comparison of SFRs estimated from $C_{24}(\lambda)$ and $L_{\rm{IR}}$\ for the GOODS-\textit{Herschel} sample from \citet{elbaz11}. Left: Comparison as a function of redshift. The values show agreement for $z\lesssim2$, beyond which the dataset is biased towards galaxies with log[$L_{\rm{IR}}$/$L_{\odot}]\gtrsim12$. The distribution of the \citet{elbaz11} sources, along with the parameters of a best-fit Gaussian to this distribution, is also shown. Right: Comparison as a function of $L_{\rm{IR}}$. The values show reasonable agreement for cases with log[$L_{\rm{IR}}$/$L_{\odot}]\lesssim12$. For cases with log[$L_{\rm{IR}}$/$L_{\odot}]\gtrsim12$, which dominate $z\gtrsim2$ for this sample, the monochromatic SFR is lower than the $L_{\rm{IR}}$\ SFR. \label{fig:C24LIR_compare}}
\end{figure*}
\begin{figure*}
\begin{center}$
\begin{array}{cc}
\includegraphics[width=3.5in]{SFRC70_zphot.eps} &
\includegraphics[width=3.5in]{SFRC70_SFRlir.eps}
\end{array}$
\end{center}
\caption{Comparison of SFRs estimated from $C_{70}(\lambda)$ and $L_{\rm{IR}}$\ for the GOODS-\textit{Herschel} sample from \citet{elbaz11}. Left: Comparison as a function of redshift when using our calibration based on a $z=0$ template (filled triangles) and the $z$-dependent template (open squares). The calibration derived from the $z$-dependent template appears to work better than the $z=0$ template and shows agreement for $z\lesssim1$, beyond which data is lacking. The distributions of the \citet{elbaz11} sources when using the $z=0$ template (filled gray) and $z$-dependent template (open red), along with the parameters of a best-fit Gaussian to these distributions, are also shown. Right: Comparison as a function of $L_{\rm{IR}}$. \label{fig:C70LIR_compare}}
\end{figure*}
\section{Conclusions}
We have presented continuous, monochromatic star formation rate (SFR) indicators over the mid-infrared wavelength range of $6-70~\micron$, using a sample of 58 star forming galaxies in the SSGSS at $z<0.2$. The continuous wavelength coverage granted with this sample has allowed for the calibration of \textit{Spitzer}, WISE, and JWST bands as SFR diagnostics covering, continuously, redshifts from $0<z<3$. We find that these diagnostics are consistent with monochromatic calibrations of SFGs in the local universe, and achieve accuracies of 30\% or better. They also appear consistent with templates of high-$z$ SFGs, with no significant evidence of variations in the shape of the SED over $6-30~\micron$ region. Subtle changes of $\mathrm{IR8}=L_{\rm{IR}}/L_{\rm{rest}}(8\mu\mathrm{m})$ with redshift appear to cause variations at $z\gtrsim2$, but currently it is unclear whether this could be due to a selection bias at these redshifts. Due to the significant changes in the FIR region beyond $30~\micron$ with redshift, the use of this region as a SFR diagnostic requires correction to our local template, however, this has been demonstrated to work well up to redshifts of at least $z\sim1$.
These powerful diagnostics are critical for future studies of galaxy evolution and allows for much easier application to large survey programs with a limited number of MIR wavelength bands. This technique is only valid for SFGs, and therefore methods are required to remove AGN and starburst from any sample before use. With the upcoming JWST mission, we hope that these diagnostics will provide important contributions as we begin to examine more typical main-sequence galaxies up to $z\sim3$.
\section*{Acknowledgments}
The authors thank the referee whose suggestions helped to clarify and improve the content of this work. We also thank the MIRI instrument team for providing the MIRI filter curves prior to their publication. AJB thanks K. Grasha, A. Kirkpatrick, A. Pope, and D. Marchesini for comments and discussion that improved the content of this paper. AJB and DC gratefully acknowledge partial support from the NASA ROSES Astrophysics Data Analysis Program, under program number NNX13AF19G.
This research has made use of the NASA/IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration (NASA).
This work is based on observations made with the \textit{Spitzer} Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA.
This work has made use of SDSS data. Funding for the SDSS and SDSS-II has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, the US Department of Energy, the National Aeronautics and Space Administration, the Japanese Monbukagakusho, the Max Planck Society and the Higher Education Funding Council for England. The SDSS website is http://www.sdss.org/.
The SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions. The Participating Institutions are the American Museum of Natural History, Astrophysical Institute Potsdam, University of Basel, University of Cambridge, Case Western Reserve University, University of Chicago, Drexel University, Fermilab, the Institute for Advanced Study, the Japan Participation Group, Johns Hopkins University, the Joint Institute for Nuclear Astrophysics, the Kavli Institute for Particle Astrophysics and Cosmology, the Korean Scientist Group, the Chinese Academy of Sciences (LAMOST), Los Alamos National Laboratory, the Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute for Astrophysics (MPA), New Mexico State University, Ohio State University, University of Pittsburgh, University of Portsmouth, Princeton University, the United States Naval Observatory and the University of Washington.
This work is based on observations made with the NASA Galaxy Evolution Explorer. \textit{GALEX} is operated for NASA by the California Institute of Technology under NASA contract NAS5-98034.
This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the NASA.
|
\section{Introduction}
The radio loud active galactic nucleus (AGN) 3C390.3 (z=0.0561) belongs to a group of 10\% AGNs with double peaked broad Balmer lines.
It is well known that 3C 390.3 shows a strong variability (to $\sim$ 5-times) in the UV/optical continuum and broad lines
\citep[see][]{ba80,ba83,yo81,nt82,pp84,cw87,vz91,zh96,wa97,di98,br98,sh01,sh10,se02,ta08,gu09,po11,di12}.
Also, 3C 390.3 has a high variability in the X-ray spectrum, where a broad Fe K$\alpha$ line is present (Inda et al. 1994;
Eracleous et al. 1996; Wozniak et al. 1998; Leighly et al.1997). The X-ray emission varies in a scale of several days, showing
the highest variability in the lower energy region \citep[][]{gl03,gu09}, while the UV/optical broad lines vary at scales of
several tens to hundred light days \citep[see e.g.][]{cw87,wa97,br98,di98,sh01,se02,sh10,di12}
Although the broad double-peaked emission lines of 3C390.3 are probably emitted from an accretion disc, there are some questions about the
nature of the broad line region (BLR), as e.g.
\cite{zh91} argued that a radial
biconical outflow is present in the BLR, but cross-correlation function (CCF) analysis shows that there is no
delay between the blue and red wings of H$\beta$, that
indicates a dominant circular motion in the BLR \citep[][]{di98,sh01,sh10,di12,zh11,zh13}.
However, besides two peaks from the disc, there is an additional central, slightly red-shifted, component in the
profiles of the broad lines \citep[][]{sh10,po11}. Also, from time to time, there are
some kind of perturbations in the disc \citep{jo10,po11}. Moreover, from a detailed study of the profiles of
the H$\alpha$ and H$\beta$ broad emission lines in the
monitored period (1995-2007), \cite{po11} have shown that the geometry of the BLR of 3C390.3 seems to be
very complex but the broad line region with
the disc-like geometry has a dominant emission. On the contrary,
spectropolarimetric observations \citep{co98,co00} indicate a BLR
model in which the H$\alpha$ emission line is formed in the biconical flow and
the polarized component of the line is scattered on the inner part of the torus.
The possible jet influence on the BLR and continuum emission has been discussed in \cite{ar07,ar10}. They found
an observational evidence for the connection between the variable optical continuum of 3C390.3 nucleus,
the compact radio emission of the jet on the sub-parsec scale and emissions of new observed jet components. To explain these correlations,
they suggested that the variable optical continuum emission originates
in the innermost part of the jet.
Note here that the double peaked broad line profiles, in this and other double-peaked AGNs,
could be explained with a number of
different models for the BLR, as e.g. supermassive binary black holes
\citep[][]{ga83,ga96,pop12}; outflowing biconical gas streams
\citep[][]{zh91}; in the accretion disc \citep[][]{pe88,ro92};
two-arm spiral waves in the accretion disc \citep[][]{cw94};
or a relativistic eccentric (elliptic) disc \citep[][]{er95}, etc.
We performed the spectropolarimetric observations of 3C 390.3.
with 6m telescope of SAO in the period 2009-2014 period with the aim
to explore the disagreement between optical monitoring results which indicate dominant disc
like emission \citep[see e.g.][and reference therein]{di98,sh01,sh10,jo10,po11,di12,zh11,zh13}
and spectropolarometric observations which are in favor of the biconical flow in the BLR
\citep{co98,co00}
Here we present the analysis of spectropolarimetric observations of 3C390.3 obtained in
24 epochs and study how the parameters of the linear polarization
changes with time in the broad H$\alpha$ emission line and continuum.
The paper is organized as following: in \S 2 we describe our observations and
data reduction, in \S 3 the results are given, in \S 4 we discuss obtained results and
finally in \S 5 we outline conclusions.
\section{Observations and data reduction}
\subsection{Observation}
In the period 2009-2014 we performed spectropolarimetric monitoring of the broad line radio galaxy 3C 390.3
(24 epochs) with the 6-meter telescope of SAO RAS using the modified spectrograph
SCORPIO \citep[][]{am05,am11} in the mode of the spectropolarimetry and polarimetry in the spectral
range $4000-8000 \ \AA$ with the spectral resolution $4-10\ \AA$. We used two types of polarization
analyzers: the single Wollaston prism (WOLL-1) separates the ordinary and
extraordinary rays in two planes 0 and 90 degrees and a double
Wollaston prism (WOLL-2) consisting of two prisms illuminating half of
the parallel beam and polarized light is separating planes in 0, 90, 45 and 135
degrees. For an unambiguous measurement of parameters of the linear polarization
in the first case it is required to obtain a sequence of four pairs of spectra at different angles of the phase plate
(0.45, 22.5 and 67.5 degrees), with WOLL-1 registering two long-slit spectra (with a slit 1-2$^{\prime\prime}$
width and 120$^{\prime\prime}$ height).
The second case records simultaneously four long-slit spectra (60$^{\prime\prime}$ height) that uniquely
define the parameters of the linearly polarized radiation.
For calibration purposes, each night we observed the polarization of standard stars
from the list of \cite{hs82} and \cite{sm92}. Additionally, we observed non-polarized stars as spectrophotometric standards.
We found that the accuracy of the linear polarization measurements is $\sim0.1\%$ \citep[more details in][]{,am05,aa12,af14}.
The typical differences between our measurements of polarization standards and these in the catalogue
were 0.1-0.2\% for polarization and
2-4 degrees for the polarization angle.
In Table \ref{tab1} the log of observations is given, where are listed: Julian date, total exposure time,
number of polarization cycles, seeing, slit, analyzer,
grating, and spectral resolution. We observed in three different modes, using gratings VPHG550G, VPHG940, and VPHG1200 which
covered the H$\beta$ and H$\alpha$ wavelength ranges. Additionally we observed the inter-stellar matter
(ISM) polarization with the filter V at
$\lambda$(max) 5500 \AA. The number of cycles denotes a number of observations for each position angle in the phase plate,
i.e., one cycle corresponds to
the observations in all four above mentioned angles. The spectral resolutions given in Table
{ \ref{tab1}} were estimated using the Full Width at Half Maximum
(FWHM) of the lines from the night-sky in the integral (AGN+night sky) spectra.
From Table \ref{tab1} it can be seen that there were large seeing variations between observations during the monitoring period.
{ This may cause changes in polarization
because the relative contribution of unpolarized host galaxy light will
change with the seeing,}
however for this galaxy it is not a major problem, since the
emission of the nucleus is very bright compared to the host galaxy.
\begin{table*}
\begin{center}
\caption[]{Log of observations.}\label{tab1}
\begin{tabular}{ccccccccc}
\hline \hline
Date &JD&Total &Num. &Seeing&Slit&Analyzer&Grating/&Spectral \\
observation&240000+&exposure, s&of cycles&arcsec&arcsec&&filter&resolution (\AA)\\
\hline
2009.09.24 & 55099 & 3600 & 6 & 1 & 2.5 & WOLL-1 & VPHG550G & 20 \\
2010.07.17 & 55394 & 2400 & 5 & 2 & 2 & WOLL-1 & VPHG940 & 7 \\
2010.11.01 & 55501 & 3600 & 5 & 1.5 & 1 & WOLL-1 & VPHG940 & 5 \\
2011.05.01 & 55682 & 2160 & 3 & 2.5 & 2 & WOLL-1 & VPHG1200 & 6 \\
2011.06.01 & 55713 & 4000 & 5 & 1.6 & 1.5 & WOLL-1 & VPHG940 & 6 \\
2011.08.26 & 55799 & 1320 & 11 & 2.5 & 2 & WOLL-2 & VPHG940 & 6 \\
2011.09.27 & 55831 & 2400 & 5 & 1.6 & 2 & WOLL-1 & VPHG1200 & 5 \\
2011.11.20 & 55885 & 2880 & 6 & 2.5 & 1 & WOLL-1 & VPHG940 & 4 \\
2012.02.01 & 55959 & 1440 & 3 & 4 & 1 & WOLL-1 & VPHG940 & 4 \\
2012.02.14 & 55971 & 4320 & 6 & 3 & 2 & WOLL-1 & VPHG940 & 6 \\
2012.04.15 & 56032 & 2880 & 6 & 3 & 2 & WOLL-1 & VPHG940 & 6 \\
2012.05.17 & 56064 & 2400 & 5 & 1.2 & 2 & WOLL-1 & VPHG1200 & 5 \\
2012.06.21 & 56099 & 2880 & 6 & 1.3 & 2 & WOLL-1 & VPHG1200 & 5 \\
2012.06.21 & 56099 & ~600 & 15 & 1.5 & image & WOLL-1 & Jonson V& - \\
2012.08.24 & 56163 & 2160 & 6 & 1.2 & 2 & WOLL-1 & VPHG940 & 6 \\
2012.09.11 & 56181 & 2400 & 8 & 2.5 & 2 & WOLL-1 & VPHG940 & 5 \\
2012.10.07 & 56207 & 3600 & 8 & 2 & 2 & WOLL-1 & VPHG940 & 6 \\
2012.11.13 & 56244 & 2400 & 5 & 1 & 2 & WOLL-1 & VPHG940 & 7 \\
2013.02.06 & 56329 & 2400 & 8 & 3 & 2 & WOLL-1 & VPHG940 & 6 \\
2013.06.15 & 56458 & 4800 & 8 & 1 & 2 & WOLL-1 & VPHG940 & 6 \\
2013.11.03 & 56599 & 1260 & 21 & 3 & 2 & WOLL-2 & VPHG940 & 6 \\
2014.02.25 & 56713 & 3480 & 29 & 2 & 1 & WOLL-2 & VPHG940 & 4 \\
2014.03.06 & 56722 & 3600 & 20 & 2 & 2 & WOLL-2 & VPHG940 & 6 \\
2014.03.25 & 56741 & 3600 & 30 & 1.5 & 2 & WOLL-2 & VPHG940 & 5 \\
2014.05.30 & 56807 & 3600 & 30 & 2 & 1 & WOLL-2 & VPHG940 & 4 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsubsection{Polarization of the ISM}
The observed linear polarization of an object is a vector composition of the ISM polarization ($\vec{P}_{ISM}$)
and polarization of the object ($\vec{P}_{AGN}$ -- in this case the radio-galaxy 3C390.3),
i.e. $\vec{P}_{obs}=\vec{P}_{AGN}+\vec{P}_{ISM}$.
The ISM polarization, as it is well known, depends on the Galactic latitude and it has strong changes in the
rate of polarization and in the polarization angle on one degree scale on celestial sphere. That is
connected with non homogeneous distribution of the ISM. In a number of papers,
the ISM polarization has been taken into account as
a function of the Galactic extinction E(B-V) for different latitudes as it is described in \cite{se75}.
However,
the problem with this method is that it does not take into account the direction of the ISM polarization vector,
i.e., the vector
$\vec{P}_{ISM}$ has direction and intensity and both quantities should be taken into account \citep[see e.g.][]{kis04}.
Therefore, here we take into account the ISM polarization vector by measuring polarization of a number
of stars around the AGN, that represents the ISM polarization. For these observations we
used the wide-field polarimetry,
as within the range of 0.45-0.8 mkm a typical change in the ISM polarization
is smaller than 10\% from the maximum,
i.e. for Galactic longitudes $>25^\circ$ the estimate of wide-field $\vec{P}_{ISM}$ is satisfactory
for this purposes.
For the ISM polarization estimation we used WOLL-1 analyzer with
rotated $\lambda/2$ phase plate, and with the 3-5$^\prime$ field. We
considered only bright surrounding stars {within} 3 arcmin around 3C390.3 in the V filter.
We found 11 {brightest} stars in the field and for each star we calculated parameters
Q and U\footnote{For more detailed procedure of polarization parameters calculation in the image mode see \cite{ar14}.}.
We calculated for each of star parameters Q and U and estimated the averaged polarization parameters of the
ISM to be: $Q_{ISM} = 0.64\pm0.23\%$ and $ U_{ISM} = -0.51\pm0.20\%$, $P_{ISM} = 0.82\pm0.22\%$ , $\varphi_{ISM} = 160.7\pm6.2^{\circ}$.
Further in the text we will use the polarization parameters corrected
for the ISM polarization.
{ Note here that in the catalogue of \cite{he00} the rate of ISM polarization is
changing around 0.5-1.1\% and angle of ISM
polarization around 159-163$^\circ$ in the radius of 5$^\circ$, that is not in contradiction with our estimates.}
The contribution of the Inter Stellar Polarization (ISP) of the host galaxy of 3C390.3
has not been considered, since, as it is mentioned above,
the nucleus is significantly brighter than the host galaxy, and there is a problem to find the host galaxy ISP even using the
[OIII] lines, because in that case one assumes that these lines are
intrinsically unpolarized, which may not be the case.
\subsection{Data reduction}
The data reduction includes standard procedure for the long-slit spectroscopy, bias, flat field, geometrical correction along
the slit, correction of the spectral line curvature, night-sky subtraction, spectral sensitivity of the instrument and
spectral wavelength calibration.
Additionally we used a comparison star in order to remove the strong atmospheric bands of O2 - B(6870\AA)
and A(7600\AA) from 3C390.3 spectrum. Note that the B band is on the blue wing of H$\alpha$. We
integrated spectra along the slit, since the procedure
of decomposition (of the observed light as a function of wavelength along the slit) of the host galaxy and AGN
increases the statistical errors. integration interval was $\pm (10^{\prime\prime}-15^{\prime\prime})$. This systematic error due to
wings image does not exceed 0.1\% in the value of the degree of polarization.
Values of normalized Stokes parameters $Q(\lambda)$, $U(\lambda)$ and total intensity $I(\lambda)$ for
WOLL-1 analyzer can be found from the simple relations \citep[see also][]{af14}:
$$Q(\lambda)={1\over{2}}\biggl({I_{0}(\lambda)-I_{90}(\lambda)\over{I_{0}(\lambda)+I_{90}(\lambda)}}\biggr)_{\phi=0}-
{1\over{2}}\biggl({I_{0}(\lambda)-I_{90}(\lambda)\over{I_{0}(\lambda)+I_{90}(\lambda)}}\biggr)_{\phi=22.5},$$
$$U(\lambda)={1\over{2}}\biggl({I_{0}(\lambda)-I_{90}(\lambda)\over{I_{0}(\lambda)+I_{90}(\lambda)}}\biggr)_{\phi=0}-
{1\over{2}}\biggl({I_{0}(\lambda)-I_{90}(\lambda)\over{I_{0}(\lambda)+I_{90}(\lambda)}}\biggr)_{\phi=67.5},$$
$$I(\lambda)=\sum_{\phi}[I_0(\lambda)+I_{90}(\lambda)]_{\phi},~~~~~~~~~\phi=0,45,22.5,67.5$$
For WOLL-2 analyzer appropriate values are determined from the relations:
$$Q(\lambda)={I_{0}(\lambda)-I_{90}(\lambda)\over{I_{0}(\lambda)+I_{90}(\lambda)}},$$
$$U(\lambda)={I_{45}(\lambda)-I_{135}(\lambda)\over{I_{45}(\lambda)+I_{135}(\lambda)}},$$
$$I(\lambda)=I_0(\lambda)+I_{90}(\lambda)+I_{45}(\lambda)+I_{135}(\lambda)$$
Then we calculated the degree of the linear polarization $P(\lambda)$ and angle of polarization plane $\varphi(\lambda)$ as:
$$P(\lambda)=\sqrt{Q(\lambda)^2+U(\lambda)^2} \ \ \ \ \varphi(\lambda)={1\over 2}{\rm arctg}[U(\lambda)/Q(\lambda)]$$
Using the procedure of the polarization parameter calculation from \cite{aa12}
we estimated the Stokes parameters ($I(\lambda)$, $Q(\lambda)$ and $U(\lambda)$),
the degree of the linear polarization $P(\lambda)$ and position angle of
polarization plane $\varphi(\lambda)$ for the rest wavelength 5500 \AA.
Also we determined the continuum flux and the shift of the broad emission
line in the polarized light relatively to the systematic velocity $\Delta V=V_{pol}-V_{sys}$.
In further analysis, the polarization parameters ($I, Q, U,P~ \rm and ~\varphi$) have been robustly estimated
as average in a spectral window of 25-30 \AA ~for all cycles of measurements,
the number of the windows for different observation data are between 5 and 10.
\begin{figure}
\centering
\includegraphics[width=8.3cm]{fig1.ps}
\caption{The observed spectrum of the unpolarized and polarized flux and its polarization parameters of 3C390.3
on May 30, 2014. From top to bottom:
the observed spectrum $I(\lambda)$, the polarized spectrum $I(\lambda)\cdot P(\lambda)$, normalized Stokes
parameters $Q(\lambda)$ and $U(\lambda)$, degree of linear polarization $P(\lambda)$ and angle of
the polarization plane $\varphi(\lambda)$. Spectra have been corrected for the spectral sensitivity and
ISM polarization. The atmospheric absorption is removed.
} \label{fig1}
\end{figure}
{ We measured the continuum flux
at rest wavelength 5100 \AA\, taking an averaged continuum in the H$\beta$ wavelength range.
To measure an averaged flux of the polarized continuum, we used a window of the half-width of 500 \AA.
To avoid the contribution of the [OIII]4959,5007 \AA\, lines we centered the window at 5500 \AA.
Fig.1 shows
a typical observed spectrum of the unpolarized and polarized flux and
its polarization parameters of 3C390.3 obtained on May 30, 2014 in the instrumental
unit without the flux calibration. As it can be seen in Fig. \ref{fig1} there practically is no difference
between the polarized continuum measured at rest 5100 \AA\ and 5500 \AA\, i.e. the difference is within
the error-bars.}
From the observed H$\alpha$ and H$\beta$ lines we removed (using Gauss decomposition)
the narrow forbidden lines and obtained the broad components. Then we subtracted
the continuum and obtained the broad emission line fluxes.
For the absolute flux calibration we used the fluxes of
the [OIII]4959+5007 \AA\, ~- forbidden lines, taking the fluxes from \cite{vz91}
(Flux([OIII]=$1.7\cdot 10^{-13}\rm erg/cm^2/s$).
\cite{se02} found that for mean seeing of $2.5''$, the stellar contribution to continuum in the H$\alpha$ region
for aperture $2\times11''$ is $7.0\cdot 10^{-16}\rm erg/cm^2/s/\AA$.
We estimated that the mean continuum flux at the rest wavelength of 5100 \AA\, for the similar
aperture is F(cnt)$ \sim (3.02 \pm 0.63)\cdot 10^{-15}\rm erg/cm^2/s/\AA$.
Thus the contribution of the stellar radiation (host galaxy) to the continuum of
the AGN 3C390.3 is $\sim$20\%. We estimated that this contribution of
the host galaxy is depolarizing the AGN continuum radiation less than $<0.25\%$.
\begin{table*}
\begin{center}
\caption[]{Observed Stokes parameters $Q$ and $U$ of the continuum, degree of polarization and polarization angle,
fluxes in continuum, broad lines H$\beta$ and H$\alpha$, polarized broad H$\alpha$ and its shift. The polarized spectra
have been corrected on the ISM polarization.
}\label{tab2}
\begin{tabular}{cccccccccc}
\hlin
JD&Q(5500),&U(5500),&$P(5500),$&$\varphi(5500),$&Flux&Flux&Flux&Flux&V$_{pol}$-V$_{sys},$\\
240000+&\%&\%&\%°rees&Continuum 5100 \AA&Broad H$\beta$&Broad H$\alpha$&Polarized H$\alpha$&km s$^{-1}$\\
\hline
55098&-0.04$\pm$0.24&-1.19$\pm$0.28&1.19$\pm$0.24&134.0$\pm$1.7&4.49$\pm$0.03&3.74$\pm$0.24&2.12$\pm$0.26&2.37$\pm$0.16&~~-800$\pm$300 \\
55394&-0.27$\pm$0.32&-1.35$\pm$0.40&1.38$\pm$0.06&129.4$\pm$5.0&1.97$\pm$0.05&3.11$\pm$0.20&1.26$\pm$0.18&1.08$\pm$0.32&-1559$\pm$441\\
55501&~1.30$\pm$0.28&-0.10$\pm$0.24&1.30$\pm$0.35&177.9$\pm$9.0&2.46$\pm$0.12&2.93$\pm$0.43&1.54$\pm$0.26&2.11$\pm$0.87&-1250$\pm$765\\
55682&~0.85$\pm$0.11&-0.15$\pm$0.10&0.86$\pm$0.11&175.0$\pm$2.0&4.13$\pm$0.05&3.61$\pm$0.18&1.41$\pm$0.18&1.21$\pm$0.27&-1416$\pm$452\\
55713&~1.06$\pm$0.14& 0.50$\pm$0.14&1.18$\pm$0.14&192.7$\pm$2.3&4.19$\pm$0.07&3.86$\pm$0.34&2.27$\pm$0.27&4.02$\pm$0.40&-1294$\pm$402\\
55799&~0.33$\pm$0.17&-1.63$\pm$0.18&1.66$\pm$0.05&140.8$\pm$0.6&2.76$\pm$0.06&4.06$\pm$0.23&2.82$\pm$0.30&7.19$\pm$0.75&-1017$\pm$182\\
55831&~0.52$\pm$0.16&-1.50$\pm$0.14&1.58$\pm$0.19&144.5$\pm$1.7&2.43$\pm$0.07&3.90$\pm$0.24&3.21$\pm$0.55&3.40$\pm$0.39&-1053$\pm$238\\
55885&-0.12$\pm$0.23&-1.57$\pm$0.28&1.57$\pm$0.27&132.9$\pm$9.3&2.68$\pm$0.12&3.56$\pm$0.46&2.84$\pm$0.30&2.92$\pm$1.10&-1273$\pm$630\\
55959&-0.72$\pm$0.08&-1.53$\pm$0.51&1.69$\pm$0.49&122.5$\pm$5.4&3.33$\pm$0.20&4.03$\pm$0.47&1.98$\pm$0.29&1.49$\pm$1.33&-1221$\pm$362\\
55971&~1.13$\pm$0.30&-1.02$\pm$0.29&1.52$\pm$0.19&158.9$\pm$3.5&3.25$\pm$0.12&3.40$\pm$0.41&2.40$\pm$0.28&2.05$\pm$0.92&-1144$\pm$473\\
56032&-0.15$\pm$0.20&-1.49$\pm$0.17&1.50$\pm$0.18&132.0$\pm$2.3&2.76$\pm$0.05&3.62$\pm$0.23&1.44$\pm$0.18&1.44$\pm$0.26&~~-993$\pm$265\\
56064&~0.74$\pm$0.19&-0.83$\pm$0.13&1.11$\pm$0.15&155.8$\pm$2.2&3.05$\pm$0.06&3.25$\pm$0.29&1.15$\pm$0.70&1.20$\pm$0.13&-1323$\pm$237\\
56099&-0.08$\pm$0.12&-0.94$\pm$0.13&0.94$\pm$0.12&132.6$\pm$1.9&3.61$\pm$0.05&3.76$\pm$0.29&1.80$\pm$0.27&1.49$\pm$0.17&-1014$\pm$227\\
56163&~0.10$\pm$0.09&-1.20$\pm$0.20&1.20$\pm$0.19&137.4$\pm$1.9&3.46$\pm$0.04&3.86$\pm$0.31&2.17$\pm$0.31&1.65$\pm$0.35&-1561$\pm$342\\
56181&~0.42$\pm$0.15&-1.02$\pm$0.13&1.10$\pm$0.16&146.3$\pm$2.0&3.62$\pm$0.05&3.67$\pm$0.26&1.87$\pm$0.26&1.46$\pm$0.26&-1585$\pm$272\\
56207&~0.90$\pm$0.16&-1.12$\pm$0.10&1.44$\pm$0.14&154.4$\pm$0.9&3.41$\pm$0.09&4.15$\pm$0.31&1.36$\pm$0.18&1.75$\pm$0.25&-1540$\pm$231\\
56244&~1.02$\pm$0.18&-1.33$\pm$0.15&1.68$\pm$0.18&153.8$\pm$2.0&2.80$\pm$0.05&3.83$\pm$0.28&1.60$\pm$0.22&1.96$\pm$0.30&-1238$\pm$254\\
56329&~0.15$\pm$0.20&-1.17$\pm$0.20&1.18$\pm$0.27&138.7$\pm$2.6&2.94$\pm$0.10&3.57$\pm$0.34&1.30$\pm$0.21&1.27$\pm$0.23&-1644$\pm$284\\
56458&~1.13$\pm$0.15&~0.01$\pm$0.20&1.13$\pm$0.17&180.4$\pm$2.5&2.53$\pm$0.06&3.42$\pm$0.24&1.08$\pm$0.11&0.97$\pm$0.18&-1148$\pm$287\\
56599&~0.21$\pm$0.14&-1.17$\pm$0.14&1.19$\pm$0.06&140.0$\pm$2.1&2.55$\pm$0.06&3.42$\pm$0.30&1.03$\pm$0.13&1.33$\pm$0.19&-1453$\pm$183\\
56713&~0.54$\pm$0.04&-1.39$\pm$0.04&1.49$\pm$0.10&145.6$\pm$1.1&2.53$\pm$0.12&3.30$\pm$0.44&1.46$\pm$0.21&1.08$\pm$0.35&-1474$\pm$431\\
56722&~0.13$\pm$0.05&-1.24$\pm$0.17&1.25$\pm$0.19&138.1$\pm$2.8&2.45$\pm$0.06&3.29$\pm$0.30&1.42$\pm$0.18&1.61$\pm$0.27&-1312$\pm$207\\
56741&~0.13$\pm$0.05&-0.97$\pm$0.11&0.98$\pm$0.14&138.8$\pm$0.5&2.51$\pm$0.06&3.29$\pm$0.32&1.44$\pm$0.19&1.02$\pm$0.26&~~-999$\pm$372\\
56807&~0.29$\pm$0.13&-1.09$\pm$0.13&1.13$\pm$0.18&142.5$\pm$0.2&2.83$\pm$0.05&3.45$\pm$0.27&1.48$\pm$0.16&1.03$\pm$0.21&-1247$\pm$231\\
\hline
\end{tabular}
\\
{UNITS: Continuum 5100 \AA~ in $10^{-15}\rm erg/cm^2/s/\AA$; Broad H$\beta$ in
$10^{-13}\rm erg/cm^2/s$; Broad H$\alpha$ in $10^{-12}\rm erg/cm^2/s$; Polarized H$\alpha$ in $10^{-14}\rm erg/cm^2/s$.
}
\end{center}
\end{table*}
\section{Results}
Observed polarization parameters in the continuum and broad H$\alpha$ line are given in Table \ref{tab2}, where
we give: the
unpolarized (at 5100\AA\,) and polarized continuum (at 5500 \AA\,) flux, the polarization parameters corrected for the ISM polarization,
the broad H$\beta$ and H$\alpha$ line fluxes, the polarized H$\alpha$ flux
and the difference between the systematic and polarized line velocity
($V_{sys}-V_{pol}$) for the broad H$\alpha$ line. Note here that, due to large error-bars in the H$\beta$ wavelength range,
we did not measure the polarization parameters for
H$\beta$.
In Fig. \ref{fig1} we show spectrum observed on May 30, 2014. The slight changes can be noticed in
the rest 23 epochs, but some characteristic structures are the same in all spectra.
From Fig. \ref{fig1} and Table \ref{tab2}, one can see several characteristics in unpolarized and polarized spectra of 3C390.3:
\begin{itemize}
\renewcommand{\labelitemi}{$-$}
\item in the unpolarized 3C390.3 spectrum the double-peaked broad emission lines of H$\alpha$ and
H$\beta$ are present; the blue peak was always brighter than the red one during the monitored period,
while the polarized component is single-peaked;
\item the degree of the linear polarization in the continuum was always somewhat higher $\sim2\%$ in the shorter
wavelengths (4000 \AA), changing to $\sim1.5\%$ at 8000 \AA, i.e. there is a trend to increase the linear polarization with decreasing wavelength (i.e. $P\sim{1/\lambda}$);
\item in the polarized light we observed the H$\alpha$ (sometimes H$\beta$) broad emission line, shifted to the blue at
$\Delta V=V_{pol}-V_{sys}=-1247\pm231$ km s$^{-1}$ with respect to the systematic velocity;
\item in the H$\alpha$ line the linear polarization $P$ is always less than $\sim$0.2\%
and has a box-like shape, i.e. P(H$\alpha$) has no pronounced structural features (there are no double peaks).
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=8.3cm]{fig2.ps}
\caption{From top to bottom: the light curves of the total (unpolarized) continuum flux at 5100 \AA\, ~
in the rest frame, variation of the degree of the linear polarization $P$ and angle polarization plane of
the continuum at 5500 \AA\, in the rest frame, and shift $\Delta$V of the polarized broad H$\alpha$
line relatively to the systematic velocity.} \label{fig2}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8.3cm]{fig3.ps}
\caption{3C 390.3 light curves - from top to bottom:
the total and polarized continuum flux at 5100 \AA\, ~
in the rest frame; the broad H$\alpha$ and H$\beta$ and the polarized broad H$\alpha$
line. Flux units:
continuum at 5100 \AA\, in $10^{-15}\rm erg/cm^2/s/\AA\,$; broad H$\beta$ in
$10^{-13}\rm erg/cm^2/s$; broad H$\alpha$ in $10^{-12}\rm erg/cm^2/s$; polarized H$\alpha$ in $10^{-14}\rm erg/cm^2/s$.
} \label{fig3}
\end{figure}
\subsection{Variability in the continuum and broad H$\alpha$ polarization}
{ The polarization in the continuum has a slight dependence on wavelength
and varied between 1\% and 2\% (see Fig. \ref{fig2}).}
In Fig. \ref{fig2} we present the variability in the continuum at 5100 \AA\, and
polarized continuum measured at 5500 \AA.
As it can be seen in Fig. \ref{fig2} and Table \ref{tab2}, the variations in the linear
polarization P\% sometimes anti-correlate with the continuum flux,
i.e. the minimum value of the continuum flux corresponds to the maximum value
of the linear polarization (for example, in the period 26.08.2011-13.11.2012
or JD2455799-JD2456244). The angle of the polarization plane of the continuum reaches
maximum values ($\sim$180 degrees) in the period 01. 11. 2010- 01. 06. 2011 or
JD2455501-JD2455713.
{The shift of the polarized broad H$\alpha$ line had no
significant change during the monitored period (almost within the error-bars).
In Fig. \ref{fig3} the light curves for the total and polarized continuum flux
at 5100 \AA\, are shown\footnote{The polarized continuum flux is obtained taking the continuum flux at 5100\AA\,
and polarization degree at 5500 \AA\,}. A fast inspection of the light curves of
the unpolarized and polarized continuum shows that they are in correlation.
The variation in the polarized continuum as well as the variation
in the broad H$\alpha$, H$\beta$ lines, and polarized
H$\alpha$ line follows the variation in the unpolarized one, but with some lags between them (Fig. \ref{fig3}).
\begin{table}
\begin{center}
\caption[]{Variability properties of the continuum end emission lines in spectra 3C390.3. UNITS: F(5100)
in ${\rm 10^{-15} erg\ cm^{-2}\ s^{-1}\ \AA^{-1}}$, F(polarized H$\alpha$) in ${\rm 10^{-14} erg\ cm^{-2}\ s^{-1}\ \AA^{-1}}$,
F(broad H$\beta$) and F(broad H$\alpha$) in ${\rm 10^{-13} erg\ cm^{-2}\ s^{-1}\ \AA^{-1}}$.
}\label{tab3}
\begin{tabular}{lccccc}
\hline \hline
\\
Feature & $<F>$ & $\sigma (F)$ & $\frac{R_(max)}{R(min)}$ & $F_{var}$& $\frac{F(H\alpha)}{F(H\beta)}$ \\
\\
\hline
P(5500), \% & ~~~~2.01 & ~~0.28 & 1.58 & 0.14 & \\
$\varphi$ (5500), deg & 151.80 & 10.90 & 1.33 & 0.07 & \\
F(5100) & ~~~~3.02 & ~~0.63 & 2.28 & 0.21 & \\
F(broad H$\beta$) & ~~~~3.72 & ~~0.72 & 2.36 & 0.19 & \\
F(broad H$\alpha$) & ~~17.72 & ~~3.87 & 3.12 & 0.22 & 4.7\\
F(polarized H$\alpha$)& ~~~~1.96 & ~~1.36 & 7.41 & 0.69 & \\
F(5100) 1995-2007& ~~~~2.31 & ~~1.07 & 5.20 & 0.46 & \\
F(broad H$\beta$) 1995-2007& ~~~~2.39 & ~~0.91 & 4.70 & 0.38 & \\
F(broad H$\alpha$) 1995-2007& ~~~~9.42 & ~~3.36 & 3.40 & 0.35 &3.94 \\
\hline
\end{tabular}
\end{center}
\end{table}
In Table \ref{tab3} we defined several parameters characterizing the variability of
the polarized parameters (P\%, $\varphi$), the continuum at 5100 \AA\ and
the broad H$\beta$, H$\alpha$ emission line fluxes and the H$\alpha$ polarized flux, using the equation given by \cite{br98}:
$$F_{var}=\sqrt{\sigma(F)^2-e^2}/F_{mean}$$
where $e^2$ is the mean square value of the individual measurement uncertainty for N observations, i.e. $e^2=\sum e(i)^2/N$.
There, N is the number of spectra, F denotes the mean flux over the
whole observing period, $\sigma(F)$ is the standard deviation, and R(max)/R(min) is the ratio of
the maximal to the minimal value of the measured parameters (or flux) in the monitored period.
The parameter $F_{var}$ is an inferred
(uncertainty-corrected) estimation of the variation amplitude with respect to the mean flux.
In Table \ref{tab3} (last 3 lines) the variability parameters for the continuum flux
at 5100 \AA\, and the broad H$\alpha$ and H$\beta$ emission line flux
are given from the paper of \cite{sh10}.
As it can be seen from Table \ref{tab3}, during the monitored period
(2009-2014) the average fluxes and their amplitudes in broad lines
and continuum were about 2 times bigger while $F_{var}$ (variability)
was 2 times smaller than in the period of spectral monitoring (1995-2007).
The variability $F_{var}$ for polarization parameters is not high
(14\% for P, 7\% for $\varphi$). However, for the polarized H$\alpha$ line $F_{var}$
is $\sim69\%$ that is considerably larger than the variability of
another parameters. {The polarization in H$\alpha$ is close to the zero level, that could be reason for the
large variability.}
The maximum of the polarized H$\alpha$ flux
($F=(7.19\pm0.75)\cdot 10^{-14} {\rm erg\ cm^{-2}\ s^{-1}}$ (Table \ref{tab2}) was observed on
26. 08. 2011. At this time
the broad H$\alpha$ and H$\beta$ emission line and polarized continuum
at 5100\AA\, were close to have the maximum fluxes, but the unpolarized (total) continuum flux was maximal
in 01. 05.-01. 06. 2011, i.e., 60-90 days before the maximum in the line flux. (Table \ref{tab2} and Fig. \ref{fig3}).
Obviously, we see a lag of response of the broad line to the continuum flux (see Fig. \ref{fig3}).
\begin{figure*}
\centering
\includegraphics[width=16cm]{fig4.ps}
\caption{Left -- the interpolated normalized light curves, from top to bottom (thick lines): the polarized continuum,
the broad H$\beta$ flux,
the broad H$\alpha$ flux, the polarized H$\alpha$ flux compared with the unpolarized
continuum (thin lines). Right -- from top to
bottom the lags between the unpolarized and polarized continuum at rest 5100 \AA; the broad H$\beta$ line, the broad H$\alpha$
line, and the polarized H$\alpha$ fluxes.
} \label{fig4}
\end{figure*}
\subsection{CCF analysis - dimension of the scattering regions}
Using the reverberation method, which is the search for correlations between the broad emission line and continuum
flux variations, it is possible to study geometry and dynamics of the BLR \citep[see][and references therein]{pp93}.
By analogy we used a long-term
spectropolarimetric monitoring of 3C390.3 to study the properties of the scattering gas geometry by reverberation method.
To do this we applied the CCF analysis using first the interpolation method \citep[so called ICCF, see][]{gs86}. We interpolated the light curves in the
unpolarized and polarized continuum (as well as in the H$\alpha$ and H$\beta$ lines), after that we used the CCF
and computed the lags relative to the CCF peak.
In Fig. \ref{fig4} (left) we plot the interpolated light curves of (from top to bottom)
the polarized continuum, the broad H$\beta$, the broad H$\alpha$ and the polarized
H$\alpha$ (thick lines) compared with the interpolated normalized light curves for the unpolarized continuum (thin line).
As it can be seen in Fig. \ref{fig4} (left) different lags
between the unpolarized continuum (thin line) and mentioned fluxes are present.
Since different time series analyses can give different lags for the same set of data
\citep[see in more details][]{ko14}, therefore we calculated the lags using Z-transformed Discrete Correlation Function
\citep[ZDCF, see][]{al13}
and SPE - Stochastic Process Estimation \citep{zu11} in addition to the ICCF method. The estimated lags are given in Table \ref{tab4}.
We found that the lags between unpolarized and polarized continuum at 5100 \AA\,~ are $\sim$10 -- 40 days;
between the unpolarized continuum and the broad H$\beta$ line $\sim$ 60 -- 80 days; between the unpolarized continuum
and broad H$\alpha$ flux $\sim$140 -- 190 days; between the unpolarized continuum and polarized H$\alpha$ is
$\sim$90 -- 160 days (see Table \ref{tab4}).
In all three methods we obtained that the lag
between the unpolarized and polarized continuum is significantly smaller than lags between the continuum and broad lines,
that indicates that the { variable} polarized continuum { is originating}
in a region which is significantly smaller than the BLR.
\begin{table}
\begin{center}
\caption[]{The lags obtained using different methods: ICCF -- interpolation method \citep{gs86},
ZDCF -- Z-transformed Discrete Correlation Function \citep{al13}
and SPE - Stochastic Process Estimation \citep{zu11}.}\label{tab4}
\begin{tabular}{lccc}
\hline
lag between: & ICCF & ZDCF & SPE \\
\hline
&&&\\
cnt - pol-cnt & 10$_{-2}^{+3}$ & 27$_{-10}^{+9}$ & 39$_{-9}^{+7}$ \\
&&&\\
cnt - broad H$\beta$ & 60$_{-8}^{+7}$ & 79$_{-15}^{+16}$ & 79$_{-9}^{+8}$ \\
&&&\\
cnt - pol. broad H$\alpha$ & 89$_{-9}^{+7}$ & 128$_{-29}^{+32}$ & 156$_{-21}^{+8}$\\
&&&\\
cnt - broad H$\alpha$ & 138$_{-35}^{+47}$ & 184$_{-12}^{+28}$ & 186$_{-6}^{+7}$ \\
\ &&&\\
\hline
\end{tabular}
\end{center}
\end{table}
Note here that our lag estimates are in good agreements with these given in literature, as e.g. \cite{sh10} obtained lags
$127^{+18}_{-18}$ days for H$\alpha$ and $93^{+20}_{-18}$ days for H$\beta$.
\begin{figure}
\centering
\includegraphics[width=8.3cm]{fig5.ps}
\caption{The variation of the continuum (open circles) in the Stokes parameters (UQ) space
with the projection of the radio jet direction (arrow).
The dimension of circles corresponds to the intensity continuum flux at 5100 \AA\, in the rest frame.
The arrow represents the direction of the radio jet of 3C390.3 in the UQ space.
The black full circles represent the H$\alpha$ broad line Stokes parameters measured in the frame of the
broad line profile.} \label{fig5}
\end{figure}
\section{Discussion}
\subsection{Polarized continuum }
Using the results of our spectropolarimetric monitoring of 3C 390.3 (data for 24 epochs during 5 years) we
found the lag $\sim10-40$ days between the unpolarized and polarized continuum flux variations. The estimated size
of the BLR is significantly larger (lag $\sim 60-190$ days). This indicates that the region { of polarized radiation} which
contributes to the variability of the polarized continuum is
more compact than the BLR in the case of 3C390.3. A similar result (lag $\sim2$ days) is obtained for Mrk 6 \citep[][]{af14},
where the BLR is estimated to be significantly larger than the scattering region. \cite{ga12} found that in the case of
NGC 4151 the lag between the
unpolarized and polarized continuum flux variations is $\sim8$ days and almost the same as the BLR lag, however
it seems that their error-bars
of the estimated lags are large \citep[see their Table 1 in][]{ga12}
The average polarization angle ($\varphi=152 \pm 11$ degrees) of the continuum at rest wavelength 5500 \AA\,
is consistent and almost aligned with the radio jet
angle $\sim144^\circ$ \citep[][]{lp91}. The variability of $\varphi$ is relatively small ($F_{var}\sim7\%$, Table \ref{tab3}),
but the degree of the linear polarization of the continuum changes as $F_{var}\sim14\%$, that may indicate an additional
continuum component to that in the nucleus.
This is in favor of the results obtained by \cite{ar10}. According to the results of VLBI monitoring
there are the observational
evidence for the connection between the variable optical continuum nucleus of 3C390.3 and the compact
radio emission of the jet on the sub-parsec scale \citep[][]{ar10}. To explain this correlation, \cite{ar10} suggested that
the variable optical continuum emission is generated in the innermost part of the jet.
{ The polarization in the 3C 390.3 continuum probably has three sources:
(i) We expect to have
the polarization from the disc (light scattering in a plane-parallel disc atmosphere), with perpendicular
polarization angle to the jet direction (and almost constant);
(ii) Scattering on the inner part of the torus, where the vector of polarization is aligned
with the radio-jet (without fast variability in the continuum polarization); and (iii)
The synchrotron continuum emission of the jet
(whose polarization vector is approximately
perpendicular to the jet direction) that probably contributes to the variability in the continuum polarized light.}
This also can
be seen in Fig. \ref{fig5} where the variability in the continuum (open circles) has arc-like structure with
respect to the jet direction (denoted with an arrow). The jet angle is taken from \cite{al88,al96}.
{ The observed decrease of the polarization in the continuum as a function of wavelength
may be caused by the Rayleigh-scattering on the torus or the inverse Compton
effect \citep[][]{bs87} in the relativistic plasma
jet. However, a detailed study of this effect requires modeling and it is out of the scope of this article.}
Another possible explanation for decreasing polarization may be wavelength
dependent dilution by the host galaxy continuum, which will be redder than the
AGN. However as we noted above the contribution of the host galaxy to the unpolarized
flux is $\sim$ 20\%, and the estimated contribution to the
polarization is around 0.25\%. This cannot cause the observed changes in the optical
polarization, which are about 1\%.
\begin{figure}
\centering
\includegraphics[width=8.3cm]{fig6.ps}
\caption{ From top to bottom: the broad H$\alpha$ line, polarized H$\alpha$ and polarization parameter $U$.
Faint dashed lines represent observations from different epochs and solid bold lines represent averaged profiles
of the unpolarized and polarized
H$\alpha$ and parameter $U$} \label{fig6}
\end{figure}
\subsection{Polarized H$\alpha$ }
There are several models proposed to explain the two-peak structure of the broad Balmer emission lines in 3C 390.3.
A widely accepted model is the formation of the broad lines in a relativistic disc. However, there is some contradictions
between these models and the polarization observations of 3C 390.3.
\cite{sm04,sm05} proposed { two mechanisms of polarization in the optical
continuum and H$\alpha$ broad emission lines in Sy1 nuclei --
equatorial (scattering on the inner part of the torus) and polar (scattering on the jet) polarization mechanisms.}
Both scattering components should be present in all
broad line AGNs, and their polarization properties can be broadly understood in terms of an orientation effect
\citep[Unified model, see]
[]{an96}.
{ In the case of equatorial scattering model, the broad emission lines which are emitted from a rotating disc (or outer regions of
accretion disc-BLR) have to show a double-peaked structure in the polarized light, similar as in the unpolarized light.
The degree of polarization will
be maximal in the broad line wings and a minimum is expected in the line center.}
The angle $\varphi$ of plane polarization is aligned with the projected disc rotation axis and hence with the
radio source axis.
In the case of 3C 390.3, we observed an unexpected difference between the unpolarized and polarized broad
line profile of H$\alpha$: the unpolarized H$\alpha$ line profile has
double-peaked structure with the blue peak at $V_r\sim-3500$ km s$^{-1}$ and red one at $V_r\sim+5000$ km s$^{-1}$.
But the polarized H$\alpha$ line profile is single-peaked and shifted to blue at $V_r\sim-1200$ km s$^{-1}$ with respect to the
narrow component (or systematic velocity). Also, the degree of the linear polarization
of H$\alpha$ is small (around 0.1\% -- 0.2\%, close to zero) and has a
box-like shape without any significant
structural details, i.e. it seems that the broad H$\alpha$ line is almost completely depolarized.
This is contrary to the equatorial
scatter model. On the other hand in 3C 390.3 polarization angle ($\varphi$) is parallel to the
pc-jet axis as it is expected in the equatorial scatter model (see Fig. \ref{fig5} - full circles).
\begin{figure}
\centering
\includegraphics[width=8.3cm]{Fig7.eps}
\caption{Sketch illustrating a possible central region of 3C 390.3. { Two-component BLR:
the double-peaked lines originate in the disc-like region, and depolarization is by warm clouds in an outflowing part of the BLR.}
The part of the variable polarized continuum seems to be emitted from the jet. } \label{fig7}
\end{figure}
It should be noted that \cite{co00} considered a model of 3C 390.3 in which H$\alpha$ photons emitted by a biconical
flow are scattered by the inner wall of the torus. Thus, the scattering plane is perpendicular to the radio jet (e.g.
E-vector parallel to pc-jet axis) and produces a single-peaked scattered H$\alpha$ line profile. However,
this model does not agree with the optical monitoring data for the BLR, where the double-peaked broad emission line is formed.
The CCF-analysis for the H$\beta$ and H$\alpha$ line wings in 3C 390 shows that there is no
a significant delay in the variation of the line wings with respect to the central part of the line, or relative to each other.
The flux in the H$\beta$ line wings and core also varied simultaneously. This result indicates a
dominant circular motion in the BLR and it is in favor of a model in which the main contribution to the broad line fluxes
originates in the accretion disc and not from the biconical flow \citep[][]{di98,di12,sh10,po11,zh11,zh13}.
\subsubsection{Polarized profile of the broad H$\alpha$ line}
In Fig. \ref{fig6} we present the main unpolarized (first panel) and polarized (second panel) broad H$\alpha$ shapes.
The polarized and unpolarized lines are normalized on their intensities
in order to compare the profile variability.
As it can be seen in Fig. \ref{fig6} there is no big change in the polarized and unpolarized line profile. The unpolarized H$\alpha$
line has a double-peaked profile,
the peaks are located (within the error-bars) as earlier measured \citep[see e.g.][]{po11}:
the blue peak is located at
V$_b \sim -3500 $ km s$^{-1}$ while
the red peak had V$_r\sim 5000$ km s$^{-1}$ for H$\alpha$. Contrary, the polarized H$\alpha$ is single-peaked and
shifted to the blue for about -1200 km s$^{-1}$.
Also, in the third
panel of Fig. \ref{fig6} we present the Stokes $U$ parameter. It is interesting that the $U$ parameter has a box-like shape,
that indicates presence
of a depolarization region, i.e. presence of the warm gas between an observer and accretion disc.
The Stokes parameters for the broad line are present in Fig. \ref{fig5} as black full circles.
It is interesting that, they are following the jet direction. It means that
there is a slight polarization from the disc, but in contrast to Mrk 6 \citep[see][]{af14},
the effect of the equatorial scattering cannot be seen.
\subsection{Polarization - Connection between disc and jet}
To speculate about the possible model that can explain the unpolarized and polarized H$\alpha$ line profile and the continuum,
let us recall several observational facts:
(i) 3C390.3 is a powerful radio source with two lobes, FRII radio galaxy with relatively strong compact core.
There are two extended lobes in position angle $PA\sim144^{\circ}$, separated by $\sim223''$ each with a hot spot at the
end \citep[][]{lp91}. Also, there is a weak well collimated thin jet in $PA\approx -37^{\circ}$, linking the core with the northern
lobe \citep[][]{lp95}. In this parsec-jet the VLBI observations at 5 GHz detected superluminal motion (with $v/c\sim4$)
\citep[][]{al88,al96}.
(ii) It seems that a part of the optical continuum is coming from the jet since the optical continuum and radio emission from a
jet-structure are correlated \citep[][]{ar07,ar10}
(iii) The unpolarized H$\alpha$ shape and its variability are in agreement with the disc emission model \citep{sh10,po11}.
However, there is a central component shifted to the red
from 300 km s$^{-1}$ to 800 km s$^{-1}$ and additionally the broad double-peaked H$\alpha$
line shows a blue shift, indicating a wind at the
disc surface with velocities from -300 km s$^{-1}$ to -800 km s$^{-1}$ \citep[][]{po11}.
Contrary, the polarized H$\alpha$ broad line profile
is single-peaked with the blue shift of 1200 km s$^{-1}$, that is not in the favor of the disc
model \citep[][]{sm04,sm05}, but seems to agree with two-side outflow model \citep{co98,co00}
To explain all these observational facts, one can take into account a complex BLR model (see Fig. \ref{fig7}), where
the disc-like BLR is covered by { an outflowing region (the wind with speed around -1200 km s$^{-1}$)
-- with a number of warm clouds which also can emit the H$\alpha$ line. This region can
depolarize the disc-like emission. The two-component BLR (disc covered by an outflowing region of
randomly distributed clouds) seems to be
a good approximation for nearby AGN \citep[][]{kz13}, as well as for a number
of single-peaked AGN \citep[see in more details][]{bo09}.}
One can expect that { the outflowing region with warm gas} can contribute to the additional
electron scattering, which may actually cause the
line emission to be polarized. However, in the outflowing region, the polarization vector is parallel to the disc plane,
and practically
depolarizes the polarized light from equatorial scattering.
{ In a scenario of the two-component model
\citep[disc+outflowing components][]{po11}, the blue-shifted H$\alpha$ component is
coming from the outflowing part of the BLR, and cannot be (self)depolarized, while the disc-like component is
depolarized by this region.
Moreover, the two-component BLR model is able to explain both,
the double-peaked unpolarized emission and depolarization of the broad H$\alpha$ line
(box-like structure in the polarized line profile, see Fig. \ref{fig1}).
Alternatively, the hot depolarization gas may have an inflowing component, that is in contradiction
with results obtained in the long term profile line analysis \citep[see][]{po11}.}
The part of the continuum is originating in this jet like part (outflow) and it may be the reason for
the wavelength dependent polarization in the continuum, and relatively high level of the
continuum polarization \citep[around 2\%, in comparison with Sy1 that is around 1\%, see][]{af14}.
\section{Conclusion}
We presented results of the spectropolarimetric monitoring of 3C390.3, obtained at 24 epochs in
the period of 2009-2014. The galaxy has been observed
with 6-m telescope of SAO observatory using SCORPIO-2 instrument with
spectral resolution of 7-10 \AA\, in the spectral range between 4000 \AA\, and 8000 \AA\,.
Also, we observed and estimated the contribution of the ISM polarization to the observed 3C 390.3
polarization. We measured polarization parameters for the continuum at 5100\AA\, (rest wavelength) and H$\alpha$
line and their variabilities, and explored the lag between the unpolarized and polarized flux.
On the basis of our investigation we can outline the following conclusions:
i) During the 5-year monitoring period we found variation in the broad line and continuum polarization parameters.
We found a lag of $10-40$ days between the polarized
and unpolarized continuum flux variation at 5100\AA\,. This lag is significantly smaller than one we
found for the BLR (lag for H$\beta$ is $60-80$ days and for H$\alpha$ is $140-190$ days) i.e.
the scattering region of the continuum probably is not the BLR (accretion disc).
It seems that the polarized continuum has a component which is coming from the disc (E vector is orientated in
the jet direction), and another, that contributes to variability
that may be synchrotron contribution to the continuum from a pc-jet component.
ii) The unpolarized double-peaked broad emission H$\beta$ and H$\alpha$ lines are observed as single peaked polarized lines.
The polarized line
is shifted to the blue for about -1200 km s$^{-1}$, and such profile was not expected in the case of the equatorial scattering.
Additionally, we observed a box-like profile
of the line polarization (that has a small value $\sim$0.1\%), that indicates depolarization by the warm gas.
Taking into account the results from
the modeling of variability, we proposed that depolarization
region is { an outflowing region} located above the disc-like emission region (that
emits double-peaked lines) and plays a role in the disc line depolarization.
\section*{Acknowledgments}
The results of observations were obtained with the 6-m BTA telescope
of the Special Astrophysical Observatory of Academy of Sciences,
operating with the financial support of the Ministry of Education
and Science of Russian Federation (state contracts no.
16.552.11.7028, 16.518.11.7073).The authors also express appreciation
to the Large Telescope Program Committee of the RAS for
the possibility of implementing the program of Spectropolarimetric
observations at the BTA. This work was supported by the Russian
Foundation for Basic Research (project N12-02-00857) and
the Ministry of Education, Science and Technological Development (Republic of Serbia) through
the project Astrophysical Spectroscopy of Extragalactic Objects
(176001). We thank to M. Gabdeeev for help in spectropolarometric observations.
L.\v C. Popovi\'c thanks to the
COST Action MP1104 'Polarization as a tool to study the Solar
System and beyond' for support. { We thank to an anonymous referee for very useful comments.}
|
\section{Introduction and summary}
QCD displays an interesting phase structure as the number of flavors $N_f$ and number of colors $N_c$ are varied.
It is natural to discuss the phase diagram in
the Veneziano limit~\cite{veneziano,vu1}:
\begin{equation} \label{Vlimit}
N_c \to \infty\ , \quad N_f \to \infty \ , \quad x \equiv \frac{N_f}{N_c}\ \ \mathrm{fixed} \ , \quad g^2 N_c \ \ \mathrm{fixed}\ ,
\end{equation}
as a function of the variable $x$ which has become continuous in this limit. The ``standard'' expectation for the diagram is shown in Fig.~\ref{fig:xphases} (at zero temperature and quark mass). We restrict to the interval $0<x< 11/2 \equiv x_\mathrm{BZ}$ where the theory is asymptotically free. Various regimes can be identified:
\begin{itemize}
\item The QCD regime $0<x<x_c$ where the infrared (IR) dynamics is similar to ordinary QCD (having $N_c=3$ and a few light quarks), with confinement and chiral symmetry breaking.
\item The walking regime with $0<x<x_c$ and $x_c-x \ll 1$ where the coupling constant of the theory varies very slowly, i.e., ``walks'', over a large range of energies.
\item The conformal window $x_c<x<x_\mathrm{BZ}$ where the theory runs to an IR fixed point (IRFP), and chiral symmetry is intact.
\end{itemize}
The existence of the conformal window is solid in the Banks-Zaks (BZ) limit $x\to x_\mathrm{BZ}$, because the value of the coupling is parametrically small and perturbation theory is trustable~\cite{bankszaks}. It is also credible that dynamics at small $x \lesssim 1$ is similar to ordinary QCD -- the $\sim 1/N_c^2$ corrections arising in the Veneziano limit are not expected to change the picture qualitatively. But the nature of the ``conformal transition'' at $x=x_c$, and the behavior of the theories near the transition, is an open question.
The phase diagram of Fig.~\ref{fig:xphases} with a walking regime is obtained in the Dyson-Schwinger approach~\cite{ds}. The transition is then of the BKT type~\cite{bkt}, associated with the so-called Miransky scaling law~\cite{miransky}. The existence of the walking regime is important, since theories in this region may have properties, which are desirable for technicolor candidates~\cite{holdom,walk2} (see also the reviews~\cite{review}). The location of the transition has been estimated by using different approaches~\cite{cw}. However, it has also been suggested that the transition is discontinuous, such that the dynamics ``jumps'' and walking is absent~\cite{Sannino:2012wy}.
There is an ongoing effort to clarify these issues by using first-principles lattice simulations~\cite{lattice,Lombardo:2014mda}, but as it turns out, obtaining reliable results in the transition region is difficult.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.65\textwidth]{141125_pd.pdf}
\end{center}
\caption{The phases of QCD in the Veneziano limit (at zero temperature and quark mass) as a function of $x=N_f/N_c$.}
\label{fig:xphases}\end{figure}
The holographic V-QCD models~\cite{jk} also have the phase diagram of Fig.~\ref{fig:xphases}, including a BKT transition an walking. The ``V'' in V-QCD refers to the Veneziano limit of~\eqref{Vlimit}, where the models are defined. V-QCD is based on two building blocks, the improved holographic QCD (IHQCD)~\cite{ihqcd} for the gluon sector, and a method for adding flavor by inserting a pair of space filling $D4-\overline{D4}$ branes~\cite{Bigazzi:2005md,ckp}. This means that the gluon sector is described by a five dimensional Einstein-dilaton gravity, and the flavor sector is described by a tachyonic Dirac-Born-Infeld action. The sectors are fully backreacted in the Veneziano limit. The structure of V-QCD at finite temperature and chemical potential has been studied in~\cite{alho,Alho:2013hsa,Iatrakis:2014txa,pionloop}. Two-point correlators and bound state masses were analyzed in~\cite{letter,Arean:2013tja}.
There have also been numerous other holographic models addressing the phenomena expected in QCD at finite values of $x$. Walking gauge theories have been modeled by using as a starting point the traditional bottom-up models for example in~\cite{butc,dietrichholo,Sparam}. Walking within the top-down framework has been found and studied in~\cite{nunez}. The conformal transition~\cite{son} was studied in a top-down setup in~\cite{kutasov}, and by using a tachyon-Dirac-Born-Infeld action in~\cite{kutasovdbi}. IHQCD has been used to study walking and IR conformal theories by modifying the holographic RG flow ``by hand'', without inclusion of dynamical fermionic degrees of freedom \cite{Jarvinen:2009fe,Alanen:2011hh}. Walking dynamics and the conformal transition have also been studied in Dynamic AdS/QCD~\cite{Alvares:2012kr,Evans:2014nfa}, which has partially similar ingredients as V-QCD.
Including effects due to finite quark masses in holographic models for QCD is important for various reasons. First of all ordinary QCD, which describes the strong interactions observed in nature, has finite quark masses. Therefore, proper understanding of the dependence of the model on the quark masses should lead to better models for ordinary QCD.
Knowledge on the dependence of quark masses is also useful when analyzing the data from lattice simulations, which are often carried out at unphysically large quark masses due to technical reasons, and therefore extrapolation in the quark mass is needed. For lattice studies which aim at uncovering the phase diagram of QCD as a function of $N_f$, extrapolations in (typically flavor independent) quark mass are particularly important in the probably most interesting region near the conformal transition where lattice simulations are demanding. For example, the so-called ``hyperscaling relations''~\cite{DelDebbio:2010ze} in the conformal window have turned out to be useful close to the conformal transition~\cite{Lombardo:2014pda}. Holography may further help to analyze the physics of the transition, because it can provide a unified description of the phases in the vicinity of $x=x_c$ where most important observables can be computed rather easily for all values of the quark mass.
Finally, it is also important to ensure that the behavior of the holographic model is realistic for all values of the quark mass. Studying limiting behavior (such as the limit of large quark mass) may lead to nontrivial constraints on the model. Fixing the behavior of the model to agree with QCD in limiting cases where field theory computations are tractable, is also expected to improve the model for all values of the quark mass and other parameters.
In this article we study in detail the dependence of holographic QCD on a flavor-independent quark mass $m_q$ in the Veneziano limit. We argue that many of the basic results, such as the dependence of the energy scales on $x$ and $m_q$, are essentially universal, i.e., independent of the details of the holographic model (given some natural assumptions which will be specified below). Additional scaling results for meson and glueball masses, the decay constants, the chiral condensate, the S-parameter, and the critical temperatures are derived for V-QCD analytically at large and small $m_q$ and in the different regimes for $x$. Explicit results are computed numerically in V-QCD and are shown to agree with the expectations from the analytic studies.
The numerical analysis of~\cite{jk,alho,Arean:2013tja} was mostly done at $m_q=0$, but some observables were already computed for a few values of $m_q$ and for a limited range of $x$. Here we extend these results to cover all relevant regimes (with $0<x<x_\mathrm{BZ}$) on the $(x,m_q)$-plane. Importantly, this extension does not require tuning the models or adding any new terms in the V-QCD action, as the value of the quark mass is determined through the boundary conditions of the model. Actually, we will use the choice for the V-QCD action with various potential terms in the action defined exactly as in~\cite{Arean:2013tja}.
We carry out a particularly detailed analysis of the dependence of the chiral condensate on $m_q$, again comparing analytic formulas with explicit numerical results in V-QCD. It is then shown how this analysis is used to study the deformation of QCD by a four-fermion operator $\sim (\bar qq)^2$. Understanding the effect of four-fermion operators is interesting for purely theoretical reasons, but also because such terms naturally arise in technicolor models due to the so-called extended technicolor interactions~\cite{review}.
This article is organized as follows. In the remaining part of introduction we summarize our main results, and discuss the status and future of the exploration of V-QCD. In Sec.~\ref{sec:vqcd}, we give a brief review of the V-QCD models. In Sec.~\ref{sec:scaling}, we discuss the universal scaling results for the energy scales of (holographic) QCD in the Veneziano limit, without explicitly referring to V-QCD. These results are confirmed numerically for V-QCD in Sec.~\ref{sec:massscales}, and extended to include the $m_q$-dependence of the bound state masses and decay constants. In Sec.~\ref{sec:condensate}, we analyze the mass dependence of the chiral condensate, and prove the Gell-Mann-Oakes-Renner (GOR) relation for the pion masses for the fully backreacted case. The results of Sec.~\ref{sec:condensate} are then used to study four-fermion deformations in Sec.~\ref{sec:4f}. The mass dependence of the S-parameter and the pion decay constant is analyzed in Sec.~\ref{sec:S}. Finally, a finite temperature is
switched on in Sec.~\ref{sec:ft}, where the scaling laws for critical temperatures are derived and compared to numerical results. The Appendix provides technical details for the derivation of the results.
The article is long but many of its sections are largely independent, so a reader only interested in a specific topic may want to jump directly to the corresponding section. There are, however, exceptions: Sec.~\ref{sec:massscales} requires studying Sec.~\ref{sec:scaling} first, and Sec.~\ref{sec:4f} requires Sec.~\ref{sec:condensate}.
\subsection{Summary of results}
Let us then summarize the main results of this article. We identify three different regimes on the ($x$,$m_q$)-plane, shown schematically in Fig.~\ref{fig:scalingregions}, where the dependence of the model on the quark mass is qualitatively different. In regime~A, the quark mass is a small perturbation. Regime~B is the ``hyperscaling'' region where the couplings constant walks, and the amount of walking is controlled by the quark mass\footnote{In this article the term ``walking regime'' always refers to the region with $x_c-x \ll 1$ where walking is found at zero quark mass, even though the coupling constant also walks in regime~B.}. Regime~C refers to the limit of large quark mass (in units of the scale of the ultraviolet (UV) RG flow $\Lambda_\mathrm{UV}$). The white regions in the plot indicate crossovers between the regimes~A, B, and~C -- there are no phase transitions at finite quark mass and zero temperature.
In this article, we discuss in detail how the diagram arises quite in general in holographic\footnote{As such, the structure of Fig.~\ref{fig:scalingregions} is not surprising, and one can argue how it arises from QCD by using arguments directly based on field theory~\cite{DelDebbio:2010ze,Dietrich:2009ns}.} models. We show how scaling results for the energy and mass scales in the various regimes can be obtained by using relatively simple assumptions and straightforward analysis (explicit results are given in Eqs.~\eqref{regimeBlowx}, \eqref{regimeBhighx}, and~\eqref{largemqscalingt}).
The most important findings are:
\begin{itemize}
\item The physics at small $m_q$, including regimes~A and~B of Fig.~\ref{fig:scalingregions}, is universal, i.e., qualitative features are independent of the details of the holographic model.
That is, the results are proven for all V-QCD models with such ``regular'' potentials that there is a BKT transition and therefore the model displays the structure of Fig.~\ref{fig:scalingregions},
but they also hold in other models which involve a BKT transition triggered by the same mechanism as in V-QCD (details on this mechanism are given in Sec.~\ref{sec:scaling} and in Appendix~\ref{app:scaling}).
\item At large $m_q$ (in regime~C) the results are model dependent, and can directly be compared to field theory results which are tractable in the limit of large $m_q$. We demonstrate that V-QCD reproduces most important features such as the decoupling of the massive quarks. Remarkably, we find that a specific subclass of the V-QCD models, which was found to be closest to QCD by analyzing the asymptotic meson spectra in~\cite{Arean:2013tja}, also works best at large quark mass. These are the models with ``potentials I'' below.
\end{itemize}
We compute numerically the meson and glueball masses in V-QCD, and show that they agree with the generic scaling results in all regimes.
Notice that Fig.~\ref{fig:scalingregions} only shows changes in the dependence of the quark mass. There are also other features: In regime~A, the theories close to $x=x_c$ are walking and therefore much different from the theories at low values of $x$, reflecting the behavior of the $m_q=0$ backgrounds, which are perturbed by the quark mass. Within regime~B, the scaling laws change in a nonanalytic manner exactly at $x=x_c$ (see Eqs.~\eqref{regimeBlowx} and~\eqref{regimeBhighx}).
We also analyze in detail how the chiral condensate depends on the quark mass. The main results are the following:
\begin{itemize}
\item In the QCD and walking regimes, there is an interesting spiral structure, which we call the Efimov spiral. We give an asymptotic formula~\cite{Iqbal:2011in} of the spiral at small $m_q$ in Eq.~\eqref{spiraleqs}, and compare this to data in Fig.~\ref{fig:spiral}. We demonstrate how the spiral is related to the subleading Efimov vacua\footnote{The terminology refers to the analogous Efimov effect in the formation of three-body bound states of identical bosons~\cite{Efimov:1970zz}.} of the theory and to the Miransky scaling law in the walking regime.
\item The standard holographic proof of the GOR relation is extended to the case of V-QCD, which requires handling of the full backreaction and the logarithmic corrections which appear in near the UV boundary due to the RG flow. The relation is given in Eq.~\eqref{GORtext} and checked numerically in Fig.~\ref{fig:GOR}.
\end{itemize}
We use the results for the chiral condensate to check how Witten's method~\cite{Witten:2001ua} for adding multi-trace deformations in holographic models by modifying the UV boundary conditions works for V-QCD. The resulting phase diagram for the case of a double-trace (four-fermion) deformation $\propto g_2 (\bar q q)^2$ at zero quark mass is shown in Fig.~\ref{fig:4fphases} in the $(x,g_2)$-plane. For positive coupling $g_2$, we see no change with respect to $g_2=0$, whereas negative $g_2$ induces an instability.
The mass dependence of the S-parameter, the pion decay constant, and related quantities is studied in detail.
\begin{itemize}
\item As any finite $m_q$ is turned on in the conformal window, the S-parameter discontinuously jumps from zero to a \order{N_f N_c} number. Except for this discontinuity, the dependence on $x$ and $m_q$ is weak.
\item We demonstrate a novel power-law scaling of the subleading terms in the S-parameter at small quark mass in the conformal window and in the walking regimes (see Figs.~\ref{fig:massdeplog} and~\ref{fig:massdepwalk}). It is argued that the power can be expressed in terms of the dimension of $\mathrm{Tr}\, F^2$ at the IRFP as in Eqs.~\eqref{massdepCW} and~\eqref{massdepwalking}.
\item We analyze the $x$-dependence of the S-parameter by writing it as a series over the contributions from the low-lying vector and axial vector meson poles (see Eq.~\eqref{Sseries}). It is shown that the increase of the S-parameter with $x$ in the QCD and walking regimes can be attributed to slower convergence of the series.
\end{itemize}
Finally we analyze the dependence of the critical temperatures of the deconfinement transition and various crossovers on $m_q$ and $x$. The scaling laws of \eqref{TcinregA} -- \eqref{TcinregC} and \eqref{Tqclaw} -- \eqref{Tmqlaw} are demonstrated numerically in Fig.~\ref{fig:FT}. The second order chiral transition, which is found at zero quark mass in the walking regime~\cite{alho}, transforms into a crossover as finite $m_q$ is turned on. We also discuss how the thermodynamics approaches that of the YM limit ($x \to 0$) as the quark mass is taken to infinity so that the quarks are decoupled.
\subsection{Outlook}
This article is part of an ongoing program~\cite{jk,alho,letter,Arean:2013tja,Alho:2013hsa,Iatrakis:2014txa,pionloop} for studying the properties of the V-QCD models, and in more general the structure of QCD in the Veneziano limit. There are several possible future directions to explore. As a continuation of this study, one could consider the case of flavor dependent quark masses, which would be interesting in order to construct more realistic models for ordinary QCD, where all quark masses are unequal. In holography, this means that the background solutions are nontrivial dependence on the flavor indices, and consequently are described in terms of a non-Abelian Dirac-Born-Infeld action. The precise definition of such a non-Abelian action is not known, so requiring the physics to be correct might lead to interesting constraints for it.
The study of the CP-odd terms of the V-QCD action is in progress at the moment~\cite{cpodd}. These terms govern the physics of the axial anomaly and the theta angle of QCD. Similarly to the analysis at zero values of theta, requiring the regularity of the solutions in the IR, and among other things the correct behavior of the asymptotic meson spectra at finite values of theta result in constraints for the potentials appearing in the CP-odd terms, in addition to the constraints obtained in the YM limit~\cite{ihqcd,data,cs}. After analyzing these constraints, we may make physically reasonable choices for these potentials, and in particular for their asymptotic behavior near the UV boundary and deep in the IR.
So far the potentials of the V-QCD have not been tuned to fit any nonperturbative QCD data, and therefore all predictions of the model are qualitative. But once the CP-odd sector of the action has been fixed, one can start fitting the parameters of the potentials both to experimental results for QCD (such as meson masses) and to lattice data (both for QCD at finite $N_f/N_c$~\cite{lattice} and for YM~\cite{panero,Lucini:2012gg}). The hope is that the overall fit, together with the other constraints for the potentials, fixes the predictions of the model to a good accuracy for all relevant values of the parameters (such as $x$, the quark mass, temperature, and chemical potential). Consequently the model would give a effective description of QCD with real predictive power, rather than being just a toy model.
\section{V-QCD} \label{sec:vqcd}
In this section, we will briefly introduce a class of bottom-up models for QCD, which we call V-QCD~\cite{jk}. The V in the name refers to the fact that the models are defined in the Veneziano limit:
\begin{equation}
N_c \to \infty\quad \mathrm{and} \quad N_f \to \infty\ , \quad \mathrm{with}\quad x\equiv \frac{N_f}{N_c} \quad \mathrm{and} \quad g^2 N_c \quad \mathrm{fixed} \ .
\end{equation}
V-QCD is based on two ``building blocks''. The first block (the glue sector of V-QCD) is IHQCD~\cite{ihqcd} which is a bottom-up model for Yang-Mills (YM) theory inspired by five-dimensional noncritical string theory. The second block (the flavor sector of V-QCD) is a framework for adding flavor via tachyonic Dirac-Born-Infeld actions~\cite{Bigazzi:2005md,ckp}. This framework has been tested previously~\cite{ikp} in the probe (or 't Hooft) limit, i.e., without including the backreaction of the flavor branes to the background. However in V-QCD, and more generally in the Veneziano limit, the flavor and glue sectors are fully backreacted.
\subsection{The V-QCD action}
Let us then discuss briefly the dictionary of V-QCD. The most relevant fields are
\begin{itemize}
\item The dilaton $\phi$. The exponential $e^{-\phi}$ is dual to the operator $\ensuremath{\mathrm{Tr}} F^2$. We will denote $\lambda=e^\phi$ below. As this notation indicates, its background value is identified as the 't Hooft coupling on the field theory side.
\item The tachyon $T_{ij}$ which is a $N_f \times N_f$ matrix in flavor space. The combination $T+T^\dagger$ is dual to the operator $\bar q_i q_j$ whereas $T-T^\dagger$ is dual to $\bar q_i \gamma_5 q_j$. For the background solutions considered here we will take $T_{ij} = \tau(r) \delta_{ij}$ so the flavor structure does not appear explicitly.
\item The left- and right-handed gauge fields $A^{L/R}_\mu$ which are dual to $\bar q \gamma_\mu(1\pm\gamma_5) q$. They are also matrices in flavor space but we have hidden the flavor indices here. These fields evaluate to zero for the backgrounds considered in this article.
\end{itemize}
In addition, the scale factor $A$ of the metric
\begin{equation}
ds^2=e^{2 A(r)} (-dt^2/f(r) +d\mathbf{x}^2+f(r) dr^2)\,
\label{bame}
\end{equation}
of the vacuum solution is identified as the logarithm of the energy scale on the field theory side. The blackening factor $f$ in the metric may be either identically equal to one or a nontrivial function of $r$.
Our convention will be that the UV boundary lies at $r=0$, and the bulk coordinate therefore runs from zero to infinity (when $f(r)\equiv 1$) or up to a horizon at a finite value of $r$ (when $f(r)$ has a nontrivial profile). The metric will be close to the AdS metric in the UV: $A \sim -\log(r/\ell)$, where $\ell$ is the (UV) AdS radius. In the UV, $r$ is therefore identified roughly as the inverse of the energy scale of the dual field theory.
The action for the V-QCD model consists of three terms:
\begin{equation}
S = S_g + S_f + S_a
\end{equation}
where $S_g$, $S_f$, and $S_a$ are the actions for the glue, flavor and CP-odd sectors, respectively. Explicit expressions for the first two terms will be given below. As discussed in \cite{jk}, only these terms contribute in the vacuum structure of the theory if the phases of the quark mass matrix and the theta angle vanish. The last term $S_a$ is important for the realization of the theta angle and the axial anomaly of QCD~\cite{ckp}. This term has been written down explicitly in~\cite{Arean:2013tja}, and it will be zero for all configurations discussed in this article.
The glue action is that of IHQCD~\cite{ihqcd}. It includes five dimensional Einstein gravity and the dilaton $\lambda=e^{\phi}$:
\begin{equation}
S_g= M^3 N_c^2 \int d^5x \ \sqrt{-\det g}\left(R-{4\over3}{
(\partial\lambda)^2\over\lambda^2}+V_g(\lambda)\right) \ .
\label{vg}\end{equation}
The flavor action is the generalized Sen's action~\cite{ckp,sen} (see also~\cite{sstachyon}),
\begin{equation}
S_f= - \frac{1}{2} M^3 N_c\ {\mathbb Tr} \int d^5x\,
\left(V_f(\lambda,T^\dagger T)\sqrt{-\det {\bf A}_L}+V_f(\lambda, TT^\dagger)\sqrt{-\det {\bf A}_R}\right)\ ,
\label{generalact}
\end{equation}
where the quantities inside the square roots are defined as
\begin{align}
{\bf A}_{L\,MN} &=g_{MN} + w(\lambda,T) F^{(L)}_{MN}
+ {\kappa(\lambda,T) \over 2 } \left[(D_M T)^\dagger (D_N T)+
(D_N T)^\dagger (D_M T)\right] \ ,\nonumber\\
{\bf A}_{R\,MN} &=g_{MN} + w(\lambda,T) F^{(R)}_{MN}
+ {\kappa(\lambda,T) \over 2 } \left[(D_M T) (D_N T)^\dagger+
(D_N T) (D_M T)^\dagger\right] \ ,
\label{Senaction}
\end{align}
with the covariant derivative
\begin{equation}
D_M T = \partial_M T + i T A_M^L- i A_M^R T\ .
\end{equation}
The trace $\mathbb Tr$ is over the flavor indices -- recall that the fields $A_{L}$, $A_{R}$ as well as $T$ are $N_f \times N_f$ matrices in the flavor space.
It is not known, in general, how the determinants over the Lorentz indices in~\eqref{generalact} should be defined when the arguments~\eqref{Senaction} contain non-Abelian matrices in flavor space. However, for our purposes such definition is not required: our background solution will be proportional to the unit matrix $\mathbb{I}_{N_f}$, as the quarks will be all massless or all have the same mass $m_q$.
In such a case, the fluctuations of the Lagrangian are unambiguous up to quadratic order.
The form of the tachyon
potential that we will use for the derivation of the spectra is
\begin{equation}
V_f(\lambda,TT^\dagger)=V_{f0}(\lambda) e^{- a(\lambda) T T^\dagger} \ .
\label{tachpot}
\end{equation}
This is the string theory tachyon potential where the constants have been allowed to depend on the dilaton $\lambda$.
For the vacuum solutions (with flavor independent quark mass) we will take $T =\tau(r) \mathbb{I}_{N_f}$ where $\tau(r)$ is real, so that
\begin{equation}
V_f(\lambda,T)=V_{f0}(\lambda) e^{- a(\lambda)\tau^2} \ .
\label{vf}\end{equation}
The coupling functions $\kappa(\lambda,T)$ and $w(\lambda,T)$ are allowed in general to depend on $T$, through such combinations that the expressions~\eqref{Senaction} transform covariantly under flavor symmetry. In this article, we will take them to be independent of $T$, emulating the known string theory results. Under these assumptions, and for the vacuum solutions (so that the gauge fields also vanish) the flavor action simplifies to
\begin{equation}
S_f= - M^3 N_cN_f\int d^5x\,\sqrt{-\det g}\ V_{f0}(\lambda)\, e^{- a(\lambda)\tau^2} \sqrt{1+g^{rr}\,\kappa(\lambda)\, (\tau')^2} \ .
\end{equation}
\subsection{Potentials and the holographic RG flow}
In order to fully fix the action, the potentials $V_g(\lambda)$, $V_{f0}(\lambda)$, $a(\lambda)$, $\kappa(\lambda)$, and $w(\lambda)$ need to be specified. It turns out~\cite{jk} that $V_g(\lambda)$ must satisfy the same constraints as in IHQCD~\cite{ihqcd}. The other potentials will be subject to analogous constraints. We review the main idea here, and the details can be found in~\cite{jk,Arean:2013tja}.
First, identification of the field $\lambda$ as the 't Hooft coupling and the scale factor $A$ as the logarithm of the field theory energy scale defines the holographic renormalization group (RG) flow and the holographic beta function for the coupling as in IHQCD,
\begin{equation}
\beta_h(\lambda) = \frac{\lambda'(r)}{A'(r)}
\end{equation}
with the understanding that the fields are evaluated on the $r$-dependent background solution. In IHQCD, the dilaton potential $V_g$ can be directly mapped to the holographic beta function~\cite{ihqcd} at any value of $r$, that is, at any energy scale.
In V-QCD, there is an additional field, the tachyon, whose background value is linked to the running quark mass. Therefore one may define a holographic gamma function, which controls the holographic RG flow of the quark mass~\cite{jk}. The mapping between the beta and gamma functions and the potentials is, in general, more complicated than in IHQCD, but simplifies in the UV.
The behavior of the potentials in the UV (where $\lambda \to 0$) is then restricted by requiring that the holographic beta and gamma functions match with their QCD counterparts in the UV. For the UV structure to be consistent, all potentials are chosen to be analytic at $\lambda=0$, and the series coefficients can be related to those of the perturbative beta and gamma functions in QCD. It turns out that the dilaton potential $V_g$ is consequently mapped to the perturbative beta function of YM theory as in IHQCD. Due to the backreaction, the beta function of QCD (in the Veneziano limit) is mapped to the combination $V_g-x V_{f0}$. This mapping leaves one undetermined parameter, the UV normalization of $V_{f0}$, which we call $W_0$. The gamma function of QCD fixes the UV behavior of the ratio $a/\kappa$. Here we will match the expansions of the potentials around $\lambda=0$ up to two loops for the beta function and up to one loop for the gamma function of QCD.
Using perturbative QCD to determine the UV behavior of the potentials may be surprising, since holography is in general not expected to work at small values of the coupling $\lambda$. The idea is, however, that by using this procedure correct boundary conditions for the more interesting IR dynamics are obtained. Notice also that the procedure can be seen as a rather mild generalization of what is usually done in bottom-up holography. For example, the bulk mass of the tachyon is typically required to satisfy the relation $-m^2\ell^2 = \Delta(4-\Delta)$ (at least in the UV), where $\ell$ is the UV AdS radius and $\Delta$ is either the dimension of the quark mass or the chiral condensate. Here this relation is effectively generalized to include loop effects, i.e., the perturbative anomalous dimension of the quark mass, which is roughly mapped to the first few coefficients in the expansion of the bulk mass of the tachyon at $\lambda=0$.
In this article, we will carry out one more check which demonstrates that our UV boundary conditions make sense. Namely, we prove that the GOR relation holds even in the backreacted case and that the UV RG flow of the quark mass and the condensate, imposed by the matching to perturbative QCD, cancels in this relation, as it should (see Sec.~\ref{sec:gmor}).
The choice of the potentials in the IR is more relevant since it affects how the nonperturbative physics of QCD is modeled. Since we are working with bottom-up models, there is a lot of freedom in choosing the potentials, and it is important to choose them such that the IR physics resembles that of QCD. The mapping to the beta functions is not useful at large values of the coupling because of scheme dependence. The asymptotic behavior of the potentials at large $\lambda$ can however be constrained heavily
by comparing to several different observables, most importantly the asymptotics of the spectra at large mass.
The IR behavior of the dilaton potential $V_g$, i.e., its asymptotics as $\lambda \to \infty$, can be fixed by requiring (among other things) confinement and correct asymptotic behavior of the glueball spectrum at large excitation numbers. The remaining parameters have been fitted to YM data~\cite{data}. Similarly, asymptotics of the meson spectra sets strict constraints on the large $\lambda$ asymptotics of the other potentials in the V-QCD action~\cite{Arean:2013tja}. The remaining degrees of freedom will be fitted to experimental and lattice data in future studies.
In this article, we will be using the choices ``potentials~I'' and ``potentials~II'' which are exactly the ones given in~\cite{letter,Arean:2013tja}, and are defined explicitly in Appendix~\ref{app:numerics}. The potentials~I reproduce more accurately qualitative features of QCD. This choice has the UV parameter $W_0$ set to a constant value $3/11$, and assumes $w=\kappa$. The potentials~II are included in order to study the model dependence of the results. This choice has the UV parameter $W_0$ set to a value which guarantees ``automatically'' the Stefan-Boltzmann normalization of pressure at high temperatures~\cite{alho}, and assumes $w=1$. Most of the results derived in this article are qualitative (e.g., scaling laws of various observables) and therefore insensitive to the details of the potentials.
\subsection{Background solutions} \label{sec:bg}
Let us discuss some general features of the background solutions of the V-QCD models, first restricting to the standard case, which has a phase diagram similar to what is usually expected to arise in QCD. Such a phase diagram is obtained if the potentials are chosen as discussed above.
\subsubsection{Zero temperature}
In this article we will mostly discuss solutions at zero temperature. In this case, the blackening factor $f$ in~\eqref{bame} is trivial, $f \equiv 1$.
To find the background, we consider $r$-dependent Ans\"atze for $\lambda$, and $A$. As pointed out above, we assume that the quark mass is flavor independent, and therefore take $T=\tau(r)\mathbb{I}_{N_f}$. We also set all other fields to zero, and look for solutions to the equations of motion (EoMs). The models are expected to have two classes of (zero temperature) vacuum solutions \cite{jk}:
\begin{enumerate}
\item Backgrounds with nontrivial $\lambda(r)$, $A(r)$ and with zero tachyon $\tau(r)=0$. These solutions have zero quark mass and intact chiral symmetry.
\item Backgrounds with nontrivial $\lambda(r)$, $A(r)$ and $\tau(r)$. These solution have broken chiral symmetry. As usual, the quark mass $m_q$ and the chiral condensate are identified as the coefficients of the normalizable and non-normalizable tachyon modes in the UV.
\end{enumerate}
In the first case, the EoMs can be integrated analytically into a single first order equation, which can easily be solved numerically. The regular solution ends on an IRFP, where the dilaton approaches a constant value, and the geometry is asymptotically AdS$_5$. In the second case, one needs to solve a set of coupled differential equations numerically. The regular solution ends in a ``good'' IR singularity~\cite{gubser}, where both the dilaton and the tachyon diverge. This kind of singularity supports extension to finite temperature and is repulsive: perturbations around the regular solution develop a nonanalyticity before reaching the singularity, which signals the fact that IR boundary conditions are uniquely fixed.
Let us first recall what happens at zero quark mass. The ratio $x=N_f/N_c$ is constrained to the range $0 \le x <11/2 \equiv x_\mathrm{BZ}$ where the upper bound was normalized to the Banks-Zaks (BZ) value in QCD, where the leading coefficient of the $\beta$-function turns positive. The standard\footnote{For some choices of potentials, also different structure can appear. In particular, there is the possibility that chiral symmetry breaking is absent in the regime of very small $x$~\cite{alho}.} phase diagram at zero quark mass has two phases separated by a phase transition at some $x=x_c$ within this range.
\begin{itemize}
\item When $x_c\le x<x_\mathrm{BZ}$, chiral symmetry is intact. The dominant vacuum solution is in the first class with the tachyon vanishing identically and an IRFP. Therefore the geometry is asymptotically AdS$_5$ in the IR.
\item When $0<x<x_c$, chiral symmetry is broken. The dominant vacuum therefore is in the second class with nonzero tachyon even though the quark mass is zero.
The geometry ends at a good IR singularity in the IR.
\end{itemize}
For potentials satisfying some reasonable requirements~\cite{jk,Arean:2013tja}, the phase transition at $x=x_c$ is due to an instability of the tachyon at the IRFP~\cite{son,Jensen:2010ga}. That is, the solution with vanishing tachyon is unstable if the bulk mass of the tachyon at the IRFP $ -m_*^2 \ell_*^2$, where $\ell_*$ is the IR AdS radius, violates the Breitenlohner-Freedman (BF) bound~\cite{Breitenlohner:1982jf}:
\begin{equation}
- m_*^2 \ell_*^2 = \Delta_*(4-\Delta_*) \le 4 \ ,
\end{equation}
where $\Delta_*$ is the dimension of the quark mass at the fixed point.
Notice that when the bound is violated, $\Delta_*$ becomes complex.
As a consequence of violating the BF bound, the phase transition at $x=x_c$ (which is only present at zero quark mass) involves BKT \cite{bkt} or Miransky \cite{miransky} scaling, for values of $x$ right below the critical one. The order parameter for the transition, the chiral condensate $\sigma \sim \langle \bar q q \rangle$ vanishes exponentially,
\begin{equation} \label{condscaling}
\sigma \sim \exp\left(-\frac{2 K}{\sqrt{x_c-x}}\right)
\end{equation}
as $x \to x_c$ from below. Here the constant $K$ is positive.
At finite quark mass, the BKT transition disappears: the background is always in the second class with finite tachyon, and the dominant vacua at all values of $x$ are smoothly connected. Therefore the geometry ends in a ``good'' singularity in the IR. In particular, the IR geometry changes in a discontinuous manner (from an IRFP to the good singularity) when a small quark mass is turned on in the conformal window ($x_c<x<x_\mathrm{BZ}$). This discontinuity causes interesting behavior of observables which will be discussed in the following sections.
The models may also have unstable subdominant vacua when $0<x<x_c$. We will discuss such vacua in more detail in Sec.~\ref{sec:condensate}.
\subsubsection{Finite temperature}
At finite temperature, one can first identify two types of background geometries~\cite{alho}:
\begin{enumerate}
\item The thermal gas solutions which have the same (and therefore temperature independent) $r$-dependence as the zero temperature solutions, and in particular $f \equiv 1$.
\item The back hole solutions which have a nontrivial $f(r)$ and end on a horizon in the IR.
\end{enumerate}
As usual the temperature is given by (the inverse of) the length of the compactified time direction, and is equal to the Hawking temperature of the black hole for the second type of solutions. Both of these geometries further split into two classes, one having zero and the other having nonzero tachyon. Therefore there can be up to four qualitatively different competing saddle points.
The finite temperature phase diagram has been studied in~\cite{alho}. The following structure was found at zero quark mass:
\begin{itemize}
\item For $x_c \le x <x_\mathrm{BZ}$, the finite temperature phase is the tachyonless black hole. A tachyonless thermal gas solution which has the same $r$-dependence as the zero temperature solution also exists at all temperatures but is subdominant. Therefore if the system is first prepared at zero temperature, any amount of heating makes the system jump to a different phase immediately. At temperatures which are much smaller than the characteristic scale of the RG flow between the two fixed points, the finite temperature backgrounds are obtained by deforming the zero temperature backgrounds, which are asymptotically AdS$_5$, only very close to the IR end. Consequently, small temperature thermodynamics is that of AdS, $p \propto T^4$, and the zero temperature transition is of 4th order.
\item For $0<x<x_c$, the low temperature phase is the tachyonic thermal gas phase which is smoothly connected to the zero temperature solution and breaks chiral symmetry. The high temperature phase is the tachyonless black hole phase. There is always a first order (``deconfinement'') transition separating the phases, but it is also possible that an intermediate, chirally broken tachyonic black hole phase appears. If this is the case, chiral symmetry is restored at a separate second order transition, which has higher critical temperature than the deconfinement transition. As $x \to x_c$ from below, the critical temperatures go to zero following Miransky scaling.
\end{itemize}
At finite quark mass, all phases are tachyonic due to the UV boundary conditions, and chiral symmetry is broken. At low temperatures the dominant vacuum is the thermal gas phase. When the system is heated it undergoes a first order transition to the black hole phase, which is interpreted as the deconfinement transition (since there is no transition linked to chiral symmetry breaking). The phase structure is therefore similar for all $0<x<x_\mathrm{BZ}$. The dependence of the critical temperature on $m_q$ and $x$ will be discussed in Sec.~\ref{sec:ft}. In addition to the phase transition, we find several crossovers which are linked to changes in the zero temperature geometry.
\section{Energy scales of (holographic) QCD in the Veneziano limit} \label{sec:scaling}
In this section we shall discuss the dependence of observables (at zero temperature) on the quark mass and $x=N_f/N_c$ for holographic models of QCD in the Veneziano limit and in general, i.e., not necessarily only for V-QCD. The results will not be proven rigorously, but we will sketch how they arise from rather natural assumptions. Comparison to quantitative results from V-QCD will be carried out in Sec.~\ref{sec:massscales}.
It is useful to discuss separately the regions with small (possibly zero) and large quark mass.
\begin{itemize}
\item At small quark mass, we shall demonstrate that under certain natural assumptions, a universal picture arises from holography, which is not dependent on the details of the model. That is, the results of this section will apply to the V-QCD models with quite generic choices of potentials but also to holographic models with more general actions, if they satisfy some natural requirements which will be given below.
\item At large quark mass, the $m_q$-dependence of energy scales and some other observables (such as meson mass gaps) can be found via arguments based on QCD. Large quark mass probes the UV structure in holography, where predictions may not be reliable. We shall see that the results from holography are model dependent in this region, and discuss what is needed to match with the known results in QCD.
\end{itemize}
In this section we will concentrate on the behavior of the various energy scales for a generic holographic model. In the remaining of this article we will analyze concrete observables such as bound state masses, the chiral condensate, and the S-parameter for V-QCD. Also the behavior of these observables at small quark mass is to large extent independent of the model details, but it is just much easier to work with the explicitly fixed V-QCD action.
We will exclude the BZ limit (i.e., the limit $x \to x_\mathrm{BZ}$) from our analysis for the moment. In this region the theory is fully under perturbative control and holographic approach may not be useful.
First we need to give rough definitions for various energy scales at zero temperature, assuming a generic holographic model. There must be a field dual to the $\bar q q$ operator, which we call the tachyon, and the geometry must be asymptotically AdS in the UV. We take the UV boundary to lie at $r=0$ as above.
Because we want to keep the discussion generic, the definitions for the energy scales will be rather sketchy. Precise definitions in the case of V-QCD will be given in Sec.~\ref{sec:massscales}.
\begin{itemize}
\item $\Lambda_\mathrm{UV}$ is the scale of the UV RG flow in QCD. In holography, it can be identified as the scale of the UV expansions of the vacuum solution. Since V-QCD also implements the UV RG flow of the coupling constant through the flow of the dilaton field, it will be most natural to define $\Lambda_\mathrm{UV}$ in terms of the UV expansion of the dilaton (see~\eqref{laUV} in the next section). In the generic treatment of this section (which assumes small quark mass, i.e., $m_q/\Lambda_\mathrm{UV} \ll 1$), this scale is most conveniently
defined in terms of the UV behavior of the tachyon instead: it is identified as the (inverse of the) boundary of the interval where the standard UV expression~\eqref{tauruns} for the tachyon holds as a good approximation.
\item $\Lambda_\mathrm{IR}$ is the soft IR scale which governs the IR expansions (or is set by an IR cutoff), in close analogy to the definition of $\Lambda_\mathrm{UV}$ in the UV. Precise definition for V-QCD will be given in~\eqref{laIR} and~\eqref{laIRFP} in Sec.~\ref{sec:massscales}.
\item The quark mass $m_q$ is defined as the source for the tachyon.
\item $\L_\tau$ is the scale of chiral symmetry breaking (whenever it is broken). In holography it can be identified as the energy scale (inverse of $r$) where the tachyon grows large (or becomes $\mathcal{O}(1)$, more precisely). As we shall see, its dependence on $m_q$ is similar to that of the constituent quark mass in nonrelativistic quark models.
\end{itemize}
Notice that we only discuss the dynamical scales appearing through the background solutions, and for example the UV AdS radius is assumed to be fixed. More precisely, the ratios of the various energy scales are determined by the dynamics of the model, whereas one of the energy scales (most conveniently $\Lambda_\mathrm{UV}$) can be taken as a fixed reference scale.
\subsection{Small quark mass}
By small quark mass we mean here that $m_q/\Lambda_\mathrm{UV} \ll 1$, and $m_q$ may also be zero. In this case our results will be essentially universal, i.e., independent of the details of the holographic model. Naturally, several assumptions must be made on the model, in order to ensure that the physics resembles that of QCD in the Veneziano limit.
Most importantly, there must be an IRFP and a conformal window, and the conformal transition at $x=x_c$ is assumed to be a BKT transition, which is naturally implemented in holography by the tachyon hitting its BF bound at the IRFP~\cite{son,Jensen:2010ga} (implemented through a geometry which is asymptotically AdS in the IR).
These assumptions are met by the V-QCD models with potentials which fulfill reasonable constraints in the UV and in the IR, and do not have any peculiar structure at intermediate energy scales~\cite{jk,Arean:2013tja}. But these assumptions can also hold more generally in holographic theories the actions of which cannot be written in the V-QCD form. Indeed similar results which we will
present here have been found in related top-down~\cite{kutasov} bottom-up~\cite{kutasovdbi,Alvares:2012kr,Evans:2014nfa} models. For simplicity we also assume that the upper edge of the conformal window lies at the QCD value, $x=11/2 \equiv x_\mathrm{BZ}$.
The model assumptions can be made explicit in terms of the tachyon background in various ranges of the bulk coordinate $r$. We will only need the solutions for small tachyon, so that the tachyon EoM is approximately linear (see Appendix~\ref{app:scaling} for more detailed derivation of these solutions). In the UV (where $r \to 0$), we assume the standard behavior determined by the dimension of the quark mass:
\begin{equation} \label{tauruns}
\tau \simeq m_q r + \sigma r^3\ , \qquad \left(r \ll \frac{1}{\Lambda_\mathrm{UV}}\right) \ ,
\end{equation}
where $\sigma$ is proportional to the chiral condensate.
Here logarithmic running of the quark mass (and the condensate) could also be included, but it is not important because it will not affect the leading scaling behavior.
In the vicinity of the fixed point when the tachyon is small and the BF bound is satisfied, corresponding to $x_c \le x < x_\mathrm{BZ}$ in QCD, we find that
\begin{equation} \label{taufp}
\tau \simeq C_m (r \Lambda_\mathrm{UV})^{\Delta_*} + C_\sigma (r \Lambda_\mathrm{UV})^{4-\Delta_*}\ , \qquad \left(\frac{1}{\Lambda_\mathrm{UV}} \ll r \ll \frac{1}{\Lambda_\tau}\right)\ ,
\end{equation}
where the precise $x$-dependence of the anomalous dimension of the quark mass $\Delta_*$ depends on the model.
When the BF bound is violated, and $x_c-x$ is small
\begin{equation} \label{tauwalks}
\tau \simeq C_w \left(r \Lambda_\mathrm{UV}\right)^2 \sin\left[\nu \log( r \Lambda_\mathrm{UV}) +\phi\right]\ , \qquad \left(\frac{1}{\Lambda_\mathrm{UV}} \ll r \ll \frac{1}{\Lambda_\tau}\right)\ ,
\end{equation}
where $\nu = \mathrm{Im}\Delta_* \simeq \pi\sqrt{(x_c-x)}/K$ as $x \to x_c$ from below. The coefficient $K$ depends on the model, and it satisfies
\begin{equation}
K = \frac{\pi}{\sqrt{\frac{d}{dx}\left[\Delta_*(\Delta_*-4)\right]_{x=x_c}}} \ .
\end{equation}
Now it is straightforward to fix the integration constants ($\sigma$, $C_m$, $C_\sigma$, $C_w$ and $\phi$) and compute the mass dependence of the various energy scales by using the following recipes:
\begin{itemize}
\item Both the normalizable and nonnormalizable terms of the tachyon solution are separately continuous\footnote{Near the IRFP in~\eqref{tauwalks} the normalizable and nonnormalizable terms cannot be separated, but the strategy is then interpreted as requiring the continuity of the tachyon and its first derivative.}. Therefore an approximate tachyon solution for all $r \ll 1/\Lambda_\tau$ can be found by gluing together the results from the different regimes~\eqref{tauruns},~\eqref{taufp}, and~\eqref{tauwalks}. Because the bulk mass of the tachyon varies smoothly, the terms of the tachyon can only jump by a factor \order{1} as we move from one regime to another.
\item The IR behavior of the tachyon in the regime $r \gg 1/\Lambda_\tau$ is nontrivial and qualitatively different from the formulas given above. Therefore, the IR boundary conditions obtained by requiring continuity at $r\simeq 1/\Lambda_\tau$ are taken to be generic, i.e., it is assumed that the conditions are not fine tuned to pick any specific solution for $r \ll 1/\Lambda_\tau$.
\end{itemize}
\subsubsection{Zero quark mass}
Let us first recall how the energy scales depend on $x$ at zero quark mass. The results for $x<x_c$ are obtained by using the above assumptions and recipes. Therefore they are also independent of the details of the model. The computation can be found in Sec.~10 of~\cite{jk} and the scaling results will also be reproduced by the analysis of Sec.~\ref{sec:condensate} below, where the chiral condensate is studied in detail (see also~\cite{kutasov}). We will only list the results here.
When $x<x_c$, $\Lambda_\mathrm{IR} \sim \Lambda_\tau$ (i.e., the tachyon becomes sizeable where the asymptotic IR geometry starts), and $\Lambda_\tau$ is not defined in the conformal window ($x_c \le x <x_\mathrm{BZ}$).
Depending on the value of $x=N_f/N_c$, there are three regions with qualitatively different behavior:
\begin{itemize}
\item In the QCD regime, meaning that $0<x<x_c$ and $x_c-x\gtrsim 1$,
there is only one scale, $\Lambda_\mathrm{UV} \sim \Lambda_\mathrm{IR}$, which corresponds to $\Lambda_\mathrm{QCD}$.
\item In the walking regime, which is found when $x<x_c$ and $x_c-x \ll 1$, the IR and UV scales are related through Miransky scaling:
\begin{equation} \label{Mscal}
\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} \sim \exp\left[\frac{K}{\sqrt{x_c-x}}\right] \ .
\end{equation}
The chiral condensate satisfies
\begin{equation}
\frac{\sigma}{\Lambda_\mathrm{UV}^3} \sim \frac{\Lambda_\mathrm{IR}^2}{\Lambda_\mathrm{UV}^2} \sim \exp\left[-\frac{2 K}{\sqrt{x_c-x}}\right] \ .
\end{equation}
\item In the conformal window ($x\ge x_c$) the tachyon is zero as chiral symmetry is intact. In this case one expects that the scales of the UV and the IR expansions are the same, $\Lambda_\mathrm{UV}\sim\Lambda_\mathrm{IR}$ (as there is no obvious mechanism which would separate the scales).
\end{itemize}
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.69\textwidth]{140714_mq_scaling_regions.pdf}
\end{center}
\caption{A schematic diagram of the different scaling regions as functions of $x$ and $m_q$.}
\label{fig:scalingregions}\end{figure}
\subsubsection{Small but finite quark mass}
Let us then generalize these results to nonzero quark mass, first assuming a small mass ($m_q/\Lambda_\mathrm{UV} \ll 1$). The detailed analysis uses the strategy formulated above, and can be found in Appendix~\ref{app:scaling}. We will only discuss the results here.
One can identify, for $m_q/\Lambda_\mathrm{UV}\ll 1$, two regimes on the $(x,m_q)$-plane, shown schematically in Fig.~\ref{fig:scalingregions}:
\begin{itemize}
\item[A] The regime where the quark mass is a small perturbation. This is possible for $0<x<x_c$ as the $m_q=0$ solution is continuously connected to the solution with finite $m_q$. In order to determine the extent of regime~A, we will perturb the solution at vanishing quark mass by adding a small $m_q$. It is a small perturbation so long as the nonnormalizable term remains small for $r \ll 1/\Lambda_\mathrm{UV}$. This is equivalent to $m_q \ll |\sigma(m_q=0)/\Lambda_\mathrm{UV}^2|$, or
\begin{equation}
\frac{m_q}{\Lambda_\mathrm{UV}} \ll 1
\end{equation}
in the QCD regime, and
\begin{equation} \label{regAcondw}
\frac{m_q}{\Lambda_\mathrm{UV}} \ll \exp\left[-\frac{2K}{\sqrt{x_c-x}}\right]
\end{equation}
in the walking regime. Apart from the perturbation caused by the quark mass, the energy scales behave as in the $m_q=0$ case discussed above. In particular, the Miransky scaling law~\eqref{Mscal} applies for $x_c-x \ll 1$.
\item[B] The ``scaling'' regime which involves walking of the coupling, and the amount of walking is determined by the quark mass. Continuity of the tachyon implies for $x\le x_c$ that (see Appendix~\ref{app:scaling} for details)
\begin{equation} \label{regimeBlowx}
\frac{m_q}{\Lambda_\mathrm{UV}} \sim \frac{\Lambda_\mathrm{IR}^2}{\Lambda_\mathrm{UV}^2} \sim \frac{\sigma}{\Lambda_\mathrm{UV}^3}\ , \qquad \left(x\le x_c\quad \mathrm{and} \quad \exp\left[-\frac{2K}{\sqrt{x_c-x}}\right] \ll \frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right) \ ,
\end{equation}
whereas in the conformal window
\begin{equation} \label{regimeBhighx}
\!\!\!\!\!\!\! \begin{array}{rcl}
\frac{m_q}{\Lambda_\mathrm{UV}} &\sim& \left(\frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}}\right)^{\Delta_*} \\ \qquad \frac{\sigma}{\Lambda_\mathrm{UV}^3} &\sim& \left(\frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}}\right)^{4-\Delta_*} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{\frac{4-\Delta_*}{\Delta_*}}\end{array} \qquad \left(x_c \le x < x_\mathrm{BZ} \quad \mathrm{and} \quad \frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right) \ .
\end{equation}
Notice that~\eqref{regimeBlowx} is obtained from~\eqref{regimeBhighx} by setting $\Delta_* = 2$, which is indeed the value of the anomalous dimension as $x \to x_c$ from above, and the scaling behavior is therefore continuous and both expressions are valid at $x=x_c$. Moreover,~\eqref{regimeBhighx} gives what is termed the ``hyperscaling'' relation for the chiral condensate (see, e.g., \cite{DelDebbio:2010ze}) and often written in terms of the anomalous dimensions $\gamma_* = \Delta_*-1$,
\begin{equation}
\frac{\sigma}{\Lambda_\mathrm{UV}^3} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{\frac{3-\gamma_*}{1+\gamma_*}} \ .
\end{equation}
Such scalings have been studied recently in a specific holographic model~\cite{Evans:2014nfa}. The behavior of $\Lambda_\tau$ is the same as for $m_q=0$ and in the regime~A, i.e.,
\begin{equation}
\Lambda_\tau \sim \Lambda_\mathrm{IR}\ .
\end{equation}
\end{itemize}
Notice that we have not discussed the $m_q$-dependence of bound state masses and decay constants, because they are difficult to analyze without specifying the details of the model. In the regime~A and at small $x$, however, it is clear that the lowest bound state masses (except for the light pions) and decay constants must be \order{\Lambda_\mathrm{IR}} because this is the only available scale (apart from the small perturbation due to the quark mass). The pions are expected to obey the GOR relation since this arises in holography quite in general.
In the regime~B, as well as in the regime~A when $x_c-x \ll 1$, the most natural expectation is that the masses and decay constants are still given by the soft IR scale \order{\Lambda_\mathrm{IR}}. This will be demonstrated for V-QCD below.
\subsection{Large quark mass}
Let us then discuss the limit $m_q/\Lambda_\mathrm{UV} \gg 1$ (regime~C in Fig.~\ref{fig:scalingregions}).
Since the quark mass is large, the tachyon grows large very close to the UV boundary where the dilaton is still small. Therefore regime~C probes the limit of small dilaton and large tachyon in the holographic model. Due to the smallness of the dilaton it is not obvious that the holographic description is reliable for all observables in this regime. But in analogy to how the UV structure of V-QCD is fixed in order to guarantee correct boundary conditions for the IR dynamics (as explained in Sec.~\ref{sec:vqcd}), it is important that the holographic model is as close to QCD as possible also at large quark mass, in order to have the best possible boundary conditions for the physics at small and \order{1} quark masses. We will therefore analyze what are the possibilities in this limit.
Deep in the UV, for $r \ll 1/m_q$, the tachyon will have the standard form of~\eqref{tauruns} which already implies scaling results for $\sigma$ and $\Lambda_\tau$, with similar assumptions as above. At $r \sim 1/m_q$ it will grow large, after which nonlinear effects are important and a generic solution cannot be written down. In particular, the RG flow is not expected to approach the IRFP at any point, and the IR structure of the tachyon solution shown above is not relevant.
Instead, additional scaling results can be derived assuming that the holographic model implements some key features of QCD (as will be the case in V-QCD). These are the RG flow of the 't Hooft coupling in QCD, which also gives the proper definition of the UV scale $\Lambda_\mathrm{UV}$ in regime~C, and the decoupling of the massive quarks at energy scales much smaller than $m_q$. Such decoupling is automatically implemented in V-QCD by the Sen-like exponential tachyon potential of the flavor action, as we will explain in Sec.~\ref{sec:massscales}.
Due to the decoupling, the RG flow is that of full QCD for $r \ll 1/m_q$ and that of YM theory for $r \gg 1/m_q$.
Continuity of the 't Hooft coupling (rather than the tachyon) is required.
The collected results for the regime~C in Fig.~\ref{fig:scalingregions} are the following (details can again be found in Appendix~\ref{app:scaling}):
\begin{itemize}
\item[C] The regime of large quark mass ($m_q/\Lambda_\mathrm{UV} \gg 1$). The form of the tachyon solution implies that
\begin{equation} \label{largemqtauscales}
\sigma \sim m_q^3 \qquad \mathrm{and} \qquad \Lambda_\tau \sim m_q\ .
\end{equation}
Here $\sigma$ will be connected to the properly renormalized chiral condensate (see Appendix~\ref{app:freeencoll} for details). The renormalization also involves scheme dependence which is important at large quark mass (see~\cite{ikp} for a discussion in the context of holography). The condensate is proportional to $m_q^3$ as in~\eqref{largemqtauscales} for generic schemes.
Further, analysis of the RG flow leads to
\begin{equation} \label{largemqscalingt}
\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{b_0/b_0^\mathrm{YM}-1}\ , \qquad \frac{m_q}{\Lambda_\mathrm{IR}} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{b_0/b_0^\mathrm{YM}}
\end{equation}
where $b_0$ ($b_0^\mathrm{YM}$) is the leading coefficient of the beta function for QCD in the Veneziano limit (YM theory at large $N_c$). That is,
\begin{equation}
\frac{b_0}{b_0^\mathrm{YM}} = 1-\frac{2x}{11} \ .
\end{equation}
\end{itemize}
Let us then briefly comment on the size of the bound state masses in regime~C. First, as the quarks are decoupled for energy scales smaller than the quark mass, the glueballs are expected to decouple from the mesons, and have a mass gap of \order{\Lambda_\mathrm{IR}}. Recall that the meson states in QCD become nonrelativistic at large $m_q$, and therefore their mass gap is $\sim 2 m_q$, and the mass splitting of the low-lying states is much smaller than the gap.
The example of V-QCD, which we shall discuss below, shows that obtaining such a mass gap in holography is nontrivial (see the analysis for V-QCD in Appendix~\ref{app:masses}), and the gap can be either \order{m_q} or \order{\Lambda_\mathrm{IR}} for actions which produce reasonable results at small $m_q$.
If an IR cutoff is placed at the point where the tachyon grows large (as in the dynamic AdS/QCD models~\cite{Alvares:2012kr}), mass gap $\propto m_q$ is obtained. In the presence of such a cutoff, $\Lambda_\tau \sim m_q$ is the only scale in the system. But in this case the splitting between the bound state masses is also expected to be \order{m_q}.
Finally let us comment on the structure of Fig.~\ref{fig:scalingregions} near the BZ region. Namely, the boundary of the regime~C bends toward higher $m_q/\Lambda_\mathrm{UV}$ as $x$ grows. Similarly the upper boundary of the regime~B bends down in the BZ region. This is a generic feature due to RG flow which becomes slower and slower in the BZ limit $x \to x_\mathrm{BZ}$. Due to slowness of the flow the separation of the energy scales $m_q$ and $\Lambda_\mathrm{UV}$ needs to be larger for the scaling results to apply. This phenomenon is discussed in more detail in Appendix~\ref{app:scaling}.
Actually the BZ regime could be identified as an additional scaling regime. Here this is not done, however, since from the point of view of holography this regime is not the most interesting one. The coupling constant is restricted from above by its small value at the IRFP, the theory is perturbatively soluble, and therefore holographic description is not expected to be useful.
\section{Scaling of bound state masses in V-QCD} \label{sec:massscales}
While we discussed above the generic behavior of (holographic) QCD in the Veneziano limit, we shall now derive explicit predictions for the V-QCD models defined in Sec.~\ref{sec:vqcd}, and demonstrate that they agree with the results of the previous section.
\subsection{Energy scales in V-QCD}
The various energy scales can be defined explicitly in terms of the background solutions. For definiteness we will write down the definitions here. We will use the UV and IR expansions which can be found in Appendix~D of~\cite{jk} and in Appendix~D of~\cite{Arean:2013tja}. The UV expansions can also be found in Appendix~\ref{app:freeen}.
\begin{itemize}
\item First, $\Lambda_\mathrm{UV}$ is the scale of the UV expansions. The precise definition is most conveniently given in terms of the UV expansion of the dilaton $\lambda$:
\begin{equation} \label{laUV}
\lambda = -\frac{1}{b_0 \log (r\Lambda_\mathrm{UV})} -
\frac{ 8 b_1 \log\left[-\log(r \Lambda_\mathrm{UV})\right]}{9 b_0^2\log(r \Lambda_\mathrm{UV})^2}+{\cal
O}\left(\frac{1}{\log(r\Lambda_\mathrm{UV})^3}\right) \ .
\end{equation}
Recall that the coefficients of the potentials in the V-QCD action were matched to those of the QCD beta function in the UV. We used this mapping to write the coefficients in the expansion~\eqref{laUV} in terms of the coefficients $b_i$ of the QCD beta function in the Veneziano limit,
\begin{equation}
\beta(\lambda) \equiv \frac{d\lambda}{d\log \mu} = -b_0 \lambda^2 +b_1 \lambda^3 + \cdots
\end{equation}
\item The IR scale $\Lambda_\mathrm{IR}$ can be defined analogously in term of the IR expansions. For the standard geometry in V-QCD with an IR singularity the definition of $\Lambda_\mathrm{IR}=1/R$ is given in limit $r \to \infty$ as~\cite{jk}
\begin{equation} \label{laIR}
\log \lambda = \frac{3}{2} \frac{r^2}{R^2} +\morder{r^0} = \frac{3}{2} \Lambda_\mathrm{IR}^2r^2 +\morder{r^0} \ .
\end{equation}
In the presence of an IRFP the definition is modified to
\begin{equation} \label{laIRFP}
\lambda \simeq \lambda_* -\left(\frac{r}{R}\right)^{-\delta} =\lambda_* -\left(r \Lambda_\mathrm{IR}\right)^{-\delta}
\end{equation}
where $\delta = \Delta_{FF}-4$ is the anomalous dimension of the $\mathrm{Tr} F^2$ operator at the fixed point.
\item The quark mass is determined\footnote{It is well known that the quark mass can be defined only up to a constant. This constant is set to one for notational simplicity.} in terms of the UV asymptotics of the tachyon ($r\to 0$),
\begin{equation}
\frac{\tau(r)}{\ell} = m_q r (-\log(r\Lambda_\mathrm{UV}))^{-\rho}\left[1+\morder{\frac{1}{\log(r\Lambda_\mathrm{UV})}}\right]\ ,
\end{equation}
where $\ell$ is the UV AdS radius and $\rho=\gamma_0/\beta_0$ is the ratio of the leading coefficients $\beta_0$ and $\gamma_0$ of the beta function and the anomalous dimension of the quark mass in QCD, respectively. Notice that in V-QCD, the logarithmic running of the quark mass is therefore included, and is matched to agree with that of QCD.
\item The scale $\L_\tau$ where the tachyon grows large is defined simply by\footnote{For potentials~II slightly modified definition is used: the number on the right hand side is set to $1/10$. This is necessary because the transition to the IR region where nonlinear terms in the tachyon are important turns out to happen when the tachyon is still smaller than one, and in the nonlinear region the tachyon grows relatively slow. Therefore using exactly~\eqref{Lambdataudef} would lead to an energy scale which does not precisely reflect the change in the dynamics.}
\begin{equation} \label{Lambdataudef}
\left.\frac{\tau}{\ell}\right|_{A=\log \Lambda_\tau} = 1 \ ,
\end{equation}
where $\tau$ and $A$ are to be evaluated on the background solution.
\end{itemize}
The background EoMs for V-QCD are invariant under the transformation
\begin{equation}
r \to \L r \ , \qquad A \to A - \log\L \ .
\end{equation}
This transformation changes all energy scales defined above by the same number, and reflects the choice of the units of energy on the field theory side. Consequently, only the ratios of the above energy scales are a priori well-defined. Notice also that $\Lambda_\mathrm{IR}$ and $\Lambda_\tau$ do not have direct counterparts in field theory, whereas $\Lambda_\mathrm{UV}$ is indeed mapped to the scale of the RG flow in field theory and $m_q$ is identified (up to a constant) as the quark mass on the field theory side. Therefore the values of $\Lambda_\mathrm{UV}$ and $m_q$ (up to the ambiguities mentioned above) can be inferred from QCD data only, whereas the determination of $\Lambda_\mathrm{IR}$ and $\Lambda_\tau$ also requires pinning down the holographic action.
There is a small issue with the above definitions which will be visible in the explicit results below. Namely, the definition of $\Lambda_\mathrm{UV}$ which is natural at generic values of $x$ is not optimal at high $x$ and in particular in the BZ limit. In this limit
one would expect $\Lambda_\mathrm{UV}$ to match the scale of the UV RG flow, but this turns out not be be the case.
Also at zero quark mass there should be only a single scale and therefore $\Lambda_\mathrm{UV}$ should equal $\Lambda_\mathrm{IR}$, but as it turns out, actually $\Lambda_\mathrm{UV}$ becomes exponentially suppressed with respect to $\Lambda_\mathrm{IR}$. Because the RG flow of the coupling is controlled by the two-loop beta function in the BZ limit, a single scale $\widetilde \L$ can be defined which has better behavior in this limit (see Appendix~B of~\cite{Kiritsis:2013xj}):
\begin{equation}
\left(\frac{b_0}{b_1 \lambda}-1\right) \ \exp\left(\frac{b_0}{b_1 \lambda}\right)\simeq \left(\widetilde \L r\right)^{-\frac{b_0^2}{b_1}}
\end{equation}
This scale is related to $\Lambda_\mathrm{UV}$ in the BZ limit by
\begin{equation}
\widetilde \L \simeq \left(\frac{b_1}{b_0^2}\right)^\frac{b_1}{b_0^2} \Lambda_\mathrm{UV} \ ,
\end{equation}
where $b_1/b_0^2 \sim (x_\mathrm{BZ}-x)^{-2}$ grows large in the BZ limit (see Appendix~\ref{app:scaling} for some more details). Notice that $\Lambda_\mathrm{UV}$ will anyhow be used as a reference scale in the numerical analysis below -- since the BZ region is not very interesting from holographic viewpoint, we have chosen to use the same definition of $\Lambda_\mathrm{UV}$ as earlier literature, even though it has unnatural behavior in this region.
\subsection{Flavor nonsinglet masses and decay constants: scaling results}
Before going to the numerical results let us argue how the scaling laws for the masses and decay constants can be derived in V-QCD. The meson masses in each sector (for vectors, axial vectors, scalars, and pseudoscalars) may be defined into two classes: the flavor singlet and nonsinglet states. The former are singlets under the vectorial $SU(N_f)$ transformation, whereas the latter are the other fluctuation modes, which correspond to quark bilinear operators involving the Hermitean traceless generators
$t_a$ of $SU(N_f)$. We restrict analytic considerations to the nonsinglet states, because the singlet mesons mix with the glueball states, which would lead to complications~\cite{Arean:2013tja,glue}. Only the main points are summarized here and the details can be found in Appendix~\ref{app:masses}.
The fluctuation equations for the scalar nonsinglet mesons can transformed in the Schr\"odinger form as detailed in Appendices~A and~B of~\cite{Arean:2013tja}. The masses for each sector are then given as eigenvalues of a Schr\"odinger equation with a certain potential term and the Schr\"odinger coordinate running from $u=0$ (UV) to $u=\infty$ (IR). In order to find the behavior of the mass gaps and splittings one then needs to study the Schr\"odinger potentials $V_S(u)$.
Let us first take finite but small quark mass ($m_q/\Lambda_\mathrm{UV} \ll 1)$, so that the regimes~A and~B are covered. We use here the V-QCD action, but results in this region are not sensitive to the details of the action. The Schr\"odinger potential for $u \ll 1/\Lambda_\mathrm{IR}$ is given by the background with small or walking dilaton and small tachyon, so that the geometry is close to AdS, and consequently $V_S(u) \sim \mathrm{const}/u^2$. For $u \gg 1/\Lambda_\mathrm{IR}$ the diverging tachyon creates the confining potential ($V_s(u) \sim \Lambda_\mathrm{IR}^4 u^2$ if the excitation spectrum is linear, $m_n^2 \sim n$ with $n$ being the excitation number). Therefore the only relevant scale is
$\Lambda_\mathrm{IR}$ and the (lowest lying) meson masses as well as the
mass
splittings between the modes are given by this scale:
\begin{equation} \label{massscalingregsAB}
m_n \sim \Lambda_\mathrm{IR} \ ,
\end{equation}
which is also consistent with the results from Dyson-Schwinger and Bethe-Salpeter approaches in regime~A~\cite{Harada:2003dc}. In units of $\Lambda_\mathrm{UV}$, by using the scaling laws from Sec.~\ref{sec:scaling} we therefore obtain the same results as at vanishing quark mass~\cite{letter,Arean:2013tja} within regime~A: $m_n \sim \Lambda_\mathrm{UV}$ in the QCD regime and
the masses go to zero obeying Miransky scaling, $m_n/\Lambda_\mathrm{UV} \sim \exp(-K/\sqrt{x_c-x})$, in the walking regime. In regime~B, we find the ``hyperscaling relations''~\cite{DelDebbio:2010ze,Evans:2014nfa} for the low-lying meson masses:
\begin{align}
\frac{m_n}{\Lambda_\mathrm{UV}} &\sim \sqrt{\frac{m_q}{\Lambda_\mathrm{UV}}} \ ,& \ &\left(x\le x_c\quad \mathrm{and} \quad \exp\left[-\frac{2K}{\sqrt{x_c-x}}\right] \ll \frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right) \ ,& \nonumber\\
\frac{m_n}{\Lambda_\mathrm{UV}} &\sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{1/\Delta_*}\ ,& \ &\left(x_c \le x < x_\mathrm{BZ} \quad \mathrm{and} \quad \frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right) \ .&
\end{align}
The sole exception to these scaling results is the pion mode (for $x<x_c$), which is massless at $m_q=0$ and obeys the GOR relation in regime~A:
\begin{equation}
m_\pi^2 f_\pi^2 \sim m_q \sigma \ ,
\end{equation}
so that it is lighter than the other meson states.
The GOR relation will be discussed in more detail in Sec.~\ref{sec:condensate}. At the level of the Schr\"odinger formalism the absence of the mass gap in the pseudoscalar sector is reflected in the negativity of the Schr\"odinger potential in the UV region. In regime~B, the pseudoscalar masses will also obey the scaling~\eqref{massscalingregsAB}.
The various decay constants are slightly more difficult to analyze. In Appendix~\ref{app:masses} it is argued that they are similarly of the order of $\Lambda_\mathrm{IR}$ when the quark mass is small.
At large quark mass (regime~C) the meson masses and decay constants depends more on the details of the action. Actually only the form of the flavor action $S_f$ in the limit of large tachyon and small dilaton is relevant, as is shown in Appendix~\ref{app:masses}. In this limit the functions $a$, $\kappa$, and $V_{f0}$ become almost constants and the action of~\eqref{generalact} takes the form of the standard DBI action with the exponential Sen potential.
The exponential dependence $V_f \propto \exp(-a(\lambda) \tau^2)$ of the flavor potential on the (squared) tachyon naturally implements the decoupling of the massive quarks. For large quark mass the tachyon is roughly proportional to $m_q r$ in the UV (for $r \ll 1/m_q$). When $r$ grows larger than $1/m_q$, that is, at energies lower than $m_q$ on the field theory side, the tachyon grows sizeable (see Appendix~\ref{app:largemq} for details) and the exponential factor in $V_f$ decays rapidly. Consequently, the flavor part of the V-QCD action becomes suppressed with respect to the glue part, and therefore the dynamics at energies below $m_q$ is governed by the gluons as expected. The decoupling of flavors will also be well visible in the numerical results for the scalar singlet states below which involve both glueball and $\bar qq$ components.
Interestingly, as shown in Appendix~\ref{app:masses}, the choice $V_f(\lambda,\tau) = V_{f0}(\lambda) \exp(-a_0\tau^2)$ of potentials~I, where the function $a(\lambda)$ is set to a constant value $a_0$, is a special case. This choice was motivated by the asymptotics of the meson trajectories at zero quark mass~\cite{jk,Arean:2013tja}, and noticeably $a(\lambda)$ is also constant in tachyon potentials obtained from string theory~\cite{Kraus:2000nj,Bigazzi:2005md,ckp}. In Appendix~\ref{app:masses} we show that this choice is essentially the only one which produces physically reasonable mass gap and splitting, if the value of $a_0$ is slightly modified from the value of potentials~I which reproduces the correct dimension of the quark mass and condensate in the UV.
We will anyhow discuss the results for potentials~I without this modification of $a_0$, because such a modification was not introduced for our numerical studies. It is found that the mass gap for the (flavor nonsinglet) mesons in all sectors (vectors, axials, scalars and pseudoscalars) is given by
\begin{equation} \label{massscalingregC}
m_\mathrm{gap} \sim m_q^\xi\ ; \qquad \xi = \frac{\left.3\ell^2\right|_{x=0}}{4 \ell^2} = \frac{3}{4} \left(1-\frac{x W_0}{12}\right)\ ,
\end{equation}
where $\ell$ is the UV AdS radius, and $\xi=1$ would be required to match with the field theory result for nonrelativistic bound states\footnote{One can have both $\xi=1$ and the correct UV dimension for the quark mass if extra terms (e.g. $\propto (1 +\# \tau^2)$) are added in the tachyon potential.}.
We stress that this formula was obtained by fixing the value of $a_0$ to produce the UV dimension of the quark mass and the condensate, as was done for the potentials~I used in the numerics. The last expression in~\eqref{massscalingregC} suggests that $\xi=1$ could be obtained by tuning the value of $W_0$. This is, however, problematic because the contribution involving $W_0$ should vanish in the probe limit $x \to 0$, and also because $W_0$ would negative, which causes $V_f$ to have a node~\cite{jk} at which the tachyon background equation becomes singular.
The mass splitting is suppressed with respect to the gap as is the case in real quarkonia. For potentials~I it scales as the inverse of the mass gap.
The decay constants of the lowest meson states are exponentially suppressed in the quark mass
\begin{equation}
f_n \sim \exp(-\# m_q^{2\xi})
\end{equation}
and they are therefore decoupled. This is found for all mesons with masses below \order{m_q}. The decay constants of the states with masses $\sim m_q$ are of the order of $\Lambda_\mathrm{IR}$.
\subsection{Numerical results}
Computing the energy scales and masses numerically is straightforward (but tedious) after the background has been constructed numerically~\cite{jk,Arean:2013tja}. The potentials~I (given explicitly in Appendix~\ref{app:numerics}) are used here unless stated otherwise. We choose three reference values $x=1$, $4$, and $4.5$, which lie in the QCD regime, walking regime, and conformal window, respectively, and plot the observables as functions of the quark mass. We also show plots where $m_q/\Lambda_\mathrm{UV}$ is fixed to $10^{-6}$ and $x$ is varied over the whole parameter space. These choices cover the most interesting structures of Fig.~\ref{fig:scalingregions}.
Notice that when plotting dimensionful parameters as a function of $x$ we are comparing different theories and a choice for the reference scale must be made.
A natural choice would be $\Lambda_\mathrm{UV}$, which roughly
corresponds to the scale where the 't Hooft coupling takes some fixed
tiny value very close to the UV boundary. In many of the plots below, however,
$\Lambda_\mathrm{IR}$ is chosen as the reference scale instead, simply because this makes the plots more easily readable.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_scales_vs_x_mq_10pm6.pdf}%
\includegraphics[width=0.49\textwidth]{141202_scales_vs_mq_x_1.pdf}
\includegraphics[width=0.49\textwidth]{141202_scales_vs_mq_x_4.pdf}%
\includegraphics[width=0.49\textwidth]{141202_scales_vs_mq_x_4p5.pdf}
\end{center}
\caption{Dependence of the energy scales on $x$ and $m_q$. Also the masses of the rho meson $m_\rho$, the lowest singlet scalar $m_\mathrm{ss}$, and the pions $m_\pi$ are shown for reference. The blue solid curves is $\Lambda_\mathrm{UV}$, the red dashed curve is $\Lambda_\tau$, the thin dotted magenta curve is the rho mass, the thin dotdashed green curve is the pions mass, and the thin long-dashed brown curve is the mass of the lowest singlet scalar in each of the plots. See text for more details.}
\label{fig:Escales}\end{figure}
\subsubsection{Energy scales}
In Fig.~\ref{fig:Escales} the dependence of the energy scales on $x$ and $m_q$ is demonstrated. We have chosen to show the scales $\Lambda_\mathrm{UV}$ and $\Lambda_\tau$ in the units of $\Lambda_\mathrm{IR}$ since this makes the details visible. Some of the bound state masses are also shown as thin lines.
The top-left plot shows the dependence of the scales on $x$ at a tiny quark mass ($m_q/\Lambda_\mathrm{UV} = 10^{-6}$). The data extend only up to $x=5.1$ because in the BZ region it is difficult to do reliable numerics. The top-right plot shows the mass dependence in the running regime ($x=1$). The bottom-left plot is in the walking regime ($x=4$, which is close to $x_c\simeq 4.0830$). The bottom-right plot is in the conformal window ($x=4.5$).
The thick solid blue curves show the ratio $\Lambda_\mathrm{UV}/\Lambda_\mathrm{IR}$ in each plot. In the top-left plot the crossover from the QCD-like regime to conformal window is clearly visible: As $x \to x_c$ from below, the ratio first grows according to the Miransky scaling law~\eqref{Mscal} in regime~A until the condition~\eqref{regAcondw} no longer holds. Thereafter the ratio saturates to roughly $\sqrt{\Lambda_\mathrm{UV}/m_q}$ in regime~B as predicted by~\eqref{regimeBlowx}, and then one moves out of regime~B at even higher $x$.
The dependence of $\Lambda_\mathrm{UV}/\Lambda_\mathrm{IR}$ on $m_q$ also follows the predicted scaling laws. In the regime~A (low $m_q/\Lambda_\mathrm{UV}$ in top-right and bottom-left plots) the ratio is constant as the quark mass is a small perturbation. In the regime~C, i.e., at large $m_q/\Lambda_\mathrm{UV}$, the ratio follows the power law of~\eqref{largemqscalingt} as best seen in the top-right plot. In the regime~B the ratio obeys the other power law of~\eqref{regimeBlowx} and~\eqref{regimeBhighx} as best seen at intermediate $m_q$ in the bottom-left plot. Notice that much of the solid blue curve was left out in the bottom row plots in order to make the details of the other curves better visible.
The thick dashed red curves show the ratio $\Lambda_\tau/\Lambda_\mathrm{IR}$. At small quark mass, including regimes~A and~B, the ratio is close to one as expected (except in the BZ limit). For the explanation of the divergence of $\Lambda_\tau/\Lambda_\mathrm{IR}$ in the BZ limit see Appendix~\ref{app:scaling}, Eq.~\eqref{LIRmqBZ}. At large quark mass (regime~C), we find $\Lambda_\tau \sim m_q$ as predicted in~\eqref{largemqtauscales}.
At high $x$ (plots in the bottom row) the convergence toward this scaling law is quite slow due to the slow running of the quark mass. It could be demonstrated by continuing the plots up to much larger $m_q/\Lambda_\mathrm{UV}$, but we have chosen not to do so in order to show the other details in the plots more clearly. Actually all variables vary slower and slower as functions of $m_q/\Lambda_\mathrm{UV}$ when $x$ is grows, and therefore we have substantially increased the range of $m_q/\Lambda_\mathrm{UV}$ in the plots with higher $x$, but this is still not enough to demonstrate the large $m_q$ scaling convincingly. The change in the $m_q$-dependence is due to the RG flow (see the end of Appendix~\ref{app:scaling} for some more details).
The dependence of $\sigma$ on $m_q$ will be discussed in Sec.~\ref{sec:condensate}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_mass_ratios_vs_x_mq_10pm6.pdf}%
\includegraphics[width=0.49\textwidth]{141202_mass_ratios_vs_mq_x_1.pdf}
\includegraphics[width=0.49\textwidth]{141202_mass_ratios_vs_mq_x_4.pdf}%
\includegraphics[width=0.49\textwidth]{141202_mass_ratios_vs_mq_x_4p5.pdf}
\end{center}
\caption{Dependence of the mass ratios between the lowest excitations of each tower on $x$ and $m_q$. The choices of $x$ and $m_q/\Lambda_\mathrm{UV}$ are as in Fig.~\protect\ref{fig:Escales}. Blue solid curve is the lowest vector ($\rho$ meson) mass, red dashed curve is the nonsinglet axial vector mass, dotted magenta curve is the lowest nonsinglet scalar mass, and the dotdashed green curve is the nonsinglet pseudoscalar (pion) mass in each plot. In addition, the long-dashed brown curve is the lowest singlet scalar mass. }
\label{fig:mratios}\end{figure}
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140716_vector_masses_vs_mq_x_1.pdf}%
\includegraphics[width=0.49\textwidth]{140716_vector_ratios_vs_mq_x_1.pdf}
\end{center}
\caption{Masses of the lowest four vector states as a function of $m_q$ at $x=1$. Left: masses in units of $\Lambda_\mathrm{IR}$. Right: masses normalized to the mass of the lowest vector meson (the $\rho$-meson).}
\label{fig:vmasses}\end{figure}
\subsubsection{Flavor nonsinglet masses and decay constants}
The flavor nonsinglet spectra can be computed as explained in~\cite{Arean:2013tja}. The results for different values of $x$ and $m_q$ are shown in Figs.~\ref{fig:Escales},~\ref{fig:mratios}, and~\ref{fig:vmasses}.
First, the $\rho$ and $\pi$ masses are shown in Fig.~\ref{fig:Escales} with thin dotted magenta and thin dotdashed green curves, respectively. As expected, the $\rho$ mass is \order{\Lambda_\mathrm{IR}} at small quark mass (regimes~A and~B) and obeys the power law~\eqref{massscalingregC} in regime~C. The pion mass is close to the $\rho$ mass in regimes~B and~C, but in regime~A it obeys the GOR relation instead, as best visible from the top-right and bottom-left plots at small $m_q$. In the top-left plot, $m_\pi/\Lambda_\mathrm{IR}$ is \order{\sqrt{m_q/\Lambda_\mathrm{UV}}} at small $x$ but then increases with $x$ in the walking regime (actually obeying the Miransky scaling law) until the ratio is \order{1} at the point of the crossover near $x=x_c$.
Fig.~\ref{fig:mratios} shows the masses of the lowest meson states (i.e., the mass gaps) in each sector normalized to the $\rho$ mass. The choices of $x$ and $m_q$ are the same as in Fig.~\ref{fig:Escales}. The masses of the lowest vector, axial, scalar, and pseudoscalar states are given by the solid blue, dashed red, dotted magenta, and dotdashed green curves, respectively (the brown curves give the scalar singlet mass gap which will be discussed below). These ratios are mostly constant and close to one as predicted by the above scaling arguments. The exception is the pion mass (dotdashed green curves) which obeys the GOR relation in regime~A. Notice that in regime~C all meson mass gaps should approach the same number (roughly $2 m_q$) as expected for nonrelativistic bound states, but we find instead that the axial and pseudoscalar gaps are larger than those of the vectors and scalars. The reasons for this are analyzed in Appendix~\ref{app:masses}. Notice also that the lowest scalar states are
lighter than vectors even at small values of $x$, which seems to be in conflict with QCD.
Such details are, however, sensitive to the choice of the potentials in the V-QCD action, and can be changed by tuning the potentials.
Finally we plot the masses of the four lowest vector states in Fig.~\ref{fig:vmasses} in order to demonstrate the dependence of the mass splittings in the spectra on $m_q$. The masses are given in units of $\Lambda_\mathrm{IR}$ (left hand plot) and normalized to the lowest mass, i.e., the $\rho$ mass (right hand plot). The splittings decrease with increasing $m_q$ in regime~C which is in qualitative agreement with the bound states becoming nonrelativistic, but the power laws are not exactly correct (see Appendix~\ref{app:masses} for details).
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_singlet_scalars_vs_mq_x_1.pdf}%
\includegraphics[width=0.49\textwidth]{141202_singlet_scalars_vs_mq_x_4.pdf}
\includegraphics[width=0.49\textwidth]{141202_singlet_scalars_vs_mq_x_4p5.pdf}
\end{center}
\caption{The masses of the five lowest scalar singlet states as functions of $m_q$ for various choices of $x$.}
\label{fig:ssmasses}\end{figure}
\subsubsection{Scalar singlet masses} \label{sec:singlets}
The singlet sector is qualitatively different from the nonsinglet sector because it also contains glueball states which mix nontrivially with the singlet meson states. Such mixing takes place in the scalar and pseudoscalar sectors. The scalar sector will be discussed in detail here while the pseudoscalar sector will be analyzed in a future publication~\cite{cpodd}.
Before going to the numerical results, let us discuss the generic features of the spectrum. The singlet mesons masses are expected to show similar\footnote{There is an exception in the singlet pseudoscalar sector: the lowest state, the $\eta'$ meson, is anomalously light at small $x$ in the Veneziano limit~\cite{Witten:1978bc,vu1}. This is also reproduced by V-QCD~\cite{cpodd}.} $m_q$ and $x$ dependence as the nonsinglet mesons above, given by~\eqref{massscalingregsAB} and~\eqref{massscalingregC}. In the singlet case there is, however, nontrivial mixing of the mesons with the glueball states, the masses of which should be independent of $m_q$ and therefore always characterized by $\Lambda_\mathrm{IR}$. In particular in regime~C the masses of the mesons become much larger than the glueballs which suggest that the meson and glueball states decouple.
Let us then demonstrate these features numerically for the scalar singlet sector in V-QCD. The mass gap of the scalar singlets is shown by the long-dashed brown curves in Figs.~\ref{fig:Escales} and~\ref{fig:mratios}. The mass gap is that of the glueballs and therefore \order{\Lambda_\mathrm{IR}} for all values of $m_q$ and $x$ as seen from Fig.~\ref{fig:Escales}. The fact that the lowest glueball mass is suppressed with respect to the rho mass $m_\rho$ in regime~C can also be seen in Fig.~\ref{fig:mratios}: the ratios $m_\mathrm{ss}/m_\rho$ given by the brown curves, where $m_\mathrm{ss}$ is the mass gap for the scalar singlet states, decrease with $m_q$ at large $m_q$.
The decoupling of the meson and glueball states is most clearly demonstrated by Fig.~\ref{fig:ssmasses}, where the masses of the five lowest scalar singlet states are plotted as a function of $m_q$ in the QCD regime ($x=1$, top-left), in the walking regime ($x=4$, top-right) and in the conformal window ($x=4.5$, bottom). When $m_q/\Lambda_\mathrm{UV} \ll 1$ the spectrum has both glueball and meson states which are nontrivially mixed. As $m_q$ increases, the meson masses increase while the glueball masses stay constant, which leads to the crossing structure seen in the plots. This would be expected to happen at $m_q/\Lambda_\mathrm{UV} \sim 1$, but as was discussed above, $\Lambda_\mathrm{UV}$ is not exactly the scale of the UV RG flow when $x$ is large, and therefore the crossing structure shifts to higher values of $m_q$ as $x$ increases. At large $m_q/\Lambda_\mathrm{UV}$ only the glueballs are left. Their decoupling from the mesons is demonstrated by the fact that the limiting values of their masses as $m_q \to \infty$ are independent of $x$, and in fact it can be
checked that they match with the glueball masses obtained in the YM limit ($x \to 0$) of V-QCD.
It was shown in~\cite{letter,Arean:2013tja} that V-QCD does not have a light ``technidilaton'' mode~\cite{walk2} (which would be the lightest scalar singlet state) as $x \to x_c$. Both the singlet and nonsinglet scalars fluctuations do have critical~\cite{son} behavior in the near conformal region (see Appendix~I in~\cite{Arean:2013tja}) and the Schr\"odinger potential for the nonsinglet scalars is negative (as was also found in~\cite{Alanen:2011hh}), but it was shown numerically that this is not enough for a technidilaton to appear. The negative result is also seen in Figs.~\ref{fig:mratios} and~\ref{fig:ssmasses}: the lowest singlet scalar does not become light with respect to the other states in the walking regime ($x=4$). We do see, however, that it is lighter in regime~A than in regime~B.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_masses_LaIR_x4.pdf}
\includegraphics[width=0.49\textwidth]{141202_masses_LaUV_x4.pdf}
\end{center}
\caption{The masses of the lowest vector (rho meson), pseudoscalar (pion), and singlet scalar states as a function of $m_q$ and in units of $\Lambda_\mathrm{IR}$ (left) and $\Lambda_\mathrm{UV}$ (right). The rho meson, pion, and singlet scalar masses are shown as the solid blue, dashed red, and doted magenta curves, respectively.}
\label{fig:massesonly}\end{figure}
\subsubsection{Masses near the conformal transition}
To conclude this section, let us add a few comments on the scaling of the bound state masses near the conformal transition. We plot the masses of the rho meson, the pion, and the lowest scalar singlet state in Fig.~\ref{fig:massesonly}. The left hand plot shows the masses in units of $\Lambda_\mathrm{IR}$ and the right hand plot shows the masses in units of $\Lambda_\mathrm{UV}$. Notice that in units of $\Lambda_\mathrm{IR}$ the pion mass deviates from the masses of the other states at extremely small $m_q$ (regime~A) where it obeys the GOR relation, but in units of $\Lambda_\mathrm{UV}$ the pion mass obeys the same power law $\sim \sqrt{m_q/\Lambda_\mathrm{UV}}$ in regimes~A and B.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.69\textwidth]{141202_gamma_at_FP_vs_x.pdf}
\end{center}
\caption{The (real part of the) anomalous dimension at the fixed point as a function of $x$ for potentials I with $W_0=3/11$ (solid blue curve) and for potentials II with SB normalized $W_0$ (dashed red curve). The kink lies at $x=x_c$ for both potentials.}
\label{fig:gammastar}\end{figure}
Knowledge of the dependence of the meson masses on $m_q$ near the walking regime is important for the lattice studies which take place at finite quark mass, and aim to locate the conformal transition at $m_q=0$~\cite{Lombardo:2014mda,Lombardo:2014pda}. As we have pointed out above, the regime~B extends even to $x<x_c$ (see Fig.~\ref{fig:scalingregions}), and the crossover between regimes~A and~B moves to lower $x$ as $m_q$ is increased. This suggests that studies at finite $m_q$ lead to an underestimate for $x_c$. One should recall, however, that the scaling in regime~B involves $\Delta_* =\gamma_*+1$ which depends strongly on $x$.
We show $\mathrm{Re} \gamma_*$, which controls the scaling exponents of the masses, as a function of $x$ for both potentials~I and~II in Fig.~\ref{fig:gammastar}. The kinks in the plots are located exactly at $x=x_c$, and $\gamma_*$ drops rapidly right above the kinks. This supports the idea that $x=x_c$ can be located by extracting $\gamma_*$ from the meson masses on the lattice.
In fact, recent lattice results for $\gamma_*$ in the conformal window report very low values~\cite{Cheng:2013eu,Lombardo:2014pda} that are in apparent contradiction of the curves in Fig.~\ref{fig:gammastar}.
Recall however that the model has not been tuned to fit any QCD data yet\footnote{One should also keep in mind that we are working in a bottom-up model, which is defined in the Veneziano limit whereas the lattice data was computed at finite $N_f$ and $N_c$.}. It is actually not difficult to construct potentials for which $\gamma_*$ drops much more rapidly when $x>x_c$.
\section{Quark mass and the chiral condensate}
\label{sec:condensate}
Let us then discuss the mass dependence of the chiral condensate. Recall that the tachyon solution in the UV reads
\begin{equation} \label{mqsigmadef}
\begin{split}
\frac{\tau(r)}{\ell} &=
m_q r \left(-\log r\Lambda_\mathrm{UV}\right)^{-\rho}\left[1+\morder{\frac{1}{\log r\Lambda_\mathrm{UV}}}\right] \\
& \ \ \ + \sigma r^3 \left(-\log r\Lambda_\mathrm{UV}\right)^{\rho}\left[1+\morder{\frac{1}{\log r\Lambda_\mathrm{UV}}}\right]
\end{split}
\end{equation}
where $\sigma$ can be identified as the chiral condensate\footnote{When $m_q$ is finite, the UV expansions in practice only define $\sigma$ up to a linear term in $m_q$. This issue will be discussed below and in Appendix~\ref{app:qbarqnorm}.} (the exact identification is studied in Appendix~\ref{app:qbarqnorm}), and $\rho$ can be expressed in terms of the leading coefficients of the beta and gamma functions as $\rho=\gamma_0/b_0 = 9/(22-4 x)$.
Notice that the analysis of previous sections was restricted to the standard, dominant vacuum. It is known~\cite{jk,Arean:2013tja}, however, that there are subdominant ``Efimov'' vacua in the QCD and walking regimes ($x < x_c$) which quite in general appear in connection to the BKT transition (see, e.g., ~\cite{Iqbal:2011in,kutasov,kutasovdbi}). These vacua are mapped to different values of the chiral condensate on the $(m_q,\sigma)$-plane: all possible regular vacua form a spiral structure, which will be called the ``Efimov spiral'' below. The results for the spiral structure will be used to analyze four-fermion deformations of QCD in Sec.~\ref{sec:4f}.
\subsection{Efimov spirals}
Let us first review the structure of the subdominant vacua. Including the solutions with finite quark mass~\cite{jk}:
\begin{itemize}
\item When $x_c\le x<x_\mathrm{BZ}$, only one vacuum exists, even at finite quark mass.
\item When $0<x<x_c$ and the quark mass is zero, there is an infinite tower of (unstable) Efimov vacua in addition to the standard, dominant solution\footnote{It is assumed here for simplicity that there is an IRFP for all positive values of $x$, and the BF bound is violated at the fixed point all the way down to $x=0$.}.
\item When $0<x<x_c$ and the quark mass is nonzero, there is an even number (possibly zero) of Efimov vacua. The number of vacua increases with decreasing quark mass for fixed $x$.
\end{itemize}
The infinite tower of Efimov vacua, which appears at zero quark mass, admits a natural enumeration $n=1,2,3,\ldots$ where $n$ is the number of tachyon nodes of the background solution as we shall demonstrate below (see also Sec.~10 and Appendix~H in~\cite{jk}). A generic feature of these backgrounds is, that they ``walk'' more than the dominant, standard vacuum, so that the scales $\Lambda_\mathrm{UV}$ and $\Lambda_\mathrm{IR}$ become well separated for all $0<x<x_c$ when $n$ is large enough. It is possible to show that
\begin{equation} \label{scalescalEf1}
\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} \sim \exp\left(\frac{\pi n}{\nu}\right)\ ,\qquad (n\to\infty)\ ,
\end{equation}
for any $0<x<x_c$. The coefficient $\nu$ will be given below in Eq.~\eqref{nudef} (see also Appendix~F in \cite{jk}). In the walking regime, one finds that
\begin{equation} \label{scalescalEf2}
\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} \sim \exp\left(\frac{K(n+1)}{\sqrt{x_c-x}}\right) \ , \qquad (x \to x_c{}^-)
\end{equation}
for any value of $n$. In particular, $n=0$ corresponds to the standard solution discussed in the previous sections, and the relation~\eqref{scalescalEf2} gives the standard Miransky scaling, whereas for $n>0$ the scaling is even faster. We also found a similar scaling result for the free energies of the solutions as $x \to x_c$ in \cite{jk}, therefore proving that the Efimov vacua are indeed subdominant in this limit, and verified this numerically for all $0<x<x_c$. In~\cite{Arean:2013tja} it was shown that the Efimov vacua are perturbatively unstable (again analytically as $x \to x_c{}^-$, and numerically for all $0<x<x_c$).
When the BF bound is violated at the IRFP, the quark mass and the condensate are known to show an oscillating behavior for the (chirally broken) backgrounds where the coupling flows very close to the fixed point~\cite{jk}. Let us first discuss how these oscillations arise from the tachyon EoM.
First, take a background at zero tachyon which reaches the fixed point as $r \to \infty$. Then consider turning on an ``infinitesimal'' tachyon. It satisfies the linearized tachyon EoM
\begin{equation}
\tau'' - \frac{3}{r} \tau' + \frac{2 \ell_*^2 a(\lambda_*)}{r^2\kappa(\lambda_*)} \tau = 0 \ , \qquad (r \to \infty) \ ,
\end{equation}
where $\lambda_*$ is the value of the coupling at the fixed point and $\ell_*$ is the IR AdS radius.
Inserting the Ansatz $\tau\sim r^{\Delta_*}$ yields
\begin{equation}
\Delta_*(4-\Delta_*) = \frac{2 \ell_*^2 a(\lambda_*)}{\kappa(\lambda_*)} = \frac{24 a(\lambda_*)}{V_\mathrm{eff}(\lambda_*) \kappa(\lambda_*)}
\end{equation}
where
\begin{equation}
V_\mathrm{eff}(\lambda) = V_g(\lambda)- x V_{f0}(\lambda)
\end{equation}
in terms of the dilaton and tachyon potentials. The BF bound is thus given by
\begin{equation}
\frac{24 a(\lambda_*)}{V_\mathrm{eff}(\lambda_*) \kappa(\lambda_*)} \le 4 \ .
\end{equation}
When the BF bound is violated we denote
\begin{equation} \label{nudef}
\nu = \mathrm{Im} \Delta_* = \sqrt{\frac{24 a(\lambda_*)}{V_\mathrm{eff}(\lambda_*) \kappa(\lambda_*)}-4} \ .
\end{equation}
In this case the asymptotic infinitesimal tachyon solution is oscillatory,
\begin{equation} \label{inftauIR}
\tau \sim r^2 \sin(\nu\log r + \phi)\ , \qquad (r \to \infty) \ .
\end{equation}
Let us consider next specific tachyon solutions $\tau_m$ and $\tau_\sigma$, where either an infinitesimal quark mass $m_q$ or a condensate $\sigma$ is turned on in the UV, respectively. These solutions are conveniently expressed in units of $\Lambda_\mathrm{UV}$, and will have the asymptotics of~\eqref{inftauIR}. We denote
\begin{align}
\label{taumdef}
\frac{\tau_m}{\ell} &\simeq \frac{m_q}{\Lambda_\mathrm{UV}} K_m \left(r \Lambda_\mathrm{UV}\right)^2\ \sin\left[\nu\log\left(r\Lambda_\mathrm{UV}\right) + \phi_m \right]\ ,& \quad &\left(r \gg \frac{1}{\Lambda_\mathrm{UV}}\right) & \\
\label{tausdef}
\frac{\tau_\sigma}{\ell} &\simeq \frac{\sigma}{\Lambda_\mathrm{UV}^3} K_\sigma \left(r \Lambda_\mathrm{UV}\right)^2\ \sin\left[\nu\log\left(r\Lambda_\mathrm{UV}\right) + \phi_\sigma \right]\ ,& \quad &\left(r \gg \frac{1}{\Lambda_\mathrm{UV}}\right) &
\end{align}
where the coefficients $K_i$ and $\phi_i$ cannot be computed analytically but it is easy to extract them from numerical solutions. One can require that $K_i>0$ and $-\pi/2 \le \phi_i<3\pi/2$.
We are, however, interested in the solutions where the tachyon is small and finite. In this case the tachyon will eventually grow large when $r \sim 1/\L_\tau \sim 1/\Lambda_\mathrm{IR}$, and drive the system away from the fixed point. The above formulas~\eqref{taumdef} and~\eqref{tausdef} then hold as approximations for $1/\Lambda_\mathrm{UV} \ll r \ll 1/\Lambda_\mathrm{IR}$.
The solution for $r \gtrsim 1/\Lambda_\mathrm{IR}$ depend on the details of the model
in the IR. However, in order to satisfy the boundary conditions imposed by the good IR singularity, the tachyon must have certain fixed normalization and phase when expressed in IR units:
\begin{equation} \label{tauIR}
\frac{\tau}{\ell} \simeq K_\mathrm{IR} \left(r \Lambda_\mathrm{IR}\right)^2\ \sin\left[\nu\log\left(r\Lambda_\mathrm{IR}\right) + \phi_\mathrm{IR} \right] \ , \qquad \left(\frac{1}{\Lambda_\mathrm{UV}} \ll r \ll \frac{1}{\Lambda_\mathrm{IR}}\right) \ .
\end{equation}
In general the IR asymptotics of the background depends on one parameter (e.g., $T_0$ for potentials~I~\cite{jk}) but as the fixed point is approached, the dependence on this parameter appears only through the ratio $\Lambda_\mathrm{UV}/\Lambda_\mathrm{IR}$ whereas $K_\mathrm{IR}$ and $\phi_\mathrm{IR}$ take fixed values (see Appendix~I in~\cite{Arean:2013tja}).
As the fixed point is approached, the result~\eqref{tauIR} must match with the sum of~\eqref{taumdef} and~\eqref{tausdef}. Therefore one finds
\begin{eqnarray}
&&\frac{m_q}{\Lambda_\mathrm{UV}} K_m \ \sin\left[\nu\log\left(r\Lambda_\mathrm{UV}\right) + \phi_m \right] + \frac{\sigma}{\Lambda_\mathrm{UV}^3} K_\sigma\ \sin\left[\nu\log\left(r\Lambda_\mathrm{UV}\right) + \phi_\sigma \right] \\\nonumber
&=& K_\mathrm{IR} \left(\frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}} \right)^2\ \sin\left[\nu\log\left(r\Lambda_\mathrm{UV}\right) + \nu\log\frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}}+ \phi_\mathrm{IR} \right]\ , \qquad \left(\frac{1}{\Lambda_\mathrm{UV}} \ll r \ll \frac{1}{\Lambda_\mathrm{IR}}\right) \ .
\end{eqnarray}
From here one can solve
\begin{equation} \label{spiraleqs}
\begin{split}
\frac{m_q}{\Lambda_\mathrm{UV}} &= \frac{K_\mathrm{IR} }{K_m}\,\frac{\sin\left(\phi_\mathrm{IR}-\phi_\sigma-\nu u \right)}{ \sin\left(\phi_m-\phi_\sigma\right)} \, e^{-2 u} \\
\frac{\sigma}{\Lambda_\mathrm{UV}^3} &= \frac{K_\mathrm{IR} }{K_\sigma }\,\frac{\sin\left(\phi_\mathrm{IR}-\phi_m-\nu u \right)}{\sin\left(\phi_\sigma-\phi_m\right)}\, e^{-2 u}
\end{split}
\end{equation}
where
\begin{equation}
u = \log\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}}\ .
\end{equation}
As $u$ varies the equations~\eqref{spiraleqs} define a spiral on the $(m_q,\sigma)$-plane, which has been studied recently at finite chemical potential in a different context (see~\cite{Iqbal:2011in}). Notice that $\nu$ does not need to be small. It can be verified numerically that the handedness of the spiral is such that its phase increases with increasing $u$ (counter clockwise direction) on the $(m_q,\sigma)$-plane. This means that
\begin{equation} \label{handedness}
\sin(\phi_m-\phi_\sigma) > 0 \ .
\end{equation}
As we shall show in Appendix~\ref{app:qbarqnorm}, this is also required in order for the standard solution (with nonzero and nodeless tachyon) to be dominant (for $x<x_c$ so that the BF bound is violated and the spiral exists).
Finally, let us point out some properties of the spiral as $x \to x_c$ from below. Notice from~\eqref{nudef} that $\nu = \morder{\sqrt{x_c-x}}$. The approximations~\eqref{taumdef} and~\eqref{tausdef} are valid for $1/\Lambda_\mathrm{UV} \ll r \ll 1/\Lambda_\mathrm{IR}$, but they should join smoothly with the UV asymptotics of the tachyon in~\eqref{mqsigmadef}. At $r \sim 1/\Lambda_\mathrm{UV}$ we find that
\begin{equation}
\frac{\tau_m}{\ell} (r \Lambda_\mathrm{UV})^{-2} \sim \frac{m_q}{\Lambda_\mathrm{UV}}\ K_m \sin\phi_m \ , \qquad
r \frac{d}{dr}\left[\frac{\tau_m}{\ell} (r \Lambda_\mathrm{UV})^{-2}\right] \sim \frac{m_q}{\Lambda_\mathrm{UV}}\ K_m \nu \cos\phi_m \ .
\end{equation}
These estimates, and similar estimates for the solution~\eqref{tausdef}, can be matched with the UV asymptotic formulas if
\begin{equation}
\frac{1}{K_m} \sim\frac{1}{K_\sigma} \sim \sin\phi_m \sim \sin\phi_\sigma \sim \sqrt{x_c-x} \ , \qquad (x \to x_c{}^-) \ .
\end{equation}
An analogous argument shows that for~\eqref{tauIR} to satisfy generic IR boundary conditions,
\begin{equation}
\frac{1}{K_\mathrm{IR}} \sim \sin\phi_\mathrm{IR} \sim \sqrt{x_c-x} \ , \qquad (x \to x_c{}^-) \ .
\end{equation}
We have found numerically that (with the convention $K_i>0$ and $-\pi/2 \le \phi_i<3\pi/2$)
\begin{equation}
\phi_m = \morder{\sqrt{x_c-x}} = \phi_\sigma \ , \qquad \phi_\mathrm{IR}-\pi = \morder{\sqrt{x_c-x}} \ ,
\end{equation}
(and $\phi_m-\phi_\sigma>0$ such that~\eqref{handedness} holds) for all potentials which we have studied. Then the first node of the mass in Eqs.~\eqref{spiraleqs} in their regime of validity $u \ll 1$ occurs at
\begin{equation} \label{firstnode}
\nu u = \nu \log \frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} = \phi_\mathrm{IR}-\phi_\sigma = \pi + \morder{\sqrt{x_c-x}} \ .
\end{equation}
This node is identified as the ``standard'' solution, where the tachyon has no nodes (in particular no nodes appear in the regime of validity of~\eqref{taumdef}, \eqref{tausdef}, and~\eqref{tauIR}). Notice that solving $\Lambda_\mathrm{UV}/\Lambda_\mathrm{IR}$ from~\eqref{firstnode} results in Miransky scaling\footnote{In principle it could be possible to satisfy the boundary conditions with $\phi_\mathrm{IR} = \morder{\sqrt{x_c-x}}$. We speculate that this happens in the model of~\cite{Evans:2013vca} where no Miransky scaling was found.} of~\eqref{Mscal}. Nodes at larger values of $u$, i.e., $\nu u \simeq (n+1) \pi$, with $n=1,2,\ldots$ are identified as the Efimov solutions, where the tachyon has $n$ nodes.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140803_Efimov_spiral_lin_x2.pdf}
\includegraphics[width=0.49\textwidth]{140803_Efimov_spiral_log_x2.pdf}
\end{center}
\caption{The dependence of $\sigma$ on the quarks mass for potentials~II with SB normalized $W_0$ at $x=2$. Left: the spiral of physical solutions on the $(m_q,\sigma)$-plane. Right: the same data after taking logarithms of both $\sigma$ and the quark mass. The red dots are numerical data while the blue curves are analytic fits. See text for details.}
\label{fig:spiral}\end{figure}
\subsection{Numerical results}
As an example, we have computed $m_q$ and $\sigma$ numerically for potentials~II with SB normalization\footnote{It is much easier to extract the Efimov spiral numerically for potentials~II than for potentials~I.} for $W_0$ at $x=2$. The results are shown in Fig.~\ref{fig:spiral}. The blue dots are our data and the red curves are given by Eqs.~\eqref{spiraleqs}. The blue line on the right hand side is a power-law fit $\sigma \simeq - C m_q^3$. The spiral structure is not well visible on the left hand plot because the distance of the curve from the origin decreases exponentially with increasing $u$ in~\eqref{spiraleqs}. In order to make the spiral structure visible, we have plotted the logarithms of (the absolute values of) $\sigma$ and $m_q$ on the right hand side of Fig.~\ref{fig:spiral}. Since the action is symmetric under $\tau \to -\tau$ there is actually also another spiral which not shown in the left hand plot but can be obtained simply by a rotation of 180 degrees around the origin.
Notice that when $m_q \ne 0$ it is difficult to define $\sigma$ unambiguously in practical calculations, because the vacuum expectation value (vev) solution of the tachyon cannot be separated from the subleading terms of the source solution.
This means that we also have to specify more carefully how $K_m$ and $\phi_m$ in~\eqref{taumdef} are defined in our numerical analysis:
we pick a reference solution which has $\sigma=0$ by definition, and consequently defines $K_m$ and $\phi_m$ (see also Appendix~\ref{app:qbarqnorm}). It is natural to expect that the $\sigma=0$ solution has a quark mass of \order{\Lambda_\mathrm{UV}} in the QCD regime. Therefore the reference solution was fixed to be the standard, dominant vacuum solution with $m_q/\Lambda_\mathrm{UV} =1$. This kind choice is important in order to avoid unnatural fine tuning effects.
The various coefficients in the solutions in~\eqref{spiraleqs} were determined as follows. Three solutions were chosen for the background for which $\sigma$ and $m_q$ are very small such that~\eqref{inftauIR} holds as a good approximation for at least two periods of oscillation. The first (second) solution was tuned to have approximately zero $\sigma$ ($m_q$) and was used to fit the constants $K_i$, $\phi_i$ in~\eqref{taumdef} [in~\eqref{tausdef}]. The third solution was used to fit the constants $K_\mathrm{IR}$ and $\phi_\mathrm{IR}$ in~\eqref{tauIR}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140803_Efimov_spiral_log_x3p6.pdf}%
\includegraphics[width=0.49\textwidth]{140803_sigma_vs_m_log_x4.pdf}
\end{center}
\caption{The dependence of $\sigma$ on the quarks mass for potentials~II with SB normalized $W_0$. Left: results in the walking regime, $x=3.6$. Right: results in the conformal window, $x=4$. The red dots are numerical data while the blue curves are fits based on Eqs.~\protect\eqref{spiraleqs}. See text for details.}
\label{fig:sigmacw}\end{figure}
In the walking regime, extra care is needed in the choice of the reference solution having $\sigma=0$. The chiral condensate is expected to have a node around such values of the quark mass where the normalizable and nonnormalizable terms in the tachyon solutions are nontrivially coupled. This happens for the standard vacuum at the crossover point between the regimes~A and~B, i.e., for
for
\begin{equation}
\frac{m_q}{\Lambda_\mathrm{UV}} \sim \exp\left[-\frac{2 K}{\sqrt{x_c-x}}\right] \ .
\end{equation}
The spiral in the walking regime ($x=3.6$, close to $x_c \simeq 3.7001$) is show on the log-log scale in Fig.~\ref{fig:sigmacw} (left). By studying numerically various quantities (for example the S-parameter, plotted below in Fig.~\ref{fig:massdepwalk}) it is found that the choice $m_q/\Lambda_\mathrm{UV} = 3 \times 10^{-11}$ lies at the crossover and is therefore chosen as the reference point with $\sigma=0$.
As the critical value $x_c$ is approached, the Efimov spiral becomes ``squeezed'', as seen by comparing plots in Fig.~\ref{fig:spiral} (right) and~\ref{fig:sigmacw} (left). In the walking regime, the consecutive solutions with vanishing $m_q$ and $\sigma$ are close so that the spiral approaches a straight line in the log-log scale. This reflects our analytic results for the spiral~\eqref{spiraleqs} derived above in the limit $x \to x_c{}^-$: in particular $\phi_m-\phi_\sigma \sim \sqrt{x_c-x} \to 0$.
We have also repeated the analysis in the conformal window. In this case the scaling of~\eqref{regimeBhighx} is expected to hold for small quark mass and that $\sigma \sim m_q^3$ for large quark mass. The results for potentials~II with SB normalization for $W_0$ at $x=4$ are shown in Fig.~\ref{fig:sigmacw} (right). The red dots are the data while the blue curves are given by the power laws mentioned above.
Again we need to specify how $\sigma$ is extracted in the numerical analysis because the tachyon vev solution cannot be separated from the subleading terms of the source solution. In the conformal window, however, the solution is simple because the values of $\sigma$ grow fast with increasing quark mass. Therefore one can choose any reference solution ``with $\sigma=0$'' at very small $m_q$, and the results are essentially independent of the choice.
Notice that the plots in the walking regime (left) and conformal window (right) do not seem too different. Indeed the curves undergo a smooth transition at $x=x_c$. The coefficients of the power laws are continuous at the transition, and the nodes where $\sigma=0$ or $m_q=0$ approach the origin of the spiral very fast as $x \to x_c$ from below. Actually the rate of the approach is given by the Miransky scaling law.
\subsection{Gell-Mann-Oakes-Renner relation} \label{sec:gmor}
The GOR relation can be obtained as usual by combining the results from two computations. The first result is the expression for the pion mass at small $m_q$, which is obtained by analyzing the fluctuation equations at small $m_q$ as done in Appendices~\ref{app:gor} and~\ref{app:fpi}. One finds that\footnote{The pion decay constant $f_\pi$ will be discussed in detail below in Sec.~\ref{sec:S}.}
\begin{equation} \label{intIres}
m_\pi^2 f_\pi^2 = 2 M^3 N_f N_c \ell^5 W_0 \kappa_0 m_q \sigma \left[1+\morder{\frac{m_q\Lambda_\mathrm{UV}^2}{\sigma}}\right]
\end{equation}
where $\kappa_0=\kappa(\lambda=0)$ and $W_0 = V_{f0}(\lambda = 0)$. The second result is the relation between $\sigma$ and the chiral condensate as $m_q \to 0$, which is obtained by deriving the renormalized on-shell action (i.e., the vacuum energy $\mathcal{E}$) with respect to $m_q$ (see Appendix~\ref{app:qbarqnorm}):
\begin{equation} \label{qbarqnorm}
\langle \bar q q \rangle = \frac{\partial \mathcal{E}}{\partial m_q} = -2 M^3 N_f N_c \ell^5 W_0 \kappa_0 \sigma\left[1+\morder{\frac{m_q\Lambda_\mathrm{UV}^2}{\sigma}}\right] \ .
\end{equation}
The combination is the GOR relation:
\begin{equation} \label{GORtext}
m_\pi^2 f_\pi^2 = -m_q \langle \bar q q \rangle \left[1+\morder{\frac{m_q\Lambda_\mathrm{UV}^2}{\sigma}}\right] \ .
\end{equation}
It can be checked that the proportionality coefficient (here minus one) is correct for our normalization of $f_\pi^2$, which differs by the factor $N_f/2$ from the standard normalization in chiral perturbation theory (see, e.g.,~\cite{Bijnens:2009qm}).
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141208_GMOR_comparison.pdf}
\includegraphics[width=0.49\textwidth]{141208_GMOR_correction.pdf}
\end{center}
\caption{Numerical test of the GOR relation. Left: Both sides of the GOR relation plotted against the quark mass. The blue curve is $m_q\langle \bar q q\rangle$ (in units of $\Lambda_\mathrm{UV}^4$), while the red dashed curve is $f_\pi^2 m_\pi^2$. Right: The leading correction to the relation. The blue curve is the numerical data, and the thin dashed red line is a linear fit ($2.5\ m_q/\Lambda_\mathrm{UV}$). Potentials I with $x=1$ and $W_0=3/11$ were used.}
\label{fig:GOR}\end{figure}
Notice that
\begin{enumerate}
\item One might expect that the logarithmic terms which appear in the UV expansion of the fields would result in correction that are only suppressed by $1/\log m_q$ in~\eqref{intIres}. Remarkably, the logarithmic terms completely cancel, as they should, since the QCD the relation holds up to linear corrections in the quark mass. Why this happens is shown in Appendix~\ref{app:gor}. The cancellation of these corrections is also consistent with the combination $m_q \langle\bar qq\rangle$ being RG invariant~\cite{Kiritsis:2014kua}.
\item The derivative in~\eqref{qbarqnorm} is nontrivial, since changing the quark mass also affects the geometry even at $m_q=0$ due to the full backreaction between the flavor and glue sectors, which may add contributions to the derivative (see Appendix~\ref{app:qbarqnorm}).
\item The relation is valid only in regime~A: the correction terms in~\eqref{GORtext} become large at the crossover between regimes~A and~B.
\end{enumerate}
We have tested the GOR relation numerically for potentials~I at $x=1$. Both sides\footnote{To be precise, we are checking equation~\eqref{intIres} -- the normalization factor of~\eqref{qbarqnorm} was assumed in the computation and subleading corrections were dropped.} of the relation are plotted as functions of the quark mass in Fig.~\ref{fig:GOR} (left) and good agreement is found. The subleading correction in Fig.~\ref{fig:GOR} (right) is clearly linear in the quark mass as it should.
\section{Four-fermion operators in V-QCD} \label{sec:4f}
After we have constructed the solutions on the $(m_q,\sigma)$-plane it is straightforward to analyze the effect of multitrace deformations following the recipe of~\cite{Witten:2001ua}. In the presence of such deformations, the coefficient of the source term, which was denoted by $m_q$ above, is no longer trivially related to the quark mass. Let us therefore denote this coefficient by $\alpha_m$ in this section. Similarly, the coefficient of the normalizable term is denoted by $\beta_m$.
One can now study the following extra terms on the field theory Lagrangian:
\begin{equation} \label{Wdef}
\frac{1}{N_fN_c } \widetilde W = -m_q \int d^4 x \mathcal{O} + \sum_{n=2}^{n_\mathrm{max}} \frac{(-1)^ng_n}{n} \int d^4 x \mathcal{O}^n \ ,
\end{equation}
where $\mathcal{O} = \bar q q/N_fN_c$.
We first replace $\mathcal{O}$ by $-c_\sigma \beta_m$ in $\widetilde W$, where $c_\sigma = 2 M^3 W_0 \kappa_0 \ell^5$ is the proportionality coefficient between $\sigma$ and $\langle \mathcal{O} \rangle$ in the absence of the multi-trace deformations\footnote{For simplicity we will omit here the difficulty of defining $\beta_m$ in practice: we take it to be directly proportional to $\langle \bar q q\rangle$ (so that $k_\tau=0$ in~\eqref{qqbarmqfin} of Appendix~\ref{app:qbarqnorm}).}.
The result is the functional
\begin{equation} \label{Wbulk}
\frac{1}{N_fN_c } W[\beta_m] = m_q c_\sigma \int d^4 x \beta_m + \sum_{n=2}^{n_\mathrm{max}} \frac{g_n c_\sigma^n}{n} \int d^4 x \beta_m^n \ .
\end{equation}
Then the boundary conditions are obtained
by setting the source term $\alpha_m$ to the value~\cite{Witten:2001ua} (see also~\cite{Mueck:2002gm,Minces:2002wp})
\begin{equation} \label{alphacond}
\alpha_m = \frac{1}{N_f N_c c_\sigma} \frac{\delta W}{\delta \beta_m(x)} = m_q +\sum_{n=2}^{n_\mathrm{max}} g_n c_\sigma^{n-1}\beta_m^{n-1} \ ,
\end{equation}
and the vev is given by
\begin{equation} \label{betacond}
\beta_m = \sigma \equiv -\frac{1}{c_\sigma} \langle \mathcal{O}\rangle = -\frac{1}{N_fN_cc_\sigma}\langle \bar q q\rangle \ .
\end{equation}
The possible vacua can then be identified by overlapping this condition with the curves of regular solutions on the $(\alpha_m,\beta_m)$-plane (shown above in Figs.~\ref{fig:spiral} and~\ref{fig:sigmacw}).
\begin{figure}[!tb]
\begin{center}
\includegraphics[height=0.4\textwidth]{140905_four_fermi_example.pdf}%
\hspace{5mm}\includegraphics[height=0.4\textwidth]{141222_four_fermi_large_cond.pdf}
\end{center}
\caption{Solutions in the presence of a double trace deformation in the QCD regime.}
\label{fig:doubletr}\end{figure}
We are most interested in the case of double trace deformation, $g_2\ne 0$ with other couplings equal to zero, since this operator becomes marginal at the critical point $x=x_c$. Let us also set $m_q=0$. In this case
\begin{equation} \label{g2constr}
\alpha_m = g_2 c_\sigma \beta_m \ .
\end{equation}
The overlap plot is shown for the phase with the Efimov spiral ($0<x<x_c$)
in Fig.~\ref{fig:doubletr}. Notice that the second branch of the spiral, obtained by reflection about the origin,
which was omitted in earlier plots (for example Fig.~\ref{fig:spiral}) was now also included\footnote{The solutions on the second branch are nontrivially related to those on the first branch only if operators with odd $n$ have been included.}.
For this special case where $g_2$ is the only nonzero coupling, as the mapping $(\alpha_m,\beta_m) \mapsto (m_q,\sigma)$ in~\eqref{alphacond} and~\eqref{betacond} implies, the change of the UV boundary conditions with respect to the standard case $g_2=0$ corresponds to changing the vertical axis to the line defined by~\eqref{g2constr} (examples are the red and dashed magenta lines in Fig.~\ref{fig:doubletr}) while the horizontal axis is kept fixed.
Notice that the solutions with $g_2=0$ lie on the vertical axis, and are denoted by the green line in Fig.~\ref{fig:doubletr}. The green dot shows the (stable) standard solution, but there is also an infinite tower of additional intersection points near the origin, which are not visible as the spiral converges very fast towards zero. These intersection points give the Efimov solutions.
In order to draw the phase diagram at nonzero $g_2$, we need to solve the free energy of each solution and find the dominant vacuum. We can start from the identity
\begin{equation} \label{Eder}
\frac{\partial\mathcal{E}}{\partial m_q} = \langle \bar q q \rangle = - N_f N_c c_\sigma \sigma \ .
\end{equation}
In Appendix~\ref{app:qbarqnorm} we show that the conditions~\eqref{alphacond} and~\eqref{betacond} are indeed consistent with~\eqref{Eder} and that the higher order expectation values satisfy
\begin{equation} \label{Egder}
\frac{n (-1)^{n+1}}{N_fN_c} \frac{\partial\mathcal{E}}{\partial g_n} = \langle \mathcal{O}^n \rangle = \langle \mathcal{O} \rangle^n
\end{equation}
in agreement with the large $N$ factorization of expectation values.
Recall from Sec.~\ref{sec:condensate} that the Efimov spiral can be parametrized as
\begin{equation} \label{spiraleqsab}
\begin{split}
\frac{\alpha_m}{\Lambda_\mathrm{UV}} &= \frac{K_\mathrm{IR} }{K_m}\,\frac{\sin\left(\phi_\mathrm{IR}-\phi_\sigma-\nu u \right)}{ \sin\left(\phi_m-\phi_\sigma\right)} \, e^{-2 u} \\
\frac{\beta_m}{\Lambda_\mathrm{UV}^3} &= \frac{K_\mathrm{IR} }{K_\sigma }\,\frac{\sin\left(\phi_\mathrm{IR}-\phi_m-\nu u \right)}{\sin\left(\phi_\sigma-\phi_m\right)}\, e^{-2 u}
\end{split}
\end{equation}
asymptotically at small values of the variable
\begin{equation}
u = \log\frac{\Lambda_\mathrm{UV}}{\Lambda_\mathrm{IR}} \ .
\end{equation}
Notice that if only $g_2$ is nonzero, $m_q$ and $\sigma$ still satisfy~\eqref{spiraleqsab} (with $\alpha_m$ ($\beta_m$) replaced by $m_q$ ($\sigma$), respectively), if the coefficients $K_i$ and $\phi_i$ are redefined. This is seen by inserting the conditions~\eqref{alphacond} and~\eqref{betacond}, and is consistent with the change of boundary conditions simply corresponding to a new choice of axes in Fig.~\ref{fig:doubletr}.
By inserting this parametrization in~\eqref{Eder} and integrating along the Efimov spiral we find that for the solutions with $m_q=0$ (see Appendix~\ref{app:qbarqnorm})
\begin{equation}
\frac{1}{N_f N_c} \left(\mathcal{E}-\mathcal{E}_0\right) = - \frac{\nu c_\sigma K_\mathrm{IR}^2 e^{-4 u}}{8 K_m K_\sigma \sin(\phi_m-\phi_\sigma)} \ ,
\end{equation}
where $\mathcal{E}_0$ is the free energy of the solution with $\alpha_m = 0 = \beta_m$. Since $\sin(\phi_m-\phi_\sigma)>0$, the dominant solution from those in the range of validity of~\eqref{spiraleqsab} is that with the largest value of $u$. In the walking regime this can be seen explicitly. Namely, we argued in Sec.~\ref{sec:condensate} that in the walking regime $\nu \simeq \pi\sqrt{(x_c-x)}/K \to 0$, and that the solutions are found at $\nu u \simeq (n+1) \pi$, with $n=0,1,2,\ldots$. Therefore
\begin{equation}
\frac{1}{N_f N_c} \left(\mathcal{E}-\mathcal{E}_0\right) \simeq - \frac{\nu c_\sigma K_\mathrm{IR}^2}{8 K_m K_\sigma \sin(\phi_m-\phi_\sigma)} \exp\left(-\frac{4 K(n+1)}{\sqrt{x_c-x}}\right) \ , \qquad (x_c-x \ll 1) \ ,
\end{equation}
which agrees with the scaling of the free energy found in~\cite{jk} (see Sec.~10 and Appendix~H there).
Also quite in general the solution with largest $|\sigma|$ is the dominant vacuum. From~\eqref{Eder} we see that the energy density is given in terms of the (oriented) area between the spiral and the horizontal axis. For clockwise oriented spirals the minimum energy is reached at the solutions furthest away from the origin.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.6\textwidth]{141222_four_fermi_phases.pdf}
\end{center}
\caption{The phase diagram of V-QCD in the presence of a four-fermion deformation (at zero quark mass and temperature). The blue horizontal line is a discontinuity at $g_2 = 0^-$. The dotted vertical line presents a BKT transition.}
\label{fig:4fphases}\end{figure}
It is then straightforward to construct the phase diagram, which is shown in Fig.~\ref{fig:4fphases} for the case of $m_q=0$ and nontrivial $g_2$.
For $x<x_c$ the diagram can be found simply by analyzing the solutions in Fig.~\ref{fig:doubletr}. When $g_2>0$ the configuration is qualitatively similar to that at $g_2=0$: the dominant vacuum for $x_c<x$ is the vacuum with the standard tachyon solution having no nodes.
When $g_2<0$,
the dominant solution is that of Fig.~\ref{fig:doubletr} (right) having sizeable $|\sigma|$. As this solution is absent for $g_2 \ge 0$, there is a discontinuity (a ``zeroth order'' transition) at $g_2 = 0 -$. It is smoothly connected, though, to the dominant solution of $g_2 > 0$ through the limit $g_2 \to \pm \infty$ (where the constraint~\eqref{g2constr} gives a horizontal line in the plots of Fig.~\ref{fig:doubletr}). There is also a subdominant solution shown in Fig.~\ref{fig:doubletr} (left), namely the continuation of the standard, dominant solution at $g_2=0$ to negative $g_2$. We will argue below that this solution is metastable for small $|g_2|$ and unstable for large $|g_2|$.
Notice that the chiral condensate is a scheme dependent quantity, and the scheme dependence is important, in particular, at large $m_q$ (see~\cite{ikp} for a discussion of the scheme dependence in the context of holography). Therefore the dominant solution $g_2<0$ appears scheme dependent at least for small $|g_2|$. We have, however, found the behavior of generic schemes $\sigma \propto m_q^3$ with a negative proportionality constant for all choices of potentials we have tried, and independently of the precise definition of $\sigma$. This is enough to guarantee that the phase diagram is that of Fig.~\ref{fig:4fphases}. Actually, as stressed in Appendix~\ref{app:masses}, the results at very large $m_q$ are essentially independent of the choices of the various functions in the V-QCD action.
In the conformal window, the situation is even simpler. For $g_2 \ge 0$ there is only the solution with zero quark mass and the condensate. When $g_2<0$, there is again a solution with large $|\sigma|$, analogous to that shown in Fig.~\ref{fig:doubletr} (right). When $g_2>0$, the transition at $x=x_c$ is similar as at $g_2=0$, i.e., a BKT transition, since changing the value of $g_2$ only amounts to changing the axes in Fig.~\ref{fig:doubletr} without affecting the structure of the spiral.
One can also show that when a finite $m_q$ is turned on, the BKT transition and the chirally symmetric phase disappear, but the discontinuity at $g_2= 0^-$ remains.
\subsection{Perturbative stability}
Finally perturbative stability of the solutions with modified boundary conditions could be analyzed following~\cite{Arean:2013tja}. Here we will only discuss which of the solutions are expected to be unstable, and will not prove the stability of any solution. Naturally, the modification of the boundary conditions for the background also implies that the boundary conditions of the fluctuations are similarly changed as the fluctuations must preserve the physical value of $m_q$.
Let us first analyze any solutions with walking, i.e., $u \gg 1$. Recall that the standard solution has been shown to be stable, while the Efimov solutions are unstable when $g_2=0$~\cite{Arean:2013tja}. The instability appeared in the scalar flavor singlet and nonsinglet sectors.
In order to study stability of the other solutions, one should look at the scalar fluctuation equations. In the walking regime, they admit simple solutions in the UV and in the vicinity of the (approximate IR) fixed point. In fact, as argued in Appendix~I of~\cite{Arean:2013tja}, the fluctuation equations (for sufficiently small mass of the fluctuation) take the same form as the EoMs for the background. This is true both in the flavor singlet and in the nonsinglet sectors. Therefore, the fluctuations in the vicinity of the (approximate) IRFP are analogous to~\eqref{tauwalks}:
\begin{equation} \label{phiwalks}
\psi_S(r) \simeq C_s r^2 \sin\left[\nu\log(r\Lambda_\mathrm{UV}) + \phi_s \right] \ , \qquad \left(\frac{1}{\Lambda_\mathrm{UV}} \ll r \ll \frac{1}{\L_\tau}\right) \ ,
\end{equation}
where $\psi_S$ is the radial wave function of any scalar fluctuation mode.
The solution necessarily has a node if this approximation is valid for more than half a period of the sine function. In terms of variable $u = \log(\Lambda_\mathrm{UV}/\Lambda_\mathrm{IR})$ the node therefore appears for $\nu u \gtrsim \pi$ (where we used the fact that $\Lambda_\mathrm{IR} \sim \L_\tau$ whenever walking is present).
Such a node of the wave function (say at zero momentum) implies an instability, for generic
UV boundary conditions for the fluctuations. It is straightforward to prove this in the case of nonsinglet scalars for which the fluctuation equations can be cast into the Schr\"odinger form, and with the standard UV boundary conditions such that the fluctuation wave function is normalizable in the UV (see also the analysis in~\cite{Anguelova:2013tha}). We shall sketch the proof here.
Denote the UV normalizable (but not necessarily IR normalizable) Schr\"odinger wave function by $\phi$ and the location of any of its nodes by $r_0$. By studying the variation of
\begin{equation}
0 = \int_0^{r_0(m^2)} \phi(r) (-\phi''(r) + V_S(r)\phi(r) - m^2 \phi(r)) dr
\end{equation}
with respect to the mass, we find that
\begin{equation} \label{nodeevol}
\frac{d r_0}{dm^2} = - \frac{1}{\phi'(r_0)^2} \int_0^{r_0} \phi^2(r) dr < 0 \ .
\end{equation}
We see that all nodes move towards the IR as $m^2$ is lowered. But as $m^2 \to -\infty$ the solution for generic UV boundary conditions is $\phi \propto \exp(|m|r)$ and has no nodes for $r \ll 1/m$. Therefore all nodes must disappear either by moving to $r=\infty$ or by merging with other nodes. However, the Schr\"odinger equation does not admit solutions with a double node for regular potentials $V_S$, so merging of the nodes is not possible, and consequently all nodes must disappear by moving to the far IR. In particular, if there the wave function has a node when $m^2=0$, the node must move to $r = \infty$ at some negative value of $m^2$. At this value, $\phi$ is IR normalizable: otherwise it could not have a node in the far IR if $m^2$ is slightly perturbed. Tachyonic normalizable mode marks the presence of instability. Putting the above observations together, we conclude that a node of the $m^2=0$ wave function marks the presence of an instability.
The above proof does not apply directly to our case because the UV boundary conditions are modified, and consequently the integral in~\eqref{nodeevol} is divergent. The divergence can be regulated by introducing a UV cutoff at $r=\epsilon$, but this results in extra counterterms on the right hand side of~\eqref{nodeevol} and its negativity is no longer obvious.
The proof can be fixed, however, in our case (i.e., scalar fluctuations and node in the region of validity of~\eqref{phiwalks}) when $x \to x_c$ from below. This is because the node becomes well separated from the UV region: the location of the node satisfies $\log (r_0/\Lambda_\mathrm{UV}) \sim 1/\sqrt{x_c-x}$ as seen from~\eqref{phiwalks}, where $\nu \sim \sqrt{x_c-x}$. For generic boundary conditions, the right hand side of~\eqref{nodeevol} is then dominated by the contributions to the integral near the node. The counterterms which cancel the divergence of the integral are essentially independent of $x$ for any reasonable boundary conditions, and therefore negligible. The rest of the proof remains unchanged.
Let us then study the solutions with $m_q=0$ as $x \to x_c$. Recall that the phase differences $\phi_\mathrm{IR}-\phi_\sigma-\pi$ and $\phi_\sigma-\phi_m$ will be $\propto \sqrt{x_c-x}$ in this limit as we argued in Sec.~\ref{sec:condensate}. The standard solution (the green dot on the vertical axis in Fig.~\ref{fig:doubletr}) is in the regime of validity of the approximations leading to~\eqref{spiraleqsab}. It is found at
\begin{equation} \label{standardvac}
\nu u = \phi_\mathrm{IR}-\phi_\sigma = \pi + \morder{\sqrt{x_c-x}} \ .
\end{equation}
Based on the above analysis, the onset of the perturbative instability is also expected at $\nu u - \pi = \morder{\sqrt{x_c-x}}$. As the standard solution is stable, the critical value of $u$ must be larger than that given in~\eqref{standardvac}. Inserting this in~\eqref{spiraleqsab}, the critical value of $g_2$ is given by the ratio $\alpha_m/(\beta_m c_\sigma)$. The factors of $\sqrt{x_c-x}$ cancel leaving a \order{1/\Lambda_\mathrm{UV}^2} number, which must be negative given the handedness of the spiral. The critical value of $g_2$ is therefore of the same order as the value corresponding to the dashed magenta line in Fig.~\ref{fig:doubletr}. The parts of the spiral which are closer to the origin from the critical points (near the intersection points with magenta dots) are unstable.
The above analysis cannot be applied in the QCD regime where $x_c-x$ is not small. This regime could be analyzed numerically. The natural expectation is that the critical value is still negative and \order{1/\Lambda_\mathrm{UV}^2}. This is supported by the fact that the Efimov vacua are unstable, which has been shown by a numerical computation~\cite{Arean:2013tja}.
To conclude this section, the phase diagram is that shown in Fig.~\ref{fig:4fphases}: For $g_2>0$, the diagram is identical to that of $g_2=0$. The result is somewhat different from the result obtained in the gauged Nambu-Jona-Lasinio model, where the four-fermion operator is also slightly different, and the lower bound of the conformal window appears to increase with $g_2$~\cite{Yamawaki:1996vr}. As any negative $g_2$ is turned on, the ``standard'' vacuum immediately becomes unstable and the dominant vacuum has much larger $|\sigma|$. For $x<x_c$ and $g_2<0$, the perturbation analysis suggests the the standard vacuum is metastable for small $|g_2|$ and becomes perturbatively unstable for larger $|g_2|$, with the critical value being \order{1/\Lambda_\mathrm{UV}^2}.
\section{S-parameter and current-current correlators} \label{sec:S}
The vector-vector and axial-axial correlators have the structure
\begin{align}
i \int d^4 x\ e^{-i q x} \langle 0|\ T\!\left\{ J_{\mu}^{a\,(V)}(x) J_{\nu}^{b\,(V)}(0)\right\} |0\rangle &=- \frac{2\delta^{ab}}{N_f}
\left( q^2 \eta_{\mu\nu}-{q_{\mu} q_{\nu}}\right)\Pi_V (q^2) &\\
i \int d^4 x\ e^{-i q x} \langle 0|\ T\!\left\{ J_{\mu}^{a\,(A)}(x) J_{\nu}^{b\,(A)}(0)\right\} |0\rangle &=- \frac{2\delta^{ab}}{N_f}\Big[
\left( q^2 \eta_{\mu\nu}-{q_{\mu} q_{\nu}}\right)\Pi_A (q^2) \\\nonumber
& \qquad \qquad \ \ + q_{\mu} q_{\nu}\Pi_L (q^2)\Big] \ , &
\end{align}
where
\begin{equation}
J_{\mu}^{a\,(V)} = \bar q \gamma_\mu t^a q \ , \qquad J_{\mu}^{a\,(A)} = \bar q \gamma_\mu\gamma_5 t^a q \ .
\end{equation}
and $t^a$ are the generators of $SU(N_f)$ with the normalization ${\mathbb Tr}\, t^a t^b =\delta^{ab}/2$.
The factors $2/N_f$ on the right hand side were added to ensure that $\Pi_V$ and $\Pi_A$ are proportional to $N_f$. The numerical factor was chosen such that these factor are equal to one when $N_f=2$, which is the smallest number for which the flavor non-singlet currents are defined. This will result in the standard normalization of the S-parameter. The sign convention is $\eta_{\mu\nu} = \mathrm{diag}(-,+,+,+)$. Notice that $\Pi_L$ vanishes for zero quark mass because then $\partial^\mu J_{\mu}^{a\,(A)} =0$.
When the spectrum is discrete (i.e., $0<x<x_c$ or $m_q$ is finite), we may formally write the correlators as sums over the contributions from the meson states:
\begin{eqnarray} \label{PiAsp}
\Pi_A &=& \frac{S_0}{q^2} + \sum_{n=1}^\infty \frac{(f_{n}^{(A)})^2}{q^2+(m_{n}^{(A)})^2} \\
\label{PiVsp}
\Pi_V &=& \sum_{n=1}^\infty \frac{(f_{n}^{(V)})^2}{q^2+(m_{n}^{(V)})^2} \\
\label{PiLsp}
\Pi_L &=& \frac{S_0}{q^2} - \sum_{n=1}^\infty \frac{(f_{n}^{(P)})^2}{q^2+(m_{n}^{(P)})^2} \ .
\end{eqnarray}
Depending on the choice of the holographic action, these series may not converge. This issue will be discussed below in the context of the S-parameter.
The residues $S_0$ of the ``spurious'' $q^2=0$ pole in~\eqref{PiAsp} and~\eqref{PiLsp} must identical for the pole to cancel in the full correlator. At zero quark mass also $S_0$ must be related to the pion decay constant, $S_0 = f_\pi^2 \equiv (f_1^{(P)})^2$, since $\Pi_L$ vanishes.
The difference of the vector-vector and axial-axial correlators involves the quantity
\begin{equation} \label{Ddef}
D(q^2) = q^2 \Pi_A(q^2)- q^2 \Pi_V(q^2) \ ,
\end{equation}
which is nontrivial only when chiral symmetry is broken. The expansion of $D(q^2)$ at $q^2=0$ defines the S-parameter:
\begin{equation} \label{Sdef}
D(q^2) = S_0 - \frac{S}{4\pi} q^2 + \frac{S_2}{4\pi}q^4 + \cdots \ .
\end{equation}
Here $S_0$ is the same coefficient which appears in~\eqref{PiAsp} and~\eqref{PiLsp}, and we will also study the higher order coefficient $S_2$ below.
\subsection{Correlators and the S-parameter in V-QCD}
Let us then recall how the correlators can be computed in V-QCD. As pointed out in Sec.~\ref{sec:vqcd}, the vector currents are dual to the gauge fields in the DBI action~\eqref{generalact}. Following the standard approach (see~\cite{Arean:2013tja} for additional details), we carry out the fluctuation analysis writing down an Ansatz which separates the spatial and radial dependence of the fluctuation modes in momentum space. The spatial (radial) wave functions of the vector, transverse axial, and longitudinal axial modes are denoted by $\mathcal{V}$, $\mathcal{A}$ and $\mathcal{P}$ ($\psi_V$, $\psi_A$, and $\psi_L$). The radial wave functions are IR normalizable and satisfy the UV boundary conditions
\begin{equation}
1=\psi_V(\epsilon,p^2)=\psi_A(\epsilon,p^2)=\psi_L(\epsilon,p^2)-\psi_P(\epsilon,p^2) \ .
\end{equation}
Here $\psi_P$ is the radial pion wave function.
The terms of the on-shell V-QCD action which are quadratic in the vector fields can then be written as
\begin{eqnarray}
S_V &=& \frac{1}{4} M^3 N_c \int \frac{d^4 p}{(2\pi)^4} \mathcal{V}_\mu^{a}(p) P^{\mu\nu} \mathcal{V}_\nu^{a}(-p) \left.V_f e^A w^2 \partial_r \psi_V(r,p^2)\right|_{r=\epsilon}\ , \\
S_A &=& \frac{1}{4} M^3 N_c \int \frac{d^4 p}{(2\pi)^4} \mathcal{A}_\mu^{a}(p) P^{\mu\nu} \mathcal{A}_\nu^{a}(-p) \left.V_f e^A w^2 \partial_r \psi_A(r,p^2)\right|_{r=\epsilon}\ , \\
S_L &=& \frac{1}{4} M^3 N_c \int \frac{d^4 p}{(2\pi)^4} \mathcal{P}^{a}(p) \mathcal{P}^{a}(-p) \left.V_f e^A p^2 w^2 \partial_r \psi_L(r,p^2)\right|_{r=\epsilon} \ ,
\end{eqnarray}
where
\begin{equation}
P^{\mu\nu} = \eta^{\mu\nu} - \frac{p^\mu p^\nu}{p^2}
\end{equation}
projects to the transverse parts of the wave functions. To compute the vector-vector correlators, we need the precise dictionary given in terms of the couplings to field theory currents:
\begin{eqnarray}
&&\int \frac{d^4p}{(2\pi)^4} J_\mu^{a\,(V)}(-p) \mathcal{V}^{\mu\, a}(p)\ , \\
&&\int \frac{d^4p}{(2\pi)^4} J_\mu^{a\,(A)}(-p)\left[P^{\mu\nu}\mathcal{A}_\nu^{a}(p) - i p^\mu \mathcal{P}^a(p)\right] \ .
\end{eqnarray}
Applying the gauge/gravity correspondence with these couplings leads to the following expressions for the form factors:
\begin{equation} \label{Pidef}
q^2 \Pi_I(q^2) = -\frac{1}{4} M^3 N_f N_c \left.V_f e^A w^2 \partial_r \psi_I(r,q^2)\right|_{r=\epsilon} \ ,
\end{equation}
where $I=V,A,L$.
Notice that for $m_q>0$ (so that $\Pi_L$ is nonzero) the wave functions $\psi_A$ and $\psi_L$ satisfy the same fluctuation equations as $q^2 \to 0$. Consequently
\begin{equation}
\lim_{q\to 0} q^2 \Pi_A(q^2) = \lim_{q\to 0} q^2 \Pi_L(q^2)
\end{equation}
and this number equals $S_0$ of~\eqref{PiAsp} and~\eqref{PiLsp} which is therefore well defined. This equality ensures the cancellation of the ``pole'' at $q^2 = 0$ as pointed out above.
By inserting~\eqref{Pidef} in the definitions~\eqref{Ddef} and~\eqref{Sdef} we obtain for the S-parameter
\begin{equation} \label{Sdiffform}
S = \pi M^3 N_f N_c V_f e^A w^2 \left[ \frac{\partial^2}{\partial r\partial q^2} \psi_A(r,q^2)-\frac{\partial^2}{\partial r\partial q^2} \psi_V(r,q^2) \right]_{r=\epsilon,\,q^2=0} \ .
\end{equation}
This formula is, however, not convenient for high precision numerical computations, because the subleading terms at $r=\epsilon$ are only suppressed by logarithms of $\epsilon$ in V-QCD, so that extremely small values of $\epsilon$ would be needed to obtain accurate results. By an analysis of the fluctuation equations, it is possible to derive more convenient integral representations for $S$ (see Appendix~\ref{app:fpi}). We find that
\begin{equation} \label{Dint}
D(q^2) = q^2 \Pi_A(q^2)- q^2 \Pi_V(q^2) = M^3 N_f N_c \int_0^\infty du\ \psi_V V_f e^{3A} \kappa \tau^2 \psi_A \ ,
\end{equation}
where $u$ is the Schr\"odinger coordinate defined in Appendix~\ref{app:fpi}, and that
\begin{align}
S &= \pi M^3 N_f N_c \int_0^\infty du \ V_f e^A w^2 \left[\psi_A^2-\psi_V^2\right]_{q^2=0} \\
&= \pi M^3 N_f N_c \int_0^\infty du \ V_f e^A w^2 \left[\psi_A^2-1\right]_{q^2=0} \ .
\label{Sparamint}
\end{align}
The formula for the S-parameter is well-known in the context of simple bottom-up models (see, e.g.,~\cite{Barbieri:2003pr}). In Appendix~\ref{app:fpi} we write it in a form which holds for very generic holographic models. We used the fact that
\begin{equation}
\left.\psi_V\right|_{q^2=0} = 1
\end{equation}
in order to obtain the last expression~\eqref{Sparamint}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_S_parameter_vs_x_mq_10m6_and_0.pdf}
\includegraphics[width=0.49\textwidth]{140717_S_parameter_vs_x_mq_10m6_and_0_potII.pdf}
\end{center}
\caption{The normalized S-parameter as a function of $x$ for $m_q/\Lambda_\mathrm{UV}=10^{-6}$ (blue solid curves) and $m_q=0$ (red dashed curves). Left: potentials~I with $W_0=3/11$. Right: potentials~II with SB normalized $W_0$.}
\label{fig:Sparam}\end{figure}
\subsection{Numerical results for the S-parameter and $f_\pi$}
Using equations~\eqref{Pidef}, \eqref{Dint}, and~\eqref{Sparamint} it is straightforward to compute the S-parameter, the pion decay constant as well as the coefficients $S_0$ and $S_2$ of~\eqref{Sdef} numerically (see~\cite{Arean:2013tja} for additional details, and~\cite{Sparam,cobi,parnachev} for analysis of the S-parameter in other holographic models). Figure~\ref{fig:Sparam} shows the results for $S$ at $m_q=0$ (dashed red curves) and at $m_q/\Lambda_\mathrm{UV}=10^{-6}$ (blue curves) for potentials~I (left) and~II (right). The numerical value of $M^3$ was fixed such that the asymptotics of the vector-vector correlator matches with perturbative QCD (see Eq. (C.10) in~\cite{Arean:2013tja}).
The most striking feature in these plots is the discontinuity of the S-parameter in the conformal window. When $x \ge x_c$, the S-parameter immediately jumps from zero to a \order{N_f N_c} number when any finite $m_q$ is turned on. The mechanism which leads to this discontinuity will be discussed in detail below (from the holographic viewpoint), but it appears rather natural: the S-parameter is \order{N_fN_c} whenever the geometry has the IR singularity, and vanishes only for zero quark mass in the conformal window where there is an IRFP instead. The result is also consistent with the analysis based on field theory at qualitative level~\cite{sannino,DelDebbio:2010ze}: the S-parameter is finite except for exactly zero mass in the conformal window.
There is, however, one striking difference~\cite{letter,Arean:2013tja} with respect to previous results: the S-parameter increases with $x$ in regime~A, whereas many earlier analyses~\cite{SinDS,walkingS,as,Sparam} suggest that the S-parameter is suppressed in the walking regime and may even vanish as $x \to x_c$. Recall, however, that the IR behavior of the potentials in the V-QCD action has not yet been fitted to QCD or lattice data, and such fits may affect the $x$-dependence of the S-parameter. By analyzing the form of the fluctuation wave functions (see Appendix~G in~\cite{Arean:2013tja}) in the integral formula~\eqref{Sparamint} one can indeed check that it is dominated in the IR (for small $m_q$ and for all values of $x$). This is consistent with the analysis of Appendix~I of~\cite{Arean:2013tja}, where the same is argued to hold for the meson masses and decay constants.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_fpi_vs_x_mq_10pm6.pdf}
\includegraphics[width=0.49\textwidth]{140717_fpi_vs_x_mq_10pm6_potII.pdf}
\end{center}
\caption{The pion decay constant as a function of $x$ for $m_q/\Lambda_\mathrm{UV}=10^{-6}$. Left: potentials~I with $W_0=3/11$. Right: potentials~II with SB normalized $W_0$. The (normalized and squared) pion decay constant is shown as the blue curves, and the values of the constant $\sqrt{S_0}$ are also shown as the red dashed curves for comparison. }
\label{fig:fpi}\end{figure}
We have also computed the pion decay constant which is defined in terms of the residue of $\Pi_L$ at the pion mass whenever $m_q \ne 0$ (see Eq.~\eqref{fpimqne0} in Appendix~\ref{app:fpi}). The results for both potentials are given in Fig.~\ref{fig:fpi}, and they are also compared\footnote{Notice that $S_0$ is UV divergent whenever the quark mass is finite, as can be seen by inserting the UV expansions of the wave functions~\cite{Arean:2013tja} in~\eqref{Pidef}, and needs to be renormalized. The divergence is $\propto m_q^2$ in agreement with the one-loop field theory computation of Appendix~\ref{app:SFT}. At small quark masses it is irrelevant how the renormalization is done because the difference between all reasonable renormalization schemes is negligible due to the smallness of the coefficient in the divergent term.} to the constant $S_0$. As expected the pion decay constant (blue curves) match with $\sqrt{S_0}$ (dashed red curves) in the QCD regime. The ratio $f_\pi^2/S_0$ decreases fast with
increasing $x$ for both
potentials when $x\gtrsim x_c$, suggesting that the pion decouples.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140716_S2_vs_x_mq_10pm6.pdf}
\includegraphics[width=0.49\textwidth]{140412_S_prime_times_constant_finite_mq_10m6.pdf}
\end{center}
\caption{The $x$ dependence of the higher order coefficient $S_2$ for $m_q/\Lambda_\mathrm{UV}=10^{-6}$ and for potentials~I with $W_0=3/11$. Left: $S_2$ in units of $\L_\mathrm{IR}$. Right: The dimensionless product $S_0 S_2$.}
\label{fig:Sprime}\end{figure}
The higher order coefficient $S_2$ (in units of $\Lambda_\mathrm{IR}$) is also shown for potentials~I in Fig.~\ref{fig:Sprime} (left) and the product $S_0S_2$ in Fig.~\ref{fig:Sprime} (right). The dependence on $x$ is similar as for the S-parameter when $x \lesssim x_c$.
We have not tried to analyze the various observables in the BZ limit $x \to x_\mathrm{BZ}$ because this region is not the most interesting one from a holographic viewpoint. In general, however, the slow RG flow in the BZ limit causes that all ratios of energy scales to be typically $\propto \exp[\#(x_\mathrm{BZ}-x)^{-2}]$. Therefore even ratios which are expected to be close to one for generic values of $x$ easily blow up in the BZ limit -- in this case $\#$ in the above relation is small, but it is difficult to define the scales such that it would be exactly zero. In view of this, it is not surprising that only the dimensionless quantities $S$ and $S_0S_2$ approach finite values in the BZ region.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.69\textwidth]{141202_S_parameter_vs_mq_various_x.pdf}
\end{center}
\caption{The mass dependence of the S-parameter for $x=1$ (blue curve), $x=4$ (dashed red curve), and $x=4.5$ (dotted magenta curve). Potentials~I with $W_0=3/11$ were used. The thin lines are extrapolations given by fits to the asymptotic behavior.
}
\label{fig:massdep}\end{figure}
Let us then analyze the mass dependence of $f_\pi$ and $S$ in more detail. Fig.~\ref{fig:massdep} shows the mass dependence of the S-parameter in the QCD-regime ($x=1$, blue curve), in the walking regime ($x=4$, dashed red curve), and in the conformal window ($x=4.5$, dotted magenta curve). The dependence on $m_q$ is relatively mild for all values of $x=0$ (apart from the discontinuity at $m_q=0$ which is only present in the conformal window and the fact that $S$ varies slower as $x$ increases which is due to the RG flow as discussed at the end of Appendix~\ref{app:scaling}). In particular the limiting value as $m_q \to \infty$ is independent of $x$. In this limit the S-parameter is expected to approach the value $N_c N_f/12\pi$ from perturbative QCD (see Appendix~\ref{app:SFT}). Even though V-QCD is not expected to reproduce perturbative results in general, the limiting value in V-QCD is numerically close to the QCD number $1/12 \pi \simeq 0.0265$.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_fpi_vs_mq_x_1.pdf}%
\includegraphics[width=0.49\textwidth]{141202_fpi_vs_mq_x_4.pdf}
\includegraphics[width=0.49\textwidth]{141202_fpi_vs_mq_x_4p5.pdf}
\end{center}
\caption{The mass dependence of $f_\pi$ for $x=1$ (top left), $x=4$ (top right), and $x=4.5$ (bottom). Potentials~I with $W_0=3/11$ were used.
}
\label{fig:fpimassdep}\end{figure}
The dependence of $f_\pi$ on $m_q$ is demonstrated in Fig.~\ref{fig:fpimassdep}. Again the different plots are in the QCD regime ($x=1$, top left plot), in the walking regime ($x=4$, top right plot), and in the conformal window ($x=4.5$, bottom plot). For small $m_q$ the dependence is weak, but when $m_q/\Lambda_\mathrm{UV} \gg 1$ the decay constant vanishes very fast with increasing $m_q$. This signals the decoupling of the pion mode in regime~C, and is consistent with the findings of Appendix~\ref{app:masses} (see Eq.~\eqref{fnpotI}).
Actually, all low-lying meson states are expected to decouple for potentials~I.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140716_S_parameter_mass_dependence_x_4p5.pdf}
\includegraphics[width=0.49\textwidth]{140716_S_parameter_mass_dependence_x_4p5_potII.pdf}
\end{center}
\caption{The mass dependence of the S-parameter for $x=4.5$ in log-log scale. The red dots are the data and the blue lines are power-law fits $\propto m_q^{0.08}$. Left: potentials~I with $W_0=3/11$. Right: potentials~II with SB normalized $W_0$.}
\label{fig:massdeplog}\end{figure}
\subsection{Scaling of the S-parameter}
We have also analyzed the S-parameter in the limit $m_q \to 0$. Numerical results are shown in the conformal window (for $x=4.5$) in Fig.~\ref{fig:massdeplog}, where the red dots are our data and the lines are given by the functions
\begin{equation} \label{Sfit}
\frac{S(m_q)-S(0^+)}{N_cN_f} = \beta_1 m_q^{\beta_2} \ ,
\end{equation}
where the parameters $\beta_{1,2}$, as well as $S(\epsilon)$, were fitted to the data. Here it is understood that
\begin{equation}
S(0^+) \equiv \lim_{m_q \to 0 +} S(m_q)
\end{equation}
which is a finite number whereas the S-parameter vanishes at zero quark mass: $S(0)=0$. Thus there indeed is a discontinuity at $m_q=0$. For both potentials $\beta_2 \simeq 0.08$ fits the data very well.
In order to understand the power law in~\eqref{Sfit}, it is useful to first discuss in more detail how the discontinuity at $m_q=0$ arises. The mechanism is the same which was studied in detail in the case of $m_q=0$ and $x \to x_c$ in section~6 and Appendix~I of~\cite{Arean:2013tja}. Here instead $x \ge x_c$ and $m_q \to 0$. In both cases the RG flow of the coupling approaches the fixed point ($\lambda=\lambda_*$) but misses it finally (due to the finite tachyon). For such flows, it is useful to divide the background to the UV and IR sections, having $\lambda<\lambda_*$ and $\lambda>\lambda_*$, respectively.
Considering flows which get closer and closer to the fixed point, the S-parameter can be computed more and more precisely in terms of the IR section. The IR part takes a fixed shape in this limit, explaining the finite value of the S-parameter. For exactly zero quark mass the IR section of the background becomes disconnected from the UV section and is therefore not present in the physical vacuum solution, which now ends at the IRFP. This is reflected in the vanishing value of the S-parameter.
By using similar arguments, we can also sketch how the power law in the mass dependence arises. It is understood to be the leading ``perturbation'' of the IR background due to the fact that the fixed point was not reached exactly. In the conformal window, the flow towards the fixed point is given by
\begin{equation}
\lambda \simeq \lambda_* - C_\lambda \left( r\Lambda_\mathrm{UV}\right)^{-\delta}\ ,
\end{equation}
where $\delta$ is related to the dimension of the $\mathrm{Tr} F^2$ operator at the fixed point. It can be computed as the derivative of the holographic beta function at the fixed point (when the tachyon is set to zero exactly)~\cite{jk,Gubser:2008yx}. One finds that
\begin{equation}
\Delta_{FF}-4 = \delta = \sqrt{4-\frac{9 V_2 \lambda_*^2}{V_0}}-2 \ ,
\end{equation}
where $V_i$ are the coefficients of the Taylor expansion of $V_\mathrm{eff}$ at $\lambda=\lambda_*$:
\begin{equation}
V_\mathrm{eff} \equiv V_g - x V_{f0} = V_0 + V_2 (\lambda-\lambda_*)^2 + \cdots \ .
\end{equation}
Notice that $V_2<0$. Flow toward the fixed point ends when $r\sim 1/\Lambda_\mathrm{IR}$. The difference of $\lambda$ with respect to the fixed point value when this happens is given by
\begin{equation} \label{massdepCW}
\lambda_*-\lambda_\mathrm{IR} \equiv \lambda_*-\lambda(r=1/\Lambda_\mathrm{IR}) \sim \left( \frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}}\right)^{\delta} \sim \left( \frac{m_q}{\Lambda_\mathrm{UV}}\right)^{\frac{\delta}{\Delta_*}} \ , \qquad \left(\frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right)
\end{equation}
where~\eqref{CWscaling} was used to obtain the last expression. This difference controls the deviation of the IR section of the background from its limiting shape as $m_q \to 0$ and correspondingly the deviation of the S-parameter from the limiting value $S(0^+)$. Indeed by using the explicit expressions for the potentials one obtains
\begin{equation}
\frac{\delta}{\Delta_*} \simeq 0.0780 \ , \qquad (x=4.5) \ ,
\end{equation}
a value in a very good agreement with the fit from above, shown in Fig.~\ref{fig:massdeplog}.
Notice that the difference $\lambda_*-\lambda_\mathrm{IR}$ not only controls the corrections to the S-parameter at small mass, but also to other quantities that can be defined in terms of the IR section of the background. Examples are decay constants and meson masses in IR units, which are therefore expected to have qualitatively similar $m_q$ dependence to the S-parameter at small masses\footnote{We have found numerically that ratios of masses or decay constants typically follow the scaling of~\eqref{massdepCW} more accurately than their values in IR units.}. Indeed, the power law of~\eqref{massdepCW} agrees with that found for the scaling corrections to generic correlators by analyzing the Wilsonian RG flow perturbatively in the vicinity of the fixed point~\cite{DelDebbio:2013qta}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140716_S_parameter_mass_dependence_x_4.pdf}
\includegraphics[width=0.49\textwidth]{140716_S_parameter_mass_dependence_x_3p6_potII.pdf}
\end{center}
\caption{The mass dependence of the S-parameter in the walking regime in log-log scale. The red dots are the data and the blue lines are power-law fits. Left: potentials~I with $W_0=3/11$ and with $x=4$. Right: potentials~II with SB normalized $W_0$ and $x=3.6$.}
\label{fig:massdepwalk}\end{figure}
The mass dependence of the S-parameter can be analyzed similarly in the walking regime ($x \to x_c{}^-$). The above calculation is approximately valid in regime B, i.e, when the quark mass controls the amount of walking. One can use~\eqref{regimeBlowx} together with~\eqref{massdepCW} to obtain
\begin{equation} \label{massdepwalking}
\lambda_*-\lambda_\mathrm{IR} \sim \left( \frac{m_q}{\Lambda_\mathrm{UV}}\right)^{\frac{\delta}{2}} \ , \qquad \left(\exp\left[-\frac{2K}{\sqrt{x_c-x}}\right]\ll \frac{m_q}{\Lambda_\mathrm{UV}} \ll 1\right) \ .
\end{equation}
Therefore one should effectively take $\Delta_* \to 2$ in~\eqref{massdepCW}, meaning that the power is continuous over the conformal transition at $x=x_c$, as $\Delta_*=2$ at the transition. For even smaller $m_q$, i.e., in regime~A, the mass term of the tachyon can be treated as a linear perturbation to the whole background, and therefore the mass dependence of the S-parameter is linear.
Both the scaling of~\eqref{massdepCW} and the linear dependence can be seen in the numerical results in the walking regime, see Fig.~\ref{fig:massdepwalk}.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{140717_S_parameter_spectral.pdf}%
\includegraphics[width=0.49\textwidth]{141202_Delta_S_spectral.pdf}
\end{center}
\caption{S-parameter computed from the spectral representation compared to the exact value. Left: the dependence of the value of $S$ on the mass cutoff in the series. The thin horizontal lines are the limiting values which match with direct computation of $S$. Right: the difference between the number obtained from the series and the exact value as a function of the cutoff. The blue, dashed red, dotted magenta, and dotdashed green curves have $x=1$, $3.5$, $4$, and $4.5$, respectively, in both plots. The filled circles and boxes denote the masses of the lowest five vector and axial mesons, respectively.}
\label{fig:Sconv}\end{figure}
\subsection{Convergence of the spectral representation}
Finally we will analyze the spectral representation of the S-parameter in order to understand better why it increases with $x$ in the QCD and walking regimes.
By inserting~\eqref{PiAsp} and~\eqref{PiVsp} in the definition for $S$, one obtains
\begin{equation} \label{Sseries}
S = 4 \pi \sum_{n=1}^\infty\left(\frac{(f_{n}^{(V)})^2}{(m_{n}^{(V)})^2}-\frac{(f_{n}^{(A)})^2}{(m_{n}^{(A)})^2}\right) \ .
\end{equation}
This series may, however, be ill defined. For example, the decay constant approach asymptotically constant values whereas the masses have linear trajectories $m_n^2 \sim n$ for potentials~I (see Appendices~E and~F in~\cite{Arean:2013tja}). One can check that~\eqref{Sseries} is convergent thanks to the asymptotic cancellation of the vector and axial terms, but it is not absolutely convergent, and therefore the result may depend on the ordering of the terms. The definition of~\eqref{Sseries}, where the states are ordered in terms of their excitation numbers $n$, would work for potentials with linear trajectories if the slopes of the vector and axial spectra are the same. This is not\footnote{The question whether the slopes should be the same in QCD is unsettled~\cite{Golterman:2001nk}, but usually they are assumed to be, which can be obtained for potentials~I by changing the IR asymptotics of $w(\lambda)$~\cite{Arean:2013tja}.} the case for potentials~I, and consequently~\eqref{Sseries} is incorrect. The solution
to this issue is simple: the contributions should be ordered by the meson
masses rather than by the excitation numbers.
It is straightforward to verify numerically that~\eqref{Sseries} converges towards the S-parameter if the terms are ordered according to the masses. It turns out to be convenient to define a mass dependent cutoff,
\begin{equation} \label{Scut}
S(m_\mathrm{cut}) = 4 \pi \sum_{n=1}^\infty\left(\frac{(f_{n}^{(V)})^2}{(m_{n}^{(V)})^2}\, f_\mathrm{cut}(m_{n}^{(V)})-\frac{(f_{n}^{(A)})^2}{(m_{n}^{(A)})^2}\, f_\mathrm{cut}(m_{n}^{(A)})\right)
\end{equation}
where
\begin{equation}
f_\mathrm{cut}(m) = \frac{1}{2}\left(1- \tanh \frac{m-m_\mathrm{cut}}{\delta m}\right) \ .
\end{equation}
A smooth cutoff function was chosen instead of a step function because it improves convergence drastically. The convergence also means that the value of the S-parameter is determined through the dynamics in the deep IR, because the same holds for the masses and decay constants (the argument can be made precise in the walking regime, see Appendix~I in~\cite{Arean:2013tja}).
The convergence of the regulated series~\eqref{Scut} towards the S-parameter is demonstrated in Fig.~\ref{fig:Sconv}. The resolution of the cutoff function was fixed to the mass difference of the two lowest vector states: $\delta m =m_2^{(V)}-m_1^{(V)}$. The speed of the convergence is best visible from the right hand plot, which shows
\begin{equation}
\Delta S = S - S(m_\mathrm{cut})
\end{equation}
as a function of the cutoff. The convergence is exponential for all values of $x$, but becomes significantly slower as $x$ increases and one moves from the QCD regime to the conformal window. The slowness of the convergence is not due to changes in the spectra. To show this, the filled circles and boxes marking the masses of the lowest five vector and axial mesons, respectively, were added in each plot. It is seen that the spectrum changes relatively little with $x$ in unit of $\Lambda_\mathrm{IR}$.
Finally let us try to extract the reason for the increase of $S$ with increasing $x$ in the region of low values of $x$ from the plots of Fig.~\ref{fig:Sconv}. Notice that the curves for $x=1$, $3.5$ and $4$ essentially overlap at small values of $m_\mathrm{cut}$ in the left hand plot. The curves for $x=1$ and $x=3.5$ deviate from that of $x=4$ as $m_\mathrm{cut}$ increases, and after deviating rapidly saturate to the final value of $S$. Therefore it appears that contributions to the S-parameter are roughly mass-independent up to a saturation scale, which increases with $x$. In order to have a good estimate for the S-parameter, a growing number of terms need to be included in the sum with increasing $x$, whereas the individual terms in the sum are of roughly constant size. Therefore the increase of $S$ with $x$ in the QCD and walking regimes can be seen to be due to slower convergence of the sum. In the conformal window, i.e., for the curve with $x=4.5$, something different happens. The convergence of the
sum is even slower, but the contributions at fixed $m_\mathrm{cut}$ are suppressed, resulting in the decrease of the S-parameter
with increasing $x$.
\section{Finite temperature phase diagram}
\label{sec:ft}
The finite temperature phase diagram has been studied in detail for IHQCD in~\cite{ihqcd2} and for V-QCD at zero quark mass and at small values of the quark mass in~\cite{alho}. Here this study is extended to large values of $m_q$ as well as very high values of $x$. The code for constructing solutions at finite temperature is available at~\cite{code}. Some extra tricks are necessary in order to obtain reliable results in the BZ region and at very large $m_q$ (see Appendix~\ref{app:numerics}).
First recall the generic structure of the ($x,T$) phase diagram~\cite{alho}, which was already reviewed in Sec.~\ref{sec:vqcd}. At zero quark mass, there is a first order deconfinement transition in the QCD and walking regimes, but there is also the possibility (depending on the choice of potentials and the value of $x$) of a chiral symmetry restoration at a separate second order transition. In the conformal window, there is a continuous phase transition at zero temperature.
When a finite quark mass is turned on, chiral symmetry is always broken, and the second order chiral transition will become a fast crossover when the quark mass is small, and completely disappear at larger quark masses.
The system is in a tachyonic thermal gas (TG) phase at small temperatures for all $0<x<x_\mathrm{BZ}$. As the system is heated, there is a first order deconfinement transition to the high temperature phase, which is implemented through a transition from TG phase to the black hole (BH) phase in holography.
The existence of the deconfinement transition requires an order parameter. While QCD at finite $N_c$ and $N_f$ has no order parameter related to deconfinement, the pressure acts as an effective order parameter at large $N_c$. This is clear in the 't Hooft limit, where the number of degrees of freedom (and consequently the pressure) is \order{N_c^0} in the low temperate phase and \order{N_c^2} in the high temperature phase. In the Veneziano limit the number of degrees of freedom is of the same order in both phases, but the phase transition may still be identified as a discontinuity of the pressure.
In fact, the pressure of the model is still exactly zero in the TG phase, because our approach does not capture the contributions corresponding to loops of pions (as well as mesons with higher masses) in this phase. Including these contributions in the model would affect the critical temperature, and potentially even alter the order of the transition~\cite{pionloop}.
\subsection{Scaling laws at finite temperature}
The critical temperature has nontrivial dependence on $m_q$, which can be analyzed analytically. Notice that the temperature brings in an additional energy scale with respect to the zero temperature solutions. Another difference is that the definition of the standard reference scale $\Lambda_\mathrm{IR}$ cannot be extended to the BH solutions in a natural manner (because the geometry now ends at a horizon rather than an IR singularity). Therefore it is understood that $\Lambda_\mathrm{IR}$ is defined below through the TG solution (or equivalently through the zero temperature solution at the same values of $m_q$ and $x$).
Let us first discuss the mass dependence of the critical temperature, which can be
inferred by using the results from~\cite{alho} and from Sec.~\ref{sec:scaling}. The temperature of the black hole can be related to the metric through the formula
\begin{equation} \label{Tdef}
T = \frac{1}{4 \pi e^{3A_h}}\left(\int_0^{r_h}\frac{dr}{e^{3A(r)}}\right)^{-1} \ ,
\end{equation}
where $A_h$ and $r_h$ are the values of the scale factor and the bulk coordinate at the horizon.
In V-QCD models $T(r_h)$ has a nontrivial minimum (for tachyonic BHs) and the transition takes place at the scale of the minimum.
Indeed, the entropy density
\begin{equation} \label{sdef}
s_\mathrm{BH} = 4 \pi M^3 N_c^2 e^{3 A_h}
\end{equation}
decreases monotonically (and fast) with $r_h$, as suggested by the UV and IR (zero temperature) expansions of $A(r)$ and as can be verified numerically. Further, the geometry of the BH solution approaches smoothly the TG solution as $r_h \to \infty$, and therefore $p_\mathrm{TG} = \lim_{r_h \to \infty} p_\mathrm{BH}$. Integrating $p_\mathrm{BH}'(r_h) = s(r_h) T'(r_h)$, a node $p_\mathrm{BH} = 0$, and consequently a first order phase transition, is found near the minimum of $T(r_h)$.
In conclusion, one should locate the minimum of $T(r_h)$ in order to determine the scaling of $T_c$. When the geometry is close to AdS,~\eqref{Tdef} implies that $T \sim 1/r_h$. This result holds both in the UV asymptotic region and when there is an approximate IRFP, i.e., walking. When the quark mass is large, there is also an approximately AdS region where the flavors are already decoupled but the dilaton $\lambda$ is still small. In summary,
\begin{equation} \label{TUV}
T \sim 1/r_h\ ,\qquad \left(r_h \ll \frac{1}{\Lambda_\mathrm{IR}}\right) \ .
\end{equation}
For $r_h \gg 1/\Lambda_\mathrm{IR}$ the temperature increases with $r_h$ as seen by studying the IR expansions (see~\cite{alho}). Therefore, the minimum of $T(r_h)$ takes place at $r_h\sim 1/\Lambda_\mathrm{IR}$. By continuity,~\eqref{TUV} implies that
\begin{equation}
T_c \sim \Lambda_\mathrm{IR}
\end{equation}
for all $m_q>0$ and $0<x<x_\mathrm{BZ}$.
The scaling results in units of $\Lambda_\mathrm{UV}$ immediately follow by using the results from Sec.~\ref{sec:scaling}:
\begin{itemize}
\item In regime A, $T_c \sim \Lambda_\mathrm{IR} \sim \Lambda_\mathrm{UV}$ for small values of $x$ and
\begin{equation} \label{TcinregA}
\frac{T_c}{\Lambda_\mathrm{UV}} \sim \frac{\Lambda_\mathrm{IR}}{\Lambda_\mathrm{UV}} \sim \exp\left[-\frac{K}{\sqrt{x_c-x}}\right]
\end{equation}
as $x \to x_c$ from below. The dependence of $T_c$ on $x$ at $m_q=0$~\cite{alho} is in qualitative agreement with analysis based on field theory (see, e.g.,~\cite{Braun:2009ns}), and the dependence on $m_q$ is expected to be a linear perturbation.
\item In regime B,
\begin{equation}
\frac{T_c}{\Lambda_\mathrm{UV}} \sim \sqrt{\frac{m_q}{\Lambda_\mathrm{UV}}}
\end{equation}
when $x \le x_c$ and
\begin{equation}
\frac{T_c}{\Lambda_\mathrm{UV}} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^\frac{1}{\Delta_*}
\end{equation}
when $x_c \le x <x_\mathrm{BZ}$.
\item In regime C,
\begin{equation} \label{TcinregC}
\frac{T_c}{\Lambda_\mathrm{UV}} \sim \left(\frac{m_q}{\Lambda_\mathrm{UV}}\right)^{1-b_0/b_0^\mathrm{YM}} \ .
\end{equation}
\end{itemize}
In addition to the phase transition, also various crossovers can be identified as the maxima of the interaction measure
\begin{equation}
\frac{\epsilon-3p}{T^4} = \frac{T s -4 p}{T^4} \ .
\end{equation}
As it turns out, such crossovers reflect the different regions of the zero temperature geometry. This can be understood by approximating the horizon as a sharp cutoff added on the zero temperature background. Substituting an AdS metric in the formulas~\eqref{Tdef} and~\eqref{sdef}, one finds that the interaction measure vanishes.
First, there is the crossover which marks the transition from the quasi-conformal or walking phase (with approximate IRFP) to the asymptotic UV phase \cite{alho,Tuominen:2012qu}. Such a crossover is found whenever there is walking, i.e., in regime B, and the $x \to x_c$ edge of regime A. The UV asymptotics is valid for $r \ll 1/\Lambda_\mathrm{UV}$, and the flow from the UV fixed point to the IRFP is characterized by $\Lambda_\mathrm{UV}$. Consequently, using~\eqref{Tdef} and~\eqref{TUV}, the crossover temperature is expected to be
\begin{equation} \label{Tqclaw}
T_\mathrm{co,qc} \sim \Lambda_\mathrm{UV}
\end{equation}
independently of $m_q$.
Second, there is a crossover at large quark mass, corresponding to the transition from the region where the quarks are decoupled to the UV asymptotic region. The decoupling of the quarks takes place at $r \sim 1/m_q$ as pointed out in Sec.~\ref{sec:scaling}. Consequently, the crossover temperature is given in terms of the quark mass:
\begin{equation} \label{Tmqlaw}
T_\mathrm{co,{\it m_q}} \sim m_q \ .
\end{equation}
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{141202_Temperatures_vs_x_mq_10pm6.pdf}%
\includegraphics[width=0.49\textwidth]{141202_Temperatures_vs_mq_x_1.pdf}
\includegraphics[width=0.49\textwidth]{141202_Temperatures_vs_mq_x_4.pdf}%
\includegraphics[width=0.49\textwidth]{141202_Temperatures_vs_mq_x_4p5.pdf}
\end{center}
\caption{The dependence of the critical transition temperatures (solid blue curves) and various crossover temperatures (thin dotted red curves) on $x$ and $m_q$. In the top left plot we also included the $m_q=0$ data, and the critical temperature of the second order chiral transition which is shown as the thick red dashed curve near $x=x_c$.}
\label{fig:FT}\end{figure}
When $T \ll m_q$ the quarks are effectively decoupled, and therefore the thermodynamics is the same as for pure YM (that is, the $x\to 0$ limit of V-QCD). Notice that $\Lambda_\mathrm{UV}$, however, is defined in terms of the UV asymptotics, i.e., effectively at infinite energy, and different from that of YM even as $m_q \to \infty$: for YM, $\Lambda_\mathrm{UV} \sim \Lambda_\mathrm{IR}$ but in the limit of large $m_q$ these scales are related through~\eqref{largemqscalingt} instead. Consequently, in order for the thermodynamics to smoothly approach YM thermodynamics as $m_q \to \infty$, dimensional quantities should be expressed in units of $\Lambda_\mathrm{IR}$ or $T_c$ rather than in units of $\Lambda_\mathrm{UV}$.
\subsection{Numerical results}
Let us then illustrate the dependence of the various critical temperatures on $x$ and $m_q$ numerically. The basic features of the phase diagram at small quark mass are demonstrated in the top-left plot of Fig.~\ref{fig:FT}. The first order ``deconfinement'' transition temperatures for potentials~I at zero and at tiny ($10^{-6}$) quark mass are shown as functions of $x$ on the logarithmic scale. The first order transitions are shown as blue curves. The curves overlap at small $x$, but as the conformal transition is approached, the curves become separated. The lower curve (which overlaps with the red dashed curve and is therefore not well visible) is the transition temperature at $m_q=0$ which goes to zero with Miransky scaling as $x \to x_c$. For $x>x_c$ there is no transition when $m_q=0$. When a tiny quark mass is turned on the transition (upper blue curve) is present for all values of $x$. the critical temperature $T_c$ decreases with $x$ inside the conformal window\footnote{In~\cite{alho} also a region
where $T_c$ increased with $x$ was seen at high $x\simeq 4.5$ (see Fig.~27 there), but this effect turns out to be due to the UV cutoff being too low in the
numerics.}.
The dashed thick red curve is the second order chiral restoration transition, which also shows Miransky scaling as $x \to x_c$ and is absent for $x>x_c$. This transition only exists for $m_q=0$ in the walking regime.
The dependence of $T_c$ on $m_q$ is also demonstrated numerically for V-QCD in Fig.~\ref{fig:FT}. The top-right, bottom-left, and bottom-right plots are in the running regime ($x=1$), walking regime ($x=4$), and in the conformal window ($x=4.5$), respectively, and $T_c$ is given by the blue curve in each plot. The power laws are in agreement with the above formulas. For example, in the regime~C with large $m_q$ we find that the exponent of~\eqref{TcinregC} is $1-b_0/b_0^\mathrm{YM} = 2x/11 \simeq 0.182$ at $x=1$ and $0.727$ at $x=4$, which is consistent with the plots.
The crossover between the quasiconformal and UV regions at $T=T_\mathrm{co,qc}$ is seen as the horizontal lines in the bottom row of Fig.~\ref{fig:FT}. The ratio $T_\mathrm{co,qc}/\Lambda_\mathrm{UV}$ is constant as expected, but the value of the constant deviates significantly from its expected value, i.e., one. This happens because $\Lambda_\mathrm{UV}$ deviates from the scale of the UV RG flow at large values of $x$, as was explained above in Sec.~\ref{sec:massscales}. The crossover due to the decoupling of the quarks at large $m_q$ at the temperature $T \sim m_q$ is best visible in the top-right plot of Fig.~\ref{fig:FT}.
In addition, in large part of the parameter space there is also a separate maximum of the interaction measure in connection to the first order transition. This kind of maxima have also been included as thin red curves in Fig.~\ref{fig:FT}, and can be found close to the blue curves denoting the transition temperatures.
\begin{figure}[!tb]
\begin{center}
\includegraphics[width=0.49\textwidth]{150119_Thermo_small_mq.pdf}%
\hspace{2mm}\includegraphics[width=0.49\textwidth]{150119_Thermo_large_mq.pdf}
\end{center}
\caption{Pressure and interaction measure as a function of temperature for small (left hand plot) and large (right hand plot) quark mass. On the left, the solid blue, dashed red, and dotted magenta curves have $m_q/\Lambda_\mathrm{UV}=0$, $10^{-12}$, and $10^{-10}$, respectively. On the right, the solid blue curves are the YM thermodynamics in our model, while the dashed red and dotted magenta curves have $m_q/\Lambda_\mathrm{UV}=10^{10}$ and $m_q=10^8$, respectively. We used potentials I with $W_0=3/11$ and $x=4$ (except for the YM curves which correspond to $x=0$). }
\label{fig:thermo}\end{figure}
Recall that in regime~C the mass gap of the mesons does not have the expected behavior $m_\mathrm{gap} \simeq 2 m_q$ in V-QCD (if Sen-like tachyon potential is assumed). Nevertheless, the crossover due to the decoupling of the quarks takes place correctly at $T_\mathrm{co,{\it m_q}} \sim m_q$. This is seen as a consequence of the decoupling of the mesons having an unphysically low mass, which was discussed above in Sec.~\ref{sec:massscales}.
Finally let us study in some details of the thermodynamic functions at very small and very large $m_q$. We plot the (normalized) energy density, pressure, and interaction measure for potentials~I with $x=4$ in Fig.~\ref{fig:thermo} (left). The solid blue, dashed red, and dotted magenta curves have $m_q/\Lambda_\mathrm{UV} = 0$, $10^{-12}$, and $10^{-10}$, respectively. The value of $x=4$ was chosen such that the model shows the separate second order chiral restoration transition, which appears as a kink in the energy density and the interaction measure of the functions at zero quark mass. The vertical thin dashed black lines mark the locations of the phase transitions. As a tiny quark mass ($m_q/\Lambda_\mathrm{UV} =10^{-12}$) is turned on, the second order transition turns into a crossover. Indeed the dashed curves follow closely the solid curves except for very close to the kink, where the curves have a smooth behavior instead. At $m_q/\Lambda_\mathrm{UV} =10^{-10}$ a much larger deviation is seen already.
The approach to YM theory can be seen by studying the thermodynamics at very large quark mass. Thermodynamic functions at sizeable $m_q$ are shown for the same choice of potentials and compared to the results for the limit of YM theory ($x \to 0$) in Fig.~\ref{fig:thermo}. As the quark mass increases, the quarks are decoupled and the thermodynamics near $T=T_c$ is expected to converge to that of YM theory. Indeed, the (dashed red) curves for $m_q/\Lambda_\mathrm{UV} = 10^{10}$ lie essentially on top of the (solid blue) curves of the YM thermodynamics\footnote{In order to show the convergence, the Planck mass $M$ was fixed to the value obtained in the YM limit (by matching the pressure to the SB law at high temperatures) also at $x=4$. If we matched the pressure separately at $x=4$, there would be a small deviation, because a fixed value of $W_0$ was used rather than the ``SB normalized'' value~\cite{alho} so that $M$ would vary as a function of $x$.}. For smaller quark mass ($m_q/\Lambda_\mathrm{UV} =10^8$, dotted magenta curves) a much
large deviation is seen
already.
\section*{Acknowledgments}\label{ACKNOWL}
I would like to thank E.~Kiritsis for advice during this project and F.~Sannino for explanations of the mass dependence of the S-parameter. In addition, I would like to thank D.~Albrecht, T. Alho, F. Aprile, B. Gout\'eraux, P. Hoyer, T. Ishii, K.~Kajantie, W. Li, T. Morita, F. Nitti, M. Panero, C. Pica, C. Rosen, R. Shrock, K. Tuominen, and R.~Zwicky for interesting discussions, relevant comments, or correspondence.
This work was partially supported by European Union's Seventh Framework Programme
under grant agreements (FP7-REGPOT-2012-2013-1) No 316165, PIF-GA-
2011-300984, the ERC Advanced Grant BSMOXFORD 228169, the EU program
Thales MIS 375734. It was also co-financed by the European Union (European Social
Fund, ESF) and Greek national funds through the Operational Program ``Education
and Lifelong Learning'' of the National Strategic Reference Framework (NSRF)
under ``Funding of proposals that have received a positive evaluation in the 3rd and
4th Call of ERC Grant Schemes'', as well as under the action ``ARISTEIA''.
I also thank the ESF network Holograv for partial support, and acknowledge the use of the computational resources of CCTP.
\newpage
|
\section{INTRODUCTION}
The correlation between the mass of the supermassive
black holes in the centers of galaxies
and the bulge stellar mass or the velocity dispersion
\citep[e.g.,][]{1995ARA&A..33..581K,1998AJ....115.2285M,2013ARA&A..51..511K}
strongly suggest that
the growth of the central supermassive black hole and
the evolution of the galaxies are connected.
Obscured Active Galactic Nuclei (AGNs) are important targets
for studying the AGN-galaxy coevolution because
a large number of host galaxies
can be investigated in detail being free from
the strong central emission from the AGN itself.
Moreover, they are of special importance
because some types of obscured AGNs
such as the ultraluminous infrared galaxies (ULIRGs)
have long been anticipated to be
in the middle of the evolutionary sequence of AGNs
\citep[e.g.,][]{1988ApJ...325...74S,1996ARA&A..34..749S,2008ApJS..175..356H}.
The ``new type AGNs'' recently found by \citet{2007ApJ...664L..79U},
where the central engine is thought to be deeply obscured
by a dust torus with an unusually small opening angle,
are also interesting for its nature and role
in the AGN evolution.
Consequently, measuring the black hole mass in obscured AGNs is important;
however, the precise measurement is not straightforward.
The position-velocity diagram of the molecular disk
surrounding the black hole,
obtained by the VLBI observation of the H$_2$O maser,
is considered to be one of the most
precise estimate for the black hole mass \citep[e.g.,][]{1995Natur.373..127M}.
However, precise alignment between the equatorial plane of
the molecular disk and the observer's line of sight is required
for the disk H$_2$O maser to be observable;
thus the number of applicable targets is very limited.
The single-epoch mass estimate based on the luminosity scaling relation
of the broad emission-line region (BLR)
\citep[e.g.,][]{1998ApJ...505L..83L,2002MNRAS.337..109M,2006ApJ...641..689V}
measuring the widths of the polarized broad Balmer emission lines
would be widely available for type-2 AGNs.
However, a significant fraction of type-2 AGNs does not
present the polarized broad Balmer emission lines
\citep[][]{2001ApJ...554L..19T}.
Moreover, this type of measurement might be unavailable
for the ULIRGs and the new-type AGNs,
whose BLR is largely covered by obscuring materials.
A precise estimate of the black hole mass
that is widely available for obscured AGNs is desired.
The fluorescent iron K$\alpha $ line
(hereafter, FeK$\alpha $ line)
is an almost ubiquitous emission line observable
for both obscured and unobscured AGNs
\citep[][]{2004ApJ...604...63Y,2007MNRAS.382..194N,2011ApJ...727...19F}.
The rest-frame energy of the narrow core of the FeK$\alpha $ emission line
is close to that of neutral or low-ionization-state irons.
Considering its narrow line width,
the emitting region is supposed to be not very close to the black hole,
such as the outer parts of the accretion disk and more distant matter.
The large equivalent width indicates both large covering fraction
and large optical depth for the emitting region,
which suggests that
the neutral FeK$\alpha $ line core originates in the dust torus
surrounding the central engine
in accordance with the unified model of AGNs \citep[e.g.,][]{1991PASJ...43..195A}.
When the width of an emission line and
the distance of its emitting region from the black hole are obtained,
the black hole mass can be estimated assuming the virial relation.
\citet{2011SCPMA..54.1354J} assumed that
the location of the emitting region of the neutral FeK$\alpha $ line core
corresponds to the near-infrared $K$-band reverberation radius
for the inner region of the dust torus,
and estimated the black hole mass of 10 type-1 AGNs
on the basis of the FeK$\alpha $ line widths
and the luminosity scaling relation of the dust reverberation radius
to the optical $V$-band luminosity by \citet{2006ApJ...639...46S}.
They examined the black hole mass on the basis of
the neutral FeK$\alpha $ line core
by comparing the black hole mass obtained from
the broad emission-line reverberation;
however, the correlation between them was statistically insignificant,
which could be attributable to small sample size and large uncertainties
in the FeK$\alpha $ line widths.
Recently, \citet{gree+10} and \citet{kosh+14} presented
the luminosity scaling relations
of the broad H$\beta $ emission-line reverberation radius
and the dust reverberation radius
to isotropic luminosity indicators
that are unbiased by the obscuration,
such as
the luminosity of the [\ion{O}{4}] $\lambda 25.89\ \mu$m emission line
$L_{{\rm [OIV]}}$,
and that of the {\it Swift} BAT hard X-ray (14--195 keV) band $L_{\rm BAT}$
\citep[e.g.,][]{mele+08,2009ApJ...700.1878R,diam+09}.
Using these radius-luminosity scaling relations,
it becomes possible to derive the radius of
the emitting region of the neutral FeK$\alpha $ line core
for obscured AGNs.
In this paper,
we estimate the black hole mass of both obscured and unobscured AGNs
on the basis of the line width of the neutral FeK$\alpha $ line core
and the luminosity scaling relation of its emitting region radius
with the [\ion{O}{4}] $\lambda 25.89\ \mu$m emission-line luminosity.
Then, the estimated black hole mass was examined by comparing them with
other black hole mass estimates that have been established.
We assume the cosmology of
$H_0=73$ km s$^{-1}$ Mpc$^{-1}$,
$\Omega _m = 0.27$, and $\Omega _\Lambda = 0.73$
according to \citet{2007ApJS..170..377S} throughout this paper.
\section{TARGETS}
We collect the line width data of
the neutral FeK$\alpha $ line core obtained by
the {\it Chandra} High Energy Transmission Grating Spectrometer
\citep[HETGS;][]{1994SPIE.2280..168M}, which affords
the best spectral resolution currently available in the FeK$\alpha$
energy band ($\sim 39$ eV, or $\sim 1860$ km s$^{-1}$
in full width at half maximum; FWHM),
and we further select target AGNs with other
mass estimates of their central supermassive black holes.
For the type-1 AGNs, seven target AGNs
whose black hole masses are measured by the reverberation
of the broad emission lines
are selected out of 12 AGNs,
for which the best FWHM constraint for
the neutral FeK$\alpha $ line core
was obtained by \citet{2010ApJS..187..581S}.
The type-1 target AGNs and
their data are listed in Table \ref{data_sy1}.
For the type-2 AGNs, four target AGNs
whose black hole masses are measured by
the VLBI observation of the H$_2$O maser
are selected out of eight AGNs,
for which the FWHM of the neutral FeK$\alpha $ line core
was measured by \citet{2011ApJ...738..147S}.
In addition, one type-2 AGN from the 12 AGNs of \citet{2010ApJS..187..581S}
and five type-2 AGNs from the eight AGNs of \citet{2011ApJ...738..147S}
are also selected,
for which the FWHMs of the polarized broad Balmer
emission lines are available to obtain
the single-epoch black hole mass estimate
on the basis of the luminosity scaling relation of the BLR.
In total, seven type-2 AGNs are selected
because three objects in the $4+6$ type-2 AGNs are common.
The type-2 target AGNs and their data are listed in Table \ref{data_sy2}.
\section{RESULTS}
\subsection{Velocity Width of the Neutral FeK$\alpha $ Line Core}
First, we compare the velocity width of the neutral FeK$\alpha $ line core
with that of the broad Balmer emission lines,
and that corresponding to the dust reverberation radius
to examine the location of its emitting region.
In Figure \ref{fwhm_fwhm}, we plotted
the FWHM of the neutral FeK$\alpha $ line core
against those of the broad H$\beta $ emission line
and the polarized broad Balmer emission lines
for the type-1 and type-2 target AGNs, respectively.
We also plotted the line that represents
the FWHM at the dust reverberation radius observed in the $K$ band,
which is estimated by scaling the FWHM of
the broad H$\beta $ emission line according to the virial relation
based on the systematic difference of
the reverberation radii of the broad H$\beta $ emission line
and the near-infrared dust emission \citep[][]{kosh+14}.
As shown in Figure \ref{fwhm_fwhm},
we find that the FWHM of the neutral FeK$\alpha $ line core
falls between that of the broad Balmer emission lines
and the corresponding value at the dust reverberation radius
for all of the type-2 AGNs, and for most of the type-1 AGNs
except the only one outlier of NGC 7469.
This suggests that
significant fraction of photons of the neutral FeK$\alpha $ line core
originate between the outer BLR and the inner dust torus in most cases.
\subsection{Luminosity Scaling Relation of the Neutral FeK$\alpha $ Line Core Emitting Region}
As a next step,
we scaled the radius-luminosity scaling relation of
the broad H$\beta $ emission line
according to the systematic difference between
the FWHMs of the neutral FeK$\alpha $ line core
and the broad H$\beta $ emission line for the type-1 AGNs
assuming the virial relation.
It is estimated
by the linear regression analysis where the slope is fixed to unity,
and
the best-fit linear regression is
$\log {\rm FWHM_{FeK\alpha }}=\log {\rm FWHM_{H\beta}}-0.186\ (\pm 0.088)$
with an additional scatter of 0.11 dex in both directions
for the reduced $\chi ^2$ to achieve unity.
Then, the emitting region radius of the neutral FeK$\alpha $ line core
is estimated as
\begin{equation}
\log {\rm r_{FeK\alpha }}=\log {\rm r_{H\beta}}-0.37\ (\pm 0.18)
\label{r_fekahb}
\end{equation}
according to the virial relation.
\citet{kosh+14} estimated the luminosity scaling relation of
the reverberation radii to the luminosity of
the [\ion{O}{4}] emission line in the form of
$\log r = \alpha + 0.5 \log L_{{\rm [OIV]}}/10^{41}\ {\rm erg}\ {\rm s}^{-1}$,
as
$\alpha = -1.94\ (\pm 0.07)$ and $\alpha = -1.28\ (\pm 0.07)$
for
the reverberation radius of the broad H$\beta $ emission line
and
the dust reverberation radius observed in the $K$ band,
respectively.
By scaling the former relation
according to Equation (\ref{r_fekahb}),
the luminosity scaling relation of the emitting region radius
of the neutral FeK$\alpha $ line core is estimated as
\begin{equation}
\log r_{{\rm FeK\alpha}} = -1.57\ (\pm 0.19) + 0.5\log L_{{\rm [OIV]}}/10^{41}\ {\rm erg}\ {\rm s}^{-1}\ .
\label{r_loiv}
\end{equation}
The radius-luminosity scaling relation
for the neutral FeK$\alpha $ line core
of Equation (\ref{r_loiv})
is located between the reverberation radii
of the broad H$\beta $ emission line
and the dust torus emission in the $K$ band
as indicated by Figure \ref{fwhm_fwhm},
and we use it
for the estimation of the black hole mass
in the next section.
\subsection{Estimating the Black Hole Mass from the Neutral FeK$\alpha $ Line Core}
Next,
we estimate the black hole mass
of both type-1 and type-2 target AGNs
using the neutral FeK$\alpha $ line core.
The emitting region radius $r_{\rm FeK\alpha}$ is estimated
from $L_{{\rm [OIV]}}$ and the radius-luminosity scaling relation of Equation (\ref{r_loiv})
because the hard X-ray flux seems to be attenuated
in some heavily obscured AGNs in the targets.
The $L_{{\rm [OIV]}}$ value is taken from \citet{liuw+10},
on the basis of the original data of \citet{mele+08} and \citet{diam+09}
for most of the targets;
however, it is recalculated from the [\ion{O}{4}] emission-line flux of \citet{diam+09},
assuming the redshift-independent distance
taken from the NASA/IPAC Extragalactic Database for very nearby objects.
In addition, the $L_{{\rm [OIV]}}$ of NGC 2110
is calculated from the data of \citet{2010ApJ...716.1151W}.
We then estimate the black hole mass assuming the virial relation as
\begin{equation}
M_{\rm BH,FeK\alpha} = f\frac{r_{\rm FeK\alpha}\sigma _{\rm FeK\alpha }^2}{G}
\label{eqn_virial}
\end{equation}
where $f$ is the virial factor and
$\sigma_{\rm FeK\alpha}$ is the velocity dispersion of the neutral FeK$\alpha $ line core.
We adopt $f=5.5$ following the references of the broad emission-line reverberation
for the type-1 target AGNs
\citep[][]{2004ApJ...615..645O,2004ApJ...613..682P,bent+06,bent+07,denn+10,grie+12}.
The velocity dispersion is calculated as $\sigma _{\rm FeK\alpha}={\rm FWHM_{\rm FeK\alpha}}/2.35$
because a Gaussian emission-line profile is used for the spectral fitting
of the neutral FeK$\alpha $ line core in
\citet{2010ApJS..187..581S,2011ApJ...738..147S}.
The $L_{{\rm [OIV]}}$ and the $M_{\rm BH,FeK\alpha}$ for the type-1 and type-2 target AGNs
are listed in Tables \ref{data_sy1} and \ref{data_sy2}, respectively.
The FWHM of the neutral FeK$\alpha$ line core
spreads between that of the broad H$\beta $ line
and that corresponding to the dust reverberation radii in most cases.
This scatter causes a systematic uncertainty in the $r_{\rm FeK\alpha}$
derived from the radius-luminosity scaling relation, and also yields
$\pm 0.3$ dex or a factor $\sim 2$ uncertainty in this black hole mass estimates,
$M_{\rm BH,FeK\alpha}$, including the measurement errors of the line widths in the data.
As in the case of NGC 7469,
in which the FWHM of the neutral FeK$\alpha$ line is
significantly larger than that of the broad H$\beta $ line,
$M_{\rm BH,FeK\alpha}$ could be sometimes overestimated.
\subsection{Comparison with Other Black Hole Mass Indicators}
Finally,
the black hole mass $M_{\rm BH,FeK\alpha}$
estimated from the neutral FeK$\alpha $ line core
is compared with
other estimates of the black hole mass for the same target
to examine the consistency between them.
For the type-1 AGNs,
we compared $M_{\rm BH,FeK\alpha}$ with
the broad emission-line reverberation mass, $M_{\rm BH,rev}$,
and for the type-2 AGNs,
we compare it with the mass $M_{\rm BH,H_2O}$ based on the H$_2$O maser
and the single-epoch mass estimate $M_{\rm BH,pol}$
based on the polarized broad Balmer lines.
The $M_{\rm BH,pol}$ is calculated in the same way as the $M_{\rm BH,FeK\alpha}$,
except that the radius-luminosity scaling relation
of the broad H$\beta $ emission line presented by \citet{kosh+14} is used.
The $M_{\rm BH,rev}$, $M_{\rm BH,H_2O}$, $M_{\rm BH,pol}$,
and their references are listed in Tables \ref{data_sy1} and \ref{data_sy2}.
In Figure \ref{mbh_Sy1_Sy2}, $M_{\rm BH,FeK\alpha}$
is plotted against
other black hole mass indicators such as
$M_{\rm BH,rev}$, $M_{\rm BH,H_2O}$, and $M_{\rm BH,pol}$
for the type-1 and type-2 target AGNs.
We then calculated the Pearson's correlation coeffient,
and also performed linear regression analysis
to $M_{\rm BH,FeK\alpha}$ and the other black hole mass indicators.
The resultant values are listed in Table \ref{fit_results},
and the best-fit linear regression lines are also presented
in Figure \ref{mbh_Sy1_Sy2}.
For the type-1 AGNs,
$M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,rev}$
are consistent within $\pm 2\sigma $ except for NGC~7469.
The best-fit linear regression where the slope is fixed to unity
is estimated as
{\footnotesize
\begin{equation}
\log (M_{\rm BH,FeK\alpha}/10^{7.5}M_{\odot}) = -0.15\ (\pm 0.11) + \log (M_{\rm BH,rev}/10^{7.5}M_{\odot})
\label{eqn_feka_rev2}\ ,
\end{equation}}
which indicates that the systematic difference between them is
within a factor of several tens percent.
However,
the correlation between $M_{\rm BH,FeK\alpha}$ and
$M_{\rm BH,rev}$ is not clear as with the preceding study
\citep{2011SCPMA..54.1354J}.
Excluding the clear outlier NGC 7469,
the Pearson's correlation coefficient is calculated
as $R=0.701$, which indicates that the confidence level
of the correlation is less than $90$\%.
The best-fit linear regression for them
is estimated as
{\footnotesize
\begin{equation}
\log (M_{\rm BH,FeK\alpha}/10^{7.5}M_{\odot}) = -0.19\ (\pm 0.17) + 1.52\ (\pm 0.92) \times \log (M_{\rm BH,rev}/10^{7.5}M_{\odot})
\label{eqn_feka_rev}
\end{equation}}
with an additional scatter in both directions
for the reduced $\chi ^{2}$ to achieve unity.
The large uncertainty in the slope of the best-fit linear regression
also indicates no or a weak correlation between them.
For the type-2 AGNs,
$M_{\rm BH,FeK\alpha}$ are also consistent with $M_{\rm BH,pol}$ within $\pm$2$\sigma$.
The best-fit linear regression for $M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$
is estimated as
{\footnotesize
\begin{equation}
\log (M_{\rm BH,FeK\alpha}/10^{7.5}M_{\odot}) = -0.13\ (\pm 0.13) + 1.05\ (\pm 0.25) \times \log (M_{\rm BH,pol}/10^{7.5}M_{\odot})\ ,
\label{eqn_feka_pol}
\end{equation}}
which indicates that the systematic difference between
$M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$ is also
within a factor of several tens percent.
In addition, $M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$ are well correlated
and show an approximately proportional relationship, since
the slope of the best-fit linear regression
is positive more than $3\sigma $ uncertainties and close to unity.
The Pearson's correlation coefficients are calculated as $R=0.831$,
which indicates that the confidence level
of the correlation is more than $95$\%.
In contrast to the good agreement with $M_{\rm BH,pol}$,
$M_{\rm BH,FeK\alpha}$ seems to be systematically larger than $M_{\rm BH,H_2O}$
although they show a clear correlation
with the Pearson's correlation coefficient
of $R=0.960$ indicating the confidence level of more than $95$\%.
The best-fit linear regression is estimated as
{\footnotesize
\begin{equation}
\log (M_{\rm BH,FeK\alpha}/10^{6.5}M_{\odot}) = 0.70\ (\pm 0.12) + 1.57\ (\pm 0.33) \times \log (M_{\rm BH,H_2O}/10^{6.5}M_{\odot})\ ,
\label{eqn_feka_h2o}
\end{equation}}
which indicates that
$M_{\rm BH,FeK\alpha}$ is systematically larger than $M_{\rm BH,H_2O}$
by about a factor of $\sim 5$.
We note that $M_{\rm BH,pol}$ is also larger
than $M_{\rm BH,H_2O}$ by a similar factor
for the three targets, NGC 1068, NGC 4388, and Circinus,
for which both $M_{\rm BH,H_2O}$ and $M_{\rm BH,pol}$ are estimated.
To summarize the results,
$M_{\rm BH,FeK\alpha}$ is consistent with
$M_{\rm BH,rev}$ for most of the type-1 AGNs
and with $M_{\rm BH,pol}$ for all of the type-2 AGNs.
$M_{\rm BH,FeK\alpha}$ is correlated well
with $M_{\rm BH,pol}$ and $M_{\rm BH,H_2O}$
for the type-2 AGNs,
although $M_{\rm BH,H_2O}$ is a systematic smaller than the others.
These results suggest that
the black hole mass estimate using the neutral FeK$\alpha$ line
is a potential indicator of the black hole mass,
especially for obscured AGNs.
\section{DISCUSSION}
\subsection{Systematic Difference of the H$_2$O maser mass from the Other Black Hole Mass Indicators}
The systematic difference of $M_{\rm BH,H_2O}$
implies an ambiguous issue.
Since $M_{\rm BH,H_2O}$ is considered to be the most
precise estimate for the black hole mass, then,
through the relations of $M_{\rm BH,H_2O}$
with $M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$,
another reliable estimate $M_{\rm BH,rev}$
is suspected to considerably overestimate the black hole mass
that is difficult to be accounted for
by the uncertainties in the virial factor $f$
($5.5$ in this paper and presented by Onken et al. 2004;
$5.2$ by Woo et al. 2010;
$2.8$ by Graham et al. 2011; $4.31$ by Grier et al. 2013).
In this section,
the systematic differences between
those black hole mass estimates
and the possible origins of these differences are discussed.
In fact,
the systematic difference between $M_{\rm BH,pol}$ and $M_{\rm BH,H_2O}$
has been indicated by preceding studies.
\citet{2010ApJ...721...26G} measured
the stellar velocity dispersion $\sigma _{\ast}$
in the centers of the galaxies hosting type-2 AGNs
whose $M_{\rm BH,H_2O}$ was obtained,
and found that $M_{\rm BH,H_2O}$ tended to be smaller
than the black hole mass that was estimated from $\sigma _{\ast}$
and the $M_{\rm BH}$--$\sigma _{\ast}$ relation of
local elliptical galaxies.
On the other hand, \citet{2008A&A...488..113Z}
estimated the radius of the BLR from the [\ion{O}{3}] luminosity
via the luminosity scaling relation of the BLR
to estimate $M_{\rm BH,pol}$ of 12 type-2 AGNs,
and found that $M_{\rm BH,pol}$ tended to be larger
than the black hole mass that was estimated from $\sigma _{\ast}$
via the $M_{\rm BH}$--$\sigma _{\ast}$ relation.
According to these results, $M_{\rm BH,pol}$ would
be systematically larger than $M_{\rm BH,H_2O}$,
as is the case in this study.
Recently, \citet{2014ApJ...789...17H} presented that
the virial factor $f$ of the reverberation-mapped AGN
is systematically different between two bulge types of the host galaxies,
as $f=6.3\pm 1.5$ for classical bulges and ellipticals,
and $f=3.2\pm 0.7$ for psuedobulges. In fact,
the black hole mass of the AGNs with psuedobulges
tends to be smaller than that with classical bulges and ellipticals
as shown in their Figure 2.
\citet{2010ApJ...721...26G} also argued that
the mass of the black hole hosted by later-type and lower-mass galaxies
tended to be less massive with larger scatter than that estimated from
the $M_{\rm BH}$--$\sigma _{\ast}$ relation for the elliptical galaxies.
Since our type-2 AGN targets with the H$_2$O maser observation
tend to have a less massive black hole than the others
if $M_{\rm BH,H_2O}$ is correct,
then, it is possible that
the systematic difference of
the target selection
partly contributes to the systematic difference
between $M_{\rm BH,H_2O}$ and the other black hole mass indicators.
Contrary to these studies,
\citet{2011ApJ...727...20K} estimated the radius of the BLR from
the absorption-corrected luminosity in the 2--10 keV X-ray band
$L_{2-10 {\rm keV}}$
via the luminosity scaling relation of the BLR
to estimate $M_{\rm BH,pol}$ of 4 type-2 AGNs
whose $M_{\rm BH,H_2O}$ was obtained by them,
and presented that $M_{\rm BH,H_2O}$ and $M_{\rm BH,pol}$
were consistent with each other.
In fact, the three targets, NGC 1068, NGC 4388, and Circinus,
for which the large systematic difference
between $M_{\rm BH,pol}$ and $M_{\rm BH,H_2O}$ is presented in this paper,
are included in the targets of \citet{2011ApJ...727...20K}.
The reason for the discrepancy between these studies is uncertain,
but a difficulty in estimating the luminosity of
the central engine $L_{2-10 {\rm keV}}$ for most of the targets
with very large X-ray absorbing column density
could be suspected,
although it was estimated with particular attention.
Assuming both $M_{\rm BH,rev}$ and $M_{\rm BH,H_2O}$
are reliable estimates for the black hole mass,
$M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$
are considered to overestimate the black hole mass
in type-2 AGNs.
The systematically larger $M_{\rm BH,FeK\alpha}$
for the type-2 AGNs implies that
the width of the neutral FeK$\alpha $ line core
is systematically larger for the large inclination angle for type-2 AGNs.
If the emitting region of the neutral FeK$\alpha $ line core
shows an equatorial motion,
then the virial factor $f$ in Equation (\ref{eqn_virial}) decreases
as the inclination angle increases.
For example, $f=2$ at the edge-on view,
which is smaller than $f=5.5$ for type-1 AGNs,
is derived assuming an equatorial circular motion.
However, it is not so small as to account for
the difference of about a factor of $\sim 5$.
Outflow motions toward the equatorial plane
might broaden the width of
the neutral FeK$\alpha $ line core for type-2 AGNs.
On the other hand,
the systematically larger $M_{\rm BH,pol}$ for the type-2 AGNs
cannot be attributed to the inclination of AGNs
from the point of view of the unified model of AGNs;
when the broad emission line is scattered to the observer
at the polar region well above the height of the dust torus,
the viewing angle of the BLR from the scattering region
would be similar to that from the observer for type-1 AGNs.
Non-virial motions such as outflows in the scattering region
are possible to be contributed, as \citet{2007Natur.450...74Y}
proposed to explain the broad H$\alpha $ emission line
in the polarized spectrum of the quasar PG 1700$+$518.
However, a common or closely related mechanism to broaden
the widths of both of the neutral FeK$\alpha $ line core and
the polarized broad Balmer emission lines
systematically for type-2 AGNs is required
to explain the approximate agreement between
$M_{\rm BH,FeK\alpha}$ and $M_{\rm BH,pol}$.
\subsection{Location of the Neutral FeK$\alpha $ Line Core Emitting Region}
As described in Section 3,
the FWHM of the neutral FeK$\alpha $ line core falls
between the FWHM of the broad Balmer emission lines and
the corresponding values at the dust reverberation radius
in most cases,
which indicates that the emitting region of
its major fraction is located
between the outer BLR and the inner dust torus.
It is suggested by previous studies for some of the target AGNs.
For NGC 4151, \citet{2002PASJ...54..373T} compared
the variation of the continuum and the excess flux
around the neutral FeK$\alpha $ line and presented that
the emitting region of the excess flux has an extent of $10^{17}$ cm,
or 40 light days, which is consistent with or slightly smaller than
the dust reverberation radius \citep{kosh+14}.
For NGC 5548, \citet{2010ApJ...710.1228L} reported that
the emitting region of the neutral FeK$\alpha $ line
was located between the two reverberation radii
of the broad H$\beta $ emission line and the near-infrared dust emission
based on its flux variation and line width.
The BLR and the dust torus are considered not to be decoupled
from each other but to constitute a continuous structure,
and the transition zone of the BLR and the dust torus
would be located between the two reverberation radii
\citep{2008ApJ...685..160N,2009ApJ...700L.109K,kosh+14}.
In addition, recent reverberation observations
of the optical \ion{Fe}{2} emission lines
presented that its reverberation radius
was comparable to or at most twice
that of the broad H$\beta $ emission line
\citep{2013ApJ...769..128B,2014ApJ...783L..34C},
which shows the presence of low ionization iron atoms there.
If the FeK$\alpha $ emitting region constitutes
a tight stratified structure with the BLR and the dust torus,
a good correlation between the FWHMs of
the neutral FeK$\alpha $ line and the broad Balmer emission lines
is expected.
However, as shown in Figure \ref{fwhm_fwhm},
there appears no clear correlation between them
as has been reported by previous studies
\citep[][]{2006MNRAS.368L..62N,2010ApJS..187..581S,2011ApJ...738..147S}.
A simple explanation for it would be that the neutral FeK$\alpha$ line
is produced in different origins such as the outer accretion disk,
the BLR, and the innermost dust torus,
and the mixing ratio is different from target to target
\citep[e.g.,][]{2004ApJ...604...63Y,2006MNRAS.368L..62N,2008MNRAS.389L..52B,2010ApJ...710.1228L,2013A&A...549A..72P}.
On the other hand,
\citet[][]{2011ApJ...738..147S}
suggested that the neutral FeK$\alpha$ line may originate
from a universal region at the same radius with respect
to the gravitational radius of the central black hole,
because its FWHM is almost constant and independent
of the black hole mass.
In any case, those uncertainties
will result in the possible
error of the black hole mass estimated from
the neutral FeK$\alpha$ line.
In this study, We used the best spectral resolution data
currently available,
but the spectral resolution and sensitivity are still limited.
As a result, the number of the targets is small
and the relatively large uncertainties remain in the FWHM data,
which would make it difficult to examine
the location of the neutral FeK$\alpha$ line emitting region
in more detail.
In the near future,
the substantial progress is expected
by the {\it ASTRO-H} X-ray satellite \citep{2010SPIE.7732E..27T},
which is capable of unprecedented energy-resolution spectroscopy
with superior sensitivity at FeK$\alpha $ band.
It will enable us to study more detail on the origin of the neutral FeK$\alpha $ line core
by examining the line profile of the neutral FeK$\alpha $ line
and its time variation,
then it will become able to measure the black hole mass
with improved accuracy
for a large number of obscured AGNs.
The hard X-ray luminosity with precise absorption correction
is obviously valuable
for the isotropic luminosity indicator
because the scaling relations of the reverberation radii
to the hard X-ray luminosity are tighter than to the [\ion{O}{4}] luminosity
\citep{kosh+14}.
\acknowledgments
We thank T. Kawaguchi, M. Yoshida, K. Kawabata, and Y. Fukazawa
for useful discussions and comments.
We also thank the anonymous referee
for valuable comments to improve the manuscript.
This research has been partly supported by
the Grants-in-Aid of Scientific Research (22540247 and 25287062)
of the Ministry of Education, Science, Culture and Sports of Japan,
and has made use of the NASA/IPAC Extragalactic Database (NED),
which is operated by the Jet Propulsion Laboratory,
California Institute of Technology,
under contract with the National Aeronautics and Space Administration.
|
\section{Introduction}
The primary intent of the ongoing nuclear collision programmes at
the Relativistic Heavy Ion Collider (RHIC) and Large Hadron Collider (LHC) energies
is to create a new state of matter called Quark Gluon Plasma (QGP).
The bulk properties of this state is governed by the light quarks ($q$) and gluons ($g$).
The heavy quarks [HQs\,$\equiv$ charm (c) and beauty (b)]
are considered as efficient tools to probe the early
state of the system~(see \cite{hfr,Rapp:2008tf,Averbeck:2013oga} for a review)
mainly for the following reasons:
(i) HQs are produced at very early stage;
(ii) The probability of creation and annihilation of the
HQs during the evolution of the fireball is small.
Hence the HQs can witness the entire evolution of
the system as they are created early and
survive the full evolution without being annihilated and/or created.
The data on the suppression of the charm quark at large momentum
($R_{AA}(p_T)$) ~\cite{stare,phenixelat,phenixe,alice}
and their elliptic flow ($v_2$)~\cite{phenixelat,alicev2} have been analyzed
by several authors (see {\it e.g.} \cite{hfr,Rapp:2008tf,Averbeck:2013oga} and references therein)
to characterize the system formed in HIC at RHIC and LHC collisions.
It has been argued that the thermalization times of
HQs are larger than the light quarks and gluons~\cite{shuryak,prlja}
by a factor $\sim M/T$~\cite{moore}, where $M$ is the mass of the HQs and $T$
is the temperature of the medium. Therefore, the evolution of
the HQs in the thermal bath of light quarks and gluons can be
treated within the ambit of Brownian motion, although a detailed investigation
of this problem within the framework of Boltzmann equation has revealed that
the evolution of charm quarks as a Brownian particle requires some corrections~\cite{fs}
(see also ~\cite{Molnar:2006ci,gossiauxv2,gre,you,Uphoff:2014cba,Song:2015sfa}).
Several theoretical attempts have been made
to study the evolution of the HQs within the framework of
Fokker Plank equation
~\cite{hfr,moore,rappv2,hvh,hiranov2,cmko,Das,alberico,jeon,bass,rappprl,ali,hees,qun,lambda,Juan_RAA,Das:2015ana}.
However, the roles of pre-equilibration phase have been ignored in these works. An attempt
has been made in the present work to study this role. This is important because the
HQs are produced in the hard collisions of the colliding partons of the nuclei and
they inevitably interact with the pre-equilibrium phase of the bulk matter.
The effect of the pre-equilibrium phase might be more significant for low-energy nuclear
collisions: for example, in the case Au+Au collisions at RHIC energy
simulations show that equilibration is achieved approximately within $1$ fm/c,
while the lifetime of the QGP phase turns out to be about $5$ fm/c \cite{Ruggieri:2013bda,Ruggieri:2013ova};
hence the lifetime of the out-of-equilibrium phase is approximately $20\%$ of the
total lifetime of the QGP.
The lifetime of the QGP phase is
of about 10 fm/c and the duration of the pre-equilibrium phase is
about 0.4 fm/c for LHC collision condition. Although, this may suggest the dwindling role of the pre-equilibrium phase
for collisions at higher energies, estimates of drag and diffusion coefficients of HQs
done in this work indicate non-negligible contributions
of the pre-equilibrium phase.
The motivation of this work is to estimate the drag and diffusion coefficients of the HQs
interacting elastically with the non-equilibrated gluons constituting
the medium.
The paper is organized as follows. In the next section we discuss the formalism
used to evaluate the drag and diffusion coefficients of the heavy quarks in the
preequilibrium stage. In Section III we summarize the
out-of-equilibrium initial conditions we implement in the calculations.
Section IV is devoted to present the results and
section V contains summary and discussions.
\section{Formalism}
The Boltzmann Transport Equation (BTE) describing the evolution of the HQs in the pre-equilibrated gluonic system can be written as:
\bea
\left[\frac{\partial}{\partial t}+\frac{\vec{p}}{E}\cdot\frac{\partial}{\partial \vec{x}}
+\vec{F}\cdot\frac{\partial}{{\partial \vec{p}}}
\right] f(\vec{x},\vec{p},t)=\left[\frac{\partial f}{\partial t}\right]_{collisions}~.
\label{boltzmann}
\eea
For binary interaction the collisional integral appearing in the right hand
side of BTE can be written as:
\be
\left[\frac{\partial f}{\partial t}\right]_{collisions}= \int d^{3}\vec{k}[w(\vec{p}+\vec{k},\vec{k})
f(\vec{p}+\vec{k})-w(\vec{p},\vec{k})f(\vec{p})].
\label{BTE}
\ee
where $w(\vec{p},\vec{k})$ is the rate of collision which encodes the change of HQs momentum
from $\vec{p}$ to $\vec{p}-\vec{k}$ is given by,
\be
\omega(p,k)=g_G\int\frac{d^3q}{(2\pi)^3}f^\prime(q)v\sigma_{p,q\rightarrow p-k,q+k}
\label{eq:omega1}
\ee
where $f^\prime$ is the phase space distribution of the particles in the bulk,
$v$ is the relative velocity between the two collision partners,
$\sigma$ denotes the cross section and $g_G$ is the statistical
degeneracy of the particles in the bulk.
Considering only the soft scattering approximation~\cite{BS}, the integro-differential
Eq.~\ref{BTE} can be reduced to the Fokker Planck equation:
\be
\frac{\partial f}{\partial t}= \frac{\partial}{\partial p_{i}}\left[A_{i}(\vec{p})f+\frac{\partial}{\partial p_{j}}[B_{ij}
(\vec{p})f]\right]~~,
\label{landaukeq}
\ee
where the kernels are defined as
\be
A_{i}= \int d^{3}\vec{k}w(\vec{p},\vec{k})k_{i}~~,
\label{eqdrag}
\ee
and
\be
B_{ij}= \frac{1}{2} \int d^{3}\vec{k}w(\vec{p},\vec{k})k_{i}k_{j}~.
\label{eqdiff}
\ee
In the limit $\mid\bf{p}\mid\rightarrow 0$, $A_i\rightarrow \gamma p_i$
and $B_{ij}\rightarrow D\delta_{ij}$ where $\gamma$ and $D$
are the drag and diffusion coefficients
respectively.
We notice that under the assumption of soft collisions, the non-linear integro-differential equation, Eq.~\ref{boltzmann}
reduces to a much simpler linear partial differential equation, Eq.~\ref{landaukeq},
provided the function $f^\prime (q)$ is known.
Now we evaluate the $\gamma$ and $D$ for HQs interacting elastically
with the bulk particles in the pre-equilibrium phase that appear in HIC before thermalization.
The $\gamma$ for the process $HQ(p_1) + g(p_2) \rightarrow HQ(p_3) + g(p_4)$ ($p_i$'s are
the respective momenta of the colliding particles)
can be expressed in terms of $A_i$ ~\cite{BS} (see also ~\cite{DKS,vc,Berrehrah:2014tva}) as:
\begin{equation}
\gamma=p_iA_i/p^2~,
\end{equation}
where $A_i$ is given by
\bea
A_i&=&\frac{1}{2E_{p_1}} \int \frac{d^3p_2}{(2\pi)^3E_{p_2}} \int \frac{d^3p_3}{(2\pi)^3E_{p_3}}
\int \frac{d^3p_4}{(2\pi)^3E_{p_4}} \nonumber \\
&&\frac{1}{g_{HQ}}
\sum {\overline {|{\cal{M}}|^2}} (2\pi)^4 \delta^4(p_1+p_2-p_3-p_4)
\nonumber \\
&&{f}(p_2)\{1\pm f(p_4)\}[(p_1-p_3)_i] \equiv \langle \langle
(p_1-p_3)_i\rangle \rangle . \nonumber \\
\label{eq1}
\eea
where $g_{HQ}$ denotes the statistical degeneracy of HQ,
$f(p_2)$ is the momentum distributions of the incident particles,
$1 \pm f(p_4)$ is the
momentum distribution (with Bose enhancement or Pauli suppression)
of the final state particles in the bath and
${\overline {|{\cal{M}}|^2}}$ represents the square modulus of the spin averaged
invariant amplitude for the elastic process, evaluated here using pQCD.
The drag coefficient in Eq.~(\ref{eq1}) is the measure of the average of the momentum transfer, $p_1-p_3$,
weighted by the interaction through $\overline{|{\cal{M}}|^2}$.
Similarly the momentum diffusion coefficient, $D$, can be defined as:
\begin{equation}
D=\frac{1}{4}\left[\langle \langle p_3^2 \rangle \rangle -
\frac{\langle \langle (p_1\cdot p_3)^2 \rangle \rangle }{p_1^2}\right]
\label{eq3}~,
\end{equation}
and it represents the averaged square momentum transfer (variance) through the interaction process mentioned above.
The following general expression has been used to
evaluate the drag and diffusion coefficients numerically with an appropriate choice of $Z(p)$,
\bea
&&\ll Z(p)\gg=\frac{1}{512\pi^4} \frac{1}{E_{p_1}} \int_{0}^{\infty}
\frac{p_2^2 dp_2 d(cos\chi)}{E_{p_2}}
\nonumber \\
&&~~~\hat{f}(p_2)\{ 1\pm f(p_4)\}\frac{\lambda^{\frac{1}{2}}(s,m_{p_1}^2,m_{p_2}^2)}{\sqrt{s}}
\int_{1}^{-1} d(cos\theta_{c.m.})
\nonumber \\
&&~~~~~~~~~~~~~~~~~~
\frac{1}{g_{HQ}} \sum {\overline {|{\cal{M}}|^2}} \int_{0}^{2\pi} d\phi_{c.m.} Z(p) \nonumber \\
\label{transport}
\eea
where $\lambda(s,m_{p_1}^2,m_{p_2}^2)=\{s-(m_{p_1}+m_{p_2})^2\}\{s-(m_{p_1}-m_{p_2})^2\}$.
\section{Initial conditions}
In most of the earlier works the distribution functions
in Eq.~(\ref{transport}) are taken as
equilibrium distribution, {\it i.e.} Fermi-Dirac for quarks and
anti-quarks and Bose-Einstein for gluons.
The transport coefficients are then corresponding to
the motion of the HQs in a thermalised medium assumed to be formed
within the time scale $\sim 1$ fm/c.
In this article instead, we will evaluate the drag and diffusion coefficients of HQ propagating through a
non-thermal gluonic system.
We achieve this by substituting the distribution functions
in Eq.~(\ref{transport}) by pre-equilibrium distribution of gluons to be specified later.
We choose the normalization of the non-equilibrium distributions
in such a way that the gluon density for the chosen distribution
and the thermal distribution at initial temperature coincides.
The initial temperature, $T_i$ is taken as
$0.34$ GeV for RHIC and $0.51$ GeV for LHC collision conditions.
These initial temperatures are chosen from simulations~\cite{Ruggieri:2013bda,Ruggieri:2013ova,Niemi:2011ix}
done to reproduce $v_2$ and spectra of light hadrons, as well as the $R_{AA}$ and
$v_2$ of heavy mesons~\cite{fs,Das:2015ana}.
We compare the drag and diffusion coefficients of HQs in the pre-equilibrium era with those
in the thermalized era at the initial temperatures just mentioned.
We now specify the out-of-equilibrium gluon distribution used in this work for evaluating HQs drag and diffusion
coefficients.
According to the general understanding of the dynamics of pre-equilibrium stage
the initial strong gluon fields (the glasma) shatters into gluon quanta in a time scale
which is of the order of the inverse of the saturation scale $Q_s$;
therefore any model of the pre-equilibrium stage will include a $Q_s$.
The first one we consider
is the classical Yang-Mills (CYM) gluon spectrum (see~\cite{Schenke:2013dpa} for details),
which assumes the initial gluon fields can be expanded in terms
of massless gluonic excitations.
Beside CYM we also consider the model known as factorized KLN
model~\cite{Drescher:2006ca,Hirano:2009ah} which includes the saturation scale
in an effective way through the unintegrated gluon distribution functions which,
for the nucleus $A$, that participates in $A+B$ collision reads:
\begin{equation}
\phi_A\left(x_A,\bm k_T^2,\bm x_\perp\right) =
\frac{1}{\alpha_s(Q_s^2)}\frac{Q_s^2}{\it{max}(Q_s^2,\bm k_T^2)}~
\end{equation}
a similar equation holds for nucleus $B$.
The momentum space gluon distribution is then given by
\begin{eqnarray}
\frac{dN}{dy d\bm p_T}&=&
\frac{{\cal N}}{p_T^2}\int d^2 x_T\int_0^{p_T}d\bm k_T
\alpha_s(Q^2)\nonumber\\
&&\times\phi_A\left(x_A,\frac{(\bm k_T + \bm p_T)^2}{4},\bm x_\perp\right)\nonumber\\
&&\times\phi_B\left(x_B,\frac{(\bm k_T - \bm p_T)^2}{4},\bm x_\perp\right)~,
\label{eq:3}
\end{eqnarray}
where ${\cal N}$ is an overall constant which is fixed by the multiplicity.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{drag.eps}\hspace{2pc}~
\includegraphics[width=17pc,clip=true]{diffc.eps}\hspace{2pc}
\caption{Drag coefficient (left panel) and diffusion coefficient (right panel)
as a function of momentum for charm quark at RHIC energy.
The results corresponding to thermal gluons (kinetic equilibrium) and thermal quarks and
gluons (chemical equilibrium) are evaluated at a temperature 340 MeV.
}
\label{fig1}
\end{center}
\end{figure}
\section{Results and discussions}
We have evaluated
the drag and diffusion coefficients of HQs propagating through a system of out-of-equilibrium
gluons formed at the very early era of HIC at the RHIC and LHC
energies. The magnitudes of the transport coefficients evaluated in the pre-equilibrium era
are compared with those obtained in the equilibrium phase
(both kinetic and chemical) keeping the number of particles fixed
in both the cases as mentioned above.
The temperature dependence given in Ref.~\cite{zantow} has been used to estimate the value
of the strong coupling, $\alpha_s$ at $T=T_i$.
The Debye screening mass, $m_D=\sqrt{8\alpha_s(N_c+N_f)T^2/\pi}$,
for a system at temperature, $T$ with $N_c$ colors and $N_f$ flavors
has been used to shield the infra-red divergence associated with the $t$-channel
scattering amplitude. $m_D$ is estimated at $T=0.34$ GeV for RHIC energy and
at $T=0.51$ GeV for LHC energy.
For the pre-equilibrium and equilibrium system same values of $m_D$ and
$\alpha_s$ have been used.
Later in this section we will show results where $m_D$ is estimated
in a self-consistent way with the underlying distribution function.
In the left panel of Fig.~\ref{fig1} we plot the drag coefficient
of the charm quark as a function of momentum in the pre-equilibrium phase for CYM and KLN
gluon distributions and compared the results with the equilibrated phase (both kinetic and chemical)
at T=0.34 GeV.
We find that the magnitude of the drag coefficient in the pre-equilibrium
phase is quite large, indeed comparable to the one in the kinetic equilibrium phase,
indicating
substantial amount of interaction of the charm quarks with the pre-equilibrated gluons.
We notice that $\gamma$ for the case of chemically equilibrated QGP (dotted line)
is smaller than the one of the purely gluonic system (solid line in Fig.~\ref{fig1}).
This can be understood by considering the fact that
to keep the total number of bath particles fixed some of the gluons in the purely gluonic system has
to be replaced by quarks in the chemically equilibrated QGP.
Because the gluon appears in more colours than quarks the cross section for $cg$ interaction
is larger than $cq$ which leads to larger drag of HQs in a pre-equilibrated gluonic system than
in a chemically equilibrated QGP.
The CYM distribution gives larger value of the drag coefficient
compared to KLN distribution. The KLN provides harder momentum distribution (compared to CYM)
hence have a smaller difference with the HQs distribution. Therefore, the
momentum transfer between the gluon (with KLN distribution) and the HQ is small. As the
drag coefficient is a measure of the momentum transfer weighted by the interaction strength,
the KLN gives rise to lower drag compared to the one obtained with the CYM distribution.
The variation of $D$ with momentum for charm quarks has been depicted in
the right panel of Fig.~\ref{fig1}.
We find that the magnitude of the $D$ for CYM initial condition is similar
to the case where the gluons are kinetically equilibrated.
However, the $D$ for the KLN initial condition is larger than CYM.
As mentioned before, the KLN has harder momentum distributions than the
CYM, hence KLN distribution
has larger variance compared to CYM. Since the momentum diffusion
is a measure of the variance acquired through interaction with
the bulk, the charm diffusion coefficient is larger for KLN distribution,
which is reflected in the results displayed in Fig.~\ref{fig1}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{DbyA.eps}\hspace{2pc}
\caption{$D/\gamma$ as a function of momentum for charm quark at RHIC energy.
The result corresponding to equilibrium cases are evaluated at a temperature 340 MeV.
}
\label{fig300}
\end{center}
\end{figure}
In Fig.\ref{fig300}, we have depicted the diffusion to drag ratio, $D/\gamma$ as function of momentum.
$D/\gamma$ can be used to understand the deviation of the calculated values from the value
obtained by using Fluctuation-Dissipation theorem (FDT)
(green line in the figure).
Since the KLN has a harder momentum distribution, results obtained from
KLN input deviates from FDT more.
In this work, the dependence of the drag/diffusion coefficient on the variation of the
phase space distribution is addressed. To make the present study
more consistent the effects of phase space on the dynamics through
Debye screening mass has been included. It is well-known that the Debye mass depends on the
equilibrium distribution as follows:
\be
m_D^2=\pi\alpha_s g_G \int \frac{d^3p}{(2\pi)^2} \frac{1}{p} (N_cf_g+N_ff_q)
\label{mdd}
\ee
It is interesting
to see how the results are affected when the equilibrium distribution
in Eq.~\ref{mdd} is replaced by the KLN or CYM distributions.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{drag_1.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{diffc_1.eps}\hspace{2pc}
\caption{Drag coefficient (left panel) and diffusion coefficient (right panel)
as a function of momentum for charm quark at RHIC energy.
Debye mass is computed self consistently using Eq.~\ref{mdd}.
The results corresponding to thermal gluons (kinetic equilibrium) and thermal quarks and
gluons (chemical equilibrium) are evaluated at a temperature 340 MeV. }
\label{fig3}
\end{center}
\end{figure}
Replacing the thermal distribution of gluons in Eq.~(\ref{mdd}) by
the KLN and CYM gluon distributions and setting $f_q=0$ for a gluonic system we estimate $m_D$ and use
this to calculate $\gamma$ and $D$.
The momentum variation of $\gamma$ and $D$ are displayed in Fig.~\ref{fig3} for charm quarks.
It is interesting to note that in this case
$\gamma$ is almost unaffected by the bulk distributions, {\it i.e.}
of KLN, CYM and thermal gluons. We find some difference between the aforementioned bulks and the
chemically equilibrated QGP, which is caused by the less number of gluons
in the latter case as discussed earlier.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{drag_b_1.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{diffc_b_1.eps}\hspace{2pc}
\caption{Drag coefficient (left panel) and diffusion coefficient (right panel)
as a function of momentum for bottom quark at RHIC energy.
Debye mass is computed selfconsistently using Eq.~\ref{mdd}.
The results corresponding to thermal gluons (kinetic equilibrium) and thermal quarks and
gluons (chemical equilibrium) are evaluated at a temperature 340 MeV. }
\label{fig5}
\end{center}
\end{figure}
The drag (left panel) and diffusion (right panel)
coefficients for $b$ quarks in the pre-equilibrium phase
have been displayed in Fig.~\ref{fig5}.
We notice that these coefficients are smaller for $b$ than $c$.
However, the qualitative variation of $b$ diffusion coefficient
with momentum is similar to $c$.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{drag_lhc_1.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{diffc_lhc_1.eps}\hspace{2pc}
\caption{Drag coefficient (left panel) and diffusion coefficient (right panel)
as a function of momentum for charm quark at LHC energy.
Debye mass is computed selfconsistently using Eq.~\ref{mdd}.
The results corresponding to thermal gluons (kinetic equilibrium) and thermal quarks and
gluons (chemical equilibrium) are evaluated at a temperature 510 MeV. }
\label{fig7}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=17pc,clip=true]{drag_lhc_b_1.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{diffc_lhc_b_1.eps}\hspace{2pc}
\caption{Drag coefficient (left panel) and diffusion coefficient (right panel)
as a function of momentum for bottom quark at LHC energy.
Debye mass is computed selfconsistently using Eq.~\ref{mdd}.
The results corresponding to thermal gluons (kinetic equilibrium) and thermal quarks and
gluons (chemical equilibrium) are evaluated at a temperature 510 MeV. }
\label{fig9}
\end{center}
\end{figure}
In the left panel of Fig.~\ref{fig7}, the momentum variation of the $\gamma$ of the $c$ quark
in the pre-equilibrium phase is depicted for LHC collision conditions.
This result is compared with the one computed for a QGP at temperature T=0.51 GeV.
Here same values of the coupling for both the
equilibrium and pre-equilibrium system have been used.
The value of $m_D$ is taken from Eq.~(\ref{mdd}) for all the cases. We
find that the magnitude of the drag
coefficient in pre-equilibrium phase is similar to that obtained for the kinetic equilibrium case,
however, for a QGP the value of drag is smaller because of the less number of gluons as mentioned earlier.
In the right panel of Fig.~\ref{fig7} we show the variation of $D$ with $p$ for the $c$ quarks at LHC energy.
Again the variation of the diffusion coefficient at LHC energy is qualitatively similar to
RHIC. Similar to RHIC the diffusion at LHC conditions is larger for the KLN distribution.
Similarly the $\gamma$ and $D$ for $b$ quarks
are plotted in Fig.~\ref{fig9} for LHC collision conditions.
Qualitatively the variation of these coefficients with $p$
is similar to RHIC. Although, the quantitative values
at LHC collision conditions are larger than RHIC as expected owing
to the larger density and temperature..
\section{Summary and discussions}
In this work we have studied the momentum variation of the drag and diffusion coefficients of HQs
in the pre-equilibrium era of heavy ion collisions at RHIC and LHC energies.
The momentum distribution of the pre-equilibrated gluons have
been taken from CYM and KLN formalisms.
This study is motivated by the fact that the effect
of the pre-equilibrium phase on the HQs suppression and elliptic flow
might be relevant for low-energy nuclear collisions.
For example, the simulations of Au+Au collision at RHIC energy
show that the equilibration is achieved approximately within a time scale of $1$ fm/c,
while the lifetime of the QGP phase turns out to be about $5$ fm/c \cite{Ruggieri:2013bda,Ruggieri:2013ova};
hence the system spends about $20\%$ of QGP life-time in the pre-equilibrium phase.
In the case of Pb-Pb collisions at LHC energy the equilibration time
is shorter and lifetime of the QGP phase is larger, hence in this case we expect
a smaller effect vis-a-vis pre-equilibrium phase.
We have compared the magnitudes of the transport coefficients computed
for equilibrated and pre-equilibrated system, keeping the number
of particles same in the two cases. We have found that
the magnitude of the transport coefficients in the pre-equilibrium phase is comparable to
the values obtained with a thermalized gluonic system.
Moreover, we have also found that the transport coefficients in the pre-equilibrium era are larger
than the ones obtained for a chemically equilibrated QGP system. This is due to the
fact that for a fixed number of particles the number of gluons are less
in the equilibrated QGP than the pre-equilibrated gluonic system and the $HQ+q$ cross section
is smaller than the $HQ+g$ cross section.
The results obtained in this work may have significant impact on the experimental
observables, for example on heavy mesons $R_{AA}$ and elliptic flow,
as well as on the suppression of single electron spectra originating from the decays
of heavy mesons and their elliptic flow. We will address these aspects in future works.
\vspace{2mm}
\section*{Acknowledgments}
We acknowledge B. Schenke and R. Venugopalan for kindly sending us the data
for the CYM spectrum. S.K.D, M.R and V.G acknowledge the support by the ERC StG under the QGPDyn
Grant n. 259684.
\section{References}
|
\section{Introduction}
In a landmark paper \cite{Rudin}, W.~Rudin and K.~T.~Smith
answered a question of J.~Korevaar by showing that, if $ X $ is a
strictly convex real Banach space of dimension $ n \neq 2 $ and,
for each finite-dimensional subspace $ \pi $ in $ X $, the best
approximation function $ P_\pi $ is linear, then $X $ is a Hilbert
space. Their theorem led to the following characterization of
ellipsoids: {\it If $K $ is a centrally symmetric compact convex
body in $n$-space (where $n$ is possibly infinite) such that, for
every $ \nu $-dimensional subspace $ \pi $ with $ 0 < \nu < n-1 $,
the union of the tangency sets of all support planes of $K $ which
are translates of $ \pi $ lies in a plane of dimension $n - \nu$,
then $ K $ is an ellipsoid.} This result is a vivid example of how
the characterization of ellipsoids is intimately tied to
characterizing the Banach spaces that are Hilbert spaces.
In \cite{JerMcA2013}, the second author with J.~Jer\'onimo showed
that a fi\-nite-di\-men\-sion\-al pointed cone in which every
bounded section has a center of symmetry is an ellipsoidal cone.
Hence these are exactly the cones over closed unit balls of
Hilbert spaces that have been translated away from the origin.
The primary goal of this paper is to generalize this result to
infinite-dimensional cones. While this may seem like a result
that would follow from a straightforward induction argument (at
least for a separable Banach space), such an approach is elusive.
We give an affirmative answer to the infinite-dimensional
generalization by carefully using the following fact: \emph{Cones
with bounded sections have dual cones with nonempty interiors}.
Fix a real normed vector space $V$. A \emph{cone} in $V$ is a
nonempty convex subset $C\subset V$ that is closed under
nonnegative scalar multiplication. The cone $C$ is \emph{pointed}
if it contains no line through the origin; that is, if $C \cap
(-C) = \braces{0}$.
Given a convex subset $K \subset V$, let $\aff(K)$ denote the
affine span of $K$, and let $\lin(K)$ be the linear span of $K$.
We write $\operatorname{int}(K)$ for the interior of $K$ in $V$ with respect to
the norm on $V$. The \emph{relative interior} $\relintr{K}$ and
the \emph{relative boundary} $\partial K$ are the interior and
boundary, respectively, of $K$ with respect to $\aff(K)$. The
\emph{cone over} $K$, denoted $\cone(K)$, is the intersection of
all cones containing $K$. A \emph{section} of $K$ is the nonempty
intersection of $K$ with a closed affine subspace of $V$. We call
a section $S$ of $K$ \emph{proper} if $S$ has codimension $1$ with
respect to $\aff(K)$.
While the concept of an ellipsoid in finite dimensions is well
known, in infinite dimensional vector spaces this requires careful
definition. We define a subset $E \subset V$ to be an
\emph{ellipsoid} if, for some $x \in \aff(E)$, there exists an
inner product on the linear space $\aff(E) - x$ such that $E - x$
is the closed unit ball corresponding to this inner product. An
\emph{ellipse} is a $2$-dimensional ellipsoid. A cone $C \subset
V$ is \emph{ellipsoidal} if some proper section of $C$ is an
ellipsoid.
A subset $S \subset V$ is \emph{centrally symmetric} if there
exists a point $x \in V$ such that $S - x = -(S - x)$. In this
case, $x$ is a \emph{center of symmetry} of $S$, and $x$ is the
unique center of symmetry of $S$ if $S$ is nonempty and bounded.
\begin{definition}[CSS Cones]
\label{defn:CSSCones}
Let $C \subset V$ be a cone. We say that $C$
satisfies the \emph{centrally symmetric sections
\textup{(}CSS\textup{)} property} if
\begin{enumerate}
\item
$C$ is closed and $\relintr{C} \ne \emptyset$,
\item
there exists a bounded proper section of $C$, and
\item
every bounded proper section of $C$ is centrally symmetric.
\end{enumerate}
We call a cone with the CSS property a \emph{CSS cone}.
\end{definition}
Our main result is that the CSS cones in $V$ are precisely the
ellipsoidal cones.
\begin{thm}[proved on p.\ \pageref{proof:CSSproof}]
\label{thm:CSSTheorem}
Let $C$ be a cone in a normed vector space $V$. Then $C$ is a
CSS cone if and only if $C$ is an ellipsoidal cone.
\end{thm}
The proof that finite-dimensional CSS cones are ellipsoidal
appeared in~\cite{JerMcA2013}. That article also established
another characterization of fi\-nite-di\-men\-sion\-al ellipsoidal
cones: Such a cone is ellipsoidal if and only if it is a so-called
\emph{FBI cone}.
\begin{definition}
\label{def:FBI}
Let $C \subset V$ be a cone. We say that~$C$ satisfies the
\emph{flat boundary intersections \textup{(}FBI\textup{)}
property} if
\begin{enumerate}
\item
$C$ is closed and $\relintr{C} \ne \emptyset$,
\item
there exists a bounded proper section of $C$, and
\item
for each $a \in \relintr{C}$, some proper section of $C$
contains $\partial C \cap \partial(a - C)$.
\end{enumerate}
We call a cone with the FBI property an \emph{FBI cone}. (See
Figure~\ref{fig:FBICone}.)
\end{definition}
\begin{figure}
\begin{center}
\includegraphics{FBI_cone.pdf}
\end{center}
\caption{An FBI cone $C$ and a translation of $-C$. The shaded
region is the convex hull of the intersection of the
boundaries. The FBI property implies that this convex hull is
contained in a hyperplane.}
\label{fig:FBICone}
\end{figure}
\begin{thm}[proved on p.~\pageref{proof:FBIimpliesEllipsoidal}]
\label{thm:FBIimpliesEllipsoidal}
Let $C$ be a cone in a normed vector space $V$. Then $C$ is a
FBI cone if and only if $C$ is an ellipsoidal cone.
\end{thm}
Unlike the proof of the CSS characterization of ellipsoidal cones,
the proof of the FBI characterization carries over with very
little change to the infinite-dimensional case. We give this
proof in Section \ref{sec:FBICones}. In \cite{JerMcA2013}, the
proof that CSS cones in $\mathbb{R}^{n}$ are ellipsoidal proceeded by
showing that CSS cones are FBI cones, so the proof that FBI cones
are ellipsoidal was key to the argument. Unfortunately, the proof
in \cite{JerMcA2013} that CSS cones are FBI cones relied on the
existence of a measure, so a different strategy is needed to prove
that CSS cones are ellipsoidal in infinite-dimensional normed
vector spaces.
As a corollary of Theorems \ref{thm:CSSTheorem} and
\ref{thm:FBIimpliesEllipsoidal}, we get two characterizations of
those normed vector spaces that are inner-product spaces.
\begin{corollary}
Let $V$ be a normed vector space. Then the following are
equivalent.
\begin{enumerate}
\item
$V$ is an inner product space.
\item
$V$ contains a CSS cone with nonempty interior.
\item
$V$ contains an FBI cone with nonempty interior.
\end{enumerate}
In particular, if a Banach space $X$ contains a CSS cone or an
FBI cone, then $X$ is a Hilbert space.
\end{corollary}
\section{Cone lemmas}
\label{sec:ConeLemmas}
In the sections that follow, we will require some lemmas regarding
cones, their duals, their sections, and the relationships between
these concepts and central symmetry. Most of these lemmas are
well known, though the property of central symmetry seems to be
little-studied in the context of infinite dimensional sections of
cones. The seminal monograph of Kre\u{\i}n and Rutman \cite{Krein}
still provides an excellent introduction to cones in linear
spaces. A more recent treatment may be found
in~\cite{AliTou2007}.
Let $V$ be a real normed vector space with norm $\norm{\cdot}$,
and let $V^{*}$ be the dual space of continuous linear functionals
on $V$ under the operator norm, also denoted by $\norm{\cdot}$.
Let a cone $C \subset V$ be given. Recall that the dual of $C$ is
the cone $\dualc{C} \subset \dual V$ defined by
\begin{equation*}
\dualc{C} \mathrel{\coloneqq} \braces{\phi \in \dual V \,:\, \text{$\phi(x)
\ge 0$ for all $x \in C$}}.
\end{equation*}
Recall also the following well-known result regarding cones in
normed vector spaces.
\begin{proposition}
Let $C$ be a cone in a normed vector space $V$. Then, under
the canonical embedding $V \hookrightarrow \ddual V$, we have
that $\operatorname{int}(C) \hookrightarrow \operatorname{int}(\ddualc C)$. Moreover,
the closure of $C$ under this embedding is $\ddualc{C}$.
\end{proposition}
We will find the following notation to be convenient: Given points
$x \in V$ and $\phi \in \dual V$, let $ S_x(\dualc C) $ and $
S_\phi(C) $ be sections of $ C^{*} $ and $ C $, respectively,
defined as follows
\begin{align*}
S_{x}(\dualc C) & \mathrel{\coloneqq} \braces{\psi \in \dualc C \,:\, \psi(x)
= 1}, \\
S_{\phi}(C) & \mathrel{\coloneqq} \braces{y \in C \,:\, \phi(y) = 1}.
\end{align*}
More generally, we will occasionally need a canonical affine
subspace in $V$ or $\dual V$ that is perpendicular to a given
affine subspace of $\dual V$ or $V$, respectively. Recall that,
if $L \subset V$ and $M \subset \dual{V}$ are linear subspaces,
then the \emph{annihilators} of $L$ and $M$ are defined by
\begin{align*}
L^{\perp} & \mathrel{\coloneqq} \braces{\phi \in \dual{V} \,:\,
\text{$\phi(x) = 0$ for all $x \in L$}},\\
\coperp{M} & \mathrel{\coloneqq} \braces{y \in V \,:\, \text{$\psi(y) = 0$
for all $\psi \in M$}}.
\end{align*}
Analogously, given an affine subspace $A \subset V$, respectively
$B \subset \dual{V}$, bounded away from the origin, define the
\emph{perpendicular affine spaces}
\begin{align*}
A^{\perp} & \mathrel{\coloneqq} \braces{\phi \in \dual{V} \,:\,
\text{$\phi(x) = 1$ for all $x \in A$}},\\
\coperp{B} & \mathrel{\coloneqq} \braces{y \in V \,:\, \text{$\psi(y) = 1$
for all $\psi \in B$}}.
\end{align*}
It is well known that, if $L \subset V$ is a closed linear
subspace and if $M \subset \dual{V}$ is a $\text{weak}^{\ast}$-closed linear
subspace, then the annihilators satisfy a duality relation:
$\coperp{(L^{\perp})} = L$ and $(\coperp{M})^{\perp} = M$.
\begin{lemma}
\label{lem:PerpIsADuality}
Let $V$ be a normed vector space. Let $A \subset V$ and $B
\subset \dual{V}$ be closed affine subspaces such that $0
\notin A$, $0 \notin B$, and $\operatorname{codim}(A)$ and $\dim(B)$ are
finite. Then $\coperp{(A^{\perp})} = A$ and
$(\coperp{B})^{\perp} = B$.
\end{lemma}
\begin{proof}
Fix $x_{0} \in A$, $\phi_{0} \in A^{\perp}$, $\psi_{0} \in B$,
and $y_{0} \in \coperp{B}$. Let $L \mathrel{\coloneqq} A - x_{0}$ and $M
\mathrel{\coloneqq} B - \psi_{0}$. Write $x_{0}^{\perp}$ and
$\coperp{\psi_{0}}$ for the annihilator of the linear span of
$x_{0}$ and $\psi_{0}$, respectively. Observe that $A^{\perp}
= (L^{\perp} \cap x_{0}^{\perp}) + \phi_{0}$ and $\coperp{B} =
(\coperp{M} \cap \coperp{\psi_{0}}) + y_{0}$. The equations to
be proved now follow from the analogous facts about
annihilators mentioned above.
\end{proof}
\begin{definition}
A convex set $K \subset V$ is \emph{linearly bounded} if the
intersection of $K$ with every affine line is a bounded line
segment. Equivalently, a linearly bounded convex set contains
no ray.
\end{definition}
The following lemma establishes a natural relationship between the
non\-empty linearly bounded sections of $ C $ and the dual cone $
C^{*} $. In particular, the dual cone of $ C$ is the set of
elements of $ \dual V $ that are positive on a linearly bounded
section of $ C $.
\begin{lemma}
\label{lem:LinearBoundedSectionGeneratesCone}
Let $C$ be a pointed cone in a normed vector space $V$, and let
$\phi \in \dual V$ be such that $S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$
is nonempty and linearly bounded. Then $\dualc C = \braces{\psi
\in \dual V \,:\, \psi(S_{\phi}) \ge 0}$.
\end{lemma}
\begin{proof}
We first show that $\phi(C \setminus \braces{0}) > 0$. For,
suppose otherwise. Then, since $C$ is pointed, there exists a
$v \in C \setminus \braces{0}$ such that $\phi(v) = 0$. Fix $w
\in S_{\phi}$. Then, for every $\lambda \ge 0$, we have that
$w + \lambda v \in C$ and $\phi(w + \lambda v) = 1$, so $w +
\lambda v \in S_{\phi}$, which implies that $S_{\phi}$ is not
linearly bounded.
Now, suppose that $\psi(S_{\phi}) \ge 0$, and let $y \in C$ be
given. If $y = 0$, then we immediately have that $\psi(y) \ge
0$. If $y \in C \setminus \braces{0}$, then $\phi(y) > 0$, so
$\frac{1}{\phi(y)}y \in S_{\phi}$. Thus,
$\psi(\frac{1}{\phi(y)}y) \ge 0$, so $\psi(y) \ge 0$.
Therefore, $\psi \in \dualc C$, as claimed. Since the converse
claim is immediate, the lemma is proved.
\end{proof}
Furthermore, there is a well-known relationship between the
bounded sections of $ C $ and the interior points of $\dualc C$.
(Cf.~\cite[Theorem 3.8.4]{Jameson}.) Given a point $v$ in $V$ or
$\dual V$, we write $B_{r}(v)$ to denote the closed ball of radius
$r$ centered at $v$.
\begin{lemma}
\label{lem:InteriorNormalsGiveBoundedSections}
Let $C$ be a pointed cone in a normed vector space $V$. Given
a nonzero functional $\phi \in \dual V$, the section $S_{\phi}
\mathrel{\coloneqq} S_{\phi}(C)$ is bounded if and only if $\phi \in
\operatorname{int}(\dualc C)$. \textup{(}More precisely, given $ r > 0$,
we have that $S_{\phi} \subset B_{r}(0)$ if and only if
$B_{1/r}(\phi) \subset \dualc C$.\textup{)}
\end{lemma}
\begin{proof}
Suppose that $S_{\phi}$ is bounded. Let $r > 0$ be such that
$S_{\phi} \subset B_{r}(0)$, and put $\epsilon \mathrel{\coloneqq} 1/r$.
Let $\psi \in B_{\epsilon}(\phi)$ be given. Observe that, for
all $x \in S_{\phi}$, we have that $\abs{1 - \psi(x)} =
\abs{\phi(x) - \psi(x)} \le \norm{\phi - \psi} \norm{x} \le
\epsilon r = 1$, and hence $\psi(x) \ge 0$. Since $S_{\phi}$
is linearly bounded, it follows from Lemma
\ref{lem:LinearBoundedSectionGeneratesCone} that $\psi \in
\dualc C$, so $\phi \in \operatorname{int}(\dualc C)$.
Conversely, suppose that $\phi \in \operatorname{int}(\dualc C)$. Let
$\epsilon > 0$ be such that $B_{\epsilon}(\phi) \subset \dualc
C$, and put $r \mathrel{\coloneqq} 1/\epsilon$. Fix $x \in S_{\phi}$.
Since the unit ball $B_{1}(0)$ has a supporting hyperplane at
$x / \norm{x}$, there exists a functional $\nu \in \dual V$
such that $\nu(x / \norm{x}) = 1$ and $\nu(B_{1}(0)) \le 1$.
That is, $\nu(x) = \norm{x}$ and $\norm{\nu} = 1$. Set $\psi
\mathrel{\coloneqq} \phi - \epsilon \nu$. Then $\norm{\phi - \psi} =
\epsilon$, so $\psi \in \dualc C$, and hence $\psi(x) \ge 0$.
Thus, $\epsilon \norm{x} = \phi(x) - \psi(x) = 1 - \psi(x) \le
1$, so $\norm{x} \le 1/\epsilon = r$, yielding the claim.
\end{proof}
It is well known that if $ X $ is a separable Banach space, then
every pointed cone $ C $ has a base, i.e.~there is a closed
bounded convex subset $ B \subset X $ such that, for every $ x \in
C\setminus\braces{0} $, there exist unique $ \lambda > 0 $ and $ y
\in B $ such that $ x = \lambda y $. The separability assumption
is indispensable, as demonstrated by the standard example of the
cone $ C $ of all non-negative real-valued functions on $ X =
\ell^2 (I) $ with respect to the counting measure, where $I$ is an
uncountable set. For this case, the set of all positive
continuous linear functionals on $ C $ is isometrically isomorphic
to $ C $, but $ C^{*} $ has no strictly positive linear
functionals, and so $ \dualc{C} $ has no interior. It readily
follows that $ C $ has no base \cite{Jameson}. It is also worth
noting that pointed cones in separable Banach spaces may not have
a bounded base. For example, if $ X = C[0,1] $ with its usual
uniform topology, then $ X $ is separable. If $ C $ is the cone
of nonnegative valued functions in $ X $, then $C$ has a base,
but, since $ C^{*} $ is the set of regular positive Borel measures
on $ [0,1] $, $\dual C$ has an empty interior. Thus $ C $ has no
bounded base.
We do not assume that our normed vector space $V$ is separable, so
there may exist pointed cones without bounded bases. However,
condition (2) in the definition of CSS cones (Definition
\ref{defn:CSSCones}) guarantees that this is not the case with a
CSS cone, because having a bounded proper section implies having a
bounded base, as may be seen in the proof of Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}. In particular, the
dual of a CSS $C$ cone always has a nonempty interior. It follows
from a series of results due to Borwein and Lewis that the
functionals in the interior of $\dualc C$ are precisely the
functionals that are strictly positive on $C \setminus \braces{0}$.
\begin{lemma}
\label{lem:BoundedSectionsStrictlySupport}
Let $C$ be a closed pointed cone in a normed vector space $V$
such that the dual cone $\dual{C}$ has nonempty interior, and
let $\phi \in \dual V$. Then $S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$
is bounded if and only if $\phi$ is strictly positive on
$C\setminus\braces{0}$.
\end{lemma}
\begin{proof}
By Lemma \ref{lem:InteriorNormalsGiveBoundedSections},
$S_{\phi}$ is bounded if and only if $\phi \in
\operatorname{int}(\dual{C})$. Since $\operatorname{int}(\dual{C}) \ne \emptyset$, it
follows from \cite[Corollary 2.14]{BorLew1992} and
\cite[Theorem 3.10]{BorLew1992} that $\phi \in \operatorname{int}{\dual{C}}$
if and only if $\phi(C\setminus{0}) > 0$.
\end{proof}
\begin{lemma}
\label{lem:PerpOfBoundedSectionIsBounded}
Let $C$ be a pointed cone in a normed vector space $V$. If
$\dualc{S}$ is a bounded finite-dimensional section of $\dualc
C$ that intersects the interior of $\dualc C$, then the section
\begin{equation*}
S \mathrel{\coloneqq} \braces{x \in C \,:\, \text{for all $\phi \in
\dualc{S}$, $\phi(x) = 1$}}
\end{equation*}
of $C$ is bounded, finite codimensional, and intersects the
interior of $C$.
\end{lemma}
\begin{proof}
We first prove that $S \cap \operatorname{int}(C) \ne \emptyset$. Let $B =
\aff(\dualc S)$. Observe that $S = \coperp{B} \cap C$, and
suppose, to get a contradiction, that $\coperp{B} \cap \operatorname{int}(C)
= \emptyset$. Then there exists a supporting hyperplane $H$ of
$C$ containing $\coperp{B}$. Let $\psi \in \dualc C
\setminus\braces{0}$ be such that $H = \ker \psi$ and $\psi(C)
\ge 0$. Fix $\phi \in \dualc S$. Then $\phi + \lambda\psi \in
\dualc C$ for all $\lambda \ge 0$. Moreover, $\phi + \lambda
\psi \in (\coperp(B))^{\perp} = B$ by Lemma
\ref{lem:PerpIsADuality}. Hence, $\phi + \lambda\psi \in
\dualc{S}$ for all $\lambda \ge 0$, so $\dualc S$ is not
bounded, a contradiction.
Let $n \mathrel{\coloneqq} \dim(\duala S)$. Since $\duala S$ intersects
the interior of $\dualc C$, there exist linear functionals
$\phi_{0}, \dotsc, \phi_{n} \in \operatorname{int}(\duala S)$ that affinely
span $\aff(\duala S)$. It is easy to see that $S =
\bigcap_{i=0}^{n} S_{\phi_{i}}$. Indeed, since $S \cap
\operatorname{int}(C) \ne \emptyset$, the affine span of $S$ is the
intersection of the affine spans of the $S_{\phi_{i}}$:
\begin{equation*}
\aff(S) = \bigcap_{i=0}^{n} \aff(S_{\phi_{i}})
\end{equation*}
Therefore, $S$ is of co-dimension $\dim(\duala S) + 1$.
Finally, $S$ is bounded by
Lemma~\ref{lem:InteriorNormalsGiveBoundedSections}.
\end{proof}
By Lemma \ref{lem:InteriorNormalsGiveBoundedSections}, a bounded
proper section $S$ of a pointed cone $C$ corresponds to an
interior point of $\dualc C$. This correspondence interchanges
dimension and codimension, since it is the correspondence between
a hyperplane and a vector normal to that hyperplane. However, if
the bounded proper section $S$ is centrally symmetric, then there
is a canonical corresponding proper section of $\dualc C$, which
is also centrally symmetric.
\begin{lemma}
\label{lem:NormalOfCentSymmSectionIsCentOfSymm}
Let $C$ be a pointed cone in a normed vector space $V$.
Suppose that $x \in C$ and $\phi \in \dualc C$ are such that
$S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$ is bounded and centrally
symmetric about $x$. Then $S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$ is
centrally symmetric about $\phi$.
\end{lemma}
\begin{proof}
Let $\psi \in S_{x}$ be given. We want to show that $\phi -
(\psi - \phi) = 2 \phi - \psi \in S_{x}$. Since it is clear
that $(2 \phi - \psi)(x) = 1$, it remains only to show that $2
\phi - \psi \in \dualc C$. By Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}, it suffices to
show that $(2 \phi - \psi)(S_{\phi}) \ge 0$, or, equivalently,
$\psi(S_{\phi}) \le 2$. To this end, let $y \in S_{\phi}$ be
given. Since $S_{\phi}$ is centrally symmetric about $x$, we
have that $2x - y \in S_{\phi} \subset C$. Thus, $\psi(2x - y)
\ge 0$, or, equivalently, $\psi(y) \le 2$, as desired.
\end{proof}
\section{CSS cones are ellipsoidal cones}
\label{sec:CSSCones}
Fix a CSS cone $C$ with nonempty interior in a normed vector space
$V$. As mentioned in the introduction, the proof that $C$ is an
ellipsoidal cone in the finite-dimensional case appeared in
\cite{JerMcA2013}. Somewhat surprisingly, there does not seem to
be a straightforward transfinite-induction argument that extends
this result to the infinite-dimensional case. We will instead
take a detour through the dual cone $\dualc{C}$.
We call a section $C'$ of $C$ a \emph{sectional subcone} of $C$ if
$C'$ contains the origin. We will give an argument by finite
induction below showing that every finite-\emph{co}dimensional
sectional subcone of $C$ is also CSS. Had we been able to extend
this induction argument to a transfinite-induction argument, we
could have applied the finite-dimensional CSS characterization of
ellipsoidal cones to prove that every finite-dimensional sectional
subcone of $C$ is ellipsoidal. The conclusion that $C$ itself is
ellipsoidal would then have followed from the Jordan--von Neumann
characterization of inner-product spaces.
Unfortunately, a direct argument by transfinite induction for the
claim that all sectional subcones of CSS cones are CSS cones
eludes us. Our strategy instead will be as follows. To show that
$C$ is ellipsoidal, we will show that its dual $\dualc{C}$ is
ellipsoidal. That $\dualc{C}$ is ellipsoidal will follow from the
Jordan--von Neumann characterization of ellipsoids once we show
that every finite-dimensional sectional subcone of $\dualc{C}$ is
ellipsoidal. To prove this, we will need to show that every
bounded finite-dimensional section $\dualc{S}$ of $\dualc{C}$ is
centrally symmetric and then apply the finite-dimensional CSS
characterization of ellipsoidal cones.
Thus, we need to find a center of symmetry $\phi$ for a given
fi\-nite-di\-men\-sion\-al section $\dualc{S}$ of $\dualc C$. To
do this, we look at the perpendicular section $$S \mathrel{\coloneqq}
\braces{y \in C \,:\, \text{$\psi(y) = 1$ for all $\psi \in
\dualc{S}$}}$$ of $C$. It follows from Lemma
\ref{lem:PerpOfBoundedSectionIsBounded} that $S$ is a bounded
section of $C$ with finite codimension. Our finite induction
argument will thus suffice to show that $S$ is centrally
symmetric, with a center of symmetry $x$. Furthermore, $S$ is
contained in a proper section $T$ of $C$ with co-dimension $1$,
which will determine a dual vector $\phi \in \dualc{C}$ via
$\phi(T) = 1$. Finally, the central symmetry of $S$ will imply
that $\dualc{S}$ is centrally symmetric about $\phi$ by Lemma
\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, establishing the
result.
\begin{lemma}
\label{lem:FiniteCodimSectionsAreCentSymm}
Let $C$ be a CSS cone with nonempty interior in a normed vector
space $V$. Let $S$ be a bounded section of $C$ such that
$\operatorname{codim}(S) < \infty$ and $S \cap \operatorname{int}(C) \ne \emptyset$. Then
$S$ is centrally symmetric.
Indeed, $S$ is contained in a proper section $T$ of $C$ such
that the center of symmetry of $T$ lies on $S$.
\end{lemma}
\begin{proof}
We begin with the case where $\operatorname{codim}(S) = 2$. We will show
that there is a bounded proper section $T$ of $C$ containing
$S$ whose center of symmetry (which exists by the CSS property)
lies on $S$.
Fix $y \in S$, and let $L \mathrel{\coloneqq} \aff(S) - y$. The image of
$C$ under the quotient map $Q \colon V \to V/L$ is a
$2$-dimensional pointed cone in which $Q(S)$ contains a single
point. Put $C' \mathrel{\coloneqq} Q(C)$ and $\braces{s} \mathrel{\coloneqq} Q(S)$.
Since quotient maps are open, $s \in \operatorname{int}(C')$. Let $\ell
\subset V/L$ be the affine line through $s$ such that $s$ is
the midpoint of $\ell \cap C'$. Let $H \mathrel{\coloneqq} Q^{-1}(\ell)$,
and put $T \mathrel{\coloneqq} H \cap C$. We claim that $T$ is a bounded
proper section of $C$ containing $S$ whose center of symmetry
is on $S$.
It is clear that $S \subset T$. To see that $T$ is bounded,
observe that $H - y$ strictly supports $C$ at $0$. That is,
$(H-y) \cap C = \braces{0}$. For, suppose that $z \in (H-y)
\cap C$. Then, since $(\ell - s) \cap C' = \braces{0}$, we have
that $Q(z) = 0$, and so $z \in (S-y) \cap C$. If $z$ were
nonzero, then $\lambda z$ would also be in $(S-y) \cap C$ for
all $\lambda > 0$, so $S$ would be unbounded, contrary to our
hypothesis. Hence, $z = 0$. Thus, by Lemma
\ref{lem:BoundedSectionsStrictlySupport}, $T$ is bounded, so
$T$ has a center of symmetry $x$, which must map to the center
of symmetry of $\ell \cap C'$ under $Q$. That is, $Q(x) = s$,
so $x \in S$, as desired. Thus, $S$ is a section of a
centrally symmetric set that contains the center of symmetry of
that set. Therefore, $S$ itself is centrally symmetric.
If $n \mathrel{\coloneqq} \operatorname{codim}(S) \ge 3$, fix a codimension-1 linear
subspace $W \subset V$ containing $S$. By the preceding
argument, $C \cap W$ is a CSS cone in which $S$ is a bounded
codimension-$(n-1)$ section intersecting the interior of $C
\cap W$. The theorem now follows from the induction hypothesis
applied to $C \cap W$.
\end{proof}
\begin{lemma}
\label{lem:CSSImpliesDualIsCoCSS}
Let $C$ be a CSS cone with nonempty interior in a normed vector
space $V$. Let $S^{\ast}$ be a bounded section of $\dualc C$
such that $\dim(S^{\ast}) < \infty$ and $S^{\ast} \cap
\operatorname{int}(\dualc C) \ne \emptyset$. Then $S^{\ast}$ is centrally
symmetric.
\end{lemma}
\begin{proof}
Let $S \mathrel{\coloneqq} \braces{y \in C \,:\, \text{$\psi(y) = 1$ for all
$\psi \in S^{\ast}$}}$. By Lemma
\ref{lem:PerpOfBoundedSectionIsBounded}, we have that $S$ is
bounded, intersects the interior of $C$, and has finite
codimension. Thus, by Lemma
\ref{lem:FiniteCodimSectionsAreCentSymm}, there exists a proper
section $T$ of $C$ containing $S$ whose center of symmetry $x$
is in $S$. Let $\phi \in C^{*}$ be such that $S_{\phi}
\mathrel{\coloneqq} S_{\phi}(C) = T$. Then, by
Lemma~\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, $\phi$ is
the center of symmetry of $S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$.
Note that $S^{\ast} \subset S_{x}$, so it remains only to show
that $\phi \in S^{\ast}$.
Indeed, since $S \cap \operatorname{int} C \ne \emptyset$, we have that
$\aff(S) = \coperp(\aff \dualc S)$. Thus, $\aff(\dualc S) =
(\aff S)^{\perp}$ by Lemma \ref{lem:PerpIsADuality}. In
particular, $\phi \in \aff(S^{\ast})$. Since $\phi \in \dualc
C$, we conclude that $\phi \in S^{\ast}$, as desired.
\end{proof}
The previous Lemma motivates the following definition.
\begin{definition}
Let $C$ be a closed pointed cone in a normed vector space $V$
with nonempty interior. We call $C$ \emph{co-CSS} if every
bounded finite-dimensional section of $C$ intersecting the
interior of $C$ is centrally-symmetric.
\end{definition}
Thus, Lemma \ref{lem:CSSImpliesDualIsCoCSS} says that the dual of
a CSS cone is co-CSS.
\begin{lemma}
\label{lem:FiniteDimSubConesOfCoCSSAreEllipsoidal}
Let $C$ be a co-CSS cone in a normed vector space $V$. Fix a
finite-dimensional subspace $L \subset V$ such that $L \cap
\operatorname{int} C \ne \emptyset$. Then the cone $L \cap C$ is
ellipsoidal.
\end{lemma}
\begin{proof}
Since $C$ is pointed, the finite-dimensional cone $L \cap C$ is
also pointed. Since $C$ is co-CSS, every bounded proper
section of $L \cap C$ is centrally symmetric. In particular,
$L \cap C$ is CSS. Therefore, by the finite-dimensional CSS
characterization of ellipsoidal cones \cite[Theorem
1.4]{JerMcA2013}, $L \cap C$ is ellipsoidal.
\end{proof}
\begin{lemma}
\label{lem:DualsOfCSSConesAreEllipsoidal}
Let $C$ be a CSS cone in a normed vector space $V$. Then the
dual cone $\dualc C$ is ellipsoidal.
\end{lemma}
\begin{proof}
Since $C$ is CSS, there exists a bounded proper section
$S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$ of $C$, where, by Lemma
\ref{lem:InteriorNormalsGiveBoundedSections}, $\phi \in
\operatorname{int}(\dualc C)$. Let $x \in C$ be the center of symmetry of
$S_{\phi}$. Then, by Lemma
\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, the section
$S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$ is centrally symmetric about
$\phi$. In addition, $S_{x}$ is a proper section of $\dualc C$
because it contains the point $\phi \in \operatorname{int}(\dualc C)$.
Furthermore, $S_{x}$ is bounded by Lemma
\ref{lem:InteriorNormalsGiveBoundedSections} because $x \in
\operatorname{int}(C) \hookrightarrow \operatorname{int}(\ddualc C)$ under the canonical
embedding. Finally, by Lemma
\ref{lem:FiniteDimSubConesOfCoCSSAreEllipsoidal}, every
finite-dimensional section of $S_{x}$ through $\phi$ is an
ellipsoid centered at $\phi$. Hence, by the Jordan--von
Neumann characterization of inner-product spaces
\cite{JorvNeu1935}, $S_{x}$ is an ellipsoid. Therefore, $\dual
C$ is ellipsoidal.
\end{proof}
We are now ready to prove our main result.
\begin{proof}[Proof of Theorem \ref{thm:CSSTheorem}]
\label{proof:CSSproof}
Every bounded proper section of an ellipsoidal cone is an
ellipsoid, and ellipsoids are centrally symmetric. Hence,
ellipsoidal cones are CSS.
To prove the converse, let a CSS cone of dimension $\ge 2$ be
given. By Lemma \ref{lem:DualsOfCSSConesAreEllipsoidal},
$\dual{C}$ is ellipsoidal. Fix an ellipsoidal proper section
$\dual{S}$ of $\dual{C}$, and let $\phi \in \dual{S}$. Thus,
we have an inner product on the codimension-$1$ linear subspace
$M \mathrel{\coloneqq} \aff(\dual{S}) - \phi$. Since $\dual{S}$ is
bounded, $\aff(\duala{S})$ does not contain the origin. Hence,
we can complete the inner product on $M$ to an inner product on
all of $\dual{V}$. Indeed, since $\dual{V}$ is already a dual
space, it is in fact a Hilbert space. The dual of an
ellipsoidal cone in a Hilbert space is ellipsoidal
\cite[p.~51]{Krein}, so $\ddual{C}$ is ellipsoidal. Since
$\ddual{C}$ is the closure of $C$ under the canonical embedding
$V \hookrightarrow \ddual{V}$, we conclude that $C$ itself is
ellipsoidal in $V$.
\end{proof}
It follows that if a normed vector space $V$ contains a CSS cone
with nonempty interior, then $V$ is an inner product space. In
particular, if a Banach space $X$ contains a full-dimensional CSS
cone, then $X$ is a Hilbert space.
\section{FBI cones are ellipsoidal cones}
\label{sec:FBICones}
We conclude by proving that every cone in a normed vector space
$V$ that satisfies the FBI property (Definition \ref{def:FBI}) is
ellipsoidal.
\begin{lemma}
\label{lem:ConeIntervalsAreBounded}
Let $C$ be a cone in a normed vector space $V$ such that
$\operatorname{int}(C)$ and $\operatorname{int}(\dualc C)$ are both nonempty. Then, for
each $a \in \operatorname{int}(C)$, the intersection $C \cap (a - C)$ is
bounded.
\end{lemma}
\begin{proof}
Since $\operatorname{int}(\dualc{C}) \ne \emptyset$, there exists a
functional $\phi \in \dualc C$ such that $S_{\phi} \mathrel{\coloneqq}
S_{\phi}(C)$ is a bounded base of $C$ by Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}. In particular,
$\ker(\phi)$ strictly supports $C$ at $0$ by Lemma
\ref{lem:BoundedSectionsStrictlySupport}. By suitably
normalizing $\phi$, we also have that $S_{\phi}$ strictly
supports $a - C$ at $a$. Thus,
\begin{equation*}
C \cap (a - C) \subset \braces{y \in C \,:\, \phi(y) \le 1}.
\end{equation*}
Since $S_{\phi}$ is bounded and is a base for $C$, it follows
that $C \cap (a - C)$ is also bounded.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:FBIimpliesEllipsoidal}]
\label{proof:FBIimpliesEllipsoidal}%
Without loss of generality, suppose that $\operatorname{int}{C} \ne
\emptyset$. Fix $x \in \operatorname{int}(C)$, and set $S \mathrel{\coloneqq}
\conv(\partial C \cap \partial(2x - C))$. By the FBI property and
Lemma \ref{lem:ConeIntervalsAreBounded}, $S$ is a bounded
section of $C$. Moreover, $S$ is centrally symmetric about
$x$. We show that every $2$-dimensional section of $S$ through
$x$ is an ellipse. Let $E$ be such a section. Observe that $0
\notin \aff E$, because $0 \in \aff E \subset \aff S$ would
imply that $\aff S = \lin S$, contrary to the boundedness of
$S$. Therefore, $C' \mathrel{\coloneqq} \cone(E)$ is a $3$-dimensional
cone. We claim that $\cone(E)$ is an FBI cone. To prove this,
let $y \in \relintr C'$, and let $\Gamma \mathrel{\coloneqq} \partial(C')
\cap \partial(y - C')$. On the one hand, $\Gamma$ is contained in
$\lin E$. On the other hand, let $H$ be the affine hyperplane
in $V$ containing $\partial C \cap \partial(y - C)$. As above, $0
\notin H$. Since $0 \in \lin E$, it follows that $\Gamma
\subset (\lin E) \cap H$, but $\lin E \not\subset H$. Hence,
$\Gamma$ is contained in a $2$-dimensional affine subspace of
$\lin E$. That is, $C'$ is an FBI cone. It follows from
\cite[Theorem 1.2]{JerMcA2013} that $E$ is an ellipse.
Therefore, by the Jordan--von Neumann characterization of
inner-product spaces \cite{JorvNeu1935}, $S$ is an ellipsoid.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
\section{Introduction}
In a landmark paper \cite{Rudin}, W.~Rudin and K.~T.~Smith
answered a question of J.~Korevaar by showing that, if $ X $ is a
strictly convex real Banach space of dimension $ n \neq 2 $ and,
for each finite-dimensional subspace $ \pi $ in $ X $, the best
approximation function $ P_\pi $ is linear, then $X $ is a Hilbert
space. Their theorem led to the following characterization of
ellipsoids: {\it If $K $ is a centrally symmetric compact convex
body in $n$-space (where $n$ is possibly infinite) such that, for
every $ \nu $-dimensional subspace $ \pi $ with $ 0 < \nu < n-1 $,
the union of the tangency sets of all support planes of $K $ which
are translates of $ \pi $ lies in a plane of dimension $n - \nu$,
then $ K $ is an ellipsoid.} This result is a vivid example of how
the characterization of ellipsoids is intimately tied to
characterizing the Banach spaces that are Hilbert spaces.
In \cite{JerMcA2013}, the second author with J.~Jer\'onimo showed
that a fi\-nite-di\-men\-sion\-al pointed cone in which every
bounded section has a center of symmetry is an ellipsoidal cone.
Hence these are exactly the cones over closed unit balls of
Hilbert spaces that have been translated away from the origin.
The primary goal of this paper is to generalize this result to
infinite-dimensional cones. While this may seem like a result
that would follow from a straightforward induction argument (at
least for a separable Banach space), such an approach is elusive.
We give an affirmative answer to the infinite-dimensional
generalization by carefully using the following fact: \emph{Cones
with bounded sections have dual cones with nonempty interiors}.
Fix a real normed vector space $V$. A \emph{cone} in $V$ is a
nonempty convex subset $C\subset V$ that is closed under
nonnegative scalar multiplication. The cone $C$ is \emph{pointed}
if it contains no line through the origin; that is, if $C \cap
(-C) = \braces{0}$.
Given a convex subset $K \subset V$, let $\aff(K)$ denote the
affine span of $K$, and let $\lin(K)$ be the linear span of $K$.
We write $\operatorname{int}(K)$ for the interior of $K$ in $V$ with respect to
the norm on $V$. The \emph{relative interior} $\relintr{K}$ and
the \emph{relative boundary} $\partial K$ are the interior and
boundary, respectively, of $K$ with respect to $\aff(K)$. The
\emph{cone over} $K$, denoted $\cone(K)$, is the intersection of
all cones containing $K$. A \emph{section} of $K$ is the nonempty
intersection of $K$ with a closed affine subspace of $V$. We call
a section $S$ of $K$ \emph{proper} if $S$ has codimension $1$ with
respect to $\aff(K)$.
While the concept of an ellipsoid in finite dimensions is well
known, in infinite dimensional vector spaces this requires careful
definition. We define a subset $E \subset V$ to be an
\emph{ellipsoid} if, for some $x \in \aff(E)$, there exists an
inner product on the linear space $\aff(E) - x$ such that $E - x$
is the closed unit ball corresponding to this inner product. An
\emph{ellipse} is a $2$-dimensional ellipsoid. A cone $C \subset
V$ is \emph{ellipsoidal} if some proper section of $C$ is an
ellipsoid.
A subset $S \subset V$ is \emph{centrally symmetric} if there
exists a point $x \in V$ such that $S - x = -(S - x)$. In this
case, $x$ is a \emph{center of symmetry} of $S$, and $x$ is the
unique center of symmetry of $S$ if $S$ is nonempty and bounded.
\begin{definition}[CSS Cones]
\label{defn:CSSCones}
Let $C \subset V$ be a cone. We say that $C$
satisfies the \emph{centrally symmetric sections
\textup{(}CSS\textup{)} property} if
\begin{enumerate}
\item
$C$ is closed and $\relintr{C} \ne \emptyset$,
\item
there exists a bounded proper section of $C$, and
\item
every bounded proper section of $C$ is centrally symmetric.
\end{enumerate}
We call a cone with the CSS property a \emph{CSS cone}.
\end{definition}
Our main result is that the CSS cones in $V$ are precisely the
ellipsoidal cones.
\begin{thm}[proved on p.\ \pageref{proof:CSSproof}]
\label{thm:CSSTheorem}
Let $C$ be a cone in a normed vector space $V$. Then $C$ is a
CSS cone if and only if $C$ is an ellipsoidal cone.
\end{thm}
The proof that finite-dimensional CSS cones are ellipsoidal
appeared in~\cite{JerMcA2013}. That article also established
another characterization of fi\-nite-di\-men\-sion\-al ellipsoidal
cones: Such a cone is ellipsoidal if and only if it is a so-called
\emph{FBI cone}.
\begin{definition}
\label{def:FBI}
Let $C \subset V$ be a cone. We say that~$C$ satisfies the
\emph{flat boundary intersections \textup{(}FBI\textup{)}
property} if
\begin{enumerate}
\item
$C$ is closed and $\relintr{C} \ne \emptyset$,
\item
there exists a bounded proper section of $C$, and
\item
for each $a \in \relintr{C}$, some proper section of $C$
contains $\partial C \cap \partial(a - C)$.
\end{enumerate}
We call a cone with the FBI property an \emph{FBI cone}. (See
Figure~\ref{fig:FBICone}.)
\end{definition}
\begin{figure}
\begin{center}
\includegraphics{FBI_cone.pdf}
\end{center}
\caption{An FBI cone $C$ and a translation of $-C$. The shaded
region is the convex hull of the intersection of the
boundaries. The FBI property implies that this convex hull is
contained in a hyperplane.}
\label{fig:FBICone}
\end{figure}
\begin{thm}[proved on p.~\pageref{proof:FBIimpliesEllipsoidal}]
\label{thm:FBIimpliesEllipsoidal}
Let $C$ be a cone in a normed vector space $V$. Then $C$ is a
FBI cone if and only if $C$ is an ellipsoidal cone.
\end{thm}
Unlike the proof of the CSS characterization of ellipsoidal cones,
the proof of the FBI characterization carries over with very
little change to the infinite-dimensional case. We give this
proof in Section \ref{sec:FBICones}. In \cite{JerMcA2013}, the
proof that CSS cones in $\mathbb{R}^{n}$ are ellipsoidal proceeded by
showing that CSS cones are FBI cones, so the proof that FBI cones
are ellipsoidal was key to the argument. Unfortunately, the proof
in \cite{JerMcA2013} that CSS cones are FBI cones relied on the
existence of a measure, so a different strategy is needed to prove
that CSS cones are ellipsoidal in infinite-dimensional normed
vector spaces.
As a corollary of Theorems \ref{thm:CSSTheorem} and
\ref{thm:FBIimpliesEllipsoidal}, we get two characterizations of
those normed vector spaces that are inner-product spaces.
\begin{corollary}
Let $V$ be a normed vector space. Then the following are
equivalent.
\begin{enumerate}
\item
$V$ is an inner product space.
\item
$V$ contains a CSS cone with nonempty interior.
\item
$V$ contains an FBI cone with nonempty interior.
\end{enumerate}
In particular, if a Banach space $X$ contains a CSS cone or an
FBI cone, then $X$ is a Hilbert space.
\end{corollary}
\section{Cone lemmas}
\label{sec:ConeLemmas}
In the sections that follow, we will require some lemmas regarding
cones, their duals, their sections, and the relationships between
these concepts and central symmetry. Most of these lemmas are
well known, though the property of central symmetry seems to be
little-studied in the context of infinite dimensional sections of
cones. The seminal monograph of Kre\u{\i}n and Rutman \cite{Krein}
still provides an excellent introduction to cones in linear
spaces. A more recent treatment may be found
in~\cite{AliTou2007}.
Let $V$ be a real normed vector space with norm $\norm{\cdot}$,
and let $V^{*}$ be the dual space of continuous linear functionals
on $V$ under the operator norm, also denoted by $\norm{\cdot}$.
Let a cone $C \subset V$ be given. Recall that the dual of $C$ is
the cone $\dualc{C} \subset \dual V$ defined by
\begin{equation*}
\dualc{C} \mathrel{\coloneqq} \braces{\phi \in \dual V \,:\, \text{$\phi(x)
\ge 0$ for all $x \in C$}}.
\end{equation*}
Recall also the following well-known result regarding cones in
normed vector spaces.
\begin{proposition}
Let $C$ be a cone in a normed vector space $V$. Then, under
the canonical embedding $V \hookrightarrow \ddual V$, we have
that $\operatorname{int}(C) \hookrightarrow \operatorname{int}(\ddualc C)$. Moreover,
the closure of $C$ under this embedding is $\ddualc{C}$.
\end{proposition}
We will find the following notation to be convenient: Given points
$x \in V$ and $\phi \in \dual V$, let $ S_x(\dualc C) $ and $
S_\phi(C) $ be sections of $ C^{*} $ and $ C $, respectively,
defined as follows
\begin{align*}
S_{x}(\dualc C) & \mathrel{\coloneqq} \braces{\psi \in \dualc C \,:\, \psi(x)
= 1}, \\
S_{\phi}(C) & \mathrel{\coloneqq} \braces{y \in C \,:\, \phi(y) = 1}.
\end{align*}
More generally, we will occasionally need a canonical affine
subspace in $V$ or $\dual V$ that is perpendicular to a given
affine subspace of $\dual V$ or $V$, respectively. Recall that,
if $L \subset V$ and $M \subset \dual{V}$ are linear subspaces,
then the \emph{annihilators} of $L$ and $M$ are defined by
\begin{align*}
L^{\perp} & \mathrel{\coloneqq} \braces{\phi \in \dual{V} \,:\,
\text{$\phi(x) = 0$ for all $x \in L$}},\\
\coperp{M} & \mathrel{\coloneqq} \braces{y \in V \,:\, \text{$\psi(y) = 0$
for all $\psi \in M$}}.
\end{align*}
Analogously, given an affine subspace $A \subset V$, respectively
$B \subset \dual{V}$, bounded away from the origin, define the
\emph{perpendicular affine spaces}
\begin{align*}
A^{\perp} & \mathrel{\coloneqq} \braces{\phi \in \dual{V} \,:\,
\text{$\phi(x) = 1$ for all $x \in A$}},\\
\coperp{B} & \mathrel{\coloneqq} \braces{y \in V \,:\, \text{$\psi(y) = 1$
for all $\psi \in B$}}.
\end{align*}
It is well known that, if $L \subset V$ is a closed linear
subspace and if $M \subset \dual{V}$ is a $\text{weak}^{\ast}$-closed linear
subspace, then the annihilators satisfy a duality relation:
$\coperp{(L^{\perp})} = L$ and $(\coperp{M})^{\perp} = M$.
\begin{lemma}
\label{lem:PerpIsADuality}
Let $V$ be a normed vector space. Let $A \subset V$ and $B
\subset \dual{V}$ be closed affine subspaces such that $0
\notin A$, $0 \notin B$, and $\operatorname{codim}(A)$ and $\dim(B)$ are
finite. Then $\coperp{(A^{\perp})} = A$ and
$(\coperp{B})^{\perp} = B$.
\end{lemma}
\begin{proof}
Fix $x_{0} \in A$, $\phi_{0} \in A^{\perp}$, $\psi_{0} \in B$,
and $y_{0} \in \coperp{B}$. Let $L \mathrel{\coloneqq} A - x_{0}$ and $M
\mathrel{\coloneqq} B - \psi_{0}$. Write $x_{0}^{\perp}$ and
$\coperp{\psi_{0}}$ for the annihilator of the linear span of
$x_{0}$ and $\psi_{0}$, respectively. Observe that $A^{\perp}
= (L^{\perp} \cap x_{0}^{\perp}) + \phi_{0}$ and $\coperp{B} =
(\coperp{M} \cap \coperp{\psi_{0}}) + y_{0}$. The equations to
be proved now follow from the analogous facts about
annihilators mentioned above.
\end{proof}
\begin{definition}
A convex set $K \subset V$ is \emph{linearly bounded} if the
intersection of $K$ with every affine line is a bounded line
segment. Equivalently, a linearly bounded convex set contains
no ray.
\end{definition}
The following lemma establishes a natural relationship between the
non\-empty linearly bounded sections of $ C $ and the dual cone $
C^{*} $. In particular, the dual cone of $ C$ is the set of
elements of $ \dual V $ that are positive on a linearly bounded
section of $ C $.
\begin{lemma}
\label{lem:LinearBoundedSectionGeneratesCone}
Let $C$ be a pointed cone in a normed vector space $V$, and let
$\phi \in \dual V$ be such that $S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$
is nonempty and linearly bounded. Then $\dualc C = \braces{\psi
\in \dual V \,:\, \psi(S_{\phi}) \ge 0}$.
\end{lemma}
\begin{proof}
We first show that $\phi(C \setminus \braces{0}) > 0$. For,
suppose otherwise. Then, since $C$ is pointed, there exists a
$v \in C \setminus \braces{0}$ such that $\phi(v) = 0$. Fix $w
\in S_{\phi}$. Then, for every $\lambda \ge 0$, we have that
$w + \lambda v \in C$ and $\phi(w + \lambda v) = 1$, so $w +
\lambda v \in S_{\phi}$, which implies that $S_{\phi}$ is not
linearly bounded.
Now, suppose that $\psi(S_{\phi}) \ge 0$, and let $y \in C$ be
given. If $y = 0$, then we immediately have that $\psi(y) \ge
0$. If $y \in C \setminus \braces{0}$, then $\phi(y) > 0$, so
$\frac{1}{\phi(y)}y \in S_{\phi}$. Thus,
$\psi(\frac{1}{\phi(y)}y) \ge 0$, so $\psi(y) \ge 0$.
Therefore, $\psi \in \dualc C$, as claimed. Since the converse
claim is immediate, the lemma is proved.
\end{proof}
Furthermore, there is a well-known relationship between the
bounded sections of $ C $ and the interior points of $\dualc C$.
(Cf.~\cite[Theorem 3.8.4]{Jameson}.) Given a point $v$ in $V$ or
$\dual V$, we write $B_{r}(v)$ to denote the closed ball of radius
$r$ centered at $v$.
\begin{lemma}
\label{lem:InteriorNormalsGiveBoundedSections}
Let $C$ be a pointed cone in a normed vector space $V$. Given
a nonzero functional $\phi \in \dual V$, the section $S_{\phi}
\mathrel{\coloneqq} S_{\phi}(C)$ is bounded if and only if $\phi \in
\operatorname{int}(\dualc C)$. \textup{(}More precisely, given $ r > 0$,
we have that $S_{\phi} \subset B_{r}(0)$ if and only if
$B_{1/r}(\phi) \subset \dualc C$.\textup{)}
\end{lemma}
\begin{proof}
Suppose that $S_{\phi}$ is bounded. Let $r > 0$ be such that
$S_{\phi} \subset B_{r}(0)$, and put $\epsilon \mathrel{\coloneqq} 1/r$.
Let $\psi \in B_{\epsilon}(\phi)$ be given. Observe that, for
all $x \in S_{\phi}$, we have that $\abs{1 - \psi(x)} =
\abs{\phi(x) - \psi(x)} \le \norm{\phi - \psi} \norm{x} \le
\epsilon r = 1$, and hence $\psi(x) \ge 0$. Since $S_{\phi}$
is linearly bounded, it follows from Lemma
\ref{lem:LinearBoundedSectionGeneratesCone} that $\psi \in
\dualc C$, so $\phi \in \operatorname{int}(\dualc C)$.
Conversely, suppose that $\phi \in \operatorname{int}(\dualc C)$. Let
$\epsilon > 0$ be such that $B_{\epsilon}(\phi) \subset \dualc
C$, and put $r \mathrel{\coloneqq} 1/\epsilon$. Fix $x \in S_{\phi}$.
Since the unit ball $B_{1}(0)$ has a supporting hyperplane at
$x / \norm{x}$, there exists a functional $\nu \in \dual V$
such that $\nu(x / \norm{x}) = 1$ and $\nu(B_{1}(0)) \le 1$.
That is, $\nu(x) = \norm{x}$ and $\norm{\nu} = 1$. Set $\psi
\mathrel{\coloneqq} \phi - \epsilon \nu$. Then $\norm{\phi - \psi} =
\epsilon$, so $\psi \in \dualc C$, and hence $\psi(x) \ge 0$.
Thus, $\epsilon \norm{x} = \phi(x) - \psi(x) = 1 - \psi(x) \le
1$, so $\norm{x} \le 1/\epsilon = r$, yielding the claim.
\end{proof}
It is well known that if $ X $ is a separable Banach space, then
every pointed cone $ C $ has a base, i.e.~there is a closed
bounded convex subset $ B \subset X $ such that, for every $ x \in
C\setminus\braces{0} $, there exist unique $ \lambda > 0 $ and $ y
\in B $ such that $ x = \lambda y $. The separability assumption
is indispensable, as demonstrated by the standard example of the
cone $ C $ of all non-negative real-valued functions on $ X =
\ell^2 (I) $ with respect to the counting measure, where $I$ is an
uncountable set. For this case, the set of all positive
continuous linear functionals on $ C $ is isometrically isomorphic
to $ C $, but $ C^{*} $ has no strictly positive linear
functionals, and so $ \dualc{C} $ has no interior. It readily
follows that $ C $ has no base \cite{Jameson}. It is also worth
noting that pointed cones in separable Banach spaces may not have
a bounded base. For example, if $ X = C[0,1] $ with its usual
uniform topology, then $ X $ is separable. If $ C $ is the cone
of nonnegative valued functions in $ X $, then $C$ has a base,
but, since $ C^{*} $ is the set of regular positive Borel measures
on $ [0,1] $, $\dual C$ has an empty interior. Thus $ C $ has no
bounded base.
We do not assume that our normed vector space $V$ is separable, so
there may exist pointed cones without bounded bases. However,
condition (2) in the definition of CSS cones (Definition
\ref{defn:CSSCones}) guarantees that this is not the case with a
CSS cone, because having a bounded proper section implies having a
bounded base, as may be seen in the proof of Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}. In particular, the
dual of a CSS $C$ cone always has a nonempty interior. It follows
from a series of results due to Borwein and Lewis that the
functionals in the interior of $\dualc C$ are precisely the
functionals that are strictly positive on $C \setminus \braces{0}$.
\begin{lemma}
\label{lem:BoundedSectionsStrictlySupport}
Let $C$ be a closed pointed cone in a normed vector space $V$
such that the dual cone $\dual{C}$ has nonempty interior, and
let $\phi \in \dual V$. Then $S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$
is bounded if and only if $\phi$ is strictly positive on
$C\setminus\braces{0}$.
\end{lemma}
\begin{proof}
By Lemma \ref{lem:InteriorNormalsGiveBoundedSections},
$S_{\phi}$ is bounded if and only if $\phi \in
\operatorname{int}(\dual{C})$. Since $\operatorname{int}(\dual{C}) \ne \emptyset$, it
follows from \cite[Corollary 2.14]{BorLew1992} and
\cite[Theorem 3.10]{BorLew1992} that $\phi \in \operatorname{int}{\dual{C}}$
if and only if $\phi(C\setminus{0}) > 0$.
\end{proof}
\begin{lemma}
\label{lem:PerpOfBoundedSectionIsBounded}
Let $C$ be a pointed cone in a normed vector space $V$. If
$\dualc{S}$ is a bounded finite-dimensional section of $\dualc
C$ that intersects the interior of $\dualc C$, then the section
\begin{equation*}
S \mathrel{\coloneqq} \braces{x \in C \,:\, \text{for all $\phi \in
\dualc{S}$, $\phi(x) = 1$}}
\end{equation*}
of $C$ is bounded, finite codimensional, and intersects the
interior of $C$.
\end{lemma}
\begin{proof}
We first prove that $S \cap \operatorname{int}(C) \ne \emptyset$. Let $B =
\aff(\dualc S)$. Observe that $S = \coperp{B} \cap C$, and
suppose, to get a contradiction, that $\coperp{B} \cap \operatorname{int}(C)
= \emptyset$. Then there exists a supporting hyperplane $H$ of
$C$ containing $\coperp{B}$. Let $\psi \in \dualc C
\setminus\braces{0}$ be such that $H = \ker \psi$ and $\psi(C)
\ge 0$. Fix $\phi \in \dualc S$. Then $\phi + \lambda\psi \in
\dualc C$ for all $\lambda \ge 0$. Moreover, $\phi + \lambda
\psi \in (\coperp(B))^{\perp} = B$ by Lemma
\ref{lem:PerpIsADuality}. Hence, $\phi + \lambda\psi \in
\dualc{S}$ for all $\lambda \ge 0$, so $\dualc S$ is not
bounded, a contradiction.
Let $n \mathrel{\coloneqq} \dim(\duala S)$. Since $\duala S$ intersects
the interior of $\dualc C$, there exist linear functionals
$\phi_{0}, \dotsc, \phi_{n} \in \operatorname{int}(\duala S)$ that affinely
span $\aff(\duala S)$. It is easy to see that $S =
\bigcap_{i=0}^{n} S_{\phi_{i}}$. Indeed, since $S \cap
\operatorname{int}(C) \ne \emptyset$, the affine span of $S$ is the
intersection of the affine spans of the $S_{\phi_{i}}$:
\begin{equation*}
\aff(S) = \bigcap_{i=0}^{n} \aff(S_{\phi_{i}})
\end{equation*}
Therefore, $S$ is of co-dimension $\dim(\duala S) + 1$.
Finally, $S$ is bounded by
Lemma~\ref{lem:InteriorNormalsGiveBoundedSections}.
\end{proof}
By Lemma \ref{lem:InteriorNormalsGiveBoundedSections}, a bounded
proper section $S$ of a pointed cone $C$ corresponds to an
interior point of $\dualc C$. This correspondence interchanges
dimension and codimension, since it is the correspondence between
a hyperplane and a vector normal to that hyperplane. However, if
the bounded proper section $S$ is centrally symmetric, then there
is a canonical corresponding proper section of $\dualc C$, which
is also centrally symmetric.
\begin{lemma}
\label{lem:NormalOfCentSymmSectionIsCentOfSymm}
Let $C$ be a pointed cone in a normed vector space $V$.
Suppose that $x \in C$ and $\phi \in \dualc C$ are such that
$S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$ is bounded and centrally
symmetric about $x$. Then $S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$ is
centrally symmetric about $\phi$.
\end{lemma}
\begin{proof}
Let $\psi \in S_{x}$ be given. We want to show that $\phi -
(\psi - \phi) = 2 \phi - \psi \in S_{x}$. Since it is clear
that $(2 \phi - \psi)(x) = 1$, it remains only to show that $2
\phi - \psi \in \dualc C$. By Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}, it suffices to
show that $(2 \phi - \psi)(S_{\phi}) \ge 0$, or, equivalently,
$\psi(S_{\phi}) \le 2$. To this end, let $y \in S_{\phi}$ be
given. Since $S_{\phi}$ is centrally symmetric about $x$, we
have that $2x - y \in S_{\phi} \subset C$. Thus, $\psi(2x - y)
\ge 0$, or, equivalently, $\psi(y) \le 2$, as desired.
\end{proof}
\section{CSS cones are ellipsoidal cones}
\label{sec:CSSCones}
Fix a CSS cone $C$ with nonempty interior in a normed vector space
$V$. As mentioned in the introduction, the proof that $C$ is an
ellipsoidal cone in the finite-dimensional case appeared in
\cite{JerMcA2013}. Somewhat surprisingly, there does not seem to
be a straightforward transfinite-induction argument that extends
this result to the infinite-dimensional case. We will instead
take a detour through the dual cone $\dualc{C}$.
We call a section $C'$ of $C$ a \emph{sectional subcone} of $C$ if
$C'$ contains the origin. We will give an argument by finite
induction below showing that every finite-\emph{co}dimensional
sectional subcone of $C$ is also CSS. Had we been able to extend
this induction argument to a transfinite-induction argument, we
could have applied the finite-dimensional CSS characterization of
ellipsoidal cones to prove that every finite-dimensional sectional
subcone of $C$ is ellipsoidal. The conclusion that $C$ itself is
ellipsoidal would then have followed from the Jordan--von Neumann
characterization of inner-product spaces.
Unfortunately, a direct argument by transfinite induction for the
claim that all sectional subcones of CSS cones are CSS cones
eludes us. Our strategy instead will be as follows. To show that
$C$ is ellipsoidal, we will show that its dual $\dualc{C}$ is
ellipsoidal. That $\dualc{C}$ is ellipsoidal will follow from the
Jordan--von Neumann characterization of ellipsoids once we show
that every finite-dimensional sectional subcone of $\dualc{C}$ is
ellipsoidal. To prove this, we will need to show that every
bounded finite-dimensional section $\dualc{S}$ of $\dualc{C}$ is
centrally symmetric and then apply the finite-dimensional CSS
characterization of ellipsoidal cones.
Thus, we need to find a center of symmetry $\phi$ for a given
fi\-nite-di\-men\-sion\-al section $\dualc{S}$ of $\dualc C$. To
do this, we look at the perpendicular section $$S \mathrel{\coloneqq}
\braces{y \in C \,:\, \text{$\psi(y) = 1$ for all $\psi \in
\dualc{S}$}}$$ of $C$. It follows from Lemma
\ref{lem:PerpOfBoundedSectionIsBounded} that $S$ is a bounded
section of $C$ with finite codimension. Our finite induction
argument will thus suffice to show that $S$ is centrally
symmetric, with a center of symmetry $x$. Furthermore, $S$ is
contained in a proper section $T$ of $C$ with co-dimension $1$,
which will determine a dual vector $\phi \in \dualc{C}$ via
$\phi(T) = 1$. Finally, the central symmetry of $S$ will imply
that $\dualc{S}$ is centrally symmetric about $\phi$ by Lemma
\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, establishing the
result.
\begin{lemma}
\label{lem:FiniteCodimSectionsAreCentSymm}
Let $C$ be a CSS cone with nonempty interior in a normed vector
space $V$. Let $S$ be a bounded section of $C$ such that
$\operatorname{codim}(S) < \infty$ and $S \cap \operatorname{int}(C) \ne \emptyset$. Then
$S$ is centrally symmetric.
Indeed, $S$ is contained in a proper section $T$ of $C$ such
that the center of symmetry of $T$ lies on $S$.
\end{lemma}
\begin{proof}
We begin with the case where $\operatorname{codim}(S) = 2$. We will show
that there is a bounded proper section $T$ of $C$ containing
$S$ whose center of symmetry (which exists by the CSS property)
lies on $S$.
Fix $y \in S$, and let $L \mathrel{\coloneqq} \aff(S) - y$. The image of
$C$ under the quotient map $Q \colon V \to V/L$ is a
$2$-dimensional pointed cone in which $Q(S)$ contains a single
point. Put $C' \mathrel{\coloneqq} Q(C)$ and $\braces{s} \mathrel{\coloneqq} Q(S)$.
Since quotient maps are open, $s \in \operatorname{int}(C')$. Let $\ell
\subset V/L$ be the affine line through $s$ such that $s$ is
the midpoint of $\ell \cap C'$. Let $H \mathrel{\coloneqq} Q^{-1}(\ell)$,
and put $T \mathrel{\coloneqq} H \cap C$. We claim that $T$ is a bounded
proper section of $C$ containing $S$ whose center of symmetry
is on $S$.
It is clear that $S \subset T$. To see that $T$ is bounded,
observe that $H - y$ strictly supports $C$ at $0$. That is,
$(H-y) \cap C = \braces{0}$. For, suppose that $z \in (H-y)
\cap C$. Then, since $(\ell - s) \cap C' = \braces{0}$, we have
that $Q(z) = 0$, and so $z \in (S-y) \cap C$. If $z$ were
nonzero, then $\lambda z$ would also be in $(S-y) \cap C$ for
all $\lambda > 0$, so $S$ would be unbounded, contrary to our
hypothesis. Hence, $z = 0$. Thus, by Lemma
\ref{lem:BoundedSectionsStrictlySupport}, $T$ is bounded, so
$T$ has a center of symmetry $x$, which must map to the center
of symmetry of $\ell \cap C'$ under $Q$. That is, $Q(x) = s$,
so $x \in S$, as desired. Thus, $S$ is a section of a
centrally symmetric set that contains the center of symmetry of
that set. Therefore, $S$ itself is centrally symmetric.
If $n \mathrel{\coloneqq} \operatorname{codim}(S) \ge 3$, fix a codimension-1 linear
subspace $W \subset V$ containing $S$. By the preceding
argument, $C \cap W$ is a CSS cone in which $S$ is a bounded
codimension-$(n-1)$ section intersecting the interior of $C
\cap W$. The theorem now follows from the induction hypothesis
applied to $C \cap W$.
\end{proof}
\begin{lemma}
\label{lem:CSSImpliesDualIsCoCSS}
Let $C$ be a CSS cone with nonempty interior in a normed vector
space $V$. Let $S^{\ast}$ be a bounded section of $\dualc C$
such that $\dim(S^{\ast}) < \infty$ and $S^{\ast} \cap
\operatorname{int}(\dualc C) \ne \emptyset$. Then $S^{\ast}$ is centrally
symmetric.
\end{lemma}
\begin{proof}
Let $S \mathrel{\coloneqq} \braces{y \in C \,:\, \text{$\psi(y) = 1$ for all
$\psi \in S^{\ast}$}}$. By Lemma
\ref{lem:PerpOfBoundedSectionIsBounded}, we have that $S$ is
bounded, intersects the interior of $C$, and has finite
codimension. Thus, by Lemma
\ref{lem:FiniteCodimSectionsAreCentSymm}, there exists a proper
section $T$ of $C$ containing $S$ whose center of symmetry $x$
is in $S$. Let $\phi \in C^{*}$ be such that $S_{\phi}
\mathrel{\coloneqq} S_{\phi}(C) = T$. Then, by
Lemma~\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, $\phi$ is
the center of symmetry of $S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$.
Note that $S^{\ast} \subset S_{x}$, so it remains only to show
that $\phi \in S^{\ast}$.
Indeed, since $S \cap \operatorname{int} C \ne \emptyset$, we have that
$\aff(S) = \coperp(\aff \dualc S)$. Thus, $\aff(\dualc S) =
(\aff S)^{\perp}$ by Lemma \ref{lem:PerpIsADuality}. In
particular, $\phi \in \aff(S^{\ast})$. Since $\phi \in \dualc
C$, we conclude that $\phi \in S^{\ast}$, as desired.
\end{proof}
The previous Lemma motivates the following definition.
\begin{definition}
Let $C$ be a closed pointed cone in a normed vector space $V$
with nonempty interior. We call $C$ \emph{co-CSS} if every
bounded finite-dimensional section of $C$ intersecting the
interior of $C$ is centrally-symmetric.
\end{definition}
Thus, Lemma \ref{lem:CSSImpliesDualIsCoCSS} says that the dual of
a CSS cone is co-CSS.
\begin{lemma}
\label{lem:FiniteDimSubConesOfCoCSSAreEllipsoidal}
Let $C$ be a co-CSS cone in a normed vector space $V$. Fix a
finite-dimensional subspace $L \subset V$ such that $L \cap
\operatorname{int} C \ne \emptyset$. Then the cone $L \cap C$ is
ellipsoidal.
\end{lemma}
\begin{proof}
Since $C$ is pointed, the finite-dimensional cone $L \cap C$ is
also pointed. Since $C$ is co-CSS, every bounded proper
section of $L \cap C$ is centrally symmetric. In particular,
$L \cap C$ is CSS. Therefore, by the finite-dimensional CSS
characterization of ellipsoidal cones \cite[Theorem
1.4]{JerMcA2013}, $L \cap C$ is ellipsoidal.
\end{proof}
\begin{lemma}
\label{lem:DualsOfCSSConesAreEllipsoidal}
Let $C$ be a CSS cone in a normed vector space $V$. Then the
dual cone $\dualc C$ is ellipsoidal.
\end{lemma}
\begin{proof}
Since $C$ is CSS, there exists a bounded proper section
$S_{\phi} \mathrel{\coloneqq} S_{\phi}(C)$ of $C$, where, by Lemma
\ref{lem:InteriorNormalsGiveBoundedSections}, $\phi \in
\operatorname{int}(\dualc C)$. Let $x \in C$ be the center of symmetry of
$S_{\phi}$. Then, by Lemma
\ref{lem:NormalOfCentSymmSectionIsCentOfSymm}, the section
$S_{x} \mathrel{\coloneqq} S_{x}(\dualc C)$ is centrally symmetric about
$\phi$. In addition, $S_{x}$ is a proper section of $\dualc C$
because it contains the point $\phi \in \operatorname{int}(\dualc C)$.
Furthermore, $S_{x}$ is bounded by Lemma
\ref{lem:InteriorNormalsGiveBoundedSections} because $x \in
\operatorname{int}(C) \hookrightarrow \operatorname{int}(\ddualc C)$ under the canonical
embedding. Finally, by Lemma
\ref{lem:FiniteDimSubConesOfCoCSSAreEllipsoidal}, every
finite-dimensional section of $S_{x}$ through $\phi$ is an
ellipsoid centered at $\phi$. Hence, by the Jordan--von
Neumann characterization of inner-product spaces
\cite{JorvNeu1935}, $S_{x}$ is an ellipsoid. Therefore, $\dual
C$ is ellipsoidal.
\end{proof}
We are now ready to prove our main result.
\begin{proof}[Proof of Theorem \ref{thm:CSSTheorem}]
\label{proof:CSSproof}
Every bounded proper section of an ellipsoidal cone is an
ellipsoid, and ellipsoids are centrally symmetric. Hence,
ellipsoidal cones are CSS.
To prove the converse, let a CSS cone of dimension $\ge 2$ be
given. By Lemma \ref{lem:DualsOfCSSConesAreEllipsoidal},
$\dual{C}$ is ellipsoidal. Fix an ellipsoidal proper section
$\dual{S}$ of $\dual{C}$, and let $\phi \in \dual{S}$. Thus,
we have an inner product on the codimension-$1$ linear subspace
$M \mathrel{\coloneqq} \aff(\dual{S}) - \phi$. Since $\dual{S}$ is
bounded, $\aff(\duala{S})$ does not contain the origin. Hence,
we can complete the inner product on $M$ to an inner product on
all of $\dual{V}$. Indeed, since $\dual{V}$ is already a dual
space, it is in fact a Hilbert space. The dual of an
ellipsoidal cone in a Hilbert space is ellipsoidal
\cite[p.~51]{Krein}, so $\ddual{C}$ is ellipsoidal. Since
$\ddual{C}$ is the closure of $C$ under the canonical embedding
$V \hookrightarrow \ddual{V}$, we conclude that $C$ itself is
ellipsoidal in $V$.
\end{proof}
It follows that if a normed vector space $V$ contains a CSS cone
with nonempty interior, then $V$ is an inner product space. In
particular, if a Banach space $X$ contains a full-dimensional CSS
cone, then $X$ is a Hilbert space.
\section{FBI cones are ellipsoidal cones}
\label{sec:FBICones}
We conclude by proving that every cone in a normed vector space
$V$ that satisfies the FBI property (Definition \ref{def:FBI}) is
ellipsoidal.
\begin{lemma}
\label{lem:ConeIntervalsAreBounded}
Let $C$ be a cone in a normed vector space $V$ such that
$\operatorname{int}(C)$ and $\operatorname{int}(\dualc C)$ are both nonempty. Then, for
each $a \in \operatorname{int}(C)$, the intersection $C \cap (a - C)$ is
bounded.
\end{lemma}
\begin{proof}
Since $\operatorname{int}(\dualc{C}) \ne \emptyset$, there exists a
functional $\phi \in \dualc C$ such that $S_{\phi} \mathrel{\coloneqq}
S_{\phi}(C)$ is a bounded base of $C$ by Lemma
\ref{lem:LinearBoundedSectionGeneratesCone}. In particular,
$\ker(\phi)$ strictly supports $C$ at $0$ by Lemma
\ref{lem:BoundedSectionsStrictlySupport}. By suitably
normalizing $\phi$, we also have that $S_{\phi}$ strictly
supports $a - C$ at $a$. Thus,
\begin{equation*}
C \cap (a - C) \subset \braces{y \in C \,:\, \phi(y) \le 1}.
\end{equation*}
Since $S_{\phi}$ is bounded and is a base for $C$, it follows
that $C \cap (a - C)$ is also bounded.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:FBIimpliesEllipsoidal}]
\label{proof:FBIimpliesEllipsoidal}%
Without loss of generality, suppose that $\operatorname{int}{C} \ne
\emptyset$. Fix $x \in \operatorname{int}(C)$, and set $S \mathrel{\coloneqq}
\conv(\partial C \cap \partial(2x - C))$. By the FBI property and
Lemma \ref{lem:ConeIntervalsAreBounded}, $S$ is a bounded
section of $C$. Moreover, $S$ is centrally symmetric about
$x$. We show that every $2$-dimensional section of $S$ through
$x$ is an ellipse. Let $E$ be such a section. Observe that $0
\notin \aff E$, because $0 \in \aff E \subset \aff S$ would
imply that $\aff S = \lin S$, contrary to the boundedness of
$S$. Therefore, $C' \mathrel{\coloneqq} \cone(E)$ is a $3$-dimensional
cone. We claim that $\cone(E)$ is an FBI cone. To prove this,
let $y \in \relintr C'$, and let $\Gamma \mathrel{\coloneqq} \partial(C')
\cap \partial(y - C')$. On the one hand, $\Gamma$ is contained in
$\lin E$. On the other hand, let $H$ be the affine hyperplane
in $V$ containing $\partial C \cap \partial(y - C)$. As above, $0
\notin H$. Since $0 \in \lin E$, it follows that $\Gamma
\subset (\lin E) \cap H$, but $\lin E \not\subset H$. Hence,
$\Gamma$ is contained in a $2$-dimensional affine subspace of
$\lin E$. That is, $C'$ is an FBI cone. It follows from
\cite[Theorem 1.2]{JerMcA2013} that $E$ is an ellipse.
Therefore, by the Jordan--von Neumann characterization of
inner-product spaces \cite{JorvNeu1935}, $S$ is an ellipsoid.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
Massive stars, when evolving off the main sequence, undergo phases of strong
mass-loss, often resulting in the formation of shells, disks or rings of
circumstellar material. One class of evolved massive stars, B[e] supergiants
(B[e]SGs), are early-type emission line stars with high-density gaseous and
dusty circumstellar disks of yet unknown origin \citep[see][for a recent
review]{2014AdAst2014E..10D}. The gaseous inner disk is
traced by the emission of low-ionized metal lines from both permitted and
forbidden transitions \citep[e.g.,][]{1985A&A...143..421Z}. Of these, the
lines of
[Ca\,II] and [O\,I]
are particularly useful. While their
intensities probe regions of high and medium densities,
where the transition between the two regimes might be settled
at an electron density of $\sim 10^{7}$\,cm$^{-3}$, their line profiles
contain the information on their kinematics \citep{2007A&A...463..627K,
2010A&A...517A..30K, 2012MNRAS.423..284A}. Farther out, molecules form via gas
phase chemistry, and intense band emission from CO molecules has been reported
for many of these stars \citep{1988ApJ...324.1071M,
1988ApJ...334..639M, 1989A&A...223..237M, 1996ApJ...470..597M,
2000A&A...362..158K, 2014ApJ...780L..10K, 2010MNRAS.408L...6L,
2012A&A...548A..72C, 2012A&A...543A..77W, 2012MNRAS.426L..56O,
2013A&A...558A..17O}. In addition, the presence of TiO band emission features
in optical spectra of several B[e]SGs has been suggested
\citep{1989A&A...220..206Z, 2012MNRAS.427L..80T}, but still lacks confirmation.
Besides atomic and molecular gas, these stars are also surrounded by warm
circumstellar dust, as is obvious from their strong infrared excess emission
\citep{1986A&A...163..119Z, 2009AJ....138.1003B, 2010AJ....140..416B},
interferometric observations \citep{2007A&A...464...81D, 2011A&A...526A.107M,
2012A&A...548A..72C}, and resolved spectral dust features
\citep{2006ApJ...638L..29K, 2010AJ....139.1993K}.
Kinematic studies of the atomic and CO gas reveal that the material
is confined to detached (sometimes multiple) rings rotating around the star on
Keplerian orbits \citep[e.g.,][]{2010A&A...517A..30K, 2013A&A...549A..28K,
2014ApJ...780L..10K, 2012MNRAS.423..284A, 2012A&A...543A..77W,
2012A&A...548A..72C, 2014arXiv1409.7550M}. Abundance studies of these rings
reveal an enrichment in the isotopic molecule $^{13}$CO, in agreement with an
evolutionary stage just beyond the main sequence \citep{2009A&A...494..253K,
2010MNRAS.408L...6L, 2013A&A...549A..28K, 2013A&A...558A..17O,
2014arXiv1409.7550M}.
Furthermore,
B[e]SGs could be the progenitors of a group of early-type blue supergiants
(BSGs) with ring nebulae such as, e.g., Sher\,25 and SBW\,1. Optical images of
these two stars display equatorial rings and bipolar lobes, and
\citet{2007AJ....134..846S} propose that as soon as the rings or disk-like
structures seen around B[e]SGs have expanded and cooled they would
look exactly like these objects.
Support for such a possible evolutionary link comes from the fact that,
like B[e]SGs, neither Sher\,25 nor SBW\,1 has evolved yet through a red
supergiant phase, indicating that in both cases, the ejection of the material
forming the disks or rings must have happened during their BSG phase.
Such a scenario has a much wider impact: as the BSGs with rings
closely resemble the progenitor star of the supernova SN 1987A,
B[e]SGs might be regarded as the most plausible candidates
for supernova explosions of the SN 1987A-type.
Notably, two B[e]SGs were recently suggested to be viable SN\,1987A-type
progenitor candidates: \object{LHA\,115-S\,18} in the Small Magellanic Cloud
\citep{2013A&A...560A..10C} and the Galactic object MWC\,137
\citep{2014arXiv1409.7550M}.
For a better understanding of
the evolution of these stars up to the supernova stage, it is hence essential
to study B[e]SGs in great detail, and in particular their mass-loss behavior
and the chemistry involved in the formation of the observed dense molecular and
dusty rings or disk-like structures.
The atmospheres of massive stars have an oxygen-rich composition, which means
that the abundance of oxygen atoms greatly exceeds that of carbon atoms.
The same trend is then expected in their disks, because they form from the
material released from the stellar surface. Consequently, carbon atoms in the
disks are locked in CO, which, as the most stable molecule with a
dissociation energy of 11.2\,eV, forms as soon as the gas temperature drops
below the dissociation temperature of CO, for which a value of
$\sim 5000$\,K is typically quoted. The excess oxygen atoms will form other
molecules, and a rich chemistry is expected at larger distances from the star.
The SiO molecule is the second most stable oxygen compound. Its binding energy
of 8.0\,eV is higher than that of other oxides or oxygen compounds such
as NO, OH, or water, but lower than that of CO, and thus SiO should form in
slightly cooler environments ($T_{\rm SiO} < T_{\rm CO} < 5000$\,K).
As with CO, the rather high gas temperature of the SiO forming region
guarantees that high vibrational levels will be excited, causing coupled
rotational-vibrational transitions, detectable via prominent band and band head
features. And in fact, SiO bands are commonly
observed in absorption in the atmospheres of Mira stars, cool giants,
and red supergiants
\citep{1976ApJ...210L.141H, 1979ApJ...230L..47G, 1997A&A...323..202A,
1999A&A...342..799A}. In these objects, SiO is an important molecule, based on
which the physical properties in the transition region between the outer
atmosphere and the innermost circumstellar envelope can be studied
\citep[e.g.,][]{2014A&A...561A..47O}.
To our knowledge, rotational-vibrational SiO bands in emission have only
been reported for
Supernova 1987A \citep{1988MNRAS.235P..19A, 1991MNRAS.252P..39R,
1994ApJ...428..769L}. These emission bands appeared 160 days
after the outburst and were not detectable beyond day 578.
The rise of the SiO features had a time-delay of 48 days with respect to
that of CO molecular bands. This is in agreement with the lower binding
energy, requiring a cooler environment for the formation of SiO molecules
by gas phase chemistry. The disappearance of the emission bands agrees
with the onset of dust formation in the ejecta, which is interpreted as
depletion of molecular SiO due to condensation of silicate dust, for which
the SiO molecule is proposed to be a major building block within
an oxygen-rich environment.
Unlike the transient phenomena of a rapidly expanding, diluting ejecta of a
supernova, the rings or disks around B[e]SGs seem to be quite stable
structures of accumulated material. This guarantees these disks
as ideal chemical laboratories to study molecule formation and
dust condensation. However, apart from CO band
emission, detected in most of these objects, and warm circumstellar dust,
no information exists about the chemical composition of these disks.
Hence, it is necessary to begin searching for other molecular emission features
that can be used to study the structure and kinematics of B[e]SG stars' disks
comprehensively.
To pioneer this work, we selected a sample of Galactic B[e]SGs, with
known CO band emission, to search for molecular rotational-vibrational
band emission from SiO.
\section{Observations}
During the period of March-May 2014, we observed four Galactic B[e]SGs in the
$L$-band wavelength region using the CRyogenic high-resolution InfraRed Echelle
Spectrograph \citep[CRIRES,][]{2004SPIE.5492.1218K} on the ESO VLT UT1-Antu
8-m telescope. Details on the objects and their observations are given in
Table\,\ref{tbl_obs}. Data were obtained at four different standard
settings with reference wavelengths of 4048.6, 4060.1, 4135.5, and 4147.0 nm,
so that the total spectra cover a continuous wavelength range from
3.988--4.176\,$\mu$m. This range contains the first four band heads of the SiO
first-overtone bands. We used the 0.4\arcsec slit, which provides a spectral
resolution of $R\sim 50\,000$, corresponding to a velocity resolution of
$\sim 6$\,km\,s$^{-1}$.
Data reduction was performed with the ESO CRIRES pipeline (version 2.3.2).
The observations were taken in an ABBA nod pattern for proper sky subtractions.
Raw frames were corrected for bad pixels, flat fields, non-linearity, and then
wavelength calibrated using OH lines in the standard star spectrum.
A telluric standard star observation was similarly reduced. For telluric
correction, a B-type standard star was observed immediately after the target at
a similar airmass. The IRAF\footnote{IRAF is distributed by the National Optical
Astronomy Observatory, which is operated by the Association of Universities for
Research in Astronomy (AURA) under cooperative agreement with the National
Science Foundation.} task \textit{telluric} was used to remove the atmospheric
features. The final spectra were normalized and continuum subtracted.
\section{Results}
We clearly detect emission from the first three band heads of the SiO
first-overtone bands in CPD-52\,9243 (Fig.\,\ref{fig_cpd-fit}), and from
the first band head in the spectra of the other three stars, CPD-57 2874,
HD\,327083, and HD\,62623 (Fig.\,\ref{fig_sio}).
In each star's spectrum, the band heads display a blue-shifted shoulder and a
red-shifted maximum with respect to the laboratory wavelength. Their
separation is marked by the bar in Figs.\,\ref{fig_cpd-fit} and \ref{fig_sio}.
Such structures are also typically observed in the band heads of CO, and are an
unequivocal manifestation of rotational broadening.
To confirm that the observed features are emission from
SiO molecules, we compute synthetic band head models. For this, we modify our
existing CO disk code \citep{2000A&A...362..158K} by implementing the SiO line
lists and Einstein transition probabilities provided by \citet{Barton2013}. For
each of our targets, information about the rotation velocity of the CO region
and disk inclination angles are listed in Table\,\ref{tbl_param}. The
rotation velocities of the CO gas have been obtained from the broadening of the
first band head of the CO first-overtone bands resolved in high-resolution
spectra, and the disk inclination angles have been determined from
interferometric observations. From the wavelength separation between the
position of the blue shoulder and the red peak of the first band head, we
obtain an estimate of the SiO rotation velocity projected to the line of
sight. This parameter is refined during the fitting procedure. The final
rotation velocities, corrected for disk inclination, are included in
Table\,\ref{tbl_param}. The high spectral resolution, combined with the
sensitivity of the synthetic spectra to rotation velocity, guarantee that these
values have a precision of $\pm 1$\,km\,s$^{-1}$.
Comparing the velocities obtained in the two molecular band emitting regions,
we find a consistently lower value for SiO, suggesting that this molecule is
formed within the Keplerian disk at larger distances from the star.
For CPD-52 9243, where three observed band heads are available, we can also
constrain temperature and column density, for which we find $T_{\rm SiO} =
2000\pm 200$\,K and $N_{\rm SiO} = (5\pm 2)\times 10^{21}$\,cm\,$^{-2}$.
With this column density, most of the individual SiO rotation-vibration
lines within the observed wavelength range are optically thick.
Although the SiO rotation velocity in this object is only marginally lower,
both temperature and column density are significantly lower than the values
obtained from the CO modeling \citep[see][]{2012A&A...548A..72C}. This
suggests that the two band head formation regions are physically disjoint.
For the other three objects, where solely the first band head is detected,
it is only possible to constrain the
rotation velocity, because different combinations of temperature and density
result in almost identical synthetic spectra. Hence, the fits shown in
Fig.\,\ref{fig_sio} are only to demonstrate that the observed features can
indeed be assigned to SiO band emission, and to determine the SiO rotation
velocity.
Under the assumption of Keplerian rotation of the disks and using the
stellar mass ranges of the objects (Table\,\ref{tbl_obs}), the radii of both
(CO and SiO) molecular rings can be computed. These values are included in
Table\,\ref{tbl_param}. Furthermore, considering that the circumstellar dust
is mainly composed of silicates, we can calculate the dust evaporation
distances. Assuming that the grains are spheres with a maximum size of $\sim
1\,\mu$m, and that the silicate evaporation temperature is $\sim 1500$\,K, the
minimum dust evaporation radii follow from the stellar luminosities (listed
in Table\,\ref{tbl_obs}) and are added to Table\,\ref{tbl_param}. Obviously,
the much hotter molecular rings reside within the dusty disks.
\section{Conclusions}
We report on the detection of SiO first-overtone band emission in the
high-resolution $L$-band spectra of four Galactic B[e]SGs, which are known to
display strong CO band emission. While SiO bands are commonly seen in
absorption from the atmospheres of cool stars and giants, they were detected in
emission, so far, only in spectra of SN\,1987\,A during the post-explosion
period from day 160 to day 578.
Hence, to our knowledge, this is the first discovery of SiO
first-overtone band emission from hot, high-density environments of
evolved massive stars.
The high spectral resolution of our data allows us to precisely determine
the rotation velocity of the SiO gas. For all of our target stars, this
velocity is lower than that determined from the CO band forming region.
Moreover, model fits to the observed three band heads in CPD-52 9243
demonstrate that, in contrast to the CO band emitting region, the emission of
the SiO bands originates from a cooler and less dense disk region, indicating
that SiO molecules form at larger distances within the Keplerian disks.
For all four objects, we find, in addition, that the SiO ring radii are smaller
than the silicate dust evaporation distances.
This finding is of great importance, because it demonstrates that molecules
provide an excellent tool to study B[e]SG stars' disks. The emission features
of both molecules detected so far, CO and SiO, carry all the essential
information about the physical properties (temperature, density) and kinematics
of their formation region. As different molecules require diverse conditions, a
huge variety can be expected to form throughout the disk, filling the space
between the hottest inner rim traced by CO bands, and the dust condensation
region. Hence, it is essential and timely to search for emission features from
many more molecules that are expected to form within the oxygen-rich
environment of these disks. Moreover, molecules such as CO and SiO are
also indispensable to study the disk structure at much larger distances. With
their numerous pure rotational transitions, these molecules are also suitable
to trace much cooler environments. The rotational transitions spread over the
submm and radio regime, making the disks of B[e]SGs ideal targets to be
observed with high spatial and spectral resolution facilities such as ALMA.
Combining the results obtained from different
molecular species tracing diverse temperature regions, and hence distances from
the star, will greatly improve our knowledge and comprehension of the disk
structure, and help to understand the disks' possible formation history.
In addition, the discovery of SiO band emission has an even wider impact.
Evolved massive stars, such as red supergiants, B[e]SGs, and supernovae
(e.g., SN\,1987A), are the main sources of silicate dust, for which SiO
molecules are
regarded as a major building block. Studying the physical properties of SiO gas
in the disks of B[e]SGs might, hence, provide indispensable insight into the
early stage of silicate dust formation within the oxygen-rich environment of
these enigmatic stars.
\acknowledgments
We thank the referee, Dr. John Bally, for his valuable comments on the
manuscript. This research made use of the NASA Astrophysics Data System (ADS)
and of the SIMBAD database, operated at CDS, Strasbourg, France.
MK acknowledges financial support from GA\v{C}R (grant number 14-21373S). The
Astronomical Institute Ond\v{r}ejov is supported by the project RVO:67985815.
LC, MLA, and AFT acknowledge financial support from the Agencia de Promoci\'on
Cient\'{\i}fica y Tecnol\'ogica (Pr\'estamo BID PICT 2011/0885), CONICET
(PIP 0300), and the Universidad Nacional de La Plata (Programa de Incentivos
G11/109), Argentina. Financial support for International Cooperation of the
Czech Republic (M\v{S}MT, 7AMB14AR017) and Argentina (Mincyt-Meys ARC/13/12 and
CONICET 14/003) is acknowledged.
|
\section{Introduction}
In 1868, the French mathematician \'E. L\'eonard Mathieu introduced a family of differential equations nowadays termed Mathieu equations in his ``\emph{memoir on vibrations of an elliptic membrane}''~\cite{Mat68}.
Mathieu's equation is related to the wave equation for the elliptic cylinder. Mathieu is notably remembered for his discovery of sporadic simple groups~\cite{ConSlo88}.
This paper is particularly concerned with the canonical form of the Mathieu Equation.
For $a\in \mathbb{R}$, $q \in \mathbb{C}$,
the Mathieu Equation is given by
\begin{equation}
\label{eq1}
\frac{\mathrm{d}^2 y}{\mathrm{d}\omega^2}
+
(a - 2 q \cos(2\omega))y
= 0.
\end{equation}
The Mathieu equation is a linear second-order differential equation with periodic coefficients. This equation was shown later to be also related to quantum mechanicals; the parameters $a$ and $q$ denote the energy level and an intensity, respectively.
For $q=0$ it reduces to the well-known harmonic oscillator, $a$ being the square of the frequency~\cite{McL64}.
The solution of~(\ref{eq1}) is the elliptic-cylindrical harmonic, known as Mathieu functions. In addition to being theoretically fascinating, Mathieu functions are applicable to a wide variety of physical phenomena, e.g., diffraction, amplitude distortion,
inverted pendulum,
stability of a floating body,
radio frequency quadrupole,
and vibration in a medium with modulated density~\cite{Rub96}.
They have also long been applied on a broad scope of waveguide problems involving elliptical geometry, including:
(i) analysis for weak guiding for step index elliptical core optical fibres~\cite{Sha95},
(ii) power transport of elliptical waveguides~\cite{HusWur97,Hen98},
(iii) evaluating radiated waves of elliptical horn antennas~\cite{Cao97},
(iv) elliptical annular microstrip antennas with arbitrary eccentricity~\cite{SunTra93},
and (v) scattering by a coated strip~\cite{HolCab92}.
The aim of this paper is to propose a new family of wavelets based on Mathieu differential equations.
Wavelets are a well-known tool for differential equation solving~\cite{Beylkin,Dahmen,Alpert}.
However, in this work, we show another connection between wavelets and differential equations:
the design of new wavelets from the solution of a differential equation.
\section{Mathieu Equations}
In general, the solutions of~(\ref{eq1}) are not periodic. However, for a given $q$, periodic solutions exist for infinitely many special values (eigenvalues) of $a$.
For several physically relevant solutions $y$ must be periodic of period $\pi$ or $2\pi$.
It is also convenient to distinguish even and odd periodic solutions, which are termed Mathieu functions of first kind. One of four simpler types can be considered: Periodic solution ($\pi$ or $2\pi$) symmetry (even or odd). For $q\neq0$, the only periodic solution $y$ corresponding to any characteristic value $a=a_r(q)$ or $a=b_r(q)$ has the following notation:
\noindent
\emph{Even periodic solution}
\begin{equation}
\label{eq2a}
\ce_r(\omega,q)
=
\sum_m
A_{r,m}
\cos m\omega
\qquad
\text{for $a = a_r(q)$},
\tag{2a}
\end{equation}
\noindent
\emph{Odd periodic solution}
\begin{equation}
\label{eq2b}
\se_r(\omega,q)
=
\sum_m
A_{r,m}
\sin m\omega
\qquad
\text{for $a = b_r(q)$},
\tag{2b}
\end{equation}\setcounter{equation}{2}
\!\!%
where the sums are taken over even (respectively odd) values of $m$ if the period of $y$ is $\pi$ (respectively $2\pi$). Given $r$, we denote henceforth $A_{r,m}$ by $A_m$, for short.
Elliptic cosine and elliptic sine functions are represented by
$\ce$ and $\se$, respectively.
Interesting relationships are found when $q\to0$, $r\neq0$~\cite{AbraSte68}:
\begin{equation}
\label{eq3}
\lim_{q\to0} \ce_r(\omega,q) = \cos(r\omega), \quad
\lim_{q\to0} \se_r(\omega,q) = \sin(r\omega).
\end{equation}
One of the most powerful results of Mathieu's functions is the Floquet's Theorem~\cite{Flo83}. It states that periodic solutions of~(\ref{eq1}) for any pair $(a, q)$ can be expressed in the form
\begin{equation}
\label{eq5}
\begin{split}
y(\omega)&=
F_\nu(\omega)=
e^{j\nu \omega}P(\omega)
\qquad\text{or}\\
y(\omega)&=
F_\nu(-\omega)=
e^{-j\nu \omega}P(-\omega),
\end{split}
\end{equation}
where $\nu$ is a constant depending on $a$ and $q$ and $P(\cdot)$ is $\pi$-periodic in $\omega$. The constant $\nu$ is called the characteristic exponent. If $\nu$ is an integer, then $F_\nu(\omega)$ and $F_\nu(-\omega)$ are linear dependent solutions.
Furthermore, $y(\omega+k\pi) = e^{j\nu k \pi}y(\omega)$ or
$y(\omega+k\pi) = e^{-j\nu k \pi}y(\omega)$, for the solution $F_\nu(\omega)$ or $F_\nu(-\omega)$, respectively. We assume that the pair $(a, q)$ is such that $|\cosh(j\nu\pi)|<1$ so that the solution $y(\omega)$ is bounded on the real axis~\cite{GradRyz65}.
The general solution of Mathieu's equation ($q\in\mathbb{R}$, $\nu$ non-integer) has the form
\begin{equation}
\label{eq6}
y(\omega) = c_1 e^{j\nu \omega}P(\omega) + c_2 e^{-j\nu \omega} P(-\omega),
\end{equation}
where $c_1$ and $c_2$ are arbitrary constants.
All bounded solutions ---those of fractional as well as integral order--- are described by an infinite series of harmonic oscillations whose amplitudes decrease with increasing frequency. In the wavelet framework we are basically concerned with even solutions of period $2\pi$. In such cases there exist recurrence relations among the coefficients~\cite{AbraSte68}:
\begin{equation}
\label{eq7}
\begin{split}
(a-1-q)A_1 - qA_3 &= 0, \\
(a-m^2)A_m - q (A_{m-2} + A_{m+2}) &= 0, \quad \text{$m\geq3$, $m$ odd.}
\end{split}
\end{equation}
In the sequel, wavelets are denoted by $\psi(t)$ and scaling functions by $\phi(t)$, with corresponding spectra $\Psi(\omega)$ and $\Phi(\omega)$, respectively.
\section{Mathieu Wavelets}
Wavelet analysis has matured rapidly over the past years and has been proved to be invaluable for scientists and engineers~\cite{Bul95}. Wavelet transforms have lately gained extensive applications in an amazing number of areas~\cite{Fou95}.
The equation $\phi(t) = \sqrt2 \sum_{n\in\mathbb{Z}} h_n \phi(2t-n)$, which is known as the \emph{dilation} or \emph{refinement equation}, is the chief relation determining a Multiresolution Analysis (MRA)~\cite{Mall00}.
\subsection{Two Scale Relation of Scaling Function and Wavelet}
Defining the spectrum of the smoothing filter $\{h_k\}$ by
$
H(\omega)
\triangleq
\frac{1}{\sqrt2}
\sum_{k\in\mathbb{Z}}
h_k
e^{-j\omega k},
$
the central equations (in the frequency domain) of a Multiresolution analysis are~\cite{Mall89}:
\begin{equation}
\label{eq9}
\Phi(\omega) = H\left(\frac{\omega}{2}\right)\Phi\left(\frac{\omega}{2}\right)
\quad\!\!\text{and}\!\!
\quad
\Psi(\omega) = G\left(\frac{\omega}{2}\right)\Phi\left(\frac{\omega}{2}\right),
\end{equation}
where
$G(\omega) \triangleq \frac{1}{\sqrt2} \sum_{k\in\mathbb{Z}} g_k e^{-j\omega k}$, is the transfer function of the detail filter.
The orthogonality condition corresponds to~\cite{Mall89}:
\begin{align}
\label{eq10a}
H(0) = 1\ \text{and}\ H(\pi)=0, \tag{8a}\\
\label{eq10b}
|H(\omega)|^2 + |H(\omega+\pi)|^2 = 1, \tag{8b} \\
\label{eq10c}
H(\omega) = - e^{-j\omega}G^{\ast}(\omega+\pi).\tag{8c}
\end{align}
\setcounter{equation}{8}
\subsection{Filters of a Mathieu MRA}
The subtle liaison between Mathieu's theory and wavelets was found by observing that the classical relationship
\begin{equation}
\Psi(\omega) =
e^{-j\omega/2}
H^\ast\left( \frac{\omega}{2}-\pi \right)
\Phi\left( \frac{\omega}{2}\right)
\end{equation}
presents a remarkable similarity to a Floquet's solution of a Mathieu's equation, since $H(\omega)$ is a periodic function.
As a first attempt, the relationship between the wavelet spectrum and the scaling function was put in the form:
\begin{equation}
\label{eq12}
\frac{\Psi(\omega)}{\Phi\left( \frac{\omega}{2}\right)}
=
e^{-j\omega/2}
H^\ast\left( \frac{\omega}{2}-\pi \right).
\end{equation}
Here, on the second member, neither $\nu$ is an integer nor $H(\cdot)$ has a period $\pi$. By an appropriate scaling of this equation, we can rewrite it as
\begin{equation}
\label{eq13}
\frac{\Psi(4\omega)}{\Phi(2\omega)}
=
e^{-j2\omega}
H^\ast(2\omega-\pi).
\end{equation}
Defining a new function
$Y(\omega) \triangleq {\Psi(4\omega)}/{\Phi(2\omega)}$,
we recognise that it
has a nice interpretation in the wavelet framework. First, we recall that $\Psi(\omega) = G\left(\frac{\omega}{2}\right)\Phi\left(\frac{\omega}{2}\right)$ so that $\Psi(2\omega) = G(\omega)\Phi(\omega)$. Therefore the function related to Mathieu's equation is exactly $Y(\omega)=G(2\omega)$. Introducing a new variable $z$, which is defined according to $2z\triangleq 2\omega-\pi$, it follows that $-Y\left( z + \frac{\pi}{2}\right) = e^{-j2z}H^{\ast}(2z)$. The characteristic exponent can be adjusted to a particular value $\nu$,
\begin{equation}
\label{eq14}
-e^{-j(\nu-2)z}
Y\left( z + \frac{\pi}{2}\right)
=
e^{-j\nu z}H^{\ast}(2z).
\end{equation}
Defining now $P(-z) \triangleq H^\ast(2z)= \sum_{k\in\mathbb{Z}}c_{2k} e^{jz2k}$, where $c_{2k} \triangleq \frac{1}{\sqrt2}h^\ast_k$, we figure out that the right-side of the above equation represents a Floquet's solution of some differential Mathieu equation. The function $P(\cdot)$ is $\pi$-periodic verifying the initial condition $P(0) = \frac{1}{\sqrt2}\sum_k h_k = 1$, as expected. The filter coefficients are all assumed to be real. Therefore, there exist a set of parameters $(a_G, q_G)$ such that the auxiliary function
\begin{equation}
\label{eq17}
y_\nu(z)
\triangleq
-e^{-j(\nu-2)z}
Y_\nu\left( z + \frac{\pi}{2} \right)
\end{equation}
is a solution of the following Mathieu equation:
\begin{equation}
\label{eq18}
\frac{\mathrm{d}^2y_\nu}{\mathrm{d}z^2}
+
(a_G - 2 q_G \cos(2z) )
y_\nu = 0,
\end{equation}
subject to $y_\nu(0) = - Y(\pi/2) = -G(\pi) = -1$ and $\cos(\pi \nu) - y_\nu(\pi) = 0$, that is, $y_\nu(\pi) = (-1)^\nu$.
In order to investigate a suitable solution of~(\ref{eq18}), boundary conditions are established for predetermined $a$, $q$. It turns out that when $\nu$ is zero or an integer, $a$ belongs to the set of characteristic values $a_r(q)$. Furthermore, $\nu=r$ is associated with $a_r(q)$. The even ($2\pi$-periodic) solution of such an equation is given by:
\begin{equation}
\label{eq19}
y_\nu(z) = - \frac{\ce_\nu(z,q)}{\ce_\nu(0,q)}.
\end{equation}
The $Y_\nu(\omega)$ function associated to $y_\nu(z)$ and related to the detail filter of a ``Mathieu MRA'' is thus:
\begin{equation}
Y_\nu(\omega) = G_\nu(2\omega) =
e^{j(\nu-2)\left( \omega - \frac{\pi}{2}\right)}
\frac{\ce_\nu\left( \omega - \frac{\pi}{2}, q\right)}{\ce_\nu(0,q)}.
\end{equation}
Finally, the transfer function of the detail filter of a Mathieu wavelet is
\begin{equation}
\label{eq21}
G_\nu(\omega) =
e^{j(\nu-2)\left( \frac{\omega - \pi}{2}\right)}
\frac{\ce_\nu\left( \frac{\omega -\pi}{2}, q\right)}{\ce_\nu(0,q)}.
\end{equation}
The characteristic exponent $\nu$ should be chosen so as to guarantee suitable initial conditions, i.e., $G_\nu(0)=0$ and $G_\nu(\pi)=1$, which are compatible with wavelet filter requirements. Therefore, $\nu$ must be odd. It is interesting to remark that the magnitude of the above transfer function corresponds exactly to the modulus of a elliptic sine~\cite{GradRyz65}:
\begin{equation}
\label{eq22}
|G_\nu(\omega)| =
\left|
\se_\nu\left( \frac{\omega}{2}, -q \right)/
\ce_\nu(0,q)
\right|.
\end{equation}
The solution for the smoothing filter $H(\cdot)$ can be found out via QMF conditions~\cite{Mall00}, yielding:
\begin{equation}
\label{eq23}
H_\nu(\omega) =
- e^{-j \nu \frac{\omega}{2}}
\frac{\ce_\nu \left( \frac{\omega}{2}, q\right) }{\ce_\nu(0,q)}.
\end{equation}
In this case, we find $H_\nu(\pi)=0$ and
\begin{equation}
\label{eq24}
|H_\nu(\omega)| =
\left|
\ce_\nu\left( \frac{\omega}{2}, q \right)/
\ce_\nu(0,q)
\right|.
\end{equation}
Given $q$, the even first-kind Mathieu function with characteristic exponent $\nu$ is given by
$
\ce_\nu(\omega,q) =
\sum_{l=0}^\infty
A_{2l+1} \cos(2l + 1) \omega
$,
in which $\ce_\nu(0,q) = \sum_{l=0}^\infty A_{2l+1}$.
The $G$ and $H$ filter coefficients of a Mathieu MRA can be expressed in terms of the values $\{ A_{2l+1}\}_{l \in \mathbb{Z}}$ of the Mathieu function as:
\begin{equation}
\label{eq26}
\frac{h_l^\nu}{\sqrt2}
=
-\frac{A_{|2l-\nu|}/2}{\ce_\nu(0,q)}
\quad\text{and}\quad
\frac{g_l^\nu}{\sqrt2}
=
(-1)^l\frac{A_{|2l+\nu-2|}/2}{\ce_\nu(0,q)}.
\end{equation}
It is straightforward to show that
$h_{-l}^\nu = h_{l+\nu}^\nu$,
$\forall l>0$.
The normalising conditions are
$\frac{1}{\sqrt2}\sum_{k=-\infty}^{\infty}h_k^\nu = -1$
and
$\sum_{k=-\infty}^{\infty}(-1)^k h_k^\nu = 0$.
\begin{figure}
\centering
\subfigure[]{\epsfig{file=fig1a.eps,width=.35\linewidth}}
\subfigure[]{\epsfig{file=fig1b.eps,width=.35\linewidth}}
\caption{Magnitude of the transfer function for Mathieu multiresolution analysis filters: smoothing filter $|H_\nu(\omega)|$ (solid line) and detail filter $|G_\nu(\omega)|$ (bold line) for a few Mathieu parameters.
(a) $\nu=3$, $q=3$, $a = 9.915506290452134$;
(b) $\nu=5$, $q=15$, $a = 31.957821252172874$.
}
\label{fig2}
\end{figure}
\begin{figure}
\centering
\subfigure[]{%
\input{fig2a1.latex}
\input{fig2a2.latex}
\input{fig2a3.latex}
}
\subfigure[]{%
\input{fig2b1.latex}
\input{fig2b2.latex}
\input{fig2b3.latex}
}
\caption{FIR-Based Approximation of Mathieu Wavelets as the number of iteration increases
(2, 4, and 6 iterations, respectively).
Filter coefficients holding $|h| <10^{-10}$ were thrown away
(19 retained coefficients per filter in both cases).
(a) Mathieu Wavelet with $\nu=3$ and $q=3$ and
(b) Mathieu Wavelet with $\nu=5$ and $q=15$.}
\label{fig3}
\end{figure}
\section{Examples}
Illustrative examples of filter transfer functions for a Mathieu MRA are shown in Fig.~\ref{fig2}, for $\nu=\text{3}$ and $\text{5}$, and a particular value of $q$ (numerical solution obtained by 5-order Runge-Kutta method). The value of $a$ is adjusted to an eigenvalue in each case, leading to a periodic solution. Such solutions present a number of $\nu$ zeroes in the interval $|\omega|<\pi$. We observe lowpass behaviour (for the filter $H$) and highpass behaviour (for the filter $G$), as expected.
Mathieu wavelets can be derived from the lowpass reconstruction filter by the cascade algorithm. %
Infinite Impulse Response filters (IIR)
should be applied since Mathieu wavelet has no compact support. However a Finite Impulse Response (FIR) approximation can be generated by discarding negligible filter coefficients, say less than $10^{-10}$.
In Fig.~\ref{fig3}, emerging pattern that progressively looks like the wavelet shape is shown for some couple of parameter $a$ and~$q$.
Waveforms were derived using the \textsc{Matlab} wavelet toolbox.
As with many wavelets there is no nice analytical formula for describing Mathieu wavelets.
\section{Conclusions}
A new and wide family of elliptic-cylindrical wavelets was introduced. It was shown that the transfer functions of the corresponding multiresolution filters are related to Mathieu equation solutions. The magnitude of the detail and smoothing filters corresponds to first-kind Mathieu functions with odd characteristic exponent. The number of zeroes of the highpass $|G(\omega)|$ and lowpass $|H(\omega)|$ filters within the interval $|\omega|<\pi$ can be appropriately designed by choosing the characteristic exponent. This seems to be the first connection found between Mathieu equations and wavelet theory. It opens new perspectives on linking wavelets and solutions of other differential equations (e.g. Associated Legendre functions).
Although there exist plenty of potential applications for Mathieu Wavelets,
none are presented:
we just disseminate the major ideas,
letting further research to be investigated.
For instance, this new family of wavelets could
be an interesting tool for analysing optical fibres due to
its ``elliptical'' symmetry.
They could as well be beneficial when examining molecular dynamics of
charged particles in electromagnetic traps such as Paul trap or the mirror trap for
neutral particles~\cite{NasFoo01,Nieetal01}.
\section*{Acknowledgements}
The authors thank Dr. Harold V. McIntosh from
\emph{Departamento de Aplicaci\'on de Microcomputadoras,
Universidad Aut\'onoma de Puebla}, Mexico for information about Mathieu's functions. They also express their indebtedness to Dr.~R.~M.~Campello de Souza (Federal University of Pernambuco, Brazil) for stimulating comments.
\nocite{*}
\bibliographystyle{ieeetr}%
|
\section{Introduction}
It is now well established that the pulsed spectrum of the Crab extends up to 400 GeV \citep{2011Sci...334...69V,2012A&A...540A..69A} and very recently new measurements of TeV photons have been announced \citep{crabtev}. Although this is the most extreme example, several other pulsars exhibit emission above 25 GeV \citep{2013ApJS..209...34A},
including the Vela pulsar with reported detection above 50 GeV \citep{2014arXiv1410.5208L}.
The Fermi-LAT data of these objects have been fitted with a sub-exponential cut-off or a broken power-law.
Such emission has been modelled as inverse Compton scattering of magnetospheric photons by the $e^{\pm}$ pairs created and accelerated in the magnetosphere \citep{2013MNRAS.431.2580L}. As an alternative, wind models have been proposed, which attribute the gamma-ray emission to the synchrotron radiation of energized plasma either very close to the light cylinder \citep{1996A&A...311..172L, 2010ApJ...715.1282B, 2013A&A...550A.101A}, or in the far wind zone \citep{2012MNRAS.424.2023P}. Inverse Compton scattering of the thermal stellar photons by the instantaneously accelerated wind has been also considered in this context \citep{2012Natur.482..507A}.
In the standard wind scenario (the \lq\lq striped wind\rq\rq\ model), the bulk speed of the flow is roughly constant and highly relativistic. The entire physics of pulsed emission is related to strong beaming effects and the presence of current sheets \citep{2002A&A...388L..29K} -- narrow regions in the wind, which are embedded in the cold, highly magnetized stripes of opposite magnetic polarity \citep{1990ApJ...349..538C}.
Radiating particles are thought to be energized during reconnection in the current layer and they are thought to have relativistic thermal distribution that satisfies a Harris equilibrium \citep{2005ApJ...627L..37P}. However, recent simulations of relativistic reconnection in plasmas with high magnetization, $\sigma\gg1$ (where $\sigma$ is the ratio of Poynting flux to bulk kinetic-energy flux), suggest that the energized particles are accelerated to a power-law distribution
\citep{2014ApJ...783L..21S, 2014arXiv1409.8262W}.
Moreover, it has been argued that in highly magnetized, relativistic systems shearing of magnetic field lines leads to formation of nearly force-free current layers rather than hot plasma sheets,
and therefore a force-free equilibrium is a more appropriate model than a Harris equilibrium \citep[and references therein]{2014PhRvL.113o5005G}.
In pulsar electrodynamics, the formation of a current layer can start close to the neutron star surface, along the last open field lines, and extend far outside the light-cylinder into the striped wind region.
We describe a current layer as approximately force-free up to the light cylinder, beyond which reconnection sets in and proceeds over many wavelengths in the wind as it propagates. Since reconnection results in energy dissipation and acceleration of particles, the energy density in the particle content increases and, together with the reconnecting field in a sheet, it maintains the pressure balance against the fields outside the layer.
This approach can be compared with the FIDO model (force-free inside -- dissipative outside the light cylinder) of \citet{2014ApJ...793...97K}, who, however, attribute the energy dissipation to a changing plasma conductivity. They have demonstrated that the model is able to reproduce the variety of Fermi-LAT pulsar lightcurves.
We focus on the spectra of gamma-ray pulsars.
In our model the spectral shape is an imprint of the acceleration regime, which can be
determined either by the radiative cooling of particles or by the size of the acceleration region (current sheet thickness).
We show that the wind scenario can be distinguished from other competing models by observations of its unique signature: synchrotron self-Compton (SSC) emission extending up to tens of TeV in the spectrum of the Crab pulsar.
In section~\ref{sec:Model}, we describe the basic picture of our wind model, including the magnetic configuration and the particle distribution function. Next, as typical examples of two acceleration regimes we investigate the emission properties of the winds from the Crab pulsar and the Vela pulsar, as explained in section~\ref{sec:Emission}. Conclusions are drawn in section \ref{sec:Conclusion}.
\section{The model}
\label{sec:Model}
\subsection{Magnetic field structure}
The wind description is based on the striped wind model of \citet{2013MNRAS.434.2636P},
valid beyond the light cylinder. In this model the wind (fluid) expands with a {constant} relativistic Lorentz factor $\Gamma=(1-\beta^2)^{-1/2}$ in the radial direction.
In the wind comoving frame the electric field vanishes. We formulate the relevant physics
in this frame and all the quantities in this frame are denoted with a prime, while non primed quantities should be understood as expressed in the lab frame.
In the fluid comoving frame the magnitude of the magnetic field is
\begin{align}
B'^2&=\beta^2B_{\rm L}^2\frac{r_{\rm L}^2}{r^2}\left(\frac{\sin^2\theta}{\Gamma^2}+\beta^2\frac{r_{\rm L}^2}{r^2}\right)\tanh^2(\psi_{\rm s}/\Delta) \nonumber \\
&=B'^2_{\rm max}\tanh^2(\psi_{\rm s}/\Delta)
\label{stripecomp}
\end{align}
where $\psi_{\rm s}=\cos\theta\cos\chi + \sin\theta\sin\chi\cos\left[\phi-\omega(t-r/\beta c)\right]$ describes the location of the current sheet of the wind, with $\chi$ being the angle between magnetic and rotational axes (obliquity), $\omega=2\pi/P$ being the pulsar frequency, $P$ -- the pulsar period, and $r_{\rm L}=c/\omega$ the light-cylinder radius. {The current sheet thickness is parametrised by $\lambda \, \Delta$, where $\lambda= 2\pi \beta\,r_{\rm L}$ is the striped wind wavelength and $\Delta\in[0;0.5]$.}
The ordered field exists mostly in stripes, vanishing in the middle of a current sheet.
We assume, however, that the current sheet is magnetized such that there is an additional field component, generated in the course of reconnection, which can be either shearing or random, and which we call \lq\lq turbulent\rq\rq. {There are two reasons for invoking this component: (1) a nonvanishing synchrotron emissivity in the middle of the current sheet (where the ordered component vanishes), (2) the turbulent nature of the field seems to be required to explain the decrease in linear polarization degree of the Crab optical pulse} \citep{2013MNRAS.434.2636P}.
We assume that only a fraction $\varepsilon_{\rm d}$ of the available magnetic energy is dissipated during reconnection
and most of the field survives as a turbulent field.
The simplest model for this turbulent component proves to be
$B'_{\rm tur}=b_0 B'_{\rm max}/\cosh(\psi_{\rm s}/\Delta)$,
{where $b_0=(1-\varepsilon_{\rm d})^{1/2}$.}
This choice ensures that in the case of no dissipation ($b_0=1$) a turbulent field provides a pressure balance between a current sheet and the field outside.
The energy dissipation and particle energization may be interpreted as a change in the fraction of plasma and field contributions to the pressure balance \citep{2009PhRvL.102m5003H}, which, in the wind comoving frame, takes the form
\eqb
\frac{B'^2_{\rm max}}{8\pi}=\frac{B'^2}{8\pi}+p+\frac{B'^2_{\rm tur}}{8\pi} \label{pbalance}
\eqe
The pressure of ultrarelativistic particles can be approximated by $p\approx u'/3$, where $u'=mc^2 \int \gamma n'(\gamma) d\gamma$ is the energy density and $n'(\gamma)$ is the distribution function.
\subsection{Particle distribution function}
So far the current sheet in a striped wind has been modelled by assuming a thermal particle distribution with a temperature determined by the synchrotron cooling \citep{2012MNRAS.424.2023P}. However, simulations suggest that during reconnection particles are energized to a broader power-law distribution.
Power law indices $s\approx1$ are attributed to the acceleration at the primary X-point, whereas indices up to $s\approx3$ result from acceleration in many X-points \citep{2004PhPl...11.1151J}. These spectral features depend also on the plasma magnetization $\sigma$ \citep{2014PhRvL.113o5005G, 2014arXiv1409.8262W}, such that for the highest values of magnetization the spectral index tends to one, $s\rightarrow1$. Radiative cooling has not been probed in simulations, but we argue that in this case a cooled distribution may be observed if the acceleration and cooling timescales are shorter than the timescale of particle injection into the acceleration process.
We suggest that depending on the pulsar period $P$ and spin down power $\dot{E}$,
particle acceleration at reconnection sites in the wind proceeds in two different regimes.
In the first -- cooling regime -- the particle spectrum is modelled as a power law with an exponential cut off:
\eqb
n'(\gamma)=n'_0\gamma^{-s}e^{-\gamma/\gamma'_{\rm rad}}
\label{distrrad}
\eqe
$\gamma'_{\rm rad}$ is the particle Lorentz factor, for which the synchrotron cooling timescale in the total magnetic field $B'_{\rm tot}=(B'^2+B_{\rm tur}'^2)^{1/2}$ is equal to the acceleration timescale in the reconnection electric field:
\begin{align}
\gamma'_{\rm rad}&=\left(\frac{6\pi e\tau}{\sigma_{\rm T}B'_{\rm tot}}\right)^{1/2} \\
&\approx3.4\times10^5 \Gamma_{1}^{3/2}P_{-2}^{1/2}\varepsilon_{\rm d,-2}^{1/2}\dot{E}_{38}^{-1/4} \left(\frac{1+\sin^2\chi}{\Gamma_1^2\hat{r}^{-2}_1+\sin^2\theta}\right)^{1/4}
\label{gammarad}
\end{align}
where we use $B_{\rm L}\approx(2\pi/P)(\dot{E}/c^3)^{1/2}(1+\sin^2\chi)^{-1/2}$ \citep{2006ApJ...648L..51S} and the notation $X=X_i\times10^i$ in CGS units.
{We have also assumed} that the reconnection electric field has a form $E'\approx \tau B'$, where $\tau\la1$ is a constant reconnection rate.
We estimate it by comparing the reconnection timescale $t'_{\rm rec}\sim\Gamma r_{\rm L}/(\tau c)$ with the wind expansion time $t'_{\rm exp}\sim r_{\rm diss}/(\varepsilon_{\rm d} \Gamma c)$ \citep{2013MNRAS.431..355G}, where $\epsilon_{\rm d}\ll 1$ determines the fraction of the available magnetic energy that is dissipated within the distance $r_{\rm diss}$. Thus, the reconnection rate can be related to the dissipation distance $\tau\approx\Gamma^2\varepsilon_{\rm d}r_{\rm L}/r_{\rm diss}$ (the dissipation distance {in units of the light cylinder radius, $\hat{r}=r_{\rm diss}/r_{\rm L}$,} is a parameter in our model, and we obtain it by fitting of the computed synchrotron flux to the pulsar spectrum, see Sect.~\ref{sec:fitting}).
For usually invoked reconnection rates ($\tau\sim0.01-0.2$) the full dissipation ($\varepsilon_{\rm d}\approx1$) would proceed over many wavelengths in the wind, possibly accelerating it \citep{2001ApJ...547..437L}. The assumption $\varepsilon_{\rm d}\ll1$ allows us to neglect the wind acceleration.
In the second regime, acceleration of particles is limited by the current sheet size, thus they escape the acceleration region before they reach the cooling regime. {This case has been investigated by} \citet{2014arXiv1409.8262W}, {who observe in their simulations formation of a power-law particle distribution with a super-exponential cut-off}
\eqb
n'(\gamma)=n'_0\gamma^{-s}e^{-\gamma^2/\gamma'^2_{\rm sl}},
\label{distrsl}
\eqe
where $\gamma'_{\rm sl}$ is the Lorentz factor, for which particle gyroradius becomes equal to a fraction $\zeta$ of {the system size.
The simulations have not been initialized with a force-free configuration, which is of interest here, and they
do not include radiative losses, nevertheless we expect that when the escape effects are dominant, a generic type of particle distribution (\ref{distrsl}) is formed, which we adopt here.
In our case, however, the current sheet contains disordered fields, and locally the acceleration lengthscale is comparable to the particle confinement scale, which we estimate as being roughly the current sheet thickness} $\Delta\lambda = \Delta 2\pir_{\rm L} \beta\approx\Delta 2\pir_{\rm L}$, thus
\begin{align}
\gamma'_{\rm sl}&= \frac{\Gamma\zeta(\Delta2\pir_{\rm L})eB'_{\rm tot}}{mc^2} \\
&\approx 2\times10^{8} \zeta_{-1} \Delta_{-1} \hat{r}_{1}^{-1} \dot{E}_{38}^{1/2} \left(\frac{\Gamma_1^2\hat{r}^{-2}_1+\sin^2\theta}{1+\sin^2\chi}\right)^{1/2}
\label{gammasl}
\end{align}
Here we have taken into account that in the lightcurves
the width of the pulses normalized to the period of the pulsar, if assumed to reflect the ratio of the sheet thickness to the wind wavelength, implies a parameter $\Delta\approx 0.1$ \citep{2005ApJ...627L..37P}. We keep it the same for each pulsar. However, a physically motivated definition of the current sheet thickness
is based on the regime in which particle acceleration proceeds. It depends solely on the combination of two pulsar observables: its spin down power and its period, since $\gamma'_{\rm rad}$ and $\gamma'_{\rm sl}$ become comparable when $\dot{E}_{38}^{3/2}/P_{-2}\sim0.002$ (where according to our fits discussed below $\Gamma_1\approx2.5$, $\hat{r}_1\approx3$).
For instance, for the Crab pulsar $\dot{E}_{38}^{3/2}/P_{-2}\approx3$ and for the Vela pulsar $\dot{E}_{38}^{3/2}/P_{-2}\approx0.002$. They are typical examples of two different scenarios.
Synchrotron emission of an accelerated particle can be observed if its gyroradius is larger than the electron inertial length, a scale on which the reconnection takes place \citep{2004PhPl...11.1151J, 2013arXiv1307.7008T}.
This defines the low energy cut-off $\gamma'_{\rm c}$ in the particle distribution, such that in both eq.~(\ref{distrrad}) and eq.~(\ref{distrsl}) $\gamma>\gamma'_{\rm c}$. In a general form the cut-off is given by $\gamma'_{\rm c}=B'/(8\pi N'mc^2)^{1/2}$, where the particle density is
\eqb N'=\int_{\gamma'_{\rm c}}^{\infty} n'(\gamma)d\gamma
\label{totdens}
\eqe
and the parameter $n'_0$ in eq.~(\ref{distrrad}) and eq.~(\ref{distrsl}) is determined by the pressure balance, eq.~(\ref{pbalance}). Calculation of $\gamma'_{\rm c}$ can be simplified by approximating particle distributions, eq.~(\ref{distrrad}) and eq.~(\ref{distrsl}), by a power law with a sharp high energy cut-off at $\gamma'_{\rm rad}$ or $\gamma'_{\rm sl}$, respectively.
Using eq.~(\ref{totdens}) and the relation
$N'(\bmath{r})=\kappa N_{\rm GJ, LC}r_{\rm L}^2/\Gamma r^2$,
where $N_{\rm GJ, LC}=\omegaB_{\rm L}/2\pi e c$ is the Goldreich-Julian particle number density at the light cylinder, one can constrain {the number density parameter} $\kappa$ of radiating particles
$\kappa\approx (e c P \hat{r}/B')\int n'(\gamma)d\gamma$.
Note that this expression determines only a lower limit on the multiplicity, because the low energy cut-off $\gamma'_{\rm c}$ is an estimation of the lowest energy of radiating particles, whose emission is observed.
\section{Pulsed emission}
\label{sec:Emission}
\subsection{The gamma-ray spectra}
Synchrotron flux produced by a particle distribution $n'(\gamma)$ is given by
\eqb
\epsilon F_{\epsilon}^{\rm syn}=\int_{V}dV\frac{D^3}{4\pi d^2}\frac{\sqrt{3} e^3 B'_{\rm tot}}{h}\epsilon'\int_{\gamma'_{\rm c}}^{\infty} d\gamma' n(\gamma') R(z)
\label{synchrfunction}
\eqe
where $z=(2 \epsilon'B_{\rm cr})/(3B'_{\rm tot}\gamma'^2)$ and $B_{\rm cr}=m^2c^3/e\hbar=4.4\times10^{13}$ G is the quantum critical field,
\eqb
R(z)=z\int_z^{\infty} d\xi K_{5/3}(\xi)\approx1.78 z^{0.297} e^{-z}.
\eqe
$\epsilon$ and $\epsilon'$ is the photon energy normalized to the electron rest mass in the lab and in the comoving frame, respectively; $d$ is the distance between the observer and the pulsar, and $D$ is the Doppler factor of a wind (for details see \citet{2005ApJ...627L..37P}).
Introducing $y=\gamma'/\gamma'_{\rm rad}$ and using the particle spectrum (\ref{distrrad}),
we obtain
\eqb
\epsilon F_{\epsilon}^{\rm syn} \propto x^{1.3}\int_{\gamma'_{\rm c}/\gamma'_{\rm rad}}^{\infty} dy e^{-(s+0.6)\ln y-y-x/y^2},
\eqe
where $x=(2 \epsilon'B_{\rm cr})/(3B'_{\rm tot}\gamma'^2_{\rm rad})$.
Asymptotic behaviour of this integral can be calculated by the steepest descent method.
We find an extremum $y_0$ of the integrand from the condition of the vanishing of its first derivative. In the exponential tail $\gamma'\gg\gamma'_{\rm rad}$, the solution can be found analytically $y_0\approx(2x)^{1/3}$. Next, we expand the integrand to the second order around $y_0$. Assuming $\gamma'_{\rm c}\ll\gamma'_{\rm rad}$, we integrate a resulting Gaussian function. An asymptotics takes the form
\eqb
\epsilon F_{\epsilon}^{\rm syn} \propto x^{1.3-(s+0.6)/3}e^{-1.9\,x^{1/3}} \label{asympexp}
\eqe
Note that this is very similar to the spectrum $\propto \exp(x^{0.35})$ obtained by \citet{2013A&A...550A.101A} by a fit.
In a similar way we calculate an asymptote of a spectrum produced by the particle distribution (\ref{distrsl}). In this case the exponent has a form $f(y)=-(s+0.6)\ln y-y^2-x/y^2$ and its derivative has a maximum at $y_0\approx x^{1/4}$.
The flux has the following asymptotic behaviour:
\eqb
\epsilon F_{\epsilon}^{\rm syn}\propto x^{1.3-(s+0.6)/4}e^{-2\,x^{1/2}} \label{asympsuperexp}
\eqe
The asymptotes (\ref{asympexp}) and (\ref{asympsuperexp}), together with fitted spectra (see next section) are shown in Fig.~\ref{fig1} and Fig.~\ref{fig2}, respectively.
\subsection{Constraining the model parameters}
\label{sec:fitting}
\figureone
\figuretwo
There are four free parameters of the model: the Lorentz factor of the wind $\Gamma$, the dissipation distance in units of the light cylinder $\hat{r}$, the particle index $s$ and the dissipation efficiency $\varepsilon_{\rm d}$.
We constrain them by matching the synchrotron flux (\ref{synchrfunction}) to the phase-averaged spectra.
To compute the fluxes in our model we integrate the wind emissivity over 3D volume (see eq.~(\ref{synchrfunction})) using spectral methods \citep{1986nras.book.....P}. According to this approach,
the integrand is expanded in a basis of orthogonal polynomials
and integrated term by term.
As bases we choose the same special functions as proposed by \citet{2013MNRAS.434.2636P}, i.e., Chebyshev polynomials in coordinate $r$, Chebyshev polynomials in $\cos\theta$, and a Fourier series in $\phi$.
This choices allow us to use a fast cosine transform to calculate the coefficients of expansion in the polynomial bases, and to apply directly the Clenshaw-Curtis quadrature rules for computing integrals (Press et al. 1986).
The model spectrum is calculated
at the peak of the lightcurve, and we obtain the phase averaged flux
using $F_{\rm av} = F_0 I$, where $F_0$ is the flux at the peak maximum,
and $I$ is the integral area under the lightcurve (assuming
that the emission peaks are normalized to 1). This approach is justified by the fact that the model spectra change very weakly with the pulsar phase (due to a weak angle dependence of the Doppler factor when the emission is boosted to the observer frame).
In Fig.~\ref{fig1} we present several fits to the Crab data from Fermi-LAT and MAGIC.
{There is no unique solution as long as the measurements in the TeV range are missing. Computed spectra result from an interplay between all four parameters of the model, which must be found simultaneously in order to give a reasonable fit.
The green, blue and orange curves show the spectra with $s=2.2$ and different $\varepsilon_{\rm d}$, for which we have found the best-fit $\Gamma$ and $\hat{r}$.
In general, smaller $\varepsilon_{\rm d}$ lead to smaller $\hat{r}$ and larger $\Gamma$, thus, by increasing $\Gamma$ and decreasing $\epsilon_{\rm d}$ we are able to fit the data with the model of emission close to the pulsar, at $\hat{r}=2$ (orange curve). This fit gives an estimate of the maximum possible Lorentz factor of the wind, on the order of $\Gamma\sim100$, similar to the value assumed in previous studies} \citep{2012ApJ...745..108A,2013ApJ...771...53M}. {Note that for the best fits with the same value of $s$ the brightness of the SSC component does not change significantly, but the maximum energy of SSC photons does. It is estimated by the Klein-Nishina limit, boosted to the observer frame:} \eqb \epsilon\approx 2\Gamma \gamma'_{\rm rad}mc^2\approx3.6\, \Gamma_{1}^{5/2}P_{-2}^{1/2}\varepsilon_{\rm d,-2}^{1/2}\dot{E}_{38}^{-1/4} {\rm TeV}. \eqe
{Red, blue and pink curves show the fits for $\varepsilon_{\rm d}=0.01$ and different $s$, for which, as before, we have also found the best-fit $\Gamma$ and $\hat{r}$. In this case, the larger is $s$, the smaller are also $\hat{r}$ and $\Gamma$. The brightness of the SSC component changes significantly, being greater for larger $s$. With the currently available VHE data the upper limit on the particle index can be estimated to be $s\approx2.4$ (red curve).}
In general, soft indices $s\ga2$ are required to reproduce the gamma-ray bump in the Crab spectrum.
The low energy cut-off in the particle distribution is estimated to be roughly $\gamma'_{\rm c}\sim50$.
A fit to the Vela spectrum requires a much harder index $s\approx1.2$ (Fig.~\ref{fig2}). In this case the SSC component is weaker by several orders of magnitude (not shown in the plot), because the energy density resides mainly in the highest energy particles, which do not produce enough of low energy synchrotron photons that can be upscattered to very high energies (VHE).
The low energy cut-off in the particle distribution is roughly $\gamma'_{\rm c}\approx2\times10^3$.
{The number density $\kappa$ of the radiating particles and the reconnection rate $\tau$ for both pulsars are compared in Table~\ref{tableone} for two (exemplary) best-fit sets of parameters: for the efficiency $\varepsilon_{\rm d}=0.01$ and for the dissipation distance $\hat{r}=2$. In all cases our fits suggest that the number density of radiating particles is much smaller in the Vela wind than in the Crab wind, but the reconnection rate is higher. According to simulations} \citep{2013ApJ...775...50T, 2014PhRvL.113o5005G} this result suggests that the Vela wind is more magnetized.
\begin{table}
\begin{tabular}{c|c|c || c|c}
& Crab & Vela & Crab & Vela \\ \hline
$\varepsilon_{\rm d}$ & {\bf 0.01} & {\bf 0.01} & $\varepsilon_{\rm d}=3\times10^{-7}$ & $\varepsilon_{\rm d}=10^{-5}$\\
$\Gamma$ & 23 & 18 & 82 & 40 \\
$\hat{r}$ & 36 & 12 & {\bf 2} & {\bf 2} \\
$\tau$ & 0.18 & 0.26 & 0.001 & 0.01 \\
$\kappa$ & $7\times10^4$ & 14 & 0.5 & 0.05
\end{tabular}
\caption{Reconnection rate $\tau$ and the number density $\kappa$ of radiating particles obtained for two exemplary fits to the spectra.}
\label{tableone}
\end{table}
\section{Conclusions}
\label{sec:Conclusion}
The spectra of gamma-ray pulsars give insight into the physics of relativistic reconnection in pulsar winds.
Particle acceleration at the reconnection sites can proceed either in the radiative cooling regime (in the most powerful pulsars with $\dot{E}_{38}^{3/2}/P_{-2}\ga 0.002$), or in the size limited regime if the pulsar is less energetic.
The synchrotron spectra behave asymptotically like $\propto\exp(\epsilon^{1/3})$
and $\propto\exp(\epsilon^{1/2})$, respectively.
In the Crab spectrum, a new SSC component at tens of TeV is expected to be
{observed by} the upcoming experiment CTA \citep{emma}.
The MAGIC Collaboration \citep{crabtev} has recently obtained new measurements of the Crab pulsar at $\sim 2$ TeV. A comparison of our results with the spectrum in the VHE range will give a constraint on the particle index $s$ of the population accelerated during reconnection in the wind.
We put {an upper limit on the Lorentz factor of the Crab wind $\Gamma\la100$ and of the Vela wind $\Gamma\la50$, which are obtained if the emission occurs at $r_{\rm diss}=2r_{\rm L}$.} These values are in good agreement with previous works on optical polarization signatures \citep{2005ApJ...627L..37P} and Crab flares \citep{2013MNRAS.436L..20B}. The {number density of radiating particles in the wind depends on the dissipation distance, but in each case our fits to the spectra imply that it is smaller in the Vela wind than in the Crab wind. The estimated reconnection rate is larger for Vela.}
Our study implies that the current sheet structure is imprinted on the pulsar gamma-ray lightcurves. Future modelling of the pulse profile change with energy, together with the phase-resolved spectroscopy, will give much more insight into the reconnection dynamics in pulsar winds.
\section*{Acknowledgments}
We are grateful to John Kirk and to the anonymous referee for helpful comments on the manuscript. We also thank Alice Harding for drawing our attention to the newest results of MAGIC. This work has been supported by the French National Research Agency (ANR) through the grant No. ANR-13-JS05-0003-01 (project EMPERE). We also benefited from the computational facilities available at Equip@Meso (Universit\'e de Strasbourg).
\bibliographystyle{mn2e}
|
\section{Introduction}
\label{sec:introduction}
Let $A\subset \R$ be a finite set.
Define the \textit{sumset}, and respectively the \textit{product set}, by
$$A+A:=\{a+b:a,b\in{A}\}$$
and
$$AA:=\{ab:a,b\in{A}\}.$$
The Erd\H{o}s-Szemer\'{e}di \cite{ES} conjecture states, for all $\epsilon>0$,
$$\max{\{|A+A|,|AA|\}}\gg{|A|^{2-\epsilon}} \,.$$
Loosely speaking, the conjecture says that any set of reals (or integers) cannot be highly structured both in a multiplicative and additive sense.
The best result in the direction is due to Solymosi \cite{soly}.
\begin{theorem}
Let $A\subset \R$ be a set.
Then
$$
\max\{ |A+A|, |AA| \} \gg |A|^{\frac{4}{3}} \log^{-\frac{1}{3}} |A| \,.
$$
\label{t:Solymosi_intr}
\end{theorem}
If one consider the set
$$
AA+A = \{ ab+c ~:~ a,b,c\in A \}
$$
then the Erd\H{o}s-Szemer\'{e}di conjecture implies that $AA+A$ has size at least $|A|^{2-\eps}$
(we assume for simplicity that $1\in A$).
In \cite{balog} Balog
formulated a weaker hypothesis that
for all $\eps>0$ one has
$$
|AA+A| \gg |A|^{2-\eps} \,.
$$
In the paper he proved the following result, which implies, in particular,
$|A A+A|\gg |A|^{3/2}$ and $|A A+A A|\gg |A||A/A|^{1/2}$.
\begin{theorem}
For every finite sets of reals $A,B,C,D \subset \R$, we have
\begin{equation}\label{f:Balog1}
|A C+ A||B C+ B|\gg |A||B||C| \,,
\end{equation}
and
\begin{equation}\label{f:Balog2}
|A C+ A D||B C+ B D|\gg |B/A||C||D| \,.
\end{equation}
More precisely, see \cite{SS1}
$$|(A\times B) \cdot \D(C)+ A\times B|\gg |A||B||C|\,$$
and
$$|(A\times B) \cdot \D(C)+ (A\times B) \cdot \D(D)|\gg |B/A||C||D|\,,$$
where
$$
\D (A) := \{ (a,a) ~:~ a \in A \} \,.
$$
\label{t:Balog}
\end{theorem}
In \cite{B_RN_S} the authors have obtained a partial answer to a "dual"\, question
on the size of $A(A+A)$.
The main result of the paper is the following new bound for $A:A+A$, $AA+A$, more precisely, see Theorem \ref{t:main_balog} below.
\begin{theorem}
Let $A\subset \R$ be a set.
Then there is $\eps_1>0$ such that
\begin{equation}\label{f:main_balog1_intr}
| A:A + A | \gg |A|^{3/2+\eps_1} \,.
\end{equation}
Moreover, there is $\eps_2>0$ with
\begin{equation}\label{f:main_balog2_intr}
|AA + A| \gg |A|^{3/2+\eps_2} \,,
\end{equation}
provided by $|A:A| \ll |AA|$.
\label{t:main_balog_intr}
\end{theorem}
Also, we prove several results on the cardinality of $AA+AA$ and $A:A + A:A$,
see Theorem \ref{t:AA+AA} and Proposition \ref{p:AA+AA_Solymosi} below.
In paper \cite{RN_Z} Roche--Newton and Zhelezov conjectured there exist absolute constants $c$, $c'$ such that for any finite $A\subset \C$
the following holds
$$
\left| \frac{A+A}{A+A} \right| \le c|A|^2 \implies |A+A| \le c' |A| \,.
$$
Similar conjectures were made for the sets $\frac{A-A}{A-A}$, $(A-A)(A-A)$, $A(A+A+A+A)$ and so on.
We finish the paper giving a partial answer to a variant of the conjecture of Roche--Newton and Zhelezov
$$
|(A+A)(A+A) + (A+A)(A+A)| \ll |A|^2 \implies |A \pm A| \ll |A| \log^{} |A| \,,
$$
see Corollary \ref{c:RN_Z}.
The main idea of the proof is
the following.
We need to estimate from below the sumset of two sets $A$ and $A:A$, say.
As in many problems of the type usual applications of Szemer\'{e}di--Trotter's theorem \cite{TV} or Solymosi's method \cite{balog} give us a lower bound of the form $|A:A+A| \gg |A|^{3/2}$.
In paper \cite{ss2} the exponent $3/2$ was improved in the particular case of sumsets of convex sets.
After that the method was developed by several authors, see e.g. \cite{KR,Li,Li2,schoen_E_3,SS1,Sh_ineq,s_mixed,s_sumsets} and others.
In \cite{s_sumsets} the author proved that the bound $|A+B| \gg |A|^{3/2+c}$, $c>0$ takes place for wide class of {\it different} sets, having roughly comparable sizes.
For example, such bound holds if $A$ and $B$ have small multiplicative doubling.
It turns out that if (\ref{f:main_balog1_intr}) cannot be improved then there is some large set $C$ such that $|AC| \ll |A|$.
This allows us to apply results from \cite{s_sumsets}.
The author is grateful to Tomasz Schoen for useful discussions.
\section{Notation}
\label{sec:definitions}
Let $\Gr$ be an abelian group and $+$ be the group operation.
In the paper we use the same letter to denote a set $S\subseteq \Gr$
and its characteristic function $S:\Gr \rightarrow \{0,1\}.$
By $|S|$ denote the cardinality of $S$.
Let $f,g : \Gr \to \C$ be two functions with finite supports.
Put
\begin{equation}\label{f:convolutions}
(f*g) (x) := \sum_{y\in \Gr} f(y) g(x-y) \quad \mbox{ and } \quad
(f\circ g) (x) := \sum_{y\in \Gr} f(y) g(y+x) \,.
\end{equation}
Let $A \subseteq \Gr$ be a set.
For any real $\a>0$ put
\begin{equation}\label{f:E_k_preliminalies_B}
\E^{+}_\a (A)=\sum_{x\in \Gr} (A \c A)^{\a} (x)
\end{equation}
be {\it the higher energy} of $A$.
In particular case $k=2$
we write
$\E^{+} (A) = \E^{+}_2 (A)$ and $\E(A,B)$ for $\sum_{x\in \Gr} (A \c A) (x) (B \c B) (x)$.
The quantity $\E^{+} (A)$ is called {\it the additive energy} of a set, see e.g. \cite{TV}.
For a sequence $s=(s_1,\dots, s_{k-1})$ put
$A^{+}_s = A \cap (A-s_1)\dots \cap (A-s_{k-1}).$
Then
$$
\E^{+}_k (A) = \sum_{s_1,\dots,s_{k-1} \in \Gr} |A^{+}_s|^2 \,.
$$
If we
have a
group $\Gr$ with a multiplication instead of addition then we use symbol $\E^{\times}_\a (A)$ for the correspondent energy of a set $A$ and write $A^{\times}_s$ for $A \cap (As^{-1}_1)\dots \cap (As^{-1}_{k-1}).$
In the case of a unique operation we write just $\E_k (A)$, $\E(A)$ and $A_s$.
Let $A,B\subseteq \Gr$ be two finite sets. {\it The magnification ratio}
$R_B [A]$ of the pair $(A,B)$ (see e.g. \cite{TV}) is defined by
\begin{equation}\label{f:R_B[A]}
R_B [A] = \min_{\emptyset \neq Z \subseteq A} \frac{|B+Z|}{|Z|} \,.
\end{equation}
A beautiful result on magnification ratio was proven by Petridis \cite{p}.
\begin{theorem}
For any $A,B,C \subseteq \Gr$, we have
\begin{equation}\label{f:Petridis_C}
|B+C+X| \le R_{B} [A] \cdot |C+X| \,,
\end{equation}
where $X\subseteq A$ and $|B+X| = R_B [A] |X|$.
\label{t:Petridis_C}
\end{theorem}
\bigskip
We conclude the section by Ruzsa's triangle inequality, see e.g. \cite{TV}.
Interestingly, that our proof (developing some ideas of papers \cite{SS1}, \cite{B_RN_S})
does not require any mapping as usual.
\begin{lemma}
Let $A,B,C\subseteq \Gr$ be any sets.
Then
\begin{equation}\label{f:triangle_my}
|C| |A-B| \le |A\times B - \Delta (C)| \le |A-C| |B-C| \,.
\end{equation}
\label{l:triangle_my}
\end{lemma}
\begin{proof}
We have
$$
|A\times B - \Delta (C)| = \sum_{z\in A-B} |B \cap (A-z) - C| \ge |A-B| |C| \,.
$$
The inequality above is trivial and the identity follows by the projection of points $(x,y) \in A\times B - \Delta (C)$, $(x,y) = (a-c,b-c)$,
$a\in A$, $b\in B$, $c\in C$ onto $z:=x-y=a-b \in A-B$.
This concludes the proof.
$\hfill\Box$
\end{proof}
\bigskip
All logarithms are base $2.$ Signs $\ll$ and $\gg$ are the usual Vinogradov's symbols.
\section{Preliminaries}
\label{sec:preliminaries}
As we discussed in the introduction our proof uses some notions from \cite{s_sumsets}.
So, let us recall the main definition of the paper.
\begin{definition}
A set $A\subset \Gr$
has
{\bf SzT--type}
(in other words $A$ is called {\bf Szemer\'{e}di--Trotter set})
with parameter $\a \ge 1$
if for any set $B\subset \Gr$ and an arbitrary $\tau \ge 1$ one has
\begin{equation}\label{f:SzT-type}
|\{ x\in A+B ~:~ (A * B)(x) \ge \tau \}| \ll c (A) |B|^\a \cdot \tau^{-3} \,,
\end{equation}
where $c (A)>0$ is a constant depends on the set $A$ only.
\label{def:SzT-type}
\end{definition}
Simple calculations (or see \cite{s_sumsets}, Lemma 7) give us some connections between various energies of SzT--type sets.
Formula (\ref{f:Li}) below
is due to Li \cite{Li}.
\begin{lemma}
Suppose that $A,B,C\subseteq \Gr$ have SzT--type with the same parameter $\alpha$.
Then
\begin{equation}\label{f:Li}
\E^{3} (A) \ll \E^2_{3/2} (A) c(A) |A|^{\alpha} \,,
\end{equation}
\begin{equation}\label{f:Li_E}
\E (A) \ll c^{1/2} (A) |A|^{1+\a/2} \,,
\end{equation}
and
$$
\sum_x (A\c A)(x) (B\c B) (x) (C\c C) (x) \ll (c(A) c(B) c(C))^{1/3} (|A| |B| |C|)^{\alpha/3}
\times
$$
\begin{equation}\label{f:Li'}
\times \log (\min\{ |A|, |B|, |C| \}) \,.
\end{equation}
\label{l:Li}
\end{lemma}
We need in Lemma 27 from \cite{SS1}.
\begin{lemma}
Any set $A\subset \R$, $\R = (\R,+)$ has SzT--type with $\a=2$ and
$
c(A) = |A| d(A)
$,
where
\begin{equation}\label{f:d(A)}
d(A) := \min_{C \neq \emptyset} \frac{|AC|^2}{|A| |C|} \,.
\end{equation}
\label{l:d(A)}
\end{lemma}
So, any set with small multiplicative doubling or, more precisely, with small quantity (\ref{f:d(A)}) has SzT--type, relatively to addition, in an effective way.
It can be checked that minimum in (\ref{f:d(A)}) is
actually attained and we left the fact to an interested reader.
Careful analysis of our proof gives that we do not need this.
Another examples of SzT--types sets can be found in \cite{s_sumsets}.
\bigskip
Now let us prove a simple result on $d(A)$, which follows from Petridis's Theorem \ref{t:Petridis_C}.
\begin{lemma}
Let $A\subseteq \R^{+}$ be a set.
Then $d(A) = d(A^{-1})$ and
\begin{equation}\label{f:d(AA)}
d(AA) \le \frac{|A|^2 d^2 (A)}{|AA||C|} \,,\quad \quad d(A:A) \le \frac{|A|^2 d^2 (A)}{|A:A||C|} \,,
\end{equation}
where $C$ is a set
where
the minimum in (\ref{f:d(A)}) is attained.
\label{l:d(AA)}
\end{lemma}
\begin{proof}
The identity $d(A) = d(A^{-1})$ is obvious.
Let us prove (\ref{f:d(AA)}).
Suppose that the minimum in (\ref{f:d(A)}) is attained at $C$.
By Theorem \ref{t:Petridis_C} there is $X\subseteq C$
such that $|AAX| \le R |AX|$, where $R = R_A [C]$ is defined by formula (\ref{f:R_B[A]}).
We have
\begin{equation}\label{tmp:24.01.2015_1}
d(AA) \le \frac{|AAX|^2}{|AA||X|} \le R^2 \frac{|AX|^2}{|AA||X|} = \frac{|AX|^4}{|AA||X|^3}
\le
\frac{|AC|^4}{|AA||C|^3}
=
\frac{d^2 (A) |A|^2}{|AA| |C|} \,.
\end{equation}
Similarly, we take $C$ such that the correspondent minimum for $d(A^{-1})$ is attained at $C$.
Further, let $Y \subseteq C$ is given by Theorem \ref{t:Petridis_C}
and put $R=R_{A} [C^{-1}]$.
Then $|(A:A) Y| \le R |A^{-1} Y| \le R|A^{-1} C|$, $R= |AY^{-1}|/|Y| \le |A C^{-1}|/|C|$
and
arguments similar to (\ref{tmp:24.01.2015_1}) can be applied.
This completes the proof.
$\hfill\Box$
\end{proof}
\begin{remark}
Actually, the proof of Lemma \ref{l:d(AA)} gives us $d(AA) \le \frac{|AC|^4}{|AA||C|^3}$,
$d(A:A) \le \frac{|AC|^4}{|A:A||C|^3}$ for any nonempty $C$.
\label{r:l_d(AA)}
\end{remark}
\bigskip
Finally, we
formulate
a full version of Theorem \ref{t:Solymosi_intr} from the introduction.
\begin{theorem}
Let $A,B\subseteq \R$ be sets, $\tau>0$ be a real number.
Then
\begin{equation}\label{f:Solymosi}
|\{ x ~:~ |A \cap xB| \ge \tau \}| \ll \frac{|A+A| |B+B|}{\tau^2} \,.
\end{equation}
In particular
\begin{equation}\label{f:Solymosi_E}
\E^{\times} (A,B) \ll |A+A| |B+B| \cdot \log (\min\{|A|, |B| \}) \,.
\end{equation}
\label{t:Solymosi}
\end{theorem}
\section{The proof of the main results}
\label{sec:proof}
Our proof relies on a partial case of Theorem 14 from \cite{s_sumsets}.
\begin{theorem}
Suppose that $A,A_*\subset \R$ have SzT--type with the same parameter $\a=2$.
Then
$$
|A \pm A_*| \gg
\max\{
d (A_*)^{-\frac{1}{3}} d(A)^{-\frac{2}{9}} |A_*|^{\frac{8}{9}} |A|^{\frac{2}{3}},
d (A)^{-\frac{1}{3}} d(A_*)^{-\frac{2}{9}} |A|^{\frac{8}{9}} |A_*|^{\frac{2}{3}} \,,
$$
\begin{equation}\label{f:p_main_diff_1}
\min\{
d (A_*)^{-\frac{2}{27}} d(A)^{-\frac{13}{27}} |A_*|^{\frac{14}{9}},
d (A)^{-\frac{2}{27}} d(A_*)^{-\frac{13}{27}} |A|^{\frac{14}{9}} \} \}
\times
(\log (|A| |A_*|))^{-\frac{2}{9}} \,.
\end{equation}
\label{t:main_diff}
\end{theorem}
Now we can prove the main result of the paper.
\begin{theorem}
Let $A\subset \R$ be a finite set.
Then
\begin{equation}\label{f:main_balog1}
| A:A + A | \gg |A|^{\frac{3}{2}+\frac{1}{82}} \cdot (\log |A|)^{-\frac{2}{41}} \,,
\end{equation}
and
\begin{equation}\label{f:main_balog2}
|AA + A| \gg
|AA|^{\frac{11}{41}} |A:A|^{-\frac{11}{41}} |A|^{\frac{62}{41}} (\log |A|)^{-\frac{2}{41}}
\,.
\end{equation}
\label{t:main_balog}
\end{theorem}
\begin{proof}
Put $l=\log |A|$.
Without loosing of generality, we can assume that $0\notin A$.
Suppose that $|A:A + A| \ll M|A|^{3/2}$, $|AA + A| \ll M|A|^{3/2}$, where $M$ is a small power of $|A|$, that is $M=|A|^\eps$ and
obtain a contradiction.
Let us begin with
(\ref{f:main_balog1})
because the proof of the second inequality requires some additional steps.
Recall the arguments from \cite{balog} or see the proof of Theorem 31 from \cite{SS1}.
Let $l_i$ be the line
$y=q_ix.$ Thus, $(x,y)\in l_i\cap A^2$ if and only if $x\in A^\m_q.$
Let $q_1,\dots,q_n\in \Pi \subseteq A/A$ be such that $q_1<q_2<\dots<q_n$.
Here $\Pi$ is a set which can vary, in principle, and at the moment we choose $\Pi$ such that $|A^\m_{q_i}|\ge 2^{-1} |A|^2/|A/A|$ for all $q_i \in \Pi$.
Thus,
$\sum_{i\in \Pi} |A^\m_{q_i}|\ge \frac12|A|^2$.
We multiply all points of $A^2$ lying on the line $l_i$ by
$\D(A^{-1})$, so we obtain $|A^\m_{q_i}:A|$ points still belonging to the line $l_i$ and then we consider sumset of
the resulting set with $l_{i+1}\cap A^2.$
Clearly, we
get
$|A^\m_{q_i}:A||A^\m_{q_{i+1}}|$ points from the set $(A:A+A)^2$
lying between the lines $l_i$ and $l_{i+1}$.
Therefore, using the definition of the number $d(A)$, we have
\begin{equation}\label{tmp:21.01.2015_1}
M^2 |A|^3 \gg |A:A + A|^2 \ge \sum_{i=1}^{n-1} | A^\times_{q_i}||A^\times_{q_{i+1}}:A|
\ge
|A|^{1/2} d^{1/2} (A) \sum_{i=1}^{n-1} | A^\times_{q_i}|^{3/2}
\gg
\end{equation}
\begin{equation}\label{tmp:21.01.2015_2}
\gg
|A|^{3/2} d^{1/2} (A) |A:A|^{-1/2} \sum_{i=1}^{n-1} | A^\times_{q_i}|
\gg
|A|^{7/2} d^{1/2} (A) |A : A|^{-1/2} \,.
\end{equation}
Thus,
\begin{equation}\label{f:d(A)_recall}
d(A) = \min_{C \neq \emptyset} \frac{|AC|^2}{|A||C|}
\ll \frac{M^4|A/A|}{|A|} \,.
\end{equation}
To estimate $d(A:A)$, $d(AA)$ we
use
Lemma \ref{l:d(AA)}, see Remark \ref{r:l_d(AA)}.
In other words, our $C$ is some $A^\times_{q_i}$, where $q_i \in \Pi$.
After that applying the first inequality of Theorem \ref{t:main_diff} with $A=A$, $A_* = A:A$
, we obtain
$$
M |A|^{3/2}
\ge
|A:A + A| \gg |A:A|^{8/9} |A|^{2/3} d^{-2/9} (A) \left( \frac{|A|^2 d^2 (A)}{|A:A||C|} \right)^{-1/3} l^{-2/9}
=
$$
$$
=
|A:A|^{11/9} d^{-8/9} (A) |C|^{1/3} l^{-2/9}
\gg
|A|^{14/9} M^{-32/9} l^{-2/9} \,,
$$
and hence $M \gg l^{-2/41} |A|^{1/82}$.
This implies (\ref{f:main_balog1}).
It remains to prove (\ref{f:main_balog2}).
In the case we multiply all points of $A^2$ lying on the line $l_i$ by
$\D(A^{})$, so we obtain $|A A^\m_{q_i}|$ points still belonging to the line $l_i$ and then we consider sumset of
the resulting set with $l_{i+1}\cap A^2.$ Clearly, we obtain $|A A^\m_{q_i}||A^\m_{q_{i+1}}|$ points from the set $(AA+A)^2$.
Thus,
\begin{equation}\label{tmp:21.01.2015_1+}
M^2 |A|^3 \gg |AA + A|^2 \ge \sum_{i=1}^{n-1} | A^\times_{q_i}||AA^\times_{q_{i+1}}|
\end{equation}
and we repeat the arguments above.
The proof gives us
\begin{equation}\label{tmp:26.01.2015_1}
|AA+A| \gg |AA|^{11/41} |A|^{-4/41} (\E^\times_{3/2} (A))^{22/41} l^{-2/41} \,.
\end{equation}
Here we have chosen the set $\Pi$ as $\sum_{q\in \Pi} |A^\times_q|^{3/2} \gg \E^\times_{3/2} (A)$
or, in other words, $|A^\times_q| \gg (\E^\times_{3/2} (A))^2 |A|^{-4}$.
Using the H\"{o}lder inequality, combining with (\ref{tmp:26.01.2015_1}), we get
$$
|AA + A| \gg |AA|^{\frac{11}{41}} |A:A|^{-\frac{11}{41}} |A|^{\frac{62}{41}} l^{-2/41} \,.
$$
This completes the proof.
$\hfill\Box$
\end{proof}
\begin{remark}
Using the full power of Theorem 14 from \cite{s_sumsets} one can obtain further results connecting $|AA:A|$, $|A:AA|$ with $|AA + A|$, $|A:A + A|$ and so on. We do not make such calculations.
\end{remark}
The same method allows us to improve the result of Balog concerning the size of $AA+AA$ and $A:A+A:A$.
\begin{theorem}
Let $A\subset \R$ be a set.
Then
\begin{equation}\label{f:main_balog1'}
| A:A + A:A | \gg
|A:A|^{\frac{14}{29}} |A|^{\frac{30}{29}} (\log |A|)^{-\frac{2}{29}}
\,,
\end{equation}
and
\begin{equation}\label{f:main_balog2'}
|AA + AA| \gg
|AA|^{\frac{19}{29}} |A:A|^{-\frac{5}{29}} |A|^{\frac{30}{29}} (\log |A|)^{-\frac{2}{29}}
\,.
\end{equation}
\label{t:AA+AA}
\end{theorem}
\begin{proof}
As in the proof of Theorem \ref{t:main_balog}, we define $l_i$ to be the line $y=q_ix$
and
$q_1,\dots,q_n\in \Pi \subseteq A/A$ be such that $q_1<q_2<\dots<q_n$ and $|A^\m_{q_i}|\ge 2^{-1} |A|^2/|A/A|$
for any $q_i \in \Pi$.
Thus, $\sum_i |A^\m_{q_i}|\ge \frac12|A|^2$.
We multiply all points of $A^2$ lying on all lines $l_i$ by
$\D(A^{-1})$, so we obtain $|A^\m_{q_i}:A|$ points still belonging to the line $l_i$ and then we consider sumset of
the resulting set with itself.
Clearly, we
get
$|A^\m_{q_i}:A||A^\m_{q_{i+1}}:A|$ points from the set $(A:A+A:A)^2$
lying between the lines $l_i$ and $l_{i+1}$. Therefore, we have
$$
\sigma^2 :=
|A:A + A:A|^2 \ge \sum_{i=1}^{n-1} | A^\times_{q_i}:A||A^\times_{q_{i+1}}:A|
\ge
$$
\begin{equation}\label{tmp:22.01.2015_1}
\ge
d(A) |A| \sum_{i=1}^{n-1} | A^\times_{q_i}|^{1/2} | A^\times_{q_{i+1}}|^{1/2}
\gg
|A|^3 d(A) \,.
\end{equation}
It gives $d(A) \ll \sigma^2 |A|^{-3}$.
Using Theorem \ref{t:main_diff} with $A=A_* = A/A$, we obtain
\begin{equation}\label{tmp:27.01.2015_1}
\sigma \gg |A/A|^{14/9} \left( \frac{|A|^2 d^2 (A)}{|A/A||C|} \right)^{-5/9} l^{-2/9}
\gg
|A/A|^{19/9} |A|^{-10/9} |C|^{5/9} d^{-10/9} (A) l^{-2/9}
\gg
\end{equation}
\begin{equation}\label{tmp:27.01.2015_2}
\gg
|A/A|^{14/9} \sigma^{-20/9} |A|^{10/3} l^{-2/9} \,.
\end{equation}
After some calculations, we
get
$\sigma \gg |A/A|^{14/29} |A|^{30/29} l^{-2/29}$.
To obtain (\ref{f:main_balog2'}) we use the previous arguments
$$
\sigma^2 := |AA + AA|^2 \ge \sum_{i\in \Pi} | AA^\times_{q_i}| |AA^\times_{q_{i+1}}|
\ge
d(A) |A| \sum_{i\in \Pi} | A^\times_{q_i}|^{1/2} | A^\times_{q_{i+1}}|^{1/2}
\gg
$$
\begin{equation}\label{tmp:22.01.2015_1+}
\gg
d(A) |A| |\Pi| \Delta \,,
\end{equation}
choosing $\Pi\subseteq A/A$ such that for any $q\in \Pi$ one has
$|A|^2 / |A:A| \ll \Delta \le |A^{\times}_q|$.
Clearly, such set $\Pi$ exists by
simple average arguments.
The
calculations
as in (\ref{tmp:27.01.2015_1})---(\ref{tmp:27.01.2015_2}) give us
\begin{equation*}\label{tmp:27.01.2015_1'}
\sigma \gg |AA|^{14/9} \left( \frac{|A|^2 d^2 (A)}{|AA|\D} \right)^{-5/9} l^{-2/9}
\gg
|AA|^{19/9} |A|^{-10/9} \D^{5/9} d^{-10/9} (A) l^{-2/9}
\gg
\end{equation*}
\begin{equation*}\label{tmp:27.01.2015_2'}
\gg
|AA|^{19/9} \sigma^{-20/9} (|\Pi| \D^{3/2})^{10/9} l^{-2/9} \,.
\end{equation*}
After some computations, we
obtain
$$
\sigma \gg |AA|^{19/29} |A:A|^{-5/29} |A|^{30/29} l^{-2/29} \,.
$$
This concludes the proof.
$\hfill\Box$
\end{proof}
\bigskip
Finally, let us obtain a result on $AA+A$, $AA+AA$ of another type.
\begin{proposition}
Let $A\subset \R$ be a set.
Then
\begin{equation}\label{f:1}
|AA+A|^4 \,,~ |A:A+A|^4 \gg |A|^{-2} (\E^\times_{3/2} (A))^2 \E^{+}_3 (A) \log^{-1} |A| \,,
\end{equation}
and
\begin{equation}\label{f:2}
|AA+AA|^2 \,,~ |A:A+A:A|^2 \gg \E^{+}_3 (A) \log^{-1} |A| \,.
\end{equation}
Moreover
\begin{equation}\label{f:1'}
|AA+A|^4 \,,~ |A:A+A|^4 \gg \frac{|A|^{10}}{|A:A| |A-A|^2} \,,
\end{equation}
\begin{equation}\label{f:2'}
|AA+AA|^2 \,,~ |A:A+A:A|^2 \gg \frac{|A|^6}{|A-A|^2} \,.
\end{equation}
\label{p:AA+AA_Solymosi}
\end{proposition}
\begin{proof}
Put $l=\log |A|$.
Using Lemma \ref{l:Li}, we obtain
for any $A,B$ and $C$
\begin{equation}\label{tmp:27.01.2015_5}
\sum_x (A\c A)(x) (B\c B) (x) (C\c C) (x) \ll |A| |B| |C| (d(A) d(B) d(C))^{1/3} \log (|A||B||C|) \,.
\end{equation}
In particular case $A=B=C$ of formula above the definition of the number $d(A)$
gives us
\begin{equation}\label{tmp:22.01.2015_2}
|A A^\times_s|^2 \,,~ |A : A^\times_s|^2 \gg |A|^{-2} |A^{\times}_s| \E^{+}_3 (A) l^{-1}
\end{equation}
for any $s\in A:A$.
Applying (\ref{tmp:21.01.2015_1}), (\ref{tmp:21.01.2015_1+}) and the last bound, we obtain (\ref{f:1}).
Using (\ref{tmp:22.01.2015_2}) one more time and Katz--Koester inclusion \cite{kk}, namely,
\begin{equation}\label{f:kk_m}
AA^{\times}_s \subseteq AA \cap s AA \,, \quad A:A^{\times}_s \subseteq (A:A) \cap s^{-1} (A:A)
\end{equation}
as well as formula (\ref{f:Solymosi}) of Solymosi's result, we get (\ref{f:2}).
Another way to prove (\ref{f:2}) is just to use formulas (\ref{tmp:22.01.2015_1}), (\ref{tmp:22.01.2015_1+}),
combining with (\ref{tmp:22.01.2015_2}).
Inequalities
(\ref{f:1'}), (\ref{f:2'}) follow similar to (\ref{f:1}), (\ref{f:2})
from a direct application of Definition \ref{def:SzT-type} and the H\"{o}lder inequality.
This completes the proof.
$\hfill\Box$
\end{proof}
\begin{remark}
Applying arguments as in the proof (\ref{f:2})
as well as
formula (\ref{f:Li_E}) of Lemma \ref{l:Li},
we obtain a similar bound, namely,
$$\E^{+} (A) \ll |A| |AA+AA|$$
(actually, using methods from \cite{Sh_ineq} one can improve the
inequality).
It is interesting to compare this estimate with Solymosi's upper bound for the multiplicative energy (\ref{f:Solymosi_E}).
Using formula (\ref{f:Li}) of Lemma \ref{l:Li}, we have also
$$
(\E^{+} (A))^{3/2} \E^{\times}_{3/2} (A) \ll \E^{+}_{3/2} (A) |A| |AA+A|^2 \,.
$$
\label{r:E^+}
\end{remark}
\bigskip
Combining inequality (\ref{f:2}) with some estimates from \cite{Sh_energy}, we obtain a result in spirit of paper \cite{RN_Z}.
\begin{corollary}
Let $A\subset \R$ be a set.
Suppose that
\begin{equation}\label{cond:RN_Z}
|(A+A)(A+A) + (A+A)(A+A)| \ll |A|^2 \quad \mbox{ and } \quad \E^{+} (A) |A-A| \ll |A|^4 \,.
\end{equation}
Then
\begin{equation}\label{f:RN_Z}
|A-A| \ll |A| \log^{4/7} |A| \,.
\end{equation}
The same holds if one replace sum onto minus and product onto division into the first condition from (\ref{cond:RN_Z}).
If just the first condition of (\ref{cond:RN_Z}) holds (with plus) then
\begin{equation}\label{f:RN_Z'}
|A \pm A| \ll |A| \log^{} |A| \,,
\end{equation}
and if it holds with minus then
\begin{equation}\label{f:RN_Z''}
|A - A| \ll |A| \log^{} |A| \,,
\end{equation}
Again, one can replace product onto division in the first condition of (\ref{cond:RN_Z}).
\label{c:RN_Z}
\end{corollary}
\begin{proof}
Let us have deal with the
situation of the sum and the product.
Another cases can be considered similarly.
By Theorem 30 from \cite{Sh_energy} and our second condition, one has
$$
\E^{+}_3 (A\pm A) \ge |A|^{45/4} |A-A|^{-1/2} (\E^{+} (A))^{-9/4}
\gg
|A|^{9/4} |A-A|^{7/4} \,.
$$
On the other hand, using formula (\ref{f:2}) from Proposition \ref{p:AA+AA_Solymosi} and
our first condition, we get
$$
|A|^4 \log |A| \gg \E^{+}_3 (A\pm A) \gg |A|^{9/4} |A-A|^{7/4}
$$
as required.
Finally, using the additive variant of Katz--Koester inclusion (\ref{f:kk_m}) (or see Proposition 29 from \cite{Sh_energy}),
we obtain
$$
|A|^3 |A \pm A| \le \E^{+}_3 (A + A) \ll |A|^4 \log |A| \,,
$$
and
$$
|A|^3 |A - A| \le \E^{+}_3 (A - A) \ll |A|^4 \log |A| \,.
$$
This completes the proof.
$\hfill\Box$
\end{proof}
|
\section{Introduction}
In game theory, coordination games are used to model situations in which players are rewarded for agreeing on a common strategy, e.g., by deciding on a common technological or societal standard. In this paper we introduce and study a very simple class of coordination games, which we call \emph{coordination games on graphs}:
\begin{quote}
We are given a finite (undirected) graph, of which the nodes correspond to the players of the game. Each player chooses a color from a set of colors available to her. The payoff of a player is the number of neighbors who choose the same color.
\end{quote}
Our main motivation for studying these games is that they
constitute a natural class of strategic games that capture the following three key characteristics:
\begin{enumerate}
\item \emph{Join the crowd property} \cite{SA15}: the payoff of each player weakly increases when more players choose her strategy.
\item \emph{Asymmetric strategy sets}: players may have different strategy sets.
\item \emph{Local dependency}: the payoff of each player depends only on the choices made by certain groups of players (i.e., neighbors in the given graph).
\end{enumerate}
The above characteristics are inherent to many applications. As a concrete example, consider a situation in which several clients have to choose between multiple competing providers offering the same service (or product), such as peer-to-peer networks, social networks, photo sharing platforms, and mobile phone providers.
Here the benefit of a client for subscribing to a specific provider increases with the number of clients who opt for this provider. Also, each client typically cares only about the subscriptions of certain other clients (e.g., friends, relatives, etc.).
In coordination games on graphs it is beneficial for each player to align her choices with the ones of her neighbors. As a consequence, the players may attempt to increase their payoffs by coordinating their choices in groups (also called \emph{coalitions}). In our studies we therefore focus on equilibrium concepts that are resilient to deviations of groups; more specifically we study \emph{strong equilibria} \cite{Aumann59} and \emph{$k$-equilibria} (also known as \emph{$k$-strong equilibria}) of coordination games on graphs.
Recall that in a strong equilibrium no coalition of players can profitably deviate in the sense that every player of the coalition strictly improves her payoff. Similarly, in a $k$-equilibrium with $k \in \{1, \dots, n\}$, where $n$ is the number of players, no coalition of players of size at most $k$ can profitably deviate.
\paragraph{Our contributions.}
The focus of this paper is on the existence, inefficiency and computability of strong equilibria and $k$-equilibria of coordination games on graphs. Our main contributions are as follows:
\bigskip\noindent
\emph{1. Existence.}
We show that Nash equilibria and 2-equilibria always exist. On the other hand, $k$-equilibria for $k \geq 3$ do not need to exist.
We therefore derive a complete characterization of the values of $k$ for which $k$-equilibria exist in our games.
We also show that strong equilibria exist if only two colors are available.
Further, we identify several graph structural properties that guarantee the existence of strong equilibria: in particular they exist if the underlying graph is a pseudoforest\footnote{Recall that in a pseudoforest each connected component has at most one cycle.}, and when every pair of cycles in the graph is edge-disjoint. Also, they exist if the graph is \emph{color complete}, i.e., if for each available color $x$ the components of the subgraph induced by the nodes having color $x$ are complete. Moreover, existence of strong equilibria is guaranteed in case the coordination game is played on a \emph{color forest}, i.e., for every color, the subgraph induced by the players who can choose that color is a forest.
We also address the following question. Given a coordination game denote its \emph{transition value} as the value of $k$ for which a $k$-equilibrium exists but a $(k+1)$-equilibrium does not. The question then is to determine for which values of $k$ a game with transition value $k$ exists. We exhibit a game with transition value 4.
In all our proofs the existence of strong equilibria is established by showing a stronger result, namely that the game has the \emph{coalitional finite improvement property}, i.e., every sequence of profitable joint deviations is finite (see Section~\ref{sec:prelim} for a formal definition).
\bigskip\noindent
\emph{2. Inefficiency.}
We also study the \emph{inefficiency} of equilibria. In our context, the \emph{social welfare} of a joint strategy is defined as the sum of the payoffs of all players. The \emph{$k$-price of anarchy} \cite{AFM09} (resp. \emph{$k$-price of stability}) refers to the ratio between the social welfare of an optimal outcome and the minimum (resp.~maximum) social welfare of a $k$-equilibrium\footnote{The $k$-price of anarchy is also commonly known as the \emph{$k$-strong price of anarchy}.}.
We show that the price of anarchy is unbounded, independently of the underlying graph structure, and the strong price of anarchy is $2$. In general, for the $k$-price of anarchy with $k \in \{2, \dots, n-1\}$ we derive almost matching lower and upper bounds of $2\frac{n-1}{k-1} - 1$ and $2\frac{n-1}{k-1}$, respectively (given a coordination game that has a $k$-equilibrium).
We also prove that the strong price of stability is 1 for the cases that there are only two colors, or the graph is a pseudoforest or color forest.
Our results thus show that as the coalition size $k$ increases, the worst-case inefficiency of $k$-equilibria decreases from $\infty$ to $2$. In particular, we obtain a constant $k$-price of anarchy for $k = \Omega(n)$.
\bigskip\noindent
\emph{3. Complexity.}
We also address several computational complexity issues.
Given a coordination game, a joint strategy $s$, and a number $k$ as input, it is co-NP complete to determine whether $s$ is a $k$-equilibrium.
However, we show that this problem can be solved in polynomial time in case the graph is a color forest.
We also give polynomial time algorithms to compute strong equilibria for the cases of color forests, color complete graphs, and pseudoforests.
\paragraph{Related work.}
Our coordination games on graphs are related to various well-studied types of games. We outline some connections below.
First, coordination games on graphs are \emph{polymatrix games}.
Recall that a polymatrix game (see \cite{How72,Jan68}) is a finite
strategic game in which the payoff for each player is the sum of the
payoffs obtained from the individual games the player plays with each
other player separately.
Cai and Daskalakis \cite{CD11} considered a special class of
polymatrix games which they call \emph{coordination-only polymatrix
games}.
These games are identical to coordination games on graphs with edge weights.
They showed that pure Nash equilibria exist and that
finding one is PLS-complete. The proof of the latter result crucially
exploits that the edge weights can be negative. Note that negative
edge weights can be used to enforce that players anti-coordinate. Our
coordination games do not exhibit this characteristic and are
therefore different from theirs.
Second, our coordination games are related to \emph{additively
separable hedonic games (ASHG)} \cite{Baner,Bogo}, which were
originally proposed in a cooperative game theory setting. Here the
players are the nodes of an edge weighted graph and form coalitions.
The payoff of a node is defined as the total weight of all edges to
neighbors that are in the same coalition. If the edge weights are
symmetric, the corresponding ASHG is said to be \emph{symmetric}.
Recently, a lot of work focused on computational issues of these games
(see, e.g., \cite{ABS10,ABS11,Gair}). Aziz and Brandt \cite{Aziz}
studied the existence of strong equilibria in these games. The
PLS-hardness result established in \cite{Gair} does not carry over to
our coordination games because it makes use of negative edge weights,
which we do not allow in our model. Note also that in ASHGs every
player can choose to enter every coalition which is not necessarily
the case in our coordination games. Such restrictions can be imposed
by the use of negative edge weights (see also \cite{Gair}) and our
coordination games therefore constitute a special case of symmetric
ASHGs with \emph{arbitrary} edge weights.
Third, our coordination games on graphs are related to
\emph{congestion games} \cite{Ros73}. In particular, they are
isomorphic to a special case of congestion games with weakly
decreasing cost functions (assuming that each player wants to minimize
her cost). Rozenfeld and Tennenholtz \cite{RT06} derived a structural
characterization of strategy sets that ensure the existence of strong
equilibria in such games. By applying their characterization to our
(transformed) games one obtains that strong equilibria exist if the
underlying graph of the coordination game is a matching or complete
(both results also follow trivially from our studies). Bil\`o et
al.~\cite{BFFM11} studied congestion games where the players are
embedded in a (possibly directed) \emph{influence graph} (describing
how the players delay each other). They analyzed the existence and
inefficiency of pure Nash equilibria in these games. However, because
the delay functions are assumed to be linearly increasing in the
number of players, these games do not cover the games we study here.
Further, coordination games on graphs are special cases of the
\emph{social network games} introduced and analyzed in \cite{AS13} (if
one uses in them thresholds equal to 0). These are games associated
with a threshold model of a social network introduced in \cite{AM11}
which is based on weighted graphs with thresholds.
Coordination games are also related to the problem of clustering, where
the task is to partition the nodes of a graph in a meaningful manner.
If we view the strategies as possible cluster names, then a Nash
equilibrium of our coordination game on a graph corresponds to a
``satisfactory'' clustering of the underlying graph. Hoefer
\cite{Hoefer2007} studied clustering games that are also polymatrix
games based on graphs. Each player plays one of two possible base
games depending on whether the opponent is a neighbor in the given
graph or not. Another more recent approach to clustering through
game theory is by Feldman, Lewin-Eytan and Naor \cite{FLN15}. In this
paper both a fixed clustering of points lying in a metric space and a
correlation clustering (in which the distance is in [0,1] and
each point has a weight denoting its `influence') is viewed as a
strategic hedonic game. However, in both references each player has
the same set of strategies, so the resulting games are not comparable
with ours.
Strategic games that involve coloring of the vertices of a graph have
also been studied in the context of the vertex coloring problem.
These games are motivated by the question of finding the chromatic
number of a graph. As in our games, the players are nodes in a graph
that choose colors. However, the payoff function differs from the one
we consider here: it is $0$ if a neighbor chooses the same color and
it is the number of nodes that chose the same color otherwise.
Panagopoulou and Spirakis \cite{PS08} showed that an efficient local
search algorithm can be used to compute a good vertex coloring.
Escoffier, Gourv\`es and Monnot \cite{EGM12} extended this work by
analyzing socially optimal outcomes and strong equilibria.
Chatzigiannakis et al.~\cite{CKPS10} studied the vertex coloring
problem in a distributed setting and showed that under certain
restrictions a good coloring can be reached in polynomial time.
Strong and $k$-equilibria in strategic games on graphs were also
studied in Gourv\`es and Monnot \cite{GM09,GM10}. These games are
related to, respectively, the \texttt{MAX-CUT} and \texttt{MAX-$k$-CUT} problems. However, they do not satisfy the join the crowd
property, so, again, the results are not comparable with ours.
To summarize, in spite of these close connections, our coordination
games on graphs are different from all classes of games mentioned
above. Notably, this is due to the fact that our games combine the
three properties mentioned above, i.e., join the crowd, asymmetric
strategy sets and local dependencies modeled by means of an undirected
graph.
Research reported here was recently followed in two different
directions. In \cite{ASW15} and \cite{SW16} coordination games
on directed graphs were considered, while in \cite{rahn:schaefer:2015}
coordination games on weighted undirected graphs were analyzed.
Both setups lead to substantially different results that are
discussed in the final section.
Finally, \cite{FF15} studied the strong price of anarchy for a general
class of strategic games that, in particular, include as special cases
our games and the \texttt{MAX-CUT} games mentioned above.
As a final remark, let us mention that the coordination games on
graphs are examples of games on networks, a vast research area
surveyed in \cite{JZ12}.
\paragraph{Our techniques.}
Most of our existence results are derived through the application of one technical key lemma. This lemma relates the change in social welfare caused by a profitable deviation of a coalition to the size of a minimum feedback edge set of the subgraph induced by the coalition\footnote{Recall that a \emph{feedback edge set} is a set of edges whose removal makes the graph acyclic.}. This lemma holds for arbitrary graphs and provides a tight bound on the maximum decrease in social welfare caused by profitable deviations. Using it, we prove our existence results by means of a generalized ordinal potential function argument. In particular, this enables us to show that every sequence of profitable joint deviations is finite. Further, we use the generalized ordinal potential function to prove that the strong price of anarchy is 1 and that strong equilibria can be computed efficiently for certain graph classes.
The non-existence proof of $3$-equilibria is based on an instance whose graph essentially corresponds to the skeleton of an octahedron and whose strategy sets are set up in such a way that at most one facet of the octahedron can be unicolored. We then use the symmetry of this instance to prove our non-existence result.
The upper bound on the $k$-price of anarchy is derived through a combinatorial argument. We first fix an arbitrary coalition of size $k$ and relate the social welfare of a $k$-equilibrium to the social welfare of an optimum within this coalition. We then extrapolate this bound by summing over all coalitions of size at most $k$. We believe that this approach might also prove useful to analyze the $k$-price of anarchy in other contexts.
\section{Preliminaries}
\label{sec:prelim}
A \emph{strategic game} $\mathcal{G} := (N, (S_i)_{i \in N}, (p_i)_{i \in N})$ consists of a set $N := \{1, \dots, n\}$ of $n > 1$ players, a non-empty set $S_i$ of \emph{strategies}, and a \emph{payoff function} $p_i : S_1 \times \cdots \times S_n \rightarrow \mathbb{R}$ for each player $i \in N$.
We denote $S_1 \times \cdots \times S_n$ by $S$,
call each element $s \in S$
a \emph{joint strategy}, and abbreviate the sequence
$(s_{j})_{j \neq i}$ to $s_{-i}$. Occasionally we write $(s_i,
s_{-i})$ instead of $s$.
We call a non-empty subset $K := \{k_1, \ldots, k_m\}$ of $N$ a \emph{coalition}. Given a joint strategy $s$ we abbreviate the sequence $(s_{k_1}, \ldots, s_{k_m})$ of strategies to $s_K$ and $S_{k_1} \times \cdots \times S_{k_m}$ to $S_{K}$. We also write $(s_K, s_{-K})$ instead of $s$. If there is a strategy $x$ such that $s_i = x$ for all players $i \in K$, we also write $(x_K, s_{-K})$ for $s.$
Given two joint strategies $s'$ and $s$ and a coalition $K$, we say that $s'$ is a \emph{deviation of the players in $K$}
from $s$ if $K = \{i \in N \mid s_i \neq s_i'\}$.
We denote this by $s \betredge{K} s'$. If in addition $p_i(s') > p_i(s)$ holds for all $i \in K$, we say that the deviation $s'$ from $s$ is \emph{profitable}. Further, we say that the players in $K$ \emph{can profitably deviate from $s$} if there exists a profitable deviation of these players
from $s$.
Next, we call a joint strategy $s$ a \emph{k-equilibrium}, where $k \in \{1, \dots, n\}$, if no coalition of at most $k$ players can profitably deviate from $s$.
Using this definition, a \emph{Nash equilibrium} is a 1-equilibrium and a \emph{strong equilibrium} \cite{Aumann59} is an $n$-equilibrium.
Given a joint strategy $s$, we call the sum $\mathit{\textit{SW}}(s)=\sum_{i \in N} p_i(s)$ the \emph{social welfare} of $s$. When the social welfare of $s$ is maximal, we call $s$ a \emph{social optimum}.
Given a finite game that has a $k$-equilibrium, its \emph{$k$-price of anarchy (resp. stability)} is the ratio $\mathit{\textit{SW}}(s)/\mathit{\textit{SW}}(s')$, where $s$ is a social optimum and $s'$ is a $k$-equilibrium with the lowest (resp. highest) social welfare\footnote{In the case of division by zero, we define the outcome as $\infty$.}. The \emph{(strong) price of anarchy} refers to the $k$-price of anarchy with $k = 1$ ($k = n$). The \emph{(strong) price of stability} is defined analogously.
A \emph{coalitional improvement path}, in short a \emph{c-improvement
path}, is a maximal sequence $(s^1, s^2, \dots)$ of joint strategies
such that for every $k > 1$ there is a coalition $K$ such that $s^k$
is a profitable deviation of the players in $K$ from
$s^{k-1}$. Clearly, if a c-improvement path is finite, its last
element is a strong equilibrium. We say that $\mathcal{G}$ has the
\emph{finite c-improvement property} (\emph{c-FIP}) if every
c-improvement path is finite. So if $\mathcal{G}$ has the c-FIP, then
it has a strong equilibrium. Further, we say that the function $P: S
\rightarrow A$ (where $A$ is any set) is a \emph{generalized ordinal
c-potential} for $\m G$ if there exists a strict partial
ordering $\succ$ on the set $A$ such that if $s \betredge{K} s'$ is a profitable deviation, then $P(s') \succ
P(s)$. A generalized ordinal potential is also called a \emph{generalized strong potential}
\cite{Harks13,Holzman97}.
It is easy to see that if a finite game admits a generalized ordinal c-potential then the game has the c-FIP.
The converse also holds: a finite game that has the c-FIP admits a generalized ordinal c-potential. The latter fact is folklore; we give a self-contained proof in Appendix~\ref{app:gen-ord-potential}.
Note that in the definition of a profitable deviation of a coalition,
we insisted that all members of the coalition change their strategies.
This requirement is irrelevant for the definitions of the $k$-equilibrium and the c-FIP, but it makes some
arguments slightly simpler.
\section{Coordination games on graphs}
\label{sec:coloring}
We now introduce the games we are interested in. Throughout the paper, we fix a finite set of colors $M$ of size $m$, an undirected graph $G = (V,E)$ without self-loops, and a \emph{color assignment} $A$. The latter is a function that assigns to each node $i$ a non-empty set $A_i \subseteq M$.
A node $j \in V$ is a \emph{neighbor} of the node $i \in V$ if $\{i,j\} \in E$.
Let $N_i$ denote the set of all neighbors of node $i$.
We define a strategic game $\mathcal{G}(G, A)$ as follows:
\begin{itemize}
\item the players are identified with the nodes, i.e., $N = V$,
\item the set of strategies of player $i$ is
$A_i$,
\item the payoff function of player $i$ is $p_i(s) := |\{ j \in N_i \mid s_i = s_j \} |$.
\end{itemize}
\noindent So each node simultaneously chooses a color from the set available to her and the payoff to the node is the number of neighbors who chose the same color.
We call these games \emph{coordination games on graphs}, from now on just \emph{coordination games}.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}
\node [label={\footnotesize $\{ \underline{a}, c\}$}] (1) at (-4, 0) {$1$};
\node [label={\footnotesize $\{a, \underline{b}\}$}] (2) at (-2, 1) {$2$};
\node [label=below:{\footnotesize $\{a, \underline{b}\}$}] (3) at (-2, -1) {$3$};
\node [label={\footnotesize $\{\underline{b}, c\}$}] (4) at (0, 1) {$4$};
\node [label=below:{\footnotesize $\{\underline{b}, c\}$}] (5) at (0, -1) {$5$};
\node [label={\footnotesize $\{c, \underline{a}\}$}] (6) at (2, 1) {$6$};
\node [label=below:{\footnotesize $\{c, \underline{a}\}$}] (7) at (2,-1) {$7$};
\node [label={\footnotesize $\{b, \underline{a}\}$}] (8) at (4, 0) {$8$};
\draw (1) -- (2);
\draw (1) -- (3);
\draw[ultra thick] (2) -- (4);
\draw[ultra thick] (2) -- (5);
\draw[ultra thick] (3) -- (4);
\draw[ultra thick] (3) -- (5);
\draw[ultra thick] (4) -- (5);
\draw (4) -- (1);
\draw (4) -- (8);
\draw (5) -- (1);
\draw (5) -- (8);
\draw (6) -- (4);
\draw (6) -- (5);
\draw (7) -- (4);
\draw (7) -- (5);
\draw[ultra thick] (8) -- (6);
\draw[ultra thick] (8) -- (7);
\path[ultra thick] (1) edge [out=-75, in=-105] node {} (8);
\end{tikzpicture}
\caption{A graph with a color set assignment. The bold edges indicate pairs of players choosing the same color.}
\label{fig:non-existence}
\end{center}
\end{figure}
\begin{example}\label{exa:1}
Consider the graph and the color assignment depicted in
Figure~\ref{fig:non-existence}.
Take the joint strategy that consists of the underlined strategies. Then the payoffs are as follows:
\begin{itemize}
\item 1 for the nodes 1, 6, 7,
\item 2 for the nodes 2, 3,
\item 3 for the nodes 4, 5, 8.
\end{itemize}
It is easy to see that the above joint strategy is a Nash equilibrium.
However, it is not a strong equilibrium because the coalition $K =
\{1, 4, 5, 6, 7\}$ can profitably deviate by choosing color $c$.
\qed
\end{example}
We now recall some notation and introduce some terminology.
Let $G = (V, E)$ be a graph. Given a set of nodes $K$, we denote by $G[K]$
the subgraph of $G$ induced by $K$ and by $E[K]$ the set of edges in $E$ that
have both endpoints in $K$. So $G[K] = (K, E[K])$. Further,
$\delta(K)$ denotes the set of edges that have one node in $K$ and
the other node outside of $K$. Also, given a subgraph $C$ of $G$ we use $V(C)$ and $E(C)$ to refer to the set of nodes and the set of edges of $C$, respectively.
Furthermore, we define $\textit{SW}_K(s) := \sum_{i \in K} p_i(s)$. Given a joint strategy $s$ we denote by
$E^+_s$ the set of edges $\{i,j\} \in E$ such that $s_i = s_j$. We call
these edges \emph{unicolored in $s$.} (In Figure \ref{fig:non-existence}, these are the bold edges.) Note that $\textit{SW}(s) = 2|E_s^+|$.
Finally, we call a subgraph \emph{unicolored in $s$} if all its nodes
have the same color in $s$.
\section{Existence of strong equilibria}
\label{sec:existence}
We begin by studying the existence of strong equilibria and $k$-equilibria of coordination games. We first prove our key lemma and then show how it can be applied to derive several existence results.
\subsection{Key lemma}
Recall that an edge set $F \subseteq E$ is a \emph{feedback edge set} of the graph $G = (V, E)$ if the graph $(V, E \setminus F)$ is acyclic.
\begin{lemma}[Key lemma]
\label{lem:tau}
Suppose $s \betredge{K} s'$ is a profitable deviation. Let $F$ be a feedback edge set of $G[K]$. Denote $\textit{SW}(s') - \textit{SW}(s)$ by $\Delta \textit{SW}$ and for a coalition $L$ denote $\textit{SW}_L(s) - \textit{SW}_L(s)$ by $\Delta \textit{SW}_L$. Then
\begin{equation}\label{eqn:deltasw}
\Delta \textit{SW} = 2 (\Delta \textit{SW}_K - |E^+_{s'} \cap E[K]| + |E^+_s \cap E[K]|)
\end{equation}
and
\begin{equation}\label{eqn:deltasw2}
\Delta \textit{SW} > 2 (|F \cap E_s^+| - |F \cap E_{s'}^+|).
\end{equation}
\end{lemma}
\begin{proof}
Let $N_K$ denote the set of neighbors of nodes in $K$ that are not
in $K$. Abbreviate $\textit{SW}(s')-\textit{SW}(s)$ to $\Delta \textit{SW}$, and
analogously, for a coalition $L$, let $\Delta
\textit{SW}_{L}=\textit{SW}_{L}(s')-\textit{SW}_{L}(s)$. The change in the social welfare
can be written as
$$
\Delta \textit{SW} = \Delta \textit{SW}_K + \Delta \textit{SW}_{N_K} + \Delta \textit{SW}_{V \setminus (K \cup N_K)}.
$$
We have
$$
\textit{SW}_K(s) = 2 |E^+_s \cap E[K]| + |E^+_s \cap \delta(K)|
$$
and analogously for $s'$. Thus
\begin{align*}
\Delta \textit{SW}_K = 2(|E^+_{s'} \cap E[K]| - |E^+_s \cap E[K]|) + |E^+_{s'} \cap \delta(K)| - |E^+_s \cap \delta(K)|.
\end{align*}
It follows that
\begin{align*}
\Delta \textit{SW}_{N_K} &= |E^+_{s'} \cap \delta(K)| - |E^+_s \cap \delta(K)| \\
&= \Delta \textit{SW}_K - 2(|E^+_{s'} \cap E[K]| - |E^+_s \cap E[K]|).
\end{align*}
Furthermore, the payoff of the players that are neither in
$K$ nor in $N_K$ does not change and hence $\Delta \textit{SW}_{V \setminus (K \cup N_K)} = 0.$
Putting these equalities together, we obtain \eqref{eqn:deltasw}.
Let $F^c = E[K] \setminus F$. Then
\begin{align*}
|E^+_s \cap E[K]| - |E^+_{s'} \cap E[K]|&= |E^+_s \cap F| -|E^+_{s'} \cap F| + |E^+_s \cap F^c| - |E^+_{s'} \cap F^c|.
\end{align*}
We know that $(K, F^c)$ is a forest because $F$ is a feedback edge set. So $|F^c| < |K|.$ Hence
\begin{align*}
|E^+_s \cap F^c| - |E^+_{s'} \cap F^c| \geq - |F^c| > - |K|.
\end{align*}
Furthermore, each player in $K$ improves his payoff when switching to
$s'$ and hence $\Delta \textit{SW}_K \geq |K|.$ So, plugging in these
inequalities in \eqref{eqn:deltasw} we get
$$
\Delta \textit{SW} > 2(|K| + |E^+_s \cap F| - |E^+_{s'} \cap F| - |K|) = 2(|E^+_s \cap F| - |E^+_{s'} \cap F|),
$$
which proves \eqref{eqn:deltasw2}.
\qed
\end{proof}
Let $\tau(K)$ be the size of a minimal feedback edge set of $G[K]$,
i.e.,
\begin{equation}\label{def:minFES}
\tau(K) = \min\{|F| \:|\: G[K] \setminus F \text{ is acyclic}\}.
\end{equation}
Equation \eqref{eqn:deltasw2} then yields that $\textit{SW}(s') - \textit{SW}(s) > - 2 \tau(K)$.
The following example shows that this bound is tight.
\begin{example} \label{exa:mona}
We define a graph $G = (V, E)$ and a color assignment as follows.
Consider a clique on $l$ nodes and let $K$ be the set of nodes.
Every $i \in K$ can choose between two colors $\{c_i, x\}$, where $c_i \neq c_j$ for every $j \neq i$. Further, every node $i \in K$ is adjacent to $(l-2)$ additional nodes of degree one, each of which has the color set $\{c_i\}$. Note that when defining a joint strategy $s$, it is sufficient to specify $s_i$ for every $i \in K$ because the remaining nodes have only one color to choose from.
Let $s:=(c_i)_{i \in K}$ and $s' := (x)_{i \in K}$. Then $s \betredge{K} s'$ is a
profitable deviation because every node in $K$ increases its payoff
from $(l-2)$ to $(l-1).$ Also $ |E^+_{s}| = l(l-2)$ and $ |E^+_{s'}| =
\frac{l(l-1)}{2}$, so
\[
|E^+_s| - |E^+_{s'}| = l \left(l-2-\frac{l-1}{2} \right) = l \left( \frac{l-1}{2}-1 \right).
\]
Furthermore, each tree on $|K|$ nodes has $|K|-1$ edges. Thus
\[
\tau(K) = |E[K]| - (|K|-1) = \frac{l(l-1)}{2} - (l-1) = l \left( \frac{l-1}{2} -1 \right) + 1.
\]
So $\textit{SW}(s') - \textit{SW}(s) = 2(|E^+_{s'}| - |E^+_s|) = -2\tau(K) + 2$. Tightness follows because the left hand side is always even.
\qed
\end{example}
\subsection{Color forests and pseudoforests}
We use our key lemma to show that coordination games on pseudoforests admit strong equilibria.
Recall that a \emph{pseudoforest} is a graph in which every connected component contains at most one cycle. For a color $x \in M$ let
\[
V_x = \{i \in V \mid x \in A_i\}
\]
be the set of nodes that can choose $x.$ If $G[V_x]$ is a forest for all $x \in M$, we call $G$ a \emph{color forest} (with respect to $A$). Note that, in particular, a forest constitutes a color forest.
Given a joint strategy $s$, we call a subgraph $G'$ of $G$ \emph{completely non-unicolored in $s$} if none of its edges is unicolored in $s$.
We first derive some corollaries from our key lemma. Throughout this section, we consider a profitable deviation $s \betredge{K} s'$ and let $\Delta \textit{SW} = \textit{SW}(s') - \textit{SW}(s)$.
\begin{corollary} \label{cor:no_unicolor_cycle}
If $\Delta \textit{SW} \leq 0$, then there is a cycle $C$ in $G[K]$ that is completely non-unicolored in $s$ and unicolored in $s'$.
\end{corollary}
\begin{proof}
Assume that the claim does not hold. Then for all cycles $C$ in $G[K]$, we can pick an edge $e_C \in E[C]$ that is unicolored in $s$ or non-unicolored in $s'$. Let
$
F = \{e_C \:|\: C \text{ is a cycle in } G[K]\}.
$
This is a feedback edge set satisfying $F \cap E_{s'}^+ \subseteq F \cap E_{s}^+$. Hence by \eqref{eqn:deltasw2},
$
\Delta \textit{SW} > 2(|E^+_s \cap F| - |E^+_{s'} \cap F|) \geq 0,
$
which is a contradiction.
\qed
\end{proof}
The next statement follows immediately from Corollary \ref{cor:no_unicolor_cycle} because unicolored cycles cannot exist in color forests. Note that forests are a special case.
\begin{theorem} \label{thm:color_forest}
Suppose that $G[K]$ is a color forest. Then $\Delta \textit{SW} > 0.$ Hence every coordination game on a color forest has the c-FIP. \qed
\end{theorem}
\begin{corollary} \label{cor:one-cycle}
If $G[K]$ is a graph with at most one cycle, then $\Delta \textit{SW} \geq 0.$
\end{corollary}
\begin{proof}
If $G[K]$ is a connected graph with exactly one cycle, then there is a feedback edge set of size $1$. Hence $\Delta \textit{SW} > -2$. Because the left hand side
is even, this implies $\Delta \textit{SW} \geq 0.$
\qed
\end{proof}
Using Corollary \ref{cor:no_unicolor_cycle} and \ref{cor:one-cycle},
we now establish the following result.
\begin{theorem} \label{thm:pseudoforest}
Every coordination game on a pseudoforest has the c-FIP.
\end{theorem}
\begin{proof}
Associate with each joint strategy $s$ the pair
\[
P(s): = \left(\textit{SW}(s),\ |\{ C \mid \text{$C$ is a unicolored cycle in $s$} \} | \right).
\]
We now claim that
$P: S \to \mathbb{R}^2$ is a generalized ordinal c-potential
when we take for the strict partial ordering $\succ$ on $P(S)$
the lexicographic ordering.
Consider a profitable deviation $s \betredge{K}
s'$. By partitioning $K$ into the subsets of different
connected components we can decompose this deviation into a sequence of profitable deviations such that each
deviating coalition induces a subgraph of a connected graph with at most one cycle.
By Corollary \ref{cor:one-cycle} the social
welfare in each of these profitable deviations weakly increases. So
$\textit{SW}(s') \geq \textit{SW}(s)$.
If $\textit{SW}(s') > \textit{SW}(s)$ then $P(s') \succ P(s)$. If $\textit{SW}(s') = \textit{SW}(s)$, then by Corollary \ref{thm:color_forest} each of these profitable deviations is by a coalition that induces a connected graph with exactly one cycle. Moreover, this cycle becomes unicolored in $s'$. Thus $P(s') \succ P(s)$.
\qed
\end{proof}
\subsection{Further applications}
The following corollary is an immediate consequence of Corollary~\ref{thm:color_forest}.
\begin{corollary} \label{cor:2-eq}
In every coordination game, every sequence of profitable deviations of coalitions of size at most two is finite. Hence Nash equilibria and 2-equilibria always exist.
\end{corollary}
\begin{corollary}
Every coordination game in which at most two colors are used has the c-FIP.
\end{corollary}
\begin{proof}
Let $s \betredge{K} s'$ be a profitable deviation. By assumption, all players in $K$ then deviate to their other option. As a consequence, every edge in $E[K]$ is unicolored in $s'$ if and only if it is unicolored in $s$. Hence
each cycle in $G[K]$ that is unicolored in $s'$ is also unicolored in
$s.$ It follows from Corollary \ref{cor:no_unicolor_cycle} that
$\textit{SW}(s') > \textit{SW}(s)$. This shows that $\textit{SW}$ is a generalized ordinal
c-potential.
\qed
\end{proof}
The existence of strong equilibria for coordination
games with two colors and symmetric strategy sets follows from
Proposition 2.2 in \cite{KBW97a}. Corollary~\ref{cor:2-eq} shows that
a stronger result holds, namely that these games have the c-FIP. This
implies that arbitrary coalitional improvement paths always converge
to a strong equilibrium.
We next derive an existence result of $k$-equilibria in graphs in which every pair of cycles is edge-disjoint. We call an edge $e$ of a graph \emph{private} if it belongs to a cycle and is node-disjoint from all other cycles.
\begin{lemma} \label{lem:disj_cycles}
Let $G$ be a graph in which every pair of cycles is edge-disjoint.
Then there exists a private edge.
\end{lemma}
\begin{proof}
Given a cycle $C$, we call a node $v \in V(C)$ an \emph{anchor point} of $C$ if $v$ can be reached from a node $v' \in V(C')$ of another cycle $C' \neq C$ without traversing an edge in $E(C)$. First we show that there always exists a cycle with at most one anchor point. Assume that the claim does not hold. Then every cycle $C$ of $G$ contains at least two distinct anchor points. Fix an arbitrary cycle $C$ of $G$ and let $v^1_C$ and $v^2_C$ be two anchor points of $C$. Start from $v^1_C$ and traverse the edges of $C$ to reach $v^2_C$. Then follow a shortest path $P$ that connects $v^2_C$ to a node $v^1_{C'} \in V(C')$ of another cycle $C' \neq C$; $P$ must exist because $v^2_C$ is an anchor point.
Note that $C$ and $C'$ share at most one node because all cycles are edge-disjoint; in particular, $P$ might have length zero and consist of a single node only. Because we choose a shortest path connecting $C$ and $C'$, $v^1_{C'}$ must be an anchor point of $C'$. By assumption, $C'$ has another anchor point $v^2_{C'}$. Repeat the above procedure with cycle $C'$ and anchor points $v^1_{C'}$ and $v^2_{C'}$. Continuing this way, we construct a path that traverses cycles of $G$. Eventually, this path must return to a previously visited cycle. So this path contains a cycle and this cycle shares at least one edge with one of the visited cycles. This contradicts the assumption that all cycles of $G$ are pairwise edge-disjoint.
Now, let $C$ be a cycle with at most one anchor point $v$ (if no such node exists, any edge in $E(C)$ is private). Then any edge $e \in E(C)$ such that $v$ is not an endpoint of $e$ is private.
\qed
\end{proof}
\begin{theorem}\label{thm:edge-disjoint}
Consider a coordination game on a graph $G$ in which every pair of cycles is edge-disjoint. Let $k$ be the minimum length of a cycle in $G$. Then every sequence of profitable deviations of coalitions of size at most $k$ is finite. In particular, the game has a 3-equilibrium.
\end{theorem}
\begin{proof}
We proceed by induction on the number $z$ of cycles. If $z = 1$, then the claim follows by Theorem~\ref{thm:pseudoforest}.
Now, let $z > 1$. Let $s \betredge{K} s'$ be a profitable deviation such that $|K| \leq k$. From \eqref{eqn:deltasw} we infer that
$$
\Delta \textit{SW} = 2 (\Delta \textit{SW}_K - |E^+_{s'} \cap E[K]| + |E^+_s \cap E[K]|)
\ge 2 (\Delta \textit{SW}_K - |E[K]|)
$$
because $|E^+_{s'} \cap E[K]| \le |E[K]|$.
Because $k$ is the minimum length of a cycle in $G$ and $|K| \leq k$, we have $|E[K]| \leq |K|.$ So $\Delta \textit{SW} \geq 0$.
Consider a sequence of profitable deviations $s_1 \betredge{K_1} s_2 \betredge{K_2} s_3 \ldots$ We show that it is finite. Because the social welfare cannot decrease and is upper bounded there is an index $l \ge 1$ such that for all $i\geq l$, $\textit{SW}(s_i) = \textit{SW}(s_l).$ We can assume without loss of generality that $l=1$. By Corollary \ref{cor:no_unicolor_cycle}, for each $i \geq 1$ there is a cycle $C_i$ in $G[K_i]$ such $C_i$ is completely non-unicolored in $s_i$ and unicolored in $s_{i+1}.$ Note that $k \leq |V(C_i)| \leq |K_i| \leq k$ and hence $K_i = V(C_i).$
By Lemma \ref{lem:disj_cycles}, there is a cycle $C$ with a private edge $e =\{u_1, u_2\} \in E(C)$. We claim that $C = C_i$ for at most one $i.$ Assume otherwise and let $i_{1}, i_{2}$ such that $C = C_{i_1} = C_{i_2}$ and $C \neq C_i$ for $i_1 < i < i_2$. Because $C = C_{i_1}$, $e$ is unicolored in $s_{i_1+1}$. We know that $C$ is the only cycle containing $u_j$ for $j=1,2$ by choice of $e$. So $u_j \notin K_i$ for $i_1 < i < i_2$ and hence $e$ is still unicolored in $i_2$. But $C$ switches from completely non-unicolored in $s_{i_2}$ to unicolored in $s_{i_2 + 1}$, a contradiction.
Since $C$ is the only cycle containing $u_j$ for $j = 1, 2$, it follows that each $u_j$ can appear at most once in a deviating coalition. So there is an index $l$ such that $u_1, u_2 \notin K_i$ for all $i>l$. Hence if we remove $e$ and call the new graph $G'$, then for all $i>l$, $s_{i} \betredge{K_i} s_{i+1}$ is a profitable deviation in $G'$. Because $G'$ has one cycle less than $G$, we can apply the induction hypothesis and conclude that the considered sequence of profitable deviations is finite.
\qed
\end{proof}
\subsection{Uniform coordination games}
\label{sec:uniform}
Next, we establish the c-FIP property for some additional classes of coordination games. We call a coordination game on a graph $G$ \emph{uniform} if for every joint strategy $s$ and for every edge $\{i, j\} \in E$ it holds that if $s_i = s_j$ then $p_i(s) = p_j(s)$.
\begin{theorem}\label{thm:consistent}
Every uniform coordination game has the c-FIP.
\end{theorem}
\begin{proof}
Given a sequence $\theta \in \mathbb{R}^n$ of reals we denote by $\theta^*$ its reordering from the largest to the smallest element. Associate with each joint strategy $s$ the sequence $(p_1(s), \dots, p_n(s))^{*}$ that we abbreviate to $p^{*}(s)$. We now claim that $p^{*}: S \to \mathbb{R}^n$ is a generalized ordinal c-potential when we take for the partial ordering $\succ$ on $p^{*}(S)$ the lexicographic ordering on the sequences of reals.
Suppose that some coalition $K$ profitably deviates from the joint strategy $s$ to $s' = (s'_K, s_{-K})$. We claim that then $p^{*}(s') \succ p^{*}(s)$.
Assume this does not hold. Rename the players such that $p^*(s') = (p_1(s'), \dots, p_n(s'))$. Let $i$ be the smallest value for which $p_{i}(s') < p_{i}(s)$. By assumption such an $i$ exists. By the choice of $i$ for all $j < i$ we have $p_{j}(s') \geq p_{j}(s)$ and also $p_{j}(s') \geq p_{i}(s')$.
Now, $p_{i}(s') < p_{i}(s)$ implies that $i \not\in K$ and hence we can write $s' = (s_i, s'_{-i})$. By the definition of the payoff functions, it follows that there exists some neighbor $j$ of $i$ with $s_j = s_i$ and $s'_j \neq s_j$. Thus, $j \in K$. By the uniformity property, $p_i(s) = p_j(s)$. So $p_j(s') > p_j(s) = p_i(s)$. Consequently, by the choice of $i$, we have $p^{*}(s') \succ p^{*}(s)$, which is a contradiction.
\qed
\end{proof}
We can capture by Theorem~\ref{thm:consistent} the following class of coordination games: We say that $G$ is \emph{color complete} (with respect to $A$) if for every $x \in M$ each component of $G[V_x]$ is complete. (Recall that $V_x = \{i \in V \mid x \in A_i\}$.)
\begin{corollary}
\label{cor:color_complete}
Every coordination game on a color complete graph has the c-FIP. In particular, every coordination game on a complete graph has the c-FIP.
\end{corollary}
The existence of strong equilibria for color complete graphs also follows from a result by Rozenfeld and Tennenholtz \cite{RT06} and the following lemma.
\begin{lemma}
\label{lem:congestion_iso}
Coordination games on color complete graphs are a special case of monotone increasing congestion games in which all strategies are singletons.
\end{lemma}
\begin{proof}
We can assume without loss of generality that for each color $x$, $G[V_x]$ is connected (otherwise, we replace $x$ with a new respective color for each component of $G[V_x]$). Then we can identify $x$ with a singleton resource, along with the payoff function $v_x: \mathbb N \to \mathbb R$ such that $v_x(k) = k-1$. Now, if a player $i$ chooses $s_i = x$ then
\[
p_i(s) = |\{j \in N_i \mid s_j = x\}| = |\{j \in V \mid s_j = x\}| - 1= v_x(|\{j \in V \mid s_j = x\}|),
\]
so the payoff in the coordination game coincides with the payoff in the associated congestion game. \qed
\end{proof}
Rozenfeld and Tennenholtz \cite{RT06} show that monotone increasing congestion games in which all strategies are singletons admit strong equilibria. Note, however, that our result above is stronger because we show that these games have the c-FIP.
\subsection{Non-existence of $3$-equilibria and existence thresholds}
\label{sec:non-existence-3eq}
We next prove that $3$-equilibria do not exist in general. Recall that $2$-equilibria always exist by Corollary~\ref{cor:2-eq}.
\begin{figure}[t]
\centering
\begin{tikzpicture}[line join=bevel,z=-5.5,scale=3]
\coordinate (A1) at (0,0,-1);
\coordinate (A2) at (-1,0,0);
\coordinate (A3) at (0,0,1);
\coordinate (A4) at (1,0,0);
\coordinate (B1) at (0,1,0);
\coordinate (C1) at (0,-1,0);
\coordinate (D1) at (-1.35,0,0);
\coordinate (D2) at (1.35,0,0);
\coordinate (D3) at (0,0,-2);
\coordinate (D4) at (0,0,2);
\draw [very thick,fill opacity=0.5,fill=green] (A2) -- (A1) -- (C1) -- cycle;
\draw [very thick,fill opacity=0.5,fill=yellow] (A2) -- (A3) -- (B1) -- cycle;
\draw [very thick,fill opacity=0.5,fill=red] (A3) -- (C1) -- (A4) -- cycle;
\draw [very thick,fill opacity=0.5,fill=blue] (A4) -- (B1) -- (A1) -- cycle;
\node at (A1) [circle,scale=0.7,fill=black] {};
\node at (A2) [circle,scale=0.7,fill=black] {};
\node at (A3) [circle,scale=0.7,fill=black] {};
\node at (A4) [circle,scale=0.7,fill=black] {};
\node at (B1) [circle,scale=0.7,fill=black] {};
\node at (C1) [circle,scale=0.7,fill=black] {};
\node at (D1) [circle,scale=0.7,fill=black] {};
\node at (D2) [circle,scale=0.7,fill=black] {};
\node at (D3) [circle,scale=0.7,fill=black] {};
\node at (D4) [circle,scale=0.7,fill=black] {};
\draw [very thick] (D1) -- (A2);
\draw [very thick] (D2) -- (A4);
\draw [very thick] (D3) -- (A1);
\draw [very thick] (D4) -- (A3);
\node [above left] at (A1) {\scriptsize $\mathbf{6},\{3,4\}$};
\node [below right] at (A4) {\scriptsize $\mathbf{5},\{2,3\}$};
\node [above right] at (A3) {\scriptsize $\mathbf{4},\{1,2\}$};
\node [below left] at (A2) {\scriptsize $\mathbf{3},\{1,4\}$};
\node [above] at (B1) {\scriptsize $\mathbf{1},\{1,3\}$};
\node [below] at (C1) {\scriptsize $\mathbf{2},\{2,4\}$};
\node [left] at (D1) {\scriptsize $\mathbf{7},\{1\}$};
\node [right] at (D2) {\scriptsize $\mathbf{9},\{3\}$};
\node [above] at (D3) {\scriptsize $\mathbf{10},\{4\}$};
\node [below] at (D4) {\scriptsize $\mathbf{8},\{2\}$};
\end{tikzpicture}
\caption{Three-dimensional illustration of the coordination game used to show that $3$-equilibria do not exist (Theorem~\ref{thm:no3eq}). The colored facets indicate the triangles that can be unicolored. The identity of the players is displayed in boldface. The strategy sets of the players are stated between curly braces.}
\label{fig:octahedron}
\end{figure}
\begin{theorem}\label{thm:no3eq}
There exists a coordination game that does not have a $3$-equilibrium.
\end{theorem}
\begin{proof
We define a coordination game $\mathcal{G}(G, A)$ as indicated in Figure~\ref{fig:octahedron}:
There are $n = 10$ players and $4$ colors. The strategy sets are as follows: $A_1 = \{1,3\}$, $A_2 = \{2,4\}$, $A_3 = \{1,4\}$, $A_4 = \{1,2\}$, $A_5 = \{2,3\}$, $A_6 = \{3,4\}$, $A_7 = \{1\}$, $A_8 = \{2\}$, $A_9 = \{3\}$, $A_{10} = \{4\}$. There are $16$ edges, defined as follows: Players $1$ and $2$ are both connected to players $3$,$4$,$5$, and $6$, accounting for $8$ of the edges. There is aditionally a cycle $(3,4,5,6,3)$, accounting for four more edges. Lastly, players $7,8,9$, and $10$ all have a single edge attached to them and are connected to players $3,4,5$, and $6$, respectively.
As can be seen from Figure~\ref{fig:octahedron}, the graph on which the game is played is essentially the skeleton of an octahedron: $12$ of the edges and $6$ of the nodes of the graph belong to this skeleton, and the four remaining edges are connected to four remaining nodes that are \emph{dummy} players (i.e., they have only one strategy that they can play).
Observe that there are eight triangles in the graph, which correspond to the eight facets of the octahedron. The strategy sets are defined such that only four out of the eight triangles of the octahedron can be unicolored. Also, this game is constructed such that if one triangle is unicolored, then the other three triangles are necessarily not unicolored.
We prove the theorem by showing that for every strategy profile of $\mathcal{G}$, there exists a profitable deviation of a set of at most $3$ players.
To simplify the proof, we make use of the many symmetries in $\mathcal{G}$, which are apparent from Figure \ref{fig:octahedron}.
Let $s$ be an arbitrary strategy profile of $\mathcal{G}$.
We distinguish two cases:
\begin{itemize}
\item If there is a triangle that is unicolored under $s$, we may assume without loss of generality that this triangle is the one corresponding to players $\{1,3,4\}$ (because of symmetry), i.e., $s_1 = s_3 = s_4 = 1$. Observe that $p_4(s) = 2$. We distinguish two cases:
\begin{itemize}
\item $p_5(s) = 2$. Then $s_5 = s_6 = 3$. If $s_2 = 4$ then player $6$ can deviate profitably to $4$. If $s_2 = 2$ then the coalition $\{2,6\}$ can deviate profitably to $4$.
\item $p_5(s) \leq 1$. If $s_2 = 2$ and $s_5 = 2$, then player $4$ can deviate profitably to $2$. If $s_2=2$ and $s_5 = 3$ then the coalition $\{4,5\}$ can deviate profitably to $2$. If $s_2 = 4$, then $s_5 = 3$, and coalition $\{2,4,5\}$ can deviate profitably to $2$.
\end{itemize}
\item If there is no triangle that is unicolored under $s$, we distinguish again two cases. By symmetry we may assume that $s_1 = 1$.
\begin{itemize}
\item $s_3 = 4$. Then $p_3(s) \leq 1$, so player $3$ can profitably deviate by changing his color to $1$.
\item $s_3 = 1$. Then $s_4 = 2$. If $p_4(s) = 1$ then player $4$ can profitably deviate by changing his color to $1$. Otherwise, $p_4(s) = 2$ and either $(s_2,s_5) = (2,3)$ or $(s_2,s_5) = (4,2)$.
\begin{itemize}
\item If $(s_2,s_5) = (2,3)$, then if also $p_5(s) = 2$ it holds that $s_6 = 3$ and therefore player $1$ can profitably deviate to $3$. If $p_5(s) = 1$, then player $5$ can profitably deviate to $2$.
\item If $(s_2,s_5) = (4,2)$, then $p_2(s) \leq 1$, so player $2$ can profitably deviate to $2$.
\end{itemize}
\end{itemize}
\end{itemize}
Note that each profitable deviation given above consists of at most three players. This concludes the proof.
\qed
\end{proof}
The coordination game given in Figure~\ref{fig:octahedron} is an example of a game that does not have a $3$-equilibrium but admits a $2$-equilibrium.
We define the \emph{transition value} of a coordination game as the value of $k$ for which a $k$-equilibrium exists but a $(k+1)$-equilibrium does not. Clearly, the instance in Figure~\ref{fig:octahedron} has an transition value of $k = 2$. An interesting question is whether one can identify instances of coordination games with a non-trivial transition value $k \ge 3$.
We next show that the coordination game given in Figure~\ref{fig:non-existence} is an instance with transition value $k = 4$.
\begin{theorem}\label{thm:non-existence}
There is a coordination game that has a transition value of $4$.
\end{theorem}
\begin{proof}
Consider the coordination game discussed in Example~\ref{exa:1} (see Figure~\ref{fig:non-existence}). We first argue that it does not admit a $5$-equilibrium. Assume for the sake of a contradiction that $s$ is a strong equilibrium of this game.
Consider players 4 and 5. Let $i, j \in \{4,5\}$, $i \neq j$ be such that $p_i(s) \leq p_j(s)$. Note that the neighbors of $i$ and $j$ (excluding $j$ and $i$, respectively) are the same. As a consequence, if $s_i \neq s_j$, then player $i$ can profitably deviate to player $j$'s color, i.e., $s'_i = s_j$. Thus players 4 and 5 have the same color in $s$, say $s_4 = s_5 = b$. (Because of the symmetry of the instance, the case $s_4 = s_5 = c$ follows analogously.)
Assume there exists a player $i \in \{2, 3\}$ with $s_i \neq b$. Then $i$ can profitably deviate by choosing $s'_i = b$. It follows that players 2 and 3 have color $s_2 = s_3 = b$.
Next, consider player 8 and suppose $s_8 = b$. Then his payoff is $p_8(s) = 2$. Further, the payoff of each of the players 1, 6 and 7 is 0 because all their neighbors have color $b$. But then the coalition $K = \{1, 6, 7, 8\}$ can profitably deviate by choosing color $a$. We conclude that $s_8 = a$.
As a consequence, for players 1, 6, and 7 we have $s_1 = s_6 = s_7 = a$ as otherwise any such player could profitably deviate by choosing $a$.
Thus, the only remaining possible configuration for $s$ is the one indicted in Figure~\ref{fig:non-existence} (by the underlined strategies). But this is not a strong equilibrium because the coalition $K = \{1, 4, 5, 6, 7\}$ can profitably deviate by choosing color $c$. This yields a contradiction and proves the non-existence of $5$-equilibria.
On the other hand, it is easy to see that the strategy profile indicated in Figure~\ref{fig:non-existence} constitutes a $4$-equilibrium. This concludes the proof.
\qed
\end{proof}
In general, we leave open the question for which $k \geq 2$ there exist coordination games with transition value $k$.
\bigskip
The above example can be adapted to show that there are coordination games that do not have the c-FIP but are \emph{c-weakly acyclic}. Recall that a game $\mathcal{G}$ is \emph{c-weakly acyclic} if for every joint strategy there exists a finite c-improvement path that starts at it. Note that a c-weakly acyclic game admits a strong equilibrium.
\begin{corollary}
There is a coordination game that does not have the c-FIP but is $c$-weakly acyclic.
\end{corollary}
\begin{proof}
Take the coordination game from Example \ref{exa:1} and modify it by adding to each color set a new, common color $d$. Then the joint strategy $s$ in which each player selects $d$ is a strong equilibrium. Moreover, for each player her payoff in $s$ is strictly higher than in any joint strategy in which she chooses another color. So $s$ can be reached from each joint strategy in just one profitable deviation, by a coalition of the players who all switch to $d$.
On the other hand, the argument presented in Example \ref{exa:1} shows that this game does not have the c-FIP.
\qed
\end{proof}
\section{Inefficiency of $k$-equilibria}
\label{sec:anarchy}
We first summarize some results concerning the strong price of stability of coordination games.
\begin{theorem}
The strong price of stability is 1 in each of the following cases:
\begin{itemize}
\item $G$ is a pseudoforest;
\item $G$ is a color forest;
\item there are only two colors.
\end{itemize}
\end{theorem}
\begin{proof}
If $G$ is a pseudoforest, a maximum of $P$ in the lexicographic ordering defined in the proof of Theorem~\ref{thm:pseudoforest} is a strong equilibrium and a social optimum. In the other two cases, the social welfare function $\textit{SW}$ is a generalized ordinal c-potential. So in both cases each social optimum is a strong equilibrium.
\qed
\end{proof}
We next study the $k$-price of anarchy of our coordination games. It is easy to see that the price of anarchy is infinite. In fact, this holds independently of the graph structure, as the next theorem shows.
\begin{theorem}
For every graph there exists strategy sets for the players such that the price of anarchy of the resulting coordination game is infinite.
\end{theorem}
\begin{proof}
Let $G = (V, E)$ be an arbitrary graph. We assign to each node $i \in V$ a color set $A_i = \{x_i, c\}$, where $x_i$ is a private color, i.e., $x_i \neq x_j$ for every $j \neq i$, and $c$ is a common color. The joint strategy $s$ in which every player chooses her private color constitutes a Nash equilibrium with $\textit{SW}(s) = 0$. On the other hand, the joint strategy $s'$ in which every player chooses the common color $c$ is a social optimum with $\textit{SW}(s') = 2|E|$.
\qed
\end{proof}
We now determine the $k$-price of anarchy and the strong price of anarchy. We define for every $j \in N$ and $K \subseteq N$ and joint strategy $s$,
\[
N_j^K(s) = \{\{i,j\} \in E \mid i \in K,\, s_i = s_j\}.
\]
Intuitively, $|N_j^K(s)|$ is the payoff $j$ derives from players in $K$ under $s$.
\begin{theorem} \label{thm:kpoa}
The $k$-price of anarchy of coordination games is between $2\frac{n-1}{k-1}-1$ and $2\frac{n-1}{k-1}$ for every $k \in \{2, \dots, n\}$. Furthermore, the strong price of anarchy is exactly $2$.
\end{theorem}
\begin{proof}
We first prove the upper bound.
By the definition of the payoff function for all joint strategies $s$
and $\sigma$, we have
$
|N_j^K(\sigma)| \leq p_j(\sigma_K, s_{-K})
$.
Suppose that the considered game has a $k$-equilibrium, say $s$,
and let $\sigma$ be a social optimum. By the definition of a
$k$-equilibrium, for all coalitions $K$ of size at most $k$ there
exists some $j \in K$ such that $p_j(\sigma_{K}, s_{-K}) \leq p_j(s)$ and
hence by the above $|N_j^K(\sigma)| \leq p_j(s).$
Fix a coalition $K = \{v_1, \ldots, v_k\}$ of size $k$. We know that there is some $j \in K$ such that $|N_j^{K}(\sigma)| \leq p_j(s)$.
Rename the nodes so that $j = v_k$. Further, there is a node $j$ such
that
$$
|N_j^{\{v_1, \ldots, v_{k-1}\}}(\sigma)| \leq p_j(s).
$$
Again we
rename the nodes so that $j = v_{k-1}$. Continuing this way
we obtain that for all $i \in \{1, \ldots, k\}$ it holds that $|N_{v_i}^{\{v_1, \ldots, v_i\}}(\sigma)| \leq p_{v_i}(s).$ Hence
\begin{align*}
p_{v_i}(\sigma) &= |N_{v_i}^{\{v_1, \ldots, v_i\}}(\sigma)| + |N_{v_i}^{V \setminus \{v_1, \ldots, v_i\}}(\sigma)|
\leq p_{v_i}(s) + |N_{v_i}^{V \setminus \{v_1, \ldots, v_i\}}(\sigma)|\\
& = p_{v_i}(s) + |N_{v_i}^{K \setminus \{v_1, \ldots, v_i\}}(\sigma)|
+ |N_{v_i}^{V \setminus K}(\sigma)|.
\end{align*}
Summing over all players in $K$ we obtain
\begin{align} \label{eqn:sw}
\textit{SW}_K(\sigma) \leq \textit{SW}_K(s) + \sum_{i = 1}^k \big( |N_{v_i}^{K \setminus \{v_1, \ldots, v_i\}}(\sigma)| + |N_{v_i}^{V \setminus K}(\sigma)|\big).
\end{align}
But
\begin{align*}
\sum_{i = 1}^k |N_{v_i}^{K \setminus \{v_1, \ldots, v_i\}}(\sigma)|
&= \sum_{i = 1}^k |\{j > i:\, \{v_i, v_j\} \in E^+_{\sigma}\}| = |E^+_\sigma \cap E[K]|
\end{align*}
and
$
\sum_{i = 1}^k |N_{v_i}^{V \setminus K}(\sigma)| = |E^+_\sigma \cap \delta(K)|.
$
Hence rewriting \eqref{eqn:sw} yields
\[
\textit{SW}_K(\sigma) \leq \textit{SW}_K(s) + |E_\sigma^+ \cap E[K]| + |E_\sigma^+ \cap \delta(K)|.
\]
It also holds that $\textit{SW}_K(\sigma) = 2|E^+_\sigma \cap E[K]| + |E^+_\sigma \cap \delta(K)|$. So we get
\[
\textit{SW}_K(\sigma) \leq \textit{SW}_K(s) + \frac{1}{2} \textit{SW}_K(\sigma) + \frac{1}{2} |E^+_\sigma \cap \delta(K)|,
\]
which implies that
\begin{align} \label{eqn:sw2}
\textit{SW}_K (\sigma) \leq 2 \textit{SW}_K(s) + |E^+_\sigma \cap \delta(K)|.
\end{align}
Now we sum over all coalitions $K$ of size $k$. Each player $i$
appears in $n-1 \choose k-1$ of such sets because it is possible to
choose $k-1$ out of $n-1$ remaining players to form a set $K$ of
size $k$ that contains $i$. Hence,
\begin{align*}
\sum_{K: |K| = k} \textit{SW}_K(\sigma) &
= \sum_{i = 1}^{n} \sum_{K: \, K \ni i} p_i(\sigma)
= \sum_{i = 1}^{n} {n-1 \choose k-1} p_i(\sigma) = {n-1 \choose k-1} \textit{SW}(\sigma).
\end{align*}
We obtain an analogous expression for the joint strategy $s$.
Furthermore, for each edge $e = \{u,v\} \in E^+_\sigma$, we can choose
$2 {n-2 \choose k-1}$ sets $K$ of size $k$ such that $e \in
\delta(K)$. Indeed, assuming that $u \in K$ and $v \notin K$, we can
choose $k-1$ out of $n-2$ remaining players to complete $K$ and hence
there exist ${n-2 \choose k-1}$ of those sets. Reversing the roles of
$u$ and $v$ and summing up yields $2 {n-2 \choose k-1}$. Hence
\[
\sum_{K: |K| = k} |E^+_\sigma \cap \delta(K)| = 2 {n-2 \choose k-1} |E^+_\sigma| = {n-2 \choose k-1} \textit{SW}(\sigma).
\]
By summing over all coalitions $K$ of size $k$, equation \eqref{eqn:sw2} yields
\[
{n-1 \choose k-1} \textit{SW}(\sigma) \leq 2 {n-1 \choose k-1} \textit{SW}(s) + {n-2 \choose k-1} \textit{SW}(\sigma).
\]
It follows that the $k$-price of anarchy is at most
$$
\frac{2{n-1 \choose k-1}}{{n-1 \choose k-1} - {n-2 \choose k-1}} = 2 \frac{n-1}{k-1}.
$$
This concludes the proof of the upper bound.
The claimed lower bounds follow from Examples~\ref{ex:lb1} and \ref{ex:lb2} given below.
\qed
\end{proof}
The following example establishes a lower bound on the $k$-price of anarchy.
\begin{example}\label{ex:lb1}
Fix $n$ and $k \in \{2, \dots, n\}$. Let $V(G)$ consist of two sets $V_1$ and $V_2$ of size $k$ and $n-k$, respectively, and define
\[
E[G] = \{\{u,v\} \mid u \in V_1, v \in V_1 \cup V_2\}.
\]
Fix three colors $a,b$ and $c$. For $v \in V_1$, let $A(v) = \{a,c\}$. For $v \in V_2$, let $A(v) = \{b,c\}$. Then the color assignment $\sigma$ in which each player chooses the common color $c$ is a social optimum. The social welfare is
\[
\textit{SW}(\sigma) = \textit{SW}_{V_1}(\sigma) + \textit{SW}_{V_2}(\sigma) = k(n-1) + (n-k)k.
\]
Next we show that the color assignment $s$ in which every node in $V_1$ chooses $a$ and every node in $V_2$ chooses $b$ is a $k$-equilibrium. Assume that there is a profitable deviation $s \betredge{K} s'$ such that $|K| \leq k$. Then all nodes in $K$ switch to $c$ and also all nodes that choose $c$ in $s'$ are in $K$. Hence for all $v \in K$, $p_v(s') = |N_v \cap K|.$ So there is a node $v \in V_1 \cap K$ because otherwise the payoff of all nodes in $K$ would remain 0. But then $p_v(s') = |N_v \cap K| \leq k = p_v(s)$, which yields a contradiction.
Note that $\textit{SW}(s) = k(k-1).$
It follows that the $k$-price of anarchy is at least
\[
\frac{\textit{SW}(\sigma)}{\textit{SW}(s)} = \frac{k(n-1) + (n-k) k}{k(k-1)} = \frac{2 (n-1) - (k-1)}{k-1} = 2\frac{n-1}{k-1}-1.
\]
\qed
\end{example}
The following example shows that the upper bound of $2$ on the strong price of anarchy ($k = n$) of Theorem~\ref{thm:kpoa} is tight.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[scale=1.5]
\node [label={\footnotesize $\{ a, b\}$}] (1) at (0,1) {$1$};
\node [label=right:{\footnotesize $\{ a, b\}$}] (2) at (1,0) {$2$};
\node [label=below:{\footnotesize $\{ a, b\}$}] (3) at (0,-1) {$3$};
\node [label=left:{\footnotesize $\{ a \}$}] (4) at (-1,0) {$4$};
\draw (1) -- (2);
\draw (2) -- (3);
\draw (3) -- (4);
\draw (1) -- (4);
\end{tikzpicture}
\end{center}
\caption{A coordination game showing that the strong price of anarchy is at least 2.}
\label{fig:graph2}
\end{figure}
\begin{example}\label{ex:lb2}
Consider the graph and the color assignment depicted in Figure~\ref{fig:graph2}.
Here $(a,a,a,a)$ is a social optimum with the social welfare 8, while
$(b,b,b,a)$ is a strong equilibrium with the lowest social welfare, 4. So the strong price of anarchy is $2$ in this example of $4$ players. By duplicating the graph $l$ times, we can draw the same conclusion for the case of $4l$ players.
\qed
\end{example}
\section{Complexity}
\label{sec:computation}
In this section we study complexity issues concerning $k$-equilibria.
\subsection{Verification}
First, we show that in general it is hard to decide whether a given joint strategy is a $k$-equilibrium.
Let $k$-\textsc{Equilibrium} denote the problem to decide, given a coordination game with a joint strategy $s$ and $k \in \{1, \dots, n\}$, whether $s$ is a $k$-equilibrium.
\begin{theorem}
$k$-\textsc{Equilibrium} is co-NP-complete.
\end{theorem}
\begin{proof}
It is easy to verify that $k$-\textsc{Equilibrium} is in co-NP: a certificate of a NO-instance is a profitable deviation of a coalition of size at most $k$.
We show the hardness by reduction of the complement of \textsc{Clique}, which is a co-NP-complete problem. Let $(G,k)$ be an instance thereof. We construct an instance of $k$-\textsc{Equilibrium} as follows. For $v \in V$ let $A_v = \{x_v, y\}$, where color $y$ and all colors $x_v, v\in V$ are distinct. Furthermore, for every node $v \in V$ we add $k-2$ nodes $u_v^1, \ldots, u_v^{k-2}$ and edges $\{v,u_v^i\}$ for $i = 1, \ldots, k-2$. These additional nodes can only choose the color $x_v.$ Let $s$ be the joint strategy in which every node $v \in V$ chooses $x_v$. We claim that this is a $k$-equilibrium if and only if $G$ has no clique of size $k.$
Suppose $G$ has a clique $K$ of size $k$. Then jointly deviating to $y$ yields to each node in $K$ a payoff of $k-1$, whereas every node has a payoff of $k-2$ in $s$. So this is a profitable deviation. For the other direction, suppose that there is a profitable deviation $s \betredge{K} s'$ by a coalition $K$ of size at most $k$. Then every node in $K$ deviates to $y$ and hence belongs to $V.$ Since every node in $K$ has a payoff of $k-2$ in $s$, $p_v(s') \geq k-1$ for all $v \in K.$ So $v$ is connected to at least $k-1$ nodes in $K$. This implies that $K$ is a clique of size $k.$
\qed
\end{proof}
We next show that for color forests the decision problem is in \text{P}. First we show that we can focus on certain profitable deviations which we call \emph{simple}: Fix a joint strategy $s$ and a coalition $K$. We call $K$ \emph{connected} if $G[K]$ is connected. A deviation $s \betredge{K} s'$ is \emph{simple} if $K$ is connected and $s' = (x_K, s_{-K})$ for some color $x.$
\begin{lemma}\label{lem:simple}
Let $s$ be a joint strategy in a coordination game. If there is a profitable deviation by a coalition of size at most $k$, then there is also a simple profitable deviation by a coalition of size at most $k$.
\end{lemma}
\begin{proof}
Let $s \betredge{K} s'$ be a profitable deviation with $|K| \leq k$. Pick an arbitrary $v \in K$ and let $x = s'_v$. Let $L$ consist of those nodes $u \in K$ for which $s'_u = x$ and $u$ is reachable in $G[K]$ from $v.$ Let $s'' = (x_L, s_{-L})$.
Then the deviation to $s''$ is simple. For all nodes $u\in L$, we have $N_u^K(s') = N_u^L(s') = N_u^L(s'')$ by the definition of $L$. Furthermore,
$
N_u^{V \setminus K}(s') \subseteq N_u^{V \setminus L}(s') \subseteq N_u^{V \setminus L}(s'').
$
Hence
\[
p_u(s') = |N_u^K(s')| + |N_u^{V \setminus K}(s')| \leq |N_u^L(s'')| + |N_u^{V \setminus L}(s'')| = p_u(s''),\]
which implies that the deviation to $s''$ is profitable for $u.$ \qed
\end{proof}
\begin{theorem}\label{thm:computation_color_forest}
Consider a coordination game on a color forest. Then there exists a polynomial-time algorithm that decides whether a given joint strategy is a $k$-equilibrium and, if this is not the case, outputs a profitable deviation of a coalition of size at most $k.$
\end{theorem}
\begin{proof}
For a statement $P$ we write below $\ensuremath{[\![} P \ensuremath{]\!]}$ to denote the variable that is 1 if $P$ is true and $0$ otherwise. For a function $f: V \to \mathbb R$ and $U \subseteq V$, let $f(U) = \sum_{v \in U} f(v).$ For a function $F: V \to 2^V$ let $F(U) = \bigcup_{v \in U} F(v).$
Let $s$ be a joint strategy. By Lemma \ref{lem:simple} it is
sufficient to check for the existence of \emph{simple} profitable
deviations by coalitions of size at most $k$. Thus, we let $x \in M$ and we
search for simple profitable deviations in which the coalition deviates to
$x$. Because a coalition in a simple deviation is connected, we can
check each connected component of $G[V_x]$ separately. Assume
without loss of generality that $G[V_x]$ itself is connected, i.e., is
a tree. Pick an arbitrary root $r$ of $G[V_x]$ and define for each node the \emph{children}, \emph{parent}, and \emph{rooted subtree} in the usual way (with respect to $r$). For each node $v
\in V_x$ let $\m C_v \subseteq V_x$ denote the set of children of $v$ and
let $\m P_v \in V_x$ denote the parent of $v$ (if $v \neq
r$). Finally, let $\m T_v$ denote the subtree of $G[V_x]$ rooted at
$v$. For each node $v$, we define $\m U(v)$, $D(v) $, $\m U^p(v)$, and $D^p(v)$ as follows.
\begin{itemize}
\item $\m U(v)$ is a connected coalition $K \subseteq \m T_v$ of minimum size such that $v \in K$ and the deviation to $(x_{K},s_{-K})$ is profitable for all nodes in $K$ (if such a coalition exists). We denote the properties it has to satisfy by ($*$).
\item $D(v) = |\m U(v)|$ if $\m U(v)$ exists and $\infty$ otherwise.
\item $\m U^p(v)$ is a connected coalition $L \subseteq \m T_v$ of minimum size such that
$v \in L$ and the deviation to $(x_{L'}, s_{-L'})$ is profitable for all nodes in $L$, where $L' := L \cup \{\m P_v\}$ (if such a coalition exists). We denote the properties it has to satisfy by ($**$). (Note that the deviation is not required to be profitable for $\m P_v$ even though $\m P_v \in L'.$)
\item $D^p(v) = |\m U^p(v)|$ if $\m U^p(v)$ exists and $\infty$ otherwise.
\end{itemize}
We can compute $D$, $D^p$, $\m U$ and $\m U^p$ using a dynamic program as follows. Let $v \in V_x$ and suppose we found these objects for all children of $v$. Let $U \subseteq \m C_v$ minimize $D^p(U)$ among all sets $U' \subseteq \m C_v$ that satisfy $D^p(U') < \infty$ and
\begin{eqnarray}\label{eqn:1}
|U'| + \ensuremath{[\![} s_{\m P_v} = x \ensuremath{]\!]} > p_v(s)
\end{eqnarray}
if such a set exists. In this case set $\m U(v) = \{v\} \cup \m U^p(U)$ and $D(v) = |\m U(v)|$. Otherwise, we set it to $\infty$.
We first prove that $K := \m U(v)$ satisfies ($*$) if $\m U(v)$ exists. The set $K$ is connected because $v \in K$ and all sets $\m U^p(u)$ are connected, for $u \in \m C_v$. It is profitable for $v$ to deviate to $x$ because of \eqref{eqn:1}. The deviation is profitable for nodes $u \in K \setminus \{v\}$ because $\m U^p(u) \subseteq K$, $\m P(u) \in K$ (by the connectivity of $K$) and $\m U^p(u)$ satisfies ($**$). Furthermore, $K$ is of minimal size amongst all coalitions that satisfy ($*$). Indeed, if $K'$ is another such coalition, then $U' := K \cap \m C_v$ satisfies \eqref{eqn:1} because it is profitable to deviate to $x$ for $v.$ It is profitable for $u \in U'$ to deviate to $x$ and hence $|K \cap \m T(u)| \geq D^p(u)$, which implies $|K'| \geq 1+ D^p(U').$ Therefore $|K'| \geq 1 + D^p(U)$ by the minimality of $U$. But $|K| = |\{v\} \cup \m U^p(U)| = 1 + D^p(U)$, which shows that $|K'| \geq |K|.$
Similarly, for $v \in V_x \setminus \{r\}$, let $W \subseteq \m C_v$ minimize $D^p(W)$ among all sets $W' \subseteq \m C_v$ that satisfy $D^p(W') < \infty$ and
\begin{equation}\label{eqn:2}
|W'| + 1 > p_v(s)
\end{equation}
if such a set exists. In this case set $\m U^p(v) = \{v\} \cup \m U^p(W)$ and $D^p(v) = |\m U^p(v)|$. Otherwise, we set $D^p(v) = \infty.$ Similar arguments as before show that if $\m U^p(v)$ exists then it indeed satisfies ($**$).
Note that we can compute $\m U(v)$ in polynomial time by sorting the nodes $u \in \m C_v$ in increasing order of $D^p(u)$ and then successively adding nodes to $\m U(v)$ until \eqref{eqn:1} is satisfied. Similarly, we can compute $\m U^p(v)$ efficiently. This shows that the algorithm runs in polynomial time.
Now, let $s \betredge{K} s'$ be a simple profitable deviation to $x$ such that $|K| \leq k$. Let $v$ be the root of $G[K]$ according to our previously fixed ordering. By the properties of the function $D$, we know that $D(v) \leq |K| \leq k$. Conversely, if $D(v) \leq k$ for some node $v$, then $\m U(v)$ of size $D(v) \leq k$ is the coalition we are looking for.
\qed
\end{proof}
\iffalse
\bigskip
\noindent
As we have seen, the problem of deciding whether a given strategy is a $k$-equilibrium is hard in general. In contrast, as we show below it is easy to decide whether it is a strong equilibrium.
Let \textsc{StrongEquilibrium} denote the problem of deciding whether, given a coordination game and a joint strategy $s$, $s$ is a strong equilibrium.
For a graph $G$ and $v \in V(G)$, denote by $\text{deg}_G(v)$ the \emph{degree} of $v$ in $G$, i.e., the number of its neighbors in $G$.
\begin{theorem}
\label{thm:SE_verification}
\textsc{StrongEquilibrium} is in \emph{P}.
\end{theorem}
\begin{proof}
We give a polynomial reduction to the following decision problem, which we call \textsc{MinDegree}. As we show in Appendix \ref{app:min-deg}, this problem is solvable in polynomial time.
\begin{quote}
Given a graph $G = (V,E)$ together with numbers $d_v$ for all nodes $v \in V$, is there a set of nodes $K \neq \emptyset$ such that every $v \in K$ has degree at least $d_v$ in the induced subgraph $G[K]$?
\end{quote}
Let $(G,s)$ be an instance of \textsc{StrongEquilibrium.} By Lemma \ref{lem:simple}, it suffices to check if there are simple deviations. So fix $x \in M$ and we check if there is a profitable deviation $s \betredge{K} (x_K, s_{-K})$ by a non-empty set $K \subseteq V$. Note that if this is the case then
$$
K \subseteq V_x \setminus \{v \in V \mid s_v = x\} =: V'_x.
$$
We consider the graph $G'$ induced by the set of nodes $V'_x$ together with the numbers
$$
d_v = p_v(s) - | \{j \in N_v \mid s_j = x\}| + 1
$$
for $v \in V'_x$.
Fix a non-empty set $K \subseteq V'_x$. Then for each $v \in K$, jointly deviating to $(x_{K}, s_{-K})$ is profitable iff it has degree at least $d_v$ in the induced subgraph $G'[K]$. Indeed,
$$
p_v(x_{K}, s_{-K}) = \text{deg}_{G'[K]}(v) +
|\{j \in N_v \mid s_j = x\}| > p_v(s)
$$
if and only if $\text{deg}_{G'[K]}(v) \geq d_v$.
So there exists a profitable simple deviation to $x$ if and only if the instance $(V_x', d_v)$ as described above is a YES-instance of \textsc{MinDegree.}
Clearly this construction is polynomial in the size of the input and we have to perform it at most $m$ times, once for each color. So this is a polynomial time reduction of \textsc{StrongEqui\-librium} to \textsc{MinDegree.}
\qed
\end{proof}
\fi
\subsection{Computing strong equilibria}
Next we focus on the problem of actually computing a strong equilibrium. As we show below, this is possible for certain graph classes.
\begin{corollary}
\label{cor:computation_color_forest}
Consider a coordination game on a color forest. Then a strong equilibrium can be computed in polynomial time.
\end{corollary}
\begin{proof}
We begin with an arbitrary initial joint strategy $s$. Putting $k = n$, by Theorem \ref{thm:computation_color_forest} there is an algorithm that decides whether $s$ is a strong equilibrium and, if this is not the case, outputs a profitable deviation $s \betredge{K} s'.$ In the first case, we output $s$; in the second case, we repeat the procedure with $s'$. We know that $\textit{SW}(s)$ is a natural number and $\textit{SW}(s') > \textit{SW}(s)$ by Corollary \ref{thm:color_forest}, so at most $\max_{s \in S} \textit{SW}(s) \leq 2|E|$ steps are necessary to reach a strong equilibrium.
\qed
\end{proof}
\begin{theorem}
Consider a coordination game on a color complete graph. Then a strong equilibrium can be computed in polynomial time.
\end{theorem}
\begin{proof}
This follows from Lemma \ref{lem:congestion_iso} and the corresponding result for monotone increasing congestion games in which all strategies are singletons, established in \cite{RT06}.
\qed
\end{proof}
\begin{theorem} \label{thm:computation_pseudoforest}
Consider a coordination game on a pseudoforest. Then a strong equilibrium can be computed in polynomial time.
\end{theorem}
\begin{proof}
We first show that for a tree, a strong equilibrium can be computed efficiently via dynamic programming. By Corollary~\ref{thm:color_forest} it suffices to compute a social optimum. Let $T$ be a tree and root it at an arbitrary node $r \in V(T)$. Given a node $i \in V(T)$, let $\m T_i$ denote the subtree of $T$ that is rooted at $i$ and let $\m C_i$ be the set of children of $i$. Given a color $s_i \in S_i$, define $d_i(s_i)$ as the maximum social welfare achievable by the nodes in $\m T_i$ if node $i$ chooses color $s_i$. Note that for each leaf $i \in V(T)$ of $T$ we have $d_i(s_i) = 0$ for all $s_i \in S_i$.
Consider a node $i \in V(T)$ that is not a leaf and assume we computed all values $d_j(s_j)$ for every $j \in \m C_i$ and $s_j \in S_j$. Define $\ensuremath{[\![} s_j = s_i \ensuremath{]\!]}$ to be 1 if $s_j = s_i$ and 0 otherwise.
We can then compute $d_i(s_i)$ for every $s_i \in S_i$ as follows:
$$
d_i(s_i) = \sum_{\text{$j \in \m C_i$}} \max_{s_j \in S_j} (d_j(s_j) + 2 \ensuremath{[\![} s_j = s_i \ensuremath{]\!]}).
$$
The intuition here is that we account for every child $j \in \m C_i$ of $i$ for the maximum social welfare achievable in $\m T_j$ plus an additional contribution of $2$ if $i$ and $j$ choose the same color.
Computing $d_i(s_i)$ for all $s_i \in S_i$ takes time at most $O(m^2|\m C_i|)$, where $m$ is the number of colors. Thus, it takes time $O(m^2 |V(T)|)$ to compute all values $d_r(s_r)$ for $s_r \in S_r$ of the root node $r$.
The optimal social welfare of the tree $T$ is then $\textit{SW}(T) = \max_{s_r \in S_r} d_r(s_r)$.
The corresponding optimal joint strategy $s^*_T$ can be determined using some standard bookkeeping.
Next suppose that $T$ is a pseudotree. Let $C = (i_1, \dots, i_k)$ be the unique cycle in $T$. Note that it might no longer be sufficient to simply compute a social optimum for $T$. Instead, the idea is to compute a social optimum $s^*_T$ of $T$ such that, if possible, $C$ is unicolored.
Note that if such a social optimum does not exist, then there is an edge in $C$ that is not unicolored. Let $\textit{SW}(j)$, $j \in \{1, \dots, k\}$, be the maximum social welfare of the tree that one obtains from $T$ by removing edge $\{i_j, i_{j+1}\}$ from $C$ (where we define $i_{k+1} = i_1$). Note that we can efficiently compute $\textit{SW}(j)$ by using the dynamic program for trees described above.\footnote{Observe that we do not enforce that the endpoints of the removed edge $\{i_j, i_{j+1}\}$ obtain different colors in the optimal solution. In fact, subsequently it will become clear that we do not have to do so.} Let $\textit{SW}_1 = \max_{j = 1, \dots, k} \textit{SW}(j)$. Computing $\textit{SW}_1$ takes time $O(k \cdot m^2 |V(T)|) = O(nm^2 |V(T)|)$.
Next assume a social optimum exists in which all nodes of $C$ are unicolored. Note that if we remove the edges on $C$ from $T$ then $T$ decomposes into $k$ trees, rooted at $i_1, \dots, i_k$. We can compute $d_{i_j}(\cdot)$ for every root $i_j$ as described above.
Let $R = \cap_{j = 1}^k S_{i_j}$ be the set of common colors of the nodes in $C$.
If all nodes in $C$ choose color $c \in R$ then we obtain a social welfare of
$$
\textit{SW}(c) = 2k + \sum_{j = 1}^k d_{i_j}(c).
$$
Let $\textit{SW}_2 = \max_{c \in R} \textit{SW}(c)$. The time needed to compute $\textit{SW}_2$ is at most $O(m^2 |V(T)| + k \cdot m)$.
Clearly, if $\textit{SW}_1 > \textit{SW}_2$ then there is no social optimum in which all nodes of $C$ have the same color. In this case, we choose an arbitrary social optimum. Otherwise, there exists a social optimum in which all nodes of $C$ have a common color. In this case, we choose such a social optimum. Let the resulting social optimum for pseudotree $T$ be $s^*_{T}$.
By proceeding this way for each pseudotree $T$ of the given pseudoforest $G$, we obtain a joint strategy $s^*$ that maximizes the social welfare and the number of unicolored cycles. By Corollary~\ref{cor:one-cycle}, $s^*$ is a strong equilibrium of $G$. The time needed per pseudotree $T$ is dominated by $O(n m^2 |V(T)|)$. The total time needed to compute $s^*$ is thus at most $O(n^2 m^2)$.
\qed
\end{proof}
\iffalse
\section{Extensions}
\label{sec:extensions}
\subsection{Edge weights}
\label{sec:weights}
A natural generalization of our model are coordination games on \emph{weighted} graphs: Given non-negative edge weights $w: E \to \mathbb R_{+}$ and a set $E'$ of edges, we write $w(E')$ for $\sum_{e \in E'} w_e$. We redefine player $i$'s payoff with respect to a joint strategy $s$ by
\[
p_i(s) := w(\{\{i,j\} \in E \mid s_i = s_j\})
\]
The weight $w_{ij}$ reflects how much player $i$ profits from coordinating with $j$. We call the resulting game \emph{weighted coordination game.}
We outline now which of our previous results continue to hold in this extension. Furthermore, we study the relationship of weighted coordination games to additively separable hedonic games.
\paragraph{Existence.}
As one can easily see, $\frac{1}{2}\textit{SW}$ continues to be a potential function for weighted coordination games, guaranteeing the existence of Nash equilibrium. This result has previously been shown for the special case of additively separable hedonic games in \cite{Bogo}.
However, the existence results of strong equilibria (Theorems \ref{thm:color_forest} and \ref{thm:pseudoforest}) and 2-equilibria (Corollary \ref{cor:2-eq}) do not hold anymore.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[scale = 2]
\node [label={\footnotesize $\{a,c\}$}] (v_0) at (0, 1) {$v_0$};
\node [label=below:{\footnotesize $\{a,b\}$}] (v_1) at (-1, 0) {$v_1$};
\node [label=below:{\footnotesize $\{b,c\}$}] (v_2) at (1, 0) {$v_2$};
\node [label=below:{\footnotesize $\{b\}$}] (u) at (-2, 0) {$u$};
\draw (v_1) -- node[auto]{$4$} (v_0);
\draw (v_0) -- node[auto]{$3$} (v_2);
\draw (v_1) -- node[auto]{$2$} (v_2);
\draw (u) -- node[auto]{$3$} (v_1);
\end{tikzpicture}
\caption{A weighted coordination game with no 2-equilibrium.}\label{fig:weighted-non-existence}
\end{center}
\end{figure}
\begin{theorem}
There is a weighted coordination game with no 2-equilibrium on a graph $G$ that is both a pseudoforest and a color forest.
\end{theorem}
\begin{proof}
Consider the coordination game depicted in Figure \ref{fig:weighted-non-existence}. Clearly, the graph is both a pseudoforest and a color forest. Assume for a contradiction that there is a 2-equilibrium $s$.
At least one of the edges of the upper triangle is unicolored in $s$ as otherwise for example the coalition $\{v_0, v_2\}$ could profitably deviate. By the choice of the color assignment, at most one of these edges can be unicolored. So let $\{v_i, v_{i+1}\}$ be the unique unicolored edge (where we calculate modulo 3). Then the coalition $\{v_{i+1}, v_{i+2}\}$ can profitably deviate by switching to the unique color they can both choose (resp., if $v_{i+2}$ already chooses this color in $s$, then $v_{i+1}$ alone can profitably deviate). So $s$ is not a 2-equilibrium. \qed
\end{proof}
By adding the missing edge between the nodes $u$ and $v_2$ in Figure \ref{fig:weighted-non-existence} and giving it a sufficiently small weight $\varepsilon > 0$, we can obtain a coordination game on a color complete graph that does not admit a 2-equilibrium. Hence Corollary \ref{cor:color_complete} does not carry over either.
\paragraph{Inefficiency.}
On the other hand, Theorem \ref{thm:kpoa} continues to hold. This is trivial for the lower bounds. The proof for the upper bounds can be generalized as follows. We replace all calculations using the cardinality of edge sets by calculations using their weights. For example, $N^K_j(s)$ is redefined as
\[
N^K_j(s) := w(\{\{i,j\} \in E \mid i \in K,\, s_i = s_j\})
\]
instead of $|\{\{i,j\} \in E \mid i \in K,\, s_i = s_j\}|$.
Then the same arguments as in the proof of Theorem \ref{thm:kpoa} continue to hold.
\paragraph{Computation.}
Consider the adapted version of the $k$-\textsc{Equilibrium} problem, which we call \textsc{Weighted $k$-Equilibrium}. It is easy to see that this problem, too, is in co-NP. Because weighted coordination games are a generalization of coordination games, the hardness result carries over. Hence \textsc{Weighted $k$-Equilibrium} is co-NP-complete.
\paragraph{Relation to hedonic games.} Weighted coordination games bear a similarity to \emph{additively separable hedonic games (ASHG)} \cite{Baner,Bogo}. In a certain sense, non-negative symmetric ASHG can be seen as a subclass of weighted coordination games, which in turn can modeled as a subclass of symmetric ASHG. The details are given in the Appendix.
\subsection{Choosing sets of colors}
The second natural generalization of coordination games is to allow players to choose non-empty \emph{sets} of colors, i.e., $S_i \subseteq 2^M \setminus \{\emptyset\}$,
and define
\[
p_i(s) := \sum_{j \in N_i} |s_i \cap s_j|.
\]
Here each player derives a utility of one per color that he shares with her neighbors. We call these games \emph{multi-color} coordination games. Clearly, coordination games are a special case in which each $S_i$ only consists of singletons.
\paragraph{Relation to weighted extension.}
Multi-colored coordination games in fact generalize weighted coordination games if the edge weights are integers.
\begin{theorem}
Let $\mathcal{G}^w = (G, A, w)$ be a weighted coordination game on a graph $G = (V, E)$ with color assignment $A$ and edge weights $w: E \rightarrow \mathbb{N}$. Then there exists a multi-color coordination game $\mathcal{G}^m = (G, \tilde{A})$ that is isomorphic to $\mathcal{G}^w$.
\end{theorem}
\begin{proof}
We construct the multi-color coordination game $\mathcal{G}^m$ as follows: We use the same graph, but a new set of colors
$$
\tilde M := \{x_{ij}^l \mid x \in M, \, \{i,j\} \in E,\ l \in \{1, \dots, w_{ij}\}\},
$$
where the colors $x_{ij}^l$ are pairwise different. Let player $i$'s new strategy set be
\[
\tilde A_i := \{\tilde x_i \mid x \in A_i\} \subseteq 2^{\tilde M}
\quad
\text{with}
\quad
\tilde x_i = \{x_{ij}^l \mid j \in N_i, \, l \in \{1, \dots, w_{ij}\}\} \subseteq \tilde M.
\]
Consider the multi-color coordination game $\mathcal G^m$ specified by $(G, \tilde A).$ We claim that the function that maps $x$ to $\tilde x$ is an isomorphism.
To see this, consider a joint strategy $s$ of $\m G^w$ and let $\tilde s$ be the corresponding joint strategy of $\m G^m.$ Then player $i$'s payoff in $\m G^w$ for $s_i = x$ is
\begin{align*}
\sum_{j \in N_i:\, s_j = x} w_{ij} = \sum_{j \in N_i:\, s_j = x} |\tilde s_i \cap \tilde s_j| = \sum_{j \in N_i} |\tilde s_i \cap \tilde s_j|,
\end{align*}
which is exactly his payoff under $\m G^m$.
\qed
\end{proof}
Note that this transformation runs in pseudopolynomial but not in polynomial time.
\paragraph{Existence.}
As before, $\frac{1}{2}\textit{SW}$ is an ordinal potential, showing that Nash equilibria always exist. Since multi-colored coordination games generalize weighted coordination games if the weights are integer, it follows from the previous section that there is a multi-colored coordination game on a pseudoforest without a strong equilibrium; hence Theorem \ref{thm:pseudoforest} does not continue to hold.
\paragraph{Inefficiency.} The inefficiency results (Theorem \ref{thm:kpoa}) carry over. Again, we reuse the initial proof for the upper bounds. This time we replace all calculations using the cardinality of edge sets $E'$ by calculations adding up $|s_i \cap s_j|$ for each edge $\{i,j\} \in E'$. For example, $N^K_j(s)$ is redefined as
\[
N^K_j(s) := \sum_{\{i,j\} \in E: \, j \in K} |s_i \cap s_j|
\]
instead of $|\{\{i,j\} \in E \mid i \in K,\, s_i = s_j\}|$.
Then the same arguments as in the proof of Thorem \ref{thm:kpoa} continue to hold.
\paragraph{Computation.}
As before, one can easily see that the adapted version of \textsc{Equilibrium} is co-NP-complete.
\subsection{A note on super-strong equilibria}
Recall that a joint strategy $s$ is a \emph{super strong equilibrium} if for all coalitions $K$ there does not exist $s'_K \in S_K$ such that $p_i(s') \geq p_i(s)$ for all $i \in K$ and $p_i(s') > p_i(s)$ for some $i \in K$, where $s' = (s'_K, s_{-K})$.
Note that studying super strong equilibria of our coordination games is not a promising direction because even in the simplest cases they do not exist. Indeed, take for instance a path with two edges and nodes having color
sets $\{a\}, \ \{a, b\}, \ \{b\}$, respectively, or a triangle with the respective color sets $\{a,b\}, \ \{b,c\}$ and $\{a,c\}$.
In both cases a super strong equilibrium does not exist.
\fi
\section{Conclusions}
\label{sec:conc}
We introduced and studied a natural class of games which we termed coordination games on graphs. We provided results on the existence, inefficiency and computation of strong equilibria for these games.
It would be interesting to prove existence of $k$-equilibria for other graph classes and to investigate the computational complexity of computing them.
Another open question is to determine the (strong) price of anarchy
when the number of colors is fixed. Yet another intriguing question
is for which $k \geq 2$ coordination games with transition value $k$
exist. In Section 4.5 we settled this question positively only for $k
= 2$ and $k = 4$. In the future we also plan to study a natural
extension of our coordination games to hypergraphs.
Another natural question that comes to one's mind is whether \emph{super strong equilibria} exist. Recall that a joint strategy $s$ is a \emph{super strong equilibrium} if for all coalitions $K$ there does not exist a deviation $s' = (s'_K, s_{-K})$ such that $p_i(s') \geq p_i(s)$ for all $i \in K$ and $p_i(s') > p_i(s)$ for some $i \in K$. It is not hard to verify that super strong equilibria are not guaranteed to exist: Consider a path consisting of two edges and assume that the nodes have color sets $\{a\}, \ \{a, b\}$ and $\{b\}$, respectively. Clearly, a super strong equilibrium does not exist for this instance.
A natural generalization of our model are coordination games on weighted graphs. Here each edge $\{i, j\}$ has a non-negative weight $w_{ij}$ specifying how much player $i$ and $j$ profit from choosing the same color. It is easy to see that $\frac{1}{2}\textit{SW}$ continues to be an exact potential function for weighted coordination games, guaranteeing the existence of a Nash equilibrium. In fact, as observed in \cite{CD11}, this is an exact potential for coordination games with arbitrary weights. Coordination games on weighted graphs are studied in more detail in \cite{rahn:schaefer:2015}. In particular, the existence results for strong equilibria (Theorems \ref{thm:color_forest} and \ref{thm:pseudoforest}) and 2-equilibria (Corollary \ref{cor:2-eq}) do not hold for these games. We refer the reader to \cite{rahn:schaefer:2015} for further studies of these games.
Another natural variation is to consider coordination games on
weighted \textit{directed graphs}. Given a directed graph
$G=(V,E)$, we say that node $j$ is a \textit{neighbour} of node
$i$ if there is an edge $(j,i)$ in $G$. Each edge $(j,i)$ has a
non-negative weight $w_{ji}$ specifying how much player $i$ profits
from choosing the same color as player $j$. The transition from
undirected to directed graphs changes the status of the games
substantially. In particular, Nash equilibria need not always exist
in these games. Moreover, the problem of determining the existence
of Nash equilibria is NP-complete. We refer the reader to
\cite{ASW15} and \cite{SW16} for further studies of these games.
\subsection*{Acknowledgements}
The fact that for the case of a ring the coordination game has the
c-FIP was first observed by Dariusz Leniowski. We thank Jos\'e Correa
for allowing us to use his lower bound in Theorem~\ref{thm:kpoa}. It
improves on our original one by a factor of $2$. We thank the
anonymous reviewers for their valuable comments. First author is also
a Visiting Professor at the University of Warsaw. He was partially
supported by the NCN grant nr 2014/13/B/ST6/01807.
\bibliographystyle{spmpsci}
|
\section{Introduction}
A major motivation for classifying mathematical objects is to produce interesting new examples which would not have been found without an extensive search. For example, it was the classification of simple Lie algebras which uncovered $E_8$ and the classification of finite simple groups which revealed the Monster group. Similarly, a major motivation in the classification of small index subfactors is to find new interesting examples.
To any finite index finite depth subfactor $N \subset M$ one can assign a rich algebraic invariant called its standard invariant. One way to describe the standard invariant is that it consists of a pair of a unitary fusion category $\mathcal{C}$ (consisting of certain $N$-$N$ bimodules and called the principal even part ) and an algebra object $A \in \mathcal{C}$ (which is the bimodule $M$). Furthermore, for finite index finite depth subfactors of the hyperfinite $II_1$ factor, the standard invariant is a complete invariant \cite{ MR1055708}. Almost all known unitary fusion categories can be constructed from finite groups or from quantum groups at roots of unity. Indeed, all subfactors of index less than $4$ have standard invariants coming from quantum $SU(2)$ via conformal inclusions, and all finite depth subfactors of index equal to $4$ have standard invariants which come from finite subgroups of $SU(2)$. It is natural to wonder whether the subfactors of index between $4$ and $5$ come from similar constructions.
There are 5 pairs of finite index finite depth subfactor standard invariants of index between $4$ and $5$:
\begin{itemize}
\item the Haagerup subfactor (and its dual),
\item the Asaeda--Haagerup subfactor (and its dual),
\item the extended Haagerup subfactor (and its dual),
\item the Goodman-de la Harpe-Jones 3311 subfactor (and its dual),
\item and the self-dual Izumi 2221 subfactor (and its complex conjugate).
\end{itemize}
\noindent The last two of these pairs come from quantum groups in an appropriate sense.
The remaining three seem to not be related to finite groups or quantum groups via well-understood constructions. Thus they are very important examples for further study since they could be ``exotic'' or ``exceptional.''
Of these three examples, the Haagerup subfactor is by far the best understood. Three different constructions of this subfactor standard invariant have been given, each of a different flavor \cite{MR1686551,MR1832764,0902.1294}. One of these constructions allows for practical calculations, for example of the $S$ and $T$ matrices of the Drinfel'd center of the even part \cite{MR1832764}. Furthermore, the principal even part of the Haagerup subfactor has a collection of invertible objects forming the group $\mathbb{Z}/3\mathbb{Z}$, and all non-invertible objects in this even part are in a single orbit of the action of $\mathbb{Z}/3\mathbb{Z}$. On the one hand, this suggests that there may be a sense in which the Haagerup subfactor can be constructed from the group $\mathbb{Z}/3\mathbb{Z}$ (or a closely related group like $S_3$, see \cite{MR2837122}), and on the other hand it suggests that the Haagerup subfactor could be generalized by replacing $\mathbb{Z}/3\mathbb{Z}$ by other finite abelian groups. We call such a subfactor a $3^A$ subfactor, since its principal graph is a star graph with $\#A$ arms where each arm has $3$ edges. The second-named author determined a numerical invariant for such subfactors given by solutions to certain polynomial equations, and by solving these equations he constructed a $3^{\mathbb{Z}/5\mathbb{Z}}$ subfactor \cite{MR1832764}. Evans--Gannon constructed $3^G$ subfactors for several other cyclic groups of odd order by solving the polynomial equations using symmetries of the associated modular data \cite{MR2837122}. Although there are still only finitely many examples known, it appears that that the Haagerup subfactor may not be exceptional after all.
The main goal of this paper is to establish a similar story for the previously mysterious Asaeda--Haagerup subfactor. The basic philosophy of our approach is that one should never study a single subfactor at a time, but instead should study all the subfactors which come from inclusions of algebra objects $A \subset B$ in a fixed unitary fusion category $\mathcal{C}$. In other words, one fixes a finite collection of bimodules over a fixed factor $R$ and looks at all subfactors $N \subset M$ where each of $N$ and $M$ can be built as direct sums of these particular bimodules. All of these subfactors fit together into a richer structure known as the maximal atlas or the Brauer--Picard groupoid \cite{MR2677836}. This approach has two advantages. First, the combinatorics of this whole collection is richer and more restrictive than the combinatorics of the individual subfactors. For example, just by looking at fusion rules one can show that certain subfactors must exist which would be very difficult to construct directly, as illustrated in \cite{GSbp}. Second, it may be that some of the other subfactors in the maximal atlas are simpler than others. In this paper we we emphasize the second point of view, and we relate the Asaeda--Haagerup subfactor to a simpler subfactor also appearing in its maximal atlas. That is, the complicated small index Asaeda--Haagerup subfactor is best thought of as a consequence of a simpler subfactor with larger index. This gives a new construction of the the Asaeda--Haagerup subfactor, which still requires a difficult computation, but which is much more illuminating in terms of understanding this subfactor.
Here is a quick overview of our new construction of the Asaeda--Haagerup subfactor. We first construct a specific $3^{\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ subfactor $\mathcal{S}$ by finding a solution to the appropriate polynomial equations, which were generalized to finite Abelian groups of possibly even order in \cite{IzumiNote}.
The subfactor $\mathcal{S}$ has index $5+\sqrt{17}$. Then we take an orbifold (or de-equivariantization) quotient of $\mathcal{S}$ by adding an extra isomorphism between the objects corresponding to the trivial and non-trivial elements of $\mathbb{Z}/2\mathbb{Z}$. This new subfactor $\mathcal{S}'$ has index $5+\sqrt{17}$, and we call its even part $\mathcal{AH}_4$. We then explicitly calculate the fusion rules for the dual even part of this subfactor (i.e. the bimodules over the larger factor) and see that it contains an object $X$ with dimension $\frac{3+\sqrt{17}}{2}$ such that $X \otimes X \cong 1 \oplus X \oplus Y$ for some non-invertible simple object $Y$. By a simple skein theory argument \cite[Thm 3.4]{GSbp}, the object $1\oplus X$ has a unique algebra structure and this algebra gives a new subfactor with index $\frac{5+\sqrt{17}}{2}$. This subfactor must be the Asaeda--Haagerup subfactor $\mathcal{AH}$, since it is easy to see that there's at most one finite depth subfactor of index $\frac{5+\sqrt{17}}{2}$. Altogether, the Asaeda--Haagerup subfactor comes as a natural consequence of the more symmetric and easier to understand subfactor $\mathcal{S}$.
\begin{figure}
$$\begin{tikzpicture}
\node (N) at (0,0) {$N$};
\node (M) at (1.5,1.5) {$M$};
\node (P) at (-1,1) {$P$};
\node (Q) at (.5,2.5) {$Q$};
\node (R) at (.5,4) {$R$};
\path (N) edge node[below,pos=.7] {$\mathcal{S}$} (M);
\path (N) edge node[below,pos=.8] {$A_3$} (P);
\path (P) edge node[above, pos=.5] {$\mathcal{S}'$} (Q);
\path (M) edge node[above=.1cm, pos=0] {$A_3$}(Q);
\path (Q) edge node[right,pos=.5]{$\mathcal{AH}$} (R);
\end{tikzpicture}
$$
\caption{Our construction of the Asaeda--Haagerup subfactor can be summarized by the above inclusions of factors, where $\mathcal{S}$ denotes our new index $5+\sqrt{17}$ subfactor, $\mathcal{S}'$ is its orbifold quotient, $\mathcal{AH}$ is the Asaeda--Haagerup subfactor of index $\frac{5+\sqrt{17}}{2}$, and $A_3$ denotes index $2$ inclusions.}
\label{fig:outline}
\end{figure}
In particular, we see that $\mathcal{AH}_4$ lies in the same higher Morita equivalence class as the even parts of the Asaeda--Haagerup subfactor $\mathcal{AH}_1$ and $\mathcal{AH}_2$. Many invariants of fusion categories, most notably the Drinfel'd center, are invariant under higher Morita equivalence. Thus instead of doing a calculation in $\mathcal{AH}_1$ or $\mathcal{AH}_2$, one can instead do the calculation in $\mathcal{AH}_4$. In particular, in a follow-up paper we will give an explicit description of the Drinfel'd center of the Asaeda--Haagerup fusion categories (a question which had remained open for 15 years) by following the method in \cite{MR1832764}. Similarly, since Etingof-Nikshych-Ostrik's classification of $G$-extensions of fusion categories is Morita invariant, constructing extensions of $\mathcal{AH}_4$ yields extensions of the other fusion categories in its Morita class. More specifically, in another follow-up paper we can determine the homotopy type of the Brauer--Picard $3$-groupoid of the Asaeda--Haagerup category, which splits as a product of Eilenberg-Maclane spaces $K(\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}, 1) \times K(\mathbb{C}^\times, 3)$. As a consequence, we get a plethora of interesting $\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$ extensions of Asaeda--Haagerup fusion categories generalizing the constructions in \cite{GJSext}, and we can use this to give a complete classification of all extensions of these fusion categories by any group.
It is natural to wonder why the Asaeda--Haagerup subfactor, which has only a $2$-fold symmetry, should be related to an $8$-fold symmetric $3^{\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ subfactor. In fact, we discovered this connection by working in the other direction: starting with the Asaeda--Haagerup subfactor and trying to describe its entire higher Morita equivalence class. An extensive combinatorial calculation suggested that $\mathcal{AH}_4$ was one of the only possible new tensor categories compatible with all the rich structure of the Asaeda--Haagerup subfactor. If $\mathcal{AH}_4$ does exist, one can then determine that it must have an equivariantization which would give a $3^{\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ subfactor. Please note that logically our construction does not require the extensive combinatorial calculations from \cite{GSbp}, but the motivation to consider $\mathcal{S}$ in the first place did come out of those calculations.
In fact, constructing $\mathcal{AH}_4$ has allowed us to complete our description of the higher Morita equivalence class of the Asaeda--Haagerup subfactors. Namely, there are exactly $6$ fusion categories, and between any two of them there are exactly $4$ Morita equivalences between them. As shown in \cite{GSbp}, the group of Morita autoequivalences is the Klein $4$-group. In fact, for $\mathcal{AH}_4$ all these autoequivalences are realized by outer automorphisms. As a consequence of this complete description of the higher Morita equivalence class, we are able to answer many other questions about the subfactors related to the Asaeda--Haagerup subfactors. For example, in a followup paper we will use this classification to find all lattices of intermediate subfactors related to Asaeda--Haagerup.
The outline of this paper is as follows. We begin in Section \ref{sec:Prelim} with some background on fusion categories, algebra objects, the Brauer--Picard groupoid, and the Asaeda--Haagerup subfactor. This section includes some expository material summarizing the key ideas of our project of understanding subfactors via the Brauer--Picard groupid.
In Section \ref{sec:BP} we further describe the Brauer--Picard groupoid of the Asaeda--Haagerup fusion categories following \cite{GSbp}. In \cite{GSbp} we saw that there were at least three fusion categories in this Morita equivalence class, that there were exactly four bimodules between any two such fusion categories, and that the BP group was the Klein $4$-group. Here we narrow things down further, showing that if $\mathcal{AH}_4$ exists then there are exactly six fusion categories in the equivalence class. These calculations are combinatorial in nature and are quite similar to those in \cite{GSbp}. However, we would like to remind the reader that our new construction of $AH$ is {\em motivated} by the results of Section \ref{sec:BP}, but strictly speaking does not {\em depend} on the results of Section \ref{sec:BP}.
Section \ref{sec:main} is the main heart of the paper. We give a direct construction of $\mathcal{S}$ as endomorphisms of an algebra using results of \cite{IzumiNote}. We then show that it has a de-equivariantization $\mathcal{AH}_4$, and show that the Asaeda--Haagerup subfactor comes from an inclusion of algebras in $\mathcal{AH}_4$.
In Section \ref{sec:apps}, we quickly sketch several applications of our main results. The full details of these applications will appear in later papers.
\textbf{Acknowledgements:}
The authors would like to thank Scott Morrison, David Penneys, and Emily Peters for helpful conversations. Pinhas Grossman was partially supported by ARC grant DP140100732. Masaki Izumi was supported in part by the Grant-in-Aid for Scientific Research (B) 22340032, JSPS. Noah Snyder was supported by DOD-DARPA grant HR0011-12-1-0009. The authors would also like to thank the American Institute of Mathematics for its hospitality during the workshop \textit{Classifying Fusion Categories}.
\section{Preliminaries}\label{sec:Prelim}
We begin with some background on fusion categories, subfactors, the Brauer--Picard groupoid, combinatorics of fusion rules, and the Asaeda--Haagerup subfactor. We assume that the reader is either familiar with the theory of fusion categories or the theory of subfactors. Thus the first subsection is aimed at readers who are familiar with subfactors but not with fusion categories, and the second subsection is aimed at readers familiar with fusion categories but not subfactors.
\subsection{Fusion categories, module categories, and bimodule categories}
\begin{definition} \cite{MR2183279}
A fusion category over an algebraically closed field $k $ is a semisimple $k$-linear rigid monoidal
category with finitely many simple objects and finite-dimensional morphism spaces such that the identity object is simple.
\end{definition}
In this paper $k$ will always be the field of complex numbers. An equivalence of fusion categories is a linear monoidal equivalence.
\begin{example}
Fix a finite collection of bifinite bimodules $M_i$ over a II${}_1$ factor $A$ which are closed under tensor product in the sense that $M_i \otimes_A M_j \cong \bigoplus_k M_k^{\oplus n_k}$. Then there is a fusion category whose objects are the direct sums of the $M_i$, the morphisms are bimodule maps, and the tensor product is $\otimes_A$.
\end{example}
A fusion category can be thought of as a higher analogue of an algebra, where instead of multiplying elements you tensor objects. Just as modules and bimodules play a crucial role in understanding algebras, module categories and bimodule categories play a similar role in understanding fusion categories.
\begin{definition}
A (left) module category ${}_{\mathcal{C}}{\mathcal{M}} $ over a fusion category $\mathcal{C} $ is a
category $\mathcal{M} $ along with a biexact bifunctor from $\mathcal{C} \times \mathcal{M}$ to $ \mathcal{M}$, along with a collection of unit and associativity isomorphisms satisfying certain coherence relations; see \cite{MR1976459} for details. In this paper we will further assume that module categories over fusion categories are semisimple.
\end{definition}
\begin{example}
If $\mathcal{C}$ comes from a collection of bimodules over $A$ as in the previous example, then one can build a left module category in the same way by finding a finite collection of simple $A$-$B$ bimodules for a factor $B$ which are closed under left multiplication by the $A$-$A$ bimodules.
\end{example}
Similarly, one may define right module categories over fusion categories and bimodules categories over pairs of fusion categories. There is a natural notion of equivalence for module or bimodule categories.
\begin{definition}
A bimodule category over a pair of fusion categories ${}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}}$ is invertible if ${}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}} \boxtimes_{\mathcal{D}} {}_{\mathcal{D}}{\mathcal{M}}^{op} {}_{\mathcal{C}} \cong {}_{\mathcal{C} } {\mathcal{C} } {}_{\mathcal{C} }$ and ${}_{\mathcal{D}}{\mathcal{M}^{op}} {}_{\mathcal{C}} \boxtimes_{\mathcal{C}} {}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}} \cong {}_{\mathcal{D} } {\mathcal{D} } {}_{\mathcal{D} }$, where $\mathcal{M}^{op} $ is the opposite bimodule category and $\boxtimes_{\mathcal{C}} $ and $ \boxtimes_{\mathcal{D}}$ are the relative tensor products of bimodule categories; see \cite{MR2677836} for details.
Two fusion categories are Morita equivalent if there is an invertible bimodule category between them; an invertible bimodule category is called a Morita equivalence.
\end{definition}
\begin{example}
If $A \subset B$ is a finite depth subfactor then the principal even part (certain $A$-$A$ bimodules) and the dual even part (certain $B$-$B$ bimodules) are Morita equivalent with the Morita equivalence given by a collection of $A$-$B$ bimodules.
\end{example}
\begin{example} \label{ex:standardmorita}
If $G$ is a finite group, let $\mathrm{Vec}(G)$ be the category of $G$-graded vector spaces and $\mathrm{Rep}(G)$ be the category of finite dimensional representations of $G$. Then $\mathrm{Vec}$ gives a Morita equivalence between $\mathrm{Vec}(G)$ and $\mathrm{Rep}(G)$ where the individual actions are given by forgetting and tensoring, but the associator for simple objects is $V = (1_g \otimes 1) \otimes V \rightarrow 1_g \otimes (1 \otimes V) = V$, with the map given by the action of $g$ on V.
This Morita equivalence can be understood from the subfactor point of view by looking at the crossed product subfactor $N \subset M = N \rtimes G$ and thinking of $\mathrm{Vec}(G)$ as $N$-$N$ bimodules, $\mathrm{Vec}$ as $N$-$M$ bimodules, and $\mathrm{Rep}(G)$ as $M$-$M$ bimodules.
\end{example}
Any simple module category over a fusion category ${}_{\mathcal{C}} \mathcal{M} $ can be given the structure of an invertible
bimodule category ${}_{\mathcal{C}} \mathcal{M} {}_{\mathcal{D}} $, where $\mathcal{D} =({}_{\mathcal{C}} \mathcal{M} )^*$ is the dual category of module endofunctors of ${}_{\mathcal{C}} \mathcal{M} $. Conversely, for any invertible bimodule category ${}_{\mathcal{C}} \mathcal{M} {}_{\mathcal{D}} $ we have $\mathcal{D} \cong ({}_{\mathcal{C}} \mathcal{M} )^*$.
Module categories can also be characterized in terms of algebras. This can be thought of as an algebraic substitute for subfactor theory where the role of factors is played by algebra objects.
\begin{definition}
An algebra in a fusion category is an object $A$ together with a unit map $I \rightarrow A $
and a multiplication map $A \otimes A \rightarrow A$ satisfying the usual relations; see \cite{MR1976459}.
\end{definition}
\begin{example}
If $N \subset M$ is a finite index finite depth subfactor, and $\mathcal{C}$ is the category of $N$-$N$ bimodules generated by $M$, then $M$ itself is an algebra object in $\mathcal{C}$. In this case, saying that $M$ is an algebra object means both that it is an algebra and that the multiplication map $M \otimes M \rightarrow M$ is a map of $N$-$N$ bimodules.
\end{example}
In a similar way one can define modules over algebras in fusion categories. If $A$ is an algebra
in $\mathcal{C} $, then the category $A $-mod of left $A$-modules in $ \mathcal{C}$ is a right module category over $\mathcal{C} $ (although not necessarily semisimple), and similarly the category of right $A$-modules is a left module category.
An algebra $A$ is called simple if its module category $A$-mod is semisimple and indecomposable, and is called a division algebra if in addition $A$ is simple as an $A$-module \cite{GSbp}. If $A$ is a simple algebra in a fusion category, then the category $A$-mod-$A$ of $A-A$ bimodules is the dual category of $A$-mod. In fact every indecomposable module category arises this way. One can define the internal hom bifunctor $\underline{\text{Hom}}(M_1,M_2) $ from ${}_{\mathcal{C}} \mathcal{M} \times {}_{\mathcal{C}} \mathcal{M} $ to $\mathcal{C} $. The internal hom satsfies $\text{Hom}(\underline{\text{Hom}} (M_1,M_2),X)\cong \text{Hom}(X \otimes M_1,M_2)$ for all $M_1,M_2 \in {}_{\mathcal{C}} \mathcal{M}$ and
$X \in {}_{\mathcal{C}}$, and the internal end $\underline{\text{End}} (M)=\underline{\text{Hom}}(M,M) $ is an algebra for every $M \in {}_{\mathcal{C}} \mathcal{M} $.
\begin{theorem}\cite{MR1976459}
Let $M \in {}_{\mathcal{C}} \mathcal{M} $ be a simple object is a simple module category over a fusion category. Then $ {}_{\mathcal{C}} \mathcal{M} $ is equivalent to the category of modules
over $ \underline{\text{End}}(M)$ in $\mathcal{C} $.
\end{theorem}
\begin{example}
If $M$ comes from an $A$-$B$ bimodule for two factors, then $\underline{\text{End}} (M)= M \otimes_B \bar{M}$ where $\bar{M}$ is the contragradient $B$-$A$ bimodule.
\end{example}
When $M$ is a simple object, the algebra $ \underline{\text{End}}(M)$ is a division algebra \cite{GSbp}.
\begin{theorem}\cite{GSbp}
The algebras $\underline{\text{End}}(X)$ and $\underline{\text{End}}(Y)$ for two objects $X$ and $Y$ in ${}_\mathcal{C} \mathcal{M}$ are isomorphic if and only if $X \cong Y g$ for some invertible object $g$ in the dual category $({}_{\mathcal{C}} \mathcal{M} )^*$.
\end{theorem}
\subsection{Subfactors}
The theory of standard invariants of finite-index subfactors has been developed both for Type II$_1$ and for properly infinite factors. The classes of standard invariants which arise in the two settings are the same, so the choice of setting is in some sense a matter of taste; however certain types of calculations or constructions may be easier in one setting or the other. In this paper we use infinite factors in order to utilize a construction of subfactors from endomorphisms of Cuntz algebras introduced by the second-named author in \cite{MR1228532}, which uses Type III factors.
There are two standard ways of building tensor categories related to an algebra $A$. The first is familiar algebraically, namely one considers a collection of bimodules over $A$ which is closed under tensor product $\otimes_A$. When $A$ is a von Neumann algebra, one needs to be a bit careful about whether one is considering Hilbert bimodules with the Connes fusion product or algebraic bimodules with the algebraic tensor product. Happily, in the finite index setting for II$_1$ factors, either choice leads to the same tensor category \cite{MR2501843}.
The second construction is called the category of sectors and is less familiar to algebraists. Nonetheless it has a nice categorical description which we give here. If $\mathcal{C}$ is a linear category then there is a monoidal category whose objects are the tensor automorphisms of $\mathcal{C}$ and whose morphisms are tensor natural transformations. The simplest kind of category is one with only one object $M$. In this setting the monoidal category of functors and natural transformations is called the category of sectors and written $\mathcal{C}(M) $.
If the underlying object is an algebra $M$, then $\mathcal{C}(M) $ can be described explicitly as follows. The objects are the endomorphisms of $M$ and the morphisms are given by
$$\text{Hom}(\rho, \sigma)=\{ v \in M : v \rho(x) = \sigma(x) v, \ \forall x \in M \} $$
for $\rho, \sigma \in \text{End}(M) $. The composition of morphisms is given by multiplication in $M$, and the tensor product of objects is given by composition of endomorphisms. Finally, the tensor product of morphisms is given by the asymmetrical formula determined by composition of natural transformations, if $v \in (\rho, \sigma)$ and $u \in (\alpha, \beta)$ then $u \otimes v \in (\alpha \circ \rho, \beta \circ \sigma)$ is given by $\beta(v) u$.
If $M$ is a Type III factor with separable predual, then any nice bimodule comes from an endomorphism, and so the category of sectors for $M$ is especially appropriate.
We use some standard definitions. We consider the category $\mathcal{C}_0(M) $ of finite-index unital normal $*$-endomorphisms of $M$.
A sector in $\mathcal{C}_0(M) $ is an isomorphism class of objects; the sector associated
to an endomorphism $\rho \in \text{End}_0(M) $ is denoted by $[\rho] $. We often simply write $ (\rho, \sigma)$ for $\text{Hom}(\rho, \sigma)$.
For a self-dual endomorphism $\rho \in \text{End}_0(M) $, let $\mathcal{C}_{\rho} $
be the tensor category tensor generated by $\rho $; i.e. it is the full subcategory of
$\mathcal{C}_0(M) $ whose objects are exactly those which are contained in some tensor power
of $\rho $. We say that $\rho $ has finite depth if $\mathcal{C}_{\rho}$ has finitely many simple objects up to isomorphism; in this case $\mathcal{C}_{\rho}$ is a fusion category.
We will often use sector notation for arbitrary tensor categories and bimodule categories, so objects are denoted by lowercase Greek letters, tensor product symbols are suppressed and $(\kappa, \lambda) := \text{dim}(\text{Hom}( \kappa,\lambda))$.
\begin{definition}\cite{MR1257245}
A Q-system is an algebra $\gamma$ in $\mathcal{C}_0(M) $ such that the unit map
$R \in (id,\gamma) $ is an isometry and the multiplication map $T \in (\gamma^2,\gamma) $ is a co-isometry.
\end{definition}
If $\gamma $ is a Q-system in $\mathcal{C}_0(M) $, then there is another properly infinite factor $N$ and a pair of finite-index unital $*$-homomorphisms $\iota: N \rightarrow M $ and $\bar{\iota}:
M \rightarrow N $ such that $\iota \bar{\iota} = \gamma $. Then $\sigma =\bar{\iota} \iota $ is a Q-system
in $\mathcal{C}_0(N) $. The homomorphisms $ \rho^n \iota, n \in \mathbb{N} $ generate a subcategory $\mathcal{M}_{\iota} $ of the category $\text{Hom}_0(N,M)$ of finite-index homomorphisms from $N$ to $M$ (where morphisms are defined exactly as for $\text{End}_0 $). Then $\mathcal{M}_{\iota} $ is an invertible $\mathcal{C}_{\gamma}-\mathcal{C}_{\sigma}$ bimodule category where the tensor product is given by composition. If $\gamma $ is a finite-depth endomorphism, then $\mathcal{C}_{\gamma}$ is a fusion category, $\mathcal{M}_{\iota} $ is equivalent to the
category of $\gamma=\underline{\text{End}} (\iota)$ modules in $\mathcal{C}_{\gamma} $, and $\mathcal{C}_{\sigma}$ is then equivalent to the category of $\gamma-\gamma $ bimodules in $\mathcal{C}_{\gamma} $.
Conversely, if $N \subseteq M $ is a finite-index subfactor with inclusion map $\iota: N \rightarrow M$, then there is a dual homomorphism $\bar{\iota}: M \rightarrow N$ such that $\sigma=\bar{\iota}\iota $ is a Q-system in $\mathcal{C}_0(N)$, called the canonical endomorphism, and $ \gamma= \iota \bar{\iota} $ is a Q-system in $\mathcal{C}_0(M)$, called the dual canonical endomorphism. The subcategories of $\mathcal{C}_0(N)$ and $\mathcal{C}_0(M)$ tensor generated by $\sigma $ and $\gamma $ are called the principal and dual even parts of the subfactor, respectively.
Finally, we recall the following calculation from \cite{ MR2418197}. Suppose $\gamma \cong id \oplus \rho $ is a finite-index endomorphism with $\rho $ a self-conjugate irreducible endomorphism not isomorphic to the identity, and let $R \in (id,\rho^2 )$ be an isometry. Then Q-systems for $\gamma $ correspond to isometries $S \in (\rho,\rho^2 )$ satisfying
$$\rho(S)R=SR, \quad \sqrt{d}R+(d-1)S^2=\sqrt{d}\rho(R)+(d-1)\rho(S)S,$$
where $d=d(\rho)=[M:\rho(M)]^{\frac{1}{2}} $ is the statistical dimension of $\rho $. Two such Q-systems corresponding to $S$ and $S'$ are equivalent iff $S=\pm S' $.
\subsection{The Brauer--Picard groupoid and related higher structures}
There are several closely related higher categorical constructions important to the study of subfactors, which we describe in this section. All of them are motivated by trying to study not just a single division algebra in a fusion category $\mathcal{C}$, but all division algebras at once. Ocneanu called this the ``maximal atlas'' and it can be formalized in several ways.
The simplest construction is to look at all division algebras in $\mathcal{C}$ (or, more generally, the Frobenius algebra objects in $\mathcal{C}$) and all bimodule objects between them. These bimodules form a $2$-category as follows: the objects are division algebras in $\mathcal{C} $, the $1$-morphisms are bimodules between division algebras, and the $2$-morphisms are bimodule maps. The composition of $1$-morphisms in this $2$-category is given by the relative tensor product over the common algebra. Furthermore, this $2$-category is rigid in the sense that every bimodule has a dual bimodule \cite{MR2075605}. This rigidity is often called Frobenius reciprocity. This construction has a strong subfactor flavor. Indeed one can think of the objects of this $2$-category as a certain finite collection of von Neumann algebras, the $1$-morphisms as a finite collection of finite index bimodules between them which are closed under tensor product, and the $2$-morphisms as bimodule maps.
The second construction looks beyond the original fusion category $\mathcal{C}$ and thus is a little more difficult to understand from a pure subfactor point of view. Any division algebra in $\mathcal{C}$ gives a Morita equivalence mod-$A$ between the fusion category $\mathcal{C}$ and the fusion category of $A$-$A$ bimodules. Thus, it is natural to study all bimodule categories between all fusion categories in the Morita equivalence class of $\mathcal{C}$. On the other hand, different algebras can give rise to the same module category, so we can consider a structure that is not defined in terms of specific algebras.
\begin{definition}\cite{MR2677836}
The Brauer-Picard $3$-groupoid of a fusion category $\mathcal{C} $ is the $3$-groupoid whose objects are fusion categories Morita equivalent to $ \mathcal{C}$, whose $1$-morphisms are invertible bimodule categories between such fusion categories, whose $2$-morphisms are equivalences between such bimodule categories, and whose $3$-morphisms are isomorphisms of such equivalences.
This $3$-groupoid can be truncated to an ordinary groupoid whose points are the fusion categories which are Morita equivalence to $\mathcal{C}$ and whose arrows are equivalence classes of Morita equivalences. The group of Morita autoequivalences of a fusion category, considered modulo equivalence of bimodule categories, forms a group, called the Brauer-Picard group. The Brauer-Picard group is an invariant of the Morita equivalence class.
\end{definition}
Typically groupoids are considered up to equivalence. For example the trivial $1$-point groupoid is equivalent to the groupoid with $n$ points and exactly one arrow between any pair of them even though they have a different number of points. We often consider a finer invariant of the Brauer--Picard groupoid. Namely, consider the groupoid whose points are fusion categories up to ordinary equivalence (not Morita equivalence!) and whose arrows are equivalence classes of Morita equivalences. By Ocneanu rigidity \cite{MR2183279}, this gives a groupoid with a finite number of points. When we say something like the Brauer--Picard groupoid has exactly $6$ points and exactly $4$ arrows between each of them, we are referring to this refined version.
An autoequivalence of a fusion category $\mathcal{C} $ is called inner
if it is equivalent as a monoidal functor to conjugation by an invertible object in $\mathcal{C} $. The group $\text{Out}(\mathcal{C}) $ of autoequivalences of a fusion category $\mathcal{C} $ modulo inner autoequivalences is a subgroup of the Brauer-Picard group, via the map sending $\alpha$ to the bimodule ${}_{\mathcal{C}} \mathcal{C} {}_{\alpha(\mathcal{C})}$ where the right action is twisted by $\alpha$.
\begin{example}
The Brauer--Picard groupoid of $\mathrm{Vec}(\mathbb{Z}/p\mathbb{Z})$ has exactly one point. The group of outer automorphisms is the group of ordinary outer automorphisms of $\mathbb{Z}/p\mathbb{Z}$ acting in the obvious way.
The Brauer--Picard group is the dihedral group $(\mathbb{Z}/p\mathbb{Z})^\times \rtimes \mathbb{Z}/2\mathbb{Z}$ where the subgroup is the group of outer automorphisms and the other coset comes from realizing $\mathrm{Vec}$ as a bimodule category over $\mathbb{Z}/p\mathbb{Z}$. Such bimodules can be realized as follows: start with the Morita equivalence between $\mathrm{Vec}(\mathbb{Z}/p\mathbb{Z})$ and $\mathrm{Rep}(\mathbb{Z}/p\mathbb{Z})$ from Example \ref{ex:standardmorita} and then pick an equivalence $\mathrm{Vec}(\mathbb{Z}/p\mathbb{Z}) \cong \mathrm{Rep}(\mathbb{Z}/p\mathbb{Z})$.
\end{example}
\begin{example}
If $G$ is a non-abelian group, then the Brauer--Picard groupoid of $\mathrm{Vec}(G)$ has at least two distinct points since $\mathrm{Vec}(G)$ and $\mathrm{Rep}(G)$ are Morita equivalent via the bimodule from Example \ref{ex:standardmorita}. Specifically the Brauer--Picard groupoid of $\mathrm{Vec}(S_3)$ has exactly two points, and exactly two Morita equivalences between each points. The nontrivial Morita autoequivalence of $\mathrm{Rep}(S_3)$ is given by $\mathrm{Rep}(S_2)$ (this corresponds to the $A_5$ subfactor with index $3$).
\end{example}
In general there may be more than one equivalence ${}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}} \boxtimes_{\mathcal{D}} {}_{\mathcal{D}}{\mathcal{N}} {}_{\mathcal{E}} \cong {}_{\mathcal{C} } {\mathcal{L} } {}_{\mathcal{E} }$. In fact, such choices of equivalence are a torsor for $\pi_2$ of the Brauer--Picard groupoid. In particular, for the Asaeda--Haagerup subfactor there is no such ambiguity, since $\pi_2 $ is trivial \cite{ GJSext}. Similarly, if $\pi_2$ were nontrivial then one has to be more careful about the associativity of these multiplication maps.
There is a third construction which unifies the above two points of view and elucidates the associativity of composition of bimodules. Although we will not use this third construction in this paper it may clarify the above constructions conceptually. One can consider a $3$-category whose objects are pairs $A \in \mathcal{C}$ of a Frobenius algebra object in a spherical fusion category, whose $1$-morphisms are pairs ${}_A M_B \in {}_\mathcal{C}\mathcal{M}_{\mathcal{D}}$ of a bimodule object in a bimodule category, whose $2$-morphisms are pairs of a bimodule functor $\mathcal{F}: \mathcal{M} \rightarrow \mathcal{N}$ and a bimodule map $f: \mathcal{F}(m) \rightarrow n$, and whose $3$-morphisms are pairs of bimodule natural transformations satisfying a compatibility condition.
\subsection{Combinatorics of fusion and module categories}
A fusion category contains a lot of information. At the first level there is the combinatorial information of how the objects tensor, but then at the second level there is the linear algebraic information of all the morphisms and their compositions and tensor product. The combinatorial data is typically a rather weak invariant of the fusion category. However, once one considers the whole Brauer--Picard groupoid, the combinatorics for tensoring all the objects in all the tensor categories and all the bimodules is a lot richer.
The fusion ring of a fusion category $\mathcal{C} $ is the based ring with basis indexed by the equivalence classes of simple objects of $\mathcal{C} $, with addition given by direct sum and multiplication given by tensor product. There is also an involution, defined on basis elements by duality in $\mathcal{C} $. There is a unique homomorphism from the fusion ring to $\mathbb{R} $ which takes basis elements to positive numbers, called the Frobenius-Perron dimension, and denoted by $d$.
In a similar way, module categories and bimodule categories determine based modules and bimodules over the fusion rings of the corresponding fusion categories, which we call fusion modules and fusion bimodules. These modules are assigned dimension functions, also denoted by $d$, as follows: for each basis element $m $, set $d(m)=\sqrt{d(\underline{\text{End}}(M))} $, where $M$ is the corresponding simple object in the module category. Then $d$ is multiplicative for the module multiplication as well, and does not depend on whether one takes the left or right internal end in invertible bimodule categories.
If $A$ is an algebra in $\mathcal{C}$ then the principal graph of $A$ is the induction-restriction graph between $\mathcal{C}$ and $A$-mod, while the dual principal graph is the induction-restriction graph between $A$-mod and $A$-mod-$A$. The principal graphs record only part of the fusion bimodule structure, namely the fusion rules for tensoring with the basic bimodules ${}_{1}A_A$ and ${}_{A}A_1$. The principal and dual graphs of a finite-depth subfactor $N \subset M $ are the graphs of the algebras $\bar{\iota} \iota $ and $ \iota \bar{\iota} $ in the principal and dual even parts, where $\iota: N \rightarrow M $ is the inclusion map.
We refer to \cite{GSbp} for the definitions of fusion modules and bimodules, and multiplicative compatibility for triples of fusion modules and bimodules; however we briefly summarize the basic idea.
Let $ {}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}}$ and ${}_{\mathcal{D}}{\mathcal{N}} {}_{\mathcal{E}}$ be invertible bimodule categories over fusion categories, and let ${}_{\mathcal{C}}{\mathcal{M}} {}_{\mathcal{D}} \boxtimes_{\mathcal{D}} {}_{\mathcal{D}}{\mathcal{N}} {}_{\mathcal{E}} \cong {}_{\mathcal{C} } {\mathcal{L} } {}_{\mathcal{E} }$. Let ${}_C M {}_D$, $ {}_D N {}_E$, and $ {}_C L {}_E $ be the
corrseponding fusion bimodules over fusion rings. Then the equivalence of bimodule categories induces a bimodule homomorphism from ${}_C M {}_D \otimes_D {}_D N {}_E $ to $ {}_C L {}_E $ such that if $m $ and $n$ are basis elements in $ {}_C M {}_D$ and
${}_D N {}_E $ respectively, then the image of $m \otimes n $ is a non-negative combination
of basis elements of $ {}_C L {}_E $; moreover, this homomorphism preserves Frobenius-Perron dimension and multiplication by duals in the fusion bimodules. The existence of such a homomorphism places strong combinatorial constraints on the triple $({}_C M {}_D, {}_D N {}_E, {}_C L {}_E )$, which can be checked tediously by hand or quickly with a computer.
\begin{remark}
As mentioned above, in general one needs to be a bit careful about the associativity of compositions of bimodule categories, and this is true even at the level of fusion bimodules. For AH, the vanishing of $\pi_2$ guarantees that the composition of bimodules is well-defined, and also that these compositions are associative at the fusion bimodule level. (Whether one can give compatible associators is a more subtle question which involves $\pi_3$ as well.) Nonetheless we will never need to use this uniqueness or associativity.
\end{remark}
\subsection{The Asaeda-Haagerup fusion categories}
The Asaeda-Haagerup subfactor, constructed in \cite{MR1686551}, is a finite depth subfactor with index $\frac{5+\sqrt{17}}{2} $
and principal and dual graph pair
$$\vpic{AHpg_unlabeled} {1.4in} , \vpic{AHdualpg} {1.4in} .$$
A subfactor with index $\frac{7+\sqrt{17}}{2} $ and principal and dual graph pair $$ \vpic{AHp1} {1.4in} , \vpic{AHp1dual} {1.4in} $$
was constructed in \cite{MR2812458}, and a third subfactor with index $\frac{9+\sqrt{17}}{2} $ and principal and dual graph pair
$$ \vpic{AHp2dual } {1.4in}, \vpic{AHp2dual } {1.4in} $$
was constructed in \cite{GSbp}.
Let $\mathcal{AH}_2$, $\mathcal{AH}_3$, and $\mathcal{AH}_1 $ be the principal even parts of these three subfactors, respectively ($\mathcal{AH}_1 $ is also the dual even part of all three subfactors). The fusion ring of $\mathcal{AH}_i$ is denoted $AH_i$ and the structure contants for these fusion rings are given in the Appendix. In \cite{GSbp} all possible fusion modules and fusion bimodules over these $AH_i$ were computed. Each of these fusion modules can be described by a small list of small matrices with small integer entries (here small means single digit) where each matrix describes the combinatorics of tensoring with one of the simple objects.
Our main result from that paper was a description of all Morita equivalences between any two of the $\mathcal{AH}_i$.
\begin{theorem}
There are exactly four invertible bimodule categories over each not-necessarily-distinct pair $\mathcal{AH}_i-\mathcal{AH}_j$, up to equivalence.
Furthermore, the Brauer-Picard group of each $\mathcal{AH}_i$ is $\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $.
\end{theorem}
The fusion rules for all $36 $ bimodule categories between the various $\mathcal{AH}_i-\mathcal{AH}_j $ were given in an online appendix to \cite{GSbp}. In addition, the combinatorial possibilities for any other fusion categories in the same Morita equivalence class were severely constrained.
\begin{theorem} \cite[Theorem 6.10]{GSbp} \label{cases}
Let $\mathcal{C}^{(k)} $ be a fusion category which is Morita equivalent to the $\mathcal{AH}_i$ (for $i =1,2,3$), but not equivalent to any of them. Then there is a triple of fusion bimodules
$(K_{AH_1}^{(k)} , L_{AH_2}^{(k)}, M_{AH_3}^{(k)} ) $ such that every $\mathcal{C}-\mathcal{AH}_1 $ Morita equivalence realizes $K_{AH_1}^{(k)} $, every $\mathcal{C}-\mathcal{AH}_2 $ Morita equivalence realizes $L_{AH_2}^{(k)}$, and every $\mathcal{C}-\mathcal{AH}_3 $ Morita equivalence realizes $L_{AH_3}^{(k)}$.
Moreover, there are exactly four possibilities for the triple $(K_{AH_1}^{(k)} , L_{AH_2}^{(k)}, M_{AH_3}^{(k)} )$. \end{theorem}
Since we will need the details of the four possiblities for $(K_{AH_1}^{(k)} , L_{AH_2}^{(k)}, M_{AH_3}^{(k)} )$, we include them in the Appendix for the convenience of the reader. We will refer to these cases using $k = 4, 5, 6, 7$. We will see that cases $(4)$, $(5)$, and $(6)$ have a somewhat similar flavor to each other, while case $(7)$ is quite different. These cases correspond to cases (c), (a), (b), and (d) respectively in \cite[Theorem 6.10]{GSbp}.
Note that the situation described by Theorem \ref{cases} is very different than what we see for $\mathcal{AH}_{1-3} $, for which the fusion bimodule of each $\mathcal{AH}_i-\mathcal{AH}_j $ invertible bimodule category are distinct from each other either as a left $AH_i $ fusion module or as a right $AH_j $ fusion module.
\section{Classification of the Morita equivalence class of the Asaeda-Haagerup fusion categories}
\label{sec:BP}
The main result of this section is the following.
\begin{theorem} \label{thm:dichotomy}
One of the following two claims is true:
\begin{itemize}
\item There are no fusion categories realizing any of cases $(4)$, $(5)$, or $(6)$.
\item There is a unique fusion category realizing each of cases $(4)$, $(5)$, and $(6)$, but there are no fusion categories realizing case $(7)$.
\end{itemize}
\end{theorem}
\begin{remark}
In Section \ref{sec:main} we will see that the latter possibility is the correct one by realizing case $(4)$.
\end{remark}
Our first goal will be to constrain the possible fusion rings for the dual fusion categories of the fusion modules $(K_{AH_1}^{(k)} , L_{AH_2}^{(k)}, M_{AH_3}^{(k)} )$.
\subsection{Finding the possible dual fusion modules of a fusion module}
Given a module category over a fusion category ${}_{\mathcal{C}} \mathcal{M}$, we would like to be able to compute the dual fusion category $\mathcal{D}=( {}_{\mathcal{C}} \mathcal{M})^*$, and in particular its fusion ring. This former computation can be difficult even when the module category is well understood.
However, just the fusion module structure of ${}_{\mathcal{C}} \mathcal{M}$ can provide strong combinatorial constraints on the fusion rules of $\mathcal{D} $, and this is sometimes sufficient to calculate the fusion ring of $\mathcal{D} $ and the dual fusion module of $\mathcal{M} {}_{\mathcal{D}} $, or at least to narrow it down to a few choices.
The idea is that Frobenius reciprocity relates the combinatorics of the two even parts of a subfactor. Suppose that $\iota$ is the inclusion $N \rightarrow M$ and that $\psi$ and $\chi$ are other $M$-$N$ sectors. By Frobenius reciprocity, we have that $(\iota \bar{\psi} , \iota \bar{\chi}) = (\bar{\iota} \iota, \bar{\chi} \psi)$. But $(\iota \bar{\psi} , \iota \bar{\chi}) $ counts the number of length two paths between $\psi$ and $\chi$ in the dual principal graph, while $(\bar{\iota} \iota, \bar{\chi} \psi)$ is asking about fusing two odd bimodules into the principal even part. Thus we can read off the number of length two paths on the dual principal graph just from understanding the principal half very well. Just knowing the odd vertices and the number of length two paths between each pair of odd vertices is often enough to determine the even vertices as well. More generally, the following lemma is an immediate consequence of Frobenius reciprocity.
\begin{lemma} \label{duallem}
Let $(K,T) $ be a fusion bimodule over the fusion rings $(A,R) $ and $(B,S)$ (where the second argument in each pair refers to the basis.) For fixed $\kappa \in T $, define the matrices $M^{\kappa}$ and $N^{\kappa}$ as follows: $$ M^{\kappa}_{ij}= \sum_{\xi \in R} \limits (\xi \kappa, \kappa) ( \xi \mu_i, \mu_j), \ \mu_i, \mu_j \in T$$ and $$ N^{\kappa}_{ij}= (\mu_i, \kappa \eta_j), \ \eta_j \in S, \mu_i \in T.$$ Then $N^{\kappa} (N^{\kappa})^t =M$.
\end{lemma}
\begin{proof}
We have $(N^{\kappa}(N^{\kappa})^t)_{ij} = \sum_k \limits (\mu_i, \kappa \eta_k)(\mu_j, \kappa \eta_k)=(\bar{\kappa} \mu_i ,\bar{ \kappa} \mu_j)=(\kappa \bar{\kappa}, \mu_j \bar{\mu_i})=\sum_{\xi \in R} \limits (\xi \kappa, \kappa) ( \xi \mu_i, \mu_j)=M^{\kappa}_{ij}$.
\end{proof}
The matrix $M^{\kappa}$ in Lemma \ref{duallem} is part of the data of the left fusion module ${}_A K $, while the matrix $N$ is the fusion matrix of $\kappa $ with respect to the right action of $B$. Thus given $ {}_A K$, we can compute all possibilities for the right fusion matrix of $\kappa $ in a fusion bimodule ${}_A K {}_B $ whose left fusion module is ${}_A K $ if we can find all decompositions of
$M^{\kappa}$ into $NN^t $ for non-negative integer matrices $N$. Any choice of $N$ determines the dimensions of the basis elements in $B$, and we may immediately discard any choices which do not allow for a basis element of dimension $1$ (the identity) as well as any choices for which there is some other $\kappa' \in T$ which does not admit a right fusion matrix corresponding to the same dimension values for $B$.
Once we have found all possible right fusion matrices for each basis element of $K $, we can try to reconstruct all possible fusion bimodules ${}_A K {}_B $ whose left fusion modules are ${}_A K $, using similar combinatorial searches as in \cite{GSbp}.
These can easily be made into computer algorithms all of which run quite quickly since the number of possibilities is small. In fact, all the results of this section were originally checked by computer, but since the calculations are all simple enough we have also checked them all by hand.
\subsection{The structure of fusion categories realizing cases (4)-(6)}
The first step toward proving Theorem \ref{thm:dichotomy} is to understand the simple objects and fusion rules for any fusion category realizing cases (4)-(6).
\begin{lemma}\label{dualgraphlemma}
Let $\mathcal{C} $ be a fusion category realizing case (4),(5), or (6) of Theorem \ref{cases}. Then
\begin{enumerate}
\item $\mathcal{C} $ has $4$ invertible objects and $4$ simple objects of dimension $4+\sqrt{17} $.
\item If $\mathcal{C}$ realizes case (4) (resp. case (5)), then there is a simple object in an invertible $\mathcal{C}-\mathcal{AH}_2$(resp. $\mathcal{C}-\mathcal{AH}_3$) bimodule category whose principal graph is
$$\vpic{newgraph} {1.5in} $$
and the adjacency matrix of whose dual graph is the first matrix appearing in the fusion module $ L_{AH_2}$ for case (4) (resp. $M_{AH_3}$ for case (5)).
\end{enumerate}
\end{lemma}
\begin{proof}
In each of the three cases, we look for a basis element of small dimension in one of the fusion modules.
In case (4), $\theta^1_{42}$ has a $2$-supertransitive principal graph (the matrix given in the appendix is the adjacency matrix of this graph). In case (5), $\theta_{53}^1$ also has a $2$-supertransitive principal graph.
We can apply Lemma \ref{duallem} to these basis elements to obtain the pictured graph; the dimensions of the simple objects in $\mathcal{C}$ are then given by the normalized Frobenius-Perron weights of the graph.
Since cases (4) and (5) are similar, we only consider case (4) in detail. We refer to the table of fusion modules in the Appendix and use the notation from there. For the fusion module $L_{AH_2}^{(4)}$ we consider the object $\theta^1_{42} $ and compute the matrix
$$M^{\theta^1_{42}}_{ij}=\sum_{\phi} (\theta^1_{42} \phi,\theta^1_{42})(\theta^i_{24} \phi, \theta^j_{24}) $$ of Lemma \ref{duallem}, where $\phi $ ranges over the simple objects of $\mathcal{AH}_2 $. Since $(\theta^1_{42} \phi,\theta^1_{42}) =1$ if $\phi $ is either $1$ or $\alpha \pi $ and $(\theta^1_{42} \phi,\theta^1_{42}) =0$ otherwise, the rows of $M^{\theta^1_{42}}$ are obtained by adding the rows corresponding to $1$ and to $\alpha \pi $ in each submatrix of the fusion module $L_{AH_2}^{(4)}$. Thus we have
$$M^{\theta^1_{42}}=
\left(
\begin{array}{cccccc}
2 & 0 & 0 & 0 & 1 & 1 \\
0 & 2 & 0 & 0 & 1 & 1 \\
0 & 0 & 2 & 0 & 1 & 1 \\
0 & 0 & 0 & 2 & 1 & 1 \\
1 & 1 & 1 & 1 & 4 & 4 \\
1 & 1 & 1 & 1 & 4 & 4 \\
\end{array}
\right) .$$
Then from the relation $ NN^t=M$ it is easy to see that up to permutations of columns, the left fusion matrix of $ \theta^1_{42}$
must be
$$N^{\theta^1_{42}}=
\left(
\begin{array}{cccccccc}
1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 \\
1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 \\
1 & 0 & 1 & 0 & 1 & 0 & 1 & 0 \\
\end{array}
\right) , $$
which is the adjacency matrix of the pictured graph.
One can also think of the above calculation as counting paths on principal graphs, which can be an easier way to arrange the calculation if you are doing it by hand. From the fusion matrix for $\theta^1_{42} $ we see that the graph for fusion with $\theta^1_{42}$ has one odd object at depth one ($\theta^1_{42}$), two at depth three ($\theta^5_{24}$ and $\theta^6_{24}$), and three at depth five ($\theta^2_{24}$, $\theta^3_{24}$, and $\theta^4_{24}$). Thus the dual graph which we are trying to compute must also have the same number of odd objects at each depth. Now, by Frobenius reciprocity, the matrix $M$ above counts the number of paths of length two between these odd vertices. In particular, the vertex at depth one is $2$-valent and has exactly one length-two path to each of the depth three vertices. Thus, the graph must agree with the pictured graph through depth three. Then the vertices at depth three are $4$-valent and since one edge is already accounted for, they must each have three edges connected to depth four vertices. Since there are four length-two paths between distinct depth three vertices (one of which is already accounted for), there must be exactly three vertices at depth four each of which is connected to both depth three vertices. Thus the graph must agree with the given one through depth four. Since each of the vertices at depth five is $2$-valent and only has one length-two path to each of the depth three vertices, each must have a single edge going to a vertex at depth six. Since these three vertices at depth five have no length two paths between them there must be three such vertices at depth six. Thus the graph must be the one in the theorem.
For case (6), we apply in a similar way Lemma \ref{duallem} to each basis element in the fusion module $M^{(6)}_{AH_3}$. The dimensions listed above are the only choice for which each basis element admits a consistent graph.
\end{proof}
\begin{remark}
In the rest of this section there are several elementary calculations like the one in the last lemma. Since they are all straightforward and similar to the first calculation, we will not go through them in detail.
\end{remark}
\begin{remark}
The principal graph of the smallest simple object in an invertible $\mathcal{C}-\mathcal{AH}_3$ bimodule category realizing case (6) must be
$$\vpic{newgraph3} {3in} .$$
\end{remark}
\begin{lemma} \label{lemah4}
Let $\mathcal{C} $ be a fusion category which admits a module category with an object having the fusion graph pictured in Lemma \ref{dualgraphlemma}. Then there is an order $4$ group $G$ such that the basis for the Grothedieck ring of $\mathcal{C}$ is indexed by $
\{\alpha_g\}_{g \in G} \cup \{\alpha_g \rho \}_{g \in G} $, satisfying $$\quad \alpha_g \alpha_h = \alpha_{g+h}, \quad \text{and }\rho^2=1+\sum_{g \in G} 2\alpha_g \rho .$$ There are exactly four possibilities for the fusion ring up to isomorphism, corresponding to four choices for the depth preserving duality involution on the basis:
\begin{enumerate}
\item the trivial involution, where $G$ is the Klein $4$-group and $\alpha_g \rho = \rho \alpha_g$,
\item the involution which swaps two invertible basis elements, where $G$ is cyclic and $\alpha_g \rho = \rho \alpha_{g^{-1}}$,
\item the involution which swaps two non-invertible basis elements, where $G$ is the Klein $4$-group, and without loss of generaltiy $\alpha_{(1,0)} \rho = \rho \alpha_{(0,1)}$, and $\alpha_{(0,1)} \rho = \rho \alpha_{(1,0)}$,
\item the involution which swaps two invertible basis elements and two noninvertible basis elements, where $G$ is cylic and $\alpha_g \rho = \rho \alpha_g$.
\end{enumerate}
\end{lemma}
\begin{proof}
It is clear from the principal graph that there are four invertible objects (which must have the fusion rules for some group of size four) and four noninvertible objects. Let $\rho$ be the non-invertible object at depth $2$. Tensoring a simple object with $\rho$ counts the number of length-$2$ paths (from the vertex reprepresenting that object to the vertex representing $\rho$) minus the number of length $0$ paths. Since each of the non-invertible objects has a length-$2$ path to a different invertible object, it follows that all non-invertible objects are of the form $\alpha_g \rho$. Furthermore, $\rho^2$ has one length-$2$ path to the identity, three length-$2$ paths to itself, and two length-$2$ paths to each of the other non-invertible objects, so $\rho^2 = 1+\sum_{g \in G} 2\alpha_g \rho$. Finally, it is clear that those are the only four possibilities for depth preserving involutions on the even vertices, and the fusion rules are determined by the duality using that $\overline{\alpha_g \rho} = \rho \overline{\alpha_g}$.
\end{proof}
In the following arguments we will frequently need to find the list of objects in a fusion category which admit the structure of a division algebra whose category of modules realizes a given fusion module. Since every division algebra is realized as the internal endomorphisms of a simple object in its category of modules, this list can be read off the data of the fusion module.
We give an example of how this works. Suppose we would like to find the list of objects in $\mathcal{AH}_3 $ which admit an algebra structure whose category of modules realizes the fusion module $M_{AH_3}^{(4)}$.
The fusion module $M_{AH_3}^{(4)}$ has $3$ basis elements, with fusion matrices
$$\left(
\begin{array}{c|ccc|ccc|ccc}
& \multicolumn{2}{l}{\theta_{43}^1 \otimes \_} & & \multicolumn{2}{l}{\theta_{43}^2 \otimes \_} & & \multicolumn{2}{l}{\theta_{43}^3 \otimes \_} & \\
& \theta_{43}^1 & \theta_{43}^2 & \theta_{43}^3 & \theta_{43}^1 & \theta_{43}^2 & \theta_{43}^3 & \theta_{43}^1 & \theta_{43}^2 & \theta_{43}^3 \\ \hline
1 & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 \\
\beta & 1 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 1 \\
\xi & 0 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 4 \\
\beta \xi & 0 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 4 \\
\xi \beta & 0 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 4 \\
\beta \xi \beta & 0 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 4 \\
\mu & 1 & 0 & 2 & 0 & 1 & 2 & 2 & 2 & 7 \\
\beta \mu & 1 & 0 & 2 & 0 & 1 & 2 & 2 & 2 & 7 \\
\nu & 0 & 0 & 2 & 0 & 0 & 2 & 2 & 2 & 6 \\
\end{array}
\right)$$
where the columns of each matrix are indexed by the basis of the fusion module and the rows by the basis of the fusion ring of $\mathcal{AH}_3 $.
To find the underlying object of the algebra $\overline{\theta_{43}^{(i)}} \theta_{43}^{(i)}$, we just look at $\theta_{43}^{(i)}$ column in the fusion matrix for $\theta_{43}^{(i)}$. Thus for any (right) module category $\mathcal{M} $ over $\mathcal{AH}_3$ which realizes the fusion module $M_{AH_3}$, there are two algebras with underlying objects $1+\beta +\mu+\beta \mu$ and one algebra with underlying object $1+\beta+4(\xi+\beta \xi+\xi\beta+\beta\xi\beta)+7(\mu+ \beta \mu)+6\nu $ whose categories of (left) modules are each equivalent to $\mathcal{M} $. The first two algebras each have dimension $2d+2$ and the latter algebra has dimension $28d+4$, where $d = 4+\sqrt{17}$. The dimensions of the basis elements of $M_{AH_3}^{(1)}$ are thus $\sqrt{2d+2}$, $\sqrt{2d+2}$, and $\sqrt{28d+4}$. The algebras $\theta_{43}^{(i)}\overline{\theta_{43}^{(i)}}$ in the dual category $ (\mathcal{M}_{\mathcal{C}} )^*$ hence also have dimensions $2d+2$ or $28d+4$.
We introduce the following notation. Let $\gamma $ be a simple algebra in a fusion category $\mathcal{C} $. Then we denote the category of $\gamma-\gamma $ bimodules in $\mathcal{C} $ by $\mathcal{C}^{\gamma} $. Recall that $C^{\gamma} $ is Morita equivalent to $\mathcal{C} $.
\begin{lemma}\label{div}
\begin{enumerate}
\item Let $\gamma $ be an algebra in a fusion category $\mathcal{C} $, and let $\delta $ be the subobject of $\gamma $ which contains each invertible object of $\mathcal{C} $ with the same multiplicity as $\gamma $ does and which does not contain any non-invertible objects. Then $\delta$ inherits an algebra structure from $\gamma $.
\item Let $\mathcal{C} $ be a fusion category and let $\gamma $, $\gamma'$ be division algebras such that $(\gamma,\gamma')=1 $ (as objects of $\mathcal{C} $). Then there is a division algebra $\delta$ in $\mathcal{C}^{{\gamma}} $ with $d(\delta)=d(\gamma)d(\gamma') $ such that $(\mathcal{C}^{\gamma} )^{\delta}$ is equivalent to $\mathcal{C}^{\gamma'} $.
\item Let $\mathcal{C} $ be a fusion category and let $ \gamma$ be a division algebra in $\mathcal{C} $ with a subalgebra $ \delta$. Then there is a division algebra $\gamma'$ in $\mathcal{C}^{\delta}$ with dimension $d(\gamma')=\frac{d(\gamma)}{d(\delta)} $ such that
$(\mathcal{C}^{\delta})^{\gamma'} $ is equivalent to $\mathcal{C}^{\gamma}$.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item This follows from the fact that the tensor product of any two invertible objects is invertible.
\item Recall that bimodules over division algebras in $\mathcal{C} $ form a rigid $2$-category; we will use multiplicative notation for composition of $1$-morphisms. Consider the bimodule
$\kappa={}_{\gamma} \gamma {}_1 \cdot {}_1 \gamma' {}_{\gamma'} $. By Frobenius reciprocity,
$(\kappa,\kappa)_{\gamma\text{-mod-}\gamma'} =(\gamma,\gamma')=1$, so $ \kappa$ is simple as a $\gamma-\gamma' $ bimodule. Then $\delta=\kappa \bar{\kappa} $ is a simple algebra in $\mathcal{C}^{\gamma} $, and $\bar{\kappa} \kappa $ is a simple algebra in $\mathcal{C}^{\gamma'} $. By multiplicativity of dimension in the $2$-category, we have $d(\delta)=d(\gamma)d(\gamma') $.
\item Let $\kappa= {}_{\gamma} \gamma {}_{\delta} $ . Then $\gamma'=\bar{\kappa} \kappa$ is a simple algebra in $\mathcal{C}^{\delta} $, $\kappa \bar{\kappa} $ is a simple algebra in $\mathcal{C}^{\gamma} $, and from the relation ${}_{\gamma} \gamma {}_1= {}_{\gamma} \gamma {}_{\delta} \cdot {}_{\delta} \delta {}_{1} $ it follows that $d(\gamma)=d(\gamma')d(\delta) $.
\end{enumerate}
\end{proof}
\begin{remark}
In the case of a fusion category coming from a subfactor, Lemma \ref{div}(3) can be interpreted as follows: the algebra $\gamma $ corresponds to a subfactor $ N \subseteq M$, the subalgebra $\delta $ corresponds to $N \subseteq P $ for an intermediate subfactor $N \subseteq P \subseteq M$, and then the ``quotient'' algebra $\gamma' $ corresponds to $P \subseteq M $.
\end{remark}
The goal of this subsection is to determine, for each of cases (4)-(6), which of the four possibilities the fusion ring must be. A priori it might be that there was more than one fusion category realizing one of the cases and that these two fusion categories have different fusion rings, but it turns out that we can determine the fusion rules from be the bimodules. First we concentrate on the invertible objects, and then the non-invertible objects.
A fusion category $\mathcal{C} $ all of whose objects are invertible is necessarily equivalent to $Vec_G^{\omega} $, where $G$ is a finite group and $\omega \in H^3(G,\mathbb{C}^{\times}) $ determines the associator. If $H$ is a subgroup of $G$, then $\bigoplus_{h \in H} k_h $ admits an algebra structure iff $\omega|_H$ is trivial; the algebra structures are then parametrized by $H^2(H, \mathbb{C}^{\times} )$. In fact, these are the only division algebras in $\omega \in H^3(G,\mathbb{C}^{\times})$ (this can be seen directly, or by applying \cite{MR1976459}).
We will need to know a little about the dual categories to the division algebras in $Vec_G^{\omega}$ when $G$ has size $4$. For $\text{Vec}_{\mathbb{Z}/4\mathbb{Z} }$ with trivial associator, the dual category over the $2$-dimensional simple algebra is $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}^\xi$ with a specific non-trivial associator which we denote $\xi$ \cite[Ex. 4.10]{MR2362670}. The associator $\xi$ is trivial when restricted to each of the $\mathbb{Z}/2\mathbb{Z}$ factors, but is non-trivial when restricted to the diagonal. In particular, $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}^\xi$ has exactly two non-trivial division algebras both of dimension $2$. On the other hand, $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ with trivial associator has three different simple $2$-dimensional algebras. By Lemma \ref{div}(2), the dual category over any one of the $2$-dimensional algebras in $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ has a $4$-dimensional simple algebra, and therefore must again be $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ with trivial associator.
If $\mathcal{C}$ is a fusion category, we denote by $\text{Inv}(\mathcal{C})$ the subcategory generated by the invertible objects. Note that by Lemma \ref{dualgraphlemma}, if $\mathcal{C} $ realizes case (4), (5), or (6), then $\text{Inv}(\mathcal{C})=\text{Vec}_{\mathbb{Z}/4\mathbb{Z}}^\omega$ or $\text{Inv}(\mathcal{C})=\text{Vec}_{\mathbb{Z}/2\mathbb{Z}\times \mathbb{Z}/2\mathbb{Z}}^\omega$, for some associator $\omega$.
\begin{lemma} \label{lemc}
Let $\mathcal{C}$ be a fusion category realizing case (4).
\begin{enumerate}
\item The associator on $\text{Inv}(\mathcal{C}) $ is trivial.
\item There is a $4$-dimensional division algebra $\gamma_4 $ in $\mathcal{C} $ such that $\mathcal{C}^{\gamma_4} $ realizes case (6).
\item For any $2$-dimensional division algebra $\gamma_2 $ in $\mathcal{C} $, the fusion category $\mathcal{C}^{\gamma_2} $ realizes case (5).
\end{enumerate}
\end{lemma}
\begin{proof}
Since $\mathcal{C}$ realizes case (4), it is the dual category to a module cateogry over $\mathcal{AH}_3$ which realizes the corresponding fusion module $M_{AH_3}^{(4)}$; therefore it contains a division algebra $ \theta_{43}^3 \overline{\theta_{43}^3}= \gamma $ of dimension $28d+4$ as in the discussion preceding Lemma \ref{div}. All objects in $\mathcal{C} $ have dimension equal to $d$ or to $1$, so $\gamma$ must contain four invertible objects.
If $\psi$ is invertible, then since $\theta_{43}^3$ and $\psi \theta_{43}^3$ are simple, $(\psi, \gamma) = (\psi\theta_{43}^3 ,\theta_{43}^3) \leq 1$. Hence, $\gamma$ must contain each of the invertible objects exactly once. The direct sum of the invertible objects then forms a subalgebra of $\gamma $ by Lemma \ref{div}.
\begin{enumerate}
\item Since there is an algebra structure on the sum of all four invertible objects, the associator on $\text{Inv}(\mathcal{C}) $ must be trivial.
\item Let $\gamma_4$ be the subalgebra of $\gamma $ corresponding to the direct sum of the invertible objects. Then $C^{\gamma_4} $ also contains four invertible objects, so it is not equivalent to $\mathcal{AH}_1-\mathcal{AH}_3$ and therefore must realize one of the four cases (4)-(7). Moreover, by Lemma \ref{div} there must be a division algebra in $\mathcal{C}^{\gamma_4} $ of dimension $\frac{28d+4}{4}=7d+1$ whose category of bimodules is equivalent to $\mathcal{AH}_3 $. But of the $M_{AH_3}^{(i)}$ for $i = 4, 5, 6, 7$, only $M_{AH_3}^{(6)}$ has a basis element with dimension $\sqrt{7d+1} $. Therefore, $\mathcal{C}^{\gamma_4} $ must realize case (6).
\item Since $\mathcal{C} $ realizes case (4), it is the dual category to a module category over $\mathcal{AH}_2$ which realizes the corresponding fusion module $L_{AH_2}^{(4)}$. Therefore, there is a division algebra in $\mathcal{C}$ with dimension $1+d$, and a non-invertible simple object $\rho$ in $\mathcal{C}$ such that $\mathcal{C}^{1+\rho }$ is equivalent to $\mathcal{AH}_2 $. Let $\gamma_2 $ be a $2$-dimensional division algebra in $\mathcal{C} $. Since $(\gamma_2,1+\rho)=1 $, by Lemma \ref{div} there is a division algebra of dimension $2(1+d) $ in $\mathcal{C}^{\gamma_2} $ whose dual category is $\mathcal{AH}_2 $. But of the four choices for $L_{AH_2}^{(i)}$, only that of case (5) has a basis element with dimension $\sqrt{2(1+d)} $. Therefore $\mathcal{C}^{\gamma_2} $ must realize case (5).
\end{enumerate}
\end{proof}
\begin{lemma} \label{lema}
Let $\mathcal{C}$ be a fusion category realizing case (5).
\begin{enumerate}
\item The associator on $\text{Inv}(\mathcal{C}) $ is non-trivial.
\item There is a $2$-dimensional division algebra $\gamma_2 $ in $\mathcal{C} $ such that $\mathcal{C}^{\gamma_2} $ realizes case (4).
\end{enumerate}
\end{lemma}
\begin{proof}
Since $\mathcal{C} $ realizes case (5), it is the dual category to a module category over $\mathcal{AH}_2$ which realizes the corresponding fusion module $L_{AH_2}^{(5)}$.
\begin{enumerate}
\item Looking at the fusion module $L_{AH_2}^{(5)}$, we see that $\mathcal{C}$ has a division algebra $\gamma $ with dimension $ 7d+1$ such that $\mathcal{C}^{\gamma} $ is equivalent to $\mathcal{AH}_2 $. Suppose that the associator on $\text{Inv}(\mathcal{C}) $ is trivial. Let $\gamma_4 $ be a simple algebra of dimension $4$ in $\mathcal{C} $. Since all simple objects in $\mathcal{C} $ have dimension $1$ or dimension $d$, $\gamma $ does not contain any non-trivial invertible objects which implies that $(\gamma,\gamma_4) =1$. By Lemma \ref{div}, $\mathcal{C}^{\gamma_4} $ must have a simple algebra $\delta $ of dimension $4(7d+1)$ such that $(\mathcal{C}^{\gamma_4})^{\delta} $ is equivalent to $\mathcal{AH}_2 $. Also, since $\mathcal{C}^{\gamma_4} $ has four invertible objects, it must realize one of the cases (4)-(7). But looking at the four possibilites for $L_{AH_2}^{(i)}$, we find that none of them contain a basis element with dimension $\sqrt{4(7d+1)} $. Therefore, $\gamma_4$ cannot exist and the associator on $\text{Inv}(\mathcal{C}) $ must be non-trivial.
\item Again looking at $L_{AH_2}^{(5)}$, we see that there is a division algebra $\gamma$ in $\mathcal{C} $ of dimension $2(d+1)$ such that $\mathcal{C}^{\gamma} $ is equivalent to $\mathcal{AH}_2 $. Then $\gamma $ must contain $2$ invertible objects, and therefore contains a $2$-dimensional subalgebra $\gamma_2 $. Then by Lemma \ref{div}, $\mathcal{C}^{\gamma_2} $ has a division algebra $\gamma' $ of dimension $ d+1$ such that $(\mathcal{C}^{\gamma_2} )^{\gamma'}$ is equivalent to $ \mathcal{AH}_2$. Therefore, $\mathcal{C}^{\gamma_2} $ realizes case (4).
\end{enumerate}
\end{proof}
\begin{lemma}\label{lemb}
Let $\mathcal{C}$ be a fusion category realizing case (6).
\begin{enumerate}
\item The associator on $\text{Inv}(\mathcal{C}) $ is trivial.
\item There is a $4$-dimensional division algebra $\gamma_4 $ in $\mathcal{C} $ such that $\mathcal{C}^{\gamma_4} $ realizes case (4).
\end{enumerate}
\end{lemma}
\begin{proof}
The proof is similar to the previous lemmas. (Namely, look at the fusion module $L_{AH_2}^{(6)}$ to find a division algebra in $ \mathcal{C}$ of dimension $4(d+1)$ which has a $4$-dimensional subalgebra $\gamma_4 $. Then there is a division algebra $\gamma' $ in $\mathcal{C}^{\gamma_4} $ of dimension $d+1$ such that $\mathcal{C}^{\gamma'} $ is equivalent to $\mathcal{AH}_2 $.)
\end{proof}
\begin{lemma}
\begin{enumerate}
\item If $\mathcal{C} $ is a fusion category realizing case (4) or case (6), then $\text{Inv}(\mathcal{C}) $ is equivalent to $Vec_{\mathbb{Z}/4\mathbb{Z}} $ with trivial associator.
\item If $\mathcal{C} $ is a fusion category realizing case (5), then $\text{Inv}(\mathcal{C}) $ is equivalent to $Vec_{\mathbb{Z}/2\mathbb{Z}\times\mathbb{Z}/2\mathbb{Z}}^\xi $.
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item First assume that $\mathcal{C} $ realizes case (4). By Lemma \ref{lemc}, $ \text{Inv}(\mathcal{C}) $ has a trivial associator, but we need to determine whether the group is cyclic or not. Furthermore, we know that there is a $2$-dimensional algebra in $\mathcal{C}$ which gives a Morita equivalence between the category $\text{Inv}(\mathcal{C})$ and $\text{Inv}(\mathcal{D})$ where $\mathcal{D} $ is a fusion category realizing case (5). But $\text{Inv}(\mathcal{D})$ has nontrivial associator by Lemma \ref{lema}. In $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ any $2$-dimensional algebra has dual category $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}$ with trivial associator, thus $\text{Inv}(\mathcal{C}) $ must be equivalent to $\text{Vec}_{\mathbb{Z}/4\mathbb{Z}}$.
Now assume $\mathcal{C}$ realizes case (6). Then by Lemma \ref{lemb}, $\text{Inv}(\mathcal{C}) $ is Morita equivalent to $\text{Inv}(\mathcal{D}) $, where $\mathcal{D} $ is a fusion category realizing case (4) and the Morita equivalence comes from a $4$-dimensional division algebra. Therefore $\text{Inv}(\mathcal{C}) $ is also equivalent to $\text{Vec}_{\mathbb{Z}/4\mathbb{Z}} $.
\item By Lemma \ref{lema}, $\text{Inv}(\mathcal{C}) $ is Morita equivalent to $\text{Inv}(\mathcal{D}) $, where $\mathcal{D} $ realizes case (4) and the Morita equivalence comes from a $2$-dimensional algebra. Since $\text{Inv}(\mathcal{C}) $ is equivalent to $\text{Vec}_{\mathbb{Z}/4\mathbb{Z}} $ with trivial associator, $\text{Inv}(\mathcal{C}) $ is equivalent to $\text{Vec}_{\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}}^\xi$ by \cite{MR2362670}.
\end{enumerate}
\end{proof}
Now we turn our attention to the non-invertible objects.
\begin{lemma} \label{graphpair}
Let $\mathcal{C} $ be a fusion category realizing case (4). Then all non-invertible simple objects in $\mathcal{C} $ are self-dual.
\end{lemma}
\begin{proof}
This can be deduced by looking at the principal/dual graph pair for $\theta_{24}^{1}$ in $L_{AH_2}^{(4)}$, where the top row corresponds to objects in $\mathcal{C}$ and the bottom row corresponds to objects in $\mathcal{AH}_2 $ (we use the label $\rho' $ in $\mathcal{AH}_2 $ to distinguish the object from the similarly named $\rho $ in $\mathcal{C} $):
$$\begin{tikzpicture}
\draw [fill] (0,0) circle [radius=.05];
\node [below] at (0,0) {$1$};
\draw [fill] (1,0) circle [radius=.05];
\node [below] at (1,0) {$\alpha$};
\draw [fill] (2,0) circle [radius=.05];
\node [below] at (2,0) {$\alpha \pi$};
\draw [fill] (3,0) circle [radius=.05];
\node [below] at (3,0) {$\pi$};
\draw [fill] (4,0) circle [radius=.05];
\node [below] at (4,0) {$\eta$};
\draw [fill] (5,0) circle [radius=.05];
\node [below] at (5,0) {$\rho' \alpha$};
\draw [fill] (6,0) circle [radius=.05];
\node [below] at (6,0) {$\alpha \rho' \alpha$};
\draw [fill] (7,0) circle [radius=.05];
\node [below] at (7,0) {$\rho'$};
\draw [fill] (8,0) circle [radius=.05];
\node [below] at (8,0) {$\alpha \rho'$};
\draw [fill] (1.5,2) circle [radius=.05];
\node [left] at (1.5,2) {$\theta_{42}^1$};
\draw [fill] (2.5,2) circle [radius=.05];
\node [left] at (2.5,2) {$\theta_{42}^2$};
\draw [fill] (3.5,2) circle [radius=.05];
\node [left] at (3.5,2) {$\theta_{42}^3$};
\draw [fill] (4.5,2) circle [radius=.05];
\node [left] at (4.5,2) {$\theta_{42}^4$};
\draw [fill] (5.5,2) circle [radius=.05];
\node [left] at (5.5,2) {$\theta_{42}^5$};
\draw [fill] (6.5,2) circle [radius=.05];
\node [left] at (6.5,2) {$\theta_{42}^6$};
\draw [fill] (.5,4) circle [radius=.05];
\node [above] at (.5,4) {$1$};
\draw [fill] (1.5,4) circle [radius=.05];
\node [above] at (1.5,4) {$\alpha_1$};
\draw [fill] (2.5,4) circle [radius=.05];
\node [above] at (2.5,4) {$\alpha_3$};
\draw [fill] (3.5,4) circle [radius=.05];
\node [above] at (3.5,4) {$\alpha_2$};
\draw [fill] (4.5,4) circle [radius=.05];
\node [above] at (4.5,4) {$\rho$};
\draw [fill] (5.5,4) circle [radius=.05];
\node [above] at (5.5,4) {$\alpha_1 \rho$};
\draw [fill] (6.5,4) circle [radius=.05];
\node [above] at (6.5,4) {$\alpha_3 \rho$};
\draw [fill] (7.5,4) circle [radius=.05];
\node [above] at (7.5,4) {$\alpha_2 \rho$};
\path (0,0) edge (1.5,2);
\path (2,0) edge (1.5,2);
\path (2,0) edge (5.5,2);
\path (2,0) edge (6.5,2);
\path (3,0) edge (5.5,2);
\path (4,0) edge (5.5,2);
\path (5,0) edge (5.5,2);
\path (6,0) edge (5.5,2);
\path (3,0) edge (6.5,2);
\path (4,0) edge (6.5,2);
\path (7,0) edge (6.5,2);
\path (8,0) edge (6.5,2);
\path (4,0) edge (2.5,2);
\path (4,0) edge (3.5,2);
\path (1,0) edge (4.5,2);
\path (3,0) edge (4.5,2);
\path (1.5,2) edge (.5,4);
\path (2.5,2) edge (1.5,4);
\path (3.5,2) edge (2.5,4);
\path (4.5,2) edge (3.5,4);
\path (1.5,2) edge (4.5,4);
\path (2.5,2) edge (5.5,4);
\path (3.5,2) edge (6.5,4);
\path (4.5,2) edge (7.5,4);
\path (5.5,2) edge (4.5,4);
\path (5.5,2) edge (5.5,4);
\path (5.5,2) edge (6.5,4);
\path (5.5,2) edge (7.5,4);
\path (6.5,2) edge (4.5,4);
\path (6.5,2) edge (5.5,4);
\path (6.5,2) edge (6.5,4);
\path (6.5,2) edge (7.5,4);
\end{tikzpicture}$$
By the usual associativity condition (cf. \cite[\S3.1]{1007.1730}), for any two vertices $\theta^i,\theta^j$ on the middle row, the number of paths between $\theta^i$ and $\theta^j$ formed by taking an edge from $\theta^i$ to the top row, switching to the dual of that vertex in the top row, and then taking an edge down to $\theta^j$, must be the same as the number of similar paths taken through the bottom row.
Since the two vertices connected to $\theta_{42}^2$ in the bottom row ($\alpha$ and $\eta$) both correspond to selfdual objects in $\mathcal{AH}_2$, there are exactly two paths from $\theta_{42}^2$ to itself through the bottom graph. Thus there must also be two paths from $\theta_{42}^2$ to itself through the top graph. The only way this can happen is if $\alpha_2 \rho$ is self-dual. Similarly, looking at $\theta_{42}^1$, we see that $\rho$ must be self-dual. There is exactly one path through the bottom graph between $\theta_{42}^2$ and $\theta_{42}^3$, so there must also be only one graph between them in the top graph. Since $\alpha_1$ and $\alpha_3$ are dual to each other there is a path going through those vertices. Thus there can not be a path going through $\alpha_1 \rho$ and $\alpha_3 \rho$, and so those vertices can not be dual to each other and so must be self-dual.
\end{proof}
By Lemma \ref{lemah4}, this completely determines the fusion ring of any fusion category realizing case (4); we call this fusion ring $AH_4$. Let $AH_5 $ be the fusion ring satisfying the fusion rules of Lemma \ref{lemah4} for $G=\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $ with all non-invertible basis elements self-dual.
\begin{lemma}
\begin{enumerate}
\item The fusion ring of any fusion category realizing cases (4) or (6) is $AH_4$.
\item The fusion ring of any fusion category realizing case (5) is $AH_5$.
\end{enumerate}
\end{lemma}
\begin{proof}
For case (5), a similar argument as in Lemma \ref{graphpair} works to show that all the non-invertible objects are self-dual. For case (6) it is a bit more complicated but it can be checked that if two of the non-invertible objects were dual to each other, the fusion ring would not admit a fusion bimodule with $AH_3 $ realizing the fusion module $M_{AH_3}^{(6)}$.
\end{proof}
\begin{remark}
Using a computer we can compute all the fusion modules over $AH_4$, of which there are $28$ up to isomorphism. Similarly, there are $126$ fusion modules over $AH_5$. We initially found many of the results below by doing computer case analysis for all these fusion modules similar to the arguments in \cite{GSbp}. In order to make this paper widely accessible, we have not used these computer arguments below and instead given alternate by-hand proofs that don't split up into so many cases.
\end{remark}
\subsection{Uniqueness for cases (4)-(6)}
\begin{theorem}\label{unq}
If there exists a fusion category realizing any of cases (4)-(6), then there exists a unique fusion category realizing each of these three cases.
\end{theorem}
\begin{proof}
If case (4) can be realized then so can case (5) and (6) as bimodules over algebras of dimension $2$ and $4$ respectively. Similarly, if case (5) or (6) can be realized, then so can (4) as bimodules over algebras of dimension $2$ and $4$ respectively. So we need only show that there is at most one fusion category realizing each case.
First we consider case (4). Let $\mathcal{C}_1, \mathcal{C}_2 $ be fusion categories realizing case (4), and let ${}_{\mathcal{C}_1} \mathcal{K} {}_{\mathcal{C}_2}$ be a Morita equivalence between them. Then there exist Morita equivalences ${}_{\mathcal{C}_1} \mathcal{K}_1 {}_{\mathcal{AH}_2}$ and ${}_{\mathcal{AH}_2} \mathcal{K}_2 {}_{\mathcal{C}_2}$ and a bimodule equivalence ${}_{\mathcal{C}_1} \mathcal{K}_1 {}_{\mathcal{AH}_2} \boxtimes_{\mathcal{AH}_2} {}_{\mathcal{AH}_2} \mathcal{K}_2 {}_{\mathcal{C}_2} \rightarrow {}_{\mathcal{C}_1} \mathcal{K} {}_{\mathcal{C}_2}$. Since any bimodule between $\mathcal{AH}_2 $ and $\mathcal{C}_1$ or $\mathcal{C}_2$ corresponds to the fusion module $L_{AH_2}^{(4)}$, there are simple objects $\kappa_1 \in \mathcal{K}_1, \kappa_2 \in \mathcal{K}_2 $ corresponding to the basis element $\theta_{42}^1$ such that $\bar{\kappa_1} \kappa_1 = 1+\rho = \kappa_2 \bar{\kappa_2}$ in $\mathcal{AH}_2$.
By Frobenius reciprocity, $\kappa_1 \kappa_2$ in $\mathcal{K} $ also has two simple summands and dimension $1+d $. Since each simple object in $\mathcal{K} $ has dimension equal to $\sqrt{i+jd } $ for some non-negative integers $i \leq 4, \ j \leq 32$, this is only possible if the dimensions of the two simple summands are $1$ and $d$, or $2$ and $d-1$. In the latter case, the internal endomorphisms of the $2$-dimensional object, would give a $4$-dimensional division algebra in $\mathcal{C}_1$ whose dual was $\mathcal{C}_2$. But this is impossible because such a dual category would have to realize case (6) and not case (4). Thus, $\mathcal{K}$ has an object of dimension one whose internal endomorphisms must be the trivial algebra. Hence, $\mathcal{C}_2$ is equivalent to the category of bimodules over the trivial algebra in $\mathcal{C}_1$ and thus is equivalent to $\mathcal{C}_1$.
It follows that cases (5) and (6) are also each realized by at most one fusion category. Any fusion category $\mathcal{C} $ which realizes case (6) has a $4$-dimensional division algebra $ \gamma_4$ such that $\mathcal{C}^{\gamma_4} $ realizes case (4). In particular, $\mathcal{C}$ must be the category of bimodules for a $4$-dimensional algebra in a the unique category realizing case (4) and thus is itself unique. Similarly, any fusion category $ \mathcal{C}$ realizing case (5) has a $2$-dimensional algebra $ \gamma_2$ such that $\mathcal{C}^{\gamma_2} $ realizes case (4), so it also must be the category of bimodules for the unique $2$-dimensional algebra in the unique fusion category realizing case (4).
\end{proof}
We will call the three (thus far hypothetical, but unique if they exist) fusion categories which realize these three cases $\mathcal{AH}_4$, $\mathcal{AH}_5$, and $\mathcal{AH}_6$.
\begin{theorem} \label{autothm}
If $\mathcal{AH}_{4-6} $ exist, then every Morita autoequivalence of each of these fusion categories is realized by an outer automorphism.
\end{theorem}
\begin{proof}
We first consider $\mathcal{AH}_4 $, which has fusion ring $ AH_4$. As in the proof of the previous lemma, any Morita autoequivalence must come from the trivial algebra, and is therefore realized by an outer automorphism.
Next we consider $ \mathcal{AH}_5$. There is a Morita equivalence between $\mathcal{AH}_4 $ and $\mathcal{AH}_5 $ which is implemented by a $2$-dimensional algebra. Since every Morita autoequivalence of $\mathcal{AH}_4 $ is realized by an outer automorphism, in fact every $\mathcal{AH}_4 -\mathcal{AH}_5 $ autoequivalence is implemented by a $2$-dimensional algebra. Again as in the proof of the previous lemma, by Frobenius reciprocity this means that every Morita autoequivalence of $\mathcal{AH}_5 $ contains an object with dimension $2$ and two simple summands. This means the summands each have dimension $1$, and the autoequivalence comes from the trivial module category, and is therefore realized by an outer automorphism. For $\mathcal{AH}_6 $ we use the same argument, starting with the $\mathcal{AH}_5 -\mathcal{AH}_6 $ Morita equivalence coming from a $2$-dimensional algebra.
\end{proof}
\begin{corollary}\label{outgroup}
If $\mathcal{AH}_{4-6} $ exist, then $\text{Out}(\mathcal{AH}_i)\cong \mathbb{Z} /2\mathbb{Z} \times \mathbb{Z} /2\mathbb{Z} $ for $i =4,5,6 $.
\end{corollary}
\subsection{Non-existence for case (7)}
In this subsection we show that if $\mathcal{AH}_{4-6}$ exist (as we will prove in Section \ref{sec:main}), then case (7) cannot be realized.
\begin{lemma}\label{15d}
Let $\mathcal{C} $ be a fusion category realizing case (7). Then every invertible $ \mathcal{C}-\mathcal{AH}_5$ bimodule category has a simple object of dimension $\sqrt{1+5d} $.
\end{lemma}
\begin{proof}
Every invertible $ \mathcal{C}-\mathcal{AH}_5$ bimodule category $\mathcal{K} $ can be expressed as a relative tensor product $\mathcal{K}=\mathcal{K}_1 \boxtimes_{\mathcal{AH}_3} \mathcal{K}_2$ of an invertible $\mathcal{C}-\mathcal{AH}_3$ bimodule category $\mathcal{K}_1$ and an invertible $\mathcal{AH}_3-\mathcal{AH}_5 $ bimodule category $\mathcal{K}_2$. Since $\mathcal{C} $ realizes case (7) and $\mathcal{AH}_5 $ realizes case (5), there are simple objects $\theta_{73}^1 \in \mathcal{K}_1$ and $\overline{\theta_{53}^1} \in \mathcal{K}_2$ such that $\overline{\theta_{73}^1} \theta_{73}^1 = 1+\beta \xi+\xi \beta+\mu $ and $\theta_{53}^1 \overline{\theta_{53}^1}= 1+ \mu $ (both objects in $\mathcal{AH}_3 $). Therefore by Frobenius reciprocity, the object $\theta_{73}^1 \overline{\theta_{53}^1} \in \mathcal{K}$ has two simple summands, and its dimension is
$ \sqrt{(2+2d)(1+d)}=\sqrt{2}(1+d)$. Since every object in $\mathcal{K} $ must have dimension
$\sqrt{i + j d} $ for nonnegative integers $i,j$, and neither summand can have integer dimension,
the only possibility is that each summand has dimension $\sqrt{1+5d} = \frac{\sqrt{2}(1+d)}{2}$.
\end{proof}
In the following two lemmas we need to discuss properties of $\mathcal{AH}_5 $. We label the simple objects by $\alpha_{g} $ and $\alpha_{g} \rho $, $g \in G=\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $, where $\rho $ is a non-invertible simple object.
\begin{lemma}\label{algah5}
The object $1+ \alpha_g \rho $ in $\mathcal{AH}_5 $ admits two algebra structures whose dual categories are $\mathcal{AH}_3$ for each $g \in G $.
\end{lemma}
\begin{proof}
Since $\mathcal{AH}_5 $ satisfies case (5), each of the four $\mathcal{AH}_5-\mathcal{AH}_3$ Morita equivalences is implemented by an algebra of dimension $1+d$ in $\mathcal{AH}_5$. Let $1+\alpha_{g_i} \rho $, $1 \leq i \leq 4 $ be algebras corresponding to the four Morita equivalences, and suppose that $g_i=g_j $ for some $i \neq j $. Then as in the proof of Theorem \ref{unq}, we can compose the two bimodule objects corresponding to the algebras $1+\alpha_{g_i} \rho $ and $1+\alpha_{g_j} \rho $ to obtain an object in a $\mathcal{AH}_3 $-Morita auto-equivalence which has dimension $1+d$ and two simple summands. Moreover, the Morita autoequiavelence must be nontrivial since $i \neq j $. However, there are no such objects in the non-trivial Morita autoequivalences of $\mathcal{AH}_3 $. Therefore all of the $g_i $ are distinct, and each $1+\alpha_{g_i} \rho $ admits an algebra structure whose dual category is $\mathcal{AH}_3 $. Since the fusion module $M_{AH_3}^{(5)}$ contains $4$ distinct basis elements of dimension $\sqrt{1+d} $, which form two equivalence classes under the action of the group of invertible objects of $\mathcal{AH}_3 $, there are in fact two such algebras for each $g \in G $.
\end{proof}
\begin{corollary}\label{outah5}
The outer automorphism group of $\mathcal{AH}_5 $ acts transitively on the noninvertible simple objects.
\end{corollary}
\begin{remark}
The situation is quite different for $ \mathcal{AH}_4$. If we let $\beta_h $, $h \in H=\mathbb{Z}/4\mathbb{Z} $ be the invertible objects of $ \mathcal{AH}_4$ and $\xi $ be a noninvertible simple object, then two of the $\mathcal{AH}_4-\mathcal{AH}_2$ Morita equivalences correspond to pairs of algebra structures on $1+\xi $ and $ 1+ \beta_2 \xi $ and the other two Morita equivalences correspond to pairs of algebra structures on $1+\beta_1 \xi $ and $ 1+ \beta_3 \xi $. The argument in the proof of the previous lemma fails here since, unlike for $\mathcal{AH}_3$, there is a nontrivial Morita autoequivalence of $\mathcal{AH}_2 $ with an object of dimension $1+d$ and two simple summands. Specifically, there is a Morita autoequivalence of $\mathcal{AH}_2 $ with simple objects of dimension $\frac{d+1}{2},\frac{d+1}{2},\frac{d+1}{2},\frac{d+1}{2},d-1,d-1,d+1 $ (see \cite{GSbp}).
\end{remark}
\begin{theorem} \label{no7}
If $\mathcal{AH}_{4-6} $ exist, there do not exist any fusion categories in the Morita equivalence class of the $\mathcal{AH}_i $ realizing case (7).
\end{theorem}
\begin{proof}
Let $\mathcal{C} $ be a fusion category realizing case (7). By Lemma \ref{15d}, there exists an object $\kappa $ of dimension $\sqrt{1+5d} $ in an
invertible $ \mathcal{C}-\mathcal{AH}_5$ bimodule category.
Then $\bar{\kappa} \kappa$ contains five summands with dimension $d$, so at least one simple summand has multiplicity at least $2$. Since the outer automorphism group acts transitively on the noninvertible simple objects, we may assume without loss of generality that that summand is $\rho$, and we have
$$(\bar{\kappa} \kappa, \rho) \geq 2 .$$
Let $\theta^1_{53} $ be a simple object in an invertible $ \mathcal{AH}_5-\mathcal{AH}_3$ bimodule category whose internal end is $1+\rho$. Then $$(\kappa \boxtimes \theta^1_{53},\kappa \boxtimes \theta^1_{53})=(\bar{\kappa}\kappa, \theta^1_{53}\overline{ \theta^1_{53}}) \geq 3$$ and $$d(\kappa \boxtimes \lambda)=\sqrt{(1+d)(1+5d)}=\sqrt{46d+6} .$$ However, the dimensions of the basis elements in $N_{AH_3} $ are
$$d(\theta^1_{73})=d(\theta^2_{73}) =\sqrt{2d+2}, \quad d(\theta^3_{73})=\sqrt{28d+4} .$$
Therefore there cannot be an object $ \mathcal{C}-\mathcal{AH}_3$ bimodule category with at least three simple summands and dimension $\sqrt{46d+6}=d(\theta^1_{73})+d(\theta^3_{73})$.
\end{proof}
\section{Existence of $\mathcal{AH}_{4-6} $} \label{sec:main}
In the previous section we established that if any of cases (4)-(6) of Theorem \ref{cases} occur, then there are exactly three additional fusion categories in the Morita equivalence class of $\mathcal{AH}_{1-3} $, which we called $\mathcal{AH}_{4-6} $.
In this section we will show that $\mathcal{AH}_{4-6} $ exist by explicitly constructing $\mathcal{AH}_{4} $ as a category of finite-index endomorphisms of a Type III factor $M$. By Lemmas \ref{lemah4}, \ref{lemc}, and \ref{graphpair}, $\mathcal{AH}_{4}$ must contain a copy of $\text{Vec}_{\mathbb{Z}/4\mathbb{Z}} $, with simple objects $\{\alpha_g\}_{g \in \mathbb{Z}/4\mathbb{Z}} $, along with a self-dual simple object $\rho $, satisfying the fusion rules
$$\alpha_g \rho \cong \rho \alpha_{-g}, \ \forall g \in \mathbb{Z}/4\mathbb{Z}, \quad \text{and }\rho^2\cong 1+2\sum_{g \in \mathbb{Z}/4\mathbb{Z}} \alpha_g \rho .$$
A general theory of such categories for arbitrary finite Abelian groups, but with the multiplicities of $\alpha_g \rho $ in $\rho^2 $ all equal to $1$ instead of $2$, has been developed by the second-named author in \cite{MR1228532, MR1832764,IzumiNote}. He determined a complete numerical invariant of such categories given by solutions of certain polynomial equations determined by the group $G$; these solutions give structure constants for endomorphisms $\{\alpha_g\}_{g \in G}$ and $ \rho $ of the Cuntz algebra $\mathcal{O}_{|G|+1} $, which can then be extended to endomorphisms of a Type III factor. In this construction one Cuntz algebra generator belongs to each intertwiner space $(\alpha_g,\rho^2) $
and one generator belongs to $(1,\rho^2) $.
We could try to perform a similar construction for $\mathcal{AH}_{4} $ by realizing $\rho $ as an endomorphism of $\mathcal{O}_{2|G|+1}=\mathcal{O}_9 $, again with one generator belonging to $(1,\rho^2) $, but this time with two generators belonging to each $(\alpha_g,\rho^2) $. However, because the intertwiner spaces are two-dimensional, it is difficult to write down equations for the structure constants of $\rho $. We therefore take a different approach, and construct instead an equivariantization of $\mathcal{AH}_{4} $, which is technically easier to deal with. We then recover $\mathcal{AH}_{4} $ by de-equivariantizing.
\subsection{Equivariantization and the orbifold construction}
We recall the notion of equivariantization of a fusion category by the action of a finite group of automorphisms. For simplicity we define equivariantization for strict categories, which include categories of endomorphisms (though as with all constructions for monoidal categories, strictness is not really important).
Let $\mathcal{C} $ be a strict fusion category, and let $G$ be a finite group acting by automorphisms $\{ \phi_g \}_{g \in G}$ on $\mathcal{C} $. A $G$-equivariant object of $\mathcal{C} $ is an object $\xi \in \mathcal{C} $, together with isomorphisms $\{ u_g: \phi_g(\xi) \rightarrow \xi \}_{g \in G}$, such that $$u_g \phi_g(u_h)=u_{gh} . $$
Morphisms and tensor product for $G$-equivariant objects can be defined in a natural way, and the $G$-equivariant objects form a fusion category, called the $G$-equivariantization of $\mathcal{C} $.
Suppose $ \mathcal{C}$ is a fusion category of finite-index endomorphisms of a Type III factor $M$, and suppose $\alpha $ is an order two automorphism of $M$ which commutes with a set of representative endomorphisms of the simple objects of $\mathcal{C} $. Since $\mathcal{C} $ is equivalent to its full subcategory of endomorphisms which commute with $\alpha $, we assume without loss of generality that $\alpha $ commutes with all endomorphisms in $\mathcal{C} $. Then $\alpha $ induces a $\mathbb{Z}/2\mathbb{Z} $-action on $\mathcal{C} $, where the action of $\alpha $ fixes objects and takes each morphism $u \in (\xi,\sigma) $ to $\alpha(u) \in (\xi, \sigma ) $. The equivariant objects with respect to this action are of the form $(\xi,u)$, with $u \in (\xi , \xi )$ satisfying $u \alpha(u)=1$. In particular for each simple object $\xi \in \mathcal{C}$, there are two inequivalent objects in the $\mathbb{Z}/2\mathbb{Z} $-equivariantization, corresponding to $u=\pm1$.
Now suppose that $\alpha$ is not in the same sector as any object of $\mathcal{C} $. Then the $ \mathbb{Z}/2\mathbb{Z}$-equivariantization of $\mathcal{C} $ induced by $\alpha $ can be realized as a category of endomorphisms on a larger factor via an orbifold construction, as follows. Let $M_1$ be the crossed product $M \rtimes_{\alpha} \mathbb{Z}/2\mathbb{Z}$, which is the von Neumann algebra generated by $M $ and a self-adjoint unitary $\lambda $ satisfying $$\lambda x \lambda=\alpha(x), \ \forall x \in M.$$
Then every equivariant object $(\xi,u)$ can be extended to an endomorphism $\tilde{\xi} $ on $M_1$ by setting
$\xi(\lambda)=u \lambda$, and the morphims between pairs of such extended endomorphisms $\tilde{\xi}_1 $ and $\tilde{\xi}_2 $ in $ \text{End}_0(M_1) $ are precisely the equivariant morphisms between $\xi_1 $ and $ \xi_2$ in $\text{End}_0(M) $.
\subsection{ $\mathbb{Z}/2\mathbb{Z}$-equivariantization of $\mathcal{AH}_4 $}
This section is purely motivational and explains why if $\mathcal{AH}_4$ exists, then it must have a $\mathbb{Z}/2\mathbb{Z}$-equivariantization $\mathcal{S}$ which is easier to understand. Logically this is not necessary, since we will directly construct this $\mathcal{S}$ and use it to construct $\mathcal{AH}_4$ via the inverse process of de-equivariantization.
The following can be shown with a considerable amount of work.
\begin{lemma}\label{lemmaeq}
Suppose $\mathcal{AH}_4 $ exists. Let $H=\mathbb{Z}/4 \mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $. Then the $\mathbb{Z}/2\mathbb{Z}$-equivariantization with respect to the action coming from conjugation by $\alpha_2 $ contains a copy of $\text{Vec}_{H} $, with simple objects $
\{\alpha_g\}_{g \in H} $, along with a simple object $\rho $, satisfying the fusion rules $$\alpha_g \rho \cong \rho \alpha_{-g}, \ \forall g \in H, \quad \text{and }\rho^2\cong 1+\sum_{g \in H} \alpha_g \rho .$$
\end{lemma}
We do not prove this lemma here, since the result is not required for anything that follows and the proof is difficult and detailed. Rather the lemma functions as motivation for constructing $\mathcal{AH}_4 $ by first constructing its $\mathbb{Z}/2\mathbb{Z}$-equivariantization. However, we note several features of this proof which will motivate some choices in the next section.
\begin{remark}
\begin{enumerate}
\item The proof proceeds by realizing (the hypothetical) $\mathcal{AH}_4 $ as the fusion category generated by automorphisms $\{\alpha_g\}_{g \in \mathbb{Z}/4\mathbb{Z}} $, along with an endomorphism $\rho $, on a Type III facor $M$ containing a Cuntz algebra $\mathcal{O}_9 $, with one generator belonging to $(1,\rho^2) $
and two generators belonging to each $(\alpha_g \rho,\rho^2) $. Then proving the lemma amounts to showing that the eigenvalues of the action of $\alpha_2 $ on each $(\alpha_g \rho,\rho^2) $ are $\{1,-1\} $.
\item Finding the eigenvalues of $\alpha_2 $ required using properties of $\mathcal{AH}_4 $ beyond just the fusion rules and the presence of a $\text{Vec}_{ \mathbb{Z}/4\mathbb{Z}} $ subcategory, namely that $1+ \rho $ admits a Q-system, that the dual category with respect to this Q-system contains a non-trivial invertible object, and that the center of $\mathcal{AH}_4 $ contains no non-trivial invertible objects (and therefore conjugation by $\alpha_2 $ is a non-trivial automorphism). These properties follow from the fact that $ \mathcal{AH}_4$ is assumed to be the dual category of $\mathcal{AH}_2 $ with repect to a Q-system of dimension $1+d$.
\item Since $\alpha_2 $ belongs to $\mathcal{AH}_4 $, we cannot simply use a crossed product by $\alpha_2 $ for the orbifold construction, as that would result in $\alpha_2 $ being identified with the identity automorphism on the Type III factor $M$. Rather, we need to twist $\alpha_2$ by an automorphism of $M$ to obtain an automorphism $ \alpha'$ which is not in the same sector as any automorphism of $\mathcal{AH}_4 $, but such that conjugation by $\alpha' $ is equivalent to conjugation by $\alpha_2 $ as an automorphism of $\mathcal{AH}_4 $.
\item From the orbifold construction, we can deduce some important structure constants of the
$\mathbb{Z}/2\mathbb{Z}$-equivariantization, which will be useful in the following subsection.
\end{enumerate}
\end{remark}
\subsection{Constructing $\mathcal{S}$}
By Lemma \ref{lemmaeq}, the $\mathbb{Z}/2\mathbb{Z}$-equivariantization of $\mathcal{AH}_4$ would have nicer fusion rules than $\mathcal{AH}_4 $ itself. We now directly construct a subfactor $\mathcal{S}$ whose principal even half has these fusion rules.
We recall the following construction from \cite{IzumiNote}, which completely characterizes unitary fusion categories which are tensor generated by $\text{Vec}_G $ for a finite Abelian group $G$ (which for us is $\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$) and a simple object $ \rho$, satisfying $$\alpha_g \rho \cong \rho \alpha_{-g}, \ \forall g \in G, \quad \text{and }\rho^2\cong 1+\sum_{g \in G} \alpha_g \rho .$$
Let $G$ be a finite Abelian group, and let $$A: G \times G \times G \rightarrow \mathbb{C} $$ $$\epsilon: G \times G \rightarrow \{-1,1\}$$
$$\eta: G \rightarrow \{ 1, e^{\frac{2\pi i }{3}}, e^{\frac{-2\pi i }{3} } \} $$ be functions. Set $A_g(h,k) =A(g,h,k)$, $\epsilon_g(h)=\epsilon(g,h)$, and $\eta_g=\eta(g) $ for $g,h,k \in G $. Let $n=|G| $ and let $$d=\displaystyle \frac{n+\sqrt{n^2+4}}{2} .$$
Consider the following system of equations:
\begin{equation} \label{e1}
\epsilon_{h+k}(g)=\epsilon_h(g)\epsilon_k(g+2h), \ \epsilon_h(0)=1
\end{equation}
\begin{equation}
\eta_{g+2h}=\eta_g, \ \eta_g^3=1
\end{equation}
\begin{equation}\label{e3}
\sum_{h \in G} A_g(h,0)=-\frac{\overline{\eta_g}}{d}
\end{equation}
\begin{equation}\label{e4}
\sum_{h \in G} A_g(h-g,k)\overline{A_{g'}(h-g',k)}=\delta_{g,g'}-\frac{\overline{\eta_g}\eta_{g'}}{d} \delta_{k,0}
\end{equation}
\begin{equation}\label{e5}
A_{g+2h}(p,q)=\epsilon_h(g)\epsilon_h(g+p)\epsilon_h(g+q)\epsilon_h(g+p+q)A_g(p,q)
\end{equation}
\begin{equation}
A_g(h,k)=\overline{A_g(k,h)}
\end{equation}
\begin{align} \label{e8}
A_g(h,k) &=A_g(-k,h-k)\eta_g \epsilon_{-k}(g+h)\epsilon_{-k}(g+k)\epsilon_{-k}(g+h+k) \\
&=A_g(k-h,-h) \overline{\eta_g} \epsilon_{-h}(g+h)\epsilon_{-h}(g+k)\epsilon_{-h}(g+h+k) \nonumber
\end{align}
\begin{align}\label{e9}
A_g(h,k)&=A_{g+h}(h,k) \eta_g \eta_{g+k} \overline{\eta_{g+h} \eta_{g+h+k}} \epsilon_h(g)\epsilon_h(g+k) \\
&= A_{g+k}(h,k) \overline{\eta_g \eta_{g+h}} \eta_{g+k} \eta_{g+h+k}\epsilon_k(g) \epsilon_k(g+h) \nonumber
\end{align}
\begin{align} \label{e10}
\sum_{l \in G} \limits & A_g(x+y,l)A_{g-p+x}(-x,l+p)A_{g-q+x+y}(-y,l+q) \\
&=A_g(p+x,q+x+y)A_{g-p}(q+y,p+x+y) \nonumber \\
& \times \eta_{g} \eta_{g+q+x} \eta_{g+p+q+y} \overline{ \eta_{g+p} \eta_{g+x+y} \eta_{g+q+x+y} } \nonumber \\
&\times \epsilon_p(g-p+x)\epsilon_{p+x}(g-p+q+y)
\epsilon_q(g-q+x+y)\epsilon_{q+y}(g-q+x) \nonumber \\
& -\displaystyle \frac{\delta_{x,0}\delta_{y,0,}}{d} \eta_g \eta_{g+p} \eta_{g+q} \nonumber
\end{align}
\begin{theorem}\label{izthm} \cite{IzumiNote}
Suppose $A$ and $\epsilon $ satisfy (\ref{e1})-(\ref{e10}).
Then there is a Type III factor $M$, an outer action $\alpha $ of $ G$ on $M$, an irreducible finite-index endomorphism $\rho $ of $M$, and $n+1$ isometries $S \in (id,\rho^2) $ and $T_g \in (\alpha_g \circ \rho,\rho^2)$ satisfying the Cuntz algebra relations such that:\\
\begin{equation} \label{e11}
\alpha_g \circ \rho = \rho \circ \alpha_{-g}, \, \forall g \in G
\end{equation}
\begin{equation}\label{e12}
\alpha_g (S)=S, \ \ \alpha_g(T_h)=\epsilon_g(h)T_{h+2g}, \, \forall g,h \in G
\end{equation}
\begin{equation}\label{e12_2}
\rho(s)=\frac{1}{d}S+\frac{1}{\sqrt{d}} \sum_{g \in G} \limits T_gT_g
\end{equation}
\begin{multline}\label{e13}
\rho(T_g)=\epsilon_g(-g)[\eta_{-g}T_{-g}SS^*+\frac{\overline{\eta_{-g}}}{\sqrt{d}}ST_{-g}^* \\ +\sum_{h,k \in G}A_{
-g}(h,k)T_{h-g}T_{h+k-g}T_{k-g}^*] ,\, \forall g\in G.
\end{multline}
The sector $[id] \oplus[\alpha_g \rho] $ admits a Q-system iff
\begin{equation} \label{e14}
A_g(h,0)=\delta_{h,0}-\frac{1}{d-1}
\, \forall h \in G.
\end{equation}
\end{theorem}
These equations were solved in \cite{IzumiNote} for all groups of order less than or equal to $5$, and solutions have been found for some groups of higher order. For all known solutions for which $[id] \oplus [\rho] $ admits a Q-system,
$[id] \oplus [\alpha_g \rho ] $ admits Q-systems for all $g \in G $. In this case we get some very useful extra relations.
\begin{lemma}\cite{IzumiNote}
If $[id] \oplus [\alpha_g \rho] $ admits a Q-system for each $g \in G $, then $\eta_g$ is identically $1$ and $\epsilon $ restricted to $G_2 \times G$ is a bicharacter, where $G_2 $ is the subgroup of order $2$ elements of $G$.
Moreover, we have
\begin{equation}\label{eq14}
|A_g(h,k) |^2=\delta_{h,0} \delta_{k,0} - \frac{\delta_{h,0}+\delta_{k,0}+\delta_{h,k}}{d-1}+\frac{d}{(d-1)^2}
\end{equation}
and
\begin{align}\label{eq15}
& A_g(-l,h)A_g(l,k)\epsilon_l(g-l)\epsilon_l(g+h-l) \nonumber \\
& -A_g(h+l,k)A_g(k-l,h) \epsilon_l(g+k-l) \epsilon_l (g+h+k-l) \nonumber \\
& = \frac{\delta_{h-k+l,0}}{d-1} \epsilon_l (g+h) -\frac{\delta_{l,0}}{d-1} \nonumber \\
\end{align}
for all $g,h,k,l \in G$.
\end{lemma}
The structure of a solution to (\ref{e1})-(\ref{e10}) seems to depend in a fundamental way on $\epsilon $, which is determined up to gauge equivalence by its restriction to $G_2 \times G $. From the orbifold construction in the preceding subsection, we can determine that in order for $\mathcal{S}$ to be related to the Asaeda--Haagerup subfactor, the gauge equivalence class of $\epsilon$
on $G=\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $ for the $ \mathbb{Z}/2\mathbb{Z}$-equivariantization of $\mathcal{AH}_4 $ should satisfy
$$\epsilon_{(0,1)}((a,b))=1, \ \epsilon_{(2,0)}((a,b))=(-1)^b, \quad \forall (a,b) \in \mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} .$$
We fix $\epsilon$ by setting $\epsilon_{(1,0)} ((2,1) ) =\epsilon_{(1,0)}((3,1))=-1$ and
$\epsilon_{(1,0)} ((a,b) )=1$ otherwise. Together with $\epsilon_{(0,1)} ((a,b))=1$ and (\ref{e1}) this determines all the $\epsilon$.
Then we can try to solve for $A_g(h,k) $ using $(\ref{e3})-(\ref{e10}) $ and possibly $(\ref{eq14})-(\ref{eq15})$. We know of no general technique for solving these equations, but nonetheless were able to find a solution in this particular case.
\begin{theorem}
There exists a solution to (\ref{e1})-(\ref{e10}) and (\ref{e14}) for the group
$G=\mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $
with $\epsilon_{(2,0)}((a,b))= (-1)^b$ and $\epsilon_{(0,1)} ((a,b))=1$ for all $(a,b) $.
\end{theorem}
\begin{proof}
We explicitly construct a solution is as follows. We define constants
$$c=\frac{1}{4}(1-d+i\sqrt{10d-2}) , \quad f=\sqrt{\frac{1}{2}(d-1-i\sqrt{26d+2})}, $$ $$ g=\frac{1}{2}\sqrt{-3d-1+i\sqrt{50d+6} }, \quad h=\frac{1}{4}(d+3-i(\sqrt{2d-10})).$$
Define $G \times G$ matrices
$$A=\frac{1}{d-1}\left(
\begin{array}{cccccccc}
d-2 & -1 & -1 & -1 & -1 & -1 & -1 & -1 \\
-1 & -1 & c & c & -f & f & -g & -g \\
-1 & \bar{c} & -1 & c & i \sqrt{d} & h & -i \sqrt{d} & \bar{h} \\
-1 & \bar{c} & \bar{c} & -1 & -\bar{f} & -\bar{g} & \bar{g} & -\bar{f} \\
-1 & -\bar{f} & -i \sqrt{d} & -f & -1 & -f & i \sqrt{d} & -\bar{f} \\
-1 & \bar{f} & \bar{h} & -g & -\bar{f} & -1 & g & -\bar{h} \\
-1 & -\bar{g} & i \sqrt{d} & g & -i \sqrt{d} & \bar{g} & -1 & -g \\
-1 & -\bar{g} & h & -f & -f & -h & -\bar{g} & -1 \\
\end{array}
\right)$$
$$
B_{(1,0)}=\left(
\begin{array}{cccccccc}
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & 1 & 1 & 1 & -1 & -1 \\
1 & 1 & 1 & 1 & 1 & -1 & 1 & -1 \\
1 & 1 & 1 & 1 & -1 & 1 & 1 & -1 \\
1 & 1 & 1 & -1 & 1 & 1 & 1 & -1 \\
1 & 1 & -1 & 1 & 1 & 1 & 1 & -1 \\
1 & -1 & 1 & 1 & 1 & 1 & 1 & -1 \\
1 & -1 & -1 & -1 & -1 & -1 & -1 & 1 \\
\end{array}
\right)
$$
$$B_{(0,1)}\left(
\begin{array}{cccccccc}
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & -1 & 1 & 1 & 1 & -1 \\
1 & 1 & -1 & -1 & 1 & 1 & -1 & -1 \\
1 & -1 & -1 & -1 & 1 & -1 & -1 & -1 \\
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & 1 & -1 & 1 & 1 & 1 & -1 \\
1 & 1 & -1 & -1 & 1 & 1 & -1 & -1 \\
1 & -1 & -1 & -1 & 1 & -1 & -1 & -1 \\
\end{array}
\right)$$ $$B_{(1,1)}=\left(
\begin{array}{cccccccc}
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
1 & 1 & -1 & 1 & 1 & 1 & 1 & -1 \\
1 & -1 & -1 & 1 & 1 & 1 & -1 & -1 \\
1 & 1 & 1 & -1 & -1 & 1 & 1 & 1 \\
1 & 1 & 1 & -1 & 1 & 1 & 1 & -1 \\
1 & 1 & 1 & 1 & 1 & 1 & -1 & -1 \\
1 & 1 & -1 & 1 & 1 & -1 & -1 & -1 \\
1 & -1 & -1 & 1 & -1 & -1 & -1 & -1 \\
\end{array}
\right)$$
We set $$A_{(0,0)}(h,k)=A(h,k), \quad A_{(1,0)}(h,k)=B_{(1,0)}(h,k) A(h,k), $$ $$\quad A_{(0,1)}(h,k)=B_{(0,1)}(h,k) A(h,k), \quad A_{(1,1)}(h,k)=B_{(1,1)}(h,k) A(h,k) $$
and use (\ref{e5}) to define the remaining $A_g(h,k)$.
We need only check that these numbers satisfy the finitely many polynomial equations (\ref{e1})-(\ref{e10}) and (\ref{e14}). Although it is possible to check any one of these equations by hand, since there are 40600 total equations (mostly from Equation \ref{e10}) we checked them by computer.
\end{proof}
\begin{remark}
The computer calculation in the above theorem has been checked twice independently. Our original program used a mix of C and mathematica and ran in several hours. A second program was written by Morrison, Penneys, and Peters in mathematica alone and is more easily human-readable but ran in a couple days. We have included along with the arxiv source to this article a improved version of their code as CheckingSolution.nb which is much faster and runs in under 3 minutes on a 2.4GHz Intel Core i5. We are grateful to Morrison, Penneys, and Peters for their help improving this code.
\end{remark}
\begin{remark}
The initial step in solving (\ref{e1})-(\ref{e10}) is determining the cocycle $\epsilon $ up to gauge equivalence. In our case we were able to determine what $\epsilon $ should be by assuming that the resulting category had an orbifold quotient which is Morita equivalent to the Asaeda-Haagerup categories. Then (\ref{e5})-(\ref{e9}) can be used to express the $A_g(h,k)$ in terms of a relatively small number of variables, which can be solved using (\ref{e4}) and (\ref{eq15}) for $g=0$. Although we were not assuming that $1+\alpha_g\rho $ admits a Q-system for all $g$, we were assuming that $1+\rho $ admits a $Q$-system, so (\ref{e14}) has to hold for $g=0$ and consequently so does (\ref{eq15}) for $g=0$. If one assumes that (\ref{e14}) holds for all $g$ (which indeed turns out to be the case), then the equations are easier to solve.
\end{remark}
We will call the fusion category corresponding to this solution $\mathcal{AH}^{\mathbb{Z}/2\mathbb{Z}}$ and the subfactor coming from $1+\rho$ we will call $\mathcal{S}$.
\subsection{De-equivariantization and the orbifold construction}
We can now construct $\mathcal{AH}_4 $ by de-equivariantizing $\mathcal{AH}^{\mathbb{Z}/2\mathbb{Z}} $; the de-equivariantization is implemented by another orbifold construction.
Let $G$ be a finite Abelian group and let $ A$ and $\epsilon$ be a solution to (\ref{e1})-(\ref{e10}). Let $0\neq z \in G_2 $ be such that $\epsilon_g(\cdot) $ is a character and $ \epsilon_{z}(z)=1$.
Let $P= M \rtimes_{\alpha_z} \mathbb{Z}/2\mathbb{Z}$ be the crossed product, which is the von Neumann algebra generated by $M$ and a unitary $\lambda $ satisfying $\lambda^2=1 $
and $\lambda x \lambda=\alpha_z(x) $ for all $x \in M $.
We extend $\alpha_g $ and $\rho $ to $P$ by setting
$\tilde{\alpha}_g(\lambda)=\epsilon_z(g) \lambda $ and $\tilde{\rho}(\lambda)=\lambda $.
Then $\tilde{\alpha}$ is a $G$-action on $P $ and $\tilde{\alpha}_z=\text{Ad } \lambda $.
\begin{theorem}\cite{IzumiNote}
\begin{enumerate}
\item $[\tilde{\rho}] $ is irreducible and self-dual. $[id] \oplus [\tilde{\rho}] $ admits a Q-system
if $[id] \oplus [\rho] $ does.
\item $[\tilde{\alpha}_g]=[\tilde{\alpha}_h] $ iff $g-h \in \{ 0,z \} $.
\item $$ [\tilde{\rho}^2]=[id]\oplus \bigoplus_{\dot{g} \in G / \{0,z\} }2 [\tilde{\alpha}_{\dot{g}} \tilde{\rho}] .$$
\end{enumerate}
\end{theorem}
We can apply this construction to $\mathcal{AH}^{\mathbb{Z}/2\mathbb{Z} } $.
\begin{theorem} \label{ahcons}
There is a Type III factor $M$ with an outer action $\alpha$ of $\mathbb{Z}/4\mathbb{Z} $, an irreducible finite-index endomorphism $\rho $ and isometries $S \in (id, \rho^2 )$ and $T_{(a,b)} \in (\alpha_{a} \rho, \rho^2 ) ,\ (a,b) \in \mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/ 2\mathbb{Z}$
such that
\begin{enumerate}
\item $\alpha_g \rho= \rho \alpha_{-g} $ and $[\rho]^2=[id]\oplus 2\sum_{g \in \mathbb{Z}/4\mathbb{Z} } \limits [\alpha_g \rho] $
\item (\ref{e12})-(\ref{e13}) hold with $A$ and $\epsilon $ as in the previous subsection for all $\alpha_g, \ g \in \mathbb{Z}/4\mathbb{Z} \subset \mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $.
\end{enumerate}
\end{theorem}
\begin{proof}
Apply the orbifold construction to $\mathcal{AH}^{\mathbb{Z}/2\mathbb{Z} } $
with $z=(0,1)$. Note that
$\alpha_{z} $ acts trivially on the original Cuntz algebra, since $z$ has order $2$ and $\epsilon_z $ is identically $1$. After replacing $T_{(a,1)} $ with $T_{(a,1)} \lambda$ for each $a \in \mathbb{Z}/4\mathbb{Z} $, it is easy to check that (\ref{e12})-(\ref{e13}) still hold.
\end{proof}
We will call the fusion category tensor generated by $\rho$ in the above theorem $\mathcal{AH}_{4'}$, since we will show below that it is $\mathcal{AH}_4$.
\begin{remark}
By looking at the details of the proof of Theorem \ref{izthm}, one can give a more direct construction of $\mathcal{AH}_{4'}$. The way that Theorem \ref{izthm} is proven is by using the structure constants $ A$, $ \epsilon$, and $\eta $ to define an endomorphism $\rho $ and a $G$-action $\alpha $ of a Cuntz algebra and then completing this Cuntz algebra to a von Neumann algebra with respect to an appropriate state. Then all of the properties in Theorem \ref{izthm} are satisfied except possibly for the action of $G$ being outer. One then needs to make an extra modification in order to enforce outerness. In our case, this extra modification is exactly cancelled out by the de-equivariantization. Since $\epsilon_z $ is trivial (which means that $\alpha_{(0,1)} $ acts trivially on the Cuntz algebra) the category that one gets on the von Neumann algebra completion of the Cuntz algebra without any modification is already $\mathcal{AH}_{4'} $. Thus it is not strictly necessary to go through the orbifold construction here.
On the other hand, without using the fact that $\mathcal{AH}_4 $ would have to come from a de-equivariantization of a generalized Haagerup category for $ \mathbb{Z}/4\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$, we would have no way of knowing that $\mathcal{AH}_4$ comes from a solution to (\ref{e1})-(\ref{e10}). Thus we found the two-step construction more natural.
\end{remark}
\subsection{Reconstruction of the Asaeda-Haagerup subfactor}
To show that the fusion category $\mathcal{AH}_{4'}$ constructed in the previous subsection satisfies the defining characterization of $\mathcal{AH}_4 $ (and therefore show that $\mathcal{AH}_{4-6}$ exist), it remains to show that $\mathcal{AH}_{4'} $ is Morita equivalent to $\mathcal{AH}_{1-3}$. We will show this by deducing that the dual category of $\mathcal{AH} _{4'}$ with respect to a Q-system for $[id] \oplus [\rho] $ is $\mathcal{AH}_2$. In the process, we will arrive at a new construction of the Asaeda-Haagerup subfactor.
Let $M$ be a Type III factor with endomorphisms $\rho $ and $\alpha_g, \ g \in \mathbb{Z}/4\mathbb{Z} $ realizing $\mathcal{AH}_{4'}$, as in Theorem \ref{ahcons}.
Then since the sector $[id] \oplus [\rho] $ admits a Q-system, there is a Type III factor $N$ and a finite-index homomorphism $\iota: N \rightarrow M $
with a dual homomorphism $\bar{\iota}: M \rightarrow N $ such that $[\iota \bar{\iota}]=[id]\oplus [\rho] $. The fusion ring of $\mathcal{AH}_{4'}$ is $AH_4 $.
\begin{lemma} \label{nti}
The fusion category $\mathcal{AH}_{2'}$ of endomorphisms of $N $ tensor generated by $\bar{\iota} \iota $ has a non-trivial invertible object $\alpha$ satisfying $\alpha_2 \iota = \iota \alpha$.
\end{lemma}
\begin{proof}
It suffices to show that $\alpha_2 \iota $ determines an equivalent $Q$-system to that of $\iota $, since the only way two different sectors can determine equivalent $Q$-systems is if they are in the same orbit under the action of the invertible objects of the dual category \cite[\S3.3]{GSbp}. Since conjugation by $\alpha_2 $ fixes $\rho $ and
$\alpha_2 $ fixes $S$ and $T_0$, this is indeed the case.
\end{proof}
\begin{lemma} \label{2graphs}
The principal graph pair of $ \iota$ is either
$$\begin{tikzpicture}
\draw [fill] (0,0) circle [radius=.05];
\node [below] at (0,0) {$1$};
\draw [fill] (1,0) circle [radius=.05];
\node [below] at (1,0) {$\alpha$};
\draw [fill] (2,0) circle [radius=.05];
\node [below] at (2,0) {$\pi_1$};
\draw [fill] (3,0) circle [radius=.05];
\node [below] at (3,0) {$\pi_2$};
\draw [fill] (4,0) circle [radius=.05];
\node [below] at (4,0) {$\eta$};
\draw [fill] (5,0) circle [radius=.05];
\node [below] at (5,0) {$\sigma_1$};
\draw [fill] (6,0) circle [radius=.05];
\node [below] at (6,0) {$\sigma_2$};
\draw [fill] (7,0) circle [radius=.05];
\node [below] at (7,0) {$\sigma_3$};
\draw [fill] (8,0) circle [radius=.05];
\node [below] at (8,0) {$\sigma_4$};
\draw [fill] (1.5,2) circle [radius=.05];
\node [left] at (1.5,2) {$\iota$};
\draw [fill] (2.5,2) circle [radius=.05];
\node [left] at (2.5,2) {$\alpha_1 \iota$};
\draw [fill] (3.5,2) circle [radius=.05];
\node [left] at (3.5,2) {$\alpha_3 \iota$};
\draw [fill] (4.5,2) circle [radius=.05];
\node [left] at (4.5,2) {$\alpha_2 \iota$};
\draw [fill] (5.5,2) circle [radius=.05];
\node [left] at (5.5,2) {$\kappa$};
\draw [fill] (6.5,2) circle [radius=.05];
\node [left] at (6.5,2) {$ \alpha_1 \kappa$};
\draw [fill] (.5,4) circle [radius=.05];
\node [above] at (.5,4) {$1$};
\draw [fill] (1.5,4) circle [radius=.05];
\node [above] at (1.5,4) {$\alpha_1$};
\draw [fill] (2.5,4) circle [radius=.05];
\node [above] at (2.5,4) {$\alpha_3$};
\draw [fill] (3.5,4) circle [radius=.05];
\node [above] at (3.5,4) {$\alpha_2$};
\draw [fill] (4.5,4) circle [radius=.05];
\node [above] at (4.5,4) {$\rho$};
\draw [fill] (5.5,4) circle [radius=.05];
\node [above] at (5.5,4) {$\alpha_1 \rho$};
\draw [fill] (6.5,4) circle [radius=.05];
\node [above] at (6.5,4) {$\alpha_3 \rho$};
\draw [fill] (7.5,4) circle [radius=.05];
\node [above] at (7.5,4) {$\alpha_2 \rho$};
\path (0,0) edge (1.5,2);
\path (2,0) edge (1.5,2);
\path (2,0) edge (5.5,2);
\path (2,0) edge (6.5,2);
\path (3,0) edge (5.5,2);
\path (4,0) edge (5.5,2);
\path (5,0) edge (5.5,2);
\path (6,0) edge (5.5,2);
\path (3,0) edge (6.5,2);
\path (4,0) edge (6.5,2);
\path (7,0) edge (6.5,2);
\path (8,0) edge (6.5,2);
\path (4,0) edge (2.5,2);
\path (4,0) edge (3.5,2);
\path (1,0) edge (4.5,2);
\path (3,0) edge (4.5,2);
\path (1.5,2) edge (.5,4);
\path (2.5,2) edge (1.5,4);
\path (3.5,2) edge (2.5,4);
\path (4.5,2) edge (3.5,4);
\path (1.5,2) edge (4.5,4);
\path (2.5,2) edge (5.5,4);
\path (3.5,2) edge (6.5,4);
\path (4.5,2) edge (7.5,4);
\path (5.5,2) edge (4.5,4);
\path (5.5,2) edge (5.5,4);
\path (5.5,2) edge (6.5,4);
\path (5.5,2) edge (7.5,4);
\path (6.5,2) edge (4.5,4);
\path (6.5,2) edge (5.5,4);
\path (6.5,2) edge (6.5,4);
\path (6.5,2) edge (7.5,4);
\end{tikzpicture} $$ or $$ \begin{tikzpicture}
\draw [fill] (0,0) circle [radius=.05];
\node [below] at (0,0) {$1$};
\draw [fill] (1,0) circle [radius=.05];
\node [below] at (1,0) {$\alpha$};
\draw [fill] (2,0) circle [radius=.05];
\node [below] at (2,0) {$\pi_1$};
\draw [fill] (3,0) circle [radius=.05];
\node [below] at (3,0) {$\pi_2$};
\draw [fill] (4,0) circle [radius=.05];
\node [below] at (4,0) {$\eta$};
\draw [fill] (6.5,0) circle [radius=.05];
\node [below] at (6.5,0) {$\sigma$};
\draw [fill] (1.5,2) circle [radius=.05];
\node [left] at (1.5,2) {$\iota$};
\draw [fill] (2.5,2) circle [radius=.05];
\node [left] at (2.5,2) {$\alpha_1 \iota$};
\draw [fill] (3.5,2) circle [radius=.05];
\node [left] at (3.5,2) {$\alpha_3 \iota$};
\draw [fill] (4.5,2) circle [radius=.05];
\node [left] at (4.5,2) {$\alpha_2 \iota$};
\draw [fill] (5.5,2) circle [radius=.05];
\node [left] at (5.5,2) {$\kappa$};
\draw [fill] (6.5,2) circle [radius=.05];
\node [left] at (6.5,2) {$ \alpha_1 \kappa$};
\draw [fill] (.5,4) circle [radius=.05];
\node [above] at (.5,4) {$1$};
\draw [fill] (1.5,4) circle [radius=.05];
\node [above] at (1.5,4) {$\alpha_1$};
\draw [fill] (2.5,4) circle [radius=.05];
\node [above] at (2.5,4) {$\alpha_3$};
\draw [fill] (3.5,4) circle [radius=.05];
\node [above] at (3.5,4) {$\alpha_2$};
\draw [fill] (4.5,4) circle [radius=.05];
\node [above] at (4.5,4) {$\rho$};
\draw [fill] (5.5,4) circle [radius=.05];
\node [above] at (5.5,4) {$\alpha_1 \rho$};
\draw [fill] (6.5,4) circle [radius=.05];
\node [above] at (6.5,4) {$\alpha_3 \rho$};
\draw [fill] (7.5,4) circle [radius=.05];
\node [above] at (7.5,4) {$\alpha_2 \rho$};
\path (0,0) edge (1.5,2);
\path (2,0) edge (1.5,2);
\path (2,0) edge (5.5,2);
\path (2,0) edge (6.5,2);
\path (3,0) edge (5.5,2);
\path (4,0) edge (5.5,2);
\path (6.5,0) edge (5.5,2);
\path (3,0) edge (6.5,2);
\path (4,0) edge (6.5,2);
\path (6.5,0) edge (6.5,2);
\path (4,0) edge (2.5,2);
\path (4,0) edge (3.5,2);
\path (1,0) edge (4.5,2);
\path (3,0) edge (4.5,2);
\path (1.5,2) edge (.5,4);
\path (2.5,2) edge (1.5,4);
\path (3.5,2) edge (2.5,4);
\path (4.5,2) edge (3.5,4);
\path (1.5,2) edge (4.5,4);
\path (2.5,2) edge (5.5,4);
\path (3.5,2) edge (6.5,4);
\path (4.5,2) edge (7.5,4);
\path (5.5,2) edge (4.5,4);
\path (5.5,2) edge (5.5,4);
\path (5.5,2) edge (6.5,4);
\path (5.5,2) edge (7.5,4);
\path (6.5,2) edge (4.5,4);
\path (6.5,2) edge (5.5,4);
\path (6.5,2) edge (6.5,4);
\path (6.5,2) edge (7.5,4);
\end{tikzpicture}.$$
\end{lemma}
\begin{proof}
The principal graph (which is the upper half of these graphs) is easy to deduce. We discuss the calculation of the dual graphs. Let $\alpha $ be the nontrivial automorphism determined by Lemma \ref{nti}. Hence, $\bar{\iota} \alpha_2 \iota=\bar{\iota}\iota \alpha$. For any $g,h \in \mathbb{Z}/4\mathbb{Z} $
we have $$(\bar{\iota}\alpha_g \iota,\bar{\iota} \alpha_h \iota )=(\alpha_g \iota \bar{\iota},\iota \bar{\iota} \alpha_h)=(\alpha_g+\alpha_g \rho,\alpha_h + \rho \alpha_h)=\delta_{g,h}+\delta_{g,-h}.$$
This means that $\iota $ and $\alpha_2 \iota $ are each connected to two vertices with no overlap and $\alpha_1 \iota $ and $\alpha_3 \iota $ are each connected to the same single vertex. It is clear that $1$ is connected to $\iota$ and $\alpha$ is connected to $\alpha_2 \iota$. We label the other vertex touching $\iota$ by $\pi_1$ and the other vertex touching $\alpha_2 \iota$ by $\pi_2$ (in the end the dual will be $\mathcal{AH}_2$, so we have named our variables appropriately). Note that, by the defining property of $\alpha$, we have that $\pi_1 \alpha = \pi_2$.
We also compute $$(\iota \pi_1,\kappa)=(\iota(1+\pi_1),\kappa)=(\iota\bar{\iota} \iota,\kappa)=1, $$ so $\pi_1 $ is connected to $\kappa $ by a single edge. In a similar way, we find that $\pi_1$, $\pi_2$, and $ \eta$ are each connected to each of $ \kappa $ and $\alpha_1 \kappa $ by a single edge.
Next we note that we must have $ (\rho\kappa ,\kappa )+(\rho\kappa,\alpha_1 \kappa )=7$, since $(\rho \kappa,\alpha_g \iota)=1 $ for each $g$ and $d(\rho)=4d(\iota) +7d(\kappa)$.
Then $$(\bar{\iota}\kappa,\bar{\iota} \kappa )+(\bar{\iota}\kappa,\bar{\iota} \alpha_1 \kappa )=(\iota\bar{\iota},\kappa \bar{\kappa} )+(\iota \bar{\iota}, \alpha_1 \kappa \bar{\kappa} )=(1+\rho,\kappa \bar{\kappa})+(1+\rho,\alpha_1 \kappa \bar{\kappa})=8.$$
Therefore the total number of paths from $\kappa $ to itself in the dual graph plus the total number of paths from $\kappa $ to $\alpha_1 \kappa $ must equal $8$, and the only two possibilities are pictured.
\end{proof}
We refer to these two cases as the ``four-sigma case'' and ``one-sigma case'', respectively. We would like to rule out the one sigma case.
\begin{remark}
It is not difficult to deduce the fusion rules for $\mathcal{AH}_{2'}$ in the one-sigma case, getting a fusion ring $R$. There does exist a consistent $AH_4-R$ fusion bimodule giving Lemma \ref{2graphs}. Therefore it is impossible to rule out this possibility from the fusion rules for this bimodule and the two fusion rings alone. As usual, we exploit the additional combinatorial structure coming from looking at multiple subfactors simultaneously. In particular, we look at the compatibility between this bimodule and the bimodule coming from the algebra $1+\alpha_1 \rho$.
\end{remark}
\begin{remark}
If one only knows the fusion rules for $\mathcal{AH}_{4'}$ and the fusion module corresponding to $\iota$ but not Lemma \ref{nti}, then there are two additional possibilities for the dual graph beyond the above two.
\end{remark}
Since $[id] \oplus [\alpha_1 \rho] $ has a Q-system, there is another module over $\mathcal{AH}_{4'}$ with a simple object $\lambda$ such that $\lambda \bar{\lambda} \cong id\oplus \alpha_1 \rho$. In other words, there is another Type III factor $N_1$ and a finite-index homomorphism $\lambda: N \rightarrow M $ with dual homomorphism $\bar{\lambda}: M \rightarrow N $ such that $\lambda \bar{\lambda} \cong id\oplus \alpha_1 \rho$.
\begin{lemma} \label{prev}
\begin{enumerate}
\item
The $d+1$-dimensional homomorphisms $ \bar{\iota}\alpha_1 \iota \in \text{End}_0(N) $ and $\bar{\lambda} \iota \in \text{Hom}_0(N,N_1) $ are irreducible.
\item The $4d$-dimensional homomorphism $ \bar{\iota} \kappa \in \text{End}_0(N)$ and $ \bar{\lambda} \kappa \in \text{Hom}_0(N,N_1)$ satisfy either:
(a) $( \bar{\iota} \kappa, \bar{\iota} \kappa)=5$ and $( \bar{\lambda} \kappa, \bar{\lambda} \kappa)=4$; or
(b) $( \bar{\iota} \kappa, \bar{\iota} \kappa)=4$ and $( \bar{\lambda} \kappa, \bar{\lambda} \kappa)=5$.
\end{enumerate}
\end{lemma}
\begin{proof}
By Frobenius reciprocity, $$( \bar{\iota} \alpha_1 \iota, \bar{\iota}\alpha_1 \iota)=
(\alpha_1 \iota \bar{\iota},\iota \bar{\iota} \alpha_1)
=(\alpha_1+\alpha_1\rho,\alpha_1+ \rho \alpha_1)=1$$ and $$(\bar{\lambda} \iota,\bar{\lambda} \iota)=
(\iota \bar{\iota}, \lambda \bar{\lambda})=(1+\rho, 1+ \alpha_1 \rho) =1.$$
Similarly, $$( \bar{\iota} \kappa, \bar{\iota} \kappa)=( \iota \bar{\iota}, \kappa\bar{\kappa})=1+(\rho \kappa, \kappa)$$ and $$( \bar{\lambda} \kappa, \bar{\lambda} \kappa)=
(\lambda \bar{\lambda}, \kappa \bar{\kappa})=1+ (\alpha_1 \rho \kappa, \kappa).$$
\end{proof}
\begin{lemma}\label{killposs}
The one-sigma case of Lemma \ref{2graphs} does not occur.
\end{lemma}
\begin{proof}
Since the objects in $\mathcal{AH}_{2'}$ would have dimensions $(1,1,d-1,d,d,d+1 )$, there is no way to combine them to get an object with dimension $4d$ and with $5$-dimensional endomorphism space. Hence, by Lemma \ref{prev}, we must have that $\bar{\iota} \kappa$ satisfies $(\bar{\iota} \kappa, \bar{\iota} \kappa) = 4$ and thus $(\bar{\lambda} \kappa, \bar{\lambda} \kappa) = 5$. We will examine the bimodule category corresponding to $\bar{\lambda} \iota$ more closely.
First note that $d(\bar{\lambda} \iota) = 5+\sqrt{17}$ which lies in the field $\mathbb{Q}(\sqrt{17})$. Since eigentheory works over any field, it follows that all the other odd vertices must also have dimension lying in $\mathbb{Q}(\sqrt{17})$. We now look more closely at $\bar{\lambda} \kappa$. Frobenius reciprocity shows that $(\bar{\lambda} \kappa, \bar{\iota} \kappa) = 1$ and $(\bar{\lambda} \kappa, \bar{\iota} \alpha \kappa) = 1$, thus $\bar{\lambda} \kappa$ breaks up as $[\bar{\iota} \kappa]$ plus $[\bar{\iota} \alpha \kappa]$ plus three other simple objects. These three objects have total dimension $(3+\sqrt{17})$ so at least one of them has dimension below $\frac{2}{3}(3+\sqrt{17})$. Thus, there's a nontrivial division algebra in $\mathcal{AH}_{2'}$ which has dimension smaller than $\frac{4}{9}(3+\sqrt{17})^2$ and whose dimension is a square in $\mathbb{Q}(\sqrt{17})$. In any division algebra the trivial appears with multiplicity one and invertible objects appear with multiplicity at most one. Thus, from the dimensions of the simple objects in $\mathcal{AH}_{2'}$ the list of possible dimensions of algebra objects in $\mathcal{AH}_{2'}$ below $\frac{4}{9}(3+\sqrt{17})^2$ is:
$$1, 2, d, d+1, d+2, d+3, 2d-1, 2d,2d+1,2d+2,2d+3,2d+4.$$
Clearly $2$ is not a square in $\mathbb{Q}(\sqrt{17})$, and none of the other numbers are squares either since their norms are the nonsquare numbers $-1$, $8$, $19$, $32$, $-19$, $-4$, $13$, $32$, $53$, and $76$, respectively. Thus, the one-sigma case cannot occur.
\end{proof}
Now that we know that we're in the four-sigma case, we want to work out the fusion rules for the $\sigma_i$ to show that $\mathcal{AH}_{2'}$ must have a $Q$-system giving the Asaeda--Haagerup subfactor.
\begin{lemma}
We have $\alpha \pi_1 = \pi_1 \alpha = \pi_2$.
\end{lemma}
\begin{proof}
Counting paths from $\iota$ to itself through the upper and lower graphs, we see that $\pi_1$ and $\pi_2$ are both selfdual. Since $\alpha$ is also self-dual, this means that $\alpha$ commutes with $\pi_1$ and $\pi_2$. Recall that $1+\pi_1 = \bar{\iota} \iota$, so $\alpha + \pi_1 \alpha = \bar{\iota} \iota \alpha = \bar{\iota} \alpha_2 \iota = \alpha + \pi_2$. Thus $\pi_1 \alpha = \pi_2$.
\end{proof}
In keeping with the suggestive notation from $AH_2$, we will let $\pi = \pi_2$ and now denote $\pi_1$ by $\alpha \pi$.
\begin{lemma}
The category $ \mathcal{AH}_{2'}$ has a self-dual simple object $\sigma$ with dimension $d(\sigma)=\displaystyle \frac{d-1}{2} $ such that $\sigma^2=1+\sigma+ \pi $.
\end{lemma}
\begin{proof}
Each of the four $\sigma_i $ has dimension $d(\sigma) =\displaystyle \frac{d-1}{2} = \frac{3+\sqrt{17}}{2}$.
Note that of $\sigma_1 $ and $\sigma_2 $, one object must be self-dual while the other one is not, in order for $\kappa $ to admit the same number of length two paths through the upper graph and through the lower graph. Similarly one of $\sigma_3 $ and $\sigma_4$ is self-dual and the other is not. Without loss of generality, suppose $\sigma_1 $ and $ \sigma_3$ are the self-dual objects while $\sigma_2$ and $\sigma_4$ are dual to each other. We would like to understand how tensoring with $\alpha$ on the left or right acts on the $\sigma_i$. We have
$$\alpha \bar{\iota}\kappa=\overline{\iota \alpha }\kappa=\overline{\alpha_2\iota}\kappa=\bar{\iota}\alpha_2 \kappa=\bar{\iota} \kappa ,$$
so tensoring with $\alpha$ on the left fixes $\sigma_1+\sigma_2 $, and similarly fixes $\sigma_3+\sigma_4$.
From the graph we have $$\alpha \pi+\pi+\eta+\sigma_1+\sigma_2= \bar{\iota}\iota \sigma_1=(1+\alpha\pi)\sigma_1=\sigma_1+\alpha\pi \sigma_1 ,$$
and hence $\alpha \pi \sigma_1=\alpha \pi+\pi+\eta+\sigma_2$.
By Frobenius reciprocity, it follows that $(\pi \sigma_1, \alpha \sigma_2 )=1$ and $(\pi, \alpha \sigma_2 \sigma_1) = 1$.
We claim that $\alpha \sigma_2=\sigma_1 $. Suppose to the contrary that $\alpha \sigma_2=\sigma_2$. Then $\sigma_2 \sigma_1$ contains $\pi $. Thus the remaining part of $\pi$ has dimension $d(\sigma)^2-d(\pi) = \frac{5+\sqrt{17}}{2}$, which means it must be the sum of an invertible object and one of the $\sigma_i$. But this is impossible because $\sigma_2 \sigma_1$ does not contain $1$ since they are not dual to each other, and does not contain $\alpha$ because $\alpha \sigma_2 = \sigma_2$ is not dual to $\sigma_1$. Therefore $\alpha \sigma_2=\sigma_1 $, and similarly $\alpha \sigma_3=\sigma_4 $.
We set $\sigma=\sigma_1 $, so that $\alpha \sigma=\sigma_2 $, $\sigma \alpha=\overline{\sigma_2}=\sigma_4 $, and $\alpha \sigma \alpha =\alpha \sigma_4 = \sigma_3 $. Now return to $\sigma^2$, we now know that $1=(\pi \sigma_1, \alpha \sigma_2) = (\pi \sigma, \sigma)$ so we must have that $\pi$ occurs in $\sigma^2$. Thus, by dimension considerations, we must have that $\sigma^2 = 1+\sigma_i+\pi$ and since $\sigma$ is selfdual while $\alpha \sigma$ and $\sigma \alpha$ are not, we must have $\sigma^2 = 1+\sigma+\pi$ or $\sigma^2 = 1+\alpha\sigma\alpha+\pi$. So our final task is the rule out the latter possibility.
Suppose that $x=\alpha \sigma \alpha $. Then from the relation
$$\sigma^2\sigma=\sigma \sigma^2 $$
we get $$\sigma \pi+ \sigma \alpha \sigma \alpha=\pi \sigma+\alpha \sigma \alpha \sigma .$$
Since $ \pi \sigma= \pi+\alpha\pi+\eta+\sigma$ and $\pi $ and $\sigma$ are both self-dual, we have $\sigma\pi= \overline{\sigma \pi} = \bar{\pi} \bar{\sigma} = \pi \sigma $. Therefore, $\sigma \alpha \sigma \alpha= \alpha \sigma \alpha \sigma$. So consider $\sigma \alpha \sigma$, since its dimension is half odd (that is it is not in $\mathbb{Z}[\sqrt{17}]$), it must have an odd number of $\sigma_i$ in it. But conjugating by $\alpha$ acts faithfully on the $\sigma_i$, so $\sigma \alpha \sigma \alpha$ and $\alpha \sigma \alpha \sigma$ cannot have the same $\sigma_i$ summands. This is a contradiction, so we must have $x = \sigma$ and $\sigma^2=1+\pi+\sigma $.
\end{proof}
This gives an alternative proof for the existence of the Asaeda-Haagerup subfactor \cite{MR1686551} by applying a short skein theoretic argument from \cite{GSbp}.
\begin{theorem}
The object $1 \oplus \sigma$ in $\mathcal{AH}_{2'}$ has an algebra structure giving the Asaeda--Haagerup subfactor.
\end{theorem}
\begin{proof}
Since the fusion ring contains a self-dual simple object $\sigma $ with $$d(\sigma)=\frac{d-1}{2}=\frac{3+\sqrt{17}}{2} $$ satisfying $\sigma^2=1+\sigma+\pi $ with $\pi $
simple, $1+\sigma $ admits a Q-system by \cite[\S3.2]{GSbp}. It is easy to see that the only possible graph pair for such a Q-system is given by the Asaeda-Haagerup principal graphs.
\end{proof}
\begin{corollary}
$\mathcal{AH}_{2'} $ is equivalent to $\mathcal{AH}_2 $.
\end{corollary}
\begin{proof}
This follows from the uniqueness of the connection on the Asaeda-Haagerup principal graph pair.
\end{proof}
\begin{theorem}
The fusion categories $\mathcal{AH}_{4-6} $ exist. The Morita equivalence class contains exactly six fusion categories up to equivalence.
\end{theorem}
\begin{proof}
Since $\mathcal{AH}_{4'} $ is Morita equivalent to $\mathcal{AH}_2 $ via a fusion category with fusion module is $L_{AH_2}^4$, it satisfies the characterizing property of $\mathcal{AH}_4$. Thus $\mathcal{AH}_4$ exists, and so by Theorem \ref{unq} each of $\mathcal{AH}_{4-6}$ exist and are unique. By Theorem \ref{no7}, there are no fusion categories realizing case (7), so by Theorem \ref{cases} the only fusion categories in the Asaeda--Haagerup Morita equivalence class are $\mathcal{AH}_i$ for $1 \leq i \leq 6$.
\end{proof}
\section{Applications} \label{sec:apps}
In this section we outline several applications of our main theorems. These applications will appear in subsequent papers.
\subsection{Drinfeld center}
The Drinfeld center, or quantum double, $Z(\mathcal{C}) $ of a spherical fusion category $\mathcal{C} $ is the braided fusion category whose objects are are pairs of an object in $X \in \mathcal{C}$ endowed with a natural half-braiding $\eta_Y: X \otimes Y \rightarrow Y \otimes X$. In the subfactor setting, the Drinfeld center can be realized as one of the even parts of the asymptotic inclusion or the Longo-Rehren construction. The category $Z(\mathcal{C}) $ is a modular tensor category, in the sense that the $S$-matrix given by the (normalized) traces of the Hopf links coming from the braiding is invertible. The $S$-matrix, along with the diagonal $T$-matrix which gives the twist of each object in the modular tensor category, is called the modular data.
Since the discovery of the ``exotic'' Haagerup and Asaeda-Haagerup in the 1990's, there has been significant interest in computing their quantum doubles. Izumi developed a general method to describe the quantum double of a fusion category from Cuntz algebra models in terms of tube algebras \cite{MR1782145}; he used this to describe the quantum doubles of the Haagerup subfactor and its generalizations \cite{MR1832764}. The modular data of the Haagerup subfactor was simplified and generalized by Evans and Gannon \cite{MR2837122}. By contrast, for the Asaeda--Haagerup subfactor, the quantum double has remained elusive.
Theorem \ref{ahcons} gives a Cuntz algebra model for the fusion category $\mathcal{AH}_4 $, and we can apply the method of \cite{MR1782145} to describe its tube algebra.
In particular, we can compute the simple objects, find their dimensions, and compute the modular data of the Asaeda-Haagerup subfactor. This calculation is complicated, so we will defer the details to a subsequent paper and provide a summary of the results here.
The rank of the quantum double of the Asaeda--Haagerup subfactor is 22. In addition to the trivial object, there are six objects of dimension $1+4(4+\sqrt{17})$, eight objects of dimension $8(4+\sqrt{17})$, one of dimension $1+8(4+\sqrt{17})$, and six of dimension $2+8(4+\sqrt{17})$. Following Evans--Gannon, as in the Haagerup case one can choose a nice ordering of the objects so that the modular data breaks up cleanly into blocks. In this ordering, the first $14$ entries of the diagonal of the $T$-matrix are given by the vector $(1,1,1,-1,1,-1,i,-i,1,1,1,1,1,1) $; the final eight entries are given by $e^{\frac{6l^2 \pi i}{17}}, \ 1 \leq l \leq 8 $. Similarly, the $S$-matrix is given by the following blocks. There is a $14 \times 14$ block given by
$$\frac{1}{8} \left(
\begin{array}{cccccccccccccc}
\frac{8}{\Lambda}& \frac{8d^2}{\Lambda} & 2 &
2 & 2 & 2 & 2 & 2 & 1 & 1 & 1 & 1 & 1 & 1 \\
\frac{8d^2}{\Lambda}& \frac{8}{\Lambda}& 2 &
2 & 2 & 2 & 2 & 2 & 1 & 1 & 1 & 1 & 1 & 1 \\
2 & 2 & 4 & -4 & 0 & 0 & 0 & 0 & 2 & 2 & -2 & -2 & -2 & -2 \\
2 & 2 & -4 & 4 & 0 & 0 & 0 & 0 & 2 & 2 & -2 & -2 & -2 & -2 \\
2 & 2 & 0 & 0 & 4 & -4 & 0 & 0 & -2 & -2 & 2 & 2 & -2 & -2 \\
2 & 2 & 0 & 0 & -4 & 4 & 0 & 0 & -2 & -2 & 2 & 2 & -2 & -2 \\
2 & 2 & 0 & 0 & 0 & 0 & -4 & 4 & -2 & -2 & -2 & -2 & 2 & 2 \\
2 & 2 & 0 & 0 & 0 & 0 & 4 & -4 & -2 & -2 & -2 & -2 & 2 & 2 \\
1 & 1 & 2 & 2 & -2 & -2 & -2 & -2 & 5 & -3 & 1 & 1 & 1 & 1 \\
1 & 1 & 2 & 2 & -2 & -2 & -2 & -2 & -3 & 5 & 1 & 1 & 1 & 1 \\
1 & 1 & -2 & -2 & 2 & 2 & -2 & -2 & 1 & 1 & 5 & -3 & 1 & 1 \\
1 & 1 & -2 & -2 & 2 & 2 & -2 & -2 & 1 & 1 & -3 & 5 & 1 & 1 \\
1 & 1 & -2 & -2 & -2 & -2 & 2 & 2 & 1 & 1 & 1 & 1 & 5 & -3 \\
1 & 1 & -2 & -2 & -2 & -2 & 2 & 2 & 1 & 1 & 1 & 1 & -3 & 5 \\
\end{array}
\right),$$
where $d=4+\sqrt{17}$ and $\Lambda=4(1+d^2) $ is the global dimension.
Then there is an $8 \times 8 $ block given by
$$S_{14+k,14+l}=-\frac{2}{\sqrt{17}} \text{cos}\left(\frac{12 \pi l k }{17}\right) , \quad 1 \leq k,l \leq 8 .$$
These blocks only interact in the first two rows/colums. For $ 15 \leq j \leq 22 $, we have
$$S_{1j}=\frac{1}{\sqrt{17}} $$
$$S_{2j}=-\frac{1}{\sqrt{17}} $$
and
$$S_{ij}=0, \ 1 \leq i \leq 14 .$$
This modular data has a very similar form to that of the Haagerup subfactor, and there is hope that it can be generalized into a series as Evans and Gannon did for the Haagerup modular data.
We note that Morrison and Walker have found a purely combinatorial method to find the dimensions of the simple objects in the quantum double of the Asaeda-Haagerup subfactor as well as the induction functors to $\mathcal{AH}_1 $ and $ \mathcal{AH}_2$ \cite{1404.3955}. However their method does not provide an explicit description of the quantum double or give the modular data.
\subsection{Graded extensions}
A grading on a fusion category $\mathcal{D}$ by a group $G$ is a decomposition $\mathcal{D} \cong \bigoplus_{g \in G}\mathcal{D}_g$ such that the tensor product of an object in $\mathcal{D}_g$ and an object in $\mathcal{D}_h$ lands in $\mathcal{D}_{gh}$. Such a grading is called faithful if the $\mathcal{D}_g$ are all nonzero. A graded extension of a fusion category $\mathcal{C}$ by a group $G$ is a faithfully $G$-graded fusion category $\mathcal{D}$ such that the identity component $\mathcal{D}_0$ is $\mathcal{C}$.
Graded extensions by the group $\mathbb{Z}/2\mathbb{Z}$ are important in subfactor theory, where it is natural to ask whether there is a $\mathbb{Z}/2\mathbb{Z}$ extension of the even part of the subfactor such that $\mathcal{D}_1$ is the odd part of the subfactor. That is, we want to know whether one can find an isomorphism of factors $M$ and $N$ which identifies the two even parts together and the two odd parts together. In planar algebraic language, this is asking about whether there is an unshaded version of the planar algebra attached to a subfactor. For example, for group subfactors, such a $\mathbb{Z}/2\mathbb{Z}$-extension is exactly a Tambara--Yamagami category.
Graded extensions of a fusion category $\mathcal{C}$ by a finite group $G$ were classified in \cite{MR2677836} using homotopy theory. The Brauer--Picard $3$-group (also called a categorical $2$-group) of $\mathcal{C}$, is the full subgroupoid of the Brauer--Picard $3$-groupoid whose only object is $\mathcal{C}$. (In other words, its only object is $\mathcal{C}$, its $1$-morphisms are Morita auto-equivalences, and its $2$ and $3$ morphisms are defined as in the Brauer--Picard groupoid.) Then, giving a $G$ extension of $\mathcal{C}$ is the same thing as giving a map of $3$-groups from $G$ to the Brauer--Picard $3$-group of $\mathcal{C}$. Furthermore, using obstruction theory from algebraic topology, such maps can be classified homologically. By the usual change-of-basepoint conjugation, a Morita equivalence between $\mathcal{C}$ and $\mathcal{D}$ induces an equivalence between their Brauer--Picard $3$-groups. Therefore, to understand the graded extensions of $\mathcal{C} $, it suffices to understand the graded extensions of any fusion category Morita equivalent to $\mathcal{C}$.
In \cite{GJSext}, we gave a partial description of the Brauer--Picard $3$-group of the Asaeda--Haagerup fusion categories. In particular, its homotopy groups are $\pi_1 = \mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$, $\pi_2 = 0$, and $\pi_3 = \mathbb{C}^\times$. For Brauer--Picard $3$-groups of fusion categories, $\pi_1$ automatically acts trivially on $\pi_3$. Thus in order to completely describe the Brauer--Picard $3$-group of Asaeda--Haagerup, we need only understand a single Postnikov $k$-invariant living in $H^4(\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}, \mathbb{C}^\times)$. Furthermore, given a map $G \rightarrow \mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$, whether a $G$-extension compatible with that map exists depends only on whether this $k$-invariant becomes trivial in $H^4(G, \mathbb{C}^\times)$. In particular, since $H^4(\mathbb{Z}/2\mathbb{Z}, \mathcal{C}^\times)$ is trivial, the obstruction automatically vanishes for $\mathbb{Z}/2\mathbb{Z}$ extensions. Therefore, any Morita autoequivalence of one the Asaeda-Haagerup fusion categories gives rise to a $\mathbb{Z}/2\mathbb{Z} $-graded extension. (There are two such extensions, since they form a torsor for $H^3(\mathbb{Z}/2\mathbb{Z}, \mathcal{C}^\times)) \cong \mathbb{Z}/2\mathbb{Z} $). A natural question is whether there exist extensions by the full Brauer-Picard group, $ \mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$. In \cite{GJSext}, we were unable to resolve this question, since unlike in the case of cyclic groups, $H^4( \mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} , \mathbb{C}^{\times})$ is nontrivial. In particular, we need to know the actual value of the $k$-invariant in this nontrivial homology group.
The results of this paper will allow us to answer these question. Instead of trying to understand $\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z}$ extensions of the original Asaeda--Haagerup fusion categories, we can instead consider extensions of the new category $\mathcal{AH}_4 $ which has a computationally accessible model as endomorphisms of the Cuntz algebra $\mathcal{O}_9 $. Furthermore, for $\mathcal{AH}_4 $, every Morita autoequivalence is realized by an outer automorphism by Theorem \ref{autothm}, so the graded components of any extension are all trivial as $\mathcal{AH}_4$-module categories. To realize such an extension, we only need to find a collection of endomorphisms of a factor containing a copy of $\mathcal{AH}_4$ and four additional automorphisms which implement the outer automorphisms of $\mathcal{AH}_4$.
In a subsequent paper we will explicitly construct the outer automorphisms of $\mathcal{AH}_4 $, and construct a (nontrivial) $\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $-graded extension of $\mathcal{AH}_4 $. By the extension theory of \cite{MR2677836}, this implies the existence of nontrivial $\mathbb{Z}/2\mathbb{Z} \times \mathbb{Z}/2\mathbb{Z} $-graded extensions for each $\mathcal{AH}_i $, $1 \leq i \leq 6 $. In particular, we can conclude that the $k$-invariant vanishes, so the Brauer--Picard $3$-group of the Asaeda--Haagerup subfactor splits as a product of Eilenberg--MacLane space $K(\mathbb{Z}/2 \times \mathbb{Z}/2, 1) \times K(\mathbb{C}^\times, 3)$. As a consequence, extensions of by any group $G$ are classified by pairs of a homomorphism from $G$ to the Klein $4$-group together with an element of $H^3(G, \mathbb{C}^\times)$.
\subsection{Intermediate subfactors}
From a complete understanding of the Brauer--Picard groupoid of $\mathcal{C}$, one can read off all division algebras in $\mathcal{C}$ as internal endomorphisms of objects in the module categories over $\mathcal{C}$. In subfactor language, this is the list of all irreducible overfactors of $N$ which can be built from a fixed finite collection of $N$--$N$ bimodules. Each division algebra in $\mathcal{C}$ corresponds to the internal endomorphisms for some simple object $\eta$ in a module category in $\mathcal{C}$, and two objects give the same algebra if they are in the same orbit under the action of invertible objects in the dual category. Thus one can first write down all the module categories (each module category gives a Morita equivalence to its dual category, and this is well-defined up to outer automorphism of the dual category) and then find the orbits of simple objects. For the Asaeda--Haagerup subfactor our results immediately imply that we have the following number of division algebras $A$ in each of the $\mathcal{AH}_i$ such that the dual category of $A$-$A$ bimodules is $\mathcal{AH}_j$. (For $1 \leq i, j \leq 3$ all of these subfactors were described in the appendix to \cite{GSbp}.)
$$
\begin{array}{c|c|c|c|c|c|c}
& \multicolumn{6}{c}{\text{\# of div. algs. in $\mathcal{AH}_i$ with dual $\mathcal{AH}_j$}} \\
& \mathcal{AH}_1 & \mathcal{AH}_2 & \mathcal{AH}_3 & \mathcal{AH}_4 & \mathcal{AH}_5 & \mathcal{AH}_6\\ \hline
\mathcal{AH}_1 & 20 & 13 & 12 & 2 & 2 & 2 \\ \hline
\mathcal{AH}_2 & 18 & 14 & 13 & 2 & 2 & 2 \\ \hline
\mathcal{AH}_3 & 18 & 12 & 14 & 2 & 2 & 2 \\ \hline
\mathcal{AH}_4 & 20 & 8 & 20 & 2 & 2 & 2 \\ \hline
\mathcal{AH}_5 & 16 & 12 & 12 & 2 & 2 & 2 \\ \hline
\mathcal{AH}_6 & 20 & 8 & 12 & 2 & 2 & 2 \\
\end{array}
$$
However, one can go further and use the composition rules in the Brauer--Picard groupoid to understand all inclusions between all of these algebras, and thus all lattices of intermediate subfactors. We did so in \cite{MR2909758} for the Haagerup fusion categories. Here's a quick sketch of where this relationship comes from. Suppose that $1 \subset A \subset B$ is an inclusion of division algebras. Let $\mathcal{D}$ be the fusion category of $A$--$A$ bimodules and $\mathcal{E}$ the fusion category of $B$-$B$ bimodules. Let $\mathcal{L}$ be the $\mathcal{C}$--$\mathcal{D}$ bimodule category of $1$--$A$ bimodules, let $\mathcal{M}$ be the $\mathcal{D}$--$\mathcal{E}$ bimodule category of $A$--$B$ bimodules, and let $\mathcal{N}$ be the $\mathcal{C}$--$\mathcal{E}$ bimodule category of $1$-$B$ bimodules. Then $B = A \otimes_A B$ shows that under the tensor product map $\mathcal{L} \boxtimes_\mathcal{D} \mathcal{M} \rightarrow \mathcal{N}$ the tensor product of two simple objects is simple. Thus intermediate subfactors yield a factorization of an object in some bimodule category as a tensor product of two other objects in bimodule categories. In the other direction, if $\iota = \eta \psi$ is a decomposition of a simple object $\iota$ in a bimodule category as a tensor product of objects in (possibly different) bimodule categories, then we get an inclusion of division rings $1 \subset \eta \bar{\eta} = \eta 1 \bar{\eta} \subset \eta \psi \bar{\psi} \bar{\eta} = \iota \bar{\iota}$. Thus there is a correspondence between inclusions of algebras and factorizations of objects in bimodule categories. In full generality, this correspondence can get a bit delicate as automorphsims of $\mathcal{D}$ (both inner and outer!) play a role, as does the group $\pi_2$ which measures the ambiguity of the tensor product.
In a later paper, we hope to give a full description of this recipe for reading off lattices of intermediate subfactors from tensor product rules between bimodule categories, and to apply this to give a classification of all intermediate subfactor lattices for all division algebras in the Asaeda--Haagerup fusion categories. It is already known that there are several interesting quadrilaterals involve $AH$, $AH+1$ and $AH+2$, and we hope to find more. Such a calculation will require computing some additional compositions which we have not yet worked out.
|
\section{Introduction}
Transition metal dichalcogenides (TMDs) have been intensively studied over past decades, and
recently attracted tremendous attention for a number of applications\cite{Radisavljevic11,Wang12,
Lopez-Sanchez13,Lin14,Pradhan14}, since TMDs have a finite energy band gap, which is more
advantageous than graphene. Among the TMDs, MoTe$_2$ has a stronger spin-orbit coupling than other
Mo related TMD materials (MoX$_2$, X = S, Se), which leads to a longer decoherence time both for
valley and spin freedom \cite{Pradhan14} and thus MoTe$_2$ is considered to be used for ideal
valleytronic devices\cite{Xiao12,Cao12}. Unlike other TMDs, MoTe$_2$ has unique structural properties,
such as a thermally-induced structural phase transition at around 900$^\circ$C from the $\alpha$-phase
diamagnetic semiconductor to the $\beta$-phase paramagnetic metal\cite{Vellinga70, Duerloo14}.
MoTe$_2$ has the smallest direct energy band gap ($\sim$ 1.1 eV) at the K point (hexagonal
corner) in the Brillouin zone (BZ) among all semiconducting TMDs, the band gap is similar to the
indirect gap ($\sim$ 1.1 eV) of silicon\cite{Streetman00,Ruppert14}. MoTe$_2$ is thus suitable for
nanoelectronics devices\cite{Lin14,Pradhan14} and for studying exciton effects at the K points\cite{Wilson69}.
Raman spectroscopy is a useful tool for characterizing TMDs since we get many information
\cite{Lee10,Sahin13,Li12} such as number of layers, electronic and phonon properties of TMDs.
Recent Raman spectroscopy studies of monolayer and few-layer MoTe$_2$ using visible light (1.9
$\sim$ 2.5 eV) show that there exists some unassigned Raman spectra which can not be assigned
to a first-order Raman process\cite{Pradhan14,yamamoto14}, and the assignment of these Raman
spectra is the subject of the present paper.
Raman spectra of TMDs have been reported for several decades\cite{Wieting71,Chen74,Wakabayashi75,
Mead77,Beal79,Sekine80a,Sekine80b,Sugai81,Sekine84,Stacy85,Sourisseau89,Sourisseau1991,Frey99}.
The first-order Raman process in TMDs has been widely investigated and is now well understood
\cite{Sugai81,Sugai82,Sekine80a,Beal79,Sanchez11,Ding11,Luo13a,Luo13b,yamamoto14}. However, the
study of second-order Raman spectra did not give consistent assignments for different TMDs\cite{Stacy85,
Chen74,Sekine84,Chakraborty12,Zhao13,Terrones14}. For example, Stacy \textit{et al.}\cite{Stacy85}
and many other groups\cite{Chen74,Feldman96,Frey99,Windom11,Li12,Chakraborty13} assigned some of
the Raman peaks of 2H-MoS$_2$ to either combination or overtone second-order phonon modes occurring
at the M point (center of hexagonal edge) in the Brillouin zone. Terrones \textit{et al.} \cite{Terrones14}
also assigned phonon modes in few-layer WSe$_2$ according to the M point phonon. On the other hand,
Sourisseau \textit{et al.} \cite{Sourisseau89,Sourisseau1991} ascribed the second-order Raman
process of 2H-WS$_2$ to the K point phonons. Berkdemir \textit{et al.}\cite{Berkdemir13} considered
that electrons are scattered by the M point phonons between the K point and I point (one point along
$\Gamma$K direction), and that this process is responsible for the double resonant Raman process in
few-layer WS$_2$. Furthermore, Sekine \textit{et al.}\cite{Sekine84} found a dispersive Raman peak
in 2H-MoS$_2$ and interpreted this peak in terms of a two-phonon Raman process at the K point. The
different assignments for both few-layer or bulk MoS$_2$ and WS$_2$, either at the M or K points, have
not been well investigated, since most of these assignments to second-order Raman peaks have been made
merely by comparing phonon frequencies with reference to some inelastic neutron scattering results\cite{Wakabayashi75,Sourisseau1991}.
Thus a detailed analysis with use of double resonance Raman theory is needed, especially for understanding
both mono- and few-layer MoTe$_2$ that are discussed in this paper.
In the present paper, we mainly study the second-order Raman spectra of monolayer MoTe$_2$,
based on {\em ab initio} electronic band calculations, density functional perturbation theory and
associated experimental Raman spectroscopy. From our electronic band structure calculation, we have
found a saddle point of the energy dispersion at the M points in the BZ, which gives rise
to two-dimensional Van Hove singularity (VHS) in the electronic density of states. The energy separation
between two VHSs (2.30 and 2.07 eV) are matched to the 2.33 eV (532 nm) and 1.96 eV (633 nm) laser
excitations in our experiment in which resonant optical absorption occurs at the M points. Consequently, inter-valley
electron-phonon scattering process occurs between two of the three inequivalent M (M, M$'$, M$''$) points,
giving rise to a double resonance Raman process. We propose that this double resonance process occurring at the
M points is essential for the interpretation of the observed second-order Raman peaks of MoTe$_2$. The
experimental second-order Raman frequencies as a function of laser excitation energy confirm our
interpretation of the double resonance Raman spectra.
Organization of the paper is as follows. In II and III, we briefly show computational details and
experimental methods, respectively. In IV, both experimental and theoretical results for mono- and few-layer
MoTe$_2$ Raman spectra are given. We assign second-order Raman peaks by comparing the calculated results with the
observed Raman spectra. In V summary of the paper is given. Further information such as layer-number dependence
is discussed in Supplemental Material.
\section{Computational details}
We performed the electronic and phonon energy dispersion calculations on monolayer and few-layer
MoTe$_2$ by using first-principles density functional theory within the local density approximation
(LDA) as implemented in the Quantum-Espresso code \cite{Giannozzi09}. The structures of monolayer
and few-layer MoTe$_2$ that are used in the present experiment are $\alpha$-MoTe$_2$\cite{yamamoto14}.
The AA$'$ and AA$'$A stacking geometries are used in the computational study for bi-layer and tri-layer
MoTe$_2$, respectively, since these stacking geometries are more stable than the other possibilities\cite{Tao14}. AA$'$
stacking refers to the geometry which has a Mo (Te) atom of one layer on top of a Te (Mo) atom of
the other layer. The few-layer MoTe$_2$ are separated by 25 \AA\ from one another in a unit cell of
the calculation to eliminate the inter-few-layer interaction. We used projector augmented-wave (PAW)
pseudopotentials\cite{blochl94} with a plane-wave cutoff energy of 65 Ry to describe the interaction
between electrons and ions. The spin-orbit split electronic band structures were calculated by
relativistic pseudopotentials derived from an atomic Dirac-like equation. The atomic coordinates are
relaxed until the atomic force is less than 10$^{-5}$ Ry/Bohr. The lattice parameters used in this paper
are listed in Table S1 and agree well with the experiment and other calculations \cite{Ding11,Puotinen61,
Kang13}. The Monkhorst-Pack scheme \cite{Monkhorst76} is used to sample the Brillouin zone (BZ) over a
17$\times$17$\times$1 $\textit{k}$-mesh for MoTe$_2$. The phonon energy dispersion relations of MoTe$_2$
are calculated based on density functional perturbation theory \cite{Baroni01}. In order to visualize
the vibrational modes at the M point, we use a 2$\times$2$\times$1 supercell in which the M point is
zone-folded to the $\Gamma$ point in the folded BZ.
The optical absorption spectrum is calculated by the real ($\epsilon$$'$) and imaginary ($\epsilon$$''$)
parts of the dielectric function as a function of photon energy, respectively, based on the PAW
methodology\cite{Gajdos06} and the conventional Kramers-Kronig transformation. The absorption coefficient
$\alpha$ is described by $\alpha$ = 4$\pi\kappa$E$_\gamma$/($\textit{hc}$), where E$_\gamma$ is the
incident laser excitation energy, $\textit{h}$ the Planck constant, $\textit{c}$ the speed of light, and $\kappa$ is
the extinction coefficient\cite{Zimmermann00}, that is, $\kappa$ = $\sqrt{(\sqrt{\epsilon'^2 +
\epsilon''^2} - \epsilon')/2}$.
\begin{figure*}[t]
\includegraphics[width=1.60\columnwidth]{Fig1}
\caption{(Color online) (a) Raman spectra of mono-, bi-, and tri-layer MoTe$_2$ under 2.33 eV
(532 nm) laser excitation energy, showing the small intensity peaks $\omega_i$ (i = 1, 2, ..., 6)
and an additional calibration peak from Si besides the typical Raman peaks of A$_{1g}$, E$_{2g}^1$
and B$_{2g}^1$ that are frequently observed in MoTe$_2$\cite{Pradhan14,yamamoto14}. (b) Raman spectra of monolayer MoTe$_2$
under 2.33 eV (532 nm), 1.96 eV (633 nm), and 1.58 eV (785 nm) laser excitation energies.}
\label{Fig1}
\end{figure*}
\section{Experimental method}
Bulk crystals of $\alpha$-MoTe$_2$ were prepared through a chemical vapor transport method\cite{Lieth77}.
Atomically thin crystals of MoTe$_2$ were mechanically exfoliated from the bulk crystals onto silicon
substrates with a 90 nm-thick oxide layer on top. The thicknesses of the atomic crystals were identified using
optical microscopy and first-order Raman spectroscopy \cite{yamamoto14}. Raman spectroscopy measurements
for monolayer MoTe$_2$ were performed using 532, 633 and 785 nm excitation lasers for discussing the observed phonon
dispersion due to double resonance Raman theory\cite{Saito01}. Raman spectroscopy studies of few-layer MoTe$_2$
under 532 nm laser excitation were also performed and the results are discussed in the supporting materials.
The grating sizes were 1200, 1800 and 2400 lines/mm for the 785, 633 and 532 nm laser excitation measurements,
respectively. The magnification of the objective lens was 100x. The accumulation times were 150-200 seconds.
The laser power was kept below 0.1 mW, 0.2 mW and 3.0 mW for the 532, 633 and 785 nm excitation lasers,
respectively, in order to avoid damage to the samples. All measurements were performed at room temperature in
the backscattering configuration. Since this paper gives greater emphasis to a theoretical understanding
of the 2nd-order Raman scattering processes, we do not present many experimental details. However, more
details about the characterization of the number of layers in our MoTe$_2$ samples can be obtained in one of
the previous works on MoTe$_2$ synthesis and characterization, based on the first-order Raman spectra\cite{yamamoto14}.
\section{Results and discussion}
Figure~\ref{Fig1}(a) shows Raman spectra of mono- to tri-layer MoTe$_2$ under the 532 nm (2.33 eV) laser excitation.
The Raman spectra show three strong peaks, including the in-plane E$^1_{2g}$ mode and the out-of-plane A$_{1g}$ mode
for monolayer and few-layer MoTe$_2$, as well as the E$_{1g}$ mode and the B$^1_{2g}$ mode for bi- and tri-layer MoTe$_2$,
that were observed previously\cite{Pradhan14,yamamoto14,note1}. These spectra are assigned to the first-order Raman spectra
of the $\Gamma$ point phonons\cite{yamamoto14}. However, we observe additional peaks with relatively small
intensities, which can not be assigned as first-order Raman spectra\cite{yamamoto14}. In Fig.~\ref{Fig1}(a),
we denote the weak Raman spectra as $\omega_i$ (i = 1, 2, ..., 6).
Since the second-order Raman spectra are dispersive as a function of excitation energy\cite{Saito01}, we performed
Raman spectroscopy measurements using different excitation laser energies. In Fig.~\ref{Fig1}(b) we show Raman spectra
of monolayer MoTe$_2$ under 2.33 eV (532 nm), 1.96 eV (633 nm) and 1.58 eV (785 nm) laser excitation energies. Here
the Raman intensities are normalized using the non-resonant Si peak intensity at 520 cm$^{-1}$ (not shown in Fig.~\ref{Fig1}(b)).
We find that the peak positions of each $\omega_i$ either upshift or downshift by up to several cm$^{-1}$ for different
laser excitation energies, suggesting that these peaks are associated with a second-order Raman process\cite{Saito01}.
\begin{figure*}[t]
\includegraphics[width=1.9\columnwidth]{Fig2}
\caption{(Color online) (a) Electronic energy band structure E(k) and the density of states (DOS), (b) the real ($\epsilon$$'$),
and imaginary ($\epsilon$$''$) parts of the dielectric function and the optical absorption coefficient $\alpha$ as a function
of photon energy E$_\gamma$, and (c) an equi-energy contour plot of the two conduction bands (c1 and c2) with the
same spin for monolayer MoTe$_2$. Energy bands are split by spin-orbit interaction, the conduction bands c1, c2 and the valence
bands which have the same spin\cite{Xiao12,Wang14} are highlighted in red solid lines in (a). When we consider energy separation
between two Van-Hove singularities (VHSs) of the DOS, the peak at 1.80 eV in the absorption spectrum corresponds to optical
transition at the M point (black double headed arrow in (a)). (c) The equi-energy contour plot for VHS E$_{band}^{c1}$ = 1.52 and
1.57 eV (left) and E$_{band}^{c2}$ = 1.72 eV (right). All equi-energy contours appear around the M point. The E$_{band}$ energies
of 1.52, 1.57 and 1.72 eV (horizontal lines in (a)) correspond to the experimental laser energies (1.58, 2.33 and 1.96 eV) in blue,
pink and orange colors.}
\label{Fig2}
\end{figure*}
Considering that the second-order process is usually connected to the electronic properties by double resonance
Raman theory\cite{Saito01}, we calculate the electronic band structure, the density of states (DOS) and the
optical absorption spectra of monolayer MoTe$_2$ as shown in Fig.~\ref{Fig2}. Our calculated energy dispersion
reproduces the main observed features\cite{Kumar12,Ding11,Ruppert14} such as the direct electronic band gap
(1.00 eV) and the spin-orbit splitting of the valence band ($\sim$230 meV) at the K point which are close to
1.10 eV and 250 meV, respectively, obtained from the optical absorption measurements\cite{Ruppert14}. It is noted that
the energy band gap is underestimated by the local density functional method. In order to fit the experimental
data, the conduction bands are upshifted by 0.10 eV, so are the DOS and optical absorption spectrum. In
Fig.~\ref{Fig2}(a), we show the calculated electronic energy band and DOS. From the parabolic bands around the
K point, two-dimensional (2D) DOS gives constant values, while at the M point we have found saddle points
of the energy dispersion in the electronic band, the corresponding DOS gives logarithmic $\log$$|${$E$ - $E_0$}$|$
divergence as is known as Van Hove Singularity (VHS) in 2D systems.
In Fig.~\ref{Fig2}(b), we show the calculated real ($\epsilon'$) and imaginary ($\epsilon''$) parts of the
dielectric constant and the optical absorption coefficient $\alpha$ as a function of photon energy. $\epsilon''$
has a nonzero value from the photon energy E$_\gamma$ = 1.10 eV which corresponds to the energy gap, and
a peak at E$_\gamma$ $\sim$ 1.80 eV. A strong optical absorption at 1.80 eV from our calculation agrees well
with the large peak at around 1.83 eV from the optical absorption measurement by Ruppert et al.\cite{Ruppert14}
on monolayer MoTe$_2$. Moreover, the electronic excitation energy of 2.30 eV (2.07 eV) from the third (top)
valence band to the lowest (third lowest) conduction band at the M point matches very well with the laser energy
2.33 eV (1.96 eV) used in our experiment, as indicated by pink (orange) arrows in the band structure of Fig.~\ref{Fig2}(a).
Thus resonant optical absorption at these laser lines is expected.
In Fig.~\ref{Fig2}(c), we plot equi-energy contour for the conduction band energies E$_{band}$ which correspond to the
laser excitation energies E$_{\gamma}$. E$_{band}$ = 1.52 and 1.57 eV for c1 and 1.72 eV for c2 correspond to the E$_{\gamma}$ =
1.58, 2.33 and 1.96 eV, respectively, as indicated by blue, pink, and orange arrows in Fig.~\ref{Fig2}(a). The conduction
bands c1, c2 and the valence bands which have the same spin\cite{Xiao12,Wang14} are highlighted in red solid lines in the
band structure. The equi-energy contour lines E$^{c1}_{band}$ and E$^{c2}_{band}$ in Fig.~\ref{Fig2}(c) both show the saddle
point shape of the energy dispersion in the vicinity of the M point.
That is, the energy band E$^{c1}_{band}$ (E$^{c2}_{band}$) starting from the M point monotonically decreases (increases) along
the M-K direction, whereas this band energy increases (decreases) along the M-$\Gamma$ direction. When we decrease laser excitation
energies from 2.33 to 1.58 eV, the contour lines in Fig.~\ref{Fig2}(c) left change the shapes around the M point, suggesting
that the area inside of the contour line for each M point decreases with increasing laser excitation energy. This situation is opposite
to the case of graphene in which the equi-energy contour is almost circular and the area increases with increasing energy.
\begin{figure}[t]
\includegraphics[width=0.80\columnwidth]{Fig3}
\caption{(Color online) (a) In the double-resonance process, photon-excited electrons are scattered inelastically by emitting phonons and make inter-valley (i.e., M$\longleftrightarrow$M$'$, M$'$$\longleftrightarrow$M$''$, and M$\longleftrightarrow$M$''$) transitions, as shown by the solid double-end arrows in the first BZ. The related phonon vectors $\vec q$ measured from the $\Gamma$ point are also given by the six dashed arrows. (b) The schematics of the double resonance process (electron-photon and electron-phonon processes). This process generates two phonons ($\hbar \omega_{p1}$, -$\vec q$) and ($\hbar \omega_{p2}$, $\vec q$).}
\label{Fig3}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=1.6\columnwidth]{Fig4}
\caption{(Color online) (a) All possible phonon vectors q$_{\textrm{\textrm{MM}}'}$ coming from the double resonance Raman process due to the laser energies of 2.33, 1.96 and 1.58 eV. The $\vec q_{\textrm{MM}'}$ is equal to either $\overline{\Gamma\textrm{M}}$ (for the laser energy E$_{\gamma}$ = 2.33 and 1.96 eV) or $\sim\frac{1}{3}\overline{\textrm{MK}}$ (for E$_{\gamma}$ = 1.96 eV) or $\sim$10\%$\overline{\textrm{MM}'}$ (for E$_{\gamma}$ = 1.58 eV), as indicated by an arrow away from the $\Gamma$ point. (b) The phonon dispersion relations of a monolayer MoTe$_2$. The second-order phonon modes are analyzed at the M$'$ point for the laser energy E$_{\gamma}$ = 2.33 eV, and then at two k points marked by two vertical dashed lines which are shifted leftward and rightward from the M point for the laser energy E$_{\gamma}$ = 1.58 eV and 1.96 eV, respectively. The red, green, blue, purple, magenta, and cyan vertical arrows are used for modes $\omega_1$, $\omega_2$, $\omega_3$, $\omega_4$, $\omega_5$, and $\omega_6$, respectively. The superscripts "+", "-", and "2" for $\omega$ refer to possible "combination modes" (i.e. $\omega^+_{4}$ = A$_{1g}$(M) + LA(M), $\omega^+_{6}$ = E$^1_{2g}$(M) + LA(M)), "difference modes" (i.e. $\omega^-_{1}$ = E$^1_{2g}$(M) $-$ LA(M), $\omega^-_{2}$ = E$^1_{2g}$(M) $-$ TA(M)) and "overtones" (i.e. $\omega^2_{1}$ = 2TA(M), $\omega^2_{3}$ = 2LA(M)), respectively.}
\label{Fig4}
\end{figure*}
If the laser excitation energy matches the optical transition energy at the M point, we expect that many photo-excited electrons at the M point are scattered to the inequivalent M$'$ or M$''$ points by an inter-valley resonant electron-phonon interaction, as shown in Fig.~\ref{Fig3}(a). The inter-valley electron-phonon process, together with the electron-photon process, gives rise to a double resonance Raman process. Selecting one process from three possible inter-valley scattering of photo-excited carriers (M$\longleftrightarrow$M$'$ or M$\longleftrightarrow$M$''$ or M$'$$\longleftrightarrow$M$''$), we show the schematic view of the double resonance process in Fig.~\ref{Fig3}(b). In Fig.~\ref{Fig3}(b), a photo-excited electron will be scattered from M to M$'$ by emitting a phonon with energy $\hbar \omega_{p1}$ and momentum $-\vec q$, which conserves the energy and momentum of both the electron and the phonon. Then the electron is scattered back to M by emitting the second phonon with energy $\hbar \omega_{p2}$ and momentum $+\vec q$. As analyzed in the Fig.~\ref{Fig3}(a), the momentum $\vec q$ of the phonon can only have a magnitude of $|\overrightarrow{\textrm{MM}'}|$ or $|\overrightarrow{\textrm{MM}''}|$ or $|\overrightarrow{\textrm{M}'\textrm{M}''}|$. When we measure the value of $|$q$|$ from the $\Gamma$ point, the $\vec q$ corresponds to $\overrightarrow{\Gamma \textrm{M}''}$ or $\overrightarrow{\Gamma \textrm{KM}'}$ or $\overrightarrow{\Gamma \textrm{M}}$, respectively. Since $\overrightarrow{\Gamma \textrm{KM}'}$ = $\overrightarrow{\Gamma \textrm{M}'}$ + $\overrightarrow{\Gamma \Gamma}$ $\equiv$ $\overrightarrow{\Gamma \textrm{M}'}$, all phonon modes related to double resonance process correspond to phonon modes at the M point.
\begin{table*}[t]
\caption{Phonon frequencies of the first- and second-order Raman spectra of monolayer MoTe$_2$ are listed for three different laser energies E$_\gamma$. A comparison for first and second-order Raman spectra is given between the experimentally unassigned modes $\omega_i$ (i = 1, 2, ..., 6) and the calculated phonon frequencies both at the M point and near the M point denoted by (off M) whose $\vec q$ lies around $\frac{1}{3}\overline{\textrm{MK}}$ and 10\%$\overline{\textrm{MM}'}$ for E$\gamma$ = 1.96 eV and 1.58 eV, respectively. The consistent dispersion behavior for the second-order Raman spectra between the experiment and the theory are underlined. The frequency is given in units of cm$^{-1}$. NA: not available.}
\begin{tabular}{p{1.0cm}|p{1.60cm}p{1.60cm}p{1.60cm}|p{1.60cm}p{2.10cm}p{1.60cm}|p{4.10cm}}
\hline
\hline
{}&{}&{}&{}&{}&{} &{} &{}\\
{}&\multicolumn{3}{c|}{Experiment}&\multicolumn{3}{c|}{Calculation}&{Phonon assignments}\\
{}&{}&{}&{}&{} &{} &{}&{}\\
\hline
{E$_\gamma$}&{2.33 eV}&{1.96 eV}&{1.58 eV}&{2.33 eV}&{1.96 eV}&{1.58 eV}&{}\\
\hline
{E$_{1g}$} & {NA }&{NA} &{NA} &{116.9}&{ }&{ } &{first-order}\\
{A$_{1g}$}&{171.1}&{171.3}&{170.0} &{172.2}&{ }&{ } &{first-order}\\
{E$^1_{2g}$} &{235.8} &{236.2}&{234.1}&{236.9}&{ }&{ } &{first-order}\\
\hline
{}&{}&{}&{}&{(at M)}&{(at M/off M)}&{(off M)}&{}\\
\hline
{$\omega_1$}&{138.2}&{140.7}&{139.0}& {\underline{137.4}}&{{137.4}/\underline{146.6}}&{\underline{140.1}}&{\underline{2TA(M)}} \\
{ }&{ }&{ }&{ }&{\underline{132.8}}&{{132.8}/\underline{133.1}}&{\underline{133.2}}&{or \underline{E$^1_{2g}$(M)-LA(M)}}\\
{$\omega_2$}&{183.7}&{183.6}&{181.5}&{\underline{178.3}} &{\underline{178.3}/{177.2}} &{\underline{177.8}}& {\underline{E$^1_{2g}$(M)-TA(M)}}\\
{$\omega_3$}&{205.5}&{205.3}&{205.9}&{\underline{197.9}} &{\underline{197.9}/{208.2}} &{\underline{200.5}}&{\underline{E$_{1g}$(M)+TA(M)}}\\
{ }&{ }&{ }&{ }&{\underline{202.7}} &{\underline{202.7}/196.0} &{\underline{200.4}}&{or {\underline{2LA(M)}}}\\
{$\omega_4$}&{$\sim$264.0}&{260-270}&{260-270}&{\underline{263.0}} &{263.0/\underline{256.6}} &{\underline{261.0}}&{\underline{A$_{1g}$(M) + LA(M)}}\\
{$\omega_5$}&{{329.0}}&{{333.2}}&{$\sim$328.0}&{\underline{323.3}} &{323.3/317.2} &{\underline{321.6}}&{2A$_{1g}$(M)}\\
{}&{}&{}&{}&{\underline{315.7}}&{315.7/\underline{323.8}} &{\underline{317.9}}&{or \underline{E$^1_{2g}$(M)+TA(M)}}\\
{$\omega_6$}&{347.4}&{348.4}&{346.5}&{\underline{348.3}} &{\underline{348.3}/348.5} &{\underline{348.0}}&{\underline{E$^1_{2g}$(M)+LA(M)}}\\
\hline
\hline
\end{tabular}
\label{Table1}
\end{table*}
In order to specify the phonons contributing to the second-order Raman process, we analyzed possible inter-valley phonon vectors $\vec q_{\textrm{MM}'}$ for the three laser energies and calculated the phonon spectra of monolayer MoTe$_2$, as shown in Fig.~\ref{Fig4}. For the laser with an energy of 2.33 eV, we analyze possible $\vec q_{\textrm{MM}'}$ wave vectors between two elliptical equi-energy contour lines (red lines) in the left panel of Fig.~\ref{Fig4}(a). The dark area for $\vec q_{\textrm{MM}'}$ shows that the most probable phonon vectors appear at the M point. Other phonon vectors, such as $\vec q_{\textrm{M}'\textrm{M}''}$ and $\vec q_{\textrm{MM}''}$, are similar to $\vec q_{\textrm{MM}'}$ (not shown). For the 1.96 and 1.58 eV excitation lasers, as already seen in Fig.~\ref{Fig2}(c), the equi-energy contours have a larger area than that for 2.33 eV. Then we expect that $\vec q_{\textrm{MM}'}$ for E$\gamma$ = 1.96 and 1.58 eV arises from the vicinity of the Brillouin zone near the M point, as shown in the central and right panels of Fig.~\ref{Fig4}(a). The relatively dark region indicates that: (1) the $\vec q_{\textrm{MM}'}$ for E$\gamma$ = 1.96 eV appear both at the M point and in the vicinity of the M point (approximately with lengths of $\frac{1}{3}\overline{\textrm{MK}}$) along the MK direction and (2) the $\vec q_{\textrm{MM}'}$ for E$\gamma$ = 1.58 eV appear in the vicinity of the M point along the M$'$M$''$ (or \textrm{MM}$''$) direction approximately with lengths of 10\%$\overline{\textrm{M}'\textrm{M}''}$ (or 6.5\%$\overline{\textrm{MM}''}$) away from the M$''$ point.
In Fig.~\ref{Fig4}(b), we show the calculated phonon dispersion relations of monolayer MoTe$_2$. The symmetry for each optical phonon band\cite{note1}, such as E$_{2g}^1$ and E$_{1g}$, is labeled in Fig.~\ref{Fig4}(b). Here LA, TA and ZA refer to the longitudinal, transverse and out-of-plane acoustic modes, respectively. We assign the second-order Raman modes at the M point for the laser energy of 2.33 eV. The strategy for the assignment is to choose any two phonon modes at the M point as one overtone or (difference) combination mode and to list in Table~\ref{Table1} those that have frequency values within the experimental $\omega_i \pm$ 6 cm$^{-1}$. These assignments are marked in Fig.~\ref{Fig4}(b) by using color arrows. The red, green, blue, purple, magenta, and cyan arrows represent overtone or combination Raman modes of $\omega_i$ (i = 1, 2, ..., 6), respectively. Downward arrows with an a "$-$" superscript represent a two-phonon process with one phonon annihilated and one phonon created to generate "difference combination mode", while upward arrows with the "+" or "2" superscripts are used for combination or overtone modes in which two phonons are created. For example, $\omega_1$ is assigned to either the $\omega^-_1$ (= E$^1_{2g}$(M) - LA(M)) or $\omega^2_1$ (= 2TA(M)), $\omega_2$ to $\omega^-_2$ (= E$^1_{2g}$(M) - TA(M)), $\omega_3$ to either $\omega^2_3$ (= 2LA(M)) or $\omega^+_3$ (= E$_{1g}$(M) + TA(M)), $\omega_4$ to $\omega^+_4$ (= A$_{1g}$(M) +LA(M)), $\omega_5$ to either the $\omega^2_5$ (= 2A$_{1g}$(M)) or $\omega^+_5$ (= E$^1_{2g}$(M) + TA(M)) and $\omega_6$ to the $\omega^+_6$ (= E$^1_{2g}$(M) + LA(M)) mode. Using the same phonon assignment, we obtain the second-order Raman frequencies for E$\gamma$ = 1.96 eV both at the M point and near the M point with a length of $\frac{1}{3}\overline{\textrm{MK}}$ and for 1.58 eV at one point away from the M point with 10\%$\overline{\textrm{M}'\textrm{M}''}$.
\begin{figure}[t]
\includegraphics[width=.70\columnwidth]{Fig5}
\caption{(Color online) Visualization of the possible Raman-active modes at the M point for the second-order process suggested in Figure 4. Both the top and side views are given and red arrows represent the vibrational directions and amplitudes of the motion of individual atoms.}
\label{Fig5}
\end{figure}
In Table~\ref{Table1}, we list phonon frequencies of the first-order (upper) and the second-order (lower) Raman spectra for both the experiment and the calculation and also the phonon assignments. The first-order experimental Raman spectra (A$_{1g}$ and E$^1_{2g}$) do not change the frequency for the laser energies (2.33 and 1.96 eV) within the experimental error (1 cm$^{-1}$) and agree with the calculations. The small frequency downshift of the first order spectra (1.7 cm$^{-1}$ at most) from 2.33 eV to 1.58 eV laser excitation is checked to be not from the laser-induced thermal effect as discussed in the Supplemental Material. Although we did not specify the origin of the downshift precisely, the origin of the change for the first order can be neglected for simplicity. The experimental second-order peak positions of $\omega_i$ are either upshifted or downshifted with the laser energy E$_\gamma$ by a small magnitude of about 1-3 cm$^{-1}$. It is pointed out that the frequency of $\omega_1$ is upshifted from 2.33 eV to 1.58 eV even though the both the first-order spectra are downshifted. This can be understood by double resonance Raman theory in which the resonance phonon frequency changes along the phonon dispersion near the M point. In the right of Table~\ref{Table1}, the second-order phonon frequencies for 2.33 and 1.58 eV laser excitations are obtained at $\vec q_{\textrm{MM}'}$ = $\overrightarrow{\Gamma\textrm{M}}$ (at M) and $\sim$10\%$\overline{\textrm{MM}'}$ (denoted by "off M"), respectively. For E$_\gamma$ = 1.96 eV, the phonon assignments are obtained either at $\vec q_{\textrm{MM}'}$ = $\overrightarrow{\Gamma\textrm{M}}$ (at M) or $\sim\frac{1}{3}\overline{\textrm{MK}}$ (off M).
\iffalse
In Table~\ref{Table1}, we list phonon frequencies of the first-order (upper) and the second-order (lower) Raman spectra for both the experiment and the calculation and also the phonon assignments. The first-order experimental Raman spectra (A$_{1g}$ and E$^1_{2g}$) have changed little with the laser energy (2.33 and 1.96 eV) and agree with the calculations. The relatively large frequency change from 2.33 (or 1.96 eV) to 1.58 eV laser excitation is excluded from the thermal effect as discussed in the Supplemental Material. The experimental second-order peak positions of $\omega_i$, except for $\omega_4$ (at the three laser excitations) and $\omega_5$ peak (at 1.58 eV) which are too weak to have the frequency value and will be discussed later, are either upshifted or downshifted with the laser energy E$\gamma$ by a small magnitude of about 1-3 cm$^{-1}$. This can be understood from a small dispersion of the phonon dispersion relation near the M point where all the phonon assignments are made. The second-order phonon assignments at 2.33 and 1.58 eV laser excitations are obtained at $\vec q_{\textrm{MM}'}$ = $\overline{\Gamma\textrm{M}}$ (at M) and $\sim$10\%$\overline{\textrm{MM}'}$ (off M), respectively. While at E$\gamma$ = 1.96 eV, the phonon assignments are obtained either at $\vec q_{\textrm{MM}'}$ = $\overline{\Gamma\textrm{M}}$ (at M) or $\sim\frac{1}{3}\overline{\textrm{MK}}$ (off M).
\fi
These assignments can be further confirmed by the observed relative Raman intensity and the phonon dispersion relation. First, in Fig.~\ref{Fig1}(b), a larger relative intensity of Raman spectra $\omega_i$ can be found at E$\gamma$ = 2.33 eV than at 1.58 eV. This supports the assignments at the M point and at $\vec q_{\textrm{MM}'}$ $\sim$10\%$\overline{\textrm{MM}'}$ for E$\gamma$ = 2.33 eV and 1.58 eV, respectively, since phonon density of states (PhDOS) at the M point is larger than that at $\vec q_{\textrm{MM}'}$ $\sim$10\%$\overline{\textrm{MM}'}$ as shown in Fig.~\ref{Fig4}(b). Similarly, for E$\gamma$ = 1.96 eV,
a relatively large (small) intensity of the Raman spectra $\omega_2$, $\omega_3$ and $\omega_6$ ($\omega_1$, $\omega_4$ and $\omega_5$) in Fig.~\ref{Fig1}(b) may arise from the large (small) PhDOS at the M point (off the M point). It is therefore reasonable to assign the $\omega_2$, $\omega_3$ and $\omega_6$ ($\omega_1$, $\omega_4$ and $\omega_5$) at the M point (off the M point), as underlined in Table~\ref{Table1}. Second, it is noted that the experimental spectra of $\omega_4$ (at the three laser energies) are too weak to determine the frequency values. The $\omega_4$ is assigned as a combination mode related to the A$_{1g}$(M) as listed in Table~\ref{Table1}. Then it is interesting to discuss why the $\omega_4$ becomes weaker and broader when the first-order A$_{1g}$($\Gamma$) mode becomes stronger by changing the laser excitation energy from 2.33 eV to 1.96 eV/1.58 eV. We expect that the photo-excited electron may choose either an intra-valley scattering process (A$_{1g}$($\Gamma$)) or an inter-valley scattering process (A$_{1g}$(M)), so that A$_{1g}$(M) competes with A$_{1g}$($\Gamma$), resulting in the alternative Raman intensity of A$_{1g}$($\Gamma$) and $\omega_4$ in the experiment. The $\omega_4$ (A$_{1g}$(M)) has no VHS of PhDOS compared with all the other $\omega$'s at the M point as seen from Fig.~\ref{Fig4}(b) which is relevant to the weak intensity of $\omega_4$, too.
\iffalse
These assignments can be further clarified from the relative Raman intensity and phonon dispersion relation by the following two points. First, in Fig.~\ref{Fig1}(b), a larger relative intensity of Raman spectra $\omega_i$ can be found at E$\gamma$ = 2.33 eV than at E$\gamma$ = 1.58 eV. This supports the assignments at the M point and at $\vec q_{\textrm{MM}'}$ $\sim$10\%$\overline{\textrm{MM}'}$ for E$\gamma$ = 2.33 eV and 1.58 eV, respectively, since phonon density of states (PhDOS) at the M point is larger than that at $\vec q_{\textrm{MM}'}$ $\sim$10\%$\overline{\textrm{MM}'}$ as shown in Fig.~\ref{Fig4}(b). Likewise, at E$\gamma$ = 1.96 eV, a relatively large (small) intensity of the Raman spectra $\omega_2$, $\omega_3$ and $\omega_6$ ($\omega_1$, $\omega_4$ and $\omega_5$) in Fig.~\ref{Fig1}(b) may arise from the large (small) PhDOS at the M point (off the M point). It is therefore reasonable to assign the $\omega_2$, $\omega_3$ and $\omega_6$ ($\omega_1$, $\omega_4$ and $\omega_5$) at the M point (off the M point), as underlined in Table~\ref{Table1}. Second, it is worth pointing out that (1) the $\omega_4$ is much weaker than the other $\omega_i$; and (2) the $\omega_4$ becomes weaker and broader with Raman intensity of first-order A$_{1g}$ mode increasing when we change the laser excitation from 2.33 eV to 1.96 eV/1.58 eV. The $\omega_4$ is assigned as a combination mode related to the A$_{1g}$(M) as listed in Table~\ref{Table1}. The A$_{1g}$(M) compared to all the other phonon modes at the M point has no VHS of PhDOS as seen from Fig.~\ref{Fig4}(b) and a relatively small value of PhDOS of the A$_{1g}$(M) may lead to the weak intensity of $\omega_4$. Moreover, the photo-excited electron may choose either an intra-valley scattering process (A$_{1g}$($\Gamma$)) or an inter-valley scattering process (A$_{1g}$(M)), so that A$_{1g}$(M) competes with A$_{1g}$($\Gamma$), resulting in the alternative Raman intensity of A$_{1g}$($\Gamma$) and $\omega_4$ in the experiment. While for the $\omega_5$, no alternative intensity of A$_{1g}$($\Gamma$) and $\omega_5$ is found in Fig.~\ref{Fig1}(b), so the contribution of A$_{1g}$(M) to $\omega_5$ can be excluded. From Table~\ref{Table1}, we see the contributions of the TA modes at the M point to the second-order Raman process. Usually, the TA modes around the $\Gamma$ point do not contribute to the electron-phonon interaction. Nevertheless, this seems not to be the case at the M point, as seen from the mode visualization in Fig.~\ref{Fig5}.
\fi
In Fig.~\ref{Fig5}, we show the normal modes of the phonons at the M point in which the phonon vibration amplitude is visualized in the 2$\times$2 supercell. As seen from Fig.~\ref{Fig4}(b), the two groups of optical phonon modes (E$_{2g}^1$ and (E$_{1g}$ or A$_{1g}$)) are well separated by about 50 cm$^{-1}$. The visualization of these M point phonon modes shows that the LO and TO E$_{2g}$(M) modes (230$\sim$300 cm$^{-1}$) and the LA(M) and TA(M) modes (60$\sim$100 cm$^{-1}$) are bond-stretching modes of the Mo-Te bond. The LO A$_{1g}$(M) and TO E$_{1g}$(M) modes (120$\sim$180 cm$^{-1}$) arises from the relative motion of two Te atoms occupying the same sublattice. The six vibrational modes in Fig.~\ref{Fig5} are Raman active, since the volume or the area of the supercell under vibration gives non-zero deformation potential. Thus all the modes can contribute to the 2nd-order overtone, combination and difference modes. Furthermore, the vibrational magnitude of the TA modes is no less than that of the LA modes, suggesting that the TA modes may induce as large deformation potential as the LA modes in the electron-phonon interaction. Usually, the TA modes around the $\Gamma$ point do not contribute to the electron-phonon interaction. Nevertheless, this seems not to be the case at the M point.
So far we have only considered the electron-involved double resonance process for MoTe$_2$ and did not discuss the hole-involved double resonance process. As Venezuela et al. \cite{Venezuela11} pointed out, the enhancement by the double resonance is strong in graphene for one-electron and one-hole scattering, since the energy denominators of the Raman intensity formula becomes zero for the symmetric $\pi$ and $\pi^*$ bands of graphene (triple resonance). In the case of TMDs, however, a similar effect is not expected since the conduction band and the valence band are not symmetric around the Fermi energy \cite{Wang14}. Since we don't have a program to calculate electron-phonon matrix element by first-principles calculations, we can't calculate Raman intensity for each double resonance process,\cite{Venezuela11} which will be a future work.
\begin{table}[t]
\caption{Laser energies that are used in the Raman spectra of TMDs. E$_g$(K), E$_g$(M) are, respectively, the optical transition energies at the K and M point, which are obtained by LDA calculations. Detail is discussed in the text.}
\begin{tabular}{lllll}
\hline
{TMDs} & {laser(eV)} &{E$_g$(K)} & {E$_g$(M)} \\
\hline
{MoS$_2$}\cite{Chen74,Stacy85} & {1.83-2.41} & {1.7} & {$>$3.0} \\
{WS$_2$}\cite{Sourisseau89,Sourisseau1991,Berkdemir13} &{1.83-2.41} & {1.9} & {$>$3.0} \\
{WSe$_2$}\cite{Zhao13,Terrones14} & {2.41} & {1.7} & {$>$3.0} \\
{MoTe$_2$}$^{\emph{this work}}$ & {1.96, 2.33} & {1.1} & {$\sim$1.8} \\
\hline
\end{tabular}
\label{Table2}
\end{table}
Finally, we briefly comment on the exciton effect on the double resonance Raman spectra of MoTe$_2$. Since the excitonic states appear at the direct energy gap at the K point of the BZ\cite{Chernikov14,Ye14}, we expect that the excitonic energy is around 1.1 eV for MoTe$_2$, which is much smaller than the three laser energies we used (1.58 - 2.33 eV). As far as we consider only MoTe$_2$, we can safely neglect the exciton effect at the K point. However, in the case of other TMDs (MoS$_2$\cite{Stacy85}, WS$_2$\cite{Sourisseau89,Sourisseau1991,Berkdemir13,ws2ki}, WSe$_2$\cite{Zhao13,Terrones14}) energy gaps at the K point (E$_g$(K)) lie in the range
of 1.7-1.9 eV as shown in Table~\ref{Table2}. The laser energies used in the experiments lie in a range from 1.83 to 2.41 eV. Thus the exciton effect needs to be considered for some laser energies. Nevertheless, when we discuss the M-point phonon of the other TMDs, we expect that the second-order Raman spectra do not appear, since the optical transition energies at the M point is larger than 3.0 eV. It would be interesting if ultra-violet laser could be used for detecting the M-point phonon modes for other TMDs, which will be a future work. In the case of MoTe$_2$, on the other hand, since E$_g$(M) $\sim$ 1.8 eV, we expect the double resonance Raman process for conventional laser energies. Although we did not consider the exciton effect at the M point explicitly as K.F. Mak \textit{et al.}\cite{Mak11} did in graphene, we believe that the present discussion of the double resonance Raman effect should work well, because only optical transition matrix elements that are common in the resonant Raman process are enhanced by the exciton effect, and the relative Raman intensity thus obtained do not change much for the M point. \cite{Jiang07a,Jiang07b}
\section{Conclusion}
In summary, we have investigated the second-order Raman spectra of MoTe$_2$ systems, based on an {\em ab initio} density functional calculation, and compared with the experimental Raman spectroscopy. The calculated electronic bands, the Van Hove singularities in the density of states, and the absorption spectra indicate a strong optical absorption at the M point in the BZ of MoTe$_2$. Using double resonance Raman theory, combination/difference or overtone mode of phonons at the M point are assigned to identify the experimental Raman modes of $\omega_i$ (i = 1, 2, ..., 6). This study facilitates a better understanding of the electron-phonon interaction and the second-order Raman process in TMD systems.
\section*{Acknowledgments}
M.Y. and K.T. sincerely acknowledge Prof. Meiyan Ni and Dr. Katsunori Wakabayashi for critical suggestions on experiments. T.Y. and Z.D.Z acknowledge the NSFC under Grant Nos. 11004201, 51331006 and the IMR SYNL-Young Merit Scholars for financial support. R.S. acknowledges MEXT Grants Nos. 25286005 and 25107005. M.S.D. acknowledges NSF-DMR Grant No. 10-04147. K.U. and K.T. acknowledge MEXT Grant No. 25107004. Calculations were performed in the HPC cluster at the Institute of Metal Research in Shenyang, China.
|
\section{{\label{sec:intro}}Introduction}
The recent advancements in quantum processing tasks viz. quantum teleportation, quantum cryptography and quantum computation are mainly based on quantum entanglement. Quantum teleportation is a quantum process in which Alice sends a qubit to a distant receiver Bob through the shared entangled channel. This process of transmission and recreation of a qubit state is possible not only in the laboratory setup but in principle over an arbitrary distance. Quantum Information Splitting (QIS) is the process for splitting information through quantum means into several parts so that no split part or parts is sufficient to recover the information but the entire set is. In case of QIS, the information can be classical or quantum in nature. The idea of teleportation has been successfully applied in the process of quantum information splitting. Since a quantum state is perfectly reconstructed with the help of other parties, the process of quantum teleportation involving more than two parties often referred in literature as quantum state sharing (QSTS).
Bennett {\it et al.}~\cite {BB93} first proposed the idea of quantum teleportation of an arbitrary single qubit using an EPR state as quantum channel. Thereafter, quantum teleportation of a single qubit state was proposed using various multipartite quantum systems. Karlsson {\it et al.}~\cite{KB98} used a tripartite GHZ state for this purpose. Pati had demonstrated a protocol for quantum teleportation using a four-partite GHZ state~\cite{P00} and an asymmetric W state~\cite{AP06}. The perfect teleportation of an arbitrary two-qubit state was first proposed by Rigolin~\cite{R05} using two Bell states. Thereafter, Yeo {\it et al.}~\cite{YC06} had shown that tensor product of two orthogonal states can also be used for this purpose. A genuinely entangled five qubit state was employed by Panigrahi {\it et al.}~\cite{P08} to demonstrate the process of perfect teleportation of arbitrary two qubit states. Two qubit teleportation was also shown by Nie {\it et al.} using five qubit cluster state~\cite{NL10} and six qubit genuinely entangled states~\cite{NJ10}.
The idea of QIS of a single qubit was first due to Hilery {\it et al.}~\cite{HBB99} using three-qubit GHZ states. Later this process was investigated by Karlsson {\it et al.}~\cite{KK99} using three particle entanglement. Cleve {\it et al.}~\cite{C99} used a process similar to error correction and Zheng used W state~\cite{Z06} for the QIS of a single qubit. The QIS of an arbitrary two-qubit state was proposed by Deng {\it et al.} using two GHZ states~\cite{D05}. QIS of two qubit states using cluster states was demonstrated by Nie~\cite{NS11}, Panigrahi~\cite{PM11,PS11} and Han~\cite{H12}. Recently, two-qubit QIS was discussed using arbitrary pure or mixed resource states~\cite{ZL11} and asymmetric multi-particle state~\cite{Z12}.
The organization of this paper is as follows. In Sec.~\ref{sec:3QIS}, we discuss the QIS of two non-equivalent forms of special three-qubit states using a pair of GHZ states as quantum channel. We describe the protocol of QIS of special type of four-qubit states in Sec.~\ref{sec:4QIS}.
\section{\label{sec:3QIS}Quantum Information Splitting of a Three-qubit State}
Any arbitrary three-qubit state cannot be used for QIS using a pair of GHZ states. It is only possible to split a special type of three-qubit state through this type of quantum channel.
Suppose, Alice, Bob and Charlie share a pair of GHZ states, $|\Psi^0_{\text{GHZ}}\rangle=\frac{1}{\sqrt 2}[|000\rangle + |111\rangle]$ among themselves, and they decide that Bob will get the state. Of the first $|\Psi_{\text{GHZ}}^0\rangle$ state, particle 1 belongs to Alice, Bob has particle 2 and Charlie has particle 3. For the the second $|\Psi_{\text{GHZ}}^0\rangle$ state, Alice has particle 1 and rest belongs to Bob. The quantum channel in the $|AABBBC\rangle$ ($A\to$ Alice, $B \to$ Bob and $C\to$ Charlie) format is then given by
\begin{equation*}
|\Psi_G\rangle =\frac{1}{2}[|000000\rangle+|010110\rangle+|101001\rangle+|111111\rangle].
\end{equation*}
In this case, a special three-qubit state $|\xi\rangle=\alpha|000\rangle+\beta|011\rangle+\gamma|100\rangle+\delta|111\rangle$ in Alice's possession, with $|\alpha|^2+|\beta|^2+|\gamma|^2+|\delta|^2=1$, can be used for QIS. The combined state may be written as
{\small
\begin{equation*}
|\xi\rangle\otimes|\Psi_G\rangle = \frac{1}{2}[\alpha|000\rangle+\beta|011\rangle+\gamma|100\rangle+\delta|111\rangle] \otimes
[|000000\rangle+|010110\rangle+|101001\rangle+|111111\rangle].
\end{equation*}
}
Alice then measures the five qubits (the first five in the above representation) in her possession in the following orthonormal basis.
In order to teleport the above mentioned restricted class of three-qubit state, the following subset of 16 states (out of 32), denoted by
$\eta_i$, where $i = 0, \ldots 15$, is sufficient.
\begin{equation*}
|\eta_{i}\rangle = \frac{1}{2}(\sigma_x^{4})^{i_3}(\sigma_x^{5})^{i_2}[|00000\rangle + (-1)^{i_1} |01101\rangle
+(-1)^{i_0} |10010\rangle + (-)^{i_1+i_0}|11111\rangle],
\end{equation*}
where $i_3i_2i_1i_0$ is the 4-bit binary representation of $i$ and
$\sigma_x^b$ denotes $\sigma_x$ operated on the $b$-th bit ($b=1, \ldots, 5$)
from the left end corresponding to each term. Alternatively, starting with
\begin{equation*}
|\eta_{0}\rangle=\frac{1}{2}[|00000\rangle+|01101\rangle+|10010\rangle+|11111\rangle],
\end{equation*}
we can represent the other states in terms of $|\eta_{0}\rangle$ as
\begin{equation*}
|\eta_{i}\rangle = (\sigma_x^{4})^{i_3}(\sigma_x^{5})^{i_2} (\sigma_z^{4})^{i_1}(\sigma_z^{5})^{i_0}) |\eta_{0}\rangle.
\end{equation*}
After the measurement, Alice communicates her measurement results via four classical bits to Bob. Charlie now measures his qubit in the Hadamard basis and communicates his result to Bob through one cbit. Alice and Charlie's measurement results, their communicated results to Bob and Bob's corresponding operations are listed in Table~\ref{tab1}. Note that Alice's information corresponding
to the state $\eta_i$ is given by the 4-bit binary representation of $i$.
Moreover, Charlie's information is 0 or 1 depending on whether his state is
$|+\rangle$ or $|-\rangle$.
\begin{table}[hb]
\tcaption{Strategy for recovering the three-qubit state}
\centerline{\footnotesize\smalllineskip
\begin{tabular}{c c c c}\\
\hline
Alice's &Charlie's &Bob's &Bob's\\
state &state &State &Operation\\
\hline
$|\eta_0\rangle$ &$|+\rangle$ &$\alpha|000\rangle+\beta|011\rangle+\gamma|100\rangle+\delta|111\rangle$ &$I \otimes I\otimes I$ \\
\hline
$|\eta_0\rangle$ &$|-\rangle$ &$\alpha|000\rangle+\beta|011\rangle-\gamma|100\rangle-\delta|111\rangle$ &$\sigma_z \otimes I\otimes I$ \\
\hline
$|\eta_1\rangle$ &$|+\rangle$ &$\alpha|000\rangle-\beta|011\rangle+\gamma|100\rangle-\delta|111\rangle$ &$I \otimes \sigma_z\otimes I $ \\
\hline
$|\eta_1\rangle$ &$|-\rangle$ &$\alpha|000\rangle-\beta|011\rangle-\gamma|100\rangle+\delta|111\rangle$ &$\sigma_z \otimes \sigma_z\otimes I $ \\
\hline
$|\eta_2\rangle$ &$|+\rangle$ &$\alpha|000\rangle+\beta|011\rangle-\gamma|100\rangle-\delta|111\rangle$ &$\sigma_z \otimes I \otimes I$ \\
\hline
$|\eta_2\rangle$ &$|-\rangle$ &$\alpha|000\rangle+\beta|011\rangle+\gamma|100\rangle+\delta|111\rangle$ &$I \otimes I \otimes I$ \\
\hline
$|\eta_3\rangle$ &$|+\rangle$ &$\alpha|000\rangle-\beta|011\rangle-\gamma|100\rangle+\delta|111\rangle$ &$\sigma_z \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_3\rangle$ &$|-\rangle$ &$\alpha|000\rangle-\beta|011\rangle+\gamma|100\rangle-\delta|111\rangle$ &$I \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_4\rangle$ &$|+\rangle$ &$\alpha|011\rangle+\beta|000\rangle+\gamma|111\rangle+\delta|100\rangle$ &$I \otimes \sigma_x \otimes \sigma_x $ \\
\hline
$|\eta_4\rangle$ &$|-\rangle$ &$\alpha|011\rangle+\beta|000\rangle-\gamma|111\rangle-\delta|100\rangle$ &$\sigma_z \otimes \sigma_x \otimes \sigma_x $ \\
\hline
$|\eta_5\rangle$ &$|+\rangle$ &$\alpha|011\rangle-\beta|000\rangle+\gamma|111\rangle-\delta|100\rangle$ &$I \otimes i\sigma_y \otimes \sigma_x $ \\
\hline
$|\eta_5\rangle$ &$|-\rangle$ &$\alpha|011\rangle-\beta|000\rangle-\gamma|111\rangle+\delta|100\rangle$ &$\sigma_z \otimes i\sigma_y \otimes \sigma_x $ \\
\hline
$|\eta_6\rangle$ &$|+\rangle$ &$\alpha|011\rangle+\beta|000\rangle-\gamma|111\rangle-\delta|100\rangle$ &$\sigma_z \otimes \sigma_x \otimes \sigma_x $ \\
\hline
$|\eta_6\rangle$ &$|-\rangle$ &$\alpha|011\rangle+\beta|000\rangle+\gamma|111\rangle+\delta|100\rangle$ &$I \otimes \sigma_x \otimes \sigma_x $ \\
\hline
$|\eta_7\rangle$ &$|+\rangle$ &$\alpha|011\rangle-\beta|000\rangle-\gamma|111\rangle+\delta|100\rangle$ &$\sigma_z \otimes i\sigma_y \otimes \sigma_x $ \\
\hline
$|\eta_7\rangle$ &$|-\rangle$ &$\alpha|011\rangle-\beta|000\rangle+\gamma|111\rangle-\delta|100\rangle$ &$I \otimes i\sigma_y \otimes \sigma_x $ \\
\hline
$|\eta_8\rangle$ &$|+\rangle$ &$\alpha|100\rangle+\beta|111\rangle+\gamma|000\rangle+\delta|011\rangle$ &$\sigma_x \otimes I \otimes I$ \\
\hline
$|\eta_8\rangle$ &$|-\rangle$ &$\alpha|100\rangle+\beta|111\rangle-\gamma|000\rangle-\delta|011\rangle$ &$i\sigma_y \otimes I \otimes I$ \\
\hline
$|\eta_9\rangle$ &$|+\rangle$ &$\alpha|100\rangle-\beta|111\rangle+\gamma|000\rangle-\delta|011\rangle$ &$\sigma_x \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_9\rangle$ &$|-\rangle$ &$\alpha|100\rangle-\beta|111\rangle-\gamma|000\rangle+\delta|011\rangle$ &$i\sigma_y \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_{10}\rangle$ &$|+\rangle$ &$\alpha|100\rangle+\beta|111\rangle-\gamma|000\rangle-\delta|011\rangle$ &$i\sigma_y \otimes I \otimes I$ \\
\hline
$|\eta_{10}\rangle$ &$|-\rangle$ &$\alpha|100\rangle+\beta|111\rangle+\gamma|000\rangle+\delta|011\rangle$ &$\sigma_x \otimes I \otimes I$ \\
\hline
$|\eta_{11}\rangle$ &$|+\rangle$ &$\alpha|100\rangle-\beta|111\rangle-\gamma|000\rangle+\delta|011\rangle$ &$i\sigma_y \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_{11}\rangle$ &$|-\rangle$ &$\alpha|100\rangle-\beta|111\rangle+\gamma|000\rangle-\delta|011\rangle$ &$\sigma_x \otimes \sigma_z \otimes I$ \\
\hline
$|\eta_{12}\rangle$ &$|+\rangle$ &$\alpha|111\rangle+\beta|100\rangle+\gamma|011\rangle+\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_x $\\
\hline
$|\eta_{12}\rangle$ &$|-\rangle$ &$\alpha|111\rangle+\beta|100\rangle-\gamma|011\rangle-\delta|000\rangle$ &$i\sigma_y \otimes \sigma_x \otimes \sigma_x $\\
\hline
$|\eta_{13}\rangle$ &$|+\rangle$ &$\alpha|111\rangle-\beta|100\rangle+\gamma|011\rangle-\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_x$\\
\hline
$|\eta_{13}\rangle$ &$|-\rangle$ &$\alpha|111\rangle-\beta|100\rangle-\gamma|011\rangle+\delta|000\rangle$ &$i\sigma_y \otimes i\sigma_y \otimes \sigma_x$\\
\hline
$|\eta_{14}\rangle$ &$|+\rangle$ &$\alpha|111\rangle+\beta|100\rangle-\gamma|011\rangle-\delta|000\rangle$ &$i\sigma_y \otimes \sigma_x \otimes \sigma_x$\\
\hline
$|\eta_{14}\rangle$ &$|-\rangle$ &$\alpha|111\rangle+\beta|100\rangle+\gamma|011\rangle+\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_x$\\
\hline
$|\eta_{15}\rangle$ &$|+\rangle$ &$\alpha|111\rangle-\beta|100\rangle-\gamma|011\rangle+\delta|000\rangle$ &$i\sigma_y \otimes i\sigma_y \otimes \sigma_x $\\
\hline
$|\eta_{15}\rangle$ &$|-\rangle$ &$\alpha|111\rangle-\beta|100\rangle+\gamma|011\rangle-\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_x $\\
\hline
\end{tabular}}
\label{tab1}
\end{table}
There is another non-equivalent distribution of particles among the three parties. In this case, of the first $|\Psi_{\text{GHZ}}^0\rangle$ state, particle 1 belongs to Alice and rest belongs to Bob and of the the second $|\Psi_{\text{GHZ}}^0\rangle$ state, Alice has particle 1, Bob has particle 2 and Charlie has particle 3. The quantum channel in this case (in AABBBC format) is given by
\begin{equation*}
|\tilde{\Psi}_G\rangle =\frac{1}{2}[|000000\rangle+|010011\rangle+|101100\rangle+|111111\rangle].
\end{equation*}
The corresponding three qubit state (in Alice's possession) used for QIS is
$|\tilde{\xi}\rangle=(\alpha|000\rangle+\beta|001\rangle+\gamma|110\rangle+\delta|111\rangle)$ with $|\alpha|^2+|\beta|^2+|\gamma|^2+|\delta|^2=1$. The combined state is given by
{\small
\begin{equation*}
|\tilde{\xi}\rangle\otimes|\tilde{\Psi}_G\rangle = \frac{1}{2}[\alpha|000\rangle+\beta|011\rangle+\gamma|100\rangle+\delta|111\rangle]\otimes
[|000000\rangle+|010011\rangle+|101100\rangle+|111111\rangle].
\end{equation*}
}
In this case, Alice performs a 5-qubit measurement in the orthonormal basis
$\tilde{\eta}_i$, $i = 0, \ldots 15$ as given below.
\begin{equation*}
|\tilde{\eta}_{i}\rangle = \frac{1}{2}(\sigma_x^{4})^{i_3}(\sigma_x^{5})^{i_2}[|00000\rangle + (-1)^{i_1} |00101\rangle
+(-1)^{i_0} |11010\rangle + (-)^{i_1+i_0}|11111\rangle],
\end{equation*}
where $i_3i_2i_1i_0$ is the 4-bit binary representation of $i$ and
$\sigma_x^b$ denotes $\sigma_x$ operated on the $b$-th bit ($b=1, \ldots, 5$)
from the left end corresponding to each term. Alternatively, starting with
\begin{equation*}
|\tilde{\eta}_{0}\rangle=\frac{1}{2}[|00000\rangle+|00101\rangle+|11010\rangle+|11111\rangle],
\end{equation*}
we can represent the other states in terms of $|\tilde{\eta}_{0}\rangle$ as
\begin{equation*}
|\tilde{\eta}_{i}\rangle = (\sigma_x^{4})^{i_3}(\sigma_x^{5})^{i_2} (\sigma_z^{4})^{i_1}(\sigma_z^{5})^{i_0}) |\tilde{\eta}_{0}\rangle.
\end{equation*}
Alice communicates her measurement results via four classical bits. Charlie communicates his measurement result in the Hadamard basis through one cbit. Alice and Charlie's measurement results, their communicated results to Bob and Bob's corresponding operations are listed in Table~\ref{tab2}. Similar to
Table~\ref{tab1}, here also Alice's information corresponding
to the state $\tilde{\eta}_i$ is given by the 4-bit binary representation of $i$. Further, Charlie's information is 0 or 1 depending on whether his state is
$|+\rangle$ or $|-\rangle$.
\begin{table}[hb]
\tcaption{Strategy for recovering the three-qubit state}
\centerline{\footnotesize\smalllineskip
\begin{tabular}{c c c c}\\
\hline
Alice's &Charlie's &Bob's &Bob's\\
state &state &State &Operation\\
\hline
$|\tilde{\eta}_0\rangle$ &$|+\rangle$ &$\alpha|000\rangle+\beta|001\rangle+\gamma|110\rangle+\delta|111\rangle$ &$I \otimes I\otimes I$ \\
\hline
$|\tilde{\eta}_0\rangle$ &$|-\rangle$ &$\alpha|000\rangle+\beta|001\rangle-\gamma|110\rangle-\delta|111\rangle$ &$I \otimes \sigma_z\otimes I$ \\
\hline
$|\tilde{\eta}_1\rangle$ &$|+\rangle$ &$\alpha|000\rangle-\beta|001\rangle+\gamma|110\rangle-\delta|111\rangle$ &$I \otimes I \otimes \sigma_z $ \\
\hline
$|\tilde{\eta}_1\rangle$ &$|-\rangle$ &$\alpha|000\rangle-\beta|001\rangle-\gamma|110\rangle+\delta|111\rangle$ &$I \otimes \sigma_z\otimes \sigma_z $ \\
\hline
$|\tilde{\eta}_2\rangle$ &$|+\rangle$ &$\alpha|000\rangle+\beta|001\rangle-\gamma|110\rangle-\delta|111\rangle$ &$I \otimes \sigma_z \otimes I$ \\
\hline
$|\tilde{\eta}_2\rangle$ &$|-\rangle$ &$\alpha|000\rangle+\beta|001\rangle+\gamma|110\rangle+\delta|111\rangle$ &$I \otimes I \otimes I$ \\
\hline
$|\tilde{\eta}_3\rangle$ &$|+\rangle$ &$\alpha|000\rangle-\beta|001\rangle-\gamma|110\rangle+\delta|111\rangle$ &$I \otimes \sigma_z \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_3\rangle$ &$|-\rangle$ &$\alpha|000\rangle-\beta|001\rangle+\gamma|110\rangle-\delta|111\rangle$ &$I \otimes I \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_4\rangle$ &$|+\rangle$ &$\alpha|001\rangle+\beta|000\rangle+\gamma|111\rangle+\delta|110\rangle$ &$I \otimes I \otimes \sigma_x $ \\
\hline
$|\tilde{\eta}_4\rangle$ &$|-\rangle$ &$\alpha|001\rangle+\beta|000\rangle-\gamma|111\rangle-\delta|110\rangle$ &$I \otimes \sigma_z \otimes \sigma_x $ \\
\hline
$|\tilde{\eta}_5\rangle$ &$|+\rangle$ &$\alpha|001\rangle-\beta|000\rangle+\gamma|111\rangle-\delta|110\rangle$ &$I \otimes I \otimes i\sigma_y $ \\
\hline
$|\tilde{\eta}_5\rangle$ &$|-\rangle$ &$\alpha|001\rangle-\beta|000\rangle-\gamma|111\rangle+\delta|110\rangle$ &$I \otimes \sigma_z \otimes i\sigma_y $ \\
\hline
$|\tilde{\eta}_6\rangle$ &$|+\rangle$ &$\alpha|001\rangle+\beta|000\rangle-\gamma|111\rangle-\delta|110\rangle$ &$I \otimes \sigma_z \otimes \sigma_x $ \\
\hline
$|\tilde{\eta}_6\rangle$ &$|-\rangle$ &$\alpha|001\rangle+\beta|000\rangle+\gamma|111\rangle+\delta|110\rangle$ &$I \otimes I \otimes \sigma_x $ \\
\hline
$|\tilde{\eta}_7\rangle$ &$|+\rangle$ &$\alpha|001\rangle-\beta|000\rangle-\gamma|111\rangle+\delta|110\rangle$ &$I \otimes \sigma_z \otimes i\sigma_y $ \\
\hline
$|\tilde{\eta}_7\rangle$ &$|-\rangle$ &$\alpha|001\rangle-\beta|000\rangle+\gamma|111\rangle-\delta|110\rangle$ &$I \otimes I \otimes i\sigma_y $ \\
\hline
$|\tilde{\eta}_8\rangle$ &$|+\rangle$ &$\alpha|110\rangle+\beta|111\rangle+\gamma|000\rangle+\delta|001\rangle$ &$\sigma_x \otimes \sigma_x \otimes I$ \\
\hline
$|\tilde{\eta}_8\rangle$ &$|-\rangle$ &$\alpha|110\rangle+\beta|111\rangle-\gamma|000\rangle-\delta|001\rangle$ &$\sigma_x \otimes i\sigma_y \otimes I$ \\
\hline
$|\tilde{\eta}_9\rangle$ &$|+\rangle$ &$\alpha|110\rangle-\beta|111\rangle+\gamma|000\rangle-\delta|001\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_9\rangle$ &$|-\rangle$ &$\alpha|110\rangle-\beta|111\rangle-\gamma|000\rangle+\delta|001\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_{10}\rangle$ &$|+\rangle$ &$\alpha|110\rangle+\beta|111\rangle-\gamma|000\rangle-\delta|001\rangle$ &$\sigma_x \otimes i\sigma_y \otimes I$ \\
\hline
$|\tilde{\eta}_{10}\rangle$ &$|-\rangle$ &$\alpha|110\rangle+\beta|111\rangle+\gamma|000\rangle+\delta|001\rangle$ &$\sigma_x \otimes \sigma_x \otimes I$ \\
\hline
$|\tilde{\eta}_{11}\rangle$ &$|+\rangle$ &$\alpha|110\rangle-\beta|111\rangle-\gamma|000\rangle+\delta|001\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_{11}\rangle$ &$|-\rangle$ &$\alpha|110\rangle-\beta|111\rangle+\gamma|000\rangle-\delta|001\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_z$ \\
\hline
$|\tilde{\eta}_{12}\rangle$ &$|+\rangle$ &$\alpha|111\rangle+\beta|110\rangle+\gamma|001\rangle+\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_x $\\
\hline
$|\tilde{\eta}_{12}\rangle$ &$|-\rangle$ &$\alpha|111\rangle+\beta|110\rangle-\gamma|001\rangle-\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_x $\\
\hline
$|\tilde{\eta}_{13}\rangle$ &$|+\rangle$ &$\alpha|111\rangle-\beta|110\rangle+\gamma|001\rangle-\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes i\sigma_y$\\
\hline
$|\tilde{\eta}_{13}\rangle$ &$|-\rangle$ &$\alpha|111\rangle-\beta|110\rangle-\gamma|001\rangle+\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes i\sigma_y$\\
\hline
$|\tilde{\eta}_{14}\rangle$ &$|+\rangle$ &$\alpha|111\rangle+\beta|110\rangle-\gamma|001\rangle-\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes \sigma_x$\\
\hline
$|\tilde{\eta}_{14}\rangle$ &$|-\rangle$ &$\alpha|111\rangle+\beta|110\rangle+\gamma|001\rangle+\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes \sigma_x$\\
\hline
$|\tilde{\eta}_{15}\rangle$ &$|+\rangle$ &$\alpha|111\rangle-\beta|110\rangle-\gamma|001\rangle+\delta|000\rangle$ &$\sigma_x \otimes i\sigma_y \otimes i\sigma_y $\\
\hline
$|\tilde{\eta}_{15}\rangle$ &$|-\rangle$ &$\alpha|111\rangle-\beta|110\rangle+\gamma|001\rangle-\delta|000\rangle$ &$\sigma_x \otimes \sigma_x \otimes i\sigma_y $\\
\hline
\end{tabular}}
\label{tab2}
\end{table}
\section{\label{sec:4QIS}Quantum Information Splitting of a Four-qubit State}
In this case, the quantum channel connecting Alice, Bob and Charlie
(in ABBBBC format) is given by $\frac{1}{2}[|000000\rangle+|000111\rangle+|111000\rangle+|111111\rangle]$. Here, Alice has particle 1, Bob has particles 2,3,4,5 and Charlie has particle 6. The four-qubit state (in Alice's possession) used for QIS is $|\zeta\rangle=\alpha(|0000\rangle +|0011\rangle)+\beta(|1100\rangle+|1111\rangle)$ with $|\alpha|^2+|\beta|^2 =\frac{1}{2}$. Alice combines this state with the qubit of the shared channel and performs
a 5-qubit measurement on the following joint state of the system:
\begin{eqnarray*}
|\zeta\rangle\otimes|\Psi_G\rangle
& = & [\alpha(|0000\rangle+|0011\rangle)+\beta(|1100\rangle+|1111\rangle)]\\
& & \otimes
\frac{1}{2}[|000000\rangle+|000111\rangle +|111000\rangle+|111111\rangle]\\
& = & \frac{\alpha}{2}[(|\nu_0\rangle+|\nu_1\rangle)(|00000\rangle+|00111\rangle)] +(|\nu_2\rangle+|\nu_3\rangle)(|11000\rangle+|11111\rangle)\\
& & +\frac{\beta}{2}[(|\nu_2\rangle-|\nu_3\rangle)(|00000\rangle+|00111\rangle)]
+(|\nu_0\rangle-|\nu_1\rangle)(|11000\rangle+|11111\rangle),
\end{eqnarray*}
where $\nu_i$'s are Alice's five-qubit measurement basis given by
\begin{equation*}
\begin{split}
|\nu_{0}\rangle=\frac{1}{2}[|00000\rangle+|00110\rangle+|11001\rangle+|11111\rangle],\\
|\nu_{1}\rangle=\frac{1}{2}[|00000\rangle+|00110\rangle-|11001\rangle-|11111\rangle],\\
|\nu_{2}\rangle=\frac{1}{2}[|00001\rangle+|00111\rangle+|11000\rangle+|11110\rangle],\\
|\nu_{3}\rangle=\frac{1}{2}[|00001\rangle+|00111\rangle+|11000\rangle+|11110\rangle].\\
\end{split}
\end{equation*}
Alice and Charlie's measurement results, corresponding classical communication and Bob's unitary operations are listed in Table~\ref{tab3}.
In our notation, Alice's information corresponding
to the state $\nu_i$ is given by the 4-bit binary representation of $i$
and Charlie's information is 0 or 1 depending on whether his state is
$|+\rangle$ or $|-\rangle$.
\begin{table*}[hb]
\tcaption{Strategy for recovering the four-qubit state}
\centerline{\footnotesize\smalllineskip
\begin{tabular}{c c c c}\\
\hline
Alice's &Charlie's &Bob's &Bob's\\
state &state &State &Operation\\
\hline
$|\nu_{0}\rangle$ &$|+\rangle$ &$\alpha(|0000\rangle+|0011\rangle)+\beta(|1100\rangle+|1111\rangle)$ & $I \otimes I\otimes I\otimes I$ \\
\hline
$|\nu_{0}\rangle$ &$|-\rangle$ &$\alpha(|0000\rangle-|0011\rangle)+\beta(|1100\rangle-|1111\rangle)$ & $I \otimes I \otimes \sigma_z\otimes I$ \\
\hline
$|\nu_{1}\rangle$ &$|+\rangle$ &$\alpha(|0000\rangle+|0011\rangle)-\beta(|1100\rangle+|1111\rangle)$ & $I \otimes \sigma_z\otimes I\otimes I$ \\
\hline
$|\nu_{1}\rangle$ &$|-\rangle$ &$\alpha(|0000\rangle-|0011\rangle)-\beta(|1100\rangle-|1111\rangle)$ & $I \otimes \sigma_z\otimes \sigma_z\otimes I$ \\
\hline
$|\nu_{2}\rangle$ &$|+\rangle$ &$\alpha(|1100\rangle+|1111\rangle)+\beta(|0000\rangle+|0011\rangle)$ & $\sigma_x\otimes \sigma_x\otimes I\otimes I$ \\
\hline
$|\nu_{2}\rangle$ &$|-\rangle$ &$\alpha(|1100\rangle-|1111\rangle)+\beta(|0000\rangle-|0011\rangle)$ & $\sigma_x\otimes \sigma_x\otimes \sigma_z\otimes I$ \\
\hline
$|\nu_{3}\rangle$ &$|+\rangle$ &$\alpha(|1100\rangle+|1111\rangle)-\beta(|0000\rangle+|0011\rangle)$ & $\sigma_x\otimes i\sigma_y\otimes I\otimes I$ \\
\hline
$|\nu_{3}\rangle$ &$|-\rangle$ &$\alpha(|1100\rangle-|1111\rangle)-\beta(|0000\rangle-|0011\rangle)$ & $\sigma_x\otimes i\sigma_y\otimes \sigma_z\otimes I$ \\
\hline
\end{tabular}}
\label{tab3}
\end{table*}
\section{\label{sec:conc}Conclusion}
Since a single arbitrary qubit can be split successfully among three parties using a GHZ state, it is quite obvious that one can split an arbitrary two-qubit states among three parties using two GHZ states. But splitting a three qubit
state, or even a special form of it, is not very obvious. In this article, we have shown that the restricted class of two non-equivalent types of three-qubit states can be used for QIS among three parties using a pair of GHZ states as the quantum channel. We also described a protocol for QIS of a restricted class of four-qubit states among the parties using the same channel. Though the realizations of various quantum processing tasks involving single and two-qubit states using various multipartite systems as resource, both theoretically and experimentally, are well understood, there is still some lack of knowledge in our understanding in multi-qubit states. We hope this article will enhance our knowledge in that direction.
\section*{Acknowledgments}
KN would like to thank the University Grants Commission of India, New Delhi for partially supporting this research under Faculty Development Programme [Grant No. F. TF.WC2-005-04/08-09 (ERO)].
\section*{References}
|
\subsubsection{} #1
\bigskip}
\newcommand{\blockn}[1]{\par #1 \bigskip}
\newcommand{\Th}[1]
{
\bigskip
\textbf{Theorem : }{\itshape #1}
\bigskip
}
\newcommand{\Prop}[1]
{
\bigskip
\textbf{Proposition : }{\itshape #1}
\bigskip
}
\newcommand{\Cor}[1]
{
\bigskip
\textbf{Corollary : }{\itshape #1}
\bigskip
}
\newcommand{\Lem}[1]
{
\bigskip
\textbf{Lemma : }{\itshape #1}
\bigskip
}
\newcommand{\Def}[1]
{
\bigskip
\textbf{Definition : }{\itshape #1}
\bigskip
}
\newcommand{\Dem}[1]{
\smallskip
\textbf{Proof : } \par
{#1} $\square$
\bigskip
}
\hyphenation{Gro-then-dieck}
\begin{document}
\pagestyle{plain}
\title{Toward a non-commutative Gelfand duality: Boolean locally separated toposes and Monoidal monotone complete $C^{*}$-categories}
\author{Simon Henry}
\maketitle
\begin{abstract}
** Draft Version ** To any boolean topos one can associate its category of internal Hilbert spaces, and if the topos is locally separated one can consider a full subcategory of square integrable Hilbert spaces. In both case it is a symmetric monoidal monotone complete $C^{*}$-category. We will prove that any boolean locally separated topos can be reconstructed as the classifying topos of ``non-degenerate" monoidal normal $*$-representations of both its category of internal Hilbert spaces and its category of square integrable Hilbert spaces. This suggest a possible extension of the usual Gelfand duality between a class of toposes (or more generally localic stacks or localic groupoids) and a class of symmetric monoidal $C^{*}$-categories yet to be discovered.
\end{abstract}
\renewcommand{\thefootnote}{\fnsymbol{footnote}}
\footnotetext{\emph{Keywords.} Boolean locally separated toposes, monotone complete C*-categories, reconstruction theorem. }
\footnotetext{\emph{2010 Mathematics Subject Classification.} 18B25, 03G30, 46L05, 46L10 .}
\footnotetext{\emph{email:} <EMAIL>}
\renewcommand{\thefootnote}{\arabic{footnote}}
\tableofcontents
\section{Introduction}
\blockn{\emph{This is a draft version. It will be replaced by a more definitive version within a few months. In the meantime any comments is welcome.}}
\blockn{This paper is part of a program (starting with the author's phd thesis) devoted to the study of the relation between topos theory and non-commutative geometry as two generalizations of topology. The central theme of this research project is the construction explained in section \ref{secPrelim} which naturally associate to any topos $\Tcal$ the $C^{*}$-category $\Hcal(\Tcal)$ of Hilbert bundles over $\Tcal$.}
\blockn{In the previous paper \cite{henry2014measure} we studied ``measure theory" of toposes and compare it to the theory of $W^{*}$-algebras (von Neumann algebras) through this construction of $\Hcal(\Tcal)$. At the end of the introduction of this previous paper one can find a table informally summing up some sort of partially dictionary between topos theory and operator algebra theory. The goal of the present work is somehow to provide a framework for making this dictionary a concrete mathematical result, by showing that a boolean locally separated topos $\Tcal$ can actually be completely reconstructed from any of the two symmetric monoidal\footnote{The term monoidal is taken with a slightly extended meaning: $\Hcal^{red}(\Tcal)$ has in general no unit object} $C^{*}$-categories $\Hcal(\Tcal)$ and $\Hcal^{red}(\Tcal)$ of Hilbert bundles and square-integrable Hilbert bundles.
}
\blockn{More precisely, we will show that if $\Tcal$ is a boolean locally separated topos, then $\Tcal$ is the classifying topos for non-degenerate\footnote{see definitions \ref{DefNondegenUnred} and \ref{DefRepReduced}.} normal symmetric monoidal representations of either $\Hcal^{red}(\Tcal)$ and $\Hcal(\Tcal)$. These are not geometric theory, not even first order theory in fact, but we will prove an equivalence of categories between points of $\Tcal$ and these representations over an arbitrary base topos, see theorems \ref{mainThunred} and \ref{mainThRed} for the precise statement.
}
\blockn{Sections \ref{secPrelim},\ref{secMCcat_of_a_bool} and \ref{secLSTandSIHS} contain some preliminaries which are mostly, but not entirely, recall of previous work.}
\blockn{Section \ref{secStatement} contains the statement of the two main theorems of the present paper: theorem \ref{mainThunred}, which is the reconstruction theorem from the ``unreduced" category $\Hcal(\Tcal)$ and theorem \ref{mainThRed} which is the reconstruction theorem from the ``reduced" category $\Hcal^{red}(\Tcal)$ of square integrable Hilbert spaces.
}
\blockn{Sections \ref{secProofRed} and \ref{secProofUnred} contain respectively the proof of the ``reduced" theorem \ref{mainThRed} and the proof that the reduced theorem imply the unreduced theorem \ref{mainThunred}.
The key arguments, which actually use the specificity of the hypothesis ``boolean locally separated" seems to all be in subsection \ref{GeomorphonObjec}, and the rest of the proof seems to us, at least in comparison, more elementary (it is mostly about some computations of multilinear algebra) and also more general (the specific hypothesis are essentially\footnote{Lemma \ref{Fsurjection2} seems to be the unique exception} not used anymore). In fact, we have unsuccessfully tried to prove a result in this spirit for several years, and results of subsection \ref{GeomorphonObjec} have always been the main stumbling block.
}
\blockn{Finally we conclude this paper with section \ref{secToward}, where we explain how the results of this paper might maybe be extended into a duality between certain monoidal symmetric $C^{*}$-categories and certain geometric objects (presumably, localic groupoids or localic stacks) which would be a common extension of the usual Gelfand duality, the Gelfand duality for $W^{*}$-algebras, the Doplicher-Roberts reconstruction theorem for compact groups, and of course the results of the present paper. This would somehow constitute a sort of non-commutative Gelfand duality. Of course the existence of such a duality, and even its precise statement are at the present time highly conjectural, but we will try to highlight what are the main difficulties on the road toward such a result.
This conjectured duality is formally extremely similar to the reconstruction theorems obtained for algebraic stacks, as for example \cite{lurie2004tannaka}, \cite{brandenburg2014tensor}.
}
\section{General preliminaries}
\label{secPrelim}
\blockn{We will make an intensive use of the internal logic of toposes (i.e. the Kripke-Joyal semantics for intuitionist logic in toposes) in this paper. A reader unfamiliar with this technique can read for example sections $14$,$15$ and $16$ of \cite{mclarty1996elementary} which give a relatively short and clear account of the subject. Other possible references are \cite[chapter 6]{borceux3}, \cite[chapter VI]{maclane1992sheaves}, or \cite[D1 and D4]{sketches}.
}
\blockn{Because this paper is mostly about boolean toposes and monotone complete $C^{*}$-algebras we will assume that the base topos (that is the category of set) satisfies the law of excluded middle, but we won't need to assume the axiom of choice.
It is reasonable to think that these results can also be formulated and proved over a non-boolean basis, but the gain in doing so would be very small: over a non-boolean basis, any boolean topos and any monotone complete $C^{*}$-algebra is automatically defined over a boolean sub-locale of the terminal object, and hence can be dealt with in a boolean framework.
This being said, a large part of the proofs will take place internally in a non boolean topos and hence will have to avoid the law of excluded middle anyway.
}
\blockn{We will also often have to juggle between internal and external logic or between the internal logic of two different toposes. We will generally precise what statement has to be interpreted ``internally in $\Tcal$" or ``externally", but we also would like to emphasis the fact that most of the time the context makes this completely clear: if an argument start by ``let $x \in X$" and that $X$ is not a set but an object of a topos $\Tcal$ it obviously means that we are working internally in $\Tcal$. This convention, is in fact completely similar with the usual use of context\footnote{``Context" is taken here in its formal ``type theoretic" meaning, i.e. the ``set" of all variable that has been declared at a given point of a proof.} in mathematics: when a mathematicians says something like ``let $x \in S$" (for $S$ a set) then what follows is actually mathematics internal to the topos $Sets_{/S}$ of sets over $S$, indeed everything being said after implicitly depends on a parameter $s \in S$, and if the conclusion actually does not depends on $s$ then it will be valid independently of the context only if $S$ was non-empty.
Our convention is hence that when we say something like ``let $x \in X$" where $X$ is an object of a topos $\Tcal$ then what follows is internal to the topos $\Tcal_{/X}$, or equivalently, internal to $\Tcal$ with a declared variable $x$, and this being true until the moment where this variable $x$ is ``removed" from the context (in classical mathematics, this is usually left implicit because, as soon as $X$ is non-empty, it is irrelevant, but in our case we will generally make it precise by saying that we are now working externally).
We will of course say explicitly in which topos we are working as soon as we think that it actually improve the readability, but we also think that this perspective makes (once we are used to it) the change from working internally into one topos from another topos as simple as introducing and forgetting abstract variables in usual mathematics and makes the text easier to read.
}
\blockn{All the toposes considered are Grothendieck toposes, in particular they all have a natural numbers object (see \cite[A2.5 and D5.1]{sketches}) ``$\mathbb{N}$" or ``$\mathbb{N}_{\Tcal}$" which is just the locally constant sheaf equal to $\mathbb{N}$.}
\blockn{A set, or an object $X$ of a topos, is said to be \emph{inhabited} if (internally) it satisfies $\exists x \in X$. For an object of a topos it corresponds to the fact that the map from $X$ to the terminal object is an epimorphism. }
\blockn{An object $X$ of a topos $\Tcal$ is said to be a bound of $\Tcal$ if subobjects of $X$ form a generating family of $\Tcal$ (i.e. is any object of $\Tcal$ admit a covering by subobject of $X$). Every Grothendieck topos admit a bound (for exemaple take the direct sum of all the representable sheaves for a given site of definition), in fact the existence of a bound together with the existence of small co-products characterize Grothendieck toposes among elementary toposes.}
\blockn{Let $\Tcal$ be an arbitrary topos. $\mathbb{C}_{\Tcal}$ is the object of ``continuous\footnote{also called Dedekind complex numbers} complex numbers" that is $\mathbb{R}_{\Tcal} \times \mathbb{R}_{\Tcal}$ where $\mathbb{R}_{\Tcal}$ is the object of continuous/Dedekind real numbers as defined for example in \cite[D4.7]{sketches}. In any topos (with a natural number object), $\mathbb{C}_{\Tcal}$ is a locale ring object.}
\blockn{When $P$ is a decidable proposition, i.e. if one does have $P$ or not-$P$, one denotes $\mathbb{I}_P$ the real number defined by $\mathbb{I}_P = 1$ if $P$ and $\mathbb{I}_P$=0 otherwise. An object $X$ is said to be \emph{decidable} if for all $x$ and $y$ in $X$ the proposition $x = y$ is decidable, in which case we denote $\delta_{x,y}$ for $\mathbb{I}_{x=y}$. i.e. $\delta_{x,y}$ is one if $x=y$ and zero otherwise. }
\blockn{By a ``Hilbert space of $\Tcal$", or a $\Tcal$-Hilbert space we mean an object $H$ of $\Tcal$, endowed with a $\mathbb{C}_{\Tcal}$-module structure and a scalar product $H \times H \rightarrow \mathbb{C}_{\Tcal}$ (linear in the second variable and anti-linear in the first), which satisfies internally all the usual axioms for being a Hilbert space, completeness being interpreted in term of Cauchy filters, or equivalently Cauchy approximations but not Cauchy sequences.
}
\blockn{$\Hcal(\Tcal)$ denotes the $C^{*}$-category\footnote{see for exemple \cite{wstarcat} for the definition of $C^{*}$-category.} whose objects are Hilbert spaces of $\Tcal$ and whose morphisms are ``globally bounded operators", that is linear maps $f : H \rightarrow H'$ which admit an adjoint and such that it exists an external number $K$ satisfying (internally) for all $x \in H$, $\Vert f(x) \Vert \leqslant K \Vert x \Vert $. The norm $\Vert f \Vert_{\infty}$ is then the smallest such constant $K$ (if we were not assuming the law of excluded middle in the base topos it would be an upper semi-continuous real number), the addition and composition of operators is defined internally , $f^{*}$ is the adjoint of $f$ internally, and this form a $C^{*}$-category.
}
\blockn{Because the tensor product of two Hilbert spaces can easily be defined, even in intuitionist mathematics, the category $\Hcal(\Tcal)$ is endowed with a symmetric monoidal structure.}
\blockn{Precise definition of a symmetric monoidal category, a symmetric monoidal functor, and a symmetric monoidal natural transformation can be found in S.MacLane's category theory books\footnote{Only in the second edition (1998).} \cite{maclane1998categories}, chapter \textsc{XI}, section $1$ and $2$.
Briefly, a symmetric monoidal category is a category endowed with a bi-functor $\_ \otimes \_ $, a specific ``unit" object $e$ and isomorphisms $e \otimes A \simeq A \otimes e \simeq A$, $(A \otimes B) \simeq (B \otimes A)$ and $(A \otimes B) \otimes C \simeq A \otimes ( B \otimes C)$ which are natural and satisfies certain coherence conditions. A symmetric monoidal functor (``braided strong monoidal functor" in MacLane's terminology) is a functor between symmetric monoidal categories with a natural isomorphism $F(A \otimes B) \simeq F(A) \otimes F(B)$ which has to satisfy a certain number of coherence and compatibility relations. Finally a symmetric monoidal transformation (the adjective symmetric is actually irrelevant for natural transformation) is a natural transformation between symmetric monoidal functor which satisfy coherence conditions, stating that, up to the previously defined natural isomorphisms, $\eta_A \otimes \eta_b$ is the same as $\eta_{A \otimes B}$ and $\eta_{e}$ is the identity.
}
\blockn{Moreover, when we are talking about (symmetric) monoidal $C^{*}$-categories or symmetric monoidal $*$-functor between such categories we are always assuming that all the structural isomorphisms are in fact isometric isomorphisms, i.e. their inverse is their adjoint. For example, it is clearly the case for $\Hcal(\Tcal)$.}
\blockn{Finally, if $f:\Ecal \rightarrow \Tcal$ is a geometric morphism between two toposes, and $H$ is a Hilbert space of $\Tcal$ then $f^{*}(H)$ is a ``pre-Hilbert" space of $\Ecal$ but fails in general to be complete and separated, we denote by $ f^{\sharp}(H)$ its separated completion. $f^{\sharp}$ is a symmetric monoidal $*$-functor from $\Hcal(\Tcal)$ to $\Hcal(\Ecal)$. }
\section{Monotone complete $C^{*}$-categories and boolean toposes}
\label{secMCcat_of_a_bool}
\blockn{We recall that a $C^{*}$-algebra is said to be \emph{monotone complete} if every bounded directed net of positive operators has a supremum. A positive linear map between two monotone complete $C^{*}$-algebras is said to be \emph{normal} if it preserves supremum of bounded directed set of positive operators.
The theory of monotone complete $C^{*}$-algebras is extremely close to the theory of $W^{*}$-algebras, in fact it is well know that a monotone complete $C^{*}$-algebra having enough normal positive linear form is a $W^{*}$-algebra (see \cite[Theorem 3.16]{takesaki2003theoryI}). }
\blockn{When $\Tcal$ is a boolean topos, $\Hcal(\topos)$ is a monotone complete $C^{*}$-category in the sense that it has bi-products and the $C^{*}$-algebra of endomorphisms of any object is monotone complete. Indeed, because $\topos$ is boolean, the supremum of a bounded net of operators can be computed internally, and as the supremum is unique it ``patches up" into an externally defined map, see \cite[section 2]{henry2014measure}.
Monotone complete $C^{*}$-categories are extremely close to the $W^{*}$-categories studied in \cite{wstarcat}, in fact most of the result of \cite{wstarcat} which does not involve the existence of normal states (or the modular time evolution) also hold for monotone complete $C^{*}$-categories. We will review some of these results:
}
\blockn{If $C$ is a monotone complete $C^{*}$-category then we define the center $Z(C)$ of $C$ as being the commutative monotone complete $C^{*}$-algebra of endomorphisms of the identity functor of $C$. In the more general situation $Z(C)$ might fail to be a set and be a proper class, but we will not be concern by this issue because we proved in \cite[3.6]{henry2014measure} that $\Hcal(\topos)$ has a generator and hence, by results of \cite{wstarcat}, the algebra $Z(C)$ can be identified with the center of the algebra of endomorphisms of this generator. }
\blockn{If $A$ is an object of a monotone complete $C^{*}$-category then we define its central support $c(A) \in Z(C)$ by:
\[ c(A)_B := \sup_{f:A^{n} \rightarrow B \atop \Vert f \Vert < 1} f f^{*} \in Hom(B,B). \]
If the monotone complete category we are working with admit bi-product then $A^{n}$ denotes the bi-product of $n$-copies of $A$ and if not, then one can still make sense of a map $(f_1,\dots , f_n)$ from $A^{n}$ to $B$ as the data of $n$ maps from $A$ to $B$, $ff^{*}$ is $\sum f_i f_i ^{*}$ a,d $\Vert f \Vert$ is defined as $\Vert f f^{*}\Vert^{1/2}$. One can check that the supremum involved in the definition of $C(A)_B$ is directed by showing that the set of such ``$ff^{*}$" is in order preserving bijection with the set of $ff^{*}$ where $f$ is an arbitrary maps from $A^{n}$ to $B$ without condition on the norm, the bijection being obtained by multiplying $f$ by a convenient function of $ff^{*}$.
Equivalently, $c(A)$ can be defined as the smallest projection $c$ in $Z(C)$ such that $c_A=Id_A$, but we will need the fact that it is a directed supremum.
}
\blockn{One says that an object $A$ is \emph{quasi-contained} in an object $B$ if $c(A) \leqslant c(B)$. An object $A$ is said to be a \emph{generator} of a monotone complete $C^{*}$-category if and only if $c(A)=1$ i.e. if every other object is quasi-contained in $A$. }
\blockn{One can for example check that if two normal functors agree on a generator and its endomorphisms then they are isomorphic: it is an easy consequence of results of \cite{wstarcat} for $W^{*}$-categories and the proof can extended to monotone complete $C^{*}$-categories easily. }
\blockn{We conclude this section by briefly mentioning what quasi-containement mean in the case of $\Hcal(\Tcal)$: }
\block{\label{QcontaininHcal}\Prop{Let $\Tcal$ be a boolean topos and $H,H' \in \Hcal(\Tcal)$ two Hilbert spaces of $\Tcal$, then $H$ is weakly contained in $H'$ if and only if there exists a set $F$ of bounded operators from $H'$ to $H$ such that internally in $\Tcal$ the functions in $F$ spam a dense subspace of $H$.}
If this is true, we will say that $H$ is covered by the maps in $F$.
\Dem{If $H$ is weakly contained in $H'$ then $c(H) \leqslant c(H')$ hence $c(H')_H=1$. Rewriting this using the definition of $c(H')$ one gets:
\[ Id_H = \sup_{f:(H')^{n} \rightarrow H \atop \Vert f \Vert < 1} f f^{*} \in Hom(H,H) \]
But as we mentioned earlier, supremums of directed nets in $\Hcal(\Tcal)$ are computed internally, this means that this supremum converge internally for the strong operator topology. In particular, for any $h\in H$ one has $ff^{*}(h)$ which is arbitrarily close to $h$ when $f$ run through the (external) set:
\[ F_0 =\{f | f: (H')^{n} \rightarrow H, \Vert f \Vert < 1 \} \]
Hence taking $F$ to be the set of ``component" of maps in $F_0$, the sum of the images of maps in $F$ spam all of $H$.
\smallskip
Conversely, assume that $H$ is spammed by a family $F$ of external maps $f:H' \rightarrow H$. For any $f:H' \rightarrow H$, and in particular, for any $f \in F$ one has $c(H')_H \circ f = f \circ c(H')_{H'} = f$. The projector $c(H')_H$ is hence (internally in $\Tcal$) equal to the identity on the image of all the maps $f \in F$ and hence on all of $H$, i.e. $c(H')_H=Id_H$ which proves that $c(H) \leqslant c(H')$.
}
}
\section{Locally separated toposes and square integrable Hilbert spaces}
\label{secLSTandSIHS}
\blockn{We recall (see \cite[chapter II]{moerdijk2000proper}) that a topos $\topos$ is said to be \emph{separated} if its diagonal map $\topos \rightarrow \topos \times \topos$ (which is localic by \cite[B3.3.8]{sketches}) is proper, i.e. if, when seen as a $(\topos \times \topos)$-locale though the diagonal maps, $\topos$ is compact.
}
\blockn{In \cite[theorem 5.2]{henry2014measure} we proved that a boolean topos is separated if and only if it is generated by internally finite objects. One will use a slightly modified form of this result :}
\block{\label{SepImpFinitness}\Th{Let $\Tcal$ be a boolean separated topos, then $\Tcal$ admit a generating family of objects $(X)$ such that for each $X$ there exists an interger $n$ such that internally in $\Tcal$, the cardinal of $X$ is smaller than $n$.}
One will say that such objects are of \emph{bounded cardinal}.
\Dem{This theorem is an immediate consequence of \cite[theorem 5.2]{henry2014measure}: $\Tcal$ is generated by a familly of internally finite object $X$, but for each object $X$ of this generating familly and for each natural number $n$ one can define $X_n := \{x \in X | |X| \leqslant n\}$ whose cardinal is internally bounded by $n$, and as $X$ is internally finite the $(X_n)_{n \in \mathbb{N}}$ form a covering family of $X$ and hence the $(X_n)_{X,n}$ form a generating family fulfilling the property announced in the theorem.} }
\blockn{An object $X$ of boolean topos is said to be \emph{separating} if the slice topos $\topos_{/X}$ is separated. A boolean topos is said to be \emph{locally separated} if it admit an inhabited separating object, or equivalently if any object can be covered by separating objects, see \cite{henry2014measure} section 5 for more details. }
\blockn{If $X$ is an object of a boolean topos $\topos$ then one can define the $\Tcal$-Hilbert space $l^{2}(X)$ of square sumable sequences indexed by $X$. It can also be done in a non boolean topos but it require $X$ to be a decidable\footnote{In order to define the scalar product of two generators or to define the sum of a sequence we need that for any $x,y \in X$ $x = y$ or $x \neq y$.} object. Internally in $\topos$, the space $l^{2}(X)$ has generators $e_x$ for $x \in X$ such that $\scal{e_x}{e_y}=\delta_{x,y}$.}
\block{\label{l2weaklycontinl2}\Prop{Let $X$ be a bound of a boolean topos $\topos$ and $Y$ be a separating object of $\topos$ then $l^{2}(Y)$ is quasi-contained in $l^{2}(X)$.}
\Dem{Let $X$ be a bound and $Y$ any separating object. As $X$ is a bound, $Y$ can be covered by maps $h:U \rightarrow Y$ with $U \subset X$. Using the fact that $\Tcal_{/Y}$ is separated and boolean, we know that it is generated by objects of bounded cardinal and the image of a map whose domain is finite is also finite, and of smaller cardinal, hence $U$ admit a covering by sub-objects with bounded cardinal in $\Tcal_{/Y}$, hence we can freely assume that $U$ is itself of bounded cardinal in $\Tcal_{/Y}$.
One can then define (for each such map $h : U \rightarrow Y$) a map $\phi_h : l^{2}(X) \rightarrow l^{2}(Y)$ by $\phi_h(e_x)=0$ if $x \notin U$ and $\phi_h(e_x) = e_{h(x)}$ if $x \in U$. The fact that the object $U$ is finite with bounded cardinal over $Y$ mean that there exists an (external) integer $n$ such that each fiber of $h$ has cardinal smaller than $n$, this is exactly what we need to know to construct the adjoint of $\phi_h$ and to prove that $\phi_h$ is bounded (and hence extend into an operator).
Now as such maps $h:U \rightarrow Y$ cover $Y$, one has internally in $\Tcal$:
``$\forall y \in Y, \exists h \in H$ such that $y$ is in the image of $h$", where $H$ denote the external set of such map $h:U\rightarrow Y$ which are finite and of bounded cardinal in $\Tcal_{/Y}$. In particular the joint image of all the $\phi_h$ for $h \in H$ contains all the generators of $l^{2}(Y)$ and hence spam a dense subspace of $l^{2}(Y)$, which concludes the proof by proposition \ref{QcontaininHcal}.
}
}
\blockn{We denote by $\Hcal^{red}(\topos)$ the full subcategory of $\Hcal(\topos)$ of objects which are weakly contained in $l^{2}(X)$ for some separating object $X$. Objects of $\Hcal^{red}(\topos)$ are said to be \emph{square-integrable}\footnote{because when $\topos$ is the topos of $G$-sets for some discrete group $G$, then $\Hcal(\Tcal)$ is the category of unitary representations of $G$ while $\Hcal^{red}(\topos)$ is precisely the category of square integrable representations of $G$}.
Results of \cite{henry2014measure} (especially section 7) suggest that the square integrable Hilbert spaces of $\Tcal$ are the one that are clearly related to the geometry of $\topos$.
Because of proposition \ref{l2weaklycontinl2}, for any separating bound $X$ of $\topos$ the Hilbert space $l^{2}(X)$ is a generator of $\Hcal^{red}(\topos)$, hence extending proposition 7.6 of \cite{wstarcat} to monotone complete $C^{*}$-category (or restricting ourselves to topos which are integrable in the sense of \cite[section 3]{henry2014measure}) gives us that when $\Tcal$ is a boolean and locally separated topos, $\Hcal^{red}(\topos)$ is equivalent to the category of reflexive Hilbert modules over $End(l^{2}(X))$. For this reason we can call this algebra ``the" (or ``a") \emph{reduced}\footnote{For example, if $\Tcal$ is the topos of $G$-sets for $G$ a discrete group one obtains the usual (reduced) von Neumann algebra of the group this way.} algebra of $\topos$ (it is unique up to Morita equivalence).
}
\blockn{Finally, The fact that it is possible to define internally the tensor product of two Hilbert spaces yields a symmetric monoidal structure on $\Hcal(\topos)$. Moreover as $l^{2}(X) \otimes l^{2}(Y) \simeq l^{2}(X \times Y)$ one can see that $\Hcal^{red}(\topos)$ is stable by tensor product\footnote{in \ref{tensorbySI} we will actually prove the stronger result that the tensor product of an arbitrary Hilbert space with a square integrable Hilbert space is square integrable}. Hence, $\Hcal^{red}(\topos)$ is also endowed with a symmetric monoidal structure, but without a unit object (unless $\Tcal$ is separated). }
\section{Statement of the main theorems}
\label{secStatement}
\block{\Def{Let $C$ be a monotone complete $C^{*}$-category, $\Ecal$ any topos (not necessary boolean) a representation of $C$ in $\Ecal$ is a $*$-functor $\rho$ from $C$ to $\Hcal(\Ecal)$. It is said to be normal if for any supremum $a = \sup a_i$ of a bounded directed net of positive operator in $C$, $\rho(a_i)$ converge internally in the weak operator topology to $\rho(a)$.}
A representation of $C$ in $\Ecal$ is also the same as a representation of $p^{*} C$ in the category of Hilbert space internally in $\Ecal$ (where $p$ is the geometric morphism from $\Ecal$ to the point). Also if $\Ecal$ is boolean, then $\Hcal(\Ecal)$ is monotone complete and a representation $\rho$ is normal in the sense of this definition if and only if it is normal as a $C*$-functor between $C^{*}$-category. }
\block{When $C$ has additional structure (for example is monoidal) we will by default assume that the representation $\rho$ preserve these structures, for exemple:
\label{DefRepUnred}\Def{If $\Tcal$ is a boolean topos and $\Ecal$ an arbitrary topos, a representation of $\Hcal(\Tcal)$ in $\Ecal$ is a normal representation of the monotone complete $C^{*}$-category $\Hcal(\Tcal)$ in $\Ecal$ such that the underlying $*$-functor is symmetric and monoidal. }}
\block{\Def{We will say that a Hilbert space $H \in \Hcal(\Ecal)$ is inhabited if one has $\exists s \in H$ such that $\Vert s \Vert >0$ internally in $\Ecal$, or equivalently, if $\exists s \in H, \Vert s \Vert =1$ holds internally in $\Ecal$.}
}
\block{\label{DefNondegenUnred}\Def{If $\Tcal$ is a boolean topos and $\Ecal$ an arbitrary topos, a representation $\rho$ of $\Hcal(\Tcal)$ is said to be non-degenerate if for any inhabited object $H$ of $\Hcal(\Tcal)$ the Hilbert space $\rho(H)$ is inhabited in $\Ecal$.}
It is important to notice that detecting whether a representation $\rho$ of $\Hcal(\Tcal)$ is non-degenerate or not can be done completely from the (monoidal) category $\Hcal(\Tcal)$ without knowing the topos $\Tcal$. Indeed, an object $H$ of $\Hcal(\Tcal)$ is inhabited in $\Tcal$ if and only if the functor $H \otimes \_$ is faithful.
As an example of a ``degenerate'' representation, one can consider $\Tcal$ the topos of $G$-sets for some infinite discrete groupe $G$, then $\Hcal(\Tcal)$ is the category of unitary representation of $G$. Let $\rho$ be the representation of $\Hcal(\Tcal)$ into $\Tcal$ defined by:
\[\rho(H) = \overline{ \{ h \in H | h \text{ belongs to a finite dimensional sub-representation } \} } \]
One has $\rho(l^{2}(G))= 0$ so it cannot be non-degenrate. One easily checks that it is a symmetric normal $*$-functor, and it is monoidal because of the following observation:
\Lem{Let $R,R' \in \Hcal(\Tcal)$ be two representations of $G$. Assume that $R \otimes R'$ contains a (non trivial) finite dimensional sub-representation, then both $R$ and $R'$ contains a non trivial finite dimensional sub-representation. }
\Dem{ Let $K \subset R \otimes R'$ a non-trivial finite dimensional sub-representation. let $K^{*}$ the dual of the representation $K$. Because $K$ is finite dimensional and non-trivial $K \otimes K^{*}$ contains a non-zero invariant vector. In particular $R \otimes R' \otimes K^{*}$ contains a non-zero invariant vector which corresponds to a non zero $G$-linear Hilbert-Schmidt $f$ from $R$ to $R'^{*}\otimes K$, $f^{*}f$ is hence a non-zero compact self-adjoint $G$-linear automorphism of $R$ which is hence going to have some non-trivial finite dimensional $G$-stable eigenspaces. This concludes the proof of the lemma.}
}
\blockn{There is in particular a representation of $\Hcal(\Tcal)$ in $\Tcal$ itself called the \emph{tautological} representation and given by the identity functor. This representation is non-degenerate. Moreover if $\rho$ is a (non-degenerate) representation of $\Hcal(\Tcal)$ in a topos $\Ecal$ and $f:\Ecal' \rightarrow \Ecal$ is any geometric morphism then $f^{\sharp}\rho$, given by composing the functor $\rho$ by the functor $f^{\sharp} : \Hcal(\Ecal) \rightarrow \Hcal(\Ecal')$ is a (non-degenerate) representation of $\Hcal(\Tcal)$ in $\Ecal'$. In particular, any geometric morphism from $f:\Ecal \rightarrow \Tcal$ induce a non-degenerate representation $f^{\sharp}$ of $\Hcal(\Tcal)$ in $\Ecal$.}
\block{\label{mainThunred}\Th{Let $\Tcal$ be a boolean locally separated topos, then $\Tcal$ is the classifying topos for non-degenerate representations of $\Hcal(\Tcal)$ and the tautological representation is the universal non-degenerate representation. More precisely, for any topos $\Ecal$ there is an equivalence of category from the category of geometric morphism $\hom(\Ecal,\Tcal)$ to the category\footnote{See \ref{Defmorphofrep} below.} of non-degenerate representations of $\Hcal(\Tcal)$ in $\Ecal$ which associate to any geometric morphism $f$ the representation $f^{\sharp}$. }
In particular, the topos $\Tcal$ is uniquely determined (up to unique isomorphism) from the symmetric monoidal $C^{*}$-category $\Hcal(\Tcal)$ of Hilbert bundles over $\Tcal$.
}
\block{\label{Defmorphofrep}This theorem, as it is stated, does not completely make sense yet because we did not say what are the morphisms of representations of $\Hcal(\Tcal)$ in $\Ecal$. It appears that the good notion of morphisms is the following:
\Def{If $\rho_1$ and $\rho_2$ are two representations of $\Hcal(\Tcal)$ in $\Ecal$ a morphism $v$ from $\rho_1$ to $\rho_2$ is a collection of isometric inclusions $v_X:\rho_1(X) \rightarrow \rho_2(X)$ for each object $X$ of $\Hcal(\Tcal)$ which is natural in $X$ and such that, up to the structural isomorphisms, $v_{\mathbb{C}_{\Tcal}}$ is the identify of $\mathbb{C}_{\Ecal}$ and $v_X \otimes v_Y$ is $v_{X \otimes Y}$. }
Because we do not assume that the $v_X$ have adjoints they are not morphisms in $\Hcal(\Ecal)$ and hence it would not make sense strictly speaking to says they form a symmetric monoidal natural transformation, but this is essentially what this definition means.
\bigskip
It can be proved directly, using the fact that $\Hcal(\Tcal)$ as a generator (as proved in \cite[3.6]{henry2014measure}), that morphisms between two representations $\rho_1$ and $\rho_2$ actually form a set and not a proper class, but we do not need to know that and one can instead obtain this result as a corollary of the theorem.
}
\block{There is also a form of this theorem for the category $\Hcal^{red}(\Tcal)$ instead of $\Hcal(\Tcal)$, which we will use as an intermediate step in the proof of theorem \ref{mainThunred}. The problem to state it directly is that $\Hcal^{red}(\Tcal)$ is not exactly a monoidal category because it does not have a unit object. In particular, one cannot ask monoidal functors to preserve the unit object, but it appears that the condition ``non-degenerate" actually completely replace the preservation of the unit object, and moreover, in this case, the definition of non-degenerate can be weakened:
\label{DefRepReduced}\Def{A non-degenerate representation $\rho$ of $\Hcal^{red}(\topos)$ in $\topos[1]$ is a normal representation of the monotone complete $C^{*}$-category $\Hcal^{red}(\topos)$ in $\Ecal$ such that:
\begin{itemize}
\item $\rho$ satisfies all the axioms of the definition of a symmetric monoidal functor not involving the unit object.
\item There exists an object $H$ of $\Hcal^{red}(\topos)$ such that $\rho(H)$ is inhabited.
\end{itemize}
}
The morphisms of such representations are defined exactly as in \ref{Defmorphofrep}.}
\blockn{Similarly to \ref{mainThunred}, there is a tautological representation of $\Hcal^{red}(\Tcal)$ in $\Tcal$ and it is possible to pullback any representation of $\Hcal^{red}(\Tcal)$ along a geometric morphism }
\block{\label{mainThRed}\Th{The tautological representation of $\Hcal^{red}(\topos)$ in $\topos$ is the universal non-degenerate representation of $\Hcal^{red}(\topos)$. I.e., for any topos $\topos[1]$ the association $p \rightarrow p^{*}$ induce an equivalence of categories between geometric morphisms from $\topos[1]$ to $\topos$ and non-degenerate representations of $\Hcal^{red}(\topos)$ in $\topos[1]$.}
}
\blockn{In the rest of the article, all the representations considered will always be implicitly assumed to be non-degenerate.}
\blockn{These two theorems together, suggest that it should be possible to reconstruct $\Hcal(\Tcal)$ directly from $\Hcal^{red}(\Tcal)$. This indeed seems to be the case, here is a sketches of proof:
We will see in \ref{tensorbySI}, that any $H$ in $\Hcal(\Tcal)$ induce by tensorisation an endofunctor:
\[ M_H:= ( H \otimes \_ ) : \Hcal^{red}(\Tcal) \rightarrow \Hcal^{red}(\Tcal) \]
which satisfies a ``multiplier" condition of the form $M_H(A\otimes B) \simeq M_H(A) \otimes B$ (this isomorphism being functorial and satisfying some coherence conditions). Conversely, if $M$ is an endofunctor of $\Hcal^{red}(\Tcal)$ satisfying the same condition than $M_H$, then for any separating object $X$ of $\Hcal$, $M_H(l^{2}(X))$ is going to be a $l^{2}(X)$-module in the sense of definition \ref{Defl2module}, hence, by proposition \ref{l2Xmodeqtoslice}, is of the form $\bigoplus_{x \in X} H_x$ for a Hilbert space $(H_x)_{x \in X}$ in $\Tcal_{/X}$, but using the coherence condition on the isomorphisms $M(A\otimes B) \simeq M(A) \otimes B$ one should be able to prove that all the $H_x$ are canonically isomorphic and hence (if $X$ is inhabited) that $M(l^{2}(X))$ if of the form $H \otimes l^{2}(X)$. One can then conclude the $M$ is exactly the tensorization by $H$ by assuming that $X$ is a separating bound and hence that $l^{2}(X)$ is a generator of $\Hcal^{red}(\Tcal)$. One also has to check that endomorphisms of $l^{2}(X)$ acts as they should on $M(l^{2}(X)) \simeq H \otimes l^{2}(X)$ but this will follow by the same ``matrix elements" argument as in the proof of \ref{proofcompatibilityRED}.
\bigskip
Hence, $\Hcal(\Tcal)$ should be, in some sense, the ``category of multiplier" of the (non-unital) monoidal category $\Hcal^{red}(\Tcal)$.
\bigskip
We decided not to include a precise form of this last result in the present version of this paper, because its proof, and in fact even its proper formulation, require some work on the precise coherence conditions required on such a ``multiplier" that seems to be out of the scope of the present paper, and we are not sure that such a statement has, despite its elegance, any interesting applications.
}
\section{From representations of $\Hcal^{red}$ to geometric morphisms}
\label{secProofRed}
\renewcommand{\thesubsubsection}{\arabic{section}.\arabic{subsection}.\arabic{subsubsection}}
\blockn{In this section we consider an arbitrary representation $\rho$ of $\Hcal^{red}(\topos)$ in a topos $\topos[1]$, and we will prove that it induces a geometric morphism from $\topos[1]$ to $\topos$. In all this section we will work internally in $\topos[1]$. Hence objects of $\topos[1]$ will be called sets and objects of $\Hcal(\topos[1])$ will simply be called Hilbert spaces, the price of this being to avoid the use of the law of excluded middle and of the axiom of choice. }
\subsection{Construction of the geometric morphism on separating objects}
\label{GeomorphonObjec}
\block{In this subsection, we fix a separating object $X \in \topos$, and we will show that $\rho(l^{2}(X))$ is of the form $l^{2}(Y)$ for a well determined decidable set $Y$. We will denote by $H$ the space $\rho(l^{2}(X))$. This space is endowed with an operator $ \Delta : H \rightarrow H \otimes H$ which comes from the operator in $\Hcal(\topos)$:
\[ \begin{array}{c c c c}
\Delta : & l^{2}(X) &\rightarrow &l^{2}(X) \otimes l^{2}(X) \\
& e_x & \mapsto & e_x \otimes e_x \end{array} \]
$\Delta$ is co-commutative and co-associative and its adjoint $\Delta^{*} : H \otimes H \rightarrow H$ defines a multiplication on $H$ denoted $ a * b :=\Delta^{*}(a \otimes b)$, which is commutative and associative. Moreover $\Delta^{*} \Delta = Id_H$ hence $\Vert a * b \Vert \leqslant \Vert a \Vert \Vert b \Vert$.
In $\topos$, the operator $\Delta^{*}$ is defined by $\Delta^{*}(e_x \otimes e_y)=e_x$ is $x=y$ and $0$ if $x \neq y$, or, tu put it another way, it is just the point-wise multiplication of $X$-indexed sequences.
In $\Ecal$, for $ h \in H$, we denote by $m_h$ the linear map on $H$ defined by $m_h(u)=h * u$. we denote by $\Vert h \Vert_{\infty} = \Vert m_h \Vert$, and the usual norm of $H$ will be denoted by $\Vert h \Vert_2$. One has $\Vert h \Vert_{\infty} \leqslant \Vert h \Vert_{2}$ because of the previous bound on the norm of a product.
Finally, let $C$ be the closure in the space of bounded linear map from $H$ to $H$ of the of algebra formed by the $m_h$ for $h \in H$. It is a commutative algebra. the term ``closure" is taken here in the sense that:
\[ C := \{ f :H \rightarrow H | \forall \epsilon > 0, \exists h \in H, \Vert f - m_h \Vert < \epsilon \} \]
In fact, it should be called ``fiberwise closure" or ``weak closure" because the resulting set is not closed in the sense the complementary of an open, see \cite{henry2014localic} for more details about these notions.
}
\block{\label{X0}Before going further we need to make a few external constructions that will also be usefull in the next subsetions.
Because there exists an object $H$ of $\Hcal(\Tcal)$ such that $\rho(H)$ is inhabited, and that the $l^{2}(X)$ for $X$ separating are generators of $\Hcal(\Tcal)$ there exists a separating object $X_0 \in \Tcal$ such that $\rho(l^{2}(X_0))$ is also inhabited. We fix such an object $X_0$.
A sub-object $U \subset X_0 \times X$ is said to be of degree smaller than $n$ over $X_0$ if internally in $\topos$ for all $x \in X_0$ there is at most $n$ distinct elements $x' \in X$ such that $(x,x') \in U$, i.e. if $U$ is of cardinal smaller than $n$ as an object of $\Tcal_{/X_0}$. It is said to be of finite degree if it is of degree smaller than $n$ for some \emph{external} natural number $n$, i.e. if, as an object of $\Tcal_{/X_0}$ its cardinal is bounded.
Theorem \ref{SepImpFinitness} applied to $\Tcal_{/X_0}$ imply that sub-objects of finite degree over $X_0$ cover $X_0 \times X$. We denote by $F$ the (external) ordered set of sub-objects of $X_0 \times X$ of finite degree. As, if $U$ and $U'$ are of finite degree then $U \vee U'$ is also of finite degree, $F$ is a directed pre-ordered set. }
\block{\label{VUPU}For $U \in F$ let (in $\topos$):
\[ \begin{array}{c c c c}
V_U :& l^{2}(X_0) &\rightarrow & l^{2}(X_0) \otimes l^{2}(X) \\
& e_x &\mapsto & \displaystyle \sum_{x' \text{ s.t. } (x,x')\in U} e_x \otimes e_{x'}
\end{array}
\]
\[ \begin{array}{c c c c}
P_U :& l^{2}(X_0) \otimes l^{2}(X) &\rightarrow & l^{2}(X_0) \otimes l^{2}(X) \\
& e_x \otimes e_{x'} & \mapsto &\displaystyle (\mathbb{I}_{(x,x') \in U } ) e_x \otimes e_{x'}
\end{array}
\]
We also denote by $V_U$ and $P_U$ their images by $\rho$, and we denote $H_0$ the space $\rho(l^{2}(X_0))$.
One easily check that:
\begin{itemize}
\item $V_U$ has an adjoint given by $V_U^{*}(e_x \otimes e_y) =e_x \mathbb{I}_{(x,y) \in U}$
\item $V_U$ is well defined and bounded (if $U$ has degree $n$, then $V_U$ has norm smaller than $\sqrt{n}$).
\item $P_U$ is a $F$-indexed net of projections whose supremum is the identity of $l^{2}(X_0) \otimes l^{2}(X)$, and as $\rho$ is normal it is also the case in $\topos[1]$ that the supremum of the $P_U$ is the identify of $H_0 \otimes H$.
\item $(Id_{H_0} \otimes \Delta^{*}) \circ (V_U \otimes Id_H )= P_U$. Indeed this is easily checked internally in $\topos$ on elements of the form $e_x \otimes e_{x'}$, extended by linearity and continuity and transported to $\topos[1]$ by $\rho$.
\end{itemize}
}
\block{\label{m_injectif}The role of these operators $V_U$ and $P_U$ is essentially to provide ``locally" an approximate unit for the product $*$. Indeed, let us fix an element $h_0 \in H_0$ such that $\Vert h_0\Vert =1$, which is always possible by assumption on $X_0$, then one can define $1_U := (h_0^{*} \otimes Id_H) V_U(h_0)$, and, because of the formula relating $V_U$ and $P_U$ one has after a short computation that for any $h \ \in H$:
\[ h*1_U = (h_0^{*} \otimes Id_{H}) (P_U(h_0 \otimes h)), \]
which (because $P_U$ converge strongly to the identity) converge in norm to $h$. In particular:
\Lem{In $\Ecal$, the association $h \mapsto m_h$ from $H$ to $C$ is injective.}
Indeed, using (internally) our approximate unit $1_U$ one gets that:
\[ h= \lim_{U \in F} h * 1_U =\lim_{U \in F} m_h(1_U). \]
}
\block{We can now prove:
\Lem{The algebra $C$ generated by the $m_h$ is a $C^{*}$-algebra.}
The term $C^{*}$-algebra is taken here in the same sense as, for example, in \cite{banaschewski2000spectral}. I.e. it is complete in the sense of Cauchy approximation or Cauchy filters, and the ``norm" is describe either by the data of the ``rational ball" $B_q$ corresponding morally to the set of $x$ such that $\Vert x \Vert x < q$ of equivalently as a function $\Vert \_ \Vert$ with value into the object of upper semi-continuous real number.
\Dem{ All we have to do is to prove that every element of $C$ have an adjoint which belongs to $C$. As the adjunction preserve the norm and that the space of bounded linear map is complete (in the norm topology) it is enough to prove it for a dense family of elements of $C$.
Let $h$ be an arbitrary element of $H$.
\bigskip
We define:
\[h_U = h * 1_U = (h_0^{*} \otimes Id_{H}) (P_U(h_0 \otimes h)) \in H \]
As mentioned earlier, $(h_U)_{U \in F}$ converges in norm to $h$.
We also define:
\[h'_U = ( h_0^{*} \otimes h^{*} \otimes Id_H) (Id_{H_0} \otimes \Delta) (V_U)(h_0) \]
We will prove that for all $a,b \in H$ one has:
\[ \scal{h'_U * a}{b} = \scal{a}{h_U*b} \]
this will show that $m_{h_U}$ has an adjoint $m_{h'_U}$ which belong to $C$. As for any $h\in H$, $m_h$ is approximated by the $m_{h_U}$ (indeed, $\Vert m_h \Vert = \Vert h \Vert_{\infty} \leqslant \Vert h \Vert _2$ ) the $m_{h_U}$ for $h \in H$ and $U \in F$ are dense in $C$, hence this will conclude the proof.
In order to do so, consider the following two operators $W_1$ and $W_2$ from $H_0 \otimes H \otimes H$ to $H_0 \otimes H$ defined by:
\[ W_1 := (Id_{H_0} \otimes \Delta^{*}) \circ (Id_{H_0} \otimes \Delta^{*} \otimes Id_{H}) \circ (V_U \otimes Id_H \otimes Id_H) \]
\[W_2:= (V_U^{*} \otimes Id_H) \circ (Id_{H_0} \otimes \Delta^{*} \otimes id_H) \circ (Id_{H_0} \otimes Id_H \otimes \Delta) \]
Two short computations show that:
\[ \scal{a}{h_U * b} = \scal{h_0 \otimes a }{W_1(h_0 \otimes h \otimes b)} \]
\[ \scal{h'_U * a}{b} = \scal{h_0 \otimes a }{W_2(h_0 \otimes h \otimes b)} \]
Moreover, both $W_1$ and $W_2$ are image by $\rho$ of operators also denoted $W_1$ and $W_2$ defined by the same formulas in $\Tcal$ (replacing $H_0$ and $H$ by $l^{2}(X_0)$ and $l^{2}(X)$) and one can easily check in $\Tcal$ that $W_1 = W_2$ indeed, if $x \in X_0, y,z \in X$ a short computation shows that:
\[ W_1 (e_x \otimes e_y \otimes e_z) = \mathbb{I}_{(y=z)} \mathbb{I}_{((x,y) \in U)}e_x \otimes e_y = W_2 (e_x \otimes e_y \otimes e_z) \]
which proves that $W_1 = W_2$ in $\Tcal$ and hence also in $\Ecal$ and hence that $\scal{a}{h_U * b} = \scal{h'_U * a}{b}$ which concludes the proof as mentioned above.
}
}
\block{\label{geomofC}One of the key observation is that this algebra $C$ is ``geometrically" attached to the representation $\rho$, in order to make this clear let us denote $C_{\rho}$ instead of $C$ in what follows.
\Lem{Let $p:\Ecal' \rightarrow \Ecal$ be any topos above $\Ecal$. One has a representation $p^{\sharp}\rho$ of $\Hcal(\Tcal)$ in $\Ecal'$, and a canonical isomorphism:
\[ C_{p^{\sharp} \rho} \simeq p^{\sharp} C_{\rho}. \]
}
\Dem{It is clear that $p^{\sharp}\rho$, obtained as the composition of $\rho :\Hcal^{red}(\Tcal) \rightarrow \Hcal(\Ecal)$ with $p^{\sharp} : \Hcal(\Ecal) \rightarrow \Hcal(\Ecal')$, is a representation in the sense of definition \ref{DefRepReduced}. Moreover, the $C^{*}$-algebra $p^{\sharp}(C_{\rho})$ is naturally acting on $p^{\sharp}(H) = p^{\sharp}\rho (l^{2}(X))$ simply because $C_{\rho}$ is acting on $H$. Moreover if $h \in H$, the operator $p^{\sharp}(m_h)$ on $p^{\sharp}(H)$ is equal to the operator $m_{p^{*}(h)}$ (this is clear from the definition of $m_h$).
Hence the image of $p^{\sharp}(C_{\rho})$ in $End(p^{\sharp}(H))$ is generated by operators of the form $m_{h}$ for $h \in p^{*}(H) \subset p^{\sharp}(H)$ but as $p^{*}(H)$ is dense in $p^{\sharp}(H)$ this proves that the image of $p^{\sharp}(C_{\rho})$ in the algebra of endomorphism of $p^{\sharp}(H)$ is indeed the algebra generated by the $m_h$ for $ h \in p^{\sharp}(H)$, i.e. the algebra $C_{p^{\sharp}\rho}$, giving a canonical surjection from $p^{\sharp} C_{\rho}$ to $C_{p^{\sharp}\rho}$. But this surjection is also isometric and hence is an isomorphism because of the relation:
\[ \Vert m_h \Vert = \lim_{n \rightarrow \infty} \Vert h * h* \dots *h \Vert ^{\frac{1}{n}}\]
which is proved in the proof of lemma \ref{LemChascontnorm} below and clearly preserved by $p^{\sharp}$.
}
In particular, if one compute $C$ in the tautological representation of $\Hcal^{red}(\Tcal)$ in $\Tcal$ it is exactly $\Ccal_0(X)$, hence if our representation $\rho$ corresponds to a pullback along some geometric morphism $f$ the previous lemma imply that $C$ will be $f^{\sharp}(\Ccal_0(X)) = \Ccal_0( f^{*}(X))$ and hence using the non-unital Gelfand duality proved in \cite{henry2014nonunital} one has that $f^{*}(X)$ can be reconstructed out of $\rho$ as the spectrum of $C$. In order to show that an arbitrary $\rho$ fo comes from a geometric morphism one needs to prove that the spectrum of $C$ is a discrete decidable locale, and we will do that by showing that $\text{Spec } C$ is locally positive (i.e. that $\text{Spec } C \rightarrow 1$ is an open map) and that the diagonal map $\text{Spec } C \rightarrow \text{Spec } C \times \text{Spec } C$ is both open (hence by \cite[C3.1.15]{sketches} that $\text{Spec } C$ is discrete) and closed (hence that $\text{Spec } C$ is decidable).
}
\block{
\label{LemChascontnorm}\Lem{The norm of any element of $C$ is a continuous real number and $\text{Spec } C$ is locally positive (or ``open", or ``overt").}
\Dem{Let $h \in H$ such that $\Vert h \Vert_2>0$, as $\lim_U m_h (1_U)= h$ this proves that $\Vert m_h \Vert >0$ in the sense that there exists a rational number $q$ such that $0<q \leqslant \Vert m_h\Vert$. Let $h^{(n)}$ denotes the $n$-th power of $h$ for the product $*$. As $m_h$ is an element of a commutative $C^{*}$-algebra one has $\Vert m_h^{n} \Vert = \Vert m_h \Vert^{n}$ hence $\Vert h^{(n)}\Vert_{\infty} = \Vert h \Vert_{\infty}^{n}$.
Moreover:
\[ \Vert h^{(n)} \Vert_2 = \Vert m_h h^{(n-1)} \Vert_2 \leqslant \Vert m_h \Vert \Vert h^{(n-1)} \Vert_2 \leqslant \Vert h \Vert_{\infty} \Vert h^{(n-1)} \Vert_2 \]
Hence by induction on $n$:
\[ \Vert h^{(n+1)} \Vert_2 \leqslant \Vert h \Vert_{\infty}^{n} \Vert h \Vert_2 \]
As $\Vert . \Vert_{\infty} \leqslant \Vert . \Vert_{2}$ one has:
\[ q\Vert h \Vert_{\infty}^{n} \leqslant \Vert h \Vert_{\infty} ^{n+1} \leqslant \Vert h^{(n+1)} \Vert_2 \leqslant \Vert h \Vert_{\infty}^{n} \Vert h \Vert_2 \]
For some rational number $q$ such that $0<q<\Vert h \Vert_{\infty}$. Taking the $n$-th root one obtains that:
\[ q^{\frac{1}{n}} \Vert h \Vert_{\infty} \leqslant \Vert h^{(n+1)} \Vert_2^{\frac{1}{n}} \leqslant \Vert h \Vert_{\infty} \Vert h \Vert_2^{\frac{1}{n}} \leqslant \Vert h \Vert_{\infty} (q')^{\frac{1}{n}} \]
And equivalently that:
\[\frac{1}{q'^{1/n}} \Vert h^{(n+1)} \Vert_2^{\frac{1}{n}} \leqslant \Vert h \Vert_{\infty} \leqslant \frac{1}{q^{1/n}} \Vert h^{(n+1)} \Vert_2^{\frac{1}{n}}\]
As $\Vert h^{(n+1)} \Vert_2^{\frac{1}{n}}$ is bounded, the difference between the upper bound and the lower bound can be made arbitrarily small, hence $\Vert h \Vert_{\infty}$ can be approximated arbitrarily close by continuous real number (and hence also by rational number), which proves that $\Vert h \Vert_{\infty}$ is a continuous real number. One also gets the identity:
\[ \lim_{n \rightarrow \infty} \Vert h^{(n+1)} \Vert_2^{\frac{1}{n}} = \Vert h \Vert_{\infty} \]
which was required in the proof of the previous lemma.
As $0$ and the $h$ such that $\Vert h \Vert_2 >0$ are dense in $H$ (for any $h \in H$ and any $\epsilon>0$ either $h$ is of positive norm or it is of norm smaller than $\epsilon$ and hence is approximated by $0$) and as the $m_h$ for $h \in H$ are dense in $C$, one has obtain a dense familly of elements of $C$ of continuous norm hence every element of $C$ has a continuous norm and by \cite[5.2]{henry2014nonunital} one can conclude that $\text{Spec } C$ is locally positive.
}}
\block{\label{const_e_chi}If one works internally in $\text{Spec } C$, one still has a representation $\rho$ of $\Hcal(\Tcal)$ (by pulling it back from the one in $\Ecal$) and by \ref{geomofC} the $C^{*}$-algebra of endomorphisms of $\rho(l^{2}(X))$ generated by this representation is isomorphic to the pullback of $C$, hence (by \cite[proposition 4.4]{henry2014nonunital} ) we have at our disposal a character $\chi : C \rightarrow \mathbb{C}$ satisfying $\exists c \in C$ such that $|\chi(c)|>0$. We will now examine what can be done with such a character, without necessary assuming that we are working internally in $\text{Spec } C$.
\bigskip
Assume one has a character $\chi$ of the $C^{*}$ algebra $C$ satisfying $\exists c \in C, |\chi(c)|>0$. The map $h \rightarrow \chi(m_h)$ is a bounded linear form on $H$, also denoted $\chi$. Also there exists a $g \in H$ such that $\chi(g)=1$.
Let $e_\chi$ be defined as $(\chi \otimes Id_H)(\Delta(g))$.
Then if $h$ is another element of $H$ one has:
\[ \scal{h}{e_\chi} = (\chi \otimes h^{*})(\Delta(g)) = \chi( (Id_H \otimes h^{*})(\Delta(g))) \]
But, for any $w \in H$ one has:
\[ \scal{w}{(Id_H \otimes h^{*})(\Delta(g))} = (w^{*} \otimes h^{*})(\Delta(g)) = \scal{\Delta^{*}(w \otimes h)}{g} = \scal{m_h w}{g} \]
hence $(Id_H \otimes h^{*})(\Delta(g)) = (m_h)^{*}(g)$ and as $\chi$ is a character of $C$ one obtain that $\chi( (Id_H \otimes h^{*})(\Delta(g))) = \overline{\chi(h)}\chi(g)=\overline{\chi(h)}$. Finally:
\[ \scal{e_\chi}{h} = \chi(h) \]
In particular $e_{\chi}$ does not depend on $g$, but only on $\chi$. This proves:
\Lem{If $\chi$ is a character of $C$ then there exists a unique element $e_\chi \in H$ such that for all $h \in H$ one has $\scal{e_\chi}{h} = \chi(m_h)$.
}
}
\block{\label{e_chi_prop}The vector $e_{\chi}$ attached to a character $\chi$ of $C$ satisfies additional properties:
\Lem{\begin{enumerate}
\item $ \Delta(e_{\chi}) = e_{\chi} \otimes e_{\chi} $
\item $ e_{\chi} * e_{\chi} = e_{\chi} $
\item $ e_{\chi}*h=\chi(h) e_{\chi} $
\item $ \Vert e_{\chi} \Vert_2 = \Vert e_{\chi} \Vert_{\infty} = 1 $
\end{enumerate}}
\Dem{
\begin{enumerate}
\item Let $a,b \in H$, then:
\[\scal{\Delta(e_{\chi})}{a \otimes b} = \scal{e_{\chi}}{a * b} = \chi(a * b)=\chi(a) \chi(b)= \scal{e_{\chi} \otimes e_{\chi}}{a \otimes b} \] which proves the results.
\item One has $e_{\chi} * e_{\chi} = \Delta^{*}(e_{\chi} \otimes e_{\chi}) = \Delta^{*}\Delta(e_{\chi}) = e_{\chi}$ because $\Delta^{*}\Delta =Id$.
\item For any $h,g \in H$ one has:
\[\scal{e_{\chi}*h}{g} = \scal{m_h (e_\chi)}{g} = \scal{e_\chi}{(m_h)^{*} g } =\chi((m_h)^{*} m_g)=\overline{\chi(h)}\chi(g).\]
In particular $\scal{e_{\chi}*h}{g} = \scal{\chi(h)e_{\chi}}{g}$ which concludes the proof.
\item $\scal{e_\chi}{e_\chi} = \chi(e_{\chi})$. But as $e_{\chi}$ is a projector for the product $*$, the value of $\chi(e_{\chi})$ is either $0$ or $1$. The norm $\Vert e_{\chi} \Vert_2$ of $e_{\chi}$ is hence either $0$ or $1$, but it cannot be $0$ (because $\chi$ is non-zero) hence it is one. This implies that $\Vert e_{\chi} \Vert_{\infty} \leqslant 1$ but has $m_{e_{\chi}} e_{\chi} = e_{\chi}$ one has $\Vert e_{\chi} \Vert_{\infty} =1$.
\end{enumerate}
}
}
\block{\label{specC_atomic}\Lem{The diagonal of $\text{Spec } C$ is both open and closed.}
\Dem{Working internally in $\text{Spec } C \times \text{Spec } C$ one has two characters $\chi_1$ and $\chi_2$ which give rise to elements $e_{\chi_1}$ and $e_{\chi_2}$ of (the completion of the pullback of) $H$ as in \ref{const_e_chi}.
\[ \scal{e_{\chi_1}}{e_{\chi_2}} = \chi_1(e_{\chi_2}) \]
Hence, as $e_{\chi_2}$ is a projector, $\scal{e_{\chi_1}}{e_{\chi_2}}$ is either $0$ or $1$. If it is $0$ then $\chi_1 \neq \chi_2$ but if it is one, as $\Vert e_{\chi_1}\Vert_2 =\Vert e_{\chi_2}\Vert_2 =1$ this imply that $\chi_1=\chi_2$. Hence, internally in $\text{Spec } C \times \text{Spec } C$ one has $\chi_1=\chi_2$ or $\chi_1 \neq \chi_2$ which proves that the diagonal of $\text{Spec } C \times \text{Spec } C$ is both open and closed.
}
}
\block{\Prop{There exists a unique (up to unique isomorphism) decidable set $X_{\rho}$ such that $H=l^{2}(X_{\rho})$ and $\Delta$ is the map defined by $\Delta(e_x)=e_x \otimes e_x$ for all $x \in X_{\rho}$. }
\Dem{
The uniqueness of $X_{\rho}$ is clear: any (decidable) object $X$ can be reconstructed out of $l^{2}(X)$ and $\Delta$ as the set of $v \in l^{2}(X)$ such that $\Vert v \Vert = 1$ and $\Delta(v) = v \otimes v$.
As $\text{Spec } C$ is open and its diagonal map is open it is a discrete locale by \cite[C3.1.15]{sketches}, i.e. $\text{Spec } C = X_{\rho}$ for some set $X_{\rho}$. As the diagonal of $\text{Spec } C$ is also closed, $ X_{\rho}$ is decidable. Each $x \in X_{\rho}$ corresponds to a character of $C$ and hence there is a corresponding vector $e_x \in H$ such that $\Delta(e_x)=e_x \otimes e_x$. Moreover the proof of \ref{specC_atomic} show that $(e_x)$ is an orthonormal family in $H$, hence $l^{2}(X_{\rho})$ embeds isometrically in $H$ in a way compatible to $\Delta$.
Let $x \in X_{\rho}$. As an element of $C \simeq \Ccal_0(X_{\rho})$, $m_{e_x}$ is the characteristic function of $\{x\}$ : indeed $\chi_y(m_{e_x}) = \scal{e_y}{e_x}=\delta_{x,y}$. Hence the characteristic function of $\{ x \}$ acts on $H$ by $m_{e_x} h = \scal{e_x}{h} e_x$ and has its image included in $l^{2}(X_{\rho})$. But any element of $C = \Ccal_0(X_{\rho})$ is a norm limit of linear combination of such characteristic function, hence the action of $C$ has its image included in $l^{2}(X_{\rho})$. But has this action contains $m_{1_U}$ which converge to the identity, this implies that $l^{2}(X_{\rho}) = H$ (in a way compatible to $\Delta$ ).
}
}
\subsection{Functoriality on separating object}
\label{secFunctoOnObj}
\block{\label{DefGammaf}Let $X$ and $Y$ be two decidable objects of any topos, and let $f:X \rightarrow Y$ be a map. Then the map $f$ induces an operator:
\[\begin{array}{c c c c}
\Gamma_f :& l^{2}(X) & \rightarrow & l^{2}(X) \otimes l^{2}(Y) \\
& e_x & \mapsto & e_x \otimes e_{f(x)}
\end{array} \]
indeed, it has an adjoint $\Gamma_f^{*}$ which send $e_x \otimes e_y$ to $e_x$ if $y=f(x)$ and to zero otherwise (which make sense constructively because $Y$ is decidable).
}
\block{\label{functorial} If $X$ and $Y$ are two objects of $\Tcal$, we already know (from the previous section) that there exists two objects $X_{\rho}$ and $Y_{\rho}$ of $\topos[1]$ such that there is canonical isomorphism:
\[ l^{2}(X_{\rho}) \simeq \rho(l^{2}(X)) \]
\[ l^{2}(Y_{\rho}) \simeq \rho(l^{2}(Y)) \]
Which are compatible with the canonical maps $\Delta_X:l^{2}(X) \rightarrow l^{2}(X) \otimes l^{2}(X)$ and $\Delta_Y:l^{2}(Y) \rightarrow l^{2}(Y) \otimes l^{2}(Y)$. Moreover $X_{\rho}$ is canonically defined, for example as the set of $v \in \rho(l^{2}(X))$ such that $\Vert v \Vert =1$ and $\Delta(v)= v \otimes v$.
The main result of this section is that:
\Prop{For any map $f:X\rightarrow Y$, there exists a unique map $f_{\rho}: X_{\rho} \rightarrow Y_{\rho}$ such that
\[ \Gamma_{f_{\rho}} = \rho(\Gamma_f) .\]
Moreover this defines a functor from the category of separating objects of $\topos$ to $\topos[1]$.}
The proof of this proposition will be concluded in \ref{proofCtrMorph}
}
\block{\label{functorialityLemma}Before proving the proposition we observe the following lemma:
\Lem{For any $f:X \rightarrow Y$ and $g:Y \rightarrow Z$ two maps between decidable objects of a topos, one has the following identities in the monoidal category of its Hilbert spaces:
\begin{equation} \label{functorialityLemmaEq1}
\Gamma_f ^{*} \Gamma_f = id_{l^{2}(X)}
\end{equation}
\begin{equation} \label{functorialityLemmaEq2} (\Delta_X \otimes id_{l^{2}(Y)} ) \circ \Gamma_f = (id_{l^{2}(X)} \otimes \Gamma_f) \circ \Delta_X
\end{equation}
\begin{equation} \label{functorialityLemmaEq3}
(id_{l^{2}(X)} \otimes \Delta_Y) \circ \Gamma_f = (\Gamma_f \otimes id_{l^{2}(Y)}) \circ \Gamma_f \end{equation}
\begin{equation} \label{functorialityLemmaEq4} \Gamma_{g \circ f} = ({\Gamma_f}^{*} \otimes id_{l^{2}(Z)}) \circ (id_{l^{2}(X)} \otimes \Gamma_g) \circ \Gamma_f
\end{equation}
}
\Dem{Each of these relations is immediately checked internally in the topos on generators.}
}
\block{\label{functorialityProp}\Prop{Let $X$ and $Y$ be two decidable objects of a topos, and let $\Gamma : l^{2}(X) \rightarrow l^{2}(X) \otimes l^{2}(Y)$ be an operator which satisfies equations (\ref{functorialityLemmaEq1}), (\ref{functorialityLemmaEq2}) and (\ref{functorialityLemmaEq3}) of lemma \ref{functorialityLemma}. Then there exists a unique map $f:X \rightarrow Y$ such that $\Gamma= \Gamma_f$. }
\Dem{We work internally in the topos. As we will prove the existence and the uniqueness of the map $f$ it will indeed corresponds to a unique external map. The uniqueness is in fact automatic, if $\Gamma_f= \Gamma_g$ then for all $x$, $e_x \otimes e_{f(x)}= e_x \otimes e_{g(x)}$ which imply that $f(x)=g(x)$. We just have to prove the existence.
\bigskip
We define for $x,x' \in X$ and $y \in Y$:
\[ \gamma_{x' y}^{x} := \scal{\Gamma(e_x)}{ e_{x'}\otimes e_y}.\]
Let also $x,a,b \in X$ and $u,v \in Y$.
On one hand:
\[ \scal{(\Delta_X \otimes id ) \circ \Gamma(e_x)}{e_a \otimes e_b \otimes e_u} = \delta_{a,b} \scal{\Gamma(e_x)}{e_a \otimes e_u} = \delta_{a,b} \gamma^{x}_{a,u} \]
on the other hand:
\[ \scal{(id \otimes \Gamma) \circ \Delta_X(e_x)}{e_a \otimes e_b \otimes e_u}= \scal{e_x \otimes e_x}{e_a \otimes \Gamma^{*}(e_b \otimes e_u)} =\delta_{x,a} \gamma^{x}_{b,u} \]
Hence, as $\Gamma$ satisfies relation (\ref{functorialityLemmaEq2}):
\[ \delta_{x,a} \gamma^{x}_{b,u} = \delta_{a,b} \gamma^{x}_{a,u} \]
taking $x=a$ one obtains:
\[ \gamma^{a}_{b,u} = \delta_{a,b} \gamma^{a}_{a,u} \]
We will denote $\gamma^{a}_u$ for $\gamma^{a}_{a,u}$. One can deduce from this that:
\[ \scal{\Gamma(e_x)}{e_a \otimes e_v} = \gamma^{x}_{a, v} =\delta_{x,a} \gamma^{x}_v= \scal{e_x}{e_a} \gamma^{x}_v \]
As this holds for any $a \in X$ one obtains by linearity and continuity that for any $m \in l^{2}(X)$ one has:
\[ \scal{\Gamma(e_x)}{m \otimes e_v} = \scal{e_x}{m} \gamma^{x}_v \]
We will now use relation (\ref{functorialityLemmaEq3}). On one hand:
\[ \scal{(id \otimes \Delta_Y) \circ \Gamma (e_x)}{e_a \otimes e_u \otimes e_v} = \delta_{u,v} \scal{\Gamma(e_x)}{ e_a \otimes e_u}= \delta_{u,v} \delta_{x,a} \gamma^{x}_{u} \]
On the other hand:
\begin{multline*}\scal{(\Gamma \otimes id) \circ \Gamma(e_x)}{e_a \otimes e_u \otimes e_v} = \scal{\Gamma(e_x)}{\Gamma^{*}(e_a \otimes e_u) \otimes e_v} \\ = \scal{e_x}{\Gamma^{*}(e_a \otimes e_u)} \gamma^{x}_v=\delta_{x,a} \gamma^{x}_u \gamma^{x}_v \end{multline*}
Hence,
\[ \delta_{u,v} \delta_{x,a} \gamma^{x}_{u} = \delta_{x,a} \gamma^{x}_u \gamma^{x}_v \]
Hence (taking $x=a$ and $u=v$) one obtains that $(\gamma^{x}_{u})^{2} = \gamma^{x}_{u}$ hence for all $x,u$ one has $\gamma^{x}_u = 0$ or $\gamma^{x}_u = 1$, also, for all $ u \neq v$ one has $\gamma^{x}_u \gamma^{x}_v=0$ hence for each $x$ there is at most one $v$ such that $\gamma^{x}_v$ is non-zero and hence equal to $1$.
Finally as, by relation (\ref{functorialityLemmaEq1}), $\Gamma^{*}\Gamma =1$, the norm of $\Gamma(e_x)$ is one, there is at least one $u$ such that $\gamma^{x}_u=1$. Hence for each $x \in X$ there exists a unique $u$ such that $\gamma^{x}_u=1$, hence there is a map $f$ defined by $f(x)=u$ and because:
\[ \scal{\Gamma(e_x)}{ e_{x'} \otimes e_y} = \gamma^{x}_{x',y} = \delta_{x,x'} \gamma_y^{x} = \delta_{x,x'} \delta_{f(x),y} \]
one indeed has $\Gamma(e_x)=e_x \otimes e_{f(x)}$ i.e. $\Gamma = \Gamma_{f}$.
}
}
\block{\label{proofCtrMorph}We can now prove proposition \ref{functorial}:
\Dem{As $\rho$ is monoidal, $\rho(\Gamma_f)$ satisfies all the relations of the lemma \ref{functorialityLemma}, hence by proposition \ref{functorialityProp} one obtain a unique map $f_{\rho}$ such that $\rho(\Gamma_f) = \Gamma_{f_{\rho}}$. The functoriality comes immediately from relation (\ref{functorialityLemmaEq4}) of lemma \ref{functorialityLemma} and the fact that as $\rho$ is monoidal it preserves this relation.
}
}
\subsection{Construction of the Geometric morphism}
\label{secCtrGeomMorph}
\blockn{In the previous subsection one has obtain from the representation $\rho$ a functor $F_{\rho}:X \mapsto X_{\rho}$ from the category of separating objects of $\topos$ to the the category of all objects of $\topos[1]$ which in some sense is compatible with $\rho$. As we assume that $\topos$ is locally separated, the category of separating object of $\topos$ endowed with the restriction of the canonical topology of $\topos$ is a site of definition for $\topos$. Moreover, this site has all finite limits (except the empty limit) and arbitrary (small) co-products, hence in order to check that the functor $F_{\rho}:X \mapsto X_{\rho}$ we have constructed extend into the $f^{*}$ part of a geometric morphism from $\topos[1]$ to $\topos$ we essentially\footnote{see the proof of \ref{CtrGeomMorph}.} need to check that it preserves finite limits, arbitrary co-products and send epimorphisms in $\topos$ to epimorphisms in $\topos[1]$.}
\block{\label{Fsurjection1}\Lem{Let $X$ and $Y$ be two separating objects of $\topos$ and $f:X\twoheadrightarrow Y$ an epimorphism such that $X$ has a bounded cardinal in $\Tcal_{/Y}$. Then $F_{\rho}(f)$ is an epimorphism in $\topos[1]$.}
\Dem{As $f$ has finite fibres of bounded cardinal there exists an operator $g:l^{2}(X) \rightarrow l^{2}(Y)$ such that $g(e_x)=e_{f(x)}$ (the norm of this map is smaller than the square root of the bound of the size of the fibres). In particular,
\[ \Gamma_f = (Id \otimes g) \circ \Delta_X \]
Moreover, as $f$ is a surjection, if one defines:
\[ h(e_y) = \frac{1}{n} \sum_{f(x)=y} e_x \]
where $n$ denotes the number of pre-images of $y$ by $f$, then $g \circ h = Id$.
Interpreting this in $\topos[1]$, the identity $\Gamma_f = (Id \otimes g) \circ \Delta_X $ characterize $\rho(g)$ as the map which send $e_x$ to $e_{f_{\rho}(x)}$ and the existence of the map $\rho(h)$ such that $\rho(g) \rho(h) = Id$ show that $f_{\rho}$ is surjective.
}
}
\block{\label{Fsurjection2}\Lem{Let $f:X \twoheadrightarrow Y$ be an arbitrary surjection between two separating objects of $\topos$ then $f_{\rho}$ is also a surjection.}
\Dem{By \ref{SepImpFinitness} applied to $\Tcal_{/Y}$, the object $X$ admit a covering by a net of subobjects $U \subset X$ such that the restriction of $f$ to $U$ has bounded cardinal in $\Tcal_{/Y}$. For each $U$ of this net, let $Y_U$ be the image of $U$ by $f$. From the previous lemma, the image by $F_{\rho}$ of the maps from $U$ to $Y_U$ are all surjective, hence all we have to prove is that the family of $F_{\rho}(Y_U)$ cover $F_{\rho}(Y)$. Now, as in the proof of lemma \ref{Fsurjection1} for each $U \subset X$ the corresponding map $i_U : l^{2}(Y_U) \rightarrow l^{2}(Y)$ is send by $\rho$ on the operator corresponding to the map from $F_{\rho}(Y_U)$ to $F_{\rho}(Y)$. The fact that the $Y_U$ cover $Y$ is translated into the fact that $i_U i_U^{*}$ converge weakly to the identity and hence is transported by $\rho$, which concludes the proof.}
}
\block{\Lem{$F_{\rho}$ preserve arbitrary\footnote{Only those indexed by decidable sets in fact, but we are not working internally in $\Ecal$ any more and we are assuming the law of excluded middle in the base topos, so this does not really matter.} co-products.}
\Dem{One has $U=\coprod_i U_i$, if and only if the $U_i$ are complemented sub-objects of $U$ with $t_i : l^{2}(U_i) \rightarrow l^{2}(U)$ the corresponding map, such that:
\[t_i^{*} t_i =Id_{l^{2}(U_i)} \]
\[t_i^{*} t_j =0 \text{ when $i \neq j$.} \]
\[\sum_i t_i t_i^{*} = Id_{l^{2}(U)} \]
where the infinite sum in the last equality mean the supremum of the the net of finite sum.
All those properties are clearly preserved by $\rho$, hence the coproduct is preserved by $F_\rho$.
}
}
\block{\Lem{$F_\rho$ preserves binary products and equalizers of pair of maps.}
\Dem{ One has $l^{2}(X \times Y) \simeq l^{2}(X) \otimes l^{2}(Y)$ and through this isomorphism, the operator $\Delta$ for $X \times Y$ is the tensor product of those for $X$ and $Y$, and this completely characterize the space $l^{2}(X \times Y)$ and its operator $\Delta$. As $\rho$ is monoidal it preserve this ``characterization" of the product and hence $F_\rho$ preserve binary product.
\bigskip
Let
\[X_0 \overset{i}{\hookrightarrow} X \overset{f,g}{\rightrightarrows} Y\]
be the equalizer of a pair of maps between separating objects of $\topos$. Then for all $x\in X$ one has:
\[\begin{array}{r c l l} \Gamma_f ^{*} \Gamma_g (e_x) &= &e_x &\text{ if $f(x)=g(x)$ } \\ & =& 0 & \text{ otherwise} \end{array} \]
Hence $\Gamma_f ^{*} \Gamma_g$ is a projection and $l^{2}(X_0)$ is isomorphic to $\Gamma_f ^{*} \Gamma_g (l^{2}(X))$. This is also preserved by $\rho$, hence $F_{\rho}$ preserves equalizer.
}
}
\block{\label{CtrGeomMorph}\Prop{There exists a geometric morphism $\chi_{\rho} : \topos[1] \rightarrow \topos$ such that $F_{\rho}$ is the restriction of $\chi_{\rho}^{*}$ to the subcategory of separating objects. }
\Dem{As the subcategory of separating object form a generating familly of $\topos$ it is a site of definition when we equiped it with the canonical topology of $\topos$. Hence, all we have to prove is that $F_{\rho}$ is flat and continuous for the canonical topology of $\topos$. It is continuous because it preserves surjections and arbitrary co-products. We just have to prove that it is flat i.e. that (internally in $\topos[1]$) :
\begin{itemize}
\item $\exists x \in F_{\rho}(X)$ for some objects $X$.
\item If $c_1 \in F_{\rho}(X_1)$ and $c_2 \in F_{\rho}(X_2)$ then there exists an object $X_3$ with two arrow $f_i : X_3 \rightarrow X_i$ and a $c_3 \in F_{\rho}(X_3)$ such that $f_i x_3 = x_i$.
\item If $c \in F_{\rho}(X)$ and there is two arrow $f,g : X \rightrightarrows Y$, such that $f.c = g.c$ then there is an arrow $h:X' \rightarrow X$ and a $c' \in F_{\rho}(X')$ such that $h.c' = c$ and $f.h=g.h$.
\end{itemize}
The second condition follows from the fact that $F_{\rho}$ commute to binary product, and the third from the fact that it commute to equalizer. The first condition is generally a consequence of the fact that the functor preserve the terminal object, but in our situation if $\topos$ is not separated there is no terminal object in our site. Instead, $F_{\rho}(X_0)$ is inhabited because $\rho(l^{2}(X_0))$ contains a vector of strictly positive norm, which concludes the proof.
}
}
\block{Finally, if $\mu : \rho \rightarrow \rho'$ is a morphism in the sense of definition \ref{Defmorphofrep}, then $\mu$ induce a natural transformation between $F_{\rho}$ and $F_{\rho'}$ because for any separating object $X$ of $\Tcal$, an $x \in \rho(l^{2}(X))$ such that $\Delta(x) = x \otimes x$ and $\Vert x \Vert =1$ will be send by $\mu$ on an $x$ with the same properties in $\rho'(l^{2}(X))$ hence inducing a map from $F_{\rho}(X)$ to $F_{\rho'}(X)$ and this is natural in $X$. And a natural transformation between the $F_{\rho}$ extend uniquely into a natural transformation between the induced geometric morphism
hence we have constructed a functor from (non-degenerate) representation of $\Hcal(\Tcal)$ in $\Ecal$ to geometric morphisms from $\Ecal$ to $\Tcal$.
}
\subsection{Proof of the ``reduced" theorem \ref{mainThRed} }
\label{ProofmainThred}
\block{If we start from a geometric morphism $\chi: \topos[1] \rightarrow \topos$ then we associate to it a representation $\rho=\chi^{\sharp}$, defined on all of $\Hcal(\topos)$ and we can restrict it to $\Hcal^{red}(\topos)$.
As for any object $X$ of $\topos$, one has $\chi^{\sharp}(l^{2}(X)) \simeq l^{2}(\chi^{*}(X))$, and for any map $f$ one has $\chi^{\sharp}(\Gamma_f)= \Gamma_{\chi^{*} f}$ (up to te canonical isomorphism), it is clear that the geometric morphism reconstructed from this representation $\rho$ will be isomorphic to $\chi$ and that there will be a bijection between morphisms of representations and natural transformation of geometric morphisms. So all we have to do to finish the proof of the theorem is the to prove the following: }
\block{\label{proofcompatibilityRED}\Prop{For any reprensation $\rho$ of $\Hcal^{red}(\topos)$ in a topos $\topos[1]$, $\rho$ is equivalent to the representation $\chi_{\rho}^{\sharp}$ induced by the geometric morphism $\chi_{\rho}:\topos[1] \rightarrow \topos$ constructed out of $\rho$. }
\newcommand{\chi_{\rho}^{\sharp}}{\chi_{\rho}^{\sharp}}
\Dem{We will first prove that $\rho$ and $\chi_{\rho}^{\sharp}$ are equivalent on the full-subcategory of Hilbert spaces of the form $l^{2}(X)$ for $X$ a separating object of $\topos$. The geometric morphism $\chi_{\rho}$ has been constructed to full-fill the relation $l^{2}(\chi^{*}(X)) \simeq \rho(l^{2}(X))$ hence one has indeed an isomorphism $\rho(l^{2}(X)) \simeq \chi_{\rho}^{\sharp}(l^{2}(X))$ for any separating object $X$ of $\topos$. Also, by construction, these isomorphisms are compatible with the maps $\Delta_X$ and $\Gamma_f$ for any separating object $X$ of $\topos$ and any maps $f:Y \rightarrow X$.
\bigskip
Moreover, if $S \subset X$, then $\rho$ and $\chi_{\rho}^{\sharp}$ also agree on the projection $P$ on $l^{2}(X)$ corresponding to the multiplication by the characteristic function of $S$. As any operator of multiplication by a bounded function on $X$ is a supremum of linear combination of such projection, $\rho$ and $\chi_{\rho}^{\sharp}$ also agree on operators on $l^{2}(X)$ defined by multiplication by complex bounded functions on $X$.
\bigskip
Let now $f :l^{2}(X) \rightarrow l^{2}(Y)$ be any bounded operator in any topos. Let $m$ the function of matrix element of $f$ on $X \times Y$, ie internally, $m(x,y):=\scal{e_y}{f(e_x)}$. The operator $M_m$ of multiplication by $m$ on $l^{2}(X) \otimes l^{2}(Y)$ is characterized by the relation :
\[ M_m = ( id_{l^{2}(X)} \otimes \Delta_Y^{*}) \circ (id_{l^{2}(X)} \otimes f \otimes id_{l^{2}(Y)}) \circ (\Delta_X \otimes id_{l^{2}(Y)}) \]
which is easily checked internally on generators.
In particular, if $X$ and $Y$ are two separating objects of $\topos$ and $f :l^{2}(X) \rightarrow l^{2}(Y)$ is any bounded operator, then the operator of multiplication by matrix elements of $\rho(f)$ is $\rho(M_m)$ because of the previous relation is preserved by $\rho$. Hence as $\rho(M_m)= \chi_{\rho}^{\sharp}(M_m)$ one can deduce that $\rho(f)$ and $\chi_{\rho}^{\sharp}(f)$ have the same matrix elements and hence are equals.
\bigskip
As $l^{2}(X)$ for $X$ a separating bound is a generator of $\Hcal^{red}(\Tcal)$,this is enough to prove that the two $*$-functors $\rho$ and $\chi_{\rho}^{\sharp}$ are isomorphic on the whole category $\Hcal^{red}(\Tcal)$ by the result mentioned in section $2$.
}
}
\section{On the category $\Hcal(\Tcal)$ and its representations}
\label{secProofUnred}
\subsection{The category $\Hcal(\Tcal_{/X})$}
\block{In this subsection, $\Tcal$ is an arbitrary topos, and $X$ is a decidable object of $\topos$. As previously, this defines an object $l^{2}(X)$ of $\Hcal(\Tcal)$ which is endowed with a multiplication $\Delta^{*} : l^{2}(X) \otimes l^{2}(X) \rightarrow l^{2}(X) $.}
\block{If $\Hcal$ is a Hilbert space of $\Tcal_{/X}$, then, internally in $\Tcal$ it corresponds to a $X$-indexed family of Hilbert spaces $(\Hcal_x)_{x \in X}$. As $X$ is decidable one can construct the orthogonal sum of such a family:
\[ \Sigma \Hcal := \oplus_{x \in X} \Hcal_{x} \]
which is a Hilbert space of $\Tcal$, and which comes endowed with an ``action" of $l^{2}(X)$:
\[ \begin{array}{c c c c}
M: &l^{2}(X) \otimes \Sigma \Hcal &\rightarrow & \Sigma\Hcal \\
& e_y \otimes (h_x)_{x \in X} & \mapsto & (\delta_{x,y} h_x)_{x \in X}
\end{array}
\]
which admit an adjoint $M^{*}((h_x)_{x \in X}) = \sum_{a}\left( e_a \otimes (\delta_x^{a} h_x)_{x \in X} \right)$, hence is indeed a morphism in $\Hcal(\Tcal)$ and satisfies $MM^{*}=Id_{\Sigma \Hcal}$ (and it is indeed an action because it is the multiplication of a sequence of vector by a sequence of scalar).
Finally, these structures satisfy the following additional relation:
\[M^{*}M = (id_{\Sigma \Hcal} \otimes \Delta^{*}) \circ (M^{*}\otimes id_{l^{2}(X)}) \]
i.e. the following diagram commute:
\[
\begin{tikzcd}[ampersand replacement=\&,scale=2]
\Sigma \Hcal \otimes l^{2}(X) \arrow{rr}{M^{*} \otimes id_{l^{2}(X)}} \arrow{d}{M} \& \& \Sigma \Hcal \otimes l^{2}(X) \otimes l^{2}(X) \arrow{d}{id_{\Sigma \Hcal} \otimes \Delta^{*}} \\
\Sigma \Hcal \arrow{rr}{M^{*}} \&\& \Sigma \Hcal \otimes l^{2}(X)\\
\end{tikzcd}
\]
Or to put it another way, $M^{*}$ is $l^{2}(X)$-linear when $\Sigma \Hcal \otimes l^{2}(X)$ is endowed with the action of $l^{2}(X)$ on the second component of the tensor product.
This is also easily checked internally on an element of the form $(h_x) \otimes e_y$.
}
\block{\label{Defl2module}\Def{A $l^{2}(X)$-module is an object $H \in \Hcal(\Tcal)$ endowed with an action $M: H\otimes l^{2}(X) \rightarrow H$ of the monoide $(l^{2}(X),\Delta^{*})$, which satisfies the two additional relation: $MM^{*}=id_H$ and $M^{*}M = (id_{H} \otimes \Delta^{*}) \circ (M^{*}\otimes id_{l^{2}(X)})$.
A morphism of $l^{2}(X)$-module is a morphism of module in the usual sense.
}
We already mentioned our main example: $\Sigma \Hcal$ is a $l^{2}(X)$-module, and one can check that if $(f_x)$ is a globally bounded family of operators $f_x:\Hcal_x \rightarrow \Hcal'_x$, then $\Sigma f$ is a morphism of $l^{2}(X)$-modules from $\Sigma \Hcal$ to $\Sigma \Hcal'$.}
\block{\label{l2Xmodeqtoslice}
\Prop{The following three categories are equivalent:
\begin{itemize}
\item The category of $l^{2}(X)$-modules, as in definition \ref{Defl2module}.
\item The category of non-degenerate representations of $C_0(X)$. I.e. the category of Hilbert spaces $H \in \Hcal(\Tcal)$ endowed with an action of $C_0(X)$ making them, internally in $\Tcal$, a non-degenerate\footnote{Non-degenerate meaning here that the image of the action map $C_0(X) \times H \rightarrow H$ spam a dense subspace of $H$.} $*$-representation of $C_0(X)$.
\item $\Hcal(\Tcal_{/X})$
\end{itemize}
}
\Dem{The equivalence between the second and the third category is proved internally and is extremely classical (and is constructive if we assume $X$ decidable): any non-degenerate representation $H$ of $C_0(X)$ decompose into a direct sum of $H_x = e_x(H)$ where $e_x$ is the characteristic function of $\{x\}$.
Suppose we start with a $l^{2}(X)$-module $H$. It is in particular a Hilbert space endowed with an action of the algebra $(l^{2}(X),*)$. As we mentioned, the axiom $M^{*}M = (id_{H} \otimes \Delta^{*}) \circ (M^{*}\otimes id_{l^{2}(X)})$ mean that the map $M^{*}: H \rightarrow H \otimes l^{2}(X) $ is $l^{2}(X)$-linear when we endow $H \otimes l^{2}(X)$ with the action of $l^{2}(X)$ on the second component only, and hence, $M$ and $M^{*}$ exhibit $H$ as a $l^{2}(X)$-linear retract of $H \otimes l^{2}(X)$. But the action of $l^{2}(X)$ on $H \otimes l^{2}(X)$ clearly extend into a non-degenerate representation of $C_0(X)$ and hence (as $M^{*}M$ is also $C_0(X)$-linear) the action of $l^{2}(X)$ on $H$ also extend into a non-degenerate representation of $C_0(X)$.
Conversely, by the equivalence of the second and the third categories, any $C_0(X)$ module is up to unique isomorphism of the form $\Sigma \Hcal$ and hence comes from a $l^{2}(X)$-module. This already proves the equivalence at the level of objects, but, as $l^{2}(X)$ is internally dense in $C_0(X)$ for the uniform norm, a map is $C_0(X)$-linear if and only if it is $l^{2}(X)$-linear and hence this concludes the proof.
}
}
\block{\label{pullbackl2module}We conclude this subsection by investigating how, if $f :X \rightarrow Y$ is any map in $\Tcal$, the ``pullback along $f$" functor from $\Hcal(\Tcal_{/Y})$ to $\Hcal(\Tcal_{/X})$ can be expressed in terms of $l^{2}(Y)$ and $l^{2}(X)$-modules and the operator $\Gamma_f : l^{2}(X) \rightarrow l^{2}(X) \otimes l^{2}(Y)$ defined in \ref{DefGammaf}.
Let $V$ be an $l^{2}(Y)$-module, consider the following map:
\[ P_{V,f} : V \otimes l^{2}(X) \overset{\Gamma_f}{\rightarrow} V \otimes l^{2}(X) \otimes l^{2}(Y) \overset{M_V}{\rightarrow} V \otimes l^{2}(X) \]
where the first map is $\Gamma_f$ acting on the second component and the second map is the structural multiplication of $l^{2}(Y)$ on $V$ after re-ordering of the terms.
\Prop{\begin{itemize}
\item $P_{V,f}$ is a $l^{2}(X)$-linear orthogonal projector of $V \otimes l^{2}(X)$. In particular, its image $P_{V,f}(V \otimes l^{2}(X) )$ is a $l^{2}(X)$-module.
\item If, as a $l^{2}(Y)$-module, one has $\displaystyle V = \sum_{y \in Y} V_y$, then:
\[ P_{V,f}(V \otimes l^{2}(X)) \simeq \sum_{x \in X} V_{f(x)}\]
as a $l^{2}(X)$-module.
\item If $h:V \rightarrow V'$ is a map of $l^{2}(Y)$-module, then $(h \otimes Id_{l^{2}(X)})$ commutes to the $P_{V,f}$ and induce the pullback of $h$ along $f$ between $P_{V,f}(V \otimes l^{2}(X))$ and $P_{V',f}(V' \otimes l^{2}(X) )$.
\end{itemize}
}
\Dem{Everything can be checked internally on generators, indeed:
\[ P_{V,f} ( v \otimes e_x) = v.e_{f(x)} \otimes e_x \]
where the point denotes the action of $l^{2}(Y)$ on $V$, hence one immediately obtains that $P_{V,f} P_{V,f} (v \otimes e_x ) = v.e_{f(x)} \otimes e_x = P_{V,f}(v \otimes e_x)$ and that for any $x' \in X$:
\[ P_{V,f}(v \otimes e_x) .e_{x'} = v.e_{f(x)} \otimes e_x.e_x' = P_{V,f}(v \otimes e_x) \delta_{x,x'} = P_{V,f}(v \otimes e_x.e_x'). \]
This concludes the first point, for the second, if $V = \sum_{y \in Y} V_y$ and if $v \in V_y$, then $P_{V,f}(v\otimes e_x)$ is $(v \otimes e_x) \delta_{f(x),y}$, hence $P_{V,f}(V \otimes l^{2}(X)$ is the subspace generated by the $(v \otimes e_x)$ for $v \in V_{f(x)}$ which is isomorphic to $\sum_{x \in X} V_{f(x)}$ (they have the same generators with the same scalar product between them).
For the third point, $h(v \otimes e_x) = h(v) \otimes e_x$ and $h$ is $l^{2}(Y)$ linear, hence:
\[ P_{V',f}(h(v) \otimes e_x) = h(v).e_{f(x)} \otimes e_x = h(v.e_{f(x)}) \otimes e_x = h(P_{V,f}(v \otimes e_x)) \]
And the action of $h$ on $P_{V,f}(v \otimes e_x)$ send $v \otimes e_x$ for $v \in V_{f(x)}$ to $h(v) \otimes e_x$ with $h(v) \in V'_{f(x)}$ hence it is indeed the pullback of $h$ along $f$.
}
}
\subsection{Tensorisation by square integrable Hilbert space}
\block{\label{LemmaSepimpSI}\Lem{Let $\Tcal$ be a separated boolean topos, then every Hilbert spaces of $\Tcal$ is square integrable.}
\Dem{Let $\Hcal$ be a Hilbert space of $\Tcal$. As $\Tcal$ is separated, theorem \ref{SepImpFinitness} implies it is generated by objects of (finite) bounded cardinal hence $\Hcal$ can be covered by maps of the form $S \rightarrow \Hcal$ with $S$ of bounded cardinal. Such a map from $S$ to $\Hcal$ can be extended by linearity into a globally bounded map from $l^{2}(S)$ to $\Hcal$ (whose norm will be smaller than $\sqrt{n}$ for any $n$ such that $|S| \leqslant n $). Hence $\Hcal$ is covered by maps from square integrable Hilbert spaces and hence is itself square integrable.}
}
\block{\label{tensorbySI}\Prop{Let $\Tcal$ be a boolean locally separated topos, $\Hcal$ a $\Tcal$-Hilbert space and $\Hcal_i$ a square integrable $\topos$-Hilbert space. Then $\Hcal \otimes \Hcal_i$ is square integrable.}
\Dem{$\Hcal_i$ being square integrable it can be covered by map from $l^{2}(X)$ with $X$ separating, hence $\Hcal \otimes \Hcal_i$ can be covered by maps of the form $\Hcal \otimes l^{2}(X)$ for $X$ separating. Hence it is enough to prove the result in the case $\Hcal_i = l^{2}(X)$ for $X$ a separating object.
Let $p:\Tcal_{/X} \rightarrow \Tcal$ the canonical map. Internally in $\Tcal$, a $\Tcal_{/X}$ Hilbert space is just a $X$-indexed family of Hilbert spaces, $p^{*}(\Hcal)$ is the constant family equal to $\Hcal$, and $l^{2}(X) \otimes \Hcal$ is the orthogonal sum of this family.
$\Tcal_{/X}$ being separated, $p^{*}(\Hcal)$ is a square integrable Hilbert space in $\Tcal_{/X}$ by lemma \ref{LemmaSepimpSI}, hence it admit a covering by globally bounded maps $l^{2}(Y_i) \rightarrow p^{*}(\Hcal)$ for $Y_i \rightarrow X$ objects of $\Tcal_{/X}$. Taking the orthogonal sum along $X$ internally in $\Tcal$ one obtains a series of maps $l^{2}(Y_i)\rightarrow \Hcal \otimes l^{2}(X)$ which also form a covering, and hence $\Hcal \otimes l^{2}(X)$ is square integrable.
}
}
\subsection{Proof of the ``unreduced" theorem \ref{mainThunred}.}
\block{Let first $\rho$ be a representation of $\Hcal(\Tcal)$ in $\Ecal$ in the sense of definition \ref{DefRepUnred}. Let $\rho^{red}$ be the restriction of $\rho$ to the full subcategory $\Hcal^{red}(\Tcal) \subset \Hcal(\Tcal)$. Clearly, $\rho^{red}$ is a representation of $\Hcal^{red}(\Tcal)$ in the sense of definition \ref{DefRepReduced}.
In particular, by theorem \ref{mainThRed}, which has been proved in \ref{ProofmainThred}, there exists a geometric morphism $f:\Ecal \rightarrow \Tcal$ such that $\rho^{red}$ is isomorphic to $f^{\sharp}$ (as symmetric monoidal $C^{*}$-functor) on $\Hcal^{red}(\Tcal)$.
\Prop{The isomorphism between $\rho^{red}$ and $f^{\sharp}$ on $\Hcal^{red}(\Tcal)$ extend canonically\footnote{In fact, uniquely as the characterization of the isomorphism given during the proof will show.} into an isomorphism of symmetric monoidal $C^{*}$-functor between $\rho$ and $f^{\sharp}$ on $\Hcal^{red}(\Tcal)$.}
\Dem{Let $H$ be any object of $\Hcal(\Tcal)$, let $X$ be any inhabited separating object of $\Tcal$ and let $H_0 = l^{2}(X) \in \Hcal^{red}(\Tcal)$. One already knows that $l^{2}(f^{*}(X)) \simeq f^{\sharp}(H_0) \simeq \rho(H_0)$. By proposition \ref{tensorbySI}, the $\Tcal$-Hilbert space $H \otimes H_0$ is in $\Hcal^{red}(\Tcal)$, hence one has an isomorphism $\rho(H \otimes H_0) \simeq f^{\sharp}(H \otimes H_0)$, and even an isomorphism $\rho(H) \otimes l^{2}(f^{*}X) \simeq f^{\sharp}(H) \otimes l^{2}(f^{*}X))$, moreover, as the isomorphism between $\rho$ and $f^{\sharp}$ on $\Hcal^{red}$ is functorial and symmetric monoidal, this last isomorphism is an isomorphism of $l^{2}(f^{*}X)$-module in the sense of definition \ref{Defl2module}. In particular, internally in $\Ecal$ it comes from a family $(\lambda_x)_{x \in f^{*}X}$ of isomorphisms between $\rho(H)$ and $f^{\sharp}(H)$.
If now $X$ and $Y$ are two inhabited separating objects of $\Tcal$ and $g:X\rightarrow Y$ any map. The previous construction gives us (internally in $\Ecal$) two families of isomorphisms $(\lambda_x)_{x \in f^{*}X}$ and $(\mu_y)_{y \in f^{*}Y}$ between $\rho(H)$ and $f^{\sharp}(H)$.
The description given in \ref{pullbackl2module} of pullbacks of $l^{2}(Y)$-modules (and their morphisms) into $l^{2}(X)$-modules (and their morphisms) along $g$ is preserved by $\rho$ and $f^{\sharp}$ just because they are monoidal symmetric, hence one obtains in $\Hcal(\Tcal)$ exactly the relation asserting the that the family of isomorphisms $(\lambda_x)$ is the pullback of the family $(\mu_y)$, i.e. for all $x\in X$, $\lambda_x = \mu_f(x)$.
Applying this observation to the two projections $X \times X \rightarrow X$ one obtain that for all $x,x' \in X \times X$ , $\lambda_x = \lambda_{x,x'} = \lambda_x'$. Hence the isomorphism $\lambda_x$ does not depend on $x$ and as $X$ is inhabited this proves (in $\Ecal$) that there is a canonical isomorphism between $\rho(H)$ and $f^{\sharp}(H)$ and hence this gives an external isomorphism $\mu_H : \rho(H) \rightarrow f^{\sharp}(H)$ (which does not depend on $X$ neither). This isomorphism is entirely characterized by the fact that for some (or any) inhabited separating object $X$ of $\Tcal$ the isomorphism $f^{\sharp}(H) \otimes l^{2}(f^{*}X) \simeq f^{\sharp}(H \otimes l^{2}(X)) \simeq \rho(H \otimes l^{2}(X)) \simeq \rho(H) \otimes l^{2}(f^{*}X) $ is equal to $\mu_H \otimes Id_{l^{2}(f^{*}X)}$.
The fact that this $\mu_H$ is indeed a monoidal symmetric natural transformation then follow simply from this characterization, and the fact that the isomorphism between $f^{\sharp}$ and $\rho$ on square integrable Hilbert spaces is itself functorial and symmetric monoidal.
}
}
\block{At this point we have proved that the two constructions forming the equivalence in theorem \ref{mainThunred} are inverse of each other at the level of object (up to canonical isomorphism). To completely conclude the proof of this theorem one need to check that the notion of morphisms on the two side are the same.
let $\phi: \rho \rightarrow \rho'$ be an isometric inclusion of symmetric monoidal representation (in the sense of definition \ref{Defmorphofrep} ), then $\phi$ restricted to $\Hcal^{red}(\Tcal)$ is a morphism between $\rho^{red}$ and $\rho'^{red}$ and hence induce a natural transformation between the corresponding geometric morphism $f$ and $f'$ that we will denote by $F(\phi)$. Conversely, any natural transformation $\mu$ between $f$ and $f'$ will induce an inclusion of $f^{\sharp} \simeq \rho$ into $f'^{\sharp} \simeq \rho'$ which we will denote by $G(\mu)$. It is immediate that $F (G(\mu)) = \mu$ (up to the canonical isomorphism). And it is already proved that the two morphisms of representations $G(F(\phi))$ and $\phi$ agree when restricted to $\Hcal^{red}(\Tcal)$. The following proposition allows to conclude.
\Prop{If $\phi$ and $\phi'$ are two morphisms of representations from $\rho$ to $\rho'$ (two non-degenerate representations of $\Hcal(\Tcal)$ in $\Ecal$ ) which agree on objects of $\Hcal^{red}(\Tcal)$ then $\rho = \rho'$.}
\Dem{Let $H$ be any object of $\Hcal(\Tcal)$ and $H'$ be any object of $\Hcal^{red}(\Tcal)$ whose image by $\rho$ in $\Hcal(\Ecal)$ contains a vector of norm $1$.
Then as $\phi$ and $\phi'$ are monoidal one has $\phi_H \otimes \phi_{H'} = \phi_{H \otimes H'}$ and similarly for $\phi'$, but as $H'$ and $H \otimes H'$ are square integrable (by assumption for the first one and by proposition \ref{tensorbySI} for the second) $\phi$ and $\phi'$ are the same on them and hence $\phi_H \otimes \phi_{H'} = \phi'_H \otimes \phi_{H'}$. Then, internally in $\Ecal$, one can take some $h \in \rho(H')$ of norm one, and composing by the corresponding maps from $\mathbb{C}$ to $\rho(H')$ and from $\rho'(H')$ to $\mathbb{C}$ one obtains that internally $\phi_H = \phi'_H$ which immediately show that the same equality hold externally and concludes the proof of the proposition, and of theorem \ref{mainThunred}.
}
}
\section{Toward a generalized Gelfand duality ?}
\label{secToward}
\blockn{The goal of this last section is to informally discus the possibility to extend the results of this paper into some sort of ``non-commutative Gelfand duality" that would relates geometric (or topological) objects with objects from the world of operators algebras.}
\blockn{A pre-requisite to understand what follows, is to understand how toposes can be fully faithfully embedded into a category of localic stacks (in fact ``geometric localic stacks", hence localic groupoids) as clearly explained in \cite{bunge1990descent}. Roughly: To any topos $\Tcal$ one can associate the stacks defined by $\hat{\Tcal}(\Lcal) = \hom(\Lcal,\Tcal)$ where $\Lcal$ is any locale\footnote{In what follows we identify a locale with its topos of sheaves, or equivalently, we actually mean ``localic topos" when we write ``locale".} and where non-invertible natural transformations in $\hom(\Lcal,\Tcal)$ have been dropped. This defines a fully faithful embeddings of the $2$-category of toposes into the $2$-category of geometric\footnote{that is the stacks that can be represented by a groupoid.} stacks on the category of locales endowed with the open surjection topology. This embeddings is the right adjoint of the functor which send any localic groupoid to the category of equivariant sheaves over it. }
\blockn{We will first explain how the two construction involved in the present ``duality" can be defined at a very high level of generality, for example the category $\Hcal(\Tcal)$ is defined for any topos $\Tcal$ without any assumption on it, but we can do even better.}
\blockn{If we start from an arbitrary monoidal symmetric $C^{*}$-category $C$ one can try to define its ``spectrum" $\text{Spec } C$ as the ``classifying topos" of symmetric monoidal $*$-representations of $C$. This $\text{Spec } C$ has in general no reason to exist as a topos, but it always makes sense as a prestack over the category of toposes or of locales: }
\blockn{\Def{Let $C$ be a symmetric monoidal $C^{*}$-category, then $\text{Spec } C$ is the pre-stack over the category of locales defined by: $(\text{Spec } C) ( \Lcal)$ is the groupoid of symmetric monoidal representations of $C$ in $\Lcal$ and symmetric monoidal $*$-isomorphisms between them, and for any morphism of locale $f:\Lcal \rightarrow \Lcal'$, $(\text{Spec } C)(f)$ is just $f^{\sharp}$.
}
}
\blockn{Defined this way, $\text{Spec } C$ is not a stack for a reasonable topology one the category of locales (except the topology of etale surjection). The problem is that Hilbert spaces, because they are not pulled-back as sheaves, does not satisfy reasonable descent properties. Fortunately, we have presented the solution to this problem in \cite{henry2014localic} : we proved that (because Hilbert space are actually pulled-back as locales) localic Hilbert spaces do descend along open surjections (in fact along any morphism which is of effective descent for locales), and form in fact the stackification of the pre-stack of Hilbert spaces.
Hence we have to change the definition of $\text{Spec } C$ by replacing $\Hcal(\Lcal)$ by the the category of localic Hilbert space of $\Lcal$ as defined in \cite{henry2014localic}, and hence $\text{Spec } C$ classifies representations of $C$ on localic Hilbert spaces, and with this definition $\text{Spec } C$ will be a stack for the topology of open surjection (as well as for any topology composed of effective descent morphisms).
}
\blockn{Still thanks to results of \cite{henry2014localic}, one can define a stacks ``$\Hcal$" of localic Hilbert spaces. One can even prove that it is a geometric stacks : its diagonal is a localic map because of \cite[section 3.5]{henry2014localic} and it admit an open surjection from the locale classifying the pre-Hilbert structure on an infinite dimensional $\mathbb{Q}$-vector space.
Moreover, for any localic stacks $\Gcal$ one can define a symmetric monoidal $C^{*}$-category $\Hcal(\Gcal)$ whose objects are morphisms of stacks from $\Gcal$ to $\Hcal$ (hence, essentially continuous fields of Hilbert spaces over $\Gcal$).
}
\blockn{At this point, and up to some set theoretical issues that we do not want to talk about here, one obtains two functors ``$\text{Spec }$" and ``$\Hcal$" which essentially form an adjoint pair between localic stacks (for example for the topology of open surjections) and symmetric monoidal $C^{*}$-category.}
\blockn{That this adjoint pair that we think can provide this generalized duality. It is not reasonable to think that every localic stacks can be reconstructed out of its category of Hilbert spaces: For example, let $\text{Set}[\mathbb{O}^{+}]$ be the classifying topos for the theory of inhabited objects (i.e. the topos of preshaves over the opposite of the category of finite non-empty sets) then one can see that $\Hcal(\text{Set}[\mathbb{O}^{+}])$ is equivalent to the monoidal $C^{*}$-category of Hilbert spaces.
It is not reasonable neither to think that any localic stacks can actually be reconstructed exactly as $\Gcal \simeq \text{Spec } \Hcal \Gcal$. As we have seen in the present paper, one needs to restrict the notion of representations considered in the definition of $\text{Spec } C$ (in our case we only considered ``non-degenerate normal" representations). }
\blockn{But for example, one can easily prove using the same kind of techniques as in section \ref{secFunctoOnObj} and \ref{secCtrGeomMorph} that for any locally decidable topos $\Tcal$ (that is a topos such that every object can be covered by decidable objects), the stack associated to $\Tcal$ is a sub-stacks of $\text{Spec } \Hcal(\Tcal)$. Hence for any locally decidable topos $\Tcal$ there exists a notion of ``good" representation of $\Hcal(\Tcal)$ such that $\Tcal$ is the classifying topos for these ``good" representations.
It opens several questions (which we don't claim are completely precise):
\begin{itemize}
\item Can we axiomatize this notion of ``good" representation ?
\item Can this axiomatization be formulated entirely in the language of symmetric monoidal $C^{*}$-category, or do we need some additional structures on $\Hcal(\Tcal)$ ?
\item Can we characterize the $C^{*}$-categories (possibly with these additional structures) that arise as $\Hcal(\Gcal)$ for some localic stacks $\Gcal$ ?
\item Can we deduce from this a duality between a sub-class of nice ``analytic" localic stacks and certain symmetric monoidal $C^{*}$-categories (possibly with additional structures ?)
\end{itemize}
}
\blockn{As mentioned, the usual Gelfand duality, the duality between commutative monotone complete $C^{*}$-algebras and boolean locales, and the Doplicher-Roberts reconstruction theorem of \cite{doplicher1989new}, can be formulated as special cases of a duality of this form. }
\blockn{We conclude by a few suggestions on the kind of ``additional strucutres" or properties of the $C^{*}$-categories of the form $\Hcal(\Gcal)$ that seems to play an important role.
\begin{itemize}
\item If $a_i$ is a net of operators between two objects $A$ and $B$ of $\Hcal(\Gcal)$ then one can define what it mean for $a_i$ to converge weakly,strongly or strongly-$*$ locally (or ``internally"). In the monotone complete case, normal representations are exactly those who preserve these notions of ``weak" convergence.
\item If $C$ is a small (possibly non-unital) $C^{*}$-category, $F :C \rightarrow \Hcal(\Gcal)$ is any $*$-functor then there is a canonical extension of $F$ from the $C^{*}$-category of $C$-Hilbert modules to $\Hcal(\Gcal)$, defined locally (``internally") as the tensorization of a Hilbert module by the representation ``$F$". This construction seems to be a $C^{*}$-categorical analogue of the classical notions of left/right Kan extension and of weighted (co)-limits, but it does not seems that there is a nice universal property characterizing this notion purely in terms of $C^{*}$-category theory\footnote{But one can give such a universal characterization if one allows the use of the weak convergence mentioned in the first point}. Moreover, the representations classified by $\text{Spec } C$ should preserve these weighted (co)limits.
Somehow categories like $\Hcal(\Gcal)$ seems to deserve the name of ``complete" $C^{*}$-category, and normal functor can also be defined as the functors preserving those weighted limits.
\item Every object of $\Hcal(\Gcal)$ has a dual object. Unfortunately, when the object is not locally of finite dimension, there is no clear characterization of what is this dual object in terms of monoidal $C^{*}$-category. Those ``dual" objects seems to play an extremely important role: they are completely central in the proof of the Doplicher-Roberts reconstruction theorem, and the net of operators $V_u$ (and its dual $V^{*}_U$) used in section \ref{GeomorphonObjec} seem to be a sort of asymptotical replacement for the evaluation and co-evaluation map identifying $l^{2}(X)$ with its own dual and this was a key ingredient in the proof of our theorem.
\end{itemize}
}
\blockn{Hence understanding how we can define what is a ``complete $C^{*}$-category" in the general case and what is the relation of this notion with the notion of ``weak convergence" (these questions are already completely understood in the case of $W^{*}$-categories, see \cite{wstarcat}), and finally understanding what can be says about ``dual objects" in this context seems to be questions that needs to be answered to be able to extend the technique of the present article to general (locally decidable) toposes, and maybe after that to localic groupoids.
Somehow, by working with boolan locally separated toposes in the present paper we did exactly what was needed to avoid those questions: Booleanness ensure that we will only work with monotone complete $C^{*}$-categories, and hence that the question of completeness and normal functors are well understood, and the locale separation hypothesis allows to obtain this net of operators $(V_U)$ which ``rigidify" the notion of dual object at least in $\Hcal^{red}$.}
|
\section{Results}
\vspace{0.5cm}
\noindent\textbf{An abstract gradient response problem.} A one-dimensional array of cells $i$ located at positions ($0<x_i<L$) is exposed to a noisy linear gradient of an input morphogen $c(x)$ spanning the range $[0,c_\mathrm{max}]$. To build intuition, we will assume the noise of input $c(x)$ to be Gaussian, of constant magnitude $\sigma_0$, and uncorrelated between cells\footnote{The assumption of uncorrelated noise is intentionally strong. In a real system, correlated noise can be introduced, for example, by variations in the total amount of morphogen deposited maternally. These fluctuations, which cannot be reduced by averaging, lead to imperfect \emph{reproducibility} of morphogen activity at a given location across multiple embryos. Much work has focused on investigating the limitations imposed on patterning by this type of fluctuations~\cite{dubuis_13,tkacik_14,petkova_14}. In contrast, our model is applicable for understanding the effects of imperfect \emph{precision} of gene expression (at a given location within the same embryo). The distinction between ``raw'' and ``accessible'' information does not rely on the assumption of uncorrelated noise.}: $c(x_i)\equiv c_i = (x_i/L)\,c_\mathrm{max} + \sigma_i,$ where $\sigma_i$ are i.i.d., drawn from a Gaussian of width $\sigma_0$ (Fig.~\ref{fig:argument}A). Cells respond to morphogen $c(x)$ by modulating gene expression through intrinsically noise-prone signal transduction and transcription/translation processes. We will model this response as a composition of three steps, three elementary operations that constitute the ``toolkit'' with which cells can access and process information contained in patterning cues: \emph{access}, \emph{amplify}, and \emph{average}.
Let $g^{\mathrm{out}}$ be a gene product whose expression is controlled by $c(x)$. The simplest readout is achieved by placing gene $g^{\mathrm{out}}$ under the control of a promoter that is responsive to $c$ and by accumulating the output protein for some time $\tau$. In our model, we express the amount of $g^{\mathrm{out}}$ produced during this time by a cell $i$ as $g_i^\mathrm{out} = F(c^\mathrm{est})$, where $c^\mathrm{est}_i$ is a noisy estimate of the true concentration $c_i$ that the system could obtain in time $\tau$ (``access''), and $F$ is some deterministic input-output function (``amplify''); for simplicity, we first consider $F$ to be pure linear amplification with coefficient $\lambda$, denoted $F_\lambda$. The ``access'' operation is the key element of our framework. Specifically, we write
$$
c^\mathrm{est}_i=c_i+\eta_i,
$$
where $\eta_i$ reflects the intrinsic stochasticity of transcription and, in principle, many other noise sources. Here we will model $\eta_i$ simply as being drawn from a Gaussian distribution of width $\eta_0$. In other words, we postulate that each ``access'' operation takes time $\tau$ and comes at the price of corrupting the signal with extra noise of magnitude $\eta_0$.
The final toolkit operation is averaging. Because patterning systems typically act over durations that are long (hours) compared to the time required to synthesize mRNA and protein (minutes), cells can perform temporal averaging by allowing stable gene products to accumulate~\cite{little_13}: if $T$ is the time available for patterning, the system can effectively perform $T/\tau$ access operations. In addition, the production of soluble factors that can be shared between cells gives rise to spatial averaging~\cite{tenWolde_09,little_13}. Both types of averaging offer the system some capacity to perform multiple measurements of the input, which we capture formally by an averaging operator $G_{N_\mathrm{eff}}$. Here $N_\mathrm{eff}$ indicates the effective number of independent measurements, so that application of $G_{N_\mathrm{eff}}$ to a morphogen, by definition, reduces expression fluctuations by a factor $1/N_\mathrm{eff}$.
We distinguish between two patterning strategies. In the first (``direct strategy''; Fig.~\ref{fig:argument}A), cell-fate-specific target genes are controlled directly by $c$ and no other patterning factors are involved. Any available averaging mechanisms are applied to $c$ itself. In the second (``two-tier'') strategy, cells perform an amplifying readout of $c$ with input-output function $F_\lambda$ to establish a spatial profile of a second factor $c^{(\lambda)}$ (Fig.~\ref{fig:argument}B). The pattering time $T$ is spent on accumulating and averaging $c^{(\lambda)}$. Mathematically, in the two scenarios, the cell-fate-specific target genes are controlled by:
\begin{align}
c^{(0)}&= G_{N_\mathrm{eff}}[c] &&\text{(direct strategy)}\label{eq:DirectStrategy}\\
c^{(\lambda)}&= G_{N_\mathrm{eff}}[F_\lambda(c+\eta)]&&\text{(two-tier strategy)}\label{eq:TwoTierStrategy}
\end{align}
We now ask: when, if ever, does the noisy amplification step of the two-tier strategy provide a benefit to the system?
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{Figure2.pdf}
\caption{\textbf{The two patterning strategies.} \textbf{A:} In the direct strategy, target genes are controlled directly by $c$. \textbf{B:} The two-tier strategy involves a second patterning factor $c^{(\lambda)}$; target genes are separated from the input by two tiers of ``access'' operations. Left, raw information content. Right, accessible information content.\label{fig:argument}}
\end{figure}
\newpage
\noindent\textbf{Standard information-theoretic considerations do not explain the benefits of amplification.}
The positional information carried by a linear morphogen $c(x)$ with dynamic range $c_\mathrm{max}$ and noise $\sigma_0$, which we call the ``raw information content'' of a gene expression profile, is given by
$$
I_\mathrm{raw}[c(x),x] = \ln\left(\frac{c_\mathrm{max}}{\sigma_0\sqrt{2\pi e}}\right)
$$
(see Supplementary Information). It depends only on the ratio $\phi=c_\mathrm{max}/\sigma_o$; for convenience, we define $\mathcal I(\phi) \equiv \ln\left(\frac{\phi}{\sqrt{2\pi e}}\right)$, which is an increasing function of $\phi$.
Let us compare the two patterning strategies from the point of view of the raw information content carried by the controlling signal. In the direct strategy~\eqref{eq:DirectStrategy}, the application of $G_{N_\mathrm{eff}}$ reduces the input noise to $\sigma_o/\sqrt{N_\mathrm{eff}}$ and so the controlling signal $c^{(0)}$ carries $I_\mathrm{raw}^{(0)}=\mathcal I\left(\frac{c_\mathrm{max}}{\sigma_0/\sqrt{N_\mathrm{eff}}}\right)$ bits of raw information. In the two-tier strategy~\eqref{eq:TwoTierStrategy}, the amplified profile $c^{(\lambda)}$ is characterized by noise
$
\xi_\lambda=\lambda \sqrt {\frac{\sigma_0^2+\eta_0^2}{N_\mathrm{eff}}},
$
and its raw information content is therefore
\begin{equation}\label{eq:amplifiedInfo}
I^{(\lambda)}_\mathrm{raw}=\mathcal I\left(\frac{\lambda c_\mathrm{max}}{\xi_\lambda}\right)=\mathcal I\left(c_\mathrm{max}\sqrt{\frac{N_\mathrm{eff}}{\sigma_0^2+\eta_0^2}}\right) < I_\mathrm{raw}^{(0)}.
\end{equation}
Averaging mitigates the loss of positional information when using a noisy readout~\cite{tenWolde_09}. If $N_\mathrm{eff}$ is sufficiently large, the amplified and averaged profile carries even more information than the original input. (Note that the information processing inequality is not violated, as it states only that the output cannot carry more information than $N_\mathrm{eff}$ independent copies of the input.) Nevertheless, applying averaging directly to the input (the direct strategy) always yields more raw information; thus, the multi-step scenario appears inferior to a direct readout.
In real systems, the three operations we treat as independent may be mechanistically linked. For example, if $c(x)$ is an intracellular factor while spatial averaging requires a small diffusible molecule, then performing an extra readout can provide access to an otherwise unavailable averaging mechanism. By assuming that the two strategies~\eqref{eq:DirectStrategy} and~\eqref{eq:TwoTierStrategy} can benefit from equal amounts of averaging, which in our model simply reduces expression noise and is obviously beneficial, we can focus specifically on the effect of signal amplification. Multi-tier patterning proceeds through rounds of amplification: small differences in input result in large differences in gene expression so as to establish increasingly sharp boundaries delimiting expression domains~\cite{gilbert}, yet in our expression~\eqref{eq:amplifiedInfo} for the information content of the amplified profile $c^{(\lambda)}$, the amplification factor $\lambda$ cancels out. Thus, considerations based on raw information content fail to explain the prevalence of signal amplification.
\vspace{0.5cm}
\noindent\textbf{The benefits of the multi-tiered strategy lie in making the ``raw'' information more accessible.}
The benefits of amplification and the advantages of the multi-tier strategy become clear when we observe that, due to the intrinsic noise in the regulatory readout, the raw information content is an inadequate measure of a morphogen's usefulness to the system. The purpose of a morphogen is to activate downstream processes; the relevant quantity is therefore not the amount of information a morphogen carries, but the amount of information it can transmit to its downstream targets. Since biological control is intrinsically noisy, the two quantities are distinct.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{Figure3.pdf}
\caption{\textbf{A:} Noisy amplification can increase accessible information even if raw information is reduced. Inner error bars are the signal variability and increase when amplification adds new noise, reducing $I_\mathrm{raw}$. Outer error bars represent the signal observed by the noisy cell machinery (corrupted by noise $\eta_0$). After amplification, the relative importance of $\eta_0$ is reduced, increasing $I_\mathrm{acc}$. \textbf{B:} The ``segmentation'' input-output function $F^z_\lambda$ for integer $\lambda$ (here $\lambda=3$) preserves the dynamic range of morphogen concentration. Locations such as those indicated by dots now have identical expression levels of $z^{(\lambda)}$ (the $y$ axis), but can be distinguished using the input morphogen $c$ (the $x$ axis on this plot).\label{fig:cartoon}}
\end{figure}
Our model was designed to make this particularly clear: since the system can never access the true concentration $c$, but only a noisy estimate $c^{\mathrm{est}}$, $I_\mathrm{raw}[c]$ is beyond the system's reach. We define \emph{accessible} information in a morphogen $I_\mathrm{acc}$ as the amount of information the system can access in time $\tau$:
\begin{equation}\label{eq:defAcc}
I_\mathrm{acc}[c]\equiv I_\mathrm{raw}[c^{\mathrm{est}}] = I_\mathrm{raw}[c+\eta],
\end{equation}
where $\eta$, again, is a Gaussian noise of magnitude $\eta_0$ within our model.
The amount of accessible information provided by the direct strategy (Fig.~\ref{fig:argument}B) is given by
\begin{equation}\label{eq:accNull}
I_\mathrm{acc}^{(0)}= \mathcal I\left(\frac{c_\mathrm{max}}{\sqrt{\frac{\sigma_0^2}{N_\mathrm{eff}}+\eta_0^2}}\right)
\end{equation}
whereas for the amplified profile $c^{(\lambda)}$ it is
\begin{equation}\label{eq:accAmpl}
I_\mathrm{acc}^{(\lambda)}=\mathcal I\left(\frac{\lambda c_\mathrm{max}}{\sqrt {\lambda^2 \frac{\sigma_0^2+\eta_0^2}{N_\mathrm{eff}}+\eta_0^2}}\right)
=\mathcal I \left(\frac{c_\mathrm{max}}{\sqrt {\frac{\sigma_0^2+\eta_0^2}{N_\mathrm{eff}}+\frac{\eta_0^2}{\lambda^2}}}\right).
\end{equation}
The amplification factor $\lambda$ no longer cancels out in ~\eqref{eq:accAmpl}; amplifying dynamic range is beneficial, since it reduces the relative importance of the intrinsic readout noise (Fig.~\ref{fig:cartoon}A). Comparing~\eqref{eq:accNull} and~\eqref{eq:accAmpl}, we find that the extra tier of noisy amplification is beneficial if and only if
\begin{equation}\label{eq:condition}
\eta_0^2\left(1-\frac 1{N_\mathrm{eff}}-\frac 1{\lambda^2}\right)>0
\end{equation}
Note that the condition~\eqref{eq:condition} is never satisfied if $N_\mathrm{eff}=1$ (no averaging) or $\lambda=1$ (no amplification). Intuitively, our argument demonstrates that the patterning system is a mechanism that invests some effort into making a careful measurement ($N_\mathrm{eff}>1$) and encodes this information in a more accessible format where steeper concentration changes ($\lambda>1$) can be interpreted with a faster, and therefore noisier readout. This mechanism is useful precisely because regulatory readout is intrinsically noisy, otherwise direct readout would have been the better strategy. In other words, to understand the purpose of the patterning system, it is essential to distinguish between the total information in a morphogen and information that can be usefully extracted and interpreted.
\vspace{0.5cm}
\noindent\textbf{Multiple tiers improve gradient interpretation even when raw information decreases.}
So far we considered the information content (raw or accessible) in each tier separately. However, in principle, downstream processes could access all patterning cues and not simply the final tier~\cite{tierIntegrate1,tierIntegrate2}. As a result, extra readout tiers can be beneficial even when they carry very little information on their own.
To see this, consider the input-output function $F^z_\lambda$ depicted in Fig.~\ref{fig:cartoon}B. In some respects, it is more realistic than the purely amplifying linear readout $F_\lambda$ considered above, since real patterning systems must operate within a limited global dynamic range of morphogen concentrations. Let $z^{(\lambda)}$ be the morphogen profile established by the new $F^z_\lambda$-shaped readout of $c$; it has noise magnitude $\xi_\lambda$ (same as the noise in $c^{(\lambda)}$), but is folded onto itself $\lambda$ times, reminiscent of the spatially reiterated expression of genes involved in \emph{Drosophila} axis segmentation. Repeatedly using the same output values at multiple positions naturally reduces mutual information between the output concentration and position:
$$
\begin{aligned}
I_\mathrm{raw}[z^{(\lambda)}]&=I_\mathrm{raw}[c^{(\lambda)}]-\ln\lambda\\
I_\mathrm{acc}[z^{(\lambda)}]&=I_\mathrm{acc}[c^{(\lambda)}]-\ln\lambda.
\end{aligned}
$$
However, the $\lambda$ locations with identical concentrations of $z^{(\lambda)}$ are made distinguishable by the original morphogen~$c$ (Fig.~\ref{fig:cartoon}B). Therefore, the \emph{joint} information that the original and the amplified profiles together provide about a cell's location is the same for $F^z_\lambda$ as it was for $F_\lambda$:
$$
I\big[\{c,z^{(\lambda)}\}, x\big]=I\big[\{c,c^{(\lambda)}\}, x\big]
$$
Replacing information content of a single profile by this joint information, our argument demonstrating that amplification increases accessible information can now be repeated verbatim~\cite{noteJointAcc}, and we again find that the extra readout is beneficial as long as~\eqref{eq:condition} is satisfied. Note, however, that on its own, $z^{(\lambda)}$ may carry \emph{less} information than the original morphogen $c$. The easiest way to see this is to compare their noise levels:
$$
\left(\frac{\xi_\lambda}{\sigma_0}\right)^2 = \frac{\lambda^2}{N_\mathrm{eff}}\left(1+\frac{\eta_0^2}{\sigma_0^2}\right)
$$
If the effect of amplification is stronger than that of averaging, we find $\xi_\lambda/\sigma_0>1$. In this scenario, the amplified profile $z^{(\lambda)}$ has the same dynamic range but lower precision than the original morphogen $c$, and therefore, on its own, carries less information (whether raw or accessible). This shows that evaluating the usefulness of a particular cue from information-theoretic standpoint can lead to misleading results, unless all other relevant cues (which are often hard to establish) are taken into account simultaneously. Here, we demonstrated that systems can benefit from multi-tiered interpretation even in cases where intermediate steps occur at a net loss of information, increasing noise.
\vspace{0.5cm}
\noindent\textbf{The multi-tier structure of \textit{Drosophila} segment patterning increases information accessibility.}
In this system, segmentation of the AP axis proceeds through four tiers of gene activity, termed maternal gradients, gap genes, pair-rule genes, and segment polarity genes~\cite{kornberg_93}. The sequential activity of each tier subdivides the naive blastoderm into smaller domains of gene expression with increasingly sharp boundaries, culminating in the designation of each row of cells with its own unique set of expressed genes (Fig.~\ref{fig:data}A). This process is subject to transcriptional noise with a large intrinsic component~\cite{little_13}, as well as several other noise sources with different signatures~\cite{tkacik_08,krivega_2012,kwak_2013,maheshri_2007}. No single value of $\eta_0$ adequately characterizes such readout noise. Nevertheless, we can gain important insight by computing $I_\mathrm{acc}^{\eta_0}[c]$ as a function of $\eta_0$, treating it as a variable parameter: the decay of $I_\mathrm{acc}^{\eta_0}[c]$ with $\eta_0$ characterizes the tolerance to added noise of the information encoded in the morphogen (or set of morphogens) $c$. Applied to gene expression data from the early \emph{Drosophila} segmentation gene network, this analysis will show how our simple model explains the use of multi-tier gradient interpretation in a real system (Fig. 4).
\begin{figure*}[t!]
\centering
\includegraphics[width=\textwidth]{Figure4.pdf}
\caption{\textbf{A:} Immunostaining of three antero-posterior (AP) axis patterning genes in the same embryo. Rather than specifying cell fate directly, the ``gap genes'' such as \emph{hunchback} (Hb; top) and \emph{Kr\"uppel} (Kr; middle) control ``pair-rule'' genes such as \emph{even-skipped} (Eve, bottom). Both tiers regulate other genes further downstream. Boxes indicate the selected region of interest (ROI), where at this time, Hb and Kr are the only relevant inputs to Eve, as shown on the cartoon. \textbf{B:} Within the ROI (shaded), Eve exhibits higher expression noise than either Hb or Kr. Expression noise computed as RMS difference between expression level of a nucleus and its immediate dorsal or ventral neighbor (see Methods), plotted against AP distance from the Hb/Kr boundary (denoted $x_0$). Error bars are standard deviation over $N=8$ embryos. \textbf{C:} Idealized morphogen profiles, restricted to the ROI. Profile shape obtained as smooth spline-fit to expression values and noise magnitudes calculated for the profiles of panel A after projection onto the AP axis. \textbf{D:} For all but the lowest readout noise magnitude, joint accessible information content in the triplet (Hb,Kr,Eve) exceeds the accessible information provided by Hb and Kr alone, even in an extreme hypothetical case when they are rendered entirely noiseless.\label{fig:data}}
\end{figure*}
We focus on a particular node in this network whereby, in early embryos, two gap genes, \textit{hb} and \textit{Kr}, regulate a pair-rule gene \textit{eve}. For $0.37<x_\mathrm{AP}<0.47$, where \textit{Kr} and \textit{hb} expression form opposing boundaries, they are jointly responsible for creating the trough between \textit{eve} stripes 2 and 3; other inputs to \textit{eve} are negligible in this region at this time~\cite{kraut_91,small_96}. Protein levels are measured simultaneously in each nucleus by a triple immunostaining experiment (Fig.~\ref{fig:data}A) in $N=8$ single embryos. We determine the expression noise of each gene by comparing levels in a given nucleus with those of its immediate dorsal and ventral neighbors (see Methods).
In the defined region of interest, \textit{eve} expression noise is higher than the respective noise in \textit{hb} or \textit{Kr} expression (Fig.~\ref{fig:data}B). The information content of \textit{eve} must therefore be lower than that carried by either of its two inputs. Due to the curvature of the embryo (Fig.~\ref{fig:data}A), the positional information of a real morphogen is only approximately related to that derived from projection onto the imaginary AP axis. Therefore, to estimate the information content for each of the three genes, we consider ``idealized'' Gaussian-noise profiles (panel C) with mean and noise obtained by smoothing the measured values in real embryos. The idealized profiles are normalized to the same maximum and are, by construction, functions of $x_\mathrm{AP}$ carrying positional information $I(c(x_\mathrm{AP}), x_\mathrm{AP})$. Restricted to the region of interest, the information content of Hb and Kr is respectively 2.6 and 2.7 bits, whereas the larger noise of Eve reduces its information content to only 2.0 bits. Why, then, does the system use Eve to regulate downstream processes, rather than utilizing Kr and Hb directly?
The answer becomes clear when we consider the accessibility of information encoded in these morphogens, namely $I_\mathrm{acc}^{\eta_0}$ as a function of $\eta_0$ (panel D). A patterning strategy lacking Eve can access only Hb and Kr. Even if some hypothetical filtering mechanism could reduce their expression noise to arbitrarily low level, the readout noise magnitude $\eta_0>0$ imposes an upper bound that $I_\mathrm{acc}^{\eta_0}[c_\mathrm{Hb}, c_\mathrm{Kr}]$ must satisfy. This corresponds to the information in a hypothetical pair of noiseless Hb and Kr and cannot be achieved in practice; it is a theoretical best-case scenario for any strategy lacking Eve.
When the readout noise $\eta_0$ is zero, $I_\mathrm{acc}^{\eta_0}$ coincides with the raw information content, which for perfectly noiseless Hb and Kr would be infinite. However, as readout noise increases, the performance bound becomes finite and drops quickly (black curve). This behavior contrasts with the joint accessible information of the triplet (Hb,Kr,Eve) (magenta) as calculated using the actual measured noise of each of the three profiles. The accessible information content in the triplet is, of course, always finite, but it is also more tolerant to readout noise: due to the steeper slopes of the Eve profile, as $\eta_0$ increases, the accessible information content of the triplet (Hb,Kr,Eve) decreases slowly; importantly, more slowly than the black curve. Therefore, a crossing point is observed, whose presence does not qualitatively depend on the specifics of the readout noise model (e.g.\ absolute noise magnitude can be replaced by fractional). Remarkably, although Eve is measurably noisier than either of its inputs, its presence enables the system to access more information than could have been extracted from Hb and Kr alone, even if these inputs could be rendered perfectly noiseless. In practice, the enhancers of the pair-rule genes also contain binding sites for maternal transcription factors~\cite{tierIntegrate1,tierIntegrate2}, which may lead to a further increase in the precision of gene expression. However, our framework demonstrates that even if Eve were regulated by Hb and Kr only, and so were fully redundant in the standard information-theoretic sense, the additional tier would still confer an advantage, because transcription is intrinsically noisy.
\section{Discussion}
The \emph{Drosophila} patterning network has been described as performing a ``transition from analog to digital specification'' of cell identity~\cite{gilbert}. The ``digital'' metaphor has its limitations: even for Eve, the graded distribution within gene expression domains contains information~\cite{dubuis_13}; nevertheless, it expresses the correct intuition that the final pattern is more tolerant to noise. Importantly, the standard information-theoretic formalism does not capture this intuition: for instance, the profile depicted in Fig.~\ref{fig:cartoon}B has the \emph{same} information content for all $\lambda$. Noise tolerance --- a critically important feature in biological systems --- becomes manifest only when the readout process is considered explicitly, for example, as we have done in our definition of accessible information. This point is implicit in the theoretical work investigating the so-called ``input noise''~\cite{tkacik_08}, but has not been emphasized. This is because in a theoretical discussion of an abstract biochemical circuit, the quantities for which information is computed are easily postulated to be the complete input and the final output; in this manner, valid theoretical results can be derived without a concern for information accessibility (for some recent examples, see~\cite{bowsher_13,deronde_14}). However, when information-theoretic arguments are applied to experimental data where the measured quantity is only an intermediate step, e.g. a transcription factor regulating downstream events, the question of information accessibility can no longer be neglected.
For example, it has been suggested that certain signaling circuits may have evolved towards optimal information transmission~\cite{tkacikBialek_14,levchenko_14}. Although the argument is plausible, applying it in practice requires caution. Consider, once again, the example of a developmental circuit. If the entire set of functional (cell-fate specific) genes were to be included into consideration, then information transmission from the input to this entire layer of functional genes would be a plausible objective function for this whole network to maximize, under some ``bounded complexity'' constraint penalizing solutions where hundreds of cell-fate specific genes are all controlled by highly complex enhancers with combinatorial, cooperative regulation. However, the usual, more economical approach does not consider the full set of hundreds of cell-fate determining genes. Instead, it recognizes that the bulk of the patterning task is accomplished by a small subset of dedicated genes that engage in complex cross-regulation to establish the pattern that all other genes can then interpret simply. If we focus only on this core subset, the ``economy of complexity'' constraint is conveniently imposed by construction. We must realize, however, that maximizing information transmission to the target genes (downstream of the patterning core) imposes a different requirement onto this core circuit than merely efficient information transfer within the core itself. Instead, the core circuit must function as a format converter, re-encoding information at its input into a format that can be accessed with a simpler and faster readout, that of a patterning cue by a functional gene.
Curiously, it has been shown that in small networks with a realistic model of noise, maximizing raw information transmission leads to network structures exhibiting features such as tiling of patterned range with amplifying input/output readouts~\cite{optimFlow1,optimFlow2,optimFlow3}, i.e.\ features that tend to also make information more accessible, even though the optimization scheme employed in these studies did not specifically consider the encoding format. This remarkable coincidence, however, should not obscure the fact that ultimately the two tasks --- maximizing information transmission and re-encoding it in a more accessible format --- could be conflicting.
Information theory is a powerful tool; its formalism does not, however, aim to replace considerations of what constitutes useful information or how it might be used by the system. As it is gaining popularity in biological applications, it is important to remember that for a channel $X\mapsto Y$, the relation between mutual information $I(X,Y)$ and the ability to use $Y$ to determine $X$ is only asymptotic: Shannon~\cite{shannon_48} proved that it is the maximum rate of error-free communication via this channel, \emph{in the limit of infinite uses} of the channel. Importantly, in development and biological signaling, the number of channel uses (e.g. integration time of the signal) is fundamentally finite~\cite{bowsher_14}. Further, Shannon's results assumed an encoder/decoder of infinite computational power~\cite{shannon_48}. This asymptotic rate is never in fact achieved in practice~\cite{macKay}, but in biological context, performance is constrained even further, since the ``encoding scheme'' is usually limited to measuring the same signal multiple times. In communication theory, this bears the name of ``repetition code'' and is formally classified as a ``bad code'', i.e.\ a code that does not attain Shannon's bound even asymptotically. This means that extracting all the ``raw'' information from a signal is impossible even in principle. For example, a signaling pathway with capacity of 1 bit is never sufficient to make a reliable binary decision~\cite{bowsher_14}, and therefore should not be conceptualized as a binary switch.
As illustrated here, making the distinction between ``raw'' and ``accessible'' information will be crucial for understanding the architecture and function of patterning and signaling circuits. More work is required: our definition of accessible information relied on a simplistic noise model; in general, quantifying the usefulness of information-bearing signals in contexts where channel uses are limited will require reinstating considerations of rate/fidelity tradeoff, which Shannon could eliminate by taking the limit of infinite-time communication. Nevertheless, information theory remains a most adequate framework to address these issues, provided it is extended to quantify both the amount and accessibility of information. Our work provides a step in this direction and demonstrates how the extended framework naturally explains a global architectural property shared by diverse patterning circuits.
\begin{acknowledgments}
We thank Ariel Amir, William Bialek, Michael Brenner, Chase Broedersz, Ted Cox, Paul Francois, Anders Hansen, Ben Machta, Gasper Tkacik, Eric Wieschaus and Ned Wingreen for helpful discussions and comments on the manuscript. This work was supported by NIH grants P50 GM071508 and R01 GM097275, NSF grants PHY-0957573, PHY-1305525, and Harvard Center of Mathematical Sciences and Applications.
\end{acknowledgments}
|
\section{introduction}
Given models $V \subseteq W$ of $ZFC$, we define the notion of generic cut of $(V, W)$, and prove some results about it. We usually assume that $W$ is a generic extension of $V$ by a set or a class forcing notion, but some of our results work for general cases.
\section{generic cut for pairs of models of $ZFC$}
Let's start with the main definition.
\begin{definition}
Suppose that $V \subseteq W$ are models of $ZFC$ with the same ordinals and $\a, \b$ are ordinals.
An $(\a, \b)-$generic cut of $(V, W)$, is a pair $\langle \vec{V}, \vec{W} \rangle,$ where
\begin{enumerate}
\item $\vec{V}=\langle V_i: i<\a \rangle$ is a $\subset$-increasing chain of generic extensions of $V$, with $V_0=V,$
\item $\vec{W}=\langle W_j: j<\b \rangle$ is a $\subset$-decreasing chain of grounds \footnote{Recall that $V$ is a ground of $W$, if $W$ is a set generic extension of $V$ by some forcing notion in $V$.} of $W$, with $W_0=W,$
\item $\forall i<\a, j<\b, V_i \subset W_j,$
\item There is no inner model $V \subseteq M \subseteq W$ of $ZFC$ such that $\forall i<\a, j<\b, V_i \subset M \subset W_j.$
\end{enumerate}
\end{definition}
Note that if $W$ is a set generic extension of $V$, then any $V_i$ is a ground of $W$ and each $W_j$ is a generic extension of $V$.
\begin{lemma}
Suppose that $V \subseteq W$ are models of $ZFC$ and there exists an $(\a, \b)$-generic cut of $(V, W).$ Then $\a \leq [W:V]^U$ and $\b \leq [W:V]_D,$ where $[W:V]^U, [W:V]_D$ are defined as in \cite{golshani}.
\end{lemma}
The next theorem shows that there are no generic cuts in the extension by Cohen forcing.
\begin{theorem}
Let $\mathbb{P}=\Add(\omega, 1)$ be the Cohen forcing for adding a new Cohen real and let $G$ be $\mathbb{P}-$generic over $V$.Then there exists no $(\a, \b)$-generic cut of $(V, V[G])$.
\end{theorem}
\begin{proof}
Assume towards a contradiction that $(\vec{V}, \vec{W})$ witnesses an $(\a, \b)$-generic cut of $(V, V[G])$. We consider several cases:
\begin{enumerate}
\item At least one of $\a$ or $\b$ is uncountable.
Suppose for example that $\a \geq \aleph_1.$ It follows that $\langle V_i: i< \aleph_1 \rangle$ is a $\subset-$increasing chain of generic extensions of $V$, which are included in $V[G]$, which contradicts \cite{golshani} Theorem 2.5$(3)$. So from now on we assume that both $\a, \b$ are countable.
\item Both of $\a=\a^{-}+1$ and $\b=\b^{-}+1$ are successor ordinals. Then we have $V \subseteq V_{\a^{-}} \subset W_{\b^{-}} \subseteq V[G],$ and there are no inner models $M$ of $V[G]$ with $V_{\a^{-}} \subset M \subset W_{\b^{-}},$ which is clearly impossible (in fact there should be $2^{\aleph_0}$ such $M$'s).
\item Both of $\a, \b < \aleph_1$ are limit ordinals.
We can imagine each $V_i, i<\a,$ is of the form $V_i=V[a_i],$ for some Cohen real $a_i$ and similarly each $W_j, j<\b,$ is of the form $W_j=V[b_j],$ for some Cohen real $b_j.$ Let $a\in V[G]$ be a Cohen real over $V$, coding all of $a_i$'s, $i<\a.$ Then for all $i<\a, j<\b, V_i \subset V[a] \subset W_j,$ a contradiction.
\item One of $\a$ or $\b$ is a limit ordinal $>0$ and the other one is a successor ordinal. Let's assume that $\a$ is a limit ordinal and $\b=\b^{-}+1$ is a successor ordinal. As $cf(\a)=\omega,$ we can just consider the case where $\a=\omega.$ Then for all $i<\omega$ we have $V_i \subset W_{\b^{-}}.$ We can assume that each $V_i$ is of the form $V_i=V[a_i],$ for some Cohen real $a_i,$ and that $W_{\b^{-}}=V[b],$ for some Cohen real $b$. Using a fix bijection $f: \omega \leftrightarrow \omega\times \omega, f\in V,$ we can imagine $b$ as an $\omega$-sequence $\langle b_i: i<\omega \rangle$ of reals which is $\Add(\omega, \omega)$-generic over $V$, so that $W_{\b^{-}}=V[\langle b_i: i<\omega \rangle].$ We can further suppose that each $b_i$ codes $a_i$ (i.e., $a_i\in V[b_i]$) Let us now define a new sequence $\langle c_i: i<\omega \rangle$ of reals in $V[\langle b_i: i<\omega \rangle],$ so that
$c_i(0)=0,$ and $c_i \upharpoonright [1, \omega)=b_i \upharpoonright [1, \omega).$ Finally let $M=V[\langle c_i: i<\omega \rangle].$ It is clear that each $V_i \subset M.$ But also $M \subset W_{\b^{-}},$ as the real $t\in W_{\b^{-}}$ defined by $t(i)=b_i(0)$ is not in $M$ (by a genericity argument). We get a contradiction.
\item One of $\a$ or $\b$ is $0$ and the other one is a limit or a successor ordinal. Then as above we can get a contradiction.
\end{enumerate}
\end{proof}
\begin{theorem}
Assume $\a, \b$ are ordinals. Then
there exists a generic extension $V[G]$ of $V$, such that there is an $(\a+1, \b+1)-$generic cut of $(V, V[G]),$
\end{theorem}
\begin{proof}
Let $\PP_1=\Add(\omega, \a),$ and let $G_1=\langle a_i: i<\a \rangle$ be a generic filter over $V$. Force over $V[G_1]$ by any forcing notion which produces a minimal extension $V[G_1][G_2]$ of $V[G_1]$. Finally force over $V[G_1][G_2]$ by $\PP_3=\Add(\omega, \b),$ and let $G_3=\langle b_j: j<\b \rangle$ be a generic filter over $V[G_1][G_2]$. Let
\begin{itemize}
\item $\vec{V}=\langle V_i: i\leq \a \rangle$, where $V_i=V[\langle a_\xi: \xi< i \rangle]$ for $i<\a,$ and $V_\a=V[G_1],$
\item $\vec{W}=\langle W_j: j\leq \b \rangle,$ where $W_j=V[G_1][G_2][\langle b_\xi: j \leq \xi <\b \rangle]$ for $j<\b,$ and $W_\b=V[G_1][G_2]$.
\end{itemize}
Then $(\vec{V}, \vec{W})$ witnesses an $(\a+1, \b+1)-$generic cut of $(V, V[G]).$
\end{proof}
\begin{remark}
If $V$ satisfies $GCH$, then we can find $V[G],$ so that it also satisfies the $GCH$; it suffices to work with $\Add(|\a|^+, \a)$ and $\Add(|\b|^+, \b)$ instead of $\Add(\omega, \a)$ and $\Add(\omega, \b)$ respectively.
\end{remark}
\begin{theorem}
Assume $\l_1, \l_2$ are infinite regular cardinals and $\k$ is a measurable cardinal above them. Then in a generic extension $V[G]$ of $V$, there exists a $(\l_1, \l_2)-$generic cut of $(V, V[G]).$
\end{theorem}
\begin{proof}
By \cite{spa}, we can find a generic extension $V[G_1]$ of $V$, by a forcing of size $<\k,$ such that in $V[G_1],$ there exists a $(\l_1, \l_2)$-gap of $P(\omega)/fin$. $\k$ remains measurable in $V[G_1],$ so let $U$ be a normal measure on $\k$ in $V[G_1],$ and force with the corresponding Prikry forcing $\PP_U,$ and let $G_2$ be $\PP_U$-generic over $V[G_1].$ Let $V[G]=V[G_1][G_2].$ By \cite{gitik},
\begin{center}
$(\{M: M$ is a model of $ZFC, V[G_1] \subseteq M \subseteq V[G] \}, \subseteq) \cong (P(\omega)/fin, \subseteq^*).$
\end{center}
Now the result should be clear, as a $(\l_1, \l_2)$-gap in $P(\omega)/fin$, produces the corresponding $(\l_1, \l_2)-$generic cut $(\vec{V}, \vec{W})$ of $(V[G_1], V[G]),$ which in turn produces the same $(\l_1, \l_2)-$generic cut of $(V, V[G])$ (by adding $V$ at the beginning of $\vec{V}$).
\end{proof}
|
\section{Introduction}
In this paper we study the problem of sequential prediction of a string of bits $ (y_1,\ldots,y_n) \triangleq y_{1:n} \in\{0,1\}^n$. At each round $t=1,\ldots,n$, the forecaster observes side information $x_t\in\X_t$, decides on the probability $\widehat{y}_t\in[0,1]$ of the event $y_t=1$, observes the outcome $y_t\in\{0,1\}$, and pays according to the logarithmic (or, \emph{self-information}) loss function
$$\loss(\widehat{y}_t,y_t) = -\ind{y=1}\log\widehat{y}_t - \ind{y_t=0}\log(1-\widehat{y}_t).$$
At each time instance $t$, the side-information set $\X_t$ is a subset of an abstract set $\X$. The subset $\X_t$ is allowed to depend on the history $h_{1:t-1}\triangleq (x_{1:t-1},y_{1:t-1})$, and the functions $\X_t:(\X\times\Y)^{t-1}\to 2^\X$ are assumed to be known to the forecaster.
The goal of the forecaster is to predict as well as a benchmark set $\F$ of functions---sometimes called ``experts''---mapping $\X$ to $[0,1]$. More specifically, the goal is to keep \emph{regret}
$$\sum_{t=1}^n \loss(\widehat{y}_t,y_t) - \inf_{f\in\F}\sum_{t=1}^n \loss(f(x_t),y_t)$$
as small as possible for all sequences $y_1,\ldots,y_n$ and $x_1,\ldots,x_n$ (satisfying $x_t\in\X_t(h_{1:t-1})$).
To illustrate the setting, consider a few examples. We may take $\X_t(h_{1:t-1})=\left\{(y_1,\ldots,y_{t-1})\right\} \subset \{0,1\}^{t-1}$ to be a singleton set containing the exact realization of the sequence so far. In this case, the choice $x_t=(y_1,\ldots,y_{t-1})$ is enforced and $f(x_t)=p_f(1|y_1,\ldots,y_{t-1})$ may be viewed as a conditional distribution; the normalized maximum likelihood forecaster is known to be minimax optimal in this extensively studied scenario (e.g. \cite[Ch. 9]{PLG}). Alternatively, we may define $\X_t(h_{1:t-1}) = \left\{y'\in \{0,1\}^{t-1}: d_{H}(y_{1:t-1},y')\leq r\right\}$ to be a set that contains histories with up to $r$ flips of the bits. In this case, the forecaster is facing a situation where history can be slightly altered in an adversarial fashion. As another example, we may take $\X_t(h_{1:t-1}) = \left\{(y_{t-k},\ldots,y_{t-1})\right\}$, in which case the forecaster competes with a set of $k$th-order stationary Markov experts. The set $\X_t$ may also be time-invariant, in which case $f$ is a memoryless expert that acts on side information. In short, the formulation we presented subsumes a wide range of interesting problems. Our goal in this paper is to understand how ``complexity'' of $\F$ affects minimax rates of regret.
The minimax regret for the problem of sequential probability assignment can be written as
\begin{align}
\label{eq:unrolled_minimax_with_swap}
V_n(\F) &= \multiminimax{\sup_{x_t\in\X_t(x_{1:t-1},y_{1:t-1})}~~ \inf_{\widehat{y}_t\in[0,1]}~~ \sup_{p_t\in [0,1]}~~ \En_{y_t \sim p_t}}_{t=1}^n \left\{\sum_{t=1}^n \loss(\widehat{y}_t,y_t) - \inf_{f\in\F}\sum_{t=1}^n \loss(f(x_t),y_t) \right\}
\end{align}
where $\En_{y_t\sim p_t}$ is a shorthand for the expectation with respect to Bernoulli $y_t$ with bias $p_t$. Following \cite{StatNotes2012}, the notation $\multiminimax{\ldots}_{t=1}^n$ represents a repeated application of the operators inside the brackets and corresponds to the unrolled minimax value of the associated game between the forecaster and Nature. Any upper bound on $V_n(\F)$ guarantees existence of a strategy that attains regret of at most that amount. In the last few years, new techniques with roots in empirical process theory have emerged for analyzing minimax values of the form \eqref{eq:unrolled_minimax_with_swap}. We bring these techniques to bear on the problem of sequential probability assignment with self-information loss.
Our point of comparison will be the study of rich classes in \cite[Section 9.10]{PLG}. Following \cite{PLG}, we employ the truncation method to deal with the unbounded loss function. To this end, fix $\delta\in(0,1/2)$, to be chosen later. For $a\in[0,1]$, let $\thresh{a}$ denote the thresholded value
\begin{align*}
\thresh{a} ~=~ \begin{cases}
\delta &\mbox{if } a<\delta \\
a &\mbox{if } a\in[\delta,1-\delta] \\
1-\delta &\mbox{if } a>1-\delta.
\end{cases}
\end{align*}
For a class $\F$, let $\F^\delta = \{\thresh{f}:f\in\F\}$ denote the class of truncated functions. It is easy to check (see \cite[Lemma 9.5]{PLG}) that
\begin{align}
\label{eq:off_delta}
V_n(\F)\leq V_n(\F^\delta)+2n\delta,
\end{align}
and we can, therefore, focus on the minimax regret with respect to $\F^\delta$. We show that $V_n(\F^\delta)$ can be upper bounded via a modified (offset) sequential Rademacher complexity, which in turn can be controlled via sequential chaining in the spirit of \cite{RakSri14a,RakSri15nonparametric}. Unlike the latter two papers, however, we do not employ symmetrization and instead use the self-information property of the loss function. We are able to mitigate the adverse dependence of $V_n(\F^\delta)$ on $\delta$ by introducing chaining with Bernstein-style terms that control the sub-Gaussian and sub-exponential tail behaviors. As an example, we recover the $n^{3/5}$ rate for monotonically increasing experts presented in \cite[Sec 9.10-9.11]{PLG}. However, our technique goes well beyond such examples of ``static'' experts. In particular, we can obtain non-trivial rates even in the setting where discretization in the style of \cite[Sec 9.10-9.11]{PLG},\cite{cesa1999minimax} leads to vacuous bounds. One such example is when experts are indexed by a unit ball in a Hilbert space (or, a high-dimensional Euclidean space) and expert's prediction depends linearly on side information. A discretization in the supremum norm of this set of experts is not finite, and thus the typical approaches to this problem fail. In contrast, we employ the ideas from empirical process theory and its sequential generalization in \cite{RakSriTew14ptrf} in order to define ``data-dependent'' notions of complexity.
Despite the improvement over the technique of \cite{PLG}, the rates attained in this paper are not always minimax optimal, as we demonstrate in Section~\ref{sec:linear}. This is in contrast to other loss functions (such as absolute, square, $q$-power, and logistic) for which matching upper and lower bounds (to within logarithmic factors) have been established recently in \cite{RakSri15nonparametric}. As mentioned in \cite{PLG}, the truncation method is crude, and we leave it as an open question whether a different technique can be employed to attain optimal rates.
We finish this introduction with a brief mention that sequential probability assignment is extensively studied in Information Theory, where regret is known as \emph{redundancy} with respect to a set of codes. The vast literature mostly investigates the case of parametric classes (see \cite{shtar1987universal,xie2000asymptotic,freund1996predicting,rissanen1986complexity,rissanen1996fisher} and the references in \cite[Ch. 9]{PLG}), with exact constants available in certain cases. We refer to \cite{mf-up-98} for a discussion of approaches to dealing with large comparator classes. Given the well-known connection to compression, it would be interesting to employ the relaxation-based algorithmic recipe of \cite{rakhlin2012relax,RakSri15nonparametric,StatNotes2012} to come up with novel data compression methods.
\section{Complexity of Large Classes of Experts}
We focus on the minimax value for the thresholded class $\F^\delta$. To state the first technical lemma, we need the definition of a tree. For an abstract set $\Z$, a $\Z$-valued complete binary tree $\z$ of depth $n$ is a collection of labeling functions $\z_t:\{0,1\}^{t-1}\to \Z$ for $t\in\{1,\ldots,n\}$. For a sequence $y=(y_1,\ldots,y_n)\in\{0,1\}^n$ (which we call \emph{a path}), we write $\z_t(y)$ for $\z_t(y_1,\ldots,y_{t-1})$. Once we take $y_1,\ldots,y_n$ to be random variables, we may view $\{\z_t\}$ as a predictable process with respect to the filtration given by $\sigma(y_1,\ldots,y_{t-1})$. \footnote{We remark that in \cite{RakSriTew14ptrf,RakSriTew10,StatNotes2012}, the trees are defined with respect to $\{\pm1\}$-valued sequences, whereas here we use the $\{0,1\}$-valued variables. The change is purely notational and all the definitions and results can be rephrased appropriately.}
We will say that an $\x$ tree is \emph{consistent} with respect to the side information set mappings $h_{1:t-1}\mapsto \X_t(h_{1:t-1})$ if for any $y\in\{0,1\}^n$, it holds that for all $t$,
$$\x_t(y) \in \X_t(\x_1(y),\ldots,\x_{t-1}(y), y_1,\ldots,y_{t-1}).$$
A consistent tree respects the sets of constraints $\X_t$ imposed by the problem. For the purposes of analyzing complexity of $\F$, it is important that the constraints are reflected in the tree $\x$.
Theorem~\ref{thm:value_upper} below relates the minimax regret with respect to $\F^\delta$ to the supremum of a stochastic process of a form similar to \emph{offset Rademacher complexity} introduced in \cite{RakSri14a}. The key difference with respect to \cite{RakSri14a} is that the stochastic process is defined with potentially biased coin flips. To prove Theorem~\ref{thm:value_upper}, we avoid symmetrization and instead exploit the fact that the logarithmic loss has the self-information property: in the maximin dual, the optimal probability assignment is given precisely by the distribution of the $y_t$ variable. We note that the symmetrization approach of \cite{RakSri14a} appears to give worse rates for the logarithmic loss function.
Let
\begin{align}
\label{def:eta}
\eta(p,a) \triangleq - \ind{a=1}p^{-1} + \ind{a=0} (1 - p)^{-1}
\end{align}
and observe that $\eta$ is zero-mean if $a$ is Bernoulli random variable with bias $p$.
\begin{theorem}
\label{thm:value_upper}
The following upper bound holds:
\begin{align*}
V_n(\F^\delta) \leq \sup_{\x, \ensuremath{\boldsymbol{\mu}}, \ensuremath{\mathbf{p}}}\En \sup_{f \in \F^\delta} \left[ \sum_{t: {\ensuremath{\mathbf{p}}_t(y)\in[\delta,1-\delta]}} \eta(\ensuremath{\mathbf{p}}_t(y),y_t) \left(\ensuremath{\boldsymbol{\mu}}_t(y) - f(\x_t(y))\right) - \frac{1}{2}\left(\ensuremath{\boldsymbol{\mu}}_t(y) - f(\x_t(y))\right)^2\right] + 2 n \delta \log(1/\delta),
\end{align*}
where $\ensuremath{\mathbf{p}},\ensuremath{\boldsymbol{\mu}}$ range over all $[0,1]$-valued trees, $\x$ ranges over consistent trees, and the stochastic process $y_1,\ldots,y_n$ is defined via $y_t|y_1,\ldots,y_{t-1} \sim \text{\sf Bernoulli}(\ensuremath{\mathbf{p}}_t(y_1,\ldots,y_{t-1}))$.
\end{theorem}
To shorten the notation in Theorem~\ref{thm:value_upper}, let $\Z = \X \times [0,1]$ and for every $f \in \F^\delta$, write $g_f(z) = g_f(x,a) = a - f(x)$. The upper bound of Theorem~\ref{thm:value_upper} can be written more succinctly as
\begin{align}
\label{eq:succinct_form}
V_n(\F^\delta) \le \sup_{\z, \ensuremath{\mathbf{p}}}\En \sup_{f \in \F^\delta} \left[ \sum_{t: {\ensuremath{\mathbf{p}}_t(y)\in[\delta,1-\delta]}} \eta(\ensuremath{\mathbf{p}}_t(y),y_t) g_f(\z_t(y)) - \frac{1}{2}g_f(\z_t(y))^2 \right] + 2 n \delta \log(1/\delta).
\end{align}
We keep in mind that the $\x$ part of $\z$ is a consistent tree. Observe that the expression above is a supremum of a collection of random variables indexed by $f\in\F^\delta$, each with a nonpositive-mean. To analyze the supremum of this stochastic process, we first consider the case when the indexing set is finite.
\begin{lemma}
\label{eq:finite_biased}
For any set $V$ consisting of $[-1,1]$ valued trees, any $[\delta,1-\delta]$-valued tree $\ensuremath{\mathbf{p}}$, and any $c>0$,
\begin{align*}
\En_y \max_{\v \in V} \left[ \sum_{t=1}^n \eta(\ensuremath{\mathbf{p}}_t(y),y_t) \v_t(y) - c\v_t(y)^2 \right] \leq \frac{\log\ |V|}{\delta \log(1 + \frac{c}{2})}
\end{align*}
where $y_t|y_1,\ldots,y_{t-1} \sim \text{\sf Bernoulli}(\ensuremath{\mathbf{p}}_t(y_1,\ldots,y_{t-1}))$. Furthermore, the same upper bound holds if $\ensuremath{\mathbf{p}}$ is any $[0,1]$-valued tree but the summation is restricted to $\{t: \ensuremath{\mathbf{p}}_t(y)\in[\delta,1-\delta]\}$.
\end{lemma}
The tight control of the expectation is possible because of the negative quadratic term that acts as a compensator. On the downside, the upper bound displays the adverse $1/\delta$ dependence. We now show a maximal inequality when the quadratic term is not present. The bound is of a Bernstein type, with the sub-Gaussian and sub-exponential behaviors. Crucially, the sub-Gaussian term scales with $1/\sqrt{\delta}$.
\begin{lemma}
\label{eq:finite_biased_l2}
For any set $V$ consisting of $[-1,1]$ valued trees, any $[\delta,1-\delta]$-valued tree $\ensuremath{\mathbf{p}}$, and any $c>0$,
\begin{align*}
\En_y \max_{\v \in V} \left[ \sum_{t=1}^n \eta(\ensuremath{\mathbf{p}}_t(y),y_t) \v_t(y) \right] \leq 5\bar{v}\sqrt{\frac{n\log |V|}{\delta}} + \frac{2v_{\text{max}}\log |V|}{\delta}
\end{align*}
where $y_t|y_1,\ldots,y_{t-1} \sim \text{\sf Bernoulli}(\ensuremath{\mathbf{p}}_t(y_1,\ldots,y_{t-1}))$, $\bar{v} = \max_{\v\in V}\max_{y} (\frac{1}{n}\sum_{t=1}^n \v_t(y)^2)^{1/2}$, and $v_{\text{max}}=\max_{\v\in V}\max_{y} |\v_t(y)|$. The same upper bound holds if $\ensuremath{\mathbf{p}}$ is any $[0,1]$-valued tree but the summation is restricted to $\{t: \ensuremath{\mathbf{p}}_t(y)\in[\delta,1-\delta]\}$.
\end{lemma}
We now pass from a finite collection to an infinite one via the sequential chaining technique \cite{RakSriTew14ptrf}. For this purpose, we recall the definition of $\ell_p$ sequential covering numbers.
\begin{definition}[\cite{RakSriTew14ptrf}]
\label{def:cover}
A set $V$ of $\mathbb{R}$-valued trees of depth $n$ is a (sequential) $\gamma$-cover (with respect to $\ell_p$, $p\geq 1$) of $\G\subseteq\mathbb{R}^\Z$ on a $\Z$-valued tree $\z$ of depth $n$ if
\begin{align}
\label{eq:def_cover}
\forall g\in\G,~~ y\in\{0,1\}^n, ~~\exists \v\in V,~~\text{s.t.}~~ \left(\frac{1}{n}\sum_{t=1}^n |\v_t(y)-g(\z_t(y))|^p \right)^{1/p}\leq \gamma.
\end{align}
The size of the smallest $\gamma$-cover is denoted by $\cN_p(\G,\gamma,\z)$. For $p=\infty$, \eqref{eq:def_cover} becomes $\max_{t}|\v_t(y)-g(\z_t(y))|\leq \gamma$.
\end{definition}
\begin{theorem}
\label{thm:biased_dudley}
Let $\G$ be a class of functions $\Z\to [-1,1]$. For any $[0,1]$-valued tree $\ensuremath{\mathbf{p}}$, any $\Z$-valued tree $\z$, any $K>0$, and $\gamma>0$,
\begin{align*}
&\En \sup_{g \in \G} \left[ \sum_{t: {\ensuremath{\mathbf{p}}_t(y)\in[\delta,1-\delta]}} \eta(\ensuremath{\mathbf{p}}_t(y),y_t) g(\z_t(y)) - K g(\z_t(y))^2 \right] \\
&\hspace{1in}\leq \frac{1}{\delta} \frac{\log\ \cN_\infty(\G,\gamma,\z)}{ \log(1 + \frac{K}{8})} + \inf_{\alpha(0,\gamma]}\left\{ \frac{4 n \alpha}{\delta} + 30\sqrt{\frac{2n}{\delta}}\int_{\alpha}^\gamma \sqrt{\log\cN_\infty(\G,\rho,\z)}d\rho + \frac{8}{\delta}\int_{\alpha}^\gamma \log\cN_\infty(\G,\rho,\z)d\rho \right\}
\end{align*}
where the stochastic process $y_1,\ldots,y_n$ is defined via $y_t|y_1,\ldots,y_{t-1} \sim \text{\sf Bernoulli}(\ensuremath{\mathbf{p}}_t(y_1,\ldots,y_{t-1}))$.
\end{theorem}
Theorem~\ref{thm:biased_dudley} is readily applied to the upper bound of Theorem~\ref{thm:value_upper} by identifying $$\G = \{g_f(z) = g_f(x,\mu) = \mu - f(x): f\in\F^\delta, \mu\in\mathbb{R}, x\in\X \}$$ and $\z_t(y) = (\x_t(y),\ensuremath{\boldsymbol{\mu}}_t(y))$. It is immediate from the definition of a cover that for any $\ensuremath{\boldsymbol{\mu}}$, $\x$, and $\z=(\x,\ensuremath{\boldsymbol{\mu}})$,
\begin{align}
\label{eq:same_cover}
\cN_p(\F^\delta,\x,\alpha) = \cN_p(\G,\z,\alpha).
\end{align}
The lower bound of Lemma~\ref{lem:lower_bound} (presented in Section~\ref{sec:lower}) and the relation between the offset Rademacher complexity and sequential fat-shattering dimension \cite{RakSriTew14ptrf,RakSri15nonparametric} yield the next theorem.
\begin{theorem}
For the case of constant sets $\X_1=\X_2=\ldots=\X$, the following are equivalent:
\begin{itemize}
\item Minimax regret is sublinear: $\frac{1}{n}V_n(\F)\to 0$ as $n\to \infty$
\item Sequential dimension $\fat_\beta(\F,\X)$ is finite for all $\beta>0$
\end{itemize}
\end{theorem}
Let us make a few remarks. First, the theorem can be easily extended to non-constant sets $\X_t$, in which case $\fat_\beta$ is defined with respect to consistent trees (as in the next section). Second, one may also phrase the equivalence through sequential covering numbers, thanks to the relations outlined in \cite{RakSriTew14ptrf,RakSri15nonparametric}.
In summary, the sequential complexities we study are intrinsic to the problem of sequential probability assignment (unlike, for instance, covering numbers with respect to the supremum norm on $\X$ --- see Section~\ref{sec:example_history} for an example). Yet, the upper bounds we derive do not quite match the lower bounds, due to the hard thresholding approach and the need to balance $n\delta$ with $V_n(\F^\delta)$ at the end of the day. It is an open problem to close the gap between the upper and lower bounds.
The upper bound of Theorem~\ref{thm:biased_dudley} is quantified as soon as we have control of sequential covering numbers. While covering numbers could be computed directly in many situations, it is often simpler to upper bound a ``scale-sensitive dimension'' of the class, defined in the next section. In Section~\ref{sec:monotonic} we present an example of such a simple calculation.
\section{Covering Numbers and Combinatorial Parameters}
\label{sec:cov_comb}
Suppose we can define a preorder $\preceq$ on the set $\X$ (that is, a binary relation that is reflexive and transitive). We say that an $\X$-valued tree $\x$ of depth $n$ is \emph{ordered} if for any path $y\in\{0,1\}^n$, it holds that $\x_t(y)\preceq \x_{t+1}(y)$ for all $t=1,\ldots,n-1$. In this section we show that the combinatorial dimensions, covering numbers, and the associated upper bounds in \cite{RakSriTew14ptrf} can be extended to ``respect'' the preorder (of course, one can always define a vacuous relation $\preceq$ and recover prior results).
\begin{definition}
\label{def:fat}
A class $\F\subset \mathbb{R}^\X$ shatters (at scale $\beta>0$) an ordered $\X'$-valued tree of depth $d$ if there exists a $\mathbb{R}$-valued witness tree $\s$ of depth $d$ such that
$$\forall y\in\{0,1\}^d,~~ \exists f\in\F,~~ \mbox{s.t.}~~ (2y_t-1)(f(\x_t(y))-\s_t(y))\geq \beta/2.$$
The largest depth of an ordered $\X'$-valued tree is denoted by $\fat_\beta^o(\F,\X')$, where the superscript $o$ stands for ``ordered''.
\end{definition}
The notion of the Littlestone's dimension $\ldim(\F,\X')$ for $\{0,\ldots,k\}$-valued function classes extends in exactly the same way to the case of ordered trees.
The main step in obtaining upper bounds on sequential covering numbers is the analogue of the Vapnik-Chervonenkis-Sauer-Shelah lemma, proved in \cite{RakSriTew14ptrf,RakSriTew10}. We now show that if we ask for a $\beta$-cover on an ordered tree $\x$, the sequential covering numbers are controlled via the ordered version $\fat_\beta^o(\F,{\sf Img}(\x))$ of the fat-shattering dimension in Definition~\ref{def:fat}.
\begin{theorem}[Extension of Theorem 4 in \cite{RakSriTew14ptrf}]
\label{thm:covering_estimate}
Let $\F\subseteq \{0,\ldots,k\}^\X$ be a class of functions with $\fat_2^o(\F,\X)=d$. Then for any $n>d$ and any ordered $\X$-valued tree $\x$,
$$\cN_\infty (\F,1/2,\x) \leq \sum_{i=0}^d {n\choose i}k^i.$$
Hence, for a class $\G\subseteq[-1,1]^\X$, for any $\beta>0$,
$$\cN_\infty(\G, \beta, \x)\leq \left(\frac{2en}{\beta}\right)^{\fat_\beta^o(\G,\X)}.$$
\end{theorem}
The following three sections are devoted to particular examples. We start by exhibiting a simple class for which sequential covering numbers are small, yet the discretization with respect to the supremum norm (typically performed to appeal to a finite-experts method) gives vacuous bounds.
\section{Example: Consistent History}
\label{sec:example_history}
We would like to illustrate that sequential covering number can be much smaller than covering numbers with respect to the supremum norm over $\X$. Consider the particular case of $\X_t(h_{1:t-1})=\{(y_1,\ldots,y_{t-1})\}$. Clearly, there is only one consistent tree, namely the one defined by $\x_t(y) = (y_1,\ldots,y_{t-1})$ for any $t$. In this case, the requirement \eqref{eq:def_cover} in Definition~\ref{def:cover} with class $\F^\delta$, consistent tree $\x$, and $p=\infty$ reads as
\begin{align}
\label{eq:def_cover2}
\forall f\in\F^\delta,~~ y\in\{0,1\}^n,~~ \exists \v\in V,~~\text{s.t.}~~ |\v_t(y_{1:t-1})-f(y_{1:t-1})| \leq \gamma.
\end{align}
We contrast this with the definition in \cite[Sec. 9.10]{PLG}, where the covering of $\F$ is done with respect to the following pointwise metric (which we normalized by $\sqrt{n}$ for uniformity):
\begin{align}
\label{def:sup_norm_distance}
d(f,g) = \sqrt{\frac{1}{n}\sum_{t=1}^n \sup_{y_{1:t}} \left( \loss(f(y_{1:t-1}), y_t) - \loss(g(y_{1:t-1}), y_t)\right)^2}.
\end{align}
To illustrate a gap in the two covering-number approaches, construct a particular class $\F$ as follows. For each element $b\in\{0,1\}^n$, define $f_b$ by
$$f_b(y_{1:t-1})=\frac{1}{4}\ind{b_{1:t-1}=y_{1:t-1}}+ \frac{1}{4}$$
and take $\F=\{f_b:b\in\{0,1\}^n\}$. In other words, on round $t$, expert $f_b$ predicts probability $1/2$ if history coincides with $b_{1:t-1}$, and $1/4$ otherwise. For two elements $f_{b},f_{b'}\in\F$, let $\kappa(b,b')=\max\{t: b_t= b_t'\}$ be the last time the two sequences agree (defined as $0$ if $b_1\neq b_1'$). Then
\begin{align*}
&\sum_{t=1}^n \sup_{y_{1:t}} \left( \loss(f_b(y_{1:t-1}), y_t) - \loss(f_{b'}(y_{1:t-1}), y_t) \right)^2 \geq \sum_{t=1}^n \left( \loss(f_b(b_{1:t-1}), 1) - \loss(f_{b'}(b_{1:t-1}), 1) \right)^2
\geq (n-\kappa(b,b'))\log(2)^2
\end{align*}
and thus there are at least $2^{n/2}$ functions at a constant distance $d(f,g)\geq c$.
In contrast, consider sequential covering in the sense of \eqref{eq:def_cover2} (and Definition~\ref{def:cover}). Take any $y\in\{0,1\}^n$ and $f_b\in\F$. The sequence of $n$ values $(f_b(\emptyset), f_b(y_1),\ldots,f_b(y_{1:t-1}),\ldots, f_b(y_{1:n-1}))$ is equal to $1/2$ until $t=\kappa(b,y)$ and $1/4$ afterwards. Let $V$ be a set of $n$ trees $\v^1,\ldots,\v^n$ labeled by $\{1/4,1/2\}$. Each $\v^i$ is defined as
$$\forall y\in\{0,1\}^n, t\in\{1,\ldots,n\},~~~~~~ \v^i_t(y) = (1/4)\ind{t\leq i-1}+1/4.$$
It is immediate that this set of $n$ trees provides an exact cover of $\F$ (at scale $0$) in the sense of Definition~\ref{def:cover}. This leads to $\mathcal{O}(\log(n)/n)$ bounds on minimax regret, while the discretization with respect to the supremum norm \eqref{def:sup_norm_distance} fails.
The above failure is endemic to approaches that attempt to discretize the set of experts before the prediction process even started. In contrast, sequential complexities can be viewed as an analogue of ``data-based'' discretization, which is known in statistical learning since the work of Vapnik and Chervonenkis in the 60's.
\section{Example: Monotonically Nondecreasing Experts}
\label{sec:monotonic}
We consider an example of a nonparametric class analyzed in \cite[p. 270]{PLG}. Let $f\in\F$ be a set of experts such that the forecasted probability does not decrease in time. To model this scenario in a general manner, we suppose that the side information $x_t = (t, x'_t)\in {\mathbb N}\times \X'_t(x_{1:t-1},y_{1::t-1})$ contains the time stamp, and $f(t+1,x'_t)\geq f(t,x''_t)$ for any $f\in\F$. The particular case of \emph{static} experts---with prediction depending only on $t$ and no other side information---has been considered in \cite{PLG}.
To invoke the results of the previous section, define a preorder on $(t,x)\in \X={\mathbb N}\times \X'$ according to the time stamp: $(t,u)\preceq (s,v)$ for any $t< s$ and $u,v\in \X'$. Suppose an ordered $\X$-valued tree $\x$ of depth $d$ is shattered, according to Definition~\ref{def:fat}, with a witness tree $\s$. We claim that the values of the witness tree must be increasing by at least $\beta$ along the path $y=(1,1,1,\ldots)$. Indeed, consider any $t\geq 1$, and let $y'=(y_{1:t},0,y_{t+2:d})$. By the definition of shattering, there must be a function that satisfies $f(\x_t(y'))\geq \s_t(y')+\beta/2$ and $f(\x_{t+1}(y'))\leq \s_{t+1}(y')-\beta/2$. Since $f(\x_t(y'))\leq f(\x_{t+1}(y'))$, we conclude that $\s_t(y)=\s_t(y')\leq \s_{t+1}(y')-\beta = \s_{t+1}(y)-\beta$. Hence, $\s_t$ increases by at least $\beta$ along the path $(1,\ldots,1)$ and thus $d\leq 1/\beta$. This quick calculation gives $\fat_\beta^o(\F,\X)\leq 1/\beta$.
In view of Theorem~\ref{thm:covering_estimate},
$$\log\cN_\infty(\F^\delta,\beta,\x)\leq (1/\beta)\log\left(2en/\beta\right)$$
In view of \eqref{eq:same_cover}, the same covering number estimate holds for $\G$. Then Theorem~\ref{thm:biased_dudley} with $\alpha=1/n$ and $\gamma=n^{-a}$ (with $a$ to be determined later) implies that
$$\int_{\alpha}^\gamma \log\cN_\infty(\G,\rho,\z)d\rho \leq C\log^2 n$$
is a lower order term, with $C$ being an absolute constant. We also have
$$\int_{\alpha}^\gamma \sqrt{\log\cN_\infty(\G,\rho,\z)}d\rho\leq C'\sqrt{\log n} \cdot \gamma^{1/2}.$$
Now, ignoring constants and logarithmic terms, this gives the overall rate of
$$\mathcal{O}^*\left(\frac{1}{\delta\gamma} + \sqrt{\frac{n\gamma}{\delta}}\right) = \mathcal{O}^*\left(n^{1/3}\delta^{-2/3}\right)$$
for the minimax regret with respect to $\F^\delta$. The terms are balanced by choosing
$\gamma=n^{-1/3}\delta^{-1/3}$. The rate with respect to $\F$ is then
$$\mathcal{O}^*\left(n\delta + n^{1/3}\delta^{-2/3}\right) = \mathcal{O}^*\left(n^{3/5}\right)$$
by choosing $\delta=n^{-2/5}$. This corresponds to the rate obtained by \cite{PLG}.
\section{Example: Linear Prediction}
\label{sec:linear}
In this section we consider the special case of $\X_1=\ldots=\X_n=\X=B_2$ and
\begin{align}
\label{eq:lin_class}
\F = \{f(x)=(\inner{w,x}+1)/2: w\in B_2\}
\end{align}
where $B_2$ is a unit Euclidean (or Hilbert) ball. Written as a function of $w$, the loss at time $t$ is (up to an additive constant $\log(2)$)
\begin{align}
\label{eq:funcs}
g_t(w) = -\ind{y_t=1}\log(1+\inner{w,x_t}) - \ind{y_t=0}\log(1-\inner{w,x_t}).
\end{align}
It is possible to estimate the sequential $\fat_\beta$ dimension of a unit Hilbert ball as $\fat_\beta = \mathcal{O}^*(1/\beta^2)$, where the $\mathcal{O}^*$ notation ignores logarithmic factors. Then Theorem~\ref{thm:biased_dudley} gives an upper bound of
$$V_n(\F^\delta) = \mathcal{O}^*\left(n^{1/2}\delta^{-1}\right),$$
and thus
$$V_n(\F) = \mathcal{O}^*\left(n^{3/4}\right).$$
Below, we exhibit an algorithm that attains regret of $\mathcal{O}^*\left(n^{1/2}\right)$, implying that the upper bounds obtained with our technique are not always tight.
\subsection{Algorithm: Regularization with Self-Concordant Barrier}
To develop an algorithm for the problem, we turn to the field of online convex optimization. We observe that functions $g_t$ defined in \eqref{eq:funcs} are convex, but not strongly convex. Moreover, the gradients of $g_t(w)$ are not bounded. We may consider a restricted set to mitigate the exploding gradient; however, a $\delta$-shrinkage of the ball $B_2$ still leaves the gradient to be of size $O(1/\delta)$. A direct gradient descent method will give the suboptimal $O(n^{3/4})$ upper bound derived above in a non-constructive way. We also mention that while the functions are exp-concave, the upper bounds for the Online Newton Step method \cite{hazan2007logarithmic} scale with the dimension of the space, which we assume to be large or infinite.
We now present an algorithm based on self-concordant barrier regularization, which appears to be of an independent interest. The algorithm answers the following question: \emph{can one obtain regret bounds for online convex optimization in terms of the maximum of function values rather than gradients?}
Consider the Follow-the-Regularized-Leader method
\begin{align}
\label{eq:ftrl}
w_{t+1} = \argmin{w\in B_2} \sum_{s=1}^t \inner{\nabla g_s(w_s), w} + \eta^{-1} R(w)
\end{align}
with the self-concordant barrier $R(w) = -\log(1-\|w\|^2)$. In accordance with the protocol of the probability assignment problem, we predict $\inner{w_{t},x_{t}}$ at round $t$ after observing $x_t$. It is shown in \cite{AbeRak09colt} that regret of \eqref{eq:ftrl} against any $w^*\in B_2$ is
\begin{align}
\label{eq:reg_bound_local}
\sum_{t=1}^n g_t(w_t) - g_t(w^*) \leq 2\eta \sum_{t=1}^n \|\nabla g_t(w_t)\|_{w_t}^{*2} + \eta^{-1} R(w^*)
\end{align}
as long as $\eta$ satisfies $\eta \|\nabla g_t(w_t)\|^*_{w_t}\leq 1/4$. Here, the \emph{local norm} is defined as
$$\|h\|^*_{w} = \sqrt{h^\ensuremath{{\scriptscriptstyle\mathsf{T}}} (\nabla^2 R(w))^{-1} h}.$$
According to the lemma below, the local norm is bounded by a constant that is independent of the dimension:
\begin{lemma}
\label{eq:bdd_local_norm}
For any $t$, the local norm of $\nabla g_t(w_t)$ is upper bounded by a constant:
$$\|\nabla g_t(w_t)\|^*_{w_t} \leq 3.$$
\end{lemma}
Together with \eqref{eq:reg_bound_local}, Lemma~\ref{eq:bdd_local_norm} implies a regret bound of
$18\eta n + \eta^{-1} R(w^*).$
Instead of taking $w^*$ at the boundary of the ball where $R(w^*)$ is infinite, we can evaluate regret against $w=(1-1/n)w^*$. For such a comparator, $R(w)=\mathcal{O}(\log n)$. By choosing $\eta$ appropriately and using an argument similar to \eqref{eq:off_delta}, we conclude that regret against any $w^*\in B_2$ is
upper bounded by
$$C\sqrt{n\log n}.$$
Importantly, $C$ is an absolute constant that does not depend on the dimension of the problem. This rate is optimal up to polylogarithmic factors. The optimality follows from Lemma~\ref{lem:lower_bound} below and an estimate on sequential covering number of a Hilbert ball \cite{RakSri14a,RakSri15nonparametric}.
\begin{lemma}
For the linear class in \eqref{eq:lin_class},
$$V_n(\F) = \mathcal{\Theta}^*(n^{1/2}).$$
\end{lemma}
The proof of Lemma~\ref{eq:bdd_local_norm} relied heavily on the ability to calculate the gradient of the loss function and match it to the inverse Hessian of the self-concordant barrier. We now give an alternative proof based on a simple and charming, yet unexpected lemma due to Nesterov (see Appendix for the short proof):
\begin{lemma}[Lemma 4 in \cite{nesterov2011barrier}]
\label{lem:nesterov}
Let $\psi$ be concave and positive on $\text{int}~ \K$. Then for any $x\in\text{int}~ \K$ we have
$$\|\nabla\psi(x)\|_x^* \leq \psi(x).$$
\end{lemma}
The lemma allows us to upper bound regret in an online convex optimization problem if we only know that the values of the functions (and not the gradients) are bounded. Consider the FTRL algorithm \eqref{eq:ftrl}, but over the shrunk ball $(1-1/n)B_2$. Suppose we can ensure $0< g_t < A$. Then $A-g_t$ is concave and positive. Hence, by above lemma
$$\|\nabla g_t(w_t)\|^*_{w_t} = \|\nabla (A-g_t(w_t))\|^*_{w_t} \leq A-g_t(w_t) \leq A$$
which provides an alternative to the bound of Lemma~\ref{eq:bdd_local_norm}.
Regret is then upper bounded by
$$\sum_{t=1}^n g_t(w_t) - \sum_{t=1}^n g_t(w^*) \leq 2\eta n A^2 + \eta^{-1} R(w^*) $$
Crucially, by employing self-concordant regularization, we avoid paying for a large gradient of cost functions at the boundary of the set. Over the shrunk set $(1-1/n)B_2$, we ensure that the values of functions $g_t$ are upper bounded by $A=O(\log n)$ even if the gradients blow up linearly with $n$. This surprising observations leads to a dimension-independent $O(\sqrt{n}\log n)$ regret bound for the Euclidean ball, and can also be used for other convex bodies and non-logarithmic loss functions when the closed-form analysis of Lemma~\ref{eq:bdd_local_norm} is not available.
\section{A Lower Bound}
\label{sec:lower}
In this section, we show that the offset sequential Rademacher complexity serves as a lower bound on the minimax regret. Hence, the complexities of the class $\F$ of experts are intrinsic to the problem. We refer to \cite{RakSri15nonparametric} for further lower bounds on the offset Rademacher complexity via the scale-sensitive dimension and sequential covering numbers.
\begin{lemma}
\label{lem:lower_bound}
The following lower bound holds:
\begin{align*}
V_n(\F) +1 \geq \sup_{\x} \En_{y} \left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n 2 (2y_t-1)(f(\x_t(y))-1/2) - 4(\log n)(f(\x_t(y))-1/2)^2
\right\}
\right]
\end{align*}
where $y_1,\ldots,y_n$ are independent with distribution ${\sf Bernoulli(1/2)}$ and the supremum is taken over consistent trees with respect to constraints $\X_t$.
\end{lemma}
\begin{proof}[\textbf{Proof of Lemma~\ref{lem:lower_bound}}]
To prove the lower bound, we proceed as in \cite{RakSri15nonparametric}. First, we observe that \eqref{eq:off_delta} holds in the other direction too:
\begin{align}
\label{eq:off_delta_opposite}
V_n(\F^\delta)\leq V_n(\F)+n\delta.
\end{align}
To see this, note that any $f$ only loses from thresholding when either $f(x_t)>1-\delta$ and $y_t=1$, or when $f(x_t)<\delta$ and $y_t=0$. In both cases, the difference in logarithmic loss is at most $-\log(1-\delta)\leq \delta$ for $\delta<1/2$. For the purposes of a lower bound, we take $\delta = 1/n$ and turn to lower-bounding $V_n(\F^{1/n})$.
As in the development leading to \eqref{eq:minimax_swapped} in the proof of the upper bound, the minimax value $V_n(\F^{1/n})$ is \emph{equal} to
\begin{align}
&\multiminimax{\sup_{x_t}\sup_{p_t\in[0,1]} \En_{y_t}}_{t=1}^n
\left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n \inf_{\widehat{y}_t\in[0,1]} \En_{y_t}\left[\loss(\widehat{y}_t,y_t)\right] - \sum_{t=1}^n \loss(f(x_t),y_t)
\right\}
\right]
\end{align}
which, by the self-information property of the loss equal to
\begin{align}
&\multiminimax{\sup_{x_t}\sup_{p_t\in[0,1]} \En_{y_t}}_{t=1}^n
\left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n \En_{y_t}\left[\loss(p_t,y_t)\right] - \sum_{t=1}^n \loss(f(x_t),y_t)
\right\}
\right]
\end{align}
By the linearity of expectation (and since the terms $\En_{y_t}\left[\loss(p_t,y_t)\right]$ do not involve $f$), we have
\begin{align}
\label{eq:lowerbd}
V_n(\F^{1/n}) = &\multiminimax{\sup_{x_t}\sup_{p_t\in[0,1]} \En_{y_t}}_{t=1}^n
\left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n \loss(p_t,y_t) - \sum_{t=1}^n \loss(f(x_t),y_t)
\right\}
\right] .
\end{align}
We now pass to the first lower bound by choosing $p_t=1/2$ for all $t$.
Consider the case $y_t=1$ and expand the loss function around $p_t=1/2$ for $z\in[1/n,1]$:
\begin{align}
\label{eq:remainder}
\loss(1/2,1) - \loss(z,1) = -\log(1/2) - (-\log(z)) = 2(z-1/2) - R(z)
\end{align}
where $R(z)$ is the remainder. We claim that the remainder can be upper bounded by a quadratic over the interval $[1/n, 1]$. To this end, consider the function
$$g(z) = -2z + (1+\log(2))+4(\log n)(z-1/2)^2$$
and note that the derivative and the value of this function at $1/2$ coincide with the derivative and the value of $-\log(z)$ at the same point. We claim that $g(z)$ dominates $-\log(z)$ on $[1/n,1]$. For $z>1/2$, this follows from $g'>(-\log)'$. The same argument holds for the interval $[1/\log (n), 1/2]$. Now, at $z=1/n$, $g(z)> -\log(z)$ and $|g'(z)|<|\log(z)'|$. The derivative relation continues to hold on the interval $[1/n,c/\log (n)]$ for large enough $c$, establishing $g>-\log$ on this interval too. The remaining interval $[c/\log(n), 1/\log(n)]$ is easily checked by the direct computation of function value. In sum,
the remainder in \eqref{eq:remainder} can be upper bounded by $R(z)\leq 4(\log n)(z-1/2)^2$.
The case of $y_t=0$ is exactly analogous, and we obtain
\begin{align}
\loss(p_t,y_t) - \sum_{t=1}^n \loss(f(x_t),y_t) &\geq 2\left[ \ind{y_t=1}(f(x_t)-1/2) + \ind{y_t=0}(-f(x_t)+1/2) \right] - 4(\log n)(f(x_t)-1/2)^2 \\
&= 2\left[ y_t(f(x_t)-1/2) + (1-y_t)(-f(x_t)+1/2) \right] - 4(\log n)(f(x_t)-1/2)^2 \\
&= 2 (2y_t-1)(f(x_t)-1/2) - 4(\log n)(f(x_t)-1/2)^2.
\end{align}
The lower bound in \eqref{eq:lowerbd} then becomes
\begin{align}
\label{eq:lowerbd}
V_n(\F^{1/n}) &\geq \multiminimax{\sup_{x_t\in\X_t(x_{1:t-1},y_{1:t-1})} \En_{y_t}}_{t=1}^n
\left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n 2 (2y_t-1)(f(x_t)-1/2) - 4(\log n)(f(x_t)-1/2)^2
\right\}
\right]\\
&= \sup_{\x} \En_{y} \left[
\sup_{f\in\F^{1/n}}
\left\{
\sum_{t=1}^n 2 (2y_t-1)(f(\x_t(y))-1/2) - 4(\log n)(f(\x_t(y))-1/2)^2
\right\}
\right]
\end{align}
where $y_1,\ldots,y_n$ are independent with distribution ${\sf Bernoulli(1/2)}$.
\end{proof}
\section{Discussion and Open Questions}
At the very first step, the analysis in this paper thresholds the class $\F$ to avoid dealing with the exploding gradient of the loss function. The authors believe that this ``hard thresholding'' approach is the source of sub-optimality, and that ``smooth'' approaches should be possible. When the class of functions has a specific structure, such as in the example of Section~\ref{sec:linear}, the exploding gradient can be mitigated in a ``smooth way'' by a regularization technique. It is not clear to the authors how to perform the ``smooth thresholding'' analysis when such a structure is not available.
Another interesting venue of investigation is the development of algorithms. It has been shown that the minimax analysis, of the type performed in this paper, can be made constructive \cite{rakhlin2012relax,RakSri15nonparametric,StatNotes2012}. It appears that the relaxation approach may yield new (and possibly computationally efficient) methods for sequential probability assignment and data compression.
|
\section{Introduction}
For multiprocessor sporadic task systems, the capacity-augmentation factor for global RM is proved to be $2.668$ in this paper. We also provide tighter analysis than the existing result by Bertogna et al. \cite{DBLP:conf/opodis/BertognaCL05} by considering both the total utilization of the task system and the maximum task utilization in the schedulability test.
\subsection{Improved Analysis for Global RM}
\label{sec:global-sporadic}
To get better analysis than the solution in Theorem~\ref{thm:multiprocessor-GRM}, we need a better sufficient test than the naive sufficient test in Lemma~\ref{lemma:gRM-sufficient}. The following lemma provides a sufficient test based on the observations by Baruah et al. \cite{DBLP:journals/rts/BaruahBMS10}, in which the proof is in Appendix B.
\begin{lemma}
\label{lemma:multiprocessor-grm-sporadic-pushforward}
Let $U_k^{\max}$ be $\max_{j=1}^{k} U_j$.
Task $\tau_k$ in a sporadic task system with implicit deadlines is schedulable by global RM on
$M$ processors if
\begin{align*}
\forall y \geq 0, \forall 0 \leq \omega_i \leq T_i, \exists t \mbox{ with } 0 < t \leq T_k+y {\;\; and \;\;} \\
U_k^{\max} \cdot(T_k+y) +
\frac{ \sum_{i=1}^{k-1} \omega_i\cdot U_i+ \ceiling{\frac{t-\omega_i }{T_i}} C_i
}{M}\leq t.
\end{align*}
\end{lemma}
The schedulability condition in Lemma~\ref{lemma:multiprocessor-grm-sporadic-pushforward} requires to test all possible $y \geq 0$ and all possible settings of $0 \leq \omega_i \leq T_i$ for the higher priority tasks $\tau_i$ with $i=1,2,\ldots,k-1$. At first glance, it may seem to be difficult to use the \framework{} framework. It requires some further annotation before adopting the framework, as shown in the proof of the following theorem.
\begin{theorem}
\label{thm:multiprocessor-grm-sporadic-tight}
Let $U_k^{\max}$ be $\max_{j=1}^{k} U_j$.
Task $\tau_k$ in a sporadic task system with implicit deadlines is schedulable by global RM on
$M$ processors if
\begin{equation}
\label{eq:schedulability-GRM-tight}
(U_k^{\max} +1) \prod_{j=1}^{k-1} (\frac{U_j}{M} + 1)\leq 2,
\end{equation}
or
\begin{equation}
\label{eq:schedulability-GRM-tight-ubound}
\sum_{j=1}^{k-1} \frac{U_j}{M}\leq \ln{\frac{2}{U_k^{\max} +1}}.
\end{equation}
\end{theorem}
\begin{proof}
We will show that the schedulability condition in the theorem holds for all possible settings of $y$ and $\omega_i$s.
Suppose that $y$ and $\omega_i$ for $i=1,2,\ldots,k-1$ are given, in which $y \geq 0$ and $0 \leq \omega_i \leq T_i$. Let $t_k$ be $T_k+y$. Now, we set $t_i$ to $\omega_i + \floor{\frac{T_k+y-\omega_i}{T_i}}T_i$ for $i=1,2,\ldots,k-1$ and reindex the tasks such that $t_1 \leq t_2 \leq \ldots \leq t_k$. The sufficient schedulability condition in Lemma~\ref{lemma:multiprocessor-grm-sporadic-pushforward} under the given $y$ and $\omega_i$s is to verify the existence of $t_j \in
\setof{t_1, t_2, \ldots t_k}$ such that
\begin{align}
&U_k^{\max}(T_k+y) + \frac{\sum_{i=1}^{k-1} \omega_i\cdot U_i+ \ceiling{\frac{t_j-\omega_i }{T_i}} C_i}{M} \\
\leq & U_k^{\max} t_k + \frac{\sum_{i=1}^{k-1} U_i t_i + \sum_{i=1}^{j-1} C_i}{M}
\leq t_j.
\end{align}
\jj{bug from here!!!}
By the definition of $t_i$ and $y \geq 0$ and the global RM scheduling (i.e., $T_k \geq T_i$), we can conclude that $t_i \geq T_k+y \geq T_i$, which implies that $C_i \leq U_i t_i$.
Hence, we know that the sufficient test is to verify whether there exists
$t_j \in
\setof{t_1, t_2, \ldots t_k}$ such that
\begin{align}
\label{eq:final-grm-k2u-precondition}
U_k^{\max} t_k + \sum_{i=1}^{k-1} \alpha_i U_i t_i + \sum_{i=1}^{j-1} \beta_i U_i t_i
\leq t_j,
\end{align}
where $\alpha_i = \frac{1}{M}$ and $\beta_i \leq \frac{1}{M}$ for $i=1,2,\ldots,k-1$.
Therefore, we reach the conclusion of the schedulability conditions in Eqs.~\eqref{eq:schedulability-GRM-tight} and \eqref{eq:schedulability-GRM-tight-ubound} by Lemma~\ref{lemma:framework-constrained} and
Lemma~\ref{lemma:framework-totalU-exclusive} under given $y$ and $\omega_i$s.\footnote{The concern that $y$ may be unbounded (so may $t_k$) can be handled by normalizing all the $t_i$ values for higher priority tasks $\tau_i$s to $t_k$ and set $t_k$ to $1$ after the normalization. }
The schedulability test in Eq.~\eqref{eq:final-grm-k2u-precondition} is independent from the settings of $y$ and $\omega_i$s. However, the setting of $y$ and $\omega_i$s affects how the $k-1$ higher priority tasks are indexed. Fortunately, it can be observed that the schedulability tests in Eqs.~\eqref{eq:schedulability-GRM-tight} and \eqref{eq:schedulability-GRM-tight-ubound} are completely independent upon the indexing of the higher priority tasks. Therefore, no matter how $y$ and $\omega_i$s are set, the schedulability conditions in Eqs.~\eqref{eq:schedulability-GRM-tight} and \eqref{eq:schedulability-GRM-tight-ubound} are the corresponding results from the \framework{} framework.
As a result, we can reach the conclusion.
\end{proof}
\begin{corollary}
\label{col:grm-tight}
The capacity augmentation factor of global RM for a sporadic system with implicit deadlines is $2.668$.
\end{corollary}
\begin{proof}
Suppose that $\sum_{\tau_i} \frac{C_i}{M T_i} \leq \frac{1}{b}$ and
$U_k \leq \max_{\tau_i} U_i \leq
\frac{1}{b}$. Therefore, by
Eq.~\eqref{eq:schedulability-GRM-tight-ubound}, we can guarantee
the schedulability of task $\tau_k$ if $\frac{1}{b} \leq
\ln{\frac{2}{1+\frac{1}{b}}}$. This is equivalent to solving $x =
\ln{\frac{2}{1+x}}$, which holds when $x\approx
2.668$ by solving the equation numerically. Therefore,
we reach the conclusion of the capacity augmentation factor
$2.668$.
\end{proof}
{\bf Remarks:} The utilization bound in Eq.~\eqref{eq:schedulability-GRM-tight-ubound} is analytically better than the best known utilization-based schedulability test $\sum_{j=1}^{k} U_j\leq \frac{M}{2}(1-U_k^{\max})+U_k^{\max}$ for global RM by Bertogna et al. \cite{DBLP:conf/opodis/BertognaCL05}, since $\frac{(1-x)}{2} \leq \ln\frac{2}{1+x}$ when $0 \leq x \leq 1$. The capacity augmentation factor $2.668$ in Corollary \ref{col:grm-tight} is the same as the result by Lundberg \cite{DBLP:conf/rtas/Lundberg02}. As shown in \cite{DBLP:conf/rtas/Lundberg02}, the factor $2.668$ is tight when the number of tasks is sufficiently large.
We would like to point out the incompleteness in the proof in \cite{DBLP:conf/rtas/Lundberg02}. In the proof of the extreme task set, the argument in Page 150 in \cite{DBLP:conf/rtas/Lundberg02} concludes that task $\tau_n$ is more difficult to be schedulable due to the increased interference of task $\tau_{n-1}$ to task $\tau_n$ after the transformation. The argument was not correctly proved since the increased interference has to be analyzed in all time points in the analysis window, whereas the analysis in \cite{DBLP:conf/rtas/Lundberg02} only considers the interference in a specific interval length. Without analyzing the resulting interference in all time points in the analysis window, task $\tau_n$ after transformation may still have chance to finish earlier due to the potential reduction of the interference at earlier time points.
We limit our attention here for the global RM scheduling. Andersson et al. \cite{DBLP:conf/rtss/AnderssonBJ01} propose the RM-US[$\varsigma$] algorithm, which gives the highest priority to tasks $\tau_i$s with $U_i > \varsigma$, and otherwise assigns priorities by using RM. Our analysis here can also be applied for the RM-US[$\varsigma$] algorithm with some modifications in the proofs by setting $\varsigma=\frac{1}{2.668}\approx 0.37482$.
\begin{appProof}{Lemma~\ref{lemma:multiprocessor-grm-sporadic-pushforward}}
This is proved by contrapositive. If task $\tau_k$ is not schedulable by global RM, we will show that there exist $y \geq 0$ and $0 \leq \omega_i \leq T_i$ such that for all $0 < t \leq T_k+y$, the condition $U_k^{\max} \cdot (T_k+y) +
\frac{ \sum_{i=1}^{k-1} \omega_i\cdot U_i+ \ceiling{\frac{t-\omega_i }{T_i}} C_i
}{M} > t$ holds. The proof is mainly based on the forwarding algorithm for the analysis of global DM by Baruah et al. in \cite{DBLP:journals/rts/BaruahBMS10}, by making some further annotations.
If $\tau_k$ is not schedulable, let $z_0$ be the first time that task $\tau_k$ misses its absolute deadline, i.e., $z_0$. Let $z_1$ be the arrival time of this job of task $\tau_k$. For notational brevity, let this job be $J_1$, which arrives at time $z_1$ and has not yet been finished at time $z_0$. By definition, we know that $z_0-z_1$ is $T_k$. Due to the scheduling policy and the implicit-deadline setting, removing (1) all the other jobs of task $\tau_k$ (except the one arriving at time $z_1$), (2) all the jobs arriving no earlier than $z_0$, and (3) lower-priority jobs does not change the unschedulability of job $J_1$. Therefore, all the jobs that are considered for the rest of the proof come from $\tau_1, \tau_2, \ldots, \tau_k$.
Now, we expand the interval of interest by using a slightly different algorithm from that proposed in \cite{DBLP:journals/rts/BaruahBMS10}, also illustrated with the notations in Figure~\ref{fig:force-forward}, as follows:
\begin{algorithm}[h]
\label{alg:feasibility-analysis2}
\begin{algorithmic}[1]\small
\FOR {$\ell\leftarrow2,3,...$}
\STATE let $J_\ell$ denote a job that
\begin{compactitem}[-]
\item arrives at some time-instant $z_\ell<z_{\ell-1}$;
\item has an absolute deadline after $z_{\ell-1}$;
\item has not completed execution by $z_{\ell-1}$; and
\item has executed for strictly less than $(z_{\ell-1}-z_\ell)\cdot \hat{U}_\ell$
units over the interval $[z_\ell,z_{\ell-1})$, where $\hat{U}_\ell$ is the utilization of the task that generates job $J_\ell$.
\end{compactitem}
\IF {there is no such a job}
\STATE $\ell\leftarrow (\ell-1)$; break;
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
Suppose that the forwarding algorithm terminates with $\ell$ equals to $\ell^*$. We now examine the schedule in the interval $(z_{\ell^*}, z_0]$. Since $J_1, J_2, \ldots, J_{\ell^*}$ belong to $\tau_1, \tau_2, \ldots, \tau_k$, we know that $\hat{U}_\ell \leq U_k^{\max}$ for $\ell=1,2,\ldots,\ell^*$. Let $\sigma_\ell$ be the total length of the time interval over $(z_\ell, z_{\ell-1}]$ during which $J_\ell$ is executed. By the choice of $J_\ell$, it follows that
\[
\sigma_\ell < (z_{\ell-1}-z_{\ell}) \cdot \hat{U}_\ell.
\]
Moreover, all the $M$ processors execute other higher-priority jobs (than $J_\ell$) at any other time points in the interval $(z_\ell, z_{\ell-1}]$ at which $J_\ell$ is not executed.
Therefore, by the above analysis, we know that the maximum amount of time from $z_{\ell^*}$ to $z_0$, in which not all the $M$ processors execute certain jobs, is at most
\begin{align*}
\sum_{\ell=1}^{\ell^*} \sigma_\ell &< \sum_{\ell=1}^{\ell^*} (z_{\ell-1}-z_{\ell}) \cdot \hat{U}_\ell\\
&\leq
\sum_{\ell=1}^{\ell^*} (z_{\ell-1}-z_{\ell}) \cdot U_k^{\max} = (z_0-z_{\ell^*}) \cdot U_k^{\max}.
\end{align*}
Now, we replace job $J_0$ with job $J_0'$ in the above schedule, where $J_0'$ is released at time $z_{\ell^*}$ with absolute deadline $z_0$ and execution time $(z_0-z_{\ell^*}) \cdot U_k^{\max}$. According to the above analysis, $J_0'$ cannot be finished before $z_{0}$ in the above schedule.
For each task $\tau_i$ for $i=1,2,\ldots,k-1$, in the above schedule, there may be one \emph{carry-in} job, denoted as $J_i^\flat$, of $\tau_i$ that arrives at time $r_i^\flat$ with $r_i^\flat < z_{\ell^*}$ and has absolute deadline after $d_i^\flat$ with $d_i^\flat > z_{\ell^*}$. According to the termination condition in the construction of $z_{\ell^*}$, if $J_i^\flat$ exists, we know that at least $(z_{\ell^*}-r_i^\flat)\cdot U_i$ amount of execution time has been executed before $z_{\ell^*}$, and the remaining execution time of job $J_i^\flat$ to be executed after $z_{\ell^*}$ is at most
$(d_i^\flat - z_{\ell^*})\cdot U_i$. If $J_i^\flat$ does not exist, then $d_i^\flat$ is set to $z_{\ell^*}$ for notational brevity. Therefore, the amount of workload $W_i'(t)$ of all the released jobs of task $\tau_i$ for $i=1,2,\ldots,k-1$ to be executed in time interval $(z_\ell^*, z_\ell^*+t)$ is at most
\begin{equation}
\label{eq:1}
W_i'(t) = (d_i^\flat - z_{\ell^*})\cdot U_i+ \ceiling{\frac{t-(d_i^\flat - z_{\ell^*}) }{T_i}} C_i.
\end{equation}
Clearly, this is more precise than the naive time demand function $W_i(t)$ in Eq.~\eqref{eq:W_i-multiprocessor}.
The assumption of the unschedulability of job $J_0'$ (due to the unschedulability of job $J_0$) under global RM implies that $J_0'$ cannot finish its computation at any time between $z_{\ell^*}$ and $z_0$. This leads to the following (necessary) condition
$U_k^{\max} (z_0-z_{\ell^*}) +
\frac{ \sum_{\tau_i \in hp(\tau_k)}W_i'(t) }{M} > t$ for all $0 < t \leq z_{\ell^*}-z_0$ for the unschedulability of job $J_0'$. Therefore, by the existence of $y=z_{0}-z_{\ell^*}$ and $\omega_i = d_i^\flat - z_{\ell^*}$ for $i=1,2,\ldots, k-1$ to force the above necessary condition, we reach the conclusion of the proof by contrapositive.
\end{appProof}
\begin{figure}[t]
\centering
\scalebox{0.45}{
\begin{tikzpicture}[x=1cm,font=\sffamily\Large,ultra thick]]
\draw[->](0,0) -- (19, 0) node[right]{$x$};
\draw[-](2,-.1) node[below]{$z_\ell$} -- (2, 1) -- (5,1) node[pos=.5,above]{$J_\ell$} -- (5,0) ;
\draw[-](4,-.1) node[below]{$z_{\ell-1}$} -- (4, 1.2) -- (7,1.2) node[pos=.5,above]{$J_{\ell-1}$};
\draw[-](10,1) node[above]{$J_{3}$} -- (13,1) -- (13,0) ;
\draw[-](12,-.1) node[below]{$z_2$} -- (12, 1.2) -- (16,1.2) node[pos=.5,above]{$J_2$} -- (16,0) ;
\draw[-](15,-.1) node[below]{$z_1$} -- (15, 1) -- (18,1) node[pos=.5,above]{$J_1$} -- (18,0)node[below]{$z_0$} ;
\node [ ] at (8,0.6) {$\cdot\cdot\cdot\cdot\cdot$};
\node [ ] at (1,0.6) {$\cdot\cdot\cdot\cdot\cdot$};
\end{tikzpicture} }
\caption{Notations of the forwarding algorithm for the analysis of global RM.}
\label{fig:force-forward}
\end{figure}
\subsection{Response Time Bound in Fixed-Priority Scheduling with Arbitrary Deadlines}
\begin{definition}[Last Release]
For a given task $\tau_i$ in the interval
$[\phi,\phi+D_n)$ for any $\phi \ge 0$, $t_i$ denotes the time of the last release of task $\tau_i$.
\begin{lemma}
\label{lemma:response-time-unschedulability}
Let $X^*$ denote the time at which the $k^{th} job$ of $\tau_n$ misses its deadline.
If the $k^{th}$ job of task $\tau_n$ misses its deadline, then
\begin{equation*}
\forall \tau_i,\quad kC_n+\sum_{i=1}^{n-1}U_it_i+\sum_{j=1}^{i-1}C_j> t_i
\end{equation*}
\end{lemma}
\begin{proof}
\end{proof}
\end{definition}
\begin{definition}[Worst-Case Response Time]
$R_n=X^*-(k-1)T_n$
\end{definition}
\begin{theorem}
The worst-case reponse time $R_n$ of task $\tau_n$ is bounded above by
\begin{align*}
R_n\le \frac{C_n+\sum_{i=1}^{n-1}C_i-\sum_{i=1}^{n-1}(U_i\sum_{j=i}^{n-1}C_j)}{1-\sum_{i=1}^{n-1}U_i}
\end{align*}
where tasks are indexed with
\begin{equation*}
\forall \tau_i \quad \frac{C_{i+1}}{U_{i+1}}<\frac{C_{i}}{U_{i}}.
\end{equation*}
\end{theorem}
\begin{proof}
We prove this theorem by finding the supremum $R_n$ subject to the conditions in
Lemma~\ref{lemma:response-time-unschedulability}.
This is a strictly convex problem. Since we have more variables than constraints,
all of the constraints have to be \emph{active} according to the extreme point theorem.
\begin{equation*}
kC_n+\sum_{i=1}^{n-1}U_i(t_1+\sum_{j=1}^{i-1}C_j)= t_1
\end{equation*}
\begin{align*}
t_1&=\frac{kC_n+\sum_{i=1}^{n-1}(U_i\sum_{j=1}^{i-1}C_j)}{1-\sum_{i=1}^{n-1}U_i}\\
&\ge 0 \quad \text{($\sum_{i=1}^nU_i<1$)}
\end{align*}
Then, the response time is bounded by
\begin{equation}
\label{eq:response-time-supermum-k}
R_n\le \frac{kC_n+\sum_{i=1}^{n-1}C_i-\sum_{i=1}^{n-1}(U_i\sum_{j=i}^{n-1}C_j)}{1-\sum_{i=1}^{n-1}U_i}-(k-1)T_n.
\end{equation}
We now prove the maximum in Eq. (\ref{eq:response-time-supermum-k}) occur for $k=1$.
\begin{align*}
\frac{d}{dk}&\left(\frac{kC_n+\sum_{i=1}^{n-1}C_i-\sum_{i=1}^{n-1}(U_i\sum_{j=i}^{n-1}C_j)}{1-\sum_{i=1}^{n-1}U_i}-(k-1)T_n\right)\\
&=\frac{C_i}{1-\sum_{i=1}^{n-1}U_i}-T_n\\
&=T_n(\frac{U_i}{1-\sum_{i=1}^{n-1}U_i}-1)\\
&=T_n(\frac{\sum_{i=1}^{n}U_i-1}{1-\sum_{i=1}^{n-1}U_i})
\end{align*}
which is negative since our assumption $\sum_{i=1}^{n}U_i<1$. Since this function is decreasing, the maximum occurs in the left bound, which means $k=1$.
By substituting $k=1$ in Eq. (\ref{eq:response-time-supermum-k}), this theorem follows.
\end{proof}
\subsection{Response Time Bound in Fixed-Priority Scheduling with Arbitrary Deadlines with unknown $C$}
\begin{lemma}
Let $X^*$ denote the time at which the $k^{th} job$ of $\tau_n$ misses its deadline.
If the $k^{th}$ job of task $\tau_n$ misses its deadline, then
\begin{equation*}
\forall \tau_i,\quad kC_n+\sum_{i=1}^{n-1}U_it_i+\sum_{j=1}^{i-1}U_jt_j> t_i
\end{equation*}
\end{lemma}
\begin{theorem}
If
\begin{equation*}
\prod_{i=1}^{n} (U_i + 1)\le2,
\end{equation*}
then the worst-case reponse time $R_n$ of task $\tau_n$ is bounded above by
\begin{equation*}
R_n\le \frac{C_n}{\frac{2}{\prod_{i=1}^{n-1} (U_i + 1)}-1}.
\end{equation*}
\end{theorem}
\begin{proof}
\begin{equation}
t_{i}U_i= t_{i+1}-t_i.
\end{equation}
\begin{align*}
&kC_n+\sum_{i=1}^{n-1}U_it_i= t_1\\
\Rightarrow &kC_n+t_n-t_1= t_1\\
\Rightarrow &kC_n+t_1\prod_{i=1}^{n-1} (U_i + 1)-t_1= t_1\\
\Rightarrow &t_1=\frac{kC_n}{2-\prod_{i=1}^{n-1} (U_i + 1)}
\end{align*}
\begin{align*}
R_n&\le kC_n+2\sum_{i=1}^{n-1}U_it_i-(k-1)T_n\\
&= kC_n+2(t_1\prod_{i=1}^{n-1} (U_i + 1)-t_1)-(k-1)T_n\\
&=kC_n+\frac{2kC_n(\prod_{i=1}^{n-1} (U_i + 1)-1)}{2-\prod_{i=1}^{n-1} (U_i + 1)}-(k-1)T_n\\
&=\frac{kC_n\prod_{i=1}^{n-1} (U_i + 1)}{2-\prod_{i=1}^{n-1} (U_i + 1)}-(k-1)T_n\\
&=\frac{kC_n}{\frac{2}{\prod_{i=1}^{n-1} (U_i + 1)}-1}-(k-1)T_n\\
\end{align*}
We now prove the maximum in Eq. (\ref{eq:response-time-supermum-k}) occur for $k=1$.
\begin{align*}
\frac{d}{dk}&\left(\frac{kC_n}{\frac{2}{\prod_{i=1}^{n-1} (U_i + 1)}-1}-(k-1)T_n\right)\\
&=\frac{C_n}{\frac{2}{\prod_{i=1}^{n-1} (U_i + 1)}-1}-T_n\\
&=T_n(\frac{U_n}{\frac{2}{\prod_{i=1}^{n-1} (U_i + 1)}-1}-1)\\
&=T_n(\frac{U_n\prod_{i=1}^{n-1} (U_i + 1)}{2-\prod_{i=1}^{n-1} (U_i + 1)}-1)\\
&=T_n(\frac{U_n\prod_{i=1}^{n-1} (U_i + 1)-2+\prod_{i=1}^{n-1} (U_i + 1)}{2-\prod_{i=1}^{n-1} (U_i + 1)})\\
&=T_n(\frac{\prod_{i=1}^{n} (U_i + 1)-2}{2-\prod_{i=1}^{n-1} (U_i + 1)})\\
\end{align*}
\end{proof}
\section{Introduction}
\label{sec:intro}
Given the emerging trend towards building complex cyber-physical
systems that often integrate external and physical devices, many
real-time and embedded systems are expected to handle a large variety
of workloads. Different formal real-time task models have been
developed to accurately represent these workloads with various
characteristics. Examples include the sporadic task
model~\cite{mok1983fundamental}, the multi-frame task
model~\cite{DBLP:dblp_journals/tse/MokC97}, the self-suspending task
model~\cite{suspension}, the directed-acyclic-graph (DAG) task model,
etc. Many of such formal models have been shown to be expressive
enough to accurately model real systems in practice. For example, the
DAG task model has been used to represent many computation-parallel
multimedia application systems and the self-suspending task model is
suitable to model workloads that may interact with I/O devices. For each of these task models, researchers tend to develop different types of techniques that result in schedulability tests with different computation complexity and performance (e.g., different utilization bounds).
\ifbool{techreport}{Over the years, real-time systems researchers have devoted a significant amount of time and efforts to efficiently analyze different task models. Many successful stories have been told. For many of the above-mentioned task models, efficient scheduling and schedulability analysis techniques have been developed (see \cite{DavisSurvey2011} for a recent survey). Unfortunately, for certain complex models such as the self-suspending task model, existing schedulability tests are rather pessimistic, particularly for the multiprocessor case (e.g., no utilization-based schedulability test exists for globally-scheduled multiprocessor self-suspending task systems). }{}
In this paper, we present \framework{}, a general schedulability analysis framework that is fundamentally based on a $k$-point effective schedulability test under fixed-priority scheduling. The key observation behind our proposed $k$-point test is the following. Traditional fixed-priority schedulability tests often have pseudo-polynomial-time (or even higher) complexity. For example, to verify the schedulability of a (constrained-deadline) task $\tau_k$ under fixed-priority scheduling in uniprocessor systems, the time-demand analysis (TDA) developed in \cite{DBLP:conf/rtss/LehoczkySD89}
can be adopted. That is, if
\begin{equation}
\label{eq:exact-test-constrained-deadline}
\exists t \mbox{ with } 0 < t \leq D_k {\;\; and \;\;} C_k +
\sum_{\tau_i \in hp(\tau_k)} \ceiling{\frac{t}{T_i}}C_i \leq t,
\end{equation}
then task $\tau_k$ is schedulable under the fixed-priority scheduling algorithm, where $hp(\tau_k)$ is the set of the tasks with higher priority than $\tau_k$, $D_k$, $C_k$, and $T_i$ represent $\tau_k$'s relative deadline, worst-case execution time, and period, respectively. TDA incurs pseudo-polynomial-time complexity to check the time points that lie in $(0, D_k]$ for Eq.~\eqref{eq:exact-test-constrained-deadline}.
To obtain sufficient schedulability tests under fixed priority
scheduling with reduced time complexity (e.g., polynomial-time), our
conceptual innovation is based on the observations by \emph{testing only a subset of such points} to
\emph{derive the minimum $C_k$ that cannot pass the
schedulability tests}. This idea is
implemented in the \framework{} framework by providing a general
$k$-point effective schedulability test, which only needs to test $k$
points under \textit{any} fixed-priority scheduling when checking
schedulability of the task with the $k^{th}$ highest priority in the
system. This $k$-point effective schedulability test can be viewed as
a ``blackbox" interface that can result in sufficient
utilization-based tests.
We show the generality of \framework{} by applying it to analyze several concrete example task models, including the constrained- and arbitrary-deadline sporadic task models, the multi-frame task model, the self-suspending task model, and the DAG task model. Note that \framework{} is not only applicable to uniprocessor systems, but also applicable to multiprocessor systems.
\noindent\textbf{Related Work.}
\ifbool{techreport}{An extensive amount of research has been conducted over the past forty years on verifying the schedulability of the classical sporadic task model in both uniprocessor and multiprocessor systems (see \cite{DavisSurvey2011} for a survey of such results). Much progress has also been made in recent years on analyzing more complex task models that are more expressive.}{}
There have been several results in the literature with respect to utilization-based, e.g., \cite{liu1973scheduling,HanTyan-RTSS97,journals/tc/LeeSP04,DBLP:conf/rtas/WuLZ05,kuo2003efficient,bini2003rate,RTSS14a}, \ifbool{techreport}{and non-utilization-based, e.g., \cite{DBLP:conf/rtss/ChakrabortyKT02,DBLP:conf/ecrts/FisherB05}, }{}
schedulability tests for the sporadic real-time task model and its generalizations in uniprocessor systems.
\ifbool{techreport}{
The approaches in \cite{DBLP:conf/rtss/ChakrabortyKT02,DBLP:conf/ecrts/FisherB05} convert the stair function $\ceiling{\frac{t}{T_i}}$ in the time-demand analysis into a linear function if $t$ in Eq.~\eqref{eq:exact-test-constrained-deadline} is large enough. The methods in \cite{DBLP:conf/rtss/ChakrabortyKT02,DBLP:conf/ecrts/FisherB05} are completely different from this paper, in which the linear function of task $\tau_i$ starts after $\frac{t}{T_i} \geq k$, in which our method is based on $k$ points, defined individually by $\tau_1, \tau_2, \ldots, \tau_k$.
}{}
Most of the existing utilization-based schedulability analyses focus on the total utilization bound. That is, if the total utilization of the task system is no more than the derived bound, the task system is schedulable by the scheduling policy. For example, the total utilization bounds derived in \cite{liu1973scheduling,HanTyan-RTSS97,DBLP:dblp_journals/tc/BurchardLOS95} are mainly for rate-monotonic (RM) scheduling, in which the results in \cite{HanTyan-RTSS97} can be extended for arbitrary fixed-priority scheduling. Kuo et al. \cite{kuo2003efficient} further improve the total utilization bound by using the notion of divisibility. Lee et al. \cite{journals/tc/LeeSP04} use linear programming formulations for calculating total utilization bounds
when the period of a task can be selected. Moreover, Wu et al. \cite{DBLP:conf/rtas/WuLZ05} adopt the Network Calculus to analyze the total utilization bounds of several task models.
The novelty of \framework{} comes from a different perspective
from these approaches
\cite{liu1973scheduling,HanTyan-RTSS97,journals/tc/LeeSP04,DBLP:conf/rtas/WuLZ05,kuo2003efficient}. We
do not specifically seek for the total utilization bound. Instead, we
look for the critical value in the specified sufficient schedulability
test while verifying the schedulability of task $\tau_k$.
A natural schedulability condition to express
the schedulability of task $\tau_k$ is a hyperbolic bound, (to be
shown in Lemma \ref{lemma:framework-constrained}), whereas the
corresponding total utilization bound can be obtained (in Lemmas
\ref{lemma:framework-totalU-constrained} and
\ref{lemma:framework-totalU-exclusive}).
The hyperbolic forms are the centric features in \framework{} analysis, in which the test by Bini et al. \cite{bini2003rate} for sporadic real-time tasks and our recent result in \cite{RTSS14a} for bursty-interference analysis are both special cases and simple implications from the \framework{} framework. With the hyperbolic forms, we are then able to provide many interesting observations with respect to the required quantitive features to be measured, like the total utilization bounds, speed-up factors, etc., not only for uniprocessor scheduling but also for multiprocessor scheduling.
For more details, we will provide further explanations at the end of Sec. \ref{sec:framework} after the framework is presented. For the studied task models to demonstrate the applicability of \framework{}, we will summarize some of the latest results on these task models in their corresponding sections.
\noindent\textbf{Contributions.} In this paper, we present a general
schedulability analysis framework, \framework{}, that can be applied
to analyze a number of complex real-time task models, on both
uniprocessors and multiprocessors.
For any
task model, if a corresponding $k$-point effective schedulability test
can be constructed, then a sufficient utilization-based test can be
derived by the \framework{} framework. We show the generality of \framework{} by applying it to several task models, in which the results are better or more general compared to the state-of-the-art:
\begin{enumerate}
\item For uniprocessor constrained-deadline sporadic task systems, the speed-up factor of our obtained schedulability test is 1.76322. This value is the same as the lower bound and upper bound of deadline-monotonic (DM) scheduling shown by Davis et al.~\cite{DBLP:journals/rts/DavisRBB09}. Our result is thus stronger (and requires a much simpler proof), as we show that the same factor holds for a polynomial-time schedulability test (not just the DM scheduler). For uniprocessor arbitrary-deadline sporadic task systems, our obtained utilization-based test works for any fixed-priority scheduling with arbitrary priority-ordering assignment.
\item For multiprocessor DAG task systems under global rate-monotonic (RM) scheduling, the capacity-augmentation
factor, as defined in ~\cite{Li:ECRTS14} and Sec.~\ref{sec:multiprocessor} in this paper, of our obtained test is
3.62143. This result is better than the best existing result, which
is 3.73, given by Li et al.~\cite{Li:ECRTS14}. Our result is also applicable for conditional sporadic DAG task systems \cite{DBLP:conf/ecrts/BaruahBM15}.
\item For multiprocessor self-suspending task systems, we obtain the \textit{first} utilization-based test for global RM.
\item For uniprocessor multi-frame task systems, our obtained utilization bound is superior to the results by Mok and Chen~\cite{DBLP:dblp_journals/tse/MokC97} analytically and Lu et al. \cite{lu2007new} in our simulations. \ifbool{techreport}{The analysis is in Appendix C, whereas the evaluation result is in Appendix D.}{Due to space limitations, the analysis is in Appendix C in the report \cite{DBLP:journals/corr/abs-1501.07084}, whereas the evaluation result is in Appendix D in \cite{DBLP:journals/corr/abs-1501.07084}.}
\end{enumerate}
Note that the emphasis of this paper is not to show that the resulting tests for different task models by applying the \framework{} framework are better than existing work. Rather, we want to show that the \framework{} framework is general, easy to use, and has relatively low time complexity, but is still able to generate good tests. By demonstrating the applicability of the \framework{} framework to several task models, we believe that this framework has great potential in analyzing many other complex real-time task models, where the existing analysis approaches are insufficient or cumbersome. To the best of our knowledge, together with \frameworkkq{} to be explained later, these are the first general schedulability analysis frameworks that can be potentially applied to analyze a large set of real-time task models under any fixed-priority scheduling algorithm in both uniprocessor and multiprocessor systems.
\noindent{\bf Comparison to \frameworkkq{}:}
The concept of testing $k$ points only is also the key in
another framework designed by us, called \frameworkkq{}
\cite{DBLP:journals/corr/abs-k2q}. Even though \frameworkkq{} and \framework{} share the
same idea by testing and evaluating only $k$ points, they are based on
completely different criteria for testing. In \framework{}, all the testings and
formulations are based on \emph{only the higher-priority task
utilizations}. In \frameworkkq{}, the testings are based \emph{not
only on the higher-priority task utilizations, but also on the
higher-priority task execution times}. The above difference
in the formulations results in completely different properties and
mathematical closed-forms. The
natural schedulability condition of \framework{} is a \emph{hyperbolic form} for testing
the schedulability, whereas the natural schedulability condition of \frameworkkq{} is a
\emph{quadratic form} for testing the schedulability or the response
time of a task.
\emph{If one framework were dominated by another or these two frameworks
were just with minor difference in mathematical formulations, it
wouldn't be necessary to separate and present
them as two different frameworks.} Both frameworks are in fact needed
and have to be applied for different cases. Due to space limitation,
we can only shortly explain their differences,
advantages, and disadvantages in this paper. For completeness, another
document has been prepared in
\cite{DBLP:journals/corr/framework-compare} to present the similarity,
the difference and the characteristics of these two frameworks in details.
Since the formulation of \framework{} is more restrictive than
\frameworkkq{}, its applicability is limited by the possibility to
formulate the tests purely by using higher-priority task utilizations
without referring to their execution times. There are cases, in which
formulating the higher-priority interference by using only task
utilizations for \framework{} is troublesome. For such cases, further
introducing the upper bound of the execution time by using
\frameworkkq{} is more precise.
\ifbool{techreport}{The above cases can be found in (1)
the schedulability tests for arbitrary-deadline sporadic task systems
in uniprocessor scheduling, (2) multiprocessor global fixed-priority
scheduling when adopting the forced-forwarding schedulability test,
etc.\footnote{ c.f. Sec. 5/6 in \cite{DBLP:journals/corr/abs-k2q} for
the detailed proofs and Sec. 5/6 in
\cite{DBLP:journals/corr/framework-compare} for the performance
comparisons.}}{}
In general, if we can formulate the schedulability
tests into the \framework{} framework by purely using higher-priority
task utilizations, it is also usually possible to formulate it into
the \frameworkkq{} framework by further introducing the task execution
times. In such cases, the same pseudo-polynomial-time (or exponential
time) test is used, and the
utilization bound or speed-up factor analysis
derived from the \framework{} framework is, in
general, tighter and better.
\ifbool{techreport}{
In a nutshell, with respect to quantitive metrics, like utilization bounds or speedup factor analysis, \framework{} is more precise, whereas \frameworkkq{} is
more general. If the exact (or very precise) schedulability test can
be constructed and the test can be converted into \framework{},
e.g., uniprocessor scheduling for
constrained-deadline task sets, then, adopting \framework{} \emph{may} lead to
tight results. By adopting \frameworkkq{}, we may be able to start
from a more complicated test with exponential-time complexity and
convert it to a linear-time approximation with better results than
\framework{}. Although \framework{} is more restrictive than
\frameworkkq{}, both of them are general enough to cover a range of
applications, ranging from uniprocessor systems to
multiprocessor systems.
For more information and comparisons, please refer to \cite{DBLP:journals/corr/framework-compare}.
}
{
}
\section{Sporadic Task and Scheduling Models}
\label{sec:model}
A sporadic task $\tau_i$ is released repeatedly, with each such
invocation called a job. The $j^{th}$ job of $\tau_i$, denoted
$\tau_{i,j}$, is released at time $r_{i,j}$ and has an absolute deadline at time
$d_{i,j}$. Each job of any task $\tau_i$ is assumed to have
execution time $C_i$. Here in this paper, whenever we refer to
the execution time of a job, we mean for the worst-case execution time
of the job since all the analyses we use are safe by only considering the worst-case execution time. Successive jobs of the same task are
required to execute in sequence. Associated with each task $\tau_i$
are a period $T_i$, which specifies the minimum time between two
consecutive job releases of $\tau_i$, and a deadline $D_i$, which
specifies the relative deadline of each such job, i.e.,
$d_{i,j}=r_{i,j}+D_i$. The utilization of a task $\tau_i$ is defined
as $U_i=C_i/T_i$.
A sporadic task system $\tau$ is said to be an implicit-deadline
system if $D_i = T_i$ holds for each $\tau_i$. A sporadic task system
$\tau$ is said to be a constrained-deadline system if $D_i \leq T_i$
holds for each $\tau_i$. Otherwise, such a sporadic task system
$\tau$ is an arbitrary-deadline system.
A task is said \emph{schedulable} by a scheduling policy if all of its
jobs can finish before their absolute deadlines. A task system is
said \emph{schedulable} by a scheduling policy if all the tasks in the
task system are schedulable. A \emph{schedulability test}
expresses sufficient conditions to ensure the feasibility of the
resulting schedule by a scheduling policy.
Throughout the paper, we will focus on fixed-priority preemptive
scheduling. That is, each task is associated with a priority level.
For a uniprocessor system, i.e., except
Sec.~\ref{sec:multiprocessor}, the scheduler always dispatches the
job with the highest priority in the ready queue to be executed. For a
uniprocessor system, it has been shown that RM
scheduling is an optimal fixed-priority scheduling policy for
implicit-deadline systems \cite{liu1973scheduling} and
DM scheduling is an optimal fixed-priority
scheduling policy for constrained-deadline
systems\cite{journals/pe/LeungW82}.
To verify the schedulability of task $\tau_k$ under
fixed-priority scheduling in uniprocessor systems, the time-demand
analysis developed in \cite{DBLP:conf/rtss/LehoczkySD89}
can be adopted, as discussed earlier. That is, if Eq.~\eqref{eq:exact-test-constrained-deadline} holds,
then task $\tau_k$ is schedulable under the fixed-priority scheduling
algorithm.
For the simplicity of presentation,
we will demonstrate how the framework works using ordinary sporadic real-time task systems
in Sec.~\ref{sec:framework} and Sec.~\ref{sec:application-FP}. We
will demonstrate more applications with respect to multi-frame tasks
\cite{DBLP:dblp_journals/tse/MokC97} in Appendix C\citetechreport{}
and with respect to multiprocessor scheduling in
Sec.~\ref{sec:multiprocessor}.
\vspace{-2mm}
\section{Analysis Flow}
\label{sec:flow}
The proposed \framework{} framework only tests the
schedulability of a specific task $\tau_k$, under the assumption
that the higher-priority tasks are already verified to be schedulable
by the given scheduling policy. Therefore, this framework has to be
applied for each of the given tasks. A task system is schedulable by
the given scheduling policy only when all the tasks in the system can
be verified to meet their deadlines. The results can be extended to
test the schedulability of a task system in linear time complexity or to allow on-line admission control in constant time complexity if the schedulability
condition (or with some more pessimistic simplifications) is
monotonic. Such extensions are provided in Appendix A for some cases.
Therefore, for the rest of this paper, we implicitly assume that all
the higher-priority tasks are already verified to be schedulable by
the scheduling policy. We will only present the schedulability test
of a certain task $\tau_k$, that is being analyzed, under the
above assumption. For notational brevity, we implicitly assume that there
are $k-1$ tasks, say $\tau_1, \tau_2, \ldots, \tau_{k-1}$ with
higher-priority than task $\tau_k$. Moreover, we only consider the cases when $k \geq 2$, since $k=1$ is pretty trivial.
\section{\framework{} Framework}
\label{sec:framework}
This section presents the basic properties of the \framework{}
framework for testing the schedulability of task $\tau_k$ in a given
set of real-time tasks (depending on the specific models given in
each application as shown later in this paper). We will first provide examples
to explain and define the \emph{$k$-point effective schedulability
test}. Then, we will provide the fundamental properties of the
corresponding utilization-based tests. Throughout this section, we
will implicitly use sporadic task systems defined in
Sec.~\ref{sec:model} to simplify the presentation. The concrete
applications will be presented in Secs.~\ref{sec:application-FP} - \ref{sec:multiprocessor}.
The $k$-point effective schedulability test is a sufficient
schedulability test that verifies only $k$ time points, defined by the
$k-1$ higher-priority tasks and task $\tau_k$. For example, instead of
testing all the possible $t$ in the range of $0$ and $D_k$ in
Eq.~\eqref{eq:exact-test-constrained-deadline}, we can simply test
only $k$ points. It may seem to be very pessimistic to only test
$k$ points. However, if these $k$ points are \textit{effective},\footnote{As to be clearly illustrated later, the $k$ points can be considered effective if they can define certain extreme cases of task parameters. For example, the ``difficult-to-schedule'' concept first introduced by Liu and Layland~\cite{liu1973scheduling} defines $k$ effective points that are used in Example \ref{example-1}. In their case~\cite{liu1973scheduling}, the selected set of the $k$ points was ``very'' effective because the tested task $\tau_k$ becomes unschedulable if $C_k$ is increased by an arbitrarily small value $\varepsilon$.} the resulting
schedulability test may be already good. We now demonstrate two
examples.
\begin{example}
\label{example-1}
{\bf Implicit-deadline task systems}: Suppose that the tasks are indexed by
the periods, i.e., $T_1 \leq \cdots \leq
T_k$. When $T_k \leq 2T_1$, task $\tau_k$ is schedulable by RM if
there exists $j \in \setof{1,2,\ldots,k}$ where
\begin{equation}
\label{eq:precodition-schedulability-sporadic}\small
C_k + \sum_{i=1}^{k-1} C_i + \sum_{i=1}^{j-1} C_i = C_k + \sum_{i=1}^{k-1} T_i U_i + \sum_{i=1}^{j-1} T_i U_i \leq T_j.
\end{equation}\normalsize \endproof
\end{example}
That is, in the above example, it is sufficient to only test $T_1,
T_2, \ldots, T_k$. The case defined in the above example is utilized
by Liu and Layland \cite{liu1973scheduling} for deriving the least
utilization upper bound $69.3\%$ for RM scheduling. We can
make the above example more generalized as follows:
\begin{example}
\label{example-2}
{\bf Implicit-deadline task systems with given ratios of periods}:
Suppose that $f T_i \leq T_k$ for a given integer $f$ with $f \geq 1$
for any higher-priority task $\tau_i$, for all $i=1,2,\ldots,k-1$.
Let $t_i$ be
$\floor{\frac{T_k}{T_i}}T_i$. Suppose that the $k-1$ higher
priority tasks are indexed such that $t_1 \leq t_2 \leq \cdots \leq
t_{k-1} \leq t_k$, where $t_k$ is defined as $T_k$. Task $\tau_k$ is
schedulable under RM if there exists $j$ with $1 \leq j \leq k$ such that
\begin{equation}
\label{eq:precodition-schedulability-sporadic-f}
C_k + \sum_{i=1}^{k-1} t_i U_i + \sum_{i=1}^{j-1} C_i \leq C_k +
\sum_{i=1}^{k-1} t_i U_i + \sum_{i=1}^{j-1} \frac{1}{f} t_i U_i\leq t_j,
\end{equation}
where the first inequality in
Eq.~\eqref{eq:precodition-schedulability-sporadic-f} is due to the
fact $C_i = T_i U_i \leq \frac{1}{f} t_i U_i$. That is, in the above example, it is sufficient to only test
$\floor{\frac{T_k}{T_1}}T_1, \floor{\frac{T_k}{T_2}}T_2, \ldots,
\floor{\frac{T_k}{T_{k-1}}}T_{k-1}, T_k$.
\endproof
\end{example}
With the above examples, for a given set $\setof{t_1, t_2, \ldots
t_k}$, we now define the $k$-point effective schedulability test as follows:
\begin{definition}
\label{def:kpoints}
A $k$-point effective schedulability test is a sufficient
schedulability test of a fixed-priority scheduling policy, that verifies the existence of $t_j \in \setof{t_1, t_2, \ldots t_k}$ with $0 < t_1 \leq t_2 \leq \cdots \leq t_k$ such that \begin{equation}
\label{eq:precodition-schedulability}
C_k + \sum_{i=1}^{k-1} \alpha_i t_i U_i + \sum_{i=1}^{j-1} \beta_i t_i U_i \leq t_j,
\end{equation}
where $C_k > 0$, $\alpha_i > 0$, $U_i > 0$, and $\beta_i >0$ are dependent upon the setting
of the task models and task $\tau_i$.
\end{definition}
For Example~\ref{example-1}, the effective values in
$\setof{t_1, t_2, \ldots t_k}$ are $T_1, T_2, \ldots, T_k$, and
$\alpha_i=\beta_i=1$ for each task $\tau_i$. For
Example~\ref{example-2}, the effective values in $\setof{t_1,
t_2, \ldots t_k}$ are with $\alpha_i=1$ and $\beta_i \leq
\frac{1}{f}$ for each task $\tau_i$.
Moreover, we only consider non-trivial cases,
in which $C_k >0$, $t_k > 0$, $0 < \alpha_i$, $0 < \beta_i$, and $0 < U_i \leq 1$ for $i=1,2,\ldots,k-1$.
The definition of the $k$-point last-release schedulability test
$C_k + \sum_{i=1}^{k-1} \alpha_i t_i U_i + \sum_{i=1}^{j-1}
\beta_i t_i U_i \leq t_j$
in Definition \ref{def:kpoints} only slightly differs from the
test
$C_k +
\sum_{i=1}^{k-1} \alpha_i t_i U_i + \sum_{i=1}^{j-1} \beta_i C_i \leq
t_j$
in the \frameworkkq{} framework \cite{DBLP:journals/corr/abs-k2q}.
However, since the tests are different, they are used for different
situations.
With these $k$ points, we are able to define the corresponding
coefficients $\alpha_i$ and $\beta_i$ in the $k$-point effective
schedulability test of a scheduling algorithm. The elegance of the
\framework{} framework is to use only the parameters $\alpha_i$ and
$\beta_i$ to analyze whether task $\tau_k$ can pass the
schedulability test. Therefore, the \framework{} framework provides
corresponding utilization-based tests \emph{automatically} if the $k$-point effective
schedulability test and the corresponding parameters $\alpha_i$ and
$\beta_i$ can be defined, which will be further demonstrated in
the following sections with several applications.
We are going to present the properties
resulting from the $k$-point effective schedulability test under given
$\alpha_i$ and $\beta_i$. In the following lemmas, we are going to
seek the extreme cases for these $k$ testing points under the given
setting of utilizations and the defined coefficients $\alpha_i$ and
$\beta_i$. To make the notations clear, these extreme testing points
are denoted as $t_i^*$ for the rest of this paper. The procedure will
derive $k-1$ extreme testing points, denoted as $t_1^*, t_2^*, \ldots,
t_{k-1}^*$, whereas $t_k^*$ is defined as $t_k$ plus a slack variable (to be defined in the proof of Lemma~\ref{lemma:framework-constrained}) for notational
brevity.
Lemmas~\ref{lemma:framework-constrained} to \ref{lemma:framework-totalU-exclusive} are useful to analyze the utilization bound, the hyperbolic bound, etc., for given scheduling strategies, when
$\alpha$ and $\beta$ can be easily defined based on the scheduling policy, with
$0 < t_k$ and
$0 < \alpha_i \leq \alpha$, and $0 < \beta_i \leq \beta$ for any
$i=1,2,\ldots,k-1$.
\begin{lemma}
\label{lemma:framework-constrained}
For a given $k$-point effective schedulability test of a scheduling
algorithm, defined in
Definition~\ref{def:kpoints},
in which $0 < t_k$ and $0 < \alpha_i \leq \alpha$, and $0 < \beta_i \leq \beta$ for any
$i=1,2,\ldots,k-1$, task $\tau_k$ is schedulable by the scheduling
algorithm if the following condition holds
\begin{equation}
\label{eq:schedulability-constrained}
\frac{C_k}{t_k} \leq \frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)} - \frac{\alpha}{\beta}.
\end{equation}
\end{lemma}
\begin{proof}
If $\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)}
- \frac{\alpha}{\beta} \leq 0$, the condition in
Eq.~\eqref{eq:schedulability-constrained} never holds since $C_k > 0$, and the
statement is vacuously true. We focus on the case $\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1}
(\beta U_j + 1)} - \frac{\alpha}{\beta} > 0$.
We prove this lemma by showing that the condition in
Eq.~\eqref{eq:schedulability-constrained} leads to the satisfactions of the
schedulability condition listed in
Eq.~(\ref{eq:precodition-schedulability}) by using contrapositive.
By taking the negation of the schedulability condition in
Eq.~(\ref{eq:precodition-schedulability}), we know that if task
$\tau_k$ is \emph{not schedulable} by the scheduling policy, then
for each $j=1,2,\ldots, k$
\begin{equation}
\label{eq:lp-init-constraints}
C_k + \alpha \sum_{i=1}^{k-1} t_i U_i + \beta\sum_{i=1}^{j-1} t_i U_i \geq
C_k + \sum_{i=1}^{k-1} \alpha_i t_i U_i + \sum_{i=1}^{j-1} \beta_i t_i U_i > t_j,
\end{equation}
due to $0 < \alpha_i \leq \alpha$, and $0 < \beta_i \leq \beta$ for any
$i=1,2,\ldots,k-1$
To enforce the condition in
Eq.~\eqref{eq:lp-init-constraints}, we are going to
show that $C_k$ must have some lower bound. Therefore, if $C_k$ is
no more than this lower bound, then task $\tau_k$ is schedulable by the
scheduling policy.
For the rest of the proof, we replace $>$ with $\geq$ in
Eq.~\eqref{eq:lp-init-constraints}, as the infimum and the minimum
are the same when presenting the inequality with $\geq$. Moreover,
we also relax the problem by replacing the constraint $t_{j+1} \geq
t_j$ with $t_j \geq 0$ for $j=1,2,\ldots,k-1$. Therefore, the
unschedulability for satisfying Eq.~\eqref{eq:lp-init-constraints}
implies that $C_k > C_k^*$, where $C_k^*$ is
defined in the following optimization problem:
\begin{subequations}\label{eq:lp-init}
\small \begin{align}
\mbox{min\;\;} & C_k^* \label{eq:precodition-schedulability-objective}\\
\mbox{s.t.\;\;} & C_k^* + \sum_{i=1}^{k-1} \alpha t_i^* U_i + \sum_{i=1}^{j-1} \beta t_i^* U_i \geq t_j^* &\forall j=1,\ldots, k-1, \label{eq:precodition-schedulability-negation-0}\\
& t_j^* \geq 0\;&\forall j=1,\ldots, k-1, \label{eq:precodition-schedulability-negation-1}\\
& C_k^* + \sum_{i=1}^{k-1} (\alpha +\beta) t_i^* U_i \geq t_k, \label{eq:precodition-schedulability-negation-2}
\end{align}
\end{subequations}
where $t^*_1, t^*_2, \ldots, t^*_{k-1}$ and $C_k^*$ are variables; and $\alpha$, $\beta$ are constants.
Let $s \geq 0$ be a slack variable such that $C_k^* +
\sum_{i=1}^{k-1} (\alpha +\beta) t_i^* U_i = t_k+s$. Therefore, $C_k^* =
t_k+s -\sum_{i=1}^{k-1} (\alpha +\beta) t_i^* U_i$. By replacing $C_k^*$ in Eq.~\eqref{eq:precodition-schedulability-negation-0}, we have
\begin{align}
& t_k+s - (\sum_{i=1}^{k-1} \alpha t_i^*U_i + \sum_{i=1}^{k-1} \beta t_i^*U_i) + \sum_{i=1}^{k-1} \alpha t_i^*U_i + \sum_{i=1}^{j-1} \beta t_i^*U_i\nonumber\\
=\; & t_k+s- \sum_{i=j}^{k-1} \beta t_i^* U_i
\geq t_j^*, \;\;\;\forall j=1,\ldots, k-1. \label{eq:precodition-schedulability-negation}
\end{align}
For notational brevity, let $t_k^*$ be $t_k+s$.
Therefore, the linear programming in Eq.~\eqref{eq:lp-init} can be rewritten as follows:
\begin{subequations}
\label{eq:lp-framework-constrained}
\begin{align}
\mbox{min } &
t_k^* -\sum_{i=1}^{k-1} (\alpha +\beta) t_i^* U_i\\
\mbox{s.t.} \;\;&
t_k^*-\beta \sum_{i=j}^{k-1} t_i^* U_i \geq t_j^*,
& \forall 1 \leq j \leq k - 1 \label{eq:lp-framework-constrained-constraints}\\
&t_j^* \geq 0
& \forall 1 \leq j \leq k - 1 \label{eq:lp-framework-constrained-boundaryconstraints}\\
&t_k^* \geq t_k \label{eq:lp-framework-constrained-slack-constraints}
\end{align}
\end{subequations}
The remaining proof is to solve the above linear programming to obtain the minimum $C_k^*$ if $\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)} - \frac{\alpha}{\beta} > 0$.
Our proof strategy is to solve the linear programming analytically as a function of $t_k^*$. This can be imagined as if $t_k^*$ is given. At the end, we will prove the optimality by considering all possible $t_k^* \geq t_k$.
This involves three steps:
\begin{itemize}
\item Step 1: we analyze certain properties of optimal solutions based on the extreme point theorem for linear programming \cite{luenberger2008linear} under the assumption that $t_k^*$ is given as a constant, i.e., $s$ is known.
\item Step 2: we present a specific solution in an \emph{extreme point}, as a function of $t_k^*$.
\item Step 3: we prove that the above extreme point solution gives the minimum $C_k^*$ if $\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)} - \frac{\alpha}{\beta} > 0$.
\end{itemize}
{\bf [Step 1:]} In this step, we assume that
$s$ is given, i.e., $t_k^*$ is specified as a constant. The original linear
programming in Eq.~(\ref{eq:lp-framework-constrained}) has $k$
variables and $2(k-1)+1$ constraints. After specifying the value
$t_k^*$ as a given constant, the new linear programming without the
constraint in
Eq.~\eqref{eq:lp-framework-constrained-slack-constraints} has only
$k-1$ variables and $2(k-1)$ constraints. Thus, according to the
extreme point theorem for linear programming
\cite{luenberger2008linear}, the linear constraints form a
polyhedron of feasible solutions. The extreme point theorem states
that either there is no feasible solution or one of the extreme
points in the polyhedron is an optimal solution when the objective
of the linear programming is finite. To satisfy
Eqs.~\eqref{eq:lp-framework-constrained-constraints}
and~\eqref{eq:lp-framework-constrained-boundaryconstraints}, we know
that $t_j^* \leq t_k^*$ for $j=1,2,\ldots,k-1$, due to $t_i^* \geq
0$, $0 < \beta$, $0 < \alpha$ and $U_i > 0$
for $i=1,2, \ldots,k-1$. As a result, the objective of the above
linear programming is $t_k^* -\sum_{i=1}^{k-1} (\alpha +\beta) t_i^* U_i \geq t_k^* -\sum_{i=1}^{k-1} (\alpha +\beta) t_k^* U_i$, which is finite (as a function of $t_k^*$, $\alpha$, and $\beta$) under the assumption $0 < \beta$, $0 < \alpha$ and $0 < \sum_{i=1}^{k-1}U_i \leq 1$.
According to the extreme point theorem, one of the extreme points is
the optimal solution of Eq.~\eqref{eq:lp-framework-constrained} for a given $t_k^*$.
There are $k-1$ variables with $2(k-1)$ constraints in
Eq.~\eqref{eq:lp-framework-constrained} for a given $t_k^*$. An extreme point must have
at least $k-1$ \emph{active} constraints in
Eqs.~\eqref{eq:lp-framework-constrained-constraints} and \eqref{eq:lp-framework-constrained-boundaryconstraints}, in which
their $\geq$ are set to equality $=$.
We now prove that an extreme point solution is feasible for Eq.~\eqref{eq:lp-framework-constrained} by setting $t_j^*$ either to $0$ or to $t_k^*-\beta \sum_{i=j}^{k-1} t_i^* U_i = t_j^* > 0$ for $j=1,2,\ldots,k-1$.
Suppose for contradiction that there exists a task $\tau_h$ with $0 = t_h^* = t_k^*-\beta \sum_{i=h}^{k-1} t_i^* U_i$.
Let $\omega$ be the
index of the next task after $\tau_h$ with $t_{\omega}^* > 0$ in this extreme point solution, i.e., $t_{h}^* = t_{h+1}^* = \cdots= t_{\omega-1}^*=0$. If $t_i^*=0$ for $h \leq i \leq k-1$, then $\omega$ is set to $k$ and $t_{\omega}^*$ is $t_k^*$. Therefore, the conditions $t_k^*-\beta \sum_{i=h}^{k-1} t_i^* U_i = t_h^*=0$ and $t_{h+1}^* = \cdots= t_{\omega-1}^*=0$ imply the contradiction that $0=t_h^* = t_k^*-\beta \sum_{i=h}^{k-1} t_i^* U_i = t_k^*-\beta \sum_{i=\omega}^{k-1} t_i^* U_i \geq t_{\omega}^* > 0$ when $\omega \leq k-1$ or $0 = t_h^* =t_k^*-\sum_{i=h}^{k-1} t_i^* U_i =t_k^*>0$ when $\omega=k$.
By the above analysis and the pigeonhole principle, a feasible extreme point solution of Eq.~\eqref{eq:lp-framework-constrained} can be represented by two sets ${\bf T}_1$ and ${\bf T}_2$ of the $k-1$ higher-priority tasks, in which $t_j^* = 0$ if $\tau_j$ is in ${\bf T}_1$ and $t_k^*-\beta \sum_{i=j}^{k-1} t_i^* U_i = t_j^* > 0$ if task $\tau_j$ is in ${\bf T}_2$. With the above discussions, we have ${\bf T}_1\cup{\bf T}_2=\setof{\tau_1, \tau_2, \ldots, \tau_{k-1}}$ and ${\bf T}_1\cap {\bf T}_2 = \emptyset$.
{\bf [Step 2:]}
For a given $t_k^*$, one special extreme point solution is to put $t_k^* - \beta \sum_{i=j}^{k-1} t_i^* U_i = t_j^*$ for every $j=1,2,\ldots,k-1$, i.e., ${\bf T}_1$ as an empty set. Therefore,
\begin{equation}
\label{eq:periodrelation}
\forall 1 \leq i \leq k - 1, \quad t_{i+1}^* - t_i^* = \beta t_i^* U_i,
\end{equation}
which implies that
\begin{equation}
\label{eq:2ndperiodrelation}
\dfrac{t_{i+1}^*}{t_i^*} = \beta U_i + 1.
\end{equation}
Moreover,
\begin{equation}
\label{eq:3rdperiodrelation}
\dfrac{t_i^*}{t_k^*}= \prod_{j=i}^{k-1}\frac{t_j^*}{t_{j+1}^*} = \frac{1}{\prod_{j=i}^{k-1} (\beta U_j + 1)}.
\end{equation}
The above extreme point solution is always feasible in the linear programming of Eq.~\eqref{eq:lp-framework-constrained} since
\begin{equation}
0 < \dfrac{t_i^*}{t_k^*}, \;\;\;\forall i=1,2,\ldots,k-1.
\end{equation}
By Eq.~\eqref{eq:periodrelation}, we have
\begin{equation}
\label{eq:alpha+beta}
(\alpha
+\beta)U_i t_i^* = (t_{i+1}^*-t_i^*) (\frac{\alpha}{\beta}+1).
\end{equation}
Therefore, in this extreme point solution, the objective function of Eq.~\eqref{eq:lp-framework-constrained} is
\begin{align}
&t_k^* - \sum_{i=1}^{k-1} (\alpha
+\beta)U_it_i^*= t_k^* - (t_k^*-t_1^*) (\frac{\alpha}{\beta}+1)\nonumber\\
= & t_k^*\left(\frac{t_1^*}{t_k^*}(\frac{\alpha}{\beta}+1)-\frac{\alpha}{\beta}\right) \overset{1}{=}
t_k^*\left(\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)}-\frac{\alpha}{\beta}\right),
\end{align}
where the last equality $\overset{1}{=}$ is due to Eq.~\eqref{eq:3rdperiodrelation} when $i$ is $1$.
{\bf [Step 3:]} From Step 1, a feasible extreme point solution is with ${\bf T}_1\cup{\bf T}_2=\setof{\tau_1, \tau_2, \ldots, \tau_{k-1}}$ and ${\bf T}_1\cap {\bf T}_2 = \emptyset$.
As a result, we can simply drop all the tasks in ${\bf T}_1$ and use the remaining tasks in ${\bf T}_2$ by adopting the same procedures from Eq.~\eqref{eq:periodrelation} to Eq.~\eqref{eq:alpha+beta} in Step 2. The resulting objective function of this extreme point solution for the linear programming in
Eq.~\eqref{eq:lp-framework-constrained} is $t_k^*(\frac{\frac{\alpha}{\beta}+1}{\prod_{\tau_j \in {\bf T}_2} (\beta U_j + 1)} - \frac{\alpha}{\beta}) \geq t_k^*(\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j + 1)}-\frac{\alpha}{\beta})$ due to the fact that $\prod_{\tau_j \in {\bf T}_2} (\beta U_j + 1) \leq \prod_{j=1}^{k-1} (\beta U_j + 1)$.
Therefore, for a given $s \geq 0$, i.e., $t_k^* \geq t_k$, all the other extreme points of
Eq.~\eqref{eq:lp-framework-constrained} were dominated by the one
specified in Step 2.
By the above analysis, if $\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta U_j +
1)}-\frac{\alpha}{\beta} > 0$, we know that
$(t_k+s)(\frac{\frac{\alpha}{\beta}+1}{\prod_{j=1}^{k-1} (\beta
U_j + 1)}-\frac{\alpha}{\beta})$ is an increasing function of
$s$, in which the minimum happens when $s$ is $0$. As a result, we
reach the conclusion of the lemma.
\end{proof}
We now provide two extended lemmas based on
Lemma~\ref{lemma:framework-constrained}, used for given $\alpha$ and
$\beta$ when $\alpha_i \leq \alpha$ and $\beta_i \leq \beta$ for
$i=1,2,\ldots,k-1$. Their proofs are in Appendix B.
\begin{lemma}
\label{lemma:framework-totalU-constrained}
For a given $k$-point effective schedulability test of a scheduling
algorithm, defined in
Definition~\ref{def:kpoints},
in which $0 < t_k$ and $0 < \alpha_i \leq \alpha$ and $0 < \beta_i \leq \beta$ for any
$i=1,2,\ldots,k-1$, task $\tau_k$ is schedulable by the scheduling
algorithm if
\begin{equation}
\label{eq:schedulability-totalU-constrained}
\frac{C_k}{t_k} + \sum_{i=1}^{k-1}U_i \leq \begin{cases}
1, & (\alpha+\beta)^{\frac{1}{k}} < 1\\
(k-1)\frac{\left((1+\frac{\beta}{\alpha})^{\frac{1}{k-1}}-1\right)}{\beta}, & (\alpha+\beta)^{\frac{1}{k}} < \alpha\\
\frac{(k-1)((\alpha+\beta)^{\frac{1}{k}}-1)+((\alpha+\beta)^{\frac{1}{k}}-\alpha)}{\beta} & \mbox{ otherwise}.
\end{cases}
\end{equation}
\end{lemma}
\begin{lemma}
\label{lemma:framework-totalU-exclusive}
For a given $k$-point effective schedulability test of a scheduling
algorithm, defined in
Definition~\ref{def:kpoints},
in which $0 < t_k$ and $0 < \alpha_i \leq \alpha$ and $0 < \beta_i \leq \beta$ for any
$i=1,2,\ldots,k-1$,
task $\tau_k$ is schedulable by the scheduling
algorithm if
\begin{equation}
\label{eq:schedulability-totalU-exclusive}
\beta \sum_{i=1}^{k-1}U_i \leq \ln(\frac{\frac{\alpha}{\beta}+1}{\frac{C_k}{t_k}+\frac{\alpha}{\beta}}).
\end{equation}
\end{lemma}
We also construct the following Lemma~\ref{lemma:framework-general}, c.f. Appendix B for the proof, as the most powerful method for a concrete task system.
Throughout the paper, we will not build theorems and corollaries based on
Lemma~\ref{lemma:framework-general}\ifbool{techreport}{ but we will
evaluate how it performs in the experimental section.}{, the
performance evaluation can be found in the report \cite{DBLP:journals/corr/abs-1501.07084}.}
\begin{lemma}
\label{lemma:framework-general}
For a given $k$-point effective schedulability test of a fixed-priority scheduling
algorithm, defined in
Definition~\ref{def:kpoints}, task $\tau_k$ is schedulable by the scheduling algorithm,
in which $0 < t_k$ and $0 < \alpha_i$ and $0 < \beta_i$ for any
$i=1,2,\ldots,k-1$,
if
the following condition holds
\begin{equation}
\label{eq:schedulability-general}
0 < \frac{C_k}{t_k} \leq 1 - \sum_{i=1}^{k-1} \frac{U_i(\alpha_i
+\beta_i)}{\prod_{j=i}^{k-1} (\beta_jU_j + 1)}.
\end{equation}
\end{lemma}
\noindent{\bf{\large Remarks and how to use the framework}:}
After presenting the \framework{} framework, here, we explain how to use the \framework{} framework and summarize how we
plan to demonstrate its applicability in several task models in the following sections.
The \framework{} framework relies on the users to index the tasks properly and define $\alpha_i$ and $\beta_i$ as constants for $i=1,2,\ldots,k-1$ based on Eq.~\eqref{eq:precodition-schedulability}. The set $\setof{t_1, t_2, \ldots, t_{k-1}}$ in Definition~\ref{def:kpoints} is used only
for the users to define those constants, where $t_k$ is usually defined to be the interval length of the original schedulability test, e.g., $D_k$ in Eq.~\eqref{eq:exact-test-constrained-deadline}.
Therefore, the \framework{} framework can only be applicable when $\alpha_i$ and $\beta_i$ are well-defined. These constants
depend on the task models and the task parameters.
The choice of good parameters $\alpha_i$ and $\beta_i$ affects the
quality of the resulting schedulability bounds in Lemmas~\ref{lemma:framework-constrained} to \ref{lemma:framework-totalU-exclusive}.
However, deriving the \emph{good} settings of $\alpha_i$ and $\beta_i$ is actually not the focus of this paper. The framework does not care how the parameters $\alpha_i$ and $\beta_i$ are obtained. The framework simply derives the bounds according to the given parameters $\alpha_i$ and $\beta_i$, regardless of the settings of $\alpha_i$ and $\beta_i$. The correctness of the settings of $\alpha_i$ and $\beta_i$ is not verified by the framework.
The ignorance of the settings of $\alpha_i$ and $\beta_i$ actually leads to the elegance and the generality of the framework, which works as long as Eq.~\eqref{eq:precodition-schedulability} can be successfully constructed for the sufficient schedulability test of task $\tau_k$ in a fixed-priority scheduling policy.
With the availability of the \framework{} framework, the hyperbolic bounds or utilization bounds can be automatically derived by adopting Lemmas \ref{lemma:framework-constrained} to \ref{lemma:framework-totalU-exclusive} as long as the safe upper bounds $\alpha$ and $\beta$ to cover all the possible settings of $\alpha_i$ and $\beta_i$ for the schedulability test in Eq.~\eqref{eq:precodition-schedulability} can be derived, regardless of the task model or the platforms.
The other approaches in \cite{journals/tc/LeeSP04,DBLP:dblp_journals/tc/BurchardLOS95,HanTyan-RTSS97} also have similar observations by testing only several time points in the TDA schedulability analysis based on Eq.~\eqref{eq:exact-test-constrained-deadline} in their problem formulations. Specifically, the problem formulations in \cite{DBLP:dblp_journals/tc/BurchardLOS95,HanTyan-RTSS97} are based on non-linear programming by selecting several good points under certain constraints. Moreover, the linear-programming problem formulation in \cite{journals/tc/LeeSP04} considers $U_i$ as variables and $t_i$ as constants and solves the corresponding linear programming analytically.
However, as these approaches in
\cite{journals/tc/LeeSP04,DBLP:dblp_journals/tc/BurchardLOS95,HanTyan-RTSS97} seek for the total utilization bounds, they have limited applications and are less flexible. For example, they are typically not applicable directly when considering sporadic real-time tasks with arbitrary deadlines or multiprocessor systems. Here, we are more flexible in the \framework{} framework.
For task $\tau_i$, after $\alpha_i$ and $\beta_i$ or their safe upper bounds $\alpha$ and $\beta$ are derived, we completely drop out the dependency to the periods and models inside \framework{}.
\ifbool{techreport}{
\begin{figure*}[t]
} {
\begin{figure}[t]
}
\begin{center}
\ifbool{techreport}{
\begin{tikzpicture}[scale=1.2,transform shape]
} {
\begin{tikzpicture}[scale=0.75,transform shape]
}
\path \practica {1}{\underline{\bf Demonstrated Applications:}
\begin{tabular}{ll}
Sec. 5.1:& Constrained-deadline sporadic tasks\\
Sec. 5.2:& Arbitrary-deadline sporadic tasks\\
App. C\ifbool{techreport}{}{ \cite{DBLP:journals/corr/abs-1501.07084}}:& Multiframe tasks\\
Sec. 6.1:& Multiprocessor DAG\\
Sec. 6.2:& Multiprocessor self-suspension\\
\end{tabular}
};
\path (p1.west)+(1.8,-3.2) node(p2)[materia, rounded rectangle,text width=6em]{{\bf $U_i, \forall i < k$\\$\alpha_i, \forall i<k$\\$\beta_i, \forall i <k$\\$C_k, t_k$}};
\path (p2.north)+(1.5,0.45) node[text width=16em]{Derive parameters\\by \underline{Definition 1}};
\path (p2.east)+(2,0) node(p3)[practica,fill=blue!20,text width=5em,text centered]{\large{\bf \framework{}\\ framework}};
\path (p3.east)+(2.5,+3) node(p4)[practica,fill=green!30,minimum width=6em,text width=5em,text centered]{Hyperbolic bound};
\path (p4.south)+(0.8,-1) node(p5)[practica,fill=green!30,minimum width=6em,text width=5em,text centered]{Other utilization bounds};
\path (p5.south)+(0,-1) node(p6)[practica,fill=green!30,minimum width=6em,text width=5em,text centered]{Extreme points test};
\path [line, ->] (p1.west) -- +(-0.3,0) node[black, rotate=90, yshift=0.3cm, xshift=-1.6cm]{\footnotesize{define $t_i, \forall i < k$ and order $k-1$ tasks}} -- + (-0.3, -3.2) -- (p2.west);
\path [line, ->] (p2.east) -- (p3.west);
\path [line, ->] (p3.east)+(0,0.3) -- node[rotate=60,yshift=0.3cm,black]{Lemma 1} (p4.west);
\path [line, ->] (p3.east)+(0,0) -- node[rotate=35,yshift=0.3cm,black]{Lemmas 2\&3} (p5.west);
\path [line, ->] (p3.east)+(0,-0.3) -- node[rotate=5,yshift=0.3cm,black]{Lemma 4} (p6.west);
\end{tikzpicture}
\end{center}
\vspace{-2mm}
\caption{The \framework{} framework. }
\label{fig:framework}
\ifbool{techreport}{
\end{figure*}
} {
\end{figure}
}
We will demonstrate the applicability and generality of \framework{} by using the most-adopted sporadic real-time task model in Sec.~\ref{sec:application-FP}, multi-frame tasks in Appendix C\citetechreport{}
and multiprocessor scheduling in Sec.~\ref{sec:multiprocessor}, as illustrated in Figure~\ref{fig:framework}.
In all these cases, we can find reasonable settings of $\alpha$ and $\beta$ to provide better results or new results for schedulability tests, with respect to the utilization bounds, speed-up factors, or capacity augmentation factors, compared to the literature. More specifically, (after certain reorganizations), we will greedily set $t_i$ as $\floor{\frac{t_k}{T_i}}T_i$ in all of the studied cases.\footnote{Setting $t_i$ as $\floor{\frac{t_k}{T_i}}T_i$ is actually the same in \cite{journals/tc/LeeSP04} for the sporadic real-time task model with implicit deadlines and the multi-frame task model when $T_i$ is given and $U_i$ is considered as a variable.}
Table \ref{tab:alpha-beta} summarizes the $\alpha_i$ and $\beta_i$ parameters derived in this paper, as well as an earlier result by Liu and Chen in \cite{RTSS14a} for self-suspending task models and deferrable servers.
\begin{table*}[t]
\renewcommand{\arraystretch}{1.5}
\centering
\scalebox{0.8}{
\begin{tabular}{|p{7cm}|c|c|p{4cm}|}
\hline
\hline
Model & $\alpha_i$ & $\beta_i$ & c.f.\\
\hline
\hline
Uniprocessor Sporadic Tasks& $\alpha_i=1$ & $0 < \beta_i \leq 1$ & Theorem~\ref{theorem:sporadic-general} and Corollary~\ref{col:arbitrary-general}\\
\hline
Multiprocessor Global RM for Different Models & $0 < \alpha_i \leq
\frac{2}{M}$ & $0 < \beta_i\leq \frac{1}{M}$ &
Theorems~\ref{thm:multiprocessor-DAG} -
\ref{thm:multiprocessor-suspension}\\
\hline
Uniprocessor Multi-frame Tasks& $0 < \alpha_i \leq 1$ & $0 < \beta_i\leq \frac{\phi_i(2)-\phi_i(1)}{\phi_i(1)}$ & \ifbool{techreport}{Theorem~\ref{thm:multiframe}}{Theorem 7 in \cite{DBLP:journals/corr/abs-1501.07084}}\\
\hline
Uniprocessor Self-Suspending Tasks& $0 < \alpha_i\leq 2$ & $0 < \beta_i\leq 1$ & Theorems 5 and 6 in \cite{RTSS14a}, implicitly\\
\hline
\end{tabular}
}
\caption{\small The applicability of \framework{} for different task models: their $\alpha_i$ and $\beta_i$ parameters, where $M$ is the number of processors in global RM scheduling and $\phi_i(\ell)$ is the maximum of the sum of the execution time of
any $\ell$ consecutive frames of task $\tau_i$ in the multi-frame model.}
\label{tab:alpha-beta}
\end{table*}
\section{Applications for Fixed-Priority Scheduling}
\label{sec:application-FP}
This section provides demonstrations on how to use the \framework{}
framework to derive efficient schedulability analysis for sporadic
task systems in uniprocessor systems. We will consider
constrained-deadline systems first in
Sec.~\ref{sec:application-constrained}
and explain how to extend to
arbitrary-deadline systems in Sec.~\ref{sec:application-arbitrary}. For the rest of this section, we will implicitly assume $C_k > 0$.
\subsection{Constrained-Deadline Systems}
\label{sec:application-constrained}
For a specified fixed-priority scheduling algorithm, let $hp(\tau_k)$ be the set
of tasks with higher priority than $\tau_k$. We now classify the task
set $hp(\tau_k)$ into two subsets:
\begin{itemize}
\item $hp_1(\tau_k)$ consists of the higher-priority tasks with periods
smaller than $D_k$.
\item $hp_2(\tau_k)$ consists of the higher-priority tasks with periods
larger than or equal to $D_k$.
\end{itemize}
For any $0 < t \leq D_k$,
we know that a safe upper bound on the interference due to higher-priority tasks is given by
\[
\sum_{\tau_i \in hp(\tau_k)} \ceiling{\frac{t}{T_i}}C_i = \sum_{\tau_i \in hp_2(\tau_k)} C_i + \sum_{\tau_i \in hp_1(\tau_k)} \ceiling{\frac{t}{T_i}}C_i .
\]
As a result, the schedulability test in
Eq.~\eqref{eq:exact-test-constrained-deadline} is equivalent to the
verification of the existence of $0 < t \leq D_k$ such that
\begin{equation}
\label{eq:exact-test-constrained-deadline-2}
C_k + \sum_{\tau_i \in hp_2(\tau_k)} C_i + \sum_{\tau_i \in hp_1(\tau_k)} \ceiling{\frac{t}{T_i}}C_i \leq t.
\end{equation}
We can then create a virtual sporadic task $\tau_k'$ with execution
time $C_k'=C_k + \sum_{\tau_i \in hp_2(\tau_k)} C_i$, relative
deadline $D_k'=D_k$, and period $T_k'=D_k$. It is
clear that the schedulability test to verify the schedulability of
task $\tau_k'$ under the interference of the higher-priority tasks
$hp_1(\tau_k)$ is the same as that of task $\tau_k$ under the
interference of the higher-priority tasks $hp(\tau_k)$.
Therefore, with the above analysis, we can use the \framework{}
framework in Sec.~\ref{sec:framework} as in the following theorem.
\begin{theorem}
\label{theorem:sporadic-general}
Task $\tau_k$ in a sporadic task system with constrained deadlines is
schedulable by the fixed-priority scheduling algorithm if
\begin{equation}
\label{eq:schedulability-sporadic-any-a}
(\frac{C_k'}{D_k}+1) \prod_{\tau_j \in hp_1(\tau_k)} (U_j + 1)\leq 2
\end{equation}
or
\begin{equation}
\label{eq:schedulability-sporadic-any-b}
\frac{C_k'}{D_k}+ \sum_{\tau_j \in hp_1(\tau_k)}U_j\leq k(2^{\frac{1}{k}}-1).
\end{equation}
\end{theorem}
\begin{proof}
For notational brevity, suppose that there are $k-1$ tasks in
$hp_1(\tau_k)$.\footnote{When $hp_2(\tau_k)$ is not empty, there are $k-1-|hp_2(\tau_k)|$ tasks in $hp_1(\tau_k)$.
To be notationally precise, we can denote the number of tasks in $hp_1(\tau_k)$ by a new symbol $k^*$. Since $k^* \leq k$, we know $k^*(2^{\frac{1}{k^*}}-1) \geq k(2^{\frac{1}{k}}-1)$ as a safe bound in Eq.~\eqref{eq:schedulability-sporadic-any-b}. However, this may make the notations too complicated. We have decided to keep it as $k-1$ for the sake of notational brevity. }
Now, we index the higher-priority tasks in
$hp_1(\tau_k)$ to form the corresponding $\tau_1, \tau_2, \ldots,
\tau_{k-1}$. The $k-1$ higher-priority tasks in $hp_1(\tau_k)$ are
ordered to ensure that the arrival times, i.e., $\floor{\frac{D_k}{T_i}}T_i$, of the last jobs no later than $D_k$
are in a non-decreasing order. That is, with the above indexing of
the higher-priority tasks in $hp_1(\tau_k)$, we have
$\floor{\frac{D_k}{T_i}}T_i \leq
\floor{\frac{D_k}{T_{i+1}}}T_{i+1}$ for $i=1,2,\ldots,k-2$.
Now, we set $t_i$ as $\floor{\frac{D_k}{T_i}}T_i$ for
$i=1,2,\ldots,k-1$, and $t_k$ as $D_k$. Due to the fact that $T_i \leq D_k$
for
$i=1,2,\ldots,k-1$, we know that $t_i > 0$.
Therefore,
for a given $t_j$ with $j=1,2,\ldots,k$, the demand requested up to time $t_j$ is at most
\begin{align}
& C_k + \sum_{\tau_i \in hp_2(\tau_k)} C_i + \sum_{\tau_i \in hp_1(\tau_k)} \ceiling{\frac{t_j}{T_i}}C_i\nonumber\\
= & C_k' + \sum_{i=1}^{k-1} \ceiling{\frac{t_j}{T_i}}C_i
\leq C_k' + \sum_{i=1}^{k-1} \frac{t_i}{T_i}C_i + \sum_{i=1}^{j-1} C_i,\nonumber
\end{align}
where the inequality comes from the indexing policy defined above,
i.e., $\ceiling{\frac{t_j}{T_i}} \leq \frac{t_i}{T_i}+1$ if $j
> i$ and $\ceiling{\frac{t_j}{T_i}} \leq
\frac{t_i}{T_i}$ if $j \leq i$.
Since $T_i < D_k$ for any task $\tau_i$ in $hp_1(\tau_k)$, we know
that $t_i \geq T_i$.
Instead of testing all the $t$ values in
Eq.~(\ref{eq:exact-test-constrained-deadline-2}), we only apply the
test for these $k$ different $t_i$ values, which is equivalent to the
test of the existence of $t_j$ such that
\begin{equation}
\label{eq:k-point-sporadic}
C_k'
+ \sum_{i=1}^{k-1} t_i U_i + \sum_{i=1}^{j-1} \frac{T_i}{t_i} t_i U_i \leq t_j.
\end{equation}
Therefore, we reach the schedulability in the form of Eq.~\eqref{eq:precodition-schedulability} under the setting of $\alpha_i$ to $1$ and $\beta_i$ to $\frac{T_i}{t_i} \leq 1$ (due to $t_i \geq T_i$), for $i=1,2,\ldots,k-1$. By taking $\beta_i \leq 1$ and $\alpha_i = 1$ for
$i=1,2,\ldots,k-1$ in Lemmas~\ref{lemma:framework-constrained}
and~\ref{lemma:framework-totalU-constrained}, we reach the
conclusion.
\end{proof}
When RM priority ordering is applied for an implicit-deadline
task system, $C_k'$ is equal to $C_k$ and $\frac{C_k'}{D_k}$ is equal
to $U_k$. For such a case, the condition in
Eq.~\eqref{eq:schedulability-sporadic-any-a} is the same as the
hyperbolic bound provided in \cite{bini2003rate}, and the condition in
Eq.~\eqref{eq:schedulability-sporadic-any-b} is the same as least
utilization upper bound in \cite{liu1973scheduling}.
The schedulability test in
Theorem~\ref{theorem:sporadic-general} may seem pessimistic at first glance. We evaluate the quality of the
schedulability test in Theorem~\ref{theorem:sporadic-general} by
quantifying the speed-up factor with respect to the optimal schedule
(i.e., EDF scheduling in such a case). We show that the speed-up
factor of the schedulability test in
Eq.~\eqref{eq:schedulability-sporadic-any-a} is $1.76322$, which is
the same as the lower bound and upper bound of DM as shown in
\cite{DBLP:journals/rts/DavisRBB09}. The speed-up factor of DM,
regardless of the schedulability tests, obtained by Davis et
al. \cite{DBLP:journals/rts/DavisRBB09} is the same as our result. Our
result is thus stronger, as we show that the factor already holds when
using the schedulability test in Eq.~\eqref{eq:schedulability-sporadic-any-a} in the following theorem.
\begin{theorem}
\label{thm:speedup-DM}
The speed-up factor of the schedulability test in
Eq.~\eqref{eq:schedulability-sporadic-any-a} under DM scheduling for
constrained-deadline tasks is $1.76322$ with respect to EDF.
\end{theorem}
The proof of Theorem~\ref{thm:speedup-DM} in the Appendix B, which is much simpler than the proof in \cite{DBLP:journals/rts/DavisRBB09}, can also be considered as an alternative proof of the speed-up factor of DM.
\begin{corollary}
\label{col:implicit-general-ratio}
Suppose that $f \cdot T_i \leq D_k$ for any higher priority task
$\tau_i$ in $hp_1(\tau_k)$, where $f$ is a positive integer. Task $\tau_k$
in a constrained-deadline sporadic task system is schedulable by the fixed-priority
scheduling algorithm if
\begin{equation}
\label{eq:schedulability-sporadic-implicit-ratio-a}
(\frac{C_k'}{f \cdot D_k}+1) \prod_{\tau_j \in hp_1(\tau_k)} (\frac{U_j}{f} + 1)\leq \frac{f+1}{f}
\end{equation}
or
\begin{equation}
\label{eq:schedulability-sporadic-implicit-ratio-b}
\frac{C_k'}{D_k}+ \sum_{\tau_j \in hp_1(\tau_k)}U_j\leq f k\left(\left({\frac{f+1}{f}}\right)^{\frac{1}{k}}-1\right).
\end{equation}
\end{corollary}
\begin{proof}
This is based on the same proof as Theorem~\ref{theorem:sporadic-general}, by taking the fact that $\frac{t_i}{T_i} = \frac{\floor{\frac{D_k}{T_i}}T_i}{T_i} \geq f$.
In the $k$-point effective schedulability test, we can set
$\alpha_i$ to $1$, $\beta_i = \frac{T_i}{t_i} \leq \frac{1}{f}$.
Therefore, we have $\alpha_i \leq \alpha=1$,
$\beta_i\leq\beta=\frac{1}{f}$, and $\frac{\alpha}{\beta}$ is $f$.
By adopting
Lemma~\ref{lemma:framework-constrained}, we know that task $\tau_k$
is schedulable under the scheduling policy if
\ifbool{techreport}{
\[}{$}
(\frac{C_k'}{D_k} + f) \prod_{{\tau_j \in hp_1(\tau_k)}} (\frac{U_j}{f} + 1)\leq f+1,
\ifbool{techreport}{
\]}{$}
which is the same as the condition in
Eq.~\eqref{eq:schedulability-sporadic-implicit-ratio-a} by dividing
both sides by $f$. By using a similar argument and applying
Lemma~\ref{lemma:framework-totalU-constrained}, we can reach the
condition in Eq.~\eqref{eq:schedulability-sporadic-implicit-ratio-b}.
\end{proof}
\ifbool{techreport}{
Note that the right-hand side of
Eq.~\eqref{eq:schedulability-sporadic-implicit-ratio-b} converges to $f
\ln(\frac{f+1}{f})$ when $k$ goes to $\infty$.}{}
\subsection{Arbitrary-Deadline Systems}
\label{sec:application-arbitrary}
We now further explore how to use the proposed framework to perform
the schedulability analysis for arbitrary-deadline task sets.
The
exact schedulability analysis for arbitrary-deadline task sets under
fixed-priority scheduling has been developed in
\cite{DBLP:conf/rtss/Lehoczky90}. The schedulability analysis is to
use a \emph{busy-window} concept to evaluate the worst-case response
time. That is, we release all the higher-priority tasks together with
task $\tau_k$ at time $0$ and all the subsequent jobs are released as
early as possible by respecting to the minimum inter-arrival time. The
busy window finishes when a job of task $\tau_k$ finishes before the
next release of a job of task $\tau_k$. It has been shown in
\cite{DBLP:conf/rtss/Lehoczky90} that the worst-case response time of
task $\tau_k$ can be found in one of the jobs of task $\tau_k$ in the
busy window.
Therefore, a simpler sufficient schedulability test for a task
$\tau_k$ is to test whether the length of the busy window is within
$D_k$. If so, all invocations of task $\tau_k$ released in the busy
window can finish before their relative deadline. Such an observation
has also been adopted in \cite{conf:/rtns09/Davis}. Therefore, a
sufficient test is to verify whether\footnote{This analysis is pretty pessimistic.
But, as our objective in this paper is to show the applicability and generality of \framework{},
how to use \framework{} to get tighter analysis for arbitrary-deadline task systems is not our focus in this paper. Some evaluations can be found in \cite{DBLP:journals/corr/framework-compare}. }
\begin{equation}
\label{eq:sufficient-test-arbitrary-deadline}
\exists t \mbox{ with } 0 < t \leq D_k {\;\; and \;\;} \ceiling{\frac{t}{T_k}}C_k +
\sum_{\tau_i \in hp(\tau_k)} \ceiling{\frac{t}{T_i}}C_i \leq t.
\end{equation}
If the condition in Eq.~\eqref{eq:sufficient-test-arbitrary-deadline} holds,
it implies that the busy window (when considering task $\tau_k$) is no
more than $D_k$, and, hence, task $\tau_k$ has worst-case response
time no more than $D_k$.
If $D_k \leq T_k$, the analysis in
Sec.~\ref{sec:application-constrained} can be directly applied. If
$D_k > T_k$, we need to consider the length of the busy-window for
task $\tau_k$ as shown above. For the rest of this section, we will
focus on the case $D_k > T_k$.
We can rewrite
Eq.~\eqref{eq:sufficient-test-arbitrary-deadline} to use a more
pessimistic condition by releasing the workload
$\ceiling{\frac{D_k}{T_k}}C_k$ at time $0$. That is, if
\begin{equation}
\label{eq:sufficient-test-arbitrary-deadline-pessimistic}
\exists t \mbox{ with } 0 < t \leq D_k {\;\; and \;\;} \ceiling{\frac{D_k}{T_k}}C_k +
\sum_{\tau_i \in hp(\tau_k)} \ceiling{\frac{t}{T_i}}C_i \leq t,
\end{equation}
then, the length of the busy window for task $\tau_k$ is no more than
$D_k$. Again, similar to the strategy
we use in Sec.~\ref{sec:application-constrained}, we classify the
tasks in $\tau_i \in hp(\tau_k)$ into two sets $hp_1(\tau_k)$ and
$hp_2(\tau_k)$ with the same definition.
Similarly, we can then create a virtual sporadic task $\tau_k'$ with
execution time $C_k' = \ceiling{\frac{D_k}{T_k}}C_k + \sum_{\tau_i \in
hp_2(\tau_k)} C_i$, relative deadline $D_k'=D_k$, and period $T_k'=D_k$.
For notational brevity, suppose that there are $k-1$ tasks in
$hp_1(\tau_k)$. Now, we index the higher-priority tasks in
$hp_1(\tau_k)$ to form the corresponding $\tau_1, \tau_2, \ldots,
\tau_{k-1}$. In the above definition of the busy window concept,
$\floor{\frac{D_k}{T_i}}T_i$ is the arrival time of the last job
of task $\tau_i$ released no later than $D_k$. The $k-1$ higher-priority
tasks in $hp_1(\tau_k)$ are ordered to ensure that the
arrival times of the last jobs before $D_k$ are in a non-decreasing
order. Moreover, $t_k$ is the specified testing point $D_k$. Instead
of testing all the $t$ values in
Eq.~\eqref{eq:sufficient-test-arbitrary-deadline-pessimistic}, we only
apply the test for these $k$ different $t_i$ values, which is
equivalent to the test of the existence of $t_j$ such that
Eq.~\eqref{eq:k-point-sporadic} holds, where $\alpha_i \leq 1$ and
$\beta_i = \frac{T_i}{t_i} \leq 1$ for $i=1,2,\ldots,k-1$, similar to the proof of Theorem~\ref{theorem:sporadic-general}.
\ifbool{techreport}{
Therefore, we can then use the \framework{} framework, i.e., Lemmas
\ref{lemma:framework-constrained} to
\ref{lemma:framework-totalU-exclusive} to test the schedulability of
task $\tau_k$. }{}
The following corollary comes from a similar argument as in
Sec.~\ref{sec:application-constrained}.
\begin{corollary}
\label{col:arbitrary-general}
Task $\tau_k$ in a sporadic arbitrary-deadline task system is schedulable by the
fixed-priority scheduling algorithm if
Eq.~\eqref{eq:schedulability-sporadic-any-a} or
\eqref{eq:schedulability-sporadic-any-b} holds,
in which there are $k-1$ higher priority tasks in $hp_1(\tau_k)$ and $C_k'$ is defined as $\ceiling{\frac{D_k}{T_k}}C_k + \sum_{\tau_i
\in hp_2(\tau_k)} C_i$.
\end{corollary}
\begin{corollary}
\label{col:utilization-RM-arbitrary}
Suppose that $f \cdot T_i \leq D_k$ for any higher task $\tau_i$ in the
task system and $f\cdot T_k \leq D_k$, where $f$ is a positive integer. Task $\tau_k$
in a sporadic task system is schedulable by using RM, i.e., $T_i \leq T_{i+1}$, if
\begin{equation}
\label{eq:schedulability-arbitrary-rm-ratio-a}
(\frac{U_k}{f}+1) \prod_{j=1}^{k-1} (\frac{U_j}{f} + 1)\leq \frac{f+1}{f}
\end{equation}
or
\begin{equation}
\label{eq:schedulability-arbitrary-rm-ratio-b}
\sum_{j=1}^{k}U_j \leq f k\left(\left({\frac{f+1}{f}}\right)^{\frac{1}{k}}-1\right).
\end{equation}
\end{corollary}
\begin{proof}
Based on the above assumption $f \cdot T_i \leq D_k$ for any higher task $\tau_i$ in the
task system, $f\cdot T_k \leq D_k$, and the rate
monotonic scheduling, we can safely set $\alpha_i$ to $1$, $\beta_i$
to $\frac{1}{f}$, and $C_k'$ to $f \cdot C_k$. Note that $\frac{C_k'}{f\cdot t_k} \leq \frac{U_k}{f}$ in this case. Therefore, by
adopting Lemma~\ref{lemma:framework-constrained} and
Lemma~\ref{lemma:framework-totalU-constrained}, we reach the
conclusion, as in the proof of Corollary~\ref{col:implicit-general-ratio}.
\end{proof}
\section{Application for Multiprocessor Scheduling}
\label{sec:multiprocessor}
It may seem, at first glance, that the \framework{} framework only
works for uniprocessor systems. We demonstrate in this section how to
use the framework in multiprocessor global RM scheduling when considering
implicit-deadline, DAG, and self-suspending task systems. The methodology can also be extended to
handle constrained-deadline systems.
In multiprocessor global scheduling, we consider that the system has
$M$ identical processors, each with the same
computation power. Moreover, there is a global queue and a global
scheduler to dispatch the jobs. We consider only global RM scheduling,
in which the priority of the tasks is defined based on RM. At any
time, the $M$-highest-priority jobs in the ready queue are dispatched
and executed on these $M$ processors.
Global RM in general does not have good utilization bounds.
However, if we constrain the total utilization
$\sum_{\tau_i} \frac{C_i}{M T_i} \leq \frac{1}{b}$
and the maximum utilization $\max_{\tau_i} \frac{C_i}{T_i} \leq
\frac{1}{b}$, it is possible to provide the schedulability guarantee
of global RM by setting $b$ to $3-\frac{1}{M}$ \cite{DBLP:conf/rtss/AnderssonBJ01,DBLP:conf/rtss/Baker03,DBLP:conf/opodis/BertognaCL05}. Such a factor $b$ has
been recently named as a \emph{capacity augmentation factor} \cite{Li:ECRTS14}.
We will use the following
time-demand function $W_i(t)$ for the simple sufficient schedulability analysis:
\begin{equation}
\label{eq:W_i-multiprocessor}
W_i(t) = (\ceiling{\frac{t}{T_i}}-1)C_i + 2C_i,
\end{equation}
which can be imagined as if the \emph{carry-in} job is fully carried into the analysis interval \cite{DBLP:conf/rtss/GuanSYY09}.
That is, we allow the first release of task $\tau_i$ to be inflated by a
factor $2$, whereas the other jobs of task $\tau_i$ have the same
execution time $C_i$.
Again, we consider testing the schedulability of task $\tau_k$ under
global RM, in which there are $k-1$ higher-priority tasks $\tau_1,
\tau_2, \ldots, \tau_{k-1}$. We have the following schedulability
condition for global RM.
\begin{lemma}
\label{lemma:gRM-sufficient}
Task $\tau_k$ is schedulable under
global RM on $M$ identical processors, if
\begin{equation}
\label{eq:gRM-sufficient}
\exists t \mbox{ with } 0 < t \leq T_k {\;\; and \;\;} C_k +\sum_{i=1}^{k-1}
\frac{W_i(t)}{M} \leq t,
\end{equation}
where $W_i(t)$ is defined in Eq.~\eqref{eq:W_i-multiprocessor}.
\end{lemma}
\begin{proof}
This has been shown in Sec. 3.2 (The Basic Multiprocessor Case)
in \cite{DBLP:conf/rtss/GuanSYY09}.
\end{proof}
\begin{theorem}
\label{thm:multiprocessor-GRM}
Task $\tau_k$ in a sporadic implicit-deadline task system is
schedulable by global RM on $M$ processors if
\begin{equation}
\label{eq:schedulability-GRM}
(\frac{C_k}{T_k}+2)\prod_{j=1}^{k-1} (\frac{U_j}{M} + 1)\leq 3,
\end{equation}
or
\begin{equation}
\label{eq:schedulability-augmentation-GRM}
\sum_{j=1}^{k-1} \frac{U_j}{M}\leq \ln{\frac{3}{\frac{C_k}{T_k}+2}}.
\end{equation}
\end{theorem}
\begin{proof}
Let $t_i$ be $\floor{\frac{T_k}{T_i}}T_i$ for $i=1,2,\ldots,k$, and
reindex the tasks such that $t_1 \leq t_2 \leq \ldots \leq t_k$. By
testing only these $k$ points in the schedulability test in
\eqref{eq:gRM-sufficient} results in a $k$-point effective
schedulability test with $\alpha_i \leq \frac{2}{M}$ and $\beta_i
\leq \frac{1}{M}$. Therefore, we can adopt the \framework{}
framework. By Lemma~\ref{lemma:framework-constrained} and
Lemma~\ref{lemma:framework-totalU-exclusive}, we have concluded the
proof.
\end{proof}
Note that Theorem~\ref{thm:multiprocessor-GRM} is not superior to the known analysis for
sporadic task systems \cite{DBLP:conf/rtss/AnderssonBJ01,DBLP:conf/rtss/Baker03,DBLP:conf/opodis/BertognaCL05}, as the schedulability condition in Lemma~\ref{lemma:gRM-sufficient} is too pessimistic. This is only used as the basis to analyze a sporadic task set
with DAG tasks in Sec.~\ref{sec:global-DAG} and self-suspending tasks in Sec.~\ref{sec:global-SSS} when adopting global RM, in which we will demonstrate similar structures as used in Theorem~\ref{thm:multiprocessor-GRM}. We also demonstrate a tighter test in Appendix F\citetechreport{}
for improving the schedulability test of global RM for sporadic tasks.
\subsection{Global RM for DAG Task Systems}
\label{sec:global-DAG}
For multiprocessor scheduling, the DAG task model has been recently studied \cite{DBLP:conf/ecrts/BonifaciMSW13}. The utilization-based analysis can be found
in \cite{Li:ECRTS14} and \cite{DBLP:conf/ecrts/BonifaciMSW13}.
Each task $\tau_i\in {\bf T}$ in a DAG task system is a parallel task.
Each task is characterized by its execution pattern, defined by a set
of directed acyclic graphs (DAGs). The execution time of a job of task
$\tau_i$ is one of the DAGs. Each node (subtask) in a DAG represents a
sequence of instructions (a thread) and each edge represents a
dependency between nodes. A node (subtask) is \emph{ready} to be
executed when all its predecessors have been executed. We will only
consider two parameters related to the execution pattern of task
$\tau_i$:
\begin{compactitem}
\item \emph{total execution time (or work)} $C_i$ of task $\tau_i$: This is
the summation of the execution times of all the subtasks
of task $\tau_i$ among all the DAGs of task $\tau_i$.
\item \emph{critical-path length} $\Psi_i$ of task $\tau_i$: This is the length of
the critical path among the given DAGs, in which each node is
characterized by the execution time of the corresponding
subtask of task $\tau_i$.
\end{compactitem}
The analysis is based on the two given parameters $C_i$ and
$\Psi_i$. Therefore, we can also allow flexible DAG structures. That
is, jobs of a task may have different DAG structures, under the total
execution time constraint $C_i$ and the critical path length
constraint $\Psi_i$. Therefore, the model can also be applied for conditional sporadic DAG task systems \cite{DBLP:conf/ecrts/BaruahBM15}.
With the above definition, we have the following lemma, in which the proof is
Appendix B\citetechreport{}.
\begin{lemma}
\label{lemma:gRM-DAG-sufficient}
Task $\tau_k$ in a sporadic DAG system with implicit deadlines is schedulable under global RM on $M$ identical
processors, if
\begin{equation}
\label{eq:gRM-DAGsufficient}
\exists t \mbox{ with } 0 < t \leq T_k {\;\; and \;\;} \Psi_k+\frac{C_k-\Psi_k}{M}+
\sum_{i=1}^{k-1}
\frac{W_i(t)}{M}\leq t,
\end{equation}
where $W_i(t)$ is defined in Eq.~\eqref{eq:W_i-multiprocessor}.
\end{lemma}
\begin{theorem}
\label{thm:multiprocessor-DAG}
Task $\tau_k$ in a sporadic DAG system with implicit deadlines is schedulable by global RM on
$M$ processors if
\begin{equation}
\label{eq:schedulability-DAG}
(\frac{\Psi_k}{T_k}+2) \prod_{j=1}^{k} (\frac{U_j}{M} + 1)\leq 3
\end{equation}
or
\begin{equation}
\label{eq:schedulability-DAG-augmentation-GRM}
\sum_{j=1}^{k} \frac{U_j}{M}\leq \ln{\frac{3}{\frac{\Psi_k}{T_k}+2}}.
\end{equation}
\end{theorem}
\begin{proof}
Based on Lemma~\ref{lemma:gRM-DAG-sufficient}, which is very similar
to Lemma~\ref{lemma:gRM-sufficient}, we can perform a
similar transformation as in Theorem~\ref{thm:multiprocessor-GRM}, in
which $\alpha_i \leq
\frac{2}{M}$ and $\beta_i \leq \frac{1}{M}$. By adopting
Lemma~\ref{lemma:framework-constrained}, we know that if
\begin{equation}
\label{eq:schedulability-DAG-weaker}
(\frac{\Psi_k + \frac{C_k-\Psi_k}{M}}{T_k}+2) \prod_{j=1}^{k-1} (\frac{U_j}{M} + 1)\leq 3,
\end{equation}
then task $\tau_k$ is schedulable. Due to the fact that $C_k-\Psi_k
\leq C_k$, we know that $(\frac{\Psi_k + \frac{C_k-\Psi_k}{M}}{T_k}+2)
\leq (\frac{\Psi_k + \frac{C_k}{M}}{T_k}+2) \leq
(\frac{\Psi_k}{T_k}+2) \cdot (\frac{U_k}{M} + 1)$. Therefore, if the
condition in Eq.~\eqref{eq:schedulability-DAG} holds, the condition in
Eq.~\eqref{eq:schedulability-DAG-weaker} also holds, which implies the
schedulability. With the result in Eq.~\eqref{eq:schedulability-DAG},
we can use the same procedure as in
Lemma~\ref{lemma:framework-totalU-exclusive} to obtain
Eq.~\eqref{eq:schedulability-DAG-augmentation-GRM}.
\end{proof}
\begin{corollary}
The capacity augmentation factor of global RM for a sporadic DAG system with implicit deadlines is $3.62143$.
\end{corollary}
\begin{proof}
Suppose that $\sum_{\tau_i} \frac{C_i}{M T_i} \leq \frac{1}{b}$ and
$\frac{\Psi_k}{T_k} \leq \max_{\tau_i} \frac{\Psi_i}{T_i} \leq
\frac{1}{b}$. Therefore, by
Eq.~\eqref{eq:schedulability-DAG-augmentation-GRM}, we can guarantee
the schedulability of task $\tau_k$ if $\frac{1}{b} \leq
\ln{\frac{3}{2+\frac{1}{b}}}$. This is equivalent to solving $x =
\ln{\frac{3}{2+x}}$, which holds when $x\approx
3.62143$ by solving the equation numerically. Therefore,
we reach the conclusion of the capacity augmentation factor
$3.62143$.
\end{proof}
\subsection{Global RM for Self-Suspending Tasks}
\label{sec:global-SSS}
The self-suspending task model extends the sporadic task model by allowing tasks to suspend themselves.
An overview of work on scheduling self-suspending task systems can be
found in \cite{RTSS14a}.
\ifbool{techreport}{In \cite{RTSS14a}, a general interference-based analysis framework was developed that can be applied to derive sufficient utilization-based tests for self-suspending task systems on uniprocessors.}{}
Similar to sporadic tasks, a self-suspending task releases jobs sporadically. Jobs alternate between computation and suspension phases. We assume that each job of $\tau_i$ executes for at most $C_i$ time units (across all of its execution phases) and suspends for at most $S_i$ time units (across all of its suspension phases). We assume that $C_i+S_i \leq T_i$ for any task $\tau_i \in \tau$; for otherwise deadlines would be missed. The self-suspending model is general: we place no restrictions on the number of phases per-job and how these phases interleave (a job can even begin or end with a suspension phase). Different jobs belong to the same task can also have different phase-interleaving patterns.
For many applications, such a general self-suspending model is needed due to the unpredictable nature of I/O operations.
We have the following lemma, in which the proof is in Appendix B\citetechreport{}.
\begin{lemma}
\label{lemma:gRM-suspension-sufficient}
Task $\tau_k$ in a self-suspending system with implicit deadlines is schedulable under global RM on $M$ identical
processors, if
\begin{equation}
\label{eq:gRM-suspension-sufficient}
\exists t \mbox{ with } 0 < t \leq T_k {\;\; and \;\;} C_k+S_k +\sum_{i=1}^{k-1}
\frac{W_i(t)}{M} \leq t,
\end{equation}
where $W_i(t)$ is defined in Eq.~\eqref{eq:W_i-multiprocessor}.
\end{lemma}
\begin{theorem}
\label{thm:multiprocessor-suspension}
Task $\tau_k$ in a sporadic self-suspending system with implicit deadlines is schedulable by global RM on
$M$ processors if
\begin{equation}
\label{eq:schedulability-suspension}
(\frac{C_k+S_k}{T_k}+2) \prod_{j=1}^{k-1} (\frac{U_j}{M} + 1)\leq 3.
\end{equation}
\end{theorem}
\begin{proof}
By Lemma~\ref{lemma:gRM-suspension-sufficient}, we can perform a
similar transformation as in Theorem~\ref{thm:multiprocessor-GRM} with $\alpha_i \leq
\frac{2}{M}$ and $\beta_i \leq \frac{1}{M}$.
\end{proof}
\vspace{-0.05in}
\section{Conclusion}
\vspace{-0.05in}
With the presented applications, we believe that the general schedulability analysis \framework{} framework for fixed-priority scheduling has high potential to be adopted for analyzing other task models in real-time systems. We constrain ourselves by demonstrating the applications for simple scheduling policies, like global RM in multiprocessor scheduling. The framework can be used, once the $k$-point effective scheduling test can be constructed. Although the emphasis of this paper is not to show that the resulting tests for different task models by applying the \framework{} framework are better than existing work, some analysis results by applying the \framework{} framework have been shown superior to the state of the art.
For completeness, another document has been prepared in
\cite{DBLP:journals/corr/framework-compare} to present the similarity,
the difference and the characteristics of \framework{} and
\frameworkkq{}. With our frameworks, some difficult
schedulability test and response time analysis problems may be solved
by building a good (or exact) exponential-time test and applying these
frameworks.
Appendix D\citetechreport{} provides some case studies with evaluation
results of some selected utilization-based schedulability
tests. Appendix E\citetechreport{} further provides some additional
properties that come directly from the \framework{} framework. More
applications can be found in partitioned scheduling \cite{DBLP:journals/corr/Chen15k}, non-preemptive
scheduling \cite{DBLP:conf/ecrts/BruggenCH15}, etc.
\begin{spacing}{0.9}
\noindent{\small
{\bf Acknowledgement}: This paper has been supported by DFG, as
part of the Collaborative Research Center SFB876
(http://sfb876.tu-dortmund.de/), and the priority program
"Dependable Embedded Systems" (SPP 1500 -
http://spp1500.itec.kit.edu). We would also like to thank Dr. Vincenzo
Bonifaci for his valuable input to improve the presentation of the paper.
}
\end{spacing}
\footnotesize
\vspace{-0.05in}
\ifbool{techreport}{
\begin{spacing}{0.98}
}{
\begin{spacing}{0.85}
\vspace{-0.03in}
}
\def-0.3pt{-0.3pt}
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{introduction}
\noindent
As a result of their cosmological distances quasars (QSOs) appear as point-like objects. Various studies
have aimed to explore the detailed inner structure of quasars, which is unresolved even for low-redshift
active galactic nuclei (AGNs), because of their remote distances and sub-parsec scales of their emission
regions. In the standard AGN paradigm the central region is divided into an accretion disk, a dusty-torus,
a jet, a broad line region (BLR), and a narrow line region (NLR). Each of these regions contribute
differently to the AGN emission spectrum.
Direct imaging of the spatial structure of AGNs is possible with current instruments mainly probing longer
scales. To date, several interferometric studies of the central engine of the brightest AGNs (e.g.
\citealt{Jaffe2004, Tristam2013, Lopez2014}) have revealed the existence of a hot, parsec-scale disk that
is surrounded by warm dust extended in the polar direction. In the optical band the geometry of the
emission line region is investigated by indirect methods. Reverberation mapping establishes the
relationship between the size and the luminosity of the BLR and yields a typical BLR size of ${\rm R}_{\rm
BLR}\sim$\,0.2\,pc \citep{Kaspi2007, Chelouche2012} for high redshift luminous quasars. Differential
microlensing allows for a constraint on the accretion disk size $\la 3\times10^{-3}$\,pc
\citep{Blackburne2011, Vicente2012} and for an estimation of the size of the BLR $\sim0.1$ pc
\citep{Sluse2011}. The observations of gamma-ray emission constrain the size of a jet constituent to a few
parsecs \citep{Abdo2010}.
Another estimate of the size of the AGN emitting regions comes from constraints derived from covering
factor analysis of intervening H$_2$ bearing clouds which happen to cover the background source only
partially. Analysis of the partial coverage of Q\,1232$+$082 by a molecular hydrogen absorption cloud
allowed \citet{Balashev2011} to estimate the size of the C\,{\sc IV} BLR, R$_{\mbox{\small C\,{\sc
IV}}}\sim\,0.16^{+0.08}_{-0.11}$\,pc.
Molecular hydrogen absorption systems, a subset of damped Ly$\alpha$ systems (DLAs) and sub-damped
Ly$\alpha$ systems (sub-DLAs), reveal diffuse and translucent interstellar clouds in high redshift
intervening galaxies \citep{Noterdaeme2008}. An analysis of H$_2$ absorption systems allows for examining
the physical conditions of diffuse clouds in distant galaxies \citep{Srianand2005, Noterdaeme2007}. It has
been shown that the gas is a part of the cold neutral medium with comparatively low kinetic temperature
($T\sim 50-500\,$K) and high densities ($n_{\rm H}$~$>$~10~cm$^{-3}$), thus compact sizes ($l \la 1\,$pc).
Comparison of 21 cm and H$_2$ absorptions suggests that the H$_2$ absorption originates from a compact gas
that probably contains only a small fraction of H\,{\sc i} measured along the line of sight
\citep{Srianand2012}. These systems are important instruments for the analysis of several cosmological
problems, as follows. (i) The discovery of HD/H$_2$ clouds at high redshift \citep{Varshalovich2001}
provides an independent way to estimate the primordial deuterium abundance (D/H) and therefore the
relative baryon density of the Universe $\Omega_{\rm b}$ which is one of the key cosmological parameters
\citep{Balashev2010, Ivanchik2010}. (ii) The comparison of H$_2$ wavelengths observed in QSO spectra with
laboratory ones (just for this quasar Q\,0528$-$250, \citet{Varshalovich1993}, \citet{Cowie1995},
\citet{Potekhin1998}, \citet{Ubachs2004}, \citet{King2011}) allows us to test the possible cosmological
variation of the proton-to-electron mass ratio $\mu=m_p/m_e$. Because the de-composition of H$_2$
absorptions into several components is crucial for studies of the fundamental constant variability
problem, we should pay attention to the partial coverage effect. It is known that taking into account the
partial coverage effects, the physical model of the absorption system differs (see
\citealt{Balashev2011}). (iii) The interpretation of the relative populations of C\,{\sc i} fine-structure
excitation levels and CO rotational levels (\citealt{Srianand2000,Noterdaeme2011}) allows us to measure
the temperature $T_{\rm CMB}(z)$ of the cosmic microwave background radiation at high redshift.
Here, we argue that it is necessary to take into account the partial coverage of quasar emission regions
by a compact intervening H$_2$ cloud in order to derive a robust fit of the absorption lines. If this
effect is not taken into account, column densities can be underestimated by a factor of up to two orders
of magnitude. The first case of such an analysis has been presented by \citealt{Balashev2011} for
Q\,1232$+$082. The second case of partial coverage has been detected by \cite{Albornoz2014} for H$_2$
bearing cloud towards the quasar Q\,0643$-$504. The third case of partial coverage for the H$_2$ cloud at
$z_{\rm abs}=2.811$ in the spectrum of Q\,0528$-$2508 is presented in this study. We analyse a new
spectrum and detect residual flux in the bottom of saturated H$_2$ lines (${\rm J=0}$ and ${\rm J=1}$
levels). In case this flux is not taken into account, saturated lines yield large $\chi^2$ values and a
multicomponent model is used instead (e.g. \citealt{King2011}).
The remainder of this paper is organized as follows. A brief description of the data is given in
Section\,\ref{data}. The principles of partial coverage are described in Section\,\ref{pc}. In
Section\,\ref{H2analysis}, we present the analysis of the H$_2$ absorption system, accounting for partial
coverage . The HD molecular lines are explored in Section\,\ref{HD}. The results are discussed in
Section\,\ref{discussion}, and we give a brief conclusion in Section\,\ref{conclusion}.
\section{Data}
\label{data}
\noindent
The molecular hydrogen was identified for the first time at high redshift in the very spectrum of
Q\,0528$-$2508 \citep{Levshakov1985}. This quasar was observed many times, in particular during the period
between 2001 and 2009 using both spectroscopic arms of Ultraviolet and Visual Echelle Spectrograph (UVES)
of the Very Large Telescope (VLT); for a description of the instrument, see \citet{Dekker2000}. The log of
the observations used in our work is shown in Table~\ref{table_obs}. These observations relate to four
programmes, three of which were carried out in 2001--2002: 66.A-0594(A) (PI: Molaro), 68.A-0600(A) (PI:
Ledoux), and 68.A-0106(A) (PI: Petitjean). The instrument settings used during these observations were a
1-arcsec slit and 2x2 CCD pixel binning in both arms, resulting in a resolving power of R$\sim$45000 in
the blue and R$\sim$43000 in the red. There was no ThAr lamp calibration attached to each of the
exposures. An additional series of observations was performed in 2008-2009 under programme 082.A-0087(A)
(PI: Ubachs). The settings for that programme were a 0.8-arcsec slit in the blue arm and a 0.7-arcsec slit
in the red. The 2x2 CCD pixel binning was also used at that time. This resulted in a resolving power of
R$\sim $60000 in the blue and R$\sim $56000 in the red. Because the goal of that programme was to set a
limit on the variation of $\mu$, ThAr lamp calibrations were also taken immediately after each
observation.
The data presented in Table 1 were reduced using the UVES Common Pipeline Library (CPL) data reduction
pipeline release 4.9.5 using the optimal extraction method\footnote{see the UVES pipeline user manual
available for download at ftp://ftp.eso.org/pub/dfs/pipelines/uves/uves-pipeline-manual-22.8.pdf}. The
inter-order background (scattered light inside the instrument) was carefully subtracted in both the
flat-field frames and the science exposures. Linear spline interpolation was used to produce a
two-dimensional background image, which was subsequently smoothed using an average boxcar. Fourth-order
polynomials were used to find the dispersion solutions. However, the errors only reflect the calibration
error at the observed wavelengths of the ThAr lines used for wavelength calibration. All the spectra were
corrected for the motion of the observatory around the barycentre of the Solar system and then converted
to vacuum wavelengths. These spectra were interpolated into a common wavelength array and generated the
weighted-mean combined spectrum using the inverse squares of errors as weights. All the available
exposures were utilized to increase the signal-to-noise ratio up to $\sim$60 per pixel in the wavelength
range of the H$_2$ absorption lines (at $z=2.811$). As shown below, this allows us to detect and study the
effect of partial coverage in details.
\begin{table}
\caption{Log of the observations}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
No. & UT Date & Program ID & Exposure & Slit \\
& & & (sec) & (arcsec) \\
\hline
1 & 03.02.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
2 & 04.02.2001 & 66.A-0594(A) & 2$\times$5655 & 1.0 \\
3 & 05.02.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
4 & 07.02.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
5 & 13.02.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
6 & 13.03.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
7 & 15.03.2001 & 66.A-0594(A) & 1$\times$5655 & 1.0 \\
8 & 17.10.2001 & 68.A-0600(A) & 1$\times$3600 & 1.0 \\
9 & 18.10.2001 & 68.A-0600(A) & 2$\times$3600 & 1.0 \\
10 & 08.01.2002 & 68.A-0106(A) & 2$\times$3600 & 1.0 \\
11 & 09.01.2002 & 68.A-0106(A) & 2$\times$3600 & 1.0 \\
12 & 10.01.2002 & 68.A-0106(A) & 2$\times$3600 & 1.0 \\
13 & 23.11.2008 & 082.A-0087(A) & 2$\times$2900 & 0.8-0.7 \\
14 & 25.11.2008 & 082.A-0087(A) & 1$\times$2900 & 0.8-0.7 \\
15 & 23.12.2008 & 082.A-0087(A) & 4$\times$2900 & 0.8-0.7 \\
16 & 25.01.2009 & 082.A-0087(A) & 1$\times$2900 & 0.8-0.7 \\
17 & 26.01.2009 & 082.A-0087(A) & 1$\times$2900 & 0.8-0.7 \\
18 & 26.02.2009 & 082.A-0087(A) & 1$\times$2900 & 0.8-0.7 \\
\hline
Total & \multicolumn{2}{c}{} & 108450 & \\
\hline
\end{tabular}
\label{table_obs}
\end{center}
\end{table}
\section{Effect of partial covering}
\label{pc}
\noindent
Partial covering implies that only a part of the background source is covered by the absorbing cloud.
Mainly this can be a consequence of the absorbing cloud size being comparable to, or even smaller than,
the projected extent of the background source.
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{CoveringFactor.eps}
\caption{\rm An illustration of the effect of partial coverage on the absorption-line profiles.
Different panels show (a) a highly saturated line with total coverage, $f_{\rm c}=1\,$,\,
(b) a highly saturated line with partial coverage, $f_{\rm c}=0.8\,$\, and
(c) a partially saturated line with the same partial coverage, $f_{\rm c}=0.8\,$. The line flux residual
(LFR) is the fraction of the QSO flux that is not intercepted by the cloud. It can be easily derived from
the spectral analysis in case (b), but it is not straightforward to detect in case (c).}
\label{Def_CF}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=0.95\textwidth]{diff_profiles5.eps}
\caption{\rm Panel (a) shows a single-component H$_2$ absorption line with $\log N=18.2$
and $b = 4$\,km\,s$^{-1}$ and a fit by a one-component model (red line). The flux at the bottom of the line
goes to zero. We add a residual flux of about 10\,per cent (for clarity) to the data in panels (b), (c), (d)
and (e). The flux at the bottom of the lines does not go to the zero level, and simultaneously the lines
have Lorentzian wings. A simple one-component model cannot describe this line (cases (b) and (c)) and
returns a large reduced $\chi^2$. However, if we add a new unresolved component to the model,
the $\chi^2$ will be significantly decreased (case (d)) but not the statistical criterion, the AICC
(see \citealt{King2011}), and the returned column density will be two order of magnitude too small.}
\label{diff_profiles}
\end{center}
\end{figure*}
Partial covering is readily detectable in the spectrum of a background quasar if the cores of the
saturated absorption lines do not reach the zero flux level. This indicates that a part of the radiation
from the QSO passes by the cloud. The covering factor characterizing partial coverage is defined by the
ratio,
\begin{equation}
f_{\rm c}=\frac{F_{\rm cloud}}{F_{\rm total}},
\end{equation}
where $F_{\rm cloud}$ is the flux that passes through the absorbing gas and $F_{\rm total}$ is the total
flux. Therefore the measured flux in the spectrum, $F(\lambda)$, can be written as:
\begin{equation}
F(\lambda) = [F_{\rm total}(\lambda) - F_{\rm cloud}(\lambda)] + F_{\rm cloud}(\lambda)
\exp[-\tau(\lambda)] ,
\end{equation}
here $\,\tau(\lambda)\,$ is the optical depth of the cloud at the wavelength $\lambda$. The line flux
residual (LFR) is the fraction of the QSO flux which is not covered by the cloud. These definitions are
illustrated in Fig.\,\ref{Def_CF}. The determination of the covering factor is trivial in the case of
highly saturated absorption lines (see Fig.\,\ref{Def_CF} panels (a) and (b)) while for a partially
saturated line (see Fig.\,\ref{Def_CF} case (c)) the analysis requires a more sophisticated procedure. In
this case it is necessary to use several absorption lines originating from the same levels but with
different values of $\lambda f$, which is the product of the oscillator strength, $f$, and the wavelength
of the transition, $\lambda$. Such an analysis has been performed by \citet{Ivanchik2010} and more
precisely by \citet{Balashev2011} for the spectrum of Q\,1232$+$082, and by \cite{Albornoz2014} for the
spectrum of Q\,0643$-$504. A similar situation was observed for HE0001$-$2340. \citet{Jones2010} have
considered the possibility of partial coverage of the BLR to explain the observed Mg\,{\sc ii} equivalent
widths. In contrast to the rare situations where partial covering occurs from intervening systems (see
also \citealt{Petitjean2000}), partial covering is typical for absorption systems associated with quasars
( e.g. \citealt{Petitjean1994,Rupke2005, Hamann2010, Muzahid2013}).
A failure to take into account the partial coverage effect in a spectroscopic analysis can lead to a
significant underestimation of the column density of an absorber. The systematic bias (of column density)
can exceed several orders of magnitude for saturated lines. As an example, consider an absorption line
which consists of one component and has the high column density. The spectrum and the corresponding
one-component model are shown in panel (a) of Fig.~\ref{diff_profiles}. In panels (b), (c), (d) and (e)
the same line is presented, but part of radiation from background source (10\,\% for clarity) passes by a
cloud. If we take into account the LFR, then the line can be properly fitted by a one component model and
we can recover the high input column density (with given accuracy) and measure the LFR value (panel (e) of
Fig.~\ref{diff_profiles}). If the residual flux is not taken into account, then the one component model --
Lorentzian (b) or Gaussian (c) profiles -- is not adequate, the reason being that a one-component Voigt
profile cannot describe the unsaturated bottom and far wings of the line simultaneously. Using additional
components, as shown in panel (d), a result with a satisfactory $\chi^2$ is obtained. However, this
solution is incorrect, because the resulting column density ($\log N = 15.8$) is much smaller than the
input one ($\log N = 18.2$). To distinguish between cases (d) and (e) we propose a new method based on an
analysis of several absorption lines with different oscillator strengths. The description of the method
and application to the analysis of H$_2$ in Q\,0528$-$250 are described in more detail later.
\section{Molecular hydrogen}
\label{H2analysis}
Molecular hydrogen lines are detected in the spectrum of Q\,0528$-$250 from the DLA at ${\rm
z_{abs}=2.811}$. Column density of neutral hydrogen in this DLA system is $\log N(\mbox{H\,{\sc i}}) =
21.35\pm0.07$ \citep{Noterdaeme2008}. H$_2$ lines correspond to transitions from rotational levels up to
${\rm J=5}$. To fit the molecular hydrogen lines, the spectrum has been normalized with a continuum
constructed by fitting the selected continuum regions devoid of any absorptions with spline.
\subsection{Number of components}
\label{numbcomp}
\begin{table}
\caption{Results of the previous analyses of the H$_2$ system at $z=2.811$ towards Q\,0528$-$250.} \label{refH2}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
\hline
Year & $\log N_{\rm tot}$ & ${\rm N_{Comp.}}$ & Resolution & Ref.\\
\hline
1985 & 16.46$\pm$0.07 & 1 & & [1] \\
1988 & 18.0 & 1 & 10 000 & [2]\\
1998 & 16.77$\pm$0.09 & 1 & 10 000 & [3]\\
2005 & 18.22$^{+0.13}_{-0.17}$ & 2 & 40 000 & [4]\\
2006 & 18.45$\pm$0.02 & -- & & [5] \\
2011 & 16.56$\pm$0.02 & 3 & 45 000 & [6]\\
{\bf 2015} & {\bf 18.28$\pm$0.02}& {\bf 2} & 45 000 & {\bf This work} \\
\hline
\hline
\multicolumn{5}{|l|}{[1] ~\citet{Levshakov1985}; [2] ~\citet{Foltz1988};}\\
\multicolumn{5}{|l|}{ [3] ~\citet{Srianand1998}; [4] ~\citet{Srianand2005};}\\
\multicolumn{5}{|l|}{[5] ~\citet{Circovic2006}; [6] ~\citet{King2011}}\\
\end{tabular}
\end{center}
\label{results}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{H2_structure.eps}
\caption{\rm Examples of H$_2$ line profiles corresponding to the transitions from ${\rm J=0-5}$ levels.
Two components are clearly seen in the ${\rm J=4}$ and 5 lines. We have found that two components are
enough to obtain a satisfying $\chi^2$. The origin of the velocity scale is taken at the redshift
of the H$_2$ component, ${\rm z_A = 2.81099.}$}
\label{H2_str}
\end{center}
\end{figure}
\begin{figure*}\begin{center}
\includegraphics[width=0.97\textwidth]{tmp1.eps}
\caption{\rm Some of the absorption lines of H$_2$ from the ${\rm J=0,1}$ levels detected in the spectrum
of Q\,0528$-$250 and the best Voigt profile fit (red line). Two components of the absorption system are shown
by blue dashed lines. Only unblended saturated lines are present. The presence of prominent Lorentzian wings
of the lines indicate high H$_2$ column density. The additional panels show a close-up of the bottom of
the corresponding lines. The dashed black horizontal line represents the zero flux level. It is clearly seen
that about 2\,per cent of the LFR is present at the bottom of the saturated H$_2$ lines. The x-axes show
the velocity offset from the centre of H$_2$ component at ${\rm z = 2.81099}$. In the profile of H$_2$
line L9R1, some pixels in the bottom of the line are lower than 2 per cent level, which is probably because
this line is blended with Ly$\alpha$ forest absorption lines.}
\label{Lwing_ex}
\end{center}
\end{figure*}
The profiles of the H$_2$ absorption lines have a complex structure that cannot be fitted with a single
component. At least two components are clearly seen in the lines corresponding to ${\rm J=4, J=5}$
rotational levels (see Fig.\ref{H2_str}). Since the first identification of the H$_2$ system in the quasar
\citep{Levshakov1985} other studies have been conducted, providing discordant results (see
Table~\ref{results}). \citet{Srianand2005} used a two-component model, while \citet{King2011} pointed out
that a three-component model is very strongly preferred over two-component model. \citet{King2011} used a
fitting procedure where they increased the number of components in the absorption system in order to
minimize the corrected Akaike information criterion, AICC\footnote{ This statistical criteria allows for a
choice of prefered model among several models with different numbers of fitting parameters. ${\rm
AICC}=\chi^2 + 2p + \frac{2p(p+1)}{n-p-1},$ where $p$ is the number of fitting parameters, $n$ is the
number of spectral points, included in an analysis.} \citep{Sugiura1978, King2011}. The resulting total
column density differs by two order of magnitude from \citet{Srianand2005}. In that case a criterion for
choosing preferred model is the consistent derived physical parameters of a cloud. It can be noted note
that by using H$_2$ column density reported by \citet{King2011} (based on the three-component model) a
$N{\rm{(HD)}}/2N{\rm{(H_2)}}$ ratio is obtained, that is about an order of magnitude higher than the
primordial one. Meanwhile, the H$_2$ column density in two-component model of \citet{Srianand2005} gives a
reasonable $N{\rm{(HD)}}/2N{\rm{(H_2)}}$ ratio, which is consistent with the typical values measured at
high redshift \citep{Balashev2010}.
A very important point for the choice of a reasonable absorption profile model in the case of
Q\,0528$-$250 is the presence of Lorentzian wings in ${\rm J=0}$, ${\rm J=1}$ line profiles (see
Fig.~\ref{Lwing_ex}). It is an indicator of the high H$_2$ column density of absorption system with $\log
N{\rm (H_2) > 18}$ which is consistent with the result reported by \citet{Srianand2005}. However, in the
new spectrum of Q\,0528$-$250 the H$_2$ lines with prominent Lorentzian wings have some residual flux at
the bottom which significantly differs from the zero flux level (see Fig.~\ref{Lwing_ex} and
Fig.~\ref{LymanForest}). As a consequence, the fit by \citet{Srianand2005} gives the large reduced
$\chi^2$. It is probable that the signal-to-noise ratio in the previous spectrum was insufficient to
detect the residual flux.
The current study shows that the residual flux detected in the bottom of saturated H$_2$ lines in the
spectrum of Q\,0528$-$250 is the result of a partial coverage effect. In this case, the profiles of H$_2$
lines can be very well fitted by two-component model. Therefore, there is no need to increase the number
of components in H$_2$ profiles (such as it is done by \citealt{King2011}) to explain the complex
structure of lines (an example is given in Fig.\,\ref{diff_profiles}). In the next three subsections, we
provide evidence of the existence of the partial coverage in the spectrum of Q\,0528$-$250.
\subsection{Zero-flux level correction}
\label{Zeroflux}
To measure the LFR in H$_2$ absorption lines we need to derive the zero flux level in the spectrum. A
non-zero flux in the core of a saturated absorption line can be the result of inaccurate determination in
between spectral orders of scattered light inside the instrument. The zero flux level in the spectrum can
be estimated using the saturated Ly$\alpha$ absorption lines which are numerous and almost uniformly
distributed over the wavelength range where H$_2$ absorption lines are located. Ly$\alpha$ lines are
associated with intergalactic clouds that are larger than several kpc, thus it is most likely that they
cover the background source completely.
Wide flat bottom lines ($\Delta \lambda >1$ \AA, i.e. $\delta v > 80$~km\,s$^{-1}$) were selected, which
guarantees that lines are saturated and therefore measured fluxes in the bottom of the lines (LFRs) are
the real zero flux level in the spectrum. These lines were selected in the spectral region
3500$-$4700\,\AA. To estimate the residual flux at the bottom of the line the procedure illustrated in
Fig.\,\ref{EQ_LFR} was implemented. We selected several pixels in a line profile for which flux is within
$f_{c}\pm1\sigma_i$, where $f_{c}$ is the flux at the centre of the line and $\sigma_i$ is the error in
pixel $i$. The residual flux was calculated as the median of the flux in the selected pixels. The width of
the line bottom was estimated as the difference between the right most and left-most selected pixels from
the line centre (see Fig.\,\ref{EQ_LFR}). The top-left panel of Fig.~\ref{LymanForest} shows the LFR
values obtained for the selected saturated Ly$\alpha$ absorption lines. The average value is found to be
$(-0.21\pm0.04)\%$. The standard deviation of the points is $\sigma\simeq 0.22\%$. The green stars show
the LFR values measured at the bottom of the Ly$\alpha$ and Ly$\beta$ lines associated with the DLA-system
at ${\rm z_{abs}=2.79}$ and with the Ly$\alpha$ line of the second DLA-system at ${\rm z_{abs} = 2.14}$.
These lines are the most saturated lines for the spectrum.
\begin{figure*}
\begin{center}
\includegraphics[width=0.97\textwidth]{zero_level_2.eps}
\caption{\rm Top left: Analysis of the zero flux level in the spectrum of Q\,0528$-$250.
The LFRs measured at the bottom of the saturated absorption lines in the Ly$\alpha$ forest are plotted
against the wavelength as filled squares. The residual fluxes are expressed in per cent of the continuum.
The green stars show the LFR measured in the Ly$\beta$ and Ly$\alpha$ absorption lines of the DLA system
at ${\rm z_{ abs}=2.79}$ and the Ly$\alpha$ absorption from the DLA system at ${\rm z_{abs}=2.14}$.
The average value of the zero flux level is shown by the red horizontal line. The scatter in the LFR at
the bottom of the lines (i.e. noise in the spectrum) is found to be at the level of 0.22 per cent.
Bottom left: a few examples of saturated Ly$\alpha$ lines. The third bottom panel is the zoom of a part
of the second panel. Top right: the residual fluxes at the bottom of the H$_2$ absorption lines from
rotational levels ${\rm J=0}$ (blue) and ${\rm J=1}$ (red) are plotted against wavelengths as filled squares.
The average value of the LFR in the H$_2$ lines is shown by the blue line. Blue circles and orange horizontal
lines show the positions of QSO broad emission lines and their widths taken from \citet{VandenBerk2001}.
Bottom right: A few saturated H$_2$ absorption lines are present. The red dashed lines show the estimated
residual flux in each H$_2$ line. The black line indicates the zero level. The first panel shows the zoom of
a part of the second panel. It can be seen that the flux in the pixels located at the bottom of saturated
H$_2$ lines systematically departs from zero. The shift is about 2\,per cent of the continuum.}
\label{LymanForest}
\end{center}
\end{figure*}
\subsection{Partial coverage of H$_2$ absorption lines}
\label{Resflux}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{ex2.eps}
\caption{\rm An illustration of the residual flux measurement procedure.
To estimate the residual flux we determine the median flux for pixels in a line profile
that are in the range of $\pm1\sigma$ from the flux at the nearest pixel to the central wavelength
$\lambda_c$. This region is shown by purple shading. We refer to the width of this region as the width
of the bottom of the line.}
\label{EQ_LFR}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=0.95\textwidth]{test_flambda.eps
\caption{\rm The residual flux at the bottom of the H$_2$ absorption lines is given versus the parameter
$\lambda f$ (where $f$ is the oscillator strength of a line). Blue and red points correspond to the H$_2$
lines from the ${\rm J=0}$ and ${\rm J=1}$ levels. The theoretical flux at the bottom of the one-component line
with a fixed column density $N$ and Doppler parameter $b$ is shown by the dashed curve. The violet and grey
curves correspond to column densities $\log N{\rm(H_2)}= 15$ and 16. It is seen that all points cannot be
described by one curve simultaneously. The blue dash-dotted line represents the calculated residual flux at
the bottom of a composite line, which consists of two components with $\log N{\rm(H_2)}= 15$ for both lines
and a velocity separation of $\delta v = 2\,{\rm km\,s^{-1}}$. The green dash-dotted line presents
the same model for three components with the same column densities and velocity separations of
$\delta v = 2$ and 4 km\,s$^{-1}$, respectively (see text). An example is shown in the right-hand top panel.
We fit a certain H$_2$ line (L2\,P1) by two models: one component with the LFR (blue line) and three components
(red line). Both models give adequate fit of the line. But model without the LFR cannot describe flux
in all lines together, as shown by the thick red line in the graph. In contrast, all points are described
by a one-component model with residual flux (see the thick blue line).}
\label{flambda}
\end{center}
\end{figure*}
To estimate the residual flux in H$_2$ lines we selected ${\rm J=0,1}$ lines without apparent blends. The
LFR was measured by the same technique as for the Ly$\alpha$ forest lines. The analysis found that the
flux in the bottom of the lines is quite constant over large velocity range $\delta v\simeq
10$\,km\,s$^{-1}$ (i.e. the dispersion of points is within the range of the average statistical error)
that is wider than the full width half maximum (FWHM) of the UVES (6\,km\,s$^{-1}$) and is comparable with
the shift between centres of the two components. Therefore, the structure of H$_2$ system has no effect on
the residual flux. The comparison of the obtained residual flux in the ${\rm J=0,1}$ H$_2$ lines with the
level of the residual flux in the Ly$\alpha$ forest lines is shown in Fig.\ref{LymanForest}. The filled
squares represent the residual flux in H$_2$ lines versus its location in the spectrum. We have estimated
the line flux residual at the level $2.22\pm0.54$\,\% of the continuum, which significantly exceeds the
zero-flux level.
Non-zero residual flux at the bottom of saturated lines can also be the result of the convolution of the
saturated lines with the instrumental function, or an imperfect data reduction, and/or of the blend of
several unresolved unsaturated lines. First, however H$_2$ lines are wide with widths larger than the FWHM
of the UVES spectrograph (6\,km/s,\,i.e.\,$\sim0.08$\,\AA), so after convolution with the instrument
function, the flux at the bottom of these lines must still go to zero. Secondly, the improper data
reduction is not a viable explanation of the residual fluxes of saturated Ly$\alpha$ lines are
consistently equal to zero. Also we tested the dependence of residual flux on the width of the bottom of
line for H$_2$ and for the Ly$\alpha$ forest lines. This is shown in Fig.\ref{AppA1}. Several lines in the
Ly$\alpha$ forest were detected; these have similar widths as the H$_2$ lines and go to the zero level.
The profiles of these lines are shown in Fig.\,\ref{AppA2}.
Lastly, the residual flux can not be a result of the composition of several unresolved unsaturated lines
because the lines have Lorentzian wings.
\subsubsection{The $\lambda f$ test}
An additional test which confirms the presence of the partial coverage effect is discussed here (and
illustrated in Fig.\,\ref{flambda}). Molecular hydrogen lines from ${\rm J = 0, 1}$ levels have very
different values of the product $\lambda f$ (from 0.6 for L0P1 to 36.0 for W1Q1). It is known that the
flux in the bottom of an absorption line decreases exponentially when $\lambda f$ increases, $F \propto
\exp(-\lambda f)$. The results of the calculation of this dependence for an absorption line in the
spectrum with VLT resolution (${\rm FWHM=6\,km\,s^{-1}}$) are shown in Fig.\,\ref{flambda} by the dashed
lines. For simplicity, the modeled line profile consists of one component. The Doppler parameter ${\rm
b=4\,km\,s^{-1}}$ and the Damping width $\Gamma_{lm}$ is the same as for the H$_2$ line L2P1. Because the
equivalent width of the line gets to the logarithmic part of the curve of growth, we do not consider the
different Doppler parameters. Two dashed curves (violet and grey) are calculated for the line with column
densities $\log N=15$ and 16. The curve is shifted from right to left as the column density $N$ increases.
For a higher column density, the residual flux equals zero for a wide range of $\lambda f$ (because the
line is saturated) and differs from zero only for small values of $\lambda f \le 5$ (the case of an
unsaturated line). In a case of a multicomponent line (i.e. the blend of two or three lines with small
column density $\log N=15$), the residual flux in the bottom of the blended line would also behave
exponentially (dashed-dot curves). However, the behavior of the residual fluxes in the bottom of the
saturated H$_2$ lines in the spectrum of Q\,0528$-$250 is quite different. The residual fluxes in the
H$_2$ lines are shown by the filled blue and red squares. The squares do not follow the expected behavior,
moreover the points scatter similarly around a median value in the whole range of $\lambda f$.
Two models were applied to obtain a consistent and correct fit to the H$_2$ lines. Model (i) considers the
best fit of the certain H$_2$ line (e.g. L2P1), which has a value of $\lambda f$ = 4.23 and ${\rm LFR =
1.7\,\%}$ of the total flux. The line has a wide flat bottom $(\delta v\sim10\,{\rm km\,s^{-1}})$ and
Lorentzian wings. To fit this line without the partial coverage, it is necessary to use several
unsaturated components in the line profile. Therefore, this model describes the line profile well. Model
(ii) takes into account the partial coverage, and the line L2P1 can be fitted by the one-component model
with high column density $\log{\rm \,N=18.2}$ and ${\rm LFR = 1.7\,\%}$ (the blue line in the right-hand
top panel). This model also describes the line profile accurately. Using only one H$_2$ line we cannot
determine the most probable model. However, if we consider several H$_2$ lines from the same J level which
cover a wide range of $\lambda f$, we will be able to discriminate it. It is known that the lines from the
same J level correspond to the same physical region of a molecular cloud; therefore the lines are
described by the same set of physical parameters (the column density ${\rm N}$ and Doppler parameter ${\rm
b}$). The difference between line profiles originating from the same J level is caused only by the
different values of $\lambda f$ for the lines. Therefore, the correct model of an absorption system must
describe all measurements of residual flux in the bottom of lines from a given J level simultaneously.
Using the best fit parameters for models (i) and (ii) we have calculated the dependence of the residual
flux on $\lambda f$ (thick red and blue curves, respectively). The thick red line cannot describe all
squares in the main panel simultaneously, whereas the blue thick line can.
\subsection{Voigt profile Fitting}
\label{VpF}
A Voigt profile fitting of the H$_2$ absorption lines was performed, taking into account the partial
coverage. To describe a complex structure of line profiles we have divided the total flux detected by an
observer into two parts: from a main source and an additional one. In addition, we consider the H$_2$
system to be composed of two components (A and B) at redshifts ${\rm z_{A} = 2.81099}$ and ${\rm z_{B} =
2.81112}$. The light from the main source ($\sim98\%$ of the total flux) is intercepted by the two H$_2$
components and does not produce any residual flux in the saturated lines. The light from the additional
source that passes by H$_2$ clouds is not absorbed and therefore produces a uniform residual flux in H$_2$
lines. The flux in the absorption line of two components A and B can be described as
\begin{equation}
\label{eq01}
F(\lambda) = (1-f_c) F_{total} + f_c F_{total}e^{-\tau_{\rm A}(\lambda)} e^{-\tau_{\rm B}(\lambda)},
\end{equation}
where $f_c$ (in relative units) is the covering factor for H$_2$ lines.
However, because the physical conditions in the A and B clouds (such us linear size, volume density, etc.)
might be different, we can expect the covering factors of quasar emission regions by two H$_2$ clouds also
to be different. In this case the construction of the H$_2$ line profiles is more complicated and we
present an analysis of this case below (see Appendix\,\ref{model2}). Here, it is important to note that
taking into account two covering factors does not allow for a better fit to the H$_2$ lines (see the
discussion in Appendix\,\ref{model2}).
Then the absorption lines for each $J$ level were described using seven fitting parameters: z$_{\rm A}$,
z$_{\rm B}$, b$_{\rm A}$, b$_{\rm B}$, N$_{\rm A}$, N$_{\rm B}$, $f_c$. We used uniform values of $f_c$
over the whole wavelength range, because the residual flux in the H$_2$ lines is almost independent of the
wavelength (see Fig.\,\ref{LymanForest}). The Doppler parameter b is a function of the rotational level
$J$. To estimate fitting parameters Markov Chain Monte Carlo (MCMC) method was implemented, and to speed
the convergence the Affine invariant ensemble sampler by \citet{Goodman2010} was used. The main advantage
of this searching algorithm is to better explore a parameter space and to avoid using the partial
derivatives of the $\chi^2$ function which eases a number of numerical issues. This allowed for more
reliable estimates of the fitting parameters in comparison with other algorithms.
\subsection{Fitting results}
The H$_2$ absorption system at $z=2.811$ has more than 130 absorption lines from ${\rm J=0}$ to ${\rm
J=5}$ rotational levels. For the current analysis, we selected H$_2$ lines that are free of any obvious
blends. The sample examined contains 99 lines. The best fit of H$_2$ lines are shown in
Fig.\ref{J0lines}$-$Fig.\ref{J5lines} ranked following wavelength positions. The best-fitting parameters
are illustrated in Table~\ref{table_res2}. The reduced $\chi^2$ is 1.08 (the number of fitting points
$\sim 1500$).
The H$_2$ absorption system is highly saturated. The total H$_2$ column densities are $18.10\pm0.02$ and
$17.82\pm0.02$ for the A and B components consequently. This is consistent with the presence of the
Lorentzian wings in the profiles of ${\rm J=0}$ and ${\rm J=1}$ levels. The obtained orto--to--para ratios
are $2.7\pm0.1$ and $3.2\pm0.3$ for the A and B components, respectively. The corresponding kinetic
temperatures of the H$_2$ clouds are ${\rm T_{01,A}=141\pm6\,K}$ and ${\rm T_{01, B} = 167\pm13\,K}$.
Fig.\,\ref{exc} shows a comparison of the H$_2$ excitation diagrams for our measurements and those of
previous work. The left-hand panel shows the result of the analysis performed by \citet{King2011}, the
right-hand panel shows the result from the present work. Because the results presented here are close to
data reported by \citet{Srianand2005}, these are not shown here. We show excitation diagrams for two H$_2$
components at ${\rm z_{A}=2.81099}$ (2.811001 found by \citet{King2011}) and ${\rm z_{B}=2.81112}$. The
third H$_2$ component found by \citet{King2011} at ${\rm z_{C}=2.8109346(11)}$ is represented by the green
stars. Although the ratio of ${\rm J=0}$ and ${\rm J=1}$ levels of third component is the same as in other
components, the excitation diagram is not physically realistic (the excitation temperatures ${\rm T_{02}}$
and ${\rm T_{13}}$ for the third component are negative). This might be the result of the incorrect model
being used. It is seen, that the discrepancy between our results and those of \citet{King2011} is larger
only for the low J levels, where the influence of the partial coverage effect on the structure of the line
profiles is significant. For high J levels, where column densities of H$_2$ are less, the results agree in
$1\sigma$. To sum up, the total column density of H$_2$ from the measurement is 18.284\,$\pm$\,0.025, that
is about two orders of magnitude larger than the value reported by \citet{King2011} 16.556\,$\pm$\,0.024.
It should be noted, that the same values of $N$ and b parameters for all transitions of one J level were
used. Using the model with taking into account the partial coverage effect we obtain the reduced
$\chi^2\simeq1$ without an increase of the statistical errors of the spectrum. This is important, because
in the previous analysis of H$_2$ system in Q\,0528$-$250 \cite{King2011} noted that without artificially
increasing the statistical errors, the reduced $\chi^2$ was $\gg1$ (see the caption of table\,6 in
\citealt{King2011}).
The value of the residual flux in H$_2$ lines is fitted as an independent parameter of an analysis. The
best value is $2.40\pm 0.07$\,\% of the continuum, which agrees with the average value of the residual
flux obtained from the analysis of ${\rm J=0,1}$ H$_2$ lines (see Section\,\ref{Resflux}).
\begin{table}
\caption{Column densities and Doppler parameters obtained from Voigt profile fitting of the H$_2$ system at ${\rm z_{abs}=2.811}$ towards Q\,0528$-$250 after taking care of partial coverage.} \label{table_res2}
\begin{tabular}{|c|c|c|c|c|}
\hline \hline
\!System\! & J & ${\rm z_{abs}}$ & $\log N\,{\rm (cm^{-2}})$ & b~(km s$^{-1}$)\\
\hline
\hline
A &0 & ~~2.8109950(20) & 17.50 $\pm$ 0.02 & ~2.66 $\pm$ 0.05 \\
&1 & ~~2.8109950(20) & 17.93 $\pm$ 0.01 & ~2.71 $\pm$ 0.05 \\
&2 & ~2.8109952(5) & 16.87 $\pm$ 0.03 & ~2.75 $\pm$ 0.03 \\
&3 & ~2.8109934(5) & 15.97 $\pm$ 0.07 & ~2.87 $\pm$ 0.07 \\
&4 & ~2.8109938(8) & 14.18 $\pm$ 0.01 & ~4.79 $\pm$ 0.11 \\
&5 & ~2.8109938(8) & 13.58 $\pm$ 0.02 & ~5.04 $\pm$ 0.39 \\
&&&&\\
B &0 & ~~2.8111240(20) & 17.16 $\pm$ 0.03 & ~1.17 $\pm$ 0.06 \\
&1 & ~~2.8111230(20) & 17.67 $\pm$ 0.02 & ~1.14 $\pm$ 0.06 \\
&2 & ~2.8111235(7) & 16.64 $\pm$ 0.03 & ~1.22 $\pm$ 0.02 \\
&3 & ~2.8111238(6) & 16.24 $\pm$ 0.06 & ~1.25 $\pm$ 0.03 \\
&4 & ~2.8111231(6) & 14.20 $\pm$ 0.01 & ~1.72 $\pm$ 0.09 \\
&5 & ~2.8111231(6) & 13.60 $\pm$ 0.02 & ~2.38 $\pm$ 0.45 \\
\hline
\hline
\end{tabular}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[width=0.97\textwidth]{Excitation_diagram.eps}
\caption{\rm The excitation diagram for H$_2$ towards Q\,0528$-$250. Here, $N_J$ is the column density of the transition from the $J$ level with $g_J$ degeneracy. Solid lines (blue, red and green) correspond to the excitation temperature ${\rm T_{01}}$ derived from $J=0$ and $J=1$ levels. The left-hand panel shows the result of the analysis performed by \citet{King2011}. The excitation diagrams for different components of the H$_2$ system are shown by blue, red and green colors. The right-hand panel shows the results from our present work. Instead of three unsaturated components we use two components with higher H$_2$ column densities.}
\label{exc}
\end{center}
\end{figure*}
\section{HD molecules}
\label{HD}
In the new spectrum (obtained by co-adding all previous and additional new observations) the molecular HD
lines associated with the H$_2$ absorption system were detected (using improved laboratory wavelengths,
\citealt{Ivanov2010}). The presence of HD lines in this system was first reported by \citet{King2011}. The
HD molecular lines are present only in the component B. Some of the HD lines are shown in
Fig.\,\ref{HDprofiles}. We have estimated the HD column density $\log{ N\rm(HD)_{\rm B}}=13.33\pm0.02$ by
analysis of the two most prominent unblended absorption lines, L4-0R(0) and L8-0R(0). The results of Voigt
profile fitting are presented in Table\,\ref{table_HD}. Other HD lines are highly blended and cannot be
used in the analysis. The obtained column density is significantly less than that required to produce
self-shielding, $\log{N\rm(HD)}\simeq\,15$, thus we can set only a lower limit to the isotopic ratio D/H
in the cloud. Because the total column density of H$_2$ in the component B is $\log
N{\rm(H_2)}=17.85\pm0.02$, we estimate ${\rm{D}}/{\rm{H}}\ge
N{\rm{(HD)}}/2N{\rm{(H_2)}}=\left(1.48\pm0.10\right)\times10^{-5}$. The obtained HD column density is
close to the result, $\log N{\rm(HD)} = 13.267 \pm 0.072$, reported by \citet{King2011}. However, taking
into account the significantly larger H$_2$ column density of the component B we have obtained a lower
value of $N{\rm(HD)}/2N{\rm(H_2)}$ than the result by \citet{King2011}. This limit is consistent with D/H
ratio obtained from the analyses of atomic species in quasar spectra (e.g. \citealt{Olive2012}). The
comparison of this result with other $N{\rm(HD)}/2N{\rm(H_2)}$ measurements at high redshift is shown in
Fig.\,\ref{HDtoH2}. The $N{\rm(HD)}/2N{\rm(H_2)}$ in this system is consistent with other values and
correspond to predictions of deuterium chemistry models of diffuse ISM clouds (e.g. \citealt{Balashev2010,
Liszt2014}).
Note that in the component B we detect HD and C\,{\sc i} whereas in the component A these species are not
present, despite the higher H$_2$ column density of this component in comparison with the component B. As
for the component A we set an upper limit for HD and C\,{\sc i} column densities which are $\log{
N\rm(HD)_{\rm A}}\le13.1$ and $\log N({\rm C\,\mbox{\sc i}})_{\rm A} \le 12.0$. The lack of HD and C\,{\sc
i} in the component A might be the result of higher local UV radiation (e.g. bright young stars near cloud
A). It can destroy HD and ionize C\,{\sc i} while the H$_2$ molecules self-shield against the local
stellar radiation because to their high column density.
\begin{table}
\caption{Best fitting parameters for HD molecular lines in the spectrum of Q\,0528$-$250.} \label{table_HD}
\begin{tabular}{|c|c|c|}
\hline \hline
${\rm z_{abs}}$ & ~~~~$\log N\, {\rm (cm^{-2})}$~~~~ & ${\rm b\, (km\,s^{-1})}$\\
\hline
~2.811121\,$\pm$\,0.000002 & ~13.33\,$\pm$\,0.02 & ~2.25 $\pm$ 0.53 \\
\hline
~{HD/2H$_2$}& ~~(1.48\,$\pm$\,0.10)\,$\times\,10^{-5}$ & ~ \\
\hline
\hline
\end{tabular}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{Q0528HD.eps}
\caption{\rm The line profiles of HD molecules in the absorption system at ${\rm z_{abs}=2.811}$ in the spectrum of Q\,0528$-$250. HD lines are detected in component B only. The fit to the W0-0R(0) line is inconsistent owing to a poor definition of the local continuum near the base of the Ly$\alpha$ line of the second DLA system at ${\rm z_{abs}=2.14}$.}
\label{HDprofiles}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.47\textwidth]{HDto2H2.eps}
\caption{\rm The measurements of $N({\rm HD})/N({\rm H_2})$ vs $N({\rm H_2})$ in absorption systems at high redshift. The data are taken from \citet{Ivanchik2015}. The estimates of $N({\rm HD})/N({\rm H_2})$ toward Q\,0528$-$250 from this work and \citet{King2011} are shown by red and blue circles, accordingly. The blue horizontal strip represents the ratio of atomic D\,{\sc i} and H\,{\sc i} measured in quasar spectra \citep{Olive2012}. The solid dashed line corresponds to constant column density of HD, $\log N({\rm HD})=15$.}
\label{HDtoH2}
\end{center}
\end{figure}
\section{Discussion}
\label{discussion}
\noindent
The presence of a residual flux at the bottom of the saturated lines of the H$_2$ system at ${\rm
z_{abs}=2.811}$ towards Q\,0528$-$250 can be interpreted according to the arguments described in the
following subsections.
\subsection{Unresolved multicomponent quasars}
At redshift ${\rm z\sim2}$, the internal structures in the emitting regions of of quasars with transverse
dimensions $\lesssim 3\, {\rm kpc}$ are unresolved by UVES observations. For example, binary quasars with
separations of the order of ${\rm \sim10\,kpc}$ \citep{Hennawi2006,Vivek2009} might remain unresolved.
Partial coverage can arise if Q\,0528$-$250 has a complex multicomponent structure and if not all of the
components are covered by the absorbing H$_2$ clouds.
The available Very Long Baseline Array (VLBA) image of Q\,0528$-$250 \citep{Kanekar2009, Srianand2012}
shows an unresolved component containing $\sim94$ per cent of the total flux in the radio band (see
table\,6 of \citealt{Srianand2012}). Another $\sim6$ per cent is probably emitted by a diffuse component.
However, it is intriguing to note the consistency of these numbers with our findings. The spatial size of
radio emission core component is $65\times380$\,pc \citep{Kanekar2009}, that is significantly larger than
the size of the H$_2$ clouds. We have also looked at the images of PKS\,0528$-$250 in 13 and 4\,cm taken
as part of National Radio Astronomy Observatories' VLBA calibrators. The sources are unresolved even at
higher resolution achieved in these images.
\subsection{Dust scattering}
Scattering by dust is characterized by a narrow radiation pattern. About 90\,per cent of the total flux is
scattered towards the observer within less than 5-deg opening angle (e.g. \citealt{Drain2003}). The size
of the scattering region can be much larger than the size of an H$_2$ cloud. This is why, even in the case
of total coverage of the QSO by an H$_2$ cloud, the scattered radiation can be registered as a residual
flux in the bottom of H$_2$ lines. The scattered flux by a dust-rich region of a DLA system can be
estimated as
\begin{equation}
F_{\rm sc} = F \left(1-e^{-\tau_{\rm DLA}}\right) \frac{\Omega_{\rm QSO}^{*}}{\Omega_{\rm QSO}} \frac{\Omega_{\rm DLA}}{\Omega_{\rm sc}},
\end{equation}
where $F$ is the flux that passes through an H$_2$ cloud and is registered by the observer, $\tau_{\rm
DLA}$ is the mean optical depth along the line of sight toward the quasar, $\Omega_{\rm QSO}$ and
$\Omega_{\rm DLA}$ are the solid angles (measured by an observer on Earth) of the QSO emission region and
the scattering region, accordingly. $\Omega_{\rm QSO}^{*}$ is the solid angle of the QSO emission region
measured by an observer at the position of the DLA and $\Omega_{\rm sc}$ is the solid angle in which most
of radiation is scattered by dust. Therefore, the residual flux in the bottom of H$_2$ lines is $LFR =
F_{sc}/F$. Assuming that a typical dimension of the scattering region is as large as $1-10\,{\rm kpc}$,
LFR can be of the order of several percent. This is consistent with observations. The scattered radiation
can produce a non-zero residual flux for metal lines also. However, this is not detected for
Q\,0528$-$250.
\subsection{Jet-cloud interaction}
The jet emission pointing toward us contributes to the continuum in the optical band. The extended part of
the jet interacts with the external ISM and can warm up a cloud which is distant from the central source.
Similar situation has been detected in the deep radio-optical-X-ray observations of the radio galaxy
PKS\,B2152$-$699 by \citep{Worrall2012}, who have reported the first high-resolution observations of the
radio jet in the direction of an optical emission-line high ionization cloud. The {\it Hubble Space
Telescope} image shows not only emission in the region of the high-ionized cloud -- interpreted as ionized
gas by \citet{Tadhunter1988} and \citet{Fosbury1998} -- but also emission associated with a radio knot.
The measured optical flux density of this knot at $4.97\,\times\,10^{14}$\,Hz is 2.5\,$\pm$\,0.2 $\mu$Jy
\citep{Worrall2012}.
This is strikingly similar to the value of the residual flux in the V optical band which is ${\rm
2.4\pm0.1\,\mu Jy}$ (using the apparent magnitude ${m_{\rm V}=17.34}$ of Q\,0528$-$250, \cite{Veron2010}).
The variability of a jet emission (during from a few days up to a month) should be high and should also
induce variability of the residual flux in the H$_2$ line. This could be tested by further observations.
\section{Conclusion}
\label{conclusion}
\noindent
We have performed a detailed analysis of the H$_2$ system at ${\rm z_{abs}=2.811}$ towards the quasar
Q\,0528$-$250. This is a well-known system and the first detection of H$_2$ molecules at high redshift
\citep{Levshakov1985}. The system has been analysed in a variety of research, but the results obtained are
not consistent (see Table \,\ref{results}). The discrepancies can be explained by the presence of a
residual flux in the bottom of H$_2$ lines, which has not been considered in previous research. We have
derived the mean value of the residual flux as $(2.22\pm0.54)$ per cent of the continuum. This is
significantly higher than the zero flux level, $(-0.21\pm0.22)$ per cent, determined by analysis of the
Ly$\alpha$ forest lines.
Taking into account the residual flux in the H$_2$ lines we have obtained a consistent fit of the H$_2$
system using a two-component model with high column densities. The derived total column densities of
components A and B are $\log N{\rm(H_2)=18.10\pm0.02}$ and $17.82\pm0.02$, respectively.
We have performed the analysis of HD absorption lines detected only in component B. The estimated column
density is $\log N{\rm (HD)}= 13.33 \pm 0.02$ and thus we derive
$N{\rm{(HD)}}/2N{\rm{(H_2)}}=(1.48\,\pm0.13)\times10^{-5}$. This value is consistent with other
measurements of $N{\rm{(HD)}}/2N{\rm{(H_2)}}$ in quasar spectra at high redshift and can be considered as
a lower limit of the primordial deuterium abundance (Balashev et al. 2010).
Some interpretations for the presence of the residual flux are being offered here: (i) a multicomponent
quasar; (ii) scattering by dust; (iii) a jet-cloud interaction. We favour the latter interpretation (iii).
However, new optical and radio observations of Q\,0528$-$250 are necessary to confirm this and reject the
others.
We argue that taking into account partial coverage effects is crucial for any analysis of H$_2$ bearing
absorption systems in particular when studying the physical state of high-redshift ISM.
\vspace{2mm}{\footnotesize {\rm Acknowledgments.} } This work is based on observations carried out at the
European Southern Observatory (ESO) under programmes ID 66.A-0594(A) (PI: Molaro), ID 68.A-0600(A) (PI:
Ledoux), ID 68.A-0106(A) (PI: Petitjean) and ID 082.A-0087(A) (PI: Ubachs) with the UVES spectrograph
installed at the Kueyen UT2 on Cerro Paranal, Chile.
The work is supported by Dynasty foundation and by
the RF President Programme(grant MK-4861.2013.2). RS and PPJ gratefully acknowledge support from the
Indo-French Centre for the Promotion of Advanced Research (IFCPAR) under Project N.4304-2.
\bibliographystyle{mn2e}
|
\section{Introduction}
Schr\"odinger's equation \cite{Schrodinger1926,Sakurai1993} is the fundamental equation of quantum mechanics that describes the evolution of a quantum state $\psi$ in continuous time:
\[ i \frac{\partial \psi}{\partial t} = H \psi. \]
We have set $\hbar = 1$, and $H$ is the Hamiltonian that characterizes the total energy of the system. For a particle of mass $m$ in free space, the Hamiltonian is simply the kinetic energy operator:
\[ H = \frac{-1}{2m} \nabla^2, \]
where $\nabla^2$ is the Laplace operator.
The continuous-time quantum walk \cite{FG1998b} is simply the discrete-space analogue of this. That is, the particle is confined to discrete positions in space, which can be expressed as the vertices of a graph, and to transitions expressed as the edges of the graph. Then the Laplace operator $\nabla^2$ is replaced by its discrete version $L = A - D$, where $A$ is the adjacency matrix of the graph ($A_{ij} = 1$ if two vertices $i$ and $j$ are adjacent, and 0 otherwise), and $D$ is the diagonal degree matrix ($D_{ii} = \text{deg}(i)$). With this discrete substitution, the Hamiltonian for a continuous-time quantum walk is
\[ H_\text{walk} = -\gamma L, \]
where we have also grouped the coefficients together into $\gamma$, which is the jumping rate (\textit{i.e.} amplitude per time) of the walk. Evolution by Schr\"odinger's equation with this Hamiltonian is precisely the definition of a continuous-time quantum walk \cite{FG1998b,CG2004}.
As an algorithmic tool, continuous-time quantum walks have been used for a variety of applications, including evaluating boolean formulas \cite{FGG2008}, identifying graph isomorphism \cite{Rudinger2012}, and performing universal computation \cite{Childs2009}. In this paper, we focus on their application to search \cite{CG2004} on regular graphs. This leads to two changes to the Hamiltonian. First, since we assume that the graph is regular, the degree matrix $D$ is simply a multiple of the identity matrix, so it can be dropped by rezeroing the energy or by factoring out a global, unobservable phase. Thus we can use the adjacency matrix $A$ instead of the graph Laplacian $L$. Second, we introduce a term that acts as a ``Hamiltonian oracle'' \cite{Mochon2007}, which marks the $k$ vertices to search for by potentials. With these two changes, the search Hamiltonian is
\begin{equation}
\label{eq:H}
H = -\gamma A - \!\!\!\!\!\! \sum_{i \in \text{marked}} \!\!\!\!\!\! | i \rangle \langle i |.
\end{equation}
The goal is to evolve the system $\ket{\psi(t)}$ by this Hamiltonian for as little time possible, from initially being in an equal superposition $\ket{s}$ over all the vertices,
\[ \ket{\psi(0)} = \ket{s} = \frac{1}{\sqrt{N}} \sum_{i = 1}^{N} \ket{i}, \]
to a state that, when measured, collapses to a marked vertex with high probability. Note that beginning in this state expresses our initial lack of knowledge of where the marked vertices might be, so it guesses each vertex with equal probability. Furthermore, this state is an eigenstate $H_\text{walk}$, so the quantum walk alone does not change our information about where the marked vertices might be---it only changes when including the oracle, as in \eqref{eq:H}.
As an example, let us consider the continuous-time quantum walk formulation of Grover's algorithm \cite{Grover1996,FG1998a}. Since Grover's algorithm solves the \emph{unstructured} search problem, one can move from any vertex to any other. So this is simply search on the complete graph of $N$ vertices \cite{CG2004}, an example of which is shown in Figure~\ref{fig:complete}.
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics{complete.pdf}
\label{fig:complete}
} \quad \quad
\subfloat[] {
\includegraphics{complete_evolution_a.pdf}
\label{fig:complete_evolution_a}
}
\caption{\textbf{(a)} The complete graph of $N = 6$ vertices, $k = 2$ of which are marked and denoted by double circles. Identically evolving vertices are identically colored and labeled. \textbf{(b)} The success probability as a function of time for search on the complete graph with $N = 1024$ and $k = 4$. The black solid curve is when $\gamma = \gamma_c = 1/N$, and the red dashed curve is when $\gamma = 2\gamma_c = 2/N$.}
\end{center}
\end{figure}
By symmetry, the marked vertices evolve identically, as do the non-marked vertices. So we respectively group identically-evolving vertices together, as shown by identical colors and labels in Figure~\ref{fig:complete}:
\[ \ket{a} = \frac{1}{\sqrt{k}} \sum_{i \in \text{red}} \ket{i}, \quad \ket{b} = \frac{1}{\sqrt{N-k}} \sum_{i \in \text{blue}} \ket{i}. \]
Then the system evolves in a two-dimensional subspace spanned by $\{\ket{a}, \ket{b}\}$, in which the search Hamiltonian \eqref{eq:H} is
\[ H = -\gamma \begin{pmatrix}
k-1 + \frac{1}{\gamma} & \sqrt{k(N-k)} \\
\sqrt{k(N-k)} & N-k-1 \\
\end{pmatrix}. \]
Following the analysis of \cite{WongDissertation}, but generalized to $k$ marked vertices, this has eigenstates
\[ \ket{\psi_{0,1}} \propto \ket{s} + \frac{1 - \gamma N \pm \mathrm{\Delta} E}{2\gamma\sqrt{kN}} \ket{a} \]
with gap in the corresponding eigenvalues $E_0$ and $E_1$
\[ \mathrm{\Delta} E = E_1 - E_0 = \sqrt{(1-\gamma N)^2 + 4k\gamma}. \]
When $\gamma$ takes its critical value of $\gamma_c = 1/N$, the eigenstates are proportional to $\ket{s} \pm \ket{a}$ with an energy gap of $\mathrm{\Delta} E = 2\sqrt{k/N}$, so the system evolves from $\ket{s}$ to $\ket{a}$ in time $\pi/\mathrm{\Delta} E = (\pi/2) \sqrt{N/k}$ (see Section 3 of \cite{Wong2015b} for an explicit calculation). This is the quadratic speedup of Grover's algorithm over a classical computer's $O(N/k)$. As a check, Figure~\ref{fig:complete_evolution_a} shows the probability of measuring the quantum walker at a marked vertex as a function of time, and it reaches $1$ at time $(\pi/2) \sqrt{1024/4} = 8\pi \approx 25.133$, as expected.
From a straightforward calculation in \cite{Wong2015c} (generalized to multiple marked vertices), $\gamma$ must be chosen within $o(1/N^{3/2})$ of its critical value of $1/N$ for the algorithm to evolve from $\ket{s}$ to the marked vertices in time $(\pi/2)\sqrt{N/k}$ for large $N$. This can be relaxed to evolve to the marked vertices with constant probability in time $\Theta(\sqrt{N/k})$ if $\gamma$ is chosen within $O(1/N^{3/2})$ of its critical value of $1/N$. If $\gamma$ is further away from its critical value than this, then the initial state $\ket{s}$ converges to an eigenstate of $H$ for large $N$. So asymptotically, the system stays in $\ket{s}$ throughout its evolution, only picking up a global phase. This is shown in Figure~\ref{fig:complete_evolution_a} with $\gamma = 2\gamma_c = 2/N$; this is far enough from the critical value of $1/N$ such that the success probability plotted as a function of time converges to a flat, horizontal line for large $N$.
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics{complete1.pdf}
} \quad \quad
\subfloat[] {
\includegraphics{complete2.pdf}
}
\caption{\label{fig:complete_move} Two equivalent ways to mark $k = 2$ vertices, denoted by double circles, on the complete graph of $N = 6$ vertices.}
\end{center}
\end{figure}
Since Grover's search problem is unstructured, the problem is unchanged no matter which $k$ vertices are marked. For example, the two configurations in Figure~\ref{fig:complete_move} are equivalent (\textit{i.e.}, isomorphic). Since the location of the marked vertices does not change the structure of the problem, the algorithm (the critical jumping rate $\gamma_c = 1/N$, the evolution, the runtime of $(\pi/2)\sqrt{N/k}$, \textit{etc.}) is unchanged.
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_one1.pdf}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_one2.pdf}
}
\caption{\label{fig:simplex_one} Two equivalent ways to mark $k = 1$ vertex, denoted by double circles, on the simplex of complete graphs with $M = 5$.}
\end{center}
\end{figure}
Even for search on non-complete graphs, which is called spatial search, it is possible to retain this property that the location of the marked vertices does not matter. The most common way is to search a vertex-transitive graph for a single marked vertex. This includes the hypercube \cite{CG2004}, arbitrary-dimensional periodic square lattices \cite{CG2004}, strongly regular graphs \cite{JMW2014}, the ``simplex of complete graphs'' \cite{MeyerWong2014}, and complete bipartite graphs \cite{Novo2015}. For example, two possible ways to mark $k = 1$ vertex on the simplex of complete graphs (defined more formally later) are shown in Figure~\ref{fig:simplex_one}, and they are clearly equivalent (\textit{i.e.}, isomorphic). Another way to make the arrangement of marked vertices irrelevant is by marking a cluster of vertices, such that moving the cluster leaves the search problem unaltered \cite{WongAmbainis2015}.
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_multimarked.pdf}
\label{fig:simplex_two1}
}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two2.pdf}
\label{fig:simplex_two2}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two3.pdf}
\label{fig:simplex_two3}
}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two4.pdf}
\label{fig:simplex_two4}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two5.pdf}
\label{fig:simplex_two5}
}
\caption{\label{fig:simplex_two}The five ways to distribute two marked vertices, indicated by double circles, on the simplex of complete graphs with $M = 5$. Identically evolving vertices are identically colored and labeled, and the labels indicate the subspace basis vectors that the vertices belong to.}
\end{center}
\end{figure}
Even though the location of a unique ($k = 1$) marked vertex on a vertex-transitive graph does not affect the search problem, when there are multiple marked vertices ($k > 1$), their locations generally do make a difference in spatial search. For example, there are five inequivalent ways to arrange $k = 2$ marked vertices on the simplex of complete graphs, and they are shown in Figure~\ref{fig:simplex_two}. The simplex of complete graphs is the $M$-simplex with each of its $M+1$ vertices replaced by a complete graph of $M$ vertices, so it has a total of $N = M(M+1)$ vertices. In this paper, we explicitly solve spatial search by continuous-time quantum walk for these five configurations, plus four more with a greater number of marked vertices. In doing so, we show that different arrangements of marked vertices can dramatically change the critical jumping rate $\gamma_c$ that is needed for the algorithm to succeed.
This shows that the algorithm is dependent on the configuration of the marked vertices. Although such a dependence has been shown for \emph{discrete}-time quantum walks \cite{Kempe2003}, this explicit demonstration seems new for \emph{continuous}-time quantum walks. This result also highlights differences between the two approaches. While $\gamma_c$ is primarily affected for continuous-time quantum walks, discrete-time quantum walks do not have this parameter. Instead, they are typically governed by a ``coin'' \cite{Meyer1996a,Meyer1996b}, and this coin is chosen in different ways to define a search problem \cite{AKR2005}. For example, with the choice from \cite{SKW2003}, one gets an algorithm that efficiently searches arbitrary-dimensional periodic square lattices with any configuration of two marked vertices \cite{AKR2005}. With more marked vertices, the runtime can change \cite{NR2015a}, even to the point of having no improvement over classically guessing \cite{AR2008}. With a different coin that yields precisely the phase flip in Grover's algorithm \cite{AKR2005,Wong2015b}, there are other exceptional configurations that cause no improvement over classical as well \cite{NR2015b}. Using Szegedy's \cite{Szegedy2004} method of defining a discrete-time quantum walk, one can also obtain a quadratic improvement in search over a classical random walk's ``extended'' hitting time when searching with multiple marked vertices \cite{KMOR2014}. This vast number of results for discrete-time quantum walks with multiple marked vertices dwarfs those for continuous-time quantum walks, of which this seems to be the first.
We choose the simplex of complete graphs for our analysis because it has enough structure to reveal interesting properties, yet enough symmetry to be analytically tractable. It was first used in quantum search to prove that connectivity is a poor indicator of fast quantum search \cite{MeyerWong2014}, and it has since been used to introduce a search algorithm that takes multiple walk-steps for each oracle query \cite{WongAmbainis2015} and to demonstrate faster search on a weighted graph \cite{Wong2015e}.
In the next section, we solve spatial search on the simplex of complete graphs with two marked vertices, whose five possible configurations were shown in Figure~\ref{fig:simplex_two}. In doing so, we show that the first case's critical jumping rate $\gamma_c$ differs from the other four. Afterwards, we solve search with a greater number of marked vertices---two cases with $M+1$ marked vertices and two cases with $2(M+1)$ marked---showing that the critical jumping rate changes in a similar manner to the $k = 2$ case. This shows that for spatial search by continuous-time quantum walk, the critical jumping rate $\gamma_c$ is dependent on the arrangement of the marked vertices.
\section{Two Marked Vertices}
The five possible configurations with $k = 2$ marked vertices were shown in Figure~\ref{fig:simplex_two}, where the marked vertices are indicated by double circles. As shown, case (a) has both marked vertices in the same complete graph, and the remaining four cases (b), (c), (d), and (e) have them in different complete graphs.
\begin{table}
\begin{center}
\caption{\label{table:simplex_two}The five cases of search with $2$ marked vertices, shown in Figure~\ref{fig:simplex_two}, with the subspace dimension and the two stage's critical jumping rate ($\gamma_c$), runtime, and evolution.}
\begin{tabular}{ccccc}
\toprule
Case & Dimension & $\gamma_{c}$ & Runtime & Evolution \\
\midrule
Figure~\ref{fig:simplex_two1} & 8D & $\frac{3}{M} + o\left(\frac{1}{M^{5/2}}\right)$ & $\frac{\pi}{6} M^{3/2}$ & $\ket{g} \rightarrow \ket{b}$ \\[1.5ex]
& & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2\sqrt{2}} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.5ex]
\midrule
Figure~\ref{fig:simplex_two2} & 4D & $\frac{2}{M} + o\left(\frac{1}{M^{5/2}}\right)$ & $\frac{\pi}{2^{5/2}} M^{3/2}$ & $\ket{g} \rightarrow \ket{b}$ \\[1.5ex]
& & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.5ex]
\midrule
Figure~\ref{fig:simplex_two3} & 8D & $\frac{2}{M} + o\left(\frac{1}{M^{5/2}}\right)$ & $\frac{\pi}{2^{5/2}} M^{3/2}$ & $\ket{g} \rightarrow \ket{b}$ \\[1.5ex]
& & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.5ex]
\midrule
Figure~\ref{fig:simplex_two4} & 11D & $\frac{2}{M} + o\left(\frac{1}{M^{5/2}}\right)$ & $\frac{\pi}{2^{5/2}} M^{3/2}$ & $\ket{g} \rightarrow \ket{b}$ \\[1.5ex]
& & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.5ex]
\midrule
Figure~\ref{fig:simplex_two5} & 13D & $\frac{2}{M} + o\left(\frac{1}{M^{5/2}}\right)$ & $\frac{\pi}{2^{5/2}} M^{3/2}$ & $\ket{m} \rightarrow \ket{b} + \ket{h}$ \\[1.5ex]
& & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{b} + \ket{h} \rightarrow \ket{a} + \ket{d}$ \\[1.5ex]
\bottomrule
\end{tabular}
\end{center}
\end{table}
As with search with a single marked vertex \cite{MeyerWong2014}, we get two-stage algorithms for each of these five cases, where the system evolves with one critical jumping rate $\gamma_{c1}$ for some time, and then with a second critical jumping rate $\gamma_{c2}$ for a (likely different) amount of time. The detailed calculations for all five cases, including the generalization of the first case to any constant number of marked vertices in a single complete graph, are in Appendix~\ref{appendix:two}, and the main results are summarized in Table~\ref{table:simplex_two}.
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two1_evolution_a.pdf}
\label{fig:simplex_two1_evolution_a}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_two2_evolution_a.pdf}
\label{fig:simplex_two2_evolution_a}
}
\caption{The success probability as a function of time for search on the simplex of complete graphs with $M = 100$ and $k = 2$ marked vertices. \textbf{(a)} Search for the first case using the correct $\gamma_c$'s and runtimes for both stages of the algorithm (black solid) and using the incorrect values from the other cases (red dashed). \textbf{(b)} Search for the second case using the correct $\gamma_c$'s and runtimes for both stages of the algorithm (black solid) and using the incorrect values from the first case (red dashed).}
\end{center}
\end{figure}
To give a sense of the calculations in Appendix~\ref{appendix:two} and explain the evolutions summarized in Table~\ref{table:simplex_two}, consider the first case in Figure~\ref{fig:simplex_two1}. The system evolves in an 8-dimensional subspace, independent of $M$, because there are only eight different kinds of vertices, as shown by the eight unique colors and labels in Figure~\ref{fig:simplex_two1}. We group identically-evolving vertices into basis vectors $\ket{a}$, $\ket{b}$, \dots, $\ket{h}$ for the 8D subspace. Then writing the search Hamiltonian \eqref{eq:H} in this 8D subspace, we find that for most values of $\gamma$, the initial state $\ket{s}$ is asymptotically an eigenvector of $H$, which means the system does not evolve except for acquiring a global, unobservable phase.
For the system to evolve, we must choose the jumping rate $\gamma$ to be within $o(1/M^{5/2})$ of its critical value $\gamma_{c1} = 3/M$ so that $\ket{s}$ experiences a ``phase transition'' \cite{CG2004}, where it becomes supported by two eigenvectors of $H$ instead of one. This can be found using degenerate perturbation theory \cite{JMW2014,Wong2014}, and it causes two eigenvectors of the Hamiltonian \eqref{eq:H} to be asymptotically proportional to $\ket{g} \pm \ket{b}$ with an energy gap of $\mathrm{\Delta} E = 6/M^{3/2}$. Since the equal superposition state $\ket{s}$ is asymptotically $\ket{g}$ for large $N$ (because the white $g$ vertices in Figure~\ref{fig:simplex_two1} overwhelming comprise most of the vertices in the graph for large $N$), near $\gamma = \gamma_c$, the system evolves from $\ket{s}$ to $\ket{b}$ in time $t_1 = \pi/\mathrm{\Delta} E = \pi M^{3/2} / 6$. From Figure~\ref{fig:simplex_two1}, this means probability has now collected at the correct complete graph (\textit{i.e.}, at the blue $b$ vertices), but is not yet at the marked red $a$ vertices within that complete graph. This is summarized in the first row of information in Table~\ref{table:simplex_two}.
For the second stage of the algorithm, we want to move the probability from $\ket{b}$ to the marked vertices $\ket{a}$. As shown in Appendix~\ref{appendix:two}, the Hamiltonian has two eigenstates that are asymptotically proportional to $\ket{b} \pm \ket{a}$ when the jumping rate $\gamma$ takes the critical value $\gamma_{c2} = 1/M$ with an energy gap of $\mathrm{\Delta} E = 2^{3/2}/\sqrt{M}$. So the system now evolves from $\ket{b}$ to $\ket{a}$ in time $t_2 = \pi / \mathrm{\Delta} E = \pi \sqrt{M}/2^{3/2}$. This causes the system to evolve from the blue $b$ vertices to the red $a$ vertices, which are marked, accomplishing the search. This is summarized in the second row of information in Table~\ref{table:simplex_two}.
The success probability for the entire evolution of the algorithm is shown in Figure~\ref{fig:simplex_two1_evolution_a} as the solid black curve. For most of the time, the system is in the first stage of the evolution, where probability asymptotically builds up at $\ket{b}$. Any buildup in $\ket{a}$ during this time is due to higher-order contributions that are negligible for large $N$. Then in the second stage of the algorithm, the probability quickly shifts from $\ket{b}$ to $\ket{a}$, indicated by the sudden spike in success probability in Figure~\ref{fig:simplex_two1_evolution_a}.
Examining search for all five configurations in Table~\ref{table:simplex_two} reveals that the first configuration behaves differently from the other four---it has a different critical jumping rate for the first stage of the algorithm $\gamma_{c1}$, and the runtime of both stages is different. Of these, the important difference is the critical jumping rate, which truly is different because $3/M$ and $2/M$ are separated by a gap of more than $o(1/M^{5/2})$. Thus evolving by the wrong configuration's value would cause the system to asymptotically stay in its initial state $\ket{s}$ throughout its evolution, only acquiring a global, unobservable phase. This is shown in Figure~\ref{fig:simplex_two1_evolution_a}, where the dashed red curve is search for the first configuration, but incorrectly using the critical $\gamma$'s and runtimes from the other cases. Note that the probability stays small, and any buildup from higher-order corrections vanishes for large $N$. Similarly, Figure~\ref{fig:simplex_two2_evolution_a} shows search with vertices in the second configuration in Figure~\ref{fig:simplex_two2}---with the correct critical $\gamma$'s and runtimes, success probability builds up as desired, but with the wrong values from the first configuration, it fails to build up. Note that the second, third, fourth, and fifth configurations corresponding to Figures~\ref{fig:simplex_two2}, \ref{fig:simplex_two3}, \ref{fig:simplex_two4}, and \ref{fig:simplex_two5}, have identical critical jumping rates and runtimes. So while rearranging marked vertices can change the critical jumping rate substantially, it could also do nothing.
Given the extensive calculations in Appendix~\ref{appendix:two} that are needed to derive the results in Table~\ref{table:simplex_two}, manually working through all configurations of $k$ marked vertices on the simplex of complete graphs is impractical using the current method. We do analyze in the next section, however, four additional configurations with a greater number of marked vertices, two of which show that moving marked vertices again changes the critical jumping rate, and two that show no change.
\section{Larger Examples}
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_onepercomplete.pdf}
\label{fig:simplex_M1a}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_fullplusone.pdf}
\label{fig:simplex_M1b}
}
\caption{\label{fig:simplex_M1}Two ways to distribute $M+1$ marked vertices, indicated double circles, on the simplex of complete graphs with $M = 5$. Identically evolving vertices are identically colored and labeled, and the labels indicate the subspace basis vectors that the vertices belong to.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\subfloat[] {
\includegraphics[width=2.0in]{simplex_twopercomplete1.pdf}
\label{fig:simplex_2M1a}
} \quad \quad
\subfloat[] {
\includegraphics[width=2.0in]{simplex_twopercomplete2.pdf}
\label{fig:simplex_2M1b}
}
\caption{\label{fig:simplex_2M1}Two ways to distribute $2(M+1)$ marked vertices, indicated double circles, on the simplex of complete graphs with $M = 5$. Identically evolving vertices are identically colored and labeled, and the labels indicate the subspace basis vectors that the vertices belong to.}
\end{center}
\end{figure}
Let us consider four examples with a larger number of marked vertices. The first two are shown in Figure~\ref{fig:simplex_M1}, and they are just two of the many ways to arrange $M+1$ marked vertices on the simplex of complete graphs; in subfigure (a), they are distributed so that each complete graph has one marked vertex, and in subfigure (b), all the marked vertices are clustered in a single complete graph, except for one. The other two configurations are shown in Figure~\ref{fig:simplex_2M1}, and they are just two of the many ways to distribute $2(M+1)$ marked vertices so that each complete graph has two marked vertices.
\begin{table}
\begin{center}
\caption{\label{table:simplex_larger}Some cases of search with multiple marked vertices with the subspace dimension, critical jumping rate ($\gamma_c$), runtime, and evolution.}
\begin{tabular}{ccccc}
\toprule
Case & Dimension & $\gamma_c$ & Runtime & Evolution \\
\midrule
Figure~\ref{fig:simplex_M1a} & 3D & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.5ex]
Figure~\ref{fig:simplex_M1b} & 7D & $1 + \frac{3}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2} \sqrt{M}$ & $\ket{g} \rightarrow \ket{b}$ \\[0.8ex]
\midrule
Figure~\ref{fig:simplex_2M1a} & 2D & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2\sqrt{2}} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\[1.1ex]
Figure~\ref{fig:simplex_2M1b} & 3D & $\frac{1}{M} + o\left(\frac{1}{M^{3/2}}\right)$ & $\frac{\pi}{2\sqrt{2}} \sqrt{M}$ & $\ket{b} \rightarrow \ket{a}$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
Unlike search with a single marked vertex \cite{MeyerWong2014} or two marked vertices in the previous section, search on these graphs are single-stage algorithms. The detailed proofs are given in Appendix~\ref{appendix:larger}, and they use the same techniques as the configurations with two marked vertices. The results are summarized in Table~\ref{table:simplex_larger}, and they show that the critical jumping rate $\gamma_c$ for the two configurations in Figure~\ref{fig:simplex_M1} differ substantially enough that the search algorithm will fail (\textit{i.e.}, the system will asymptotically stay in its initial state) if the wrong $\gamma_c$ is used. In Figure~\ref{fig:simplex_2M1}, however, the critical jumping rate $\gamma_c$ and runtime are the same. So again, we see that rearranging the marked vertices can change the critical jumping rate substantially, but it also might not change it at all.
\section{Conclusion}
We have shown that when there are multiple marked vertices, their configuration on a graph can affect the critical jumping rate $\gamma_c$ of the continuous-time quantum walk. Previous work on search by continuous-time quantum walk has avoided this effect by restricting search to a single marked vertex on vertex-transitive graphs, or by clustering marked vertices together. This highlights a difference between continuous- and discrete-time quantum walks as they are used for search, since discrete-time quantum walks do not have the jumping rate as a parameter.
This work leaves open how general configurations of multiple marked vertices affect search on the simplex of complete graphs, since our analysis only examined nine specific arrangements. As a curious observation, our results follow a pattern: the critical jumping rate $\gamma_c$ is affected by the number of marked vertices in each complete graph, not by their arrangement within them. That is, going from Figure~\ref{fig:simplex_M1}a to Figure~\ref{fig:simplex_M1}b moves marked vertices \emph{across} complete graphs, changing the number within each, which changes $\gamma_c$. On the other hand, going from Figure~\ref{fig:simplex_2M1}a to Figure~\ref{fig:simplex_2M1}b moves marked vertices \emph{within} complete graphs, and this rearrangement within complete graphs makes no difference to $\gamma_c$. This behavior is consistent with two marked vertices in Figure~\ref{fig:simplex_two}, and it might stem from each complete graph being sufficiently connected within itself that the arrangement of marked vertices within them does not matter. Whether this pattern holds in general is a subject of further investigation, as is how continuous-time quantum walks search other graphs with multiple marked vertices.
\begin{acknowledgements}
Thanks to Andris Ambainis for useful discussions. This work was supported by the European Union Seventh Framework Programme (FP7/2007-2013) under the QALGO (Grant Agreement No.~600700) project, and the ERC Advanced Grant MQC.
\end{acknowledgements}
\bibliographystyle{qinp}
|
\section{Introduction}
\subsection{Background: counting query answers and complexity}
The computational problem of evaluating a formula (of some logic)
on a finite relational structure is of central interest in
database theory and logic.
In the context of database theory, this problem is
often referred to as \emph{query evaluation}, as
it models the posing of a query to a database, in a well-acknowledged way:
the formula is the query,
and the structure represents the database.
We will refer to the results of such an evaluation as
\emph{answers};
logically, these are the satisfying assignments of the formula on the structure.
The particular case of this problem where the formula is a sentence
is often referred to as \emph{model checking},
and even in just the case of first-order sentences,
can capture a variety of well-known decision problems
from all throughout computer science~\cite{FlumGrohe06-parameterizedcomplexity}.
In this article, we are concerned with the counting version of
this problem in first-order logic,
namely, given a first-order formula and a finite structure,
output the \emph{number} of answers
(see for example~\cite{PichlerSkritek11-counting,GrecoScarcello14-counting,DurandMengel13-structuralcounting,ChenMengel15-pp-icdt}
for previous studies).
In the sequel,
we assume by default that (unless otherwise specified)
all formulas and sentences
are first-order and relational.
This problem
of \emph{counting query answers}
generalizes model checking, which can be viewed as the
particular case thereof where one is given a sentence and structure,
and wants to decide if the number of answers is $1$ or $0$,
corresponding to whether or not the empty assignment is satisfying.
In addition to this counting problem's basic and fundamental interest,
it can be pointed out that all practical query languages
supported by database management systems have
a counting operator.
Indeed, it has been argued that database queries with counting
are at the basis of decision support systems
that handle large data volume~\cite{GrecoScarcello14-counting}.
As has been previously articulated,
a typical situation in the database setting is the evaluation of a relatively short formula on a relatively large structure. Consequently, it has been argued that, in measuring the time complexity of query evaluation tasks, one could reasonably allow a slow (non-polynomial-time) preprocessing of the formula, so long as the desired evaluation can be performed in polynomial time following the preprocessing~\cite{PapadimitriouYannakakis99-database,FlumGrohe06-parameterizedcomplexity}.
Relaxing polynomial-time computation to allow arbitrary preprocessing of a \emph{parameter} of a problem instance yields, in essence, the notion of \emph{fixed-parameter tractability}.
This notion is at the core of
and is the primary tractability notion in
\emph{parameterized complexity theory}, which provides a taxonomy for classifying problems where each instance has an associated parameter.
Following this motivation,
whenever the problem of counting query answers
(or the model checking problem) is considered using
parameterized complexity, in this article,
the formula is taken to be the parameter.
The problem of counting query answers
is well-known to be computationally intractable.
It is possible to restrict this problem by
considering restricted classes of queries,
and then trying to understand which classes of queries
are computationally well-behaved in that they give
rise to a tractable case of the general problem.
Precisely, for a class $\Phi$ of first-order formulas,
define $\mathsf{count}(\Phi)$ to be the problem
where an instance is a formula $\phi \in \Phi$ paired
with a finite structure $\mathbf{B}$,
and the output is the number of answers of $\phi$ on $\mathbf{B}$.
We hence have a family of problems, one problem
$\mathsf{count}(\Phi)$
for each such formula class $\Phi$,
and one can inquire which of these problems are
tractable (and which are not).
We will also have cause to consider the case of
model checking, so,
when $\Phi$ is a class of first-order sentences,
define $\mathsf{MC}(\Phi)$ to be the \emph{model checking}
problem where an instance is a sentence $\phi \in \Phi$
paired with a finite structure $\mathbf{B}$,
and the output is yes or no depending on whether or not
$\mathbf{B}$ satisfies $\phi$.
The complexity of the problem family $\mathsf{MC}(\Phi)$ has
been considered in numerous papers; see
for example~\cite{Grohe07-otherside,AdlerWeyer09-treewidth,ChenDalmau12-decomposingquantified,ChenMarx13-blocksorted,Chen14-existentialpositive,ChenMueller14-hierarchy,Chen14-frontier}.
\subsection{Background: width}
\emph{Width} is a syntactic measure of logical formulas.
The \emph{width} of a first-order formula
$\phi$ is defined
as the maximum number of free variables over all subformulas
of $\phi$.
In studies of the problem family $\mathsf{MC}(\Phi)$,
width has emerged as a crucially relevant measure;
we now turn to explain how.
Say that a class $\Phi$ of first-order sentences has
\emph{bounded width} if there exists a constant $k \geq 1$
such that each $\phi \in \Phi$ has width at most $k$.
It is now well-known that bounded width sentence classes are
computationally desirable for model checking, made precise as follows.
\begin{observation} \cite{Vardi95-boundedvariable}
\label{obs:mc-ptime}
Suppose that $\Phi$ is a sentence class having bounded width.
The problem $\mathsf{MC}(\Phi)$
is polynomial-time decidable, via the algorithm that (given a sentence and a structure) simply computes the set of satisfying assignments for each subformula, inductively.
\end{observation}
As suggested above,
it is known that a problem is fixed-parameter tractable
if, after performing preprocessing on the parameter,
an instance
can be resolved in polynomial time~\cite[Theorem 1.37]{FlumGrohe06-parameterizedcomplexity}.
For a problem $\mathsf{MC}(\Phi)$, one can readily observe
that if each sentence $\phi \in \Phi$ can be algorithmically
translated to a sentence lying in a class having bounded width,
then $\mathsf{MC}(\Phi)$ is fixed-parameter tractable as a consequence
of the previous observation. This can be formalized as follows.
\begin{observation} \cite{Chen14-existentialpositive}
\label{obs:mc-fpt}
Suppose that $\Phi$ is a sentence class.
The following condition, which we will call the
\emph{classical condition}, is sufficient for
the problem $\mathsf{MC}(\Phi)$ to be fixed-parameter tractable:
there exists a sentence class $\Phi'$ having bounded width
and an algorithm $f$ that computes, for each $\phi \in \Phi$,
a logically equivalent sentence $f(\phi)$ that is in $\Phi'$.
\end{observation}
Research on the problem family $\mathsf{MC}(\Phi)$ has succeeded in
obtaining classifications on classes $\Phi$ of bounded arity
where
the quantifiers and connectives are restricted
(see for examples~\cite{Grohe07-otherside,ChenMarx13-blocksorted,Chen14-existentialpositive}).
An example relevant to the present article
is work of
Chen~\cite{Chen14-existentialpositive},
who studied existential positive logic
(by which mean the positive fragment of first-order logic
consisting of formulas built from
atoms, $\wedge$, $\vee$, and $\exists$),
and,
making crucial use of a hardness result by
Grohe~\cite{Grohe07-otherside},
observed that when $\Phi$ is a class of existential positive sentences
having bounded arity,
the problem $\mathsf{MC}(\Phi)$ is fixed-parameter tractable
if and only if the classical condition applies to $\Phi$.
(Here and elsewhere in our discussion, we assume the standard
complexity-theoretic hypothesis that FPT $\neq$ W[1].)
That is, the sufficient condition for fixed-parameter tractability
identified by Observation~\ref{obs:mc-fpt} is
the \emph{exclusive} explanation for fixed-parameter tractability,
in the described setting of existential positive logic.
Let us remark that the query preprocessing algorithm $f$ here
(in Observation~\ref{obs:mc-fpt})
is related to and akin to the database notion of a query optimizer that computes a query execution plan.
Existential positive logic is a natural fragment of
first-order logic~\cite{Rossman08-hompreservation} and is studied heavily
in database theory.
Existential positive formulas include and are
semantically equivalent to so-called
\emph{unions of conjunctive queries}, also known as
\emph{select-project-join-union queries},
which have been argued to be the most
common database queries~\cite{AbiteboulHullVianu95-foundationsdatabases}.
Recently, the present authors
generalized the mentioned dichotomy
on existential positive sentence classes,
by presenting
a classification theorem~\cite{ChenMengel16-ep-classification-arxiv}
describing the fixed-parameter tractable problems
of the form $\mathsf{count}(\Phi)$, where $\Phi$ is a bounded arity class
of existential positive formulas. The classification theorem is
in fact a trichotomy theorem, which demonstrates that
the studied problems $\mathsf{count}(\Phi)$ can exhibit three types of complexity behavior.
\subsection{Contributions}
As a means of introducing our contributions,
we here wish to highlight a
conceptual point:
the applicability of the above \emph{classical condition}
indicates that for the model checking problem,
\emph{first-order logic itself can be used as a model of computation
in which desirable, efficient algorithms can be expressed.}
This condition posits the existence of an algorithm that translates
first-order sentences to a polynomial-time evaluable format;
and, the particular format used therein is
that of a first-order sentence!
Let us highlight that here, logic can be viewed as playing
two complementary roles: on the one hand, the
computational problems of interest are phrased directly in terms of logic;
on the other hand, appropriate algorithmic solutions to this problem
are themselves describable by logical sentences.
Inspired by this perspective of
\emph{logic as a useful model of computation},
the present work was motivated by the desire to develop a
logic that could serve as a useful model of computation for the
problem of counting query answers---analogously to
how first-order logic itself
serves as a useful model of computation for the model checking problem.
Let us point out some desiderata that such a logic ought to fulfill.
First, recall that in the problem of counting query answers,
an instance is a first-order formula paired with a structure, and the
output is the \emph{number} of answers.
The hope, then,
would be to be able to translate a first-order formula $\phi$
to a sentence $\psi$ in the logic such that evaluating $\psi$
on a structure returns the number of answers to $\phi$ on the structure;
hence, in the logic, the evaluation of a sentence on a structure
ought to return a \emph{numerical quantity}, instead of
a propositional value as in usual first-order logic.
Second, the logic must accommodate the fact that, while
the classical condition explains all tractable cases of
model checking in existential positive logic (in the sense made precise),
there are classes $\Phi$ of queries on which the classical condition
does not hold, but for which the problem $\mathsf{count}(\Phi)$
is fixed-parameter tractable.
For instance, consider
the class $\Theta = \{ \theta_n ~|~ n \geq 1 \}$
where
$\theta_n(x_1, \ldots, x_n) = U_1(x_1) \wedge \cdots \wedge U_n(x_n)$.
The number of answers for $\theta_n$ on any structure $\mathbf{B}$
is straightforwardly verified to be $\prod_{i=1}^n |U_i^{\mathbf{B}}|$;
given $\theta_n$ and $\mathbf{B}$, this quantity can indeed be computed
in polynomial time, but it can be verified that the classical condition
does not apply to $\Theta$
\paragraph{$\sharp$-logic.}
In this article, we introduce and study a logic,
which we call \emph{$\sharp$-logic}
and which possesses the sought-after characteristics just described.
In $\sharp$-logic,
the evaluation of a \emph{$\sharp$-sentence}
(a type of formula in $\sharp$-logic) on a structure
returns an integer value.
From the discussed viewpoint of bounded width as
an explanation for the tractability of model checking,
the relationship of the counting query answers problem
to $\sharp$-logic is strongly analogous to
the relationship of model checking
to usual first-order logic. The following parallel of
Observation~\ref{obs:mc-ptime} holds.
(Note that the \emph{width} of a formula in $\sharp$-logic
will be defined in a natural way.)
\begin{observation}
\label{obs:count-ptime}
Suppose that $\Psi$ is a $\sharp$-sentence class having bounded width.
The problem $\mathsf{count}(\Psi)$
is polynomial-time computable.
\end{observation}
As expressed, one purpose of $\sharp$-logic is to allow for the translation of
a first-order formula $\phi$ to a $\sharp$-sentence $\psi$
such that $\psi$ \emph{represents} $\phi$ in that,
for any structure $\mathbf{B}$,
the number of answers to $\phi$ on $\mathbf{B}$
is equal to the quantity that results from evaluating $\psi$ on $\mathbf{B}$.
The following is an immediate consequence of
the previous observation, and a parallel of
Observation~\ref{obs:mc-fpt}.
\begin{observation}
\label{obs:count-fpt}
Suppose that $\Phi$ is a first-order formula class.
The following condition, which we will call the \emph{counting condition},
is sufficient for the problem $\mathsf{count}(\Phi)$ to be
fixed-parameter tractable:
there exists a $\sharp$-sentence class $\Psi$ having bounded width
and an algorithm $f$ that computes,
for each $\phi \in \Phi$, a $\sharp$-sentence $f(\phi)$
that is in $\Psi$ and
that
represents $\phi$.
\end{observation}
Of course, Observation~\ref{obs:count-fpt} is only of interest
if the \emph{counting condition} applies to interesting
first-order formula classes $\Phi$, that is, only if this condition
allows one to explain the fixed-parameter tractability of
such problems $\mathsf{count}(\Phi)$.
We in fact show that for \emph{any} problem $\mathsf{count}(\Phi)$
that is tractable according to the
mentioned classification theorem on existential positive logic,
the counting condition applies to $\Phi$.
That is, the
counting condition is the
\emph{exclusive} explanation for the tractability of
$\mathsf{count}(\Phi)$ in this existential positive setting,
providing an analog to the result that
the classical condition is the
exclusive explanation for the tractability of
$\mathsf{MC}(\Phi)$ in the existential positive setting.
On a conceptual level, we view our result as strong evidence that,
for the problem of counting query answers, $\sharp$-logic
is a useful, expressive model of computation in which relevant,
efficient algorithms can be presented.
This result is obtained as an immediate consequence of two theorems:
\begin{itemize}
\item We show that when such a problem
$\mathsf{count}(\Phi)$ is
tractable, then there exists a bounded width class $\Psi$
of $\sharp$-sentences such that each $\phi \in \Phi$
has a representation in $\Psi$
(Theorem~\ref{thm:shepequi}).
\item We prove that there is a \emph{minimization algorithm} that,
given an existential positive formula, computes a representation
of minimum width
(Theorem~\ref{thm:shepcompute}).
\end{itemize}
The latter theorem can be read as demonstrating that
$\sharp$-logic is well-characterized and well-understood
as a model of computation: conceiving of a $\sharp$-sentence
representation of an existential positive formula
as a computational procedure for counting query answers,
this theorem provides a minimization algorithm that always outputs
an \emph{optimal} procedure for a given existential positive formula,
where optimality here is measured using width.
In short, our presentation and study of $\sharp$-logic
forwards the discussed use of logic as a means
for expressing computationally desirable procedures;
in particular, $\sharp$-logic allows for the direct expression
of procedures for counting query answers.
\paragraph{Counting homomorphisms.}
The problem of counting the number of homomorphisms
from a given \emph{source} structure $\mathbf{A}$
to a given \emph{target} structure $\mathbf{B}$
arises and has been studied in numerous contexts~\cite{GroheThurley11-counting-and-partition,Bulatov13-counting}.
This problem can be viewed as the special case of counting query
answers where the formula is a quantifier-free conjunction of atoms;
there is indeed a correspondence that allows one to pass
from a source structure $\mathbf{A}$ to such a formula $\phi_{\mathbf{A}}$
which
originates from the classical work~\cite{ChandraMerlin77-optimal}
(and which is explained in Section~\ref{sect:preliminaries}).
This problem is now well-known to be polynomial-time tractable
under a constant treewidth bound on the permitted
source structures; indeed, the corresponding algorithm,
which performs dynamic programming over a tree decomposition of $\mathbf{A}$,
has received
a textbook treatment~\cite[Section 5.3]{FominKratsch10-exact-exponential-algs}.
We discuss how,
from a tree decomposition for a structure $\mathbf{A}$,
one can polynomial-time compute a $\sharp$-sentence $\psi$
that has
width at most the width of the given decomposition (plus one),
and
that represents $\mathbf{A}$ in the sense that
evaluating $\psi$ on an arbitrary structure $\mathbf{B}$
always yields the number of homomorphisms from $\mathbf{A}$ to $\mathbf{B}$
(Example~\ref{ex:tree-decomposition-to-sh-formula}).
Combining this result with Observation~\ref{obs:count-ptime},
we obtain that the algorithm of this well-known polynomial-time
tractability result
can be expressed in $\sharp$-logic.
Indeed, we believe that the resulting $\sharp$-sentences
accurately, faithfully, and cleanly describe the execution
of this algorithm.
\paragraph{A dual perspective on a classical Lov\'{a}sz theorem.}
Fix a relational signature $\tau$; in the following discussion,
all structures are finite and on $\tau$.
Let $\mathbf{B}$ be a structure,
let $\mathrm{str}[\tau]$ denote the class of finite structures on $\tau$,
and let $L(\mathbf{B})$ be the vector from $\mathbb{Q}^{\mathrm{str}[\tau]}$
that maps a structure $\mathbf{A} \in \mathrm{str}[\tau]$
to the number of homomorphisms
from $\mathbf{A}$ to $\mathbf{B}$. A classical theorem of Lov\'{a}sz
states that, for any two structures $\mathbf{B}$, $\mathbf{B}'$,
it holds that $L(\mathbf{B}) = L(\mathbf{B}')$ if and only if
$\mathbf{B}$ and $\mathbf{B}'$ are isomorphic~\cite{Lovasz67-operations-structures}.
When one is concerned with homomorphisms from
one structure $\mathbf{A}$ to another structure $\mathbf{B}$,
sometimes, the structure $\mathbf{A}$ is referred to as the
\emph{left-hand} structure and the structure $\mathbf{B}$
is referred to as the \emph{right-hand} structure.
The vectors studied by Lov\'{a}sz indicate,
for a structure $\mathbf{B}$, the number of homomorphisms
coming \emph{from} each possible left-hand side structure.
One can naturally formulate a dual vector, as follows.
For any structure $\mathbf{A}$, define $R(\mathbf{A})$
to be the vector from $\mathbb{Q}^{\mathrm{str}[\tau]}$
that maps a structure $\mathbf{B} \in \mathrm{str}[\tau]$
to the number of homomorphisms from $\mathbf{A}$ to $\mathbf{B}$.
That is, the vector $R(\mathbf{A})$ indicates,
for a structure $\mathbf{A}$,
the number of homomorphisms \emph{to}
each possible right-hand side structure.
Our previous work~\cite{ChenMengel16-ep-classification-arxiv} implied
a dual of Lov\'{a}sz's theorem, namely, that
for any two structures $\mathbf{A}$, $\mathbf{A}'$,
it holds that $R(\mathbf{A}) = R(\mathbf{A}')$ if and only if
$\mathbf{A}$ and $\mathbf{A}'$ are isomorphic.
In the present work, we prove and use a
natural generalization of this fact,
namely, that for any finite sequence
$\mathbf{A}_1, \ldots, \mathbf{A}_k$ of pairwise non-isomorphic
structures, the vectors $R(\mathbf{A}_i)$ are
linearly independent (Theorem~\ref{thm:linearcombination}).
This linear independence theorem is used as a key tool
to establish the correctness of our minimization algorithm.
We believe that it should play an important role in future
studies of counting query answers, and that
the techniques and concepts that its proof requires
may be of independent and future interest.
Indeed, to prove this theorem,
we extend notions and techniques from the work
of Lov\'{a}sz~\cite{Lovasz67-operations-structures};
for example, we introduce and crucially use a notion of multivariate
polynomial associated to a primitive positive formula.
\subsection{Discussion}
Logics with counting mechanisms have been considered
in finite model theory and descriptive complexity;
one well-known example is the counting logic studied
by Immerman and Lander~\cite{ImmermanLander90-describing-graphs}.
A typical motivation in these areas for studying such logics
is the desire to extend first-order logic so as to allow for
the capture of properties not expressible in first-order logic.
This motivation contrasts somewhat with ours here;
our objective is to introduce logics that allow for
the relatively direct expression of useful algorithms
for the problem of counting query answers.
We believe that it could be of interest to try to understand
the relationship (if any) between existing counting logics
and the logics studied in the present work.
We wish to emphasize that, as regards our present motivations,
our logic\footnote{Precisely,
the \emph{$\sharp\mathsf{EP}$-formulas}, the
fragment of our logic that we focus on.}
trades off expressivity and computability properties
in an extremely desirable fashion.
On the one hand, the algorithms
for the tractable cases of $\mathsf{count}(\Phi)$
(where $\Phi$ is existential positive)
can be expressed in our logic, as described above;
on the other hand, our \emph{minimization algorithm}
described above evidences that the measure of \emph{width}
can be computably minimized in our logic, and is
thus in a certain sense well-characterized.
Let us remark here that
previous work established that
there is no algorithm for minimizing width
in positive first-order logic~\cite[Section 5]{BovaChen14-width-ep}.
We believe that our introduction of $\sharp$-logic
may open up further research directions.
One particular question for future research that we may pose
is whether or not there are
Ehrenfeucht-Fra\"{i}ss\'{e} style games for proving
inexpressibility in bounded width fragments of our logic.
\section{Preliminaries}
\label{sect:preliminaries}
\paragraph{Logic.}
We assume basic familiarity with the syntax and semantics of first-order logic.
In this article, we focus on relational first-order logic
where equality is not built-in to the logic.
Hence, each \emph{vocabulary/signature} under discussion consists
only of relation symbols.
We assume structures under discussion to be \emph{finite}
(that is, have finite universe);
nonetheless,
we sometimes describe structures as \emph{finite} for emphasis.
We use the letters $\mathbf{A}$, $\mathbf{B}$, $\ldots$ to denote structures,
and the corresponding letters $A$, $B$, $\ldots$
to denote their respective universes.
When $\mathbf{A}, \mathbf{B}$ are structures over the same signature $\tau$,
a \emph{homomorphism} from $\mathbf{A}$ to $\mathbf{B}$
is a mapping $h: A \to B$ such that, for each $R \in \tau$
and each tuple $(a_1, \ldots, a_k) \in R^{\mathbf{A}}$,
it holds that $(h(a_1), \ldots, h(a_k)) \in R^{\mathbf{B}}$.
We use the term \emph{fo-formula} to refer to a first-order formula.
An \emph{ep-formula} (short for \emph{existential positive formula})
is a fo-formula built from
\emph{atoms}
(by which we refer to predicate applications of the form $R(v_1, \ldots, v_k)$, where $R$ is a relation symbol and the $v_i$ are variables),
conjunction ($\wedge$), disjunction ($\vee$),
and existential quantification ($\exists$).
A \emph{pp-formula}
(short for \emph{primitive positive formula})
is an ep-formula where disjunction does not occur.
An fo-formula is \emph{prenex} if it has the form
$Q_1 v_1 \ldots Q_n v_n \theta$ where $\theta$ is quantifier-free,
that is, if all quantifiers occur in the front of the formula.
The set of free variables of a formula $\phi$ is denoted by
$\mathrm{free}(\phi)$ and is defined as usual;
a formula $\phi$ is a \emph{sentence} if $\mathrm{free}(\phi) = \emptyset$.
We define an \emph{ep-sentence} to be an ep-formula that is
a sentence, and define \emph{fo-sentence} and \emph{pp-sentence} similarly.
We now present some definitions and conventions that
are not totally standard.
A primary concern in this article
is in counting satisfying assignments of fo-formulas
on a finite structure.
The count is sensitive to the set of variables over which
assignments are considered; and, we will
\emph{sometimes} want to count
relative to a set of variables that is strictly larger than
the set of free variables.
Hence, we will often associate with each fo-formula $\phi$
a set $V$ of variables called the \emph{liberal variables},
denoted by $\mathrm{lib}(\phi)$, for which it is required that
$\mathrm{lib}(\phi) \supseteq \mathrm{free}(\phi)$.
We generally assume that the variables in
$\mathrm{lib}(\phi) \setminus \mathrm{free}(\phi)$ are not used in $\phi$.
To indicate that $V$ is the set of liberal variables of $\phi$,
we often use the notation $\phi(V)$;
we also use $\phi(v_1, \ldots, v_n)$, where the $v_i$
are a listing of the liberal variables.
Relative to a formula $\phi(V)$, when~$\mathbf{B}$ is a structure,
we will use $\phi(\mathbf{B})$ to denote the set of
assignments $f: V \to B$ such that $\mathbf{B}, f \models \phi$.
We call an fo-formula $\phi$ \emph{free} if $\mathrm{free}(\phi) \neq \emptyset$,
and
\emph{liberal} if
$\mathrm{lib}(\phi)$ is defined and $\mathrm{lib}(\phi) \neq \emptyset$.
\begin{example}
Consider the formula $\phi(x,y,z) = E(x,y) \vee F(y,z)$.
Define $\psi(x,y,z) = E(x,y)$ and $\psi'(x,y,z) = F(y,z)$.
The notation is intended to indicate that
$\mathrm{lib}(\phi) = \mathrm{lib}(\psi) = \mathrm{lib}(\psi') = \{ x, y, z \}$.
In the context of studying $\phi$, it is natural to define
$\mathrm{lib}(\psi)$ and $\mathrm{lib}(\psi')$ to be $\{ x, y, z \}$;
under this definition,
it holds that $\phi(\mathbf{B}) = \psi(\mathbf{B}) \cup \psi'(\mathbf{B})$,
but this would not hold in the case that
(say) $\mathrm{lib}(\psi)$ was defined as $\{ x, y \}$
(which set is equal to $\mathrm{free}(\psi)$).
\end{example}
\paragraph{pp-formulas.}
It is well-known~\cite{ChandraMerlin77-optimal}
that there is a correspondence between prenex pp-formulas
and relational structures.
In particular,
each prenex pp-formula $\phi(S)$ (on signature $\tau$)
with $\mathrm{lib}(\phi) = S$
may be viewed as a
pair $(\mathbf{A}, S)$
consisting of a structure $\mathbf{A}$ (on $\tau$)
and a set $S$;
the universe $A$ of $\mathbf{A}$ is
the union of $S$ with the variables appearing in $\phi$,
and the following condition
defines the relations of $\mathbf{A}$: for each $R \in \tau$,
a tuple $(a_1, \ldots, a_k) \in A^k$
is in $R^{\mathbf{A}}$
if and only if $R(a_1, \ldots, a_k)$ appears in $\phi$.
In the other direction,
such a pair $(\mathbf{A}, S)$ can be viewed as a prenex pp-formula $\phi(S)$
where all variables in $A \setminus S$ are quantified
and the atoms of $\phi$ are defined according to the above condition.
A basic known fact~\cite{ChandraMerlin77-optimal}
that we will use
is that
when $\phi(S)$ is a pp-formula corresponding to the pair $(\mathbf{A}, S)$,
$\mathbf{B}$ is an arbitrary structure, and $f: S \to B$ is an arbitrary map,
it holds that
$\mathbf{B}, f \models \phi(S)$ if and only if
there is an extension $f'$ of $f$ that is a homomorphism
from~$\mathbf{A}$ to $\mathbf{B}$.
\emph{We will freely interchange between the
structure view
and the usual notion
of a prenex pp-formula.}
For a prenex pp-formula specified as a pair $(\mathbf{A},S)$,
we typically assume that $S \subseteq A$.
\begin{example}
\label{ex:pp-views}
Consider the pp-formula
$\phi(u,v,w,x) = \exists y( E(u,v) \wedge F(w,y))$.
To convert $\phi$ to a structure $\mathbf{A}$,
we take the universe $A$ of $\mathbf{A}$ to be the union of
$\mathrm{lib}(\phi)$ with the variables appearing in $\phi$,
so $A = \{ u, v, w, x, y \}$.
The relations of $\mathbf{A}$ are as defined above,
so $E^{\mathbf{A}} = \{ (u,v) \}$ and $F^{\mathbf{A}} = \{ (w,y) \}$.
The resulting pair representation of $\phi$
is $(\mathbf{A}, \{ u, v, w, x \})$.
\end{example}
\begin{comment}
Two structures are \emph{homomorphically equivalent} if
each has a homomorphism to the other.
A structure is a \emph{core} if it is not homomorphically equivalent to a proper substructure of itself. A structure $\mathbf{B}$ is a
\emph{core} of a structure $\mathbf{A}$ if $\mathbf{B}$ is a substructure of $\mathbf{A}$ that is a core and is homomorphically equivalent to $\mathbf{A}$.
It is known that all cores of a structure are isomorphic and hence one sometimes speaks of \emph{the} core of a structure.
For a prenex pp-formula $(\mathbf{A},S)$
on signature $\tau$,
we define its \emph{augmented structure},
denoted by
$\mathsf{aug}(\mathbf{A},S)$,
to be the structure
over the expanded vocabulary $\tau\cup \{R_a\mid a\in S\}$ (understood to be a disjoint union)
where $R_a^{\mathsf{aug}(\mathbf{A},S)} =\{a\}$;
we define the \emph{core} of the pp-formula $(\mathbf{A},S)$
to be the core of~$\mathsf{aug}(\mathbf{A},S)$.
The following fundamental facts on pp-formulas
will be used throughout.
\begin{theorem}[follows from \cite{ChandraMerlin77-optimal}]\label{thm:ChandraMerlin}
Let $(\mathbf{A},V)$, $(\mathbf{B},V)$ be prenex pp-formulas.
The formula $(\mathbf{B},V)$ logically entails
the formula $(\mathbf{A},V)$ if and only if
there exists a homomorphism from
$\mathsf{aug}(\mathbf{A},V)$ to $\mathsf{aug}(\mathbf{B},V)$.
The formulas $(\mathbf{A}, V)$, $(\mathbf{B},V)$ are
logically equivalent if and only if
they have isomorphic cores, or equivalently, when
$\mathsf{aug}(\mathbf{A},V)$ and $\mathsf{aug}(\mathbf{B},V)$ are homomorphically equivalent.
\end{theorem}
\end{comment}
\paragraph{Graphs.}
Throughout the paper, all graphs under discussion
should be assumed to be undirected by default.
To every prenex pp-formula $(\mathbf{A}, S)$
we assign a graph
whose vertex set is $A \cup S$
and where two vertices are connected by an edge if they appear together in a tuple of a relation of $\mathbf{A}$.
A prenex pp-formula $(\mathbf{A}, S)$ is called \emph{connected} if its graph is connected.
A prenex pp-formula $(\mathbf{A}', S')$ is a
\emph{component} of a prenex pp-formula $(\mathbf{A}, S)$
over the same signature $\tau$
if there exists a set $C$ that forms a connected component
of the graph of $(\mathbf{A}, S)$, where
$S' = S \cap C$, and
for each relation $R \in \tau$,
a tuple $(a_1, \ldots, a_k)$ is in $R^{\mathbf{A}'}$
if and only if
$(a_1, \ldots, a_k) \in R^{\mathbf{A}} \cap C^k$.
Note that when this holds, the graph of $(\mathbf{A}', S')$
is the connected component of the graph of $(\mathbf{A}, S)$ on vertices $C$.
We will use the fact that, if $\phi(V)$ is a prenex pp-formula
and $\phi_1(V_1), \ldots, \phi_k(V_k)$ is a list of its components,
then for any finite structure $\mathbf{B}$,
it holds that
$|\phi(\mathbf{B})| = \prod_{i=1}^k |\phi_i(\mathbf{B})|$.
\begin{example}
Consider the formula $\phi$ from
Example~\ref{ex:pp-views}.
The connected components of the graph of $\phi$
can be readily verified to be
$\{ u, v \}$, $\{ w, y \}$, and
$\{ x \}$.
Hence, the pp-formula $\phi$ has three components,
which can be readily verified to be
$\phi_1(u,v) = E(u,v)$,
$\phi_2(w,y) = \exists y F(w,y)$, and
$\phi_3(x) = \top$;
here, $\top$ denotes the empty conjunction
\end{example}
\paragraph{Treewidth.}
We give some basic facts about tree decompositions
and treewidth; see for example~\cite{FlumGrohe06-parameterizedcomplexity} for more details.
A tree decomposition of a graph $G=(V,E)$ is a pair $(T, (B_t)_{t\in V(T)})$ where $T$ is a tree and $(B_t)_{t\in V(T)}$ is a family of subsets of $V$ such that (1)
for every $v\in V$, the set $\{t\in V(T)\mid v\in B_t\}$ is non-empty and connected in $T$, and
(2)
for every edge $uv\in E$, there is a $t\in V(T)$ such that $u,v\in B_t$.
The \emph{width} of a tree decomposition $(T, (B_t)_{t\in V(T)})$ is
defined as
$\max\{|B_t|: t\in V(T)\}-1$.
The treewidth $\mathsf{tw}(G)$ of $G$ is the minimum width over all the tree decompositions of $G$.
Computing tree decompositions of minimal width is fixed-parameter tractable parameterized by the
treewidth~\cite{Bodlaender96}.
A tree decomposition is called \emph{nice}, if every $t\in V(T)$ is of one of the following types:
\begin{itemize}
\item (leaf) $t$ has no children and $|B_t|=1$.
\item (introduce) $t$ has one child $t'$ and $B_t= B_{t'}\cup \{v\}$ for a vertex $v\in V\setminus B_{t'}$.
\item (forget) $t$ has one child $t'$ and $B_t= B_{t'}\setminus \{v\}$ for a vertex $v\in B_{t'}$.
\item (join) $t$ has two children $t_1, t_2$ with $B_t= B_{t_1}= B_{t_2}$.
\end{itemize}
It is well-known that a
width $k$ tree decomposition of $G$ can be converted to a
width $k$ nice tree decomposition, in polynomial time.
\section{$\sharp$-logic}
In this section, we present the syntax and semantics of $\sharp$-logic,
as well as some associated terminology.
Syntactically, $\sharp$-logic consists of \emph{$\sharp$-formulas};
each $\sharp$-formula $\phi$
has an associated set of free variables,
denoted by $\mathrm{free}(\phi)$,
as well as an associated set of \emph{closed} variables,
denoted by $\mathrm{closed}(\phi)$.
Let $\mathbf{B}$ be a structure, let $\phi$ be a $\sharp$-formula
over the signature of $\mathbf{B}$,
and let $h: \mathrm{free}(\phi) \to B$ be an assignment.
Semantically,
evaluating $\phi$ with respect to $\mathbf{B}$ and $h$
returns an integer value, as opposed to a propositional value
(as for a fo-formula); to present the symantics
of $\sharp$-logic,
we will notationally use $[\mathbf{B}, \phi]$
to denote the mapping taking an assignment
$h: \mathrm{free}(\phi) \to B$ to the integer value
provided to the triple consisting of
$\phi$, $\mathbf{B}$, and $h$.
Before turning to the formal definition of $\sharp$-logic,
we provide an example that previews some aspects
of this logic.
\begin{example}
We consider the pp-formula $\phi(x,y,z) = E(x,y) \wedge F(y,z)$.
The first type of $\sharp$-formula is a \emph{casting}
of a fo-formula; define $\psi$ to be
$C(E(x,y), \{ x, y \})$, which is a $\sharp$-formula
with free variables $\{ x, y \}$.
Let $\mathbf{B}$ be a structure, and let
$h: \{ x, y \} \to B$ be an assignment;
$[\mathbf{B}, \psi](h)$ is equal to $1$ or $0$
depending on whether or not $\mathbf{B}, h \models E(x,y)$ (respectively).
Once one has a $\sharp$-formula, it is possible to define
a further $\sharp$-formula by \emph{projecting} free variables.
For example, $P \{ y \} \psi$ is a $\sharp$-formula
with free variables $\{ x \}$.
When $g: \{ x \} \to B$ is an assignment,
the value $[\mathbf{B}, P \{ y \} \psi](g)$ is the
number of extensions $h: \{ x, y \} \to B$ of $g$ such that
$\mathbf{B}, h \models E(x,y)$.
In an analogous fashion, one may define
$\psi'$ to be the $\sharp$-formula
$C(F(x,z), \{ x, z \})$;
then,
when $g: \{ x \} \to B$ is an assignment,
the value $[\mathbf{B}, P \{ z \} \psi'](g)$ is the
number of extensions $h': \{ x, z \} \to B$ of $g$
such that
$\mathbf{B}, h' \models F(x, z)$.
Now, let $g: \{ x \} \to B$ be an assignment.
Observe that the number of extensions
$h^+: \{ x, y, z \} \to B$
of $g$ such that $\mathbf{B}, h^+ \models E(x,y) \wedge F(x,z)$
is equal to the product of $[\mathbf{B}, P \{ y \} \psi](g)$
and $[\mathbf{B}, P \{ z \} \psi'](g)$.
A product connective $\times$ is provided by $\sharp$-logic,
and said product is equal to
$[\mathbf{B}, (P \{ y \} \psi) \times (P \{ z \} \psi')](g)$.
Finally, $\theta = P \{ x \} ((P \{ y \} \psi) \times (P \{ z \} \psi'))$
is a $\sharp$-formula with no free variables;
letting $\epsilon$ denote the empty assignment,
$[\mathbf{B}, \theta](\epsilon)$ will be equal to
the sum, over all assignments $g: \{ x \} \to B$,
of $[\mathbf{B}, (P \{ y \} \psi) \times (P \{ z \} \psi')](g)$,
which is equal to the number of assignments $f: \{ x, y, z \} \to B$
such that $\mathbf{B}, f \models \phi$.
Phrased in terminology introduced below, the $\sharp$-formula
$\theta$ \emph{represents} the fo-formula $\phi(x,y,z)$.
\end{example}
\subsection{Syntax}
We define $\sharp$-formulas inductively, as follows.
\newcommand{\vspace{0pt}}{\vspace{0pt}}
\begin{itemize}
\item
$C(\phi, L)$ is a $\sharp$-formula
if $\phi$ is a fo-formula \\and $L \supseteq \mathrm{free}(\phi)$.
Define
$\mathrm{free}(C \phi) = L$ and $\mathrm{closed}(C \phi) = \emptyset$.
\vspace{0pt}
\item
$PV \phi$ is a $\sharp$-formula
if $\phi$ is a $\sharp$-formula and $V$ is
a set of variables with
$V \cap \mathrm{closed}(\phi) = \emptyset$.
Define $\mathrm{free}(PV \phi) = \mathrm{free}(\phi) \setminus V$\\
and $\mathrm{closed}(PV \phi) = V \cup \mathrm{closed}(\phi)$.
\vspace{0pt}
\item
$EV \phi$ is a $\sharp$-formula
if $\phi$ is a $\sharp$-formula
and $V$ is a set of variables
with $V \cap (\mathrm{free}(\phi) \cup \mathrm{closed}(\phi)) = \emptyset$.
Define $\mathrm{free}(EV \phi) = V \cup \mathrm{free}(\phi)$\\
and $\mathrm{closed}(EV \phi) = \mathrm{closed}(\phi)$.
\vspace{0pt}
\item
$\phi \times \phi'$ is a $\sharp$-formula
if $\phi$ and $\phi'$ are $\sharp$-formulas
with $\mathrm{free}(\phi) = \mathrm{free}(\phi')$
and $\mathrm{closed}(\phi) \cap \mathrm{closed}(\phi') = \emptyset$.
Define $\mathrm{free}(\phi \times \phi') = \mathrm{free}(\phi)$\\
and $\mathrm{closed}(\phi \times \phi') = \mathrm{closed}(\phi) \cup \mathrm{closed}(\phi')$.
\vspace{0pt}
\item
$\phi + \phi'$ is a $\sharp$-formula
if $\phi$ and $\phi'$ are $\sharp$-formulas
with $\mathrm{free}(\phi) = \mathrm{free}(\phi')$.
Define $\mathrm{free}(\phi + \phi') = \mathrm{free}(\phi)$\\
and $\mathrm{closed}(\phi + \phi') = \mathrm{closed}(\phi) \cup \mathrm{closed}(\phi')$.
\vspace{0pt}
\item $n$ is a $\sharp$-formula if $n \in \mathbb{Z}$.
Define $\mathrm{free}(n) = \emptyset$\\
and $\mathrm{closed}(n) = \emptyset$.
\end{itemize}
A formula $C(\phi,L)$ can be thought of as the casting
of a fo-formula $\phi$ into a $\sharp$-formula;
the $P$ quantifier can be thought of as projecting or
closing variables; and the $E$ quantifier can be thought of
as expanding the set of free variables.
The connectives $\times$ and $+$ perform the usual arithmetic
operations.
Let $\psi$ be a $\sharp$-formula.
A \emph{$\sharp$-subformula} $\theta$ of $\psi$
is defined as a $\sharp$-formula
such that, when $\psi$ and $\theta$ are viewed as trees,
$\theta$ appears as a subtree of $\psi$.
A \emph{fo-subformula} of $\psi$ is a
subformula of any fo-formula $\phi$
such that $\psi$ contains a $\sharp$-subformula of the form
$C(\phi, L)$.
A \emph{subformula} of $\psi$ is a $\sharp$-subformula or fo-subformula
of $\psi$.
We define $\mathsf{width}(\psi)$ to be the maximum of $|\mathrm{free}(\theta)|$
over all subformulas $\theta$ of $\psi$,
and $\sh\textup{-}\width(\psi)$ to be the maximum of $|\mathrm{free}(\theta)|$
over all $\sharp$-subformulas $\theta$ of $\psi$.
We say that $\psi$ is a \emph{$\sharp$-sentence} if $\mathrm{free}(\psi) = \emptyset$.
We define a \emph{$\sharp\mathsf{PP}$-formula} to be a $\sharp$-formula
where, in each $\sharp$-subformula of the form $C(\phi,L)$,
$\phi$ is a pp-formula; the notion of
\emph{$\sharp\mathsf{EP}$-formula} is defined analogously.
We define a \emph{$\sharp\mathsf{PP}$-sentence} to be a $\sharp\mathsf{PP}$-formula
that is a $\sharp$-sentence, and we define
a \emph{$\sharp\mathsf{EP}$-sentence} similarly.
\subsection{Semantics}
We define the semantics of our logic as follows.
\begin{itemize}
\item When $C(\phi,L)$ is a $\sharp$-formula,\\
$[\mathbf{B},\phi](h) = 1$ if $\mathbf{B}, h \models \phi$;
$[\mathbf{B},\phi](h) = 0$ otherwise.
\vspace{0pt}
\item When $PV \phi$ is a $\sharp$-formula,\\
$[\mathbf{B},PV \phi](h) = \sum_{h'} [\mathbf{B},\phi](h')$,
where the sum is over all extensions
$h': \mathrm{free}(\phi) \cup V \to B$
of $h$.
\vspace{0pt}
\item When $EV \phi$ is a $\sharp$-formula,\\
$[\mathbf{B}, EV \phi](h) = [\mathbf{B}, \phi](h \upharpoonright \mathrm{free}(\phi))$.
\vspace{0pt}
\item When $\phi \times \phi'$ is a $\sharp$-formula,\\
$[\mathbf{B}, \phi \times \phi'](h) = [\mathbf{B},\phi](h) \cdot [\mathbf{B},\phi'](h)$.
\vspace{0pt}
\item When $\phi + \phi'$ is a $\sharp$-formula,\\
$[\mathbf{B}, \phi + \phi'](h) = [\mathbf{B},\phi](h) + [\mathbf{B},\phi'](h)$.
\vspace{0pt}
\item When $n$ is a $\sharp$-formula,\\
$[\mathbf{B}, n](h) = n$.
\end{itemize}
We consider two $\sharp$-formulas $\phi$, $\phi'$ with
$\mathrm{free}(\phi) = \mathrm{free}(\phi')$ to be \emph{logically equivalent} if
for each structure $\mathbf{B}$, it holds that
$[\mathbf{B}, \phi] = [\mathbf{B}, \phi']$.
A $\sharp$-sentence $\psi$ \emph{represents} or is a \emph{representation} of
a fo-formula $\phi(V)$ if
for each finite structure $\mathbf{B}$,
it holds that $|\phi(\mathbf{B})| = [\mathbf{B}, \psi](\epsilon)$,
where $\epsilon$ is the empty assignment.
For simplicity, when $\psi$ is a $\sharp$-sentence,
we will typically write
$[\mathbf{B},\psi]$ in place of
$[\mathbf{B}, \psi](\epsilon)$.
We will use the term \emph{$\sharp\mathsf{PP}$-representation} to refer to
a representation that is a $\sharp\mathsf{PP}$-formula,
and define \emph{$\sharp\mathsf{EP}$-representation} similarly.
We make the basic observation that each fo-formula
has a representation.
\begin{prop}
For each fo-formula $\phi(V)$,
the $\sharp$-sentence $PV C(\phi,V)$
is a representation of $\phi(V)$.
\end{prop}
\subsection{Examples}
\begin{example}
Define
the formula $\phi(x_0,x_1,x_2,y_0,y_1,y_2)$
as $\phi_0 \wedge \phi_1 \wedge \phi_2$,
where
$\phi_i = \exists z_i T_i(x_i, x_{i+1}, y_i, z_i)$
for each $i \in \{ 0,1,2 \}$;
here, the quantity $i+1$ appearing in $x_{i+1}$ is computed modulo $3$.
These formulas are over the vocabulary $\{ T_0, T_1, T_2 \}$
having three relation symbols, each of arity $4$.
Define $\psi_i = P \{ y_i \} C(\phi_i, \{ x_0, x_1, x_2, y_i \})$
for each $i \in \{ 0,1,2 \}$.
Observe that when $\mathbf{B}$ is a structure and $h: \{ x_0, x_1, x_2 \} \to B$
is a map,
$[\mathbf{B}, \psi_i](h)$ gives the number of extensions
$h': \{x_0,x_1,x_2,y_i\} \to B$ of $h$ satisfying $\phi_i$ on $\mathbf{B}$.
We have $\mathrm{free}(\psi_0) = \mathrm{free}(\psi_1) = \mathrm{free}(\psi_2) = \{x_0,x_1,x_2\}$
and $\mathrm{closed}(\psi_i) = \{y_i\}$.
Now consider $\psi = (\psi_0 \times \psi_1) \times \psi_2$.
It can be verified that,
for a structure $\mathbf{B}$ and a map $h: \{ x_0, x_1, x_2 \} \to B$,
$[\mathbf{B}, \psi](h)$ gives the number of extensions
$h': \{ x_0, x_1, x_2, y_0, y_1, y_2 \} \to B$ of $h$
satisfying $\phi$ on $\mathbf{B}$.
It follows that the $\sharp$-sentence $P\{x_0,x_1,x_2\} \psi$
is a representation of $\phi$.
Note that this representation of $\phi$
has width equal to $\max(3,\mathsf{width}(\psi_0),\mathsf{width}(\psi_1),\mathsf{width}(\psi_2))=4$.
As a further remark, consider, as an example, the subformula
$C(\phi_0,\{x_0,x_1,x_2,y_0\})$ of $\psi_0$.
It holds that $x_2 \notin \mathrm{free}(\phi_0)$,
and so
$E\{x_2\} C(\phi_0,\{x_0,x_1,y_0\})$
is a $\sharp$-formula and is logically equivalent to
$C(\phi_0,\{x_0,x_1,x_2,y_0\})$.
\end{example}
\begin{example}
\label{ex:tree-decomposition-to-sh-formula}
Consider a prenex pp-formula $\phi$ without quantifiers
whose structure view has the form $(\mathbf{A}, A)$.
Suppose that $(T, (B_t))$ is a width $k$ tree decomposition
of the graph of $(\mathbf{A}, A)$;
without loss of generality, we may assume that this tree decomposition
is nice, and that the root node $r$ of $T$ has $B_r = \emptyset$.
We explain how to give a representation of $\phi$
having width $\leq k + 1$.
Say that a node $u$ of $T$ is \emph{below} a node $t$ of $T$
if $t$ occurs on the unique path from $u$ to the root of $T$
(this is understood to hold in particular when $u = t$).
When $t$ is a node of $T$, define $B_{\leq t}$
to be the union of $B_u$ over all nodes $u$ that are below $t$.
We show that, for each node $t$ of $T$,
there exists a $\sharp$-formula $\psi_t$ such that:
\begin{itemize}
\item $\mathrm{free}(\psi_t) = B_t$,
\item $\mathrm{closed}(\psi_t) = B_{\leq t} \setminus B_t$, and
\item for any structure $\mathbf{D}$,
the value $[\mathbf{D}, \psi_t](h)$ is equal to the number
of extensions $h': B_{\leq t} \to D$ of $h$ such that
$(\mathbf{D}, h')$ satisfies
each atom $R(a_1, \ldots, a_k)$ of $\phi$
whose variables
all fall into a bag $B_u$, with $u$ below $t$.
\end{itemize}
We give a $\sharp$-formula depending on the type of the node $t$;
we use the notation from the definition of nice tree decomposition.
\begin{itemize}
\item (introduce)
$\psi_t = (Ev \psi_{t'}) \times C(\alpha_1, B_t) \times \cdots \times C(\alpha_m, B_t)$
where the $\alpha_i$ are the atoms of $\phi$
whose variables fall into $B_t$.
Note that the presence of the $C(\alpha_i, B_t)$
ensures that $[\mathbf{D}, \psi_t](h) = 0$ if $(\mathbf{D}, h)$ does not satisfy
all of the $\alpha_i$.
\item (leaf) $\psi_t$ is defined as in the previous case,
except $\psi_{t'}$ is taken to be $C(\top, \emptyset)$,
where $\top$ is the true value.
\item (forget) $\psi_t = P \{ v \} \psi_{t'}$.
\item (join) $\psi_t = \psi_{t_1} \times \psi_{t_2}$.
In this case,
$\psi_{t_1} \times \psi_{t_2}$
is a $\sharp$-formula since
$\mathrm{free}(\psi_{t_1}) = B_{t_1} = B_{t_2} = \mathrm{free}(\psi_{t_2})$,
and, by the definition of tree decomposition,
$B_{\leq t_1} \setminus B_{t_1} = \mathrm{closed}(\psi_{t_1})$ and
$B_{\leq t_2} \setminus B_{t_2} = \mathrm{closed}(\psi_{t_2})$ are disjoint.
\end{itemize}
The desired representation is $\psi_r$.
The claim on the width of $\psi_r$ holds, as each
$\sharp$-subformula $\psi'$ of $\psi_r$
has $\mathrm{free}(\psi') = B_w$ for a node $w$ of $T$,
and each fo-subformula of $\psi_r$ is
either $\top$ or an atom
whose variables fall into a bag $B_t$.
\end{example}
\section{Main theorem statements}
The treewidth of a prenex pp-formula $\phi$,
denoted $\mathsf{tw}(\phi)$, is defined as the treewidth of the graph of $\phi$.
The following notions are adapted from~\cite{ChenMengel15-pp-icdt,ChenMengel16-ep-classification-arxiv}.
Let $(\mathbf{A}, S)$ be a prenex pp-formula with graph $G$.
An \emph{$\exists$-component} of $(\mathbf{A}, S)$
is a graph of the form $G[V']$ where there exists
$V \subseteq A$ that is a connected component of $G[A \setminus S]$
and $V'$ is the union of $V$ with all vertices in $S$ having
an edge to $V$. Define $\mathsf{contract}(\mathbf{A}, S)$ to be the
graph on vertices $S$ obtained by starting from $G[S]$
and adding an edge between any two vertices that appear together
in an $\exists$-component of $(\mathbf{A}, S)$.\footnote{
Note that in previous articles~\cite{ChenMengel15-pp-icdt,ChenMengel16-ep-classification-arxiv}, this graph was defined in terms of the
core of the given formula.
}
A \emph{core} of $(\mathbf{A}, S)$ is a prenex pp-formula $(c(\mathbf{A}),S)$
where $c$ is an $\mathbf{A}$-endomorphism
fixing each $s \in S$
that has minimum image size;
by $c(\mathbf{A})$, we mean the structure with universe $c(A)$
and where $R^{c(\mathbf{A})} = \{ c(t) ~|~ t \in R^{\mathbf{A}} \}$;
here, $c(t)$ denotes the tuple
obtained by applying $c$ to each entry of $t$.
It is known that any core of a prenex pp-formula $\phi$
is logically equivalent to $\phi$~\cite{ChandraMerlin77-optimal}.
We speak of \emph{the core} of a pp-formula, as it
is unique up to isomorphism; this follows from the
basic theory of cores~\cite{HellNesetril92-core}.
A class $\Phi$ of prenex pp-formulas satisfies
the \emph{contraction condition}
if the class containing each graph $\mathsf{contract}(\psi)$,
where $\psi$ is the core of a formula in $\Phi$,
has bounded treewidth;
$\Phi$ satisfies the
\emph{tractability condition} if it satisfies the contraction
condition and the cores of formulas in $\Phi$ have bounded treewidth.
Previous work showed that,
for bounded arity $\Phi$,
$\mathsf{count}(\Phi)$ is fixed-parameter
tractable if $\Phi$ satisfies the tractability condition;
interreducible with the parameterized clique problem
if $\Phi$ satisfies the contraction condition but not the tractability condition; and as hard as the parameterized counting clique problem
otherwise (see~\cite{ChenMengel15-pp-icdt} for a precise statement).
We first study representations of pp-formulas, obtaining
the following theorems.
\begin{theorem}\label{thm:shppequi}
Let $\Phi$ be a class of prenex pp-formulas.
\begin{itemize}
\item
The class $\Phi$ satisfies the tractability condition
if and only if
there exists $k \geq 1$ such that
each formula in $\Phi$ has a
$\sharp \mathsf{PP}$-representation $\phi'$ having $\mathsf{width}(\phi') \leq k$.
\item
The class $\Phi$ satisfies the contraction condition
if and only if
there exists $k \geq 1$ such that
each formula in $\Phi$ has a
$\sharp\mathsf{PP}$-representation $\phi'$ having $\sh\textup{-}\width(\phi') \leq k$.
\end{itemize}
\end{theorem}
\begin{theorem}\label{thm:shppcompute}
There exists an algorithm that,
given a prenex pp-formula $\phi$,
outputs a $\sharp\mathsf{PP}$-representation $\psi$ of $\phi$ of minimum width.
\end{theorem}
Building on this understanding of pp-formulas,
we are then able to achieve general versions of these theorems
for ep-formulas.
Previous work showed that for any class $\Phi$ of ep-formulas,
there exists a class $\Phi^+$ of prenex pp-formulas such that
the problems $\mathsf{count}(\Phi)$ and $\mathsf{count}(\Phi^+)$
are interreducible
(see~\cite{ChenMengel16-ep-classification-arxiv} for a precise statement),
and hence (for example)
whether or not the tractability condition holds on
$\Phi^+$ determines whether or not $\mathsf{count}(\Phi)$
is fixed-parameter tractable.
\begin{theorem}\label{thm:shepequi}
Let $\Phi$ be a class of ep-formulas.
\begin{itemize}
\item
The class $\Phi^+$ satisfies the tractability condition
if and only if
there exists $k \geq 1$ such that
each formula in $\Phi$ has a
$\sharp \mathsf{EP}$-representation $\phi'$ having $\mathsf{width}(\phi') \leq k$.
\item
The class $\Phi^+$ satisfies the contraction condition
if and only if
there exists $k \geq 1$ such that
each formula in $\Phi$ has a
$\sharp\mathsf{EP}$-representation $\phi'$ having $\sh\textup{-}\width(\phi') \leq k$.
\end{itemize}
\end{theorem}
\begin{theorem}\label{thm:shepcompute}
There exists an algorithm that,
given an ep-formula $\phi$,
outputs a $\sharp\mathsf{EP}$-representation $\psi$ of $\phi$
of minimum width.
\end{theorem}
\section{Independence theorem}
\label{sect:independence-theorem}
In this section, we establish a key tool for reasoning
about $\sharp\mathsf{EP}$-formulas.
Define two fo-formulas
$\phi(V)$, $\phi'(V')$
to be \emph{counting equivalent}
if, for each structure $\mathbf{B}$, it holds that
$|\phi(\mathbf{B})| = |\phi'(\mathbf{B})|$;
note that a decidable characterization
of counting equivalence on pp-formulas
is known (see Theorem~\ref{thm:renamingandcountingold}).
In the scope of this article, define a \emph{linear combination}
to be an expression of the form
$\sum_{i=1}^m c_i |\phi_i(V_i)|$,
where each $c_i$ is a non-zero rational number
and the $\phi_i(V_i)$
are pp-formulas that are pairwise not counting equivalent.
Each linear combination $\ell$ naturally induces a mapping
$\ell(\cdot)$
from finite structures to $\mathbb{Q}$.
The following theorem will be key for our understanding of
equivalence of $\sharp\mathsf{EP}$-formulas.
\begin{theorem} (Independence theorem)
\label{thm:linearcombination}
For any non-empty linear combination $\ell$,
there exists a finite structure $\mathbf{D}$ such that
$\ell(\mathbf{D}) \neq 0$.
\end{theorem}
We devote the rest of this section to proving this theorem.
By multiplying all values $c_i$ by a multiple of their denominators,
we can and will assume that each value $c_i$ is an integer.
In order to establish this theorem,
we will make use of
an action of a univariate polynomial
on a structure;
this action has the following commutativity property.
\begin{lemma}
\label{lemma:commutativity}
(Lov\'{a}sz \cite{Lovasz67-operations-structures})
For
each connected liberal pp-formula $\phi$,
each univariate polynomial $p$ with positive integer coefficients
and each finite structure $\mathbf{B}$, it holds that
$|\phi(p(\mathbf{B}))| = p(|\phi(\mathbf{B})|)$.
\end{lemma}
The following lemma shows that
certain pp-formulas can be controlled independently of each other,
made precise as follows.
\begin{lemma}
\label{lemma:orthogonality}
Let $\phi_1(S_1), \ldots, \phi_n(S_n)$ be connected liberal pp-formulas
(over signature $\tau$)
that are pairwise not counting equivalent.
Then for every $m \geq 2$, there exist structures
$(\mathbf{B}_{(a_1, \ldots, a_n)} ~|~ (a_1, \ldots, a_n) \in [m]^n)$
and injective functions
$f_1, \ldots, f_n: [m] \to \mathbb{N}$
such that for each $(a_1, \ldots, a_n) \in [m]^n$ and
each $i \in [n]$,
it holds that $|\phi_i(\mathbf{B}_{(a_1, \ldots, a_n)})| = f_i(a_i)$.
Moreover, when $\mathbf{A}$ is any structure on which
$|\phi_i(\mathbf{A})| > 0$ for each $i \in [n]$,
each structure
$\mathbf{B}_{(a_1, \ldots, a_n)}$ can be chosen
in the form
$\mathbf{A} \times \mathbf{E}$,
where $\mathbf{E}$ is a structure such that
$|\phi(\mathbf{E})| > 0$ for all pp-formulas over $\tau$.
\end{lemma}
\newcommand{\llbracket}{\llbracket}
\newcommand{\rrbracket}{\rrbracket}
We now introduce a highly useful notion, that of
\emph{component polynomial}.
Fix a set $V$ of liberal variables.
Denote by $\mathcal{E}$ the set of counting equivalence classes of
liberal connected pp-formulas (with liberal variables from $V$).
A \emph{component polynomial} $q$ is a multivariate polynomial
with integer coefficients over variables
$\{ X_e ~|~ e \in \mathcal{E} \}$.
For any finite structure $\mathbf{B}$,
we define the value of $q$ evaluated on $\mathbf{B}$,
denoted by $q \llbracket \mathbf{B} \rrbracket$,
as the integer value obtained by evaluating $q$
when each $X_e$ is given the value $|\phi_e(\mathbf{B})|$,
for a formula $\phi_e \in e$.
The following is our main theorem on component polynomials.
\begin{theorem}
\label{thm:component-polynomials}
When $q$ is a non-zero component polynomial,
there exists a finite structure $\mathbf{B}$ such that
$q\llbracket \mathbf{B} \rrbracket \neq 0$.
Moreover, when
$\phi_1(S_1), \ldots, \phi_n(S_n)$ are representatives
of the equivalence classes $e_1, \ldots, e_n \in \mathcal{E}$
whose corresponding variables $X_{e_i}$ appears in $q$,
the structure $\mathbf{B}$ may be picked as
a structure of the form provided by Lemma~\ref{lemma:orthogonality}.
\end{theorem}
In order to establish this theorem,
we will make use of the following known fact
concerning multivariate polynomials;
see for example~\cite[Lemma 2.1]{Alon99-nullstellensatz} for a proof.
\begin{prop}
\label{prop:non-zero-multivariate-polynomial}
Let $p(x_1, \ldots, x_n)$ be a multivariate polynomial
in $n$ variables over a field $F$.
For each $i \in [n]$, let $d_i$ denote the degree of $p$
as a polynomial in $x_i$,
and suppose that $T_i \subseteq F$ is a set of size $d_i+1$ or greater.
Then, if $p$ is not the zero polynomial,
there exists a point $(t_1, \ldots, t_n) \in T_1 \times \cdots \times T_n$
such that $p(t_1, \ldots, t_n) \neq 0$.
\end{prop}
\begin{proof} (Theorem~\ref{thm:component-polynomials})
Let $\phi_1(S_1), \ldots, \phi_n(S_n)$ be as described in the theorem statement.
Let $m \geq 2$ be a value that exceeds the degree of
each of the variables $X_{e_1}, \ldots, X_{e_n}$ in $q$, and
apply Lemma~\ref{lemma:orthogonality}
to obtain structures
\begin{center}
$(\mathbf{B}_{(a_1, \ldots, a_n)} ~|~ (a_1, \ldots, a_n) \in [m]^n)$
\end{center}
and the corresponding functions
$f_1, \ldots, f_n: [m] \to \mathbb{N}$.
Evaluating $q$ on these structures amounts
to evaluating $q$ when the variables
$(X_{e_1}, \ldots, X_{e_n})$ are given values in
$f_1([m]) \times \cdots \times f_n([m])$.
By Proposition~\ref{prop:non-zero-multivariate-polynomial},
$q$ must evaluate to a non-zero value on one of these structures.
\end{proof}
\begin{proof} (Theorem~\ref{thm:linearcombination})
Denote $\ell$ by $\sum_{i=1}^m c_i |\phi_i(V_i)|$
and let $(\mathbf{A}_1, V_1), \ldots, (\mathbf{A}_m, V_m)$ be the pairs
corresponding to the formulas $\phi_1(V_1), \ldots, \phi_m(V_m)$.
By rearranging the indices, we may assume for the sake of notation
that $\mathbf{A}_1, \ldots, \mathbf{A}_k$ are homomorphically equivalent
structures (where $k \in [m]$)
and that for no $i$ with $k < i \leq m$ does $\mathbf{A}_i$ have
a homomorphism to $\mathbf{A}_1$.
For any structure $\mathbf{B}$, it holds that one of the values
$|\phi_1(\mathbf{B})|, \ldots, |\phi_k(\mathbf{B})|$
is non-zero if and only if all of them are.
Now, for each $i$,
define $\widehat{\phi_i}(V_i)$ from $\phi_i(V_i)$
by removing non-liberal components,
that is, by removing each atom whose variables fall into
a non-liberal component.
Note that for every $i \in [k]$ and for every structure $\mathbf{B}$
such that $|\phi_i(\mathbf{B})|>0$, we have that $|\widehat{\phi_i}(\mathbf{B})|=|\phi_i(\mathbf{B})|$. Since the $\phi_i$ are pairwise not counting equivalent, it follows that the $\widehat{\phi_i}$ are pairwise not counting equivalent.
For each formula $\widehat{\phi_i}(V_i)$,
by considering its liberal connected components,
we may define $r_i$ to be a component polynomial
which is a product of variables from $\{ X_e ~|~ e \in \mathcal{E} \}$
such that $|\widehat{\phi_i}(\mathbf{B})| = r_i\llbracket \mathbf{B} \rrbracket$
for all finite structures $\mathbf{B}$.
The products $r_1, \ldots, r_k$ are pairwise distinct,
so $r = c_1 r_1 + \cdots + c_k r_k$ is a non-zero component polynomial.
By applying Lemma~\ref{lemma:orthogonality} with $\mathbf{A} = \mathbf{A}_1$
and then invoking Theorem~\ref{thm:component-polynomials},
we obtain a finite structure $\mathbf{D}$ of the form $\mathbf{A} \times \mathbf{C}$
such that $r \llbracket \mathbf{D} \rrbracket \neq 0$. Since there is by assumption a homomorphism from $\mathbf{A}_i$ to $\mathbf{A}$ and there is by Theorem 5.4 a homomorphism from $\mathbf{A}_i$ to $\mathbf{C}$, we have $|\phi_i(\mathbf{D})| > 0$ for every $i\in [k]$. Consequently, $|\phi_i(\mathbf{D})| = |\widehat{\phi_i}(\mathbf{B})|$ by the observation from above.
Since no structure $\mathbf{A}_i$ with $k < i \leq m$ maps homomorphically
to $\mathbf{A}$, we have
$|\phi_{k+1}(\mathbf{D})| = \cdots = |\phi_{m}(\mathbf{D})| = 0$
and hence $\ell(\mathbf{D}) = r \llbracket \mathbf{D} \rrbracket \neq 0$.
\end{proof}
\section{Quantifier aware width}
\label{sect:quantifier-aware-width}
In this section, we introduce a new width measure of pp-formulas which we call \emph{quantifier aware width}
and show that it is related to the width of $\sharp\mathsf{PP}$-formulas.
We here assume all tree decompositions of pp-formulas to be nice. So let $(T, (B_t)_{t\in T})$ be a nice tree decomposition of a pp-formula $\phi$. For every variable $x$ of $\phi$ let $\mathsf{top}(x)$ be the vertex $t$ of $T$ that is highest in $T$ such that $x\in B_t$.
We call a tree decomposition of $\phi$ \emph{quantifier aware} if for every $\exists$-component $C$ of $\phi$ and for all $x\in V(C)\setminus \mathrm{free}(\phi)$ and all $y\in V(C)\cap \mathrm{free}(\phi)$, we have that $\mathsf{top}(y)$ is on the path from $\mathsf{top}(x)$ to the root of $T$. We call the \emph{quantifier aware width} of a pp-formula $\phi$, denoted by $\mathsf{qaw}(\phi)$, the minimal treewidth of a quantifier aware tree decomposition of $\phi$ plus~$1$.
We call a $\sharp\mathsf{PP}$-formula $\phi$ \emph{basic}
if it does not contain $+$ nor subformulas of the form $n$,
where $n \in \mathbb{Z}$.
In a sense made precise in Appendix~\ref{app:qaw}, we will show the following three properties of quantifier aware width.
\begin{itemize}
\item For every class of pp-formulas, having bounded quantifier aware width is equivalent to the tractability condition
\item There is an FPT-algorithm parameterized by the quantifier aware width that given a pp-formula computes an equivalent pp-formula of minimal width.
\item Every pp-formula $\phi$ has a representation by a basic $\sharp\mathsf{PP}$-formula $\phi'$ and for every basic $\sharp\mathsf{PP}$-formula $\phi'$ there is a pp-formula $\phi$ ``representing'' it. Moreover, these translations can be performed so that $\mathsf{qaw}(\phi) = \mathsf{width}(\phi')$.
\end{itemize}
\section{Proofs of main theorems}
The following lemma shows that if we are only interested in $\sharp\mathsf{PP}$-representations of pp-formulas, we may restrict ourselves to basic $\sharp\mathsf{PP}$-formulas.
\begin{lemma}\label{lem:basicisenoughsketch}
Let $\phi$ be a pp-formula and let $\phi'$ be a $\sharp\mathsf{PP}$-representation of $\phi$. Then there is a basic $\sharp\mathsf{PP}$-sentence $\phi''$ that is also a $\sharp\mathsf{PP}$-representation of $\phi$ such that $\mathsf{width}(\phi'')\le \mathsf{width}(\phi')$ and $\sh\textup{-}\width(\phi'')\le \sh\textup{-}\width(\phi')$.
\end{lemma}
\begin{proof}(sketch) Using a normalization result for $\sharp\mathsf{PP}$-formulas (see Appendix~\ref{app:normalization}) and the results of Section~\ref{sect:quantifier-aware-width}, we can find numbers $c_i\in \mathbb{Z}$ and pp-formulas $\phi_i''$ such that for all structures $\mathbf{B}$
\begin{align}\label{eq:linearcombsketch}[\mathbf{B}, \phi'] = \sum_{i=1}^{\ell'} c_i' |\phi_i''(\mathbf{B})|,\end{align} where the $\phi_i''$ are pairwise not counting equivalent and have $\mathsf{qaw}(\phi_i'')\le \mathsf{width}(\phi')$.
Note that (\ref{eq:linearcombsketch}) is a linear combination.
Now note that $[\mathbf{B}, \phi']=|\phi(\mathbf{B})|$, which gives another linear combination. If follows with Theorem~\ref{thm:linearcombination} that
(\ref{eq:linearcombsketch}) consists only of one summand with coefficient $1$. Now applying the results of Section~\ref{sect:quantifier-aware-width} yields the result.
\end{proof}
\begin{proof}[of Theorem~\ref{thm:shppequi}](sketch)
Observing that by Lemma~\ref{lem:basicisenoughsketch} we may ourselves restrict to basic $\sharp\mathsf{PP}$-sentences, the statements can be shown with the results of Section~\ref{sect:quantifier-aware-width} in a rather straightforward fashion.
\end{proof}
\begin{proof}[of Theorem~\ref{thm:shppcompute}](sketch)
By Lemma~\ref{lem:basicisenoughsketch} we may assume that the desired $\sharp\mathsf{PP}$-representa\-tion is basic. But then the results of Section~\ref{sect:quantifier-aware-width} directly yield the theorem.
\end{proof}
\begin{proof}[of Theorem~\ref{thm:shepequi}](sketch)
Let first $\Phi^*$ satisfy the tractability condition. Then there is a constant $k$ such that for all cores $\psi$ of pp-formulas in $\Phi^*$ we have $\mathsf{tw}(\psi)\le k$ and $\mathsf{tw}(\mathsf{contract}(\psi))+1 \le k$.
With the results of Section~\ref{sect:quantifier-aware-width}
it follows that $\mathsf{qaw}(\psi)\le 2k$ for such cores $\psi$.
Using techniques from~\cite{ChenMengel16-ep-classification-arxiv}, we can express for each $\phi\in \Phi$ and each structure $\mathbf{B}$ the $|\phi(\mathbf{B})|$ as a polynomial in unknowns of the form $|\psi'(\mathbf{B})|$ where crucially the polynomial only depends on $\phi$ and not on $\mathbf{B}$. We can then substitute every term $|\psi'(.)|$ by a $\sharp\mathsf{PP}$-formula of width at most $2k$ with the results of Section~\ref{sect:quantifier-aware-width}. This yields a $\sharp\mathsf{PP}$-formula of width at most $2k$.
For the other direction, assume that there is a constant $k$ such that each ep-formula in $\Phi$ has a $\sharp\mathsf{EP}$-representation $\phi'$ with $\mathsf{width}(\phi)\le k$. Let $\psi$ be a pp-formula from $\Phi^*$. We will show that $\psi$ is equivalent to a pp-formula of quantifier aware width at most $k$ which completes the proof with Section~\ref{sect:quantifier-aware-width}.
We first choose $\phi\in \Phi$ such that $\psi$ appears in $\phi^*$. Let $\phi'$ be the $\sharp\mathsf{EP}$-representation of $\phi$ of width at most $k$. We first construct a linear combination $\ell_1(.)$ as in (\ref{eq:linearcombsketch}). Note that all summands of $\ell_1(.)$ have quantfier aware width at most $k$ as in the proof of Lemma~\ref{lem:basicisenoughsketch}.
We then construct a second representation of $|\phi(\mathbf{B})|$ as a linear combination $\ell_2(.)$. As in the first direction, for every structure we can express
$|\phi(\mathbf{B})|$ as a polynomial in unknowns of the form $|\theta(\mathbf{B})|$. Arithmetic simplifications and elimination of counting equivalent terms gives the second linear combination $\ell_2(.)$ computing $|\phi(\mathbf{B})|$. We then argue that a term of the form $c\cdot |\psi(\mathbf{B})|$ must appear in $\ell_2(.)$.
We have that $\ell_1$ and $\ell_2$ compute $|\phi(\mathbf{B})|$ and are hence equal.
With Theorem~\ref{thm:linearcombination} it follows that $\ell_1$ contains a term that is counting equivalent to $\psi$.
Since all summands of $\ell_1(.)$ have quantifier aware width at most $k$, the claim follows.
\end{proof}
\begin{proof}[of Theorem~\ref{thm:shepcompute}] (sketch)
Using Section~\ref{sect:quantifier-aware-width}
and normalization results (Appendix~\ref{app:normalization}), $\phi$ can be turned into a linear combination as in (\ref{eq:linearcombsketch}). Moreover, starting with any such representation yields the same linear combination up to counting equivalence of the summands by Theorem~\ref{thm:linearcombination}. Now turning this linear combination into a $\sharp\mathsf{PP}$-formula yields a $\sharp\mathsf{PP}$-representation and minimizing the width of the summands with Theorem~\ref{thm:shppcompute} gives a representation of optimal width.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}
The primary goal of tomography is to determine the internal structure of an object without cutting it, namely using data obtained by methods that leave the object under investigation undamaged.
These data can be obtained by exploiting the interaction between the object and various kinds of probes including X-rays, electrons,
and many others. After its interaction with the object under investigation, the probe is detected to produce what we call a projected distribution or tomogram, see Fig. \ref{fig:profiles}.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{profiles.pdf}
\caption{Distribution and two tomograms. }
\label{fig:profiles}
\end{figure}
Tomography is a rapidly evolving field for its broad impact on issues of fundamental nature and for its important applications such as the development of diagnostic tools relevant to disparate fields, such as engineering, biomedical and archaeometry. Moreover, tomography can be a powerful tool for many reconstruction problems coming from many areas of research, such as imaging, quantum information and computation, cryptography, lithography, metrology and many others, see Fig.~\ref{fig:Tomography}.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{Tomography.png}
\caption{Reconstruction problems from diverse fields may be united within the framework of tomography.}
\label{fig:Tomography}
\end{figure}
From the mathematical point of view the reconstruction problem can be formulated as follows: one wants to recover an unknown function through the knowledge of an appropriate family of integral transforms.
It was proved by J. Radon~\cite{Rad1917} that a smooth function $f(x,y)$ on $\mathbb{R}^2$ can be determined explicitly by means of its integrals over the lines in $\mathbb{R}^2$. Let $\mathcal{R}{f}(X,\theta)$ denote the integral of $f$ along the line $x \cos\theta + y \sin\theta = X$ (tomogram). Then
\begin{equation}
\label{eq:Radoninversion}
f(x,y)=(-\Delta)^{\frac{1}{2}}\int_{0}^{2\pi} \mathcal{R}{f}(x \cos\theta + y \sin\theta, \theta)\;
\frac{\mathrm{d}\theta}{4\pi },
\end{equation}
where $\Delta=\frac{\partial^2}{\partial x^2} + \frac{\partial^2}{\partial y^2}$ is the Laplacian on $\mathbb{R}^2$, and its square root is defined by Fourier transform (see Theorem~\ref{thm:inversioneformula}).
We now observe that the formula above has built in a remarkable duality: first one integrates over the set of points in a line, then one
integrates over the set of lines passing through a given point. This formula can be extended to the $N$-dimensional case by computing the integrals of the function $f$ on all possible hyperplanes.
This suggests to consider the transform $f \mapsto \mathcal{R}{f}$ defined as follows. If $f$ is a function on $\mathbb{R}^N$ then $\mathcal{R}{f}$ is the function defined on the space of all possible $(N-1)$-dimensional planes in $\mathbb{R}^N$ such that, given a hyperplane $\lambda$, the value of $\mathcal{R}{f}(\lambda)$ is given by the integral of $f$ along $\lambda$. The function $\mathcal{R}{f}$ is called \emph{Radon transform} of $f$.
There exist several important generalizations of the
Radon transform by
John \cite{John}, Gel'fand \cite{Gelf}, Helgason \cite{Helgason} and Strichartz \cite{Strichartz}.
More recent analysis has been boosted by Margarita and Volodya Man'ko and has focused on symplectic transforms \cite{Mancini95}, on the deep relationship with classical systems and classical dynamics \cite{Olga97,tomogram,m2}, on the formalism of
star product quantization \cite{MarmoPhysScr,MarmoJPA,Marmoopen}, and on the study of
marginals along curves that are not straight lines
\cite{ManMenPhysD,tomocurved}.
In quantum mechanics the Radon transform of the Wigner function
\cite{Wig32,Moyal,Hillary84}
was considered in the tomographic approach to the study
of quantum states \cite{Ber-Ber,Vog-Ris} and experimentally
realized with different particles and in diverse situations.
For a review on the modern mathematical aspects of classical and quantum tomography see~\cite{tomolectures}.
Good reviews on recent tomographic applications can be found in~\cite{Deans} and in~\cite{Jardabook}, where particular emphasis is given on maximum likelihood methods, that enable one to extract the maximum reliable
information from the available data can be found.
As explained above, from the mathematical point of view, the internal structure of the object is described by an unknown function $f$ (density), that is connected via an operator to some measured quantity $g$ (tomograms). The tomographic reconstruction problem can be stated as follows: for given data $g$, the task is to find $f$ from the operator equation $\mathcal{R}{f}=g$.
There are many problems related to the implementation of effective tomographic techniques due to the instability of the reconstruction process. There are two principal reasons of this instability. The first one is the ill-posedness of the reconstruction problem: in order to obtain a satisfactory estimate of the unknown function it is necessary an extremely precise knowledge of its tomograms, which is in general physically unattainable~\cite{tomothick}. The second reason is the discrete and possibly imperfect nature of data that allows to obtain only an approximation of the unknown function.
The first question is whether a partial information still determines the function uniquely. A negative answer is given by a theorem of Smith, Solomon and Wagner~\cite{SmithSolmonWagner}, that states: ``A function $f$ with compact support in the plane is uniquely determined by any infinite set, but by no finite set of its tomograms''. Therefore, it is clear that one has to abandon the request of uniqueness in the applications of tomography. Thus, due to the ill-posedness of reconstruction problem and to the loss of uniqueness in the inversion process, a regularization method has to be introduced to stabilize the inversion.
A powerful approach is the introduction of a Mumford-Shah (MS) functional, first introduced in a different context for image denoising and segmentation~\cite{MS}. The main motivation is that, in many practical applications, one is not only interested in the reconstruction of the density distribution $f$, but also in the extraction of some specific features or patterns of the image. An example is the problem of the determination of the boundaries of inner organs. By minimizing the MS functional, one can find not only (an approximation of) the function but also its sharp contours. Very recently a MS functional for applications to tomography has been introduced in the literature~\cite{RamlauRing1, RamlauRing2, RondiSantosa}. Some preliminary results in this context are already available but there are also many interesting open problems and promising results in this direction, as we will try to explain in the second part of this article.
The article is organized as follows. Section~\ref{Sec:Radon} contains a short introduction to the Radon transform, its dual map and the inversion formula. Section~\ref{Sec:ill} is devoted to a brief discussion on the ill-posedness of the tomographic reconstruction and to the introduction of regularization methods. In Section~\ref{Sec:MS} a MS functional is applied to tomography as a regularization method. In particular, in Subsection~\ref{Subsec:MS} the piecewise constant model and known results are discussed together with a short list of some interesting open problems.
Finally, in Section~\ref{Sec:3Dinterpretation} we present an electrostatic interpretation of the regularization method based on the MS functional,
which motivates us to introduce an improved regularization method, based on the Blake-Zisserman functional~\cite{BZ}, as a relaxed version of the previous one.
\section{The Radon transform: definition and inversion formula}\label{Sec:Radon}
Consider a body in the plane $\mathbb{R}^2$, and consider a beam of particles (neutrons, electrons, X-rays, etc.) emitted by a source. Assume that the initial intensity of the beam is $I_{0}$. When the particles pass through the body they are absorbed or scattered and the intensity of the beam traversing a length $\Delta s$ decreases by an amount proportional to the density of the body $\mu$, namely
\begin{equation}
\Delta I /I=-\Delta s\, \mu(s),
\end{equation}
so that
\begin{equation}
I(s)=I_{0}\exp \left(-\int_{0}^{s}\mu(r)\;\mathrm{d} r \right).
\end{equation}
A detector placed at the exit of the body measures the final intensity $I(s)$ and then from
\begin{equation}
-\ln \frac{I(s)}{I_0} =\int_{0}^{s} \mu(r)\; \mathrm{d} r
\end{equation}
one can record the value of the density $\mu$ integrated on a line.
If another ray with a different direction is considered, with the same procedure one obtains the value of the integral of the density on that line.
The mathematical model of the above setup is the following:
Given a
smooth function $f(x)$ on the plane, $x\in\mathbb{R}^2$,
and a line $\lambda$, consider its tomogram, given by
\begin{equation}\label{sharp lambda}
\mathcal{R}{f}(\lambda)=\int_{\lambda} f(x)\; \mathrm{d} m(x),
\end{equation}
where $\mathrm{d} m$ is the Euclidean measure on the line $\lambda$.
In this way, we have defined an operator $\mathcal{R}$ that maps a smooth function $f$ on the plane $\mathbb{R}^2$ into a function $\mathcal{R}{f}$ on $\PM^{2}$, the manifold of the lines in $\mathbb{R}^2$.
We ask the following question: If we know the family of tomograms $(\mathcal{R}{f}(\lambda))_{\lambda \in \PM^{2}}$, can we reconstruct the density function $f$? The answer is affirmative and in the following we will see how to obtain this result.
Let us generalize the above definitions to the case of an $N$-dimensional space.
Let $f$ be a function defined on $\mathbb{R}^N$, integrable on each hyperplane in $\mathbb{R}^N$ and let $\PM^N$ be the manifold of all hyperplanes in $\mathbb{R}^N$. The Radon transform of $f$ is defined
by Eq.~(\ref{sharp lambda}),
where $\mathrm{d} m$ is the Euclidean measure on the hyperplane $\lambda$.
Thus we have an operator $\mathcal{R}$, the \emph{Radon transform}, that maps a function $f$ on $\mathbb{R}^N$ into a function $\mathcal{R}{f}$ on $\PM^{N}$, namely $f \mapsto \mathcal{R}{f}$. Its dual transform, also called \emph{back projection operator}, $g \mapsto \mathcal{I}{g}$ associates to a function $g$ on $\PM^{N}$ the function $\mathcal{I}{g}$ on $\mathbb{R}^N$ given by
\begin{equation}\label{dual map}
\mathcal{I}{g}(x)=\int_{x \in \lambda} g(\lambda) \; \mathrm{d}\mu(\lambda),
\end{equation}
where $\mathrm{d}\mu$ is the unique probability measure on the compact set $\{\lambda \in \PM^{N}|x \in \lambda\}$ which is invariant under the group of rotations around $x$.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{radon.pdf}
\caption{Parametrization of the hyperplane $\lambda$ using its signed distance $X$ from the origin and a unit vector $\xi$ perpendicular to $\lambda$.}
\label{fig:radon nd}
\end{figure}
Let us consider the following covering of $\PM^N$
\begin{equation}\label{double covering}
\mathbb{R} \times \mathbb{S}^{N-1} \to \PM^N, \qquad (X,\xi)\mapsto \lambda,
\end{equation}
where $\mathbb{S}^{N-1}$ is the unit sphere in $\mathbb{R}^N$. Thus, the
equation of the hyperplane $\lambda$ is
\begin{equation}
\lambda=\{x\in\mathbb{R}^N \, | \, X-\xi \cdot x=0\},
\end{equation}
with $a\cdot b$ denoting the Euclidean inner product of $a, b\in\mathbb{R}^N$. See Fig. \ref{fig:radon nd}.
Observe that the pairs $(X,\xi),(-X,-\xi) \in \mathbb{R} \times \mathbb{S}^{N-1}$ are mapped into the same hyperplane $\lambda \in \PM^N$. Therefore (\ref{double covering}) is a double covering of $\PM^N$.
Thus $\PM^N$ has a canonical manifold structure with respect to which this covering mapping is differentiable. We identify continuous (differentiable) functions $g$ on $\PM^N$ with continuous (differentiable) functions $g$ on $\mathbb{R} \times \mathbb{S}^{N-1}$ satisfying $g(X,\xi)=g(-X,-\xi)$.
We will momentarily work in the Schwartz space $\mathcal{S}(\mathbb{R}^N)$ of complex-valued rapidly decreasing functions on $\mathbb{R}^N$.
In analogy with $\mathcal{S}(\mathbb{R}^N)$ we define $\mathcal{S}(\mathbb{R} \times \mathbb{S}^{N-1})$ as the space of $C^{\infty}$ functions $g$ on $\mathbb{R} \times \mathbb{S}^{N-1}$ which for any integers $m\geq 0$, any multiindex $\alpha \in \mathbb{N}^N$, and any differential operator $D$ on $\mathbb{S}^{N-1}$ satisfy
\begin{equation}
\sup_{X \in \mathbb{R} , \xi \in \mathbb{S}^{N-1}}\left| |X|^m \frac{\partial^\alpha(Dg)}{\partial X^\alpha}(X,\xi)\right| < +\infty.
\end{equation}
The space $\mathcal{S}(\PM^N)$ is then defined as the set of $g \in \mathcal{S}(\mathbb{R} \times \mathbb{S}^{N-1})$ satisfying $g(-X,-\xi)=g(X,\xi)$.
Now we want to obtain an inversion formula, namely we want to prove that one can recover a function $f$ on $\mathbb{R}^N$ from the knowledge of its Radon transform.
In order to get this result we need a preliminary lemma, whose proof can be found in~\cite{tomolectures}, which suggests an interesting physical interpretation.
\begin{lemma}\label{sharp flat}
Let $f \in \mathcal{S}(\mathbb{R}^N)$ and $V(x)=1/|x|$, $x \in \mathbb{R}^N$, $x\neq0$ . Then
\begin{equation}
\mathcal{I}(\mathcal{R}{f})(x)=a_{N}\, f\ast V,
\label{eq:electid}
\end{equation}
where $a_N$ depends only on the dimension $N$, and
$*$ denotes the convolution product,
\begin{equation}
(f\ast g)(x) =\int_{\mathbb{R}^N} f(y)\, g(x-y) \; \mathrm{d} y .
\end{equation}
\end{lemma}
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{potential.pdf}
\caption{$\mathcal{I}(\mathcal{R}{f})(x)$ is the potential at $x$ generated by the charge distribution $f$.}
\label{fig:potential}
\end{figure}
A physical interpretation of Lemma~\ref{sharp flat} is the following: if $f$ is a charge distribution, then the potential at the point $x$ generated by that charge is exactly $\mathcal{I}(\mathcal{R}{f}(x))$, see Fig.~\ref{fig:potential}. Notice, however, that the potential of a point charge scales always as the inverse distance \emph{independently} of the dimension $N$, and thus it is Coulomb only for $N=3$. The only dependence on $N$ is in the strength of the elementary charge $a_N$. This fact is crucial: indeed, the associated Poisson equation involves an $N$-dependent (fractional) power of the Laplacian, which appears in the inversion formula for the Radon transform.
\begin{theorem}\label{thm:inversioneformula}
Let $f \in \mathcal{S}(\mathbb{R}^N)$. Then
\begin{equation}\label{inversion formula0}
f(x)= \frac{1}{2(2\pi)^{N-1}}(-\Delta)^{\frac{N-1}{2}} \mathcal{I}(\mathcal{R}{f})(x)
\end{equation}
where $(-\Delta)^\alpha$, with $\alpha>0$, is a pseudodifferential operator whose action is
\begin{equation}
((-\Delta)^\alpha f) (x)= \int_{\mathbb{R}^N} |k|^{2\alpha}\, \hat{f}(k)\, \mathrm{e}^{\mathrm{i} k \cdot x}\; \frac{\mathrm{d} k}{(2\pi)^N},
\label{eq:fractionalLap}
\end{equation}
where $\hat{f}$ is the Fourier transform of $f$,
\begin{equation}
\hat{f}(k) = \int_{\mathbb{R}^N} f(x)\, \mathrm{e}^{-\mathrm{i} k \cdot x}\; \mathrm{d} x.
\label{eq:Fouriertrans}
\end{equation}
\end{theorem}
The proof of Theorem~\ref{thm:inversioneformula} can be found in~\cite{tomolectures}.
Equation~(\ref{inversion formula0}) says that, modulo the final action of $(-\Delta)^{\frac{N-1}{2}}$, the function $f$ can be recovered from its Radon transform $\mathcal{R}{f}$ by the application of the dual mapping $\mathcal{I}$: first one integrates over the set of points in a hyperplane and then one integrates over the set of hyperplanes passing through a given point. Explicitly we get
\begin{equation}\label{inversion formula}
f(x) =\frac{1}{2(2\pi)^{N-1}}(-\Delta)^{\frac{N-1}{2}}\int_{\mathbb{S}^{N-1}} \mathcal{R}{f}(\xi \cdot x, \xi)\; \mathrm{d}\xi,
\end{equation}
which has the following remarkable interpretation. Note that if one fixes a direction $\xi \in \mathbb{S}^{N-1}$, then the function $\mathcal{R}{f}(\xi \cdot x,\xi)$ is constant on each plane perpendicular to $\xi$, i.e.\ it is a (generalized) plane wave. Therefore, Eq.~(\ref{inversion formula}) gives a representation of $f$ in terms of a continuous superposition of plane waves. A well-known analogous decomposition is given by Fourier transform. When $N=2$, one recovers the inversion formula~(\ref{eq:Radoninversion}) originally found by Radon \cite{Rad1917}.
\section{Instability of the inversion formula with noisy data}\label{Sec:ill}
We have defined the Radon transform of any function $f \in \mathcal{S}(\mathbb{R}^N)$ as $\mathcal{R}{f}$. The following theorem~\cite{Helgason}
contains the characterization of the range of the Radon linear operator $\mathcal{R}$ and the extension of $\mathcal{R}$ to the space of square integrable functions $L^2(\mathbb{R}^N)$.
\begin{theorem}
\label{th:Radonbijection}
The Radon transform $\mathcal{R}$ is a linear one-to-one mapping of $\mathcal{S}(\mathbb{R}^N)$ onto $\mathcal{S}_H(\PM^N)$, where the space $\mathcal{S}_H(\PM^N)$ is defined as follows: $g \in \mathcal{S}_H(\PM^N)$ if and only if $g \in \mathcal{S}(\PM^N)$ and for any integer $k \in \mathbb{N}$ the integral
\begin{equation}
\int_{\mathbb{R}} g(X,\xi)\, X^k \; \mathrm{d} X
\end{equation}
is a homogeneous polynomial of degree $k$ in $\xi_1, \dots, \xi_N$. Moreover, the Radon operator $\mathcal{R}$ can be extended to a continuous operator from $L^2(\mathbb{R}^N)$ and $L^2(\PM^{N})$.
\end{theorem}
In medical imaging, computerized tomography is a widely used technique
for the determination of the density $f$ of a sample from measurements
of the attenuation of X-ray beams sent through the material along
different angles and offsets. The measured data $g$ are connected to the density
$f$ via the Radon Transform $\mathcal{R}{}$.
To compute the density distribution $f$ the equation $g = \mathcal{R}{f}$
has to be inverted. Unfortunately it is a well known fact that $\mathcal{R}$ is not continuously invertible on $L^2(\PM^{N})$~\cite{Helgason}, and this imply that the problem of inversion is ill-posed.
For this reason, regularization methods have to be introduced to stabilize the inversion in the presence of data noise.
We discuss ill-posed problems only in the framework of linear problems in
Hilbert spaces~\cite{Natterer}. Let $\mathscr{H}, \mathscr{K}$ be Hilbert spaces and let $A$ be a linear bounded operator
from $\mathscr{H}$ into $\mathscr{K}$. The problem
\begin{equation}\label{prob:inverseproblem}
\textrm{given $g\in \mathscr{K}$, find $f \in \mathscr{H}$ such that $Af=g$}
\end{equation}
is called well-posed by Hadamard (1932) if it is uniquely solvable for each $g \in \mathscr{K}$
and if the solution depends continuously on $g$. Otherwise, (\ref{prob:inverseproblem}) is called ill-posed.
This means that for an ill-posed problem the operator $A^{-1}$ either does not exist,
or is not defined on all of $\mathscr{K}$, or is not continuous. The practical difficulty with an
ill-posed problem is that even if it is solvable, the solution of $Af = g$ need not be
close to the solution of $Af = g^\epsilon$ if $g^\epsilon$ is close to $g$.
In general $A^{-1}$ is not a continuous operator. To restore continuity we introduce the notion of a regularization of $A^{-1}$. This is a family $(T_\gamma )_{\gamma >0}$ of linear continuous operators $T_\gamma: \mathscr{K} \to \mathscr{H}$ which are defined on all $\mathscr{K}$ and for which
\begin{equation}
\lim_{\gamma \to 0} \; \;T_\gamma g= A^{-1} g
\end{equation}
on the domain of $A^{-1}$. Obviously $\|T_\gamma\| \to + \infty$ as $\gamma \to 0$ if $A^{-1}$ is not bounded. With the help of a regularization we can solve (\ref{prob:inverseproblem}) approximately in the following sense. Let $g^\epsilon \in \mathscr{K}$ be an approximation to $g$ such that $\|g -g^\epsilon\| \leq \epsilon$ . Let $\gamma(\epsilon)$ be such that, as $\epsilon \to 0$,
\begin{equation}
\gamma(\epsilon) \to 0,
\qquad \|T_{\gamma(\epsilon)}\| \epsilon \to 0.
\end{equation}
Then, as $\epsilon \to 0$,
\begin{eqnarray}
\|T_{\gamma(\epsilon)} g ^\epsilon -A^{-1} g\| & \leq & \|T_{\gamma(\epsilon)} (g ^\epsilon-g) \|
\nonumber\\ & &
+ \| (T_{\gamma(\epsilon)}-A^{-1}) g\| \nonumber \\
& \leq & \|T_{\gamma(\epsilon)}\| \epsilon + \| (T_{\gamma(\epsilon)}-A^{-1} )g\| \to 0. \nonumber \\
\end{eqnarray}
Hence, $T_{\gamma(\epsilon)} g^\epsilon$ is close to $A^{-1} g$ if $g^\epsilon$ is close to $g$. The number $\gamma$ is called a \emph{regularization parameter}. Determining a good
regularization parameter is one of the crucial points in the application of regularization methods.
There are several methods for constructing a regularization as the truncated singular value decomposition, the method of Tikhonov-Phillips or some iterative methods~\cite{Natterer}. In the following section we present a regularization method based on the minimization of a Mumford-Shah type functional.
\section{Mumford-Shah functional for the simultaneous segmentation and reconstruction of a function}\label{Sec:MS}
In many practical applications one is not only interested in the reconstruction
of the density distribution $f$ but also in the extraction of some specific
features within the image which represents the density distribution of the
sample. For example, the planning of surgery might require the determination
of the boundaries of inner organs like liver or lung or the separation
of cancerous and healthy tissue.
Segmenting a digital image means finding its \emph{homogeneous regions} and its \emph{edges}, or \emph{boundaries}. Of course, the homogeneous regions are supposed to correspond to meaningful parts of objects in the real world, and the edges to their apparent contours. The Mumford-Shah variational model is one of the principal models of image segmentation. It defines the segmentation problem as a joint smoothing/edge detection problem: given an image $g(x)$, one seeks simultaneously a ``piecewise smoothed image'' $u(x)$ with a set $\Gamma$ of abrupt discontinuities, the ``edges'' of $g$.
The original Mumford-Shah functional~\cite{MS}, is the following:
\begin{eqnarray}\label{defn:JMS}
MS(\Gamma,u)&=&\|u-g\|_{L^2(D)}^2+\alpha \int_{D \setminus \Gamma} |\nabla u(x)|^2\; \textrm{d}x \nonumber \\
&& + \beta \mathcal{H}^{N-1}(D \cap \Gamma),
\end{eqnarray}
where
\begin{itemize}
\item $D \subset \mathbb{R}^N$ is an open set (\emph{screen});
\item $\Gamma \subset \mathbb{R}^N$ is a closed set (\emph{set of edges});
\item $u: D \to \mathbb{R}$ (\emph{cartoon});
\item $\nabla u$ denotes the distributional gradient of $u$;
\item $g \in L^2(D)$ is the datum (\emph{digital image});
\item $\alpha, \beta >0$ are parameters (\emph{tuning parameters});
\item $\mathcal{H}^{N-1}$ denotes the $(N-1)$-dimensional Hausdorff measure.
\end{itemize}
The squared $L^2$ distance in (\ref{defn:JMS}) plays the role of a fidelity term: it imposes that the cartoon $u$ approximate the image $g$. The second term in the functional imposes that the cartoon $u$ be piecewise smooth outside the edge set $\Gamma$. In other word this term favors sharp contours rather than zones where a thin layer of gray is used to pass smoothly from white to black or viceversa. Finally the third term in the functional imposes that the contour $\Gamma$ be ``small'' and as smooth as possible. What is expected from the minimization of this functional is a sketchy, cartoon-like version of the given image together with its contours. See Fig.~\ref{fig:cartooneye}.
The minimization of the $MS$ functional represents a compromise between accuracy and segmentation. The compromise depends on the tuning parameters $\alpha$ and $\beta$
which have different roles. The parameter $\alpha$ determines how much the cartoon $u$ can vary, if $\alpha$ is small some variations of $u$ are allowed, while as $\alpha$ increases $u$ tends to be a piecewise constant function. The parameter $\beta$ represents a scale parameter of the functional and measure the amount of contours: if $\beta$ is small, a lot of edges are allowed and we get a fine segmentation. As $\beta$ increases, the segmentation gets coarser.
For more details on the model see the original paper~\cite{MS}, and the book~\cite{JMMS}.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{image_edge_cartoon}
\caption{Left: image of an eye ($g$). Center: contours of the image in the Mumford-Shah model (edges $\Gamma$). Right: piecewise smooth function approximating the image (cartoon $u$) \cite{Yuille}.}
\label{fig:cartooneye}
\end{figure}
The minimization of the $MS$ functional in (\ref{defn:JMS}) is performed among the admissible pairs $(\Gamma,u)$ such that $\Gamma$ is closed and $u \in C^1(D \setminus \Gamma)$. It is worth noticing that in this model there are two unknowns: a scalar function $u$ and the set $\Gamma$ of its discontinuities. For this reason this category of problems is often called ``Free Discontinuities Problem''. Existence of minimizers of the $MS$ functional in~(\ref{defn:JMS}) was proven by De Giorgi, Carriero, Leaci in~\cite{DGCL} in the framework of bounded variation functions without Cantor part (space SBV) introduced by Ambrosio and De Giorgi in~\cite{DGA}. Further regularity properties for optimal segmentation in the Mumford-Shah model were shown in~\cite{DMMS,AmbFusPal,LMS, BD}.
Here we present a variation of the MS functional, adapted to the inversion problem of the Radon transform. More precisely, we consider a regularization method that quantifies the edge sets together with images, i.e. a procedure that gives simultaneously a reconstruction and a segmentation of $f$ (assumed to be supported in $D\subset\mathbb{R}^N$) directly from the measured tomograms $g$, based on the minimization of the Mumford-Shah type functional
\begin{eqnarray}\label{def:MSfunctional}
J_{MS}(\Gamma,f)&=&\|\mathcal{R}{f}-g\|_{L^2(\PM^N)}^2 \nonumber \\
& &+\alpha \int_{D \setminus \Gamma} |\nabla f(x)|^2\; \textrm{d}x
+ \beta \mathcal{H}^{N-1}(\Gamma).
\end{eqnarray}
The only difference between the functionals $MS$ and $J_{MS}$ is the first term, i.e. the fidelity term, that ensures that the reconstruction for $f$ is close enough to a solution of the equation $\mathcal{R}{f} = g$, whereas the other terms play exactly the same role explained for the functional $MS$. As explained above, in addition to the reconstruction of the density $f$, we are interested in the reconstruction of its singularity set $\Gamma$,
i.e. the set of points where the solution $f$ is discontinuous. The main difference with respect to the standard Mumford-Shah functional~(\ref{defn:JMS}) is that we have to translate the information about the set of sharp discontinuities of $g$ (and hence on the space of the Radon transform) into information about the strong discontinuities of $f$.
\subsection{The piecewise constant model}\label{Subsec:MS}
Here we will review the results obtained by Ramlau and Ring~\cite{RamlauRing1} concerning the minimization of (\ref{def:MSfunctional}) restricted to piecewise constant functions $f$, and then consider some interesting open problems.
For medical applications, it is often a good approximation to
restrict the reconstruction to densities $f$ that are constant with respect to a
partition of the body, as the tissues of inner organs, bones, or muscles have
approximately constant density.
We introduce the space $PC_m(D)$ as the space of piecewise constant functions that attain at most $m$ different function values, where $D$ is an open and bounded subset of $\mathbb{R}^N$. In other words, each $f \in PC_m(D)$ is a linear combination of $m$ characteristic
functions $\chi_{\Omega_k}$ of sets $(\Omega_k)_{k=1,\dots,m}$ which satisfy
$$
\sum_{k=1}^m\chi_{\Omega_k}=\chi_{D} \qquad \mathrm{a.e.}
$$
We assume that the $\Omega_k$'s are open relatively to $D$ and we set $\Gamma_k=\partial \Omega_k$ for the boundary of $\Omega_k$ with respect to the topology relative to the open domain $D$.
In this situation the edge set will be given by the union of the boundaries of $\Omega_k$'s. For technical reasons it is necessary to assume a \emph{nondegeneracy} condition on the admissible partitions of $D$:
\begin{equation}\label{nondegeneracy-cond}
\bm{\Omega}=(\Omega_k)_{k=1,\dots,m}\quad \textrm{is admissible if}\quad \mathcal{L}^N(\Omega_k)\geq \delta,
\end{equation}
for some $\delta >0$, for all $k=1,\dots,m$, where $\mathcal{L}^{N}$ denotes the Lebesgue measure on $\mathbb{R}^N$.
It turns out to be convenient to split the information encoded in a typical function,
\begin{equation}
f=\sum_{k=1}^m f_k\chi_{\Omega_k} \in PC_m(D),
\end{equation}
into a ``geometrical'' part described by the $m$-tuple of pairwise disjoint
sets $\bm{\Omega}=(\Omega_k)_{k=1,\dots,m}$ which cover $D$ up to a set of measure zero and a ``functional'' part given by the $m$-tuple of values $\bm{f}=(f_k)_{k=1,\dots,m}$. We also use the notation $\bm{\Gamma}=(\Gamma_k)_{k=1,\dots,m}$, for the boundaries $\Gamma_k= \partial \Omega_k$ of $\Omega_k$.
As usual when dealing with inverse problems, we have to assume that the data $g$ are not exactly known, but that we are only given noisy measured tomograms $g^\epsilon$ of a (hypothetical) exact data set $g$ with $\|g^\epsilon-g\|_{L^2(\PM^N)} \leq \epsilon$.
If we restrict the functional (\ref{def:MSfunctional}) to functions in $PC_m(D)$ we obtain that the second term (involving the derivatives of $f$) disappears, therefore it remains to minimize the functional
\begin{equation}\label{functional:Jbeta}
J_\beta(\bm{\Omega},\bm{f})=\|\mathcal{R}{f}-g^\epsilon\|^2_{L^2(\PM^N)}+\beta \sum_{k=1}^m \mathcal{H}^{N-1}(\Gamma_k),
\end{equation}
over $PC_m(D)$, with respect to the functional variable $\bm{f}$ (a vector of $m$ components) and the geometric variable $\bm{\Omega}$ (a partition of the domain $D$ with at most $m$ distinct regions satisfying the non degeneracy condition (\ref{nondegeneracy-cond})).
So the problem is to find $\tilde{f} \in PC_m(D)$ such that
\begin{equation}\label{eqn:minimizerJbeta}
\tilde{f}=\sum_{k=1}^m \tilde{f}_k\chi_{\tilde{\Omega}_k} \in PC_m(D),
\end{equation}
where
\begin{equation}\label{argmin:Jbeta}
(\tilde{\bm{\Omega}},\tilde{\bm{f}}) = \textrm{arg} \min_{(\bm{\Omega} , \bm{f})}\; J_\beta(\bm{\Gamma},\bm{f}).
\end{equation}
It is clear that $\tilde{f}$ will depend on the regularization parameter $\beta$ and on the error level $\epsilon$.
Now we can state the results concerning the functional $J_\beta$ in (\ref{functional:Jbeta}). There are several technical details necessary for the precise statement and proof of the theorems, for which we refer to the original paper~\cite{RamlauRing1}.
Here we will give a simplified version of the theorems with the purpose of explain the main goal, without too many technical details. The first result is about the existence of minimizers of the functional $J_\beta$ in (\ref{functional:Jbeta}).
\begin{theorem}
For all $g^\epsilon \in L^2(\PM^N)$ there exists a minimizer $(\tilde{\bm{\Omega}}^\epsilon_\beta,\tilde{\bm{f}}_\beta^\epsilon)$ of the functional $J_\beta$ in (\ref{functional:Jbeta}), with $\beta>0$.
\end{theorem}
The second result regards the stable dependence of the minimizers of the functional $J_\beta$ in (\ref{functional:Jbeta}) on the error level $\epsilon$.
\begin{theorem}
Let $(g^{\epsilon_n})_{n \in \mathbb{N}}$ be a sequence of functions in $L^2(\PM^N)$ and let $g^\epsilon \in L^2(\PM^N)$. For all $n \in \mathbb{N}$, let $(\tilde{\bm{\Omega}}^{\epsilon_n}_\beta,\tilde{\bm{f}}_\beta^{\epsilon_n})$ denote the minimizers of the functional $J_\beta$ with initial data $g^{\epsilon_n}$. If $g^{\epsilon_n} \to g^\epsilon$ in $L^2(\PM^N)$, as $n \to +\infty$, then there exists a subsequence of $(\tilde{\bm{\Omega}}^{\epsilon_n}_\beta,\tilde{\bm{f}}_\beta^{\epsilon_n})$ such that
$$
(\tilde{\bm{\Omega}}^{\epsilon_{n_j}}_\beta,\tilde{\bm{f}}_\beta^{\epsilon_{n_j}}) \to (\tilde{\bm{\Omega}}^\epsilon_\beta,\tilde{\bm{f}}_\beta^\epsilon)
$$
as $j \to + \infty$, and $(\tilde{\bm{\Omega}}^\epsilon_\beta,\tilde{\bm{f}}_\beta^\epsilon) $ is a minimizer of $J_\beta$ with initial data $g^\epsilon$. Moreover, the limit of each convergent subsequence of $(\tilde{\bm{\Omega}}^{\epsilon_n}_\beta,\tilde{\bm{f}}_\beta^{\epsilon_n})$ is a minimizer of $J_\beta$ with initial data $g^\epsilon$.
\end{theorem}
Finally the last Theorem is a regularization result.
\begin{theorem}
Let $f^\ast \in PC_m(D)$ be given,
$$
f^\ast=\sum_{k=1}^m f_k^\ast \chi_{\Omega_k^\ast},
$$
and let $g^\ast=\mathcal{R}{f}^\ast$. Assume we have noisy data $g^\epsilon \in L^2(\PM^N)$ with $\|g^\epsilon-g^\ast\|_{L^2(\PM^N)}\leq \epsilon$. Let us choose the parameter $\beta=\beta(\epsilon)$ satisfying the conditions $\beta(\epsilon) \to 0$ and $\epsilon^2/ \beta(\epsilon) \to 0$ as $\epsilon \to 0$. For any sequence $\epsilon_n \to 0$, let
$(\tilde{\bm{\Omega}}^{n},\tilde{\bm{f}}^{n})$
denote the minimizers of the functional $J_{\beta(\epsilon_n)}$ with initial data $g^{\epsilon_n}$ and regularization parameter $\beta=\beta(\epsilon_n)$. Then there exists a convergent subsequence of
$(\tilde{\bm{\Omega}}^{n},\tilde{\bm{f}}^{n})$. Moreover, for every convergent subsequence with limit $(\tilde{\bm{\Omega}},\tilde{\bm{f}})$ the function
$$
\tilde{f}=\sum_{k=1}^m \tilde{f}_k \chi_{\tilde{\Omega}_k} \in PC_m(D)
$$
is a solution of the equation $\mathcal{R}{f}=g^\ast$ with a minimal perimeter. Moreover if $f^\ast$ is the unique solution of this equation then the whole sequence converges
$$
\tilde{f}^{n}\to f^\ast, \quad \textrm{in $L^2(D)$},
$$
when $n\to+\infty$.
\end{theorem}
Finally, let us list some open problems in this context:
\begin{itemize}
\item Is the nondegeneracy condition (\ref{nondegeneracy-cond}) necessary?
\item Can one find an a priori optimal value for the number $m$ of different values?
\item Is it possible to give an a priori estimate on the $L^{\infty}$-norm of the solution (maximum principle)?
\item And finally, it would be very important for applications to prove the existence of minimizers of the functional $J_{MS}$ not restricted to piecewise constant functions $f$.
\end{itemize}
We observe that all these problems are quite natural, and have been completely solved in the case of the standard Mumford-Shah functional $MS$ in (\ref{defn:JMS}), see e.g.~\cite{JMMS,Fusco}.
\section{Electrostatic interpretation of $J_{MS}$}\label{Sec:3Dinterpretation}
In this section we restrict our attention to the $3$-dimensional case. We propose an electrostatic interpretation of the regularization method based on the functional $J_{MS}$ discussed in the previous section. The intent is to give a physical explanation of the fidelity term $\|\mathcal{R}{f}-g\|^2_{L^2(\PM^3)}$ in the functional (\ref{def:MSfunctional}), that provide the intuition for an improved regularization method.
For $N=3$, the inversion formula~(\ref{inversion formula0}) and the electrostatic identity~(\ref{eq:electid}) particularize, respectively, as follows: for all $f \in \mathcal{S}(\mathbb{R}^3)$ one gets
\begin{equation}\label{inversionformula_again}
f=-\frac{1}{2(2\pi)^2}\Delta\mathcal{I}(\mathcal{R}{f})
\end{equation}
and
\begin{equation}
\mathcal{I}(\mathcal{R}{f})=a_3 f \ast V,
\end{equation}
where $V(x)=1/|x|$ and $a_3$ is a constant.
We present two preliminary Lemmas.
\begin{lemma}\label{lemma:normRf}
For all real valued $f \in \mathcal{S}(\mathbb{R}^3)$ one has
\begin{equation}
\|\mathcal{R}{f}\|^2=\frac{1}{2(2\pi)^2}\|\nabla\mathcal{I}(\mathcal{R}{f})\|^2.
\end{equation}
\end{lemma}
\begin{proof}
We know that $f=-\frac{1}{2(2\pi)^2} \Delta \mathcal{I}(\mathcal{R}{f})$, therefore
\begin{eqnarray}
&& \int_{\mathbb{R} \times \mathbb{S}^2}| \mathcal{R} f(X , \xi)|^2 \; \textrm{d}X\; \textrm{d}\xi
=\int_{\mathbb{R}^3} f(x)\, \mathcal{I}(\mathcal{R}{f}) (x)\; \textrm{d}x \nonumber \\
&&= \int_{\mathbb{R}^3} \frac{1}{2(2\pi)^2}[-\Delta \mathcal{I}(\mathcal{R}{f})] (x)\, \mathcal{I}(\mathcal{R}{f}) (x) \; \textrm{d}x \nonumber \\
&&=\frac{1}{2(2\pi)^2} \int_{\mathbb{R}^3}| \nabla \mathcal{I}(\mathcal{R}{f}) (x)|^2\; \textrm{d}x .
\nonumber
\end{eqnarray}
\end{proof}
\begin{lemma}\label{lemma:electricfield}
For all real valued $f \in \mathcal{S}(\mathbb{R}^3)$
define
\begin{equation}
E=- \frac{1}{2(2\pi)^2} \nabla \mathcal{I}(\mathcal{R}{f})
\end{equation}
and
\begin{equation}
\varphi= \frac{1}{2(2\pi)^2} \mathcal{I}(\mathcal{R}{f}).
\end{equation}
Then
\begin{equation}
f= \nabla \cdot E =- \Delta \varphi.
\label{ephif}
\end{equation}
\end{lemma}
\begin{proof}
\begin{eqnarray*}
\nabla \cdot E &=& - \frac{1}{2(2\pi)^2} \nabla \cdot \nabla \mathcal{I}(\mathcal{R}{f})
= -\Delta \varphi \nonumber \\
&=& \frac{1}{2(2\pi)^2}(-\Delta) \mathcal{I}(\mathcal{R}{f})
= f,
\end{eqnarray*}
where we used the inversion formula~(\ref{inversionformula_again}).
\end{proof}
Now we consider a measured tomogram $g : \PM^3 \to \mathbb{R}$ and let us assume that $g=\mathcal{R}{f_0}$ for some $f_0 : \mathbb{R}^3 \to \mathbb{R}$. By Lemma \ref{lemma:normRf}-\ref{lemma:electricfield} it follows immediately that the fidelity term $\|\mathcal{R}{f}-g\|_{L^2(\PM^3)}^2$ can be rewritten as follows:
\begin{eqnarray}
\|\mathcal{R}{f}-g\|_{L^2(\PM^3)}^2 &= & 2(2\pi)^2 \|E-E_g\|^2_{L^2(\mathbb{R}^3)} \nonumber \\
& =& 2(2\pi)^2\|\nabla \varphi - \nabla \varphi_g\|^2_{L^2(\mathbb{R}^3)},
\end{eqnarray}
where
\begin{equation}
E=- \frac{1}{2(2\pi)^2} \nabla \mathcal{I}(\mathcal{R}{f}), \quad
E_g=- \frac{1}{2(2\pi)^2} \nabla \mathcal{I}(g)
\end{equation}
are the corresponding electric fields, while
\begin{equation}
\varphi= \frac{1}{2(2\pi)^2} \mathcal{I}(\mathcal{R}{f}), \quad
\varphi_g= \frac{1}{2(2\pi)^2} \mathcal{I}(g)
\label{eq:potentials}
\end{equation}
are the corresponding potentials.
With respect to the standard Mumford-Shah functional $MS$ in (\ref{defn:JMS}), the new fidelity term in the functional $J_{MS}$ in (\ref{def:MSfunctional}) controls the distance between the Radon transform of $f$ and the tomographic data $g$. The relevant difference with respect to the original functional is that the function $f$ and its Radon transform $\mathcal{R}{f}$ are defined in different spaces.
Let us try to interpret the fidelity term $\|\mathcal{R}{f}-g\|_{L^2(\PM^3)}^2 $ from a physical point of view. A key ingredient for this goal is the electrostatics formulation of the Radon transform. This formulation can be summarized as follows: if we consider, in dimension $3$, a function $f$,
we can think at it as a charge distribution density; if we apply to $f$ first the Radon operator $\mathcal{R}$ and then its adjoint $\mathcal{I}$ we obtain, up to a constant, the electrostatic potential generated by the charge distribution $f$.
This formulation can be stated in any dimension $N$: the difference with general potential theory in dimension $N$ is that, in tomography, the potential produced by a point charge always scales like $1/|x|$, which is the case of electrostatic potential only in dimension $3$. From the electrostatic formulation of the Radon transform we can prove that the fidelity term in the functional $J_{MS}$ actually imposes that the electric field produced by the charge distribution $f$ must be close to the ``measured electric field''. Therefore we conclude that the term $ \|\mathcal{R}{f}-g\|_{L^2(\PM^3)}^2$ is a fidelity term in this weaker sense.
Using this property based on the electrostatic interpretation of the tomographic reconstruction, we can try to minimize some appropriate functionals in the new variables $E$ (electric field) or $\varphi$ (electric potential) and then compute the corresponding $f$ (charge density).
We manipulate the functional $J_{MS}$ as follows:
\begin{eqnarray}\label{eqnaeeay_manipulation}
&& J_{MS}(\Gamma,f)\nonumber \\
&=& \|\mathcal{R}{f}-g\|^2_{L^2(\PM^3)}+\alpha \int_{\mathbb{R}^3 \setminus \Gamma} |\nabla f(x)|^2 \; \textrm{d}x + \beta \mathcal{H}^2(\Gamma) \nonumber \\
&=&2 (2\pi)^2 \|E-E_g\|^2+\alpha \int_{\mathbb{R}^3 \setminus \Gamma} |\nabla (\nabla \cdot E)(x)|^2 \; \textrm{d}x \nonumber \\
&& + \beta \mathcal{H}^2(\Gamma) \nonumber \\
&=&2 (2\pi)^2 \|E-E_g\|^2+\alpha \int_{\mathbb{R}^3 \setminus \Gamma} | \Delta E|^2 \; \textrm{d}x + \beta \mathcal{H}^2(\Gamma) \nonumber \\
&=& F(\Gamma,E)
\end{eqnarray}
where $F$ is a new functional depending on a vector function $E$ and on a set $\Gamma$, and we used the fact that $\nabla \wedge E =0$, since $E$ is conservative.
We observe that the functional
\begin{eqnarray}
F(\Gamma,E) &=& 2 (2\pi)^2 \|E-E_{g}\|^2+\alpha \int_{\mathbb{R}^3 \setminus \Gamma} |\Delta E|^2 \; \textrm{d}x
\nonumber\\
& & + \beta \mathcal{H}^2(\Gamma)
\end{eqnarray}
is a second order functional for a vector field $E$ in which appears the measure of the set $\Gamma$ that is the set of discontinuities of $f$ and thus is the set of discontinuities of $\nabla \cdot E$.
In the functional $F$ we recognize some similarities with a famous second-order free-discontinuity problem: the Blake-Zisserman model. This model is based on the minimization of the Blake-Zisserman functional
\begin{eqnarray}\label{defn:JBZ}
BZ(\Gamma_0,\Gamma_1,v)&=&\|v-v_0\|_{L^2(D)}^2 \nonumber \\
&& + \alpha \int_{D \setminus (\Gamma_0 \cup \Gamma_1)} |\Delta v(x)|^2\; \textrm{d}x \nonumber \\
&&+ \beta \;\mathcal{H}^{N-1}(\Gamma_0 \cap D)\nonumber \\
&& + \gamma\; \mathcal{H}^{N-1}((\Gamma_1 \setminus \Gamma_0) \cap D),
\end{eqnarray}
among admissible triplets $(\Gamma_0,\Gamma_1,v)$, where
\begin{itemize}
\item $D \subset \mathbb{R}^N$ is an open set;
\item $\Gamma_0, \Gamma_1 \subset \mathbb{R}^N$ are closed sets;
\item $\Gamma_0$ is the set of discontinuities of $v$ (jump set), and $\Gamma_1$ the the set of discontinuities of $\nabla v$ (crease set);
\item $v: D \to \mathbb{R}$, $v \in C^2(D \setminus (\Gamma_0 \cup \Gamma_1)) \cap C(D \setminus \Gamma_0)$ is a scalar function;
\item $\Delta v$ denotes the distributional Laplacian of $v$;
\item $v_0 \in L^2(D)$ is the datum (grey intensity levels of the given image);
\item $\alpha, \beta, \gamma >0$ are parameters;
\item $\mathcal{H}^{N-1}$ denotes the $(N-1)$-dimensional Hausdorff measure.
\end{itemize}
The Blake-Zisserman functional allows a more precise segmentation than the Mumford-Shah functional in the sense that also the curvature of the edges of the original picture is approximated. On the other hand, minimizers may not always exist, depending on the values of the parameters $\beta, \gamma$ and on the summability assumption on $v_0$.
We refer to \cite{BZ, CLT1, CLT2, CLT3, CLT4, CLT5} for motivation and analysis of variational approach to image segmentation and digital image processing. In particular see \cite{CLT1, CLT2, Coscia} for existence of minimizer results and \cite{CLT3} for a counterexample to existence and \cite{CLT6, CLT7} for results concerning the regularity of minimizers.
Equation~(\ref{eqnaeeay_manipulation}) implies that the functional $J_{MS}$ can be rewritten in terms of the vector field $E$ and of the discontinuities set of $\nabla \cdot E$, i.e. the set of creases of $E$, using the terminology of the Blake-Zisserman model. The fact that in the functional $F$ the discontinuities set of $E$ is not present depends on the fact that we are assuming that the charge density $f$ in the functional $J_{MS}$ do not concentrate on surfaces or on lines. If we admit concentrated charge layers we can consider the Blake-Zisserman model for the vector function $E$ as a relaxed version of the Mumford-Shah model for the charge $f$. In other words we propose to investigate the connections between minimizers of $J_{MS}$ and minimizers of the higher order functional $J_{BZ}$:
\begin{eqnarray}\label{defn:JBZE}
J_{BZ}(\Gamma_0,\Gamma_1,E)&=&\|E-E_g\|_{L^2(\mathbb{R}^3)}^2 \nonumber \\
&& + \alpha \int_{\mathbb{R}^3 \setminus (\Gamma_0 \cup \Gamma_1)} |\Delta E(x)|^2\; \textrm{d}x \nonumber \\
&&+ \beta \;\mathcal{H}^{2}(\Gamma_0 )\nonumber \\
&& + \gamma\; \mathcal{H}^{2}(\Gamma_1 \setminus \Gamma_0),
\end{eqnarray}
with the additional constraint $\nabla \wedge E=0$.
The main advantage of this approach is that the functional $J_{BZ}$ is a purely differential functional, while the functional $J_{MS}$ is an integro-differential one. We expect that some results about the Blake-Zisserman model that could be rephrased into tomographic terms would provide immediately new results in tomography. Conversely all the peculiar tomographic features as the intrinsic vector nature of the variable $E$, the fact that its support cannot be bounded and the extra-constraint $\nabla \wedge E = 0$, motivate new research directions in the study of free-discontinuities problems. For example, an interesting result in this context would be the determination of a good hypothesis on the datum $E_g$ that ensure that the charge density $f$ do not concentrate. \
We conclude this section with some comments:
\begin{itemize}
\item We proved that the measured data $g$ are actually the measured electric field produced by the unknown charge density, so the term $\|\mathcal{R}{f}-g\|^2_{L^2(\PM^3)}$ in the functional is a fidelity term in a weak sense.
\item The problem of the reconstruction of the charge can be rephrased into a reconstruction problem for the electric field. The electric field is an irrotational vector field, so the new minimization problem is actually a constrained minimization. In order to avoid this constraint one could reformulate the reconstruction problem in terms of the electric potential $\varphi$ ($E=-\nabla \varphi$) obtaining a third-order functional in which the fidelity term is
\begin{equation}
\|\nabla \varphi - \nabla \varphi_g\|^2_{L^2(\mathbb{R}^3)},
\end{equation}
where the potentials are given by~(\ref{eq:potentials}).
\item All this considerations hold true in dimension~$3$. In a generic dimension $n \geq 3$ the situation is quite different because the inversion formula for the Radon transform involves a (possibly fractional) power of the Laplacian. In this case the electrostatic description of tomography given in this section fails. In order to restore it, it is necessary to consider another Radon-type transform which involves integrals of $f$ over linear manifolds with codimension $d$ such that $(n-d)/2=1$, i.e. $d= n-2$, see e.g. \cite{Helgason,tomolectures}.
\end{itemize}
\section*{Acknowledgments}
We thank G. Devillanova, G. Florio and F. Maddalena for for helpful discussions.
This work was supported by ``Fondazione
Cassa di Risparmio di Puglia'' and by the Italian National Group of Mathematical Physics (GNFM-INdAM).
\section*{References}
|
\section{Introduction}
\subsection{{\em The model}}
Let us consider the {\em Boltzmann-Gibbs} measure :
\begin{equation}\label{gib}
\mu(dx)=Z_{\mu}^{-1}e^{-\beta V(x)}dx,
\end{equation}
where $x\in \mathcal{D}^N$ denotes the position of $N$ particles in $\mathcal{D}$. The space $\mathcal{D}$ is called the configuration space. One should think of $\mathcal{D}$ as a subset of $\mathbb{R}^n$, or the $n$-dimensional torus $\mathbb{T}^n$ (where $\mathbb{T}=\mathbb{R}/\mathbb{Z}$ denotes the one dimensional torus). The potential energy function $V:\mathcal{D} \longrightarrow \mathbb{R}$ associates to the positions of the particles $x\in \mathcal{D}$ its energy $V(x)$. In addition, $Z_{\mu}=\displaystyle \int_{\mathcal{D}} e^{-\beta V(x)}dx$ (assumed to be finite) is the normalization constant and $\beta=1/(k_{B}T)$ is proportional to the inverse of the temperature $T$, $k_{B}$ being the Boltzmann constant.
The probability measure $\mu$ is the equilibrium measure sampled by the particles in the canonical statistical ensemble. A typical dynamics that can be used to sample this measure is the {\em Overdamped Langevin Dynamics}:
\begin{equation}\label{over}
dX_t=-\nabla V(X_t)dt+\sqrt{\frac{2}{\beta}}dW_t,
\end{equation}
where $X_t\in \mathcal{D}^N$ and $W_t$ is a $Nn$-dimensional standard Brownian motion. Under loose assumptions on $V$, the dynamics $(X_t)_{t\geq 0}$ is ergodic with respect to the measure $\mu$, which means: for any smooth test function $\varphi$,
\begin{equation} \label{ergo}
\displaystyle\lim _{T\rightarrow +\infty}\frac{1}{T}\int_0^T\varphi(X_t)dt=\int\varphi d\mu,
\end{equation}
i.e. trajectory averages converge to canonical averages.
\subsection{{\em Metastability, reaction coordinate and free energy}}\label{meta}
In many cases of interest, there exists regions of the configuration space where the dynamics \eqref{over} remains trapped for a long time, and jumps only occasionally to another region, where it again remains trapped for a long time. This typically occurs when there exist high probability regions separated by very low probability areas. The regions where the process $(X_t)_{t \ge 0} $ remains trapped for very long times, are called metastable.
Because of the metastability, trajectorial averages \eqref{ergo} converge very slowly to their ergodic limit. Many methods have been proposed to overcome this difficulty, and we concentrate here on the Adaptive Biasing Force (denoted ABF) method (see [\refcite{dav:01,hen:04}]). In order to introduce the ABF method, we need another ingredient: a {\em reaction coordinate} (also known as an order parameter), $\xi=(\xi_1,...,\xi_m):\mathcal{D}\longrightarrow \mathbb{R}^m$, $\xi(x)=z$, where $m< nN$. Typically, in \eqref{over}, the time-scale for the dynamics on $\xi(X_t)$ is larger than the time-scale for the dynamics on $X_t$ due to the metastable states, so that $\xi$ can be understood as a function such that $\xi(X_t)$ is in some sense a slow variable compared to $X_t$. We can say that $\xi$ describes the metastable states of the dynamics associated to the potential $V$.
For a given configuration $x$, $\xi(x)$ represents some macroscopic information. For example, it could represent angles or bond lengths in a protein, positions of defects in a material, etc ...
In any case, it is meant to be a function with values in a small dimensional space (i.e. $m\leq 4$), since otherwise, it is difficult to approximate accurately the associated free energy which is a scalar function defined on the range of $\xi$ (see equation \eqref{free0} below). The choice of a "good" reaction coordinate is a highly debatable subject in the literature. One aim of the mathematical analysis conducted here or in previous papers (see for example [\refcite{tony:08}]) is to quantify the efficiency of numerical algorithms once a reaction coordinate has been chosen.
The image of the measure $\mu$ by $\xi$ is defined by:
\begin{equation}\label{imxi}
\xi*\mu:= \displaystyle\exp(-\beta A(z))dz ,
\end{equation}
where $A$ is the so-called free energy associated with the reaction coordinate $\xi$. By the co-formula (see [\refcite{tony:08}], Appendix A), the following formula for the free energy can then be obtained: up to an additive constant,
\begin{equation}\label{free0}
A(z)=-\beta^{-1}\,{\rm ln}(Z_{\Sigma_{z}}),
\end{equation}
where $Z_{\Sigma_{z}}=\displaystyle\int_{\Sigma_{z}}e^{-\beta V(x)}\delta_{\xi(x)-z}(dx)$, the submanifold $\Sigma_{z}$ is defined by
$$\Sigma_{z}=\{x=(x_1,...,x_n)\in\mathcal{D}\,|\,\xi(x)=z\},$$
and $\delta_{\xi(x)-z}(dx)$ represents a measure with support $\Sigma_{z}$, such that $\delta_{\xi(x)-z}(dx)dz=dx$ (for further details on delta measures, we refer to [\refcite{tony:10}], Section~3.2.1). We assume henceforth that $\xi$ and $V$ are such that $Z_{\Sigma_{z}}<\infty$, for all $z\in \mathbb{R}^m$.
The idea of free energy biasing methods, such as the adaptive biasing force method (see [\refcite{dav:01,hen:04}]) or the Wang Landau algorithm (see [\refcite{wan:01}]), is that, if $\xi$ is well chosen, the dynamics associated to $V-A\circ\xi$ is less metastable than the dynamics associated to $V$. Indeed, from the definition of the free energy $\eqref{imxi}$, for any compact subspace $\mathcal{M}\subset \mathbb{R}^m$, the image of $\tilde{Z}^{-1}\mathrm{e}^{(-\beta(V-A\circ \xi)(x))}1_{\xi(x)\in\mathcal{M}}$ by $\xi$ is the uniform law $\displaystyle\frac{1_{\mathcal{M}}}{|\mathcal{M}|}$, where $\tilde{Z}=\displaystyle\int_{Z_{\Sigma_{z}}}\mathrm{e}^{(-\beta(V-A\circ \xi)(x))}1_{\xi(x)\in\mathcal{M}}$ and $|\mathcal{M}|$ denotes the Lebesgue measure on $\mathcal{M}$. The uniform law is typically easier to sample than the original measure $\xi*\mu$. If the function $\xi$ is well chosen (i.e. if the dynamics in the direction orthogonal to $\xi$ is not too metastable), the free energy can be used as a biasing potential to accelerate the sampling of the dynamics (see [\refcite{tony:08}]). The difficulty is of course that the free energy $A$ is unknown and difficult to approximate using the original dynamics \eqref{over} because of metastability. Actually, in many practical cases, it is the quantity of interest that one would like to approximate by molecular dynamics simulations (see [\refcite{Chi:07,tony:10}]). The principle of adaptive biasing methods is thus to approximate $A$ (or its gradient) on the fly in order to bias the dynamics and to reduce the metastable features of the original dynamics \eqref{over}.
\subsection{\textit{Adaptive biasing force method (ABF)}}
In order to introduce the ABF method, we need a formula for the derivatives of $A$. The so called {\em mean force} $\nabla A(z)$, can be obtained from \eqref{free0} as (see [\refcite{tony:10}], Section~3.2.2):
\begin{equation}\label{mean}
\nabla A(z)=\displaystyle \int_{\Sigma_{z}}f(x)d\mu_{\Sigma_{z}},
\end{equation}
where $d\mu_{\Sigma_{z}}$ is the probability measure $\mu$ conditioned to a fixed value $z$ of the reaction coordinate:
\begin{equation}\label{codms}
d\mu_{\Sigma_{z}}=Z_{\Sigma_{z}}^{-1}\mathrm{e}^{-\beta V(x)}\delta_{\xi(x)-z}(dx),
\end{equation}
and $f$ is the so-called {\em local mean force} defined by
\begin{equation}\label{cmf}
f_i=\displaystyle \sum _{j=1}^m G_{i,j}^{-1}\nabla \xi_j.\nabla V-\beta^{-1} {\rm div}\left(\sum _{j=1}^m G_{i,j}^{-1}\nabla \xi_j\right),
\end{equation}
where $G=(G_{i,j})_{i,j=1,...,m}$, has components $G_{i,j}=\nabla \xi_i\cdot\nabla \xi_j$.
This can be rewritten in terms of conditional expectation as: for a random variable $X$ with law $\mu$ (defined by \eqref{gib}),
\begin{equation}
\nabla A(z)=\mathbb{E}(f(X)|\xi(x)=z).
\end{equation}
We are now in position to introduce the standard adaptive biasing force (ABF) technique, applied to the overdamped Langevin dynamics \eqref{over}:
\begin{equation}\label{abf0}
\left \lbrace
\begin{aligned}
dX_t &=- \left ( \nabla V-\sum_{i=1}^mF_t^i\circ\xi\nabla\xi_i+ \nabla(W\circ \xi) \right )(X_t) dt+\sqrt{2\beta^{-1}}dW_t,\\
F_t^i(z)&=\displaystyle\mathbb{E}[f_i(X_t) |\xi(X_t)=z],\,i=1,...,m,
\end{aligned}
\right.
\end{equation}
where $f$ is defined in \eqref{cmf}. Compared with the original dynamics \eqref{over}, two modifications have been made to obtain the ABF dynamics \eqref{abf0}:
\begin{enumerate}
\item First and more importantly, the force $\displaystyle\sum_{i=1}^mF_t^i\circ\xi\nabla\xi_i$ has been added to the original force $-\nabla V$. At time $t$, $F_t$ approximates $\nabla A$ defined in \eqref{mean}.
\item Second, a potential $W\circ\xi$ has been added. This is actually needed in the case when $\xi$ lives in an unbounded domain. In this case, a so-called confining potential $W$ is introduced so that the law of $\xi(X_t)$ admits a longtime limit $Z_W^{-1}\mathrm{e}^{-\beta W(z)}dz$ (see Remark~\ref{conlaw} at the end of Section~\ref{pr}), where $Z_W=\displaystyle\int_{\mathrm{Ran}(\xi)}\mathrm{e}^{-\beta W}$ is assumed to be finite. When $\xi$ is living in a compact subspace of $\mathbb{R}^m$, there is no need to introduce such a potential and the law of $\xi(X_t)$ converges exponentially fast to the uniform law on the compact subspace (as explained in Section~\ref{meta} and Section~\ref{pr}). Typically, $W$ is zero in a chosen compact subspace $\mathcal{M}$ of $\mathbb{R}^m$ and is harmonic outside $\mathcal{M}$. For example, in dimension two, suppose that $\xi=(\xi_1,\xi_2)$ and $\mathcal{M}=[\xi_{min},\xi_{max}]\times[\xi_{min},\xi_{max}]$, then $W$ can be defined as:
\begin{equation}\label{conf}
W(z_1,z_2)=\displaystyle\sum_{i=1}^2 1_{z_i\geq \xi_{max}}(z_i-\xi_{max})^2+\sum_{i=1}^21_{z_i\leq \xi_{min}}(z_i-\xi_{min})^2.
\end{equation}
\end{enumerate}
It is proven in [\refcite{tony:08}] that, under appropriate assumptions, $F_t$ converges exponentially fast to $\nabla A$. In addition, for well chosen $\xi$, the convergence to equilibrium for \eqref{abf0} is much quicker than for \eqref{over}. This can be quantified using entropy estimates and Logarithmic Sobolev Inequalities, see [\refcite{tony:08}].
Notice that even though $F_t$ converges to a gradient ($\nabla A$), there is no reason why $F_t$ would be a gradient at time $t$. In this paper, we propose an alternative method, where we approximate $\nabla A$, at any time $t$, by a gradient denoted $\nabla A_t$. The gradient $\nabla A_t$ is defined as the Helmholtz projection of $F_t$. One could expect improvements compared to the original ABF method since the variance of $\nabla A_t$ is then smaller than the variance of $F_t$ (since $A_t$ is a scalar function). Reducing the variance is important since the conditional expectation in \eqref{abf0} is approximated by empirical averages in practice.
\subsection{\textit{Projected adaptive biasing force method (PABF)}}
A natural algorithm to reconstruct $A_t$ from $F_t$, consists in solving the following Poisson problem:
\begin{equation}\label{poi0}
\Delta A_t={\rm div}F_t\quad \mbox{on}\,\mathcal{M},
\end{equation}
with appropriate boundary conditions depending on the choice of $\xi$ and $\mathcal{M}$. More precisely, if $\xi$ is periodic and $\mathcal{M}$ is the torus $\mathbb{T}^m$, then we are working with periodic boundary conditions. If $\xi$ is with values in $\mathbb{R}^m$ and $\mathcal{M}$ is a bounded subset of $\mathbb{R}^m$, then Neumann boundary conditions are needed (see Remark \ref{neu} at the end of Section~\ref{conv}). To solve this Poisson problem, standard methods such as finite difference methods, finite element methods, spectral methods or Fourier transforms can be used. Note that \eqref{poi0} is the Euler equation associated with the minimization problem:
\begin{equation}\label{min}
A_t=\displaystyle \mathop{\mbox{argmin}}_{g\in H^1(\mathcal{M})/\mathbb{R}}\int_{\mathcal{M}} |\nabla g-F_t|^2,
\end{equation}
where $\displaystyle H^1(\mathcal{M})/\mathbb{R}=\left\{ g\in H^1(\mathcal{M})\,|\,\int_{\mathcal{M}}g=0\right\}$ denotes the subspace of $H^1(\mathcal{M})$ of zero average functions.
In view of \eqref{min}, $A_t$ can be interpreted as the function such that its gradient is the closest to $F_t$.
Solving \eqref{poi0} amounts to computing the so-called {\em Helmholtz-Hodge} decomposition of the vector field $F_t$ as (see [\refcite{gira:86}], Section~3):
\begin{equation}\label{helm0}
F_t=\nabla A_t+R_t,\quad \mbox{on}\,\mathcal{M},
\end{equation}
where $R_t$ is a divergence free vector field.
Finally, the {\em projected ABF dynamics} we propose to study is the following non linear stochastic differential equation:
\begin{equation}\label{ABF}
\left\lbrace
\begin{aligned}
dX_t&=-\nabla(V-A_t\circ\xi+W\circ \xi)(X_t)dt+\!\sqrt{2\beta^{-1}}dW_t,\\
\Delta A_t&={\rm div}F_t\quad \mbox{on}\,\mathcal{M}, \mbox{ with appropriate boundary conditions},\\
F_t^i(z)&=\displaystyle\mathbb{E}[f_i(X_t)|\xi(X_t)=z],\,i=1,...,m.
\end{aligned}
\right.
\end{equation}
Compared with the standard ABF dynamics \eqref{abf0}, the only modification is that the mean force $F_t$ is replaced by $\nabla A_t$, which is meant to be an approximation of $\nabla A$ at time $t$.
The main theoretical result of this paper is that $A_t$ converges exponentially fast to the free energy $A$ in $\mathcal{M}$ (at least in a specific setting and for a slightly modified version of \eqref{ABF}, see Section~\ref{lcpa} for more details). Moreover, we illustrate numerically this result on a typical example. From a numerical point of view, the interest of the method is that the variance of the projected estimated mean force (i.e. $\nabla A_t$) is smaller than the variance of the estimated mean force (i.e. $F_t$). We observe numerically that this variance reduction enables a faster convergence to equilibrium for PABF compared with the original ABF.
The paper is organized as follows. In Section~\ref{lcpa}, the longtime convergence of the projected ABF method is proven. Section~\ref{num} is devoted to a numerical illustration of the interest of the projected ABF compared to the standard ABF approach. Finally, the proofs of the results presented in Section~\ref{lcpa} are provided in Section~\ref{pro}.
\section{Longtime convergence of the projected ABF method}\label{lcpa}
For the sake of simplicity, we assume in this Section that $\mathcal{D}=\mathbb{T}^n$ and that $\xi(x)=(x_1,x_2)$. Then $\xi$ lives in the compact space $\mathcal{M}=\mathbb{T}^2$ and we therefore take $W=0$. The free energy can be written as:
\begin{equation}\label{free}
A(x_1,x_2)=-\beta^{-1}\,{\rm ln}(Z_{\Sigma_{(x_1,x_2)}}),
\end{equation}
where $Z_{\Sigma_{(x_1,x_2)}}=\displaystyle\int_{\Sigma_{(x_1,x_2)}}e^{-\beta V(x)}dx_3...dx_n$ and $\Sigma_{(x_1,x_2)}=\{x_1,x_2\}\times \mathbb{T}^{n-2}$. The mean force becomes:
\begin{equation}\label{meanf}
\nabla A(x_1,x_2)=\displaystyle \int_{\Sigma_{(x_1,x_2)}}f(x)d\mu_{\Sigma_{(x_1,x_2)}},
\end{equation}
where $f=(f_1,f_2)=(\partial_1 V,\partial_2 V)$ and the conditional probability measure $d\mu_{\Sigma_{(x_1,x_2)}}$ is:
$$d\mu_{\Sigma_{(x_1,x_2)}}=Z_{\Sigma_{(x_1,x_2)}}^{-1} {\rm e}^{-\beta V}dx_3...dx_n.$$
Finally, the vector field $F_t(x_1,x_2)$ writes $\displaystyle \int_{\Sigma_{(x_1,x_2)}}fd\mu_{\Sigma_{(x_1,x_2)}}(t,.)$, or equivalently:
$$F_t^i(x_1,x_2)=\mathbb{E}\big(\partial_iV(X_t)|\xi(X_t)=(x_1,x_2)\big),\,i=1,2.$$
\subsection{Helmholtz projection}\label{he}
In section \ref{heldec}, weighted Helmholtz-Hodge decomposition of $F_t$ is presented. In section \ref{minpb}, the associated minimization problem and projection operator are introduced.
Let us first fix some notations. For $x\in\mathbb{R}^n$ and $1\leq i<j\leq n$, $x_i^j$ denotes the vector $(x_i,x_{i+1},...,x_j)$ and $dx_i^j$ denotes $dx_i\,dx_{i+1}\,...\,dx_j$. Moreover, $\nabla_{x_1^2}$, ${\rm div}_{x_1^2}$ and $\Delta_{x_1^2}$ represent respectively the gradient, the divergence and the laplacian in dimension two for the first two variables $(x_1,x_2)$. Likewise, $\nabla_{x_3^n}=(\partial_{3},...,\partial_n)^T$ represents the gradient vector starting from the third variable of $\mathbb{R}^n$.
\subsubsection{\textit{Helmholtz decomposition}}\label{heldec}
The space $\mathbb{T}^2$ is a bounded and connected space. For any smooth positive probability density function $\varphi:\mathbb{T}^2\rightarrow \mathbb{R}$, let us define the weighted Hilbert space: $L^2_{\varphi}(\mathbb{T}^2)=\{f:\mathbb{T}^2\rightarrow \mathbb{R},\,\int_{\mathbb{T}^2}|f|^2\varphi<\infty\}$. Let us also introduce the Hilbert space $H_{\varphi}({\rm div};\mathbb{T}^2)=\{g\in L^2_{\varphi}(\mathbb{T}^2)\times L^2_{\varphi}(\mathbb{T}^2),\,{\rm div}_{x_1^2}( g)\in L^2(\mathbb{T}^2)\}$. It is well-known that any vector field $F_t:\mathbb{T}^2\rightarrow \mathbb{R}^2$ $\in\,H_1({\rm div},\mathbb{T}^2)$ can be written (see [\refcite{gira:86}], Section~3 for example) as (Helmholtz decomposition): $F_t=\nabla_{x_1^2} A_t+R_t$, where $R_t$ is a divergence free vector field. We will need a generalization of the standard Helmholtz decomposition to the weighted Hilbert spaces $L^2_{\varphi}(\mathbb{T}^2)$ and $H_{\varphi}({\rm div};\mathbb{T}^2))$:
\begin{equation}\label{hel}
F_t\varphi=\nabla_{x_1^2} (A_t)\varphi+R_t,
\end{equation}
s.t. ${\rm div}_{x_1^2}(R_t)=0$. This weighted Helmholtz decomposition is required to simplify calculations when studying the longtime convergence (see Remark~\ref{simpdi} in Section~\ref{difprf} for more details). Recall the space: $H^1(\mathbb{T}^2)/\mathbb{R}=\left\{ g\in H^1(\mathbb{T}^2)\,|\,\int_{\mathbb{T}^2}g=0\right\}$. The function $A_t$ is then the solution to the following problem:
\begin{equation}\label{sca}
\displaystyle\int_{\mathbb{T}^2}\nabla_{x_1^2} A_t\cdot\nabla_{x_1^2} g\,\varphi=\int_{\mathbb{T}^2}F_t\cdot\nabla_{x_1^2} g\,\varphi,\quad \forall g\in H^1(\mathbb{T}^2)/\mathbb{R},
\end{equation}
which is the weak formulation of the Poisson problem:
\begin{equation} \label{poi}
{\rm div}_{x_1^2}(\nabla_{x_1^2} A_t\varphi(t,.))={\rm div}_{x_1^2}(F_t\varphi(t,.)),
\end{equation}
with periodic boundary conditions. Using standard arguments (Lax-Milgram theorem), it is straight forward to check that \eqref{sca} admits a unique solution $A_t\in H^1(\mathbb{T}^2)/\mathbb{R}$.
\subsubsection{\textit{Minimization problem and projection on a gradient}}\label{minpb}
\begin{proposition}\label{propmin}
Suppose that $F_t \in H_{\varphi}({\rm div}; \mathbb{T}^2)$. Then for any smooth positive probability density function $\varphi$, the equation \eqref{sca} is the Euler Lagrange equation associated with the following minimization problem:\\
\begin{equation}\label{minw}
A_t=\displaystyle \min_{h\in H^1(\mathbb{T}^2)/\mathbb{R}}\int_{\mathbb{T}^2} |\nabla_{x_1^2} h-F_t|^2\varphi= \min_{h\in H^1(\mathbb{T}^2)/\mathbb{R}}||\nabla_{x_1^2} h-F_t||^2_{L^2_{\varphi}(\mathbb{T}^2)}.
\end{equation}
Furthermore, $A_t$ belongs to $H^2(\mathbb{T}^2)$.
\end{proposition}
\begin{proof}
Let us introduce the application $I:H^1(\mathbb{T}^2)/\mathbb{R}\rightarrow \mathbb{R}^+$, defined by $I(g)=||\nabla_{x_1^2} h-F_t||^2_{L^2_{\varphi}(\mathbb{T}^2)}$. It is easy to prove that $I$ is $\alpha$-convex and coercive, i.e. $\displaystyle\lim_{\|g\|_{H^1}\!\rightarrow+\infty}I(g)=+\infty$. Thus $I$ admits a unique global minimum $A_t\in \displaystyle H^1(\mathbb{T}^2)/\mathbb{R}$. Furthermore, $\forall\varepsilon>0,\forall g\in H^1(\mathbb{T}^2)$,
\begin{equation}\label{epsl}
\begin{aligned}
I(A_t+\varepsilon g)&=\displaystyle\int_{\mathbb{T}^2}|\nabla_{x_1^2}( A_t+\varepsilon g)-F_t|^2\varphi\\
&=\displaystyle\int_{\mathbb{T}^2}|\nabla_{x_1^2} A_t-F_t|^2\varphi-2\varepsilon\int_{\mathbb{T}^2} (\nabla_{x_1^2} A_t-F_t)\cdot\nabla_{x_1^2} g\,\varphi+\varepsilon^2\int_{\mathbb{T}^2}|\nabla_{x_1^2} g|^2\varphi\\
&=I(A_t)-2\varepsilon\displaystyle\int_{\mathbb{T}^2}(\nabla_{x_1^2} A_t-F_t)\cdot\nabla_{x_1^2} g\,\varphi+\varepsilon^2\int_{\mathbb{T}^2}|\nabla_{x_1^2} g|^2\varphi.
\end{aligned}
\end{equation}
Since $A_t$ is the minimum of $I$, then $I(A_t+\varepsilon g)\geq I(A_t),\,\forall\varepsilon>0,\forall g\in H^1(\mathbb{T}^2)$. By considering the asymptotic regime $\varepsilon\rightarrow 0$ in the last equation, one thus obtains the equation \eqref{sca}:
\begin{equation*}
\displaystyle\int_{\mathbb{T}^2}\nabla_{x_1^2} A_t\cdot\nabla_{x_1^2} g\,\varphi=\int_{\mathbb{T}^2} F_t\cdot\nabla_{x_1^2} g\,\varphi,\,\forall g\in H^1(\mathbb{T}^2)/\mathbb{R}.
\end{equation*}
This is the weak formulation of the following problem \eqref{poi} in $H^1(\mathbb{T}^2)/\mathbb{R}$:
$${\rm div}_{x_1^2}(\nabla_{x_1^2} A_t\varphi(t,.))={\rm div}_{x_1^2}(F_t\varphi(t,.)).$$
Since $\varphi$ is a smooth positive function, then $\exists \,\delta>0$, s.t. $\varphi>\delta$. Furthermore, since ${\rm div}_{x_1^2}(F_t\varphi(t,.))\in L^2(\mathbb{T}^2)$, thus $\Delta_{x_1^2} A_t\in L^2(\mathbb{T}^2).$ Therefore, using standard elliptic regularity results, $A_t\in H^2(\mathbb{T}^2)$.
\end{proof}
For any positive probability density function $\varphi$, the estimation vector field $\nabla_{x_1^2}A_t$ is the projection of $F_t$ onto a gradient. In the following, we will use the notation:
\begin{equation}\label{projop}
\mathcal{P}_{\varphi} (F_t)=\nabla_{x_1^2}A_t,
\end{equation}
where the projection operator $\mathcal{P}_{\varphi}$ is a linear projection defined from $H_{\varphi}({\rm div}; \mathbb{T}^2)$ to $ H^1(\mathbb{T}^2)\times H^1(\mathbb{T}^2)$. Notice in particular that $\mathcal{P}_{\varphi}\circ\mathcal{P}_{\varphi}=\mathcal{P}_{\varphi}$.
\subsection{The projected ABF (PABF) method}\label{pr}
We will study the longtime convergence of the following PABF dynamics:
\begin{equation}\label{pabf}
\left\lbrace
\begin{aligned}
dX_t&=-\nabla\big( V-A_t\circ\xi\big)(X_t)dt+\sqrt{2\beta^{-1}}dW_t,\\
\nabla_{x_1^2} A_t&=\mathcal{P}_{\psi^{\xi}} (F_t),\\
F_t^i(x_1,x_2)&=\displaystyle\mathbb{E}[\partial_i V(X_t)|\xi(X_t)=(x_1,x_2)],\,i=1,2,
\end{aligned}
\right.
\end{equation}
where $\mathcal{P}_{\psi^{\xi}} $ is the linear projection defined by \eqref{projop} and $W_t$ is a standard $nN$-dimensional Brownian motion. Thanks to the diffusion term $\sqrt{2\beta^{-1}}dW_t$, $X_t$ admits a smooth density $\psi$ with respect to the the Lebesgue measure on $\mathbb{T}^n$ and $\psi^{\xi}$ then denotes the marginal distribution of $\psi$ along $\xi$:
\begin{equation}\label{psixi}
\psi^{\xi}(t,x_1,x_2)=\displaystyle\int_{\Sigma_{(x_1,x_2)}}\psi(t,x)dx_3^n.\\
\end{equation}
The dynamics \eqref{pabf} is the PABF dynamics \eqref{ABF} with $\xi(x)=(x_1,x_2)$, $W=0$ and a weighted Helmholtz projection. The weight $\psi^{\xi}$ is introduced to simplify the convergence proof (see Remark~\ref{simpdi} in Section~\ref{pro}).
\begin{remark}\label{condme}
If the law of $X_t$ is $\psi(t,x)dx$ then the law of $\xi(X_t)$ is $\psi^{\xi}(t,x_1,x_2)dx_1dx_2$ and the conditional distribution of $X_t$ given $\xi(X_t)=(x_1,x_2)$ is (see \eqref{codms} for a similar formula when $\psi=Z^{-1}_{\Sigma_{(x_1,x_2)}} \mathrm{e}^{-\beta V}$):\\
\begin{equation}\label{cond}
d\mu_{t,x_1,x_2}=\displaystyle\frac{\psi(t,x)dx_3^n}{\psi^{\xi}(t,x_1,x_2)}.
\end{equation}
Indeed, for any smooth functions $f$ and $g$,
$\begin{aligned}
\mathbb{E}(f(\xi(X_t))g(X_t))&=\displaystyle \int_{\mathbb{T}^n}f(\xi(x))g(x)\psi(t,x)dx\\
&=\int_{\mathbb{T}^2}\int_{\Sigma_{(x_1,x_2)}}f\circ \xi g\psi dx_3^ndx_1dx_2\\
&=\int_{\mathbb{T}^2}f(x_1,x_2)\frac{\int_{\Sigma_{(x_1,x_2)}} g\psi dx_3^n}{\psi^{\xi}(x_1,x_2)}\psi^{\xi}(x_1,x_2)dx_1dx_2.
\end{aligned}$\\
$\lozenge$
\end{remark}
Let us now introduce the non linear partial differential equation (the so-called {\em Fokker-Planck equation}) which rules the evolution of the density $\psi(t,x)$ of $X_t$ solution of \eqref{pabf}:\\
\begin{equation}\label{Fok}
\left\{
\begin{array}{l}
\displaystyle\partial_t\psi={\rm div}\left (\left(\nabla
\displaystyle V-\sum_{i=1}^2\partial_iA_t\circ\xi\nabla\xi_i\right )\psi+\beta^{-1}\nabla\psi\right ),\quad \mbox{for}\, (t,x)\in[0,\infty[\times \mathbb{T}^n,\\
\forall t\geq 0,\,{\rm div}(\nabla A_t\psi^{\xi}(t,.))={\rm div}(F_t\psi^{\xi}(t,.)),\, \mbox{in }\mathbb{T}^2\mbox{ with periodic boundary conditions,}\\
\displaystyle \forall t\geq 0,\,\forall (x_1,x_2)\in \mathbb{T}^2,\, F_t^i(x_1,x_2)=\frac{\displaystyle\int_{\Sigma_{(x_1,x_2)}}\partial_iV\psi dx_3^n}{\psi^{\xi}(x_1,x_2)},\,i=1,2.\\
\end{array}
\right.
\end{equation}
The first equation of \eqref{Fok} rewrites:
\begin{equation}\label{reform}
\partial_t\psi={\rm div}[\nabla V\psi+\beta^{-1}\nabla\psi]-\partial_1((\partial_1A_t)\psi)-\partial_2((\partial_2A_t)\psi).
\end{equation}
Suppose that ($\psi,F_t,A_t)$ is a solution of \eqref{Fok} and let us introduce the expected long-time limits of $\psi$, $\psi^{\xi}$ (defined by \eqref{psixi}) and $\mu_{t,x_1,x_2}$ (defined by \eqref{cond}) respectively:
\begin{enumerate}
\item $\psi_{\infty}=e^{-\beta(V-A\circ\xi)}$;
\item $\psi_{\infty}^{\xi}=1$ (uniform law);
\item \begin{equation}\label{ecm}
\mu_{\infty,x_1,x_2}=Z_{\Sigma_{(x_1,x_2)}}^{-1}e^{-\beta V}dx_3^n.
\end{equation}
\end{enumerate}
Notice that the probability measure $\psi_{\infty}^{\xi}(x_1,x_2)dx_1dx_2$ is the image of the probability measure $\psi_{\infty}(x)dx$ by $\xi$ and that $ \mu_{\infty,x_1,x_2}=\mu_{\Sigma_{(x_1,x_2)}}$ defined in \eqref{codms}. Furthermore, we have that
$$\displaystyle\int_{\mathbb{T}^n}\psi_{\infty}=1,\displaystyle\int_{\mathbb{T}^2}\psi_{\infty}^{\xi}=1 \mbox{ and }\forall (x_1,x_2)\in \mathbb{T}^2,\quad\displaystyle\int_{\Sigma_{(x_1,x_2)}}d\mu_{\infty,x_1x_2}=1.$$
\begin{remark}\label{conlaw}
In the case when $W\neq0$, the first equation of the Fokker-Plank problem~\eqref{Fok} becomes:
$$\partial_t\psi={\rm div}\left (\nabla\left(V-A_t\circ\xi-W\circ\xi \right)\psi+\beta^{-1}\nabla\psi\right ).$$
The expected long-time limits of $\psi$, $\psi^{\xi}$ and $\mu_{t,x_1,x_2}$ are respectively:
\begin{enumerate}
\item $\psi_{\infty}=Z_W^{-1}e^{-\beta(V-A\circ\xi-W\circ\xi)}$;
\item $\psi_{\infty}^{\xi}=Z_W^{-1}e^{-\beta W}$;
\item $\mu_{\infty,x_1,x_2}=Z_{\Sigma_{(x_1,x_2)}}^{-1}e^{-\beta V}dx_3^n$,
\end{enumerate}
where $\displaystyle Z_W=\int_{\mathbb{T}^2}e^{-\beta W}$.
\end{remark}
\subsection{Precise statements of the longtime convergence results}\label{res}
In section~\ref{entr}, some well-known results on entropy techniques are presented. For a general introduction to logarithmic Sobolev inequalities, their properties and their relation to long-time behaviours of solutions to partial differential equations, we refer to [\refcite{ane:00,arn:01,Vil:03}]. Section~\ref{conv} presents the main theorem of convergence.
\subsubsection{\textit{Entropy and Fisher information}}\label{entr}
Define the relative entropy $H(.|.)$ as follows: for any probability measures $\mu$ and $\nu$ such that $\mu$ is absolutely continuous with respect to $\nu$ (denoted $\mu \ll\nu$),
$$H(\mu | \nu)=\displaystyle\int{\rm ln}\left(\frac{d\mu}{d\nu}\right)d\mu.$$
Abusing the notation, we will denote $H(\varphi|\psi)$ for $H(\varphi(x) dx|\psi(x) dx)$ in case of probability measures with densities.
Let us recall the Csiszar-Kullback inequality (see [\refcite{arn:01}]):
\begin{equation}\label{kul}
\|\mu-\nu\|_{TV}\leq \sqrt{2H(\mu|\nu)},
\end{equation}
where $\|\mu-\nu\|_{TV}=\displaystyle\sup_{\|f\|_{L^{\infty}}\leq1}\left\{\int fd(\mu-\nu)\right\}$ is the total variation norm of the signed measure $\mu-\nu$. When both $\mu$ and $\nu$ have densities with respect to the Lebesque measure, $\|\mu-\nu\|_{TV}$ is simply the $L^1$ norm of the difference between the two densities. The entropy $H(\mu|\nu)$ can be understood as a measure of how close $\mu$ and $\nu$ are.
Now, let us define the Fisher information of $\mu$ with respect to $\nu$:
\begin{equation}\label{fish}
I(\mu | \nu)=\displaystyle\int\left|\nabla{\rm ln}\left(\frac{d\mu}{d\nu}\right)\right|^2d\mu.
\end{equation}
The Wasserstein distance is another way to compare two probability measures $\mu$ and $\nu$ defined on a space $\Sigma$,
$$W(\mu,\nu)=\sqrt{\displaystyle\inf_{\pi\in\prod(\mu,\nu)}\int_{\Sigma\times\Sigma}d_{\Sigma}(x,y)^2d\pi(x,y)},$$
where the geodesic distance $d_{\Sigma}$ on $\Sigma$ is defined as: $\forall \, x,y\in \Sigma,$
$$d_{\Sigma}(x,y)=\inf\left\{\sqrt{\int_0^1|\dot{w}(t)|^2dt}\,\mid \,w\in C^1([0,1],\Sigma),\,w(0)=x,w(1)=y\right\},$$
and $\prod(\mu,\nu)$ denotes the set of coupling probability measures, namely probability measures on $\Sigma\times\Sigma$ such that their marginals are $\mu$ and $\nu$: $\forall\pi\in\prod(\mu,\nu),\int_{\Sigma\times\Sigma}\phi(x)d\pi(x,y)=\int_{\Sigma}\phi d\mu $ and $\int_{\Sigma\times\Sigma}\psi(y)d\pi(x,y)=\int_{\Sigma}\psi d\nu .$
\begin{definition}
We say that a probability measure $\nu$ satisfies a logarithmic Sobolev inequality with constant $\rho>0$ (denoted LSI($\rho$)) if for all probability measure $\mu$ such that $\mu \ll\nu$ ,
$$H(\mu | \nu)\leq\frac{1}{2\rho}I(\mu | \nu).$$
\end{definition}
\begin{definition}
We say that a probability measure $\nu$ satisfies a Talagrand inequality with constant $\rho>0$ (denoted T($\rho$)) if for all probability measure $\mu$ such that $\mu \ll\nu$,
$$W(\mu,\nu)\leq\sqrt{\frac{2}{\rho}H(\mu | \nu)}.$$
\end{definition}
\begin{remark}
We implicitly assume in the latter definition, that the probability measures have finite moments of order 2. This is the case for the probability measures used in this paper.
\end{remark}
The following lemma is proved in [\refcite{Ott:00}], Theorem~1:
\begin{lemma}\label{tala}
If $\nu$ satisfies $LSI(\rho)$, then $\nu$ satisfies $T(\rho)$.
\end{lemma}
Recall that $X_t$ solution to \eqref{pabf} has a density $\psi(t,.)$. In the following, we denote the {\em Total Entropy} by
\begin{equation}\label{tent}
E(t)=H(\psi(t,.)|\psi_{\infty})=\int_{\mathbb{T}^n}\ln(\psi/\psi_{\infty})\psi ,
\end{equation}
the {\em Macroscopic Entropy} by
\begin{equation}\label{maent}
E_M(t)=H(\psi^{\xi}(t,.)|\psi^{\xi}_{\infty})=\int_{\mathbb{T}^2}\ln(\psi^{\xi}/\psi^{\xi}_{\infty})\psi^{\xi},
\end{equation}
and the {\em Microscopic Entropy} by
\begin{equation}\label{mient}
E_m(t)=\displaystyle\int_{\mathcal{M}}e_m(t,x_1,x_2)\psi^{\xi}(t,x_1,x_2)dx_1dx_2,
\end{equation}
where $e_m(t,x_1,x_2)=H(\mu_{t,x_1,x_2}|\mu_{\infty,x_1,x_2})$. The following result is straightforward to check:
\begin{lemma}\label{entsum} It holds, $\forall t\geq 0$,
$$E(t)=E_M(t)+E_m(t).$$
\end{lemma}
Note that the Fisher information of $\mu_{t,x_1,x_2}$ with respect to $\mu_{\infty,x_1,x_2}$ can be written as (see \eqref{fish}):
$$I(\mu_{t,x_1,x_2}|\mu_{\infty,x_1,x_2})=\displaystyle \int_{\Sigma_{(x_1,x_2)}}|\nabla_{x_3^n}\ln(\psi(t,.)/\psi_{\infty})|^2d\mu_{t,x_1,x_2}.$$
\subsubsection{\textit{Convergence of the PABF dynamics \eqref{Fok}}}\label{conv}
The following proposition shows that the density function $\psi^{\xi}$ satisfies a simple diffusion equation.
\begin{proposition}\label{diff}
Suppose that $(\psi,F_t,A_t)$ is a smooth solution of \eqref{Fok}. Then $\psi^{\xi}$ satisfies:
\begin{equation}\label{dif2}
\left\{\begin{array}{rl}
\partial_t\psi^{\xi}&=\beta^{-1}\Delta_{x_1^2}\psi^{\xi}, \mbox{ in }[0,\infty[\times \mathbb{T}^2,\\
\psi^{\xi}(0,.)&=\psi^{\xi}_0,\mbox{ on } \mathbb{T}^2.
\end{array}
\right.
\end{equation}
\end{proposition}
\begin{remark}\label{rmdif}
If $ \psi^{\xi}_0= 0$ at some points or is not smooth, then $F$ at time $0$ may not be well defined or $I(\psi^{\xi}(0,.)/\psi^{\xi}_{\infty})$ may be infinite. Since, by Proposition \ref{diff}, $ \psi^{\xi}$ satisfies a simple diffusion equation these difficulties disappear as soon as $t > 0$. Therefore, up to considering the problem for $t>t_0>0$, we can suppose that $ \psi^{\xi}_0$ is a smooth positive function.
We also have that for all $t>0$, $\psi^{\xi}(t,.)>0$, $\displaystyle\int_{\mathbb{T}^2}\psi^{\xi}=1$ and $\psi^{\xi}(t,.)\in C^{\infty}(\mathbb{T}^2).$
\end{remark}
\begin{remark}
In the case where $W\neq 0$, the probability density function $\psi^{\xi}$ satisfies the modified diffusion equation:
$$\partial_t\psi^{\xi}=\nabla_{x_1^2}\cdot \left ( \beta^{-1}\nabla_{x_1^2} \psi^{\xi}+\psi^{\xi}\nabla_{x_1^2}W \right).$$
\end{remark}
Here are two simple corollaries of Proposition~\ref{diff}.
\begin{corollary}\label{fishco}
There exists $t_0>0$ and $I_0>0$ (depending on $\psi^{\xi}_0$), such that
$$\forall t>t_0,\quad I(\psi^{\xi}(t,.)|\psi^{\xi}_{\infty})<I_0e^{-\beta^{-1}8\pi^2t}.$$
\end{corollary}
\begin{corollary}\label{mic}
The macroscopic entropy $E_M(t)$, defined by \eqref{maent}, converges exponentially fast to zero:
$$\forall t>t_0,\quad E_M(t)\leq \frac{I_0}{8\pi^2}e^{-\beta^{-1}8\pi^2t},$$
where $I_0$ is the constant introduced in Corollary~\ref{fishco}.
\end{corollary}
The assumptions we need to prove the longtime convergence of the biasing force $\nabla A_t$ to the mean force $\nabla A$ are the following:
\begin{description}
\item[[H1\!\!]] $V\in C^2(\mathbb{T}^n)$ and satisfies:
$$\exists \gamma >0,\,\forall \,3\leq j\leq n,\forall x\in \mathbb{T}^n,\quad \max(|\partial_1\partial_j V(x)|,|\partial_2\partial_j V(x)|)\leq \gamma .$$
\item[[H2\!\!]] $V$ is such that $\exists\rho>0$, $\forall (x_1,x_2)\in \mathbb{T}^2$, $\mu_{\infty,x_1,x_2}=\mu_{\Sigma(x_1,x_2)}$ defined by \eqref{ecm} satisfies $LSI(\rho).$
\end{description}
The main theorem is:
\begin{theorem}\label{theo1}
Let us assume [H1] and [H2]. The following properties then hold:
\begin{enumerate}
\item The microscopic entropy $E_m$ converges exponentially fast to zero:
\begin{equation}\label{Em}
\exists C>0,\exists \lambda>0,\forall t\geq 0,\quad \sqrt{E_m(t)}\leq Ce^{-\lambda t}.
\end{equation}
Furthermore, if $\rho\neq 4\pi^2$, then $\lambda=\beta^{-1}\min (\rho,4\pi^2)$ and\\
$\displaystyle C=\sqrt{E_m(0)}+\frac{\gamma}{\beta^{-1}|\rho-4\pi^2|}\sqrt{\frac{I_0}{2\rho}}$.
If $\rho=4\pi^2$, then for all $\lambda<\beta^{-1}\rho$, there exists a positive constant $C$ such that \eqref{Em} is satisfied.
\item $\sqrt{E(t)}$ and $\|\psi(t,.)-\psi_{\infty}\|_{L^1( \mathbb{T}^n)}$ both converge exponentially fast to zero with rate $\lambda$.
\item The biasing force $\nabla_{x_1^2} A_t$ converges to the mean force $\nabla_{x_1^2} A$ in the following sense:
\begin{equation}\label{At}
\displaystyle\forall\, t\geq 0,\,\int_{\mathbb{T}^2}|\nabla_{x_1^2} A_t-\nabla_{x_1^2} A|^2\psi^{\xi}(t,x_1,x_2)dx_1dx_2\leq\frac{8\gamma^2}{\rho}E_m(t).
\end{equation}
\end{enumerate}
\end{theorem}
The proofs of the results presented in this section are provided in Section~\ref{pro}.
\begin{remark}
We would like to emphasize that our arguments hold under the assumption of existence of regular solutions. In particular, we suppose that the density $\psi(t,.)$ is sufficiently regular so that the algebric manipulations in the proofs (see Section~\ref{pro}) are valid.
\end{remark}
\begin{remark}
This remark is devoted to show how the rate of convergence of the dynamics \eqref{over} is improved thanks to PABF method. First of all, we mention a classical computation to get a rate of convergence for \eqref{over}. Precisely, if one denotes $\varphi (t,.)$ the probability density function of $X_t$ satisfying \eqref{over}, and $\varphi_{\infty}=Z_{\mu}^{-1}\mathrm{e}^{-\beta V}$ its longtime limit, then by standard computations (see for example [\refcite{arn:01}]), one obtains:
$$ \frac{d}{dt}H(\varphi(t,.)|\varphi_{\infty}) =-\beta^{-1}I(\varphi(t,.)|\varphi_{\infty}).$$
Therefore, if $\varphi_{\infty}$ satisfies $LSI(R)$, then one obtains the estimate
\begin{equation}\label{overcv}
\exists R>0,\,\forall t>0,\quad H(\varphi(t,.)|\varphi_{\infty})\leq H(\varphi_0|\varphi_{\infty})\,\mathrm{e}^{-2\beta^{-1}Rt}.
\end{equation}
From \eqref{kul}, we obtain that $\|\varphi(t,.)-\varphi_{\infty}\|_{L^1( \mathbb{T}^n)}$ converges exponentially fast to zero with rate $\beta^{-1}R$. The constant $R$ is known to be small if the metastable states are separated by large energy barriers or if high probability regions for $\mu$ are separated by large regions with small probability (namely $\mu$ is a multimodal measure).
Second, by Theorem~\ref{theo1}, one can show that $\nabla A_t$ converges exponentially fast to $\nabla A$ in $L^2(\psi^{\xi}_{\infty}(x_1,x_2)dx_1dx_2)-$norm at rate $\lambda=\beta^{-1}\min (\rho,4\pi^2)$. Indeed, since $\psi^{\xi}_{\infty}=1$,
\begin{align*}
\int_{\mathbb{T}^2}|\nabla A_t-\nabla A|^2dx_1dx_2&=\int_{\mathbb{T}^2}|\nabla A_t-\nabla A|^2\frac{\psi^{\xi}(t,x_1,x_2)}{\psi^{\xi}(t,x_1,x_2)}dx_1dx_2\\
&\leq \frac{8\gamma^2}{(1-\varepsilon)\rho}E_m(t),
\end{align*}
where $\varepsilon >0$ such that $\psi^{\xi}(t,x_1,x_2)\geq 1-\varepsilon$ (for more details refer to the proof of Corollary~\ref{fishco} in Section~\ref{pro}).
This result must be compared with \eqref{overcv}. More precisely, $\lambda$ is related to the rate of convergence $4\pi^2$ at the macroscopic level, for equation \eqref{dif2} satisfied by $\psi^{\xi}$, and the rate of convergence $\rho$ at the microscopic level, coming from the logarithmic Sobolev inequalities satisfied by the conditional measures $\mu_{\infty,x_1,x_2}$. Of course, $\rho$ depends on the choice of the reaction coordinate. In our framework, we could state that a "good reaction coordinate" is such that $\rho$ is as large as possible. Typically, for good choices of $\xi$, $\lambda \gg R$, and the PABF dynamics converges to equilibrium much faster than the original dynamics \eqref{over}. This is typically the case if the conditional measures $\mu_{\infty,x_1,x_2}$ are less multimodal than the original measure $\mu$.
\end{remark}
\begin{remark}(Extension to other geometric setting)\label{neu}\\
The results of Theorem~\ref{theo1} are easily generalized to the following setting:
If $\mathcal{D}=\mathbb{R}^n$, $\xi(x)=(x_1,x_2)$ and $\mathcal{M}$ is a compact subspace of $\mathbb{R}^n$, then choose a confining potential $W$ (defined in \eqref{conf}) such that $Z_{W}=\int \mathrm{e}^{-\beta W}<+\infty$, $Z_{W}^{-1}\mathrm{e}^{-\beta W}$ satisfies $LSI(r^*)$ (for some $r^*>0$) and $W$ is convex potential, then Corollary~\ref{fishco} is satisfied with rate $2\beta^{-1}(r^*-\varepsilon)$, for any $\varepsilon\in (0,r^*)$ (refer to Corollary~1 in [\refcite{tony:08}] for further details). In this case, Neumann boundary conditions are needed to solve the Poisson problem \eqref{poi}:
\begin{equation}\label{poisext}
\left\lbrace
\begin{aligned}
{\rm div}(\nabla A_t\psi^{\xi}(t,.))&={\rm div}(F_t\psi^{\xi}(t,.))& \mbox{in}\,\mathcal{M},\\
\frac{\partial A_t}{\partial n}&=F_t.n &\text{on}\,\partial \mathcal{M},
\end{aligned}
\right.
\end{equation}
where $n$ denotes the unit normal outward to $ \mathcal{M}$. The convergence rate $\lambda$ of Theorem~\ref{theo1} becomes $\beta^{-1}\min(\rho,r^*-\varepsilon)$. Neumann boundary conditions come from the minimization problem \eqref{minw} associated to the Euler-Lagrange equation. The numerical applications in Section~\ref{num} are performed in this setting.
\end{remark}
\begin{remark}(Extension to more general reaction coordinates)\\
In this section, we have chosen $\xi(x_1,...,x_n)=(x_1,x_2)$. The results can be extended to the following settings:
\begin{enumerate}
\item In dimension one, the Helmholtz projection has obviously no sense. If $\mathcal{D}=\mathbb{T}^n$ and $\xi(x)=x_1$, then $F_t$ converges to $A'$, which is a derivative of a periodic function and thus $\displaystyle\int_{\mathbb{T}}A'=0$. Since $\displaystyle\int_{\mathbb{T}}F_t$ is not necessary equal to zero, one can therefore take a new approximation $A_t^{'}=F_t-\displaystyle \int_{\mathbb{T}}F_t$, which approximates $A'$ at any time $t$. The convergence results of this section can be extended to this setting, to show that $A'_t$ converges exponentially fast to $A'$.
\item More generally, for a reaction coordinate with values in $\mathbb{T}^m$, the convergence results presented in this paper still hold under the following orthogonality condition:
\begin{equation}\label{ortho}
\forall i\neq j,\,\nabla \xi_i\cdot\nabla \xi_j=0.
\end{equation}
In the case when \eqref{ortho} does not hold, it is possible to resort to the following trick used for example in metadynamics (refer to [\refcite{bu:06,tony:07}]). The idea is to introduce an additional variable $z$ of dimension $m$, and an extended potential $V_{\xi}(x,z)=V(x)+\frac{\kappa}{2}|z-\xi(x)|^2$, where $\kappa$ is a penalty constant. The reaction coordinate is then chosen as $\xi_{meta}(x,z)=z$, so that the associated free energy is:
$$A_{\xi}(z)=-\beta^{-1}\displaystyle \ln \int_{\mathcal{D}}e^{-\beta V_{\xi}(x,z)}dx,$$
which converges to $A(z)$ (defined in \eqref{free0}) when $\kappa$ goes to infinity. The extended PABF dynamics can be written as:
$$
\left\lbrace
\begin{aligned}
dX_t &= - \left( \nabla V(X_t) + \kappa \sum_{i=1}^m (\xi_i(X_t) - Z_{i,t}) \nabla \xi_i (X_t) \right) \, dt + \sqrt{2 \beta^{-1}} dW_t,\\
dZ_t &= \kappa (\xi(X_t) - \nabla E_t (Z_t)) \, dt + \sqrt{2 \beta^{-1}} d \overline{W}_t,\\
\nabla E_t &= {\mathcal P}_{\psi^{\xi_{meta}}} \left( G_t \right),\\
G_t(z) &= \mathbb{E}(\xi(X_t)|Z_t=z),
\end{aligned}
\right.
$$
where $\overline{W}_t$ is a $m-$dimensional Brownian motion independent of $W_t$. The results of Theorem~\ref{theo1} apply to this extended PABF dynamics.
\end{enumerate}
\end{remark}
\section{Numerical experiments}\label{num}
\subsection{Presentation of the model}
We consider a system composed of $N$ particles $(q_i)_{0\leq i\leq N-1}$ in a two-dimensional periodic box of side length $L$. Among these particles, three particles (numbered $0$, $1$ and $2$ in the following) are designated to form a trimer, while the others are solvent particles. In this model, a trimer is a molecule composed of three identical particles linked together by two bonds (see Figure~\ref{tri}).
\subsubsection{Potential functions}
All particles, except the three particles forming the trimer, interact through the purely repulsive WCA pair potential, which is the Lennard-Jones (LJ) potential truncated at the LJ potential minimum:
$$V_{WCA}(d)=\left\{
\begin{array}{cc}
\varepsilon+4\varepsilon\left[ \left(\frac{\sigma}{d}\right)^{12}-\left(\frac{\sigma}{d}\right)^{6} \right]&\qquad \mbox{if} \, d \leq d_0,\\
0 & \qquad \mbox{if} \, d \geq d_0 ,\\
\end{array}
\right.
$$
where $d$ denotes the distance between two particles, $\varepsilon$ and $\sigma$ are two positive parameters and $d_0 = 2^{1/6}\sigma$.
A particle of the solvent and a particle of the trimer also interact through the potential $V_{WCA}$. The interaction potential between two particles of the trimer ($q_0/q_1$ or $q_1/q_2$) is a double-well potential (see Figure~\ref{pot2}):
\begin{equation}\label{vs}
V_{S}(d)=h\left[ 1- \frac{(d-d_1-\omega)^2}{\omega^2}\right]^{2},
\end{equation}
where $d_1$, $h$ and $\omega$ are positive parameters.\\
The potential $V_{S}$ has two energy minima. The first one, at $d=d_1$, corresponds to the compact bond. The second one, at $d=d_1+2\omega$, corresponds to the stretched bond. The height of the energy barrier separating the two states is $h$.
In addition, the interaction potential between $q_0$ and $q_2$ is a Lennard-Jones potential:
$$V_{LJ}(d)=4\varepsilon '\left[\left(\frac{\sigma'}{d}\right)^{12}-\left(\frac{\sigma'}{d}\right)^{6} \right],$$
where $\varepsilon '$ and $\sigma '$ are two positive parameters.
Finally, the three particles of the trimer also interact through the following potential function on the angle $\theta$ formed by the vectors $\overrightarrow{q_1q_0}$ and $\overrightarrow{q_1q_2}$:
$$V_{\theta_0}(\theta)=\frac{k_{\theta}}{2}\left(\cos(\theta)-\cos(\theta_0) \right)^2,$$
where $\theta_0$ is the equilibrium angle and $k_{\theta}$ is the angular stiffness. Figure~\ref{tri} presents a schematic view of the system.
\begin{figure}[htp]
\centering
\begin{tabular}{ccc}
\includegraphics[width=40mm]{compact.pdf}&
\includegraphics[width=40mm]{mixed.pdf}&
\includegraphics[width=40mm]{stretched.pdf}\\
\end{tabular}
\caption{Trimer ($q_0, q_1, q_2$). Left: compact state; Center: mixed state; Right: stretched state.}\label{tri}
\end{figure}
The total energy of the system is therefore, for $q \in (L\mathbb{T})^{2N}$:
\begin{align*}
V(q)=&\displaystyle \sum_{3\leq i<j\leq N-1}V_{WCA}(|q_i-q_{j}|)+\sum_{i=0}^{2}\sum_{j=3}^{N-1}V_{WCA}(|q_i-q_{j}|)\\
&+ \sum_{i=0}^{1} V_S(|q_i-q_{i+1}|) + V_{LJ}(|q_0-q_{2}|) + V_{\theta_0}(\theta).
\end{align*}
\begin{figure}[htbp]
\centerline{\psfig{figure=dw.pdf,height=50mm}}
\caption{Double-well potential \eqref{vs}, with $d_1=2^{1/6}$, $\omega=2$ and $h=2$.}\label{pot2}
\end{figure}
\subsubsection{Reaction coordinate and physical parameters}
The reaction coordinate describes the transition from compact to stretched state in each bond. It is the normalised bond length of each bond of the trimer molecule. More precisely, the reaction coordinate is $\xi=(\xi_1,\xi_2)$, with $\xi_1(q)=\frac{|q_0- q_1|-d_0}{2\omega}$ and $\xi_2(q)=\frac{|q_1- q_2|-d_0}{2\omega}$. For $i=1,2$, the value $\xi_i=0$ refers to the compact state (i.e. $d=d_0$) and the value $\xi_i=1$ corresponds to the stretched state (i.e. $d=d_0+2\omega$).
We apply ABF and PABF dynamics to the trimer problem described above. The inverse temperature is $\beta = 1$, we use $N = 100$ particles ($N-3$ solvent particles and the trimer) and the box side length is $L=15$. The parameters describing the WCA and the Lennard-Jones interactions are set to $\sigma = 1$, $\varepsilon=1$, $\sigma '=1$, $\varepsilon '=0.1$, $d_0=2^{1/6}$, $d_1=2^{1/6}$ and the additional parameters for the trimer are $\omega = 2$ and $h = 2$. The parameters describing the angle potential are: $\theta_0$ such that $\cos(\theta_0)=1/3$ and $k_{\theta}=1$ (we refer to [\refcite{rap:97}], Section 10.4.2, for the choice of such parameters). The initial condition on the trimer is as follows: Both bonds $q_0q_1$ and $q_1q_2$ are in compact state, which means that the distance between $q_0$ and $q_1$ and the distance between $q_1$ and $q_2$ are equal to $d_0$. Moreover, the initial bond angle is $\theta_0$.
\subsubsection{Numerical methods and numerical parameters}
Standard and projected ABF methods are used with $N_{replicas} = 100$ replicas of the system evolving according to the overdamped Langevin dynamics discretized with a time-step $\Delta t = 2.5\times 10^{-4}$.
The reaction coordinate space of interest is taken of the form $\mathcal{M}= [\xi_{min}, \xi_{max}]\times [\xi_{min}, \xi_{max}] $, where $\xi_{min}=-0.2$ and $\xi_{max}=1.2$. $\mathcal{M}$ is discretized into $N_{bins}\times N_{bins} = 50\times 50=2500$ bins of equal sizes and $\delta=\delta_x=\delta_y=\frac{\xi_{max}-\xi_{min}}{N_{bins}}=0.028$ denotes the size of each bin along both axes.
To implement the ABF and PABF method, one needs to approximate $F_t^i(x,y)=\displaystyle\mathbb{E}[f_i(x,y)|\xi(x,y)=(\xi_1(x,y),\xi_2(x,y))],\,i=1,2$. The mean force $F_t$ is estimated in each bin as a combination of plain trajectorial averages and averages over replicas. It is calculated at each time as an average of the local mean force in the bin over the total number of visits in this bin. More precisely, at time $t$ and for $l=1,2$, the value of the mean force in the $(i,j)^{th}$ bin is:
\begin{equation}\label{F_t}
F_t^{l}(i,j)=\displaystyle \frac{\displaystyle\sum_{t'\leq t}\sum_{k=1}^{N_{replicas}}f_{l}(q_{k,t'})1_{\{indx(\xi(q_{k,t'}))=(i,j)\}}}{\displaystyle\sum_{t'\leq t}\sum_{k=1}^{N_{replicas}}1_{\{indx(\xi(q_{k,t'}))=(i,j)\}}},
\end{equation}
where $q_{k,t}$ denotes the position $(x_k,y_k)$ at time $t$, $f=(f_1,f_2)$ is defined in \eqref{cmf} and $indx(\xi(q_{k,t'}))$ denotes the number of the bin where $\xi(q_{k,t'})$ lives, i.e.
$$indx(\xi(q))=\displaystyle\left(\left[\frac{\xi_1(q)-\xi_{min}}{\delta}\right]_+,\left[\frac{\xi_2(q)-\xi_{min}}{\delta}\right]_+\right),\,\forall q\in\mathcal{M}.$$
If the the components of the index function (i.e. $indx$) are either equal to $-1$ or to $N_{bins}$, it means that we are outside $\mathcal{M}$ and then the confining potential is non zero, and defined as:
$$W(\xi(q_{k,t}))=\displaystyle\sum_{i=1}^2\left[ 1_{\{\xi_i(q_{k,t})\geq \xi_{max}\}}(\xi_i(q_{k,t})-\xi_{max})^2+1_{\{\xi_i(q_{k,t})\leq \xi_{min}\}}(\xi_i(q_{k,t})-\xi_{min})^2\right].$$
To construct the PABF method, the solution to the following Poisson problem with Neumann boundary conditions are approximated:
\begin{equation}\label{poistri}
\left\{
\begin{array}{llll}
\Delta A&=&{\rm div}F&\quad\mbox{in }\mathcal{M}= [\xi_{min}, \xi_{max}]\times [\xi_{min}, \xi_{max}],\\
\frac{\partial A}{\partial n}&=&F\cdot n&\quad \mbox{on }\partial\mathcal{M},
\end{array}
\right.
\end{equation}
where $n$ denotes the unit normal outward to $\mathcal{M}$. We use for simplicity in the numerical experiments the standard Helmholtz problem, without the weight $\psi^{\xi}$. Problem~\eqref{poistri} is solved using finite element method of type $Q^1$ on the quadrilateral mesh defined above, with nodes $(x_i,y_j)$, where $x_i=\xi_{min}+i\delta$ and $y_j=\xi_{min}+j\delta$, for $i,j=0,.., N_{bins}$. The space $\mathcal{M}$ is thus discretized into $N_T=N_{bins}^2$ squares, with $N_s=(N_{bins}+1)^2$ nodes. The associated variational formulation is the following:\\
\begin{equation*}
\left\{
\begin{array}{l}
\mbox{Find}\, A\in H^1(\mathcal{M})/\mathbb{R}\mbox{ such that}\\
\displaystyle \int_{\mathcal{M}}\nabla A\cdot\nabla v=\int_{\mathcal{M}}F\cdot\nabla v,\,\forall\, v\in H^1(\mathcal{M})/\mathbb{R}.
\end{array}
\right.
\end{equation*}
\subsection{ Comparison of the methods}
In this section, we compare results obtained with three different simulations: without ABF, with ABF and with projected ABF (PABF). First, it is observed numerically that both ABF methods overcome metastable states. Second, it is illustrated how PABF method reduces the variance of the estimated mean force compared to ABF method. As a consequence of this variance reduction, we observe that the convergence of $\nabla A_t$ to $\nabla A$ with the PABF method is faster than the convergence of $F_t$ to $\nabla A$ with the ABF method.
\subsubsection{Metastability}
To illustrate numerically the fact that ABF methods improve the sampling for metastable processes, we observe the variation, as a function of time, of the two metastable distances (i.e. the distance between $q_0$ and $q_1$, and the distance between $q_1$ and $q_2$). On Figures~\ref{dst1} and \ref{dst2}, the distance between $q_1$ and $q_2$ is plotted as a function of time for three dynamics: without ABF, ABF and PABF.
Both ABF methods allow to switch faster between the compact and stretched bond and thus to better explore the set of configurations. Without adding the biasing term, the system remains trapped in a neighborhood of the first potential minimum (i.e. $d_0\simeq 1.12$) region for $20$ units of time at least (see Figure~\ref{dst1}), while when the biasing term is added in the dynamics, many jumps between the two local minima are observed (see Figure~\ref{dst2}).
\begin{figure}[htp]
\centering
\includegraphics[width=62mm]{dist_without_abf.pdf}\\
\caption{Without ABF.}\label{dst1}
\end{figure}
\begin{figure}[htp]
\centering
\begin{tabular}{ccc}
\includegraphics[width=62mm]{dist_abf.pdf}&
\includegraphics[width=62mm]{dist_proj.pdf}
\end{tabular}
\caption{Left: ABF; Right: PABF.}\label{dst2}
\end{figure}
\subsubsection{Variance reduction}
Since we use Monte-Carlo methods to approximate $\nabla A_t$, the variance is an important quantity to assess the quality of the result. The following general proposition shows that projection reduces the variance.
\begin{proposition}\label{vared}
Let $F$ be a random function from $ \mathbb{T}^2$ into $ \mathbb{R}^2$ and belongs to $H(\mathrm{div},\mathbb{T}^2)$, and define $\mathcal{P}=\mathcal{P}_1$ (i.e. without weight) the projection on gradient vector fields defined in Section \ref{minpb}. Then, the variance of $\mathcal{P}(F)$ is smaller than the variance of $F$:
$$\displaystyle \int_ \mathcal{M}{\rm Var}(\mathcal{P} (F))\leq \int_ \mathcal{M}{\rm Var}(F),$$
where, for any vector field $F$, $\mathrm{Var}(F)=\mathbb{E}(|F|^2)-\mathbb{E}(|F|)^2$ and $|F|$ being the Euclidian norm.
\end{proposition}
\begin{proof}
Let $F$ be a random vector field of $ H({\rm div},{\mathbb{T}^2})$. Let us introduce $\mathcal{P}(F)\in H^1(\mathbb{T}^2)\times H^1(\mathbb{T}^2)$ its projection. Notice that by the linearity of the projection $\mathcal{P}(\mathbb{E}(F))=\mathbb{E}(\mathcal{P}(F))$. By definition of $\mathcal{P}(F)$, one gets:
$$\displaystyle\int_{\mathbb{T}^2}(F-\mathcal{P}(F))\cdot\nabla h=0,\,\forall h\in H^1(\mathbb{T}^2).$$
Therefore, using Pythagoras and the fact that $\mathcal{P}(F)$ is agradient,
$$\displaystyle\int_{\mathbb{T}^2}|F|^2=\int_{\mathbb{T}^2}|F-\mathcal{P}(F)|^2+\int_{\mathbb{T}^2}|\mathcal{P}(F)|^2$$
and
$$\displaystyle\int_{\mathbb{T}^2}|F-\mathbb{E}(F)|^2=\int_{\mathbb{T}^2}|F-\mathbb{E}(F)-\mathcal{P}(F-\mathbb{E}(F))|^2+\int_{\mathbb{T}^2}|\mathcal{P}(F-\mathbb{E}(F))|^2.$$
Using the linearity of $\mathcal{P}$, we thus obtain
$$\displaystyle\int_{\mathbb{T}^2}{\rm Var}(F)=\int_{\mathbb{T}^2}{\rm Var}(F-\mathcal{P}(F))+\int_{\mathbb{T}^2}{\rm Var}(\mathcal{P}(F)),$$
which concludes the proof.
\end{proof}
We illustrate the improvement of the projected method in terms of the variances of the biasing forces, by comparing ${\rm Var}(\nabla A_t)={\rm Var}(\partial_1 A_t)+{\rm Var}(\partial_2 A_t)$ (for the PABF method) and ${\rm Var}(F_t)={\rm Var}(F_t^1)+{\rm Var}(F_t^2)$ (for the ABF method). Figure~\ref{var} shows that the variance for the projected ABF method is smaller than for the standard ABF method.
We have $N_{bins}\times N_{bins}=2500$ variable for each term (i.e. $\partial_1A_t$, $\partial_2 A_t$, $F_t^1$ and $F_t^2$). The variances are computed using $20$ independent realizations as follows:
$$\displaystyle {\rm Var}( F_t^1)=\frac{1}{2500}\displaystyle \sum_{i,j=1}^{50}\frac{1}{20}\sum_{k=1}^{20}F_t^{1,k}(x_i,y_j)^2 - \frac{1}{2500}\displaystyle \sum_{i,j=1}^{50}\left (\frac{1}{20}\sum_{k=1}^{20}F_t^{1,k}(x_i,y_j)\right )^2.$$
Note that four averages are involved in this formula: an average with respect to the space variable, an average over the $20$ Monte-Carlo realizations, an average over replicas and a trajectorial average (the last two averages are more explicit in \eqref{F_t}). Notice that since the variance of the biasing force is smaller with PABF, one may expect better convergence in time results. This will be investigated in Section~\ref{error} and Section~\ref{distrint}.
\begin{figure}[htp]
\centering
\includegraphics[width=100mm]{variance.pdf}
\caption{Variances as a function of time.}\label{var}
\end{figure}
\subsubsection{Free energy error}\label{error}
We now present, the variation, as a function of time, of the normalized $L^2-$ distance between the real free energy and the estimated one, in both cases: ABF and PABF methods. As can be seen in Figure~\ref{err}, in both methods, the error decreases as time increases. Moreover, this error is always smaller for the projected ABF method than for the ABF method.
\begin{figure}[htp]
\centering
\includegraphics[width=100mm]{error.pdf}
\caption{ Free energy error as a function of time.}\label{err}
\end{figure}
\subsubsection{Distribution}\label{distrint}
Another way to illustrate that the projected ABF method converges faster than the standard ABF method is to plot the density function $\psi^{\xi}$ as a function of time (see Figure~\ref{dtr0}-\ref{dtr25}). It is illustrated that, as time increases, the probability of visiting all bins (of the reaction coordinate space $\mathcal{M}$) increases.
It is observed that, for the projected ABF method, the state where both bonds are stretched is visited earlier (at time $5$) than for the standard ABF method (at time $20$). The convergence to uniform law along $(\xi_1,\xi_2)$ is faster with the projected ABF method.
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\includegraphics[width=40mm, angle=270 ]{int1_0,025.pdf}&
\includegraphics[width=40mm, angle=270 ]{int2_0,025.pdf}\\
\end{tabular}
\caption{At time $0.025$. Left: $\int \psi^{\xi}(x_1,x_2)dx_2$; Right: $\int \psi^{\xi}(x_1,x_2)dx_1$.}\label{dtr0}
\end{figure}
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\includegraphics[width=40mm, angle=270]{int1_5.pdf}&
\includegraphics[width=40mm, angle=270]{int2_5.pdf}\\
\end{tabular}
\caption{At time $5$. Left: $\int \psi^{\xi}(x_1,x_2)dx_2$; Right: $\int \psi^{\xi}(x_1,x_2)dx_1$.}\label{dtr5}
\end{figure}
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\includegraphics[width=40mm, angle=270]{int1_10.pdf}&
\includegraphics[width=40mm, angle=270]{int2_10.pdf}\\
\end{tabular}
\caption{At time $10$. Left: $\int \psi^{\xi}(x_1,x_2)dx_2$; Right: $\int \psi^{\xi}(x_1,x_2)dx_1$.}\label{dtr10}
\end{figure}
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\includegraphics[width=40mm, angle=270]{int1_20.pdf}&
\includegraphics[width=40mm, angle=270]{int2_20.pdf}\\
\end{tabular}
\caption{At time $20$. Left: $\int \psi^{\xi}(x_1,x_2)dx_2$; Right: $\int \psi^{\xi}(x_1,x_2)dx_1$.}\label{dtr20}
\end{figure}
\newpage
\begin{figure}[htp]
\centering
\begin{tabular}{cc}
\includegraphics[width=40mm, angle=270]{int1_25.pdf}&
\includegraphics[width=40mm, angle=270]{int2_25.pdf}\\
\end{tabular}
\caption{At time $25$. Left: $\int \psi^{\xi}(x_1,x_2)dx_2$; Right: $\int \psi^{\xi}(x_1,x_2)dx_1$.}\label{dtr25}
\end{figure}
\section{Proofs}\label{pro}
The proofs are inspired from [\refcite{tony:08}]. One may assume that $\beta=1$ up to the following change of variable: $\tilde{t}=\beta^{-1}t$, $\tilde{\psi}(\tilde{t},x)=\psi(t,x),$ $\tilde{V}(x)=\beta V(x)$. Recall, that we work in $\mathcal{D}=\mathbb{T}^n$, $\mathcal{M}=\mathbb{T}^2$ and $\forall x=(x_1,...,x_n)\in~\mathbb{T}^n$, $\xi(x)=(x_1,x_2)$.
\subsection{\textit{Proof of Proposition \ref{diff}}}\label{difprf}
Let $g:\mathbb{T}^2\rightarrow\mathbb{R}$ be a function in $H^1(\mathbb{T}^2).$
\begin{align*}
\frac{d}{dt}\displaystyle\int_{\mathbb{T}^{2}}\psi^{\xi}gdx_1dx_2&=\frac{d}{dt}\displaystyle\int_{\mathbb{T}^n}\psi g\circ\xi dx_1^n\\
&=\displaystyle\int_{\mathbb{T}^n}{\rm div}[(\nabla V-\sum_{i=1}^2\partial_iA_t\circ\xi\nabla\xi_i)\psi+\nabla\psi]g\circ\xi dx_1^n\\
&=-\displaystyle\int_{\mathbb{T}^n}\sum_{j=1}^2[(\nabla V-\sum_{i=1}^2\partial_iA_t\circ\xi\nabla\xi_i)\psi+\nabla\psi].\nabla\xi_j\partial_jg\circ\xi dx_1^n\\
&=-\sum_{i=1}^2\int_{\mathbb{T}^n}[(\nabla V.\nabla\xi_i\psi+\nabla\psi.\nabla\xi_i]\partial_ig\circ\xi dx_1^n\\
&\quad+\sum_{i=1}^2\int_{\mathbb{T}^n}\partial_iA_t\circ\xi\psi\partial_ig\circ\xi dx_1^n.
\end{align*}
Applying Fubini, it holds:
\begin{align*}
\frac{d}{dt}\displaystyle\int_{\mathbb{T}^{2}}\psi^{\xi}gdx_1dx_2&=-\displaystyle\sum_{i=1}^2\int_{\mathbb{T}^{2}}\int_{\Sigma_{(x_1,x_2)}}[\partial_i V\psi+\partial_i\psi]dx_3^n\partial_ig(x_1,x_2) dx_1dx_2\\
&\quad +\displaystyle\sum_{i=1}^2\int_{\mathbb{T}^{2}}\int_{\Sigma_{(x_1,x_2)}}\partial_iA_t(x_1,x_2)\psi dx_3^n\partial_ig(x_1,x_2) dx_1dx_2\\
&=-\displaystyle\sum_{i=1}^2\int_{\mathbb{T}^{2}} F_t^i\psi^{\xi}\partial_ig(x_1,x_2) dx_1dx_2-\displaystyle\sum_{i=1}^2\int_{\mathbb{T}^{2}}\partial_i\psi^{\xi}\partial_ig(x_1,x_2) dx_1dx_2\\
&\quad +\displaystyle\sum_{i=1}^2\int_{\mathbb{T}^{2}}\partial_iA_t(x_1,x_2)\psi^{\xi}\partial_ig(x_1,x_2) dx_1dx_2\\
&=\displaystyle\int_{\mathbb{T}^{2}}\Delta\psi^{\xi}g(x_1,x_2) dx_1dx_2,
\end{align*}
where we used \eqref{sca} with $\varphi=\psi^{\xi}(t,.)$. This is the weak formulation of:
$$\partial_t\psi^{\xi}=\Delta\psi^{\xi},\, \text{on}\,[0,\infty[\times \mathbb{T}^2.$$
\begin{remark}\label{simpdi}
The reason why we consider the weighted Helmholtz decomposition~\eqref{hel} with $\varphi=\psi^{\xi}(t,.)$ in the PABF dynamics \eqref{pabf} instead of the standard one~\eqref{helm0} is precisely to obtain this simple diffusion equation on the function $\psi^{\xi}$. This is will also be useful in the proof of Lemma~\ref{neg} below.
\end{remark}
\subsection{\textit{Proof of Corollary~\ref{fishco}}}
Let $\phi=\psi^{\xi}$ and $\phi_{\infty}=\psi^{\xi}_{\infty}=1$. It is known that $\forall t\geq0$ and $\forall (x_1,x_2)\in \mathbb{T}^2$, $\phi$ satisfies:
\begin{equation}\label{diff2}
\partial_t\phi=\Delta_{x_1^2}\phi.
\end{equation}
Moreover (See Remark~\ref{rmdif}), it is assumed that
and is such that $$\displaystyle\int_{\mathbb{T}^2}\phi(0,.)=1 \,\mbox{and}\,\phi(0,.)\geq0.$$
Let us show that $\forall\,t\geq 0,\forall k>0,\,\|\phi(t,.)-1\|_{H^k({\mathbb{T}^2})}\leq \|\phi(0,.)-1\|_{H^k({\mathbb{T}^2})}e^{-8\pi^2 t}.$
First, we prove that $\phi$ converges to $1$ in $L^2({\mathbb{T}^2})$,
\begin{align*}
\frac{1}{2}\frac{d}{dt}\displaystyle\int_{\mathbb{T}^2} | \phi-1|^2&=\displaystyle\int_{\mathbb{T}^2} \partial_t\phi(\phi-1)\\
&=\displaystyle\int_{\mathbb{T}^2}\Delta\phi(\phi-1)\\
&=-\displaystyle\int_{\mathbb{T}^2}\nabla\phi\nabla(\phi-1)\\
&=-\displaystyle\int_{\mathbb{T}^2}|\nabla\phi|^2\\
&\leq -4\pi^2\displaystyle\int_{\mathbb{T}^2}|\phi-1|^2,
\end{align*}
where we have used the Poincar\'e-Wirtinger inequality on the torus $\mathbb{T}^2$, applied to $\phi$: for any function $f\in H^1(\mathbb{T}^2)$,
$$\displaystyle \int_{\mathbb{T}^2}\left(f-\int_{\mathbb{T}^2}f\right)^2\leq \frac{1}{4\pi^2}\int_{\mathbb{T}^2}|\nabla f|^2.$$
We therefore obtain, $\|\phi(t,.)-1\|^2_{L^2(\mathbb{T}^2)}\leq\|\phi(0,.)-1\|^2_{L^2(\mathbb{T}^2)}e^{-8\pi^2t}$.
Second, we prove that $\partial_i\phi$ converges to $0$ in $L^2({\mathbb{T}^2})$. For $i=1,2$, $\partial_i\phi$ satisfies~\eqref{diff2}: $\partial_t(\partial_i\phi)=\Delta_{x_1^2}(\partial_i\phi)$, with periodic boundary conditions. As above,
\begin{align*}
\frac{1}{2}\frac{d}{dt}\displaystyle\int_{\mathbb{T}^2} | \partial_i\phi|^2&=\displaystyle\int_{\mathbb{T}^2} \partial_t(\partial_i\phi)\partial_i\phi\\
&=\displaystyle\int_{\mathbb{T}^2}\Delta(\partial_i\phi)\partial_i\phi\\
&=-\displaystyle\int_{\mathbb{T}^2}|\nabla(\partial_i\phi)|^2.
\end{align*}
Using again Poincar\'e-Wirtinger inequality on $\partial_i\phi$,
\begin{align*}
\frac{1}{2}\frac{d}{dt}\displaystyle\int_{\mathbb{T}^2} | \partial_i\phi|^2&\leq-4\pi^2\displaystyle\int_{\mathbb{T}^2}\left(\partial_i\phi-\displaystyle\int_{\mathbb{T}^2}\partial_i\phi\right)^2\\
&=-4\pi^2\displaystyle\int_{\mathbb{T}^2}|\partial_i\phi|^2.
\end{align*}
Where we used $\int_{\mathbb{T}^2}\partial_i\phi=0$, since $\phi$ is periodic on $\mathbb{T}^2$. Therefore, it holds $\|\partial_i\phi(t,.)\|^2_{L^2(\mathbb{T}^2)}\leq\|\partial_i\phi(0,.)\|^2_{L^2(\mathbb{T}^2)}e^{-8\pi^2t}$.
Third, one can prove by induction that all higher derivatives of $\phi$ converge exponentially fast to $0$, with rate $8\pi^2$ and the following estimation is then proven:
$$\forall\,t\geq 0,\forall k>0,\quad\|\phi(t,.)-1\|^2_{H^k({\mathbb{T}^2})}\leq \|\phi(0,.)-1\|_{H^k({\mathbb{T}^2})}^2e^{-8\pi^2 t}.$$
As $\displaystyle H^k(\mathbb{T}^2)\hookrightarrow L^{\infty}(\mathbb{T}^2),\,\forall k>1$, then $\exists c>0$,
$$\|\phi-1\|^2_{ L^{\infty}}\leq c\|\phi-1\|^2_{H^k}\leq ce^{-8\pi^2 t}.$$
Therefore, $\forall \varepsilon>0$, $\exists t_0>0$, $\forall x\in \mathbb{T}^2$, $\forall t>t_0$, $\phi(t,x)\geq 1-\varepsilon$. Finally, $\forall t>t_0$\\
$I(\psi^{\xi}|\psi^{\xi}_{\infty})=\displaystyle\int_{\mathbb{T}^2}\frac{|\nabla_{x_1^2}\phi|^2}{\phi}\leq\frac{1}{1-\varepsilon}\displaystyle\int_{\mathbb{T}^2}|\nabla_{x_1^2}\phi|^2\leq \frac{\|\nabla_{x_1^2}\phi(0,.)\|^2_{L^2(\mathbb{T}^2)}}{1-\varepsilon}e^{-8\pi^2t}.$
\subsection{\textit{Proof of Corollary \ref{mic}}}
We have that $\psi^{\xi}_{\infty}=1$ satisfies $LSI(r)$, for some $r>0$ (see Chapter~3, Section~3 in [\refcite{ane:00}]). Referring to Proposition~\ref{diff} and since $\psi^{\xi}$ is a probability density function, one gets:
\begin{align*}
\frac{d}{dt}E_M&=\displaystyle\int_{\mathbb{T}^2}\partial_t\left(\psi^{\xi}\ln(\psi^{\xi})\right)\\
&=\displaystyle\int_{\mathbb{T}^2}\partial_t\psi^{\xi}\ln(\psi^{\xi})+\int_{\mathbb{T}^2}\partial_t\psi^{\xi}\\
&=\int_{\mathbb{T}^2}\Delta\psi^{\xi}\ln(\psi^{\xi})\\
&=-\int_{\mathbb{T}^2}|\nabla_{x_1^2}\ln(\psi^{\xi})|^2\psi^{\xi}\\
&=-I(\psi^{\xi}|\psi^{\xi}_{\infty})\\
&\leq-2r H(\psi^{\xi}|\psi^{\xi}_{\infty})\\
&=-2r E_M.
\end{align*}
Therefore, $E_M$ converges exponentially fast to zero. Referring to Corollary~\ref{fishco} and since $E_M$ converges to zero, we have that for any $t>t_0$,
\begin{align*}
\displaystyle -E_M(t)=\int_t^{\infty}\frac{d}{ds}E_M(s)ds&=-\int_t^{\infty}I(\psi^{\xi}|\psi^{\xi}_{\infty})ds\\
&\geq -I_0\int_t^{\infty}\mathrm{e}^{-8\pi^2s}ds\\
&=-\frac{I_0}{8\pi^2}\mathrm{e}^{-8\pi^2t},
\end{align*}
which yields the desired estimation.
\subsection{\textit{Proof of Theorem \ref{theo1}}}
To prove our main result, several intermediate lemmas are needed.
\begin{lemma}\label{dif}
$\forall t\geq 0$, $\forall (x_1,x_2)\in \mathbb{T}^2$ and for $i=1,2$, we have:
$$(F_t^i-\partial_iA)(x_1,x_2)=\displaystyle\left(\int_{\Sigma_{(x_1,x_2)}}\partial_i\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}dx_3^n\right)(x_1,x_2)-\left(\partial_i\ln(\psi^{\xi}/\psi^{\xi}_{\infty})\right)(x_1,x_2).$$
\end{lemma}
\begin{proof}
\begin{align*}
&\displaystyle\int_{\Sigma_{(x_1,x_2)}}\partial_i\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}dx_3^n-\partial_i\ln(\psi^{\xi}/\psi^{\xi}_{\infty})\\
&=\displaystyle\int_{\Sigma_{(x_1,x_2)}}\partial_i\ln(\psi)\frac{\psi}{\psi^{\xi}}dx_3^n-\int_{\Sigma_{(x_1,x_2)}}\partial_i\ln(\psi_{\infty})\frac{\psi}{\psi^{\xi}}dx_3^n-\partial_i\ln(\psi^{\xi})+\partial_i\ln(\psi^{\xi}_{\infty})\\
&=\displaystyle\frac{1}{\psi^{\xi}}\int_{\Sigma_{(x_1,x_2)}}\partial_i\psi dx_3^n+\int_{\Sigma_{(x_1,x_2)}}(\partial_i V-\nabla\xi_i\partial_iA\circ\xi)\frac{\psi}{\psi^{\xi}}dx_3^n-\partial_i\ln(\psi^{\xi})\\
&=\displaystyle\frac{\partial_i\psi^{\xi}}{\psi^{\xi}}+F_t^i-\partial_iA-\frac{\partial_i\psi^{\xi}}{\psi^{\xi}}\\
&=\displaystyle F_t^i-\partial_iA.
\end{align*}
\end{proof}
\begin{lemma}\label{difabs}
Suppose that [H1] and [H2] hold, then for all $t\geq0$, for all $(x_1,x_2)\in {\mathbb{T}^2}$ and for $i=1,2$, we have:
$$\displaystyle |F_t^i(x_1,x_2)-\partial_iA(x_1,x_2)|\leq \gamma\sqrt{\frac{2}{\rho}e_m(t,x_1,x_2)}.$$
\end{lemma}
\begin{proof}
For any coupling measure $\pi\displaystyle\in\prod(\mu_{t,x_1,x_2},\mu_{\infty,x_1,x_2})$ defined on $\Sigma_{(x_1,x_2)}\times\Sigma_{(x_1,x_2)}$, it holds:
\begin{align*}
\displaystyle |F_t^i-\partial_iA|&=\left|\displaystyle\int_{\Sigma_{(x_1,x_2)}\times\Sigma_{(x_1,x_2)}}(\partial_iV(x)-\partial_iV(x'))\pi(dx,dx')\right|\\
&=\|\nabla_{x_3^n}\partial_iV\|_{L^{\infty}}\sqrt{\displaystyle\int_{\Sigma_{(x_1,x_2)}\times\Sigma_{(x_1,x_2)}}d_{\Sigma_{(x_1,x_2)}}(x,x')^2\pi(dx,dx')}.
\end{align*}
Taking now the infimum over all $\pi\displaystyle\in\prod(\mu(t,.|(x_1,x_2)),\mu(\infty,.|(x_1,x_2)))$ and using Lemma \ref{tala}, we obtain
\begin{align*}
\displaystyle |F_t^i-\partial_iA|&\leq\displaystyle \gamma W(\mu(t,.|(x_1,x_2)),\mu^{\xi}(\infty,.|(x_1,x_2)))\\
&\leq\displaystyle \gamma\sqrt{\frac{2}{\rho}H(\mu^{\xi}(t,.|(x_1,x_2)),\mu^{\xi}(\infty,.|(x_1,x_2)))}\\
&=\displaystyle \gamma\sqrt{\frac{2}{\rho}e_m(t,(x_1,x_2))}.\\
\end{align*}
\end{proof}
\begin{lemma}\label{em}
Suppose that [H2] holds, then for all $t\geq0$,
$$E_m(t)\leq\displaystyle\frac{1}{2\rho}\int_{\mathbb{T}^n}|\nabla_{x_3^n}\ln(\psi(t,.)/\psi_{\infty})|^2\psi.$$
\end{lemma}
\begin{proof}
Using [H2],
\begin{align*}
E_m&=\displaystyle\int_{\mathbb{T}^2}e_m\psi^{\xi}dx_1dx_2\\
&=\displaystyle\int_{\mathbb{T}^2}H(\mu(t,.|(x_1,x_2))|\mu(\infty,.|(x_1,x_2)))\psi^{\xi}dx_1dx_2\\
&\leq \displaystyle\int_{\mathbb{T}^2}\frac{1}{2\rho}\int_{\Sigma_{x_1^2}}|\nabla_{x_3^n}\ln(\psi(t,.)/\psi_{\infty})|^2dx_3^n\frac{\psi(t,.)}{\psi^{\xi}(t,x_1,x_2)}dx_1dx_2\\
&=\displaystyle\frac{1}{2\rho}\int_{\mathbb{T}^n}|\nabla_{x_3^n}\ln(\psi(t,.)/\psi_{\infty})|^2\psi dx_1^n.\\
\end{align*}
\end{proof}
\begin{lemma}\label{neg}
It holds for all $t\geq0$,
$$\displaystyle\int_{\mathbb{T}^n}(\partial_1A_t-F_t^1)[\partial_1\ln(\psi/\psi_{\infty})]\psi+\int_{\mathbb{T}^n}(\partial_2A_t-F_t^2)[\partial_2\ln(\psi/\psi_{\infty})]\psi\leq 0.$$
\end{lemma}
\begin{proof}
Using Fubini,
\begin{align*}
&\displaystyle\int_{\mathbb{T}^n}(\partial_1A_t-F_t^1)[\partial_1\ln(\psi/\psi_{\infty})]\psi+\int_{\mathbb{T}^n}(\partial_2A_t-F_t^2)[\partial_2\ln(\psi/\psi_{\infty})]\psi\\
&=\displaystyle\int_{\mathbb{T}^2}(\partial_1A_t-F_t^1)\int_{\Sigma_{(x_1,x_2)}}[\partial_1\ln(\psi/\psi_{\infty})]\psi+\int_{\mathbb{T}^2}(\partial_2A_t-F_t^2)\int_{\Sigma_{(x_1,x_2)}}[\partial_2\ln(\psi/\psi_{\infty})]\psi.
\end{align*}
For the first term, we have
\begin{align*}
\displaystyle\int_{\Sigma_{(x_1,x_2)}}[\partial_1\ln(\psi/\psi_{\infty})]\psi&=\displaystyle\int_{\Sigma_{(x_1,x_2)}}(\partial_1\ln\psi)\psi-\int_{\Sigma_{(x_1,x_2)}}(\partial_1\ln\psi_{\infty})\psi\\
&=\partial_1\psi^{\xi}+\int_{\Sigma_{(x_1,x_2)}}\partial_1(V-A)\psi\\
&=(\partial_1\ln\psi^{\xi})\psi^{\xi}+F_t^1\psi^{\xi}-\partial_1A\psi^{\xi}.
\end{align*}
Similarly, we have
$$\displaystyle\int_{\Sigma_{(x_1,x_2)}}[\partial_2\ln(\psi/\psi_{\infty})]\psi=(\partial_2\ln\psi^{\xi})\psi^{\xi}+F_t^2\psi^{\xi}-\partial_2A\psi^{\xi}.$$
Therefore, one gets
\begin{align*}
&\displaystyle\int_{\mathbb{T}^n}(\partial_1A_t-F_t^1)[\partial_1\ln(\psi/\psi_{\infty})]\psi+\int_{\mathbb{T}^n}(\partial_2A_t-F_t^2)[\partial_2\ln(\psi/\psi_{\infty})]\psi\\
&=\displaystyle\int_{\mathbb{T}^2}(\partial_1A_t-F_t^1)(\partial_1\ln\psi^{\xi})\psi^{\xi}+\int_{\mathbb{T}^2}(\partial_2A_t-F_t^2)(\partial_2\ln\psi^{\xi})\psi^{\xi}\\
&\quad -\displaystyle\int_{\mathbb{T}^2}(\partial_1A_t-F_t^1)^2\psi^{\xi}-\int_{\mathbb{T}^2}(\partial_2A_t-F_t^2)^2\psi^{\xi}
\end{align*}
which concludes the assertion since the first line is equal to zero (by \eqref{sca} with $\varphi=\psi^{\xi}(t,.)$) and the second line is non positive. Again the weighted Helmholtz decomposition helps in simplifying terms.
\end{proof}
\noindent\textbf{Proof of Theorem~\ref{theo1}:}\\
Now we will prove the exponentially convergence of $E_m(t)$. Recall \eqref{reform}:
$$\partial_t\psi=\displaystyle{\rm div}(\nabla V\psi+\nabla\psi)-\partial_1((\partial_1A_t)\psi)-\partial_2((\partial_2A_t)\psi),$$
which is equivalent to
$$\partial_t\psi=\displaystyle{\rm div}\left(\psi_{\infty}\nabla\left(\frac{\psi}{\psi_{\infty}}\right)\right)+\partial_1[(\partial_1A-\partial_1A_t)\psi]+\partial_2[(\partial_2A-\partial_2A_t)\psi].$$
Using \eqref{tent}, \eqref{maent} and \eqref{diff2}, one obtains
\begin{align*}
\displaystyle\frac{dE}{dt}&=-\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi +\int_{\mathbb{T}^n}\!(\partial_1A_t-\partial_1A)[\partial_1\ln(\psi/\psi_{\infty})]\psi\\
&\quad+ \displaystyle\int_{\mathbb{T}^n}\!(\partial_2A_t-\partial_2A)[\partial_2\ln(\psi/\psi_{\infty})]\psi,\\
\displaystyle\frac{dE_M}{dt}&=-\displaystyle\int_{\mathbb{T}^2}|\nabla\ln(\psi^{\xi})|^2\psi^{\xi}.
\end{align*}
\noindent Using then Lemma \ref{entsum} and Lemma \ref{dif}, one gets
\begin{align*}
\displaystyle\frac{dE_m}{dt}&=\displaystyle\frac{dE}{dt}-\displaystyle\frac{dE_{M}}{dt}\\
&=-\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi+\int_{\mathbb{T}^n}\!(\partial_1A_t-\partial_1A)\partial_1\ln(\psi/\psi_{\infty})\psi\\
&\quad+ \displaystyle\int_{\mathbb{T}^n}\!(\partial_2A_t-\partial_2A)\partial_2\ln(\psi/\psi_{\infty})\psi +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}\\
&=-\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi\\
&\quad \displaystyle+\int_{\mathbb{T}^n}\!(\partial_1A_t-F_t^1)[\partial_1\ln(\psi/\psi_{\infty})]\psi+ \int_{\mathbb{T}^n}\!(F_t^1-\partial_1A)[\partial_1\ln(\psi/\psi_{\infty})]\psi\\
&\quad \displaystyle+\int_{\mathbb{T}^n}\!(\partial_2A_t-F_t^2)[\partial_2\ln(\psi/\psi_{\infty})]\psi+ \int_{\mathbb{T}^n}\!(F_t^2-\partial_2A)[\partial_2\ln(\psi/\psi_{\infty})]\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}.
\end{align*}
\noindent Lemma \ref{neg} then yields
\begin{align*}
\displaystyle\frac{dE_m}{dt}&\leq-\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi\\
&\quad \displaystyle+\int_{\mathbb{T}^n}(F_t^1-\partial_1A)\partial_1\ln(\psi/\psi_{\infty})\psi+\int_{\mathbb{T}^n}\!(F_t^2-\partial_2A)\partial_2\ln(\psi/\psi_{\infty})\psi \\
&\quad +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}.
\end{align*}
Using lemma \ref{dif} and Fubini, one then obtains
\begin{align*}
\displaystyle\frac{dE_m}{dt}&\leq-\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}\left[\int_{\Sigma_{(x_1,x_2)}}\!\!\!\!\!\!\!\!\!\!\partial_1\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}\right]\int_{\Sigma_{(x_1,x_2)}}\!\!\!\!\!\!\!\!\!\!\partial_1\ln(\psi/\psi_{\infty})\psi
-\displaystyle\int_{\mathbb{T}^n}\!\!\!\!\partial_1\ln(\psi^{\xi})\partial_1\ln(\psi/\psi_{\infty})\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}\left[\int_{\Sigma_{(x_1,x_2)}}\!\!\!\!\!\!\!\!\!\!\partial_2\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}\right]\int_{\Sigma_{(x_1,x_2)}}\!\!\!\!\!\!\!\!\!\!\partial_2\ln(\psi/\psi_{\infty})\psi
-\displaystyle\int_{\mathbb{T}^n}\!\!\!\!\partial_2\ln(\psi^{\xi})\partial_2\ln(\psi/\psi_{\infty})\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}\\
&\leq -\displaystyle\int_{\mathbb{T}^n}|\nabla\ln(\psi/\psi_{\infty})|^2\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}\left[\int_{\Sigma_{(x_1,x_2)}}\partial_1\ln(\psi/\psi_{\infty})\psi\right]^2\frac{1}{\psi^{\xi}}-\displaystyle\int_{\mathbb{T}^n}\partial_1\ln(\psi^{\xi})\partial_1\ln(\psi/\psi_{\infty})\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}\left[\int_{\Sigma_{(x_1,x_2)}}\partial_2\ln(\psi/\psi_{\infty})\psi\right]^2\frac{1}{\psi^{\xi}}-\displaystyle\int_{\mathbb{T}^n}\partial_2\ln(\psi^{\xi})\partial_2\ln(\psi/\psi_{\infty})\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}.
\end{align*}
Applying Cauchy-Schwarz on the first terms of the second and third lines, we obtain
\begin{align*}
\displaystyle\frac{dE_m}{dt}&\leq -\displaystyle\int_{\mathbb{T}^n}|\nabla_{x_3^n}\ln(\psi/\psi_{\infty})|^2\psi\\
&\quad -\displaystyle\int_{\mathbb{T}^n}\partial_1\ln(\psi^{\xi})\partial_1\ln(\psi/\psi_{\infty})\psi-\int_{\mathbb{T}^n}\partial_2\ln(\psi^{\xi})\partial_2\ln(\psi/\psi_{\infty})\psi\\
&\quad +\displaystyle\int_{\mathbb{T}^2}|\partial_1\ln\psi^{\xi}|^2\psi^{\xi}+\int_{\mathbb{T}^2}|\partial_2\ln\psi^{\xi}|^2\psi^{\xi}\\
&\leq-\displaystyle\int_{\mathbb{T}^n}|\nabla_{x_3^n}\ln(\psi/\psi_{\infty})|^2\psi-\displaystyle\int_{\mathbb{T}^2}\partial_1\ln(\psi^{\xi})\left[\int_{\Sigma_{(x_1,x_2)}}\!\!\!\!\!\!\!\!\!\!\partial_1\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}-\partial_1\ln\psi^{\xi}\right]\psi^{\xi}\\
&\quad -\displaystyle\int_{\mathbb{T}^2}\partial_2\ln(\psi^{\xi})\left[\int_{\Sigma_{(x_1,x_2)}}\partial_2\ln(\psi/\psi_{\infty})\frac{\psi}{\psi^{\xi}}-\partial_2\ln\psi^{\xi}\right]\psi^{\xi}.
\end{align*}
Applying Lemma~\ref{em}, Lemma~\ref{dif}, Cauchy-Schwarz, Lemma~\ref{difabs} and Corollary~\ref{fishco},
\begin{align*}
\displaystyle\frac{dE_m}{dt}&\leq -2\rho E_m+\displaystyle\sqrt{\int_{\mathbb{T}^2}|\partial_1\ln(\psi^{\xi})|^2\psi^{\xi}}\sqrt{\int_{\mathbb{T}^2}\frac{2}{\rho}e_m(t,(x_1,x_2))\psi^{\xi}}\\
&\quad +\displaystyle\sqrt{\int_{\mathbb{T}^2}|\partial_2\ln(\psi^{\xi})|^2\psi^{\xi}}\sqrt{\int_{\mathbb{T}^2}\frac{2}{\rho}e_m(t,(x_1,x_2))\psi^{\xi}}\\
&\leq \displaystyle-2\rho E_m+2\gamma\sqrt{\frac{2}{\rho}E_m}\sqrt{\int_{\mathbb{T}^2}|\nabla_{x_1^2}\ln(\psi^{\xi})|^2\psi^{\xi}}\\
&\leq\displaystyle-2\rho E_m+2\gamma\sqrt{\frac{2}{\rho}E_m}\sqrt{I(\psi^{\xi}/\psi^{\xi}_{\infty})}\\
&\leq\displaystyle-2\rho E_m+2\gamma\sqrt{\frac{2}{\rho}E_m}\sqrt{I_0}e^{-4\pi^2t}.
\end{align*}
Finally we obtain
\begin{align*}
\frac{d}{dt}\sqrt{E_m(t)}&\leq\displaystyle-\rho\sqrt{E_m(t)}+\gamma\sqrt{\frac{I_0}{2\rho}}e^{-4\pi^2t}.
\end{align*}
First, if $\rho\neq 4\pi^2$, using Gronwall inequality, one obtains
\begin{align*}
\sqrt{E_m(t)}&\leq\sqrt{E_m(0)}\,\displaystyle e^{-\rho t}+\gamma\sqrt{\frac{I_0}{2\rho}}\displaystyle\int_0^te^{\rho(-t+s)}e^{-4\pi^2s}ds\\
&\leq \displaystyle\sqrt{E_m(0)}\,e^{-\rho t}+\gamma\sqrt{\frac{I_0}{2\rho}}\frac{e^{-\rho t}}{\rho-4\pi^2}\left(e^{(\rho-4\pi^2)t}-1\right)\\
&\leq \displaystyle\sqrt{E_m(0)}\,e^{-\rho t}+\gamma\sqrt{\frac{I_0}{2\rho}}\frac{e^{-\rho t}}{|\rho-4\pi^2|}e^{(\rho-4\pi^2)t}.
\end{align*}
\begin{align*}
\sqrt{E_m(t)}&\leq\displaystyle\sqrt{E_m(0)}\,e^{-\rho t}+\gamma\sqrt{\frac{I_0}{2\rho}}\displaystyle\int_0^te^{-\rho t}ds\\
&\leq \displaystyle\left(\sqrt{E_m(0)}+\gamma\sqrt{\frac{I_0}{2\rho}}t\right)e^{-\rho t},
\end{align*}
which leads the desired estimation \eqref{Em}.
Using this above convergence, Corollary \ref{mic} and Lemma \ref{entsum}, it is then easy to see that $E$ converges exponentially fast to zero. Using \eqref{kul}, one obtains the convergence of $\psi$ to $\psi_{\infty}$ since:
$$\|\psi-\psi_{\infty}\|_{L^1({\mathbb{T}^n})}\leq\sqrt{2H(\psi|\psi_{\infty})}=\sqrt{2E}.$$
The second point of the theorem is checked. Finally, we are now in position to prove the last point of Theorem \ref{theo1}. Using \eqref{minw} and Lemma \ref{difabs},
\begin{align*}
\|\nabla A_t-\nabla A\|_{L^2_{\psi^{\xi}}(\mathbb{T}^2)}^2&\leq 2\|\nabla A_t-F_t\|_{L^2_{\psi^{\xi}}(\mathbb{T}^2)}^2+2\|F_t-\nabla A\|_{L^2_{\psi^{\xi}}(\mathbb{T}^2)}^2\\
&\leq 4\|F_t-\nabla A\|_{L^2_{\psi^{\xi}}(\mathbb{T}^2)}^2\\
&\leq 8\frac{\gamma^2}{\rho}E_m.
\end{align*}
\section*{Acknowledgment}
The authors thank Gabriel Stoltz for very helpful discussions on the numerical experiments. Houssam Alrachid would like to thank the Ecole des Ponts ParisTech and CNRS Libanais for supporting his PHD thesis. The work of Tony Leli\`evre is supported by the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement number 614492.
|
\section{Introduction}
The $L(h,k)$-labeling problem (or $(h,k)$-coloring problem) is that of vertex labeling of an undirected graph $G$ with non-negative integers such that for every $u,v \in V(G)$, $uv \in E(G)$, $\left|l(u)-l(v)\right|\geq h$ and for all $u,v\in V(G), d(u,v)=2$, $\left|l(u)-l(v)\right|\geq k$.
The difference between the largest label and the smallest label assigned is called the span. The aim of $L(h,k)-$labeling is to obtain the smallest non negative integer $\lambda_h^k(G),$ such that there exists an $L(h,k)$-labeling of $G$ with no label on $V(G)$ greater than $\lambda_h^k(G)$.
Motivated by Hales' 1980 paper \cite{H11}, which provided a new model for frequency assignment problems as a graph coloring problem, Griggs and Yeh \cite{GY10} formulated the $L(2,1)$ problem to model the channel assignment problem. The general notion of $L(h,k)$- labeling was first presented by Georges and Mauro \cite{GM7} in 1995. The topic has since then been an object of extensive research for various graphs. Calamonerri's survey paper \cite{C2} contains known results on $L(h,k)$-labeling of graphs.
$L(1,1)$-labeling (or strong labeling condition) of a graph is a labeling of $G$ such that vertices with a common neighbor are assigned distinct labels. The usual labeling (or proper vertex coloring) condition is that adjacent vertices have different colors, but for $L(1,1)$, also all neighbors of any vertex are colored differently. This is equivalent to a proper vertex-coloring of the square of a graph $G$. Note that a proper $k$- coloring of a graph is a mapping $\alpha: V(G) \rightarrow \{1, \cdots, k \}$ such that for all $uv \in E(G)$ $\alpha(u) \neq \alpha(v) $ and the square $G^2$ of $G$ has vertex $V(G)$ with an edge between two vertices which are adjacent in $G$ or have a common neighbor in $G$. The chromatic number $\chi(G)$ of $G$ is the smallest $k$ for which $G$ admits a $k$-coloring. Therefore, $\chi(G^2)=\lambda_1^1(G)+1$ for a graph $G$.
Labeling of graph powers is often motivated by applications in frequency assignment and has attracted much attention [See for example, \cite{AH}]. $L(1,1)$-labeling has applications in computing approximation to sparse Hessian matrices, design of collision-free multi-hop channel access protocols in radio networks segmentation problem for files in a network and drawings of graphs in the plane \cite{{B88},{KL},{M1},{MMZ}} to mention a few.
For graphs $G$ and $H$, the direct product $G \times H$ have vertex set $V(G) \times V(H)$ where two vertices $(x_1, x_2)$ and $(y_1, y_2)$ are adjacent if and only if $(x_1, y_1) \in E(G)$ and $(x_2, y_2) \in E(H)$. This product is one of the most important graph products with potential applications in engineering, computer science and related disciplines. \cite{JKV13}. The $L(h,k)$- labeling of direct product of graphs was investigated in \cite{{AA1},{CPP3},{HJ},{JKV},{KRS1},{SW1},{SW18},{WP19}}.
In particular, Jha et al \cite{JKV} gave upper bounds for $\lambda_1^k$-labeling of the multiple direct product and cartesian product of cycles with some conditions on $k$ and the length of the cycles. They also presented some cases where we have exact values. In addition by using backtracking algorithm, they computed $\lambda_1^d (C_m \times C_n)$ for $2\leq d \leq 4 $ and $4 \leq m,n\leq 10$. Since every $L(2,1)$-labeling is an $L(1,1)$-labeling, then $\lambda_1^1(G) \leq \lambda_1^2(G)$. Therefore, their results for $d=2$ provided upper bounds for $L(1,1)$-labeling of $C_m \times C_n$ for $4 \leq m,n\leq 10$. The only result for $\lambda_1^1$-labeling for direct product of two cycles in the paper is that if $m, n \equiv 0 mod 5$ then $\lambda_1^1(C_m \times C_n) = 4$.
In this paper, we solve the $L(1,1)$-labeling problem for direct product of cycles $C_m$, $C_n$, $m, n \geq 3 $, except for $m \in \left\{16,18,22,26,32,36,46\right\}$, $n \in \left\{14,16,18,26,28,34\right\}$ and for these outstanding cases we conjecture that $\lambda_1^1(C_m \times C_n) =5$.
The paper is organized as follows: We give some preliminaries in Section 2 and obtain the $\lambda_1^1$ labeling numbers for $C_m \times C_n$ for $m\geq 3$ and $n =4 \ {\rm and}\ 6$ and some of their multiples in Section 3. Section 4 deals with labeling of direct product of bigger cycles.
\section{Preliminaries}
Let $G$ be a finite simple undirected graph with at least two vertices. For subgraph $V' \subseteq V(G)$, we denote by $L(V')$ the set of $L(1,1)-$labeling on $V'$ and for a non-negative integer, say, $k$, we take $[k(\epsilon)]$ as the set of even integers and zero in $[k]$ while $[k(o)]$ is the set of odd integers in $[k]$. Suppose further that $v \in V(G)$, we denote $d_v$ as the degree of $v$.
The following results, remarks and definitions are needed in the work.
\begin{thm} \label{a}\cite{IK12} Graph $G\times H$ is connected if and only if $G$ and $H$ are connected and at least one of $G$ and $H$ is non-bipartite.
\end{thm}
\begin{rem}\begin{itemize}
\item[(i)] \label{b} Let $G=C_m \times C_n$, where $m,n$ are even positive integers. Then, $G$= $G_1 \cup G_2$, where $G_1$ and $G_2$ are the connected components of $C_m \times C_n$, where
\begin{center}
$\left\{V(G_1)=u_iv_j: i\in [(m-1)(\epsilon)], j\in [(n-1)(\epsilon)] \; {\rm or} \; \right.$ \ \\ $\left. i \in [(m-1)(o)]; j\in [(n-1)(0)]\right\}$ and
$\left\{V(G_2)=u_iv_j: i\in [(m-1)(\epsilon)], j\in [(n-1)(o)] \;{\rm or} \; \right.$ \ \\ $\left. i \in [(m-1)(o)]; j\in [(n-1)(\epsilon)]\right\}.$
\end{center}
Note that $G_1$ and $G_2$ are isomorphic and it is demonstrated in the graph $C_4 \times C_6$ below\\
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-13,6.5) {Fig. 1: The components of $C_4 \times C_6$};
\node [minimum size=0cm,draw,circle] (0) at (-16,7) {};
\node [minimum size=0cm,draw,circle] (1) at (-14,7) {};
\node [minimum size=0cm,draw,circle] (2) at (-12,7) {};
\node [minimum size=0cm,draw,circle] (3) at (-10,7) {};
\node [minimum size=0cm,draw,circle] (4) at (-15,8) {};
\node [minimum size=0cm,draw,circle] (5) at (-13,8) {};
\node [minimum size=0cm,draw,circle] (6) at (-11,8){};
\node [minimum size=0cm,draw,circle] (7) at (-16,9) {};
\node [minimum size=0cm,draw,circle] (8) at (-14,9) {};1
\node [minimum size=0cm,draw,circle] (9) at (-12,9) {};
\node [minimum size=0cm,draw,circle] (10) at (-10,9) {};
\node [minimum size=0cm,draw,circle] (11) at (-15,10) {};
\node [minimum size=0cm,draw,circle] (12) at (-13,10) {};
\node [minimum size=0cm,draw,circle] (13) at (-11,10) {};
\node [minimum size=0cm,draw,circle] (14) at (-16,11) {};
\node [minimum size=0cm,draw,circle] (15) at (-14,11) {};
\node [minimum size=0cm,draw,circle] (16) at (-12,11) {};
\node [minimum size=0cm,draw,circle] (17) at (-10,11) {};
\node [minimum size=0cm,draw,circle] (0a) at (-15,7) {};
\node [minimum size=0cm,draw,circle] (1a) at (-13,7) {};
\node [minimum size=0cm,draw,circle] (2a) at (-11,7) {};
\node [minimum size=0cm,draw,circle] (3a) at (-16,8) {};
\node [minimum size=0cm,draw,circle] (4a) at (-14,8) {};
\node [minimum size=0cm,draw,circle] (5a) at (-12,8) {};
\node [minimum size=0cm,draw,circle] (6a) at (-10,8) {};
\node [minimum size=0cm,draw,circle] (7a) at (-15,9) {};
\node [minimum size=0cm,draw,circle] (8a) at (-13,9) {};
\node [minimum size=0cm,draw,circle] (9a) at (-11,9) {};
\node [minimum size=0cm,draw,circle] (10a) at (-16,10) {};
\node [minimum size=0cm,draw,circle] (11a) at (-14,10) {};
\node [minimum size=0cm,draw,circle] (12a) at (-12,10) {};
\node [minimum size=0cm,draw,circle] (13a) at (-10,10) {};
\node [minimum size=0cm,draw,circle] (14a) at (-15,11) {};
\node [minimum size=0cm,draw,circle] (15a) at (-13,11) {};
\node [minimum size=0cm,draw,circle] (16a) at (-11,11) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (3) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (6) to (10);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (10) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [thin=1.00] (13) to (17);
\draw [very thick=3.00] (0a) to (3a);
\draw [very thick=3.00] (0a) to (4a);
\draw [very thick=3.00] (1a) to (4a);
\draw [very thick=3.00] (1a) to (5a);
\draw [very thick=3.00] (2a) to (5a);
\draw [very thick=3.00] (2a) to (6a);
\draw [very thick=15.00] (3a) to (7a);
\draw [very thick=1.00] (4a) to (7a);
\draw [very thick=1.00] (4a) to (8a);
\draw [very thick=1.00] (5a) to (8a);
\draw [very thick=1.00] (5a) to (9a);
\draw [very thick=1.00] (6a) to (9a);
\draw [very thick=1.00] (7a) to (10a);
\draw [very thick=1.00] (7a) to (11a);
\draw [very thick=1.00] (8a) to (11a);
\draw [very thick=1.00] (8a) to (12a);
\draw [very thick=1.00] (9a) to (12a);
\draw [very thick=1.00] (9a) to (13a);
\draw [very thick=1.00] (10a) to (14a);
\draw [very thick=1.00] (11a) to (14a);
\draw [very thick=1.00] (11a) to (15a);
\draw [very thick=1.00] (12a) to (15a);
\draw [very thick=1.00] (12a) to (16a);
\draw [very thick=1.00] (13a) to (16a);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
}}
\item[(ii)] \label{c} Suppose $G=C_m \times C_n$ such that $G=G'\cup G''$, where $G', G''$ are components of $G$, then, $\lambda_1^1(G)=max\left\{\lambda_1^1(G'),\lambda_1^1(G'')\right\}.$
\item[(iii)] \label{d} Let $G=C_m \times C_n$, where $m$ is even and $n$ odd positive integers. Then, $G \equiv G_1 $, where $G_1$ is any of the two connected components of $C_m \times C_{2n}$.
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-13.5,6) {Fig. 2: $C_4 \times C_5$ is isomorphic to a component of $C_4 \times {10}$};
\node [minimum size=0cm,draw,circle] (0) at (-16,7) {${0,0}$};
\node [minimum size=0cm,draw,circle] (1) at (-14,7) {$ {0,2}$};
\node [minimum size=0cm,draw,circle] (2) at (-12,7) {$ {0,4}$};
\node [minimum size=0.65cm,draw,circle] (4) at (-15,8) {};
\node [minimum size=0.65cm,draw,circle] (5) at (-13,8) {};
\node [minimum size=0cm,draw,circle] (6) at (-11,8) {$ {1,0}$};
\node [minimum size=0cm,draw,circle] (7) at (-16,9) {$ {2,0}$};
\node [minimum size=0.65cm,draw,circle] (8) at (-14,9) {};
\node [minimum size=0.65cm,draw,circle] (9) at (-12,9) {};
\node [minimum size=0.65cm,draw,circle] (11) at (-15,10) {};
\node [minimum size=0.65cm,draw,circle] (12) at (-13,10) {};
\node [minimum size=0cm,draw,circle] (13) at (-11,10) {$ {3,0}$};
\node [minimum size=0cm,draw,circle] (14) at (-16,11) {$ {0,0}$};
\node [minimum size=0cm,draw,circle] (15) at (-14,11) {$ {0,2}$};
\node [minimum size=0cm,draw,circle] (16) at (-12,11) {$ {0,4}$};
\node [minimum size=0cm,draw,circle] (0a) at (-15,7) {$ {0,1}$};
\node [minimum size=0cm,draw,circle] (1a) at (-13,7) {$ {0,3}$};
\node [minimum size=0cm,draw,circle] (2a) at (-11,7) {$ {0,0}$};
\node [minimum size=0cm,draw,circle] (3a) at (-16,8) {$ {1,0}$};
\node [minimum size=0.65cm,draw,circle] (4a) at (-14,8) {};
\node [minimum size=0.65cm,draw,circle] (5a) at (-12,8) {};
\node [minimum size=0.65cm,draw,circle] (7a) at (-15,9) {};
\node [minimum size=0.65cm,draw,circle] (8a) at (-13,9) {};
\node [minimum size=0cm,draw,circle] (9a) at (-11,9) {$ {2,0}$};
\node [minimum size=0cm,draw,circle] (10a) at (-16,10) {$ {3,0}$};
\node [minimum size=0.65cm,draw,circle] (11a) at (-14,10) {};
\node [minimum size=0.65cm,draw,circle] (12a) at (-12,10) {};
\node [minimum size=0cm,draw,circle] (14a) at (-15,11) {$ {0,1}$};
\node [minimum size=0cm,draw,circle] (15a) at (-13,11) {$ {0,3}$};
\node [minimum size=0cm,draw,circle] (16a) at (-11,11) {$ {0,0}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [very thick=3.00] (0a) to (3a);
\draw [very thick=3.00] (0a) to (4a);
\draw [very thick=3.00] (1a) to (4a);
\draw [very thick=3.00] (1a) to (5a);
\draw [very thick=3.00] (2a) to (5a);
\draw [very thick=15.00] (3a) to (7a);
\draw [very thick=1.00] (4a) to (7a);
\draw [very thick=1.00] (4a) to (8a);
\draw [very thick=1.00] (5a) to (8a);
\draw [very thick=1.00] (5a) to (9a);
\draw [very thick=1.00] (7a) to (10a);
\draw [very thick=1.00] (7a) to (11a);
\draw [very thick=1.00] (8a) to (11a);
\draw [very thick=1.00] (8a) to (12a);
\draw [very thick=1.00] (9a) to (12a);
\draw [very thick=1.00] (10a) to (14a);
\draw [very thick=1.00] (11a) to (14a);
\draw [very thick=1.00] (11a) to (15a);
\draw [very thick=1.00] (12a) to (15a);
\draw [very thick=1.00] (12a) to (16a);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
}}
\end{itemize}
\end{rem}
Let $P_m$ be a path of length $m-1$. The following results are from \cite{AA1}:
\begin{cor}\label{e} For $m \geq 3$, $\lambda_1^1(P_m \times C_6)=5$
\end{cor}
\begin{lem} \label{f} For $m \geq 5$, $n \geq 9$, $n \not\equiv 0 \mod 5$ $\lambda_1^1(P_m \times C_n) \geq 5$
\end{lem}
A useful lower bound on $L(1,1)$-labeling for any graph $G$ is contained in the following Lemma:
\begin{lem}\cite{GM7} If $G$ is a graph with maximum degree $\triangle$, and $G$ includes a vertex with $\triangle$ neighbors, each of which is of degree $\triangle$, then $\lambda_1^1(G) \geq \triangle$
\end{lem}
From the lemma, we have for $m, n \geq 3$ $\lambda_1^1(C_m \times C_n) = 4 $
\section{Labeling of $C_m \times C_n, n= 4,6$}
In this section, we investigate the $\lambda-$ numbers of graph product $C_m \times C_4$ and $C_m \times C_6$, where $m \geq 3$.
Let $G'$ be the connected component of the product graph under consideration.
\begin{lem} \label{cc1} For $m \geq 4,$ and even, $\lambda_1^1(C_m \times C_4) \geq 5.$
\end{lem}
\begin{proof} Let $G' \subset C_m \times C_4$, where $m \geq 4$ and even. Suppose $V_i, V_{i+1}, V_{i+2} \subset V(G')$. Let $G'_1$ be the subgraph of $G'$ induced by $V_{i+j}$, for all $j \in [2].$ Then, $V(G'_1)$=$\{ u_iv_0, u_iv_2, u_{i+1}v_1, \ \\ u_{i+1}v_2, u_{i+2}v_0, u_{i+2}v_2 \}$. Now it is clear that the diameter of $G'_1 $ is 2. Thus for every pair $v_1,v_2 \in V(G'_1),$ $d(v_1,v_2) \leq 2$. Thus, $l(v_1) \neq l(v_2)$ for all $v_1,v_2 \in V(G'_1).$ Now, $\left|V(G'_1)\right|=6$. Therefore $\lambda_1^1(C_m \times C_4) \geq 5$.
\end{proof}
\begin{rem} \label{cc2} Note that if $G' \subset C_m \times C_4$, $m \geq 4$ with $m$ even and $v_i \in V_i,$ for some $i, \; V_i \subset V(G'),$ such that $l(v_i)=\alpha_i \in [m],$ with $\lambda_1^1(G')=m,$ then $\alpha_1 \notin L\left\{V_{i-2}V_{i-1}V_{i+1}V_{i+2}\right\}.$
\end{rem}
\begin{thm} \label{cc3} $\lambda_1^1(C_4 \times C_4)=7$
\end{thm}
\begin{proof} Let $G' \subset C_4 \times C_4$ and $V_i \subset V(G')$ for each $i \in [3]$. Clearly, $V(G')= \cup^3_{i=0}V_i$. Let $G'_1$ be a subgraph of $G'$ such that $G'_1$ is induced by $V_0,V_1,V_2.$ By the proof of Lemma \ref{cc1}, $\left|V(G'_1)\right|=6$ and suppose $\alpha'_k,\alpha''_k, \in L(V_3)$, then by remark \ref{cc2},$\alpha'_k,\alpha''_k, \notin L(V(G'_1)).$ Thus there exists $\alpha'_k,\alpha''_k, \notin [5]$ such that $\left\{\alpha'_k,\alpha''_k\right\}=L(V_3)$, and $\alpha'_k \neq \alpha''_k$ since $d(v'_3,v''_3)=2$ for $v'_3,v''_3 \in V_3.$ Thus, $\left|L(\cup^3_{i=0}V_i)\right|$=$\left|L(V(G'))\right|=6+2$. Therefore, $\lambda_1^1(C_4 \times C_4)=\lambda_1^1(G')=7.$
\end{proof}
Next we present the necessary and sufficient condition under which $\lambda_1^1(C_m \times C_4)$ is 5.
\begin{thm} \label{cc4} For $ m \geq 4, m$ even, $\lambda_1^1(C_m \times C_4)=5$ if and only if $m \equiv 0 \; mod\;6 $.
\end{thm}
\begin{proof} Let $m=6n,\ n \in \mathbb N$. By Lemma \ref{cc1}, $\lambda_1^1(G) \geq 5$. Therefore, $\lambda_1^1(G')\geq 5$, where $G' \subset C_m \times C_4$. Let $G''$ be the connected component of $C_6 \times C_4$. By Corollary \ref{cc2}, $L(V_0) \cap L(V_1)= \emptyset$, $L(V_1) \cap L(V_2)= \emptyset$, $L(V_2) \cap L(V_0)= \emptyset$. Now, set $L(V_0)=L(V_3)$, $L(V_1)=L(V_4)$,$L(V_2)=L(V_5)$. But $L(V_5) \cap L(V_0)= \emptyset$, $L(V_5) \cap L(V_1)= \emptyset$. Thus, $\lambda_1^1(G'') \leq 5$ and $\lambda_1^1(C_6 \times C_4) = 5$. Thus by re-occurrence along $C_n$ and $C_m$, $m=0\;\mod\;6$ implies $\lambda^1_1(G)=5$.
Conversely, suppose $\lambda^1_1(G)=5.$ Let $G'$ be a connected component of $G=C_m \times C_4, \; m\geq 4, m$ even. Then, $\lambda_1^1(G')=5.$ Now, assume that $m \not\equiv 0 \mod 6 $ , then $m=6n'+2$ or $m=6n'+4$ where $n' \in \mathbb{N} \cup 0$. For $n'=0$, $G=C_4 \times C_4$, for which $\lambda_1^1(G')=7$ by Theorem \ref{cc3}. \\
Case i: For $m=6n'+2, \; n'\in \mathbb{N},$ let $V_0,V_1,V_2$ be subsets of $V(G')$. By Corollary \ref{cc2}, $L(V_0) \cap L(V_1)= \emptyset$, $L(V_1) \cap L(V_2)= \emptyset$ and $L(V_2) \cap L(V_0)= \emptyset$. Now let $G'_1$ be the subgraph of $G'$ induced by $V_0,V_1,V_2$. Since $L(V_3) \cap L(V_2)= \emptyset$ for $V_3 \subset V(G')$ and $\lambda_1^1(G')=5,$ then $L(V_3)=L(V_0).$ Let $V_4 \subset V(G').$ Then $L(V_4) \cap L(V_3)=\emptyset$ and $L(V_4) \cap L(V_2)=\emptyset$. Thus $L(V_4)=L(V_1)$. Let $V_5 \subset V(G').$ Then $L(V_5) \cap L(V_4)=\emptyset$ and $L(V_5) \cap L(V_3)=\emptyset$ and therefore $L(V_4)=L(V_1)$. The scheme continues in such a way that $L(V_i)=L(V_i+3)$ for all $i \in [m-1],$ that is $L(V_0)=L(V_3)=L(V_6)= \cdots =L(V_{6n'})$, $L(V_1)=L(V_4)=L(V_{7})= \cdots =L(V_{6n'+1})$, $L(V_2)=L(V_5)=L(V_8)= \cdots =L(V_0)$. Now, for all $v_a \in V_0$ and $v_b \in V_{6n'},$ $d(v_a,v_b)=2$. For all $v_c \in V_1,$ $v_d \in V_{6n'+1}$, $d(v_c,v_d)=2$ and finally, $L(V_0) \cap L(V_2)=\emptyset.$ Thus a contradiction.\\
Case ii: For $m=6n'+4$, $n' \in \mathbb N$, similar argument as in $m=6n+2$ applies. Thus, $\lambda_1^1(C_m \times C_4)=5$ if and only $m\equiv 0 \;\mod\; 6$
\end{proof}
\begin{cor} \label{cc5} \label{h} Let $m \equiv 0\; mod \;6 $ and $n \equiv 0\; mod \;4$. Then, $\lambda_1^1(C_m \times C_n)=5$
\end{cor}
\begin{proof}The claim follows from Theorem \ref{cc4} and the re-occurrence of the optimal labeling of $C_{6n'} \times C_4$, $n'\in \mathbb N$.
\end{proof}
\begin{cor} \label{cc6}For all $m\not\equiv 0 \mod 6$, $\lambda_1^1(C_m \times C_4) \geq 6.$
\end{cor}
\begin{thm} \label{cc7} For $C_8 \times C_4,$ $\lambda_1^1(C_8 \times C_4)=7$
\end{thm}
\begin{proof} Suppose $G'$ is a connected component of $C_8 \times C_4$ and suppose that $\lambda_1^1(G')=6.$ By Corollary \ref{cc2}, $\left|L\left\{V_0,V_1,V_2\right\}\right|=6$. Likewise, for all $\alpha_k \in L(V_0),$ $\alpha_k \notin L(V_7)$ and $\alpha_k \notin L(V_6)$. Also, for all $\alpha_j \in L(V_1)$, $\alpha_j \notin L(V_7)$. Suppose $L(V_2)=L(V_7)$, then by Corollary \ref{cc2}, if $\alpha_a, \alpha_b \in L(V_2),$ then $\alpha_a,\alpha_b \notin L \left\{V_3,V_4,V_5,V_6\right\}$. Since $\lambda_1^1(G')=6$, then there exists only five members of [6] that labels $V_3,V_4,V_5,V_6$. However, this contradicts Lemma \ref{cc1}. Thus, $L(V_2) \neq L(V_7)$. Now suppose one of $\alpha_a,\alpha_b \in L(V_2),$ say $\alpha_a$, labels some vertex $v_1 \in L(V_7)$, then there exists some $\alpha'_a \in [6]$ such that $\alpha'_a \notin L\left\{V_0,V_1,V_2\right\}$ such that $\alpha'_a = l(v_2) \in V_7$, with $v_1 \neq V_2.$ Now let $\alpha_c, \alpha_d \in L(V_0).$ Suppose $L(V_3)=L(V_0)$. Then by Corollary \ref{cc2}, $\alpha_a,\alpha_b \notin L(V_4,V_5,V_6)$, which contradicts Lemma \ref{cc1} since $\left|L(V_4,V_5,V_6)\right|=6$ and $[6]\backslash 2=5.$ Then, $\alpha'_a \in L(V_3)$ and also also one of $\alpha_a, \alpha'_b \in L(V_3)$. Further, by Corollary \ref{cc2}, $\alpha_a, \alpha'_b \notin L(V_4,V_5,V_6)$. Thus, $\lambda^1_1(G')\geq 7.$ Conversely, $\lambda^1_1(G')\leq 7$ follows directly from re-occurrence of the labeling of $C_4 \times C_4$. Thus, $\lambda^1_1(G')=\lambda^1_1(C_8 \times C_4)\geq 7$.
\end{proof}
The next result focuses on the $\lambda_1^1-$number of $C_m \times C_4$, for $m \geq 9$. Theorem \ref{cc6} have already established the lower bound for $\lambda_1^1-$number of $C_m \times C_4$ to be $6$ if $m$ is not a multiple of $6$. So we only need to label $C_{10} \times C_4$ with [6] such that it combines perfectly with the labeling of $C_6 \times C_4$ with [5] to establish general bound for all cases except when $m=14$ which is dealt with separately.
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-6.5,3.5) {$Fig. \; 3$: $5-L(1,1)$- labeling of $C_4 \times C_6$};
\node [minimum size=0cm,draw,circle] (0) at (-8,4) {$0$};
\node [minimum size=0cm,draw,circle] (1) at (-7,4) {$1$};
\node [minimum size=0cm,draw,circle] (2) at (-6,4) {$2$};
\node [minimum size=0cm,draw,circle] (3) at (-5,4) {$0$};
\node [minimum size=0cm,draw,circle] (4) at (-7.5,4.5) {$2$};
\node [minimum size=0cm,draw,circle] (5) at (-6.5,4.5) {$0$};
\node [minimum size=0cm,draw,circle] (6) at (-5.5,4.5) {$1$};
\node [minimum size=0cm,draw,circle] (7) at (-8,5) {$3$};
\node [minimum size=0cm,draw,circle] (8) at (-7,5) {$4$};1
\node [minimum size=0cm,draw,circle] (9) at (-6,5) {$5$};
\node [minimum size=0cm,draw,circle] (10) at (-5,5) {$3$};
\node [minimum size=0cm,draw,circle] (11) at (-7.5,5.5) {$5$};
\node [minimum size=0cm,draw,circle] (12) at (-6.5,5.5) {$3$};
\node [minimum size=0cm,draw,circle] (13) at (-5.5,5.5) {$4$};
\node [minimum size=0cm,draw,circle] (14) at (-8,6) {$0$};
\node [minimum size=0cm,draw,circle] (15) at (-7,6) {$1$};
\node [minimum size=0cm,draw,circle] (16) at (-6,6) {$2$};
\node [minimum size=0cm,draw,circle] (17) at (-5,6) {$0$};
\node [minimum size=0cm,] at (-0.5,3.5) {$Fig.\;4$: $6-L(1,1)$- labeling of $C_4 \times C_{10}$};
\node [minimum size=0cm,draw,circle] (18) at (-3,4) {$0$};
\node [minimum size=0cm,draw,circle] (19) at (-2,4) {$1$};
\node [minimum size=0cm,draw,circle] (20) at (-1,4) {$5$};
\node [minimum size=0cm,draw,circle] (21) at (0,4) {$1$};
\node [minimum size=0cm,draw,circle] (22) at (1,4) {$2$};
\node [minimum size=0cm,draw,circle] (23) at (2,4) {$0$};
\node [minimum size=0cm,draw,circle] (24) at (-2.5,4.5) {$2$};
\node [minimum size=0cm,draw,circle] (25) at (-1.5,4.5) {$3$};
\node [minimum size=0cm,draw,circle] (26) at (-0.5,4.5) {$4$};
\node [minimum size=0cm,draw,circle] (27) at (0.5,4.5) {$0$};
\node [minimum size=0cm,draw,circle] (28) at (1.5,4.5) {$1$};
\node [minimum size=0cm,draw,circle] (29) at (-3,5) {$3$};
\node [minimum size=0cm,draw,circle] (30) at (-2,5) {$4$};
\node [minimum size=0cm,draw,circle] (31) at (-1,5) {$0$};
\node [minimum size=0cm,draw,circle] (32) at (0,5) {$3$};
\node [minimum size=0cm,draw,circle] (33) at (1,5) {$5$};
\node [minimum size=0cm,draw,circle] (34) at (2,5) {$3$};
\node [minimum size=0cm,draw,circle] (35) at (-2.5,5.5) {$5$};
\node [minimum size=0cm,draw,circle] (36) at (-1.5,5.5) {$6$};
\node [minimum size=0cm,draw,circle] (37) at (-0.5,5.5) {$2$};
\node [minimum size=0cm,draw,circle] (38) at (0.5,5.5) {$6$};
\node [minimum size=0cm,draw,circle] (39) at (1.5,5.5) {$4$};
\node [minimum size=0cm,draw,circle] (40) at (-3,6) {$0$};
\node [minimum size=0cm,draw,circle] (41) at (-2,6) {$1$};
\node [minimum size=0cm,draw,circle] (42) at (-1,6) {$5$};
\node [minimum size=0cm,draw,circle] (43) at (0,6) {$1$};
\node [minimum size=0cm,draw,circle] (44) at (1,6) {$2$};
\node [minimum size=0cm,draw,circle] (45) at (2,6) {$0$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (3) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (6) to (10);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (10) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [thin=1.00] (13) to (17);
\draw [thin=1.00] (18) to (24);
\draw [thin=1.00] (19) to (24);
\draw [thin=1.00] (19) to (25);
\draw [thin=1.00] (20) to (25);
\draw [thin=1.00] (20) to (26);
\draw [thin=1.00] (21) to (26);
\draw [thin=1.00] (21) to (27);
\draw [thin=1.00] (22) to (27);
\draw [thin=1.00] (22) to (28);
\draw [thin=1.00] (23) to (28);
\draw [thin=1.00] (24) to (29);
\draw [thin=1.00] (24) to (30);
\draw [thin=1.00] (25) to (30);
\draw [thin=1.00] (25) to (31);
\draw [thin=1.00] (26) to (31);
\draw [thin=1.00] (26) to (32);
\draw [thin=1.00] (27) to (32);
\draw [thin=1.00] (27) to (33);
\draw [thin=1.00] (28) to (33);
\draw [thin=1.00] (28) to (34);
\draw [thin=1.00] (29) to (35);
\draw [thin=1.00] (30) to (35);
\draw [thin=1.00] (30) to (36);
\draw [thin=1.00] (31) to (36);
\draw [thin=1.00] (31) to (37);
\draw [thin=1.00] (32) to (37);
\draw [thin=1.00] (32) to (38);
\draw [thin=1.00] (33) to (38);
\draw [thin=1.00] (33) to (39);
\draw [thin=1.00] (34) to (39);
\draw [thin=1.00] (35) to (40);
\draw [thin=1.00] (35) to (41);
\draw [thin=1.00] (36) to (41);
\draw [thin=1.00] (36) to (42);
\draw [thin=1.00] (37) to (42);
\draw [thin=1.00] (37) to (43);
\draw [thin=1.00] (38) to (43);
\draw [thin=1.00] (38) to (44);
\draw [thin=1.00] (39) to (44);
\draw [thin=1.00] (39) to (45);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
}}
\begin{thm} \label{cc8} Let $m',m'' \in \mathbb N \cup 0$, with $10m'+6n''$ not a multiple of 6. Then $\lambda_1^1(C_{10m'+6m''} \times C_4)=6.$
\end{thm}
\begin{proof} By Corollary \ref{cc6}, $\lambda_1^1(C_m \times C_n) \geq 6$ for all $m$ not multiple of 6. The claim follows required combinations of Figures $3$ and $4$ above which shows that $\lambda_1^1(C_{10m'+6m''} \times C_4)\leq 6$ for $10m'+6m''$ not a multiple of 6.
\end{proof}
Clearly, every even number $m \geq 10, \; m\neq 14$ can be obtained from $10m'+6m''$ defined above. Therefore, we can conclude that for all $m \geq 9,$ $\lambda_1^1(C_m \times C_n)=6$ for all $m$ that is not a multiple of 6 if we can establish that the $\lambda_1^1-$number of $C_{14} \times C_4$ is $6$. We show this in the next result.
\begin{thm} \label{cc9} $\lambda_1^1(C_{14} \times C_4)=6$
\end{thm}
\begin{proof} $\lambda_1^1(C_{14} \times C_4)=6$
\end{proof}
We have now completely determined the $\lambda_1^1-$numbers of $(C_m \times C_4)$ for all $m \geq 3$. In what follows, we investigate the values of $\lambda_1^1(C_m \times C_6)$.
\begin{prop} \label{cc10} Let $G'$ be a connected component of $C_m \times C_6,$ $m \in \mathbb N, m \geq 3.$ Let $V_i \subseteq V(G'), \;i \in [m-1].$ Then,
$(i)$\;$\lambda_1^1(C_m \times C_6) \geq 5.$
$(ii)$ \;Given $v_a, v_b \in V_i, $ $d(v_a,v_b) \leq 2.$
$(iii)$\;For all $V_i \subseteq V(G'), \; \left|L(V_i)\right|=3$
$(iv)$\;suppose $\alpha_k \in L(V_i),$ then $\alpha_k \notin L(V_{i+2}).$
\end{prop}
\begin{proof}The proof of the claims above are as follows:\\
$(i)$ $C_m \times C_6$ contains $P_m \times C_6$. Now from Corollary \ref{e}, $\lambda_1^1(P_m \times C_6)=5.$ Therefore $\lambda_1^1(C_m \times C_6) \geq 5$.\\
$(ii)$ Let $V_i \subseteq V(G').$ $V_i=\left\{u_iv_j,u_iv_{j+2},u_iv_{j+4}\right\}$, where $j \in \left\{0,1\right\}$. Now, since $C_m$ is a cycle,then $d(u_iv_j, u_iv_{j+4})=2$. Clearly, $d(u_iv_j, u_iv_{j+2})=2$, $d(u_iv_{j+2}, u_iv_{j+4})=2$ and thus the claim.\\
$(iii)$ This is quite obvious.\\
$(iv)$ It is obvious that for all $u_iv_j \in V_i$ and $u_{i+2}v_k \in V_{i+2}$, $d(u_iv_j,u_{i+2}v_k)=2$. Therefore, $L(V_i) \cap L(V_{i+2})= \emptyset.$
\end{proof}
The next result describes a property of $L(1,1)$-labeling of $C_m \times C_6$
\begin{lem} \label{cc12} Let $\alpha_k \in L(V_i)$, $i \in [m-1]$, $V_i \subseteq V(G')$, then $\alpha_k$ labels some vertex $v_{i+1} \in V_{i+1}$. In other words, $L(V_i)$ labels $V_{i+1}$.
\end{lem}
\begin{proof} Let $\left\{u_iv_j,u_iv_{j+2},u_iv_{j+4}\right\}=V_i$ and $\left\{u_{i+1}v_{j+1},u_{i+1}v_{j+3},u_{i+1}v_{j+5}\right\}=V_{i+1}$ Clearly $d(u_iv_j,u_{i+1}v_{j+3})=d(u_iv_{j+2},u_{i+1}v_{j+5})=d(u_iv_{j+4},u_{i+1}v_{j+1})=3$. Therefore, suppose $\alpha_k = l(v_i)$, for some $v_i \in V_i,$ then, there exists some unique $v_{i+k} \in V_{i+1}$ such that $l(v_i)=l(v_{i+k})$,with $\left|(i-(i+k))\right|=3$. (The uniqueness of $v_{i+k}$ results from Proposition \ref{cc10}(b).)
\end{proof}
\begin{cor}\label{cc13} If $L(V_i)=L(V_{i+1})$, then, $L(V_{i+2}) \cap L(V_i) = \emptyset$ and $L(V_i) \cap L(V_{i+3})= \emptyset$.
\end{cor} It is obvious from Proposition \ref{cc10}(d).
\begin{cor}\label{cc14} $\lambda_1^1(C_m \times C_6)=5$ if and only if $m \equiv 0\; mod \;4$
\end{cor}
\begin{proof} Let $m \equiv 0\mod4.$ For $m=4,$ clearly $\lambda_1^1(C_4 \times C_6)=5$, which is obtained from Corollary \ref{cc6}. Now in the case of the general $m \equiv 0\mod 4$, by re-occurrence of the labeling of $C_4 \times C_6$ along $C_m$, it follows that $\lambda_1^1(C_m \times C_6)=5$. Conversely, suppose that $\lambda_1^1(C_m \times C_6)=5$. We show that $m \equiv 0 \mod 4$. Let $L(V_i)=\left\{\alpha_i,\alpha_j,\alpha_k\right\}$. By Proposition \ref{cc10} $(c)$,$\alpha_i \neq \alpha_j \neq \alpha_k \neq \alpha_i,$ that is, $\left|L(V_i)\right|=3$. Suppose $L(V_0)=L(V_1)$ by Lemma \ref{cc12}, then by Corollary \ref{cc13}, $L(V_2)\cap L(V_1) = \emptyset$ and $L(V_3)\cap L(V_1) = \emptyset$. Since $\lambda_1^1(C_m \times C_6)=5,$ Then $L(V_2)=L(V_3)=[5] \backslash L(V_0)$. This scheme continues such that $L(V_0)=L(V_4)=L(V_5)$; $L(V_2)=L(V_6)=L(V_7) \cdots= L(V_0)=L(V_{m-4})=L(V_{m-3})$; and $L(V_0)=L(V_4)=L(V_8)=\cdots =L(V_{4(n)})$, $n \in \mathbb N$, where $4n=(m-1)+1=m$ since $C_m$ is a cycle. Thus $m \equiv 0\mod4.$
\end{proof}
The implication of the last result is that the lower bound for the $\lambda_1^1-$number of graph product $C_m \times C_6$, $m \geq 3$ is 6 except for when $m \equiv 0 \mod 4$, in which case the optimal $ \lambda_1^1-$number reduces by $1$.
Now we consider particular cases where the lower bound is strictly greater than 6.
\begin{thm} \label{cc15}$\lambda_1^1(C_6 \times C_6)=8$
\end{thm}
\begin{proof} Suppose that $\lambda_1^1(C_6 \times C_6)=7$. Let $\left\{V_i\right\} \subseteq V(G'), $ for all $i \in [5]$. By proposition \ref{cc10} (d), $L(V_0) \cap L(V_2)= \emptyset$; $L(V_0) \cap L(V_4)= \emptyset$ and $L(V_4) \cap L(V_2)= \emptyset$. Now, $L(V_2) \subseteq[7] \backslash L(V_0)$ and $L(V_4) \subseteq [7] \backslash L(V_0)$. Note that $\left|[7] \backslash L(V_0)\right|=5$. Now set $[7] \backslash L(V_0)=[A']$. $L(V_4) \subseteq [A'] \backslash L(V_2)$ since $L(V_4) \cap L(V_2)= \emptyset$. Now, $\left|[A'] \backslash L(V_2)\right|=2$. However, by Proposition \ref{cc10} (c), $\left|L(V_4)\right|=3$. Therefore a contradiction and hence $\lambda(C_6 \times C_6) \geq 8.$ The labeling in Figure $6$ confirms that $\lambda_1^1(C_6 \times C_6)\leq 8$, and thus, $\lambda_1^1(C_6 \times C_6)=8$.
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
{\tiny{
\node [minimum size=0cm,] at (-6.5,4) {$Fig. \; 6$: $8-L(1,1)$- labeling of $C_6 \times C_6$};
\node [minimum size=0cm,draw,circle] (0) at (-8,5) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-7,5) {$ 1$};
\node [minimum size=0cm,draw,circle] (2) at (-6,5) {$ 2$};
\node [minimum size=0cm,draw,circle] (3) at (-5,5) {$ 0$};
\node [minimum size=0cm,draw,circle] (4) at (-7.5,5.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (5) at (-6.5,5.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (6) at (-5.5,5.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (7) at (-8,6) {$ 3$};
\node [minimum size=0cm,draw,circle] (8) at (-7,6) {$ 4$};
\node [minimum size=0cm,draw,circle] (9) at (-6,6) {$ 5$};
\node [minimum size=0cm,draw,circle] (10) at (-5,6) {$ 3$};
\node [minimum size=0cm,draw,circle] (11) at (-7.5,6.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (12) at (-6.5,6.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (13) at (-5.5,6.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (14) at (-8,7) {$ 6$};
\node [minimum size=0cm,draw,circle] (15) at (-7,7) {$ 7$};
\node [minimum size=0cm,draw,circle] (16) at (-6,7) {$ 8$};
\node [minimum size=0cm,draw,circle] (17) at (-5,7) {$ 3$};
\node [minimum size=0cm,draw,circle] (18) at (-7.5,7.5) {$ 8$};
\node [minimum size=0cm,draw,circle] (19) at (-6.5,7.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (20) at (-5.5,7.5) {$ 7$};
\node [minimum size=0cm,draw,circle] (21) at (-8,8) {$ 0$};
\node [minimum size=0cm,draw,circle] (22) at (-7,8) {$ 1$};
\node [minimum size=0cm,draw,circle] (23) at (-6,8) {$ 2$};
\node [minimum size=0cm,draw,circle] (24) at (-5,8) {$ 0$};
}}\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (3) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (6) to (10);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (10) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [thin=1.00] (13) to (17);
\draw [thin=1.00] (14) to (18);
\draw [thin=1.00] (15) to (18);
\draw [thin=1.00] (15) to (19);
\draw [thin=1.00] (16) to (19);
\draw [thin=1.00] (16) to (20);
\draw [thin=1.00] (17) to (20);
\draw [thin=1.00] (18) to (21);
\draw [thin=1.00] (18) to (22);
\draw [thin=1.00] (19) to (22);
\draw [thin=1.00] (19) to (23);
\draw [thin=1.00] (20) to (23);
\draw [thin=1.00] (20) to (24);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
\end{proof}
\begin{thm} \label{cc16} $\lambda_1^1(C_{10} \times C_6)=7$
\end{thm}
\begin{proof}Let $G'$ be a connected component of $C_{10} \times C_6.$ and suppose that $\lambda_1^1(G')=6$. Let $V_0 \subset V(G')$ such that $L(V_0) \subset [6].$ By Proposition \ref{cc10} (c), $L(V_0) \cap L(V_2)= \emptyset$. Therefore $L(V_0) \subset [6] \backslash L(V_0),$ where $\left|L(V_0)\right|=3$. Thus, $\left|[6]\backslash L(V_0)\right|=4$. Now, for all $v_0 \in V_0$ and $v_8 \in V_8,$ $V_0,V_8 \subset V(G')$, $d(v_0,v_8)=2,$ since $C_{10}$ is a cycle of length $10$. Therefore, $L(V_8) \subset [6] \backslash L(V_0).$ Now, suppose that, $L(V_8)=L(V_2),$ by Proposition \ref{cc10}, then there exists $\alpha_k \in [6]$ such that $\alpha_k \notin L(V_0)$, and$\alpha_k \in L(V_2).$ Thus $L(V_4) \subset L(V_0) \cup \alpha_k$ and $L(V_6) \subset L(V_0) \cup \alpha_k$. Now, $\left|L(V_0) \cup \alpha_k \right|=4$. By Proposition \ref{cc10}, (d), $L(V_6) \cap L(V_4)= \emptyset$. Thus $\left|L(V_6) \cup L(V_4)\right|=6,$ which is a contradiction.
Now, suppose $L(V_8) \neq L(V_0)$ then it is not difficult to see that there exists $\alpha_a, \alpha_b \in [6] \backslash L(V_0)$ such that $L(V_8) \cap L(V_2)=\left\{\alpha_a, \alpha_b\right\}$. Thus $L(V_8)=\left\{\alpha_a,\alpha_b,\alpha_c\right\}$ and $L(V_2)=\left\{\alpha_a,\alpha_b,\alpha_d\right\}$ such that $L(V_2) \cup L(V_8)=[6]\backslash L[V_2]$.
Now by Proposition \ref{cc10} (d) still, $L(V_4) \subseteq [6]\backslash L(V_2)=L(V_0) \cup \alpha_k,$ such that $\alpha_k \notin L(V_0)$, $\alpha_k \in L(V_2).$ $L(V_0) \subseteq [6]\backslash L(V_8)=L(V_0) \cup \alpha_j$ for $\alpha_j \notin L(V_2),$ $\alpha_k \notin L(V_0), \; \alpha_j \neq \alpha_k$. Thus, $\left|L(V_0) \cup \left\{\alpha_j\cup \alpha_k\right\}\right|=5$. By $\left|L(V_6 \cup L(V_4))\right|=6$, and for all $v_6 \in V_6,$ and $v_4 \in V_4$, $d(v_4,v_6)=2.$ Thus a contradiction and hence $\lambda(G') \geq 7.$ \\Conversely, we consider the $7-L(1,1)$-labeling of $C_{10} \times C_6$ in Figure $5$ below.
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
{\tiny{
\node [minimum size=0cm,] at (-5.5,-9) {$Fig. \; 7 $: $7-L(1,1)$- labeling of $C_{10} \times C_6$};
\node [minimum size=0cm,draw,circle] (0) at (-8,-8) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-8,-7) {$ 1$};
\node [minimum size=0cm,draw,circle] (2) at (-8,-6) {$ 2$};
\node [minimum size=0cm,draw,circle] (3) at (-8,-5) {$ 0$};
\node [minimum size=0cm,draw,circle] (4) at (-7.5,-7.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (5) at (-7.5,-6.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (6) at (-7.5,-5.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (7) at (-7,-8) {$ 6$};
\node [minimum size=0cm,draw,circle] (8) at (-7,-7) {$ 7$};1
\node [minimum size=0cm,draw,circle] (9) at (-7,-6) {$ 3$};
\node [minimum size=0cm,draw,circle] (10) at (-7,-5) {$ 6$};
\node [minimum size=0cm,draw,circle] (11) at (-6.5,-7.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (12) at (-6.5,-6.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (13) at (-6.5,-5.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (14) at (-6,-8) {$ 1$};
\node [minimum size=0cm,draw,circle] (15) at (-6,-7) {$ 4$};
\node [minimum size=0cm,draw,circle] (16) at (-6,-6) {$ 5$};
\node [minimum size=0cm,draw,circle] (17) at (-6,-5) {$ 1$};
\node [minimum size=0cm,draw,circle] (18) at (-5.5,-7.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (19) at (-5.5,-6.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (20) at (-5.5,-5.5) {$ 7$};
\node [minimum size=0cm,draw,circle] (21) at (-5,-8) {$ 2$};
\node [minimum size=0cm,draw,circle] (22) at (-5,-7) {$ 7$};
\node [minimum size=0cm,draw,circle] (23) at (-5,-6) {$ 0$};
\node [minimum size=0cm,draw,circle] (24) at (-5,-5) {$ 2$};
\node [minimum size=0cm,draw,circle] (25) at (-4.5,-7.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (26) at (-4.5,-6.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (27) at (-4.5,-5.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (28) at (-4,-8) {$ 6$};
\node [minimum size=0cm,draw,circle] (29) at (-4,-7) {$ 5$};
\node [minimum size=0cm,draw,circle] (30) at (-4,-6) {$ 3$};
\node [minimum size=0cm,draw,circle] (31) at (-4,-5) {$ 6$};
\node [minimum size=0cm,draw,circle] (32) at (-3.5,-7.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (33) at (-3.5,-6.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (34) at (-3.5,-5.5) {$ 7$};
\node [minimum size=0cm,draw,circle] (35) at (-3,-8) {$ 0$};
\node [minimum size=0cm,draw,circle] (36) at (-3,-7) {$ 1$};
\node [minimum size=0cm,draw,circle] (37) at (-3,-6) {$ 2$};
\node [minimum size=0cm,draw,circle] (38) at (-3,-5) {$ 0$};
}}\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (3) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (6) to (10);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (10) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [thin=1.00] (13) to (17);
\draw [thin=1.00] (14) to (18);
\draw [thin=1.00] (15) to (18);
\draw [thin=1.00] (15) to (19);
\draw [thin=1.00] (16) to (19);
\draw [thin=1.00] (16) to (20);
\draw [thin=1.00] (17) to (20);
\draw [thin=1.00] (18) to (21);
\draw [thin=1.00] (18) to (22);
\draw [thin=1.00] (19) to (22);
\draw [thin=1.00] (19) to (23);
\draw [thin=1.00] (20) to (23);
\draw [thin=1.00] (20) to (24);
\draw [thin=1.00] (21) to (25);
\draw [thin=1.00] (22) to (25);
\draw [thin=1.00] (22) to (26);
\draw [thin=1.00] (23) to (26);
\draw [thin=1.00] (23) to (27);
\draw [thin=1.00] (24) to (27);
\draw [thin=1.00] (25) to (28);
\draw [thin=1.00] (25) to (29);
\draw [thin=1.00] (26) to (29);
\draw [thin=1.00] (26) to (30);
\draw [thin=1.00] (27) to (30);
\draw [thin=1.00] (27) to (31);
\draw [thin=1.00] (28) to (32);
\draw [thin=1.00] (29) to (32);
\draw [thin=1.00] (29) to (33);
\draw [thin=1.00] (30) to (33);
\draw [thin=1.00] (30) to (34);
\draw [thin=1.00] (31) to (34);
\draw [thin=1.00] (32) to (35);
\draw [thin=1.00] (32) to (36);
\draw [thin=1.00] (33) to (36);
\draw [thin=1.00] (33) to (37);
\draw [thin=1.00] (34) to (37);
\draw [thin=1.00] (34) to (38);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
\end{proof}
\begin{thm} \label{cc18} $\lambda_1^1(C_{14} \times C_6)=6$
\end{thm}
\begin{proof} Since $14$ is not a multiple of $4$ and by Corollary \ref{cc14}, $\lambda_1^1(C_{14} \times C_6) \geq 6.$
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
{\tiny{
\node [minimum size=0cm,] at (5.5,-9) {$Fig. \; 8$: $5-L(1,1)$- labeling of $C_{14} \times C_6$};
\node [minimum size=0cm,draw,circle] (0) at (2,-8) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (2,-7) {$ 1$};
\node [minimum size=0cm,draw,circle] (2) at (2,-6) {$ 2$};
\node [minimum size=0cm,draw,circle] (3) at (2,-5) {$ 0$};
\node [minimum size=0cm,draw,circle] (4) at (2.5,-7.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (5) at (2.5,-6.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (6) at (2.5,-5.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (7) at (3,-8) {$ 6$};
\node [minimum size=0cm,draw,circle] (8) at (3,-7) {$ 3$};1
\node [minimum size=0cm,draw,circle] (9) at (3,-6) {$ 4$};
\node [minimum size=0cm,draw,circle] (10) at (3,-5) {$ 6$};
\node [minimum size=0cm,draw,circle] (11) at (3.5,-7.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (12) at (3.5,-6.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (13) at (3.5,-5.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (14) at (4,-8) {$ 5$};
\node [minimum size=0cm,draw,circle] (15) at (4,-7) {$ 0$};
\node [minimum size=0cm,draw,circle] (16) at (4,-6) {$ 1$};
\node [minimum size=0cm,draw,circle] (17) at (4,-5) {$ 5$};
\node [minimum size=0cm,draw,circle] (18) at (4.5,-7.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (19) at (4.5,-6.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (20) at (4.5,-5.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (21) at (5,-8) {$ 2$};
\node [minimum size=0cm,draw,circle] (22) at (5,-7) {$ 3$};
\node [minimum size=0cm,draw,circle] (23) at (5,-6) {$ 6$};
\node [minimum size=0cm,draw,circle] (24) at (5,-5) {$ 2$};
\node [minimum size=0cm,draw,circle] (25) at (5.5,-7.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (26) at (5.5,-6.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (27) at (5.5,-5.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (28) at (6,-8) {$ 4$};
\node [minimum size=0cm,draw,circle] (29) at (6,-7) {$ 0$};
\node [minimum size=0cm,draw,circle] (30) at (6,-6) {$ 3$};
\node [minimum size=0cm,draw,circle] (31) at (6,-5) {$ 4$};
\node [minimum size=0cm,draw,circle] (32) at (6.5,-7.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (33) at (6.5,-6.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (34) at (6.5,-5.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (35) at (7,-8) {$ 6$};
\node [minimum size=0cm,draw,circle] (36) at (7,-7) {$ 1$};
\node [minimum size=0cm,draw,circle] (37) at (7,-6) {$ 2$};
\node [minimum size=0cm,draw,circle] (38) at (7,-5) {$ 6$};
\node [minimum size=0cm,draw,circle] (39) at (7.5,-7.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (40) at (7.5,-6.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (41) at (7.5,-5.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (42) at (8,-8) {$ 3$};
\node [minimum size=0cm,draw,circle] (43) at (8,-7) {$ 4$};
\node [minimum size=0cm,draw,circle] (44) at (8,-6) {$ 5$};
\node [minimum size=0cm,draw,circle] (45) at (8,-5) {$ 3$};
\node [minimum size=0cm,draw,circle] (46) at (8.5,-7.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (47) at (8.5,-6.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (48) at (8.5,-5.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (49) at (9,-8) {$ 0$};
\node [minimum size=0cm,draw,circle] (50) at (9,-7) {$ 1$};
\node [minimum size=0cm,draw,circle] (51) at (9,-6) {$ 2$};
\node [minimum size=0cm,draw,circle] (52) at (9,-5) {$ 0$};
}}\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (4);
\draw [thin=1.00] (1) to (4);
\draw [thin=1.00] (1) to (5);
\draw [thin=1.00] (2) to (5);
\draw [thin=1.00] (2) to (6);
\draw [thin=1.00] (3) to (6);
\draw [thin=1.00] (4) to (7);
\draw [thin=1.00] (4) to (8);
\draw [thin=1.00] (5) to (8);
\draw [thin=1.00] (5) to (9);
\draw [thin=1.00] (6) to (9);
\draw [thin=1.00] (6) to (10);
\draw [thin=1.00] (7) to (11);
\draw [thin=1.00] (8) to (11);
\draw [thin=1.00] (8) to (12);
\draw [thin=1.00] (9) to (12);
\draw [thin=1.00] (9) to (13);
\draw [thin=1.00] (10) to (13);
\draw [thin=1.00] (11) to (14);
\draw [thin=1.00] (11) to (15);
\draw [thin=1.00] (12) to (15);
\draw [thin=1.00] (12) to (16);
\draw [thin=1.00] (13) to (16);
\draw [thin=1.00] (13) to (17);
\draw [thin=1.00] (14) to (18);
\draw [thin=1.00] (15) to (18);
\draw [thin=1.00] (15) to (19);
\draw [thin=1.00] (16) to (19);
\draw [thin=1.00] (16) to (20);
\draw [thin=1.00] (17) to (20);
\draw [thin=1.00] (18) to (21);
\draw [thin=1.00] (18) to (22);
\draw [thin=1.00] (19) to (22);
\draw [thin=1.00] (19) to (23);
\draw [thin=1.00] (20) to (23);
\draw [thin=1.00] (20) to (24);
\draw [thin=1.00] (21) to (25);
\draw [thin=1.00] (22) to (25);
\draw [thin=1.00] (22) to (26);
\draw [thin=1.00] (23) to (26);
\draw [thin=1.00] (23) to (27);
\draw [thin=1.00] (24) to (27);
\draw [thin=1.00] (25) to (28);
\draw [thin=1.00] (25) to (29);
\draw [thin=1.00] (26) to (29);
\draw [thin=1.00] (26) to (30);
\draw [thin=1.00] (27) to (30);
\draw [thin=1.00] (27) to (31);
\draw [thin=1.00] (28) to (32);
\draw [thin=1.00] (29) to (32);
\draw [thin=1.00] (29) to (33);
\draw [thin=1.00] (30) to (33);
\draw [thin=1.00] (30) to (34);
\draw [thin=1.00] (31) to (34);
\draw [thin=1.00] (32) to (35);
\draw [thin=1.00] (32) to (36);
\draw [thin=1.00] (33) to (36);
\draw [thin=1.00] (33) to (37);
\draw [thin=1.00] (34) to (37);
\draw [thin=1.00] (34) to (38);
\draw [thin=1.00] (35) to (39);
\draw [thin=1.00] (36) to (39);
\draw [thin=1.00] (36) to (40);
\draw [thin=1.00] (37) to (40);
\draw [thin=1.00] (37) to (41);
\draw [thin=1.00] (38) to (41);
\draw [thin=1.00] (39) to (42);
\draw [thin=1.00] (39) to (43);
\draw [thin=1.00] (40) to (43);
\draw [thin=1.00] (40) to (44);
\draw [thin=1.00] (41) to (44);
\draw [thin=1.00] (41) to (45);
\draw [thin=1.00] (42) to (46);
\draw [thin=1.00] (43) to (46);
\draw [thin=1.00] (43) to (47);
\draw [thin=1.00] (44) to (47);
\draw [thin=1.00] (44) to (48);
\draw [thin=1.00] (45) to (48);
\draw [thin=1.00] (46) to (49);
\draw [thin=1.00] (46) to (50);
\draw [thin=1.00] (47) to (50);
\draw [thin=1.00] (47) to (51);
\draw [thin=1.00] (48) to (51);
\draw [thin=1.00] (48) to (52);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
\end{proof}
We can conclude that for $C_m \times C_6,$ if $m$ is even, and $m \not\equiv 0 \mod 4$, then $\lambda_1^1(C_m \times C_6) \geq 6$. For $m \geq 14$, this class of direct product graphs can be obtained from $C_{14+4m'}$, where $m'$ is a non-negative integer.
\begin{thm} \label{cc19} For $m=14+4m'$, where $m'$ is a non-negative integer, $\lambda_1^1(C_m \times C_6)=6$
\end{thm}
\begin{proof} Since for any non-negative integer $m'$, $m=14+4m' \not\equiv 0 \mod \; 4,$ then by \ref{cc14}, $\lambda_1^1(C_m \times C_n) \geq 6$. By combining the labeling in Figure 8 and $m'-$multiple of the labeling in Figure $3$, we have that $\lambda_1^1(C_m \times C_n)\leq 6$ and the result follows.
\end{proof}
Note that the following result was established in \cite{AA1}
\begin{thm}\label{ccc20} Let $m', n' \equiv 0 mod 10$ and $A=\left\{12,14,16,18\right\}$. Then, for all $k \in A and m, n$, $\lambda_1^1(C_{m'} \times C_{k+n'})=5$. Also, let $m,n \equiv 0 mod 5$, then $\lambda_1^1(C_m \times C_n)=4.$
\end{thm}
\section{Labeling of $C_m \times C_n,\ n\geq 8$}
In this section, we obtain the $\lambda_1^1-$numbers of graph product $C_m \times C_n$, where $n,m \geq 8$.
Now we establish the $\lambda_1^1-$number for $C_m \times C_8$. Since labeling of product graphs is commutative, we restrict our work in this section to $m \geq 8$ since the cases for smaller graphs have been taken care of in the last sections.
The following result are helpful to reveal some useful properties of $L(1,1)$- labeling of $C_m \times C_8$.
\begin{lem} \label{cc20} Let $G'$ be a connected component of $C_m \times C_8, \;m \geq 4.$ Suppose there exist $v_a,v_b \in V_i$ such that $\alpha_k= l(v_a)=l(v_b) \in[p]$, $p\in \mathbb N.$ then $\alpha_k \notin L(V_{i+1} \cup V_{i+2})$. Furthermore, $\alpha_k \notin L(V_{i-1} \cup V_{i-2})$.
\end{lem}
\begin{proof} Claim: Let $\alpha_k=l(v_a)=l(v_b),$ $v_a,v_b \in V_i,$ $V_i \subseteq V(G').$ Then, $d(v_a,v_b)=4$.
\\ Reason: Clearly, $\left|V_i\right|=4$ for $i \in [m-1]$. Let $V_i=\left\{u_iv_0,u_iv_2,u_iv_4,u_iv_6\right\}$. So, $d(u_iv_0,u_iv_2)$=$d(u_iv_2,u_iv_4)$=$d(u_iv_4,u_iv_6)$=$2$. also, $d(u_iv_6,u_iv_4)=2$ since $C_8$ is a cycle. However, $d(u_iv_0,u_iv_4)$=$d(u_iv_2,u_iv_6)=4$. Thus $v_a=v_iv_0$ and $v_b=v_iv_4$ or $v_a=v_iv_2$ and $v_b=v_iv_6$.
Now, suppose $v_a=u_iv_0$ and $v_b=u_iv_4.$ Let $V_{i+1}=\left\{u_{i+1}v_1,u_{I+1}v_3,u_{i+1}v_5,u_{i+1}v_7\right\}$. Then that $d(u_iv_0,u_{i+1}v_1)=1=d(u_iv_0,u_{i+1}v_7)$ follows from the definition of $C_m \times C_{n=8}.$ Likewise, $d(u_iv_4,u_{i+1}v_3)=1=d(u_iv_4,u_{i+1}v_5)$. Therefore, $\alpha_k \in L(V_{i+1}).$ Also, let $V_{i+2}=\left\{u_{i+2}v_0,u_{i+2}v_2,u_{i+2}v_4,u_{i+2}v_6\right\}$. Then $d(u_iv_0,u_{i+2}v_0)=2$=$d(u_iv_0,u_{i+2}v_2)$ and $d(u_iv_0,u_{i+2}v_6)$=2=$d(u_iv_0,u_{i+2}v_2)$ since $C_{m=8}$ is a cycle. Now $d(u_iv_4,u_{i+2}v_{2(4,6)})=2$ and therefore $\alpha_k \notin L(V_{i+2})$. This argument is valid for $V_{i-1}$ and $V_{1-2}$.
\end{proof}
The consequence of Lemma \ref{cc20} is that if a label is assigned to two vertices on $V_i \subset V(G')$, then the label could no longer be assigned to another vertex on the vertex sets two step above or below it. The next result is similar.
\begin{prop} \label{cc21} Suppose $v_i \in V_i$ and $v_{i+2} \in V_{i+1}$ such that $\alpha_k =l(v_i)=l(v_{i+1}),$ then, $d(v_i,v_{i+1})=3$.
\end{prop}
\begin{proof} Suppose $v_i=u_iv_0$ without loss of generality, then $d(v_i,u_{i+1}v_{1(7)})=2.$ Now. $d(v_i, u_{i+1}v_3)=3$ and $d(v_i, u_{i+1}v_5)=3$.
\end{proof}
\begin{lem}
\label{cc22} Suppose $V_1,V_{i+1} \subset V(G')$ where $G'$ is a connected component of $C_m \times C_8$. Let $\alpha_k \in L(V_i) \cap L(V_{i+1})$ then $\alpha_k \notin L(V_{i-1}) \cup L(V_{i+2})$.
\end{lem}
\begin{proof}
By Proposition \ref{cc21}, suppose that $\alpha_k=l(u_iv_j)$ and that $\alpha_k \in L(V_{i+1})$ then, $\alpha_k=l(u_{i+1}v_{j+3})$ or $\alpha_k=l(u_{i+1}v_{j+5})$. Without loss of generality, suppose that in fact, $\alpha_k=l(u_{i+1}v_j+5)$. Then $d(u_iv_j,u_{i+2}v_{j(j+2,j+6)})=2$. Meanwhile, $d(u_{i+1}v_{j+3(j+5)},u_{i+2}v_{j+4})=1$. Thus, $\alpha_k \notin L(V_{i+2})$. Similar argument holds for $\alpha_k \notin L(V_{i-1})$.
\end{proof}
By to Lemma \ref{cc22}, it is quite clear that if $\alpha_k$ belongs $L(V_i) \cap L(V_i+1)$, then $\alpha_k$ does not belong to $L(V_{i-1} \cup V{i+2})$. A similar result is as follows:
\begin{lem} \label{cc23} Suppose $\alpha_k \in L(V_i) \cap L(V_{i+2}) \subset V(G')$, where $G'$ is a connected component of $C_m \times C_8$, $m \geq 4$, then $\alpha_k \notin L(V_{i+1})$
\end{lem}
\begin{proof} Let, $v_i \in V_i$ be $u_iv_0$. Note that, $d(u_iv_0,u_{i+2}v_{0(2,6)})=2$ since $C_8$ is a cycle. Then, the remaining vertex $v_{i+2} \in V_{i+2}$ such that $l(v_{i+2})=\alpha_k$ is $u_{i+2}v_4$ and $d(u_iv_0,u_{i+2}v_4)=4$. Now $d(v_i,u_{i+1}v_{1(7)})=1$ and $d(u_{i+2}v_4,u_{i+1}v_{3(5)})=1$. Thus, $\alpha_k \notin L(V_{i+1})$.
\end{proof}
The consequence of Lemma \ref{cc23} is that if two vertices on $V_i$ and $V_{i+2}$ share the same label, then that label can not be shared by another vertex on $V_{i+1}$ given that $V_i,V_{i+1}$ and $V_{i+2}$ are all in $V(G')$.
Next we establish the lower bound of $\lambda^1_1(C_m \times C_8)$ where $m \geq 8$ and $m\equiv 2 \mod 6$. We require the following definition.
Let $G'$ be a connected component of $G$. Then, $V_{\alpha_k} $ is the class of all vertices on $V(G')$ labeled $\alpha_k$.
\begin{lem} \label{cc24}For $m\geq 8,m \equiv 2\; mod \; 6$, $\lambda_1^1(C_m \times C_8) \geq 6$.
\end{lem}
\begin{proof}Case 1: Let $\alpha_k \in L(V(G'))$ such that if $\alpha_k \in L(V_i),\; V_i \subset V(G'),$ $i \in [m-1]$, then there exist $v'_i,v''_i \in V_i$ such that $l(v'_i)= \alpha_k=l(v''_i)$. Let $\bar{V}$ be a class of all $V_i \in V(G')$ such that $\alpha_k \in L(V_i)$. Now suppose, without loss of generality, that $V_0 \in \bar{V}$. By this and Lemma \ref{cc20}, and by assuming that $\alpha_k $ labels $V(G')$ optimally, suppose $V_i \in \bar{V}$, then $i\equiv 0\; \mod\; 3, i \neq m-2.$ Since $m\equiv 2 \; \mod \; 6$, then there exists $n' \in \mathbb N$, such that $m=6n'+2$. Note that $m-5=(6n'+2)-5=3(2n'-1)$. Thus, $V_{m-5} \in \bar{V}.$ By Lemma \ref{cc20} and since $V_0,V_{m-5} \in \bar{V}$, then $\alpha_k \notin L(V_{m-4} \cup V_{m-3}\cup V_{m-2}\cup V_{m-1})$. Thus $\bar{V}=\left\{V_0,\cdots,V_{m-5}\right\}$. Set $\bar{V'}=\bar{V}\backslash\left\{V_0\right\}$. Since $\left|\bar{V'}\right|=2n'-1$, then $\left|V'\right|=2n'$. Now, $\left|V(G')\right|=(6n'+2)4$. Clearly $\left|V_{\alpha_k}\right|=2(2n')=4n'$. Hence, $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}$=$\frac{6n'+2}{n'} > 6$.
Case 2: Suppose that for all triple $V_i,V_{i+1}V_{i+2} \subset V(G)$, $\alpha_k \in L(V_i \cap V_{i+1})$ and by Lemma \ref{cc22} $\alpha_k \notin V_{i+2}$. Without loss of generality, we select the initial triple to be $V_0,V_1,V_2$, such that $\alpha_k \in L(V_0 \cap V_1)$, $\alpha_k \notin V_2$; (and $\alpha_k \in L(V_3 \cap V_4)$,$\alpha_k \notin L(V_5) \cdots )$. Therefore, $\alpha_k \notin V_i $ for all $i \in [m-1]$ such that $i+1\equiv \;0 \mod \; 3.$ Now, $m \equiv 2\; \mod\; 6$ implies there exists $n' \in \mathbb N$ such that $m=6n'+2$. Thus, $m-2 \equiv 0\; \mod \;3$ and hence $\alpha_k \notin L(V_{m-3}).$ Now, since $\alpha_k \in L(V_0) \cap L(V_1)$, then $\alpha_k \notin L(V_{m-1})$ by Lemma \ref{cc22} and since $C_m$ is a cycle. By Lemma \ref{cc23}, it is possible for $\alpha_k \in L(V_{m-2})$ since $\alpha_k \notin L(V_{m-3})$. Thus we, for maximality, assume that $\alpha_k \in L(V_{m-2})$. Now, Let $\bar{V_{\alpha_k}}$=$\left\{V_0,\cdots,V_{m-3}\right\} \subset V(G')$. Then $\left|\bar{V_{\alpha_k}}\right|=\left[\frac{(m-3)+1}{3}\right]=2(\frac{6n'}{3})=4n'$, where $n' \in \mathbb N$. Thus, for all $V_i \in V(G'),$ $\left|V_{\alpha_k}\right|=4n'+1$ since $\alpha_k \in L(V_{m-2})$. Thus $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}=\frac{(6n'+2)4}{4n'+1} > 6$.
Case 3. Suppose that, by Lemma \ref{cc23}, $\alpha_k \in L(V_i,V_{i+2}, V_{i+4}, \cdots, V_{i-2})$. Clearly, $\left|V_{\alpha_k}\right|=\frac{m}{2}$, since $m$ is even. Now, $m=2\; \mod \;6$ implies that there exists $n' \in \mathbb N$ such that $m=6n'+2$. Therefore $\left|V_{\alpha_k}\right|=3n'+1$. Now $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}=\frac{(6n'+2)4}{3n'+1} > 8$.
It is easy therefore to see that combination of the Cases 1-3 will still result in $\frac{\left|V(G')\right|}{V_{\alpha_k}} \geq 7.$ Thus for all $\alpha_k \in [p]$, where $\alpha_k \in [p]$, $\lambda_1^1(G')=p$,$\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}\geq 7$. Suppose $\lambda^1_1(G')=p=5$ and the maximum number of vertices in $G'$ that $\alpha_k \in [p]$ labels for all $\alpha_k \in [p]$ is $V_{\alpha_k}$, then $(p+1)V_{\alpha_k} \geq \left|V(G')\right|$ implies that $p+1 \geq $ $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}$. This implies that $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|} \leq p+1$. Now, since $p=5$, then $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|} \leq 6$, which is a contradiction since in fact, $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|} \geq 7.$ \\Thus $\lambda_1^1(C_m \times C_8) \geq 6$ for all $m \equiv 2\; \mod \; 6$.
\end{proof}
Next, we consider the second case of $m \equiv 4 \; \mod \; 6$.
\begin{lem} \label{cc25} For $m \equiv 4\; mod\; 6$, $\lambda_1^1(C_m \times C_8) \geq 6$.
\end{lem}
\begin{proof} Case 1: Let $G'$ be a connected component of $C_m \times C_8$, $m \equiv 4 \mod 6$ and let $\bar{V}$ be a set of $V_i \subset V(G')$ such that for all $i$ there exist $v'_i,v''_i \in V_i$ such that $l(v'_i)=\alpha_k=l(v''_i)$. Now suppose $V_0 \in \bar{V}.$ By Lemma \ref{cc20}, $\alpha_k \notin L(V_1 \cup V_2)$. Since $\bar{V},$ contains all possible $V_i \subset V(G)$ and since $V_0\in \bar{V}$, then for all $i \equiv 0\;\mod\;3,$ $V_i \in \bar{V}$ except for $i=m-1$ since $C_m$ is a cycle and $V_0 \in \bar{V}$. We know that $m=6n'+4$, $n' \in \mathbb N$ and thus, $m-4=0 \; \mod\;3$, which implies that $V_{m-4} \in \bar{V}$. Set $\bar{V'}=\left\{V_3,\cdots ,V_{m-4}\right\}$. Thus, $\left|V'\right|=\frac{m-4}{3}=\frac{6n'+4-4}{3}=2n'$. Now, $\bar{V}=\bar{V}' \cup V_0$. Thus $\left|\bar{V}\right| 2n'+1$ and $\left|V_{\alpha_k}\right|=2(2n'+1)=4n'+2$. Now, $\left|V(G')\right|=4(6n'+4)=24n'+16.$ Finally, $\frac{\left|V(G)\right|}{\left|V_{\alpha_k}\right|}=\frac{24n'+16}{4n'+2} > 6 $
Case 2: Suppose that for all triple $V_i,V_{i+1},V_{i+2} \subset V(G)$, $\alpha_k \in L(V_i) \cap L(V_{i+1})$ and $\alpha_k \notin L(V_{i+2})$. We can select the initial triple as $V_0,V_1,V_2$, that is, $\alpha_k \in L(V_0) \cap L(V_1)$ and $\alpha_k \notin L(V_2)$ (and subsequently, $\alpha_k \in L(V_3) \cap L(V_4)$ and $\alpha_k \notin L(V_5) \cdots $). Thus, $\alpha_k \notin V_i$ for all $i$ such that $i+1 \equiv 0\; \mod \; 3.$ Now since $m \equiv 4\; \mod \; 6$ there exists $n' \in \mathbb N$ such that $m\equiv 6n'+4$. Clearly, $m-1=6n'+3=3(2n'+1)\equiv 0\;\mod\; 3.$ However, $\alpha_k \notin L(V_{m-1})$ since $C_m$ is a cycle and by the Lemma \ref{cc22}. Therefore let $\bar{V}=\left\{V_0,V_1,V_2,\cdots,V_{m-2} \right\} \subseteq V(G')$. Then $\left|\bar{V}\right|=m-2+1=m-1$. Clearly $\left|V_{\alpha_k}\right|=2 \frac{\left|\bar{V}\right|}{3} = \frac{2(m-1)}{3}$. The last equation implies that $\frac{2(6n'+4-1)}{3}=$$\frac {2\cdot3(2n'+1)}{3}=2(n'+1)$, $n' \in \mathbb N.$ Now,$\left|V(G')\right|=4m=4(6n'+16)=24n'+16$. Therefore, $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}=\frac{24n'+16}{4n'+2} > 6$.
Case 3: This follows similar argument as in Case 3, in the proof of Lemma \ref{cc24}.
Therefore for $m$ even, $m\equiv 4\; \mod \;6,$ $\lambda_1^1(G') \geq 6$ follows similar argument as in proof of Lemma \ref{cc24}.
\end{proof}
\begin{cor} \label{cc26}For all $m$ even, $m \not\equiv \ 0 mod 6$, $\lambda_1^1(C_m \times C_8) \geq 6.$
\end{cor}
\begin{proof} It follows from combining the results in Lemmas \ref{cc24} and \ref{cc25}.
\end{proof}
Next we obtain the $\lambda_1^1$ number of a special case of Corollary \ref{cc26}.
\begin{thm} \label{cc27} $\lambda_1^1(C_8 \times C_8)=7$
\end{thm}
\begin{proof} By following the the process in the proof of Lemma \ref{cc24}, we have that $\left|V_{\alpha_k}\right|=4$, for $V_{\alpha_k} \subseteq V(G')$, where $G'$ is a connected component of $C_8 \times C_8$ and therefore $\frac{\left|V(G')\right|}{\left|V_{\alpha_k}\right|}=\frac{32}{4}=8$. Thus $\lambda_1^1(C_8 \times C_8)\geq 7$. From an earlier result,
$\lambda_1^1(C_4 \times C_8)=7$. By copying re-occurrence of the labeling of $C_4 \times C_8$, we have that $\lambda_1^1(C_8 \times C_8)\leq 7$ and the result follows.
\end{proof}
In what follows, we extend our result to $m \geq 10$.
\begin{thm}\label{cc28} Let $m \in \left\{10, 14\right\}$. Then $\lambda_1^1(C_{m} \times C_8)=6$
\end{thm}
\begin{proof}
By Corollary \ref{cc26}, $\lambda_1^1(C_{m} \times C_8) \geq 6$ for all $m \in \left\{10, 14\right\}$.
Conversely, we show that for $m \in \left\{10, 14\right\}$, $\lambda_1^1(C_m \times C_8) \leq 6$ by labeling their connected component as shown below.
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-6.0,16) {$Fig. \; 9$: $6-L(1,1)$-Labeling of $C_{10} \times C_8$};
\node [minimum size=0cm,draw,circle] (0) at (-4,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-5,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (2) at (-6,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (3) at (-7,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (4) at (-8,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (6) at (-4.5,17.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (7) at (-5.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (8) at (-6.5,17.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (9) at (-7.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (11) at (-4,18) {$ 1$};
\node [minimum size=0cm,draw,circle] (12) at (-5,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (13) at (-6,18) {$ 1$};
\node [minimum size=0cm,draw,circle] (14) at (-7,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (15) at (-8,18) {$ 1$};
\node [minimum size=0cm,draw,circle] (17) at (-4.5,18.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (18) at (-5.5,18.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (19) at (-6.5,18.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (20) at (-7.5,18.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (22) at (-4,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (23) at (-5,19) {$ 4$};
\node [minimum size=0cm,draw,circle] (24) at (-6,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (25) at (-7,19) {$ 4$};
\node [minimum size=0cm,draw,circle] (26) at (-8,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (28) at (-4.5,19.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (29) at (-5.5,19.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (30) at (-6.5,19.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (31) at (-7.5,19.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (33) at (-4,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (34) at (-5,20) {$ 0$};
\node [minimum size=0cm,draw,circle] (35) at (-6,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (36) at (-7,20) {$ 0$};
\node [minimum size=0cm,draw,circle] (37) at (-8,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (39) at (-4.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (40) at (-5.5,20.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (41) at (-6.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (42) at (-7.5,20.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (44) at (-4,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (45) at (-5,21) {$ 5$};
\node [minimum size=0cm,draw,circle] (46) at (-6,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (47) at (-7,21) {$ 5$};
\node [minimum size=0cm,draw,circle] (48) at (-8,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (50) at (-4.5,21.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (51) at (-5.5,21.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (52) at (-6.5,21.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (53) at (-7.5,21.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (55') at (-4,22) {$ 0$};
\node [minimum size=0cm,draw,circle] (56') at (-5,22) {$ 4$};
\node [minimum size=0cm,draw,circle] (57') at (-6,22) {$ 0$};
\node [minimum size=0cm,draw,circle] (58') at (-7,22) {$ 4$};
\node [minimum size=0cm,draw,circle] (59') at (-8,22) {$ 0$};
\node [minimum size=0cm,] at (0,16) {$Fig. \; 10$: $6-L(1,1)$-Labeling of $C_{14} \times C_{8}$};
\node [minimum size=0cm,draw,circle] (70) at (2,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (71) at (1,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (72) at (0,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (73) at (-1,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (74) at (-2,17){$ 0$};
\node [minimum size=0cm,draw,circle] (76) at (1.5,17.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (77) at (0.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (78) at (-0.5,17.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (79) at (-1.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (81) at (2,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (82) at (1,18) {$ 6$};
\node [minimum size=0cm,draw,circle] (83) at (0,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (84) at (-1,18) {$ 6$};
\node [minimum size=0cm,draw,circle] (85) at (-2,18){$ 2$};
\node [minimum size=0cm,draw,circle] (87) at (1.5,18.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (88) at (0.5,18.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (89) at (-0.5,18.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (90) at (-1.5,18.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (92) at (2,19) {$ 1$};
\node [minimum size=0cm,draw,circle] (93) at (1,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (94) at (0,19) {$ 1$};
\node [minimum size=0cm,draw,circle] (95) at (-1,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (96) at (-2,19){$ 1$};
\node [minimum size=0cm,draw,circle] (98) at (1.5,19.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (99) at (0.5,19.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (100) at (-0.5,19.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (101) at (-1.5,19.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (103) at (2,20) {$ 6$};
\node [minimum size=0cm,draw,circle] (104) at (1,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (105) at (0,20) {$ 6$};
\node [minimum size=0cm,draw,circle] (106) at (-1,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (107) at (-2,20){$ 6$};
\node [minimum size=0cm,draw,circle] (109) at (1.5,20.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (110) at (0.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (111) at (-0.5,20.5){$1$};
\node [minimum size=0cm,draw,circle] (112) at (-1.5,20.5) {$4$};
\node [minimum size=0cm,draw,circle] (114) at (2,21) {$ 0$};
\node [minimum size=0cm,draw,circle] (115) at (1,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (116) at (0,21) {$ 0$};
\node [minimum size=0cm,draw,circle] (117) at (-1,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (118) at (-2,21){$ 0$};
\node [minimum size=0cm,draw,circle] (120) at (1.5,21.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (121) at (0.5,21.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (122) at (-0.5,21.5){$ 5$};
\node [minimum size=0cm,draw,circle] (123) at (-1.5,21.5){$ 3$};
\node [minimum size=0cm,draw,circle] (125') at (2,22) {$ 6$};
\node [minimum size=0cm,draw,circle] (126') at (1,22) {$ 4$};
\node [minimum size=0cm,draw,circle] (127') at (0,22) {$ 6$};
\node [minimum size=0cm,draw,circle] (128') at (-1,22) {$ 4$};
\node [minimum size=0cm,draw,circle] (129') at (-2,22) {$ 6$};
\node [minimum size=0cm,draw,circle] (130) at (1.5,22.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (131) at (0.5,22.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (132) at (-0.5,22.5){$ 0$};
\node [minimum size=0cm,draw,circle] (133) at (-1.5,22.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (134) at (2,23) {$ 5$};
\node [minimum size=0cm,draw,circle] (135) at (1,23) {$ 1$};
\node [minimum size=0cm,draw,circle] (136) at (0,23) {$ 5$};
\node [minimum size=0cm,draw,circle] (137) at (-1,23) {$ 1$};
\node [minimum size=0cm,draw,circle] (138) at (-2,23) {$ 5$};
\node [minimum size=0cm,draw,circle] (139) at (1.5,23.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (140) at (0.5,23.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (141) at (-0.5,23.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (142) at (-1.5,23.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (143) at (2,24) {$ 0$};
\node [minimum size=0cm,draw,circle] (144) at (1,24) {$ 4$};
\node [minimum size=0cm,draw,circle] (145) at (0,24) {$ 0$};
\node [minimum size=0cm,draw,circle] (146) at (-1,24) {$ 4$};
\node [minimum size=0cm,draw,circle] (147) at (-2,24){$ 0$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (6);
\draw [thin=1.00] (1) to (6);
\draw [thin=1.00] (1) to (7);
\draw [thin=1.00] (2) to (7);
\draw [thin=1.00] (2) to (8);
\draw [thin=1.00] (3) to (8);
\draw [thin=1.00] (3) to (9);
\draw [thin=1.00] (4) to (9);
\draw [thin=1.00] (6) to (11);
\draw [thin=1.00] (6) to (12);
\draw [thin=1.00] (7) to (12);
\draw [thin=1.00] (7) to (13);
\draw [thin=1.00] (8) to (13);
\draw [thin=1.00] (8) to (14);
\draw [thin=1.00] (9) to (14);
\draw [thin=1.00] (9) to (15);
\draw [thin=1.00] (11) to (17);
\draw [thin=1.00] (12) to (17);
\draw [thin=1.00] (12) to (18);
\draw [thin=1.00] (13) to (18);
\draw [thin=1.00] (13) to (19);
\draw [thin=1.00] (14) to (19);
\draw [thin=1.00] (14) to (20);
\draw [thin=1.00] (15) to (20);
\draw [thin=1.00] (17) to (22);
\draw [thin=1.00] (17) to (23);
\draw [thin=1.00] (18) to (23);
\draw [thin=1.00] (18) to (24);
\draw [thin=1.00] (19) to (24);
\draw [thin=1.00] (19) to (25);
\draw [thin=1.00] (20) to (25);
\draw [thin=1.00] (20) to (26);
\draw [thin=1.00] (22) to (28);
\draw [thin=1.00] (23) to (28);
\draw [thin=1.00] (23) to (29);
\draw [thin=1.00] (24) to (29);
\draw [thin=1.00] (24) to (30);
\draw [thin=1.00] (25) to (30);
\draw [thin=1.00] (25) to (31);
\draw [thin=1.00] (26) to (31);
\draw [thin=1.00] (28) to (33);
\draw [thin=1.00] (28) to (34);
\draw [thin=1.00] (29) to (34);
\draw [thin=1.00] (29) to (35);
\draw [thin=1.00] (30) to (35);
\draw [thin=1.00] (30) to (36);
\draw [thin=1.00] (31) to (36);
\draw [thin=1.00] (31) to (37);
\draw [thin=1.00] (33) to (39);
\draw [thin=1.00] (34) to (39);
\draw [thin=1.00] (34) to (40);
\draw [thin=1.00] (35) to (40);
\draw [thin=1.00] (35) to (41);
\draw [thin=1.00] (36) to (41);
\draw [thin=1.00] (36) to (42);
\draw [thin=1.00] (37) to (42);
\draw [thin=1.00] (39) to (44);
\draw [thin=1.00] (39) to (45);
\draw [thin=1.00] (40) to (45);
\draw [thin=1.00] (40) to (46);
\draw [thin=1.00] (41) to (46);
\draw [thin=1.00] (41) to (47);
\draw [thin=1.00] (42) to (47);
\draw [thin=1.00] (42) to (48);
\draw [thin=1.00] (44) to (50);
\draw [thin=1.00] (45) to (50);
\draw [thin=1.00] (45) to (51);
\draw [thin=1.00] (46) to (51);
\draw [thin=1.00] (46) to (52);
\draw [thin=1.00] (47) to (52);
\draw [thin=1.00] (47) to (53);
\draw [thin=1.00] (48) to (53);
\draw [thin=1.00] (50) to (55');
\draw [thin=1.00] (50) to (56');
\draw [thin=1.00] (51) to (56');
\draw [thin=1.00] (51) to (57');
\draw [thin=1.00] (52) to (57');
\draw [thin=1.00] (52) to (58');
\draw [thin=1.00] (53) to (58');
\draw [thin=1.00] (53) to (59');
\draw [thin=1.00] (70) to (76);
\draw [thin=1.00] (71) to (76);
\draw [thin=1.00] (71) to (77);
\draw [thin=1.00] (72) to (77);
\draw [thin=1.00] (72) to (78);
\draw [thin=1.00] (73) to (78);
\draw [thin=1.00] (73) to (79);
\draw [thin=1.00] (74) to (79);
\draw [thin=1.00] (76) to (81);
\draw [thin=1.00] (76) to (82);
\draw [thin=1.00] (77) to (82);
\draw [thin=1.00] (77) to (83);
\draw [thin=1.00] (78) to (83);
\draw [thin=1.00] (78) to (84);
\draw [thin=1.00] (79) to (84);
\draw [thin=1.00] (79) to (85);
\draw [thin=1.00] (81) to (87);
\draw [thin=1.00] (82) to (87);
\draw [thin=1.00] (82) to (88);
\draw [thin=1.00] (83) to (88);
\draw [thin=1.00] (83) to (89);
\draw [thin=1.00] (84) to (89);
\draw [thin=1.00] (84) to (90);
\draw [thin=1.00] (85) to (90);
\draw [thin=1.00] (87) to (92);
\draw [thin=1.00] (87) to (93);
\draw [thin=1.00] (88) to (93);
\draw [thin=1.00] (88) to (94);
\draw [thin=1.00] (89) to (94);
\draw [thin=1.00] (89) to (95);
\draw [thin=1.00] (90) to (95);
\draw [thin=1.00] (90) to (96);
\draw [thin=1.00] (92) to (98);
\draw [thin=1.00] (93) to (98);
\draw [thin=1.00] (93) to (99);
\draw [thin=1.00] (94) to (99);
\draw [thin=1.00] (94) to (100);
\draw [thin=1.00] (95) to (100);
\draw [thin=1.00] (95) to (101);
\draw [thin=1.00] (96) to (101);
\draw [thin=1.00] (98) to (103);
\draw [thin=1.00] (98) to (104);
\draw [thin=1.00] (99) to (104);
\draw [thin=1.00] (99) to (105);
\draw [thin=1.00] (100) to (105);
\draw [thin=1.00] (100) to (106);
\draw [thin=1.00] (101) to (106);
\draw [thin=1.00] (101) to (107);
\draw [thin=1.00] (103) to (109);
\draw [thin=1.00] (104) to (109);
\draw [thin=1.00] (104) to (110);
\draw [thin=1.00] (105) to (110);
\draw [thin=1.00] (105) to (111);
\draw [thin=1.00] (106) to (111);
\draw [thin=1.00] (106) to (112);
\draw [thin=1.00] (107) to (112);
\draw [thin=1.00] (109) to (114);
\draw [thin=1.00] (109) to (115);
\draw [thin=1.00] (110) to (115);
\draw [thin=1.00] (110) to (116);
\draw [thin=1.00] (111) to (116);
\draw [thin=1.00] (111) to (117);
\draw [thin=1.00] (112) to (117);
\draw [thin=1.00] (112) to (118);
\draw [thin=1.00] (114) to (120);
\draw [thin=1.00] (115) to (120);
\draw [thin=1.00] (115) to (121);
\draw [thin=1.00] (116) to (121);
\draw [thin=1.00] (116) to (122);
\draw [thin=1.00] (117) to (122);
\draw [thin=1.00] (117) to (123);
\draw [thin=1.00] (118) to (123);
\draw [thin=1.00] (120) to (125');
\draw [thin=1.00] (120) to (126');
\draw [thin=1.00] (121) to (126');
\draw [thin=1.00] (121) to (127');
\draw [thin=1.00] (122) to (127');
\draw [thin=1.00] (122) to (128');
\draw [thin=1.00] (123) to (128');
\draw [thin=1.00] (123) to (129');
\draw [thin=1.00] (125') to (130);
\draw [thin=1.00] (126') to (130);
\draw [thin=1.00] (126') to (131);
\draw [thin=1.00] (127') to (131);
\draw [thin=1.00] (127') to (132);
\draw [thin=1.00] (128') to (132);
\draw [thin=1.00] (128') to (133);
\draw [thin=1.00] (129') to (133);
\draw [thin=1.00] (130) to (134);
\draw [thin=1.00] (130) to (135);
\draw [thin=1.00] (131) to (135);
\draw [thin=1.00] (131) to (136);
\draw [thin=1.00] (132) to (136);
\draw [thin=1.00] (132) to (137);
\draw [thin=1.00] (133) to (137);
\draw [thin=1.00] (133) to (138);
\draw [thin=1.00] (134) to (139);
\draw [thin=1.00] (135) to (139);
\draw [thin=1.00] (135) to (140);
\draw [thin=1.00] (136) to (140);
\draw [thin=1.00] (136) to (141);
\draw [thin=1.00] (137) to (141);
\draw [thin=1.00] (137) to (142);
\draw [thin=1.00] (138) to (142);
\draw [thin=1.00] (139) to (143);
\draw [thin=1.00] (139) to (144);
\draw [thin=1.00] (140) to (144);
\draw [thin=1.00] (140) to (145);
\draw [thin=1.00] (141) to (145);
\draw [thin=1.00] (141) to (146);
\draw [thin=1.00] (142) to (146);
\draw [thin=1.00] (142) to (147);
\end{pgfonlayer} {edgelayer}
\end{tikzpicture}
\end{center}
}}
\end{proof}
In the next result we show that for all $m \equiv 2\;\mod\;6$ and $m \equiv 4\;\mod\;6$, $m\geq 14$, $\lambda_1^1(C_m \times C_8)=6$
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-4,12) {$Fig. \; 11$: $5-L(1,1)$-Labeling of $C_{6} \times C_8$};
\node [minimum size=0cm,draw,circle] (0) at (-2,13) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-3,13) {$ 4$};
\node [minimum size=0cm,draw,circle] (2) at (-4,13) {$ 0$};
\node [minimum size=0cm,draw,circle] (3) at (-5,13) {$ 4$};
\node [minimum size=0cm,draw,circle] (4) at (-6,13) {$ 0$};
\node [minimum size=0cm,draw,circle] (6) at (-2.5,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (7) at (-3.5,13.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (8) at (-4.5,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (9) at (-5.5,13.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (11) at (-2,14) {$ 3$};
\node [minimum size=0cm,draw,circle] (12) at (-3,14) {$ 1$};
\node [minimum size=0cm,draw,circle] (13) at (-4,14) {$ 3$};
\node [minimum size=0cm,draw,circle] (14) at (-5,14) {$ 1$};
\node [minimum size=0cm,draw,circle] (15) at (-6,14) {$ 3$};
\node [minimum size=0cm,draw,circle] (17) at (-2.5,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (18) at (-3.5,14.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (19) at (-4.5,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (20) at (-5.5,14.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (22) at (-2,15) {$ 2$};
\node [minimum size=0cm,draw,circle] (23) at (-3,15) {$ 5$};
\node [minimum size=0cm,draw,circle] (24) at (-4,15) {$ 2$};
\node [minimum size=0cm,draw,circle] (25) at (-5,15) {$ 5$};
\node [minimum size=0cm,draw,circle] (26) at (-6,15) {$ 2$};
\node [minimum size=0cm,draw,circle] (28) at (-2.5,15.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (29) at (-3.5,15.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (30) at (-4.5,15.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (31) at (-5.5,15.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (33) at (-2,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (34) at (-3,16) {$ 4$};
\node [minimum size=0cm,draw,circle] (35) at (-4,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (36) at (-5,16) {$ 4$};
\node [minimum size=0cm,draw,circle] (37) at (-6,16) {$ 0$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (6);
\draw [thin=1.00] (1) to (6);
\draw [thin=1.00] (1) to (7);
\draw [thin=1.00] (2) to (7);
\draw [thin=1.00] (2) to (8);
\draw [thin=1.00] (3) to (8);
\draw [thin=1.00] (3) to (9);
\draw [thin=1.00] (4) to (9);
\draw [thin=1.00] (6) to (11);
\draw [thin=1.00] (6) to (12);
\draw [thin=1.00] (7) to (12);
\draw [thin=1.00] (7) to (13);
\draw [thin=1.00] (8) to (13);
\draw [thin=1.00] (8) to (14);
\draw [thin=1.00] (9) to (14);
\draw [thin=1.00] (9) to (15);
\draw [thin=1.00] (11) to (17);
\draw [thin=1.00] (12) to (17);
\draw [thin=1.00] (12) to (18);
\draw [thin=1.00] (13) to (18);
\draw [thin=1.00] (13) to (19);
\draw [thin=1.00] (14) to (19);
\draw [thin=1.00] (14) to (20);
\draw [thin=1.00] (15) to (20);
\draw [thin=1.00] (17) to (22);
\draw [thin=1.00] (17) to (23);
\draw [thin=1.00] (18) to (23);
\draw [thin=1.00] (18) to (24);
\draw [thin=1.00] (19) to (24);
\draw [thin=1.00] (19) to (25);
\draw [thin=1.00] (20) to (25);
\draw [thin=1.00] (20) to (26);
\draw [thin=1.00] (22) to (28);
\draw [thin=1.00] (23) to (28);
\draw [thin=1.00] (23) to (29);
\draw [thin=1.00] (24) to (29);
\draw [thin=1.00] (24) to (30);
\draw [thin=1.00] (25) to (30);
\draw [thin=1.00] (25) to (31);
\draw [thin=1.00] (26) to (31);
\draw [thin=1.00] (28) to (33);
\draw [thin=1.00] (28) to (34);
\draw [thin=1.00] (29) to (34);
\draw [thin=1.00] (29) to (35);
\draw [thin=1.00] (30) to (35);
\draw [thin=1.00] (30) to (36);
\draw [thin=1.00] (31) to (36);
\draw [thin=1.00] (31) to (37);
\end{pgfonlayer} {edgelayer}
\end{tikzpicture}
\end{center}
}}
\begin{thm}\label{cc29} Let $m \not\equiv 0\; \mod\; 6$, $m\geq 10$ and even. Then $\lambda_1^1(C_m \times C_8)=6$.
\end{thm}
\begin{proof}By Corollary \ref{cc26}, we see that for $m \not\equiv 0\; \mod\; 6$, $m\geq 16$, $\lambda_1^1(C_m \times C_8) \geq 6$. Now, by combining the $m'-$copies of the labeling in Figure 1, with the $n'-$copies of labeling in Figure 11, $m', n' \in \mathbb N$ we have that $\lambda_1^1(C_{10m'+6n'} \times C_8) \leq 6$, with $10m'+6n' \equiv 4\;\mod\;6$. By combining the labeling in Figure 10 with the $n'-$copies labeling Figure 11, $n' \geq 1$, $n' \in \mathbb N$ we have that $\lambda_1^1(C_{14+6n'} \times C_8) \leq 6$, with $14+6n' \equiv 4\;\mod\;6$. Thus, $C_m \times C_8 \leq 6$ for all $m\geq 16, m \not\equiv 0\;\mod\;6.$ Note that if $n', m'=0$, then we have the $L(1,1)$-labeling of $C_m \times C_8$, where $m\in \left\{10,14\right\}$, which are done in Theorem \ref{cc28}
\end{proof}
In what comes next, we obtain the $\lambda_1^1(C_m \times C_{10})$. Our result will be based on that of $P_m \times C_n$.
\begin{lem} For all $m\geq 9, \ n\geq 12$, $\lambda_1^1(C_m \times C_n) \geq 5$.
\end{lem}
\begin{proof} It is easy to see that $P_m \times C_n \subseteq C_m \times C_n$. Therefore the claim follows from Lemma \ref{f}.
\end{proof}
Now that the lower bound has been shown for $C_m \times C_n$, for specific lengths of cycles, we proceed to establish the optimal $L(1,1)$-numbers for various graphs in this class. In the case of $C_m \times C_{10}$, see Theorem \ref{ccc20}.
\begin{thm} For $m\geq 3$, $\lambda_1^1(C_m\times C_{12})=5$
\end{thm}
\begin{proof}
For all $m \equiv 0\mod 4,$ or $m \equiv 0\mod 6$, and by commutativity of $C_m \times C_n$ the claim follows from Corollary \ref{h}. We now need to show the result for $m\not\equiv 0\mod 4$, $m\not\equiv 0\mod6$. It is easy to see that such number, $m'$, is obtainable from this formula: $m'=p+2,$ where $p\in \mathbb{N},\ p \equiv 0 \mod 4, 0 \mod 6$. The first of such number is 14. We need a $5-$labeling of $C_{14} \times C_{12}$.
{\tiny{
\begin{center}
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] at (-5,16) {$Fig. \; 12$: $5-L(1,1)$-Labeling of $C_{14} \times C_{12}$};
\node [minimum size=0cm,draw,circle] (0) at (-2,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-3,17) {$ 1$};
\node [minimum size=0cm,draw,circle] (2) at (-4,17) {$ 2$};
\node [minimum size=0cm,draw,circle] (3) at (-5,17) {$ 3$};
\node [minimum size=0cm,draw,circle] (4) at (-6,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (5) at (-7,17) {$ 5$};
\node [minimum size=0cm,draw,circle] (6) at (-8,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (7) at (-2.5,17.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (8) at (-3.5,17.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (9) at (-4.5,17.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (10) at (-5.5,17.5){$ 5$};
\node [minimum size=0cm,draw,circle] (11) at (-6.5,17.5){$ 0$};
\node [minimum size=0cm,draw,circle] (12) at (-7.5,17.5){$ 1$};
\node [minimum size=0cm,draw,circle] (13) at (-2,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (14) at (-3,18) {$ 4$};
\node [minimum size=0cm,draw,circle] (15) at (-4,18) {$ 5$};
\node [minimum size=0cm,draw,circle] (16) at (-5,18) {$ 0$};
\node [minimum size=0cm,draw,circle] (17) at (-6,18) {$ 1$};
\node [minimum size=0cm,draw,circle] (18) at (-7,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (19) at (-8,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (20) at (-2.5,18.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (21) at (-3.5,18.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (22) at (-4.5,18.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (23) at (-5.5,18.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (24) at (-6.5,18.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (25) at (-7.5,18.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (26) at (-2,19) {$ 0$};
\node [minimum size=0cm,draw,circle] (27) at (-3,19) {$ 1$};
\node [minimum size=0cm,draw,circle] (28) at (-4,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (29) at (-5,19) {$ 3$};
\node [minimum size=0cm,draw,circle] (30) at (-6,19) {$ 4$};
\node [minimum size=0cm,draw,circle] (31) at (-7,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (32) at (-8,19) {$ 0$};
\node [minimum size=0cm,draw,circle] (33) at (-2.5,19.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (34) at (-3.5,19.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (35) at (-4.5,19.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (36) at (-5.5,19.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (37) at (-6.5,19.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (38) at (-7.5,19.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (39) at (-2,20) {$ 4$};
\node [minimum size=0cm,draw,circle] (40) at (-3,20) {$ 6$};
\node [minimum size=0cm,draw,circle] (41) at (-4,20) {$ 0$};
\node [minimum size=0cm,draw,circle] (42) at (-5,20) {$ 1$};
\node [minimum size=0cm,draw,circle] (43) at (-6,20) {$ 2$};
\node [minimum size=0cm,draw,circle] (44) at (-7,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (45) at (-8,20) {$ 4$};
\node [minimum size=0cm,draw,circle] (46) at (-2.5,20.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (47) at (-3.5,20.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (48) at (-4.5,20.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (49) at (-5.5,20.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (50) at (-6.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (51) at (-7.5,20.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (52) at (-2,21) {$ 1$};
\node [minimum size=0cm,draw,circle] (53) at (-3,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (54) at (-4,21) {$ 3$};
\node [minimum size=0cm,draw,circle] (55) at (-5,21) {$ 4$};
\node [minimum size=0cm,draw,circle] (56) at (-6,21) {$ 5$};
\node [minimum size=0cm,draw,circle] (57) at (-7,21) {$ 4$};
\node [minimum size=0cm,draw,circle] (58) at (-8,21) {$ 1$};
\node [minimum size=0cm,draw,circle] (59) at (-2.5,21.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (60) at (-3.5,21.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (61) at (-4.5,21.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (62) at (-5.5,21.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (63) at (-6.5,21.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (64) at (-7.5,21.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (65) at (-2,22) {$ 5$};
\node [minimum size=0cm,draw,circle] (66) at (-3,22) {$ 0$};
\node [minimum size=0cm,draw,circle] (67) at (-4,22) {$ 1$};
\node [minimum size=0cm,draw,circle] (68) at (-5,22) {$ 2$};
\node [minimum size=0cm,draw,circle] (69) at (-6,22) {$ 3$};
\node [minimum size=0cm,draw,circle] (70) at (-7,22) {$ 4$};
\node [minimum size=0cm,draw,circle] (71) at (-8,22) {$ 5$};
\node [minimum size=0cm,draw,circle] (72) at (-2.5,22.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (73) at (-3.5,22.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (74) at (-4.5,22.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (75) at (-5.5,22.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (76) at (-6.5,22.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (77) at (-7.5,22.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (78) at (-2,23) {$ 2$};
\node [minimum size=0cm,draw,circle] (79) at (-3,23) {$ 3$};
\node [minimum size=0cm,draw,circle] (80) at (-4,23) {$ 4$};
\node [minimum size=0cm,draw,circle] (81) at (-5,23) {$ 5$};
\node [minimum size=0cm,draw,circle] (82) at (-6,23) {$ 0$};
\node [minimum size=0cm,draw,circle] (83) at (-7,23) {$ 1$};
\node [minimum size=0cm,draw,circle] (84) at (-8,23) {$ 2$};
\node [minimum size=0cm,draw,circle] (85) at (-2.5,23.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (86) at (-3.5,23.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (87) at (-4.5,23.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (88) at (-5.5,23.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (89) at (-6.5,23.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (90) at (-7.5,23.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (91) at (-2,24) {$ 0$};
\node [minimum size=0cm,draw,circle] (92) at (-3,24) {$ 1$};
\node [minimum size=0cm,draw,circle] (93) at (-4,24) {$ 2$};
\node [minimum size=0cm,draw,circle] (94) at (-5,24) {$ 3$};
\node [minimum size=0cm,draw,circle] (95) at (-6,24) {$ 4$};
\node [minimum size=0cm,draw,circle] (96) at (-7,24) {$ 5$};
\node [minimum size=0cm,draw,circle] (97) at (-8,24) {$ 0$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (7);
\draw [thin=1.00] (1) to (7);
\draw [thin=1.00] (1) to (8);
\draw [thin=1.00] (2) to (8);
\draw [thin=1.00] (2) to (9);
\draw [thin=1.00] (3) to (9);
\draw [thin=1.00] (3) to (10);
\draw [thin=1.00] (4) to (10);
\draw [thin=1.00] (4) to (11);
\draw [thin=1.00] (5) to (11);
\draw [thin=1.00] (5) to (12);
\draw [thin=1.00] (6) to (12);
\draw [thin=1.00] (7) to (13);
\draw [thin=1.00] (7) to (14);
\draw [thin=1.00] (8) to (14);
\draw [thin=1.00] (8) to (15);
\draw [thin=1.00] (9) to (15);
\draw [thin=1.00] (9) to (16);
\draw [thin=1.00] (10) to (16);
\draw [thin=1.00] (10) to (17);
\draw [thin=1.00] (11) to (17);
\draw [thin=1.00] (11) to (18);
\draw [thin=1.00] (12) to (18);
\draw [thin=1.00] (12) to (19);
\draw [thin=1.00] (13) to (20);
\draw [thin=1.00] (14) to (20);
\draw [thin=1.00] (14) to (21);
\draw [thin=1.00] (15) to (21);
\draw [thin=1.00] (15) to (22);
\draw [thin=1.00] (16) to (22);
\draw [thin=1.00] (16) to (23);
\draw [thin=1.00] (17) to (23);
\draw [thin=1.00] (17) to (24);
\draw [thin=1.00] (18) to (24);
\draw [thin=1.00] (18) to (25);
\draw [thin=1.00] (19) to (25);
\draw [thin=1.00] (20) to (26);
\draw [thin=1.00] (20) to (27);
\draw [thin=1.00] (21) to (27);
\draw [thin=1.00] (21) to (28);
\draw [thin=1.00] (22) to (28);
\draw [thin=1.00] (22) to (29);
\draw [thin=1.00] (23) to (29);
\draw [thin=1.00] (23) to (30);
\draw [thin=1.00] (24) to (30);
\draw [thin=1.00] (24) to (31);
\draw [thin=1.00] (25) to (31);
\draw [thin=1.00] (25) to (32);
\draw [thin=1.00] (26) to (33);
\draw [thin=1.00] (27) to (33);
\draw [thin=1.00] (27) to (34);
\draw [thin=1.00] (28) to (34);
\draw [thin=1.00] (28) to (35);
\draw [thin=1.00] (29) to (35);
\draw [thin=1.00] (29) to (36);
\draw [thin=1.00] (30) to (36);
\draw [thin=1.00] (30) to (37);
\draw [thin=1.00] (31) to (37);
\draw [thin=1.00] (31) to (38);
\draw [thin=1.00] (32) to (38);
\draw [thin=1.00] (33) to (39);
\draw [thin=1.00] (33) to (40);
\draw [thin=1.00] (34) to (40);
\draw [thin=1.00] (34) to (41);
\draw [thin=1.00] (35) to (41);
\draw [thin=1.00] (35) to (42);
\draw [thin=1.00] (36) to (42);
\draw [thin=1.00] (36) to (43);
\draw [thin=1.00] (37) to (43);
\draw [thin=1.00] (37) to (44);
\draw [thin=1.00] (38) to (44);
\draw [thin=1.00] (38) to (45);
\draw [thin=1.00] (39) to (46);
\draw [thin=1.00] (40) to (46);
\draw [thin=1.00] (40) to (47);
\draw [thin=1.00] (41) to (47);
\draw [thin=1.00] (41) to (48);
\draw [thin=1.00] (42) to (48);
\draw [thin=1.00] (42) to (49);
\draw [thin=1.00] (43) to (49);
\draw [thin=1.00] (43) to (50);
\draw [thin=1.00] (44) to (50);
\draw [thin=1.00] (44) to (51);
\draw [thin=1.00] (45) to (51);
\draw [thin=1.00] (46) to (52);
\draw [thin=1.00] (46) to (53);
\draw [thin=1.00] (47) to (53);
\draw [thin=1.00] (47) to (54);
\draw [thin=1.00] (48) to (54);
\draw [thin=1.00] (48) to (55);
\draw [thin=1.00] (49) to (55);
\draw [thin=1.00] (49) to (56);
\draw [thin=1.00] (50) to (56);
\draw [thin=1.00] (50) to (57);
\draw [thin=1.00] (51) to (57);
\draw [thin=1.00] (51) to (58);
\draw [thin=1.00] (52) to (59);
\draw [thin=1.00] (53) to (59);
\draw [thin=1.00] (53) to (60);
\draw [thin=1.00] (54) to (60);
\draw [thin=1.00] (54) to (61);
\draw [thin=1.00] (55) to (61);
\draw [thin=1.00] (55) to (62);
\draw [thin=1.00] (56) to (62);
\draw [thin=1.00] (56) to (63);
\draw [thin=1.00] (57) to (63);
\draw [thin=1.00] (57) to (64);
\draw [thin=1.00] (58) to (64);
\draw [thin=1.00] (59) to (65);
\draw [thin=1.00] (59) to (66);
\draw [thin=1.00] (60) to (66);
\draw [thin=1.00] (60) to (67);
\draw [thin=1.00] (61) to (67);
\draw [thin=1.00] (61) to (68);
\draw [thin=1.00] (62) to (68);
\draw [thin=1.00] (62) to (69);
\draw [thin=1.00] (63) to (69);
\draw [thin=1.00] (63) to (70);
\draw [thin=1.00] (64) to (70);
\draw [thin=1.00] (64) to (71);
\draw [thin=1.00] (65) to (72);
\draw [thin=1.00] (66) to (72);
\draw [thin=1.00] (66) to (73);
\draw [thin=1.00] (67) to (73);
\draw [thin=1.00] (67) to (74);
\draw [thin=1.00] (68) to (74);
\draw [thin=1.00] (68) to (75);
\draw [thin=1.00] (69) to (75);
\draw [thin=1.00] (69) to (76);
\draw [thin=1.00] (70) to (76);
\draw [thin=1.00] (70) to (77);
\draw [thin=1.00] (71) to (77);
\draw [thin=1.00] (72) to (78);
\draw [thin=1.00] (72) to (79);
\draw [thin=1.00] (73) to (79);
\draw [thin=1.00] (73) to (80);
\draw [thin=1.00] (74) to (80);
\draw [thin=1.00] (74) to (81);
\draw [thin=1.00] (75) to (81);
\draw [thin=1.00] (75) to (82);
\draw [thin=1.00] (76) to (82);
\draw [thin=1.00] (76) to (83);
\draw [thin=1.00] (77) to (83);
\draw [thin=1.00] (77) to (84);
\draw [thin=1.00] (78) to (85);
\draw [thin=1.00] (79) to (85);
\draw [thin=1.00] (79) to (86);
\draw [thin=1.00] (80) to (86);
\draw [thin=1.00] (80) to (87);
\draw [thin=1.00] (81) to (87);
\draw [thin=1.00] (81) to (88);
\draw [thin=1.00] (82) to (88);
\draw [thin=1.00] (82) to (89);
\draw [thin=1.00] (83) to (89);
\draw [thin=1.00] (83) to (90);
\draw [thin=1.00] (84) to (90);
\draw [thin=1.00] (85) to (91);
\draw [thin=1.00] (85) to (92);
\draw [thin=1.00] (86) to (92);
\draw [thin=1.00] (86) to (93);
\draw [thin=1.00] (87) to (93);
\draw [thin=1.00] (87) to (94);
\draw [thin=1.00] (88) to (94);
\draw [thin=1.00] (88) to (95);
\draw [thin=1.00] (89) to (95);
\draw [thin=1.00] (89) to (96);
\draw [thin=1.00] (90) to (96);
\draw [thin=1.00] (90) to (97);
\end{pgfonlayer} {edgelayer}
\end{tikzpicture}
\end{center}
}}
Note that $\cup_{i=0}^4\left\{V_i\right\}$ in Figure 12 above forms a component of a $C_4 \times C_{12}$ and $L(\cup_{i=0}^4\left\{V_i\right\})=[5]$. Therefore there exist an independent $5-L(1,1)$-labeling of $C_4 \times C_{12}$ in the $L(1,1)$-labeling of $C_{14} \times C_{12}$.
Thus $\lambda_1^1(C_{10+4m'} \times C_{12})=5$, $m'\in \mathbb N$. Since for all $p\in \mathbb N$, $p$ can be expressed as $10+4m'$, then the required result holds.
\end{proof}
\begin{cor}
For all $m\equiv 0\; mod \; 14$, $n\equiv 0 \; mod\; 12$, $m,n \not\equiv 0\; mod \; 5,$ $\lambda_1^1(C_m \times C_n)=5$.
\end{cor}
The next result establishes an optimal $L(1,1)$-labeling of $C_m \times C_n$ of a certain size. This resolves all cases of large enough $m$ and $n$.
\begin{thm}For $m',m'', n', n'' \in \mathbb{Z}_+$ $\lambda_1^1(C_{10m'+14m''} \times C_{10n'+12n''})= 5$.
\end{thm}
\begin{proof}
From earlier results, $\lambda_1^1(C_m \times C_n) \geq 5$ for $C_m \times C_n$ defined in the statement above. Each of the quadrant in Figure $13$ represents special $5-L(1,1)$-labelings of $C_{10} \times C_{10}$, $C_{10} \times C_{12}$, $C_{10} \times C_{14}$ and $C_{12} \times C_{14}$ respectively. Clearly, these labelings form a $5-L(1,1)$-labeling of $C_{10+14} \times C_{10+12}$. Thus, for $m',m'', n', n'' \in \mathbb{Z}_+$, $\lambda_1^1(C_{10m'+14m''} \times C_{10n'+12n''})\leq 5$.
\newpage
{\tiny{
\pgfdeclarelayer{nodelayer}
\pgfdeclarelayer{edgelayer}
\pgfsetlayers{nodelayer,edgelayer}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [minimum size=0cm,] (0e) at (-7,15.25) {};
\node [minimum size=0cm,draw,circle] (0) at (-2,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (1) at (-3,16) {$ 1$};
\node [minimum size=0cm,draw,circle] (2) at (-4,16) {$ 2$};
\node [minimum size=0cm,draw,circle] (3) at (-5,16) {$ 4$};
\node [minimum size=0cm,draw,circle] (4) at (-6,16) {$ 5$};
\node [minimum size=0cm,draw,circle] (5) at (-7,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (6) at (-2.5,16.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (7) at (-3.5,16.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (8) at (-4.5,16.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (9) at (-5.5,16.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (10) at (-6.5,16.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (11) at (-2,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (12) at (-3,17) {$ 5$};
\node [minimum size=0cm,draw,circle] (13) at (-4,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (14) at (-5,17) {$ 1$};
\node [minimum size=0cm,draw,circle] (15) at (-6,17) {$ 3$};
\node [minimum size=0cm,draw,circle] (16) at (-7,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (17) at (-2.5,17.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (18) at (-3.5,17.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (19) at (-4.5,17.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (20) at (-5.5,17.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (21) at (-6.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (22) at (-2,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (23) at (-3,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (24) at (-4,18) {$ 4$};
\node [minimum size=0cm,draw,circle] (25) at (-5,18) {$ 5$};
\node [minimum size=0cm,draw,circle] (26) at (-6,18) {$ 0$};
\node [minimum size=0cm,draw,circle] (27) at (-7,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (28) at (-2.5,18.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (29) at (-3.5,18.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (30) at (-4.5,18.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (31) at (-5.5,18.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (32) at (-6.5,18.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (33) at (-2,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (34) at (-3,19) {$ 0$};
\node [minimum size=0cm,draw,circle] (35) at (-4,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (36) at (-5,19) {$ 3$};
\node [minimum size=0cm,draw,circle] (37) at (-6,19) {$ 4$};
\node [minimum size=0cm,draw,circle] (38) at (-7,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (39) at (-2.5,19.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (40) at (-3.5,19.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (41) at (-4.5,19.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (42) at (-5.5,19.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (43) at (-6.5,19.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (44) at (-2,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (45) at (-3,20) {$ 4$};
\node [minimum size=0cm,draw,circle] (46) at (-4,20) {$ 5$};
\node [minimum size=0cm,draw,circle] (47) at (-5,20) {$ 0$};
\node [minimum size=0cm,draw,circle] (48) at (-6,20) {$ 2$};
\node [minimum size=0cm,draw,circle] (49) at (-7,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (50) at (-2.5,20.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (51) at (-3.5,20.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (52) at (-4.5,20.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (53) at (-5.5,20.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (54) at (-6.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (55') at (-2,21) {$ 0$};
\node [minimum size=0cm,draw,circle] (56') at (-3,21) {$ 1$};
\node [minimum size=0cm,draw,circle] (57') at (-4,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (58') at (-5,21) {$ 4$};
\node [minimum size=0cm,draw,circle] (59') at (-6,21) {$ 5$};
\node [minimum size=0cm,draw,circle] (60') at (-7,21) {$ 0$};
\node [minimum size=0cm,] (0g) at (-1,21) {};
\node [minimum size=0cm,] (0gg) at (-4.5,15.5) {$C_{10} \times C_{10}$};
\node [minimum size=0cm,draw,circle] (55a) at (-2,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (56a) at (-3,14.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (57a) at (-4,14.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (58a) at (-5,14.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (59a) at (-6,14.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (60a) at (-7,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (61a) at (-2.5,14) {$ 5$};
\node [minimum size=0cm,draw,circle] (62a) at (-3.5,14) {$ 0$};
\node [minimum size=0cm,draw,circle] (63a) at (-4.5,14) {$ 1$};
\node [minimum size=0cm,draw,circle] (64a) at (-5.5,14) {$ 2$};
\node [minimum size=0cm,draw,circle] (65a) at (-6.5,14) {$ 4$};
\node [minimum size=0cm,draw,circle] (66a) at (-2,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (67a) at (-3,13.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (68a) at (-4,13.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (69a) at (-5,13.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (70a) at (-6,13.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (71a) at (-7,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (72a) at (-2.5,13) {$ 1$};
\node [minimum size=0cm,draw,circle] (73a) at (-3.5,13) {$ 2$};
\node [minimum size=0cm,draw,circle] (74a) at (-4.5,13) {$ 4$};
\node [minimum size=0cm,draw,circle] (75a) at (-5.5,13) {$ 3$};
\node [minimum size=0cm,draw,circle] (76a) at (-6.5,13) {$ 0$};
\node [minimum size=0cm,draw,circle] (77a) at (-2,12.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (78a) at (-3,12.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (79a) at (-4,12.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (80a) at (-5,12.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (81a) at (-6,12.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (82a) at (-7,12.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (83a) at (-2.5,12) {$ 4$};
\node [minimum size=0cm,draw,circle] (84a) at (-3.5,12) {$ 5$};
\node [minimum size=0cm,draw,circle] (85a) at (-4.5,12) {$ 0$};
\node [minimum size=0cm,draw,circle] (86a) at (-5.5,12) {$ 1$};
\node [minimum size=0cm,draw,circle] (87a) at (-6.5,12) {$ 2$};
\node [minimum size=0cm,draw,circle] (88a) at (-2,11.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (89a) at (-3,11.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (90a) at (-4,11.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (91a) at (-5,11.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (92a) at (-6,11.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (93a) at (-7,11.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (94a) at (-2.5,11) {$ 0$};
\node [minimum size=0cm,draw,circle] (95a) at (-3.5,11) {$ 1$};
\node [minimum size=0cm,draw,circle] (96a) at (-4.5,11) {$ 2$};
\node [minimum size=0cm,draw,circle] (97a) at (-5.5,11) {$ 4$};
\node [minimum size=0cm,draw,circle] (98a) at (-6.5,11) {$ 5$};
\node [minimum size=0cm,draw,circle] (99a) at (-2,10.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (100a) at (-3,10.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (101a) at (-4,10.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (102a) at (-5,10.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (103a) at (-6,10.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (104a) at (-7,10.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (105a) at (-2.5,10) {$ 3$};
\node [minimum size=0cm,draw,circle] (106a) at (-3.5,10) {$ 4$};
\node [minimum size=0cm,draw,circle] (107a) at (-4.5,10) {$ 5$};
\node [minimum size=0cm,draw,circle] (108a) at (-5.5,10) {$ 0$};
\node [minimum size=0cm,draw,circle] (109a) at (-6.5,10) {$ 1$};
\node [minimum size=0cm,draw,circle] (110a) at (-2,9.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (111a) at (-3,9.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (112a) at (-4,9.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (113a) at (-5,9.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (114a) at (-6,9.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (115a) at (-7,9.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (116a) at (-2.5,9) {$ 5$};
\node [minimum size=0cm,draw,circle] (117a) at (-3.5,9) {$ 1$};
\node [minimum size=0cm,draw,circle] (118a) at (-4.5,9) {$ 2$};
\node [minimum size=0cm,draw,circle] (119a) at (-5.5,9) {$ 3$};
\node [minimum size=0cm,draw,circle] (120a) at (-6.5,9) {$ 4$};
\node [minimum size=0cm,draw,circle] (121a) at (-2,8.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (122a) at (-3,8.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (123a) at (-4,8.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (124a) at (-5,8.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (125a) at (-6,8.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (126a) at (-7,8.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (127a) at (-2.5,8) {$ 2$};
\node [minimum size=0cm,draw,circle] (128a) at (-3.5,8) {$ 3$};
\node [minimum size=0cm,draw,circle] (129a) at (-4.5,8) {$ 5$};
\node [minimum size=0cm,draw,circle] (130a) at (-5.5,8) {$ 0$};
\node [minimum size=0cm,draw,circle] (131a) at (-6.5,8) {$ 1$};
\node [minimum size=0cm,] (0d) at (-1,7.5) {};
\node [minimum size=0cm,draw,circle] (132a) at (-2,7.5) {$0$};
\node [minimum size=0cm,draw,circle] (133a) at (-3,7.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (134a) at (-4,7.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (135a) at (-5,7.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (136a) at (-6,7.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (137a) at (-7,7.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (55) at (0,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (56) at (1,16) {$ 5$};
\node [minimum size=0cm,draw,circle] (57) at (2,16) {$ 4$};
\node [minimum size=0cm,draw,circle] (58) at (3,16) {$ 3$};
\node [minimum size=0cm,draw,circle] (59) at (4,16) {$ 2$};
\node [minimum size=0cm,draw,circle] (60) at (5,16) {$ 1$};
\node [minimum size=0cm,draw,circle] (60b) at (6,16) {$ 0$};
\node [minimum size=0cm,draw,circle] (61) at (0.5,16.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (62) at (1.5,16.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (63) at (2.5,16.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (64) at (3.5,16.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (65) at (4.5,16.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (65b) at (5.5,16.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (66) at (0,17) {$ 4$};
\node [minimum size=0cm,draw,circle] (67) at (1,17) {$ 3$};
\node [minimum size=0cm,draw,circle] (68) at (2,17) {$ 2$};
\node [minimum size=0cm,draw,circle] (69) at (3,17) {$ 1$};
\node [minimum size=0cm,draw,circle] (70) at (4,17) {$ 0$};
\node [minimum size=0cm,draw,circle] (71) at (5,17) {$ 5$};
\node [minimum size=0cm,draw,circle] (71b) at (6,17){$ 4$};
\node [minimum size=0cm,draw,circle] (72) at (0.5,17.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (73) at (1.5,17.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (74) at (2.5,17.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (75) at (3.5,17.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (76) at (4.5,17.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (76b) at (5.5,17.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (77) at (0,18) {$ 2$};
\node [minimum size=0cm,draw,circle] (78) at (1,18) {$ 1$};
\node [minimum size=0cm,draw,circle] (79) at (2,18) {$ 0$};
\node [minimum size=0cm,draw,circle] (80) at (3,18) {$ 5$};
\node [minimum size=0cm,draw,circle] (81) at (4,18) {$ 4$};
\node [minimum size=0cm,draw,circle] (82) at (5,18) {$ 3$};
\node [minimum size=0cm,draw,circle] (82b) at (6,18){$ 2$};
\node [minimum size=0cm,draw,circle] (83) at (0.5,18.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (84) at (1.5,18.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (85) at (2.5,18.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (86) at (3.5,18.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (87) at (4.5,18.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (87b) at (5.5,18.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (88) at (0,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (89) at (1,19) {$ 4$};
\node [minimum size=0cm,draw,circle] (90) at (2,19) {$ 3$};
\node [minimum size=0cm,draw,circle] (91) at (3,19) {$ 2$};
\node [minimum size=0cm,draw,circle] (92) at (4,19) {$ 1$};
\node [minimum size=0cm,draw,circle] (93) at (5,19) {$ 0$};
\node [minimum size=0cm,draw,circle] (93b) at (6,19) {$ 5$};
\node [minimum size=0cm,draw,circle] (94) at (0.5,19.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (95) at (1.5,19.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (96) at (2.5,19.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (97) at (3.5,19.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (98) at (4.5,19.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (98b) at (5.5,19.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (99) at (0,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (100) at (1,20) {$ 2$};
\node [minimum size=0cm,draw,circle] (101) at (2,20) {$ 1$};
\node [minimum size=0cm,draw,circle] (102) at (3,20) {$ 0$};
\node [minimum size=0cm,draw,circle] (103) at (4,20) {$ 5$};
\node [minimum size=0cm,draw,circle] (104) at (5,20) {$ 4$};
\node [minimum size=0cm,draw,circle] (104b) at (6,20) {$ 3$};
\node [minimum size=0cm,draw,circle] (105) at (0.5,20.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (106) at (1.5,20.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (107) at (2.5,20.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (108) at (3.5,20.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (109) at (4.5,20.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (109b) at (5.5,20.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (110) at (0,21) {$ 0$};
\node [minimum size=0cm,draw,circle] (111) at (1,21) {$ 5$};
\node [minimum size=0cm,draw,circle] (112) at (2,21) {$ 4$};
\node [minimum size=0cm,draw,circle] (113) at (3,21) {$ 3$};
\node [minimum size=0cm,draw,circle] (114) at (4,21) {$ 2$};
\node [minimum size=0cm,draw,circle] (115) at (5,21) {$ 1$};
\node [minimum size=0cm,draw,circle] (115b) at (6,21) {$ 0$};
\node [minimum size=0cm,] (115cc) at (3,15.5) {$C_{10}\times C_{12}$};
\node [minimum size=0cm,] (0f) at (6,15.25) {};
\node [minimum size=0cm,draw,circle] (0c) at (0,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (1c) at (1,14.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (2c) at (2,14.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (3c) at (3,14.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (4c) at (4,14.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (5c) at (5,14.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (6c) at (6,14.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (7c) at (0.5,14) {$ 4$};
\node [minimum size=0cm,draw,circle] (8c) at (1.5,14) {$ 3$};
\node [minimum size=0cm,draw,circle] (9c) at (2.5,14) {$ 2$};
\node [minimum size=0cm,draw,circle] (10c) at (3.5,14) {$ 1$};
\node [minimum size=0cm,draw,circle] (11c) at (4.5,14) {$ 0$};
\node [minimum size=0cm,draw,circle] (12c) at (5.5,14){$ 5$};
\node [minimum size=0cm,draw,circle] (13c) at (0,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (14c) at (1,13.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (15c) at (2,13.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (16c) at (3,13.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (17c) at (4,13.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (18c) at (5,13.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (19c) at (6,13.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (20c) at (0.5,13) {$0 $};
\node [minimum size=0cm,draw,circle] (21c) at (1.5,13) {$ 5$};
\node [minimum size=0cm,draw,circle] (22c) at (2.5,13) {$ 4$};
\node [minimum size=0cm,draw,circle] (23c) at (3.5,13) {$ 3$};
\node [minimum size=0cm,draw,circle] (24c) at (4.5,13) {$ 2$};
\node [minimum size=0cm,draw,circle] (25c) at (5.5,13){$ $1};
\node [minimum size=0cm,draw,circle] (26c) at (0,12.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (27c) at (1,12.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (28c) at (2,12.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (29c) at (3,12.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (30c) at (4,12.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (31c) at (5,12.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (32c) at (6,12.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (33c) at (0.5,12) {$ 3$};
\node [minimum size=0cm,draw,circle] (34c) at (1.5,12) {$ 2$};
\node [minimum size=0cm,draw,circle] (35c) at (2.5,12) {$ 1$};
\node [minimum size=0cm,draw,circle] (36c) at (3.5,12) {$ 0$};
\node [minimum size=0cm,draw,circle] (37c) at (4.5,12) {$ 5$};
\node [minimum size=0cm,draw,circle] (38c) at (5.5,12) {$ 4$};
\node [minimum size=0cm,draw,circle] (39c) at (0,11.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (40c) at (1,11.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (41c) at (2,11.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (42c) at (3,11.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (43c) at (4,11.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (44c) at (5,11.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (45c) at (6,11.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (46c) at (0.5,11) {$ 5$};
\node [minimum size=0cm,draw,circle] (47c) at (1.5,11) {$ 4$};
\node [minimum size=0cm,draw,circle] (48c) at (2.5,11) {$ 3$};
\node [minimum size=0cm,draw,circle] (49c) at (3.5,11) {$ 2$};
\node [minimum size=0cm,draw,circle] (50c) at (4.5,11) {$ 1$};
\node [minimum size=0cm,draw,circle] (51c) at (5.5,11) {$ 0$};
\node [minimum size=0cm,draw,circle] (52c) at (0,10.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (53c) at (1,10.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (54c) at (2,10.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (55c) at (3,10.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (56c) at (4,10.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (57c) at (5,10.5) {$ 6$};
\node [minimum size=0cm,draw,circle] (58c) at (6,10.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (59c) at (0.5,10) {$ 1$};
\node [minimum size=0cm,draw,circle] (60c) at (1.5,10) {$ 0$};
\node [minimum size=0cm,draw,circle] (61c) at (2.5,10) {$ 5$};
\node [minimum size=0cm,draw,circle] (62c) at (3.5,10) {$ 4$};
\node [minimum size=0cm,draw,circle] (63c) at (4.5,10) {$ 3$};
\node [minimum size=0cm,draw,circle] (64c) at (5.5,10) {$ 2$};
\node [minimum size=0cm,draw,circle] (65c) at (0,9.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (66c) at (1,9.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (67c) at (2,9.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (68c) at (3,9.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (69c) at (4,9.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (70c) at (5,9.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (71c) at (6,9.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (72c) at (0.5,9) {$ 4$};
\node [minimum size=0cm,draw,circle] (73c) at (1.5,9) {$ 3$};
\node [minimum size=0cm,draw,circle] (74c) at (2.5,9) {$ 2$};
\node [minimum size=0cm,draw,circle] (75c) at (3.5,9) {$ 1$};
\node [minimum size=0cm,draw,circle] (76c) at (4.5,9) {$ 0$};
\node [minimum size=0cm,draw,circle] (77c) at (5.5,9) {$ 5$};
\node [minimum size=0cm,draw,circle] (78c) at (0,8.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (79c) at (1,8.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (80c) at (2,8.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (81c) at (3,8.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (82c) at (4,8.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (83c) at (5,8.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (84c) at (6,8.5){$ 3$};
\node [minimum size=0cm,draw,circle] (85c) at (0.5,8) {$ 1$};
\node [minimum size=0cm,draw,circle] (86c) at (1.5,8) {$ 0$};
\node [minimum size=0cm,draw,circle] (87c) at (2.5,8) {$ 5$};
\node [minimum size=0cm,draw,circle] (88c) at (3.5,8) {$ 4$};
\node [minimum size=0cm,draw,circle] (89c) at (4.5,8) {$ 3$};
\node [minimum size=0cm,draw,circle] (90c) at (5.5,8) {$ 2$};
\node [minimum size=0cm,draw,circle] (91c) at (0,7.5) {$ 0$};
\node [minimum size=0cm,draw,circle] (92c) at (1,7.5) {$ 5$};
\node [minimum size=0cm,draw,circle] (93c) at (2,7.5) {$ 4$};
\node [minimum size=0cm,draw,circle] (94c) at (3,7.5) {$ 3$};
\node [minimum size=0cm,draw,circle] (95c) at (4,7.5) {$ 2$};
\node [minimum size=0cm,draw,circle] (96c) at (5,7.5) {$ 1$};
\node [minimum size=0cm,draw,circle] (97c) at (6,7.5) {$ 0$};
\node [minimum size=0cm,] (97cc) at (3,7.0) { $C_{14}\times C_{12}$};
\node [minimum size=0cm,] (97cc) at (-4.5,7.0) {$C_{14}\times C_{10}$};
\node [minimum size=0cm,] (97h) at (-0.5,6) {Fig. $13$: $5-L(1,1)-$labeling of $C_{10m'+14m''} \times C_{10n'+14n''}$, for all $m',m'', n', n'' \in \left\{0,1\right\}$.};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [thin=1.00] (0) to (6);
\draw [thin=1.00] (1) to (6);
\draw [thin=1.00] (1) to (7);
\draw [thin=1.00] (2) to (7);
\draw [thin=1.00] (2) to (8);
\draw [thin=1.00] (3) to (8);
\draw [thin=1.00] (3) to (9);
\draw [thin=1.00] (4) to (9);
\draw [thin=1.00] (4) to (10);
\draw [thin=1.00] (5) to (10);
\draw [thin=1.00] (6) to (11);
\draw [thin=1.00] (6) to (12);
\draw [thin=1.00] (7) to (12);
\draw [thin=1.00] (7) to (13);
\draw [thin=1.00] (8) to (13);
\draw [thin=1.00] (8) to (14);
\draw [thin=1.00] (9) to (14);
\draw [thin=1.00] (9) to (15);
\draw [thin=1.00] (10) to (15);
\draw [thin=1.00] (10) to (16);
\draw [thin=1.00] (11) to (17);
\draw [thin=1.00] (12) to (17);
\draw [thin=1.00] (12) to (18);
\draw [thin=1.00] (13) to (18);
\draw [thin=1.00] (13) to (19);
\draw [thin=1.00] (14) to (19);
\draw [thin=1.00] (14) to (20);
\draw [thin=1.00] (15) to (20);
\draw [thin=1.00] (15) to (21);
\draw [thin=1.00] (16) to (21);
\draw [thin=1.00] (17) to (22);
\draw [thin=1.00] (17) to (23);
\draw [thin=1.00] (18) to (23);
\draw [thin=1.00] (18) to (24);
\draw [thin=1.00] (19) to (24);
\draw [thin=1.00] (19) to (25);
\draw [thin=1.00] (20) to (25);
\draw [thin=1.00] (20) to (26);
\draw [thin=1.00] (21) to (26);
\draw [thin=1.00] (21) to (27);
\draw [thin=1.00] (22) to (28);
\draw [thin=1.00] (23) to (28);
\draw [thin=1.00] (23) to (29);
\draw [thin=1.00] (24) to (29);
\draw [thin=1.00] (24) to (30);
\draw [thin=1.00] (25) to (30);
\draw [thin=1.00] (25) to (31);
\draw [thin=1.00] (26) to (31);
\draw [thin=1.00] (26) to (32);
\draw [thin=1.00] (27) to (32);
\draw [thin=1.00] (28) to (33);
\draw [thin=1.00] (28) to (34);
\draw [thin=1.00] (29) to (34);
\draw [thin=1.00] (29) to (35);
\draw [thin=1.00] (30) to (35);
\draw [thin=1.00] (30) to (36);
\draw [thin=1.00] (31) to (36);
\draw [thin=1.00] (31) to (37);
\draw [thin=1.00] (32) to (37);
\draw [thin=1.00] (32) to (38);
\draw [thin=1.00] (33) to (39);
\draw [thin=1.00] (34) to (39);
\draw [thin=1.00] (34) to (40);
\draw [thin=1.00] (35) to (40);
\draw [thin=1.00] (35) to (41);
\draw [thin=1.00] (36) to (41);
\draw [thin=1.00] (36) to (42);
\draw [thin=1.00] (37) to (42);
\draw [thin=1.00] (37) to (43);
\draw [thin=1.00] (38) to (43);
\draw [thin=1.00] (39) to (44);
\draw [thin=1.00] (39) to (45);
\draw [thin=1.00] (40) to (45);
\draw [thin=1.00] (40) to (46);
\draw [thin=1.00] (41) to (46);
\draw [thin=1.00] (41) to (47);
\draw [thin=1.00] (42) to (47);
\draw [thin=1.00] (42) to (48);
\draw [thin=1.00] (43) to (48);
\draw [thin=1.00] (43) to (49);
\draw [thin=1.00] (44) to (50);
\draw [thin=1.00] (45) to (50);
\draw [thin=1.00] (45) to (51);
\draw [thin=1.00] (46) to (51);
\draw [thin=1.00] (46) to (52);
\draw [thin=1.00] (47) to (52);
\draw [thin=1.00] (47) to (53);
\draw [thin=1.00] (48) to (53);
\draw [thin=1.00] (48) to (54);
\draw [thin=1.00] (49) to (54);
\draw [thin=1.00] (50) to (55');
\draw [thin=1.00] (50) to (56');
\draw [thin=1.00] (51) to (56');
\draw [thin=1.00] (51) to (57');
\draw [thin=1.00] (52) to (57');
\draw [thin=1.00] (52) to (58');
\draw [thin=1.00] (53) to (58');
\draw [thin=1.00] (53) to (59');
\draw [thin=1.00] (54) to (59');
\draw [thin=1.00] (54) to (60');
\draw [thin=1.00] (55) to (61);
\draw [thin=1.00] (56) to (61);
\draw [thin=1.00] (56) to (62);
\draw [thin=1.00] (57) to (62);
\draw [thin=1.00] (57) to (63);
\draw [thin=1.00] (58) to (63);
\draw [thin=1.00] (58) to (64);
\draw [thin=1.00] (59) to (64);
\draw [thin=1.00] (59) to (65);
\draw [thin=1.00] (60) to (65);
\draw [thin=1.00] (60) to (65b);
\draw [thin=1.00] (60b) to (65b);
\draw [thin=1.00] (61) to (66);
\draw [thin=1.00] (61) to (67);
\draw [thin=1.00] (62) to (67);
\draw [thin=1.00] (62) to (68);
\draw [thin=1.00] (63) to (68);
\draw [thin=1.00] (63) to (69);
\draw [thin=1.00] (64) to (69);
\draw [thin=1.00] (64) to (70);
\draw [thin=1.00] (65) to (70);
\draw [thin=1.00] (65) to (71);
\draw [thin=1.00] (65b) to (71);
\draw [thin=1.00] (65b) to (71b);
\draw [thin=1.00] (66) to (72);
\draw [thin=1.00] (67) to (72);
\draw [thin=1.00] (67) to (73);
\draw [thin=1.00] (68) to (73);
\draw [thin=1.00] (68) to (74);
\draw [thin=1.00] (69) to (74);
\draw [thin=1.00] (69) to (75);
\draw [thin=1.00] (70) to (75);
\draw [thin=1.00] (70) to (76);
\draw [thin=1.00] (71) to (76);
\draw [thin=1.00] (71) to (76b);
\draw [thin=1.00] (71b) to (76b);
\draw [thin=1.00] (72) to (77);
\draw [thin=1.00] (72) to (78);
\draw [thin=1.00] (73) to (78);
\draw [thin=1.00] (73) to (79);
\draw [thin=1.00] (74) to (79);
\draw [thin=1.00] (74) to (80);
\draw [thin=1.00] (75) to (80);
\draw [thin=1.00] (75) to (81);
\draw [thin=1.00] (76) to (81);
\draw [thin=1.00] (76) to (82);
\draw [thin=1.00] (76b) to (82);
\draw [thin=1.00] (76b) to (82b);
\draw [thin=1.00] (77) to (83);
\draw [thin=1.00] (78) to (83);
\draw [thin=1.00] (78) to (84);
\draw [thin=1.00] (79) to (84);
\draw [thin=1.00] (79) to (85);
\draw [thin=1.00] (80) to (85);
\draw [thin=1.00] (80) to (86);
\draw [thin=1.00] (81) to (86);
\draw [thin=1.00] (81) to (87);
\draw [thin=1.00] (82) to (87);
\draw [thin=1.00] (82) to (87b);
\draw [thin=1.00] (82b) to (87b);
\draw [thin=1.00] (83) to (88);
\draw [thin=1.00] (83) to (89);
\draw [thin=1.00] (84) to (89);
\draw [thin=1.00] (84) to (90);
\draw [thin=1.00] (85) to (90);
\draw [thin=1.00] (85) to (91);
\draw [thin=1.00] (86) to (91);
\draw [thin=1.00] (86) to (92);
\draw [thin=1.00] (87) to (92);
\draw [thin=1.00] (87) to (93);
\draw [thin=1.00] (87b) to (93);
\draw [thin=1.00] (87b) to (93b);
\draw [thin=1.00] (88) to (94);
\draw [thin=1.00] (89) to (94);
\draw [thin=1.00] (89) to (95);
\draw [thin=1.00] (90) to (95);
\draw [thin=1.00] (90) to (96);
\draw [thin=1.00] (91) to (96);
\draw [thin=1.00] (91) to (97);
\draw [thin=1.00] (92) to (97);
\draw [thin=1.00] (92) to (98);
\draw [thin=1.00] (93) to (98);
\draw [thin=1.00] (93) to (98b);
\draw [thin=1.00] (93b) to (98b);
\draw [thin=1.00] (94) to (99);
\draw [thin=1.00] (94) to (100);
\draw [thin=1.00] (95) to (100);
\draw [thin=1.00] (95) to (101);
\draw [thin=1.00] (96) to (101);
\draw [thin=1.00] (96) to (102);
\draw [thin=1.00] (97) to (102);
\draw [thin=1.00] (97) to (103);
\draw [thin=1.00] (98) to (103);
\draw [thin=1.00] (98) to (104);
\draw [thin=1.00] (98b) to (104);
\draw [thin=1.00] (98b) to (104b);
\draw [thin=1.00] (99) to (105);
\draw [thin=1.00] (100) to (105);
\draw [thin=1.00] (100) to (106);
\draw [thin=1.00] (101) to (106);
\draw [thin=1.00] (101) to (107);
\draw [thin=1.00] (102) to (107);
\draw [thin=1.00] (102) to (108);
\draw [thin=1.00] (103) to (108);
\draw [thin=1.00] (103) to (109);
\draw [thin=1.00] (104) to (109);
\draw [thin=1.00] (104) to (109b);
\draw [thin=1.00] (104b) to (109b);
\draw [thin=1.00] (105) to (110);
\draw [thin=1.00] (105) to (111);
\draw [thin=1.00] (106) to (111);
\draw [thin=1.00] (106) to (112);
\draw [thin=1.00] (107) to (112);
\draw [thin=1.00] (107) to (113);
\draw [thin=1.00] (108) to (113);
\draw [thin=1.00] (108) to (114);
\draw [thin=1.00] (109) to (114);
\draw [thin=1.00] (109) to (115);
\draw [thin=1.00] (109b) to (115);
\draw [thin=1.00] (109b) to (115b);
\draw [thin=1.00] (55a) to (61a);
\draw [thin=1.00] (56a) to (61a);
\draw [thin=1.00] (56a) to (62a);
\draw [thin=1.00] (57a) to (62a);
\draw [thin=1.00] (57a) to (63a);
\draw [thin=1.00] (58a) to (63a);
\draw [thin=1.00] (58a) to (64a);
\draw [thin=1.00] (59a) to (64a);
\draw [thin=1.00] (59a) to (65a);
\draw [thin=1.00] (60a) to (65a);
\draw [thin=1.00] (61a) to (66a);
\draw [thin=1.00] (61a) to (67a);
\draw [thin=1.00] (62a) to (67a);
\draw [thin=1.00] (62a) to (68a);
\draw [thin=1.00] (63a) to (68a);
\draw [thin=1.00] (63a) to (69a);
\draw [thin=1.00] (64a) to (69a);
\draw [thin=1.00] (64a) to (70a);
\draw [thin=1.00] (65a) to (70a);
\draw [thin=1.00] (65a) to (71a);
\draw [thin=1.00] (66a) to (72a);
\draw [thin=1.00] (67a) to (72a);
\draw [thin=1.00] (67a) to (73a);
\draw [thin=1.00] (68a) to (73a);
\draw [thin=1.00] (68a) to (74a);
\draw [thin=1.00] (69a) to (74a);
\draw [thin=1.00] (69a) to (75a);
\draw [thin=1.00] (70a) to (75a);
\draw [thin=1.00] (70a) to (76a);
\draw [thin=1.00] (71a) to (76a);
\draw [thin=1.00] (72a) to (77a);
\draw [thin=1.00] (72a) to (78a);
\draw [thin=1.00] (73a) to (78a);
\draw [thin=1.00] (73a) to (79a);
\draw [thin=1.00] (74a) to (79a);
\draw [thin=1.00] (74a) to (80a);
\draw [thin=1.00] (75a) to (80a);
\draw [thin=1.00] (75a) to (81a);
\draw [thin=1.00] (76a) to (81a);
\draw [thin=1.00] (76a) to (82a);
\draw [thin=1.00] (77a) to (83a);
\draw [thin=1.00] (78a) to (83a);
\draw [thin=1.00] (78a) to (84a);
\draw [thin=1.00] (79a) to (84a);
\draw [thin=1.00] (79a) to (85a);
\draw [thin=1.00] (80a) to (85a);
\draw [thin=1.00] (80a) to (86a);
\draw [thin=1.00] (81a) to (86a);
\draw [thin=1.00] (81a) to (87a);
\draw [thin=1.00] (82a) to (87a);
\draw [thin=1.00] (83a) to (88a);
\draw [thin=1.00] (83a) to (89a);
\draw [thin=1.00] (84a) to (89a);
\draw [thin=1.00] (84a) to (90a);
\draw [thin=1.00] (85a) to (90a);
\draw [thin=1.00] (85a) to (91a);
\draw [thin=1.00] (86a) to (91a);
\draw [thin=1.00] (86a) to (92a);
\draw [thin=1.00] (87a) to (92a);
\draw [thin=1.00] (87a) to (93a);
\draw [thin=1.00] (88a) to (94a);
\draw [thin=1.00] (89a) to (94a);
\draw [thin=1.00] (89a) to (95a);
\draw [thin=1.00] (90a) to (95a);
\draw [thin=1.00] (90a) to (96a);
\draw [thin=1.00] (91a) to (96a);
\draw [thin=1.00] (91a) to (97a);
\draw [thin=1.00] (92a) to (97a);
\draw [thin=1.00] (92a) to (98a);
\draw [thin=1.00] (93a) to (98a);
\draw [thin=1.00] (94a) to (99a);
\draw [thin=1.00] (94a) to (100a);
\draw [thin=1.00] (95a) to (100a);
\draw [thin=1.00] (95a) to (101a);
\draw [thin=1.00] (96a) to (101a);
\draw [thin=1.00] (96a) to (102a);
\draw [thin=1.00] (97a) to (102a);
\draw [thin=1.00] (97a) to (103a);
\draw [thin=1.00] (98a) to (103a);
\draw [thin=1.00] (98a) to (104a);
\draw [thin=1.00] (99a) to (105a);
\draw [thin=1.00] (100a) to (105a);
\draw [thin=1.00] (100a) to (106a);
\draw [thin=1.00] (101a) to (106a);
\draw [thin=1.00] (101a) to (107a);
\draw [thin=1.00] (102a) to (107a);
\draw [thin=1.00] (102a) to (108a);
\draw [thin=1.00] (103a) to (108a);
\draw [thin=1.00] (103a) to (109a);
\draw [thin=1.00] (104a) to (109a);
\draw [thin=1.00] (105a) to (110a);
\draw [thin=1.00] (105a) to (111a);
\draw [thin=1.00] (106a) to (111a);
\draw [thin=1.00] (106a) to (112a);
\draw [thin=1.00] (107a) to (112a);
\draw [thin=1.00] (107a) to (113a);
\draw [thin=1.00] (108a) to (113a);
\draw [thin=1.00] (108a) to (114a);
\draw [thin=1.00] (109a) to (114a);
\draw [thin=1.00] (109a) to (115a);
\draw [thin=1.00] (110a) to (116a);
\draw [thin=1.00] (111a) to (116a);
\draw [thin=1.00] (111a) to (117a);
\draw [thin=1.00] (112a) to (117a);
\draw [thin=1.00] (112a) to (118a);
\draw [thin=1.00] (113a) to (118a);
\draw [thin=1.00] (113a) to (119a);
\draw [thin=1.00] (114a) to (119a);
\draw [thin=1.00] (114a) to (120a);
\draw [thin=1.00] (115a) to (120a);
\draw [thin=1.00] (116a) to (121a);
\draw [thin=1.00] (116a) to (122a);
\draw [thin=1.00] (117a) to (122a);
\draw [thin=1.00] (117a) to (123a);
\draw [thin=1.00] (118a) to (123a);
\draw [thin=1.00] (118a) to (124a);
\draw [thin=1.00] (119a) to (124a);
\draw [thin=1.00] (119a) to (125a);
\draw [thin=1.00] (120a) to (125a);
\draw [thin=1.00] (120a) to (126a);
\draw [thin=1.00] (121a) to (127a);
\draw [thin=1.00] (122a) to (127a);
\draw [thin=1.00] (122a) to (128a);
\draw [thin=1.00] (123a) to (128a);
\draw [thin=1.00] (123a) to (129a);
\draw [thin=1.00] (124a) to (129a);
\draw [thin=1.00] (124a) to (130a);
\draw [thin=1.00] (125a) to (130a);
\draw [thin=1.00] (125a) to (131a);
\draw [thin=1.00] (126a) to (131a);
\draw [thin=1.00] (127a) to (132a);
\draw [thin=1.00] (127a) to (133a);
\draw [thin=1.00] (128a) to (133a);
\draw [thin=1.00] (128a) to (134a);
\draw [thin=1.00] (129a) to (134a);
\draw [thin=1.00] (129a) to (135a);
\draw [thin=1.00] (130a) to (135a);
\draw [thin=1.00] (130a) to (136a);
\draw [thin=1.00] (131a) to (136a);
\draw [thin=1.00] (131a) to (137a);
\draw [thin=1.00] (0c) to (7c);
\draw [thin=1.00] (1c) to (7c);
\draw [thin=1.00] (1c) to (8c);
\draw [thin=1.00] (2c) to (8c);
\draw [thin=1.00] (2c) to (9c);
\draw [thin=1.00] (3c) to (9c);
\draw [thin=1.00] (3c) to (10c);
\draw [thin=1.00] (4c) to (10c);
\draw [thin=1.00] (4c) to (11c);
\draw [thin=1.00] (5c) to (11c);
\draw [thin=1.00] (5c) to (12c);
\draw [thin=1.00] (6c) to (12c);
\draw [thin=1.00] (7c) to (13c);
\draw [thin=1.00] (7c) to (14c);
\draw [thin=1.00] (8c) to (14c);
\draw [thin=1.00] (8c) to (15c);
\draw [thin=1.00] (9c) to (15c);
\draw [thin=1.00] (9c) to (16c);
\draw [thin=1.00] (10c) to (16c);
\draw [thin=1.00] (10c) to (17c);
\draw [thin=1.00] (11c) to (17c);
\draw [thin=1.00] (11c) to (18c);
\draw [thin=1.00] (12c) to (18c);
\draw [thin=1.00] (12c) to (19c);
\draw [thin=1.00] (13c) to (20c);
\draw [thin=1.00] (14c) to (20c);
\draw [thin=1.00] (14c) to (21c);
\draw [thin=1.00] (15c) to (21c);
\draw [thin=1.00] (15c) to (22c);
\draw [thin=1.00] (16c) to (22c);
\draw [thin=1.00] (16c) to (23c);
\draw [thin=1.00] (17c) to (23c);
\draw [thin=1.00] (17c) to (24c);
\draw [thin=1.00] (18c) to (24c);
\draw [thin=1.00] (18c) to (25c);
\draw [thin=1.00] (19c) to (25c);
\draw [thin=1.00] (20c) to (26c);
\draw [thin=1.00] (20c) to (27c);
\draw [thin=1.00] (21c) to (27c);
\draw [thin=1.00] (21c) to (28c);
\draw [thin=1.00] (22c) to (28c);
\draw [thin=1.00] (22c) to (29c);
\draw [thin=1.00] (23c) to (29c);
\draw [thin=1.00] (23c) to (30c);
\draw [thin=1.00] (24c) to (30c);
\draw [thin=1.00] (24c) to (31c);
\draw [thin=1.00] (25c) to (31c);
\draw [thin=1.00] (25c) to (32c);
\draw [thin=1.00] (26c) to (33c);
\draw [thin=1.00] (27c) to (33c);
\draw [thin=1.00] (27c) to (34c);
\draw [thin=1.00] (28c) to (34c);
\draw [thin=1.00] (28c) to (35c);
\draw [thin=1.00] (29c) to (35c);
\draw [thin=1.00] (29c) to (36c);
\draw [thin=1.00] (30c) to (36c);
\draw [thin=1.00] (30c) to (37c);
\draw [thin=1.00] (31c) to (37c);
\draw [thin=1.00] (31c) to (38c);
\draw [thin=1.00] (32c) to (38c);
\draw [thin=1.00] (33c) to (39c);
\draw [thin=1.00] (33c) to (40c);
\draw [thin=1.00] (34c) to (40c);
\draw [thin=1.00] (34c) to (41c);
\draw [thin=1.00] (35c) to (41c);
\draw [thin=1.00] (35c) to (42c);
\draw [thin=1.00] (36c) to (42c);
\draw [thin=1.00] (36c) to (43c);
\draw [thin=1.00] (37c) to (43c);
\draw [thin=1.00] (37c) to (44c);
\draw [thin=1.00] (38c) to (44c);
\draw [thin=1.00] (38c) to (45c);
\draw [thin=1.00] (39c) to (46c);
\draw [thin=1.00] (40c) to (46c);
\draw [thin=1.00] (40c) to (47c);
\draw [thin=1.00] (41c) to (47c);
\draw [thin=1.00] (41c) to (48c);
\draw [thin=1.00] (42c) to (48c);
\draw [thin=1.00] (42c) to (49c);
\draw [thin=1.00] (43c) to (49c);
\draw [thin=1.00] (43c) to (50c);
\draw [thin=1.00] (44c) to (50c);
\draw [thin=1.00] (44c) to (51c);
\draw [thin=1.00] (45c) to (51c);
\draw [thin=1.00] (46c) to (52c);
\draw [thin=1.00] (46c) to (53c);
\draw [thin=1.00] (47c) to (53c);
\draw [thin=1.00] (47c) to (54c);
\draw [thin=1.00] (48c) to (54c);
\draw [thin=1.00] (48c) to (55c);
\draw [thin=1.00] (49c) to (55c);
\draw [thin=1.00] (49c) to (56c);
\draw [thin=1.00] (50c) to (56c);
\draw [thin=1.00] (50c) to (57c);
\draw [thin=1.00] (51c) to (57c);
\draw [thin=1.00] (51c) to (58c);
\draw [thin=1.00] (52c) to (59c);
\draw [thin=1.00] (53c) to (59c);
\draw [thin=1.00] (53c) to (60c);
\draw [thin=1.00] (54c) to (60c);
\draw [thin=1.00] (54c) to (61c);
\draw [thin=1.00] (55c) to (61c);
\draw [thin=1.00] (55c) to (62c);
\draw [thin=1.00] (56c) to (62c);
\draw [thin=1.00] (56c) to (63c);
\draw [thin=1.00] (57c) to (63c);
\draw [thin=1.00] (57c) to (64c);
\draw [thin=1.00] (58c) to (64c);
\draw [thin=1.00] (59c) to (65c);
\draw [thin=1.00] (59c) to (66c);
\draw [thin=1.00] (60c) to (66c);
\draw [thin=1.00] (60c) to (67c);
\draw [thin=1.00] (61c) to (67c);
\draw [thin=1.00] (61c) to (68c);
\draw [thin=1.00] (62c) to (68c);
\draw [thin=1.00] (62c) to (69c);
\draw [thin=1.00] (63c) to (69c);
\draw [thin=1.00] (63c) to (70c);
\draw [thin=1.00] (64c) to (70c);
\draw [thin=1.00] (64c) to (71c);
\draw [thin=1.00] (65c) to (72c);
\draw [thin=1.00] (66c) to (72c);
\draw [thin=1.00] (66c) to (73c);
\draw [thin=1.00] (67c) to (73c);
\draw [thin=1.00] (67c) to (74c);
\draw [thin=1.00] (68c) to (74c);
\draw [thin=1.00] (68c) to (75c);
\draw [thin=1.00] (69c) to (75c);
\draw [thin=1.00] (69c) to (76c);
\draw [thin=1.00] (70c) to (76c);
\draw [thin=1.00] (70c) to (77c);
\draw [thin=1.00] (71c) to (77c);
\draw [thin=1.00] (72c) to (78c);
\draw [thin=1.00] (72c) to (79c);
\draw [thin=1.00] (73c) to (79c);
\draw [thin=1.00] (73c) to (80c);
\draw [thin=1.00] (74c) to (80c);
\draw [thin=1.00] (74c) to (81c);
\draw [thin=1.00] (75c) to (81c);
\draw [thin=1.00] (75c) to (82c);
\draw [thin=1.00] (76c) to (82c);
\draw [thin=1.00] (76c) to (83c);
\draw [thin=1.00] (77c) to (83c);
\draw [thin=1.00] (77c) to (84c);
\draw [thin=1.00] (78c) to (85c);
\draw [thin=1.00] (79c) to (85c);
\draw [thin=1.00] (79c) to (86c);
\draw [thin=1.00] (80c) to (86c);
\draw [thin=1.00] (80c) to (87c);
\draw [thin=1.00] (81c) to (87c);
\draw [thin=1.00] (81c) to (88c);
\draw [thin=1.00] (82c) to (88c);
\draw [thin=1.00] (82c) to (89c);
\draw [thin=1.00] (83c) to (89c);
\draw [thin=1.00] (83c) to (90c);
\draw [thin=1.00] (84c) to (90c);
\draw [thin=1.00] (85c) to (91c);
\draw [thin=1.00] (85c) to (92c);
\draw [thin=1.00] (86c) to (92c);
\draw [thin=1.00] (86c) to (93c);
\draw [thin=1.00] (87c) to (93c);
\draw [thin=1.00] (87c) to (94c);
\draw [thin=1.00] (88c) to (94c);
\draw [thin=1.00] (88c) to (95c);
\draw [thin=1.00] (89c) to (95c);
\draw [thin=1.00] (89c) to (96c);
\draw [thin=1.00] (90c) to (96c);
\draw [thin=1.00] (90c) to (97c);
\draw [thin=1.00] (0g) to (0d);
\draw [thin=1.00] (0e) to (0f);
\end{pgfonlayer}
\end{tikzpicture}
}}
\end{proof}
\begin{cor} For $m \geq 48$ and $n \geq 40$, $\lambda_1^1(C_m \times C_n)=5$.
\end{cor}
The last corollary gave the values of $m,n$ beyond which $\lambda_1^1(C_m \times C_n)=5$. However, there are smaller product graphs whose $L(1,1)$-number is $5$ as demonstrated in the next corollary.
\begin{cor} \label{cor1} For all $m,n \geq 14$, $m \notin \left\{14,16,18,22,26,32,36,46\right\}$, $n \notin \left\{14,16,18,26,28,34\right\}$, $\lambda_1^1(C_m \times C_n)=5$.
\end{cor}
For some of pairs $\left\{m',n'\right\}$, in the two sets defined above, namely $\left\{16,18\right\}$,\\$\left\{18,28\right\}$,$\left\{32,18\right\}$,$\left\{36,16\right\}$,$\left\{36,18\right\}$,$\left\{36,2 8\right\},$ $\lambda_1^1(C_m \times C_n)=5$. This is obvious from earlier results. For the remaining pairs, it can easily be confirmed, by manual labeling, that $\lambda_1^1(C_m \times C_n) \leq 6$. However, we observe that there could be a better upper bound and therefore, present the following conjecture:
\begin{conj}\label{conj1} For $m,n \geq 12$, $m,n \not\equiv 0\mod 5$, $\lambda_1^1(C_m \times C_n)=5$.
\end{conj}
Solving this conjecture only requires confirming the $5-L(1,1)$-labeling for $C_m \times C_n$, where $m,n$ are the remaining pair yet to be confirmed in the sets in Corollary \ref{cor1}.
The results obtained is summarized in the table below:
\begin{center}
$\begin{array}{|c|c|c|} \hline
m & n & \lambda_1^1(C_m \times C_n)\\ \hline
4 & 4,5,8,10 & 7\\ \hline
4 & n\not\equiv 0 \mod3 \ \& \ n \geq 11 & 6\\ \hline
m \equiv 0 \mod 4 & n\equiv 0\mod 3 & 5\\ \hline
m\equiv 0 \mod 5 & n \equiv 0 \mod 5 & 4 \\ \hline
6 & 3, 6 & 8\\ \hline
6 & 5, 10 & 7 \\ \hline
6 & 7, 9, 11, 14+4n', n' \geq 0 & 6\\ \hline
8 & 8 & 7 \\ \hline
8 & n \geq 10 , n \not\equiv 0 \mod 3 & 6 \\ \hline
m \equiv 0\mod 10 & n\geq 11, n \not\equiv 0 \mod 5 & 5\\ \hline
12 & n\geq 12 & 5 \\ \hline
m \geq 40, even & n \geq 48 & 5 \\ \hline
\end{array}$
\end{center}
|
\section{Motivation}
In Carnot groups, the following theorem has been established.
\begin{thm}\cite{B:HG, W:W, B:MP} \label{subelliptic}
Let $\Omega$ be a bounded domain in a Carnot group and let
$v:\partial \Omega \to \mathbb{R}$ be a continuous
function.
Then the Dirichlet problem
\begin{eqnarray*}
\left\{ \begin{array}{cc}
\Delta_{\infty}u = 0 & \textmd{in\ \ \ } \Omega \\
u = v & \textmd{on\ \ \ } \partial \Omega
\end{array} \right.
\end{eqnarray*}
has a unique viscosity solution $u_\infty$.
\end{thm}
Our goal is to prove a parabolic version of Theorem \ref{subelliptic} for a class of equations (defined in the next section), namely
\begin{conj}
Let $\Omega$ be a bounded domain in a Carnot group and let $T>0$. Let $\psi\in C(\overline{\Omega})$ and let $g \in C(\Omega \times [0,T))$ Then the Cauchy-Dirichlet problem
\begin{equation}\label{maincon}
\left\{ \begin{array}{cl}
u_t - \Delta^h_{\infty}u =0& \hspace{10pt} in \hspace{5pt} \Omega \times (0,T),\\
u(x,0) =\psi(x) & \hspace{10pt} on \hspace{5pt} \overline{\Omega} \\
u(x,t)=g(x,t) & \hspace{10pt} on \hspace{5pt} \partial \Omega \times (0,T)
\end{array}\right.
\end{equation}
has a unique viscosity solution $u$.
\end{conj}
In Sections 2 and 3, we review key properties of Carnot groups and parabolic viscosity solutions. In Section 4, we prove uniqueness and Section 5 covers existence.
\section{Calculus on Carnot Groups}
We begin by denoting an arbitrary Carnot group in $\mathbb{R}^N$ by $G$ and its
corresponding Lie Algebra by $g$. Recall that $g$ is nilpotent and
stratified, resulting in the decomposition $$g= V_1 \oplus V_2 \oplus
\cdots \oplus V_l$$ for appropriate vector spaces that satisfy the Lie
bracket relation $[V_1,V_j]=V_{1+j}.$ The Lie Algebra $g$ is associated
with the group $G$ via the exponential map $\exp: g \to G.$ Since
this
map is a diffeomorphism, we can choose a basis for $g$ so that it is
the identity map. Denote this basis by $$X_1,X_2,\ldots ,
X_{n_1},Y_1,Y_2,\ldots ,Y_{n_2},Z_1,Z_2, \ldots , Z_{n_3}$$ so that
\begin{eqnarray*}
V_1 & = & \textmd{span}\{X_1,X_2,\ldots ,X_{n_1}\} \\
V_2 & = & \textmd{span}\{Y_1,Y_2,\ldots ,Y_{n_2}\} \\
V_3 \oplus V_4 \oplus \cdots \oplus V_l & = &
\textmd{span}\{Z_1,Z_2,\ldots ,Z_{n_3}\}.
\end{eqnarray*}
We endow $g$ with an inner product $\ip{\cdot}{\cdot}$
and related norm $\|\cdot\|$ so that this basis is orthonormal.
Clearly, the Riemannian dimension of $g$ (and so $G$) is $N
= n_1+n_2+n_3$.
However, we will also consider the homogeneous dimension of $G$, denoted $\mathcal{Q}$,
which is given by $$\mathcal{Q}=\sum_{i=1}^l i \cdot \dim V_i.$$
Before proceeding with the calculus, we recall the group and metric space properties. Since
the exponential map is the identity, the group law is the
Campbell-Hausdorff formula (see, for example, \cite{BO:LAG}). For our
purposes, this formula is given by
\begin{equation}\label{CCH}
p \cdot q =
p+q+\frac{1}{2}[p,q]+R(p,q)
\end{equation}
where $R(p,q)$ are terms of order $3$ or
higher. The identity element of $G$ will be denoted by $0$ and called the
origin. There is also a natural metric on $G$, which is
the Carnot-Carath\'{e}odory distance, defined for the points $p$ and $q$
as follows:
\begin{equation*}
d_C(p,q)= \inf_{\Gamma} \int_{0}^{1} \| \gamma '(t) \| dt
\end{equation*}
where the set $ \Gamma $
is the set of all curves $ \gamma $ such that $ \gamma (0) = p,
\gamma (1) = q $ and $\gamma '(t) \in V_1$.
By Chow's theorem (see, for example,
\cite{BR:SRG}) any two points
can be connected by such a curve, which means $ d_C(p,q) $ is an honest
metric. Define a Carnot-Carath\'{e}odory ball of radius $r$ centered at a
point $p_0$ by $$B(p_0,r)=\{p\in G : d_C(p,p_0) < r\}.$$
In addition to the Carnot-Carath\'{e}odory metric, there is a smooth (off the origin) gauge. This gauge is defined for a point $p=(\zeta_1, \zeta_2, \ldots, \zeta_l)$
with $\zeta_i \in V_i$ by
\begin{equation}\label{gaugedef}
\mathcal{N}(p)=\bigg(\sum_{i=1}^l \|\zeta_{i}\|^{\frac{\displaystyle 2l!}{\displaystyle i}}
\bigg )^\frac{\displaystyle 1}{\displaystyle 2l!}
\end{equation}
and it induces a metric $d_{\mathcal{N}}$ that is bi-Lipschitz equivalent to the
Carnot-Carath\'{e}odory metric and is given by
$$d_{\mathcal{N}}(p,q) = \mathcal{N}(p^{-1}\cdot q).$$ We define a gauge ball
of radius $r$ centered at a
point $p_0$ by $$B_{\mathcal{N}}(p_0,r)=\{p\in G : d_{\mathcal{N}}(p,p_0) < r\}.$$
In this environment, a smooth function $u: G \to \mathbb{R}$ has the horizontal derivative given by
$$\nabla_0 u=(X_1u,X_2u, \ldots, X_{n_1}u )$$ and the symmetrized horizontal second derivative matrix,
denoted by $(D^2u)^\star$, with entries
\begin{eqnarray*}
((D^2u)^\star)_{ij} = \frac{1}{2} (X_iX_ju+X_jX_iu)
\end{eqnarray*}
for $i,j=1,2,\ldots , n_1.$ We also consider the semi-horizontal derivative given by
$$\nabla_1 u=(X_1u,X_2u, \ldots, X_{n_1}u, Y_1u,Y_2u, \ldots, Y_{n_2}u).$$
Using the above derivatives, we define the $h$-homogeneous infinite Laplace operator for $h\geq 1$ by
\begin{equation*}
\Delta^h_{\infty}f = \|\nabla_0 f\|^{h-3} \sum_{i,j=1}^{n_1} X_ifX_jfX_iX_jf = \|\nabla_0 f\|^{h-3}\ip{(D^2f)^{\star}\nabla_0f}{\nabla_0f}.
\end{equation*}
Given $T>0$ and a function $u:G \times [0,T] \to \mathbb{R}$, we may define the analogous subparabolic infinite Laplace operator by
$$u_t-\Delta^h_{\infty}u$$ and we consider the corresponding equation
\begin{equation}\label{eqmain}
u_t-\Delta^h_{\infty}u=0.
\end{equation}
We note that when $h\geq 3$, this operator is continuous. When $h=3$, we have the subparabolic infinite Laplace equation analogous to the infinite Laplace operator in \cite{B:MP}. The Euclidean analog for $h=1$ has been explored in \cite{JK:JK} and the Euclidean analog for $1<h<3$ in \cite{PV}.
We recall that for any open set $\mathcal{O} \subset G$, the
function $f$ is in the horizontal Sobolev space $W^{1,\texttt{p}}(\mathcal{O})$ if
$f$ and $ X_if$ are in $ L^\texttt{p}(\mathcal{O}) $ for $i=1,2,\ldots , n_1$. Replacing $ L^\texttt{p}(\mathcal{O})$ by $L_{loc}^\texttt{p}(\mathcal{O})$, the space $ W_{loc}^{1,\texttt{p}}(\mathcal{O}) $ is defined similarly. The space $W_{0}^{1,\texttt{p}}(\mathcal{O})$ is the closure in $W^{1,\texttt{p}}(\mathcal{O})$ of smooth functions with
compact support. In addition, we recall a function $u:G\to
\mathbb{R}$ is $\operatorname{\mathcal{C}^2_{\textmd{sub}}}$ if $\nabla_1u$ and $X_iX_ju$ are continuous for
all $i,j=1,2,\ldots n_1$. Note that $\operatorname{\mathcal{C}^2_{\textmd{sub}}}$ is not equivalent to (Euclidean) $C^2$.
For spaces involving time, the space $C(t_1,t_2;X)$ consists of all continuous functions $u:[t_1,t_2]\to X$ with $\max_{t_1\leq t\leq t_2} \|u(\cdot,t)\|_X <\infty$. A similar definition holds for $L^\texttt{p}(t_1,t_2;X)$.
Given an open box $\mathcal{O}=(a_1,b_1)\times (a_2,b_2)\times \cdots \times (a_N,b_N)$, we define the parabolic space $\mathcal{O}_{t_1,t_2}$ to be $\mathcal{O} \times [t_1,t_2]$. Its parabolic boundary is given by
$\partial_{\operatorname{par}}\mathcal{O}_{t_1,t_2} = (\overline{\mathcal{O}}\times \{t_1\}) \cup (\partial \mathcal{O}\times (t_1,t_2])$.
Finally, recall that if $G$ is a Carnot group with homogeneous dimension $\mathcal{Q}$, then $G \times \mathbb{R}$ is again a Carnot group of homogeneous dimension $\mathcal{Q} + 1$ where we have added an extra vector field $\frac{\partial}{\partial t}$ to the first layer of the grading. This allows us to give meaning to notations such as $W^{1,2}(\mathcal{O}_{t_1,t_2})$ and $\operatorname{\mathcal{C}^2_{\textmd{sub}}}(\mathcal{O}_{t_1,t_2})$ where we consider $\nabla_0u$ to be $\left(X_1u, X_2u, \ldots, X_{n_1}u, \frac{\partial u}{\partial t}\right)$.
\section{Parabolic Jets and Viscosity Solutions}
\subsection{Parabolic Jets}
In this subsection, we recall the definitions of the parabolic jets, as given in \cite{BM}, but included here for completeness. We define the parabolic superjet of $u(p,t)$ at the point $(p_0,t_0) \in \mathcal{O}_{t_1,t_2}$, denoted $P^{2,+}u(p_0,t_0)$, by using triples $(a,\eta,X) \in \mathbb{R} \times V_1\oplus V_2 \times S^{n_1}$ so that $(a,\eta,X) \in P^{2,+}u(p_0,t_0)$ if
\begin{eqnarray*}
u(p,t) & \leq & u(p_0,t_0) + a(t-t_0) + \ip{\eta}{\widehat{p_0^{-1}\cdot p}}+\frac{1}{2}\ip{X\overline{p_0^{-1}\cdot p}}{\overline{p_0^{-1}\cdot p}}\\
& & + o(|t-t_0|+|p_0^{-1}\cdot p|^2)\ \ \textmd{as}\ \ (p,t)\rightarrow (p_0,t_0).
\end{eqnarray*}
We recall that $S^k$ is the set of $k\times k$ symmetric matrices and $n_i=\dim V_i$. We define $\overline{p_0^{-1}\cdot p}$ as the first $n_1$ coordinates of
$p_0^{-1}\cdot p$ and $\widehat{p_0^{-1}\cdot p}$ as the first $n_1+n_2$ coordinates of
$p_0^{-1}\cdot p$.
This definition is an extension of the superjet definition for subparabolic equations in the Heisenberg group \cite{B:HP}. We define the subjet $P^{2,-}u(p_0,t_0)$ by
$$P^{2,-}u(p_0,t_0)=-P^{2,+}(-u)(p_0,t_0).$$
We define the set theoretic closure of the superjet, denoted $\overline{P}^{2,+}u(p_0,t_0)$, by requiring $(a,\eta,X) \in \overline{P}^{2,+}u(p_0,t_0)$ exactly when there is a sequence \\
$(a_n,p_n,t_n,u(p_n,t_n),\eta_n,X_n)\to (a,p_0,t_0,u(p_0,t_0),\eta,X)$ with the triple \\ $(a_n,\eta_n,X_n)\in P^{2,+}u(p_n,t_n)$. A similar definition holds for the closure of the subjet.
We may also define jets using appropriate test functions. Given a function $u:\mathcal{O}_{t_1,t_2}\to \mathbb{R}$
we consider the set $\mathcal{A}u(p_0,t_0)$ given by
\begin{equation*}
\mathcal{A}u(p_0,t_0) = \{\phi \in \operatorname{\mathcal{C}^2_{\textmd{sub}}}(\mathcal{O}_{t_1,t_2}):u(p,t)-\phi(p,t)\leq u(p_0,t_0)-\phi(p_0,t_0)=0\ \forall (p,t)\in \mathcal{O}_{t_1,t_2}\}.
\end{equation*}
consisting of all test functions that touch $u$ from above at $(p_0,t_0)$. We define
the set of all test functions that touch from below, denoted $\mathcal{B}u(p_0,t_0)$,
similarly.
The following lemma relates the test functions to jets. The proof is identical to Lemma 3.1 in \cite{B:HP}, but uses the (smooth) gauge $\mathcal{N}(p)$ instead of Euclidean distance.
\begin{lemma}
$$P^{2,+}u(p_0,t_0)=\{(\phi_t(p_0,t_0),\nabla \phi(p_0,t_0), (D^2\phi(p_0,t_0))^\star): \phi \in \mathcal{A}u(p_0,t_0)\}.$$
\end{lemma}
\subsection{Jet Twisting}
We recall that the set $V_1=\span \{X_1,X_2,\ldots, X_{n_1}\}$ and notationally, we will always denote $n_1$ by $n$. The vectors $X_i$ at the point $p\in G$ can be written as
$$X_i(p)=\sum_{j=1}^N a_{ij}(p)\frac{\partial}{\partial x_j}$$
forming the $n\times N$ matrix $\mathbb{A}$ with smooth entries $\mathbb{A}_{ij}=a_{ij}(p)$. By linear independence of the $X_i$, $\mathbb{A}$ has rank $n$. Similarly, $$Y_i(p)=
\sum_{j=1}^N b_{ij}(p)\frac{\partial}{\partial x_j}$$ forming the $n_2 \times N$ matrix $\mathbb{B}$ with smooth entries $\mathbb{B}_{ij}=b_{ij}$. The matrix $\mathbb{B}$ has rank $n_2$.
The following lemma differs from \cite[Corollary 3.2]{B:MP} only in that there is now a parabolic term. This term however, does not need to be twisted. The proof is then identical, as only the space terms need twisting.
\begin{lemma}\label{twist}
Let $(a,\eta, X) \in \overline{P}^{2,+}_{\operatorname{eucl}}u(p,t)$. (Recall that $(\eta,X) \in \mathbb{R}^N \times S^N$.) Then $$(a, \mathbb{A} \cdot \eta \oplus \mathbb{B} \cdot \eta, \
\mathbb{A}X\mathbb{A}^T+\mathbb{M}) \in \overline{P}^{2,+}u(p,t).$$ Here the entries of the (symmetric) matrix $\mathbb{M}$ are given by
\begin{eqnarray*}
\mathbb{M}_{ij} & = &\begin{cases}
\displaystyle \sum_{k=1}^N\displaystyle\sum_{l=1}^N \bigg(a_{il}(p)\frac{\displaystyle\partial}{\displaystyle\partial x_l}a_{jk}(p)+a_{jl}(p)\frac{\displaystyle\partial a_{ik}}{\displaystyle\partial x_l}(p)\bigg)\eta_k & i \neq j, \\
& \\
\displaystyle\sum_{k=1}^N\displaystyle\sum_{l=1}^N a_{il}(p)\frac{\displaystyle\partial a_{ik}}{\displaystyle\partial x_l}(p)\eta_k & i = j.
\end{cases}
\end{eqnarray*}
\end{lemma}
\subsection{Viscosity Solutions}
We consider parabolic equations of the form
\begin{equation}\label{main}
u_t+F(t,p,u,\nabla_1 u,(D^2u)^{\star})=0
\end{equation}
for continuous and proper $F:[0,T]\times G \times \mathbb{R} \times g \times S^{n} \to \mathbb{R}$. \cite{CIL:UGTVS}
We recall that $S^{n}$ is the set of $n \times n$ symmetric matrices (where $\dim V_1=n$) and the derivatives $\nabla_1 u$ and $(D^2u)^{\star}$ are taken in the space variable $p$.
We then use the jets to define subsolutions and supersolutions to Equation \eqref{main} in the usual way.
\begin{definition}\label{1}
Let $(p_0,t_0)\in \mathcal{O}_{t_1,t_2}$ be as above. The upper semicontinuous function $u$ is a \emph{parabolic viscosity subsolution} in $\mathcal{O}_{t_1,t_2}$ if for all $(p_0,t_0) \in \mathcal{O}_{t_1,t_2}$ we have $(a,\eta,X) \in \overline{P}^{2,+}u(p_0,t_0)$ produces
$$a+F(t_0,p_0,u(p_0,t_0),\eta,X)\leq 0.$$
A lower semicontinuous function $u$ is a \emph{parabolic viscosity supersolution} in $\mathcal{O}_{t_1,t_2}$ if for all $(p_0,t_0) \in \mathcal{O}_{t_1,t_2}$ we have $(b,\nu,Y) \in \overline{P}^{2,-}u(p_0,t_0)$ produces
$$b+F(t_0,p_0,u(p_0,t_0),\nu,Y)\geq 0.$$
A continuous function $u$ is a \emph{parabolic viscosity solution} in $\mathcal{O}_{t_1,t_2}$ if it is both a parabolic viscosity subsolution and parabolic viscosity supersolution.
\end{definition}
\begin{remark}
In the special case when $F(t,p,u,\nabla_1 u,(D^2u)^{\star})=F^h_\infty(\nabla_0 u,(D^2u)^{\star})=-\Delta^h_{\infty}u$, for $h\geq 3$, we use the terms ``parabolic viscosity h-infinite supersolution", etc.
\end{remark}
In the case when $1\leq h<3$, the definition above is insufficient due to the singularity occurring when the horizontal gradient vanishes. Therefore, following \cite{JK:JK} and \cite{PV}, we define viscosity solutions to
Equation \eqref{eqmain} when $1\leq h<3$ as follows:
\begin{definition}\label{2}
Let $\mathcal{O}_{t_1,t_2}$ be as above. A lower semicontinuous function
$v:\mathcal{O}_{t_1,t_2}\to\mathbb{R}$ is a \emph{parabolic viscosity h-infinite supersolution} of $u_t-\Delta^h_\infty u=0$ if whenever $(p_0,t_0)\in \mathcal{O}_{t_1,t_2}$ and $\phi\in \mathcal{B}u(p_0,t_0)$, we have
\begin{eqnarray*}\left\{\begin{array}{cc}
\phi_t(p_0,t_0)-\Delta^h_\infty \phi(p_0,t_0) \geq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) \neq 0 \\ \mbox{} & \mbox{} \\
\phi_t(p_0,t_0)-\displaystyle \min_{\|\eta\|=1}\ip{(D^2\phi)^\star(p_0,t_0)\;\eta}{\eta} \geq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) = 0 \ \textmd{and\ \ } h=1 \\
\phi_t(p_0,t_0) \geq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) = 0 \ \textmd{and\ \ } 1<h<3
\end{array}\right.
\end{eqnarray*}
An upper semicontinuous function
$u:\mathcal{O}_{t_1,t_2}\to\mathbb{R}$ is a \emph{parabolic viscosity h-infinite subsolution} of $u_t-\Delta^h_\infty u=0$ if whenever $(p_0,t_0)\in \mathcal{O}_{t_1,t_2}$ and $\phi\in \mathcal{A}u(p_0,t_0)$, we have
\begin{eqnarray*}\left\{\begin{array}{cc}
\phi_t(p_0,t_0)-\Delta^h_\infty \phi(p_0,t_0) \leq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) \neq 0 \\ \mbox{} & \mbox{} \\
\phi_t(p_0,t_0)-\displaystyle \max_{\|\eta\|=1}\ip{(D^2\phi)^\star(p_0,t_0)\;\eta}{\eta} \leq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) = 0 \ \textmd{and\ \ } h=1 \\
\phi_t(p_0,t_0) \leq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) = 0 \ \textmd{and\ \ } 1<h<3
\end{array}\right.
\end{eqnarray*}
A continuous function is a \emph{parabolic viscosity h-infinite solution} if it is both a parabolic viscosity h-infinite subsolution and parabolic viscosity h-infinite subsolution.
\end{definition}
\begin{remark}
When $1<h<3$, we can actually consider the continuous operator
\begin{eqnarray}\label{relax}
F^h_\infty(\nabla_0 u,(D^2u)^{\star}) =
\left\{\begin{array}{cl}
-\|\nabla_0u\|^{h-3}\ip{(D^2u)^{\star}\nabla_0u}{\nabla_0u}=-\Delta^h_\infty u & \nabla_0u \neq 0 \\
0 & \nabla_0u = 0. \end{array}\right.
\end{eqnarray}
Definitions \ref{1} and {2} would then agree. (cf. \cite{PV})
\end{remark}
We also wish to define what \cite{J:PD} refers to as parabolic viscosity solutions. We first need to consider the set $$\mathcal{A}^-u(p_0,t_0)=\{\phi \in \mathcal{C}^2(\mathcal{O}_{t_1,t_2}): u(p,t)-\phi(p,t) \leq u(p_0,t_0)-\phi(p_0,t_0)=0\ \textmd{for}\ p \neq p_0, t < t_0\}$$
consisting of all functions that touch from above only when $t<t_0$.
Note that this set is larger than $\mathcal{A}u$ and corresponds physically to the past alone playing a role in determining the present. We define $\mathcal{B}^-u(p_0,t_0)$ similarly.
We then have the following definition.
\begin{definition}
An upper semicontinuous function $u$ on $\mathcal{O}_{t_1,t_2}$ is a \emph{past parabolic viscosity subsolution} in $\mathcal{O}_{t_1,t_2}$ if $\phi\in \mathcal{A}^-u(p_0,t_0)$ produces
$$\phi_t(p_0,t_0)+F(t_0,p_0,u(p_0,t_0),\nabla_1 \phi(p_0,t_0),(D^2\phi(p_0,t_0))^{\star}) \leq 0.$$
An lower semicontinuous function $u$ on $\mathcal{O}_{t_1,t_2}$ is a \emph{past parabolic viscosity supersolution} in $\mathcal{O}_{t_1,t_2}$ if $\phi\in \mathcal{B}^-u(p_0,t_0)$ produces
$$\phi_t(p_0,t_0)+F(t_0,p_0,u(p_0,t_0),\nabla_1 \phi(p_0,t_0),(D^2\phi(p_0,t_0))^{\star}) \geq 0.$$
A continuous function is a \emph{past parabolic viscosity solution} if it is both a past parabolic viscosity supersolution and subsolution.
\end{definition}
We have the following proposition whose proof is obvious.
\begin{prop}\label{onedir}
Past parabolic viscosity sub(super-)solutions are parabolic viscosity sub(super-)solutions. In particular, past parabolic viscosity h-infinite sub(super-)solutions are parabolic viscosity h-infinite subsub(super-)solutions for $h \geq 1$.
\end{prop}
\subsection{The Carnot Parabolic Maximum Principle}
In this subsection, we recall the Carnot Parabolic Maximum Principle and key corollaries, as proved in \cite{BM}.
\begin{lemma}[Carnot Parabolic Maximum Principle]\label{cpmp}
Let $u$ be a viscosity subsolution to Equation \eqref{main} and $v$ be
a viscosity supersolution to Equation \eqref{main} in the bounded parabolic set $\Omega \times (0,T)$ where $\Omega$ is a (bounded) domain and let $\tau$ be a positive real parameter. Let $\phi(p,q,t)=\varphi(p\cdot q^{-1},t)$ be a $C^2$ function in the space variables $p$ and $q$ and a $C^1$ function in $t$.
Suppose the local maximum
\begin{equation}\label{Mdef}
M_\tau \equiv \max_{\overline{\Omega}\times \overline{\Omega}\times [0,T]}\{u(p,t)-v(q,t)-\tau\phi(p,q,t)\}
\end{equation}
occurs at the interior point $(p_\tau, q_\tau, t_\tau)$ of the parabolic set $\Omega\times \Omega \times (0,T)$.
Define the $n \times n$ matrix $W$ by
\begin{eqnarray*}
W_{ij}=X_i(p)X_j(q)\phi(p_\tau,q_\tau, t_\tau).
\end{eqnarray*}
Let the $2n \times 2n$ matrix $\mathfrak{W}$ be given by
\begin{eqnarray}
\mathfrak{W}=\begin{pmatrix}
0 & \frac{1}{2}(W-W^T) \\
\frac{1}{2}(W^T-W) & 0
\end{pmatrix}
\end{eqnarray}
and let the matrix $\mathcal{W}\in S^{2N}$ be given by
\begin{eqnarray}\label{Adef}
\mathcal{W}=\begin{pmatrix}
D^2_{pp}\phi(p_\tau,q_\tau, t_\tau) & D^2_{pq}\phi(p_\tau,q_\tau, t_\tau) \\
& \\
D^2_{qp}\phi(p_\tau,q_\tau, t_\tau) & D^2_{qq}\phi(p_\tau,q_\tau, t_\tau)
\end{pmatrix}_.
\end{eqnarray}
Suppose $$\lim_{\tau \rightarrow \infty}\tau\phi(p_\tau,q_\tau, t_\tau)=0.$$ Then for each $\tau>0$,
there exists real numbers $a_1$ and $a_2$, symmetric matrices $\mathcal{X}_{\tau}$ and
$\mathcal{Y}_{\tau}$ and
vector $\Upsilon_{\tau} \in V_1 \oplus V_2$, namely
$\Upsilon_{\tau}=\nabla_{1}(p)\phi(p_{\tau}, q_{\tau}, t_\tau)$,
so that the following hold:
\begin{enumerate}
\item [A)] $(a_1,\tau \Upsilon_{\tau},\mathcal{X}_{\tau}) \in \overline{P}^{2,+}u(p_\tau,t_\tau)$ and $(a_2,\tau \Upsilon_{\tau},\mathcal{Y}_{\tau}) \in \overline{P}^{2,-}v(q_\tau,t_\tau).$
\item [B)] $a_1-a_2=\phi_t(p_\tau, q_\tau, t_\tau).$
\item [C)] For any vectors $\xi, \epsilon \in V_1$, we have
\begin{eqnarray}
\ip{\mathcal{X}_{\tau}\xi}{\xi} - \ip{\mathcal{Y}_{\tau}\epsilon}{\epsilon} & \leq &
\tau \ip{(D^2_p\phi)^\star(p_\tau, q_{\tau}, t_\tau) (\xi-\epsilon)}{(\xi-\epsilon)}+\tau\ip{\mathfrak{W}(\xi\oplus\epsilon)}{(\xi\oplus\epsilon)}\nonumber\\
& & \mbox{}+\tau \|\mathcal{W}\|^2\|\mathbb{A}(\hat{p})^T\xi\oplus\mathbb{A}(\hat{q})^T\epsilon\|^2. \label{fake}
\end{eqnarray}
In particular,
\begin{equation}
\ip{\mathcal{X}_{\tau}\xi}{\xi} - \ip{\mathcal{Y}_{\tau}\xi}{\xi}\lesssim
\tau \|\mathcal{W}\|^2\|\xi\|^2.\label{real}
\end{equation}
\end{enumerate}
\end{lemma}
\begin{corollary}\label{cpmpcor}
Let $\phi(p, q, t) = \phi(p,q)=\varphi(p\cdot q^{-1})$ be independent of $t$ and a non-negative function. Suppose $\phi(p,q)=0$ exactly when $p=q$. Then
$$\lim_{\tau\to\infty} \tau \phi(p_\tau, q_\tau)=0.$$ In particular, if
\begin{equation}\label{phidef}
\phi(p, q, t) = \frac{1}{m}\sum_{i=1}^{N}\big((p\cdot q^{-1})_i \big)^m
\end{equation}
for some \textbf{even} integer $m\geq4$ where $(p\cdot q^{-1})_i$ is the $i$-th component of the Carnot group multiplication group law, then for the vector $\Upsilon_\tau$ and matrices $\mathcal{X}_\tau, \mathcal{Y}_\tau$, from the Lemma, we have
\begin{enumerate}
\item [A)] $(a_1,\tau \Upsilon_\tau,\mathcal{X}_{\tau}) \in \overline{P}^{2,+}u(p_\tau,t_\tau)$ and $(a_1,\tau \Upsilon_\tau,\mathcal{Y}_{\tau}) \in \overline{P}^{2,-}v(q_\tau,t_\tau).$
\item [B)] The vector $\Upsilon_\tau$ satisfies $$\| \Upsilon_\tau\| \sim \phi(p_\tau, q_\tau)^{\frac{m-1}{m}}.$$
\item [C)] For any fixed vector $\xi \in V_1$, we have
\begin{equation}
\ip{\mathcal{X}_{\tau}\xi}{\xi} - \ip{\mathcal{Y}_{\tau}\xi}{\xi}\lesssim
\tau \|\mathcal{W}\|^2\|\xi\|^2 \lesssim \tau(\phi(p_\tau, q_\tau))^{\frac{2m-4}{m}}\|\xi\|^2.\label{real2}
\end{equation}
\end{enumerate}
\end{corollary}
\section{Uniqueness of viscosity solutions}
We wish to formulate a comparison principle for the following problem.
\begin{problem}\label{mainprob}
Let $h \geq 1$. Let $\Omega$ be a bounded domain and let $\Omega_T=\Omega\times [0,T)$.
Let $\psi \in C(\overline{\Omega})$ and $g \in C(\overline{\Omega_T})$. We consider the following boundary and initial value problem:
\begin{eqnarray}\label{problem}
\left\{\begin{array}{rl}
u_t+F^h_\infty(\nabla_0 u, (D^2u)^\star) = 0 & \textmd{in}\ \ \Omega \times (0,T) \hspace{.8in}(E)\\
u(p,t)=g(p,t) & p \in \partial \Omega,\ t \in [0,T)\hspace{.5in} (BC)\\
u(p,0) = \psi(p) & p \in \overline{\Omega}\hspace{1.25in}(IC)
\end{array}\right.
\end{eqnarray}
We also adopt the definition that a subsolution $u(p,t)$ to Problem \ref{problem} is a viscosity subsolution to (E), $u(p,t) \leq g(p,t)$ on $\partial \Omega$ with $0 \leq t < T$ and $u(p,0) \leq \psi(p)$ on $\overline{\Omega}$. Supersolutions and solutions are defined in an analogous matter.
\end{problem}
Because our solution $u$ will be continuous, we offer the following remark:
\begin{remark}
The functions $\psi$ and $g$ may be replaced by one function $g\in C(\overline{\Omega_T})$. This combines conditions (E) and (BC) into one condition
\begin{equation}
u(p,t)=g(p,t),\ \ \ (p,t) \in \partial_{\operatorname{par}} \Omega_T \hspace{.5in} (IBC)
\end{equation}
\end{remark}
\begin{thm}\label{pinf}
Let $\Omega$ be a bounded domain in $G$ and let $h\geq 1$. If $u$ is a parabolic viscosity subsolution and $v$ a parabolic viscosity supersolution to Problem \eqref{problem} then $u \leq v$ on $\Omega_T\equiv\Omega\times [0,T) $.
\end{thm}
\begin{proof}
Our proof follows that of \cite[Thm. 8.2]{CIL:UGTVS} and so we discuss only the main parts.
For $\varepsilon > 0$, we substitute $\tilde{u}=u-\frac{\displaystyle\varepsilon}{T-t}$ for $u$ and prove the theorem for
\begin{eqnarray}
u_t+F^h_\infty(\nabla_0 u,(D^2u)^\star) \leq -\frac{\varepsilon}{T^2} < 0 \\
\lim_{t \uparrow T}u(p,t) = -\infty \ \ \textmd{uniformly on }\ \ \overline{\Omega}
\end{eqnarray}
and take limits to obtain the desired result.
Assume the maximum occurs at $(p_0,t_0)\in \Omega \times (0,T)$ with $$u(p_0,t_0)-v(p_0,t_0)= \delta >0.$$
\textbf{Case 1: $\boldsymbol{h>1.}$} \\
Let $H\geq h+3$ be an even number. As in Equation \eqref{phidef}, we let $$\phi(p, q) = \frac{1}{H}\sum_{i=1}^{N}\big((p\cdot q^{-1})_i \big)^H$$
where $(p\cdot q^{-1})_i$ is the $i$-th component of the Carnot group multiplication group law. Let
$$M_\tau=u(p_\tau,t_\tau)-v(q_\tau,t_\tau)-\tau\phi(p_\tau,q_\tau)$$ with $(p_\tau,q_\tau,t_\tau)$ the maximum point in $\overline{\Omega} \times \overline{\Omega} \times [0,T)$ of
$u(p,t)-v(q,t)-\tau \phi(p,q)$.
If $t_\tau=0$, we have
$$0 < \delta \leq M_\tau \leq \sup_{\overline{\Omega}\times\overline{\Omega}}(\psi(p)-\psi(q)-\tau \phi(p,q))$$ leading to a contradiction for large $\tau$. We therefore conclude $t_\tau >0$ for large $\tau$. Since $u \leq v$ on $\partial \Omega \times [0,T)$ by Equation (BC) of Problem \eqref{problem}, we conclude that for large $\tau$, we have $(p_\tau,q_\tau,t_\tau)$ is an interior point. That is,
$(p_\tau,q_\tau,t_\tau) \in \Omega \times \Omega \times (0,T)$.
Using Corollary \ref{cpmpcor} Property A, we obtain
\begin{eqnarray*}
(a,\tau \Upsilon(p_\tau,q_\tau), \mathcal{X}_\tau) & \in & \overline{P}^{2,+}u(p_\tau,t_\tau) \\
(a,\tau \Upsilon(p_\tau,q_\tau), \mathcal{Y}_\tau) & \in & \overline{P}^{2,-}v(q_\tau,t_\tau)
\end{eqnarray*}
satisfying the equations
\begin{eqnarray*}
a+F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{X}_\tau) & \leq & -\frac{\varepsilon}{T^2} \\
a+F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{Y}_\tau) & \geq & 0.
\end{eqnarray*}
If there is a subsequence $\{p_\tau,q_\tau\}_{\tau>0}$ such that $p_\tau\neq q_\tau$, we subtract, and using Corollary \ref{cpmpcor}, we have
\begin{eqnarray} \label{festimate}
0 < \frac{\varepsilon}{T^2} & \leq & (\tau\Upsilon(p_\tau,q_\tau))^{h-3} \tau^2\bigg(\ip{\mathcal{X}_\tau\Upsilon(p_\tau,q_\tau)}{\Upsilon(p_\tau,q_\tau)}-\ip{\mathcal{Y}_\tau\Upsilon(p_\tau,q_\tau)}{\Upsilon(p_\tau,q_\tau)}\bigg)\nonumber \\
& \lesssim & \tau^h\big( \varphi(p_\tau,q_\tau)^{\frac{H-1}{H}}\big)^{h-3}(\varphi(p_\tau,q_\tau))^{\frac{2H-4}{H}} (\varphi(p_\tau,q_\tau))^{\frac{2H-2}{H}} \\
& = & \tau^h(\varphi(p_\tau,q_\tau))^{\frac{Hh+H-h-3}{H}}= (\tau\varphi(p_\tau,q_\tau))^h
\varphi(p_\tau,q_\tau)^{\frac{H-h-3}{H}}.
\end{eqnarray}
Because $H>h+3$, we arrive at a contradiction as $\tau \rightarrow \infty$.
If we have $p_\tau = q_\tau$, we arrive at a contradiction since $$F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{X}_\tau)=F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{Y}_\tau)=0.$$
\noindent\textbf{Case 2: $\boldsymbol{h=1.}$} \\
We follow the proof of Theorem 3.1 in \cite{JK:JK}. We let
$$\varphi(p, q, t, s ) = \frac{1}{4}\sum_{i=1}^{N}\big((p\cdot q^{-1})_i \big)^4+\frac{1}{2}(t-s)^2$$
and let $(p_\tau,q_\tau,t_\tau,s_\tau)$ be the maximum of $$u(p,t)-v(q,s)-\tau \phi(p, q,t ,s)$$
Again, for large $\tau$, this point is an interior point. If we have a sequence where $p_\tau \neq q_\tau$, then Lemma \ref{twist} yields
\begin{eqnarray*}
(\tau (t_\tau-s_\tau),\tau \Upsilon(p_\tau,q_\tau), \mathcal{X}_\tau) & \in & \overline{P}^{2,+}u(p_\tau,t_\tau) \\
(\tau (t_\tau-s_\tau),\tau \Upsilon(p_\tau,q_\tau), \mathcal{Y}_\tau) & \in & \overline{P}^{2,-}v(q_\tau,s_\tau)
\end{eqnarray*}
satisfying the equations
\begin{eqnarray*}
\tau (t_\tau-s_\tau)+F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{X}_\tau) & \leq & -\frac{\varepsilon}{T^2} \\
\tau (t_\tau-s_\tau)+F^h_\infty(\tau\Upsilon(p_\tau,q_\tau), \mathcal{Y}_\tau) & \geq & 0.
\end{eqnarray*}
As in the first case, we subtract to obtain
\begin{eqnarray*}
0 < \frac{\varepsilon}{T^2} & \leq & (\tau\Upsilon(p_\tau,q_\tau))^{-2} \tau^2\bigg(\ip{\mathcal{X}_\tau\Upsilon(p_\tau,q_\tau)}{\Upsilon(p_\tau,q_\tau)}-\ip{\mathcal{Y}_\tau\Upsilon(p_\tau,q_\tau)}{\Upsilon(p_\tau,q_\tau)}\bigg) \\
& \lesssim & \varphi(p_\tau,q_\tau)^{-\frac{3}{2}}(\tau\varphi(p_\tau,q_\tau) \varphi(p_\tau,q_\tau)^{\frac{3}{2}}) = \tau\varphi(p_\tau,q_\tau).
\end{eqnarray*}
We arrive at a contradiction as $\tau\to\infty$.
If $p_\tau = q_\tau$, then
$v(q,s)-\beta^v(q,s)$ has a local minimum at $(q_\tau,s_\tau)$ where
$$\beta^v(q,s)=- \frac{\tau}{4}\sum_{i=1}^{N}\big((p_\tau\cdot q^{-1})_i \big)^4-\frac{\tau}{2}(t_\tau-s)^2.$$ We then have
$$0<\varepsilon(T-s_\tau)^{-2} \leq \beta^v_s(q_\tau,s_\tau)-\min_{\|\eta\|=1}\ip{(D^2\beta^v)^\star (q_\tau,s_\tau)\;\eta}{\eta}.$$
Similarly, $u(p,t)-\beta^u(p,t)$ has a local maximum at $(p_\tau,t_\tau)$ where
$$\beta^u(p,t)=\frac{\tau}{4}\sum_{i=1}^{N}\big((p\cdot q_\tau^{-1})_i \big)^4+\frac{\tau}{2}(t-s_\tau)^2.$$
We then have
$$0 \geq \beta^u_t(p_\tau,t_\tau)-\max_{\|\eta\|=1}\ip{(D^2\beta^u)^\star (p_\tau,t_\tau)\;\eta}{\eta}$$
and subtraction gives us
\begin{eqnarray*}
0<\varepsilon(T-s_\tau)^{-2} & \leq &\max_{\|\eta\|=1}\ip{(D^2\beta^u)^\star (p_\tau,t_\tau)\;\eta}{\eta}-\min_{\|\eta\|=1}\ip{(D^2\beta^v)^\star (q_\tau,s_\tau)\;\eta}{\eta} \\
& & \mbox{}+ \beta^v_s(q_\tau,s_\tau)-\beta^u_t(p_\tau,t_\tau) \\
& = & \tau\max_{\|\eta\|=1}\ip{(D_{pp}^2\varphi(p\cdot q_\tau^{-1}))^\star (p_\tau,t_\tau)\;\eta}{\eta} \\
& & \mbox{}-\tau\min_{\|\eta\|=1}\ip{(D^2_{qq}\varphi(p_\tau\cdot q^{-1}))^\star (q_\tau,s_\tau)\;\eta}{\eta} \\
& & \mbox{}+ \tau(t_\tau-s_\tau)- \tau(t_\tau-s_\tau) \\
& = & 0.
\end{eqnarray*}
Here, the last equality comes from the fact that $p_\tau=q_\tau$ and the definition of $\varphi(p\cdot q^{-1})$.
\end{proof}
The comparison principle has the following consequences concerning properties of solutions:
\begin{corollary}\label{h-infinite}
Let $h\geq 1$.
The past parabolic viscosity h-infinite solutions are exactly the parabolic viscosity h-infinite solutions.
\end{corollary}
\begin{proof}
By Proposition \ref{onedir}, past parabolic viscosity h-infinite sub(super-)solutions are parabolic viscosity h-infinite sub(super-)solutions. To prove the converse, we will follow the proof of the subsolution case found in
\cite{J:PD}, highlighting the main details. Assume that $u$ is not a past parabolic viscosity h-infinite subsolution.
Let $\phi \in \mathcal{A}^-u(p_0,t_0)$ have the property that
$$\phi_t(p_0,t_0)-\Delta^h_\infty \phi(p_0,t_0) \geq \epsilon > 0$$ for a small parameter $\epsilon$. We may assume $p_0$ is the origin.
Let $r > 0$ and define $S_r= B_{\mathcal{N}}(r) \times (t_0-r,t_0)$ and let $\partial S_r$ be its parabolic boundary. Then the function $$\tilde{\phi}_r(p,t)= \phi(p,t)+(t_0-t)^{8l!}-r^{8l!}+(\mathcal{N}(p))^{8l!}$$ is a classical supersolution for sufficiently small $r$. We then observe that $u \leq \tilde{\phi}_r$ on $\partial S_r$ but $u(0,t_0) > \tilde{\phi}(0,t_0)$. Thus, the comparison prinicple, Theorem \ref{pinf}, does not hold. Thus, $u$ is not a parabolic viscosity h-infinite subsolution. The supersolution case is identical and omitted.
\end{proof}
The following corollary has a proof similar to \cite[Lemma 3.2]{JK:JK}.
\begin{corollary}\label{moddef}
Let $u: \Omega_T \to \mathbb{R}$ be upper semicontinuous. Let $\phi\in \mathcal{A}u(p_0,t_0)$. If
\begin{eqnarray}\label{modhyp}
\left\{\begin{array}{cc}
\phi_t(p_0,t_0)-\Delta^1_\infty \phi(p_0,t_0) \leq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) \neq 0 \\
\phi_t(p_0,t_0) \leq 0 & \textmd{when}\ \ \nabla_0 \phi(p_0,t_0) = 0, (D^2\phi)^\star(p_0,t_0)=0
\end{array}\right.
\end{eqnarray}
then $u$ is a viscosity subsolution to (E) of Problem \eqref{problem}.
\end{corollary}
We also have the following function estimates with respect to boundary data.
\begin{corollary}
Let $h\geq1$. Let $g_1, g_2 \in C(\overline{\Omega_T})$ and $u_1$, $u_2$ be parabolic viscosity solutions to Equation \ref{problem}
with boundary data $g_1$ and $g_2$, respectively. Then
$$\sup_{(p,t)\in\Omega_T}|u_1(p,t)-u_2(p,t)|\leq \sup_{(p,t)\in\partial_{\operatorname{par}}\Omega_T}|g_1(p,t)-g_2(p,t)|. $$
\end{corollary}
\begin{proof}
The function $u^+(p,t)=u_2(p,t)+\sup_{(p,t)\in\partial_{\operatorname{par}}\Omega_T}|g_1(p,t)-g_2(p,t)|$ is a parabolic viscosity supersolution with boundary data $g_1$ and the function $u^-(p,t)=u_2(p,t)-\sup_{(p,t)\in\partial_{\operatorname{par}}\Omega_T}|g_1(p,t)-g_2(p,t)|$ is a parabolic viscosity subsolution with boundary data $g_1$. Moreover, $u^- \leq u_1 \leq u^+$ on $\partial_{\operatorname{par}}\Omega_T$ and by Theorem \ref{pinf}
$u^- \leq u_1 \leq u^+$ in $\Omega_T$.
\end{proof}
\begin{corollary}
Let $h\geq1$. Let $g\in C(\overline{\Omega_T})$. Then every parabolic viscosity solution to Problem \ref{problem}
satisfies $$\sup_{(p,t)\in\Omega_T}|u(p,t)|\leq \sup_{(p,t)\in\partial_{\operatorname{par}}\Omega_T}|g(p,t)|$$
\end{corollary}
\begin{proof}
The proof is similar to the previous corollary, but using the functions $u^\pm(p,t)=\pm\sup_{(p,t)\in\partial_{\operatorname{par}}\Omega_T}|g(p,t)|$ instead.
\end{proof}
\section{Existence of Viscosity Solutions}
\subsection{Parabolic Viscosity Infinite Solutions: The Continuity Case} As above, we will focus on the equations of the form \eqref{main}
for continuous and proper $F:[0,T]\times G \times \mathbb{R} \times g \times S^{n_1} \to \mathbb{R}$ that possess a comparison principle such as Theorem \ref{pinf} or \cite[Thm. 3.6]{BM}. We will use Perron's method combined with the Carnot Parabolic Maximum Principle to yield the desired existence theorem. In particular, the following proofs are similar to those found in \cite[Chapter 2]{Giga} except that the Euclidean derivatives have been replaced with horizontal derivatives and the Euclidean norms have been replaced with the gauge norm.
\begin{lemma}\label{lem1}
Let $\mathcal{L}$ be a collection of parabolic viscosity supersolutions to \eqref{main} and let $u(p,t) = \inf\{v(p,t): v \in \mathcal{L}\}$. If $u$ is finite in a dense subset of $\Omega_T=\Omega\times [0,T)$ then $u$ is a parabolic viscosity supersolution to \eqref{main}.
\end{lemma}
\begin{proof}
First note that $u$ is lower semicontinous since every $v \in \mathcal{L}$ is. Let $(p_0,t_0) \in \Omega_T$ and $\phi \in \mathcal{A}u(p_0,t_0)$. Now let
$$\psi(p,t) = \phi(p,t) - \left(d_\mathcal{N}(p_0,p)\right)^{2l!} - |t-t_0|^2$$
and notice that $\psi \in \mathcal{A}u(p_0,t_0)$.
Then
\begin{eqnarray*}
(u-\psi)(p,t) - \left(d_\mathcal{N}(p_0,p)\right)^{2l!} - |t-t_0|^2 &=& (u-\phi)(p,t) \\&\geq& (u-\phi)(p_0,t_0) \\&=& (u-\psi)(p_0,t_0) \\&=& 0
\end{eqnarray*}
yields
\begin{equation}\label{bound}
(u-\psi)(p,t) \geq \left(d_\mathcal{N}(p_0,p)\right)^{2l!} + |t-t_0|^2.
\end{equation}
Since $u$ is lower semicontinuous, there exists a sequence $\{(p_k,t_k)\}$ with $t_k < t_0$ converging to $(p_0,t_0)$ as $k \to \infty$ such that
$$(u-\psi)(p_k,t_k) \to (u-\psi)(p_0,t_0) = 0.$$
Since $u(p,t) = \inf \left\{ v(p,t): v \in \mathcal{L}\right\}$, there exists a sequence $\left\{v_k\right\} \subset \mathcal{L}$ such that $v_k(p_k,t_k) < u(p_k,t_k) + 1/k$ for k = 1,2, \ldots. Since $v_k \geq u$, $\eqref{bound}$ gives us
\begin{equation}\label{bound2}
(v_k - \psi)(p,t) \geq (u-\psi)(p,t) \geq\left(d_\mathcal{N}(p_0,p)\right)^{2l!} + |t-t_0|^2.
\end{equation}
Let $B \subset \Omega$ denote a compact neighborhood of $(p_0,t_0)$. Since $v_k-\psi$ is lower semicontinuous, it attains a minimum in $B$ at a point $(q_k,s_k) \in B$. Then by $\eqref{bound}$ and $\eqref{bound2}$ we have
$$ (u-\psi)(p_k,t_k) + 1/k > (v_k - \psi)(p_k,t_k) \geq (v_k - \psi)(q_k,s_k) \geq \left(d_\mathcal{N}(p_0,q_k)\right)^{2l!} + |s_k-t_0|^2 \geq 0$$
for sufficiently large $k$ such that $(p_k,t_k) \in B$. By the squeeze theorem, $(q_k,s_k) \to (p_0,t_0)$ as $k \to \infty$.
Let $\eta = \psi - \left(d_\mathcal{N}(q_k,p)\right)^{2l!} - |s_k-t|^2$. Then $\eta \in \mathcal{A}v_k(q_k,s_k)$ and we have that
$$\eta_t(q_k,s_k)+F(s_k, q_k, v_k(q_k,s_k),\nabla_1\eta(q_k,s_k), (D^2 \eta(q_k,s_k))^\star) \geq 0.$$
This implies
$$\psi_t(q_k,s_k)+F(s_k, q_k, v_k(q_k,s_k),\nabla_1\psi(q_k,s_k), (D^2 \psi(s_k,s_k))^\star)\geq 0.$$
Letting $k \to \infty$ yields
$$\phi_t(p_0,t_0)+F(t_0,p_0,u(p_0,t_0)\nabla_1\phi(p_0,t_0), (D^2\phi(p_0,t_0))^\star)\geq 0.$$
and that $u$ is a parabolic viscosity supersolution as desired.
\end{proof}
A similar argument yields the following.
\begin{lemma}\label{closedsup}
Let $\mathcal{L}$ be a collection of parabolic viscosity subsolutions to \eqref{main} and let $u(p,t) = \sup\{v(p,t): v \in \mathcal{L}\}$. If $u$ is finite in a dense subset of $\Omega_T$ then $u$ is a parabolic viscosity subsolution to \eqref{main}.
\end{lemma}
For the following lemmas, we need to recall the following definition.
\begin{definition}
The \textit{upper and lower semi-continuous envelopes} of a function $u$ are given by
$$u^*(p,t):= \lim_{r \downarrow 0}\sup \{u(q,s) : |q^{-1}p|_g + |s-t| \leq r\}$$
and
$$u_*(p,t):= \lim_{r \downarrow 0}\inf \{u(q,s) : |q^{-1}p|_g + |s-t| \leq r\},$$
respectively.
\end{definition}
\begin{lemma}\label{maximal}
Let $h$ be a parabolic viscosity supersolution to \eqref{main} in $\Omega_T$. Let $\mathcal{S}$ be the collection of all parabolic viscosity subsolutions $v$ of \eqref{main} satisfying $v \leq h$. If for $\hat{v} \in \mathcal{S}$, $\hat{v}_*$ is not a parabolic viscosity supersolution of $\eqref{main}$ then there is a function $w \in \mathcal{S}$ and a point $(p_0,t_0)$ such that $\hat{v}(p_0,t_0) < w(p_0,t_0)$.
\end{lemma}
\begin{proof} Let $\hat{v} \in \mathcal{S}$ such that $\hat{v}_*$ is not a parabolic viscosity supersolution of \eqref{main}. Then there exists $(\hat{p},\hat{t}) \in \Omega_T$ and $\phi \in \mathcal{A}\hat{v}_*(\hat{p},\hat{t})$ such that
\begin{equation}\label{contradict}
\phi_t(p,t)+F(t,p,\hat{v}_*(p,t),\nabla_1\phi(p,t), (D^2\phi(p,t))^\star) > 0.
\end{equation}
Let
$$\psi(p,t) = \phi(p,t) - \left(d_\mathcal{N}(\hat{p},p)\right)^{2l!} - |t-\hat{t}|^2$$
and notice that $\psi \in \mathcal{A}\hat{v}_*(\hat{p},\hat{t})$.
As in Lemma \ref{lem1},
\begin{equation}\label{bounded1}
(\hat{v}_*-\psi)(p,t) \geq \left(d_\mathcal{N}(\hat{p},p)\right)^{2l!} + |t-\hat{t}|^2.
\end{equation}
Let $B$ denote a compact neighborhood of $(\hat{p},\hat{t})$ and let $$B_{k\epsilon} = B \cap \left\{ (p,t) : \left(d_\mathcal{N}(\hat{p},p)\right)^{2l!} \leq k\epsilon \text{ and } |t-\hat{t}|^2 \leq k\epsilon \right\}.$$ Since $\hat{v} \in \mathcal{S}$, we have that $\hat{v} \leq h$ and thus $\psi(\hat{p},\hat{t})=\hat{v}_*(\hat{p},\hat{t}) \leq \hat{v}(\hat{p},\hat{t}) \leq h(\hat{p},\hat{t})$. However, if $\psi(\hat{p},\hat{t}) = h(\hat{p},\hat{t})$, then $\psi \in \mathcal{A}h(\hat{p},\hat{t})$ and inequality $\eqref{contradict}$ would be contradictory. Thus,
$$\psi(\hat{p},\hat{t}) < h(\hat{p},\hat{t}).$$
Since $\psi$ is continuous and $h$ is lower semicontinuous, there exists $\epsilon > 0$ such that
$$\psi(p,t) + 4\epsilon \leq h(p,t)$$
for $(p,t) \in B_{2\epsilon}$. Notice that $\psi + 4\epsilon$ is a subsolution of \eqref{main} on the interior of $B_{2\epsilon}$. Further, by $\eqref{bounded1}$
\begin{equation}\label{bounded2}
\hat{v}(p,t) \geq \hat{v}_*(p,t) \geq \psi(p,t) + 4\epsilon \text{ for } (p,t) \in B_{2\epsilon} \backslash B_{\epsilon}.
\end{equation}
We now define $\omega$ by
\begin{equation*}
\omega = \left\{ \begin{array}{ll}
\max\{\psi(p,t) + 4\epsilon,\hat{v}(p,t)\} & (p,t) \in B_\epsilon \\
\hat{v}(p,t) & (p,t) \in \Omega_T\backslash B_\epsilon
\end{array} \right.
\end{equation*}
But by $\eqref{bounded2}$
$$\omega(p,t) = \max\{\psi(p,t) + 4\epsilon,\hat{v}(p,t)\} \text{ for } (p,t) \in B_{2\epsilon},$$
not just for $(p,t) \in B_\epsilon$. Then by Lemma \ref{closedsup}, $\omega$ is a subsolution in the interior of $B_{2\epsilon}$
and thus a subsolution in $\Omega_T$. Therefore, $\omega \in \mathcal{S}$. Since
$$0 = (\hat{v}_*-\psi)(\hat{p},\hat{t}) = \lim_{r \downarrow 0} \inf\left\{ (\hat{v} - \psi)(p,t) : (p,t) \in B_r \right\}$$
there is a point $(p_0,t_0) \in B_\epsilon$ that satisfies
$$\hat{v}(p_0,t_0) - \psi(p_0,t_0) < 4\epsilon$$
which yields
$$\hat{v}(p_0,t_0) < \psi(p_0,t_0) + 4\epsilon = \omega(p_0,t_0).$$
Thus, we have constructed $\omega \in \mathcal{S}$ that satisfies $\hat{v}(p_0,t_0) < \omega(p_0,t_0)$.
\end{proof}
We then have the following existence theorem concerning parabolic viscosity solutions.
\begin{thm}\label{exist}
Let $f$ be a parabolic viscosity subsolution to \eqref{main} and $g$ be a parabolic viscosity supersolution to \eqref{main} satisfying $f \leq g$ on $\Omega_T$ and $f_* = g^*$ on $\partial_{\operatorname{par}}\mathcal{O}_{0,T}$. Then there is a parabolic viscosity solution $u$ to \eqref{main} satisfying $u \in C(\overline{O_T})$. Explicitly, there exists a unique parabolic viscosity infinite solution to Problem \ref{mainprob} when $h>1$.
\end{thm}
\begin{proof}
Let
$$S = \{\nu : \hspace{3pt} \nu \text{ is a parabolic viscosity subsolution to \eqref{main} in } \Omega_T \text{ with } \nu \leq g \text{ in } \Omega_T\}$$
and
$$ u(p,t) = \sup \{ \nu(p,t) : \hspace{3pt} \nu \in S\}.$$
Since $f \leq g$, the set $S$ is nonempty. Notice that $f\leq u \leq g$ by construction. By Lemma $\eqref{closedsup}$, $u$ is a parabolic viscosity subsolution. Suppose $u_*$ is not a parabolic viscosity supersolution. Then by Lemma \ref{maximal}, there exists a function $w \in S$ and a point $(p_0,t_0) \in \Omega_T$ such that $u(p_0,t_0) < w(p_0,t_0)$. But this contradicts the definition of $u$ at $(p_0,t_0)$. Thus $u_*$ is a parabolic viscosity supersolution. By our assumptions on $f$ and $g$ on $\partial_{\operatorname{par}}\mathcal{O}_{0,T}$,
$$u=u^* \leq g^* = f_* \leq u_*$$
on $\partial_{\operatorname{par}}\mathcal{O}_{0,T}$. Then by the (assumed) comparison principle, $u \leq u_*$ on $\Omega_T$. Thus we have $u$ is a parabolic viscosity solution such that $u \in C(\overline{O_T})$.
\end{proof}
\subsection{The $h=1$ case}
We begin by recalling the definition of upper and lower relaxed limit of a function. \cite{CIL:UGTVS, Giga}.
\begin{definition}
For $\varepsilon>0$, consider the function $h_\varepsilon:O_T\subset G\to\mathbb{R}$. The \emph{upper relaxed limit} $\overline{h}(p,t)$ and the lower relaxed limit $\underline{h}(p,t)$ are given by
\begin{eqnarray*}
\overline{h}(p,t) & = & \limsup_{\hat{p}\to p,\hat{t}\to t, \varepsilon\to 0} h_\varepsilon(\hat{p}, \hat{t}) \\
& = & \lim_{\varepsilon\to 0} \sup_{0<\delta<\varepsilon}\{h_\delta(\hat{p},\hat{t}): O_T\cap B_\varepsilon(\hat{p},\hat{t})\} \\
\textmd{and\ } \underline{h}(p,t) & = & \liminf_{\hat{p}\to p,\hat{t}\to t, \varepsilon\to 0} h_\varepsilon(\hat{p}, \hat{t}) \\
& = & \lim_{\varepsilon\to 0} \inf_{0<\delta<\varepsilon}\{h_\delta(\hat{p},\hat{t}): O_T\cap B_\varepsilon(\hat{p},\hat{t})\} \\
\end{eqnarray*}
\end{definition}
Taking the relaxed limits as $h\to 1^+$ of the operator $F^h_\infty(\nabla_0 u,(D^2u)^{\star})$ in Equation \ref{relax}, we have via the continuity of the operator
\begin{eqnarray*}
\overline{F}^1_\infty(\nabla_0 u,(D^2u)^{\star}) = \underline{F}^1_\infty(\nabla_0 u,(D^2u)^{\star})=
\left\{\begin{array}{cl}
-\|\nabla_0u\|^{-2}\ip{(D^2u)^{\star}\nabla_0u}{\nabla_0u} & \nabla_0u \neq 0 \\
0 & \nabla_0u = 0. \end{array}\right.
\end{eqnarray*}
We give this operator the label $\mathcal{F}(\nabla_0 u,(D^2u)^{\star})$.
Consider the relaxed limits $\overline{u}(p,t)$ and $\underline{u}(p,t)$ of the sequence of unique (continuous) viscosity solutions to Problem \ref{mainprob} $\{u_h(p,t)\}$ as $h\to 1^+$. By \cite[Thm 2.2.1]{Giga}, we have $\overline{u}(p,t)$ is a viscosity subsolution and $\underline{u}(p,t)$ is a viscosity supersolution to
$$u_t+\mathcal{F}(\nabla_0 u,(D^2u)^{\star})=0.$$
We have the following comparison principle, whose proof is similar to Theorem \ref{pinf} in the case to $h=1$ and is omitted.
\begin{lemma}
Let $\Omega$ be a bounded domain in $G$. If $\mathfrak{u}$ is a parabolic viscosity subsolution and $\mathfrak{v}$ a parabolic viscosity supersolution to $$u_t+\mathcal{F}(\nabla_0 u,(D^2u)^{\star})=0.$$ then $\mathfrak{u} \leq \mathfrak{v}$ on $\Omega_T\equiv\Omega\times [0,T) $.
\end{lemma}
\begin{corollary}
$\overline{u}(p,t)=\underline{u}(p,t)$.
\end{corollary}
\begin{proof}
By construction, $\underline{u}(p,t)\leq \overline{u}(p,t)$. By the Lemma, $\underline{u}(p,t)\geq \overline{u}(p,t)$.
\end{proof}
Using the corollary, we will call this common relaxed limit $u^1(p,t)$. By \cite[Chapter 2]{Giga} and \cite[Section 6]{CIL:UGTVS}, it is continuous and the sequence $\{u_h(p,t)\}$ converges locally uniformly to $u^1(p,t)$ as $h\to 1^+$.
We then have the following theorem.
\begin{thm}\label{oneexist}
There exists a unique parabolic viscosity infinite solution to Problem \ref{mainprob} when $h=1$.
\end{thm}
\begin{proof}
Let $\{u_h(p,t)\}$ and $u^1(p,t)$ be as above. Let $\{h_j\}$ be a subsequence with $h_j\to 1^+$ where
$u_h(p,t)\to u^1(p,t)$ uniformly. We may assume $h_j<3$.
Let $\phi\in\mathcal{A}u_1(p_0,t_0)$. Using the uniform convergence, there is a sequence $\{p_j,t_j\}\to (p_0,t_0)$ so that $\phi \in \mathcal{A}u_{h_j}(p_j,t_j)$. If $\nabla_0\phi(p_0,t_0)\neq 0$, we have $\nabla_0\phi(p_j,t_j)\neq 0$ for sufficiently large $j$.
We then have
$$\phi_t(p_j,t_j)-\Delta^{h_j}_\infty \phi(p_j,t_j) \leq 0$$ and letting $j\to\infty$ yields
$$\phi_t(p_0,t_0)-\Delta^1_\infty \phi(p_0,t_0) \leq 0.$$
Suppose $\nabla_0\phi(p_0,t_0)=0$. By Corollary \ref{moddef}, we may assume $(D^2\phi)^\star(p_0,t_0)=0$. Suppose passing to a subsequence if needed, we have $\nabla_0\phi(p_j,t_j)\neq 0$.
Then
\begin{equation*}
\phi_t(p_j,t_j)-\displaystyle \max_{\|\eta\|=1}\ip{(D^2\phi)^\star(p_j,t_j)\;\eta}{\eta} \leq
\phi_t(p_j,t_j)-\Delta^{h_j}_\infty \phi(p_j,t_j) \leq 0.
\end{equation*}
Letting $j\to\infty$ yields
$$\phi_t(p_0,t_0)=\phi_t(p_j,t_j)-\displaystyle \max_{\|\eta\|=1}\ip{(D^2\phi)^\star(p_0,t_0)\;\eta}{\eta}\leq 0.$$
In the case $\nabla_0\phi(p_j,t_j)= 0$, since $h_j<3$, we have $\phi_t(p_j,t_j) \leq 0$ and letting $j\to\infty$ yields $\phi_t(p_0,t_0)\leq 0$. We conclude that $u_1$ is a parabolic viscosity h-infinite subsolution. Similarly,
$u_1$ is a parabolic viscosity h-infinite supersolution.
\end{proof}
\section{ The limit as $t\to\infty$.}
We now focus our attention on the asymptotic limits of the parabolic viscosity h-infinite solutions. We wish to show that for $1 \leq h$, we have the (unique) viscosity solution to $u_t-\Delta^h_\infty u=0$ approaches the viscosity solution of $-\Delta^h_\infty u=0$ as $t\to\infty$. Our goal is the following theorem:
\begin{thm}\label{final} Let $h>1$ and $u \in C(\overline{\Omega} \times [0,\infty))$ be a viscosity solution of
\begin{equation}\label{finaleqn}
\left\{ \begin{array}{cl}
u_t -\Delta^h_\infty u =0 & \hspace{10pt} in \hspace{5pt} \Omega \times (0,\infty),\\
u(p,t) =g(p) & \hspace{10pt} on \hspace{5pt} \partial_{\operatorname{par}}(\Omega \times (0, \infty))
\end{array}\right.
\end{equation}
with $g:\overline{\Omega} \to \mathbb{R}$ continuous and assuming that $\partial \Omega$ satisfies the property of positive geometric density (see \cite[pg. 2909]{J:PD}). Then $u(p,t) \to U(p)$ uniformly in $\Omega$ as $t \to \infty$ where $U(p)$ is the unique viscosity solution of $-\Delta^h_\infty U = 0$ with the Dirichlet boundary condition $lim_{q \to p} U(q) = g(p)$ for all $p \in \partial \Omega$.
\end{thm}
We first must establish the uniqueness of viscosity solutions to the limit equation. Note that for future reference, we include the case $h=1$.
\begin{thm}
Let $1\leq h< \infty$ and let $\Omega$ be a bounded domain. Let $u$ be a viscosity subsolution to $\Delta^h_\infty u =0$
and let $v$ be a viscosity supersolution to $-\Delta^h_\infty u =0$.
Then,
\begin{equation*}
\sup_{p \in \overline{\Omega}} (u(p)-v(p)) = \sup_{p \in \partial \Omega} (u(p)-v(p)).
\end{equation*}
\end{thm}
\begin{proof}
Let $u$ be a viscosity subsolution to $-\Delta^h_\infty u =0$. Then choose $\phi \in \operatorname{\mathcal{C}^2_{\textmd{sub}}}(\Omega)$ such that $0 = \phi(p_0) - u(p_0) < \phi(p) - u(p)$ for $p \in \Omega$, $p \neq p_0$.
If $\|\nabla_0 \phi(p_0)\| = 0$, then $-\ip{(D^2\phi)^\star(p_0)\nabla_0\phi(p_0)}{\nabla_0\phi(p_0)} = 0 \leq 0$.
If $\|\nabla_0 \phi(p_0)\| \neq 0$, we then have
$$-\Delta^h_\infty \phi(p_0) = -\|\nabla_0\phi(p_0)\|^{h-3}\ip{(D^2\phi)^\star(p_0)\nabla_0\phi(p_0)}{\nabla_0\phi(p_0)} \leq 0.$$
Dividing, we have $-\ip{(D^2\phi)^\star(p_0)\nabla_0\phi(p_0)}{\nabla_0\phi(p_0)} \leq 0$. In either case, $u$ is a viscosity subsolution to $-\Delta^3_\infty u=0$. Similarly, $v$ is a viscosity supersolution to $-\Delta^3_\infty u=0$. The theorem follows from the corresponding result for $-\Delta^3_\infty u=0$ in \cite{B:MP, B:HG, W:W}.
\end{proof}
We state some obvious corollaries:
\begin{corollary}
Let $1\leq h< \infty$ and let $g:\partial \Omega \to \mathbb{R}$ be continuous. Then there exactly one solution to
\begin{equation*}
\left\{ \begin{array}{cl}
-\Delta^h_\infty u =0 & \hspace{10pt} in \hspace{5pt} \Omega \\
u=g & \hspace{10pt} on \hspace{5pt} \partial \Omega.
\end{array}\right.
\end{equation*}
\end{corollary}
\begin{corollary}\label{threeisone}
Let $1\leq h< \infty$ and let $g:\partial \Omega \to \mathbb{R}$ be continuous. The unique viscosity solution to
\begin{equation*}
\left\{ \begin{array}{cl}
-\Delta^h_\infty u =0 & \hspace{10pt} in \hspace{5pt} \Omega \\
u=g & \hspace{10pt} on \hspace{5pt} \partial \Omega
\end{array}\right.
\end{equation*}
is the unique viscosity solution to
\begin{equation*}
\left\{ \begin{array}{cl}
-\Delta^3_\infty u =0 & \hspace{10pt} in \hspace{5pt} \Omega \\
u=g & \hspace{10pt} on \hspace{5pt} \partial \Omega.
\end{array}\right.
\end{equation*}
\end{corollary}
Our method of proof for Theorem \ref{final} follows that of \cite[Theorem 2]{J:PD}, the core of which hinges on the construction of a parabolic test function from an elliptic one. In order to construct such a parabolic test function, we need to examine the homogeneity of Equation \eqref{finaleqn}. A quick calculation shows that for a fixed $h>1$, $k^{\frac{1}{h-1}}u(x,kt)$ is a $\operatorname{\mathcal{C}^2_{\textmd{sub}}}$ solution to Equation \eqref{finaleqn} if $u(x,t)$ is a $\operatorname{\mathcal{C}^2_{\textmd{sub}}}$ solution. A routine calculation then shows parabolic viscosity h-infinite solutions share this homogeneity. We use this property in the following lemma, the proof of which can be found in \cite[pg. 170]{DiB}. (Also, cf. \cite[Lemma 6.2] {BM} and \cite{J:PD}.)
\begin{lemma}\label{bounduth}
Let $u$ be as in Theorem \ref{final} and $h>1$. Then for every $(x,t) \in \Omega \times (0,\infty)$ and for $0 < \mathcal{T}< t$, we have
\begin{equation*}
|u(x,t-\mathcal{T}) - u(x,t)| \leq \frac{2 ||g||_{\infty,\Omega}}{h-1}\left(1-\frac{\mathcal{T}}{t}\right)^{\frac{h}{1-h}}\frac{\mathcal{T}}{t}
\end{equation*}
\end{lemma}
\begin{proof} \textbf{[Theorem \ref{final}]} Fix $h>1$. Let $u$ be a viscosity solution of \eqref{finaleqn}. The results of \cite[Chapter III]{DiB} imply that the family $\{u(\cdot,t): t \in (0,\infty)\}$ is equicontinuous. Since it is uniformly bounded due to the boundedness of $g$, Arzela-Ascoli's theorem yields that there exists a sequence $t_j \to \infty$ such that $u(\cdot,t_j)$ converge uniformly in $\overline{\Omega}$ to a function $U \in C(\overline{\Omega})$ for which $U(p) = g(p)$ for all $p \in \partial \Omega$. Since it is known from \cite[Lemma 5.5]{B:MP} that the Dirichlet problem for the subelliptic $\texttt{p}$-Laplace equation possesses a unique solution, it is enough to show that $U$ is a viscosity $\texttt{p}$-subsolution to $-\Delta_\texttt{p} U = 0$ on $\Omega$. With that in mind, let $p_0 \in \Omega$ and choose $\phi \in \operatorname{\mathcal{C}^2_{\textmd{sub}}}(\Omega)$ such that $0 = \phi(p_0) - U(p_0) < \phi(p) - U(p)$ for $p \in \Omega$, $p \neq p_0$. Using the uniform convergence, we can find a sequence $p_j \to p_0$ such that $u(\cdot,t_j) - \phi$ has a local maximum at $p_j$. Now define
$$\phi_j(p,t) = \phi(p) + C\left(\frac{t}{t_j}\right)^{\frac{h}{1-h}}\frac{t_j-t}{t_j},$$
where $C = 2||g||_{\infty,\Omega}/(h-1)$. Note that $\phi_j(p,t) \in \operatorname{\mathcal{C}^2_{\textmd{sub}}}(\Omega \times (0, \infty))$. Then using Lemma \ref{bounduth},
\begin{eqnarray*}
u(p_j,t_j) - \phi_j(p_j,t_j) &=& u(p_j,t_j) - \phi(p_j) \geq u(p,t_j) - \phi(p)\\
&\geq& u(p,t) - \phi(p) - C\left(\frac{t}{t_j}\right)^{\frac{h}{1-h}}\frac{t_j-t}{t_j}\\
&=& u(p,t) - \phi_j(p,t)
\end{eqnarray*}
for any $p \in \Omega$ and $0<t < t_j$. Thus we have that $\phi_j$ is an admissible test function at $(p_j,t_j)$ on $\Omega \times [0,T]$. Therefore,
$$(\phi_j)_t(p_j,t_j)-\Delta^h_{\infty}\phi_j(p_j,t_j) \leq 0.$$
This yields
$$- \Delta^h_{\infty}\phi(p_j) \leq \frac{C}{t_j}.$$
The theorem follows by letting $j \to \infty$.
\end{proof}
Combining the results of the previous sections, we have the following theorem:
\begin{thm}
The following diagram commutes:
$$\begin{CD}
u^{h,t}_t-\Delta^h_{\infty}u^{h,t}=0 @>>{h\to 1^+}>u^{1,t}_t-\Delta^1_{\infty}u^{1,t}=0 \\
@VV{t\to \infty}V @VV{t\to\infty}V \\
-\Delta^h_{\infty}u^{h,\infty}=0 @>>{h\to 1^+}> -\Delta^1_{\infty}u^{1,\infty}=0
\end{CD}$$
\end{thm}
\begin{proof}
By Theorem \ref{finaleqn}, Corollary \ref{threeisone}, and Theorem \ref{oneexist}, the top, bottom and left limits exist, with the left limit being a uniform limit. By results of iterated limits (see, for example, \cite{Ba}), we have the fourth limit exists, as does the full limit. In particular,
$$\lim_{\stackrel{h \to 1^+}{t\to\infty}}u^{h,t}=\lim_{h \to 1^+}\lim_{t\to\infty}u^{h,t}=\lim_{t\to\infty}\lim_{h \to 1^+}u^{h,t}=u^{1,\infty}$$
\end{proof}
|
\section{Introduction}
In statistics in general and perhaps especially in spatial statistics we
often find ourselves with distributions known only up to an
unknown normalising constant. Calculating this
normalising constant typically involves high dimensional
summation or integration. This is the case for the class of
discrete Markov random fields (MRF).
A common situation in spatial statistics is that we have some
unobserved latent field $x$
for which we have noisy observations $y$. We model $x$ as an MRF with
unknown parameters
$\theta$ for which we want to do inference of some kind. If we are
Bayesians we could
imagine adding a prior for our parameters $\theta$ and
studying the posterior distribution
$p(\theta |y)$. A frequentist approach could involve finding a maximum
likelihood estimator for
our parameters.
Without
the normalising constant, these become non-trivial tasks.
There are a number of techniques that have been proposed to overcome this
problem. The normalising constant can be estimated by running a
Markov chain Monte Carlo (MCMC) algorithm,
which can then be combined with various techniques to produce maximum
likelihood estimates, see for instance
\citet{art52}, \citet{art111} and \citet{art134}.
Other approaches take advantage of the fact that exact sampling
can be done, see \citet{art120}. In the present report however, we focus
on the class of deterministic methods, where deterministic in this
setting is referring to that repeating the estimation process
yields the same estimate.
In \citet{art100} the authors devise
a computationally efficient algorithm for handling so called
general factorisable models of
which MRFs are a common example. This algorithm, which we refer to from
here on as
the variable elimination algorithm, grants a large computational saving
in calculating the
normalising constant by exploiting the factorisable structure of
the models. For MRFs
defined on a lattice this allows for calculation of the normalising
constant on lattices with
up to around 20 rows for models with first order neighbourhoods.
In \citet{art105} and \citet{art117} the authors construct
approximations for larger lattices by doing
computations for a number of sub-lattices using the algorithm
in \citet{art100}.
The energy function of a binary MRF is an example of a so called
pseudo-Boolean function. In general, a pseudo-Boolean function is a
function of the following type,
$f:\{0,1\}^n\rightarrow \mathbb{R}$.
A full representation of a
pseudo-Boolean function requires $2^n$ terms. Finding approximate
representations of pseudo-Boolean functions that require fewer
coefficients is a well studied field, see
\citet{art137} and \citet{art138}. In \citet{book35} the authors show
how any pseudo-Boolean function can be expressed as a
binary polynomial in $n$ variables. \citet{art130} expressed the energy
function of MRFs in this manner and by dropping small terms during the
variable elimination algorithm constructed an approximate MRF.
Our approach and the main contribution of this paper is to apply and
extend approximation theory for pseudo-Boolean function to design
an approximate variable elimination algorithm. By approximating the
binary polynomial representing the distribution before summing
out each variable we get an algorithm less restricted by the
dependence structure of the model, thus capable of handling MRFs
defined on large lattices and MRFs with larger neighbourhood structures.
For the MRF application this
approximation defines an approximation to the normalising constant, and
as a by-product we also get a partially ordered Markov model (POMM)
approximation to the MRF. For the POMM approximation
we can calculate the normalising
constant and evaluate the likelihood, as well as generate
realizations. We also discuss how to modify our approximation strategy
to instead get upper and lower bounds for the normalising constant,
and how this in turn can be used to construct an interval in which
the maximum likelihood estimate must lie. We also discuss another
variant of the approach which produces an approximate version of the Viterbi
algorithm \citep{col5}.
The article has the following layout. In Section \ref{sec:pbf} we define
pseudo-Boolean functions and give a number of approximation theorems
for this function class. Thereafter, in Section \ref{sec:mrf} we
introduce binary MRFs and the variable elimination algorithm,
and in Section \ref{sec:mrfApprox} we apply
the approximation theorems for pseudo-Boolean functions to define our
approximative variable elimination algorithm for binary MRFs.
In Section \ref{sec:marg} we define a modified variant of the approximate
variable elimination algorithm through which we obtain upper and lower bounds for
the normalising constant of a binary MRF, and we discuss how to modify the
approximation algorithm to obtain an
approximate version of the Viterbi algorithm. We briefly discuss some
implementational issues in Section \ref{sec:implementation}, and
in Section \ref{sec:examples} we present simulation
and data examples. Finally,
in Section \ref{sec:cr} we provide closing remarks.
\section{\label{sec:pbf}Pseudo-Boolean functions}
In this section we introduce the class of pseudo-Boolean functions and discuss
various aspects of approximating pseudo-Boolean functions partly based on results of
\citet{art137} and \citet{art138}.
\subsection{\label{sec:pbfdef}Definition and notation}
Let $x=(x_1,\hdots,x_n) \in \Omega = \{0,1\}^n$ be a vector of
binary variables and let $N=\{1,\hdots,n\}$ be the corresponding list of indices. Then for any subset $\Lambda
\subseteq N$ we associate an incidence vector $x$ of length $n$ whose $k$th
element is $1$ if $k\in \Lambda$ and $0$ otherwise. We refer to
an element of $x$, $x_k$, as being "on" if it has value $1$ and "off" if it is $0$. A pseudo-Boolean
function $f$, of dimension $\mbox{dim}(f) = n$, is a function that
associates a real value to each vector, $x \in \{0,1\}^n$, i.e $f:\{0,1\}^n\rightarrow \mathbb{R}$.
\cite{book35} showed that any
pseudo-Boolean function can be expressed uniquely as a binary polynomial,
\begin{equation}
f(x) = \sum_{\Lambda \subseteq N} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k,
\label{eq:pb1}
\end{equation}
where $\beta^{\Lambda}$ are real
coefficients which we refer to as interactions. We define the degree of
$f$, $\mbox{deg}(f)$ as the degree of the
polynomial.
In general the
representation of a function in this manner requires $2^n$
coefficients. In some cases one or more
$\beta^{\Lambda}$ might be zero and in this case a reduced
representation of the pseudo-Boolean function can be defined by
excluding some or all the terms in the sum in \eqref{eq:pb1} where
$\beta^\Lambda = 0$. Thus we get,
\begin{equation}\label{eq:canonical}
f(x) = \sum_{\Lambda \in S} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k,
\end{equation}
where $S$ is a set of subsets of $N$ at least containing all $\Lambda
\subseteq N$ for which $\beta^\Lambda \neq 0$. We then say that $f(x)$ is
represented on $S$. Moreover, we say that our
representation of $f$ is dense if for all $\Lambda \in S$, all subsets
of $\Lambda$ are also included in $S$. The minimal dense representation of
$f$ is thereby \eqref{eq:canonical} with,
\begin{equation}
S = \{\lambda \subseteq N:\beta^\Lambda \neq 0 \text{ for some }
\Lambda \supseteq \lambda \}.
\label{eq:Sdef}
\end{equation}
Throughout this report we restrict the attention to dense
representations of pseudo-Boolean functions.
We also need notation for some subsets of $S$ and $\Omega$.
For $\lambda\in S$ we define $S_\lambda = \{\Lambda
\in S: \lambda \subseteq \Lambda\}$, the set of all
interactions that include $\lambda$. For example if
$n=3$ and $S=\{ \emptyset,\{ 1\},\{ 2\},\{ 3\},\{ 1,2\}, \{ 1,3\},
\{ 2,3\},\{ 1,2,3\}\}$, we have
$S_{\{1,2\}} = \{\{1,2\},\{1,2,3\}\}$.
Equivalently for the set $\Omega$, for $\lambda\in S$ we define
$\Omega_\lambda = \{x \in \Omega: x_k=1
,\text{ }\forall k \in \lambda\}$, the set of all states $x$ where
$x_k$ are on for all $k\in\lambda$. For the same $S$ as above we have
for instance $\Omega_{\{1,2\}} = \{(1,1,0),(1,1,1)\}$. We
will use also the complements of these two
subsets, $S_\lambda^c = S \setminus S_\lambda$ and $\Omega_\lambda^c = \Omega \setminus
\Omega_\lambda$.
Lastly we define $S_\lambda^0 = \{\Lambda \in S: \lambda
\cap \Lambda = \emptyset\}$ and $\Omega_\lambda^0 = \{x \in \Omega:x_k=0,\text{ }
\forall k \in \lambda\}$. If we think of the sets $S_\lambda$ and
$\Omega_\lambda$ as the sets where $\lambda$ is on, then $S_\lambda^0$
and $\Omega_\lambda^0$ are the sets where $\lambda$ is off. Again, using the same example $S$
as above we have for
instance $S_{\{1,2\}}^0 = \{\emptyset,3\}$ and $\Omega_{\{1,2\}}^0 =
\{\{0,0,0\},\{0,0,1\}\}$. Note that in general $S_\lambda^c \neq
S_\lambda^0$ and equivalently $\Omega_\lambda^c \neq \Omega_\lambda^0$.
\subsection{\label{sec:pbfapprox}Approximating pseudo-Boolean functions}
For a general pseudo-Boolean function, the number of possible interactions in our representation grows
exponentially with the dimension $n$. It is therefore natural to ask if we can
find an approximate representation which requires less memory. We could choose
some set $\tilde{S} \subseteq S$ to define our approximation, thus choosing
which interactions to retain, $\tilde{S}$, and which to remove, $S
\setminus \tilde{S}$. For a given $\tilde{S}$ our interest lies in the best such approximation
according to some criteria. We define
$A_{\tilde{S}}\{f(x)\} = \tilde{f}(x) = \sum_{\Lambda \in \tilde{S}} \tilde{\beta}^{\Lambda} \prod_{k \in \Lambda}x_k$ as the
operator which returns the approximation that, for some given approximation set $\tilde{S}$, minimises the
error sum of squares ($\mbox{SSE}$),
\begin{equation}\label{eq:SSE}
\mbox{SSE}(f,\tilde{f}) = \sum_{x \in \Omega} \left\{f(x)-\tilde{f}(x) \right\}^2.
\end{equation}
We find the best approximation by
taking partial derivatives with respect to $\tilde{\beta}^\lambda$ for all
$\lambda \in \tilde{S}$ and setting these expressions equal to zero. This gives us a system of linear equations,
\begin{equation}
\sum_{x \in \Omega_\lambda}\tilde{f}(x) = \sum_{x \in \Omega_\lambda}\left(\sum_{\Lambda \in \tilde{S}} \tilde{\beta}^{\Lambda} \prod_{k \in
\Lambda}x_k\right) = \sum_{x \in \Omega_\lambda}f(x), \text{ } \forall \text{ } \lambda
\in \tilde{S}.
\label{eq:equations}
\end{equation}
Existence and uniqueness of a solution is assured since we have
$|\tilde{S}|$ unknown variables and the same number of linearly
independent equations. If $\tilde{S} \supseteq S$ the best approximation is clearly
the function itself, $\tilde{f}(x) = f(x)$.
It is common practice in statistics and approximation theory in
general to approximate higher order terms by lower order terms. A natural way to design an
approximation would be to let $\tilde{S}$
include all interactions of degree less than or equal to some value
$k$. In \citet{art137} the authors focus on approximations of this type
and proceed to show how the resulting system of linear equations through clever
reorganisation can be transformed into a lower triangular system. They solve this
for $k=1$ and $k=2$ as well as proving a
number of useful properties. \citet{art138} solve this for general $k$.
In the present article we consider the situation where $\tilde{S}$
is a dense subset of $S$. So if $\lambda \in
\tilde{S}$, then all $\Lambda \subset \lambda$ must also be included
in $\tilde{S}$.
Clearly the approximation using all interactions up to degree $k$ is a special
case of our class of approximations. Our motivation for studying this
particular design of $\tilde{S}$ will become clear as we study the
variable elimination algorithm in Section \ref{sec:mrf}.
We now give some useful properties of this
approximation. The two first theorems are from \citet{art137}.
\begin{theorem}
\label{th:1}
The above approximation $A_{\tilde{S}}\{f(x)\}$ is a
linear operator, i.e. for any constants $a,b \in \mathbb{R}$ and
pseudo-Boolean functions $g(x)$ and $h(x)$ represented on $S$,
we have that $A_{\tilde{S}}\{a g(x) + b h(x)\} =
a A_{\tilde{S}}\{g(x)\} + b A_{\tilde{S}}\{h(x)\}$.
\end{theorem}
\begin{theorem}
\label{th:2}
Assume we have two approximations of $f(x)$, $A_{\tilde{S}}\{f(x)\}$ and $A_{\tilde{\tilde{S}}}\{f(x)\}$,
such that $\tilde{\tilde{S}} \subseteq \tilde{S}
\subseteq S$. Then $A_{\tilde{\tilde{S}}}[A_{\tilde{S}}\{f(x)\}] = A_{\tilde{\tilde{S}}}\{f(x)\}$.
\end{theorem}
Proofs can be found in \citet{art137} and in \citet{phd6}.
Since each interaction term in a pseudo-Boolean function is a pseudo-Boolean function in itself,
Theorem \ref{th:1} is important because it means that we can approximate a pseudo-Boolean function
by approximating each of the interaction terms involved in the function
individually. Also, since the best approximation of a pseudo-Boolean
function is itself, we only need to worry about how to approximate the
interaction terms we want to remove. Theorem \ref{th:2} shows that a sequential scheme for calculating the
approximation is possible.
Next we give two theorems characterising the properties of the error introduced by the approximation.
\begin{theorem}\label{th:3}
Assume again that we have two approximations of $f(x)$, $A_{\tilde{S}}\{f(x)\}$
and $A_{\tilde{\tilde{S}}}\{f(x)\}$, such that $\tilde{\tilde{S}} \subseteq \tilde{S}
\subseteq S$. Letting $\tilde{f}(x) = A_{\tilde{S}}\{f(x)\}$ and
$\tilde{\tilde{f}}(x) = A_{\tilde{\tilde{S}}}\{f(x)\}$, we
then have $\mbox{SSE}(f,\tilde{\tilde{f}}) = \mbox{SSE}(f,\tilde{f}) +
\mbox{SSE}(\tilde{f},\tilde{\tilde{f}})$.
\end{theorem}
\begin{theorem}
\label{th:4}
Given a pseudo-Boolean function $f(x)$ and an approximation $\tilde{f}(x)$
constructed as described, the error sum of squares can be written as,
\begin{equation}
\sum_{x \in \Omega} \left\{f(x) - \tilde{f}(x) \right\}^2 = \sum_{\Lambda \in
S\setminus \tilde{S}}\left[ \beta^{\Lambda} \sum_{x \in \Omega_{\Lambda}}
\{f(x)-\tilde{f}(x)\} \right]
\end{equation}
\end{theorem}
Proofs of Theorems \ref{th:3} and \ref{th:4} are in Appendices \ref{app:proof3} and
\ref{app:proof4}, respectively. Note that Theorem \ref{th:4} tells us that the error can be
expressed as a sum over the $\beta$'s that we remove when constructing our
approximation. Note also the special case where $\tilde{S} = S \setminus \lambda$
for some $\lambda\in S$, i.e we remove only one interaction $\beta^{\lambda}$. Then,
\begin{equation}\label{eq:errorRemoveOne}
\sum_{x \in \Omega} \left\{f(x) - \tilde{f}(x) \right\}^2 = \beta^{\lambda} \left[
\sum_{x \in \Omega_{\lambda}}
\{f(x)-\tilde{f}(x)\} \right].
\end{equation}
With these theorems in hand we can go from $S$ to $\tilde{S}$ by removing all
nodes in $S \setminus \tilde{S}$. Theorems \ref{th:1} and \ref{th:2}
allow us to remove these interactions sequentially one
at a time. We start by removing the interaction (or one of, in the case
of several) $\beta^\lambda,\lambda\in S\setminus \widetilde{S}$
with highest degree and approximate it
by the set containing all $\Lambda \subset \lambda$. In other words,
if the interaction has degree $k=|\lambda|$ we design the $k-1$ order approximation of
that interaction term. \citet{art138} gives us the expression for
this,
\begin{equation}
\tilde{\beta}^{\Lambda} = \left\{
\begin{array}{ll}
\beta^{\Lambda} +
(-1)^{|\lambda|-1-|\Lambda|} \left(\frac{1}{2}\right)^{|\lambda| -
|\Lambda|} \beta^{\lambda} & \quad \text{if~~}\forall \Lambda \subset \lambda, \\
\beta^{\Lambda} & \quad \text{otherwise.} \\
\end{array} \right.
\label{eq:sol}
\end{equation}
We then proceed by removing one interaction at a time until we reach the set of interest,
$\tilde{S}$. The approximation error in one step of this procedure
is given by (\ref{eq:errorRemoveOne}) and the total approximation
error is given as the sum of the errors in each of the approximation steps.
\subsection{\label{sec:pairwise}Second order interaction removal}
In this section we discuss pseudo-Boolean function approximation
for a specific choice of $\tilde{S}$, which is of particular interest
for the variable elimination algorithm. We show how we can construct a new
way of solving the resulting system of equations and term this approximation the
second order interaction removal (SOIR) approximation.
For $i,j\in N$, $i\neq j$ and $\{ i,j\} \in S$, assume we
have $\tilde{S} = S_{\{i,j\}}^c$. In other words we want to remove all interactions
involving both $i$ and $j$ and approximate these by lower order
interactions.
To find this approximation we could of course proceed as in the
previous section, sequentially
removing one interaction at the time until we reach our desired
approximation. However, the following theorem gives explicit
expressions for both the approximation and
the associated error.
\begin{theorem}\label{th:6}
Assume we have a pseudo-Boolean function $f(x)$ represented on
a dense set $S$.
For $i,j\in N$, $i\neq j$ and $\{ i,j\}\in S$, the least squares approximation of $f(x)$
on $\tilde{S} = S_{\{ i,j\}}^c$ is given by
\begin{equation}\label{eq:ftilde}
\tilde{f}(x)=A_{\tilde{S}}\{f(x)\} = \sum_{\Lambda\in\tilde{S}} \tilde{\beta}^\Lambda \prod_{k\in\Lambda}x_k,
\end{equation}
where $\tilde{\beta}^\Lambda$ for $\Lambda\in \tilde{S}$ is given as
\begin{equation}\label{eq:soir}
\tilde{\beta}^\Lambda = \left\{ \begin{array}{ll}
\beta^\Lambda - \frac{1}{4}\beta^{\Lambda\cup\{ i,j\}} & \mbox{~~if $\Lambda \cup \{i,j\} \in S$}, \\
\beta^\Lambda + \frac{1}{2}\beta^{\Lambda\cup\{ i\}} & \mbox{~~if $\Lambda\cup\{ i\}\in S$
and $\Lambda\cup\{ j\}\not\in S$}, \\
\beta^\Lambda + \frac{1}{2}\beta^{\Lambda\cup \{ j\}} & \mbox{~~if $\Lambda\cup\{ i\}\not\in S$
and $\Lambda\cup\{ j\}\in S$}, \\
\beta^\Lambda & \mbox{~~otherwise.}
\end{array}\right.
\end{equation}
The associated approximation error is
\begin{equation}\label{eq:soirerror}
f(x)-\tilde{f}(x) = \left(x_ix_j + \frac{1}{4} - \frac{1}{2}x_i -
\frac{1}{2}x_j\right)\sum_{\Lambda \in
S_{\{i,j\}}}\left(\beta^{\Lambda}\prod_{k \in \Lambda \setminus \{i,j\}}x_k \right).
\end{equation}
\end{theorem}
A proof is given in Appendix \ref{app:proof6}.
Clearly, the approximation solution of this theorem
corresponds to the solution we would get using the sequential scheme discussed above,
but (\ref{eq:soir}) is much faster to calculate and the above theorem has the advantage
of giving us a nice explicit expression for the error.
Note that the absolute value of the parenthesis outside the sum in (\ref{eq:soirerror}) is
always $\frac{1}{4}$ and thus the absolute value of $f(x)-\tilde{f}(x)$ does not depend on $x_i$ or $x_j$.
We thereby get the following expression for the
error sum of squares,
\begin{align}
\mbox{SSE}(f,\tilde{f}) = \sum_{x \in \Omega} \{f(x)-\tilde{f}(x)\}^2 = \frac{1}{4}\sum_{x \in
\Omega_{\{i,j\}}}\left(\sum_{\Lambda \in
S_{\{i,j\}}}\beta^{\Lambda}\prod_{k \in \Lambda \setminus \{i,j\}}x_k\right)^2.
\end{align}
\subsection{\label{sec:maxmin}Upper and lower bounds for pseudo-Boolean functions}
In this section we construct upper and lower bounds for
pseudo-Boolean functions.
We denote the upper and lower bounds
by $f_U(x)$ and $f_L(x)$, respectively, i.e. we require
$f_L(x)\leq f(x)\leq f_U(x)$ for all $x\in\Omega$. Just like for the
SOIR approximation $\tilde{f}(x)$, we require that all interactions
involving both $i$ and $j$ are removed from $f_U(x)$ and
$f_L(x)$. Using the expression for the SOIR approximation error
in (\ref{eq:soirerror}) we have
\begin{equation}
f(x) = \tilde{f}(x) + \left(x_ix_j+\frac{1}{4} -\frac{1}{2}x_i-\frac{1}{2}x_j\right)
\sum_{\Lambda\in S_{\{ i,j\}}}\left( \beta^\Lambda \prod_{k\in\Lambda\setminus\{ i,j\}}x_k\right).
\end{equation}
The terms that are constant or linear in $x_i$ and $x_j$ can be kept unchanged in $f_U(x)$
and $f_L(x)$, whereas we need to find bounds for the interaction part
\begin{equation}\label{eq:g}
g(x) = x_ix_j\sum_{\Lambda\in S_{\{ i,j\}}} \left( \beta^\Lambda\prod_{k\in\Lambda\setminus\{ i,j\}}x_k\right).
\end{equation}
As $x_j\in \{ 0,1\}$, an upper bound for $g(x)$
which is linear in $x_i$ and constant
as a function of $x_j$ is
\begin{equation}\label{eq:upper1}
g(x) \leq x_i\max\left\{ 0,
\sum_{\Lambda\in S_{\{ i,j\}}}\left( \beta^\Lambda \prod_{k\in\Lambda\setminus\{ i,j\}}x_k\right)\right\}.
\end{equation}
An upper bound which is linear $x_j$ and constant as
a function of $x_i$ is correspondingly found by using that $x_i\in \{ 0,1\}$.
Moreover, any convex linear combination of these two bounds
is also a valid upper bound for $g(x)$.
Similar reasoning for a lower bound produces the same type of expressions,
except that the $\max$ operators are replaced by
$\min$ operators.
The bounds defined above are clearly valid upper and lower bounds,
and by construction they have no interactions involving both $i$ and $j$.
However, in the next section we need to find the canonical forms, given by
(\ref{eq:canonical}), of the bounds and then the computational complexity of
constructing these representations is important.
Focusing on the bound in (\ref{eq:upper1}),
the $\max$ function is a function of
$d_{ij}=|\{ \Lambda\in S_{\{ i,j\}} : |\Lambda|=3\}$ variables, so
we need to compute the values of $2^{d_{ij}}$ interaction coefficients.
This is
computationally feasible only if $d_{ij}$ is small enough. If $d_{ij}$
is too large we need to consider computationally cheaper, and coarser,
bounds. To see how this can be done, first
note that for any $r\in N\setminus \{ i,j\}$
and $\{ i,j,r\}\in S_{\{ i,j\}}$ we have
$S_{\{ i,j,r\}}\subset S_{\{ i,j\}}$ and thereby the $g(x)$ defined
in (\ref{eq:g}) can alternatively be expressed as
\begin{equation}\label{galt}
g(x) = x_i x_j x_r \sum_{\Lambda\in S_{\{ i,j,r\}}} \left(
\beta^\Lambda \prod_{k\in \Lambda\setminus\{ i,j,r\}} x_k\right) +
x_i x_j \sum_{\Lambda\in S_{\{ i,j\}} \setminus S_{\{ i,j,r\}}} \left(
\beta^\Lambda \prod_{k\in\Lambda\setminus\{ i,j\}}\right).
\end{equation}
An alternative upper bound can then be defined by following a similar strategy
as above, but for each of the two terms in (\ref{galt}) separately. This gives
the upper bound
\begin{equation}\label{boundalt}
f(x) \leq \sum_{\Lambda\in S_{\{ i,j\}}^c} \beta^\Lambda \prod_{k\in \Lambda} x_k +
x_i\max\left\{ 0,\sum_{\Lambda\in S_{\{ i,j,r\}}}\left(
\beta^\Lambda \prod_{k\in \Lambda\setminus\{ i,j,r\}}x_k\right)\right\}
\end{equation}
\[
+ x_i\max \left\{0,\sum_{\Lambda\in S_{\{ i,j\}}\setminus S_{\{ i,j,r\}}}\left(
\beta^\Lambda\prod_{k\in \Lambda\setminus\{ i,j\}} x_k\right)\right\},
\]
and the corresponding lower bound is again given by the same type of
expression except that the $\max$ operators are
replaced by $\min$ operators. One should note the canonical form of the bound in
(\ref{boundalt}) can be done for each $\max$ term separately. Moreover,
as both $\max$ terms in (\ref{boundalt}) are functions of strictly less than $d_{ij}$ variables
the computational complexity of finding the canonical representation of the bound in
(\ref{boundalt}) is smaller than the complexity of the
corresponding operation for the bound in (\ref{eq:upper1}).
However, for one or both of the $\max$ functions in (\ref{boundalt}), the task of transforming it
into the canonical form may still be too computationally expensive. If so, the process of splitting
a sum into a sum of two sums must be repeated. For example, if the first $\max$ function is
problematic, one needs to locate an $s\in N\setminus\{ i,j,r\}$ so that $\{ i,j,r,s\}\in S_{\{ i,j,r\}}$ and the sum
over $\Lambda\in S_{\{ i,j,r\}}$ can be split into a sum over $\Lambda \in S_{\{ i,j,r,s\}}$ and
a sum over $\Lambda \in S_{\{ i,j,r\}}\setminus S_{\{ i,j,r,s\}}$ and finding bounds as before.
By repeating this process
sufficiently many times one will eventually end up with a bound consisting of a sum of $\max$
terms that can be transformed to the canonical form in a reasonable computation time.
\section{\label{sec:mrf}MRFs and the variable elimination algorithm}
In this section we give a short introduction to binary MRFs. In particular we explain how the
variable elimination algorithm can be applied to this class of models and
point out its computational limitation. For a general
introduction to MRFs see \citet{art3} or \citet{book10} and for more
on the connection between binary MRFs and pseudo-Boolean functions see
\citet{art130}. For more on the variable elimination algorithm and
applications to MRFs see \citet{art100} and \citet{art105}.
\subsection{Binary Markov random fields}
Assume we have a vector of $n$ binary variables $x = \{x_1,\hdots,x_n\} \in \Omega =
\{0,1\}^n$, $N=\{1,\hdots,n\}$. Let $\mathcal{N} =
\{\mathcal{N}_1,\hdots,\mathcal{N}_n\}$ denote a neighbourhood system where
$\mathcal{N}_k$ denotes the set of indices of nodes that are neighbours of node
$k$. As usual we require a symmetrical neighbourhood system, so if $i
\in \mathcal{N}_j$ then $j \in \mathcal{N}_i$, and by convention a node
is not a neighbour of itself. Then $x$ is a binary MRF with respect to a neighbourhood
system $\mathcal{N}$ if $p(x) > 0$ for all $x \in \Omega$ and the full conditionals $p(x_k|x_{-k})$ have
the Markov property,
\begin{equation}
p(x_k|x_{-k}) = p(x_k|x_{\mathcal{N}_k}) \text{ } \forall \text{ }x
\in \Omega \mbox{~and~}k\in N,
\label{fullconditionals}
\end{equation}
where $x_{\mathcal{N}_k} = (x_i:i \in \mathcal{N}_k)$. A clique $\Lambda$ is
a set $\Lambda \subseteq N$ such
that for all pairs $i,j\in \Lambda$ we have $i \in \mathcal{N}_j$.
A clique is a maximal clique if it is not a subset of another clique. The set of all
maximal cliques we denote by $\mathcal{C}$. The Hammersley-Clifford theorem
\citep{art3,col7} tells us
that we can express the distribution of $x$ either through the full
conditionals in \eqref{fullconditionals} or through clique potential functions,
\begin{equation}\label{eq:p}
p(x) = \frac{1}{c} \exp\{U(x)\} = \frac{1}{c} \exp\left\{\sum_{\Lambda \in \mathcal{C}}U_\Lambda(x_\Lambda)\right\},
\end{equation}
where $c$ is a normalising constant, $U_\Lambda(x_\Lambda)$ is a
potential function for a clique $\Lambda$ and $x_\Lambda=(x_i:i\in\Lambda)$. $U(x)$ is commonly referred to as the
energy function. From the previous sections we know that $U(x)$ is a
pseudo-Boolean function and can be expressed as,
\begin{equation}
U(x) = \sum_{\Lambda \subseteq N} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k =\sum_{\Lambda \in S} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k,
\label{eq:ux}
\end{equation}
where $S$ is defined as in \eqref{eq:Sdef}. For a given energy
function $U(x)$, \citet{art130} show how the interactions $\beta^\Lambda$ can be calculated
recursively by evaluating $U(x)$. Moreover, \citet{art130} show that $\beta^\Lambda=0$
whenever $\Lambda$ is not a clique. From this we understand that
it is important that we represent $U(x)$ on $S$ as defined in (\ref{eq:Sdef}), and not
use the full representation.
\subsection{\label{sec:exactVEA}The variable elimination algorithm}
As always the problem when evaluating the likelihood or generating
samples from MRFs is that $c$ is a function of the model parameters and in
general unknown. Calculation involves a sum over $2^n$
terms,
\begin{equation}
c = \sum_{x \in \Omega} \exp\left\{U(x)\right\} = \sum_{x \in \Omega} \exp\left(\sum_{\Lambda
\in S} \beta^{\Lambda} \prod_{k \in \Lambda}x_k\right).
\label{eq:normsum1}
\end{equation}
The variable elimination algorithm \citep{art100,art105}
calculates the sum in \eqref{eq:normsum1} by
taking advantage of the fact that we
can calculate this sum more efficiently by factorising the
un-normalised distribution. We now cover this
recursive procedure.
Clearly we can always split the set $S$ into two parts, $S_{\{i\}}$ and
$S_{\{i\}}^c$ where $i\in N$. Thus we can
split the energy function in \eqref{eq:ux} into a sum of two sums,
\begin{equation}
U(x) = \sum_{\Lambda \in S_{\{i\}}^c} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k + \sum_{\Lambda \in S_{\{i\}}} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k.
\label{uxsplit}
\end{equation}
Note that the first sum contains no interaction terms involving
$x_i$. Letting $x_{-i} = (x_1,\hdots,$ $x_{i-1},x_{i+1},\hdots,x_n)$, we
note that this is essentially equivalent to factorising $p(x) =
p(x_i|x_{-i})$ $p(x_{-i})$, since
\begin{equation}
p(x_i|x_{-i}) \propto \exp\left(\sum_{\Lambda \in S_{\{i\}}} \beta^{\Lambda} \prod_{k \in
\Lambda}x_k \right).
\end{equation}
By summing out $x_i$ from $p(x)$ we get
the distribution of $p(x_{-i})$. Taking advantage of the split in
\eqref{uxsplit} we can write this as,
\begin{equation}\label{eq:marginal}
p(x_{-i}) = \sum_{x_i}p(x) = \frac{1}{c}\exp\left(\sum_{\Lambda \in S_{\{i\}}^c}
\beta^{\Lambda} \prod_{k \in \Lambda}x_k \right) \sum_{x_i} \exp\left(\sum_{\Lambda \in S_{\{i\}}}
\beta^{\Lambda} \prod_{k \in \Lambda}x_k \right).
\end{equation}
The last sum over $x_i$ can be expressed as the exponential of a new
binary polynomial, i.e.
\begin{equation}
\exp\left(\sum_{\Lambda \subseteq \mathcal{N}_i} \check{\beta}^{\Lambda} \prod_{k \in
\Lambda}x_k\right) = \sum_{x_i} \exp\left(\sum_{\Lambda \in S_{\{i\}}}
\beta^{\Lambda} \prod_{k \in \Lambda}x_k \right),
\label{eq:newpbf}
\end{equation}
where the interactions $\check{\beta}^\Lambda$ can be sequentially
calculated by evaluating the sum over $x_i$ in \eqref{eq:newpbf} as
described in \citet{art130}.
Note that this new function is a pseudo-Boolean
function potentially of full degree. The number of non-zero interactions in this
representation could be up to $2^{|\mathcal{N}_i|}$. Summing out $x_i$
leaves us with a new MRF with a new neighbourhood system. This is the first step in a sequential
procedure for calculating the normalising constant $c$. In each step
we sum over one of the remaining variables by splitting the energy
function as above. Repeating this procedure until we have summed out
all the variables naturally yields the normalising constant.
The computational bottleneck for this algorithm occurs when
representing the sum in \eqref{eq:newpbf}. Assume we have summed out
variables $x_{1:i-1} = (x_1,\hdots,x_{i-1})$, have an MRF with a neighbourhood system
$\check{\mathcal{N}} = \{\check{\mathcal{N}}_i,\hdots,\check{\mathcal{N}}_n\}$ and want
to sum out $x_i$. If
$\check{\mathcal{N}}_i$ is too large we run into trouble
with the sum corresponding to
\eqref{eq:newpbf} since this requires us to compute and store up to
$2^{\check{\mathcal{N}}_i}$ interaction terms. In models where
$\check{\mathcal{N}}_i$ increases as we sum out variables the
exponential growth causes us to run into problems very quickly. As a practical example of
this consider the Ising model defined on a lattice. Assuming we sum out variables in the
lexicographical order, the size of the
neighbourhood will grow to the number of rows in our lattice.
This thus restricts the number of rows in the lattice to less or
equal to $20$ for practical purposes.
\section{\label{sec:mrfApprox}Construction of an approximate variable
elimination algorithm}
In this section we include the approximation results of Section
\ref{sec:pbf} in the variable elimination algorithm described in the previous
section to obtain an approximate, but computationally more efficient variant of
the variable elimination algorithm.
To create an algorithm that is computationally viable we must seek to
control $|\check{\mathcal{N}}_i| = \eta_i$ as we sum out variables. If this
neighbourhood becomes too large, we run into problems both with
memory and computation time. Our idea is to construct an
approximate representation of the MRF before summing out each
variable. The approximation is chosen so that
$\eta_i \leq \nu$, where $\nu$ is an input to our
algorithm. Given a design for the approximation we then want to
minimise the error sum of squares of our energy function.
Assume we have an MRF and have (approximately) summed out variables $x_{1:i-1}$, so we
currently have an MRF with a neighbourhood structure
$\check{\mathcal{N}}$ and energy function $\check{U}(x_{i:n}) = \sum_{\Lambda \in \check{S}}
\check{\beta}^{\Lambda}\prod_{k \in \Lambda}x_k $, so,
\begin{equation}
c = \sum_{x_{i:n}} \exp\left\{\check{U}(x_{i:n}) \right\}.
\end{equation}
If $\eta_i$ is too large we run into problems when summing over
$x_i$. Our strategy for overcoming this problem is first to create an
approximation of the energy function
$\check{U}(x_{i:n})$,
\begin{equation}
\check{U}(x_{i:n}) = \sum_{\Lambda \in \check{S}}
\check{\beta}^{\Lambda}\prod_{k \in \Lambda}x_k \approx
\tilde{U}(x_{i:n}) = \sum_{\Lambda \in \tilde{S}}
\tilde{\beta}^{\Lambda}\prod_{k \in \Lambda}x_k.
\end{equation}
We control the size of $\eta_i$, by designing our approximation set
$\tilde{S}$ and thus the new approximate neighbourhood $\tilde{\mathcal{N}}$ in such a
way that $|\tilde{\mathcal{N}}_i| = \tilde{\eta}_i \leq
\nu$. Assuming we can do this, we could construct an approximate
variable elimination algorithm where we check the size of the neighbourhood
$\eta_i$ before summing out each variable. If this is greater than
the given $\nu$ we approximate the energy function
before summation. This leaves two questions; how do we choose the set
$\tilde{S}$ and how do we define the approximation?
The two questions are obviously linked, however we start by looking
more closely at how we may choose the set $\tilde{S}$. Our
tactic is to reduce $\eta_i$ by one at the time.
To do this we need to design $\tilde{S}$ in such
a way that $i$ and some node $j$ are no longer neighbours.
Doing this is equivalent to requiring all interactions
$\tilde{\beta}^{\Lambda}$, involving both $i$ and $j$ to be zero. As before
we denote the subset of all interactions involving $i$ and $j$ as
$\check{S}_{\{i,j\}} \subseteq \check{S}$ and construct our
approximation set as in Section \ref{sec:pairwise}, defining $\tilde{S} = \check{S}\setminus
\check{S}_{\{i,j\}}$. Our approximation is defined by the
equations corresponding to \eqref{eq:equations} and using the results
from Section \ref{sec:pairwise}, the solution is easily available.
We can then imagine a scheme where we reduce
$\eta_i$ one at a time until we reach our desired size $\nu$. This
leaves the question of how to choose $j$. One could calculate the
$\mbox{SSE}$ for all possibilities of $j$ and choose the value of $j$
that has the minimum $\mbox{SSE}$. However, this may be
computationally very expensive and would in many cases dominate the
total computation time of our algorithm. Instead we propose to compute
an approximate upper bound for $|f(x)-\widetilde{f}(x)|$ for all
values of $j$ and to select the value of $j$ that minimises this
approximate bound. We define the approximate bound by first defining
a modified version $f^\star(x)$ of $f(x)$, where we set all
first, second and third order interactions for $f^\star(x)$
equal to corresponding quantities for $f(x)$, and set all
fourth and higher order interactions for $f^\star(x)$ equal to zero.
We then define the approximate upper bound as the exact
upper bound for $|f^\star(x)-\widetilde{f}^\star(x)|$. As $f^\star(x)$
has no non-zero fourth or higher order interactions upper and lower
bounds for $f^\star(x)-\widetilde{f}^\star(x)$ can readily be computed
as discussed in Section \ref{sec:maxmin}.
One should note that Theorem \ref{th:2} means that after reducing $\eta_i$ by $\eta_i-\nu$
our approximation is still optimal for the given selection of
$j$'s. However, there is no guarantee that our selection of $j$'s is
optimal, and it is possible that we could have obtained a better set of
$j$'s by looking at the error from
reducing $\eta_i$ by more than one at the time.
Using this approximate variable elimination algorithm
we can define a corresponding approximate model through a product of the approximate
conditional distributions,
\begin{equation}\label{eq:pommapprox}
\tilde{p}(x) = \tilde{p}(x_1|x_{2:n}) \hdots \tilde{p}(x_{n-1}|x_{n})\tilde{p}(x_n),
\end{equation}
which is a POMM \citep{art119}. One of the aspects we wish
to investigate in the results section is to what extent this distribution can
mimic some of the attributes of the original MRF. Clearly, to sample from
$\tilde{p}(x)$ is easy via a backward pass, first simulating $x_n$ from
$\tilde{p}(x_n)$, thereafter simulating $x_{n-1}$ from $\tilde{p}(x_{n-1}|x_n)$
and so on. One should note that two versions (\ref{eq:pommapprox}) can be defined.
The first version is obtained by taking $\widetilde{p}(x_i|x_{i+1:n})$ to be the
resulting conditional distribution after we have (approximately) summed out $x_{1:i-1}$.
In $\widetilde{p}(x_i|x_{i+1:n})$ we then have no guarantee for how many of the
elements in $x_{i+1:n}$ the variable $x_i$ really depends on, and it may be
computationally expensive to compute the normalising constant of this
conditional distribution for all values of $x_{i+1:n}$.
The alternative is to let $\widetilde{p}(x_i|x_{i+1:n})$ be the
resulting conditional distribution after one has both (approximately) summed
out $x_{1:i-1}$ and done the necessary approximations so that $x_i$ is linked with
at most $\nu$ of the elements in $x_{i+1:n}$. Then we know that $x_i$ is linked to at most
$\nu$ of the variables in $x_{i+1:n}$ in the conditional distribution also, and we have
an upper limit for the computational complexity of computing the
normalising constant in the conditional distributions. Which of the two
versions of $\widetilde{p}(x)$ one should use depends on what one intends to use
$\widetilde{p}(x)$ for. One would expect the first version to be the
best approximation of $p(x)$, but for some applications it may be computationally
infeasible. We discuss this issue further in the examples in Section
\ref{sec:examples}.
\section{\label{sec:marg}Bounds and alternative marginalisation operations}
In this section we consider some variations of the approximate algorithm defined above.
We fist discuss how the results in Section \ref{sec:maxmin} can be used
to modify the procedure to get upper and lower bounds for the normalising constant.
Thereafter we
consider how approximations (or bounds) for some other quantities can be found by
replacing the summation operation in the above algorithm
with alternative marginalisation operations.
\subsection{\label{sec:bound}Bounds for the normalising constant}
The approximate value for the normalising constant, $c$, found by the algorithm
in Section \ref{sec:mrfApprox} comes without any measure of precision.
Using the results of Section \ref{sec:maxmin} we can modify the algorithm described
above and instead find upper or lower bounds for $c$.
Our point of origin for finding a bound $c_L$ (or $c_U$) such that $c_L\leq c$
(or $c_U \geq c$) is the approximate variable elimination algorithm described
in the previous section. An iteration of this algorithm consists of two steps.
First the energy function is replaced by an approximate
energy function and, second, we sum over the chosen
variable. To construct an upper or lower bound we simply change the first step.
Instead of replacing the energy function by an approximation we
replace it with a lower (or upper) bound. To define such a bound we
adopt the strategy discussed in Section \ref{sec:maxmin}.
Letting $x_i$ denote the next variable to sum over, we first have to
decide which second order interaction to remove, i.e. the value of $j$
in Section \ref{sec:maxmin}. For this we follow
the same strategy as for the approximate variable elimination algorithm
discussed above. Then we use the bound in (\ref{eq:upper1})
whenever $d_{ij}\leq \nu$, where $d_{ij}$ is as defined in
Section \ref{sec:pairwise} and $\nu$ is the same input parameter to the
algorithm as in Section \ref{sec:mrfApprox}. If $d_{ij} > \nu$ we use a coarser
bound as discussed in Section \ref{sec:mrfApprox}. In the definition of these
coarser bounds it remains to specify how to choose the
value of $r$ in (\ref{boundalt}), and if necessary also the value of
$s$ and so forth. For definiteness we here describe how we
select the value of $r$, but follow the same strategy for all such choices.
For the choice of $r$ we adopt a similar strategy as for $j$ in Section
\ref{sec:mrfApprox}, but now include all interactions up to order
four in the definition of $f^\star(x)$. When selecting a value
for $s$ we include in $f^\star(x)$ interactions up to order five and so fourth.
\subsection{Alternative marginalisation operations}
The exact variable elimination algorithm finds
the normalising constant of an MRF by
summing over each variable in turn. As also discussed in \citet{book34} for the
junction tree algorithm, other quantities of interest can be
found by replacing the summation operation by alternative marginalisation
operations. Two quantities of particular interest in our setting is to find
the state $x$ which maximises $U(x)$, and to compute moments of $x$. In the
following we first consider the maximisation problem and thereafter
the computation of moments
\subsubsection{\label{sec:maximisation}Maximisation}
By replacing the summation over $x_i$ in the exact variable elimination
algorithm with a maximisation over $x_i$, the algorithm returns
the maximal value of $\exp\{ U(x)\}$ over $x\in\Omega$. By a following
backward scan it is also possible to find the value of $x$ which
maximises $\exp\{ U(x)\}$. The forward and
backward passes are together known as the Viterbi
(1967)\nocite{art140} algorithm. Just as summation over $x_i$ becomes
computationally infeasible if the number of neighbours to
node $i$ is too large, the maximisation over $x_i$ is also
computationally infeasible in this situation.
To see how to construct an approximate Viterbi algorithm,
recall that our approximate variable elimination algorithm consists
of two steps in each iteration. First to approximate the energy function and
then to sum over a variable. To get an approximate Viterbi
algorithm we simply replace the second step, so instead of summing over a
variable we take the maximum over that variable. Note that a lower or
upper bound for $\exp\{ U(x)\}$ can be found by replacing the
approximation with a lower or upper bound as discussed
in Section \ref{sec:bound}.
\subsubsection{Moments}
Consider the problem of computing a moment $\mbox{E}\{\psi(x)\}$,
where $x$ is distributed
according to an MRF $p(x)$, and $\psi(x)$ is a given function of $x\in\Omega$.
Inserting the expression for $p(x)$ in (\ref{eq:p}) we get
\begin{equation}\label{eq:moment}
\mbox{E}\{\psi(x)\} = \frac{1}{c}\sum_{x\in\Omega} \exp\{U(x)+\ln\psi(x)\}.
\end{equation}
Thereby $\mbox{E}\{\psi(x)\}$ can be found by running the exact variable
elimination algorithm twice, first as described in Section
\ref{sec:exactVEA} to find $c$ and thereafter with the energy function
redefined as $U(x)+\ln\psi(x)$ to find the sum in (\ref{eq:moment}).
In general the computational complexity of the second run of the
variable elimination algorithm is much higher than for the
first run, but if the pseudo-Boolean function $\psi(x)$ can be
represented on the same set $S$ as the energy function $U(x)$ both
runs are of the same complexity.
We obtain an approximation to $\mbox{E}\{\psi(x)\}$ simply by adopting the
approximate variable elimination algorithm instead of the exact one.
To obtain a lower (upper) bound for $\mbox{E}\{\psi(x)\}$ we can divide
a lower (upper) bound for the sum in (\ref{eq:moment}) with an upper
(lower) bound for $c$.
\section{\label{sec:implementation}Some implementational issues}
When implementing the exact and approximate algorithms discussed above we
need to use one (or more) data structure(s) for storing our representation
of pseudo-Boolean functions. The operations we need to perform on
pseudo-Boolean functions fall into two categories. The first
is to compute and store all interaction parameters of a given
pseudo-Boolean function without any (known) Markov structure. The
computation of the interaction parameters $\check{\beta}^\Lambda$ defined
by (\ref{eq:newpbf}) is of this form, and so is the corresponding operation
for the $\max$ terms in (\ref{eq:upper1}) and (\ref{boundalt}). When doing this type
of operations we are simply numbering all the interaction parameters
in some order and storing their values in a vector. The values are then
fast to assess and the necessary computations can be done efficiently.
The second type of operation we need to do on a pseudo-Boolean function
consists of operations on functions defined as in (\ref{eq:canonical}),
functions for which a lot of the interaction parameters are zero. The
approximation operation defined in Theorem \ref{th:6} is of this type.
We then need to adopt a data structure which stores an
interaction parameter $\beta^\Lambda$ only if $\Lambda\in S$. We use a
directed acyclic graph (DAG) for this, where we have one node for
each $\Lambda\in S$ and a node $\lambda\in S$ is a child of
another node $\Lambda\in S$ if and only if
$\lambda = \Lambda \cup \{ i\}$ for some $i\in N\subset\Lambda$.
An illustration for $N=\{ 1,2,3,4\}$ and $S=\{ \emptyset,\{ 1\},\{ 2\},\{ 3\},
\{ 4\},\{ 1,2\},\{ 1,3\},\{ 2,3\},\{ 1,2,3\}\}$ is shown in Figure
\ref{fig:dag}.
\begin{figure}
\begin{center}
\vspace*{-0.9cm}
\begin{picture}(400,200)(-200,0)
\put(0,0){\circle{23}}
\put(0,0){\makebox(0,0){$\emptyset$}}
\put(-9.57,6.38){\vector(-3,2){55.86}}
\put(9.57,6.38){\vector(3,2){55.86}}
\put(-5.14,10.29){\vector(-1,2){14.71}}
\put(5.14,10.29){\vector(1,2){14.71}}
\put(-75,50){\circle{23}}
\put(-75,50){\makebox(0,0){$1$}}
\put(-75,61.5){\vector(0,1){27}}
\put(-66.87,58.13){\vector(1,1){33.74}}
\put(-25,50){\circle{23}}
\put(-25,50){\makebox(0,0){$2$}}
\put(-33.13,58.13){\vector(-1,1){33.74}}
\put(-16.87,58.13){\vector(1,1){33.74}}
\put(-14.71,55.14){\vector(2,1){79.43}}
\put(25,50){\circle{23}}
\put(25,50){\makebox(0,0){$3$}}
\put(16.87,58.13){\vector(-1,1){33.74}}
\put(25,61.5){\vector(0,1){27}}
\put(75,50){\circle{23}}
\put(75,50){\makebox(0,0){$4$}}
\put(75,61.5){\vector(0,1){27}}
\put(-75,100){\circle{23}}
\put(-75,100){\makebox(0,0){$12$}}
\put(-66.86,108.13){\vector(1,1){33.74}}
\put(-25,100){\circle{23}}
\put(-25,100){\makebox(0,0){$13$}}
\put(-25,111.5){\vector(0,1){27}}
\put(25,100){\circle{23}}
\put(25,100){\makebox(0,0){$23$}}
\put(16.87,108.13){\vector(-1,1){33.74}}
\put(75,100){\circle{23}}
\put(75,100){\makebox(0,0){$24$}}
\put(-25,150){\circle{23}}
\put(-25,150){\makebox(0,0){$123$}}
\end{picture}
\end{center}
\caption{\label{fig:dag}The directed acyclic graph used to represent a
pseudo-Boolean function represented on $S=\{ \emptyset,\{ 1\},\{ 2\},\{ 3\},
\{ 4\},\{ 1,2\},\{ 1,3\},\{ 2,3\},\{ 1,2,3\}\}$.}
\end{figure}
The value of $\beta^\Lambda$ is stored in the
node $\Lambda$, and the arrows in the figure are represented as pointers.
To assess the value of an interaction parameter $\beta^\Lambda$ in
this data structure we need to follow the pointers from the root
to node $\Lambda$. This is clearly less efficient than in the
vector representation discussed above, but this is the cost one has to
pay to reduce the memory requirements. One should also
note that such a DAG representation is very convenient when
computing the approximation defined by Theorem \ref{th:6}. What
one needs to do is first to clip out the subgraph which
corresponds to $S_{\{ i,j\}}$. Thereafter one should traverse that
subgraph and for each node in the subgraph add the required quantity
to three interaction parameters in the remaining DAG, as specified by
(\ref{eq:soir}).
\section{\label{sec:examples}Simulation and data examples}
In this section we first present the results of a number of simulation
exercises to evaluate the quality of the approximations and bounds.
Thereafter we present some simulation examples to demonstrate possible
applications of the approximations and bounds we have introduced.
Finally, we use our approximation in the evaluation of a data set of
cancer mortality from the United States.
In all the examples we adopt the approximation and the bounds
defined in Sections \ref{sec:mrfApprox} and \ref{sec:marg}.
\subsection{\label{sec:models}Models}
In the simulation examples we consider two classes of MRFs.
The first class we consider is the Ising model \citep{art17}.
The energy function can then be expressed as
\begin{equation}\label{eq:Ising}
U(x) = \theta\sum_{i\sim j} I(x_i=x_j),
\end{equation}
where the sum is over all first order neighbourhood pairs, $\theta$
is a model parameter, and $I(x_i=x_j)$ is the indicator function
and takes value $1$ if $x_i=x_j$ and $0$ otherwise. We present results
for $\theta=0.4$, $0.6$, $0.8$ and $-\ln (\sqrt{2}-1)$, where the last
value is the critical value in the infinite lattice case.
Representing
$U(x)$ as a binary polynomial is done by recursively calculating
the first and second order interactions, for details see
\citet{art130}.
In the second class of MRFs we assume a third order neighbourhood.
Then each node sufficiently far away from the borders has 12 neighbours
and there exists two types of maximal cliques, see Figure
\ref{fig:maxCliques}.
\begin{figure}
\begin{center}
\begin{tabular}{ccc}
\begin{picture}(50,50)(0,-10)
\put(0,-5){\line(0,1){10}}
\put(10,-15){\line(0,1){30}}
\put(20,-25){\line(0,1){50}}
\put(30,-25){\line(0,1){50}}
\put(40,-15){\line(0,1){30}}
\put(50,-5){\line(0,1){10}}
\put(20,-25){\line(1,0){10}}
\put(10,-15){\line(1,0){30}}
\put(0,-5){\line(1,0){50}}
\put(0,5){\line(1,0){50}}
\put(10,15){\line(1,0){30}}
\put(20,25){\line(1,0){10}}
\put(20,-4.5){\line(1,0){10}}
\put(20,-4){\line(1,0){10}}
\put(20,-3.5){\line(1,0){10}}
\put(20,-3){\line(1,0){10}}
\put(20,-2.5){\line(1,0){10}}
\put(20,-2){\line(1,0){10}}
\put(20,-1.5){\line(1,0){10}}
\put(20,-1){\line(1,0){10}}
\put(20,-0.5){\line(1,0){10}}
\put(20,0){\line(1,0){10}}
\put(20,0.5){\line(1,0){10}}
\put(20,1){\line(1,0){10}}
\put(20,1.5){\line(1,0){10}}
\put(20,2){\line(1,0){10}}
\put(20,2.5){\line(1,0){10}}
\put(20,3){\line(1,0){10}}
\put(20,3.5){\line(1,0){10}}
\put(20,4){\line(1,0){10}}
\put(20,4.5){\line(1,0){10}}
\end{picture} &
~~~~~~~~ &
\begin{tabular}{cc}
\begin{picture}(20,50)(0,-20)
\put(0,0){\line(1,0){20}}
\put(0,10){\line(1,0){20}}
\put(0,20){\line(1,0){20}}
\put(0,0){\line(0,1){20}}
\put(10,0){\line(0,1){20}}
\put(20,0){\line(0,1){20}}
\end{picture} &
\begin{picture}(30,50)(0,-20)
\put(10,-5){\line(0,1){30}}
\put(20,-5){\line(0,1){30}}
\put(0,5){\line(1,0){30}}
\put(0,15){\line(1,0){30}}
\put(0,5){\line(0,1){10}}
\put(30,5){\line(0,1){10}}
\put(10,-5){\line(1,0){10}}
\put(10,25){\line(1,0){10}}
\end{picture}
\end{tabular} \\
(a) & & (b)
\end{tabular}
\end{center}
\caption{\label{fig:maxCliques}(a) The third-order
neighbourhood structure used in the higher-order interaction MRF. The white nodes
are neighbour to the black node. (b) The corresponding two types types of maximal cliques.}
\end{figure}
In contrast to the Ising model this model includes interactions of higher order
than two, and we denote it the higher-order interaction MRF.
For the two types of maximal cliques we adopt potential functions that are
invariant under rotation, reflection and when interchanging $0$ and $1$.
With these restrictions the potential functions for the $2\times 2$ and five node
cliques can take four and six values, respectively. We define two models of this
type and the potentials for the various clique configurations are given in
Figure \ref{fig:pot}.
\begin{figure}
\begin{center}
\begin{tabular}{c|cccccccccc}
Configuration & \begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){10}}
\put(0,6.0){\line(1,0){10}}
\put(0,6.5){\line(1,0){10}}
\put(0,7.0){\line(1,0){10}}
\put(0,7.5){\line(1,0){10}}
\put(0,8.0){\line(1,0){10}}
\put(0,8.5){\line(1,0){10}}
\put(0,9.0){\line(1,0){10}}
\put(0,9.5){\line(1,0){10}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\put(5,0.5){\line(1,0){5}}
\put(5,1.0){\line(1,0){5}}
\put(5,1.5){\line(1,0){5}}
\put(5,2.0){\line(1,0){5}}
\put(5,2.5){\line(1,0){5}}
\put(5,3.0){\line(1,0){5}}
\put(5,3.5){\line(1,0){5}}
\put(5,4.0){\line(1,0){5}}
\put(5,4.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\put(5,3){\line(1,0){5}}
\put(5,3.5){\line(1,0){5}}
\put(5,4){\line(1,0){5}}
\put(5,4.5){\line(1,0){5}}
\put(5,5){\line(1,0){5}}
\put(5,5.5){\line(1,0){5}}
\put(5,6){\line(1,0){5}}
\put(5,6.5){\line(1,0){5}}
\put(5,7){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\put(5,8){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(5,10){\line(1,0){5}}
\put(5,10.5){\line(1,0){5}}
\put(5,11){\line(1,0){5}}
\put(5,11.5){\line(1,0){5}}
\put(5,12){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\put(5,8){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(5,10){\line(1,0){5}}
\put(5,10.5){\line(1,0){5}}
\put(5,11){\line(1,0){5}}
\put(5,11.5){\line(1,0){5}}
\put(5,12){\line(1,0){5}}
\put(10,3){\line(1,0){5}}
\put(10,3.5){\line(1,0){5}}
\put(10,4){\line(1,0){5}}
\put(10,4.5){\line(1,0){5}}
\put(10,5){\line(1,0){5}}
\put(10,5.5){\line(1,0){5}}
\put(10,6){\line(1,0){5}}
\put(10,6.5){\line(1,0){5}}
\put(10,7){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\put(5,3){\line(1,0){5}}
\put(5,3.5){\line(1,0){5}}
\put(5,4){\line(1,0){5}}
\put(5,4.5){\line(1,0){5}}
\put(5,5){\line(1,0){5}}
\put(5,5.5){\line(1,0){5}}
\put(5,6){\line(1,0){5}}
\put(5,6.5){\line(1,0){5}}
\put(5,7){\line(1,0){5}}
\put(5,8){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(5,10){\line(1,0){5}}
\put(5,10.5){\line(1,0){5}}
\put(5,11){\line(1,0){5}}
\put(5,11.5){\line(1,0){5}}
\put(5,12){\line(1,0){5}}
\end{picture} &
\begin{picture}(15,15)(0,-3)
\put(5,-2.5){\line(0,1){15}}
\put(10,-2.5){\line(0,1){15}}
\put(0,2.5){\line(1,0){15}}
\put(0,7.5){\line(1,0){15}}
\put(0,2.5){\line(0,1){5}}
\put(15,2.5){\line(0,1){5}}
\put(5,-2.5){\line(1,0){5}}
\put(5,12.5){\line(1,0){5}}
\put(5,8){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(5,10){\line(1,0){5}}
\put(5,10.5){\line(1,0){5}}
\put(5,11){\line(1,0){5}}
\put(5,11.5){\line(1,0){5}}
\put(5,12){\line(1,0){5}}
\put(5,-2){\line(1,0){5}}
\put(5,-1.5){\line(1,0){5}}
\put(5,-1){\line(1,0){5}}
\put(5,-0.5){\line(1,0){5}}
\put(5,0){\line(1,0){5}}
\put(5,0.5){\line(1,0){5}}
\put(5,1){\line(1,0){5}}
\put(5,1.5){\line(1,0){5}}
\put(5,2){\line(1,0){5}}
\end{picture} \\ \hline
Model 1 & $0.5$ & $0.0$ & $0.0$ & $-1.0$ & $0.0$ & $-1.5$ & $0.0$ & $0.0$ & $-0.5$ & $-0.5$ \\
Model 2 & $0.75$ & $0.0$ & $0.0$ & $-1.5$ & $0.0$ & $-2.0$ & $0.0$ & $0.0$ & $-1.0$ & $-1.0$
\end{tabular}
\end{center}
\caption{\label{fig:pot}Potential values for the various clique configurations in the higher-order
MRF models. The potentials are invariant under rotation, reflection and inversion of the colours.}
\end{figure}
A realization from each of the models, generated by
Gibbs sampling, is shown in Figure \ref{fig:homrf-real}.
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{homrf2-real} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{homrf3-real} \\
Model 1 & Model 2 \\[-0.5cm]
\end{tabular}
\end{center}
\caption{\label{fig:homrf-real}Realizations from the two higher-order MRF models on a $100\times 100$ lattice,
generated
by Gibbs sampling. Potential functions for the models are defined in Figure
\ref{fig:pot}.}
\end{figure}
\subsection{\label{sec:evaluation}Empirical evaluation}
In this section we first consider the quality of the approximation of
the normalising constant $c$ and the corresponding bounds.
Thereafter we evaluate to what extent
our $\widetilde{p}(x)$ defined in Section \ref{sec:mrfApprox} can be
used as an approximation to the corresponding $p(x)$.
We first compute the approximate normalising constant
$\widetilde{c}$ and the corresponding upper and lower bounds
$c_U$ and $c_L$ for $\nu=1,2,\ldots,18$ for each of our four
$\theta$ values. The results are presented in Figure \ref{fig:IsingC}.
\begin{figure}
\begin{center}
\begin{tabular}{cccc}
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApprox4Zoom} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApprox6Zoom} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApprox8Zoom} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApproxcZoom} \\
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApproxLength4} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApproxLength6} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApproxLength8} &
\includegraphics[height=3.0cm,width=2.5cm,angle=-90]{cApproxLengthc} \\
$\theta=0.4$ & $\theta=0.6$ & $\theta=0.8$ & $\theta=-\ln(\sqrt{2}-1)$\\[-0.4cm]
\end{tabular}
\end{center}
\caption{\label{fig:IsingC}Results for the Ising model: The upper row shows
the approximation $\ln (\widetilde{c})$ (solid) and the corresponding
upper and lower bounds $\ln (c_U)$ and $\ln(c_L)$ (dashed) for $\nu=8,\ldots,18$,
and the lower row shows $\ln (c_U)-\ln (c_L)$ for $\nu=1$ to $18$.}
\end{figure}
The upper row shows the approximation $\ln (\widetilde{c})$ together
with the bounds $\ln (c_L)$ and $\ln (c_U)$ for $\nu=8$ to $19$,
whereas the lower row shows $\ln (c_U)-\ln (c_L)$ for $\nu=1$ to $19$.
For a given value of $\nu$, computation of an approximation takes about the
same time for all values of $\theta$, and computation of a bound takes about the
same time as evaluating the corresponding approximation. The computation time
for one evaluation as a function of $\nu$ is shown in Figure
\ref{fig:cpuTime}(a).
\begin{figure}[t]
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxTime} &
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxTimeh} \\
(a) Ising & (b) higher-order MRF\\[-0.4cm]
\end{tabular}
\end{center}
\caption{\label{fig:cpuTime}The natural logarithm of the computation times (in seconds)
used to compute an approximation or a bound in (a) the Ising model, and (b) the
higher-order interaction MRF.}
\end{figure}
Figure \ref{fig:homrfC} shows similar results for the higher-order
MRF models, again for a $100\times 100$ lattice, and
corresponding computation times are shown in Figure \ref{fig:cpuTime}(b).
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxM2Zoom} &
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxM3Zoom} \\
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxLengthM2} &
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{cApproxLengthM3} \\
Model 1 & Model 2\\[-0.4cm]
\end{tabular}
\end{center}
\caption{\label{fig:homrfC}Results for the higher-order MRF model: The upper row shows
the approximation $\ln (\widetilde{c})$ (solid) and the corresponding
upper and lower bounds $\ln (c_U)$ and $\ln(c_L)$ (dashed) for $\nu=6,\ldots,18$,
and the lower row shows $\ln (c_U)-\ln (c_L)$ for $\nu=1$ to $18$.}
\end{figure}
Not surprisingly, we see that the quality of the approximation and bounds are best
for models with weak interactions.
Next we evaluate the quality of the POMM approximation $\widetilde{p}(x)$ given in
(\ref{eq:pommapprox}), still on a $100\times 100$ lattice.
To do this we consider an independent proposal
Metropolis--Hastings algorithm
where the MRF $p(x)$ is the target distribution and $\widetilde{p}(x)$ is used as
proposal distribution. We use the acceptance rate in such an algorithm to measure
the quality of the approximation. It should be emphasised that we do not
propose this Metropolis--Hastings as a way to sample from $p(x)$, we just
use the acceptance rate of this algorithm to measure the quality of our approximation.
It should be noted that in this evaluation test
we do not need to compute the normalising constant of the conditional distributions
in (\ref{eq:pommapprox}) for all values of the conditioning variables, so we apply
the first (and best) POMM approximation variant discussed in the paragraph following
(\ref{eq:pommapprox}).
To estimate the quantity we generate
$1000$ independent samples from $p(x)$ and corresponding $1000$ independent
samples from $\widetilde{p}(x)$, compute the Metropolis--Hastings acceptance
probability for each pair and use the average of these numbers as our estimate.
For the Ising model we generate perfect samples from $p(x)$ on a
$100\times 100$ lattice by first
sampling perfectly from the dual random cluster model by coupling from the
past \citep{art27}. For the higher-order MRF model we are not able to generate
perfect samples, so we instead run a long Gibbs sampling algorithm,
and obtain (essentially) independent realizations by sub-sampling this chain. The results for
the Ising and higher-order MRF models are given in Figure \ref{fig:MHaccept}.
\begin{figure}
\begin{center}
\begin{tabular}{ccc}
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{MHAcceptIsing} &
~~~~~~&
\includegraphics[height=4.5cm,width=3.0cm,angle=-90]{MHAcceptHigher} \\
(a) Ising & & (b) higher-order MRF\\[-0.4cm]
\end{tabular}
\end{center}
\caption{\label{fig:MHaccept}Estimated acceptance rates, for $\nu=1,\ldots,18$,
for an independent proposal
Metropolis--Hastings algorithm with target distribution given by the MRF $p(x)$
on a $100\times 100$ lattice
and proposal distribution given by the corresponding POMM approximation
$\widetilde{p}(x)$ in (\ref{eq:pommapprox}). (a) Results for the
Ising model, from top to bottom the curves are for $\theta=0.4$, $\theta=0.6$, $\theta=0.8$ and
$\theta=-\ln(\sqrt{2}-1)$. (b) Results for the higher-order MRFs, the upper and
lower curves are for Model 1 and Model 2, respectively, defined in Figures
\ref{fig:maxCliques} and \ref{fig:pot}.}
\end{figure}
For $\theta=0.4$ and $0.6$ in the Ising model we see that we get very good approximations
even for quite small values of $\nu$, For $\theta=0.8$ a large values for $\nu$ is
needed to get high acceptance rate. For $\theta=-\ln(\sqrt{2}-1)$ the acceptance
rate ends at $30\%$ for $\nu=18$, and remembering that this is for a block update
of $100\times 100$ variables we think this is quite impressive. For the higher-order
MRFs, the acceptance rate for Model 1 becomes very high for the higher values of
$\nu$, whereas the results for Model 2 resemble the results
for $\theta = -\ln(\sqrt{2}-1)$ in the Ising model.
\subsection{\label{sec:applications}Some possible applications}
In this section we present some simulation examples that demonstrate some
possible applications of the proposed approximations and bounds. All
the examples are for MRFs defined on a rectangular lattice, but similar
applications for MRFs defined on graphs are of course also possible.
\subsubsection{Maximum likelihood estimation}
Assume we have observed an image $x$ which we suppose is a realization from
an MRF $p(x|\theta)$, where $\theta$ is a scalar parameter. The $p(x|\theta)$
may for example be the Ising model. If we want to estimate $\theta$ the maximum
likelihood estimator (MLE) is a natural alternative. As discussed in the
introduction the computation of the MLE is complicated by the intractable
normalising constant of the MRF. Figure \ref{mle-0.6}
\begin{figure}[t]
\begin{center}
\begin{tabular}{ccc}
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-2.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-4.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-6.eps} \\[-0.1cm]
$\nu=2$ & $\nu=4$ & $\nu=6$ \\[-0.3cm]
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-8.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-10.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-12.eps} \\[-0.1cm]
$\nu=8$ & $\nu=10$ & $\nu=12$ \\[-0.3cm]
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-14.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-16.eps} &
\includegraphics[height=4.0cm,width=2.2cm,angle=-90]{mle-0.6-18.eps} \\[-0.1cm]
$\nu=14$ & $\nu=16$ & $\nu=18$\\[-0.2cm]
\end{tabular}
\end{center}
\caption{\label{mle-0.6}Maximum likelihood example: The identification of an
interval in which the MLE for $\theta$ in an Ising model must be located. The
estimation is based on a sample from the Ising model with
$\theta=0.6$. The plots shows computed upper and lower bounds for
the log-likelihood function for different values of $\nu$. The horizontal dotted
line is the maximum of the computed lower bounds and the two vertical
dotted lines defines the interval in which the maximum likelihood estimate
must be located.}
\end{figure}
illustrates how upper
and lower bounds for the normalising constant can be used to identify
an interval in which the MLE for $\theta$ must be located. To produce the curves
in the figure we first simulated an $x$
from the Ising model with parameter
$\theta=0.6$, on an $100\times 100$ rectangular lattice.
Assume we want to find the MLE of $\theta$ based on this $x$.
We first computed upper and lower bounds for the log-likelihood function by
replacing the intractable normalising constant with corresponding upper and
lower bounds with $\nu=2$ for $11$ values of $\theta$ on a mesh from $0$ to $2$. The
maximum log-likelihood value must clearly be higher than the maximum of the lower
bound values. Assuming
the log-likelihood function to be concave we can then identify an interval in
which the MLE of $\theta$ must lie. Defining a new mesh of $11$ $\theta$ values over this
interval we repeated the process for $\nu=4$ and obtain an improved interval, see the
upper middle plot in Figure \ref{mle-0.6}.
We then repeated this process for $\nu=6,8,\ldots,18$, and the final interval for the MLE
shown in the lower right plot in Figure \ref{mle-0.6} was
$(0.6055,0.6123)$. It is important not to confuse this interval with
confidence or credibility intervals, the interval computed here is just
an interval which, with certainty, includes the MLE of $\theta$.
It should be noted that computations of the
$2\cdot 11=22$ bounds for the normalising constants can be done in
parallel, and for larger values of $\nu$ this is essential
for the strategy to be practical.
In corresponding simulation experiments with true
$\theta$ values equal to $0.4$, $0.8$ and $-\ln(\sqrt{2}-1)$ we obtained
the final intervals $(0.4113,0.4115)$, $(0.7716,0.8490)$ and
$(0.8234,0.9250)$, respectively. Our bounds are less tight for higher values of $\theta$
and naturally this gives longer intervals for the MLEs when the true $\theta$ value
is larger.
\subsubsection{\label{sec:mpe}Maximum posterior estimation}
The approximate Viterbi algorithm discussed in Section \ref{sec:maximisation}
can be used in image analysis applications. Suppose the constructed scene in Figure
\ref{fig:mpe}(a), which is in an $89\times 85$ lattice and which we denote by $x$,
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{sceneTrue.eps}\\[-0.7cm]
\end{tabular}
\end{center}
\caption{\label{fig:mpe}Maximum posterior estimation example: True scene used in the maximum posterior
estimation and fully Bayesian model examples.}
\end{figure}
is an unobserved true scene, and assume that a corresponding noisy version $y$ of $x$ is
observed. Here we will use a $y$
generated from $x$ by drawing each element of $y$ independently from a normal
distribution with unit standard deviation and with means $0$ and $1$ for the
white and black areas of $x$, respectively. Assuming the likelihood parameters
to be known, and assigning an MRF prior $p(x)$ to $x$, we can estimate $x$ from
$y$ by maximising the posterior distribution with respect to $x$, i.e.
\begin{equation}
\widehat{x} = \mbox{arg}\max_x\left\{ p(x)p(y|x)\right\}.
\end{equation}
To compute $\widehat{x}$ is computationally intractable,
but by adopting the procedure discussed in Section \ref{sec:maximisation}
we obtain an approximation to $\widehat{x}$. We have done this for the six priors
defined above and for values of $\nu$ from $1$ to $18$. The scenes in Figure
\ref{fig:mpeResults}
\begin{figure}
\begin{center}
\begin{tabular}{@{}c@{}c@{}c@{}c@{}}
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-I4-18.eps} &
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-I6-18.eps} &
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-I8-18.eps} &
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-Ic-18.eps}\\[-0.4cm]
$0.4$ & $0.6$ & $0.8$ & $-\ln(\sqrt{2}-1)$\\[-0.5cm]
&
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-H2-18.eps} &
\includegraphics[height=3.3cm,width=3.3cm,angle=-90]{xhat-H3-18.eps} &
\\[-0.4cm]
& Model 1 & Model 2 & \\[-0.5cm]
\end{tabular}
\end{center}
\caption{\label{fig:mpeResults}Maximum posterior estimation example: Approximations
to $\widehat{x}$ when $\nu=18$ for the four Ising priors (upper row) and
the two higher order MRFs (lower row).}
\end{figure}
show the results for $\nu=18$. It is interesting to observe that the
results for the Ising prior with $\beta=0.6$ and $0.8$ are very similar
to the results for the higher-order interaction MRFs Models 1 and 2,
respectively.
Comparing the results for different values of
$\nu$ we find that for the Ising model with $\beta=0.4$ the results are
identical for all $\nu=5$ to $18$. For the other seven priors the resulting
scenes continued to vary slightly for all the values we used for $\nu$, but
in all cases the differences became smaller for higher values of $\nu$.
For example, for all these seven priors less than $0.6\%$ of the nodes were
assigned different values when $\nu=16$ and $\nu=18$ were used. When
comparing the results with the true scene the Ising model with $\beta=0.8$
has the lowest number of misclassifications, with the higher-order MRF Model 2
slightly behind.
\subsubsection{\label{sec:fullyBayesian}Fully Bayesian modelling}
Assume again that we have observed a noisy scene $y$ corresponding to a unobserved
true image $x$. An alternative to the strategy for estimating $x$ from $y$ discussed
above is to adopt a fully Bayesian approach. Here we consider the same simulated $y$
as we did in Section \ref{sec:mpe}. Then we consider the unobserved $x$ to
be a sample from an MRF prior $p(x|\theta)$, where $\theta$ is a parameter vector. For
$p(x|\theta)$ we try both the Ising prior (\ref{eq:Ising}) and the higher order MRF defined in
Section \ref{sec:models}. The higher order MRF has the ten parameters specified in Figure \ref{fig:pot},
but to make the model identifiable we fix one parameter corresponding to each of the two
maximal cliques. We set the
potentials to zero for the configurations where all nodes in a maximal clique are equal.
Thus, $\theta$ gets eight elements in this prior. To fully specify the Bayesian
model we need to adopt a parametric form for the likelihood $p(y|x,\varphi)$, where
$\varphi$ is a parameter vector, and to specify priors for $\theta$ and $\varphi$. For the
likelihood we assume the product of normals given in Section \ref{sec:mpe} and
$\varphi$ contains a mean value and a standard deviation for each of the two possible
values of the elements in $x$, i.e. $\varphi = [\mu_0,\mu_1,\sigma_0,\sigma_1]$.
To avoid problems if all elements of $x$ are assigned to the same value we need to
adopt proper priors for the likelihood parameters. For $\mu_0$ and $\mu_1$ we use independent
normal priors with zero mean and standard deviation ten, but
to make the model identifiable we add the restriction $\mu_0\leq \mu_1$. For
$\sigma_0,\sigma_1\geq 0$ we a priori assume independent exponential densities
with means equal to ten.
To simulate from the resulting posterior $p(x,\theta,\varphi|y) \propto
p(\theta)p(\varphi)p(x|\theta)p(y|x,\varphi)$ is computationally infeasible due
to the normalising
constant of $p(x|\theta)$, but we can simulate from the approximation
$\widetilde{p}(x,\theta,\varphi|y) \propto
p(\theta)p(\varphi)\widetilde{p}(x|\theta)p(y|x,\varphi)$, where $\widetilde{p}(x|\theta)$
is defined as in (\ref{eq:pommapprox}) and here we adopt the second POMM
approximation variant discussed in the paragraph following (\ref{eq:pommapprox}).
. To simulate from $\widetilde{p}(x,\theta,\varphi|y)$
we adopt a Metropolis--Hastings algorithm and alternate between single site updates for the
likelihood parameters and the for
the components of $x$ and a joint block update for the MRF parameters and $x$. To do
Gibbs updates for the likelihood parameters is not feasible, but we use proposals
distributions which are close to the full conditionals. More precisely, for each of $\mu_0$ and $\mu_1$
we find the full conditionals when ignoring the restriction $\mu_0<\mu_1$ and use
these as proposals distributions, and for $\sigma_0$ and $\sigma_1$ we propose
potential new values by proposing values for the corresponding variances
from their full conditionals when a very vague inverse gamma prior is assumed.
The proposals are then accepted of rejected according to the standard Metropolis--Hastings
procedure. In practice essentially all proposals are accepted.
In the single site updates for the components of $x$
we generate the potential new state by changing the value of a randomly
chosen element in $x$, and in the block update we use a proposal distribution
\begin{equation}
q(\theta^\prime,x^\prime|\theta,\varphi,x) =
q(\theta^\prime|\theta) \widetilde{p}(x^\prime|y,\theta^\prime,\varphi^\prime),
\end{equation}
where $q(\theta^\prime|\theta)$ is a
random walk proposal for $\theta$, and
$\widetilde{p}(x^\prime|y,\theta^\prime,\varphi^\prime)$ is the POMM approximation
of the posterior MRF $p(x^\prime|y,\theta^\prime,\varphi^\prime)\propto \widetilde{p}(x|\theta^\prime)
p(y|x^\prime,\varphi^\prime)$. In the random walk proposals for $\theta$ we sample the
elements independently from normal distributions centered at the current values and with all
standard deviations equal to $0.1$ both for the Ising and higher-order MRF prior
cases. We define one MH iteration to consist of one update for each of the
likelihood parameters,
$89\cdot 85$ single site updates for elements of $x$ and one block update as defined above.
Using $\nu=8$ to define
the POMM approximations, Figure \ref{fig:fullyhomrf}
\begin{figure}
\begin{center}
\begin{tabular}{cccc}
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{xHat-Fully-Ising} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-Ising-param-2} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-Ising-param-3} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-Ising-param-1} \\
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{xHat-Fully-Higher} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-higher-param-10} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-higher-param-11} &
\includegraphics[height=3.0cm,width=3.0cm,angle=-90]{Fully-higher-param-7}\\[-0.6cm]
\end{tabular}
\end{center}
\caption{\label{fig:fullyhomrf}Fully Bayesian model example: Results for a fully Bayesian model, using
an Ising prior (upper row) and the higher-order interaction MRF prior (lower row). The leftmost column shows
the maximum marginal posterior probability estimate of the underlying scene. The two following columns show
estimated posterior distributions for $\mu_0$ and $\mu_1$, respectively. The rightmost column
shows estimated posterior distribution for $\theta$ for the Ising prior and for the eighth parameter
in Figure \ref{fig:pot} for the higher-order interaction prior.}
\end{figure}
summarises some simulation results. The left column shows the estimated true scene using the
marginal posterior probability estimator for each of the two priors. We observe that the
higher-order MRF looks less noisy, and the fraction of wrongly classified values are
$4.7\%$ and $4.2\%$ for the Ising and higher-order MRF priors, respectively. The
corresponding means of the marginal posterior probabilities for the wrong class labels are $0.102$ and
$0.083$, respectively. Thus, the posterior probability for the wrong class is on average
$18.5\%$ lower when using the higher-order MRF prior than with the Ising prior. The two
middle columns in Figure \ref{fig:fullyhomrf} show the estimated posterior distributions
for $\mu_0$ and $\mu_1$ and we observe that the true values, zero for the second column and one
for the third, are far out in the tail of the posterior when using the Ising prior and more centrally
located when using the higher-order MRF. All of this demonstrates the potential advantage of using
a higher-order MRF prior. The last column in Figure \ref{fig:fullyhomrf} shows the
estimated posterior distribution for $\theta$ when using the Ising prior, and the for the eighth
parameter in Figure \ref{fig:pot} for the higher-order interaction prior. We observe that the
variance in the posterior for $\theta$ is quite small, but its value is also
well above the phase transition limit.
The posterior variability for the parameter in the
higher-order interaction prior, however, is very large, and the same is true for the other
parameters in this prior. This may indicate that the higher-order interaction prior is
over-parameterised and that better results could perhaps have been obtained by adopting a prior
with fewer parameters. The best alternative would perhaps be to put a prior also on the
interaction order of the model, so that the complexity of the model could adapt automatically to the
problem in focus.
\subsubsection{Perfect sampling by rejection sampling}
The last application of our approximation and bounds we discuss is how it can be
used to construct a rejection sampling algorithm generating perfect samples
from some MRFs $p(x)$. Let $p(x)=\frac{1}{c}\exp\{ U(x)\}$
be a given binary MRF, and let $\widetilde{p}(x)$
denote the corresponding POMM approximation.
One can then imagine a rejection sampling
algorithm generating candidate samples from $\widetilde{p}(x)$ and accepting a
candidate $x$ with probability
\begin{equation}\label{eq:rejectionSampling}
\alpha(x) = k\cdot\frac{p(x)}{\widetilde{p}(x)} = \widetilde{k}\cdot \frac{\exp\{ U(x)\}}{\widetilde{p}(x)},
\end{equation}
where $k$ is a constant so that $\alpha(x)\leq 1$ for all $x$, and $\widetilde{k}=k/c$.
Clearly the optimal choice for
$\widetilde{k}$ is
\begin{equation}
\widetilde{k}_{\mbox{\tiny opt}} = \min_x\left\{ \widetilde{p}(x)e^{-U(x)}\right\},
\end{equation}
but to compute $\widetilde{k}_{\mbox{\tiny opt}}$ is computationally intractable. However, noting that
$\ln\left\{\widetilde{p}(x)e^{-U(x)}\right\}$ is a pseudo-Boolean function we can find a lower bound
$\ln \widetilde{k}_{\mbox{\tiny bound}} \leq \ln \widetilde{k}_{\mbox{\tiny opt}}$ as discussed in
Section \ref{sec:mpe}, and use $\widetilde{k}=\widetilde{k}_{\mbox{\tiny bound}}$ in
(\ref{eq:rejectionSampling}).
It should be noted this procedure implies that we need to compute the normalising constant
of the conditional distributions in (\ref{eq:pommapprox}) for all values of
the conditioning variables, so when defining the POMM approximation $\widetilde{p}(x)$
we must use the second approximation variant discussed in the paragraph
following (\ref{eq:pommapprox}).
The acceptance rate of such a rejection sampling procedure will depend both
on the quality of the approximation $\widetilde{p}(x)$ and the ratio
$\widetilde{k}_{\mbox{\tiny bound}}/\widetilde{k}_{\mbox{\tiny opt}}$. One should note that the goal
in this setting is not to get a very high rejection sampling acceptance rate, but rather to
find a good trade off between the acceptance rate and the computation time for generating the proposals,
for example by finding the smallest value of $\nu$ that give an acceptance rate above a given threshold.
We have tried the procedure on the posterior distributions defined in Section \ref{sec:mpe}.
For the posteriors based on Ising priors with $\beta=0.4$, $0.6$, $0.8$ and $-\ln(\sqrt{2}-1)$
we needed $\nu=5$, $7$, $9$ and $11$, respectively, to obtain acceptance rates above $0.1$,
and to get acceptance rates above $0.5$ the corresponding values for $\nu$ are
$6$, $8$, $11$ and $13$. Perfect samples from these four posterior distributions can of course
alternatively be generated by the coupling from the past procedure of \citet{art27}. As the
rejection sampling procedure requires an initiation step of establishing $\widetilde{p}(x)$
and computing $\widetilde{k}_{\mbox{\tiny bound}}$, coupling from the past is the
most efficient alternative whenever only one or a few independent samples are required,
but if a large number of samples are wanted the rejection sampling algorithm
becomes the best alternative. We have tried the rejection sampling procedure also for
the posteriors when adopting the higher order models defined in Section \ref{sec:models} as priors,
but for this prior we were not able to get useful acceptance rates within reasonable computation times.
\subsection{\label{sec:dataExample}Real data example}
In this section we consider a United States cancer mortality
map compiled by \citet{book36}. Figure \ref{fig:mortalityMap} shows
\begin{figure}
\begin{center}
\includegraphics[height=6.0cm,width=6.0cm,angle=-90]{realdata}
\end{center}
\vspace*{-0.8cm}
\caption{\label{fig:mortalityMap}Real data example: U.S. liver and gallbladder cancer mortality map
for white males during 1950-1959. Black and white squares denote counties with high
and low cancer mortality rates, respectively. The outer box shows the boundary of the extended lattice.}
\end{figure}
the mortality map for liver and gallbladder cancers for white males in 1950-1959,
where black and white squares denote counties with high and low cancer mortality rates,
respectively. The data set is previously analysed in
\citet{art148} and \citet{art149}, see also the discussion in
\citet{book37}. In these studies a free boundary autologistic model
is compared with a model where the values in the nodes are assumed to
be independent, and the conclusions in both studies are that the spatial model
gives a much better fit to the data than the independence model.
We define a fully Bayesian model for the data set and a priori assign
equal probabilities for three possible models. The two first models are the
independence and autologistic models adopted in the previous studies,
whereas the third model is an MRF with a $3\times 3$ neighbourhood
and higher-order interactions.
To reduce the effect of the boundary assumption we include the observed
nodes in a larger rectangular lattice as illustrated in Figure
\ref{fig:mortalityMap}, and adopt a free boundary assumption
for the extended lattice. We let $z$ denote a vector of the observed values and
$y$ a vector of the unobserved values, and let $x=(y,z)$ be a vector of
all the values in the extended lattice. The dimensions of the rectangular
lattice is chosen so that every observed node is at least $10$ nodes
away from the border, and the extended lattice is then $78\times 87$.
Of the nodes in the extended lattice $34.6\%$ is observed. In the following
we first give a more precise specification of the three possible models we allow
on the extended lattice and thereafter define prior distributions
for the parameters in these models.
The independence model has only one parameter, which we denote by $\theta$,
and
\begin{equation}
p_1(x|\theta) \propto \exp\left( \theta\sum_{i\in S} x_i\right).
\end{equation}
In the autologistic model we assume a first-order neighbourhood and assume the
horizontal and vertical interactions to be equal. The model then has two
parameters, which we denote by $\theta_0$ and $\theta_1$ and set
$\theta = (\theta_0,\theta_1)$. Our MRF is then given by
\begin{equation}
p_2(x|\theta) \propto \exp\left\{ \theta_0\sum_{i\sim j}I(x_i\neq x_j)
+ \theta_1\sum_{i\sim j}I(x_i=1 \cap x_j=1)\right\}.
\end{equation}
In the $3 \times 3$ neighbourhood MRF the maximal cliques are $2\times 2$ blocks of
nodes. Also for this model we assume the potential function $V_\Lambda(x_\Lambda)$
to be invariant under rotation of the values in $x_C$. The
rotational invariance restriction groups the $2^{2\cdot 2}=16$ possible
configurations of $x_C$ into the six groups illustrated in
Figure \ref{fig:model3}.
\begin{figure}
\begin{center}
\begin{tabular}{c|cccccc}
$x_C$ & \begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){10}}
\put(0,6.0){\line(1,0){10}}
\put(0,6.5){\line(1,0){10}}
\put(0,7.0){\line(1,0){10}}
\put(0,7.5){\line(1,0){10}}
\put(0,8.0){\line(1,0){10}}
\put(0,8.5){\line(1,0){10}}
\put(0,9.0){\line(1,0){10}}
\put(0,9.5){\line(1,0){10}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\put(5,0.5){\line(1,0){5}}
\put(5,1.0){\line(1,0){5}}
\put(5,1.5){\line(1,0){5}}
\put(5,2.0){\line(1,0){5}}
\put(5,2.5){\line(1,0){5}}
\put(5,3.0){\line(1,0){5}}
\put(5,3.5){\line(1,0){5}}
\put(5,4.0){\line(1,0){5}}
\put(5,4.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\put(5,5.5){\line(1,0){5}}
\put(5,6.0){\line(1,0){5}}
\put(5,6.5){\line(1,0){5}}
\put(5,7.0){\line(1,0){5}}
\put(5,7.5){\line(1,0){5}}
\put(5,8.0){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9.0){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(0,0.5){\line(1,0){5}}
\put(0,1.0){\line(1,0){5}}
\put(0,1.5){\line(1,0){5}}
\put(0,2.0){\line(1,0){5}}
\put(0,2.5){\line(1,0){5}}
\put(0,3.0){\line(1,0){5}}
\put(0,3.5){\line(1,0){5}}
\put(0,4.0){\line(1,0){5}}
\put(0,4.5){\line(1,0){5}}
\end{picture} &
\begin{picture}(10,15)(0,0)
\put(0,0){\line(1,0){10}}
\put(0,5){\line(1,0){10}}
\put(0,10){\line(1,0){10}}
\put(0,0){\line(0,1){10}}
\put(5,0){\line(0,1){10}}
\put(10,0){\line(0,1){10}}
\put(0,5.5){\line(1,0){5}}
\put(0,6.0){\line(1,0){5}}
\put(0,6.5){\line(1,0){5}}
\put(0,7.0){\line(1,0){5}}
\put(0,7.5){\line(1,0){5}}
\put(0,8.0){\line(1,0){5}}
\put(0,8.5){\line(1,0){5}}
\put(0,9.0){\line(1,0){5}}
\put(0,9.5){\line(1,0){5}}
\put(5,5.5){\line(1,0){5}}
\put(5,6.0){\line(1,0){5}}
\put(5,6.5){\line(1,0){5}}
\put(5,7.0){\line(1,0){5}}
\put(5,7.5){\line(1,0){5}}
\put(5,8.0){\line(1,0){5}}
\put(5,8.5){\line(1,0){5}}
\put(5,9.0){\line(1,0){5}}
\put(5,9.5){\line(1,0){5}}
\put(0,0.5){\line(1,0){5}}
\put(0,1.0){\line(1,0){5}}
\put(0,1.5){\line(1,0){5}}
\put(0,2.0){\line(1,0){5}}
\put(0,2.5){\line(1,0){5}}
\put(0,3.0){\line(1,0){5}}
\put(0,3.5){\line(1,0){5}}
\put(0,4.0){\line(1,0){5}}
\put(0,4.5){\line(1,0){5}}
\put(5,0.5){\line(1,0){5}}
\put(5,1.0){\line(1,0){5}}
\put(5,1.5){\line(1,0){5}}
\put(5,2.0){\line(1,0){5}}
\put(5,2.5){\line(1,0){5}}
\put(5,3.0){\line(1,0){5}}
\put(5,3.5){\line(1,0){5}}
\put(5,4.0){\line(1,0){5}}
\put(5,4.5){\line(1,0){5}}
\end{picture}
\\ \hline
$V_\Lambda(x_\Lambda;\theta)$ & $0$ & $\theta^0$ & $\theta^1$ & $\theta^2$ & $\theta^3$ & $\theta^4$ \\[-0.4cm]
\end{tabular}
\end{center}
\caption{\label{fig:model3}The real data example: The six possible configurations in a $2\times 2$
clique (up to rotation), and the corresponding values of the
potential function $V_\Lambda(x_\Lambda;\theta)$ in the $3\times 3$ neighbourhood MRF.}
\end{figure}
Without loss of generality we can set the potential value for the all zero
configuration to zero, and we define one parameter for each of the other five
configuration sets, again as shown in Figure \ref{fig:model3}. The
parameter vector in the model is thereby
$\theta=(\theta^0,\ldots,\theta^4)$ and the MRF is given as
\begin{equation}
p_3(x|\theta) \propto \exp\left\{ \sum_{\Lambda\in{\cal C}} U_\Lambda(x_\Lambda;\theta)\right\},
\end{equation}
where $U_\Lambda(x_\Lambda;\theta)$ is as specified in Figure \ref{fig:model3}.
As specified above we assume a prior probability $p(m)=1/3$ for each model $m=1, 2, 3$.
Given any of the three models we let the prior for the associated parameters,
$p_m(\theta)$, be a Gaussian distribution where the components of $\theta$ are
independent and Gaussian distributed with zero mean and some variance $\sigma^2$.
A priori we do not expect very strong interactions for this type of data so we
set $\sigma^2=1$. The resulting posterior distribution of interest is
$p(m,\theta|z)$, but the unobserved vector $y$ makes it hard to simulate
from this distribution so in the simulation we focus on
$p(m,\theta,y|z) \propto p(m)p_m(\theta)p_m(y,z|\theta)$.
For $m=2$ and $3$ the MRF $p_m(y,z|\theta)$ of course contains a computationally intractable
normalising constant, so as in Section \ref{sec:fullyBayesian} we replace
$p_m(y,z|\theta)$ with a corresponding approximation $\widetilde{p}_m(y,z|\theta)$
as defined in (\ref{eq:pommapprox}). As in the fully Bayesian example in Section
\ref{sec:fullyBayesian} we here use the second POMM approximation variant
discussed in the paragraph following (\ref{eq:pommapprox}) as this alternative
is somewhat faster than the alternative POMM approximation.
The distribution we want to sample is then
\begin{equation}
\widetilde{p}(m,\theta,y|z) \propto p(m)p_m(\theta)\widetilde{p}_m(y,z|\theta),
\end{equation}
and to sample from this distribution we adopt a reversible jump MCMC algorithm
\citep{art22}. To get an acceptable trans-dimensional mixing in
the MCMC algorithm we found it necessary to include more auxiliary variables.
For each of the three models $k=1,2,3$ we added a parameter $\theta_k$ and
vector $y_k$, where $\theta_1\in\mathbb{R}$, $\theta_2\in\mathbb{R}^2$ and
$\theta_3\in\mathbb{R}^5$ and each of the $y_k$'s are vectors of the same
size as $y$. We assume
$(\theta_k,y_k)\sim p_k(\theta_k) \widetilde{p}_k(y_k|\theta_k,z)$
independently for each $k$ and independent of $(m,\theta,y)$, where
$\widetilde{p}_k(y_k|\theta_k,z)$ is the POMM approximation of
$p_k(y_k|\theta_k,z)\propto p_k(y_k,z|\theta_k)$.
To simulate from
\begin{equation}
\widetilde{p}(m,\theta,y,\theta_1,y_1,\theta_2,y_2,\theta_3,y_3|z) =
\widetilde{p}(m,\theta,y|z)\prod_{k=1}^3 p_k(\theta_k)\widetilde{p}_k(y_k|\theta_k,z)
\end{equation}
we adopt the reversible jump setup with
three types of proposals. The first proposal is a random walk proposal
for each of $\theta$, $\theta_1$, $\theta_2$ and $\theta_3$. A separate update is
performed for each of these four variables, so a proposed new value for one of them
is accepted or rejected before a change for another of the vectors
is proposed. The components in the potential new vector are generated
independently from Gaussian distributions centered at the current value
and with the same variance $\tau^2$ for all components. By trial and error
we tuned the value of the proposal variance and ended up using $\tau=0.025$.
The second proposal type generate potential new values for each of $y$, $y_1$, $y_2$ and
$y_3$. Again a proposal followed by an acceptance or rejection is done for
each of the vectors separately. For $y_k,k=1,2,3$ the potential new
value is generated from $\widetilde{p}_k(y_k|\theta_k,z)$, and the
potential new value for $y$ is generated from $\widetilde{p}_m(y|\theta_m,z)$.
One can note that the updates for $y_1$, $y_2$ and $y_3$ are
Gibbs updates, whereas the update for $y$ is not. The acceptance rate
for the $y$ updates is, however, very close to one. The third update type
is the only trans-dimensional update. Here new values are proposed
for $(m,\theta,y)$ and for one of $(\theta_1,y_1)$, $(\theta_2,y_2)$
and $(\theta_3,y_3)$. First a new value $m^\prime$ is proposed for the model indicator
$m$. If $m=1$ or $3$ we always set $m^\prime=2$, and if $m=2$ we set
$m^\prime=1$ or $3$ with probabilities a half for each.
Thereafter potential new values for $\theta$ and $y$ is
deterministically set as $\theta^\prime=\theta_{m^\prime}$ and
$y^\prime=y_{m^\prime}$. Finally potential new values for $\theta_{m^\prime}$
and $y_{m^\prime}$ are sampled as $\theta^\prime_{m^\prime}\sim\mbox{N}(\theta_{m^\prime},
\tau^2 I)$ and $y^\prime_{m^\prime}\sim \widetilde{p}_{m^\prime}(y_m|\theta^\prime_{m^\prime},z)$,
respectively. Here we use the same variance $\tau^2$ as in the random walk proposals
discussed above, and $I$ is the identity matrix of the suitable dimension.
The proposed new values are then accepted or rejected according to the
standard reversible jump MCMC acceptance probability. In particular it is
easy to show that the Jacobian determinant occurring in the expression for
the acceptance probability equals unity for this kind of proposal.
To check the convergence and mixing properties of the reversible
jump MCMC algorithm we ran three independent runs with different
starting values. For each of $k=1,2$ and $3$ we had a chain where we
initially set $m=k$ and ran 200 iterations without the transdimensional move.
The idea here is that the chains should essentially converge conditional on
the value of $m$ before we allow the value of $m$ to vary. After these
$200$ initial iterations we ran the three chains for an additional $4300$ iterations,
now including the transdimensional move. The chain starting with $m=1$
then quickly changed to having $m=2$ and $m=3$ and never returned to the state
$m=1$. The chains starting with $m=2$ and $m=3$ never visited
the state $m=1$. All three chains had several jumps back and fourth
between $m=2$ and $m=3$ as can be seen from Figure \ref{fig:mchains}.
\begin{figure}
\begin{center}
\includegraphics[height=6.0cm,width=3.0cm,angle=-90]{trace1}
\end{center}
\vspace*{-0.3cm}
\caption{\label{fig:mchains}Real data example: For the Markov chain simulation starting
at $m=1$, trace plot of the simulated values for $m$.}
\end{figure}
In the analysis of the Markov chain runs we discard the first $500$ iterations of each run
as burn-in, and use all three runs to estimate the quantities of interest.
We estimate the posterior probability for model $m=3$ by the fraction of times the chains
are visiting this state, and get $\widehat{p}(m=3|z)=0.680$. To evaluate the uncertainty
in this number we split the chains into intervals of $100$ iterations and estimate
$p(m=3|z)$ based on each of these intervals. These estimates are close to being
uncorrelated so, considering them as independent,
we use them to construct an approximate $95\%$ credibility $t$-interval for $p(m=3|z)$.
The interval becomes $[0.619,0.741]$, clearly showing that $m=3$ is the model
with highest a posteriori probability. This demonstrates that even for data sets
where the interactions are quite weak there can be a need for including
higher-order interactions. Figures \ref{fig:resultModel2} and
\begin{figure}
\begin{center}
\begin{tabular}{cc}
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta2-0} &
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta2-1} \\
$\theta_0$ & $\theta_1$
\end{tabular}
\end{center}
\vspace*{-0.7cm}
\caption{\label{fig:resultModel2}Real data example: Histograms of the simulated posterior
values for $\theta_0$ and $\theta_1$ when $m=2$.}
\end{figure}
\ref{fig:resultModel3} show estimates of the marginal distributions for the model
\begin{figure}
\begin{center}
\begin{tabular}{ccc}
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta3-0} &
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta3-1} &
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta3-2} \\
$\theta_0$ & $\theta_1$ & $\theta_2$ \\[-0.4cm]
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta3-3} &
\includegraphics[height=4.0cm,width=3.0cm,angle=-90]{theta3-4} \\
$\theta_3$ & $\theta_4$ \\[-0.6cm]
\end{tabular}
\end{center}
\caption{\label{fig:resultModel3}Real data example: Histograms of the simulated posterior
values for $\theta_0,\ldots,\theta_4$ when $m=3$.}
\end{figure}
parameters for model $m=2$ and $m=3$, respectively. One should note that the
spread of all of these parameters are well within the variability of the corresponding standard
normal prior, so our prior here should not be very influential on these results. We also
observe that all parameters, both for model $m=2$ and $3$, are significantly smaller than zero.
This reflects that zero is the dominating value of the data set and that we in both models have chosen
to set the potential for maximal cliques with only zero values to the reference value of zero.
\section{\label{sec:cr}Closing remarks}
In this report we have shown how we can derive an approximate
forward-backward algorithm by studying how to approximate the
pseudo-Boolean energy function during the summation process. This approximation can then be used to
work with statistical models such as MRFs. It
allows us to produce approximations and bounds of the normalising constant and
likelihood for models that would normally be
too computationally heavy to work with directly. It also
gives POMM approximations of MRFs that can be used as a surrogate for
MRFs in more complicated model setups.
We have demonstrated
the accuracy of the approximation and bounds through simple experiments with the
Ising model and a higher-order interaction model,
and demonstrated some potential applications by simulation experiments.
We have also applied our approximation to a real life data set and here
demonstrated that higher-order interactions may be important even
for data sets with weak interactions.
We round off now with some possible
future extensions as well as some closing remarks.
The approximation we have defined was inspired by the work in
\citet{art130} and there are many parallels between the two.
There the energy function was represented as a
binary polynomial and small interactions was dropped while running
the forward-backward algorithm. This worked reasonably well, but
one may worry that dropping many smalls interactions may produce
an approximation no better than when dropping one large interaction.
Moreover, for models with strong interactions the
approximation in \citet{art130} would either include too many terms and thus explode in
run-time, or if the cutoff level was set low enough to run the
algorithm, exclude so many of the interactions that the approximation became
uninteresting. In a sense the work in this report has been an effort to
deal with these issues. We have a clearly defined approximation
criterion and the construction of the algorithm allows for much
more direct control over the run time. Also, by not just dropping small
terms, but approximating the pseudo-Boolean function in such a way
that we minimise the error sum of squares we manage to get better
approximations of the models with stronger interactions.
In our setting the sample space of the pseudo-Boolean
function has a probability measure on it, however in our discussion of
approximating pseudo-Boolean functions we have considered each state
in the sample space as equally important. Assuming we are interested
in approximating the normalising constant, this is probably far from
optimal and is reflected in our results. As we can see for the Ising model
our approximation
works better the smaller the interaction parameter $\theta$. Our initial
approximation attempts to spread the error of removing interactions as
evenly as possible among the states. Ideally, we would like to have small errors in states
with a corresponding high probability, and larger errors in states
with low probability. To get this approximation we should replace
the $\mbox{SSE}$ in (\ref{eq:SSE}) with a
weighted error sum of squares.
This problem has also been studied in the literature, see
for instance \citet{art151,art150}. However, unlike
the unweighted case an explicit solution is not readily available for
a probability density like the MRF. The iterative method of removing
interactions does not work here, nor can we group the equations like
we do with the SOIR approximation.
In all our examples we consider MRFs defined on a lattice and assume
the potential functions to be translational invariant. Our
approximations and bounds are, however, also valid for MRFs
defined on an arbitrary graph and applications of the approximation and bounds
in such situations is something we want to explore in the future.
|
\section{Introduction}
\label{sec:intro}
Soliton and black hole solutions of Einstein-Yang-Mills (EYM) theory have been studied extensively for over twenty years (see, for example, Ref.~\onlinecite{Volkov1} for a review).
The first solutions found were spherically symmetric, purely magnetic, asymptotically flat, solitons \cite{Bartnik1} and black holes \cite{Bizon1} in four-dimensional
${\mathfrak {su}}(2)$ EYM.
Discrete families of solutions were found numerically and their existence was later proven (see Refs.~\onlinecite{Breitenlohner1, Smoller} for some analytic work).
The purely magnetic gauge field is described by a single function $\omega $, which has at least one zero. The families of solutions are characterized by the event horizon radius $r_{h}$ (with $r_{h}=0$ corresponding to soliton solutions) and the number of zeros of the function $\omega $.
Both the soliton and black hole families of solutions are unstable under linear, spherically symmetric perturbations \cite{Straumann1},
with the number of unstable perturbation modes of the solutions being twice the number of zeros of
of $\omega $ \cite{Lavrelashvili1,Volkov2}.
Many generalizations of the original spherically symmetric ${\mathfrak {su}}(2)$ solitons and black holes have been considered in the literature (some of which are reviewed in Ref.~\onlinecite{Volkov1}).
For example, numerical solutions have been found which retain the spherical symmetry of the original solutions but enlarge the gauge group to ${\mathfrak {su}}(N)$ (see, for example, Ref.~\onlinecite{Galtsov1}).
The solution space is more complicated with the larger gauge group, but solutions still exist in discrete families.
Furthermore, all asymptotically flat, spherically symmetric, soliton and black hole solutions with arbitrary gauge group are unstable under linear, spherically symmetric perturbations \cite{Brodbeck}.
The model can also be generalized by considering space-times which are not asymptotically flat or which have more than four dimensions. In four-dimensional space-time, discrete families of spherically symmetric soliton and black hole solutions of ${\mathfrak {su}}(2)$ EYM also exist in asymptotically de Sitter space-time \cite{Volkov3}, but, like their asymptotically flat counterparts, they are unstable \cite{dSstab}.
If one considers higher-dimensional space-times, in order to have spherically symmetric finite mass solutions, the YM action must be modified by the addition
of higher-order curvature terms \cite{Volkov4}.
With these additional terms, soliton and black hole solutions have been found in both asymptotically flat and asymptotically de Sitter space-times \cite{HD}.
About ten years after the discovery of four-dimensional, spherically symmetric, purely magnetic, asymptotically flat, solitons and black holes in ${\mathfrak {su}}(2)$ EYM, their analogues in four-dimensional asymptotically anti-de Sitter (adS) space-time were found \cite{Winstanley1,Bjoraker,Breitenlohner2}.
The purely magnetic ${\mathfrak {su}}(2)$ gauge field is still described by a single function $\omega $, but now continuous families of solutions are found, which are indexed by the event horizon radius $r_{h}$ as before (including $r_{h}=0$ for solitons), the negative cosmological constant $\Lambda $ and the value of the gauge field function on the horizon $\omega _{h}$ (there is an alternative parameter for soliton solutions, which governs the behaviour of the magnetic gauge field function near the origin).
One striking feature of the families of solutions is the existence, for sufficiently large $\left| \Lambda \right| $, of solutions where the magnetic gauge field function $\omega $ has no zeros.
These solutions where $\omega $ is nodeless are particularly important because at least some of them are stable under linear, spherically symmetric, perturbations \cite{Winstanley1,Bjoraker,Breitenlohner2}.
The existence, for sufficiently large $\left| \Lambda \right| $, of soliton and black hole solutions which are stable under linear, non-spherically symmetric, perturbations has also been proven \cite{Sarbach1, Sarbach2}.
In this paper we consider only four-dimensional, spherically symmetric solutions, but asymptotically adS generalizations to higher-dimensions \cite{HDadS}
or non-spherically symmetric space-times \cite{adSgen,adSthermo} do exist.
One generalization which has received a great deal of attention in the literature over the past seven years is topological EYM black holes in adS, in particular the relevance of black holes with planar event horizons to models of holographic superconductors (see, for example, the recent review \cite{holreview} for more details and references).
Purely magnetic black holes with non-spherical event horizon topology in ${\mathfrak {su}}(2)$ EYM in adS appeared in the literature soon after their spherically symmetric counterparts \cite{adStop}.
Unlike the situation in asymptotically flat space-time \cite{ershov}, in asymptotically adS space-time ${\mathfrak {su}}(2)$ EYM black holes and solitons can
have nontrivial electric and magnetic fields.
While spherically symmetric dyonic solutions (both solitons and black holes) in ${\mathfrak {su}}(2)$ EYM in adS were found soon after the purely magnetic black holes \cite{Bjoraker}, topological dyonic solutions have been studied only more recently.
Gubser \cite{gubser} considered four-dimensional dyonic ${\mathfrak {su}}(2)$ EYM black holes in adS with planar event horizons.
He found a second-order phase transition between the embedded planar Reissner-Nordstr\"om-adS black hole and a black hole with a nontrivial
YM field condensate.
Planar EYM black holes in adS have subsequently been widely studied as models of $p$-wave superconductors \cite{pufu} (see also Refs.~\onlinecite{holreview,holographic}
for a selection of work in this area).
Returning to four-dimensional, spherically symmetric, purely magnetic, asymptotically adS solutions, a natural question is whether the above stable ${\mathfrak {su}}(2)$ solitons and black holes have generalizations with a larger ${\mathfrak {su}}(N)$ gauge group.
The answer is affirmative: such solutions have been found numerically for gauge groups ${\mathfrak {su}}(3)$ and ${\mathfrak {su}}(4)$ \cite{Baxter2}.
For the larger ${\mathfrak {su}}(N)$ gauge group, the purely magnetic gauge field is described by $N-1$ functions $\omega _{j}$ (see section \ref{sec:ansatz} below).
As in the ${\mathfrak {su}}(2)$ case, there are continuous families of solutions, parameterized by the negative cosmological constant $\Lambda $, the event horizon radius $r_{h}$ (with $r_{h}=0$ for soliton solutions) and $N-1$ parameters describing the form of the gauge field functions either on the event horizon or near the origin.
Numerically it is found that, if $\left| \Lambda \right| $ is sufficiently large, then there are solutions in which all the gauge field functions $\omega _{j}$ have no zeros.
For general $N$, the existence of such nodeless, spherically symmetric, purely magnetic, asymptotically adS, soliton and black hole solutions of ${\mathfrak {su}}(N)$ EYM has been proven for
sufficiently large $\left| \Lambda \right| $ \cite{Baxter3}.
In this paper we address the question of whether these soliton and black hole solutions of ${\mathfrak {su}}(N)$ EYM in which all the magnetic gauge field functions have no zeros are stable.
The outline of the paper is as follows.
In section \ref{sec:EYM} we introduce ${\mathfrak {su}}(N)$ EYM with a negative cosmological constant and the ansatz \cite{Kunzle1} for a spherically symmetric gauge potential. We derive the field equations describing static, purely magnetic, configurations and the perturbation equations for linear, spherically symmetric, perturbations.
With an appropriate choice of gauge, the perturbation equations decouple into two sectors: the sphaleronic and gravitational sectors.
These are considered in sections \ref{sec:sphaleronic} and \ref{sec:gravitational}
respectively.
Finally we present our conclusions in section \ref{sec:conc}.
\section{The Einstein-Yang-Mills equations}
\label{sec:EYM}
\subsection{Action, metric and gauge potential}
\label{sec:ansatz}
The action for four-dimensional ${\mathfrak {su}}(N)$ Einstein-Yang-Mills (EYM) theory with a negative cosmological constant $\Lambda <0$ is:
\begin{equation}
S_{\mathrm {EYM}} = \frac {1}{2} \int d ^{4}x {\sqrt {-g}} \left[
R - 2\Lambda - \Tr \, F_{\mu \nu }F ^{\mu \nu }
\right] ,
\label{eq:action}
\end{equation}
where $R$ is the Ricci scalar, $F_{\mu \nu }$ is the non-Abelian gauge field
and ${\mathrm {Tr}}$ denotes a Lie algebra trace.
Throughout this paper, the metric has signature $\left( -, +, +, + \right) $ and
we use units in which $4\pi G = 1 = c$.
In addition, the gauge coupling constant is fixed to be equal to unity.
Varying the action (\ref{eq:action}) yields the field equations:
\begin{eqnarray}
T_{\mu \nu } & = & R_{\mu \nu } - \frac {1}{2} R g_{\mu \nu } +
\Lambda g_{\mu \nu },
\nonumber \\
0 & = & D_{\mu } F_{\nu }{}^{\mu } = \nabla _{\mu } F_{\nu }{}^{\mu }
+ \left[ A_{\mu }, F_{\nu }{}^{\mu } \right] ;
\label{eq:fieldeqns}
\end{eqnarray}
where the Yang-Mills stress-energy tensor is
\begin{equation}
T_{\mu \nu } = \Tr \left[ F_{\mu \lambda } F_{\nu }{}^{\lambda }
- \frac {1}{4} g_{\mu \nu }
F_{\lambda \sigma} F^{\lambda \sigma } \right] ,
\label{eq:Tmunu}
\end{equation}
which involves a Lie-algebra trace.
The Yang-Mills gauge field $F_{\mu \nu }$ is given in terms of the gauge potential
$A_{\mu }$ by
\begin{equation}
F_{\mu \nu } = \partial _{\mu }A_{\nu } - \partial _{\nu }A_{\mu } +
\left[ A_{\mu },A_{\nu } \right] .
\end{equation}
Our focus in this paper is on equilibrium, static, spherically symmetric, soliton and black hole solutions of the field equations (\ref{eq:fieldeqns}) and time-dependent, spherically symmetric, perturbations of those equilibrium solutions.
We therefore consider a time-dependent, spherically symmetric, geometry, whose metric in standard Schwarzschild-like co-ordinates takes the form
\begin{equation}
ds^{2} = - \mu S^{2} \, dt^{2} + \mu ^{-1} \, dr^{2} +
r^{2} \, d\theta ^{2} + r^{2} \sin ^{2} \theta \, d\phi ^{2} ,
\label{eq:metric}
\end{equation}
where the metric functions $\mu (t,r)$ and $S(t,r)$ depend on the co-ordinates $t$ and $r$ only.
Since we have a negative cosmological constant $\Lambda $, it is useful to write the metric function $\mu (t,r)$ in the form
\begin{equation}
\mu (t,r) = 1 - \frac {2m(t,r)}{r} - \frac {\Lambda r^{2}}{3}.
\label{eq:mu}
\end{equation}
In our later analysis we will also find it useful to define another function
$\Delta (t,r)$ such that
\begin{equation}
S(t,r) = \exp \Delta (t,r).
\end{equation}
With this metric ansatz the relevant components of the Einstein tensor are:
\begin{eqnarray}
G_{tt} & = &
-\frac{\mu S^2}{r^2}\left(\mu'r-1+\mu\right) ,
\nonumber
\\
G_{tr} & = & -\frac{\dot{\mu}}{\mu r} ,
\nonumber
\\
G_{rr} & = & \frac{1}{\mu Sr^2}\left(\mu'Sr+2S'\mu r-S+\mu S\right) ,
\label{eq:EinsteinTensor}
\end{eqnarray}
where here and throughout this paper we use a dot to denote $\partial /\partial t$
and a prime to denote $\partial /\partial r$.
Note that we do not need to consider the $G_{\theta \theta }$ or $G_{\phi \phi }$
components of the Einstein tensor as the field equations involving these components follow from those involving the components in (\ref{eq:EinsteinTensor}) by the Bianchi identities.
We make the following ansatz for a time-dependent, spherically symmetric,
${\mathfrak {su}}(N)$ gauge potential \cite{Kunzle1}:
\begin{eqnarray}
A & = &
{\mathcal {A}} \, dt + {\mathcal {B}} \, dr +
\frac {1}{2} \left( C - C^{H} \right) \, d\theta
- \frac {i}{2} \left[ \left(
C + C^{H} \right) \sin \theta + D \cos \theta \right] \, d\phi ,
\label{eq:gaugepot}
\end{eqnarray}
where ${\mathcal {A}}$, ${\mathcal {B}}$, $C$ and $D$ are all $\left( N \times N \right) $ matrices depending on the co-ordinates $(t,r)$ only and
$C^{H}$ is the Hermitian conjugate of $C$.
With this gauge potential ansatz, the non-zero components of the gauge field are:
\begin{eqnarray}
F_{tr} & = & \dot{B}-A' ,
\nonumber
\\
F_{t\theta} & = & \frac{1}{2}\left\{ (C-C^H)\dot{}+[A,C-C^H]\right\} ,
\nonumber
\\
F_{t\phi} & = & -\frac{i}{2}\left\{(C+C^H)\dot{}+[A,C+C^H]\right\}\sin\theta ,
\nonumber
\\
F_{r\theta} & = & \frac{1}{2}\left\{(C-C^H)'+[B,C-C^H]\right\} ,
\nonumber
\\
F_{r\phi} & = & -\frac{i}{2}\left\{(C+C^H)'+[B,C+C^H]\right\}\sin\theta ,
\nonumber
\\
F_{\theta\phi} & = & -\frac{i}{2}\left\{ [C,C^H]-D\right\}\sin\theta.
\label{eq:Fmunu}
\end{eqnarray}
In computing the component $F_{\theta \phi }$ we have made use of the identities
\cite{Kunzle1}
\begin{equation}
[D,C] =2C ,
\qquad
[D, C^H]=-2C^H.
\label{eq:Dcomm}
\end{equation}
The matrices ${\mathcal {A}}$ and ${\mathcal {B}}$ are diagonal and traceless, and we define functions $\alpha _{j}(t,r)$ and $\beta _{j}(t,r)$ for $j=1,\ldots N$ such that
\begin{eqnarray}
{\mathcal {A}} & = & i \, \Diag \left( \alpha _{1}(t,r), \ldots
\alpha _{N}(t,r) \right) ,
\nonumber \\
{\mathcal {B}} & = & i \, \Diag \left( \beta _{1}(t,r), \ldots
\beta _{N}(t,r) \right),
\end{eqnarray}
where the fact that these two matrices must be traceless means that
\begin{equation}
\sum _{j=1}^{N} \alpha _{j} (t,r) = 0 = \sum _{j=1}^{N} \beta _{j}(t,r).
\end{equation}
The matrix $C$ is upper triangular, with non-zero entries only immediately above the main diagonal. These entries are given in terms of functions $\omega _{j}(t,r)$
and $\gamma _{j}(t,r)$ for $j=1,\ldots , N-1$ by
\begin{equation}
C_{j,j+1}=\omega _{j}(t,r) e^{i\gamma _{j}(t,r)}.
\label{eq:Celements}
\end{equation}
Finally, the matrix $D$ is a constant diagonal matrix \cite{Kunzle1}:
\begin{equation}
D=\Diag \left(N-1,N-3,\ldots,-N+3,-N+1\right) .
\label{eq:matrixD}
\end{equation}
\subsection{Static solutions}
\label{sec:static}
For static solutions, all field variables depend only on the radial co-ordinate $r$.
We denote static equilibrium functions with a bar (e.g.~${\ov {\omega }}_{j}$) to distinguish them from the time-dependent perturbations which we shall consider shortly.
The static equilibrium solutions in which we are interested are purely magnetic, which means that we set the electric gauge field functions ${\ov {\alpha }}_{j}(r)\equiv 0$ for all $j=1, \ldots ,N$.
The remaining gauge freedom is then used to set all the functions
${\ov {\beta }}_{j}(r)\equiv 0$ for $j=1,\ldots ,N$ \cite{Kunzle1}.
From now on we assume that none of the magnetic gauge functions
${\ov {\omega }}_{j}(r)$ are identically zero.
In asymptotically flat space, other families of solutions have been found when this assumption is relaxed \cite{Galtsov1}.
Assuming that none of the ${\ov {\omega }}_{j}(r)$ are identically zero, one of the Yang-Mills equations becomes \cite{Kunzle1}:
\begin{equation}
{\ov {\gamma }}_{j}(r) =0, \qquad j=1,\ldots , N-1 ,
\end{equation}
and the gauge field is described by the $N-1$ magnetic gauge field functions
${\ov {\omega }}_{j}(r)$, $j=1,\ldots , N-1$.
We comment that our ansatz (\ref{eq:gaugepot})
is by no means the only possible choice in
${\mathfrak {su}}(N)$ EYM (in Ref.~\onlinecite{Bartnik2} all irreducible models are explicitly listed for $N\le 6$, and
techniques for finding {\em {all}} spherically symmetric ${\mathfrak {su}}(N)$ gauge potentials are developed).
\subsubsection{Static field equations}
\label{sec:staticeqns}
For purely magnetic, static equilibrium solutions as described above, the field equations (\ref{eq:fieldeqns}) simplify as follows.
The Einstein equations take the form:
\begin{equation}
{\ov {m}} ' = {\ov {\mu }} (r) {\ov {\Gamma }} + r^{2} {\ov {\Pi }},
\qquad
{\ov {\Delta }} ' = \frac {2{\ov {\Gamma }}}{r} ,
\label{eq:Ee}
\end{equation}
where
\begin{eqnarray}
{\ov {\Gamma }} & = & \sum _{j=1}^{N-1} {\ov {\omega }}_{j}'^{2},
\nonumber \\
{\ov {\Pi }} & = & \frac {1}{4r^{4}} \sum _{j=1}^{N} \left[ \left(
{\ov {\omega }}_{j}^{2} - {\ov {\omega }}_{j-1}^{2}
-N - 1+ 2j\right) ^{2} \right] .
\label{eq:Pistatic}
\end{eqnarray}
The $N-1$ Yang-Mills equations take the form
\begin{equation}
0 = r^{2} {\ov {\mu }} {\ov {\omega }}_{j}'' + \left(
2{\ov {m}} - 2r^{3} {\ov {\Pi}} - \frac {2\Lambda r^{3}}{3} \right)
{\ov {\omega }}_{j} + W_{j} {\ov {\omega }}_{j} ,
\label{eq:YMe}
\end{equation}
where
\begin{equation}
W_{j} = 1-{\ov {\omega }}_{j}^{2} + \frac {1}{2} \left(
{\ov {\omega }}_{j-1}^{2} + {\ov {\omega }}_{j+1}^{2} \right) .
\label{eq:Wdef}
\end{equation}
The field equations (\ref{eq:Ee}, \ref{eq:YMe}) are invariant under the transformation
\begin{equation}
{\ov {\omega }}_{j} (r) \rightarrow -{\ov {\omega }}_{j}(r)
\label{eq:omegatransform}
\end{equation}
independently for each $j$, and also under the substitution:
\begin{equation}
j\rightarrow N-j.
\end{equation}
\subsubsection{Boundary conditions}
\label{sec:boundary}
The field equations (\ref{eq:Ee}, \ref{eq:YMe}) are singular at the origin $r=0$,
a black hole event horizon $r=r_{h}$ (where ${\ov {\mu }}(r_{h})=0$) and at
infinity $r\rightarrow \infty $.
Below we briefly outline the form of the equilibrium field functions in a neighbourhood of the singular points.
The existence of local solutions near these singular points, with the forms below, is proven in Ref.~\onlinecite{Baxter3}.
\paragraph{Origin}
The form of the static field functions near the origin is rather complicated.
In particular, to completely specify the gauge field in a neighbourhood of the origin, a power series up to $O(r^{N})$ is required,
involving $N-1$ initial parameters. These $N-1$ parameters, together with the cosmological constant $\Lambda $, completely determine the solution in a neighbourhood of the origin.
The details of this power series can be found in Ref.~\onlinecite{Baxter3}
(following the analysis of Ref.~\onlinecite{Kunzle2} for the asymptotically flat case).
For our analysis in this paper, we only require the leading order behaviour
of the static field functions, which is:
\begin{eqnarray}
{\ov {m}}(r) & = & m_{3}r^{3} + O(r^{4}),
\nonumber \\
{\ov {S}}(r) & = & S_{0} + S_{2}r^{2} + O(r^{3}),
\nonumber \\
{\ov {\omega }}_{j}(r) & = & \pm \left[ j \left( N - j\right) \right] ^{\frac {1}{2}}
+ O(r^{2}),
\label{eq:origin1}
\end{eqnarray}
where $m_{3}$, $S_{0}$ and $S_{2}$ are constants.
Without loss of generality, we take the positive sign in the form of
${\ov {\omega }}_{j}(r)$ due to the invariance of the field equations under the
transformation (\ref{eq:omegatransform}).
\paragraph{Event horizon}
Assuming there is a non-extremal black hole event horizon at $r=r_{h}$, the metric
function ${\ov {\mu }}(r)$ will have a single zero there.
This fixes the value of ${\ov {m}}(r_{h})$ to be
\begin{equation}
{\ov {m}}( r_{h} ) = \frac {r_{h}}{2} - \frac {\Lambda r_{h}^{3}}{6}.
\end{equation}
In a neighbourhood of the horizon, the field variables have the form
\begin{eqnarray}
{\ov {m}}(r) & = & {\ov {m}} (r_{h}) + {\ov {m}}' (r_{h}) \left( r - r_{h} \right)
+ O \left( r- r_{h} \right) ^{2} ,
\nonumber \\
{\ov {\omega }}_{j} (r) & = & {\ov {\omega }}_{j}(r_{h})
+ {\ov {\omega }}_{j}' (r_{h})
\left( r - r_{h} \right)
+ O \left( r -r_{h} \right) ^{2},
\nonumber \\
{\ov {S}}(r) & = & {\ov {S}}(r_{h}) + {\ov {S}}'(r_{h}) \left( r-r_{h} \right)
+ O\left( r - r_{h} \right) ,
\label{eq:horizon}
\end{eqnarray}
where ${\ov {m}}'(r_{h})$, ${\ov {\omega }}_{j}'(r_{h})$ and ${\ov {S}}'(r_{h})$ can be written
in terms of the constants ${\ov {\omega }}_{j}(r_{h})$ and $S(r_{h})$ by using the field equations (\ref{eq:Ee}, \ref{eq:YMe}).
Again, due to the invariance of the field equations under the transformation
(\ref{eq:omegatransform}), we may take ${\ov {\omega }}_{j}(r_{h})>0$ without loss of generality.
The $N-1$ initial parameters $\omega _{j}(r_{h})$, together with the cosmological constant $\Lambda $ and event horizon radius $r_{h}$, completely
determine the solution of the field equations in a neighbourhood of the horizon \cite{Baxter3}.
\paragraph{Infinity}
As $r\rightarrow \infty $, the field variables have the form:
\begin{eqnarray}
{\ov {m}}(r) & = & M + O \left( r^{-1} \right) ,
\nonumber \\
{\ov {S}}(r) & = & 1 + O\left( r^{-1} \right) ,
\nonumber \\
{\ov {\omega }}_{j}(r) & = & {\ov {\omega }}_{j,\infty } + c_{j}r^{-1} +
O \left( r^{-2} \right) ,
\label{eq:infinity}
\end{eqnarray}
where $M$, ${\ov {\omega }}_{j,\infty }$ and $c_{j}$ are constants.
\subsubsection{Embedded solutions}
\label{sec:embedded}
Despite the complexity of the static field equations (\ref{eq:Ee}, \ref{eq:YMe}),
there are some embedded solutions which will be useful in our later analysis.
\paragraph{Schwarzschild-adS}
If we set
\begin{equation}
{\ov {\omega }}_{j} (r) \equiv \pm {\sqrt {j\left( N-j \right) }},
\end{equation}
for all $j=1,\ldots , N-1$ then the components of the gauge field strength tensor (\ref{eq:Fmunu}) vanish identically.
In this case the stress-energy tensor (\ref{eq:Tmunu}) therefore also vanishes and
we obtain the Schwarzschild-adS black hole solution
with
\begin{equation}
{\ov {m}}(r) \equiv M, \qquad
{\ov {S}}(r) \equiv 1,
\end{equation}
where $M$ is a constant representing the mass of the black hole.
Setting $M=0$ gives pure adS space-time as a solution of the field equations.
\paragraph{Reissner-Nordstr\"om-adS}
Alternatively, if we set
\begin{equation}
{\ov {\omega }}_{j}(r) \equiv 0
\end{equation}
for all $j=1,\ldots, N-1$ then the gauge field strength tensor (\ref{eq:Fmunu}) does not vanish as $F_{\theta \phi }$ has a contribution from the nonzero matrix
$D$ (\ref{eq:matrixD}).
In this case we obtain the magnetically charged Reissner-Nordstr\"om black hole solution with
\begin{equation}
{\ov {m}}(r) = M - \frac {Q^{2}}{2r}, \qquad
{\ov {S}}(r) \equiv 1,
\end{equation}
where the magnetic charge $Q$ is fixed to be
\begin{equation}
Q^{2} = \frac {1}{6} N \left( N+1 \right) \left( N-1 \right) .
\end{equation}
\paragraph{Embedded ${\mathfrak {su}}(2)$ solutions}
The above two solutions are effectively Abelian embedded solutions.
For all $N>2$, there is another class of embedded solutions, corresponding to
${\mathfrak {su}}(2)$ non-Abelian solutions.
To obtain these solutions, we write the $N-1$ magnetic gauge field functions
${\ov {\omega }}_{j}(r)$ in terms of a single magnetic gauge field function
${\ov {\omega }}(r)$ as follows:
\begin{equation}
{\ov {\omega }}_{j}(r) = \pm {\ov {\omega }}(r) {\sqrt {j\left( N - j\right)}} .
\end{equation}
It is shown in Ref.~\onlinecite{Baxter3} that, by a suitable rescaling of the other field variables, in this case the static field equations (\ref{eq:Ee}, \ref{eq:YMe}) reduce to those for the ${\mathfrak {su}}(2)$ case with ${\ov {\omega }}(r)$ as the single magnetic gauge field function.
Therefore any solution of the ${\mathfrak {su}}(2)$ field equations can be embedded as a solution of the ${\mathfrak {su}}(N)$ field equations.
In particular, setting ${\bar {\omega }}(r) \equiv 1$ gives the Schwarzschild-adS solution of the embedded ${\mathfrak {su}}(2)$ field equations and setting
${\bar {\omega }}(r) \equiv 0$ gives the magnetically charged Reissner-Nordstr\"om-adS black hole.
\subsubsection{Non-embedded solutions}
\label{sec:numerical}
Genuinely ${\mathfrak {su}}(N)$ static soliton and black hole solutions, which do not fall into one of the categories described in section \ref{sec:embedded}, have been studied in some detail already in the literature \cite{Baxter2}. Therefore in this section we very briefly describe some of the key features of the solutions which are required for our subsequent analysis.
As discussed in the introduction, continuous families of solutions of the field equations (\ref{eq:Ee}, \ref{eq:YMe}) are found numerically. The solutions are parameterized by the cosmological constant $\Lambda $, the event horizon radius $r_{h}$ (we can consider $r_{h}=0$ to represent soliton solutions) and, for
${\mathfrak {su}}(N)$, there are a further $N-1$ parameters which describe the gauge field (see section \ref{sec:boundary}).
For black holes, these $N-1$ parameters are simply the values of the gauge field functions on the horizon $\omega _{j}(r_{h})$ (\ref{eq:horizon}).
For soliton solutions the situation is more complicated, details of the parameters in this case can be found in Refs.~\onlinecite{Baxter2,Baxter3}.
In Ref.~\onlinecite{Baxter3} the existence of genuinely ${\mathfrak {su}}(N)$ solutions of the static field equations (\ref{eq:Ee}, \ref{eq:YMe}) in a neighbourhood of the above embedded ${\mathfrak {su}}(2)$ solutions was proven for all $N$, for sufficiently large $\left| \Lambda \right| $.
In this article we focus on those ${\mathfrak {su}}(N)$ solutions which are close to embedded ${\mathfrak {su}}(2)$ solutions and for which all the gauge field
functions $\omega _{j}(r)$ have no zeros.
In Refs.~\onlinecite{Baxter2, Baxter1} we have presented various phase space plots for ${\mathfrak {su}}(3)$ and ${\mathfrak {su}}(4)$
which demonstrate numerically the existence of regions of these nodeless soliton and black hole solutions.
Here we simply plot, in figures \ref{fig:1} and \ref{fig:2}, examples of nodeless soliton and black hole solutions for ${\mathfrak {su}}(3)$ and
${\mathfrak {su}}(4)$ respectively, referring the reader to Ref.~\onlinecite{Baxter2} for further details of the phase space of solutions.
\begin{figure}
\includegraphics[width=7.5cm]{Figure1a.eps}
\includegraphics[width=7.5cm]{Figure1b.eps}
\caption{Example nodeless solutions for ${\mathfrak {su}}(3)$ EYM with $\Lambda = -10$. In each case we plot the gauge field functions $\omega _{1}(r)$
and $\omega _{2}(r)$ (the typical behaviour of the metric functions can be found in the example solutions plotted in Ref.~\onlinecite{Baxter2}).
In (a) we show a soliton solution, and in (b) a black hole solution with $r_{h}=1$.}
\label{fig:1}
\end{figure}
\begin{figure}
\includegraphics[width=7.5cm]{Figure2a.eps}
\includegraphics[width=7.5cm]{Figure2b.eps}
\caption{Example nodeless solutions for ${\mathfrak {su}}(4)$ EYM with $\Lambda = -10$. In each case we plot the gauge field functions $\omega _{1}(r)$, $\omega _{2}(r)$ and $\omega _{3}(r)$ (the typical behaviour of the metric functions can be found in the example solutions plotted in Ref.~\onlinecite{Baxter2}).
In (a) we show a soliton solution, and in (b) a black hole solution with $r_{h}=1$.}
\label{fig:2}
\end{figure}
\subsection{Perturbation equations}
\label{sec:perturbation}
In this paper we are interested in linear, spherically symmetric, perturbations of the static equilibrium solutions discussed in section \ref{sec:static}.
Our particular interest is in time-periodic, bound state, perturbations which vanish at either the origin or event horizon (as applicable, for soliton and
black hole solutions respectively) and at infinity.
To derive the perturbation equations, we write our time-dependent field variables as
a sum of the static equilibrium quantities (denoted by a bar, e.g.~${\ov {\mu }}(r)$) plus small perturbations
(denoted by a $\delta $, e.g.~$\delta \mu (t,r)$) as follows:
\begin{eqnarray}
\mu (t,r) & = & {\ov {\mu }}(r) + \delta \mu (t,r),
\nonumber \\
S(t,r) & = & {\ov {S}}(r) + \delta S(t,r),
\nonumber \\
m(t,r) & = & {\ov {m}}(r) + \delta m(t,r),
\nonumber \\
\Delta (t,r) & = & {\ov {\Delta }}(r) +\delta \Delta (t,r),
\nonumber \\
\alpha _{j} (t,r) & = & \delta \alpha _{j} (t,r),
\nonumber \\
\beta _{j}(t,r) & = & \delta \beta _{j}(t,r),
\nonumber \\
\omega _{j}(t,r) & = & {\ov {\omega }}_{j}(r) + \delta \omega _{j}(t,r),
\nonumber \\
\gamma _{j}(t,r) & = & \delta \gamma _{j}(t,r).
\label{eq:perturbations}
\end{eqnarray}
Recall from section \ref{sec:static} that the gauge field functions $\alpha _{j}$,
$\beta _{j}$ and $\gamma _{j}$ all vanish for static equilibrium solutions, but here we consider non-zero perturbations of these parts of the gauge potential.
The perturbation equations are found by substituting the field variables in the form (\ref{eq:perturbations}) into the components of the Einstein tensor (\ref{eq:EinsteinTensor}) and gauge field (\ref{eq:Fmunu}), and then working out the field equations (\ref{eq:fieldeqns}). We work only to first order in the perturbations and simplify the resulting equations using the static equilibrium field equations (\ref{eq:Ee}, \ref{eq:YMe}).
First of all, the linearized Einstein perturbation equations become
\begin{subequations}
\begin{eqnarray}
\delta \mu ' & = & \frac {1}{r} \left[
-\delta \mu - 2{\ov {\Gamma }}\delta \mu - 2{\ov {\mu }} \delta \Gamma
- 2r^{2} \delta \Pi
\right] ,
\label{eq:deltamup}
\\
\delta {\dot {\mu }} & = & -\frac {2{\ov {\mu }}}{r} \delta H ,
\label{eq:deltamudot}
\\
\delta \Delta ' & = & \frac {2}{r} \delta \Gamma ,
\label{eq:deltadeltap}
\end{eqnarray}
\end{subequations}
where
\begin{eqnarray}
\delta \Gamma & = &
2 \sum _{j=1}^{N-1} {\ov {\omega }}_{j}' \delta \omega _{j}',
\nonumber \\
\delta \Pi & = &
\frac {1}{r^{4}} \sum _{j=1}^{N} \left[
{\ov {\omega }}_{j}^{2} - {\ov {\omega }}_{j-1}^{2} - N - 1 +2j
\right]
\left[
{\ov {\omega }}_{j} \delta \omega _{j} - {\ov {\omega }}_{j-1} \delta \omega _{j-1}
\right],
\nonumber \\
\delta H & = & \sum _{j=1}^{N-1} 2{\ov {\omega }}_{j}'\delta {\dot {\omega }}_{j},
\label{eq:deltaPi}
\end{eqnarray}
and we remind the reader that we are using a dot to denote $\partial /\partial t$
and a prime to denote $\partial /\partial r$.
Given the form of the elements of the matrix $C$ (\ref{eq:Celements}), it is useful to consider the following combinations of the perturbations
$\delta \omega _{j}(t,r)$
and $\delta \gamma _{j}(t,r)$, for $j=1,\ldots N-1$:
\begin{eqnarray}
\delta \psi _{j} (t,r) & = & \delta \omega _{j}(t,r) + i{\ov {\omega }}_{j}(r)
\delta \gamma _{j}(t,r) ,
\nonumber \\
\delta \psi _{j}^{*} (t,r) & = & \delta \omega _{j}(t,r) - i{\ov {\omega }}_{j}(r)
\delta \gamma _{j}(t,r) ,
\end{eqnarray}
in terms of which the entries of the matrix $C$ are, to first order in the perturbations:
\begin{equation}
C_{j,j+1} = {\ov {\omega }}_{j}(r) + \delta \psi _{j}(t,r).
\end{equation}
In terms of $\delta \psi _{j}$, $\delta \psi _{j}^{*}$, the linearized
Yang-Mills perturbation equations are:
\begin{widetext}
\begin{subequations}
\begin{eqnarray}
0 & = &
{\ov {\mu }}\left[
\delta {\dot{\beta}}_{j}'-\delta \alpha_j''
+\left( \delta{\dot{\beta}}_{j}-\delta \alpha_j' \right)
\left( \frac{2}{r}-\frac{{\ov {S}}'}{{\ov {S}}}\right)
\right]
+
\frac{1}{2r^2}
\left[{\ov {\omega}}_{j}
\left( \delta{\dot {\psi }}_j-\delta {\dot {\psi}}_j^* \right)
\right. \nonumber \\ & & \left.
-{\ov {\omega }}_{j-1}
\left( \delta {\dot {\psi }}_{j-1}-\delta {\dot {\psi }}_{j-1}^* \right)
+ 2{\ov {\omega }}_j^{2}
\left( \delta\alpha_j-\delta\alpha_{j+1} \right)
-2{\ov {\omega }}_{j-1}^{2}
\left( \delta\alpha_{j-1}-\delta\alpha_j \right) \right] ,
\label{eq:tYMpert}
\\
0 & = &
\frac{1}{{\ov {\mu }} {\ov {S}}^2}
\left(\delta {\ddot {\beta }}_j -\delta {\dot {\alpha }}_j' \right)
+\frac{1}{2r^2}
\left[
{\ov {\omega }}_j \left( \delta\psi_j' -\delta\psi_j^{*'} \right)
-{\ov {\omega}}_{j-1} \left( \delta\psi_{j-1}'-\delta\psi_{j-1}^{*'} \right)
-{\ov {\omega }}_j' \left(\delta\psi_j-\delta\psi_j^* \right)
\right. \nonumber \\ & & \left.
+{\ov {\omega }}_{j-1}' \left( \delta\psi_{j-1}-\delta\psi_{j-1}^* \right)
+2{\ov {\omega }}_{j-1}^{2} \left( \delta\beta_j-\delta\beta_{j-1} \right)
-2{\ov {\omega }}_j^{2}\left(\delta\beta_{j+1}-\delta\beta_j \right)
\right],
\label{eq:rYMpert}
\\
0 & = &
-\frac{1}{2{\ov {\mu }} {\ov {S}}^2}
\left[ \delta {\ddot {\psi }}_{j}
+{\ov {\omega }}_j \left( \delta {\dot {\alpha }}_j
-\delta {\dot {\alpha}}_{j+1} \right) \right]
+\frac {1}{2}{\ov {\omega }}_{j}'' \delta \mu
+\frac{{\ov {\mu }}}{2} \left[
\delta\psi_j''
+{\ov {\omega }}_j \left( \delta\beta_j'-\delta\beta_{j+1}' \right)
\right. \nonumber \\ & & \left.
+2{\ov {\omega}}_j' \left( \delta\beta_j-\delta\beta_{j+1} \right) \right]
+ \frac{{\ov {\omega }}_{j}'}{2}\left[
\delta\mu'+{\ov {\mu }} \delta \left(\frac{S'}{S}\right)
+ \frac{{\ov {S}}'}{{\ov {S}}}\delta\mu \right]
+ \frac {{\ov {\mu }}'}{2} \left[
\delta\psi'_j+{\ov {\omega }}_j \left(
\delta\beta_j-\delta\beta_{j+1} \right) \right]
\nonumber \\ & &
+\frac{1}{2r^2} \left[
-{\ov {\omega }}_j^{2} \left(\delta\psi_j +\delta\psi_j^* \right)
+\frac{1}{2}{\ov {\omega }}_{j}{\ov {\omega }}_{j+1}
\left( \delta\psi_{j+1}+\delta\psi_{j+1}^* \right)
+\frac{1}{2}{\ov {\omega }}_{j} {\ov {\omega }}_{j-1}
\left( \delta\psi_{j-1}+\delta\psi_{j-1}^* \right)
\right. \nonumber \\ & & \left.
+W_{j} \delta\psi_j
\right] ,
\label{eq:qrealYMpert}
\\
0 & = &
-\frac{1}{2{\ov {\mu }} {\ov {S}}^2}
\left[
-\delta {\ddot {\psi }}_j^* +{\ov {\omega }}_j
\left( \delta {\dot {\alpha }}_j-\delta {\dot {\alpha }}_{j+1} \right)
\right]
-\frac {1}{2}{\ov {\omega }}_{j}'' \delta \mu
+\frac{{\ov {\mu }}}{2} \left[
-\delta\psi_j^{*''}
+{\ov {\omega }}_j \left( \delta\beta_j'-\delta\beta_{j+1}' \right)
\right. \nonumber \\ & & \left.
+2{\ov {\omega }}_j' \left( \delta\beta_j-\delta\beta_{j+1} \right) \right]
-\frac{{\ov {\omega }}_{j}'}{2} \left[
\delta\mu'+{\ov {\mu }} \delta \left(\frac{S'}{S}\right)
+\frac{{\ov {S}}'}{{\ov {S}}} \delta\mu
\right]
+ \frac {{\ov {\mu }}'}{2} \left[
-\delta\psi_j^{*'}+{\ov {\omega }}_j \left( \delta\beta_j-\delta\beta_{j+1} \right) \right]
\nonumber \\ & &
+\frac{1}{2r^2}\left[
{\ov {\omega }}_j^{2} \left( \delta\psi_j+\delta\psi_j^* \right)
-\frac{1}{2}{\ov {\omega }}_{j} {\ov {\omega }}_{j+1}
\left( \delta\psi_{j+1}+\delta\psi_{j+1}^* \right)
-\frac{1}{2}{\ov {\omega }}_{j} {\ov {\omega }}_{j-1}
\left( \delta\psi_{j-1}+\delta\psi_{j-1}^* \right)
\right. \nonumber \\ & & \left.
- W_{j} \delta\psi_j^*
\right] .
\label{eq:qimYMpert}
\end{eqnarray}
\end{subequations}
\end{widetext}
The equations (\ref{eq:tYMpert}--\ref{eq:rYMpert}) come from the $t$ and $r$ Yang-Mills equations, respectively, and there are $N$ of each of these equations, corresponding to $j=1,\ldots N$.
The equations (\ref{eq:qrealYMpert}--\ref{eq:qimYMpert}) come from taking the real and imaginary parts of the $\theta $ Yang-Mills equation (the $\phi $ Yang-Mills equation gives the same pair of equations), assuming that all perturbations are real, and there are $N-1$ of each of these equations, corresponding to $j=1,\ldots N-1$.
Our time-dependent, spherically symmetric gauge field ansatz (\ref{eq:gaugepot}) has a residual gauge degree of freedom. For a diagonal matrix ${\mathfrak {g}}(t,r)$, consider the following gauge transformation:
\begin{eqnarray}
{\mathcal {A}} & \rightarrow & {\mathcal {A}}+{\mathfrak {g}}^{-1}
{\dot{{\mathfrak {g}}}} ,
\nonumber \\
{\mathcal {B}} & \rightarrow & {\mathcal {B}}
+ {\mathfrak {g}}^{-1} {\mathfrak {g}}' ,
\nonumber \\
C-C^H & \rightarrow & {\mathfrak {g}}^{-1}\left( C-C^H \right) {\mathfrak {g}} ,
\nonumber \\
C+C^H & \rightarrow &
{\mathfrak {g}}^{-1}\left( C+C^H \right) {\mathfrak {g}} ,
\label{eq:gaugetrans}
\end{eqnarray}
under which the gauge field transforms as
\begin{equation}
F_{\mu \nu } \rightarrow {\mathfrak {g}}^{-1} F_{\mu \nu } {\mathfrak {g}}.
\end{equation}
We choose the diagonal matrix ${\mathfrak {g}}$ so that
${\mathcal {A}}+{\mathfrak {g}}^{-1} {\dot{{\mathfrak {g}}}} =0 $, which enables us to set the perturbations $\delta \alpha _{j}(t,r) \equiv 0$ for all $j=1,\ldots N$.
With this choice of gauge, the perturbation equations
(\ref{eq:deltamup}--\ref{eq:deltadeltap}, \ref{eq:tYMpert}--\ref{eq:qimYMpert})
decouple into two sectors.
The first sector contains the Yang-Mills perturbations $\delta \beta _{j}$,
$j=1,\ldots , N$ and $\delta \gamma _{j}$, $j=1,\ldots , N-1$ and does not contain any metric perturbations.
This sector is known as the {\em {sphaleronic}} sector \cite{Lavrelashvili1}.
This terminology arises from the fact that the ${\mathfrak {su}}(2)$ EYM solitons \cite{Bartnik1} and black holes \cite{Bizon1} in asymptotically flat space
possess instabilities in this sector \cite{Lavrelashvili1,Volkov2,sphalstab} analogous to the unstable mode of the Yang-Mills-Higgs sphaleron \cite{sphaleron}.
The second sector contains the perturbations of the metric functions $\delta \mu $ and $\delta \Delta $ and the Yang-Mills perturbations $\delta \omega _{j}$,
$j=1,\ldots , N-1$.
This sector is known as the {\em {gravitational}} sector.
As the static equilibrium solutions are purely magnetic and spherically symmetric, they are invariant under a parity transformation. As a result of this additional symmetry, the two decoupled sectors of perturbations transform in a particular way under a parity transformation: the perturbations in the sphaleronic sector have odd parity and change sign under a parity transformation; the perturbations in the gravitational sector have even parity and do not change under a parity transformation.
In the analysis of the sphaleronic and gravitational perturbation sectors in sections \ref{sec:sphaleronic} and \ref{sec:gravitational} respectively, we will change our independent radial variable to the usual `tortoise' co-ordinate $r_{*}$, defined by
\begin{equation}
\frac {dr_{*}}{dr} = \frac {1}{{\ov {\mu }}{\ov {S}}}.
\label{eq:tortoise}
\end{equation}
For perturbations of static soliton solutions, we choose the constant of integration
such that $r_{*}=0$ at the origin where $r=0$. In this case $r_{*}$ has a maximum value, $r_{c}$, as $r\rightarrow \infty $.
For perturbations of static black hole solutions, we choose the constant of integration such that $r_{*} \rightarrow 0 $ as $r\rightarrow \infty $, and then $r_{*}\rightarrow -\infty $ as the event horizon is approached, $r\rightarrow r_{h}$.
\section{Sphaleronic sector perturbations}
\label{sec:sphaleronic}
The sphaleronic sector consists of the odd parity Yang-Mills perturbations
$\delta \beta _{j}$ ($j=1,\ldots , N$) and $\delta \gamma _{j}$ ($j=1,\ldots , N-1$).
In the gauge $\delta \alpha _{j} \equiv 0$, $j=1,\ldots , N$, the sphaleronic sector perturbation equations are (\ref{eq:tYMpert}, \ref{eq:rYMpert}) and a third perturbation equation which comes from adding equations (\ref{eq:qrealYMpert}) and
(\ref{eq:qimYMpert}). The equations are:
\begin{subequations}
\begin{eqnarray}
0 & = & {\ov {\mu }} \left[
\delta {\dot{\beta }}_j' + \delta {\dot{\beta }}_j
\left(
\frac{2}{r}-\frac{{\ov {S}}'}{{\ov {S}}}
\right) \right]
+\frac{1}{r^2} \left[
{\ov {\omega }}_j^{2} \delta {\dot{\gamma }}_{j}
-{\ov {\omega }}_{j-1}^{2} \delta {\dot {\gamma}}_{j-1}
\right] ,
\label{eq:sphalsector1}
\\
0 & = &
-\frac{1}{{\ov {\mu}} {\ov {S}}^2}
\delta {\ddot {\beta }}_j
+\frac{1}{r^2} \left[
-{\ov {\omega }}_j^{2} \delta\gamma_j'+{\ov {\omega }}_{j-1}^{2}\delta\gamma_{j-1}'
-{\ov {\omega }}_{j-1}^{2}
\left( \delta\beta_j-\delta\beta_{j-1} \right)
+{\ov {\omega }}_j^{2}\left( \delta\beta_{j+1}-\delta\beta_j \right) \right] ,
\nonumber \\ & &
\label{eq:sphalsector2}
\\
0 & = & -\frac{1}{{\ov {\mu }} {\ov {S}}^2}{\ov {\omega }}_j
\delta {\ddot {\gamma }}_j +{\ov {\mu }}{\ov {\omega }}_j\delta\gamma_j''
+{\ov {\mu}}{\ov {\omega }}_{j} \left( \delta\beta_j'-\delta\beta_{j+1}' \right)
\nonumber \\ & &
+\left[
2{\ov {\mu }}{\ov {\omega }}_j' +{\ov {\mu }}'{\ov {\omega }}_j
+{\ov {\mu }}{\ov {\omega }}_j \frac{{\ov {S}}'}{{\ov {S}}}\right]
\left[
\delta\gamma_j'+\delta\beta_j-\delta\beta_{j+1}\right] .
\label{eq:sphalsector3}
\end{eqnarray}
\end{subequations}
\subsection{Sphaleronic sector perturbation equations in matrix form}
\label{sec:smatrix}
We first simplify these equations by changing our radial co-ordinate from $r$ to the tortoise co-ordinate $r_{*}$ (\ref{eq:tortoise}) and by introducing new dependent
variables $\delta \epsilon _{j}$ ($j=1,\ldots , N$) and $\delta \Phi _{j}$ ($j=1,\ldots , N-1$) by:
\begin{equation}
\delta \epsilon _{j} = r{\sqrt {\ov {\mu }}} \, \delta \beta _{j},
\qquad
\delta \Phi _{j} = {\ov {\omega }}_{j} \delta \gamma _{j}.
\label{eq:newsvars}
\end{equation}
The perturbation equations (\ref{eq:sphalsector1}--\ref{eq:sphalsector3}) then take the form
\begin{subequations}
\begin{eqnarray}
0 & = &
\drs \delta {\dot {\epsilon }}_j
+
\left(
\frac{2{\ov {\mu }} {\ov {S}}}{r}
-\frac{\drs {\ov {S}}}{{\ov {S}}}
\right)
\delta {\dot {\epsilon }}_j
+\frac {{\ov {S}}{\sqrt {\ov {\mu }}}}{r}
\left(
{\ov {\omega }}_j \delta {\dot {\Phi }}_j
-{\ov {\omega }}_{j-1} \delta {\dot {\Phi}}_{j-1}
\right) ,
\label{eq:gaussc1}
\\
\delta {\ddot {\epsilon }}_j & = &
h
\left[
-{\ov {\omega }}_j \drs \delta \Phi_j
+ {\ov {\omega }}_{j-1} \drs \delta \Phi_{j-1}
+ \left( \drs {\ov {\omega }}_j \right) \delta \Phi_j
-\left( \drs {\ov {\omega }}_{j-1} \right)
\delta\Phi_{j-1}
\right]
\nonumber \\ & &
+h^{2}
\left[
{\ov {\omega }}_j^{2} \left( \delta \epsilon _{j+1}- \delta \epsilon _j \right)
-{\ov {\omega }}_{j-1}^{2} \left( \delta \epsilon _j - \delta \epsilon _{j-1} \right)
\right],
\\
\delta {\ddot {\Phi }}_j & = &
\drs ^{2}\delta \Phi_j
-\frac{\drs ^{2}{\ov {\omega }}_j}{{\ov {\omega }}_{j}} \delta \Phi _{j}
+ h{\ov {\omega }}_{j}
\drs \left( \delta \epsilon _j-\delta \epsilon _ {j+1} \right)
+
\left[ \drs \left( h {\ov {\omega }}_{j} \right) + h\drs {\ov {\omega }}_{j} \right]
\left( \delta \epsilon _j - \delta \epsilon _{j+1} \right) ,
\nonumber \\ & &
\label{eq:sphsectorint3}
\end{eqnarray}
\end{subequations}
where we have introduced the quantity
\begin{equation}
h = \frac {{\ov {S}}{\sqrt {\ov {\mu }}}}{r}.
\end{equation}
We now express the perturbation equations in matrix form by defining
\begin{eqnarray}
{\bmath {\delta \epsilon }} & = & \left( \delta \epsilon _{1},\ldots ,
\delta \epsilon _{N}
\right) ^{T} ,
\nonumber \\
{\bmath {\delta \Phi }} & = & \left( \delta \Phi _{1},\ldots , \delta \Phi _{N-1}
\right) ^{T},
\end{eqnarray}
in terms of which the perturbation equations (\ref{eq:gaussc1}--\ref{eq:sphsectorint3}) take the form
\begin{subequations}
\begin{eqnarray}
0 & = &
\drs \left( h^{-1} {\bmath {\delta }}{\dot {\bmath {\epsilon }}} \right)
+ {\mathcal {F}} {\bmath {\delta }}{\dot {\bmath {\Phi }}} ,
\label{eq:gaussc2}
\\
{\bmath {\delta }} {\ddot {{\bmath {\epsilon }}}} & = &
h^{2} {\mathcal {K}} {\bmath {\delta \epsilon }}
- h \left[
{\mathcal {F}} \drs {\bmath {\delta \Phi }} - \left( \drs {\mathcal {F}} \right)
{\bmath {\delta \Phi }} \right] ,
\label{eq:sphsectorvector1}
\\
{\bmath {\delta }}{\ddot {{\bmath {\Phi }}}}
& = &
\drs ^{2} {\bmath {\delta \Phi }}
+ h {\mathcal {F}}^{T} \drs {\bmath {\delta \epsilon }}
+ {\mathcal {X}} {\bmath {\delta \epsilon }}
+ {\mathcal {W}} {\bmath {\delta \Phi }} ,
\label{eq:sphsectorvector2}
\end{eqnarray}
\end{subequations}
where we have defined an $N\times \left( N -1 \right) $ matrix ${\mathcal {F}}$,
an $N \times N$ matrix ${\mathcal {K}}$, an $\left( N -1 \right) \times \left(
N-1 \right)$ matrix ${\mathcal {W}}$ and an $\left( N-1 \right) \times N$ matrix
${\mathcal {X}}$ as follows:
\begin{subequations}
\begin{eqnarray}
{\mathcal {F}} & = &
\left(
\begin{array}{ccccc}
{\ov {\omega }}_1&0&0&\cdots&0\\
-{\ov {\omega}}_1&{\ov {\omega }}_2&0&\cdots&0\\
0&-{\ov {\omega }}_2&{\ov {\omega }}_3&\cdots&0\\
\vdots&\vdots&\vdots&\ddots&{\ov {\omega }}_{N-1}\\
0&0&0&0&-{\ov {\omega }}_{N-1}\\
\end{array}
\right) ,
\\
{\mathcal {K}} & = &
\left(
\begin{array}{ccccc}
-{\ov {\omega }}_1^{2}& {\ov {\omega }}_1^{2}&0&\cdots&0\\
{\ov {\omega }}_1^{2}&-{\ov {\omega }}_1^{2}-{\ov {\omega }}_2^{2}
& {\ov {\omega }}_2^{2}&\cdots&0\\
0&{\ov {\omega }}_2^{2}&-{\ov {\omega }}_2^{2}
-{\ov {\omega }}_3^{2}&\cdots&0\\
\vdots&\vdots&\vdots&\ddots&{\ov {\omega }}_{N-1}^{2}\\
0&0&0&{\ov {\omega }}_{N-1}^{2}&- {\ov {\omega }}_{N-1}^{2}\\
\end{array}
\right) ,
\nonumber \\ & &
\label{eq:Kdef}
\\
{\mathcal {W}} & = &
h^{2}
\,
\Diag \left(
W_{1}, \ldots , W_{N-1} \right) ,
\label{eq:Wmatrixdef}
\\
{\mathcal {X}} & = &
2h\drs {\mathcal {F}}^{T}
+ \left( \drs h \right) {\mathcal {F}}^{T},
\end{eqnarray}
\end{subequations}
where the quantities $W_{j}$, $j=1,\ldots , N-1$ are given by (\ref{eq:Wdef}).
Finally, we introduce a vector ${\bmath {\Psi }}$ of dimension $2N-1$ by
\begin{equation}
{\bmath {\Psi }} = \left(
\begin{array}{c}
{\bmath {\delta \epsilon }} \\
{\bmath {\delta \Phi }}
\end{array}
\right) ,
\end{equation}
in terms of which the first perturbation equation (\ref{eq:gaussc2}) takes the form
\begin{equation}
{\mathcal {G}} {\dot {{\bmath {\Psi }}}} \equiv
\drs
\left[ h^{-1}
\left(
\begin{array}{cc}
{\mathcal {I}}_{N} & 0 \\
0 & 0
\end{array}
\right)
{\dot {\bmath {\Psi }}}
\right]
+ \left(
\begin{array}{cc}
0 & {\mathcal {F}} \\
0 & 0
\end{array}
\right)
{\dot {{\bmath {\Psi }}}}
= 0,
\label{eq:gausscfinal}
\end{equation}
and the remaining equations (\ref{eq:sphsectorvector1}--\ref{eq:sphsectorvector2}) can be compactly written as
\begin{equation}
-{\ddot {{\bmath {\Psi }}}} =
{\mathcal {U}} {\bmath {\Psi }},
\label{eq:sphalpertfinal}
\end{equation}
where we have defined the operator
\begin{eqnarray}
{\mathcal {U}} {\bmath {\Psi }} & \equiv &
-\left(
\begin{array}{cc}
0 & 0 \\
0 & {\mathcal {I}}_{N-1}
\end{array}
\right)
\drs ^{2} {\bmath {\Psi }}
- h \left(
\begin{array}{cc}
0 & -{\mathcal {F}} \\
{\mathcal {F}}^{T} & 0
\end{array}
\right)
\drs {\bmath {\Psi }}
- \left(
\begin{array}{cc}
h^{2}{\mathcal {K}} &
h \drs {\mathcal {F}} \\
{\mathcal {X}} & {\mathcal {W}}
\end{array}
\right)
{\bmath {\Psi }}
\label{eq:Udef}
\end{eqnarray}
and ${\mathcal {I}}_{n}$ denotes the $n\times n$ identity matrix.
It is straightforward to show that the operator ${\mathcal {U}}$ (\ref{eq:Udef}) is
real and symmetric when acting on perturbations which vanish at either the origin or event horizon (as applicable) and at infinity.
However, as noted in Ref.~\onlinecite{Brodbeck}, the operator ${\mathcal {U}}$ is not elliptic and so the perturbation equation (\ref{eq:sphalpertfinal}) is not currently in hyperbolic form.
For time-periodic perturbations for which
${\bmath {\Psi }}(t,r)= e^{i\sigma t}{\bmath {\Psi }}(r)$, the perturbation equations (\ref{eq:sphalpertfinal}) take the form
\begin{equation}
\sigma ^{2}{\bmath {\Psi }} = {\mathcal {U}} {\bmath {\Psi }}.
\label{eq:sphalperiodic}
\end{equation}
If we can show that the operator ${\mathcal {U}}$ is a positive operator, then the eigenvalues $\sigma ^{2}$ must also be positive.
This means that $\sigma $ is real and the perturbations are periodic in time. In this case small perturbations remain small and there are
no unstable modes in the sphaleronic sector.
Our aim for the remainder of this section will be to show that there are at least some equilibrium ${\mathfrak {su}}(N)$ soliton and black
hole solutions for which ${\mathcal {U}}$ is a positive operator.
\subsection{The Gauss constraint}
\label{sec:Gaussc}
The first of the linearized Yang-Mills perturbation equations (\ref{eq:gausscfinal}) is known as the {\em {Gauss constraint}}.
A lengthy calculation reveals that the Gauss constraint propagates, in other words the perturbation equations (\ref{eq:sphsectorvector1}--\ref{eq:sphsectorvector2})
imply that
\begin{equation}
{\mathcal {G}} {\ddot {\bmath {\Psi }}} =0
\end{equation}
independently of the Gauss constraint.
Equivalently, we may write \cite{Brodbeck}
\begin{equation}
{\mathcal {G}}{\mathcal {U}} =0.
\label{eq:gaussprop}
\end{equation}
Following Ref.~\onlinecite{Brodbeck}, we integrate (\ref{eq:gausscfinal}) with respect to time, and choose the constant of integration (in this case an arbitrary function of $r$) so that:
\begin{equation}
{\mathcal {G}} {\bmath {\Psi }} =0 ,
\label{eq:gausscstrong}
\end{equation}
which we will call the {\em {strong Gauss constraint}} \cite{Brodbeck}.
Suppose we have a vector of perturbations ${\bmath {\Psi }}$ which satisfy the strong Gauss constraint at initial time $t=0$, and which initially satisfy the Gauss constraint (\ref{eq:gausscfinal}).
By virtue of (\ref{eq:gaussprop}), this vector of perturbations will satisfy the strong Gauss constraint at all subsequent times.
Now consider any vector of perturbations ${\bmath {\Psi }}$ (satisfying the perturbation equations) and write it as the sum of two parts: the first, ${\bmath {\Psi }}_{1}$, satisfying the strong Gauss constraint and the second, ${\bmath {\Psi }}_{2}$, failing to satisfy the strong Gauss constraint.
It is shown in Ref.~\onlinecite{Brodbeck} that the second vector of perturbations, ${\bmath {\Psi }}_{2}$,
is pure gauge, having the form
\begin{equation}
{\bmath {\Psi }}_{2} = {\mathcal {G}}^{\dagger } {\bmath {\Upsilon}}
\end{equation}
where
\begin{equation}
{\mathcal {G}}^{\dagger }= -h^{-1}\left(
\begin{array}{cc}
{\mathcal {I}}_{N} & 0 \\
0 & 0
\end{array}
\right)
\drs +
\left(
\begin{array}{cc}
0 & 0 \\
{\mathcal {F}}^{T} & 0
\end{array}
\right)
\end{equation}
is the adjoint of the operator ${\mathcal {G}}$ (\ref{eq:gausscfinal}).
Such perturbations correspond to infinitesimal gauge transformations of the form (\ref{eq:gaugetrans}) with
(for small $\varepsilon $)
\begin{equation}
{\mathfrak {g}}=\exp \left( - \varepsilon {\tilde {{\bmath {\Upsilon}}}} \right)
\label{eq:gSGC}
\end{equation}
where ${\tilde {{\bmath {\Upsilon }}}}$ is an $N\times N$ matrix of the form
\begin{equation}
{\tilde {\bmath {\Upsilon}}} = \Diag
\left(
\Upsilon _{1}, \ldots , \Upsilon _{N}
\right)
\end{equation}
with $\Upsilon _{j}$, $j=1,\ldots, N$ the first $N$ elements in
${\bmath {\Upsilon }}$.
Therefore a vector of perturbations ${\bmath {\Psi }}$ which satisfy the Gauss constraint but not the strong Gauss constraint can be gauge-transformed
to a vector of perturbations satisfying the strong Gauss constraint.
Without loss of generality, we may therefore restrict attention to physical perturbations satisfying the strong Gauss constraint, which is essentially an initial condition.
Since this is a gauge transformation of initial data only, the matrix ${\mathfrak {g}}$ (\ref{eq:gSGC}) is time-independent, and so, by (\ref{eq:gaugetrans}),
this transformation preserves the gauge condition $\delta \alpha _{j}\equiv 0$.
\subsection{An alternative form of the operator ${\mathcal {U}}$}
\label{sec:altU}
In order to prove the existence of static solutions which have no unstable modes in the sphaleronic sector governed by the equations (\ref{eq:sphalpertfinal}), in the next subsection we will want to show that the symmetric operator ${\mathcal {U}}$ (\ref{eq:Udef}) is positive.
In this subsection we will write the operator in an alternative form which will enable us to find static equilibrium solutions for which
${\mathcal {U}}$ is a positive operator.
In particular, following Ref.~\onlinecite{Brodbeck}, we seek operators $\chi $ and ${\mathcal {V}}$ such that we may write
\begin{equation}
{\mathcal {U}} = \chi ^{\dagger } \chi + {\mathcal {V}}
- {\mathcal {G}}^{\dagger } h^{2} {\mathcal {G}},
\label{eq:Vdef}
\end{equation}
where the operator ${\mathcal {G}}$ is given by (\ref{eq:gausscfinal}).
We then find
\begin{eqnarray}
{\mathcal {G}}^{\dagger }h^{2} {\mathcal {G}}
& = &
-\left(
\begin{array}{cc}
{\mathcal {I}}_{N} & 0 \\
0 & 0
\end{array}
\right)
\drs ^{2}
+ h\left(
\begin{array}{cc}
0 & -{\mathcal {F}} \\
{\mathcal {F}}^{T} & 0
\end{array}
\right)
+
\left(
\begin{array}{cc}
h^{-1} \drs ^{2} h &
-\drs \left( h{\mathcal {F}} \right) - \left( \drs h \right) {\mathcal {F}}
\\
-\left( \drs h \right) {\mathcal {F}}^{T} &
h^{2} {\mathcal {F}}^{T} {\mathcal {F}}
\end{array}
\right) .
\nonumber \\ & &
\label{eq:GdaggerG}
\end{eqnarray}
Next define the operator $\chi $ as
\begin{equation}
\chi = \drs + h{\mathcal {Z}} ,
\label{eq:chidef}
\end{equation}
where ${\mathcal {Z}}$ is some $\left( 2N-1 \right) \times \left( 2N-1 \right) $ matrix which does not contain any derivative operators and which is to be determined.
Then
\begin{equation}
\chi ^{\dagger } \chi
= -\drs ^{2} + h\left( {\mathcal {Z}}^{T} - {\mathcal {Z}} \right) \drs
+\left[ h^{2} {\mathcal {Z}}^{T} {\mathcal {Z}} - \drs
\left( h{\mathcal {Z}} \right) \right] .
\label{eq:chidaggerchi}
\end{equation}
From now on we assume that ${\mathcal {Z}}$ is symmetric, so that there is no
first order derivative operator in $\chi ^{\dagger }\chi $.
Writing the matrix ${\mathcal {Z}}$ in the form
\begin{equation}
{\mathcal {Z}} = \left(
\begin{array}{cc}
{\mathcal {Z}}_{11} & {\mathcal {Z}}_{12} \\
{\mathcal {Z}}_{12}^{T} & {\mathcal {Z}}_{22}
\end{array}
\right) ,
\end{equation}
where ${\mathcal {Z}}_{11}$ is a symmetric $N \times N $ matrix,
${\mathcal {Z}}_{12}$ is an $N\times \left( N -1 \right)$ matrix and
${\mathcal {Z}}_{22}$ is a symmetric $\left( N-1 \right) \times \left( N-1 \right) $
matrix, using (\ref{eq:GdaggerG}, \ref{eq:chidaggerchi}) we find that the matrix
${\mathcal {V}}$ defined in (\ref{eq:Vdef}) has the form
\begin{equation}
{\mathcal {V}} = \left(
\begin{array}{cc}
{\mathcal {V}}_{11} & {\mathcal {V}}_{12} \\
{\mathcal {V}}_{21} & {\mathcal {V}}_{22}
\end{array}
\right) ,
\end{equation}
where
\begin{eqnarray}
{\mathcal {V}}_{11} & = &
-h^{2} {\mathcal {K}} + h^{2} \left( {\mathcal {Z}}_{11}^{T} {\mathcal {Z}}_{11} + {\mathcal {Z}}_{12} {\mathcal {Z}}_{12}^{T} \right)
-\drs \left( h{\mathcal {Z}}_{11} \right)
+ h^{-1} \left(
\drs ^{2} h \right) {\mathcal {I}}_{N} ,
\nonumber \\
{\mathcal {V}}_{12} & = &
-2h\drs {\mathcal {F}} + h^{2} \left( {\mathcal {Z}}_{11}^{T} {\mathcal {Z}}_{12}
+{\mathcal {Z}}_{12} {\mathcal {Z}}_{22}^{T} \right)
-\drs \left( h {\mathcal {Z}}_{12} \right)
-2\left( \drs h \right) {\mathcal {F}},
\nonumber \\
{\mathcal {V}}_{21} & = &
-2\left( \drs h \right) {\mathcal {F}}^{T} - 2h \drs {\mathcal {F}}^{T}
+h^{2}\left( {\mathcal {Z}}_{12}^{T} {\mathcal {Z}}_{11}
+ {\mathcal {Z}}_{22}^{T} {\mathcal {Z}}_{12}^{T} \right)
- \drs \left( h{\mathcal {Z}}_{12}^{T} \right) ,
\nonumber \\
{\mathcal {V}}_{22} & = &
-{\mathcal {W}} + h^{2} \left( {\mathcal {Z}}_{12}^{T} {\mathcal {Z}}_{12}
+ {\mathcal {Z}}_{22}^{T} {\mathcal {Z}}_{22} \right)
-\drs \left( h{\mathcal {Z}}_{22} \right)
+ h^{2} {\mathcal {F}}^{T} {\mathcal {F}}.
\end{eqnarray}
We are free to choose the matrix ${\mathcal {Z}}$ so as to simplify the form of
${\mathcal {V}}$.
We first make the choices
\begin{equation}
{\mathcal {Z}}_{11}=h^{-2} \left( \drs h \right) {\mathcal {I}}_{N},
\qquad
{\mathcal {Z}}_{22}=0.
\end{equation}
In this case the form of ${\mathcal {V}}_{12}$ simplifies to
\begin{equation}
{\mathcal {V}}_{12} =
-2\drs \left( h{\mathcal {F}} \right) - h\drs {\mathcal {Z}}_{12} ,
\end{equation}
which vanishes if we choose ${\mathcal {Z}}_{12}$ such that
\begin{equation}
{\mathcal {Z}}_{12} = -2 \int _{r_{*}=r_{*,{\mathrm {min}}}}^{r_{*}}
h^{-1} \drs \left( h{\mathcal {F}} \right) \, dr_{*} ,
\end{equation}
where $r_{*,{\mathrm {min}}}=0$ for equilibrium static soliton solutions and
$r_{*,{\mathrm {min}}}=-\infty $ for equilibrium static black hole solutions.
With this choice of ${\mathcal {Z}}_{12}$, it is straightforward to see that
${\mathcal {V}}_{21}$ also vanishes.
The matrix ${\mathcal {V}}$ is then block diagonal, with its diagonal entries being
\begin{eqnarray}
{\mathcal {V}}_{11} & = &
-h^{2} {\mathcal {K}} + h^{2} {\mathcal {Z}}_{12} {\mathcal {Z}}_{12}^{T}
+ \left( h^{-1} \drs h \right) ^{2} {\mathcal {I}}_{N},
\nonumber \\
{\mathcal {V}}_{22} & = &
-{\mathcal {W}} + h^{2} {\mathcal {Z}}_{12}^{T} {\mathcal {Z}}_{12}
+ h^{2} {\mathcal {F}}^{T} {\mathcal {F}}.
\label{eq:Ventries}
\end{eqnarray}
\subsection{Conditions for no instabilities in the sphaleronic sector}
\label{sec:sphalstabconds}
For physical perturbations satisfying the strong Gauss constraint (\ref{eq:gausscstrong}), the
operator ${\mathcal {U}}$ (\ref{eq:Vdef}) appearing in the sphaleronic sector perturbation equations (\ref{eq:sphalperiodic}) reduces to
\begin{equation}
{\mathcal {U}} = \chi ^{\dagger } \chi + {\mathcal {V}}.
\label{eq:Ureduced}
\end{equation}
Since the matrix ${\mathcal {V}}$ is symmetric, the operator ${\mathcal {U}}$ is symmetric and real.
From the form of the operator $\chi ^{\dagger} \chi $ (\ref{eq:chidaggerchi}), the operator (\ref{eq:Ureduced}) is elliptic.
Furthermore, since $\chi ^{\dagger }\chi $ is a positive operator, to show that ${\mathcal {U}}$ is a positive operator it suffices to show that ${\mathcal {V}}$ is a positive matrix.
The matrix ${\mathcal {V}}$ is block diagonal, and hence positive if its two non-zero diagonal blocks ${\mathcal {V}}_{11}$ and ${\mathcal {V}}_{22}$
(\ref{eq:Ventries}) are positive.
Let us begin with ${\mathcal {V}}_{11}$. The second and third terms in
${\mathcal {V}}_{11}$ are manifestly positive, so it remains to consider the term $-h^{2} {\mathcal {K}}$ where the matrix ${\mathcal {K}}$ is given by
(\ref{eq:Kdef}).
For an arbitrary vector ${\bmath {x}}=\left( x_{1},\ldots , x_{N} \right) ^{T}$, we have
\begin{eqnarray}
-{\bmath {x}}^{T} {\mathcal {K}} {\bmath {x}} & = &
{\ov {\omega }}_{1}^{2} x_{1}^{2}
+ \left( {\ov {\omega }}_{1}^{2} + {\ov {\omega }}_{2}^{2} \right) x_{2}^{2}
+ \ldots
+
\left( {\ov {\omega }}_{N-2}^{2} + {\ov {\omega }}_{N-1}^{2} \right) x_{N-1}^{2}
+ {\ov {\omega }}_{N-1}^{2} x_{N}^{2}
\nonumber \\ & &
- 2{\ov {\omega }}_{1}^{2} x_{1}x_{2} - 2{\ov {\omega }}_{2}^{2} x_{2}x_{3} - \ldots
- 2{\ov {\omega }}_{N-1}^{2} x_{N-1}x_{N}
\nonumber \\ & = &
{\ov {\omega }}_{1}^{2} \left( x_{1}-x_{2} \right) ^{2}
+ {\ov {\omega }}_{2}^{2} \left( x_{2}-x_{3} \right) ^{2}
+ \ldots
+
{\ov {\omega }}_{N-1}^{2} \left( x_{N-1}-x_{N} \right) ^{2}
\nonumber \\
& \ge & 0.
\end{eqnarray}
Therefore ${\mathcal {V}}_{11}$ is positive.
For ${\mathcal {V}}_{22}$, again the second and third terms are manifestly positive.
The first term, $-{\mathcal {W}}$ (\ref{eq:Wmatrixdef}) is a diagonal matrix, which will be positive if and only if its entries are positive.
For this to be the case, we require $W_{j}\le 0$ for $j=1,\ldots, N-1$, where
the quantities $W_{j}$ are defined in (\ref{eq:Wdef}).
This gives the following set of inequalities to be satisfied by the static equilibrium solutions for all $r$:
\begin{eqnarray}
{\ov {\omega }}_{1}^{2} & \ge &
1 + \frac {1}{2} {\ov {\omega }}_{2}^{2} ,
\nonumber \\
{\ov {\omega }}_{2}^{2} & \ge &
1 + \frac {1}{2} \left( {\ov {\omega }}_{1}^{2} + {\ov {\omega }}_{3}^{2} \right) ,
\nonumber \\
\vdots & &
\nonumber \\
{\ov {\omega }}_{j}^{2} & \ge &
1 + \frac {1}{2} \left( {\ov {\omega }}_{j-1}^{2} + {\ov {\omega }}_{j+1}^{2}
\right) ,
\nonumber \\
\vdots & &
\nonumber \\
{\ov {\omega }}_{N-1}^{2} & \ge &
1+\frac {1}{2} {\ov {\omega }}_{N-2}^{2}.
\label{eq:sstabineqs}
\end{eqnarray}
If the inequalities (\ref{eq:sstabineqs}) are satisfied for all $r$, then we can deduce that ${\mathcal {V}}_{22}$ is positive.
Therefore the matrix ${\mathcal {V}}$ is positive, and hence the operator ${\mathcal {U}}$ is a positive operator.
We deduce that physical solutions of the sphaleronic sector perturbation equations (\ref{eq:sphalperiodic}) must have $\sigma ^{2}$ positive,
so $\sigma $ is real and the perturbations are periodic in time.
Therefore small perturbations remain small and the equilibrium solutions have no instabilities in the sphaleronic sector.
We emphasize that the inequalities (\ref{eq:sstabineqs}) are sufficient for an equilibrium solution to have no unstable modes in the sphaleronic sector;
we have no expectation that these inequalities are necessary for stability.
Our interest in this paper is in proving the existence of stable soliton and black hole solutions of the ${\mathfrak {su}}(N)$ EYM equations.
Therefore we will have achieved this aim, at least for the sphaleronic sector, if we can find equilibrium solutions satisfying (\ref{eq:sstabineqs}) for all $r$.
\subsection{Special cases}
\label{sec:Sspecial}
Before proving the existence of non-trivial ${\mathfrak {su}}(N)$ equilibrium solutions which have no instabilities in the sphaleronic sector,
in this subsection we consider the embedded solutions discussed in section \ref{sec:embedded}.
\subsubsection{Schwarzschild-adS}
\label{sec:sSadS}
Setting ${\ov {\omega }}_{j} \equiv {\sqrt {j\left( N- j\right)}}$ and $m(r) \equiv M$,
we find that $W_{j}\equiv 0$ for all $j=1,\ldots , N-1$ and so the matrix
${\mathcal {W}}$ vanishes identically.
In this case the matrix ${\mathcal {V}}$ is manifestly positive and the operator
${\mathcal {U}}$ is positive when acting on physical perturbations.
Therefore the Schwarzschild-adS solution, as expected, has no instabilities in the sphaleronic sector.
\subsubsection{Reissner-Nordstr\"om-adS}
\label{sec:sRNadS}
If ${\ov {\omega }}_{j} \equiv 0$ for all $j=1,\ldots ,N-1$, then the sphaleronic sector perturbation equations reduce greatly. The perturbations $\delta \Phi _{j}$
(\ref{eq:newsvars}) vanish identically, leaving only the $\delta \epsilon _{j}$ perturbations. The only $\delta \epsilon _{j}$ perturbations which then satisfy the strong Gauss constraint have the form
\begin{equation}
\delta \epsilon _{j}= z_{j} h = z_{j} \frac {{\ov {S}}{\sqrt {\ov {\mu }}}}{r},
\end{equation}
where $z_{j}$ are arbitrary constants.
These perturbations do not vanish at the origin or the event horizon or infinity unless $z_{j}=0$ for all
$j=1,\ldots, N$.
For bound state perturbations which vanish at either the origin or the event horizon (as applicable) and at infinity, the only possibility is
$z_{j}=0$ and hence $\delta \epsilon _{j}\equiv 0$.
This means that there is no dynamics in the sphaleronic sector when the static equilibrium solution is embedded, magnetically-charged, Abelian
Reissner-Nordstr\"om-adS. The only allowed perturbations of the gauge potential correspond to gauge transformations.
\subsubsection{Embedded ${\mathfrak {su}}(2)$ solutions}
\label{sec:ssu2}
Setting ${\ov {\omega }}_{j} \equiv {\ov {\omega }}(r){\sqrt {j\left( N-j\right) }}$,
the sphaleronic sector perturbation equations simplify considerably.
In particular, the matrix ${\mathcal {W}}$ (\ref{eq:Wmatrixdef}) reduces to
\begin{equation}
{\mathcal {W}} = h^{2} \left( 1- {\ov {\omega }}^{2} \right) {\mathcal {I}}_{N-1}.
\end{equation}
The inequalities (\ref{eq:sstabineqs}) are then all satisfied if
\begin{equation}
{\ov {\omega }}(r)^{2} \ge 1
\label{eq:su2sstabineqs}
\end{equation}
for all $r$, in which case the operator ${\mathcal {U}}$ is positive and there are no instabilities in the sphaleronic sector.
It remains to prove the existence of embedded ${\mathfrak {su}}(2)$ solutions satisfying (\ref{eq:su2sstabineqs}) for all $r$.
Embedded ${\mathfrak {su}}(2)$ black holes are parameterized by the radius of the event horizon $r_{h}$, the cosmological constant $\Lambda $ and
${\ov {\omega }}(r_{h})$.
Fix $r_{h}$ and choose ${\ov {\omega }}(r_{h})>1$.
From the Yang-Mills equation (\ref{eq:YMe}) in the ${\mathfrak {su}}(2)$ case, we have ${\ov {\omega }}'(r_{h})>0$, so that ${\ov {\omega }}$ is an increasing
function of $r$ close to the horizon.
Embedded ${\mathfrak {su}}(2)$ solitons are described by a single parameter $b$, such that, near the origin,
\begin{equation}
{\ov {\omega }}(r) = 1 + br^{2} + O(r^{3}),
\end{equation}
where, without loss of generality, we are assuming that ${\ov {\omega }}(0)>0$.
Choose $b>0$, so that ${\ov {\omega }}(r)>1$ in a neighbourhood of the origin and ${\ov {\omega }}'(r)>0$ for $r$ sufficiently small.
Also from (\ref{eq:YMe}), we see that the gauge function ${\ov {\omega }}$ cannot have a maximum if ${\ov {\omega }}>1$.
Therefore ${\ov {\omega }}$ will be an increasing function of $r$ for all $r\ge r_{h}$ for our black hole solution and all $r>0$ for the soliton solution.
Therefore (\ref{eq:su2sstabineqs}) will be satisfied and these solutions will have no instabilities in the sphaleronic sector.
The same argument applies if ${\ov {\omega }}<-1$ either at the horizon or near the origin.
In Ref.~\onlinecite{Winstanley1} it is proven that ${\mathfrak {su}}(2)$ black holes have no instabilities in the sphaleronic sector of ${\mathfrak {su}}(2)$ EYM
perturbations as long as the gauge function $\omega (r)$ has no zeros, with no further conditions on $\omega (r)$.
While the proof in Ref.~\onlinecite{Winstanley1} is for black holes only, the argument carries over trivially to the soliton case.
We note that here, we have a stronger sufficient condition (\ref{eq:su2sstabineqs}) for ${\mathfrak {su}}(2)$ solutions embedded in ${\mathfrak {su}}(N)$
EYM to have no instabilities in the ${\mathfrak {su}}(N)$ sphaleronic sector.
This is to be expected since the ${\mathfrak {su}}(N)$ EYM sphaleronic sector has more degrees of freedom ($2N-2$, comprising
$N$ functions $\beta _{j}$ whose sum must vanish and $N-1$ functions $\gamma _{j}$) than the ${\mathfrak {su}}(2)$
EYM sphaleronic sector (which has just two).
\subsection{Existence of static solutions with no sphaleronic sector instabilities}
\label{sec:sphstab}
We now turn to proving the existence of non-trivial ${\mathfrak {su}}(N)$ EYM solitons and black holes having no instabilities in the sphaleronic sector.
From the above analysis, all that is required is to show the existence of equilibrium solutions satisfying the inequalities (\ref{eq:sstabineqs}).
The argument for both soliton and black hole solutions is straightforward, based on results from Ref.~\onlinecite{Baxter3}.
Due to the symmetry of the field equations (\ref{eq:Ee}, \ref{eq:YMe}) under the transformation (\ref{eq:omegatransform}) it is sufficient to
consider gauge field functions such that $\omega _{j}>0$ near either the origin if we are considering a soliton solution or the event horizon if we are considering a black hole solution.
First we define the open region ${\mathcal {R}}$ which is the set of all positive values of the equilibrium
gauge field functions ${\ov {\omega }}_{j}>0$, $j=1,\ldots ,N-1$
such that $W_{j}<0$ (so that the inequalities (\ref{eq:sstabineqs}) are strictly satisfied for all points in ${\mathcal {R}}$).
From the argument in the previous subsection, there are embedded ${\mathfrak {su}}(2)$ soliton and black hole solutions such that the values of the gauge
field functions lie in ${\mathcal {R}}$ for all $r$.
From Proposition 9 of Ref.~\onlinecite{Baxter3}, there are genuinely (that is, non-embedded) ${\mathfrak {su}}(N)$ soliton and black hole solutions whose initial parameters (either near the origin or the event horizon, see section \ref{sec:boundary} for details)
lie in a neighbourhood of the initial parameters for the embedded ${\mathfrak {su}}(2)$ solitons and black holes.
Propositions 3 and 6 of Ref.~\onlinecite{Baxter3} tell us that the equilibrium gauge field functions ${\ov {\omega }}_{j}$ are analytic functions of the initial parameters and the radial co-ordinate.
Fix $r_{h}$ for the black hole solutions under consideration and set $r_{1}\gg \max \{ 1,r_{h} \}$ (with $r_{h}=0$ for soliton solutions).
Then, by analyticity, providing our ${\mathfrak {su}}(N)$ solutions have initial parameters sufficiently close to the initial parameters for the embedded
${\mathfrak {su}}(2)$ solutions, the gauge field functions ${\ov {\omega }}_{j}(r)$ for the ${\mathfrak {su}}(N)$ solutions will remain close to the embedded ${\mathfrak {su}}(2)$ solutions for all $r\le r_{1}$ and hence also within the
region ${\mathcal {R}}$ for all $r\le r_{1}$.
Providing we have chosen $r_{1}$ sufficiently large, for $r>r_{1}$ we are in the asymptotic large $r$ regime discussed in section 4.2 of Ref.~\onlinecite{Baxter3}.
The upshot of that analysis is that, by taking $r_{1}$ sufficiently large, the change in the gauge field functions as $r\rightarrow \infty $ from $r=r_{1}$
can be made arbitrarily small.
Therefore, since ${\mathcal {R}}$ is an open region, the gauge field functions ${\ov {\omega }}_{j}$ for our ${\mathfrak {su}}(N)$ solutions will remain
inside ${\mathcal {R}}$ for all $r>r_{1}$.
By way of illustration, in figures \ref{fig:3} and \ref{fig:4}, we show how $-W_{j}$ (\ref{eq:Wdef}) depend on $r$ for the example soliton and black hole
solutions plotted in figures \ref{fig:1} and \ref{fig:2} respectively.
In all cases we see that $-W_{j}\ge 0$ for all $r$, so that these example solutions have no instabilities in the sphaleronic sector.
\begin{figure}
\includegraphics[width=7.5cm]{Figure3a.eps}
\includegraphics[width=7.5cm]{Figure3b.eps}
\caption{$-W_{1}(r)$ and $-W_{2}(r)$ for the example nodeless ${\mathfrak {su}}(3)$ solutions shown in figure \ref{fig:1}, (a) soliton and (b) black hole.
In both cases $-W_{j}\ge 0$ for all $r$, so that the inequalities (\ref{eq:sstabineqs}) are satisfied and the solutions have no instabilities in the
sphaleronic sector.}
\label{fig:3}
\end{figure}
\begin{figure}
\includegraphics[width=7.5cm]{Figure4a.eps}
\includegraphics[width=7.5cm]{Figure4b.eps}
\caption{$-W_{1}(r)$, $-W_{2}(r)$ and $-W_{3}(r)$ for the example nodeless ${\mathfrak {su}}(4)$ solutions shown in figure \ref{fig:2}, (a) soliton and (b) black hole.
In both cases $-W_{j}\ge 0$ for all $r$, so that the inequalities (\ref{eq:sstabineqs}) are satisfied and the solutions have no instabilities in the
sphaleronic sector.}
\label{fig:4}
\end{figure}
\section{Gravitational sector perturbations}
\label{sec:gravitational}
The gravitational sector consists of the even-parity Yang-Mills perturbations
$\delta \omega _{j}$ ($j=1,\ldots, N-1$) and the metric perturbations $\delta \mu $ and $\delta \Delta $.
The governing equations are the perturbed Einstein equations (\ref{eq:deltamup}--\ref{eq:deltadeltap}) together with the linearized Yang-Mills equations formed by subtracting equations (\ref{eq:qrealYMpert}) and
(\ref{eq:qimYMpert}), namely:
\begin{eqnarray}
0 & = &
-\delta {\ddot {\omega }}_j
+ {\ov {\mu }}^2 {\ov {S}}^2 \delta\omega''_j
+ {\ov {\mu }} {\ov {S}}^2
\left(
{\ov {\mu }}' + {\ov {\mu }} \frac{{\ov {S}}'}{{\ov {S}}}\right) \delta\omega_j'
+{\ov {\mu }} {\ov {S}}^2 {\ov {\omega }}_j'
\left[
\delta \mu' + {\ov {\mu }} \delta \left( \frac{S'}{S} \right)
+\delta \mu \frac{{\ov {S}}'}{{\ov {S}}}\right]
+{\ov {\mu }} {\ov {S}}^{2}{\ov {\omega }}_j''\delta\mu
\nonumber \\ & &
+\frac{{\ov {\mu }} {\ov {S}}^2}{r^2}
\left[
W_j \delta\omega_j -2{\ov {\omega }}_j^{2}\delta\omega_j
+{\ov {\omega }}_{j+1} {\ov {\omega }}_j \delta\omega_{j+1}
+{\ov {\omega }}_{j-1} {\ov {\omega }}_j \delta\omega_{j-1}
\right] .
\label{eq:deltaomegaeqn1}
\end{eqnarray}
\subsection{Metric perturbations}
\label{sec:metricperts}
The linearized Einstein equations (\ref{eq:deltamup}--\ref{eq:deltadeltap}) can be
used to eliminate the metric perturbations from the remaining gravitational sector
perturbation equations (\ref{eq:deltaomegaeqn1}). The linearized Einstein equation (\ref{eq:deltamudot}) can be immediately integrated to give
\begin{equation}
\delta \mu = -\frac {4{\ov {\mu }}}{r}
\sum _{j=1}^{N-1} {\ov {\omega }}_{j}' \, \delta \omega _{j}
+ \delta Y (r),
\end{equation}
where $\delta Y(r)$ is an arbitrary function of $r$ alone.
Using the linearized Einstein equation (\ref{eq:deltamup}) gives, after some lengthy algebra
\begin{equation}
\delta Y' = -\frac {1}{r} \left( 1 + 2{\ov {\Gamma }} \right) \delta Y ,
\label{eq:Yprime}
\end{equation}
where ${\ov {\Gamma }}$ is given by (\ref{eq:Pistatic}).
Integrating (\ref{eq:Yprime}) we find
\begin{equation}
\delta Y(r) = Y_{0} \exp \left( - \int _{r_{0}}^{r} \frac {1}{r'} \left[ 1+
2{\ov {\Gamma }}(r') \right] dr' \right)
\end{equation}
where $Y_{0}$ is a constant. The lower limit on the integral $r_{0}=0$ if we are considering perturbations of a static soliton solution, $r_{0}=r_{h}$ if we are considering perturbations of a static black hole solution.
Since we require our perturbations to vanish at either the origin or black hole event horizon, as relevant, it must be the case that $Y_{0}=0$ and hence
$\delta Y(r) \equiv 0$.
We therefore have
\begin{equation}
\delta \mu = -\frac {4{\ov {\mu }}}{r}
\sum _{j=1}^{N-1} {\ov {\omega }}_{j}' \, \delta \omega _{j}.
\label{eq:deltamu}
\end{equation}
\subsection{Gravitational sector perturbation equations in matrix form}
\label{sec:gravmatrix}
Now that we have the form (\ref{eq:deltamu}) of the metric perturbation
$\delta \mu $, together with (\ref{eq:deltadeltap}) for the perturbation $\delta \Delta '$, we can eliminate the metric perturbations from the gravitational sector perturbation equations (\ref{eq:deltaomegaeqn1}).
First we consider the quantity
\begin{equation}
\delta \mu ' + {\ov {\mu }} \delta \left( \frac {S'}{S} \right) +
\frac {{\ov {S}}'}{{\ov {S}}} \delta \mu
= -\frac {1}{r} \left( \delta \mu + 2r^{2} \delta \Pi \right),
\label{eq:gravuseful}
\end{equation}
where $\delta \Pi $ is given in (\ref{eq:deltaPi}). The right-hand-side of
(\ref{eq:gravuseful}) depends only on the perturbations $\delta \omega _{j}$ and not on their derivatives.
Next we define a vector of perturbations as follows:
\begin{equation}
{\bmath {\delta \omega }} = \left(
\delta \omega _{1}, \ldots , \delta \omega _{N-1} \right) ^{T}.
\end{equation}
Changing the radial co-ordinate to the tortoise co-ordinate (\ref{eq:tortoise}),
the gravitational sector perturbation equations (\ref{eq:deltaomegaeqn1}) take the form
\begin{equation}
-{\bmath {\delta }} {\ddot {\bmath {\omega }}}
=-\drs ^{2} {\bmath {\delta \omega }} + {\mathcal {M}} {\bmath {\delta \omega }} .
\label{eq:gravmatrix}
\end{equation}
The $\left( N-1 \right) \times \left( N-1 \right)$ matrix ${\mathcal {M}}$
depends only on the static equilibrium solutions and does not contain any derivative operators.
To simplify the entries of ${\mathcal {M}}$, we make extensive use of the static equilibrium field equations (\ref{eq:Ee}, \ref{eq:YMe}).
After a lengthy calculation, we can write the entries of the symmetric matrix ${\mathcal {M}}$ as follows. There are three different types of entry which have different forms: (i) the diagonal entries ${\mathcal {M}}_{j,j}$, (ii) entries immediately above and below the diagonal ${\mathcal {M}}_{j,j+1}$ and (iii) other entries not on the diagonal nor immediately above or below it ${\mathcal {M}}_{j,k}$ ($k\neq j, j+1$). We give these entries explicitly below, where there is no summation:
\begin{eqnarray}
{\mathcal {M}}_{j,j} & = &
-\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} \left[
W_{j} - 2{\ov {\omega }}_{j}^{2} \right]
- \frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }}_{j} \right) ^{2}
- \frac {8{\ov {S}}}{r^{3}}
W_{j} {\ov {\omega }}_{j} \drs {\ov {\omega }}_{j},
\nonumber
\\
{\mathcal {M}}_{j, j+1} & = &
-\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}}{\ov {\omega }}_{j} {\ov {\omega }}_{j+1}
-\frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }}_{j} \right) \left( \drs {\ov {\omega }}_{j+1} \right)
- \frac {4{\ov {S}}}{r^{3}}
\left[ W_{j} {\ov {\omega }}_{j} \drs {\ov {\omega }}_{j+1}
+ W_{j+1} {\ov {\omega }}_{j+1} \drs {\ov {\omega }}_{j} \right] ,
\nonumber
\\
{\mathcal {M}}_{j, k} & = &
-\frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }}_{j} \right) \left( \drs {\ov {\omega }}_{k} \right)
- \frac {4{\ov {S}}}{r^{3}}
\left[ W_{j} {\ov {\omega }}_{j} \drs {\ov {\omega }}_{k}
+ W_{k} {\ov {\omega }}_{k} \drs {\ov {\omega }}_{j} \right] ,
\label{eq:Mentries}
\end{eqnarray}
where we have defined
\begin{equation}
{\mathcal {Q}} =
\frac {1}{{\ov {\mu }}} \drs {\ov {\mu }} + \frac {1}{{\ov {S}}} \drs {\ov {S}}
+ \frac {{\ov {\mu }}{\ov {S}}}{r}.
\end{equation}
We now consider time-periodic perturbations for which ${\bmath {\delta \omega }}(t,r)=e^{i\sigma t} {\bmath {\delta \omega }}(r)$, and then the gravitational
sector perturbation equations (\ref{eq:gravmatrix}) are:
\begin{equation}
\sigma ^{2} {\bmath {\delta \omega }}
=-\drs ^{2} {\bmath {\delta \omega }} + {\mathcal {M}} {\bmath {\delta \omega }} .
\label{eq:gravtimepert}
\end{equation}
Since (\ref{eq:gravtimepert}) has the form of a standard Schr\"odinger-like equation, the operator on the right-hand-side of (\ref{eq:gravtimepert})
is positive if the matrix ${\mathcal {M}}$ is positive.
If this is the case, then $\sigma ^{2}$ is real and there are no unstable modes in the gravitational sector.
\subsection{Special cases}
\label{sec:Gspecial}
The gravitational sector perturbation equations (\ref{eq:gravmatrix}) are rather complicated in general, so first we consider some special cases.
\subsubsection{Schwarzschild-adS}
\label{sec:gSadS}
Setting ${\ov {\omega }}_{j} \equiv {\sqrt {j\left( N -j \right)}}$ for
$j=1,\ldots ,N-1$, the entries in the matrix ${\mathcal {M}}$ in the gravitational sector perturbation equations reduce to
\begin{eqnarray}
{\mathcal {M}}_{j,j} & = &
\frac {2{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} j\left( N - j \right) ,
\nonumber \\
{\mathcal {M}}_{j,j+1} & = &
-\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}}
{\sqrt {j\left( N-j \right) \left( j+1 \right) \left( N -j - 1\right) }} ,
\nonumber \\
{\mathcal {M}}_{j,k} & = & 0, \qquad k \neq j, j+1.
\end{eqnarray}
Therefore we have
\begin{equation}
{\mathcal {M}} = \frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} {\mathcal {E}}_{N-1},
\end{equation}
where ${\mathcal {E}}_{N-1}$ is the constant matrix with entries
\begin{eqnarray}
{\mathcal {E}}_{N-1,j,k} & = & {\sqrt {j\left( N - j\right)}}
{\sqrt {k\left( N-k\right) }}
\left[ 2\delta _{j,k} - \delta _{j+1,k} - \delta _{j-1,k} \right] .
\label{eq:MN-1def}
\end{eqnarray}
It is shown in Ref.~\onlinecite{Baxter2} that the eigenvalues of the matrix
${\mathcal {E}}_{N-1}$ are $k\left( k+1\right)$ for $k=1,\ldots N-1$,
so that the matrix ${\mathcal {E}}_{N-1}$ (and therefore the matrix ${\mathcal {M}}$)
is positive.
The net result of this is that, as anticipated, the embedded Schwarzschild-adS solution has no instabilities in the gravitational sector.
\subsubsection{Reissner-Nordstr\"om-adS}
\label{sec:gRNadS}
In this case ${\ov {\omega }}_{j} \equiv 0$ for all $j=1,\ldots , N-1$,
and the matrix ${\mathcal {M}}$ reduces to
\begin{equation}
{\mathcal {M}} = -\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} {\mathcal {I}}_{N-1}.
\end{equation}
Therefore the matrix ${\mathcal {M}}$ is negative definite everywhere and, as expected, the embedded
magnetically charged Reissner-Nordstr\"om-adS solution is unstable.
\subsubsection{Embedded ${\mathfrak {su}}(2)$ solutions}
\label{sec:gsu2}
We now have ${\ov {\omega }}_{j} \equiv {\ov {\omega }}(r) {\sqrt {j\left( N-j \right) }}$
for $j=1,\ldots ,N-1$, and the entries of the matrix ${\mathcal {M}}$ (\ref{eq:Mentries}) take the form
\begin{widetext}
\begin{eqnarray}
{\mathcal {M}}_{j,j} & = &
\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} \left( {\ov {\omega }} ^{2} -1 \right)
+ j \left( N - j \right) \left[
\frac {2{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} {\ov {\omega }}^{2}
- \frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r } \left( \drs {\ov {\omega }} \right) ^{2}
- \frac {8{\ov {S}}}{r^{3}}
\left( 1 -{\ov {\omega }} ^{2} \right) {\ov {\omega }} \drs {\ov {\omega }} \right],
\nonumber \\
{\mathcal {M}}_{j,j+1} & = &
- {\sqrt {j\left( N-j \right) \left( j+1 \right) \left( N-j-1 \right) }}
\left[
\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} {\ov {\omega }} ^{2}
+\frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }} \right) ^{2}
+ \frac {8{\ov {S}}}{r^{3}}
\left( 1 - {\ov {\omega }}^{2} \right) {\ov {\omega }} \drs {\ov {\omega }} \right]
,
\nonumber \\
{\mathcal {M}}_{j,k} & = &
- {\sqrt {j\left( N-j \right) k \left( N-k \right) }}
\left[
\frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }} \right) ^{2}
+ \frac {8{\ov {S}}}{r^{3}}
\left( 1- {\ov {\omega }}^{2} \right) {\ov {\omega }} \drs {\ov {\omega }} \right].
\end{eqnarray}
\end{widetext}
In this case it is helpful to consider the matrix ${\mathcal {M}}$ as a sum of three parts:
\begin{equation}
{\mathcal {M}}={\mathcal {N}}_{1} + {\mathcal {N}}_{2} + {\mathcal {N}}_{3},
\end{equation}
where
\begin{eqnarray}
{\mathcal {N}}_{1} & = &
\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}} \left( {\ov {\omega }}^{2} -1 \right)
{\mathcal {I}}_{N-1},
\nonumber \\
{\mathcal {N}}_{2} & = &
\frac {{\ov {\mu }}{\ov {S}}^{2}}{r^{2}}
{\ov {\omega }}^{2} {\mathcal {E}}_{N-1},
\nonumber \\
{\mathcal {N}}_{3} & = &
\left[
-\frac {4{\mathcal {Q}}}{{\ov {\mu }}{\ov {S}}r }
\left( \drs {\ov {\omega }} \right) ^{2}
- \frac {8{\ov {S}}}{r^{3}}
\left( 1- {\ov {\omega }} ^{2} \right) {\ov {\omega }} \drs {\ov {\omega }}
\right] {\widetilde {\mathcal {E}}}_{N-1} ,
\nonumber \\
\end{eqnarray}
where the constant matrix ${\mathcal {E}}_{N-1}$ is given in (\ref{eq:MN-1def})
and the constant matrix ${\widetilde {{\mathcal {E}}}}_{N-1}$ has entries
\begin{equation}
{\widetilde {{\mathcal {E}}}}_{N-1,j,k}=
{\sqrt {j\left( N-j \right) k \left( N-k \right) }}.
\end{equation}
The first matrix, ${\mathcal {N}}_{1}$, is positive if ${\ov {\omega }}(r) ^{2} \ge 1$ for all $r$, which is the same sufficient
condition as we found previously for no instabilities in the sphaleronic sector (see section \ref{sec:ssu2}).
Since the matrix ${\mathcal {E}}_{N-1}$ is positive, the second matrix
${\mathcal {N}}_{2}$ is also positive.
The positivity of the third matrix, ${\mathcal {N}}_{3}$ is less clear-cut.
However, it has been shown \cite{Winstanley1} that
${\sqrt {\left| \Lambda \right| }}{\ov {\omega }}' \rightarrow 0$ for all $r$ as $\left| \Lambda \right| \rightarrow \infty $.
Therefore, for sufficiently large $\left| \Lambda \right| $, the third matrix ${\mathcal {N}}_{3}$ is negligible compared with ${\mathcal {N}}_{1}$
and ${\mathcal {N}}_{2}$.
Hence, for sufficiently large $\left| \Lambda \right| $, if ${\ov {\omega }}(r)^{2}\ge 1$ for all $r$, the matrix ${\mathcal {M}}$ is positive and there
are no instabilities for embedded ${\mathfrak {su}}(2)$ solutions in the gravitational sector of ${\mathfrak {su}}(N)$ EYM perturbations.
In Ref.~\onlinecite{Winstanley1} it shown that ${\ov {\omega }}(r)^{2}\ge 1/3$ is a sufficient condition for ${\mathfrak {su}}(2)$ black holes to have no instabilities
in the gravitational sector of ${\mathfrak {su}}(2)$ EYM perturbations provided $\left| \Lambda \right| $ is sufficiently large.
As with the sphaleronic sector perturbations, for embedded ${\mathfrak {su}}(2)$ solutions in ${\mathfrak {su}}(N)$ EYM, we have a stronger
sufficient condition for the absence of unstable modes.
This is to be expected because of the greater number of degrees of freedom in the ${\mathfrak {su}}(N)$ gravitational sector perturbations than in the
${\mathfrak {su}}(2)$ gravitational sector perturbations.
We emphasize that the condition of ${\ov {\omega }}(r)^{2}\ge 1$ for all $r$ and sufficiently large $\left| \Lambda \right| $ is a sufficient condition, and there may be solutions which do not satisfy this condition but which are nonetheless stable.
\subsection{An alternative form of the gravitational sector perturbation equations}
\label{sec:gravalt}
In order to show, in the next section, that there exist both soliton and black hole non-embedded solutions of the static equilibrium field equations which have no instabilities in the gravitational sector, we shall follow the method of Ref.~\onlinecite{Sarbach2} and employ a nodal theorem due to Amann and Quittner \cite{Amann} which allows one to count the number of bound states of a Schr\"odinger-like equation.
In this subsection we state Amann and Quittner's result and cast our gravitational sector perturbation equations (\ref{eq:gravtimepert}) in the form required for the application of the theorem in section \ref{sec:gravstab}. We will need to consider solitons and black holes separately.
\subsubsection{The nodal theorem}
\label{sec:nodal}
Amann and Quittner's theorem \cite{Amann} is concerned with the number of bound states of a radial Schr\"odinger-like operator.
Let ${\mathfrak {D}}$ be the linear differential operator
\begin{equation}
{\mathfrak {D}} {\bmath {u}} = -\frac {d}{d\rho } \left[ {\mathfrak {A}} (\rho )
\frac {d}{d\rho } {\bmath {u}} \right]
+ \left[ \frac {1}{\rho ^{2}} {\mathfrak {B}}(\rho ) + {\mathfrak {C}} (\rho )
\right] {\bmath {u}},
\label{eq:operator}
\end{equation}
acting on $n$-dimensional vectors ${\bmath {u}}(\rho )$, where $\rho \in \left[ 0,
\infty \right) $ lies on the half-line.
The $n\times n$ matrices ${\mathfrak {A}}(\rho )$, ${\mathfrak {B}}(\rho )$ and
${\mathfrak {C}}(\rho )$ are assumed to be real, symmetric, smooth and uniformly bounded on $\left[ 0, \infty \right) $.
It is further assumed that ${\mathfrak {A}}(\rho )$ is uniformly positive definite on $\left[ 0, \infty \right) $, that is,
there is a constant ${\mathfrak {a}}>0$ such that
\begin{equation}
{\mathfrak {A}}(\rho ) \ge {\mathfrak {a}} >0 \qquad
{\mbox {for $0\le \rho < \infty $}},
\label{eq:Acond}
\end{equation}
and that ${\mathfrak {B}}(0)$ is non-negative.
The theorem is concerned with the eigenvalue problem
\begin{equation}
{\mathfrak {D}} {\bmath {u}} = \lambda {\bmath {u}}, \qquad
{\bmath {u}}(0)=0,
\label{eq:eigenvalue}
\end{equation}
where ${\bmath {u}} \in L_{2}\left( (0,\infty ), {\mathbb {R}}^{n} \right) $.
Following Ref.~\onlinecite{Amann}, we further assume that the bottom of the essential spectrum of ${\mathfrak {D}}$ is positive and that the eigenvalue problem (\ref{eq:eigenvalue}) has only finitely many negative eigenvalues.
Sufficient conditions for this assumption to be valid are \cite{Amann}:
\begin{equation}
{\mathfrak {C}}(\rho ) \rightarrow 0 {\mbox { as $\rho \rightarrow \infty $}}
\label{eq:Ccond1}
\end{equation}
and
\begin{equation}
{\mathfrak {C}}(\rho ) \rho^{2} + {\mathfrak {B}}(\rho ) \ge - {\mathfrak {b}}
\label{eq:Ccond2}
\end{equation}
for some ${\mathfrak {b}}< 1/4$ and all sufficiently large $\rho $.
The statement of Amann and Quittner's theorem involves an auxiliary problem, which we now state.
Choose $n$ linearly independent real, constant, $n$-dimensional vectors
${\bmath {e}}_{j}$, $j=1,\ldots, n$.
Let ${\mathfrak {c}}>0$ and let
\begin{equation}
{\mathfrak {U}}_{{\mathfrak {c}}}= \left[ {\bmath {u}}_{1}, \ldots ,
{\bmath {u}}_{n} \right]
\end{equation}
be the $n\times n$ matrix whose columns are the solutions of the $n$ initial
value problems
\begin{equation}
{\mathfrak {D}} {\bmath {u}}_{j} = 0, \quad
{\mathfrak {c}}< \rho < \infty,
\quad
{\bmath {u}}_{j} ({\mathfrak {c}}) =0, \quad
\frac {d}{d\rho }{\bmath {u}}_{j}({\mathfrak {c}}) = {\bmath {e}}_{j},
\label{eq:IVP}
\end{equation}
for $j=1,\ldots ,n$.
We then define a scalar function ${\mathfrak {F}}(\rho )$ by
\begin{equation}
{\mathfrak {F}}(\rho )= \det {\mathfrak {U}}_{\mathfrak {c}}(\rho ).
\label{eq:Fdef}
\end{equation}
We are now in a position to quote Amann and Quittner's theorem \cite{Amann}:
\begin{theorem}
If ${\mathfrak {c}}>0$ is sufficiently small and ${\mathfrak {d}}>{\mathfrak {c}}$
is sufficiently large, the number of zeros (counted with multiplicities) in the interval $({\mathfrak {c}},{\mathfrak {d}})$ of the function ${\mathfrak {F}}(\rho )$
equals the number of negative eigenvalues of (\ref{eq:eigenvalue}) (counted with multiplicities).
\end{theorem}
In order to apply this theorem, we need to cast the gravitational sector perturbation equations (\ref{eq:gravtimepert}) in the form (\ref{eq:eigenvalue}), choose a suitable co-ordinate $\rho $ and check that the matrices ${\mathfrak {A}}(\rho )$,
${\mathfrak {B}}(\rho )$ and ${\mathfrak {C}}(\rho )$ satisfy the required conditions
(\ref{eq:Acond}, \ref{eq:Ccond1}, \ref{eq:Ccond2}) together with the requirement that ${\mathfrak {B}}(0)$ is non-negative.
To do this, as in Ref.~\onlinecite{Sarbach2}, we need to consider soliton and black hole solutions separately. We consider black holes first as this case is simpler.
\subsubsection{Black holes}
\label{sec:nodalBHs}
Following Ref.~\onlinecite{Sarbach2}, for static black hole solutions we take
\begin{equation}
\rho = -r_{*} \in \left[ 0,\infty \right)
\end{equation}
where $r_{*}$ is the tortoise co-ordinate (\ref{eq:tortoise}).
This means that $\rho \rightarrow 0$ corresponds to $r\rightarrow \infty $,
and $\rho \rightarrow \infty $ corresponds to $r\rightarrow r_{h}$, approaching the event horizon.
We choose the $\left( N-1 \right) \times \left( N-1 \right) $ matrices appearing in the differential operator ${\mathfrak {D}}$
(\ref{eq:operator}) as follows:
\begin{equation}
{\mathfrak {A}}(\rho ) = {\mathcal {I}}_{N-1},
\qquad
{\mathfrak {B}}(\rho ) =0,
\qquad
{\mathfrak {C}}(\rho ) = {\mathcal {M}},
\end{equation}
where the matrix ${\mathcal {M}}$ has entries (\ref{eq:Mentries}).
Taking ${\mathfrak {a}}=\frac {1}{2}>0$, the condition (\ref{eq:Acond}) is automatically satisfied, and, furthermore, ${\mathfrak {B}}(0)=0$ is non-negative.
It remains therefore to check the conditions on the matrix ${\mathfrak {C}}$
(\ref{eq:Ccond1}--\ref{eq:Ccond2}).
We first examine the behaviour of the matrix ${\mathfrak {C}}(\rho )$ as
$\rho \rightarrow 0$, that is, $r\rightarrow \infty $.
Using the boundary conditions (\ref{eq:infinity}),
and noting that, as $r\rightarrow \infty $,
\begin{equation}
\drs {\ov {\omega }}_{j} = \frac {\Lambda c_{j}}{3} +O(r^{-1}),
\qquad
{\mathcal {Q}} = -\Lambda r + O(1),
\end{equation}
we find that the leading-order behaviour of ${\mathfrak {C}}$ is given by the entries (\ref{eq:Mentries})
\begin{eqnarray}
{\mathcal {M}}_{j,j} & = &
\frac {\Lambda }{3}\left[ 1- 3{\ov {\omega }}_{j,\infty }^{2}
+\frac {1}{2} \left( {\ov {\omega }}_{j+1,\infty }^{2}
+{\ov {\omega }}_{j-1,\infty }^{2} \right) \right]
+ O(r^{-1}),
\nonumber \\
{\mathcal {M}}_{j,j+1} & = &
\frac {\Lambda }{3} {\ov {\omega }}_{j,\infty }{\ov {\omega }}_{j+1,\infty }
+ O(r^{-1}),
\nonumber \\
{\mathcal {M}}_{j,k} & = &
O(r^{-2}).
\end{eqnarray}
Therefore the matrix ${\mathfrak {C}}(\rho )$ remains bounded as
$r\rightarrow \infty $, that is, $\rho \rightarrow 0$.
As $\rho \rightarrow \infty $, we have $r\rightarrow r_{h}$.
Using the boundary conditions (\ref{eq:horizon}), we find that
\begin{equation}
\drs {\ov {\omega }}_{j} = O(r-r_{h}), \qquad
{\mathcal {Q}} = {\ov {S}}(r_{h}) {\ov {\mu }}'(r_{h}) + O(r-r_{h}),
\end{equation}
and that the entries of the matrix ${\mathfrak {C}}$ are all $O(r-r_{h})$ as
$r\rightarrow r_{h}$, so that ${\mathfrak {C}}(\rho ) \rightarrow 0$ as $\rho \rightarrow \infty $, satisfying (\ref{eq:Ccond1}).
This also means that the matrix ${\mathfrak {C}}(\rho )$ is uniformly bounded
on $\left[ 0, \infty \right) $.
To check whether (\ref{eq:Ccond2}) is satisfied, we first note that, for $r\sim r_{h}$,
\begin{equation}
\rho = -r_{*} \sim -\rho _{h}\ln \left( r- r_{h} \right) ,
\end{equation}
where $\rho _{h}$ is a positive constant.
Therefore, as $\rho \rightarrow \infty $,
\begin{equation}
{\mathfrak {C}}(\rho ) \sim O\left( e^{-\frac {\rho }{\rho _{h}}} \right) .
\end{equation}
Therefore $\rho ^{2} {\mathfrak {C}}(\rho ) \rightarrow 0$ as
$\rho \rightarrow \infty $.
Therefore there exists a $\rho _{1}$ such that for all $\rho > \rho _{1}$,
\begin{equation}
\frac {1}{8} < \rho ^{2} {\mathfrak {C}} (\rho ) < -\frac {1}{8} .
\end{equation}
Hence we have satisfied (\ref{eq:Ccond2}) with ${\mathfrak {b}}=1/8$.
Therefore we have cast the gravitational sector perturbation equations in the form (\ref{eq:eigenvalue}) required for the application of the nodal theorem, and all the conditions required by the theorem are satisfied.
We comment that this case was simpler to deal with than the situation in Ref.~\onlinecite{Sarbach2}, because in that paper terms arising from non-spherically symmetric perturbations cannot be included in ${\mathfrak {C}}(\rho )$.
\subsubsection{Solitons}
\label{sec:nodalsolitons}
For solitons, the system of gravitational perturbation equations needs further transformation before it is in the form required for the application of the nodal theorem.
Following Ref.~\onlinecite{Sarbach2}, we define
\begin{equation}
\rho = r^{-\frac {1}{2}},
\end{equation}
so that $\rho \rightarrow 0$ corresponds to $r\rightarrow \infty $ and
$\rho \rightarrow \infty $ corresponds to the origin.
We also make a transformation of the perturbations:
\begin{equation}
{\bmath {\delta \omega }} = {\mathfrak {X}}{\bmath {v}},
\end{equation}
where
\begin{equation}
{\mathfrak {X}} = r^{\frac {3}{4}} \left(
{\ov {\mu }}{\ov {S}} \right) ^{-\frac {1}{2}} .
\end{equation}
The gravitational sector perturbation equations now take the form
\begin{equation}
4{\mathfrak {X}}^{4} \lambda {\bmath {v}}
= -\frac {d^{2}{\bmath {v}}}{d\rho ^{2}} + {\widetilde {\mathcal {M}}}{\bmath {v}},
\label{eq:gravmodified}
\end{equation}
where
\begin{equation}
{\widetilde {{\mathcal {M}}}}= 4{\mathfrak {X}}^{4}
{\mathcal {M}}
- {\mathfrak {X}} \frac {d}{d\rho } \left( \frac {1}{{\mathfrak {X}}^{2}}
\frac {d{\mathfrak {X}}}{d\rho } \right) {\mathcal {I}}_{N-1}.
\end{equation}
Comparing with (\ref{eq:operator}), as with the black hole case we take
\begin{equation}
{\mathfrak {A}}(\rho ) ={\mathcal {I}}_{N-1}.
\end{equation}
To fix the matrices ${\mathfrak {B}}(\rho )$ and ${\mathfrak {C}}(\rho )$, we
need to study the behaviour of the matrix ${\widetilde {\mathcal {M}}}$ as
$\rho \rightarrow 0$.
As $\rho \rightarrow 0$, $r\rightarrow \infty $ and
\begin{equation}
{\mathfrak {X}} = \frac {{\sqrt {3}}}{{\sqrt {-\Lambda }}}r^{-\frac {1}{4}}
+O\left( r^{-\frac {5}{4}} \right)
= \frac {{\sqrt {3}}}{{\sqrt {-\Lambda }}}\rho ^{\frac {1}{2}}
+O\left( \rho ^{\frac {5}{2}} \right) .
\label{eq:xrho0}
\end{equation}
Therefore, as $\rho \rightarrow 0$,
\begin{equation}
{\mathfrak {X}} \frac {d}{d\rho } \left( \frac {1}{{\mathfrak {X}}^{2}}
\frac {d{\mathfrak {X}}}{d\rho } \right)
= -\frac {3}{4} \rho ^{-2} + O(1).
\end{equation}
From the analysis of the previous subsection,
we know that ${\mathcal {M}}=O(1)$ as $r\rightarrow \infty $.
This suggests that we should take
\begin{equation}
{\mathfrak {B}}(\rho ) = -\rho ^{2} {\mathfrak {X}} \frac {d}{d\rho } \left( \frac {1}{{\mathfrak {X}}^{2}}
\frac {d{\mathfrak {X}}}{d\rho } \right) {\mathcal {I}}_{N-1},
\qquad
{\mathfrak {C}}(\rho ) = 4{\mathfrak {X}}^{4}
{\mathcal {M}} .
\label{eq:BCsoliton}
\end{equation}
With this choice, we have
\begin{equation}
{\mathfrak {B}}(0) = \frac {3}{4}{\mathcal {I}}_{N-1},
\end{equation}
which is non-negative as required.
To check the other conditions (\ref{eq:Ccond1}, \ref{eq:Ccond2}) on the matrices ${\mathfrak {B}}(\rho )$ and
${\mathfrak {C}}(\rho )$, we need to examine their behaviour as $\rho \rightarrow \infty $, that is, $r\rightarrow 0$.
In this case, using the boundary conditions (\ref{eq:origin1})
\begin{equation}
{\mathfrak {X}} = \frac {1}{{\sqrt {S_{0}}}}
r^{\frac {3}{4}}+O\left( r^{\frac {7}{4}} \right)
= \frac {1}{{\sqrt {S_{0}}}}
\rho ^{-\frac {3}{2}}+O\left( \rho ^{-\frac {7}{2}} \right) ,
\label{eq:Xrhoinf}
\end{equation}
which gives
\begin{equation}
{\mathfrak {X}} \frac {d}{d\rho } \left( \frac {1}{{\mathfrak {X}}^{2}}
\frac {d{\mathfrak {X}}}{d\rho } \right)
= -\frac {3}{4} \rho ^{-2} + O(\rho ^{-4}).
\label{eq:Xdrhoinf}
\end{equation}
Therefore ${\mathfrak {B}}(\rho ) \rightarrow \frac {3}{4}{\mathcal {I}}_{N-1}$ as
$\rho \rightarrow \infty $
and the matrix ${\mathfrak {B}}(\rho )$ is uniformly bounded on
$\left[ 0, \infty \right) $.
We next turn to the behaviour of the matrix ${\mathfrak {C}}(\rho )$ as $\rho \rightarrow 0$ and $r\rightarrow \infty $.
We have already seen that ${\mathcal {M}} = O(1)$ as $r\rightarrow \infty $.
Then, using the definition of ${\mathfrak {C}}(\rho )$ (\ref{eq:BCsoliton}) and the behaviour of ${\mathfrak {X}}$ as $\rho \rightarrow 0$ (\ref{eq:xrho0}),
the matrix ${\mathfrak {C}}(\rho ) \rightarrow 0$ as $\rho \rightarrow 0$.
Therefore it remains to investigate the properties of ${\mathfrak {C}}(\rho )$
as $\rho \rightarrow \infty $ and $r\rightarrow 0$.
Using the boundary conditions (\ref{eq:origin1}),
we first note that, as $r\rightarrow 0$,
\begin{equation}
\drs {\ov {\omega }}_{j} = O(r), \qquad
W_{j} =O(r^{2}), \qquad
{\mathcal {Q}} = \frac {S_{0}}{r} + O(1).
\end{equation}
Therefore the behaviour of the entries of the matrix ${\mathcal {M}}$ (\ref{eq:Mentries}) as $r\rightarrow 0$ is:
\begin{eqnarray}
{\mathcal {M}}_{j,j} & = &
\frac {2S_{0}^{2}}{r^{2}} j\left( N-j\right) +O(1),
\nonumber \\
{\mathcal {M}}_{j,j+1} & = & - \frac {S_{0}^{2}}{r^{2}}
{\sqrt {j\left( N- j\right) \left( j +1 \right) \left( N-j-1 \right) }}
+O(1),
\nonumber \\
{\mathcal {M}}_{j,k} & = & O(1).
\label{eq:Mrhoinf}
\end{eqnarray}
Using the behaviour of ${\mathfrak {X}}$ as $\rho \rightarrow \infty $ (\ref{eq:Xrhoinf}), we then have
\begin{equation}
{\mathfrak {C}} (\rho ) = O(\rho ^{-2}) \rightarrow 0 {\mbox { as $\rho \rightarrow \infty $}},
\end{equation}
so (\ref{eq:Ccond1}) is satisfied.
Looking at the remaining condition (\ref{eq:Ccond2}), using the asymptotic forms
(\ref{eq:Xrhoinf}, \ref{eq:Xdrhoinf}, \ref{eq:Mrhoinf}), we see that, as $\rho \rightarrow \infty $,
\begin{equation}
{\mathfrak {B}}(\rho ) + \rho ^{2} {\mathfrak {C}}(\rho )
= \frac {3}{4} {\mathcal {I}}_{N-1}
+4{\mathcal {E}}_{N-1}
+O(\rho ^{-2})
\end{equation}
where the matrix ${\mathcal {E}}_{N-1}$ is given by (\ref{eq:MN-1def}).
Since we know that the matrix ${\mathcal {E}}_{N-1}$ has only positive eigenvalues,
we deduce that ${\mathfrak {B}}(\rho )+\rho ^{2}{\mathfrak {C}}(\rho)$ is positive for sufficiently large $\rho $ and therefore (\ref{eq:Ccond2}) is satisfied.
Finally in this subsection we note that the eigenvalue problem we have in the gravitational sector (\ref{eq:gravmodified}) is not exactly of the form
(\ref{eq:eigenvalue}) required for the application of the nodal theorem.
However, since ${\mathfrak {X}}^{4} \ge 0$ everywhere, this will not be a major difficulty in our analysis in section \ref{sec:gravstab}.
\subsection{Existence of static solutions with no gravitational sector instabilities}
\label{sec:gravstab}
We are now in a position to prove, in this subsection, the existence of non-trivial (that is, non-embedded) ${\mathfrak {su}}(N)$ solitons and black holes which have no instabilities in the gravitational sector.
For both solitons and black holes, our argument will use Amann and Quittner's nodal theorem \cite{Amann}.
For black holes, in section \ref{sec:nodalBHs} we have written the gravitational sector perturbation equations in the standard form (\ref{eq:eigenvalue})
required for the application of the nodal theorem.
In order to show that black hole solutions have no instabilities in the gravitational sector, it therefore suffices to show that the function
${\mathfrak {F}}(\rho )$ (\ref{eq:Fdef}) has no zeros on an interval $\rho \in ({\mathfrak {c}},{\mathfrak {d}})$, for small
${\mathfrak {c}}$ and large ${\mathfrak {d}}$.
For soliton solutions, the argument is a little more involved.
First of all, the gravitational sector perturbation equations (\ref{eq:gravmodified}) take the form
\begin{equation}
{\mathfrak {G}} \lambda {\bmath {v}}
= {\mathfrak {D}}{\bmath {v}}
= -\frac {d^{2}{\bmath {v}}}{d\rho ^{2}} + {\widetilde {\mathcal {M}}}{\bmath {v}},
\label{eq:eigenvaluemod}
\end{equation}
where ${\mathfrak {G}}=4{\mathfrak {X}}^{4}$ is a positive function, whereas the nodal theorem applies to the eigenvalue problem
${\mathfrak {D}}{\bmath {v}}=\lambda {\bmath {v}}$ (\ref{eq:eigenvalue}).
Suppose that we are able to show that there exist ${\mathfrak {su}}(N)$ soliton solutions for which the function ${\mathfrak {F}}(\rho )$ (\ref{eq:Fdef}) has
no zeros in the interval $\rho \in ({\mathfrak {c}},{\mathfrak {d}})$.
Then, applying the nodal theorem, the eigenvalue problem (\ref{eq:eigenvalue}) has no negative eigenvalues.
This means that the operator ${\mathfrak {D}}$ is a positive operator.
Then, if ${\mathfrak {D}}$ is a positive operator, it must be the case that the eigenvalue problem (\ref{eq:eigenvaluemod}) also cannot have
any negative eigenvalues because ${\mathfrak {G}}$ is a positive function.
The upshot is that, for the soliton case as for the black hole case, if we can show that the function ${\mathfrak {F}}(\rho )$ (\ref{eq:Fdef})
has no zeros in
an appropriate interval, then there are no instabilities in the gravitational sector.
From the existence theorems in Ref.~\onlinecite{Baxter3}, we know that the equilibrium ${\mathfrak {su}}(N)$ solutions of the field equations are analytic in $r$, $\Lambda $ and the parameters at the origin or event horizon characterizing either soliton or black hole solutions.
The matrices ${\mathfrak {A}}$, ${\mathfrak {B}}$ and ${\mathfrak {C}}$ appearing in the operator ${\mathfrak {D}}$ (\ref{eq:operator})
are analytic functions of the equilibrium field functions ${\ov {\mu }}(r)$, ${\ov {S}}(r)$ and ${\ov {\omega }}_{j}(r)$ and $r$ (and hence $\rho $)
for values of $\rho $ in our interval of interest $({\mathfrak {c}}, {\mathfrak {d}})$.
Standard existence theorems for ordinary differential equations (see, for example, Ref.~\onlinecite{codlev}) then tell us that the solutions ${\bmath {u}}_{j}$
of the initial value problems (\ref{eq:IVP}) are also analytic functions of $\rho $, $\Lambda $ and the initial parameters at either the origin or
event horizon. Therefore, the function ${\mathfrak {F}}(\rho )$ (\ref{eq:Fdef}) is also analytic in $\rho $, $\Lambda $ and the initial parameters at the origin or event horizon.
In section \ref{sec:ssu2}, we proved the existence of embedded ${\mathfrak {su}}(2)$ solitons and black holes
for which ${\ov {\omega }}(r)^{2}>1$ for all $r$.
There we also showed that these embedded ${\mathfrak {su}}(2)$ solutions have no instabilities in the sphaleronic sector.
From section \ref{sec:gsu2} they also have no instabilities in the gravitational sector.
Pick such an embedded ${\mathfrak {su}}(2)$ solution (either a soliton or a black hole).
Fix ${\mathfrak {c}}$ to be very small and ${\mathfrak {d}}$ to be very large.
Then, from the nodal theorem, the function ${\mathfrak {F}}$ will have no zeros on the interval $({\mathfrak {c}},{\mathfrak {d}})$ for this particular embedded ${\mathfrak {su}}(2)$ soliton or black hole.
From the existence theorems in Ref.~\onlinecite{Baxter3}, there exist genuinely ${\mathfrak {su}}(N)$ solutions in a neighbourhood of this embedded
${\mathfrak {su}}(2)$ solution.
Since ${\mathfrak {F}}(\rho )$ is analytic in the parameters at the origin or event horizon which characterize the ${\mathfrak {su}}(N)$ solutions,
providing the ${\mathfrak {su}}(N)$ solutions are sufficiently close to the embedded ${\mathfrak {su}}(2)$ solution, the function ${\mathfrak {F}}(\rho )$
will continue to have no zeros in the interval $({\mathfrak {c}},{\mathfrak {d}})$ for the ${\mathfrak {su}}(N)$ solitons or black holes.
Therefore, if we consider ${\mathfrak {su}}(N)$ solutions sufficiently close to this stable embedded ${\mathfrak {su}}(2)$ solution,
using the nodal theorem (and
considering the operator ${\mathfrak {D}}$ for soliton solutions as described above), we have therefore proven that these ${\mathfrak {su}}(N)$ solutions
have no instabilities in the gravitational sector.
In section \ref{sec:sphstab}, we showed that ${\mathfrak {su}}(N)$ solutions in a neighbourhood ${\mathfrak {N}}_{1}$ of the above stable embedded ${\mathfrak {su}}(2)$ solution
have no instabilities in the sphaleronic sector.
Having, in the current section, shown that ${\mathfrak {su}}(N)$ solutions in another neighbourhood ${\mathfrak {N}}_{2}$ of the above
embedded ${\mathfrak {su}}(2)$ solution
have no instabilities in the gravitational sector, we can deduce that those ${\mathfrak {su}}(N)$ solutions in the intersection of ${\mathfrak {N}}_{1}$
and ${\mathfrak {N}}_{2}$ are stable under linear, spherically symmetric perturbations.
\section{Conclusions}
\label{sec:conc}
In this paper we have proven the existence of non-trivial, purely magnetic, spherically symmetric,
${\mathfrak {su}}(N)$ Einstein-Yang-Mills solitons and black holes in asymptotically
anti-de Sitter space (with a negative cosmological constant $\Lambda $) which are stable under linear, spherically symmetric perturbations.
The equilibrium solutions we consider are purely magnetic and spherically symmetric and the Yang-Mills field is described by $N-1$ functions
${\ov {\omega }}_{j}(r)$.
With an appropriate choice of gauge, the perturbation equations for linear, spherically symmetric, perturbations decouple into two sectors: the sphaleronic
sector and the gravitational sector.
The sphaleronic sector, involving only gauge field perturbations, is easier to analyze and is considered in section \ref{sec:sphaleronic}.
We find a series of inequalities (\ref{eq:sstabineqs}) on the equilibrium functions ${\ov {\omega }}_{j}(r)$ which are sufficient for there to be no instabilities in the sphaleronic sector.
We first proved the existence of embedded ${\mathfrak {su}}(2)$ solutions which satisfy these inequalities, before showing that ${\mathfrak {su}}(N)$
solutions in a neighbourhood of these stable embedded ${\mathfrak {su}}(2)$ solutions also have no instabilities in the sphaleronic sector.
The gravitational sector is studied in section \ref{sec:gravitational}.
The metric perturbations can be eliminated to leave a set of equations for gauge field perturbations.
Our approach to proving stability in this sector follows Ref.~\onlinecite{Sarbach2}, making use of a nodal theorem for a multidimensional Schr\"odinger system \cite{Amann}.
Again we can prove the existence of ${\mathfrak {su}}(N)$ solutions, in a neighbourhood of stable embedded ${\mathfrak {su}}(2)$ solutions, which have no instabilities in the gravitational sector.
A natural question is how the stable ${\mathfrak {su}}(N)$ EYM black holes whose existence we have proven in this paper fit into the context of the ``no-hair'' conjecture as formulated by Bizon \cite{Bizon2}:
\begin{quotation}
Within a given matter model, a {\em {stable}} stationary black hole is uniquely
determined by global charges.
\end{quotation}
It is argued in Ref.~\onlinecite{Shepherd} that, for sufficiently large $\left| \Lambda \right| $, there exist $N-1$ non-Abelian magnetic global charges which
uniquely characterize ${\mathfrak {su}}(N)$ EYM black holes, at least for large event horizon radius $r_{h}$ and in a region of the parameter space which
contains embedded ${\mathfrak {su}}(2)$ black holes.
For both the sphaleronic and gravitational sectors, our proof of the existence of stable ${\mathfrak {su}}(N)$ EYM black holes (and solitons)
is valid for large $\left| \Lambda \right| $.
Combining our results in this paper with those in Ref.~\onlinecite{Shepherd}, we have evidence that at least some large stable
${\mathfrak {su}}(N)$ EYM black holes are uniquely determined by global charges, in accordance with Bizon's ``no-hair'' conjecture
(see also Ref.~\onlinecite{Winstanley}).
In this paper we have considered only purely magnetic, spherically symmetric solitons and black holes.
The existence of ${\mathfrak {su}}(N)$ purely magnetic topological black holes has been proven \cite{Baxter4}, and
solutions found numerically for the ${\mathfrak {su}}(3)$ gauge group \cite{Baxter5}.
Very recently it has been shown that the argument we have presented here can be extended to show the stability of some of
these ${\mathfrak {su}}(N)$ purely magnetic topological black holes \cite{Baxter6}.
Dyonic solitons and black holes
in ${\mathfrak {su}}(2)$ EYM have been found numerically \cite{Bjoraker} and the existence of solutions where
both the electric and magnetic gauge field functions have no zeros has been proven \cite{Nolan}.
Dyonic solutions have also been found numerically for the larger gauge group ${\mathfrak {su}}(3)$ \cite{Shepherd1}.
Recently the existence of dyonic soliton and black hole solutions of the ${\mathfrak {su}}(N)$ field equations has been proven \cite{Baxter7}.
The existence of stable ${\mathfrak {su}}(2)$ dyonic solutions has been proven very recently \cite{Nolan1} and it would be interesting to investigate
whether our results in this paper on the existence of stable purely magnetic solitons and black holes in ${\mathfrak {su}}(N)$ EYM in
anti-de Sitter space can be extended to dyonic solutions.
Finally, we comment that in this paper our focus has been the classical stability of ${\mathfrak {su}}(N)$ purely magnetic EYM black holes and solitons in anti-de Sitter space.
We have considered only linear, spherically symmetric perturbations.
The extension of our results to general linear perturbations is likely to be extremely challenging technically (see Refs.~\onlinecite{Sarbach1,Sarbach2} for the
${\mathfrak {su}}(2)$ case) and we would expect that at least some of the solutions which are stable under spherically symmetric linear perturbations will
remain stable when general linear perturbations are considered.
Going beyond classical stability, recent work has considered the thermodynamics of
purely magnetic ${\mathfrak {su}}(2)$ EYM black holes in anti-de Sitter space \cite{thermo} (see also Refs.~\onlinecite{adSthermo,Shepherd}).
In the ${\mathfrak {su}}(2)$ EYM case, for generic (non-integer) magnetic charge there are two branches of asymptotically anti-de Sitter black holes,
one of which is thermodynamically stable.
It would be interesting to extend the work of Ref.~\onlinecite{thermo} to the larger ${\mathfrak {su}}(N)$ gauge group.
\begin{acknowledgments}
We thank Brien Nolan for many helpful discussions on this project, and the anonymous referee for helpful comments.
The work of E.W. is supported by the Lancaster-Manchester-Sheffield
Consortium for Fundamental Physics under STFC grant ST/L000520/1.
\end{acknowledgments}
|
\section{Introduction}\label{sec:intro}
We consider the continuous, separable, differentiable and convex resource
allocation problem with a single resource constraint. The problem is
formulated as follows: Let $J := \{ 1,2, \dotsc ,n \}$. Let $\phi_j:\mathbb{R}
\rightarrow \mathbb{R}$ and $g_j:\mathbb{R} \rightarrow \mathbb{R}$, $j \in
J$, be convex and continuously differentiable. Moreover, let $b \in
\mathbb{R}$ and $-\infty < l_j<u_j < \infty$, $j \in J$. Consider the problem
to
\begin{subequations}
\begin{align}
\minimize_x &\quad \phi(\mathbf{x}) := \sum_{j \in J} \phi_j(x_j),
\\ \subto & \quad g(\mathbf{x}):= \sum _{j \in J} g_j(x_j) \le
b, \label{problemb} \\ & \quad l_j \le x_j \le u_j, \quad j \in
J. \label{problemc}
\end{align}
\label{problem}
\end{subequations}
We also consider the problem where the inequality constraint (\ref{problemb})
is replaced by an equality, i.e.,
\begin{subequations}
\begin{align}
\minimize_x &\quad \phi(\mathbf{x}) := \sum_{j \in J}
\phi_j(x_j),\\ \subto & \quad g(\mathbf{x}):= \sum_{j \in J} a_jx_j =
b, \label{problemeqb} \\ & \quad l_j \le x_j \le u_j, \quad j \in
J, \label{problemeqc}
\end{align}
\label{problemeq}
\end{subequations}
where $a_j \neq 0$, $j \in J$, and the sign is the same for all $j \in
J$. Further, we assume that there exists an optimal solution to problems
(\ref{problem}) and (\ref{problemeq}). For brevity, in the following
discussions we define $X_j := [l_j, u_j]$, $j \in J$.
Problems (\ref{problem}) and (\ref{problemeq}) arise in many areas, e.g., in
search theory (\cite{Koo99}), economics (\cite{Mar52}), stratified sampling
(\cite{BRS99}), inventory systems (\cite{MaK93}), and queuing manufacturing
networks (\cite{BiT89}). Further, these problems occur as subproblems in
algorithms that solve the integer resource allocation problem
(\cite[Section~4.7]{Mje83}, \cite[pp. 72--75]{IbK88}, and \cite{BrS02}),
multicommodity network flows (\cite[Section~4.2]{Sho85}), and several
others. Moreover, problems (\ref{problem}) and (\ref{problemeq}) can be used
as subproblems when solving resource allocation problems with more than one
resource constraint (\cite{Mje83,FeZ83}), and to solve extensions of problems
(\ref{problem}) and (\ref{problemeq}) to a nonseparable objective function
$\phi$ (\cite{Mje83,DSV07}); The books \cite{Mje83,IbK88,Lus12} describe
several extensions, such as to minmax/maxmin objectives, multiple time
periods, substitutable resources, network constraints, and integer decision
variables.
Many numerical studies of the problems (\ref{problem}) and (\ref{problemeq})
have been performed; for example, see
\cite{BiH81,NiZ92,RJL92,KoL98,Kiw07,Kiw08a,Kiw08b}. Our numerical study is
however timely and well motivated, since except for those by Kiwiel
\cite{Kiw07,Kiw08a,Kiw08b}, where the quadratic knapsack problem is studied,
none of the earlier approaches study large-scale versions of problem
(\ref{problem}) or (\ref{problemeq}). There are also several algorithms (e.g.,
\cite[Section 1.4]{NiZ92} and \cite{Ste01}) which are claimed to be promising,
but have not been evaluated in a large-scale study. Only one earlier study
(\cite{KoL98}) evaluates the performance of algorithms for the problems
(\ref{problem}) and (\ref{problemeq}) with respect to variations in the
portions of the variables whose values are at a lower or upper bound at the
optimal solution (see Section \ref{sec:dop}), and this is done for modest size
instances ($n = 10^4$) only. Further, no study has been done on the
computational complexity for non-quadratic versions of the problems
(\ref{problem}) or (\ref{problemeq}). Our numerical study also incorporates
improvements of the relaxation algorithm, as presented in
Sections~\ref{sec:bh4}--\ref{sec:tandf}, and utilizes performance profiles
(\cite{DoM02}).
As a final note on the computational tests, we only consider problem instances
where the dual variable corresponding to the resource constraint
(\ref{problemb}), respectively (\ref{problemeqb}), can be found in closed
form; otherwise, we would need to implement a numerical method in some of the
steps, e.g., a Newton method. We consider only customized algorithms for the
problem at hand, since we presume that they perform better than more general
algorithms under the above assumption.
Patriksson \cite{Pat08} presents a survey of the history and applications of
problems (\ref{problem}) and (\ref{problemeq}). Since its publication several
related articles have been published; the survey of \cite{Pat08} is therefore
complemented in Section~\ref{sec:extsur}. Section~\ref{sec:rank} presents a
framework of breakpoint algorithms, resulting in three concrete
representatives. Section~\ref{sec:relaxation} presents a framework of
relaxation algorithms, and ultimately six concrete example methods. In Section
\ref{sec:zen_algor} we describe a quasi-Newton method, due to Nielsen and
Zenios~\cite{NiZ92}, for solving the problem
(\ref{problemeq}). Section~\ref{sec:method} describes the numerical study. A
test problem set is specified and the performance profile used for the
evaluation is defined. In Section~\ref{sec:compexp}, we analyze the results
from the numerical study. The structure is such that we first compare the
relaxation algorithms, second the pegging process, and third the best
performing algorithms among these two with the quasi-Newton method. Finally,
we draw overall conclusions.
\section{Extension of the survey in \cite{Pat08} } \label{sec:extsur}
We here extend the survey in \cite{Pat08}, using the same taxonomy, and sorted
according to publication date.
\begin{description}
\item[{\cite{Mje83}}] {\sc K.\ M.\ Mjelde}, {\em Methods of the Allocation of
Limited Resources}, Section 4.7
\begin{description}
\item[{\sf (Problem)}] $\phi_j \in C^2$; linear equality ($a_j=1$); $l_j=0$
\item[{\sf (Methodology)}] The ranking algorithm of \cite{LuG75}
\item[{\sf (Citations)}] Applications in capital budgeting
(\cite{Han68,Shi77}), cost-effectiveness problems
(\cite{Kir68,Pac70,Mje78}), health care (\cite{Fet73}), marketing
(\cite{LuG75}), multiobjective optimization (\cite{Geo67}), portfolio
selection (\cite{JuD75}), production (the internal report leading to
\cite{BiH79}), reliability (\cite{Bod69}), route-planning for ships or
aircraft (\cite{DBR66}), search (\cite{ChC58}), ship loading (\cite{Kyd69}),
and weapons selection (\cite{Dan67})
\item[{\sf (Notes)}] A monograph on resource allocation problems containing a
comprehensive overview of the resource allocation problem, including
extensions to several resources, non-convex or non-differentiable
objectives, integral decision variables, fractional programming
formulations, etcetera.
\end{description}
\end{description}
\begin{description}
\item[{\cite{Sho85}}] {\sc N.\ Z.\ Shor}, {\em Minimization Methods for
Nondifferentiable Functions}, Section 4.2
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{1}{2}(x_j-y_j)^2$; linear equality
($a_j=1$); $l_j=0$
\item[{\sf (Methodology)}] Pegging
\item[{\sf (Citations)}] \cite{ShI69}, in which the motivating linear
programming application is described
\item[{\sf (Notes)}] The problem arises within the framework of a right-hand
side allocation algorithm for a large-scale linear program.
\end{description}
\end{description}
\begin{description}
\item[{\cite{HuZ05}}] {\sc Z.-S.\ Hua and B.\ Zhang}, {\em Direct algorithm for
separable continuous convex quadratic knapsack problem} (in Chinese)
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$; linear
inequality ($a_j > 0$); $l_j=0$
\item[{\sf (Methodology)}] Pegging
\item[{\sf (Citations)}] Algorithms for the problem
(\cite{PaK90,MeR00,BrS02,BrS02b}) as well as for the case of integer
variables
\item[{\sf (Notes)}] A numerical illustration ($n = 6$).
\end{description}
\end{description}
\begin{description}
\item[{\cite{DaF06}}] {\sc Y.-H.\ Dai and R.\ Fletcher}, {\em New algorithms
for singly linearly constrained quadratic programs subject to lower and
upper bounds}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$, $q_j>0$;
$g_j$ convex in $C^2$ with $g'(x_j) > 0$
\item[{\sf (Methodology)}] A combination of a bracketing algorithm on the
Lagrangian dual derivative, and a secant algorithm for the Lagrangian dual
problem
\item[{\sf (Citations)}] Algorithms for the problem
(\cite{HKL80,Bru84,CaM87,PaK90})
\item[{\sf (Notes)}] The problem arises as a subproblem in a gradient
projection method for a general quadratic programming problem over a scaled
simplex.
\end{description}
\end{description}
\begin{description}
\item[{\cite{LiS06}}] {\sc D.\ Li and X.\ Sun}, {\em Nonlinear Integer
Programming}, Chapter 6: {\rm Nonlinear Knapsack Problems}, Section 6.1:
{\rm Continuous--Relaxation-based Branch--and--Bound Methods}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ and $g_j$ increasing; $g_j$ convex in $C^2$
with $g'_j > 0$
\item[{\sf (Methodology)}] Multiplier search
\item[{\sf (Citations)}] Multiplier search methods (\cite{BrS95}), pegging
methods (\cite{BrS02,BrS02b})
\item[{\sf (Notes)}] The problem arises as a subproblem in branch--and--bound
methods for the integer programming version of the problem, such as for the
quadratic knapsack problem, stratified sampling, manufacturing capacity
planning, linearly constrained redundancy optimization in reliability
networks, and linear cost minimization in reliability networks.
\end{description}
\end{description}
\begin{description}
\item[{\cite{DSV07}}] {\sc K.\ Dahiya, S.\ K.\ Suneja and V.\ Verma}, {\em
Convex programming with a single separable constraint and bounded variables}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$, $q_j>0$;
$g_j$ convex in $C^2$ with $g'(x_j) > 0$; studies also the special case of a
linear equality
\item[{\sf (Methodology)}] Iterative descent process using strictly convex
quadratic separable approximations of a nonseparable original objective $f
\in C^2$; subproblems solved using the pegging algorithm of \cite{Ste01}
\item[{\sf (Citations)}] General references on convex programming over box
constraints; \cite{HKL80,DFL86,PaK90} for example algorithms for separable
convex programming
\item[{\sf (Notes)}] Numerical QP ($n=6$, $g_j$ quadratic) illustration;
numerical comparison with an augmented Lagrangian algorithm for a small
problem ($n=2$).
\end{description}
\end{description}
\begin{description}
\item[{\cite{Kiw07}}] {\sc K.\ C.\ Kiwiel}, {\em On linear-time algorithms for
the continuous quadratic knapsack problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$, $q_j>0$;
linear equality
\item[{\sf (Methodology)}] Breakpoint search algorithm applying median search
of all breakpoints
\item[{\sf (Citations)}] Breakpoint search algorithms:
\cite{Bru84,CaM87,PaK90,MSMJ03}; sorting and searching methods:
\cite{Knu98,Kiw05}
\item[{\sf (Notes)}] Develops a general $O(n)$ breakpoint algorithm;
shows that the algorithms of \cite{PaK90,MSMJ03} may fail even on small
examples; presents a modification of the breakpoint removal in the algorithm
of \cite{CaM87}. Numerical experiments ($n \in [50 \cdot 10^3, 2 \cdot
10^6]$) for uncorrelated, weakly, and strongly correlated data; the new
algorithm wins in CPU time over those in \cite{Bru84} and \cite{CaM87} by
23\%, and 21\%, respectively, on average.
\end{description}
\end{description}
\begin{description}
\item[{\cite{Kiw08a}}] {\sc K.\ C.\ Kiwiel}, {\em Breakpoint searching
algorithms for continuous quadratic knapsack problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$, $q_j>0$;
linear equality
\item[{\sf (Methodology)}] A family of breakpoint search algorithms that
include several choices of breakpoints for a median search and updates of
quantities used for evaluating the piecewise linear implicit constraint
function at the median point
\item[{\sf (Citations)}] Applications in resource allocation
(\cite{BiH81,BrS97,HoH95}), hierarchical production planning (\cite{BiH81}),
network flows and transportation (\cite{HKL80,ShM90,Ven91,NiZ92,CoH94}),
constrained matrix problems (\cite{CDZ86}), quadratic integer programming
(\cite{BSS95,BSS96,HoH95}), Lagrangian relaxation (\cite{HWC74}), and
quasi-Newton methods (\cite{CaM87}); $O(n \log n)$ sorting algorithms for
the solution of the Lagrangian dual problem (\cite{HWC74,HKL80}), $O(n)$
algorithms based on median search
(\cite{Bru84,CaM87,MaD89,PaK90,CoH94,HoH95,MMP97}) and approximate median
search methods with $O(n)$ average-case performance (\cite{PaK90}); primal
pegging algorithms with $O(n^2)$ worst-case performance
(\cite{Zip80,BiH81,Mic86,Ven91,RJL92,BSS96})
\item[{\sf (Notes)}] Develops several variants of $O(n)$ breakpoint search
algorithms, including some ideas earlier proposed in, e.g.,
\cite{PaK90,CoH94,HoH95,MMP97}; remarks that the more complex choices made
in \cite{MaD89,PaK90,CoH94,HoH95,MMP97} also means that for some simple
cases cycling may occur, and also provides convergent modifications for each
of them. Numerical experiments ($n \in [50 \cdot 10^3, 2 \cdot 10^6]$) for
uncorrelated and weakly and strongly correlated data, and for both exact and
inexact computations of the median; comparisons made with the $O(n)$
versions from \cite{Bru84,CaM87}, reporting that a version (Algorithm 3.1)
using exact medians is about 20\% faster than the other ones; refers to an
as yet unavailable technical report from 2006 for more extensive tests and
comparisons with pegging methods.
\end{description}
\end{description}
\begin{description}
\item[{\cite{Kiw08b}}] {\sc K.\ C.\ Kiwiel}, {\em Variable fixing algorithms
for continuous quadratic knapsack problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$, $q_j>0$;
linear equality
\item[{\sf (Methodology)}] Pegging
\item[{\sf (Citations)}] Applications: same references as in \cite{Kiw08a};
breakpoint search methods
(\cite{HWC74,HKL80,CaM87,MaD89,PaK90,CoH94,HoH95,MMP97,MSMJ03,Kiw07};
pegging methods (\cite{LuG75,BiH79,BiH81,Sho85,Mic86,Ven91,RJL92,BSS96})
\item[{\sf (Notes)}] Develops a basic pegging algorithm and proposes several
implementational choices for the solution of the reduced problem and the
updates; shows that the algorithms in \cite{Mic86,RJL92,BSS96} fail on a
small example, and that there is a gap in the convergence analysis in
\cite{BiH81} (which also is closed) that affects algorithms whose analyses
rest on that in \cite{BiH81} (e.g., \cite{Ven91}); provide more efficient
versions of several of these methods, including the introduction of
incremental updates which reduce computations, and a more efficient stopping
criterion. Numerical experiments ($n \in [50 \cdot 10^3, 2 \cdot 10^6]$) for
uncorrelated and weakly and strongly correlated data; comparisons made with
the breakpoint search method of \cite{Kiw08a} which uses exact medians; on
average the latter is 14\% slower while at the same time it has a more
stable run time; it is remarked that the advantage of pegging over
breakpoint search has been reported also in \cite{Ven91,RJL92}.
\end{description}
\end{description}
\begin{description}
\item[{\cite{ZhH08}}] {\sc B.\ Zhang and Z.\ Hua}, {\em A unified method for a
class of convex separable nonlinear knapsack problems}
\begin{description}
\item[{\sf (Problem)}] $\phi'_j/g'_j$ monotone and invertible, $g'_j$
positive; consider $\sum_{j=1}^n g_j(x_j) \lhd \; b$ where $\lhd \in \{ \le,
=, \ge \}$
\item[{\sf (Methodology)}] Pegging algorithm using binary search on the value
of $\phi'_j/g'_j$
\item[{\sf (Citations)}] Applications: resource allocation
(\cite{LuG75,Zip80,BiH81,Hoc94}), the singly constrained multi-product
newsvendor problem (\cite{HaW63,Erl00,AMM04}), production/inventory problems
(\cite{BSSW94,BSS95,BrS02}), stratified sampling (\cite{BSS95,BRS99,BrS02}),
``core subproblem'' (\cite{JuD75,AHKL80,MoT89,RJL92,BSS95,BSS96});
algorithms: breakpoint methods (\cite{BrS95}, \cite{Ste01}, \cite{LuG75})
and relaxation methods (\cite{KoL98}, \cite{BrS02}, \cite{BrS02b})
\item[{\sf (Notes)}] Claims (without a proof) that the complexity of the
algorithm is $O(n)$, but the algorithm presented makes use of a mean-value
method for the evaluation of the breakpoints which in the worst case results
in $2n-1$ iterations. We also note that each iteration consist of $O(n)$
operations, whence the complexity is $O(n^2)$. Numerical experiments ($n \in
[10, 10^4]$) ($\phi$ quadratic, $g_j$ linear, $\lhd$ chosen uniformly
randomly).
\end{description}
\end{description}
\begin{description}
\item[{\cite{WaW12}}] {\sc A.\ De Waegenaere and J.\ L.\ Wielhouwer}, {\em A
breakpoint search approach for convex resource allocation problems with
bounded variables}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ and $g_j$ convex; $g_j$ strictly monotone and
$g_j([\phi_j'/g_j']^{-1})$ is either strictly increasing or strictly
decreasing for all $j$
\item[{\sf (Methodology)}] Breakpoint search algorithm using a refined pegging
method (5-sets pegging); generalizes the quadratic breakpoint algorithm in
\cite{PaK90} and its extension in \cite{Kiw08a} such that it applies for the
problem setting in \cite{BrS02}
\item[{\sf (Citations)}] Generalizes the quadratic breakpoint algorithm in
\cite{PaK90} and its extension in \cite{Kiw08a} such that it is valid for
$f_j$ and $g_j$ as in \cite{BrS02}. Applications in resource allocation
\cite{PaK90,NiZ92,BrS95,VeW95,BSS96,BrS02,WaW02,CLZ03,BSSV06,Pat08}
\item[{\sf (Notes)}] Discuss their algorithms' advantages compared to other
articles presented in \cite{Pat08}.
\end{description}
\end{description}
\begin{description}
\item[{\cite{KiW12}}] {\sc G.\ Kim and C.-H.\ Wu},
{\em A pegging algorithm for separable continuous nonlinear knapsack problems
with box constraints}
\begin{description}
\item[{\sf (Problem)}] $\phi'_j$ invertible: linear equality ($a_j > 0$)
\item[{\sf (Methodology)}] Pegging; improves the pegging algorithm in
\cite{BiH81} by allowing the primal box constraints to be checked
implicitly, using bounds on their Lagrange multipliers
\item[{\sf (Citations)}] Applications to portfolio selection \cite{Mar52},
multicommodity network flows \cite{AHKL80}, transportation \cite{OhK84},
production planning \cite{Tam80}
\item[{\sf (Notes)}] Compares their methodology with the method in
\cite{BiH81} on random test problems. On randomized quadratic continuous
knapsack problems ($n \in [5 \cdot 10^3, 2 \cdot 10^6]$) their algorithms
wins by 8--10\%, except for the smallest problems where the method in
\cite{BiH81} wins by $12.5\%$. Two other types of test problems are also
investigated, wherein the quadratic continuous knapsack problem arises as a
subproblem: quadratic network flows, and portfolio optimization. In the
former case, the algorithm (based on conjugate gradients) is taken from
\cite{Ara00}. Here, the pegging algorithm proposed wins over that in
\cite{BiH81} by 10--48\%, with $n \in [200, 5 \cdot 10^3]$. In the
portfolio optimization algorithm, which is based on a progressive hedging
method described in \cite{Ara00}, the quadratic continuous knapsack problem
arises as a subproblem. Here, the range of $n$ is not completely disclosed;
however, the speed-up over the method in \cite{BiH81} is reported to be
21--25\%.
\end{description}
\end{description}
\begin{description}
\item[{\cite{Lus12}}] {\sc H.\ Luss}, {\em Equitable Resource Allocation:
Models, Algorithms and Applications}, Chapter 2: {\em Nonlinear Resource
Allocation}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ strictly convex; $g_j(x_j) = x_j$
\item[{\sf (Methodology)}] Pegging
\item[{\sf (Citations)}] \cite{Koo53} as an origin; \cite{KoL98} for
computational examples; \cite{Pat08} as a survey
\item[{\sf (Notes)}] This book extends the resource allocation problem
(discussed only in Chapter 2) in several ways, including equitable
optimization through the use of minmax/maxmin objectives, multiple time
periods, substitutable resources, network constraints, and integer decision
variables.
\end{description}
\end{description}
\begin{description}
\item[{\cite{ZhC12}}] {\sc B.\ Zhang and B.\ Chen}, {\em Heuristic and exact
solution method for convex nonlinear knapsack problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ strictly convex; $g_j(x_j) = x_j$
\item[{\sf (Methodology)}] The problem at hand is a subproblem in a
branch--and--bound procedure for the solution of an integer-restricted
version of the problem
\item[{\sf (Citations)}] Applications to the newsvendor problem
(\cite{Erl00,AMM04}), resource allocation (\cite{BiH81,Hoc94}), production
(\cite{BSS95}), and stratified sampling (\cite{BRS99}); efficient methods
for the continuous relaxation (\cite{BrS95,KoL98,Ste01,ZhH08}); heuristics
for the integer program based on rounding of the solution to the continuous
relaxation (\cite{BrS95,HuZ06}); algorithms for the integer program based on
the solution of continuous problems and branch--and--bound
(\cite{BrS95,BrS02})
\item[{\sf (Notes)}] Utilizing the algorithm from \cite{ZhH08} to solve the
continuous relaxations (and rounding to produce feasible solutions), the
authors develop a branch--and--bound algorithm. It is compared with the
methods from \cite{BrS95} as well as with branch--and--bound algorithms
utilizing a variety of tree search principles, on instances with quadratic
objectives, according to problem generation principles from \cite{BrS02} ($n
\in \{10, 15, 500, 1000, 2000\}$).
\end{description}
\end{description}
\begin{description}
\item[{\cite{BGRS13}}] {\sc L.\ Bay{\'o}n, J.\ M.\ Grau, M.\ M.\ Ruiz, and
P.\ M.\ Su{\'a}rez}, {\em An exact algorithm for the continuous quadratic
knapsack problem via infimal convolution}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ strictly convex quadratic; linear equality
($a_j = 1$)
\item[{\sf (Methodology)}] Sorting of breakpoints
\item[{\sf (Citations)}] Previous algorithms for the problem
(\cite{MeT93,CoH94})
\item[{\sf (Notes)}] Two numerical applications: (1) an economic dispatch
problem ($n = 5$), and an academic example ($n \in [200, 10^4]$); in the
latter example the results are favourably compared with the Matlab solver
QUADPROG.
\end{description}
\end{description}
\begin{description}
\item[{\cite{DHH13}}] {\sc T.\ A.\ Davis, W.\ W.\ Hager, and
J.\ T.\ Hungerford}, {\em The separable convex quadratic knapsasck problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$; $q_j>0$;
linear equality
\item[{\sf (Methodology)}] Breakpoint search utilizing a heap data structure,
based on an initial multiplier estimate using a secant Newton method
\item[{\sf (Citations)}] Applications (\cite{HKL80,CaM87,ShM90,NiZ92,DaF06});
multiplier algorithms (\cite{HKL80,Bru84,CaM87,PaK90,MSMJ03}); pegging
methods (\cite{BiH81,Sho85,Mic86,Ven91,RJL92,BSS96,Kiw08b}); quasi-Newton
methods (\cite{DaF06,CMS14})
\item[{\sf (Notes)}] Numerical experiments ($n = 6.25 \cdot 10^6$) on random
test problems, examining (a) the best number of initial (secant) Newton
iterations, and (b) the performance against a primal pegging algorithm.
Overall, breakpoint search is favourable on its own given a good initial
mulitiplier estimate; otherwise, 3 or 4 iterations of the secant Newton
method is a good initialization procedure.
\end{description}
\end{description}
\begin{description}
\item[{\cite{FrG13}}] {\sc A.\ Frangioni and E.\ Gorgone}, {\em A library for
continuous convex separable quadratic knapsack problems}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$; $q_j>0$;
$g_j(x_j) = x_j$
\item[{\sf (Methodology)}] Compares a breakpoint algorithm with CPLEX and
concludes that the breakpoint algorithm outperforms CPLEX
\item[{\sf (Citations)}] Applications in resource allocation and algorithms
(\cite{Pat08})
\item[{\sf (Notes)}] Presents an open source library for the continuous convex
separable quadratic knapsack problem and concludes that the library can be
useful for further studies of the problem at hand.
\end{description}
\end{description}
\begin{description}
\item[{\cite{ZCCS13}}] {\sc T.\ Zhu, W.\ Chen, J.\ Chen, and W.\ Sun}, {\em
Direct algorithm for continuous separable knapsack problem (in Chinese)}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = \frac{q_j}{2}x^2_j - r_jx_j$; $q_j>0$;
$g_j(x_j) = a_jx_j$
\item[{\sf (Methodology)}] Pegging
\item[{\sf (Citations)}] Algorithms for the more general integer version of
the problem; algorithms by Bretthauer and Shetty
(\cite{BrS95,BrS02,BrS02b}); other specialized algorithms for the problem
(\cite{BiH81,PaK90,RJL92,Ste01,Kiw08b})
\item[{\sf (Notes)}] A detailed numerical example ($n = 8$); favourable
comparisons with a Matlab solver ($n \in \{50, 100, 200\}$).
\end{description}
\end{description}
\begin{description}
\item[{\cite{Zha13}}] {\sc L.\ Zhang}, {\em A Newton-type algorithm for
solving problems of search theory}
\begin{description}
\item[{\sf (Problem)}] $\phi_j(x_j) = -a_j(1 - {\rm exp}(-c_j x_j))$, $a_j,
c_j > 0$; $X$ is a scaled unit simplex
\item[{\sf (Methodology)}] The KKT conditions, as defined in
(\ref{eq:sep_conv_ineq_constr_kkt}), are relaxed into a system of non-smooth
equations through the utilization of the Fischer--Burmeister (\cite{Fis92})
smoothing function; a Newton-like algorithm is then employed for these
equations for a sequence of values of the smoothing parameter. The algorithm
is shown to asymptotically and superlinearly converge to the unique optimal
solution
\item[{\sf (Citations)}] Survey (\cite{Pat08}); methodologies (\cite{Ste04});
applications (\cite{Koo99})
\item[{\sf (Notes)}] Three sets of numerical experiments. Main experiment ($n
\in [10^2, 10^4]$) on randomized problems; compares with the pegging
algorithm in \cite{Ste04}, noting that the proposed algorithm is faster. The
proposed methodology is however terminated based on a nonzero value of the
Fischer--Burmeister smoothing function, whence the final solution need not
be feasible or optimal. Second set of experiments on a problem taken from
\cite{WiG69} ($n = 4$), showing no comparisons. Third experiment on data
from the Bureau of Water Conservatory ($n = 5$), showing no comparisons.
\end{description}
\end{description}
\begin{description}
\item[{\cite{CMS14}}] {\sc R.\ Cominetti, W.\ F.\ Mascarenhas, and
P.\ J.\ S.\ Silva}, {\em A Newton's method for the continuous quadratic
knapsack problem}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ strictly convex quadratic; linear equality
\item[{\sf (Methodology)}] An approximative Newton method for the Lagrangian
dual problem utilizing a secant globalization and variable fixing
\item[{\sf (Citations)}] Applications to resource allocation
(\cite{BiH81,HoH95,BrS97}), multicommodity network flows
(\cite{HKL80,NiZ92,ShM90}), Lagrangian relaxation using subgradient methods
(\cite{HWC74}), quasi-Newton updates with bounds (\cite{CaM87}), semismooth
Newton methods (\cite{FeM04}); related methods (\cite{DaF06,Kiw08b}), where
the latter method is shown to be equivalent to the proposed one when there
are only lower bounds {\sl or} lower bounds
\item[{\sf (Notes)}] Numerical experiments ($n \in [50 \cdot 10^3, 2 \cdot
10^6]$) comparing the proposed method to a secant method (\cite{DaF06}),
breakpoint search (\cite{Kiw08b}), and median search (\cite{Kiw08a} on
problems with uncorrelated, weakly correlated, and correlated data. The
proposed Newton method is overall better---about 30\% better on the larger
instances. A further test is made on the classification problems described
in \cite{DaF06} arising in the training in support vector machines; as the
Hessian is non-diagonal, a projected gradient method is used, leading to
subproblems of the type considered. Here, Newton's method is superior.
\end{description}
\end{description}
\begin{description}
\item[{\cite{WrR14}}] {\sc S.\ E.\ Wright and J.\ J.\ Rohal}, {\em Solving the
continuous nonlinear resource allocation problem with an interior point
method}
\begin{description}
\item[{\sf (Problem)}] $\phi_j$ and $g_j$ convex, and in $C^2$ on an open set
containing $[l_j, u_j]$. Further, $\phi_j$ is decreasing on $[l_j, u_j]$ and
$g_j$ is increasing on $[l_j, u_j]$, and $\sum_{j \in J} g_j(l_j) < b <
\sum_{j \in J} g_j(u_j)$. Test instances include resource renewal
[$\phi_j(x_j) := c_jx_j (e^{-1/x_j} - 1)$ for $x_j > 0$, $\phi_j(x_j) :=
-x_j$ for $x_j < 0$, and $g_j$ linear], weighted $p$-norm over a ball
[$\phi_j(x_j) := c_j(x_j - y_j)^p$ and $g_j(x_j) := |x_j|^r$ with $p, r
\in \{2, 2.5, 3, 4\}$], sums of powers [$\phi_j(x_j) := c_j|x_j -
y_j|^{p_j}$, and $g_j(x_j) := |x_j|^{r_i}$], convex quartic over a
simplex [$\phi_j$ a fourth-power polynomial, and $g_j(x_j) := x_j$], and
log-exponential [$\phi_j(x_j) := \ln [\sum_i {\rm exp}(a_{ij}x_j +
d_{ij})]$, $g_j$ linear]
\item[{\sf (Methodology)}] A damped feasible interior-point Newton method for
the solution of the KKT conditions
\item[{\sf (Citations)}] Survey (\cite{Pat08}); application to resource
renewal (\cite{MeR00}); methodologies
(\cite{Bru84,PaK90,RJL92,MeR00,Kiw08a})
\item[{\sf (Notes)}] The algorithm is introduced for problems where
subsolutions are not available in closed form. Shows that the linear system
defining the Newton search direction is solvable in $O(n)$ time. Numerical
experiments ($n \in [10^2, 10^6]$) conclude that the interior point method
wins over breakpoint search, often by an order of magnitude.
\end{description}
\end{description}
\section{Breakpoint algorithms}\label{sec:rank}
Algorithms based on the Lagrangian relaxation of the explicit constraint
(\ref{problemb}) have an older history than the relaxation algorithms. This is
probably due to the fact that the relaxation algorithm quite strongly rests on
the Karush--Kuhn--Tucker (KKT) conditions, which did not become widely
available until the end of the 1940s and early 1950s with the work of F.\ John
\cite{Joh48}, W.\ Karush \cite{Kar39}, and H.~W. Kuhn and A.~W. Tucker
\cite{KuT51}. Lagrangian based algorithms have been present much longer and
the famous ``Lagrange multiplier method'' for equality constrained
optimization is classic in the calculus curriculum. Indeed, Lagrange
multiplier techniques for our problem (\ref{problem}) date back at least as
far as to \cite{Bec52}; see \cite{Pat08} for a survey of the history of the
problem.
Considering problem (\ref{problem}) and introducing the Lagrange multiplier
$\mu$ for constraint (\ref{problemb}), we obtain the following conditions for
the optimality of $\mathbf{x}^*$ in (\ref{problem}):
\begin{subequations}
\label{eq:sep_conv_ineq_constr_kkt}
\begin{align}
\mu ^* \geq 0, \quad g(\mathbf{x}^*) & \leq b, \quad \mu^* (
g(\mathbf{x}^*)-b) = 0,
\label{eq:sep_conv_ineq_constr_kkt1} \\
x_j^* & \in X_j, \qquad j \in J, \label{eq:sep_conv_ineq_constr_kkt2}
\end{align}
and
\begin{align}
x_j^* = l_j, & \qquad \text{ if } \phi_j'(x_j^*) \geq - \mu^* g_j'(x_j^*),
\qquad j \in J, \label{eq:sep_conv_ineq_constr_kkt3} \\ x_j^* = u_j, & \qquad
\text{ if } \phi_j'(x_j^*) \leq - \mu^* g_j'(x_j^*), \qquad j \in
J, \label{eq:sep_conv_ineq_constr_kkt4} \\ l_j \leq x_j^* \leq u_j, & \qquad
\text{ if } \phi_j'(x_j^*) = - \mu^* g_j'(x_j^*), \qquad j \in J.
\label{eq:sep_conv_ineq_constr_kkt5}
\end{align}
\end{subequations}
For a fixed optimal value $\mu^*$ of the Lagrange multiplier the conditions
(\ref{eq:sep_conv_ineq_constr_kkt3})--(\ref{eq:sep_conv_ineq_constr_kkt5}) are
the optimality conditions for the minimization over $\mathbf{x} \in
\prod_{j=1}^n X_j$ of the Lagrangian function defined on $\prod_{j=1}^n X_j
\times \mathbb{R}_+$,
\[
L(\mathbf{ x}, \mu) := -b\mu + \sum_{j = 1}^n \{ \phi_j(x_j) + \mu g_j(x_j) \}.
\]
Given $\mu \geq 0$ its minimization over $\mathbf{ x} \in \prod_{j = 1}^n X_j$
separates into $n$ problems, yielding the Lagrangian dual function
\begin{equation}
\label{eq:sep_conv_gen_dual}
q(\mu) := -b \mu + \sum_{j = 1}^n \mathop{\rm minimum}_{x_j \in X_j} \, \{
\phi_j(x_j) + \mu g_j(x_j) \}.
\end{equation}
By introducing additional properties of the problem, we can ensure that the
function $q$ is not only concave but finite on $\mathbb{R}_+$ and
moreover differentiable there. Suppose, for example, that for each $j$,
$\phi_j(\cdot) + \mu g_j(\cdot)$ is weakly coercive on $X_j$ for every $\mu
\geq 0$ [that is, that either $X_j$ is bounded or that for every $\mu \geq 0$,
$\phi_j(x_j) + \mu g_j(x_j)$ tends to infinity whenever $x_j$ tends to $\pm
\infty$], and that $\phi_j$ is strictly convex on $X_j$. In this case
the derivative $q'$ exists on $\mathbb{R}_+$ and equals
\[
q'(\mu) = \phi'_j(x_j(\mu)) + \mu g'_j(x_j(\mu)),
\]
where $\mathbf{ x}(\mu)$ is the unique minimum of the Lagrange function
$L(\cdot , \mu)$ over $\prod_{j=1}^n X_j$. Thanks to this simple form of the
dual derivative, the maximum $\mu^*$ of $q$ over $\mathbb{R}_+$ can be characterized
by the complementarity conditions (\ref{eq:sep_conv_ineq_constr_kkt1}), and
the conditions (\ref{eq:sep_conv_ineq_constr_kkt}) are the primal--dual
optimality conditions for the pair of primal--dual convex programs.
If we assume that $\mu^* \neq 0$, we search for $\mu^* > 0$ such that
$q'(\mu^*) = 0$ [or, in other words, $g(\mathbf{ x}(\mu^*)) = b$], that is, we
need to solve a special equation in the unknown entity $\mu$, where the
function $q'$ is implicitly defined, but is known to be decreasing since $q$
is concave. This equation can of course be solved through the use of any
general such procedure [for example, {\sl bisection search} takes two initial
values $\overline{\mu}$ and $\underline{\mu}$ with $q'(\overline{\mu}) < 0$
and $q'(\underline{\mu}) > 0$, then iteratively cancels part of the initial
interval given the sign of $q'$ at its midpoint $(\overline{\mu} +
\underline{\mu})/2$], but the structure of $q'$ makes specialized algorithms
possible to utilize.
From the above optimality conditions for the Lagrangian minimization problem,
we obtain that
\begin{equation}\label{xmu}
x_j (\mu) = \begin{cases} l_j, & \text{if } \mu \geq \mu_j^l :=
-\phi'_j(l_j)/g'_j(l_j), \cr u_j, & \text{if } \mu \leq \mu_j^u := -
\phi'_j(u_j)/g'_j(u_j), \qquad j \in J. \cr x_j, & \text{if }
\phi'_j(x_j) + \mu g'_j(x_j) = 0,
\end{cases}
\end{equation}
In a rudimentary algorithm we order these indices (or, breakpoints) $\mu_j^l$
and $\mu_j^u$ in an increasing (for example) order into $\{ \mu_1, \dotsc,
\mu_N \}$, where $N \leq 2n$ due to the possible presence of ties. Finding
$\mu^*$ then amounts to finding an index $\jmath^*$ such that that
$q'(\mu_{\jmath*}) > 0$ and $q'(\mu_{\jmath* + 1}) < 0$; then we know that
$\mu^* \in (\mu_{\jmath*}, \mu_{\jmath* +1})$ and $q'(\mu^*) = 0$. Hence from
equation (\ref{xmu}), we know for all $j$ if $x_j^* = l_j$, $x_j^* = u_j$ or
$l_j < x_j^* < u_j$. Now by fixing all variables $x_j^*$ that equal the
corresponding lower or upper bound we can ignore the bound constraint
(\ref{problemc}) and find an analytical solution of the problem.
Two decisions thus need to be made: how to find the index $\jmath^*$, and how
to perform the interpolation. Starting with the former, the easiest means is
to run through the indices in ascending or descending order to find the index
where $q'$ changes sign. If we have access to the indices $\jmath^+$ and
$\jmath^-$ for which $q'(\mu_{\jmath^+}) > 0$ while $q'(\mu_{\jmath^-}) < 0$,
then we can choose the midpoint index, check the corresponding sign of $q'$,
and reduce the index set accordingly. Given the sorted list, we can also find
this index in some randomized fashion.
As remarked above, algorithms such as bisection search can be implemented
without the use of the breakpoints, and therefore without the use of sorting,
as long as an initial interval can somehow be found; also general methods for
solving the equation $q'(\mu) = 0$, such as the secant method or regula falsi,
can be used even without an initial interval; notice however that $q \not\in
C^2$, whence a pure Newton method is not guaranteed to be well-defined.
While the sorting operation used in the ranking and bisection search methods
takes $O(n \log n)$ time, it is possible to lower the complexity by choosing
the trial index based on the {\sl median} index, which is found without the
use of sorting; the complexity of the algorithm is then reduced to $O(n)$. It
is not clear, however, that the latter must always be more efficient, since
the ``$O$'' definition calls for $n$ to be ``large enough''.
We also remark that in the case when the problem (\ref{problem})
arises as a subproblem in an iterative method, as the method converges the
data describing the problem will tend to stabilize. This fact motivates the
use of reoptimization of the problems, which most obviously can be done by
using the previous value of the Lagrange multiplier as a starting point and/or
utilizing the previous ordering of the breakpoints; in the latter case, the
$O(n \log n)$ sorting complexity will eventually drop dramatically.
In Section \ref{sec:equalityconstraint} we consider the breakpoint algorithm
for the equality problem (\ref{problemeq}). In Section \ref{sec:pegbreak} we
describe three pegging methods and in Sections
\ref{sec:medians}--\ref{sec:MB5} we apply these pegging methods to the
breakpoint algorithm. Finally, in Section \ref{sec:conbreak} we briefly discuss
the convergence and time complexity of the breakpoint algorithm.
\subsection{Equality constraints}\label{sec:equalityconstraint}
We now consider problem (\ref{problemeq}) where the inequality of the primal
constraint is replaced by an equality. For the problem to be convex, the
resource constraint (\ref{problemeqb}) has to be affine, i.e.,
$g(\mathbf{x}):=\sum_{j \in J} a_jx_j - b$. Beside the resource constraint,
the Lagrangian and the optimality conditions will take the same form as for
problem (\ref{problem}) but with one important difference; $\mu$ is
unrestricted in sign, whence the condition
(\ref{eq:sep_conv_ineq_constr_kkt1}) is replaced by ``$g(\mathbf{x}^*) = b$''.
\subsection{The pegging process}\label{sec:pegbreak}
The origin of the pegging process is found in the relaxation algorithm; see,
e.g., \cite{BiH81}. The purpose of pegging is to predict properties of the
primal variables in the optimal solution from an arbitrary dual value. In
Sections \ref{sec:taop}, \ref{sec:3set}, and \ref{sec:5set} we show how to
predict if an optimal primal variable value equals its lower or upper bound,
or is strictly within any of the bounds.
\subsubsection{2-sets pegging}\label{sec:taop}
If we can determine if a variable equals its lower bound at the optimal
solution then we add its variable index to a set $L$ and reduce the original
problem. Similarly, if we know that a variable equals its upper bound at the
optimal solution then we add the variable index to the set $U$. Using the sets
$L$ and $U$ when solving problem (\ref{problem}) or (\ref{problemeq}) will be
referred to as 2-sets pegging.
Assume that we have a lower limit $\underline{\mu}$ and an upper limit
$\overline{\mu}$ on the optimal dual value, that is, $\underline{\mu} \le
\mu^* \le \overline{\mu}$. From (\ref{xmu}) we can define the sets
$L(\underline{\mu}) := \{ j \in J \mid \underline{\mu} \ge -\phi_j^\prime(l_j)
/ g_j^\prime(l_j) \}$ and $U(\overline{\mu}) := \{ j \in J \mid \overline{\mu}
\le -\phi_j^\prime(u_j) / g_j^\prime(u_j) \}$. Let $J^k := J \setminus \{
L(\underline{\mu}) \cup U(\overline{\mu}) \}$ and let $b^k := b - \sum_{j \in
L (\underline{\mu})} g_j(l_j) -\sum_{j \in U(\overline{\mu})}
g_j(u_j)$. Hence we can define a subproblem of problem (\ref{problem}) as
follows:
\begin{subequations}
\begin{align}
\minimize_x &\quad \phi(\mathbf{x}) := \sum _{j\in J^k}
\phi_j(x_j), \\ \subto & \quad g(\mathbf{x}):= \sum
_{j\in J^k} g_j(x_j) \le b^k, \label{problemb2} \\
& \quad l_j \le x_j \le u_j, \quad j \in J^k. \label{problemc2}
\end{align}
\label{problem2}
\end{subequations}
Similarly we can define a subproblem of problem (\ref{problemeq}) as follows:
\begin{subequations}
\begin{align}
\minimize_x &\quad \phi(\mathbf{x}) := \sum _{j \in J^k}
\phi_j(x_j),\\ \subto & \quad g(\mathbf{x}):= \sum _{j\in
J^k} a_jx_j = b^k, \label{problemeqb2} \\
& \quad l_j \le x_j \le u_j, \quad j \in J^k. \label{problemeqc2}
\end{align}
\label{problemeq2}
\end{subequations}
\noindent Consider problem (\ref{problem2}). Assuming that $\mu^*>0$, the
constraint (\ref{problemb}) has to be fulfilled with equality. For any given
dual variable $\mu^k$ we can determine the primal solution $\mathbf{x}^k(\mu
^k)$ of problem (\ref{problem2}) from condition (\ref{xmu}). We know from
Section \ref{sec:rank} that all optimality conditions
(\ref{eq:sep_conv_ineq_constr_kkt}) except the resource constraint
(\ref{problemb}) are satisfied. Moreover, we know that the resource constraint
has to be fulfilled with equality, in order for the solution to be
optimal. Substituting $\mathbf{x}^k$ into the resource constraint will be
referred to as explicit evaluation; this leaves us with three cases, namely
\begin{subequations}
\begin{align}
\sum_{j\in J^k}g_j(x_j^k) & = b^k \label{caseopt}, \\
\sum_{j\in J^k}g_j(x_j^k) & < b^k \label{caselow}, \mathrm{ or } \\
\sum_{j\in J^k}g_j(x_j^k) & > b^k \label{casehigh}.
\end{align}
\end{subequations}
If (\ref{caseopt}) is fulfilled for $\mathbf{x}^k$ then all optimality
conditions are met and $\mathbf{x}^* = \mathbf{x}^k$. Consider next the case
(\ref{caselow}); clearly $\mathbf{x}^k$ is not optimal but we know that
$\mathbf{x}^*$ is such that $\sum_{j \in J^k} g_j(x_j^*) > \sum_{j \in J^k}
g_j(x_j^k)$ since $\sum_{j \in J^k} g_j(x_j^*) = b^k$. The function $g_j$ is
convex and differentiable and can increase in one interval and decrease in
another (consider, e.g., $g_j(x_j) = x^2_j$), which implies that no
predictions can be made of the size of $x_j^*$ relative to that of $x_j^k$.
Hence, we need $g_j$ to be monotone. For problem (\ref{problem}), Bretthauer
and Shetty \cite[Section 2]{BrS02} consider four cases equivalent to the
following:
\begin{description}
\item[Case 1:] For all $j\in J$, $g_j$ is decreasing and
$\mu(x_j) := -\phi'_j (x_j) / g'_j (x_j)$ is increasing in $x_j$.
\item[Case 2:] For all $j\in J$, $g_j$ is increasing and
$\mu(x_j)$ is decreasing in $x_j$.
\item[Case 3:] For all $j\in J$, $g_j$ is decreasing and
$\mu(x_j)$ is decreasing in $x_j$.
\item[Case 4:] For all $j\in J$, $g_j$ is increasing and
$\mu(x_j)$ is increasing in $x_j$.
\end{description}
If Case 3 or 4 holds it is possible to find a closed form of the optimal
solution to the problem (\ref{problem}), see \cite[Proposition
10]{BrS02}. Considering problem (\ref{problemeq}), Cases 3 and 4 cannot
occur, since the resource constraint is affine, i.e., $\mu(x_j) := -\phi'_j
(x_j) / a_j$. Hence, only Cases 1 and 2 are of interest here.
Note that if $\mu(x_j)$ is increasing in $x_j$ then $x_j(\mu)$ is
nondecreasing and vice versa. This is an essential property for the validity
of the pegging process. We can indeed state the following proposition (a
similar one can be stated for problem (\ref{problemeq}), but without the
assumption $\mu^* > 0$):
\begin{proposition}[pegging for Cases 1 and 2]\label{pegglow}
Consider problem (\ref{problem}) and assume that $\mu^* > 0$.
(i) If Case 1 holds, and if (\ref{caselow}) holds, then $x_j^* = l_j$ for all
$j \in L(\mu^k)$.
(ii) If Case 1 holds, and if (\ref{casehigh}) holds, then $x_j^* = u_j$ for
all $j \in U(\mu^k)$.
(iii) If Case 2 holds, and if (\ref{caselow}) holds, then $x_j^* = u_j$ for
all $j \in U(\mu^k)$.
(iv) If Case 2 holds, and if (\ref{casehigh}) holds, then $x_j^* = l_j$ for
all $j \in L(\mu^k)$.
\end{proposition}
\begin{proof}
A proof of (i) is given; the proofs for (ii), (iii), and (iv) are
analogous. From (\ref{caselow}) we have that
\[ \sum_{j\in J^k}g_j(x_j(\mu^k)) < b^k.\]
In Case 2, for all $j$ $g_j$ is increasing and $x_j(\mu)$ is nonincreaing,
which implies that $g_j(x_j(\mu))$ is nonincreasing in $\mu$ for all
$j$. Hence, we have that $\mu^k = \overline{\mu} \ge \mu^*$ which implies that
$x^k_j \le x^*_j$ for all $j$ since $x_j(\mu)$ is nonincreasing in $\mu$ for
all $j$. Hence, for $j \in U(\mu^k)$ we know that $x^{k+1}_j = u_j = x_j^*$,
i.e., we can peg $j \in U(\mu^k)$.
\end{proof}
\subsubsection{3-sets pegging}\label{sec:3set}
As in the 2-sets pegging principle of Section \ref{sec:taop} we determine if a
variable takes the value of the lower or upper bound at the optimal
solution. Additionally for the 3-sets pegging we determine if a variable
belongs to the open interval between the lower and upper bound, i.e., if
$x_j^* \in (l_j,u_j)$. Assume that we know that $\mu^* \in (\mu_j^{u},
\mu_j^l)$; then it follows from (\ref{xmu}) that $x_j^* \in (l_j,u_j)$ and
there is no need to check if $x_j^k$ equals the lower or upper bound which
will reduce future calculations. 3-sets pegging is used for the quadratic
knapsack problem (\ref{quadratic}) in \cite[Section 3]{Kiw08a}. The method
described in \cite[Section 3]{Kiw08a} can be generalized according to the
following proposition:
\begin{proposition}[relax primal variables from lower and upper
bounds]\label{minkiwiel1}
Assume that Case 1 or 2 in Section \ref{sec:taop} holds and that for some
values of $\underline{\mu}$ and $\overline{\mu}$, $\underline{\mu} < \mu^* <
\overline{\mu}$ holds. If $\underline{\mu}, \overline{\mu} \in
[\mu_j^l,\mu_j^u]$ holds for some $j \in \mathbf{J}^k$ then $l_j <
x_j^* < u_j$.
\end{proposition}
\begin{proof}
Assume that Case 2 holds, i.e., $\mu = -\phi'_j/g_j'$ is decreasing. (The
proof for Case 1 is analogous.) Assume that $\underline{\mu}, \overline{\mu}
\in [\mu_j^u,\mu_j^l]$ holds for some $j \in \mathbf{J}^k$. Since
$-\phi'_j/g_j'$ is decreasing we have that $\underline{\mu}, \overline{\mu}
\in [\mu_j^u,\mu_j^l]$ implies that $\mu^* \in (\mu_j^u,\mu_j^l)$, and from
(\ref{xmu}) it then follows that $l_j < x_j^* < u_j$.
\end{proof}
\subsubsection{5-sets pegging} \label{sec:5set}
As in the 3-sets pegging principle in Section \ref{sec:3set} we determine if a
variable takes the value of the lower or upper bound or if the variable
strictly belongs to the interval between the lower and upper bound. For 5-sets
pegging we also determine if a variable is larger than the lower bound or
smaller than the upper bound. 5-sets pegging for problem (\ref{problem}) is
used in \cite{WaW12}, generalizing a method from \cite{Kiw08a}. Assuming that
we know that $x_j^* < u_j$, there is no need to check if $x_j^k$ equals the
upper bound; this might reduce future calculations. The proof of the following
proposition follows from the monotonicity of $g_j$ and $x_j(\mu)$ (see
\cite{WaW12} for a proof).
\begin{proposition}[relax primal variables from lower or upper bound]
\label{wawa}
Assume that Case 1 or 2 in Section \ref{sec:taop} holds. Let $j \in
\mathbf{J}^k$. If $\mu^* < \overline{\mu} \leq \mu^l_j$, then $x_j^* >l_j$ and
if $\mu^*> \underline{\mu} \geq \mu^u_j$ then $x_j^* <u_j$.
\qed
\end{proposition}
\subsection{Algorithm: Median search of Breakpoints with 2-sets pegging
(MB2)}\label{sec:medians}
Consider Case 1 or 2 in Section \ref{sec:taop} for problem
(\ref{problem}). Let $\rm{median}(\cdot)$ denote a function which provides the
median of a finite vector; let $\mu_m$ be the median breakpoint, and define
the total use of the resource due to variables that equal the lower and upper
bounds as $\beta_l := \sum_{ j\in \{ N \mid \mu_j^l \le \mu_m \} } g_j(l_j)$,
and $\beta_u := \sum_{j\in \{ N \mid \mu_j^u \ge \mu_m \}} g_j(u_j)$,
respectively. In the spirit of \cite[Section 3.1]{BrS02b} and \cite[Algorithm
3.1]{Kiw07}, we present the following algorithm:
\begin{tabular}{l}
\\ \textbf{Initialization:} \\ $\qquad$ Set $N := J$, $k:=1$, and $b^k := b$.
\\ $\qquad$ Compute breakpoints $\mu^l:= ( \mu^l_j )_{j \in N }$, $\mu^u:= (
\mu^u_j )_{j \in N }$ as in (\ref{xmu}), and let $\mu^k := \left( -\infty,
\left( \mu^l \right)^\intercal, \left( \mu^u\right)^\intercal, \infty
\right)^\intercal$.
\\
\textbf{Step 0} (check if $\mu = 0$ is optimal):
\\
$\qquad$ If $\sum_{j \in N} g_j (x_j(0)) \le b$, for $x_j(\mu)$ is determined
from (\ref{xmu}), then $\mu^*=0$, $x_j^* = x_j(0)$ for $j \in N$. Stop.
\\
\textbf{Iterative algorithm:}
\\
\textbf{Step 1} (stopping test):
\\
$\qquad$ If $\mu^k = \emptyset$ then find $\mathbf{x}^*$ and $\mu^*$ from
problem (\ref{problem}) relaxed from lower and upper bounds.
\\
$\qquad$ Otherwise, let $\mu_m := \rm{median}(\mu^k)$.
\\
\textbf{Step 2} (compute explicit reference):
\\
$\qquad$ Determine $ \delta := \sum_{ j\in \{ N \mid \mu_j^u < \mu_m < \mu_j^l
\} }
g_j(x_j(\mu_m))
+ \beta_u + \beta_l$, where $x_j(\mu)$ is determined from (\ref{xmu}).
\\
$\qquad$ If $\delta >b^k$, then go to Step 3.1. \\ $\qquad$ If $\delta < b^k$,
then go to Step 3.2. \\
$\qquad$ Otherwise $(\delta = b^k)$ let $\mu^* := \mu_m$, find
$\mathbf{x}^*$ from (\ref{xmu}), and stop.
\\
\textbf{Step 3.1} (update and fix lower bounds):
\\
$\qquad$ For all $j\in N$: If $\mu_m \ge \mu_j^l$ then let
$N := N \setminus \{j\}$ and $x_j^* := l_j$.
\\
$\qquad$ Let $\mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \mid \mu_m < \mu^k_j \}
}$,
$b^{k+1} := b^{k}- \beta_l$, and $k := k+1$.
\\
$\qquad$ Go to Step 1.
\\
\textbf{Step 3.2} (update and fix upper bounds):
\\
$\qquad$ For all $j\in N$: If $\mu_m \le \mu_j^u$ then let $N :=
N \setminus \{j\}$ and $x_j^* := u_j$.
\\
$\qquad$ Set $ \mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \mid \mu_m >
\mu^k_j \}}$,
$b^{k+1} := b^{k}- \beta_u$, and $k := k +1$.
\\ $\qquad$ Go to Step 1.
\\ \\
\end{tabular}
\noindent For problem (\ref{problemeq}), the algorithm is similar except Step
0 vanishes.
\subsection{Algorithm: Median search of Breakpoints with 3-sets pegging (MB3)}
\label{sec:rmm}
If we can determine if the value of a variable $x_j$ is strictly within the
bounds for all $\mu$ such that $\underline{\mu} < \mu < \overline{\mu}$, then
we don't have to check if $x_j$ violates the bounds when we determine $x_j$
from (\ref{xmu}) in future iterations (see Proposition \ref{minkiwiel1}). This
might save us some operations. Define a set of indices for the lower and
upper limit being within the interval of the lower and upper breakpoint for a
variable: $M := \{ \, j \in N \mid \underline{\mu},\overline{\mu} \in
[\mu_j^l,\mu_j^u] \, \}$. From Proposition \ref{minkiwiel1} we have that if
$j\in M$ then $l_j < x_j^* < u_j$. Hence if $j\in M$ we do not have to check
if $x_j$ violates the bounds in future iterations. But we should have in mind
that to determine if $j\in M$ requires some extra operations. If we let the
initial values of the limits be $\underline{\mu} = -\infty$ and
$\overline{\mu} = \infty$, then $\underline{\mu},\overline{\mu} \notin
[\mu_j^l,\mu_j^u]$ and we note that there is no need to check if $j\in M$, as
long as $\underline{\mu} = -\infty$ or $\overline{\mu} = \infty$. Hence, to
avoid unnecessary operations we start by checking if $j\in M$ when
$\underline{\mu} > -\infty$ or $\overline{\mu} < \infty$; this is in contrast
to the algorithms in \cite[Section 3]{Kiw08a} and \cite{WaW12}.
We finally define the contribution to the resource constraint from the
variables including $M$: $\gamma(\mu) := \sum_{j\in M } g_j(x_j(\mu))$. Note
that the value of $\gamma$ depends on the value of the parameters in $\phi_j$,
$g_j$ and $\mu$. If we can separate the parameters from the multiplier $\mu$,
i.e., if $\gamma$ is additively and/or multiplicatively separable [that is,
$\gamma (\mu , \cdot) = \gamma_1(\mu) \gamma_2(\cdot) + \gamma_3 (\cdot)$]
then the values of $\gamma_2$ and $\gamma_3$ can be calculated successively so
that no calculations are done more than once. Consider, for example, the
negative entropy function, $\phi_j(x_j) := x_j\log(x_j/a_j-1)$ and the
resource constraint function $g_j(x_j):=x_j$. Then, $\gamma(\mu,\mathbf{a})
:=\sum_{j\in M} g_j(x_j^{k}(\mu)) = \sum_{j\in M} a_je^{-\mu} =
\gamma_2(\mathbf{a})\gamma_1(\mu)$, where $\gamma_1(\mu) = e^{-\mu}$ and
$\gamma_2(\mathbf{a}) = \sum_{j\in M} a_j$, i.e., we update $\gamma_2$ in
Steps 3.1 and 3.2 such that if $\underline{\mu},\overline{\mu} \in
[\mu_j^u,\mu_j^l]$ then $\gamma_2 := \gamma_2 + a_j$. For the quadratic
knapsack problem, this approach is applied in \cite[Section 3]{Kiw08a}. We
next present an algorithm that applies the usage of $M$; Steps 0 and 1 are
similar to the algorithm MB2 in Section \ref{sec:medians}:
\begin{tabular}{l}
\\ \textbf{Initialization:}
\\
$\qquad$ Set $N := J$, $k:=1$, $b^k := b$, $\gamma := 0$, $M := \emptyset$,
$\underline{\mu} := -\infty$, and $\overline{\mu} := \infty$.
\\
$\qquad$ Compute breakpoints $\mu^l:= ( \mu^l_j )_{j \in N }$, $\mu^u:=
( \mu^u_j )_{j \in N }$ as in (\ref{xmu}), and let $\mu^k :=
\left( \begin{matrix} \mu^l \\ \mu^u \end{matrix} \right)$.
\\
\textbf{Iterative algorithm:}
\\
\textbf{Step 2} (compute explicit reference):
\\ $\qquad$ Determine $ \delta := \sum_{j\in \{ N \mid \mu_j^u < \mu_m < \mu_j^l \}}
g_j(x_j(\mu_m)) + \gamma(\mu_m) + \beta_u + \beta_l$,
\\
$\qquad$ where $x_j(\mu)$ is determined from (\ref{xmu}).
\\
$\qquad$ If $\delta >b^k$, then go to Step 3.1.
\\
$\qquad$ If $\delta < b^k$, then go to Step 3.2.
\\
$\qquad$ Otherwise $(\delta = b^k)$ set $\mu^* := \mu_m$; find optimal
$\mathbf{x}^*$ from (\ref{xmu}), and stop.
\\
\textbf{Step 3.1} (update and fix lower bounds):
\\
$\qquad$ Let $\underline{\mu}=\mu_m$.
\\
$\qquad$ For $j\in N$: If $\underline{\mu} \ge \mu_j^l$ then let $N
:= N \setminus \{j\}$ and $x_j^* := l_j$. \\ $\qquad \qquad\quad\; $
$\quad$ If $\underline{\mu},\overline{\mu} \in [\mu_j^u,\mu_j^l]$
then let $N := N
\setminus \{j\}$ and $M := M \cup \{j\}$; update $\gamma_{1}$ and $\gamma_{2}$.
\\
$\qquad$ Let $\mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \cup M \mid \mu_m < \mu^k_j \}
}$, $b^{k+1} := b^{k}-
\beta_l$, and $k := k+1$.
\\ $\qquad$ Update $\gamma$ and go to Step 1.
\\
\textbf{Step 3.2} (update and fix upper bounds):
\\
$\qquad$ Let $\overline{\mu} := \mu_m$.
\\
$\qquad$ For
$j\in N$: If $\overline{\mu} \le \mu_j^u$ then let $N := N \setminus
\{j\}$ and $x_j^* := u_j$.
\\ $\qquad \qquad\quad\; $ $\quad$ If $\underline{\mu},\overline{\mu} \in
[\mu_j^u,\mu_j^l]$ then let $N := N \setminus \{j\}$ and
$M := M \cup \{j\}$; update $\gamma_{1}$ and $\gamma_{2}$.
\\
$\qquad$ Let $\mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \cup M \mid \mu_m > \mu^k_j
\}}$, $b^{k+1} := b^{k} - \beta_u$, and $k := k +1$.
\\ $\qquad$ Go to Step 1.
\\
\end{tabular}
\begin{remark} If $\gamma$ is not separable then the updates of $\gamma_1$
and/or $\gamma_2$ in Steps 3.1 and 3.2 vanish.
\end{remark}
\subsection{Algorithm: Median search of Breakpoints with 5-set pegging
(MB5)}\label{sec:MB5}
As in the algorithm MB2 in Section \ref{sec:medians} we peg the variables
whose values equal the lower or upper bounds, and as in MB3 we determine if a
variable is strictly within the bounds. Further, we determine if a variable is
smaller than the upper bound or larger than the lower bound as in Section
\ref{sec:5set}. Define a set $L_-$ of indices where the optimal solution is
known to be strictly smaller than the upper bound in the optimal solution,
i.e., $L_- := \{j \in {N} \mid x_j^*<u_j \}$, and similarly define a set $U_+$
of indices where the variable is known to be larger than the lower bound in
the optimal solution, i.e., $U_+:=\{j \in {N} \mid x_j^*>l_j\}$. Define,
respectively, the total use of the resource due to variables whose values
equal the lower and upper bounds as $\beta_l := \sum_{j \in \{ N \cup L_- \mid
\mu_j^l \le \mu_m\} } g_j(l_j)$ and $\beta_u := \sum_{j\in \{ N \cup U_+
\mid \mu_j^u \ge \mu_m \} } g_j(u_j)$. We present an algorithm that applies
5-sets pegging where Steps 0 and 1 are similar to the algorithm MB2 in Section
\ref{sec:medians}:
\begin{tabular}{l}
\\
\textbf{Initialization:}
\\
$\qquad$ Set $N := J$, $k:=1$, $b^k := b$, $\gamma :=0$, and
$M = L_- = U_+ := \emptyset$,
$\underline{\mu} := -\infty$, and $\overline{\mu} := \infty$.
\\
$\qquad$ Compute breakpoints $\mu^l:= ( \mu^l_j )_{j \in N }$, $\mu^u:=
( \mu^u_j )_{j \in N }$ as in (\ref{xmu}), and let $\mu^k :=
\left( \begin{matrix} \mu^l \\ \mu^u \end{matrix} \right)$.
\\
\textbf{Iterative algorithm:}
\\
\textbf{Step 2} (compute explicit reference):
\\ $\qquad$ Determine $ \delta := \sum_{j\in \{ N \cup U_+ \cup L_- \mid
\mu_j^u < \mu_m < \mu_j^l \} }
g_j(x_j(\mu_m)) + \gamma(\mu_m) + \beta_u + \beta_l$,
\\
$\qquad$
where $x_j(\mu)$ is determined from (\ref{xmu}).
\\
$\qquad$ If $\delta >b^k$, then go to Step 3.1.
\\
$\qquad$ If $\delta < b^k$, then go to Step 3.2. \\ $\qquad$
Otherwise $(\delta = b^k)$ set $\mu^* := \mu_m$ and find optimal
$\mathbf{x}^*$ from (\ref{xmu}), and stop.
\\
\textbf{Step 3.1} (update and fix lower bounds):
\\
$\qquad$ Let $\underline{\mu} := \mu_m$
\\
$\qquad$ For $j\in N$: If $ \underline{\mu} \ge \mu_j^l$ then let $N
:= N \setminus \{j\}$ and $x_j^* := l_j$.
\\
$\qquad$ For $j \in L_-$: If $\underline{\mu} \ge \mu_j^l$ then let $L_- :=
L_- \setminus \{j\}$ and $x_j^* := l_j$.
\\
$\qquad$ For $j\in N$: If $\underline{\mu} \ge \mu_j^u$ then let $N
:= N \setminus \{j\}$ and $L_- := L_- \cup \{j\}$.
\\
$\qquad$ For $j\in U_+$: If $\underline{\mu} \in [\mu_j^u,\mu_j^l]$ and
$\overline{\mu}\ge\mu^u_j$
then let $U_+ := U_+ \setminus \{j\}$ and $M := M \cup \{j\}$; update
$\gamma_{1}$, $\gamma_{2}$.
\\
$\qquad$ Let $\mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \cup M \cup L_- \cup U_+
\mid \mu_m < \mu^k_j \}}$,
$b^{k+1} := b^{k}-
\beta_l$, and $k := k+1$.
\\
$\qquad$ Go to Step 1.
\\
\textbf{Step 3.2} (update and fix upper bounds):
\\
$\qquad$ Let $\overline{\mu} := \mu_m$.
\\
$\qquad$ For $j\in N$: If $ \overline{\mu} \le \mu_j^u$ then let
$N := N \setminus \{j\}$ and $x_j^* := u_j$.
\\
$\qquad$ For $j \in U_+$: If $\overline{\mu} \le \mu_j^u$ then let
$U_+ := U_+ \setminus \{j\}$ and $x_j^* := u_j$.
\\
$\qquad$ For $j\in N$: If $\overline{\mu} \le \mu_j^l$ then let
$N := N \setminus \{j\}$ and $U_+ := U_+ \cup \{j\}$.
\\
$\qquad$ For $j\in L_-$: If $\overline{\mu} \in [\mu_j^u,\mu_j^l]$ and
$\underline{\mu}\le\mu^l_j$
then let $L_- := L_- \setminus \{j\}$ and $M := M \cup \{j\}$;
update $\gamma_{1}$, $\gamma_{2}$.
\\
$\qquad$ Let $\mu^{k+1} := ( \mu^k_j )_{ j \in \{ N \cup M \cup L_-
\cup U_+ \mid \mu_m > \mu^k_j \}}$,
$b^{k+1} := b^{k}-
\beta_u$, and $k := k+1$.
\\
$\qquad$ Go to Step 1.
\\
\\
\end{tabular}
\noindent Note that when we determine $x_j(\mu_m)$ for $j \in U_+$ we do not
need to check if $x_j=l_j$, and for $j\in L_-$ we do need to check if
$x_j=u_j$. This will in some cases save us some operations. Further when we
determine if $x_j \in M$ we only need to check if this holds for $j \in U_+$
or $j \in L_-$ depending on if we update the lower or upper bound of the dual
variable. This differs from the algorithm in \cite{WaW12} since that algorithm
finds $M$ from $U_+ \cup L_-$. Note that we make use of information from
earlier iterations when updating $\delta$. When updating $\delta$ in
\cite{WaW12} information from earlier iteration is negligible hence some
operations are repeated.
\begin{remark}
Consider iteration $k$. In Step 3.1 we only need to check if $\mu_m \in
[\mu_j^u,\mu_j^l]$ and $\overline{\mu}\ge\mu^u_j$ for $j\in U_+$ since we know
from earlier iterations that if $j \in N$ then $\overline{\mu} \nleq
\mu^l_j$. Similar holds for Step 3.2.
\end{remark}
\subsection{Convergence and time complexity of breakpoint
algorithms}\label{sec:conbreak}
Similar to \cite{Kiw08a} and \cite{WaW12} it is possible to show that the
breakpoint algorithms converge to the optimal solution. Consider the time
complexity for algorithm MB2, MB3 and MB5. Assuming that the median function
$\rm{median(\cdot)}$ is linear, we have $C_i n$ operations in each of the
Steps 0 through 3.2 for some constants $C_i$ for $i \in \{ 0; 1; 2; 3.1;
3.2\}$ corresponding to Step 0 through 3.2 respectively. Since we use a
median search function the number of iterations will terminate in $\log_2 (n)$
iterations (in the worst case). Hence the time complexity of the algorithm is
$O (n \log_2 (n))$. However in \cite{Bru84} a proof for a time complexity of
$O (n)$ is given for the breakpoint algorithm solving the quadratic knapsack
problem (\ref{quadratic}).
\section{Relaxation algorithms}\label{sec:relaxation}
In a relaxation algorithm for the problem (\ref{problem}) we iteratively solve
the problem relaxed from constraints (\ref{problemc}), i.e., we solve the
following problem:
\begin{subequations}
\begin{align}
\minimize_x &\quad \phi(\mathbf{x}) := \sum_{j \in J^k} \phi_j(x_j),
\\ \subto & \quad g(\mathbf{x}):= \sum_{j \in J^k} g_j(x_j) \le b.
\end{align}
\label{problemrelax}
\end{subequations}
\noindent From problem (\ref{problemrelax}) we obtain a solution
$\hat{\mathbf{x}}$. Together with $\hat{\mathbf{x}}$ we also obtain an
estimate $\hat{\mu}$ of the multiplier value $\mu^*$ from the optimality
condition. Then we adjust the solution $\hat{\mathbf{x}}$ for constraints
(\ref{problemc}) by determining $x_j$ from
\begin{equation}\label{xprim}
x_j := \begin{cases} l_j, & \text{if } \hat{x}_j \leq l_j,
\cr u_j, & \text{if } \hat{x}_j \geq u_j,
\cr \hat{x}_j, & \text{if } l_j<\hat{x}_j<u_j.
\end{cases}
\end{equation}
At the beginning of the algorithm we must determine whether constraint
(\ref{problemb}) is satisfied with equality at an optimal solution. From the
optimality condition (\ref{eq:sep_conv_ineq_constr_kkt1}) we have that if the
inequality constraint ({\ref{problemb}}) is satisfied then $\mu^*=0$. Hence,
for $\hat{\mu} = 0$ we find $\hat{x}_j$ by solving $\phi_j(\hat{x}_j) +
\hat{\mu} g_j(\hat{x}_j)= 0 $; if $\sum_{j \in J} g_j(x_j) \le b$, where the
value of $x_j$ is determined from ({\ref{xprim}}), then we have found the
optimal solution to problem (\ref{problem}). Otherwise, we know that $\mu^* >
0$ and that the inequality constraint (\ref{problemb}) can be regarded as an
equality. Hence, we solve the problem (\ref{problemrelax}), obtaining a
solution $\hat{\mathbf{x}}$. Let
\[
L(\hat {\mathbf{x}}) := \{ \, j \in J^k \mid \hat{x}_j \le l_j \, \},
\qquad U(\hat {\mathbf{x}}) := \{ \, j \in J^k \mid \hat{x}_j \ge u_j
\, \}
\]
denote the sets of variables that are either out of bounds at $\hat{\mathbf{
x}}$ or equal a lower, respectively an upper, bound.
In order to simplify the remaining discussion, we consider Case 2 in Section
\ref{sec:taop}, i.e., $\mu(x_j)$ to be monotonically decreasing and $g_j$ is
increasing; Case 1 is treated analogously. Calculate the total
deficit and excess at $\hat{\mathbf{ x}}$, respectively, as
\begin{equation}\label{nabladelta}
\nabla := \sum_{j \in L} (g_j(l_j) - g_j(\hat{x}_j)), \qquad
\Delta := \sum_{j \in U} (g_j(\hat{x}_j) - g_j(u_j)).
\end{equation}
Now, if $\Delta > \nabla$ then we set $x_j^* = u_j$ for $j \in U(\hat{\mathbf
{x}})$; if $\Delta < \nabla$ we set $x_j^* = l_j$ for $j \in L(\hat{\mathbf{
x}})$; otherwise $\Delta = \nabla$ and we have found the optimal
solution. If $\Delta \neq \nabla$, then we reduce the problem by removing the
fixed variables, and adjusting the right-hand side of the constraint
(\ref{problemb}) to reflect the variables fixed. If any free variables are
left, we re-solve problem (\ref{problemrelax}) and repeat the procedure,
otherwise we have obtained an optimal solution.
The rationale behind this procedure is quite simple and natural: Suppose that
$\Delta > \nabla$ holds. We have that $\hat{\mu} = -
\phi_j'(\hat{x}_j)/g_j'(\hat{x})$ for $j \in J^k \setminus \{ L \cup U
\}$. Let $s \in U(\hat{\mathbf{ x}})$ and $i \in J^k \setminus U(\hat{\mathbf{
x}})$. Since the functions $-\phi_j' / g_j'$ are decreasing, it follows
that
\[
- \frac{\phi_s'(u_s)}{g'_s(u_s)} \geq
- \frac{\phi_s'(\hat{x}_s)}{g'_s(\hat{x}_s)} = \hat{\mu} =
- \frac{\phi_i'(\hat{x}_i)}{g_i'(\hat{x}_i)} \geq
- \frac{\phi_i'(u_i)}{g'_i(u_i)}.
\]
Denote by $b_+$ the right-hand side in the following iteration given that
$\Delta > \nabla$ holds: $b_+ := b - \sum_{j \in U(\hat{ x})} g_j(u_j)$. Also
let $(\hat{\mathbf{x}}', \hat{\mu}')$ denote a pair of relaxed optimal
primal--dual solutions in the following iteration. We must have that $\hat\mu'
\leq \hat\mu$, since
\[
\sum_{j \in J^k \setminus U(\hat{x}) } g_j(\hat{x}_j) = b - \sum_{j \in U
(\hat{x})} g_j(\hat{x}_j) \leq b - \sum_{j \in U (\hat{ x})} g_j(u_j) = b_+
= \sum_{j \in J^k \setminus U(\hat{ x}) } g_j(\hat{x}_j');
\]
hence, for at least one $j \in J \setminus U(\hat{\mathbf{x}})$ we have that
$\hat{x}_j' \geq \hat{x}_j$, and therefore,
\[
\hat{\mu}' = - \frac{\phi_j'(\hat{x}_j')}{g_j'(\hat{x}_j')} \leq -
\frac{\phi_j'(\hat{x}_j)}{g_j'(\hat{x})} = \hat{\mu}
\]
follows. This derivation was first described by Bitran and Hax \cite{BiH81}.
Since in each iteration at least one variable is pegged to an optimal value,
the algorithm is clearly finite. The most serious disadvantage of the
algorithm may be the requirement that the problem without the variable bounds
present must have an optimal solution. The computational efficiency of this
method is also determined by whether or not it is possible to provide an
explicit formula for each $\hat{x}_j$ in terms of the multiplier.
\subsection{Explicit/Implicit evaluation}\label{sec:peme}
For the breakpoint algorithm we determine $\mathbf{x}$ from (\ref{xmu}) for an
arbitrary multiplier $\mu$; then we explicitly evaluate the optimality of
$\mathbf{x}$ by substituting it into the resource constraint
({\ref{problemb}}). The explicit evaluation leaves us with one out of three
possible scenarios: (\ref{caseopt})--(\ref{casehigh}). For the relaxation
algorithm the traditional method to evaluate a solution to the problem
(\ref{problemrelax}) is to calculate the total deficit and excess at
$\mathbf{\hat{x}}$ (see Section \ref{sec:relaxation}). Also this evaluation
leaves us with 3 possible scenarios, namely
\begin{subequations}\label{relaxbit2}
\begin{align}
\Delta & = \nabla \label{caseopt3}, \\
\Delta & > \nabla \label{caselow3}, \\
\Delta & < \nabla \label{casehigh3}.
\end{align}
\end{subequations}
Evaluating the optimality from (\ref{caseopt3})--(\ref{casehigh3}) will be
referred to as an {\em implicit evaluation}. For the relaxation algorithm we
next show that the implicit and explicit evaluations are equivalent.
Propositions \ref{prop1brs} and \ref{prop2brs} below state the relations
between explicit and implicit evaluation. The proof of Proposition
\ref{prop2brs} is similar to the proof of Proposition \ref{prop1brs}.
\begin{proposition}[relation between explicit and implicit evaluation for
$g_j$ monotonically increasing]\label{prop1brs} If for all $j\in J$ $g_j$ is
monotonically increasing, then the explicit evaluation
(\ref{caseopt})--(\ref{casehigh}) is equivalent to the implicit evaluation
(\ref{caseopt3})--(\ref{casehigh3}), i.e., (\ref{caseopt3})
$\Longleftrightarrow$ (\ref{caseopt}), (\ref{caselow3})
$\Longleftrightarrow$ (\ref{caselow}), and (\ref{casehigh3})
$\Longleftrightarrow$ (\ref{casehigh}).
\end{proposition}
\begin{proof}
For all $j \in J$, let $\hat{x}_j$ be the solution to ({\ref{problemrelax}}).
Let $ \nabla$ and $\Delta$ be defined as in (\ref{nabladelta}). We
have from (\ref{xprim}) that
\begin{align*}
\sum_{j\in J}g_j(x_j) = & \sum_{j\in J \setminus \{
U \cup L \} } g_j(\hat{x}_j) + \sum_{j\in
L} \left \{ g_j(l_j)-g_j(\hat{x}_j) + g_j(\hat{x}_j)\right
\} \nonumber
\\
& + \sum_{j\in U}\left \{ g_j(u_j) -
g_j(\hat{x}_j) + g_j(\hat{x}_j) \right \} \nonumber \\ = &
\sum_{j\in J \setminus \{ U \cup L \} }
g_j(\hat{x}_j) + \sum_{j\in L} \left \{
g_j(l_j)-g_j(\hat{x}_j)\right \} + \sum_{j\in L}
g_j(\hat{x}_j) \nonumber \\ & - \sum_{j\in U} \left \{ g_j(u_j)
- g_j(\hat{x}_j) \right \} + \sum_{j\in U} g_j(\hat{x}_j)
\nonumber
\\
= & \sum_{j\in J} g_j(\hat{x}_j) + \nabla - \Delta \nonumber
\\
= & \; b + \nabla - \Delta, \nonumber
\end{align*}
where the last equality follows from the fact that $\hat{\mathbf{x}}$ is the
solution to the relaxed problem. Hence, $\hat{\mathbf{x}}$ must satisfy the
resource constraint. We know that $\Delta,\nabla \ge 0$ since $g_j$ is
increasing. Hence, if $\Delta = \nabla$ then (\ref{caseopt}) holds, if $\Delta
>\nabla$ then (\ref{caselow}) holds, and if $\Delta < \nabla$ then
(\ref{casehigh}) holds.
\end{proof}
\begin{proposition}[relation between explicit and implicit evaluation for
$g_j$ monotonically decreasing]\label{prop2brs} If $g_j$ is monotonically
decreasing for all $j\in J$, then the explicit evaluation
(\ref{caseopt})--(\ref{casehigh}) is equivalent to the implicit evaluation
(\ref{caseopt})--(\ref{casehigh}), i.e., (\ref{caseopt3})
$\Longleftrightarrow$ (\ref{caseopt}), (\ref{casehigh3})
$\Longleftrightarrow$ (\ref{caselow}), and (\ref{caselow3})
$\Longleftrightarrow$ (\ref{casehigh}).
\end{proposition}
\subsection{Primal/Dual evaluation of boundaries}\label{sec:primDualEval}
In the breakpoint algorithms in Section \ref{sec:rank}, we use the dual
variable to determine if $x_j$ equals a bound or if it lies in between the
bounds; see (\ref{xmu}). In relaxation algorithms the traditional way to solve
the problem (\ref{problem}) is to determine the primal optimal solution
$\hat{x}_j$ of the relaxed problem (\ref{problemrelax}) and then simply check
if $\hat{x}_j$ is within the lower and upper bounds; see (\ref{xprim}). An
alternative is to find the optimal dual variable $\hat{\mu}$ of the relaxed
problem (\ref{problemrelax}) and then (similar to the breakpoint algorithm in
Section \ref{sec:rank}) determine the primal variables from (\ref{xmu}). Of
course this requires us to determine the breakpoints. However, for the
variables that violate the bounds (\ref{problemc}) we don't have to determine
$x_j$ from the relation $\phi_j(x_j) + \mu g_j(x_j)= 0$. Hence, instead of
evaluating $L(\hat{\mathbf{x}})$ and $U(\hat{\mathbf{x}})$ from the primal
variable as in Section \ref{sec:relaxation}, we can evaluate $L(\hat{\mu})$
and $U(\hat{\mu})$ from the dual variable as in Section \ref{sec:taop}.
Define $\Phi_j(x_j) := \phi_j^\prime(x_j) / g_j^\prime(x_j)$ and assume that
there exists an inverse to $\Phi_j$. It follows from the optimality
conditions (\ref{xmu}) that $\Phi_j(\hat{x}_j(\hat\mu)) = -\hat\mu$,
$\Phi_j(l_j) = \phi_j^\prime(l_j) / g_j^\prime(l_j)$ and $\Phi_j(u_j) =
\phi_j^\prime(u_j) / g_j^\prime(u_j)$. Hence we have that:
\[
L(\hat {\mu}) := \{ \, j \in J \mid \hat{\mu} \ge -\phi_j^\prime(l_j) /
g_j^\prime(l_j) \, \} \Longleftrightarrow L(\hat {\mathbf{x}}) := \{ \, j \in
J \mid \hat{x}_j \le l_j \, \},
\]
and
\[
U(\hat {\mu}) := \{ \, j \in J \mid \hat{\mu} \le -\phi_j^\prime(u_j) /
g_j^\prime(u_j) \, \} \Longleftrightarrow U(\hat {\mathbf{x}}) := \{ \, j \in
J \mid \hat{x}_j \ge u_j \, \}.
\]
\subsection{Implementation choices for the relaxation algorithm}
\label{sec:pdei}
Considering the performance of a relaxation algorithm two decisions need to be
made: should the relaxed solution of the problem be evaluated implicitly from
$\nabla$ and $\Delta$ or explicitly from $\sum g_j$ (see Section
\ref{sec:peme})? Should the algorithm solve the primal or dual relaxed problem
(see Section \ref{sec:primDualEval})? An overview of the relaxation algorithm
and its possible realizations is shown in Figure \ref{relaxationOverview}. The
leftmost path in the figure is the classic primal relaxation algorithm of
Bitran and Hax \cite{BiH81}, the rightmost path in the figure is implemented
in \cite{Ste01}, and the algorithm that applies the right path in Step 1 and
the left path in Step 2 is implemented in \cite{KiW12}. Beside these two paths
no other paths have been explored. Our intention is to evaluate the
theoretically and practically best performing paths. Since no earlier studies
have applied 3- or 5-sets pegging for the relaxation algorithm, our intention
is to apply these two more sophisticated pegging methods.
\begin{figure}[h!]
\centering
\psfrag{C}{\small \bf ~~Check if inequality feasible}
\psfrag{F}{\small \bf ~~~~~~~~~~Find breakpoints}
\psfrag{P}{\small \bf ~~~~Solve relaxed primal problem}
\psfrag{D}{\small \bf ~~~~Solve relaxed dual problem}
\psfrag{I}{\small \bf ~~~~~~~~~~~Implicit evaluation}
\psfrag{E}{\small \bf ~~~~~~~~~~~Explicit evaluation}
\psfrag{A}{\small \bf \!\!\!\!\!\! Apply 2-, 3-, or 5-sets pegging.
If optimal stop, else go to Step 1}
\psfrag{a}{\small \bf Step 1}
\psfrag{b}{\small \bf Step 2}
\psfrag{c}{\small \bf Step 3}
\includegraphics[width=12cm]{Constr_Rel_Alg.eps}
\caption{The diagram shows different possibilities to consider when
constructing a relaxation algorithm.} \label{relaxationOverview}
\end{figure}
In Section \ref{sec:bih} we present an algorithm corresponding to the leftmost
path in Figure \ref{relaxationOverview}; in Section \ref{sec:mod2} we present
an algorithm which utilizes the rightmost path in Step 1 of Figure
\ref{relaxationOverview} and then changes to the left path in Step 2 of the
figure; in Section \ref{sec:ste} an algorithm corresponding to the rightmost
path of Figure \ref{relaxationOverview} is presented; and in Section
\ref{sec:bh4} an algorithm that utilizes the rightmost path in Step 1 of
Figure \ref{relaxationOverview} and then utilizes the theoretically best path
in Step 2 is presented. All algorithms in Sections
\ref{sec:bih}--\ref{sec:bh4} utilize 2-sets pegging. In Section
\ref{sec:tandf} we describe how 3- and 5-sets pegging can be utilized in
these.
\subsubsection{Algorithm: Primal determination with Implicit evaluation of
the Relaxed problem (PIR2)}\label{sec:bih}
We first assume that Case 1 in Section \ref{sec:taop} holds for problem
(\ref{problem}). Define parameters to calculate the total deficit and excess,
as $\alpha^k_+ := \sum_{j \in U^k} g_j(\hat{x}_j)$, $\alpha^k_- := \sum_{j \in
L^k} g_j(\hat{x}_j)$, $\beta^k_- := \sum_{j \in L^k} g_j(l_j)$ and
$\beta^k_+ := \sum_{j \in U^k} g_j(u_j)$. Hence we have that $\Delta^k =
\alpha^k_+ -\beta^k_+ $ and $\nabla^k = \beta^k_-- \alpha^k_-$. In the next
iteration when we reduce the resource $b^k$ we hence don't have to
re-calculate the part of the pegged variables which define $\beta^k_+$ or
$\beta^k_-$ (see Steps 3.1 and 3.2). We present an algorithm for problem
(\ref{problem}) which is similar to the algorithm in Section 2 of
\cite{BrS02}:
\begin{tabular}[h]{l}
\\
\textbf{Initialization: } Set $k:=1, \; J^k := J$,
and $b^k := b$.
\\
\textbf{Step 0} (check if $\mu = 0$ is optimal):
\\
$\qquad$ Let $\mu=0$ and find the solution $\mathbf{\hat{x}}$ to the relaxed
problem (\ref{problemrelax}), i.e., solve $\phi_j'(\hat{x}_j) + \mu
g_j'(\hat{x}_j)$, $j \in J^k$.
\\
$\qquad$ If $\sum_{j \in J^k} g_j (x_j) \le b$, for $x_j$ determined from
(\ref{xprim}), then $\mu^*=0$, $x_j^* = x_j$ for $j \in J^k$, and stop.
\\
\textbf{Iterative algorithm: }
\\
\textbf{Step 1} (solve relaxed primal problem):
\\ $\qquad$ For $j\in J^k$, find $\hat{x}_j^k$ by solving the relaxed problem
(\ref{problemrelax}).
\\
$\qquad$ Set $L:=\emptyset$ and $U:=\emptyset$.
\\
\textbf{Step 2} (implicit evaluation): \\ $\qquad$
Determine $U(\hat{\mathbf{x}}^k)$ and $L(\hat{\mathbf{x}}^k)$ while computing
$\Delta^k := \alpha^k_+ -\beta^k_+ $ and $\nabla^k := \beta^k_-- \alpha^k_-$.
\\
$\qquad$ If $\Delta^k > \nabla^k$, then go to Step 3.1. \\ $\qquad$ If
$\Delta^k < \nabla^k$, then go to Step 3.2.
\\
$\qquad$ If $\Delta^k = \nabla^k$, then set $x_j^* := l_j$ for $j \in
L(\hat{\mathbf{x}}^k)$, $x_j^* := u_j$ for $j \in U(\hat{\mathbf{x}}^k)$,
\\ $\qquad$ $\quad$
$x_j^* := x_j^k$ for $j \in J^k \setminus \{ L(\hat{\mathbf{x}}^k) \cup
U(\hat{\mathbf{x}}^k) \}$, and stop.
\\
\textbf{Step 3.1} (peg lower bounds): \\
$\qquad$ Set $x_j^* := l_j$ for $j \in L$, $b^{k+1} := b^k - \beta^k_-$ and
$J^{k+1} := J^k \setminus L$.
\\ $\qquad$ If
$J^{k+1} := \emptyset$ then stop, else set $k := k+1$ and go to Step 1.
\\
\textbf{Step 3.2} (peg upper bounds):
\\ $\qquad$ Set $x_j^* := u_j$ for
$j \in U$, $b^{k+1} := b^k - \beta^k_+$ and $J^{k+1} :=
J^k \setminus U$. \\ $\qquad$ If $J^{k+1} :=
\emptyset$ then stop else set $k:=k+1$ and go to Step 1. \\
\\
\end{tabular}
\noindent We need to clarify some of the steps of the algorithm. In Step 1, we
find $\hat{\mathbf{x}}^{k+1}$ from, or partly from, $\hat{\mathbf{x}}^{k}$.
Assume, for example, that $\phi_j(x_j) = x_j\log(x_j/a_j-1)$ and
$g_j(x_j)=x_j$; then, $x_j^{k+1} = a_j b^{k+1} / \sum_{j\in J^{k+1}} a_j =
b^{k+1} / (\omega -\sum_{j\in K(\hat{\mathbf{x}}^k)} a_j$), where $\omega =
\sum_{j\in J^{k}} a_j$ and $K := U$ if the upper bound was pegged at iteration
$k$ and $K := L$ if the lower bound was pegged at iteration $k$. If $|K| < |
J^{k+1} |$ then this will save us some operations. A similar update of
$\hat{\mathbf{x}}$ for the quadratic knapsack problem is performed in
\cite[Section 3]{RJL92}, \cite{BSS96} and \cite[Section 5.1]{Kiw08b}.
As in \cite[Algorithm 3.1]{Kiw08b}, our algorithm will stop if $\Delta^k =
\nabla^k$, while the algorithm in \cite[Section 2]{BrS02} stops only if
$L(\hat{x}^k) \cup U(\hat{x}^k) = \emptyset$. Moreover, in Steps 3.1 and 3.2,
we peg the variables that violate the bounds and calculate $b^k$ explicitly,
while in \cite[Section 2]{BrS02} the index $j$ is added to the set of violated
bounds ($L$ or $U$) and $b^k$ is calculated as $b - \sum_{j \in L} g_j(l_j)
- \sum_{j \in U} g_j(u_j)$.
According to Proposition \ref{pegglow}, if Case 2 in Section \ref{sec:taop}
holds, Step 2 in the algorithm is modified as follows (an analogous
modification can be defined for the relaxation algorithms in Sections
\ref{sec:mod2} through \ref{sec:tandf}):
\begin{tabular}{l}
\\ \textbf{Step 2'} (implicit evaluation):
\\
$\qquad$ Determine
$U(\hat{\mathbf{x}}^k)$ and $L(\hat{\mathbf{x}}^k)$ while computing $\Delta^k
= \alpha^k_+ -\beta^k_+ $ and $\nabla^k = \beta^k_-- \alpha^k_-$.
\\
$\qquad$
If $\Delta^k > \nabla^k$, then go to Step 3.2. \\ $\qquad$ If $\Delta^k <
\nabla^k$, then go to Step 3.1.
\\
$\qquad$ If $\Delta^k = \nabla^k$, then set $x_j^* :=
l_j$ for $j \in L(\hat{\mathbf{x}}^k)$, $x_j^* := u_j$ for $j \in
U(\hat{\mathbf{x}}^k)$,
\\
$\qquad$ $\quad$ $x_j^* := x_j^k$ for $j \in J^k
\setminus \{ L(\hat{\mathbf{x}}^k) \cup U(\hat{\mathbf{x}}^k) \}$, and stop.
\\ \\
\end{tabular}
\begin{remark}
For the equality problem (\ref{problemeq}), $\mu$ is unrestricted; hence the
algorithm for problem (\ref{problemeq}) will be similar to the above, except
we ignore Step 0.
\end{remark}
\begin{remark}
In Step 1 we have to calculate $\hat{x}_j$ $| J^k|$ times. In Step 2 we need
to find $x_j$ which needs at most $2| J^k |$ comparisons. We also have to
calculate $\nabla^k$ and $\Delta^k$, which implies $2| L \cup U |$ operations.
\end{remark}
\subsubsection{Algorithm: Dual determination with Implicit evaluation of the
Relaxed problem (DIR2)}\label{sec:mod2}
Instead of evaluating the primal variables in Step 1 as in the algorithm PIR2
in Section \ref{sec:bih}, we can evaluate the dual variable $\mu$. Note that
if we evaluate the dual variable, then we have to determine the breakpoints in
the initialization. Our modification of the algorithm in Section \ref{sec:bih}
takes the following form (Steps 0 and 3 are same as in PIR2 and are therefore
not repeated):
\begin{tabular}{l}
\\
\textbf{Initialization: }
\\
$\qquad$ Set $k:=1$, $J^k := J$, and $b^k := b$.
\\
$\qquad$ Calculate breakpoints
$\mu^l:= ( \mu^l_j )_{j \in J^k}$, and $\mu^u:= ( \mu^u_j )_{j \in J^k} $ as in (\ref{xmu}).
\\
\textbf{Step 1} (solve relaxed dual problem):
\\
$\qquad$ Find the optimal dual variable $\hat{\mu}^k$ of the relaxed problem (\ref{problemrelax}).
\\
\textbf{Step 2} (implicit evaluation): \\ $\qquad$
Determine $L(\hat{\mu}^k)$ and $U(\hat{\mu}^k)$ while
computing $\Delta^k := \alpha^k_+ -\beta^k_+ $ and $\nabla^k :=
\beta^k_-- \alpha^k_-$. \\
$\qquad$ If $\Delta^k > \nabla^k$, then go to Step 3.1.
\\
$\qquad$ If $\Delta^k < \nabla^k$, then go to Step 3.2.
\\
$\qquad$ If $\Delta^k = \nabla^k$, then set $x_j^* := l_j$ for $j \in L(\hat{\mu}^k)$, $x_j^* := u_j$ for $j \in U(\hat{\mu}^k)$
\\
$\qquad$ $\quad$ $x_j^* := x_j^k$ for $j \in J \setminus \{ L(\hat{\mu}^k) \cup
U(\hat{\mu}^k) \}$, and stop.
\\
\end{tabular}
\begin{remark}
In Step 1 we need to find the optimal dual solution $\mu^k$ to the relaxed
problem. In Step 2 we need at most $2|J^k|$ comparisons but we only need to
calculate $\hat{x}_j(\mu)$ for $j \in \{ L \cup U \}$. For the evaluation we
need to calculate $\nabla^k$ and $\Delta^k$, which implies $2| L \cup U |$
operations.
\end{remark}
\subsubsection{Algorithm: Dual determination with Explicit evaluation of the
Relaxed problem (DER2) }\label{sec:ste} As in the algorithm DIR2 in Section
\ref{sec:mod2} we evaluate the dual variable but instead of evaluating the
relaxed solution implicitly we do it explicitly. Define $\beta_l := \sum_{j\in
L} g_j(l_j)$ and $\beta_u :=\sum_{j\in U} g_j(u_j)$. The algorithm follows
(Steps 0 and 3 are same as in PIR2 in Section \ref{sec:bih} and are therefore
not repeated):
\begin{tabular}{l}
\\ \textbf{Initialization: }
\\
$\qquad$ Set $k:=1$, $J^k := J$, and $b^k := b$.
\\
$\qquad$ Calculate breakpoints
$\mu^l:= ( \mu^l_j )_{j \in J^k }$, $\mu^u:= ( \mu^u_j )_{j \in J^k }$ as in
(\ref{xmu}).
\\
\textbf{Step 1} (solve relaxed dual problem):
\\
$\qquad$ Find the optimal dual variable $\hat{\mu}^k$ of the relaxed problem
(\ref{problemrelax}).
\\
\textbf{Step 2} (explicit evaluation):
\\
$\qquad$ Determine $U(\hat{\mu}^k)$ and $L(\hat{\mu}^k)$ and
calculate $\delta (\hat\mu^k) := \sum_{j \in J^k \setminus \{
U(\hat\mu^k) \cup L(\hat\mu^k) \} } g_j(x_j(\hat\mu^k)) + \beta_l + \beta_u$.
\\
$\qquad$ If $\delta (\hat\mu^k) > b^k$, then go to Step 3.1.
\\
$\qquad$ If $\delta (\hat\mu^k) < b^k$, then go to Step 3.2.
\\
$\qquad$ If $\delta (\hat\mu^k) = b^k$, then set $x_j^* := l_j$ for $j \in
L(\hat{\mu}^k)$, $x_j^* := u_j$ for $j \in U (\hat{\mu}^k)$
\\
$\qquad$ $\quad$ $x_j^* := x_j^k$ for $j \in J \setminus \{ L(\hat{\mu}^k) \cup
U(\hat{\mu}^k) \}$, and stop.
\\ \\
\end{tabular}
\noindent
The algorithm uses the principle of the algorithm in \cite[Algorithm 1]{Ste01}.
\begin{remark}
In Step 1 we need to find the value of $\hat\mu^k$ from the relaxed
problem. In Step 2 we need at most $2|J^k|$ comparisons and we need to
calculate $x^k_j$ for $j \in J^k \setminus \{ L \cup U \}$. For the evaluation
we need to calculate $\delta(\hat\mu^k)$, which implies $|J^k|$ operations.
\end{remark}
\subsubsection{Algorithm: Dual determination modification with blended
evaluation of the Relaxed problem (DBR2)} \label{sec:bh4}
Consider the implicit evaluation in Section \ref{sec:mod2}. We calculate
$\nabla^k$ and $\Delta^k$ from $\nabla^k := \sum_{j\in L(\hat\mu^k)} g_j(l_j)
- \sum_{j\in U(\hat\mu^k)} g_j(\hat{x}^k_j)$ and $\Delta^k := \sum_{j\in
U(\hat\mu^k)} g_j(\hat{x}^k_j) - \sum_{j\in L(\hat\mu^k)} g_j(u_j)$, which
implies $2P|L^k \cup U^k |$ operations, where $P$ is an integer associated
with the number of operations it takes to calculate $g_j(x_j)$. Moreover, we
have to determine $\hat{x}_j$ for $j \in \{ L^k \cup U^k \}$, which implies
$Q|L^k \cup U^k |$ operations, where $Q$ is an integer associated with the
number of operations it takes to determine $x_j(\hat\mu)$.
Now consider the explicit evaluation: We have to calculate $\delta (\hat\mu^k)
:= \sum_{j \in J^k \setminus \{ U(\hat\mu^k) \cup L(\hat\mu^k) \} }
g_j(x_j(\hat\mu^k)) + \sum_{j \in L(\hat\mu^k)} g_j(l_j) + \sum_{j \in
U(\hat\mu^k)} g_j(u_j)$, which implies $P|J^k|$ operations. Moreover, we
have to determine $\hat{x}_j$ for $j \in J^k \setminus \{ L^k \cup U^k \}$,
which implies $Q|J^k \setminus \{ L^k \cup U^k \} |$ operations.
Hence, if $(P+Q)|J^k| < (2P+2Q)|U(\mu^k) \cup L(\mu^k)|$ or, equivalently,
$|J^k| < 2|U(\mu^k) \cup L(\mu^k)|$, then using the explicit evaluation of the
relaxed solution $\mathbf{\hat{x}}$ in Step 2 would require less operations,
and it would be more successful to use the algorithm in Section
\ref{sec:ste}. If however $|J^k| > 2|U(\mu^k) \cup L(\mu^k)|$, then there will
be less operations if we use the algorithm in Section \ref{sec:mod2}. So, we
propose a new algorithm that utilizes the cardinalities of the sets $J^k$,
$U(\hat{\mu}^k)$ and $L(\hat{\mu}^k)$: from the cardinalities we make the
decision whether to use an explicit or implicit evaluation in Step 2. We
consider the following modification of the algorithm PIR2 in Section
\ref{sec:bih}:
\begin{tabular}{l}
\\ \textbf{Initialization: }
\\
$\qquad$ Set $k:=1$, $J^k := J$, and $b^k := b$.
\\
$\qquad$ Calculate breakpoints
$\mu^l:= ( \mu^l_j )_{j \in J^k }$, $\mu^u:= ( \mu^u_j )_{j \in J^k }$ as in
(\ref{xmu}).
\\
\textbf{Step 1: } (solve relaxed dual problem):
\\
$\qquad$ Find the optimal dual variable $\hat{\mu}^k$ of the relaxed problem
(\ref{problemrelax}).
\\
\textbf{Step 1.1: } (implicit or explicit evaluation?):
\\
$\qquad$
Determine $U(\hat{\mu}^k)$ and $L(\hat{\mu}^k)$.
\\
$\qquad$ If $|J^k| < 2|U(\hat{\mu}^k) \cup L(\hat{\mu}^k)|$ then use explicit
evaluation (continue with algorithm DER2 in
\\
$\qquad$
Section \ref{sec:ste}), otherwise use implicit evaluation (continue with
algorithm DIR2 in Section \ref{sec:mod2}).
\\ \\
\end{tabular}
\subsubsection{3- and 5-sets pegging for relaxation algorithms}\label{sec:tandf}
From the proof of convergence of the relaxation algorithm (see Section
\ref{sec:proofrel}) we have that the algorithm improves the lower or the upper
bound for the dual variable in each iteration. Hence, similar to the
breakpoint algorithm (see Sections \ref{sec:rmm} and \ref{sec:MB5}) it is
possible to apply 3- and 5-sets pegging for the dual relaxation algorithms
DIR2, DER2 and DBR2. Similar to the relaxation algorithm we will denote a
relaxation algorithm that uses 3-sets-pegging with a suffix ''3'', e.g., DBR3,
and one that uses 5-sets-pegging with a suffix ''5'', e.g., DBR5. The
implementation of 3- and 5-sets pegging is similar for the three dual
relaxation algorithms; therefore only DIR3 and DIR5 are given (Step 0 is
similar as for PIR in Section \ref{sec:bih}):
\begin{tabular}{l}
\\
\textbf{DIR3: } \\
\textbf{Initialization: }
\\
$\qquad$ Set $k:=1$, $J^k := J$, $M^k := \emptyset $, $b^k := b$, $\underline{\mu} = -\infty$, and $\overline{\mu} = \infty$.
\\
$\qquad$ Calculate breakpoints
$\mu^l:= ( \mu^l_j )_{j \in J^k }$, $\mu^u:= ( \mu^u_j )_{j \in J^k}$ as in (\ref{xmu}).
\\
\textbf{Step 1} (solve relaxed dual problem):
\\
$\qquad$ Find the optimal dual variable $\hat{\mu}^k$ of the relaxed problem (\ref{problemrelax}).
\\
\textbf{Step 2} (implicit evaluation): \\ $\qquad$
Determine $L(\hat{\mu}^k)$ and $U(\hat{\mu}^k)$ from $J^k$ and
compute $\Delta^k = \alpha^k_+ -\beta^k_+ $ and $\nabla^k =
\beta^k_-- \alpha^k_-$. \\
$\qquad$ If $\Delta^k > \nabla^k$, then go to Step 3.1.
\\
$\qquad$ If $\Delta^k < \nabla^k$, then go to Step 3.2.
\\
$\qquad$ If $\Delta^k = \nabla^k$, then set $x_j^* = l_j$ for $j \in L(\hat{\mu}^k)$, $x_j^* = u_j$ for $j \in U(\hat{\mu}^k)$
\\
$\qquad$ $\quad$ $x_j^* = x_j(\hat{\mu}^k)$ for $j \in \{ J^k \cup M^k \} \setminus \{ L(\hat{\mu}^k) \cup
U(\hat{\mu}^k) \}$, and stop.
\\
\textbf{Step 3.1} (peg lower bounds):
\\
$\qquad$ Update lower bound, $\underline{\mu} := \hat{\mu}^k$, and resource $b^{k+1} := b^k - \beta^k_-$. \\
$\qquad$ Set $x_j^* := l_j$ for $j \in L^k$, $J^{k+1} := J^k \setminus L^k$ and $M^{k+1} := M^{k}$.
\\
$\qquad$ For $j \in J^{k+1}$: If $\underline{\mu},\overline{\mu} \in [\mu_j^u,\mu_j^l]$
then $J^{k+1} := J^{k+1}
\setminus \{j\}$, $M^{k+1} := M^{k+1} \cup \{j\}$.
\\
$\qquad$ If $J^{k+1} = \emptyset$ then find optimal solution and stop, else set $k := k+1$ and go to Step 1.
\\
\textbf{Step 3.2} (peg upper bounds):
\\
$\qquad$ Update upper bound, $\bar{\mu} := \hat{\mu}^k$, and resource $b^{k+1} := b^k - \beta^k_+$.
\\
$\qquad$ Set $x_j^* := u_j$ for $j \in U^k$, $J^{k+1} := J^k \setminus U^k$ and $M^{k+1} := M^{k}$.
\\
$\qquad$ For $j \in J^{k+1}$: If $\underline{\mu},\overline{\mu} \in [\mu_j^u,\mu_j^l]$
then $J^{k+1} := J^{k+1}
\setminus \{j\}$, $M^{k+1} := M^{k+1} \cup \{j\}$.
\\
$\qquad$ If $J^{k+1} = \emptyset$ then find optimal solution and stop, else set $k := k+1$ and go to Step 1.
\\
\\
\end{tabular}
\noindent In Steps 3.1 and 3.2, if $J^{k+1} = \emptyset$ then we can find the
optimal solution from $M$ since we know that for all $j \in M$ it holds that
$l_j < x_j < u_j$, i.e., we don't have to consider the constraint
(\ref{problemc}). Further, in Step 3.1 (and analogusly for Step 3.2) when
searching for $j\in J^{k+1}$ such that $\underline{\mu},\overline{\mu} \in
[\mu_j^u,\mu_j^l]$, we only have to consider $j\in J^{k+1} \setminus U^k
$, since we know that if $j \in U^k$ then $\hat{\mu}^k \le \mu_j^u$. Note that
for the case where $\hat{\mu}^k = \mu_j^u$ the corresponding index $j$ will
continue to belong to $J^{k+1}$. Finally we note that we don't have to check
if $\underline{\mu},\overline{\mu} \in [\mu_j^u,\mu_j^l]$ if $\underline{\mu}
= -\infty$ and/or $\overline{\mu} = \infty$.
\begin{tabular}{l}
\\
\textbf{DIR5: } \\
\textbf{Initialization: }
\\
$\qquad$ Set $k:=1$, $J^k := J$, $M^k = U^k_+ = L^k_- := \emptyset $, $b^k := b$, $\underline{\mu} = -\infty$, and $\overline{\mu} = \infty$.
\\
$\qquad$ Calculate breakpoints
$\mu^l:= ( \mu^l_j )_{j \in J^k }$, $\mu^u:= ( \mu^u_j )_{j \in J^k }$ as in (\ref{xmu}).
\\
\textbf{Step 1} (solve relaxed dual problem):
\\
$\qquad$ Find the optimal dual variable $\hat{\mu}^k$ of the relaxed problem (\ref{problemrelax}).
\\
\textbf{Step 2} (implicit evaluation): \\ $\qquad$
Determine $L(\hat{\mu}^k)$ and $U(\hat{\mu}^k)$ from $J^k \cup L_-^{k} \cup U_+^{k}$,
compute $\Delta^k = \alpha^k_+ -\beta^k_+ $ and $\nabla^k =
\beta^k_-- \alpha^k_-$. \\
$\qquad$ If $\Delta^k > \nabla^k$, then go to Step 3.1.
\\
$\qquad$ If $\Delta^k < \nabla^k$, then go to Step 3.2.
\\
$\qquad$ If $\Delta^k = \nabla^k$, then set $x_j^* = l_j$ for $j \in L(\hat{\mu}^k)$, $x_j^* = u_j$ for $j \in U(\hat{\mu}^k)$
\\
$\qquad$ $\quad$ $x_j^* = x_j(\hat{\mu}^k)$ for $j \in \{ J^k \cup M^k \cup L^k_- \cup U^k_+ \} \setminus \{ L(\hat{\mu}^k) \cup U(\hat{\mu}^k) \}$, and stop.
\\
\textbf{Step 3.1} (peg lower bounds):
\\
$\qquad$ Update lower bound, $\underline{\mu} := \hat{\mu}^k$, and resource $b^{k+1} := b^k - \beta^k_-$.
\\
$\qquad$ Set $x_j^* := l_j$ for $j \in L^k$, $J^{k+1} := J^k \setminus L^k$ and $L_-^{k+1} := L_-^{k} \setminus L^k$.
\\
$\qquad$ For $j \in \{ J^{k+1} \setminus U^k_+ \}$: If $\mu_m \le \mu_j^l$ then $J^{k+1} := J^{k+1} \setminus \{j\}$ and $L_-^{k+1} := L_-^{k+1} \cup \{j\}$.
\\
$\qquad$ For $j \in \{ U_+^k \setminus U^k \}$: If $\underline{\mu}\le\mu^l_j$
then $U_+^k := U_+^k \setminus \{j\}$ and $M^k := M^k \cup \{j\}$.
\\
$\qquad$ Set $M^{k+1} := M^{k}$ and $U_+^{k+1} := U_+^{k}$. \\
$\qquad$ If $J^{k+1} \cup L_-^{k+1} \cup U_+^{k+1}= \emptyset$ then find optimal solution and stop,
\\
$\qquad$ $\quad$ else set $k := k+1$ and go to Step 1.
\\
\textbf{Step 3.2} (peg upper bounds):
\\
$\qquad$ Update upper bound, $\bar{\mu} := \hat{\mu}^k$, and resource $b^{k+1} := b^k - \beta^k_+$.
\\
$\qquad$ Set $x_j^* := u_j$ for $j \in U^k$, $J^{k+1} := J^k \setminus U^k$ and $U_+^{k+1} := U_+^{k} \setminus U^k$.
\\
$\qquad$ For $j \in \{ J^{k+1} \setminus L^k_- \}$: If $\mu_m \le \mu_j^l$ then $J^{k+1} := J^{k+1} \setminus \{j\}$ and $U_+^{k+1} := U_+^{k+1} \cup \{j\}$.
\\
$\qquad$ For $j \in \{ L_-^k \setminus L^k \}$: If $\underline{\mu}\le\mu^l_j$
then $L_-^k := L_-^k \setminus \{j\}$ and $M^k := M^k \cup \{j\}$.
\\
$\qquad$ Set $M^{k+1} := M^{k}$ and $L_-^{k+1} := L_-^{k}$. \\
$\qquad$ If $J^{k+1} \cup L_-^{k+1} \cup U_+^{k+1}= \emptyset$ then find optimal solution and stop,
\\
$\qquad$ $\quad$ else set $k := k+1$ and go to Step 1.
\\
\\
\end{tabular}
\begin{remark}
Note that for DBR2 in Section \ref{sec:bh4} we determine whether we should
continue with DIR or DER from $| J^k | < 2 | U(\hat{\mu}^k) \cup
L(\hat{\mu}^k) |$. This condition needs to be modified for the 3- and the
5-sets algorithms DBR3 and DBR5. This is done by substituting the
condition mentioned by $| J^k \cup M^k | < 2 | U^k(\hat{\mu}^k) \cup L^k(\hat{\mu}^k) |$
and $| J^k \cup M^k \cup L_-^k \cup U_+^k | < 2 | U^k(\hat{\mu}^k) \cup
L^k(\hat{\mu}^k) |$ for 3- and 5-sets pegging, respectively.
\end{remark}
\subsection{Optimality of the relaxation algorithms} \label{sec:proofrel}
For the inequality problem (\ref{problem}), optimality and validation for the
pegging process for the primal relaxation algorithm PIR2 was established by
Bretthauer and Shetty \cite[Propositions 1--9]{BrS02}. Let $k^*$ be the
iteration where the algorithm terminates. Then Bretthauer and Shetty state
that the primal relaxation algorithm generates the following solution for
problem (\ref{problemeq}):
\begin{subequations}\label{aopt}
\begin{align}
\mu &= \mu^{k^*} = -\phi_j^\prime(x_j^{k^*})/ g_j(x_j^{k^*}), & j &\in
J^{k^*}, \label{aopt1}
\\
\rho_j &= \phi_j'(l_j^{k^*}) + \mu^{k^*} g_j'(x_j^{k^*}), & j &\in L, \label{aopt2}
\\
\rho_j &= 0, & j &\in J^{k^*} \cup U, \label{aopt3}
\\
\lambda_j &= -\phi_j'(u_j^{k^*}) - \mu^{k^*} g_j'(x_j^{k^*}), & j &\in
U, \label{aopt4}
\\
\lambda_j &= 0, & j &\in J^{k^*} \cup L, \label{aopt5}
\\
x_j &= l_j, & j &\in L, \label{aopt6}
\\
x_j &= u_j, & j &\in U, \label{aopt7}
\\
x_j &= x_j^{k^*}, & j &\in J^{k^*}. \label{aopt8}
\end{align}
\end{subequations}
\noindent Bretthauer and Shetty establish that this solution fulfills all KKT
conditions, and therefore is optimal.
The optimality and convergence for the equality problem (\ref{problemeq}) can
be established similarly to the proof for the inequality problem given in
\cite{BrS02} except for the proof for feasibility of the dual variables
corresponding to the lower and upper bounds (Propositions 8 and 9 in
\cite{BrS02}). Additionally we do not have to prove that $\mu^* \ge 0$
(Proposition 4 in \cite{BrS02}) for the equality problem. Therefore we give a
complementary proof for the feasibility of the dual variables corresponding to
the lower and upper bounds. For the equality problem (\ref{problemeq}) we have
that $g_j(x_j) = a_jx_j$ and in the proof we will assume that $a_j > 0$.
\begin{lemma}\label{nd3}
Consider the equality problem (\ref{problemeq}) and algorithm PIR2 in Section
\ref{sec:bih}.
(a) If $\nabla^k > \Delta^k$ then $\mu^{k^*} \ge \mu^k$.
(b) If $\nabla^k < \Delta^k$ then $\mu^{k^*} \le \mu^k$.
\end{lemma}
\begin{proof}
The proof is similar to that of Proposition 7 in \cite{BrS02}.
\end{proof}
\begin{proposition}[feasibility of dual variables corresponding to
bounds]\label{mythm} For problem (\ref{problemeq}), the solution (\ref{aopt})
generated by PIR2 in Section \ref{sec:bih} satisfies the feasibility of the
dual variables ($\rho_j$ and $\lambda_j$) corresponding to the lower and
upper bounds.
\end{proposition}
\begin{proof}
(i) For $j \in J^{k^*} \cup U$, we have from (\ref{aopt3})
that $\rho_j = 0$.
(ii) For $j \in L$, we have from (\ref{aopt3}) that $\rho_j = \phi_j'(l_j) +
\mu^{k^*} a_j$. We know that all variables $x_j$ with $j \in L$ were pegged in
the iterations $k$ where $\nabla^k > \Delta^k$. For these iterations $k$ we
have $\hat{x}_j^k \le l_j$. Further, from the convexity of $\phi_j$ and the
assumption that $a_j>0$, we have that $\mu (x_j) = -\phi_j'(x_j)/a_j$ is
decreasing in $x_j$. Hence,
\begin{equation}
\frac{\rho_j}{a_j} = \frac{\phi_j'(l_j)}{a_j} + \mu^{k^*} \ge
\frac{\phi_j'(\hat{x}_j^k)}{a_j} + \mu^{k^*} = -\mu^k + \mu^{k^*} \ge 0,
\end{equation}
where the last inequality follows from Lemma \ref{nd3}(a).
(iii) For $j \in J^{k^*} \cup L$, we have from
(\ref{aopt3}) that $\lambda_j = 0$.
(iv) For $j \in U$, we have from (\ref{aopt5}) that $\lambda_j = -
\phi_j'(u_j) - \mu^{k^*} a_ju_j$. We know that all variables $x_j$ with $j \in
U$ were pegged in the iterations where $\nabla^k < \Delta^k$. For these
iterations $k$ we have $\hat{x}_j^k \ge u_j$. Further, from the convexity of
$\phi_j$ and the assumption that $a_j>0$, we have that $\mu (x_j) =
-\phi_j'(x_j)/a_j$ is decreasing in $x_j$. Hence,
\begin{equation}
\frac{\lambda_j}{a_j} = -\frac{\phi_j'(u_j)}{g'_j(u_j)} - \mu^{k^*} \ge
-\frac{\phi_j'(\hat{x}_j^k)}{g'_j(\hat{x}_j^k)} - \mu^{k^*} = \mu^k -
\mu^{k^*} \ge 0,
\end{equation}
where the last inequality follows from Lemma \ref{nd3}(b).
\end{proof}
\noindent The algorithms in Sections \ref{sec:mod2}, \ref{sec:ste}, and
\ref{sec:bh4} will also converge to the optimal solution since they are
equivalent to PIR.
\begin{remark}
If we introduce the additional assumption that $\phi_j$ and $g_j$ are twice
differentiable and that $g_j^\prime >0$, an alternative convergence result for
the dual relaxation algorithm is found in \cite{Ste01}.
\end{remark}
\subsection{Time complexity}
Consider algorithm PIR2 in Section \ref{sec:bih}. In Step 0 we need at most
$2n$ comparisons to determine the primal variables and $C_0n$ operations, for
a constant $C_0$, to determine if the solution is feasible or not. In Step 1,
we solve the relaxed problem, which gives $C_1n$ operations for a constant
$C_1$. In Step 2 we perform at most $2n$ comparisons to determine the lower
and upper sets and we compute $\Delta^k$ and $\nabla^k$, which gives at most
$C_2n$ operations for a constant $C_2$. Steps 3.1 and 3.2 give at most $2n+1$
operations. Further, in the worst case the algorithm only pegs one primal
variable in each iteration, which results in $n$ iterations. Hence, we conclude
that the algorithm has a complexity of $O(n^2)$.
\section{A quasi-Newton algorithm (NZ)} \label{sec:zen_algor}
Nielsen and Zenios \cite[Section 1.4]{NiZ92} develop a quasi-Newton method for
finding the dual optimal solution $\mu^*$ of the problem (\ref{problemeq}). It
is assumed that the objective term $\phi_j$ is strictly convex with a
derivative $\phi^\prime_j$ whose range is $\mathbb{R}$. They compare their
numerical method with three linesearch methods \cite{HKL80,CeL81,Tse90}. Their
results show that their numerical method always performs well compared with
the other algorithms. They implement their algorithms on a massively parallel
computer. This is not our intention. But since the algorithm seems to perform
well on parallel computers it makes sense to evaluate it on non-parallel
computers.
Let $f_j$, $j \in J$, be the inverse of $\phi^\prime_j$ such that
\begin{equation}
\text{for } j \in J, \quad \phi_j(f_j(\mu)) = \mu, \quad \mu \in \mathbb{R}.
\end{equation}
Similar to (\ref{xmu}) we conclude that
\begin{equation}
x_j(\mu) = \max{\{l_j,\min{\{f_j(a_j\mu),u_j\}} \}} \label{min_x}.
\end{equation}
The heart of the algorithm is, like in the breakpoint algorithm, to find $\mu$
such that the primal constraint (\ref{problemb}) is fulfilled. In other words
find $\mu$ such that
\begin{equation}
\Psi(\mu) := b - \sum_{j \in J} a_jx_j(\mu) = 0.
\label{psi}
\end{equation}
Nielsen and Zenios \cite[Section 1.4]{NiZ92} define two functions $\Phi_j^+$
and $\Phi_j^-$:
\begin{equation}
\Phi_j^+(\mu) := \left\{
\begin{array}{l l l }
\min{\{f_j(a_j\mu),u_j\}}, & \quad \mathrm{if }\;a_j>0,\\
\max{\{l_j,f_j(a_j\mu)\}}, & \quad \mathrm{if }\; a_j<0,
\end{array} \right.
j \in J,
\end{equation}
and
\begin{equation}
\Phi_j^-(\mu) := \left\{
\begin{array}{l l l }
\min{\{f_j(a_j\mu),u_j\}}, & \quad \mathrm{if }\;a_j<0,\\
\max{\{l_j,f_j(a_j\mu)\}}, & \quad \mathrm{if }\; a_j>0.
\end{array} \right.
j \in J,
\end{equation}
Note that for $j \in J$, if $a_j > 0$ and $f_j$ is concave and increasing
then $\Phi^+_j$ is concave and if $a_j > 0$ and $f_j$ is convex and decreasing
then $\Phi^-_j$ is convex. Further, we define two sets of indices such that
$J^+ := \{ \, j \in J \mid a_j>0 \, \}$ and $J^- :=\{ \, j \in J \mid a_j<0 \,
\} $. Define two approximation of $\Psi$ such that
\begin{subequations}
\begin{align}
\Psi^+(\mu) & := b - \sum_{j \in J} a_j \Phi^+(\mu) \nonumber \\ & =
b- \sum_{j\in J^+} a_j \min{\{g(a_j\mu),u_j\}} - \sum _{j\in J^-}
\max{\{l_j,g(a_j\mu)\}},
\end{align}
\end{subequations}
and
\begin{subequations}\label{rox1}
\begin{align}
\Psi^-(\mu) & := b - \sum_{j \in J} a_j \Phi^-(\mu) \nonumber \\ & = b-
\sum_{j\in J^-} a_j \min{\{g(a_j\mu),u_j\}} - \sum _{j\in
J^+} \max{\{l_j,g(a_j\mu)\}}.
\end{align}
\end{subequations}
Note that if $a_j>0$ and $f_j$ is concave then $\Psi^+$ is convex
and if $a_j>0$ and $f_j$ is convex then $\Psi^-$ is concave. Define the sub- and
superdifferentials of $\Psi^+$, respectively, $\Psi^-$, as
\begin{subequations}\label{rox2}
\begin{align}
\partial \Psi^+(\mu) &:= \{ \, d \in \mathbb{R} \mid
(\Psi^+(\mu^\prime ) - \Psi^+(\mu) \ge d(\mu^\prime - \mu) \quad
\forall \mu^\prime \in \mathbb{R} \, \}, \\ \partial \Psi^-(\mu) &:=
\{ \, d \in \mathbb{R} \mid (\Psi^-(\mu^\prime) - \Psi^-(\mu) \le
d(\mu^\prime - \mu) \quad \forall \mu^\prime \in \mathbb{R} \, \}.
\end{align}
\end{subequations}
Further, define $\mu^*_\varepsilon$ and $\mathbf{x}_\varepsilon^*$ as the approximate
dual and primal solution such that $|\Psi(\mu^*_\varepsilon)| < \varepsilon$ where
$\varepsilon > 0$. The algorithm (NZ) follows (\cite[Linesearch 4]{NiZ92}):
\begin{tabular}{l}
\\ \textbf{Initialization: }
Set $\varepsilon >0$, $k = 0$, $\mu^0 \in \mathbb{R}$.
\\
\textbf{Iterative algorithm:}
\\
\textbf{Step 1} (compute step size):
\\
$\quad$ If $\Psi(\mu^k) > \varepsilon$ then
\\
$\qquad \Delta \mu ^{k+1} := -\frac{\Psi(\mu^k)}{d^k}$ where $d^k \in \partial
\Psi^+(\mu^k)$; go to Step 2,
\\
$\quad$ else if $\Psi(\mu^k) < -\varepsilon$ then
\\
$\qquad \Delta \mu ^{k+1} := -\frac{\Psi(\mu^k)}{d^k}$ where $d^k
\in \partial \Psi^-(\mu^k)$; go to Step 2,
\\
$\quad$ else \\ $\qquad$ Set
$\mu^*_\varepsilon := \mu^k$ and determine $\mathbf{x}_\varepsilon^*$ from
(\ref{min_x}). Stop.
\\
\textbf{Step 2} (dual variable update):
\\
$\quad$ set
$\mu^{k+1} := \mu^k + \Delta \mu ^{k+1} $.
\\
\textbf{Step 3: } Set $k := k +1$ and go to
Step 1. \\ \\
\end{tabular}
\noindent The algorithm converges to a value $\mu^*_\varepsilon$, such that
$|\Psi(\mu^*_\varepsilon)| < \varepsilon$ if the objective function components
$\phi_j$ is such that the corresponding function $\Psi^+(\mu)$ is convex or if
the corresponding function $\Psi^-(\mu)$ is concave \cite[Proposition
8]{NiZ92}. For some problems, the inverse of the derivative might however
result in imaginary values. One solution to this problem is to consider the
equivalent maximization problem of (\ref{problemeq}), i.e., to $\maximize_x
-\phi(x)$.
\begin{remark}
In practice we will choose $\varepsilon > 0$ and the algorithm will in most
cases stop such that $| \Psi(\mu^k) | > 0$. Hence we will end up with an
approximate solution of the optimal dual variable $\mu^*$. The map from the
dual space to the primal might not be linear. Hence the primal error might be
larger than we expect, i.e., $|x _\varepsilon^* - x^*| >> | \mu _\varepsilon^*
- \mu^* | $. However, there are methods for generating primal optimal
solutions from any Lagrangian dual vector (see for example
\cite{LMOP07}). Another plausible method to find the optimal solution from the
approximate solution $\mu_{\varepsilon}^*$ is to use a breakpoint or a
relaxation algorithm, starting from $\mu_{\varepsilon}^*$.
\end{remark}
\section{Method of algorithm evaluation}\label{sec:method}
This section serves to provide an overview of the procedure for the numerical
study. In Section \ref{sec:limitations} we define problem instances for the
numerical study. Some theory on how the problem instances can be designed
follows in Section \ref{sec:dop}. In Section \ref{sec:performance} we give a
brief overview of performance profiles (\cite{DoM02}) which are used for the
evaluation of the numerical study. Finally, in Section \ref{sec:plc}, we
describe the computational environment.
\subsection{Problem set}\label{sec:limitations}
For the numerical study we consider five common special cases of problem
(\ref{problemeq}); it also covers the inequality problem (\ref{problem}) when
$\mu^* > 0$. Only finite values of the lower an upper bounds
(\ref{problemeqc}) are considered, i.e., for $j \in J$, $l_j > -\infty$ and
$u_j < \infty$. The five problem cases are briefly specified next:
\paragraph{Quadratic problem}
The convex separable quadratic problem is the special case of
(\ref{problemeq}), where
\begin{equation}\label{quadratic}
\phi _j ( x_j) = \frac{w_j}{2}x^2_j - c_jx_j \; \text{ for } j \in J,
\end{equation}
where $w_j, c_j > 0$, $j \in J$. Numerical studies of algorithms for problem
(\ref{quadratic}) are widely explored; e.g., see
\cite{NiZ92,Kiw07,Kiw08a,Kiw08b}. In our numerical study the parameters are
randomized such that $a_j \in [1,30]$, $w_j \in [1,20]$, $c_j \in [1,25]$,
$l_j \in [0,3]$ and $u_j \in (3,11]$.
\paragraph{Stratified sampling}
If we have a large population and would like to perform a statistical research
among the population, it is practically infeasible to examine every single
individual in the population. Instead we can stratify the population into $n$
strata. An example of a stratum might be people of a certain age. Let $M$ be
the number of individuals in the entire population and $M_j$ the number of
individuals in strata $j$. If we want to minimize the variance of the entire
population we need to allocate the number of samples $x_j$ from each strata
from:
\begin{equation} \label{stratified}
\phi_j(x_j) = \omega_j \frac{(M-x_j)\rho ^2}{(M-1)x_j} \; \text{ for } j \in J,
\end{equation}
where $\omega_j = M_j/M$ and $\rho_j$ is an estimate of the variance for
strata $j$. From $b$ in the resource constraint we specify the total sample
size. In our numerical study the parameters are
randomized such that $a_j \in [1,30]$, $m_j \in [5,30]$, $c_j \in [1,4]$, $l_j
\in [1,3]$ and $u_j \in (3,15]$.
\paragraph{Sampling}
According to \cite{BRS99}, often the objective terms for sampling problems can
be written as
\begin{equation}\label{production}
\phi_j (x_j) = c_j/x_j \; \text{ for } j \in J.
\end{equation}
In our numerical study the parameters are randomized such that $a_j \in
[1,4]$, $c_j \in [5,30]$, $l_j \in [0,3]$ and $u_j \in (3,6]$.
\paragraph{The theory of search}
In the theory of search problem it is determined how a resource $b$ of time
should be spent to find an object among $n$ subdivisions of an area with the
largest probability. It is assumed that we know the probability $m_j$ for an
object to be found in subdivision $j$. The objective component $\phi_j$
describes the probability of finding the object in subdivision $j$ and takes
the form:
\begin{equation}\label{TEO}
\phi_j (x_j) = m_j(e^{-b_jx_j}-1) \; \text{ for } j \in J.
\end{equation}
This problem is widely explored by Koopman \cite{Koo53,Koo99} and the problem
is possible to apply to a large variation of search problems, e.g., searching
for refugees fleeing from Cuba \cite{Sto81}. In our numerical study the
parameters are randomized in the following intervals: $m_j \in [0.5, -8]$,
$b_j \in [0.1, 3]$, $a_j \in [1, 3]$, $l_j \in [0, 0.1]$ and $u_j \in (0.1,
5]$.
\paragraph{Negative entropy function}
The negative entropy function mentioned in \cite{NiZ92}:
\begin{equation} \label{negentropy}
\phi_j(x_j) = x_j\log{(\frac{x_j}{a_j} - 1)} \; \text{ for } j \in J.
\end{equation}
In our numerical study the parameters are
randomized such that $c_j \in [50,250]$, $l_j \in [20,100]$ and $u_j \in
(30,210]$.
\subsection{Design of problem instances} \label{sec:dop}
Similarly to the numerical study in \cite{KoL98}, we divide our set of test
problem instances into groups containing different portions of the activities
within the lower and upper bounds at the optimal solution. Let $H := \{ \, j
\in J \mid l_j < x^*_j < u_j \, \}$. Then, the percentage is determined as
$|H| / n$. To motivate this approach, we refer to the variance of CPU times for
different portions of active activities in \cite{KoL98}. Further, similar to
the numerical study in \cite{Kiw07}, we consider different problem sizes $n$
since the theoretical CPU time for different algorithms vary between $O(n)$
and $O(n^2)$.
For the problem set, we pseudo randomize the parameters $l_j$, $u_j$ and all
the parameters associated with $\phi_j$ and $g_j$. In the numerical study we
use a linear resource constraint such that $g_j(x_j) = a_jx_j$, where $a_j>0$
for all $j \in J$. This simplifies the design of the problem set, since
$\phi_j$ is convex and $l_j < u_j$ for all $j \in J$. We have that
\begin{equation}\label{designprob2}
x_j^* = \left\{
\begin{array}{l l l }
x_j^*, & \quad \text{if } \mu^* = -\phi_j^\prime(x_j^\prime)/ a_j, \\ l_j,
& \quad \text{if } \mu^* \ge -\phi_j^\prime(l_j) / a_j \ge
-\phi_j^\prime(u_j) / a_j, \\ u_j, & \quad \text{if } \mu^* \le
-\phi_j^\prime(u_j) / a_j ) \le -\phi_j^\prime(l_j) / a_j.
\end{array} \right.
\end{equation}
\noindent By using the properties of (\ref{designprob2}) we can determine $H$
such that $|H| / n = y$ for any $y \in [0,1] $.
\subsection{Performance profiles} \label{sec:performance}
Dolan and Mor\'{e} \cite{DoM02} propose a performance profile for the
evaluation of optimization software. The method is briefly summarized as
follows: Assume that we have a set $A$ of algorithms that consist of $n_{a}$
algorithms and a problem set $P$ that consists of $n_p$ problem instances. Let
$t_{p,a}$ denote the time it takes for algorithm $a\in A$ to solve problem
$p\in P$. A performance ratio $r_{p,a}$ is introduced:
\begin{equation}
r_{p,a}(t_{p,a}) := \frac{t_{p,a}}{\min \{t_{p,l} \mid l\in A \}}.
\end{equation}
For a problem $p$, the performance ratio is a measure of how fast algorithm
$a$ is relative to the fastest algorithm solving problem $p$. Fix a constant
$r_M$ such that $r_M \ge r_{p,a}$ for all $p\in P$, $a \in A$
and let $r_{p,a} = r_M$ if algorithm $a$ fails to solve problem $p$. Further,
we introduce the distribution
\begin{equation}
\rho_a (\tau) = \frac{1}{n_p} | \{p\in P \mid r_{p,a} \le \tau \} |,
\end{equation}
for each algorithm, where $|\cdot|$ denote the cardinality of a set and $\tau
\in [1, r_M]$. For algorithm $a$, the distribution $\rho_a$ describes the
percentage of problem instances that are solved at least as fast as $\tau$
times the fastest algorithm for problem $p$. Note that $\rho_a(1)$ is the
percentage for algorithm $a$ being the fastest. Moreover, $\lim_{\tau
\rightarrow r_M} \rho_s(\tau)$ is the probability that algorithm $a$ will
solve a problem $p$ in $P$. If we have a large problem set $P$ then $\rho_a
(\tau )$ will not be affected much by a small change in $P$ (\cite[Theorem
1]{DoM02}).
\subsection{Program language, computer and code}\label{sec:plc}
The algorithms are implemented in Fortran 95, compiled with {\em gfortran}
under mac OS X 10.8.2 (2.5 GHz Intel Core i5, 4 GB 1600 MHz DDR3).
\section{Computational experiments} \label{sec:compexp}
In this section we present the results from numerical experiments of the
problems defined in Section \ref{sec:limitations}. If nothing else is
mentioned, 100 problem instances of each problem is evaluated for each problem
size. Further, the problem instances are designed such that different values
of $|H| / n$ are considered, see Section \ref{sec:dop}.
The development of numerical experiments for the problem (1) is illustrated in
Table \ref{decade}, where we cite the size of the largest test problem
reported during each decade; the table is an extension of Table 2 in
\cite{Pat08}.
\begin{table}[h!]\
\centering
\caption{Largest problem instances solved for each algorithm class through the
decades.}
\begin{tabular}{@{} l | r r r r r r r @{}} \label{decade}
Decade & 50s & 60s & 70s & 80s & 90s & 00s & 10s \\ [1ex] \hline Breakpoint
algorithm & 2 & 60 & 12 & 200 & $10^4$ & $2 \cdot 10^6$ & $30 \cdot 10^6$
\\[1ex] Relaxation algorithm & -- & -- & 200 & 200 & $10^4$ & $2
\cdot 10^6$ & $110 \cdot 10^6$ \\[1ex]
\end{tabular}
\end{table}
In Section \ref{sec:Bhprim} we present performance profiles for the relaxation
algorithms defined in Section \ref{sec:relaxation}. In Section
\ref{sec:resrank} we evaluate the pegging process for both the breakpoint and
the relaxation algorithm defined in Section \ref{sec:rank}. In Section
\ref{sec:final}, the best performing relaxation algorithm and breakpoint
algorithm are compared with the quasi-Newton algorithm in Section
\ref{sec:zen_algor}. In Section \ref{sec:acr} we give a critical review of the
numerical study. Finally, in Section \ref{sec:conclusion}, we make some
overall conclusions.
In the following Figures~\ref{bhvsdualbh}--\ref{finall} we show performance
profiles, as defined in Section~\ref{sec:performance}, for three competing
sets of algorithms. To summarize the appearance of these plots, for each plot
and corresponding competing set of algorithms, the graph for one algorithm
shows the portion (between $0$ and $1$ on the y-axis) of the problem instances
considered that are solved within $\tau$ (on the x-axis) times the fastest
algorithm in the test for problem $p$. In particular, the value of $\rho_a(1)$
is the portion of the problems in which algorithm $a$ is the fastest, and
$\lim_{\tau \rightarrow r_M} \rho_a(\tau)$ is the probability that algorithm
$a$ will solve a problem $p$ in $P$. (The latter information is particularly
illustrative and relevant for Figure~\ref{finall}.)
\subsection{Evaluation of the relaxation algorithm}\label{sec:Bhprim}
We evaluate the different paths in Figure \ref{relaxationOverview}. First,
recall that PIR2 determines the primal variables while DIR2 determines the
dual variable. As can be seen in the leftmost performance profile in
Figure~\ref{bhvsdualbh}, DIR2 is the fastest in $67.8\%$ of the problem
instances solved. The result is what we can expect from theory, since DIR2
does not need as many operations in Step 1; see Sections \ref{sec:bih} and
\ref{sec:mod2}. PIR2 is faster in $32.2\%$ of the problems solved; the latter
cases stem mainly from to the negative entropy problem (\ref{negentropy}) but
also from the theory of search problem (\ref{TEO}) and the stratified sampling
problem (\ref{stratified}) when $n$ and $| H | /n$ is small. The reason for
PIR2 to be faster in almost all cases for the negative entropy problem is due
to the computationally simple expression of the primal variables $x_j^k(\mu^k)
= \frac{a_j}{\sum_{j\in J^k} c_j}$ while the dual variable is evaluated from
$\mu^k = \log{ \sum_{j \in J^k} c_j } - \log{b}$. Also the computations of the
breakpoints might be a larger part of the total time when the equations for
the primal/dual variables are simple.
\begin{figure}[h!]
\centering
\includegraphics[width=3.12in]{PIRvsDIR.eps}
\includegraphics[width=3.12in]{DIRvsDERvsDBR.eps}
\caption{The leftmost figure shows the performance profiles for PIR2 and
DIR2 and the rightmost for DER2, DIR2 and DBR2. The numerical experiment
was done according to the description in Section \ref{sec:dop} for $n = 50
\cdot 10^3; 100 \cdot 10^3; 200 \cdot 10^3; 500 \cdot 10^3; 10^6; 2 \cdot
10^6$ and for the problems in Section \ref{sec:limitations} . The
algorithms were implemented uniformly.}
\label{bhvsdualbh}
\end{figure}
We recall that DER2 applies explicit evaluation, DIR2 applies implicit
evaluation and DBR2 applies the theoretically most profitable evaluation. The
results are in favour of DBR2 and DIR2; see the rightmost performance profile
in Figure \ref{bhvsdualbh}. DBR2 is fastest in $70.6\%$, DIR2 is fastest in
$25.9\%$ and DER2 is fastest in $4.7\%$ of the problems solved. (Note that
$70.6\% + 25.9\% + 4.7\% = 100.2 \%$; in some cases two algorithms are equally
fast.)
Figure \ref{bhvsdualbh} also shows that the performance of DIR2 and DBR2 are
quite similar. For just a few cases DIR2 is more than 1.30 times slower than
the fastest algorithm while for as few cases DBR2 is only 1.2 times slower
than the fastest algorithm. On average DBR2 performs somwhat better than DIR2.
\paragraph{Conclusion}
For the problem set considered, it is more profitable to evaluate the dual
variable even if we have to compute all the breakpoints at the beginning of
the algorithm. Hence for the problem set considered DIR2 seems to outperform
PIR2.
For the problem set considered, the results show that in most cases it is more
profitable to evaluate a solution $\mathbf{x}^k$ implicitly. The small
difference between the performance of DBR2 and DIR2 implies that in most cases
$|J^k|>2|L(\mu^k) \cup U(\mu^k)|$. We conclude that DBR2 performs slightly
better than DIR2 which agrees with the theory in Section \ref{sec:bh4}.
\subsection{Evaluation of the pegging process}\label{sec:resrank}
We compare the three pegging approaches for the breakpoint and relaxation
algorithms described in Sections \ref{sec:medians}--\ref{sec:MB5} and
\ref{sec:tandf}, respectively. For the breakpoint algorithm median search is
implemented similarly to \cite[Section 8.5]{PTVF92}.
The performance profiles in Figure \ref{rmrmm} show that 5-sets pegging is
nearly always the fastest (MB5 in $95\%$ and DBR5 in $94\%$ of the problem
instances solved). Additionally, when 5-sets pegging is not fastest it is
never more than $10 \%$ slower than the fastest algorithm. It is obvious that
for the few cases when 2-sets pegging performs best are when few optimal
primal variables equals the lower or upper bounds ($| H | /n$ is small). This
follows from the fact that 2-sets pegging does not check if the variables
belong to the additional sets that are used in 3- and 5-sets pegging.
\begin{figure}[h!]
\centering
\includegraphics[width=3.12in]{MBvsMMB.eps}
\includegraphics[width=3.12in]{DBR235.eps}
\caption{The leftmost figure shows the performance profile for MB2, MB3 and
MB5 and the rightmost for DBR2, DBR3 and DBR5. The numerical experiment
was done according to the description in Section \ref{sec:dop} for $n = 50
\cdot 10^3; 100 \cdot 10^3; 200 \cdot 10^3; 500 \cdot 10^3; 10^6; 2 \cdot
10^6$ and for the problems in Section \ref{sec:limitations}. The
algorithms were implemented uniformly. }\label{rmrmm}
\end{figure}
\paragraph{Conclusion} In general 5-sets pegging is the most profitable.
\begin{remark}
Numerical experiments for a breakpoint algorithm applying bisectional search
of a sorted sequence of breakpoints, using quicksort as in \cite[Section
5.2.2]{Knu98}, was also performed. The algorithm was on average 1.5 times
slower than MB2. Also the algorithm in \cite{Zip80}, where the upper bounds
are relaxed, was evaluated. The algorithm performed poorly, on average 2.9
times as slow as the breakpoint algorithm applying sorting.
\end{remark}
\subsection{A comparison between relaxation, breakpoint, and quasi-Newton
methods}\label{sec:final}
In this section we compare the best performing relaxation algorithm (DBR5),
the best performing breakpoint algorithm (MB5), and the numerical quasi-Newton
method (NZ) in Section \ref{sec:zen_algor}. For NZ we use the stopping
criterion $| \frac{\sum_{j\in J} a_jx_j}{b} - 1 | < 0.01$ which is weaker than
the stopping criterion in \cite{NiZ92} ($<10^{-4}$). This will of course give
us an approximate optimal solution only. The initial value of the dual
variable $\mu$ is set to the mean of the breakpoints $\mu^0 = ( \sum_{j\in J}
\phi_j(l_j) / a_j + \sum_{j\in J} \phi_j(u_j) / a_j) / (2n)$. If the algorithm
does not converge within 100 CPU seconds, we start over with the mean of the
breakpoints corresponding to the lower breakpoints $\mu^0 = ( \sum_{j\in J}
\phi_j(l_j) / a_j) / n$. Similarly, if the algorithm does not converge within
100 CPU seconds then we start over with the mean of the breakpoints
corresponding to the upper breakpoints: $\mu^0 = ( \sum_{j\in J} \phi_j(u_j) /
a_j) / n$. If the algorithm does not terminate within $300$ CPU seconds then
we terminate the algorithm and consider the problem as unsolved.
The reader may have noticed that the breakpoint algorithms in Sections
\ref{sec:medians}--\ref{sec:MB5} are given without the use of the sets for the
pegging process ($L$, $U$, $M$, $L_-$, $U_+$) while the relaxation algorithms
in Section \ref{sec:bih}--\ref{sec:tandf} are given with the use of these
sets. Of course the notation using these sets, as for the relaxation
algorithms, are theoretically more elegant. However, our implementations of
the algorithms in Fortran 95 showed that for breakpoint algorithms it was in
general more profitable to not use the pegging sets.
Figure \ref{finall} shows the performance profile for the algorithms. In
general, we can see a significant advantage of using DBR5 since it is fastest
for $99.0 \%$ of the problem instances solved. Moreover, when DBR5 is not the
fastest algorithm it is almost as fast as the fastest. From the figure we also
see that MB5 is more than 2.7 times slower than the fastest algorithm in $50
\%$ of the problem instances solved.
\begin{figure}[h!]
\centering
\includegraphics[width=3.1in]{final_all.eps}
\includegraphics[width=3.1in]{fin_all_large.eps}
\caption{Performance profile for DBR5, MB5 and NZ. (Left) The numerical
experiment was done according to the description in Section \ref{sec:dop}
for $n = 50 \cdot 10^3$, $100 \cdot 10^3$, $200 \cdot 10^3$, $500 \cdot
10^3$, $10^6$, and $2 \cdot 10^6$ for the problems in
Section~\ref{sec:limitations}. (Right) The numerical experiment was done
according to the description in Section \ref{sec:dop} but with 10 problem
instances and for $n = 4 \cdot 10^6$, $6 \cdot 10^6$, $8 \cdot 10^6$, $10
\cdot 10^6$, $15 \cdot 10^6$, $20 \cdot 10^6$, $25 \cdot 10^6$, and $30
\cdot 10^6$ for the problems in
Section~\ref{sec:limitations}. }\label{finall}
\end{figure}
The Newton method terminated the fastest for only $1 \%$ of the test problems,
which probably is due to good initial values. Also we should have in mind that
NZ terminates with an approximate solution only. The Newton method performs
relatively well for the quadratic problem (\ref{quadratic}), the theory of
search problem (\ref{TEO}), and the negative entropy problem
(\ref{negentropy}) when $|H|/n>0.8$. It is not very successful for the
stratified sampling problem (\ref{stratified}) and the sampling problem
(\ref{production}). We can see from the performance profile that in $28\%$ of
the problem instances tested, NZ is more than 5.5 times slower than DBR5 (the
fastest algorithm). In $5.3 \%$ $(158/3000)$ of the problem instances the
algorithm does not solve the problem; these problem instances mainly stem from
the stratified sampling problem (\ref{stratified}) and the sampling problem
(\ref{production}) when $|H|/n< 0.3$.
By comparing the performance profiles in Figure \ref{finall}, we can see that
they are very similar. In other words, relative to each other, the performance
of the algorithms seems to be similar for different problem sizes $n$.
\begin{figure}[h!]
\centering \includegraphics[width=3.5in]{htgraf.eps}
\caption{The average CPU time for $n=10^6$ is plotted as a function of the
average portion of the optimal variables that obtain a value between the
lower and upper bounds. }\label{htgraf}
\end{figure}
Considering the breakpoint algorithm MB5, Figure \ref{htgraf} shows that the
CPU time increases when the number of optimal primal variables strictly within
the bounds $|H|/n$ grows for the problems in the problem set. Concerning the
relaxation algorithm DBR5 the result shows no significant dependence of
$|H|/n$, neither does NZ.
\begin{figure}[h!]
\centering
\includegraphics[width=3.1in]{timecomp.eps}
\includegraphics[width=3.1in]{timecomp2.eps}
\caption{The average CPU time is plotted as a function of the size of the
problem $n$ for DBR5 and MB5. The values in the left figure are the mean
of 100 problem instances solved and the values in the right figure are the
mean of 10 problem instances solved.}\label{naverage}
\end{figure}
In Figure \ref{naverage}, we can see that the CPU time for MB5 and DBR5 is
linear for $n \in [50 \cdot 10^3, 2 \cdot 10^6]$ respectively $n \in [50 \cdot
10^3, 30 \cdot 10^6]$. This is impressive, since DBR5 has a worst-case time
complexity of $O(n^2)$. For problem sizes a bit larger than $30 \cdot 10^6$
variables the CPU time tends to increase faster. However, additional
computations on a more powerful computer have shown that, for the relaxation
algorithm, the quadratic problem has a linear time complexity up to $n= 90
\cdot 10^6$, the sampling problem up to $n = 100 \cdot 10^6$, the theory of
search problem up to $n = 70 \cdot 10^6$ and the negative entropy problem up
to $n = 110 \cdot 10^6$. Hence we would also like to stress that the linearity
probably is constrained by the memory of the computer rather than the algebra
in the algorithm.
Tables \ref{tqua}--\ref{tneg} show the mean CPU time in seconds, for solving
the problems in the problem set for DBR5, NZ and MB5. The tables show the
great advantage of using DBR5.
\begin{table}[h!]
\centering
\caption{The average CPU times algorithms MB5, DBR5 and NZ for solving the
quadratic problem (\ref{quadratic}). Each value is the mean of 100
randomized computations and the CPU times are given in seconds.}
\begin{tabular}{@{} c c c c c c c @{} } \label{tqua}
$n$ & 50,000 & 100,000 & 200,000 & 500,000 & 1,000,000 &
2,000,000\\ [1ex] \hline DBR5 & 0.0031 & 0.0065 & 0.0136 & 0.0346 & 0.0702 &
0.1438 \\[1ex] NZ & 0.0083 & 0.0179 & 0.0378 & 0.1018 & 0.1923 & 0.3822
\\[1ex] MB5 & 0.0097 & 0.0205 & 0.0428 & 0.1144 & 0.2346 & 0.4807 \\[1ex]
\hline
\end{tabular}
\end{table}
\begin{table}[h!]\
\centering
\caption{The average CPU times algorithms MB5, DBR5 and NZ for solving the
stratified sampling problem (\ref{stratified}). Each value is the mean of
100 randomized computations and the CPU times are given in seconds. For NZ
only the cases when NZ found the optimal solution have been considered.}
\begin{tabular}{@{} c c c c c c c @{}} \label{tstr}
$n$ & 50,000 & 100,000 & 200,000 & 500,000 & 1,000,000 &
2,000,000\\ [1ex] \hline DBR5 & 0.0040 & 0.0082 & 0.0167 & 0.0436 & 0.0875 &
0.1773 \\[1ex] NZ & 0.0379 & 0.0952 & 0.1529 & 0.3560 & 0.9026 & 1.8299
\\[1ex] MB5 & 0.0098 & 0.0214 & 0.0438 & 0.1178 & 0.2390 & 0.4981 \\[1ex]
\hline
\end{tabular}
\end{table}
\begin{table}[h!]
\centering
\caption{The average CPU times algorithms MB, DBR5 and NZ for solving the
sampling problem (\ref{production}). Each value is the mean of 100
randomized computations and the CPU times are given in seconds. For NZ only
the cases when NZ found the optimal solution have been considered.}
\begin{tabular}{@{} c c c c c c c @{}} \label{tpro}
$n$ & 50,000 & 100,000 & 200,000 & 500,000 & 1,000,000 &
2,000,000\\ [1ex] \hline DBR5 & 0.0032 & 0.0069 & 0.0146 & 0.0375 & 0.0788 &
0.1597 \\[1ex] NZ & 0.1069 & 0.2403 & 0.5670 & 1.4319 & 3.8036 & 6.3949
\\[1ex] MB5 & 0.0089 & 0.0192 & 0.0397 & 0.1066 & 0.2213 & 0.4522 \\[1ex]
\hline
\end{tabular}
\end{table}
\begin{table}[h!]
\centering
\caption{The average CPU times algorithms MB5, DBR5 and NZ for solving the
theory of search problem (\ref{TEO}). Each value is the mean of 100
randomized computations and the CPU times are given in seconds.}
\begin{tabular}{@{} c c c c c c c @{}} \label{ttos}
$n$ & 50,000 & 100,000 & 200,000 & 500,000 & 1,000,000 &
2,000,000\\ [1ex] \hline DBR5 & 0.0063 & 0.0132 & 0.0259 & 0.0679 & 0.1382 &
0.2778 \\[1ex] NZ & 0.0187 & 0.0434 & 0.0940 & 0.2669 & 0.4511 & 0.9258
\\[1ex] MB5 & 0.0127 & 0.0273 & 0.0549 & 0.1454 & 0.2983 & 0.6090 \\[1ex]
\hline
\end{tabular}
\end{table}
\begin{table}[h!]
\centering
\caption{The average CPU times algorithms MB5, DBR5 and NZ for solving the
negative entropy problem (\ref{negentropy}). Each value is the mean of 100
randomized computations and the CPU times are given in seconds.}
\begin{tabular}{@{} c c c c c c c @{} } \label{tneg}
$n$ & 50,000 & 100,000 & 200,000 & 500,000 & 1,000,000 & 2,000,000\\ [1ex]
\hline DBR5 & 0.0036 & 0.0074 & 0.0157 & 0.0406 & 0.0837 & 0.1698 \\[1ex] NZ
& 0.0071 & 0.0171 & 0.0310 & 0.0819 & 0.1709 & 0.3191 \\[1ex] MB5 & 0.0096 &
0.0206 & 0.0427 & 0.1131 & 0.2326 & 0.4774 \\[1ex] \hline
\end{tabular}
\end{table}
\paragraph{Summary}
Even if NZ only finds an approximate solution and even if the stop criterion
is weak, NZ is outperformed by DBR5. Hence, we see no reason to evaluate
methods for finding the exact optimal solution $\mathbf{x}^*$ from the
approximate solution $\mathbf{x}^*_\varepsilon$ generated by NZ. In many
cases, NZ fails to solve the stratified sampling problem (\ref{stratified})
and the sampling problem (\ref{production}). This is probably due to the
stiffness of the problems.
MB5 performs very well for problems where $|H|/n$ is small. This is due to the
pegging process: since we peg a large part of $x_j^k \in J^k$ in each
iteration, the problem is reduced in the next iteration. This follows from the
reduction of breakpoints by half in each iteration; note that this does not
hold for the relaxation algorithm. This might be the reason for MB5:s
dependency of $|H|/n$. However, DBR5 performs best for almost all of the
problem instances solved and it is worth noting that DBR5 never performs
poorly.
\subsection{A critical review}\label{sec:acr}
When we set up the initial values for NZ, we tried different initial values
and concluded that the one used was the most profitable on average. However,
especially if we customize initial values for each problem, there may still be
potential improvements available.
Considering the results in Section \ref{sec:Bhprim}, we consider the dual
algorithm (DIR2) to outperform the primal algorithm (PIR2) since it performs
better in general. Then we continued to evaluate several modification of the
dual algorithm while similar modifications of the primal relaxation algorithm
were not evaluated, i.e., we don't evaluate blended evaluation or 3- and 5-sets
pegging for the primal algorithm. However since the plausible modification of
the primal algorithm would take the same form as for the modification of the
dual algorithm it is assumed that a modified dual algorithms will outperform a
similar modified primal algorithm.
\section{Conclusion}\label{sec:conclusion}
We have complemented the survey in \cite{Pat08} on the resouce allocation
problem at hand, and introduced, and critically evaluated, new implementations
of breakpoint and relaxation algorithms for its solution.
The results show that our new implementations (DIR2 and DBR2) of the
relaxation algorithm outperform the earlier algorithms (PIR2 and DER2). Hence
we should evaluate the dual variable for the relaxation algorithm, i.e., DIR2
outperforms PIR2. Moreover, it is more profitable in theory, as well as in
practice, to apply blended evaluation, i.e., DBR2 outperforms DER2 and
DIR2. Our results, as well as the results in \cite{RJL92,KoL98,Kiw08b}, imply
that the relaxation algorithm is to prefer when a closed form of the dual
variable $\mu$ can be found.
We introduced 3- and 5-sets pegging for the relaxation algorithm and showed
that it is most profitable to apply 5-sets pegging, which also holds for the
breakpoint algorithm.
For the problems considered, MB5 and DBR5 have a practical time complexity of
$O(n)$ also for very large values of $n$. We also showed that the relaxation
algorithm DBR5 performs better than both the Newton-like algorithm NZ and the
breakpoint algorithm MB5. Potential future improvements include the
implementation of a pegging method for a Newton-like algorithm and/or a hybrid
of the different algorithms (NZ, DBR5 and MB5), and, hence, it would be of
interest to compare the best algorithms from our study to these.
The findings made herein can most certainly be profitably utilized also in the
efficient solution of the more complex versions of the resource allocation
problem discussed, for example, in the books \cite{Mje83,IbK88,Lus12}.
\bibliographystyle{alpha}
|
\section{Introduction}
A finite word $u$ is \emph{bordered}, if there exists a non-empty
word $z$ which is a proper prefix and a suffix of $u$. The
following theorem states a well-known connection between
periodicity and unbordered factors:
\begin{theorem}\emph{\cite{EhrSil,Saari}} \label{Theorem1}
An infinite word $w$ is purely periodic if and only if
there exists a constant $C$ such that every factor $v$ of $w$ with $|v|\geq C$ is bordered.
\end{theorem}
In this paper we are interested in extending this result to
the abelian and weak abelian settings.
Two finite words $u$ and $v$ are \emph{abelian equivalent} if and only if for
each letter $a,$ the number of occurrences of $a$ in $u $ (denoted $|u|_a)$ is equal to $|v|_a$.
In other words, $u$ and $v$ are permutations of one another.
An infinite word $w$ is called \emph{abelian ultimately periodic} if $w = u v_1v_2v_3 \cdots$, where
the $v_i$ are pairwise abelian equivalent.
An infinite word $w$ is called \emph{weakly abelian ultimately periodic}
if $w = u v_1v_2v_3 \cdots$, where $v_i$ are finite words with the same frequencies of letters for $i\geq 1$.
In other words, a word is weakly abelian ultimately periodic if it can be
factored into some prefix $u$ followed by infinitely many words of
possibly different lengths with the same letter frequencies. For
more on weak abelian periodicity see \cite{wap}.
We say that a finite
word $u$ is \emph{abelian bordered} if $u$ contains a non-empty proper
prefix which is abelian equivalent to a suffix of $u.$
Abelian bordered words were recently considered in
\cite{CCCI,RRS}. A finite word $u$ is \emph{weakly abelian
bordered} if it has a non-empty proper prefix and a suffix with
the same letter frequencies.
In Sections \ref{wab2} and \ref{wabk}, we consider weak abelian
borders in the binary and non-binary cases, respectively. We prove
that the condition on finitely many weak abelian unbordered
factors implies weak abelian periodicity plus some additional
restrictions (Theorems \ref{wap} and \ref{wap1}).
In Section~\ref{ab}, we consider infinite words having only finitely many abelian unbordered factors.
We do not know if such words are necessarily abelian periodic.
However, we are able to show that such words have bounded abelian complexity.
In Section~\ref{5}, we provide an answer to a question from
\cite{akp} concerning abelian critical factorization theorem.
In the last section we summarize the results and propose some open problems.
\section{Preliminaries}
In this section we give some basics on words following terminology
from \cite {Lo} and introduce our notions.
Given a finite non-empty set $\Sigma$ (called the alphabet), we
denote by $\Sigma^*$ and $\Sigma^{\omega}$, respectively, the set
of finite words and the set of (right) infinite words over the
alphabet $\Sigma$. Given a finite word $u = u_1 u_2 \cdots u_n$
with $n \geq 1$ and $u_i \in \Sigma$, we denote the length $n$ of $u$ by $|u|$.
The empty word will be denoted by $\varepsilon$ and we set $|\varepsilon| = 0$.
Given the words $w$, $x$, $y$, $z$ such that $w = xyz$, $x$ is called a \emph{prefix},
$y$ is a \emph{factor} and $z$ a \emph{suffix} of $w$.
The word $x$ is a \emph{proper prefix} if $0<|x|<|w|$.
We let $w[i, j] = w_iw_{i+1} \cdots w_j$ denote the factor starting at position $i$ and ending at position $j$
of $w=w_1w_2w_3\cdots$, where the $w_k\in\Sigma$.
An infinite word $w$ is \emph{ultimately periodic} (or briefly \emph{periodic})
if for some finite words $u$ and $v$ it holds $w=uv^{\omega} = uvv\cdots$; $w$ is
\emph{purely periodic} if $u=\varepsilon$.
An infinite word is \emph{aperiodic} if it is not ultimately periodic.
Given a finite word $u = u_1 u_2 \cdots u_n$ with $n \geq 1$ and
$u_i \in \Sigma$, for each $a \in \Sigma$, we let $|u|_a$ denote
the number of occurrences of the letter $a$ in $u$.
Two words $u$ and $v$ in $\Sigma^*$ are \emph{abelian equivalent},
denoted $u\sim_{ab} v$, if and only if $|u|_a = |v|_a$ for all $a\in \Sigma$.
It is easy to see that abelian equivalence is an equivalence relation on $\Sigma^*$.
An infinite word $w$ is called \emph{abelian ultimately periodic}
(or briefly \emph{abelian periodic})
if $w = u v_1v_2v_3 \cdots$, where $v_i \sim_{ab} v_j$ for all integers $i, j \geq 1$.
For a finite non-empty word $w\in\Sigma^*$, we define the \emph{frequency}
$\rho_a(w)$ of a letter $a\in\Sigma$ in $w$ as $\rho_a(w)=\frac{|w|_a}{|w|}$.
For an infinite word $w\in\Sigma^\omega$ and a letter $a\in\Sigma$, if the limit
\[
\lim_{n\to\infty}\rho_a(w[1..n])
\]
exists, then we define the \emph{frequency} $\rho_a(w)$ of $a$ in
$w$ to be the latter limit.
The infinite word $w$ has
\emph{uniform letter frequencies} if, for every letter $a$ of $w$, the
ratio $\frac{|w[k..k+n]|_a}{ n+1 }$ has a limit $\rho_a(w)$ when $n\to \infty$, uniformly in $k$.
In general the existence of uniform
frequencies is stronger than (prefix) frequencies, but in our
consideration in fact all the words have uniform frequencies.
Now we provide some background on weak abelian periodicity from
\cite{wap}.
\begin{definition}
An infinite word $w$ over an alphabet $\Sigma$ is called \emph{weakly abelian ultimately periodic} (or briefly \emph{weakly abelian periodic})
if $w = u v_1v_2v_3 \cdots$, where $\rho_a(v_i) = \rho_a(v_j)$
for all $a\in\Sigma$ and all integers $i, j\geq 1$.
\end{definition}
An infinite word $w$ is called \emph{bounded weakly abelian
periodic}, if it is weakly abelian periodic with bounded lengths of
blocks, i.e., there exists $C$ such that for every $i$ we have $|v_i|\leq C$.
\begin{example}
The word $(01)(0^21^2)(0^31^3)(0^41^4)\cdots$ is weakly abelian periodic
but not bounded weakly abelian periodic.
\end{example}
We make use of the following geometric interpretation
of weak abelian periodicity.
We translate an infinite word $w= w_1 w_2 \cdots \in
\Sigma^{\omega}$ to a graph visiting points of the lattice
$\mathbb{Z}^{|\Sigma|}$ by interpreting letters of $w$ as drawing
instructions. In the binary case we associate $0$
with a move by the vector $\textbf{v}_0=(1,-1)$, and
$1$ with a move by $\textbf{v}_1=(1,1)$.
We start at the origin $(x_0,y_0)=(0,0)$.
At step $n$, we are at a point $(x_{n-1},y_{n-1})$
and we move by a vector corresponding to the letter $w_{n}$, so
that we come to a point $(x_{n},
y_{n})=(x_{n-1},y_{n-1})+\textbf{v}_{w_n}$,
and the two points $(x_{n-1}, y_{n-1})$ and $(x_{n}, y_{n})$
are connected with a line segment.
We let $G_w$ denote the
corresponding graph, and $G_w(n)=(x_n,y_n)$.
\begin{lemma}
A weakly abelian periodic word $w$ has a graph with infinitely many integer points
on some line with a rational slope.
\end{lemma}
We remark that instead of the vectors $(1,-1)$ and $(1,1)$,
one could use any other pair of noncollinear vectors $\textbf{v}_0$
and $\textbf{v}_1$.
For a $k$-letter alphabet one can consider a similar graph in
$\mathbb{Z}^k$ with linearly independent integer vectors
$\textbf{v}_1, \dots, \textbf{v}_k$.
For example, one can take as $\textbf{v}_i$ a unit vector with $1$ at position $i$
and all other coordinates equal to $0$.
In the binary case and for vectors
$\textbf{v}_0=(1,-1)$ and $\textbf{v}_1=(1,1)$, it will be
convenient for us to use a function $ g_w\colon n\mapsto y_n $, so
that $G_w(n) = (n, g_w(n))$.
It can also be regarded as a piecewise linear function with line
segments connecting integer points (see an example on Figure 1).
\begin{figure}[h]
\begin{center} \unitlength=1pt
\begin{picture}(330,60)
\multiput(10,0)(10,0){32}{\line(0,1){60}}
\multiput(0,10)(0,10){5}{\line(1,0){330}}
\thicklines \put(0,30){\vector(1,0){330}}
\put(10,0){\vector(0,1){60}} \put(10,30){\circle*{5}}
\put(10,30){\line(1,-1){10}} \put(20,20){\line(1,1){20}}
\put(40,40){\line(1,-1){10}} \put(50,30){\line(1,1){10}}
\put(60,40){\line(1,-1){20}} \put(80,20){\line(1,1){20}}
\put(100,40){\line(1,-1){20}} \put(120,20){\line(1,1){10}}
\put(130,30){\line(1,-1){10}} \put(140,20){\line(1,1){20}}
\put(160,40){\line(1,-1){10}} \put(170,30){\line(1,1){10}}
\put(180,40){\line(1,-1){20}} \put(200,20){\line(1,1){10}}
\put(210,30){\line(1,-1){10}} \put(220,20){\line(1,1){20}}
\put(240,40){\line(1,-1){20}} \put(260,20){\line(1,1){20}}
\put(280,40){\line(1,-1){20}} \put(300,20){\line(1,1){10}}
\put(310,30){\line(1,-1){10}}
\end{picture}
\caption{The graph of the Thue-Morse word with $\textbf{v}_0=(1,-1)$,
$\textbf{v}_1=(1,1)$. }\label{table}
\end{center}
\end{figure}
\section{Weak abelian borders: binary alphabet}\label{wab2}
We proceed with the following theorem relating weak abelian
borders and weak abelian periodicity, giving one way analog of the
characterization of periodicity in terms of unbordered factors (Theorem~\ref{Theorem1}):
\begin{theorem} \label{wap}
Let $w$ be an infinite binary word.
If there exists a constant $C$ such that every factor $v$ of $w$ with $|v|\geq C$
is weakly abelian bordered, then $w$ is bounded weakly abelian periodic.
Moreover, its graph lies between two rational lines
and has points on each of these two lines with bounded gaps.
\end{theorem}
Remark that following the terminology from \cite{wap},
a bounded weakly abelian periodic word is necessarily of bounded width.
Namely, a word $w$ is of \emph{bounded width} if there exist two lines with the same slope,
so that the graph of $w$ lies between these two lines.
Formally, there exist numbers $\alpha, \beta, \beta'$
so that $\alpha x+\beta \leq g_w(x)\leq \alpha x+\beta'$.
We note that in the case of weak abelian periodicity $\alpha$ must be rational.
We will also need the notion of balance.
A word $w\in\Sigma^{\omega}$ is called $K$\emph{-balanced}, if for each
letter $a$ and two factors $u, v$ of $w$ such that $|u|=|v|$ the
inequality $||u|_a-|v|_a|\leq K$ holds.
We simply say that w is \emph{balanced} if it is $K$-balanced for some $K$.
We remark that balance implies the existence of letter frequencies:
\begin{proposition} \cite[Proposition~2.4]{Berthe-Delecroix}\label{24}
An infinite word $w \in \Sigma^{\omega}$ is balanced if and
only if it has uniform letter frequencies and there exists a
constant $B$ such that for any factor $u$ of $w$ and for all letters $a\in
\Sigma$, we have $\big| |u|_a - \rho_a (w) |u|\big| \leq B$.
\end{proposition}
In fact, the latter proposition implies that balance is equivalent to
the bounded width property.
Indeed, the bounded width property implies balance.
The converse is obtained by applying the
above proposition to the prefixes of the word.
\begin{corollary}
An infinite word is balanced if and only if it is of bounded width.
\end{corollary}
The proof of Theorem \ref{wap} heavily relies on the graph representation of the word.
It consists of several lemmas restricting the form of the word.
We first prove that if a factor $w[i+1..j]$
of $w$ is such that the graph of the word between
$i$ and $j$ lies above or below the line segment connecting the
points $(i, g_w(i))$ and $(j, g_w(j))$, then $w[i+1..j]$ is weakly abelian unbordered:
\begin{lemma}\label{wab_factor}
Let $w$ be an infinite binary word, and $i$, $j$ be integers, $i<j$.
If $g_w(k) > \frac{g_w(j) - g_w(i) }{j-i}k
+ \frac{g_w(i)j - g_w(j) i}{j-i}$ for each $ i< k< j$,
then the factor $w[i+1..j]$ is weakly abelian unbordered.
\end{lemma}
We also note that a symmetric
assertion holds for the case when the graph lies below this line:
simply invert the inequality in the statement of the lemma.
\begin{figure}[h]
\begin{center}
\unitlength=1pt
\begin{picture}(100,60)
\multiput(10,0)(10,0){9}{\line(0,1){60}}
\multiput(0,10)(0,10){5}{\line(1,0){100}}
\thicklines
\put(20,20){\line(1,1){20}}
\put(40,40){\line(1,-1){10}}
\put(50,30){\line(1,1){10}}
\put(60,40){\line(1,-1){10}}
\thinlines
\put(20,20){\line(5,1){50}}
\end{picture}
\caption{Illustration of Lemma \ref{wab_factor}:
weakly abelian unbordered factor 11010.}
\label{fig_lemma1}
\end{center}
\end{figure}
\begin{proof}
For any proper prefix $p$ and any proper suffix $s$ of $w[i+1..j]$
we have $\rho_1(p)-\rho_0(p)> \frac{g_w(j) - g_w(i)}{j-i}$
and $\rho_1(s)-\rho_0(s)< \frac{g_w(j) - g_w(i)}{j-i}$.
Hence there is no weak abelian border.
\end{proof}
However, the reciprocal of the previous lemma does not hold.
A counterexample is given by the weakly abelian unbordered word $1110000011000$:
see Figure~\ref{counterex}.
\begin{figure}[htbp]
\begin{center}
\unitlength=1pt
\begin{picture}(160,90)
\multiput(10,0)(10,0){15}{\line(0,1){90}}
\multiput(0,10)(0,10){8}{\line(1,0){160}}
\thicklines
\put(20,40){\line(1,1){30}}
\put(50,70){\line(1,-1){50}}
\put(100,20){\line(1,1){20}}
\put(120,40){\line(1,-1){30}}\thinlines
\put(20,40){\line(13,-3){130}}
\end{picture}
\caption{The reciprocal of Lemma \ref{wab_factor} is not true:
weakly abelian unbordered factor $1110000011000$.}
\label{counterex}
\end{center}
\end{figure}
\begin{lemma}\label{bal}
Let $w$ be an infinite binary word with finitely many weakly abelian unbordered factors.
Then $w$ is balanced.
\end{lemma}
We remind that by balance we mean $K$-balance for some $K$.
\begin{proof}
Suppose the converse.
Unbalance implies that for every $K$ there exist $i$, $j$, $l$ with $i<l<j$, such that
\begin{equation}\label{eq_bal}
\left|g_w(l) - \frac{g_w(j) - g_w(i) }{j-i} l
- \frac{g_w(i)j - g_w(j) i}{j-i} \right| \geq K.
\end{equation}
In other words, the point $(l, g_w(l))$ lies at distance at least $K$
from the line connecting the points $i$ and $j$ (distance measured vertically).
Indeed, unbalance implies that there exist two factors $u$ and $v$ of the same length
in which the number of occurrences of the letter $0$ (and hence also the letter $1$) differs by more than $K$.
Consider the factor $w[i+1..j]$ starting with $u$ and ending with $v$.
It is not hard to see that it satisfies the inequality
\eqref{eq_bal} for either $l=i+|u|$ or $l=j-|v|$.
Take the shortest such factor (i.e., choose $i$, $j$ satisfying
the condition \eqref{eq_bal} with the smallest $j-i$).
Then the graph $G_w$ does not intersect the line segment
connecting the points $(i, g_w(i))$ and $(j, g_w(j))$,
for otherwise there exists a shorter factor satisfying \eqref{eq_bal}.
Hence due to Lemma \ref{wab_factor}, it remains to notice that the lengths
of such factors grow as $K$ grows (in fact, $j-i>K$).
\end{proof}
Since balance implies the existence of letter frequencies,
Lemma \ref{bal} implies that a word satisfying the conditions of
Theorem \ref{wap} has letter frequencies. We will now prove
that the frequencies are rational.
\begin{lemma}\label{freq}
Let $w$ be an infinite binary word with finitely many weakly abelian unbordered factors.
Then the letter frequencies are rational.
\end{lemma}
\begin{proof}
Balance implies bounded width, i.e.,
there exist $\alpha,$ $\beta',$ $\beta''$ such that the graph of
$w$ lies between two lines: $ \alpha x+\beta'\leq g_w(x)\leq \alpha x+\beta''$.
Here $\alpha$ is related to the letter frequencies as follows:
$\rho_0=\frac{1-\alpha}{2}$ and
$\rho_1=\frac{1+\alpha}{2}$.
So, we need to prove that $\alpha$ is rational.
Assume the converse.
Take the lowest (in the sense of the smallest $\beta$) line $y=\alpha x + \beta$ such that there
are points of the graph arbitrarily close to it.
We build the line (and prove its existence) as follows.
Take $\varepsilon>0$, divide the stripe $ \alpha x+\beta'\leq
y\leq \alpha x+\beta''$ into stripes of width at most
$\varepsilon$ (width can be measured vertically, although this is not important).
At least one of the stripes contains infinitely
many points $(i, g_w(i))$ of the graph $G_w$.
Choose the lowest such stripe $ \alpha x+\beta'_1\leq y \leq \alpha x+\beta''_1$.
Divide it into stripes of width $\varepsilon_1< \varepsilon$,
continue the process with $\varepsilon_n\to 0$.
The value of $\beta$ in the desired line is obtained as a limit
$\beta=\lim_{n\to\infty}\beta'_n=\lim_{n\to\infty}\beta''_n$.
To prove that $\alpha$ is rational, we consider three cases:
\textbf{Case 1:}
There are no points of the graph $G_w$ below the line $y=\alpha x + \beta$.
In this case take $\varepsilon>0$ and consider the stripe
$\alpha x+\beta\leq y\leq \alpha x+\beta+\varepsilon$.
There are infinitely many points in this
stripe by the definition of the line $y=\alpha x + \beta$.
Consider any two consecutive points $(n_1, g_w(n_1))$ and $(n_2,
g_w(n_2))$ from the stripe (consecutive in the sense that for any
$n$, $n_1<n<n_2$, we have $g_w(n)>\alpha x+\beta+\varepsilon$).
By Lemma \ref{wab_factor}, the factor $w[n_1+1..n_2]$ is weakly abelian
unbordered, since the points of the graph $G_w$ between $n_1$ and
$n_2$ lie above the line $y= \alpha x+\beta+\varepsilon$, which in
turn lies above the line segment connecting the points $(n_1,
g_w(n_1))$ and $(n_2, g_w(n_2))$.
Now we will prove that the minimal possible length $(n_2-n_1)$
grows as $\varepsilon\to 0$. Indeed, take the shortest factor
$w[n_1+1..n_2]$ with $n_1$, $n_2$ consecutive points in the stripe
such that the slope of the line segment connecting the points
$(n_1, g_w(n_1))$ and $(n_2, g_w(n_2))$ is the closest to
$\alpha$, i.e., $\big|\alpha - |\frac{g_w(n_2)-g_w(n_1)}{n_2-n_1}|\big|$
is minimal among the factors of the length $(n_2-n_1)$ with both
ends in the stripe.
If there are several such factors, choose one
of them, say the leftmost.
Since $\alpha$ is irrational, it is not equal to the slope of the line connecting points
$(n_1,g_w(n_1))$ and $(n_2, g_w(n_2))$.
So, take $\varepsilon'<\big||g_w(n_2)-g_w(n_1)|-\alpha(n_2-n_1)\big|$,
then no two points of the graph corresponding to positions in the word at
distance $(n_2-n_1)$ can simultaneously be in the stripe $ \alpha
x+\beta\leq y\leq \alpha x+\beta+\varepsilon'$.
So, the minimal distance between points in the new (thinner) stripe of width
$\varepsilon'$ increased, which corresponds to weak abelian
unbordered factors of length more than $n_2-n_1$.
Continuing this line of reasoning
and taking smaller and smaller $\varepsilon$, we get arbitrary
long weakly abelian unbordered factors.
\textbf{Case 2:} There are finitely many points of the graph
$G_w$ below the line $y=\alpha x + \beta$ (but the number of
such points is nonzero). In this case take the last point $(n_0,
g_w(n_0))$ below the line: $g_w(n_0)\leq\alpha n_0 + \beta$, and
$g_w(n)>\alpha n + \beta$ for each $n>n_0$. Consider a prefix of
length $N>n_0$ of $w$, and the part of the graph between $n_0$ and
$N$. Choose a point $n'$, $n_0<n'\leq N$, such that the point
$(n', g_w(n'))$ is the closest to the line among the points $(n,
g_w(n))$, $n_0<n\leq N$:
$$|g_w(n')-\alpha n' - \beta| = \min_{n_0<n\leq N} |g_w(n)-\alpha n -
\beta|.$$
Then due to Lemma
\ref{wab_factor}, the factor $w[n_0+1..n']$ is weak abelian
unbordered. Indeed, the points of the graph $G_w$ between $n_0$
and $n'$ lie above the line $y= \alpha x+g_w(n')-\alpha n'$ (the
line parallel to $y=\alpha x + \beta$ and passing through the
point $(n', g_w(n'))$), which in turn lies above the line segment
connecting the points $(n_0, g_w(n_0))$ and $(n', g_w(n'))$:
\[
g_w(n)\geq\alpha n+g_w(n')-\alpha n'
> \frac{g_w(n') -g_w(n_0) }{n'-n_0}n
+ \frac{g_w(n_0)n' - g_w(n') n_0}{n'-n_0}
\]
for $n_0<n<n'$. Since there are points of the
graph arbitrary close to the line $y=\alpha x + \beta$, then,
increasing $N$, we will find larger and larger values $n'$ and
hence find arbitrary long weakly abelian unbordered factors.
\textbf{Case 3:} There are infinitely many points of the graph
$G_w$ below the line $y=\alpha x + \beta$.
Since we chose the lowest line with the property that there are infinitely many
points arbitrary close to the line, it follows that for any
$\varepsilon>0$ there are only finitely many points below the line
$y=\alpha x + \beta-\varepsilon$ (if any).
Take $\varepsilon$ such that there are points below the line $y=\alpha x +
\beta-\varepsilon$.
By the condition of Case 3 there are infinitely many points
in the stripe $\alpha x + \beta -\varepsilon \leq y \leq \alpha x + \beta$.
For each two consecutive points in the stripe the corresponding factor is weakly
abelian unbordered due to Lemma \ref{wab_factor}.
As $\varepsilon\to 0$, the lengths of such factors grow for
irrational $\alpha$ (the proof of this is similar to the proof of
the similar fact from Case 1).
Therefore, we came to a contradiction in each case. So the
frequency must be rational.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{wap}.]
Lemmas \ref{bal} and \ref{freq} imply that the graph of the word lies between two lines
of rational slope $\alpha$: $\alpha n + \beta'\leq
g_w(n)\leq \alpha n + \beta''$ for each $n$.
We can choose such lines
with the minimal $\beta''$ and maximal $\beta'$.
To prove the theorem it remains to show that the graph of the word has points
on each of these two lines with bounded gaps.
We provide a proof for the line $y=\alpha x + \beta'$, for the other line the proof is symmetric.
First we prove that there are points of the graph on the line.
Clearly, for rational $\alpha$ there is a fixed minimal distance
from the integer points of the lattice $\mathbb{Z}^2$ to the line
(note that for irrational slope this does not hold). If there are
no points of the graph on the line, since the slope is rational,
instead of $y=\alpha x + \beta'$ we could choose a different line
$y=\alpha x + \beta'''$ with the same slope passing through the
point of the graph closest to the line, and with
$\beta'''>\beta'$, which contradicts the choice of $\beta'$ as the
maximal possible.
Secondly, there are infinitely many points on the line.
Otherwise
consider a line $y=\alpha x + \beta'''$ with $\beta'''>\beta'$, with
infinitely many points of the graph on the line and with the
smallest such $\beta'''$.
Note that here we need again $\alpha$ to be rational.
Then there exists $n_0$ such that for all $n>n_0$ we
have $g_w(n)\geq \alpha n + \beta'''$, i.e., starting from $n_0$
the graph lies above the line.
Taking the smallest possible $n_0$
satisfying the condition, we get that the point $(n_0, g_w(n_0))$
lies below the line.
Hence for each $n'>n_0$ corresponding to a
point on the line, i.e., with $g_w(n')=\alpha n' + \beta'''$, we
have that the factor $w[n_0+1..n']$ is weakly abelian unbordered by
Lemma \ref{wab_factor}.
Indeed, all the points between $n_0$ and
$n''$ lie above the line $y=\alpha x + \beta'''$ and hence above
the line segment connecting the points $(n_0,g_w(n_0))$ and
$(n',g_w(n'))$ of the graph corresponding to the beginning and the
end of the factor.
Finally, we need to prove that the points on the line $y=\alpha x
+ \beta'$ appear in bounded gaps.
This follows from the fact that for two consecutive points $n_1,$ $n_2$ on the line $y=\alpha x +
\beta'$ the factor $w[n_1+1..n_2]$ is weakly abelian unbordered by
Lemma \ref{wab_factor}.
\end{proof}
\begin{remark} Except for the case $\rho_0=\rho_1=1/2$
(equal frequencies), we do not know whether the converse of
Theorem 2 is true (see Question~\ref{question2}).
In the case of equal
frequencies the converse holds, and it could be easily seen from
the graph of the word.
In fact, we will prove that all
sufficiently long factors have weak abelian borders with
frequencies $1/2$.
Frequencies equal to $1/2$ correspond to
$\alpha=0$, so the graph lies between two horizontal lines
$y=\beta'$ and $y=\beta''$, and intersects each of them with
bounded gaps.
Denote the maximal gap by $C$.
Consider a factor $w[i+1..j]$ of length $>C$: $j-i> C$.
Then there exist
$n',n''$, $i<n',n''<j$, such that $g_w(n')=\beta'$,
$g_w(n'')=\beta''$. Without loss of generality $n'<n''$. For each
integer $\beta$, $\beta'\leq\beta\leq\beta''$, there exists $n$,
$n'\leq n \leq n''$, such that $g_w (n)=\beta$.
In other words, in
the interval $[n', n'']$ the graph passes through any integer
horizontal line between $\beta'$ and $\beta''$. In particular,
there exist $i'$, $j'$, $i<n'\leq i',j'\leq n''<j$, such that
$g_w(i)=g_w(i')$, $g_w(j)=g_w(j')$.
So, the prefix $w[i+1..i']$ and the suffix $w[j'+1..j]$
give a weak abelian border with the frequencies $1/2$.
But we do not see how to generalize this
observation to the case of nonequal frequencies, since in general
the graph can ``jump'' through several rational non-horizontal
lines and it does not have to pass through all rational lines
between points on lower and upper lines.
\end{remark}
\section{Weak abelian borders: nonbinary alphabet}\label{wabk}
Now we continue with non-binary alphabets and prove the analogue of
Theorem \ref{wap}.
Basically, a non-binary word with finitely many
weakly abelian unbordered factors must also be balanced, weakly
abelian periodic and have rational letter frequencies.
To state the analogue of the additional condition in Theorem \ref{wap}
(graph lying between two parallel rational lines and bounded weak
abelian periodicity along those lines) we need some notation.
First, in the non-binary case, instead of bounded width we define cylinder set.
Let $\emph{\textbf{l}}$ denote a line in $\mathbb{R}^k$,
$\emph{\textbf{l}}\colon x_i=x_i^0+a_i t$, $i=1,
\dots, k$, and let $M$ be a constant.
A $k$-dimensional \emph{cylinder} $C (\emph{\textbf{l}},M)$
is defined as
\[
C (\emph{\textbf{l}},M)
= \{ \bar{x} \in \mathbb{R}^k | d(\emph{\textbf{l}}, \bar{x})\leq M\}
\]
where $d(\emph{\textbf{l}}, \bar{x})$ denotes the distance between the
point $\bar{x}$ and the line $\emph{\textbf{l}}$.
The line $\emph{\textbf{l}}$ is the \emph{axis} of the cylinder.
Now we will give an analogue of the two parallel lines such that the
graph lies between those two lines and has points on each of those
lines with bounded gaps.
If the graph $G_w$ of an infinite word
$w$ over a $k$-letter alphabet belongs to some cylinder $C
(\emph{\textbf{l}},M)$, we define a \emph{tangential hyperplane}
to the graph of $G_w$ as a hyperplane $H\colon b_1 x_1 + \dots + b_k x_k = d$ parallel to
the line $\emph{\textbf{l}}$ such that for each point
$\bar{x}=(x_1, \dots, x_k)\in G_w$ one has $b_1 x_1 + \dots + b_k
x_k \geq d$, and in addition there is a point $\bar{x}'=(x_1',\ldots,x_n')\in G_w$
such that $b_1 x'_1 + \dots + b_k x'_k = d$.
We remark that a tangential hyperplane to the graph $G_w$ does not have to be a
``proper'' tangential hyperplane to the cylinder $C
(\emph{\textbf{l}},M)$, but it is a hyperplane parallel to it.
In the remaining part of the text, by ``tangential'' we always mean tangential to
the graph. Next, we define a \emph{tangential line}
$\emph{\textbf{l}}'$ to $G_w$, $\emph{\textbf{l}}'\colon x_i=x_i'^0+a_i t$, $i=1, \dots, k$, as a line which belongs to
some tangential hyperplane $H$ to $G_w$, $H\colon b_1 x_1 + \dots +
b_k x_k = d$, such that for each point $\bar{x}\in G_w \cap H$ one
has also $\bar{x}\in \emph{\textbf{l}}'$.
In other words, a tangential line $\emph{\textbf{l}}'$ is defined as a line parallel
to the cylinder lying in some tangential hyperplane, such that the
only points of the graph belonging to the hyperplane also belong to the line.
We remark that this line can be inside the cylinder.
Now we are ready to state the non-binary analogue of Theorem
\ref{wap}:
\begin{theorem} \label{wap1}
Let $w$ be an infinite word over a $k$-letter alphabet.
If there exists a constant $C$ such that every factor $v$ of $w$ with $|v|\geq C$
is weakly abelian bordered, then $w$ is bounded weakly abelian periodic.
Moreover, its graph $G_w$ belongs to a $k$-dimensional cylinder with axis with rational coefficients,
and each tangential line to $G_w$ has points of $G_w$ on it with bounded gaps.
\end{theorem}
We emphasize that in fact Theorem \ref{wap} is a particular case of Theorem \ref{wap1}.
We state and prove it separately for several reasons.
First, for clarity reasons: the statement of Theorem \ref{wap} is easier and more intuitive.
Second, key lemmas in non-binary case are simple consequences of those in the binary case.
On the other hand, the proof of the body of the theorem is more involved,
so there is not much repetition in the proofs.
First we prove that Lemmas \ref{bal} and \ref{freq} also hold for
the non-binary case:
\begin{lemma}\label{bal1}
If an infinite word
has finitely many weakly abelian unbordered factors,
then it is balanced.
\end{lemma}
\begin{proof}
Assume that $w$ is an infinite word
with finitely many weakly abelian unbordered factors.
For $a\in\alph(w)$, consider a morphism $h_a\colon a\mapsto a,\
b\mapsto c$ for all $b\neq a$ where $c$ is a letter not belonging to $\alph(w)$.
We call such a morphism a \emph{projection}.
Clearly, if two words $u$ and $v$ have the same letter frequencies,
then so do $h_a(u)$
and $h_a(v)$, for each $a\in \alph(w)$.
Hence if $u$ is weakly abelian bordered, then so is $h_a(u)$.
Then $h_a(w)$ has finitely many weakly abelian unbordered factors.
Now, since a projection of $w$ is a binary word (over the
alphabet $\{a,c\}$), we can apply Lemma \ref{bal}, hence $h_a(w)$
is also balanced:
There exists a constant $K_a$, such that for
each two factors $u, v$ of $h_a(w)$ such that $|u|=|v|$ the
inequality $||u|_a-|v|_a|\leq K_a$ holds.
Taking $K=\max_{a\in \Sigma} K_a$, we get that $w$ itself is $K$-balanced by definition.
\end{proof}
\begin{lemma}\label{freq1}
If an infinite word has finitely many weakly abelian unbordered factors,
then its letter frequencies are rational.
\end{lemma}
\begin{proof}
Clearly, the frequency of any letter $a$ exists in an
infinite word $x$ if and only if the frequency of the letter $a$
exists in its projection $h_a(x)$, and these frequencies are equal.
Since the projection is a binary word, Lemma \ref{freq}
implies that $w$ has letter frequencies and they are rational.
\end{proof}
We let $(\bar{b}, \bar{x})$ denote the scalar product of vectors
$\bar{b}$ and $\bar{x}\colon (\bar{b}, \bar{x}) = b_1 x_1 + \dots + b_k x_k $.
We make use of the following analogue of Lemma~\ref{wab_factor}.
\begin{lemma}\label{wab_factor1}
Let $w$ be an infinite word, and $i$, $j$ be integers, $i<j$.
If there exists a hyperplane $H\colon b_1 x_1 + \dots + b_k x_k = d$ such that
$G_w(i), G_w(j) \in H$ and $(\bar{b},
G_w(l))>d$ for each $ i< l< j$ (or $(\bar{b}, G_w(l))<d$ for each $ i< l< j$),
then the factor $w[i+1..j]$ is weakly abelian unbordered.
\end{lemma}
\begin{proof}
Two factors $w[i_1+1..j_1]$ and $w[i_2+1..j_2]$
share the same letter frequencies if the
vectors $G_w(j_1)-G_w(i_1)$ and $G_w(j_2)-G_w(i_2)$ connecting
their ends are collinear.
Moreover, they differ by a positive multiple.
For any proper prefix $w[i+1..p]$ and any proper suffix
$w[s+1..j]$ of $w[i+1..j]$
we have $(G_w(p)-G_w(i), \bar{b}) >0$ and
$(G_w(j)-G_w(s), \bar{b})<0$, which implies that $w[i+1..p]$
and $w[s+1..j]$ do not have the same letter frequencies.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{wap1}.]
Like in the binary case, balance
implies that
$G_w$ belongs to some $k$-dimensional cylinder $C (\emph{\textbf{l}},M) $.
The coefficients $a_i$ of the axis $\emph{\textbf{l}}$ of the cylinder
are related to the letter frequencies and can be defined via
$\{{\bf{v}}_1, \dots, {\bf{v}}_{k}\}$.
Since frequencies are rational and all ${\bf{v}}_i$ have integer entries, the
coefficients $a_i$ are also rational.
In fact, the graph of the word lies on finitely many lines with rational coefficients
parallel to the cylinder axis.
Clearly, in dimension two this is equivalent to bounded width.
First notice that such tangential lines always exist.
We can choose a cylinder with the minimal $M$.
Similarly to the binary
case and using the fact that the slope vector $\bar{a}$ of the
cylinder is rational it is not hard to show that there exists a
point $\bar{x}' \in G_w$ with $d(\emph{\textbf{l}}, \bar{x}') = M.$
Consider the line $\emph{\textbf{l}}'$ parallel to
$\emph{\textbf{l}}$ and passing through one of such points,
$\emph{\textbf{l}}'\colon \bar{x} = \bar{x}' + \bar{a} t $.
Taking the tangential plane $H$ to the cylinder passing through $\bar{x}'$,
we see that $\emph{\textbf{l}}'$ is a tangential line to the graph
$G_w$. We remark that this tangential plane $H$ is both a
tangential plane to the graph $G_w$ and a ``proper'' tangential
plane to the cylinder $C (\emph{\textbf{l}},M) $.
Although there can be other tangential lines to the graph $G_w$ lying inside the
cylinder, so that the corresponding tangential planes to $G_w$ are
not tangential to the cylinder.
To prove the theorem it remains to show that for a tangential line
$\emph{\textbf{l}}''\colon \bar{x} = \bar{x}'' + \bar{a} t $ the points
of the graph lie on this line with bounded gaps.
Like in the binary case, we first show that there are infinitely
many points of the graph on the line $\emph{\textbf{l}}''$.
Suppose that there are only finitely many of those. Take the
corresponding tangential hyperplane $H''\colon b_1 x_1 + \dots + b_k
x_k = d''$ and consider the hyperplane $H'\colon b_1 x_1 + \dots + b_k
x_k = d'$ parallel to $H''$ such that it has infinitely many
points of $G_w$ on it (since $G_w$ lies on finitely many lines
parallel to $\emph{\textbf{l}}''$ due to rational coefficients, it
also lies on finitely many hyperplanes parallel to $H''$). So,
there are finitely many points of $G_w$ such that $b_1 x_1 + \dots
+ b_k x_k < d'$, and infinitely many with $b_1 x_1 + \dots + b_k x_k = d'$.
Now take the point $\bar{x}^{(n_0)}=G_w(n_0)$ on the line with
the largest $n_0$ such that $b_1 x^{(n_0)}_1 + \dots + b_k x^{(n_0)}_k
< d'$. Then for each $n$ with $b_1 x^{(n)}_1 + \dots + b_k x^{(n)}_k = d'$
for $\bar{x}^{(n)}=G_w(n)$ the factor $w[n_0+1..n]$ is weakly abelian
unbordered. Indeed, arguing as in Lemma \ref{wab_factor1}, for any
proper prefix $w[i+1..p]$ and any proper suffix $w[s+1..j]$ of
$w[i+1..j]$ we have $(G_w(p)-G_w(i), \bar{b}) >0$ and
$(G_w(j)-G_w(s), \bar{b})<0$, which implies that $w[i+1..p]$ cannot
be weakly abelian equivalent to $w[s+1..j]$.
So, we proved that there are infinitely many points of the graph on a tangential line.
Finally, we need to prove that the points on the tangential line
$\emph{\textbf{l}}''$ appear in bounded gaps.
It follows from the fact that for two consecutive points $n_1,$ $n_2$ on
$\emph{\textbf{l}}''$ the factor $w[n_1+1..n_2]$ is weakly abelian
unbordered by Lemma \ref{wab_factor1}.
\end{proof}
\section{Abelian borders}\label{ab}
In this section we consider the connections between abelian
periodicity and the condition of having finitely many abelian unbordered factors.
First we show that finitely many abelian unbordered
factors is not a sufficient condition for periodicity:
\begin{proposition}\label{prop_abborders}
There exists an infinite aperiodic word $w$ and constants $C, D$
such that every factor $v$ of $w$ with $|v|\geq C$ has an abelian
border of length at most $D$.
\end{proposition}
\begin{proof}
Any aperiodic word $w \in \{ 0101 00110011, 0101 0011 0011 0011 \}^\omega$
satisfies the condition with $C=15$ and $D=14$.
The proof is a fairly straightforward and not too technical case study:
If $u$ is a sufficiently long abelian unbordered factor, then it
must begin or end with either $0100$ or with $1101$ because the other factors of
length $4$ have equal frequencies.
So there are only these two cases to consider: the factor
starts with $0100$ (the case of ending with $1101$ is symmetric) or
with $1101$ (the case of ending with $0100$ is symmetric).
Consider first the case of a factor beginning with $0100$.
If it ends with $0$, we have an abelian border of length $1$;
if it ends with $01$, we have an abelian border of length $2$.
So we need to analyze the remaining case of the form $0100 \cdots 0011$.
Considering extension of the prefix $0100$ to the right, one has
after $01$ abelian period $4$ with equal frequencies.
Considering extension of the suffix $0011$ to the left, one has abelian period
$4$ with equal frequencies.
So, extending to the left the suffix, one will get an abelian border as soon as there is a block $0101$.
Thus, there exists an abelian border of length $4k+2$, where $k$ is at most 3.
Now consider the case of a factor beginning with $1101$.
It continues uniquely to the right with $01001100$, so that $110101001100$ is the prefix of the factor.
If it ends with $1$, we have an abelian border of length $1$,
so we need to consider the cases of ending with $00$ and $10$.
If it ends with $00$, then it has either a suffix $0100$, or
$1100$. The suffix $0100$ extends uniquely to the left to
$11010100$, giving an abelian border of length $8$. The
suffix $1100$ extends to either $11001100$ giving an abelian border of length
$8$, or to $110101001100$, giving an abelian border of length $12$.
If it ends with $10$, then it has a suffix $010$ (otherwise there is an abelian border of length $3$), which
extends uniquely to $1010$. This suffix in turn extends to
$01010$ (otherwise there is an abelian border of length $5$),
which extends to $1101010$, giving an abelian border of length
$7$. This concludes the proof.
\end{proof}
We observe that all infinite word $w \in \{ 0101 00110011, 0101 0011 0011 0011 \}^\omega$
occurrring in the proof of Proposition \ref{prop_abborders}
are abelian periodic with period $2$.
In general, we do not know whether an infinite word containing
only finitely many abelian unbordered factors is necessarily abelian periodic.
However, we begin by showing that abelian periodicity alone
does not in general imply only finitely many abelian unbordered factors.
In other words, at least one direction of Theorem~1 fails in the abelian setting.
\begin{proposition}\label{pal}
Let $x$ be a uniformly recurrent aperiodic word containing infinitely many distinct palindromes.
Then $x$ admits an infinite number of abelian unbordered factors.
\end{proposition}
\begin{proof}
Let $N$ be any positive integer.
We claim that $x$ admits a factor of the form $aUb$ where $a$ and $b$ are distinct letters
and $U$ a palindrome with $|U|\geq N$.
In fact, since $x$ contains infinitely many palindromes,
there exists a palindromic factor $u$ of $x$ with $|u|\in \{N,N+1\}$.
By uniform recurrence there exists a constant $M$ such that every factor of $x$
of length $M$ contains an occurrence of $u$.
Since $x$ is aperiodic, there exists a word $v$ with $|v|=3M$ and distinct letters $c$ and $d$
such that both $cv$ and $dv$ are factors of $x$.
In the usual terminology, $v$ is a left special factor of $x$.
Let $\eta$ denote the center of the first occurrence of $u$ in the prefix of $v$ of length $M$.
So $\eta$ is either an integer or a half integer depending on the parity of $|u|$.
Let $U$ be the longest palindromic factor of $v$ centered at $\eta$.
It follows that $2M\geq|U|\geq |u| \geq N$.
If $U$ is a not a prefix of $v$, then there exist distinct letters $a$ and $b$
such that the word $aUb$ is a factor of $v$ centered at $\eta$.
On the other hand, if $U$ is a prefix of $v,$ then let $b$ be a letter such that $Ub$ is a prefix of $v$.
Since both $cv$ and $dv$ are factors of $x$ we can find a letter $a$ different from $b$ such that $aUb$ is a factor of $x$.
Having established our claim, the result of the proposition now follows
immediately since any word of the form $aUb$ with $a\neq b$ and $U$ a palindrome is clearly abelian unbordered.
\end{proof}
As an immediate corollary we have:
\begin{corollary}\label{cor2}
There exist abelian periodic words having infinitely many abelian unbordered factors.
\end{corollary}
\begin{proof}
Let $x\in\{0,1\}^\omega$ be any binary uniformly recurrent aperiodic word
containing infinitely many palindromes.
For instance $x$ can be taken to be the Thue-Morse infinite word or any Sturmian word.
Let $\tau$ denote the morphism $0\mapsto 0110$, $ 1\mapsto 1001$.
Then $\tau(x)$ is abelian periodic (with abelian period $2$).
Moreover, it is readily verified that $\tau(x)$ is uniformly recurrent,
aperiodic, and contains infinitely many distinct palindromes.
By Proposition~\ref{pal}, it follows that $\tau(x)$
admits infinitely many abelian unbordered factors.
\end{proof}
As mentioned earlier, we do not know whether the other direction
of Theorem~\ref{Theorem1}
holds in the abelian context,
that is whether a word having only finitely many abelian unbordered factors
is necessarily abelian periodic.
However we are able to establish the abelian analogue of the following
weaker version of Theorem~\ref{Theorem1}:
\begin{theorem}\label{comp}
Let $x$ be an infinite word having only finitely many unbordered factors.
Then there exists a constant $N$ such that $x$ contains at most $N$ factors
of each given length $n\geq 1$.
In other words, $x$ has bounded factor complexity.
\end{theorem}
The reason this is a weaker version of Theorem~\ref{Theorem1}
is because bounded complexity is equivalent to ultimate periodicity.
For this reason the converse of Theorem~\ref{comp} does not hold in general.
For instance $x=01111\cdots$ has bounded complexity,
yet contains infinitely many unbordered prefixes.
We now give several new proofs of Theorem~\ref{Theorem1}
which may be of independent interest.
One direction is of course trivial:
namely, if x is purely periodic then all sufficiently long factors of x are bordered.
For the other direction, as in the original proof by Ehrenfeucht and Silberger,
we break the proof into two parts;
in a first part we show that if x is any infinite word with the property
that all sufficiently long factors are bordered, then x is ultimately periodic.
We give three proofs of this first part which to our knowledge are all new.
However, each of the following proofs relies on one
or more key features which do not extend to the abelian context.
See the remark following the proofs.
In a second part we show that x must be purely periodic.
Also our proof of the second part is new and greatly simplifies
the original proof by Ehrenfeucht and Silberger.
\begin{proof}[Proof of Theorem~\ref{Theorem1}]
Let $\mathcal{U}(x)$ denote the set of unbordered factors of $x$.
Clearly if $x$ is purely periodic then $\mathcal{U}(x)$ is finite.
Conversely, assume $\mathcal{U}(x)$ is finite.
We first show that $x$ is ultimately periodic.
A first quantitative proof of this was given by K. Saari in which
he shows that for any aperiodic word $x$ and any positive integer $n$,
the number $\mathcal{L}_x(n)$ of Lyndon words of length $\leq n$ occurring in $x$
is greater or equal to $\mathcal{L}_{{\mathbf f}}(n)$ where $\mathbf f$ is the Fibonacci word.
See Theorem~3 and Remark~3 in \cite{Saari}.
A second proof, also using Lyndon words proceeds as follows:
If $x$ is aperiodic, then by Proposition~4.8 in \cite{SS},
$x$ contains an infinite Lyndon word $x'$ in its shift orbit closure,
i.e., there exists an ordering of the alphabet of $x$ relative to which
$x'$ is lexicographically smaller than all its proper suffixes.
Since $x'$ begins in infinitely many unbordered prefixes,
it follows that $x$ contains infinitely many unbordered factors.
A third proof uses the so-called Duval conjecture,
first proved by the second author together with D. Nowotka in \cite{HN},
which states that if $v$ is a word of length $2n$
such that the prefix $u$ of $v$ of length $n$ is unbordered
and every factor of $v$ of length greater than $n$ is bordered, then $v=uu$.
Assume $\mathcal{U}(x)$ is finite and let $u\in \mathcal{U}(x)$ be of maximal length.
Then by the Duval conjecture, every occurrence of $u$ in $x$ is an occurrence of $u^2$.
Hence $x$ is ultimately periodic with period $u$.
A fourth proof, perhaps the simplest in that it is completely self contained,
proceeds as follows: Assume $\mathcal{U}(x)$ is finite.
By replacing $x$ if necessary by some suffix of $x$,
we can assume that each $u\in \mathcal{U}(x)$ is recurrent in $x$,
i.e., occurs in $x$ an infinite number of times.
Let $u\in \mathcal{U}(x)$ be of maximal length.
Let $v$ be any first return to $u$ in $x$, i.e.,
$vu$ is a factor of $x$ which begins and ends in $u$ and $|vu|_u=2$.
Then since $u$ is unbordered, $|v|\geq|u|$ i.e., $u$ is a prefix of $v$.
If $|v|>|u|$, then $v$ is bordered (by maximality of $|u|)$.
Let $v'$ denote the shortest border of $v$.
Then $v'\in \mathcal{U}(x)$ from which it follows that $|v'|\leq|u|$
and hence $v'$ is a proper prefix of $u$ (since $u$ is not a suffix of $v)$.
But then the factor $v'u$ of $x$ is an unbordered factor of $x$
of length greater than $|u|$, a contradiction.
Hence, if $v$ is a first return to $u$, then $|v|=|u|$, i.e., $v=u$.
It follows that $x$ is ultimately periodic with period $u$.
We next show that if $x=x_1x_2x_3\cdots $ is ultimately periodic
and all sufficiently long factors of $x$ are bordered, then $x$ is purely periodic.
Let $x'$ denote the left infinite word $\cdots x_3x_2x_1$ obtained by reflecting $x$.
By assumption there exists a primitive word $u$
such that $\prescript{\omega\mkern -2mu}{}u$
is a prefix of $x'$
(where $\prescript{\omega\mkern -2mu}{}u$
denotes the left infinite word $\cdots uuu$).
Clearly if $u=a$ for some letter $a$, then $x'=\prescript{\omega\mkern -2mu}{}a$.
Thus we can assume that $u$ contains at least two distinct letters.
Let $<$ be a linear order of $\alph(x)$ and let $<'$ denote the opposite order,
i.e., $a<'b$ if and only if $b<a$.
Let $M$ ($m$, respectively) denote the Lyndon conjugate of $u$
relative to $<$ ($<'$, respectively).
Without loss of generality we can write
$x'=\prescript{\omega\mkern -4mu}{}Mz=\prescript{\omega\mkern -2mu}{}myz$
for some suffix $z$ of $x'$ and some suffix $y$ of $M$.
We will show that $z$ is a prefix of $M^\omega$
and hence that $x'$ is purely periodic.
Let $M_z$ denote the prefix of $M^{\omega}$ of length $|z|$
and $m_{yz}$ the prefix of $m^{\omega}$ of length $|yz|$.
Since $M$ is unbordered and all sufficiently long factors of $x'$ are bordered,
it follows that $z$ is a product of prefixes of $M$ and hence that $z\leq M_z$.
Similarly, $yz\leq' m_{yz}$ or equivalently $yz\geq m_{yz}$.
Since $m^\omega =yM^\omega$, we deduce that $z\geq M_z$
and hence $z=M_z$ as required.
\end{proof}
\begin{remark}
We note that none of the above proofs readily extend to the abelian context.
In fact, we know of no suitable analogue of Lyndon words in the abelian setting.
Also, the Duval conjecture fails miserably in the abelian setting.
More precisely, there exists an infinite word $x$ with the following properties:
i) $x$ begins in an unbordered factor $u,$
ii) all factors of $x$ of length greater than $|u|$ are abelian bordered,
iii) the prefix of $x$ of length $2|u|$ is not an abelian square.
To see this, let $y$ be any infinite word in $\{ 0101 00110011, 0101 0011 0011 0011 \}^\omega$
beginning in $(0101 0011 0011 0011 )^2$.
Let $x$ denote the second shift of $y,$ i.e., $x=(01)^{-1}y$.
Then $x$ begins in the unbordered factor $u=01001100110011$ of length $14$.
As in the proof of Proposition~\ref{prop_abborders},
every factor of $x$ of length greater than $14$ is abelian bordered.
Yet it is easily checked that the prefix of length $28$ of $x$ is not an abelian square, i.e.,
not of the form $uu'$ with $u'$ abelian equivalent to $u$.
Finally, the fourth proof uses the following fact which is no longer true in the abelian setting,
namely that if $v$ is a proper prefix of an unbordered word $u,$ then $vu$ is unbordered.
For instance, we can take the prefix $v=01$ of the abelian unbordered word $u=010011$.
Yet $vu=01010011$ is abelian bordered.
\end{remark}
We end by establishing the following abelian analogue of Theorem~\ref{comp}:
\begin{theorem}Let $x$ be an infinite word having only finitely many abelian unbordered factors.
Then there exists a constant $N$ such that $x$ contains
at most $N$ abelian equivalence classes of factors of each given length $n\geq 1$.
In other words, $x$ has bounded abelian factor complexity.
\end{theorem}
\begin{proof} Since the condition that $x$ has only finitely many
abelian unbordered factors implies that $x$ has only finitely many
weakly abelian unbordered factors, it follows that all
conclusions we derived earlier in the weak abelian setting still hold.
In particular $x$ is balanced. Thus $x$ has bounded abelian complexity (see Lemma~3 in \cite{RSZ}).
\end{proof}
\section{A non-abelian periodic word with bounded abelian square at each position}\label{5}
In \cite{akp}, the following open question was proposed: Let $w$
be an infinite word and $C$ be an integer such that each position
in $w$ is a centre of an abelian square of length at most $C$. Is
$w$ abelian periodic? We answer this question negatively by
providing an example (actually, a family of examples).
Consider a family of infinite words of the following form:
$$( 000101010111000111000 (111000)^* 111010101 )^{\omega}$$
A straightforward case study shows that words of this form have an
abelian square of length at most $12$ at each position. It is not
hard to see that this family contains abelian aperiodic words.
\section{Conclusions and open questions}
We conclude with the summary of our results and propose two open problems.
\begin{question}\label{question1}
Let $w$ be an infinite word and $C$ a constant such that every
factor $v$ of $w$ with $|v|\geq C$ is abelian bordered. Does it
follow that $w$ is abelian periodic?
\end{question}
Recall that we showed that there exist examples of words
satisfying these conditions which are not periodic (Proposition
\ref{prop_abborders}), but all examples we have are abelian
periodic.
\medskip
The next question asks whether the converse of Theorem \ref{wap1}
(and in particular Theorem \ref{wap}) is true:
\begin{question} \label{question2}
Let $w$ be an infinite bounded weakly abelian periodic word over a $k$-letter alphabet
such that its graph $G_w$ belongs to a $k$-dimensional cylinder with axis with rational coefficients,
and each tangential line to $G_w$ has points of $G_w$ on it with bounded
gaps. Does it follow that $w$ has only finitely many weakly
abelian unbordered factors?
\end{question}
We remark that in the case of binary alphabets and letter frequencies equal to $1/2$
the answer to this question is positive
(see Remark after Theorem \ref{wap}).
Our main results and open questions are summarized in Figure 1.
\small
\begin{figure}[h]
\begin{center}
\unitlength=3.7pt
\begin{picture}(50, 50)(0,-5)
\gasset{Nw=62,Nh=5,Nmr=2.5,curvedepth=0}
\thinlines
\put(5,0){ \node(A1)(-10,40){finitely many unbordered factors}
\gasset{Nw=45,Nh=5,Nmr=2.5,curvedepth=0}
\node(A2)(60,40){periodicity}
\drawedge(A1,A2){ Th.\ref{Theorem1}}
\drawedge(A2,A1){}
\gasset{Nw=62,Nh=5,Nmr=2.5,curvedepth=0}
\node(A3)(-10,20){finitely many abelian unbordered factors}
\gasset{Nw=45,Nh=5,Nmr=2.5,curvedepth=4}
\node(A4)(60,20){abelian periodicity $+$ Condition 1}
\drawedge(A3,A4){? Q.\ref{question1}}
\drawedge(A4,A3){ Cor.\ref{cor2}}
\drawline[AHnb=0](29,14)(33,18)
\drawline[AHnb=0](33,14)(29,18)
\gasset{Nw=62,Nh=5,Nmr=2.5,curvedepth=0}
\drawedge(A3,A2){Pr.\ref{prop_abborders}}
\drawline[AHnb=0](28,29.5)(32,33.5)
\drawline[AHnb=0](32,29.5)(28,33.5)
\drawedge(A1,A3){}
\drawedge(A2,A4){}
\node(A5)(-10,0){finitely many weakly abelian unbordered factors}
\gasset{Nw=45,Nh=5,Nmr=2.5,curvedepth=0}
\node(A6)(60,0){WAP $+$ Condition 1}
\gasset{Nw=45,Nh=5,Nmr=2.5,curvedepth=4}
\drawedge(A5,A6){ Th.\ref{wap1}}
\drawedge(A6,A5){? Q.\ref{question2}}
\gasset{Nw=45,Nh=5,Nmr=2.5,curvedepth=0}
\drawedge(A3,A5){}
\drawedge(A4,A6){}
}
\end{picture}
\end{center}
\caption{Results and open questions.
WAP means ``weakly abelian periodic".
Condition 1: The graph $G_w$ of the word belongs
to a $k$-dimensional cylinder with axis with rational coefficients,
and each tangential line to $G_w$ has points of $G_w$ on it with bounded
gaps.}
\label{table}
\end{figure}
\normalsize
|
\section{INTRODUCTION}
Two decades after the first confirmed discovery of
planets outside the solar system \citep{wolszczan94,mayor95},
we know more than 1500 confirmed planets
of different mass, size and surface temperature that are
orbiting around stars of different spectral types.
Many of the gas giant planets discovered are
orbiting so close to their parent stars that tidal effects of
the star and atmospheric erosion due to strong stellar
irradiation play dominant role in determining their physical
properties and evolution \citep{lammer03,baraffe04,
hubbard07,erkaev07,penz08,sanz11}. On the other
hand, many small and possibly rocky planets are recently
discovered that may have surface temperature similar to that
of the Earth. Therefore, the focus has rapidly changed into
detecting planets that may have favorable environment to
harbor life.
Classically, a habitable zone is defined as the one that has
favorable ambient temperature to keep water in liquid state
\citep{huang59,
hart78, kasting93, selsis07}. In the recent years
a good number of planets are detected within the habitable zone
around stars of various spectral types \citep{udry07,vogt10,pepe11,borucki11,
borucki12,bonfils13}.
It is known that about 75\% of the stars in the
extended solar neighborhood are M dwarfs \citep{reid02}.
Analysis of data obtained by using the transit method indicates
the occurence rate of small and potentially rocky planets with radius ranging
from 0.5 to 2.0 $R_{\oplus}$ ($R_{\oplus}$ is the radius of the
Earth) around M dwarf stars is about 0.51 per star
\citep{dressing13, koppa13} while that by using
the radial velocity method is 0.41 per star \citep{bonfils13}.
It is believed that a good fraction of these rocky planets are situated
within the habitable zone of their star. Therefore, there may exist a large
number of habitable planets in the solar neighborhood and
in our galaxy.
However, a habitable planet should have appropriate
amount of hydrogen, nitrogen, oxygen gas
and water molecules that support a habitable environment.
Various thermal and non-thermal mechanisms cause hydrogen to
escape the atmosphere of a terrestrial planet. Thermal mechanisms
include hydrodynamic escape and Jeans escape. The loss of
hydrogen from a planetary atmosphere is limited either at the
homopause by diffusion or at the exobase by energy. Diffusion
causes substantial hydrogen loss during the early evolutionary
period of a terrestrial planet.
In today's Earth, the atmosphere is collisionless above the
tropopause and hence the barometric laws break down.
Therefore, Jeans law is not applicable.
Jean's law is also not applicable if the atmosphere is not
under hydrostatic equilibrium. This may occur by the
absorption of stellar Extreme Ultraviolet (EUV) radiation
with wavelengths ranging from 100 to 920 $\AA $. Strong EUV
irradiation heats up the hydrogen rich
thermosphere so significantly that the internal energy of the
gas becomes greater than the gravitational potential energy.
This leads to the expansion of the atmosphere and powers
hydrodynamic escape of hydrogen which can drag off heavier
elements if the escape flux is high enough.
The Earth does not receive
sufficiently strong EUV irradiation from the Sun at present. Also
the total hydrogen mixing ratio in the present stratosphere is
as small as $10^{-5}$. Therefore, in today's Earth,
hydrogen escape is limited by diffusion through the homopause.
The bolometric luminosity is one of the important observable stellar
properties that determines the distance of the Habitable Zone (HZ) or the
circumstellar region in which a planet can have appropriate
temperature to sustain water in liquid state.
On the other hand, the EUV luminosity determines the rate at
which hydrogen escapes from the atmosphere of a planet in
the HZ. If the bolometric luminosity of the star is low, the
HZ is closer to the star and hence a planet in the HZ is
exposed to stronger EUV irradiation.
If the EUV luminosity of such faint
stars is sufficiently high then a habitable planet would lose
substantial amount of hydrogen and heavier gases. It
would lead to oxidation of the surface and accumulation of oxygen
in the atmosphere, scarcity of hydrocarbon and significant
reduction in surface water. All these would
make the planet uninhabitable. If a habitable planet undergoes
runway greenhouse, water vapor would reach the stratosphere
where it would get photolyzed and subsequently should escape
the atmosphere.
In this paper, I present an analytical expression for the
upper limit of the ratio between the EUV and the bolometric
luminosities of a planet-hosting star of any spectral type which
serves as an essential condition to ensure the presence of
sufficient amount of hydrogen, nitrogen, oxygen, water etc. that
supports the habitability of a rocky planet in the HZ of the
star. Since both $L_{EUV}$ and $L_b$ are observable quantities,
it will help to determine if a planet in the habitable zone of a
star is really habitable or not. Thus the limit should serve as a ready
reckoner for eliminating candidate habitable planets.
\section{CRITERIA FOR PLANETARY HABITABILITY}
The distance of a habitable planet from its
parent star is given by \citep{kasting93,koppa13a}
\begin{eqnarray}
d=\left(\frac{L_B}{L_\odot S_{eff}}\right)^{1/2}d_E
\label{hzd}
\end{eqnarray}
where $L_B$ and $L_\odot$ are the bolometric luminosity
of the star and the Sun respectively, $d_E$ is the distance between the Earth
and the Sun (1 AU) and $S_{eff}$ is the normalized effective stellar flux
incident on the planet.
Let $L_{EUV}$ be the EUV luminosity at the surface of a star. If
$\epsilon$ is the efficiency at which the EUV is absorbed by
the planetary atmosphere then the sphere-averaged and
efficiency corrected heating rate in the planetary
thermosphere due to stellar EUV irradiation is
\begin{eqnarray}
S=\frac{L_{EUV}}{L_B}\left(\frac{\epsilon S_{eff} L_\odot}{4\pi d^2_E}\right).
\label{ratio}
\end{eqnarray}
The heating efficiency is less than one because
part of the total incident EUV energy drives ionization,
dissociation and other reactions without striping out the atoms
or molecules from the atmosphere.
In order to remain habitable, $S$ must be low enough so that
the rate of hydrogen escape from the planetary surface due to
EUV irradiation does not exceed a critical value that
may cause the heavier constituents including oxygen to be
entrained with the outflow of hydrogen. If the mixing ratio
of hydrogen is low in the lower atmosphere, hydrogen escape
should be limited by diffusion. Therefore the critical rate
of energy limited loss should be comparable to that of diffusion
limited escape under such circumstance.
Let $S_c$ be the value of
$S$ corresponding to the critical rate of hydrogen-loss.
Therefore, the necessary condition to prevent
hydrogen loss at critical rate by EUV irradiation can
be written as :
\begin{eqnarray}
\frac{L_{EUV}}{L_{B}}< \frac{4\pi d_E^2}{\epsilon S_{eff} L_\odot}S_c.
\label{cond}
\end{eqnarray}
Now we need to derive $S_c$, the sphere-averaged,
efficiency-corrected and energy-limited critical EUV flux that
causes hydrogen to escape at the critical rate.
\section{ENERGY-LIMITED EUV FLUX}
In order to derive the energy-limited critical EUV flux $S_c$,
the formalism given by \cite{watson81} is adopted.
\cite{watson81} applied this formalism to estimate
the mass loss of Earth's hydrogen exosphere due to solar EUV
irradiation. It was also used by \cite{lammer03} for explaining
the observed extended atmosphere for the
close-in hot and giant transiting planet HD 209458b
\citep{vidal03}. The formalism is derived from the
usual steady state equations of mass momentum and energy
conservation of a dynamically expanding, non-viscous gas of
constant molecular weight in which the pressure is isotropic.
Unlike the Jean's treatment, this formalism can be applied to
a dense thermosphere of hydrogen with a fixed temperature at
the lower boundary located sufficiently above the homopause
such that the mixing ratios of heavier gases are negligible as
compared to that of hydrogen. All the EUV energy is absorbed in
a narrow region with visible optical depth less than unity
and no EUV energy is available bellow this region due to
complete absorption.
The rate of mass loss estimated by using this method is within
a factor of few \citep{hubbard07} of the same provided by other
models including those that involve detailed numerical solutions
\citep{baraffe04,tian05,yelle04}. However, \cite{murray09} argued
that for the irradiated EUV flux greater than $10^4$ erg
cm $^{-2}$ s$^{-1}$, mass loss ceases to be energy limited and
becomes radiation/recombination limited. As a consequence the
formalism prescribed by \cite{watson81} is not applicable in that
case.
The two equations that provide the maximum rate of
hydrogen escape and the expansion of the atmosphere
due to EUV heating are given as \citep{watson81} :
\begin{eqnarray}
\xi=\frac{2}{q+1}\left[\frac{(\lambda_1/2)^{(1+q)/2}+1}{
\lambda_0-\lambda_1}\right]^2
\label{watson1}
\end{eqnarray}
and
\begin{eqnarray}
\beta=\xi\lambda_1^2\left[\lambda_0-\left\{\frac{2}{(1+q)\xi}\right\}^{1/2}\right].
\label{watson2}
\end{eqnarray}
In the above equations the dimensionless parameters $\xi$, $\beta$,
$\lambda_{0}$ and $\lambda_{1}$ are related with the physical parameters by
\begin{eqnarray}
\xi=F_m\left(\frac{k^2T_0}{\kappa_0GM_Pm_H}\right),
\label{xi}
\end{eqnarray}
\begin{eqnarray}
\beta=S\left(\frac{GM_Pm_H}{\kappa_0 k T^2_0}\right),
\label{beta}
\end{eqnarray}
\begin{eqnarray}
\lambda_{0,1}=\frac{GM_Pm_H}{k T_0 r_{0,1}},
\label{radius}
\end{eqnarray}
where $r_0=R_P$ is the radius of the planet where the visible
optical depth is one and the temperature is $T_0$,
$r_1$ is the radius of the region
where all the EUV energy is absorbed. The optical depth of the atmosphere
to EUV energy at the lower boundary $r_0$ is much greater
than unity. $F_m$ is the sphere averaged
flux of escaping particles (escape flux of particles per
steradian per second), $k$ is the Boltzmann constant, $M_P$ is
the total mass of the planet, $m_H$ is the mass of hydrogen
atom and $\kappa$ is the thermal conductivity of the gas which is
parameterized by $\kappa=\kappa_0\tau^q$ where $\tau$ is
the visible optical depth of the thermosphere such that
$\tau(\lambda_0)=1$.
\section{ANALYTICAL AND NUMERICAL SOLUTIONS}
Equation~(\ref{watson1}) and equation~(\ref{watson2}) need to be
solved numerically for an arbitrary value of $q$. However,
for $q=1$ we obtain analytical solutions for $\lambda_1$ and
$\beta$ and hence $S$. Thus for $q=1$, equation~(\ref{watson1})
gives
\begin{eqnarray}
\lambda_1=\frac{2(\xi^{1/2}\lambda_0-1)}{1+2\xi^{1/2}}.
\label{sol1}
\end{eqnarray}
Substituting equation~(\ref{sol1}) in equation~(\ref{watson2}),
equation~(\ref{beta}) gives,
\begin{eqnarray}
S(q=1)=4\xi\left(\frac{\xi^{1/2}\lambda_0-1}{1+2\xi^{1/2}}
\right)^2\left(\lambda_0-\frac{1}{\xi^{1/2}}\right)
\left(\frac{\kappa_0 k T^2_0}{GM_Pm_H}\right).
\label{sol2}
\end{eqnarray}
Now, $S=S_c$ when $F_m=F_c$ in the expression for
$\xi$ given in equation~(\ref{xi}) where $F_c$ is the critical
rate of hydrogen escape. Hydrogen diffuses before it reaches the
thermobase if the background gas of heavier species is static.
This happens if the heavier species are not absorbed or
cannot escape at the surface. The diffusion limit for hydrogen
is achieved when the heavier gases attain the
maximum upward velocity such that the background becomes
non-static \citep{hunten73}. Therefore if $F_c$ is greater than or
equal to the diffusion limit, heavier species would escape.
The rate of hydrogen loss limited by diffusion is given by
\cite{hunten73, zahnle90}
\begin{eqnarray}
F_H(diffusion)=\frac{bg(m_s-m_H)}{kT_0(1+X_s/X_H)}
\end{eqnarray}
where $b=4.8\times10^{17}(T/K)^{0.75}$ cm$^{-1}$s$^{-1}$ is
the binary diffusion coefficient for the two species
\citep{zahnle86}, $X_H$ and $X_s$ are the molar mixing ratio at
the exobase for hydrogen and the heavier atom with mass $m_s$
respectively and $g$ is the surface gravity of the planet.
Therefore we set
\begin{eqnarray}
F_m=F_c=R_P^2\frac{bg(m_s-m_H)}{kT_0(1+X_s/X_H)}{\rm
particles~sr^{-1}s^{-1}}
\label{fm}
\end{eqnarray}
which gives the cross over mass $m_s=m_H+(kT_0F_c/R_P^2)/(bgX_H)$
such that any species with mass less
than or equal to $m_s$ would be efficiently dragged by the
escaping hydrogen \citep{hunten87}. In the present derivation
I consider $m_s=m_O=16m_H$ such that the critical rate of
hydrogen loss would enable oxygen and other atoms lighter
than oxygen to be dragged with the escaping gas. I assume that
both hydrogen and oxygen are atomic near the exobase due to fast
dissociation of the molecules by photolysis \citep{murray09}.
Therefore, from equation~(\ref{cond}) we obtain, for $q=1$
\begin{eqnarray}
\frac{L_{EUV}}{L_B}< \frac{16\pi d_E^2\kappa_0 k T^2_0}{\epsilon GM_Pm_H
(1-e^2)^{1/2}S_{eff}L_\odot}
\frac{\xi^{1/2}(\lambda_0\xi^{1/2}-1)^3}{(2\xi^{1/2}+1)^2},
\label{result}
\end{eqnarray}
where
$\lambda_0$, the parameter for the planetary radius is given by
equation~(\ref{radius}) and
\begin{eqnarray}
\xi=R_P^2\frac{bg(m_O-m_H)}{1+X_O/X_H}\frac{k}{\kappa_0GM_Pm_H}.
\end{eqnarray}
The term $(1-e^2)^{1/2}$ in the denominator is
introduced in order to include elliptical orbit with
eccentricity $e$.
The above analytical expression is derived by taking $q=1$. However, for a
neutral gas, $q\simeq 0.7$ \citep{banks73}.
Therefore, I solve equation~(\ref{watson1}) and
equation~(\ref{watson2})
numerically for different values of $q\leq 1.0$ keeping all other
parameters fixed. Figure~1 presents the values of $S_c$ for
different values of $q$ and shows that $S_c (q=1) < S_c (q<1)$.
Therefore, equation~(\ref{cond}) implies that the inequality
given by equation~(\ref{result}) is valid for any value of
$q\leq 1$.
It is worth mentioning that the bolometric luminosity of
a main sequence star increases with time. For a solar-type star,
the bolometric luminosity increases by about 10 \% in every
1 Gyr. The Sun was about 30\% fainter in the visible
light during the first billion years after its birth. On the other hand the EUV luminosity is governed
by the coronal activities of a star. Since a star rotates faster
during younger age resulting into greater coronal activities,
the EUV emission rate or the EUV luminosity was higher in the
past and it decreases with time. Consequently, the rate of mass
loss from the planetary surface was higher in the past.
Since, the value of $L_{EUV}/L_{B}$ in the past was greater than
its present value, the condition derived here is valid for a
constant rate of mass loss corresponding to the present value of
$L_{EUV}/L_B$. A time-dependent solution would have provided more
stringent condition.
\section{APPLICATION TO HABITABLE EXOPLANETS }
The thermal conductivity of hydrogen $\kappa_0=4.45\times10^4$
ergs cm$^{-1}$s$^{-1}$K$^{-1}$ \citep{hanley70}. The size and the
mass of the planets are usually determined from the observed
parameters of the planet. Transit method provides the size and the
orbital inclination angle of the planet and radial velocity method provides
the projected mass of the planet for known stellar mass.
Therefore, the two unknown parameters in equation~(\ref{result})
are $T_0$ and $S_{eff}$.
The lower boundary $r_0$ is located some distance
above the homopause and the temperature $T_0$ at the lower boundary may be
fixed at the equlibrium temperature of the planet \citep{watson81}. However,
as shown in Figure~2, the above limit is weakly dependent on $T_0$ because
the parameter $\lambda_0$ is inversely proportional to $T_0$. Therefore,
without loss of generality, $T_0$ can be fixed at 254 K, the equilibrium
temperature of the Earth. This makes the above expression independent of
planetary albedo because derivation of equilibrium temperature requires the
value of planetary albedo. With $T_0=254$K and taking
$X_O=1/3$ and $X_H=2/3$, I
obtain $F_c=1.43\times10^{13}$ cm$^{-2}$ s$^{-1}$ which is
two orders of magnitude higher than that calculated by
\cite{watson81} for the Earth irradiated by solar EUV radiation.
This ensures the cross over mass $m_s$ is equal to $m_O$.
The total mass of hydrogen ($M_H$) in the
atmosphere, Ocean and in the crust of the Earth is about
$1.9\times10^{23}$ gm \citep{anders77, sharp09}.
Therefore at this critical rate, an Earth like planet would
lose all of its surface hydrogen in about 50 Myr. Water
molecules would be photo-dissociated into hydrogen and oxygen
atoms that would subsequently escape the planet. Even if the
amount of water is not reduced sufficiently, such a significant
loss of hydrogen and other heavier gases such as nitrogen and
oxygen would make the environment drastically different than
that of the present Earth.
\cite{koppa13a} have presented relationships between the stellar effective
temperature $T_{eff}$ and the normalized stellar flux $S_{eff}$ impinging on
the top of the atmosphere of an Earth-like planet at the habitable zone.
These relationships are applicable in the range 2600 K$\leq T_{eff} \leq $
7200 K. According to these relationships, $S_{eff}$ is minimum at
$T_{eff}=2600$ K and
increases with the increase in $T_{eff}$. $S_{eff}$ also varies for different
atmospheric conditions that determines the inner and the outer
boundaries of the habitable zone. While condition like "Recent
Venus", "Runway Greenhouse" or "Moist Greenhouse" determines the inner
boundary, condition such as "Maximum Greenhouse" or
"Early Mars" determines the outer boundary of the habitable zone. Since,
$L_{EUV}/L_B$ in equation~(\ref{result}) is inversely proportional to
$S_{eff}$, I use the minimum value of $S_{eff}$
corresponding to $T_{eff}=2600$K. This makes the upper limit on $L_{EUV}/L_B$
independent of the effective temperature of the star.
As $S_{eff}$ increases with the
increase in $T_{eff}$, the value of $S_{eff}$ corresponding
to the actual $T_{eff}$ of the star would only tighten the limit. It is
worth mentioning here that according to the relationships provided by
\cite{koppa13a}, $S_{eff}$ increases only by 2-3 times with the increase
in $T_{eff}$ from 2600 K to 7200 K.
Taking $T_0=254$ K- the equilibrium temperature of the Earth,
$R_P=6.378\times10^8$ cm - radius of the Earth,
$M_P=5.9726\times10^{27}$ gm - mass of the Earth,
and $\epsilon=0.15$ \citep{watson81}, I derive, from equation~(\ref{result}),
the upper limit on $L_{EUV}/L_B$ for planets within the outer boundary of
the habitable zone. The values are presented in Table~1.
For the same set of parameters the numerical solution
of equation~(\ref{watson1}) and equation~(\ref{watson2}) for $q=0.7$ is also
presented in Table~1.
\begin{table}[h]
\begin{center}
\caption{Upper limit on $L_{EUV}/L_B$ for $S_{eff}$ corresponding to $T_{eff}=
2600$K under atmospheric conditions that determines the outer boundary of
the habitable zone. For $q=0.7$ the results are obtained numerically.}
\begin{tabular}{ccccc}
\hline \hline
Conditions & $S_{eff}$ & & $L_{EUB}/L_B$ & \\
& & q=1.0 & & q=0.7 \\
\hline
Maximum Greenhouse & 0.215 & $1.33\times10^{-4}$ & & $1.56\times10^{-4}$ \\
Early Mars & 0.199 & $1.44\times10^{-4}$ & & $1.69\times10^{-4}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
The critical EUV flux $S_c$ is less than 10 erg cm$^{-2}$ s$^{-1}$ which ensures
energy-limited escape. Also, equation (\ref{sol1}) gives
the expanded radius $r_1=1.57R_P$ for $T_0=254$K.
\section{DISCUSSIONS AND CONCLUSIONS}
Estimating the stellar EUV luminosities is difficult not only
because the energy gets completely absorbed at the uppermost
layer of the Earth's atmosphere but also due to photoelectric
absorption by the interstellar medium along the line of sight.
Using the ROSAT space telescope, \cite{hodgkin94} estimated
the EUV luminosities of a large number of nearby stars with
spectral type ranging from F to M . Here, I use the bolometric
and EUV luminosities of planet-hosting stars of spectral type
ranging from A to M, given by \cite{sanz11} who
derived the EUV and bolometric luminosities
by using the data from ROSAT,
XMM-Newton and Chandra space telescopes. I have considered
only those stars that are older than 1 Gyr.
It is found that out of all such stars, only two of
the planet-hosting M stars satisfies the
criteria presented in this work. The value of
$L_{EUV}/L_{B}$ for 2MASS 1207 (spectral type M8) is
$1.23\times 10^{-3}$. For GJ 317 (M3.5), GJ 674
(M2.5) and for GJ 176 (M2.5V), the values of $L_{EUV}/L_B$ are
$1.78\times10^{-3}$, $2.95\times10^{-4}$, $4.26\times10^{-4}$ respectively.
Therefore, none
of these M stars can have a planet with habitable environment.
GJ 832 (M1.5V) has $L_{EUV}/L_B=10^{-4}$ and so it marginally satisfies
the habitability criteria. However, since the effective temperature of this
M1.5V spectral type star is higher than 2600K, this planet should also
be considered uninhabitable.
On the other hand the values of $L_{EUV}/L_B$ for
GJ 436 (M2.5) and GJ 876 (M4V) are $1.58\times 10^{-5}$ and
$1.9\times 10^{-5}$ respectively.
Therefore, these two M stars satisfies the EUV habitability
criteria provided in Table~1.
This is expected because M stars are the faintest among
stars of all spectral types. As a consequence the HZ
of M stars is located very near to the stars and hence
a planet in the habitable zone is exposed to strong EUV radiation.
It is worth mentioning here that a rocky planet in the habitable zone
of M stars may lose much of its volatiles during the formation
because the pre-main sequence phase of such stars
are comparatively longer. However, if the planet accretes
sufficiently large amount of water during formation or if it
were formed far away from the star and then migrated to the HZ,
it may remain habitable \citep{lissa07}. But strong EUV
irradiation should make it uninhabitable.
However, planet hosting stars of all other spectral types listed
by \cite{sanz11} have $L_{EUV}/L_B$ much lower than
the upper limits provided here.
The values of $L_{EUV}/L_B$ for GJ 86 (K1V) is $9.77\times 10^{-5}$. Therefore,
according to the present upper limit, this star marginally satisfies the
habitability criteria of having sufficient amount of water.
On the other hand, the values of $L_{EUV}/L_B$
for HD 87883 (K0V), $\epsilon$
Eridani (K2V), HD 46375 (K1V), HD 93083 (K3V),
HD 130322 (K0V), HD 189733 (K1-K2) and HD 218566 (K3V) are
$2.75\times 10^{-5}$, $2.19\times 10^{-5}$,
$1.02\times 10^{-5}$, $1.479\times10^{-5}$,
$1.95\times 10^{-5}$, $2.4\times 10^{-5}$ and
$1.66\times 10 ^{-5}$ respectively. So, all of them
satisfies the upper limit for $L_{EUV}/L_B$.
Hence, the upper limit
presented here is important in deciding the habitability of
planets around M stars. Note that the numerical values of
the parameters involved in the present habitability condition do
not differ much from that of the Earth. The density of a rocky
planet should be about 4.0-5.5 gmcm$^{-3}$. Therefore, the
radius of a rocky planet does not exceed 1.6 times the Earth's
radius and the mass must be less than 5 times the Earth's mass \citep{rogers15}.
More massive planets are Neptune and Jupiter types of gaseous
planets \citep{marcy14}. Therefore, a large value of
$L_{EUV}/L_B$ for M stars may rule out the presence
of habitable planets around them.
Owing to our poor knowledge on the actual environment of
exoplanets, the habitability of rocky planets around the habitable zone
remains ambiguous and incomplete. However, we must impose as many
conditions as can be determined directly and easily in order
to narrow down our search for planets that may possibly harbor
life. The present result provides such an important condition as it
depends on two observable quantities of the planet hosting star -
the bolometric and EUV luminosities.
Spectra of habitable planets around M-type stars
that do not satisfy the EUV criteria presented here
should show lack of hydrogen, nitrogen and oxygen in their
atmosphere and hence can confirm the limit presented here.
I thank the reviewer for a critical reading of the manuscript and for
providing useful suggestions and comments.
\clearpage
|
\section{Introduction}
\subsection{Toral nodal intersections}
Let $\mathbb T^d={\mathbb R}^d/{\mathbb Z}^d$ be the standard flat $d$-dimensional torus and
$\curve \subset \mathbb T^d$ a fixed reference curve\footnote{By a curve we always mean a parameterized, compact, immersed curve. }. Given a real-valued
eigenfunction $F(x)$ of the Laplacian
\begin{equation*}
-\Delta F=4\pi^2\zeigen \cdot F \;,
\end{equation*}
we wish to study the number of intersections
\begin{equation*}
\Zc(F):=\#\{x\in \mathbb T^d: F(x)=0\}\cap \curve
\end{equation*}
of the nodal set of $F$ with the reference curve $\curve$ as a
function of the corresponding eigenvalue.
In dimension $d=2$, if the curve is smooth and has nowhere vanishing
curvature, then deterministically for {\em every} eigenfunction, the
number of nodal intersections satisfies \cite{BRGAFA, BRNI}
\begin{equation*}
\frac{\zeigen^{1/2}}{(\log \zeigen)^{5/2}} \ll \Zc(F) \ll
\zeigen^{1/2};
\end{equation*}
here and everywhere $f\ll g $ (equivalently $ f=O\left(g\right) $) means that
there exists a constant $ C>0 $ such that $ |f|\le C|g| $.
For the upper bound we require that $\curve$ is real-analytic, in
particular it is shown that $\curve$ is not contained in the nodal
set for $\zeigen$ sufficiently large. One can improve the lower
bound to $\Zc(F)\gg \zeigen^{1/2}$ conditionally on a certain
number-theoretic conjecture \cite{BRNI}.
In this note we deal with dimension $d=3$. If we consider the
intersection of the nodal set with a fixed real-analytic reference
{\em surface} $\Sigma\subset \mathbb T^3$ with nowhere-zero Gauss-Kronecker
curvature, then for $\zeigen$ sufficiently large, $\Sigma$ is not
contained in the nodal set \cite{BRINV}, the length of the
intersection of $\Sigma$ with the nodal set is $\ll \sqrt{\zeigen}$
\cite{BRGAFA}, and the nodal intersection is non-empty
\cite{BRGAFA}. However, for the intersection of the nodal set with a
fixed reference {\em curve} (where we expect only finitely many
points), the following examples indicate that one cannot expect to
have any deterministic bounds on the number of nodal intersections.
\begin{example}
Take the eigenfunctions of the form $$F_k(x_1,x_2,x_3)=\sin(2\pi
kx_1);$$ their nodal surfaces are the planes $\left\{x\in
\mathbb T^3:x_1\in \frac 1 {2k} {\mathbb Z} \right\}$ and any curve lying on the
plane $x_1=1/2\pi$ does not intersect these, whereas a curve on the
plane $x_1=0$ is lying inside all of these nodal surfaces. We
observe that the curves in this example are {\em planar}.
\end{example}
\begin{example}
Let $F_{0}\left(x,y\right)$ be an eigenfunction on the
two-dimensional torus with eigenvalue $4\pi^2\zeigen_{0}^{2}$, and
$S_{0}$ a curved segment contained in the nodal set, admitting an
arc-length parameterization $\gamma_{0}: [0,L ]\to S_{0}$, with
curvature $\kappa_0(t)=|\gamma_0''(t)|>0$. For $n\ge 0$ let
$F_{n}\left(x,y,z\right)=F_{0}\left(x,y\right)\cos\left(2\pi
nz\right)$, an eigenfunction on $\mathbb T^{3}$ with eigenvalue
$4\pi^2(\zeigen_0^2+n^{2})$. Let $\curve$ be the parametric curve
$\gamma\left(t\right)=\left(\gamma_{0}(\frac t{\sqrt{2}}),\frac
t{\sqrt{2}}\right)$. A computation shows that the curvature is
$\kappa(t) = \frac 12 \kappa_0(\frac t{\sqrt{2}})>0$, and that the
torsion is $\tau(t) = \pm\frac 1{2 }\kappa_0(\frac t{\sqrt{2}})\neq
0$,
so that $\curve$ is non-planar. Clearly $\curve $ is contained in
the nodal set of $F_{n}$ for all $n$. Thus even in the non-planar
case, we can have the reference curve $\curve$ contained in the
nodal set for arbitrarily large $\zeigen$.
\end{example}
\begin{question} Does there exist a non-planar curve
$\curve$ with no nodal intersections, $\zeigen$ arbitrarily large?
\end{question}
\subsection{Arithmetic random waves}
As there is no deterministic bound on $\Zc(F)$ in dimension $3$, we
investigate what happens for ``typical" eigenfunctions. Let
$\vE(\zeigen)$ be the set of lattice points lying on a sphere
\begin{equation*}
\vE(\zeigen) = \{\vec x\in {\mathbb Z}^3: |\vec x|^2=\zeigen\},
\end{equation*}
and let
\begin{equation*}
N= N_\zeigen:=\#\vE(\zeigen).
\end{equation*}
The Laplace spectrum on $\mathbb T^3$ is of high multiplicities, with the
dimension of an eigenspace corresponding to an eigenvalue
$4\pi^2\zeigen$ being of size $N_\zeigen\approx \zeigen^{1/2 \pm
o(1)}$. The eigenfunctions corresponding to the eigenvalue
$4\pi^2\zeigen$ are of the form
\begin{equation}
\label{eq:F arith wave def} F(x)=F_{\zeigen}\left( x\right) =
\dfrac{1}{\sqrt{N_\zeigen}} \sum_{\mu\in
\mathcal{E}(E)}{a_{\mu}e^{2\pi i \left <\mu , x\right >}}.
\end{equation}
We consider random Gaussian eigenfunctions (``arithmetic random
waves" \cite{KKW}) by taking the coefficients $ a_\mu $ to be
standard complex Gaussian random variables, independent save for the
relations $ a_{-\mu} = \overline{a_{\mu}}$, making $F$ real-valued.
\subsection{Statement of the main results}
\begin{theorem}\label{thmA}
Let $\curve \subset \mathbb T^3$ be a smooth curve of length $L$, with
nowhere zero curvature. Assume further that one of the following
holds:
\begin{enumerate}
\item $\curve$ has nowhere-vanishing torsion;
\item $\curve$ is planar (so that the torsion vanishes identically).
\end{enumerate}
Then for all $\epsilon>0$, as $\zeigen \to \infty$ along integers
$\zeigen\not\equiv 0,4,7\bmod 8$, the number of nodal intersections
satisfies
\begin{equation*}
\lim_{\substack{E\to \infty\\ E\not\equiv 0,4,7 \bmod 8}}
\operatorname{Prob}\left(\left|\frac{\Zc(F)}{\sqrt{\zeigen}}-\frac{2}{\sqrt{3}}L
\right|>\epsilon\right) \to 0.
\end{equation*}
\end{theorem}
Note that the condition $\zeigen\not\equiv 0,4,7\bmod 8$ is natural,
as otherwise $\zeigen=4^a \zeigen'$ with $\zeigen'\not\equiv 7 \bmod
8$ (if $\zeigen'\equiv 7\bmod 8$ then $\zeigen$ is not a sum of
three squares and hence does not yield a Laplace eigenvalue); then
an eigenfunction of eigenvalue $4\pi^2\zeigen$ is necessarily of the
form $F(x)=H(2^a x)$ with $H$ an eigenfunction of eigenvalue
$\zeigen'$. Hence any question on the nodal set of $F$ reduces to the
corresponding question on the nodal set of eigenfunctions with
eigenvalue $E'=E/4^a$ (which may be trivial, e.g. if $E=4^a$).
To prove Theorem~\ref{thmA} we compute the expected value of $\Zc$
with respect to the Gaussian measure defined on the eigenspace as
above to be
\begin{equation*}
\mathbb E(\Zc) = \frac{2}{\sqrt{3}}L\sqrt{\zeigen},
\end{equation*}
and give an upper bound for the variance:
\begin{theorem}\label{thmvar}
Let $\curve \subset \mathbb T^3$ be a smooth curve, with nowhere-zero
curvature. Assume also that either $ \curve $ has nowhere-vanishing
torsion, or $ \curve $ is planar. Then for $\zeigen\not\equiv
0,4,7\bmod 8$,
\begin{equation}
\label{eq:Var<<eig^-delta} \operatorname{Var}
\left(\frac{\Zc}{\sqrt{\zeigen}}\right) \ll \frac
1{\zeigen^{\delta}}
\end{equation}
for all $\delta<1/3$ in case $\curve$ has nowhere vanishing torsion, and
all $\delta<1/4$ in case $\curve$ is planar.
\end{theorem}
\begin{remark}
\label{rem:variance analytic} If $\curve$ is real-analytic and
non-planar, so that the torsion is not identically zero, but may
vanish at finitely many points, the result
\eqref{eq:Var<<eig^-delta} above is valid with some
$\delta=\delta_\curve>0$, see \S~\ref{sec:thmvar proof}.
\end{remark}
\subsection{Outline of the paper and the key ideas}
We prove the approximate Kac-Rice formula (briefly explained in
\S~\ref{sec:Kac-Rice short}) in \S~\ref{sec:Kac-Rice},
followed by a study of certain oscillatory integrals on the curve in
\S~\ref{sec:oscillary integrals}. The arithmetic heart of the paper
is \S~\ref{sec:second_moment}, where we bound the second moment of
the covariance function and its derivatives, following some
background on the arithmetic of sums of three squares in
\S~\ref{sec:sums of 3 squares}, expanded on in Appendix~\ref{sec:Appendix}.
\vspace{3mm}
A similar result to Theorem \ref{thmvar}
was proved in the two-dimensional case for $\curve
\subset \mathbb T^2$ having nowhere-zero curvature \cite{RW2014}. In that
case the authors found that the precise asymptotic behaviour of the
nodal intersections variance is {\em non-universal}, namely
dependent on both the angular distribution of the lattice points $
\mathcal{E}\left (E\right ) $ and the geometry of $\curve$. In the
$3$-dimensional case we were only able to obtain an upper bound
\eqref{eq:Var<<eig^-delta} on the variance, which implies the
``almost-all" statement of Theorem~\ref{thmA}. These two cases
differ both in terms of analytic and arithmetic ingredients; the
arithmetic of ternary quadratic forms differs significantly from
that of binary quadratic forms.
\subsection{An approximate Kac-Rice formula}
\label{sec:Kac-Rice short}
We end this Introduction with a discussion of a key step in our work, the approximate Kac-Rice formula.
By restricting the arithmetic random waves \eqref{eq:F arith wave
def} along $\curve$, the problem of nodal intersections count is
reduced to evaluating the number of zeros (zero crossings) of a
random (nonstationary) Gaussian process. The Kac-Rice formula is a
standard tool or meta-theorem for expressing the (factorial) moments
of the zero crossings number of a Gaussian process precisely in
terms of certain explicit integrals with integrands depending on the
given process.
It is then easy to evaluate the expected number of zeros
precisely and explicitly via an evaluation of a standard Gaussian
expectation.
For the variance, or the second factorial moment, the validity of
the Kac-Rice formula is a very subtle question with a variety of
sufficient conditions known in the literature (e.g. ~\cite{CL,AW}).
While the classical treatise ~\cite{CL} requires the non-degeneracy
of the (Gaussian) distribution of the values of the given process at
two points together with their derivatives, only the non-degeneracy
of the values distribution (at two points) is required for the more
modern treatment ~\cite{AW}, which, to the best knowledge of the
authors, is the weakest known sufficient condition for the validity
of Kac-Rice. Unfortunately, even this weaker condition may fail in
our case.
In order to treat this situation (\S \ref{sec:Kac-Rice}) we
divide the interval into many small subintervals of length
commensurable to the wavelength $\frac{1}{\sqrt{\zeigen}}$
and decompose the total variance as a sum over pairs of subintervals
of zero number covariances \eqref{eq:var=sum cov ij}. We were able
to prove the validity of Kac-Rice for {\em most} of the pairs of
subintervals that includes all the diagonal pairs (i.e. the variance
of nodal intersections along sufficiently small curves), and bound
the contribution of the other pairs via a simple Cauchy-Schwartz
argument. The price is that along the way we incur an error term,
hence yielding an {\em approximate} Kac-Rice (Proposition
\ref{prop:approx Kac-Rice}) reducing a variance computation to an
estimate for some moments of the covariance function and its
derivatives; such a strategy was also used in \cite{RW2014, CMW}.
The remaining part of this paper is concerned with proving such an
estimate on the second moment of the covariance function and a
couple of its derivatives.
We finally record that we may omit a certain technical assumption
made in our previous work on the two-dimensional case \cite{RW2014}
by using the more general form of the Kac-Rice formula as in
~\cite{AW} (vs. ~\cite{CL}). Given an integer $m$ expressible as a
sum of $2$ squares we defined the probability measure
$$\tau_{m}=\sum\limits_{\|\lambda\|^{2}=m}\delta_{\lambda/\sqrt{m}}$$ on the unit circle $\mathcal{S}^{1}\subseteq {\mathbb R}^{2}$
supported on all the lattice points $\lambda\in{\mathbb Z}^{2}$ lying on the
centered radius-$\sqrt{m}$ circle in ${\mathbb R}^{2}$ projected to the unit
circle. In \cite[Theorem 1.2]{RW2014} we evaluated the
variance of the number of nodal intersections under the assumption
that the Fourier coefficient $\widehat{\tau_{m}}(4)$ is bounded away
from $\pm 1$ (i.e. $\tau_{m}$ bounded away from the singular
measures $\frac{1}{4}(\delta_{\pm 1}+\delta_{\pm i})$ and its
$\pi/4$-tilted version). Using ~\cite{AW} makes that assumption no
longer necessary.
\section{An approximate Kac-Rice formula}
\label{sec:Kac-Rice}
\subsection{The Kac-Rice premise}
Let $d\geq 3$, and $ \curve\subseteq \mathbb T^d $ be a smooth curve. Let $
F $ be the arithmetic random wave
\begin{equation*}
F(x)= \dfrac{1}{\sqrt{N_\zeigen}} \sum_{\mu\in
\mathcal{E}(E)}{a_{\mu}e^{2\pi i \left <\mu , x\right >}}
\end{equation*}
(with the obvious generalization of all the previous notation to
higher dimensions $ d\ge 3 $). We wish to study the number $
\mathcal{Z} (F) $ of nodal intersections of $F$ with $\curve$ by
restricting $ F $ to $ \curve $ as follows.
Let $\gamma:[0,L]\rightarrow {\mathbb R}$ be a unit speed parameterization of
$ \curve $. We restrict $ F $ to $ \curve $ by defining the random
Gaussian process
\begin{equation}
\label{eq:f(t)=F(gamma(t))} f(t)=F \left (\gamma (t)\right )
\end{equation}
on $ [0,L] $. It is then obvious that the nodal intersections number
$\mathcal{Z} (F)$ is equal to the number of the zeros of $ f $. The
Kac-Rice formula (see e.g. \cite{CL}, \cite{AW}) is a standard tool
(meta-theorem) for evaluating the expected number and higher
(factorial) moments of zeros of a ``generic" process: let
$X:I\rightarrow {\mathbb R}$ be a (a.s. $C^{1}$-smooth, say) random Gaussian
process on an interval $I\subseteq {\mathbb R}$, and $\Zc=\Zc_{I;X}$ the
number of zeros of $X$ on $I$. For $m\ge 1$ and distinct points $t_{1},\ldots,
t_{m}\in I$ denote $\varphi_{t_{1},t_{2},\ldots, t_{m}}(u_{1},\ldots
u_{m})$ to be the (Gaussian) probability density function of the
random vector $(X(t_{1}),\ldots X(t_{m}))\in {\mathbb R}^{m}$. Then, under
appropriate assumptions on $X$, the $m$-th factorial moment of $\Zc$
is given by
\begin{equation}
\label{eq:Kac_Rice}
\begin{split}
\mathbb E \left[ \Zc^{[m]} \right] = &\int_{I^m}
K_{m}(t_{1},\ldots,t_{m}) \, \mbox{d}t_{1}\ldots \mbox{d}t_{m},
\end{split}
\end{equation}
where
\[ \Zc^{[m]} := \begin{cases} \Zc(\Zc-1)\cdots (\Zc-m+1) & 1 \le m\le \Zc \\ 0 & \mbox{otherwise}. \end{cases}, \]
and $K_{m}$, given by
\begin{equation}
\label{eq:Km def}
\begin{split}
K_{m}(t_{1},\ldots t_{m}) &=
\varphi_{t_{1},\ldots,t_{m}}(0,\ldots,0)\times\\&\times\mathbb E \left[
\left \vert X'(t_1)\dots X'(t_m) \right \vert \, \Big \vert
X(t_1)=0, \dots, X(t_m)=0 \right],
\end{split}
\end{equation}
is the $m$-th zero-{\em intensity} of $\Zc$. Note that for the
Gaussian case
\begin{equation}
\label{eq:phi(0...0) expr} \varphi_{t_{1},\ldots,t_{m}}(0,\ldots,0)
= \frac{1}{(2\pi)^{m/2}\sqrt{\det{A}}}
\end{equation}
with $A$ the covariance matrix of the values
$(X(t_{1}),\ldots,X(t_{m}))$, provided that $\det{A}\ne 0$, or,
equivalently, that the distribution of $(X(t_{1}),\ldots,X(t_{m}))$
is non-degenerate.
The validity of the meta-theorem \eqref{eq:Kac_Rice} was established
under a number of various scenarios. Originally the result
\eqref{eq:Kac_Rice} was proven to hold ~\cite{CL} provided that for
all distinct points $t_{1},\ldots,
t_{m}\in I$ the distribution of the Gaussian
vector $(X(t_{1}),\ldots, X(t_{m}),X'(t_{1}),\ldots
X'(t_{m}))\in{\mathbb R}^{2m}$ is non-degenerate. This non-degeneracy
condition was relaxed\footnote{This fortunate fact simplifies our
treatment of the approximate Kac-Rice formula below (though it is
possible to work ~\cite{RW2014} with the more restrictive version to
obtain the same results).} ~\cite{AW}, as in the following theorem:
\begin{theorem}[~\cite{AW}, Theorem $6.3$]
\label{thm:Kac-Rice} Let $X:I\rightarrow{\mathbb R}$ be a Gaussian process
having $C^1 $ paths and $m\ge 1$. Assume that for every $m $
pairwise distinct points $ t_1,\ldots ,t_{m}\in I $, the joint
distribution of $ (X(t_1), \dots X(t_m))\in {\mathbb R}^{m} $ is
non-degenerate. Then \eqref{eq:Kac_Rice} holds.
\end{theorem}
The cases $m=1,2$ are of our particular interest (see the following
sections). The main problem is that for $m=2$ even the weaker
non-degeneracy hypothesis in Theorem \ref{thm:Kac-Rice} may not be
satisfied for our process $f$ as in \eqref{eq:f(t)=F(gamma(t))}; to
resolve this issue we will decompose the interval $I=[0,L]$ into
small subintervals, and apply Kac-Rice for each pair of the
subintervals to develop ``approximate Kac-Rice" formula following an
idea from \cite{RW2014} in the two-dimensional case (see
\S~\ref{sec:var=sum covar}).
The covariance function of the centered Gaussian random field $ F $
reads\[ r_F(x,y) := \mathbb{E}\left [F(x)F(y)\right ] =
\dfrac{1}{N_\zeigen} \sum_{\mu\in \mathcal{E}}\cos\left (2\pi \left
<\mu ,y-x\right >\right )\] for $ x,y\in \mathbb{T}^d $. As $F$ is
stationary ($r_F$ depending on $ y-x $ only), we may think of $r_F$
as a function of one variable on $\mathbb{T}^d $. The covariance
function of $ f $ is \[ r(t_1,t_2) = r_f(t_1,t_2) := \mathbb{E}\left
[f(t_1)f(t_2)\right ] = r_F\left (\gamma(t_1)-\gamma(t_2)\right ).
\] Therefore, $ f $ is a centered unit variance Gaussian process
(non-stationary); $ r(t_1,t_2)\ne \pm1 $ if and only if the joint
distribution of $ f(t_1),f(t_2) $ is non-degenerate, so the
probability density $ \varphi_{t_1,t_2} $ of the Gaussian random
vector $ \left ( f(t_1),f(t_2) \right ) $ exists. Denote
\[ r_1:=\frac{\partial r}{\partial t_1}, r_2:=\frac{\partial r}{\partial t_2}, r_{12}:=\frac{\partial ^2 r}{\partial t_1 \partial t_2}, \]and let
\begin{equation}
\label{eq:r_2} \Rc_{2}(\zeigen) := \int\limits_{[0,L]^{2}}
\left(r^{2}+(r_{1}/\sqrt{\zeigen})^2 +
(r_{2}/\sqrt{\zeigen})^{2}+(r_{12}/\zeigen)^2\right) \, \mbox{d}t_1
\mbox{d}t_2
\end{equation}
be the sum of second moments of $r$ and its few normalized
derivatives along $\curve$; we will control the various quantities
via $\Rc_{2}$ (see Proposition \ref{prop:approx Kac-Rice} below).
Later we will show that $\Rc_{2}(\zeigen)$ is decaying with
$\zeigen$ (\S~\ref{sec:second_moment}).
\begin{proposition}[Approximate Kac-Rice formula]
\label{prop:approx Kac-Rice} We have
\[ \operatorname{Var} \left (\frac{\mathcal{Z}}{\sqrt{\zeigen}}\right) = O( \Rc_{2}(\zeigen))\]
with $\Rc_{2}$ is given by \eqref{eq:r_2}.
\end{proposition}
The rest of this section is dedicated to the proof of Proposition
\ref{prop:approx Kac-Rice}, finally given in
\S~\ref{sec:var=sum covar}, following some preparations.
\subsection{Expectation}
Since $ f $ is a centered unit variance Gaussian process with $ C^1
$ paths (in, particular, the non-degeneracy condition of Theorem
\ref{thm:Kac-Rice} is automatically satisfied), we may use the
Kac-Rice formula \eqref{eq:Kac_Rice}; for $m=1$ it reads
\begin{equation}
\label{eq:E[Z]=int(K1)} \mathbb E[\Zc] = \int\limits_{I}K_{1}(t) \, \mbox{d}t
\end{equation}
with \[ K_1(t)= \frac{1}{\sqrt{2 \pi}}\cdot \mathbb E\left [ \left
|f'(t)\right | \Big| \, f(t)=0 \right ] ,\] the ``zero density" of
$f$ (here \eqref{eq:phi(0...0) expr} reads
$\varphi_{t}(0)=\frac{1}{\sqrt{2\pi}}$). Let $ \Gamma $ be the
covariance matrix of $ \left (f(t),f'(t)\right ) $: \[ \Gamma(t)=
\begin{pmatrix}
r(t,t) & r_1(t,t) \\ r_2(t,t) & r_{12}(t,t)
\end{pmatrix}. \]
\begin{lemma}
For a smooth curve of length $ L $, the expectation of nodal intersections number is given by
\begin{equation*}
\mathbb E[\Zc] = L \frac{2}{\sqrt{d}}\cdot \sqrt{\zeigen}
\end{equation*}
\end{lemma}
\begin{proof}
Since $ f $ is unit variance, it is immediate that for every $ t\in
[0,L] $ we have $ r_1(t,t)=r_2(t,t)=0 $, so \[
\Gamma(t)=\begin{pmatrix} 1&\\&\alpha
\end{pmatrix}, \] where $$\alpha := r_{12} (t,t) = -\dot{\gamma}(t)^tH_{r_F}(0,0)\dot{\gamma}(t)$$ and $ H_{r_F}$ is the Hessian of $ r_F $ (see e.g. \cite{RW2014}). Since $ H_{r_F}(0,0) $ is a scalar matrix \cite{RW} \[ H_{r_F}(0,0)= \dfrac{-4}{d}\pi ^2\zeigen\ \cdot I_d \] it follows that $$ \alpha = \dfrac{4}{d}\pi ^2\zeigen. $$ The distribution of $ f'(t)$ conditional on $ f(t)=0 $ is centered Gaussian with variance $ \alpha $. Recall that for $ X \sim N(0,\sigma^2) $ we have $ \mathbb E(|X|)= \sigma \sqrt{2/ \pi}$, so
\begin{equation} \label{eq:K_1}
K_1(t)=\dfrac{1}{\pi}\sqrt{\alpha}=
\dfrac{2}{\sqrt{d}}\sqrt{\zeigen}
\end{equation}
independent of $x$, and the statement of the lemma follows upon
substituting \eqref{eq:K_1} into \eqref{eq:E[Z]=int(K1)}.
\end{proof}
\subsection{Variance}
For $m=2$ the Kac-Rice formula \eqref{eq:Kac_Rice} reads
\begin{equation}
\label{eq:Kac_Rice2} \mathbb E[\Zc^{2}-\Zc]=\int\limits_{I\times I}
K_{2}(t_{1},t_{2}) \, \mbox{d}t_{1}\mbox{d}t_{2}
\end{equation}
with $K_{2}$, the ``$2$-point correlation function", defined for
$r(t_{1}, t_{2})\ne \pm 1$ as (see \eqref{eq:Km def} and
\eqref{eq:phi(0...0) expr})
\[ K_2(t_1,t_2)=\frac{1}{2\pi \sqrt{1-r^{2}}} \cdot \mathbb{E}\left[\left\vert f'(t_1)\right\vert \cdot \left \vert f'(t_2)\right \vert \Big| \, f(t_1)=f(t_2)=0 \right ], \]
holding (Theorem \ref{thm:Kac-Rice}) provided that for all $t_{1}\ne
t_{2}$ we have $r(t_{1},t_{2}) \ne \pm 1$ (equivalently, the
distribution of $(f(t_{1}),f(t_{2}))$ is non-degenerate).
Equivalently, the zero number variance is given by (cf.
\eqref{eq:E[Z]=int(K1)})
\begin{equation}
\label{eq:Kac_Rice var}
\operatorname{Var}\left (\mathcal{Z}\right ) =\int\limits_{I\times I}\left ( K_2(t_1,t_2) - K_1(t_1)K_1(t_2) \right ) \, \mbox{d}t_1 \mbox{d} t_2 + \mathbb{E}\left [\mathcal{Z}\right ].
\end{equation}
As it was mentioned above, in our case the assumption that $
r(t_1,t_2)\ne \pm 1$ for all $t_1 \ne t_2 $ may not be satisfied,
and, as explained in \cite{RW2014}, it is easy to construct an
example of a curve where the Kac-Rice formula for the second
factorial moment \eqref{eq:Kac_Rice2} does not hold. To resolve this
situation we will divide the interval $I=[0,L]$ into small
subintervals, and note that the proof of ~\cite{AW} Theorem $6.3$
yields that if $J_{1},J_{2}\subseteq I$ are two {\em disjoint}
subintervals, then (recall that we denoted $\Zc_{J}$ to be the
number of zeros of $f$ on a subinterval $J\subseteq I$)
\begin{equation}
\label{eq:Kac-Rice covar disjoint} \mathbb E[\Zc_{J_{1}}\cdot
\Zc_{J_{2}}]=\int\limits_{J_{1}\times J_{2}}
K_{2}(t_{1},t_{2}) \, \mbox{d}t_{1}\mbox{d}t_{2},
\end{equation}
provided that for all $t_{1}\in J_{1}$, $t_{2}\in J_{2}$,
$$r(t_{1},t_{2})\ne \pm 1.$$
The $2$-point correlation function was evaluated ~\cite{RW2014}
explicitly to be
\begin{equation}
\label{eq:K2_explicit} K_{2}(t_{1},t_{2}) = \frac{1}{\pi^{2}
(1-r^{2})^{3/2}}\cdot \mu\cdot
(\sqrt{1-\rho^{2}}+\rho\arcsin{\rho}),
\end{equation}
where
\begin{equation}
\label{eq: mu def} \mu = \mu_{\zeigen}(t_{1},t_{2}) =
\sqrt{\alpha(1-r^{2})-r_{1}^{2}}\cdot \sqrt{\alpha
(1-r^{2})-r_{2}^{2}},
\end{equation}
and
\begin{equation}
\label{eq:rho def} \rho = \rho_{\zeigen}(t_{1},t_{2}) =
\frac{r_{12}(1-r^{2})+rr_{1}r_{2}}{\sqrt{\alpha (1-r^{2})-r_{1}^{2}}
\cdot \sqrt{\alpha (1-r^{2})-r_{2}^{2}}}
\end{equation}
(it follows from the derivation of \eqref{eq:K2_explicit} that
$|\rho|\le 1$).
\subsection{Proof of Proposition \ref{prop:approx Kac-Rice}}
\label{sec:var=sum covar}
Before giving the proof of Proposition \ref{prop:approx Kac-Rice} we
will have to do some preparatory work. To overcome the
above-mentioned obstacle we let $ c_0 $ be a sufficiently small
constant to be chosen below, and decompose the interval $[0,L]$ into
small intervals of length roughly $c_{0}\cdot
\frac{1}{\sqrt{\zeigen}}$ so that we can apply Kac-Rice on the
corresponding diagonal cubes. To be more concrete, let $ k = \lfloor
L \cdot \frac{\sqrt{\zeigen}}{c_0} \rfloor + 1 $ and $ \delta_0 =
\frac{L}{k} $, and divide the interval $ [0,L] $ into the
subintervals $ I_i = \left [ (i-1)\delta_0, i\delta_0 \right ] $
where $ i=1,\dots, k $. Note that $ \delta_0 \asymp
\frac{1}{\sqrt{\zeigen}}. $ With $ \mathcal{Z}_i $ denoting the
number of zeros of $ f $ on $ I_i $ $\left ( i=1,\dots , k\right )$
we have
\begin{equation}
\label{eq:var=sum cov ij} \operatorname{Var}(\Zc) =
\sum\limits_{i,j}{\rm{Cov}}(\Zc_{i},\Zc_{j}).
\end{equation}
Our first goal is to give an upper bound for the individual summands
in \eqref{eq:var=sum cov ij}; to this end we need the following
lemmas, whose proofs are postponed till \S~\ref{sec:aux lem
Tayl proof}:
\begin{lemma}
\label{lem:diag_cubes}
There exists a constant $ c_0 > 0 $ sufficiently small, such that \\
for all $ t_1\ne t_2 \in [0,L] $ with $ \left \vert t_2 - t_1 \right
\vert < c_0/\sqrt{\zeigen} $ we have $$ r(t_1,t_2)\ne\pm 1 .$$
\end{lemma}
\begin{lemma}[Uniform bound on the $2$-point correlation function around the diagonal]
\label{lem:ptwise_bnd} For all $0<|t_{2}-t_{1}|< c_{0}/
\sqrt{\zeigen}$ we have
\begin{equation*}
K_{2}(t_{1},t_{2}) = O(\zeigen).
\end{equation*}
\end{lemma}
\begin{corollary}
\label{col:cov_bound} We have
\begin{equation*}
{\rm{Cov}}(\Zc_{i},\Zc_{j}) = O(1),
\end{equation*}
uniformly for all $i,j$ and $\zeigen$ (the implied constant is
universal).
\end{corollary}
\begin{proof}[Proof of Corollary \ref{col:cov_bound} assuming lemmas \ref{lem:diag_cubes}-\ref{lem:ptwise_bnd}.]
By Lemma \ref{lem:diag_cubes}, $ r(t_1,t_2) \ne \pm 1 $ for all $ t_1 \neq t_2 $ in every
diagonal cube $ I_i^2. $ Hence (Theorem \ref{thm:Kac-Rice} applied
on the interval $I_{i}$ corresponding to a diagonal cube $I_i^2$)
we can apply Kac-Rice \eqref{eq:Kac_Rice var} to compute the
variance of $\Zc_{i}$:
\[\operatorname{Var}\left (\mathcal{Z}_i\right ) =\int\limits_{I_i^2}\left ( K_2(t_1,t_2) - K_1(t_1)K_1(t_2) \right ) \, \mbox{d}t_1 \mbox{d} t_2 + \mathbb{E}\left [\mathcal{Z}_i\right ].
\] By Kac-Rice we have \[ \mathbb{E}\left [ \mathcal{Z}_i \right ] = \int\limits_{I_i}K_1(t) \, \mbox{d}t = \delta_0 \cdot \frac{2}{\sqrt{d}}\sqrt{\zeigen}; \] using Lemma \ref{lem:ptwise_bnd} we conclude that $$ \operatorname{Var}\left (\mathcal{Z}_i\right ) \ll \zeigen\delta_0^2+\sqrt{\zeigen}\delta_0 \ll 1 .$$
This proves the statement of the corollary for $ i=j $; the result
for arbitrary $ i,j $ follows from the above and the Cauchy-Schwartz
inequality
\begin{equation*}
{\rm{Cov}}(\Zc_{i},\Zc_{j}) \le \sqrt{\operatorname{Var}\left (\mathcal{Z}_i\right
)\cdot \operatorname{Var}\left (\mathcal{Z}_j\right )}.
\end{equation*}
\end{proof}
\begin{definition}(Singular and nonsingular cubes.)
\label{def:singular}
\begin{enumerate}
\item Let
$$S_{ij} = I_{i}\times I_{j} = [i\delta_{0}, (i+1)\delta_{0}] \times [j\delta_{0}, (j+1)\delta_{0}]$$ be a cube in $[0,L]^{2}$.
We say that $S_{ij}$ is a {\em singular} if it contains a point
$(t_{1},t_{2})\in S_{ij}$ satisfying $$|r (t_{1},t_{2})| > 1/2. $$
\item The union of all the singular cubes is the singular set
$$B=B_{\zeigen} = \bigcup\limits_{S_{ij}\text{ singular}}S_{ij}.$$
\end{enumerate}
\end{definition}
Note that since $r/\sqrt{E}$ is a Lipschitz function with a
universal constant (independent of $ \zeigen $), if $S_{ij}$ is a
singular cube, then $$|r(t_1,t_2)|>1/4$$ everywhere on $S_{ij}$,
provided that $c_{0}$ is chosen sufficiently small. Using the above
it is easy to obtain the following bound on the number of singular
cubes:
\begin{lemma}\label{lem:3.7}
The number of singular cubes is bounded above by \[ E\cdot
\int\limits_{[0,L]^{2}} r^{2}(t_1,t_2) \, \mbox{d}t_1
\mbox{d}t_2.\]
\end{lemma}
\begin{proof}
Using the Chebyshev-Markov inequality, we see that \[ \operatorname{meas} \left (
B\right ) \ll \int\limits_{[0,L]^{2}} r^{2}(t_1,t_2) \, \mbox{d}t_1
\mbox{d}t_2.\] The statement of this lemma follows from the fact
that the volume of each cube is $ \asymp 1/\zeigen $.
\end{proof}
With Lemma \ref{lem:3.7} together with Corollary \ref{col:cov_bound}
it is easy to bound the contribution to \eqref{eq:var=sum cov ij} of
all $(i,j)$ corresponding to singular cubes $S_{ij}$ (i.e.
``singular contribution"), see the proof of Proposition
\ref{prop:approx Kac-Rice} below. Next we will deal with the
nonsingular contribution. Here the Taylor expansion of $K_{2}$ as a
function of $r$ and its scaled derivatives around
$r=r_{1}=r_{2}=r_{12}=0$ (up to the quadratic terms) is valid; it
will yield the following result, whose proof will be given postponed
in \S~\ref{sec:aux lem Tayl proof}.
\begin{lemma}
\label{lem:K2=O(r^2 der)} For $(t_{1},t_{2})$ outside the singular
set we have
\begin{equation}
\label{eq:K2 r2 bnd}
\left|K_{2}(t_{1},t_{2})-K_{1}(t_{1})K_{1}(t_{2}) \right| = E\cdot
O\left( r^{2}+(\frac{r_{1}}{\sqrt{\zeigen}})^2 +
(\frac{r_{2}}{\sqrt{\zeigen}})^{2}+(\frac{r_{12}}{\zeigen})^2
\right).
\end{equation}
\end{lemma}
We are finally in a position to give a proof to the main result of
this section.
\begin{proof}[Proof of Proposition \ref{prop:approx Kac-Rice}]
First, it is easy to bound the total contribution of the singular
set to \eqref{eq:var=sum cov ij} (i.e. all $i,j$ with $S_{ij}$
singular): Lemma~\ref{lem:3.7} and Corollary \ref{col:cov_bound}
imply that it is bounded by
\begin{equation}
\label{eq:nonsing contr} \sum\limits_{(i,j):\: S_{ij}\subseteq
B}{\rm{Cov}}(\Zc_{i},\Zc_{j})= O\left (\zeigen \cdot
\int\limits_{[0,L]^{2}} r^{2}(t_1,t_2) \, \mbox{d}t_1 \mbox{d}t_2
\right ).
\end{equation}
Next we deal with the indexes $(i,j)$ corresponding to nonsingular
$S_{ij}$.
We observe that, by the definition, for such a nonsingular cube
$S_{ij}$, necessarily for every $(t_{1},t_{2})\in S_{ij}$,
$$r(t_1,t_2)\ne \pm 1.$$ As this is a sufficient condition for the
application of Kac-Rice formula \eqref{eq:Kac-Rice covar disjoint}
for computation of ${\rm{Cov}}(\Zc_{i},\Zc_{j})$, bearing in mind
\eqref{eq:K2 r2 bnd} it yields that for $S_{ij}$ nonsingular (this
in particular implies $i\ne j$),
\begin{equation*}
\begin{split}
{\rm{Cov}}(\Zc_{i},&\Zc_{j}) = \int\limits_{S_{ij}}(K_{2}(t_{1},t_{2})-K_{1}(t_{1})K_{1}(t_{2})) \, \mbox{d}t_1\mbox{d}t_2 \\
&= O\left( E \cdot \int\limits_{S_{ij}} \left(
r^{2}+(r_{1}/\sqrt{\zeigen})^2 +
(r_{2}/\sqrt{\zeigen})^{2}+(r_{12}/\zeigen)^2 \right) \, \mbox{d}t_1
\mbox{d}t_2 \right).
\end{split}
\end{equation*}
Hence the total contribution of the nonsingular set to
\eqref{eq:var=sum cov ij} is $O(\zeigen \cdot \Rc_{2}(\zeigen))$. As
the total contribution of the singular set to \eqref{eq:var=sum cov
ij} was bounded in \eqref{eq:nonsing contr}, and obviously
$$\int\limits_{[0,L]^{2}} r^{2}(t_1,t_2) \, \mbox{d}t_1 \mbox{d}t_2 \le \Rc_{2}(\zeigen),$$
this concludes the proof of Proposition \ref{prop:approx Kac-Rice}.
\end{proof}
\subsection{Proofs of the auxiliary lemmas \ref{lem:diag_cubes}, \ref{lem:ptwise_bnd}
and \ref{lem:K2=O(r^2 der)}}
\label{sec:aux lem Tayl proof}
\begin{proof}[Proof of Lemma \ref{lem:diag_cubes}]
For $ t_1\in [0,L] $ fixed, we compute the Taylor expansion of $
r(t_1,t_2) $ around $ t_2=t_1 $. Recall that
\begin{equation*}
r(t_1,t_2) = r_F\left (\gamma(t_1)-\gamma(t_2)\right ) =
\dfrac{1}{N} \sum_{\mu\in \mathcal{E}}\cos\left (2\pi \left <\mu
,\gamma(t_1)- \gamma(t_2)\right > \right).
\end{equation*}
Thus, $ r(t_1,t_1)=1 $, $ r_2(t_1,t_1)=0 $ and $ r_{22}(t_1,t_1) =
\dot{\gamma}(t_1)^tH_{r_F}(0)\dot{\gamma}(t_1) = -\alpha $. Moreover, we
clearly have $ r_{222}(t_1,t_2) = O(E^{3/2}) $, and therefore \[
r(t_1,t_2) = 1 - \frac{\alpha}{2}(t_2-t_1)^2 + O\left ( \left (
\sqrt{E}(t_2-t_1) \right )^3\right ). \] Hence, for $ t_2-t_1 \ll
1/\sqrt{E} $ we have
\begin{equation} \label{eq:r_asymp}
\begin{split}
1-r^2(t_1,t_2) &= \alpha (t_2-t_1)^2 + O\left ( \left ( \sqrt{E}(t_2-t_1) \right )^3\right ) \\
&= \alpha(t_2-t_1)^2 \left ( 1+ O\left ( \sqrt{E}(t_2-t_1) \right
)\right ),
\end{split}
\end{equation}
so there is a constant $ c_0>0 $ sufficiently small, such that $
1-r^2(t_1,t_2) $ is strictly positive for $ 0<|t_2-t_1|<c_0 /
\sqrt{E} $.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:ptwise_bnd}]
The function \begin{equation} \label{eq:G_rho} G(\rho):=\frac{2}{\pi
} \left ( \sqrt{1-\rho^2}+\rho\arcsin{\rho} \right )
\end{equation} satisfies $ \frac{2}{\pi}\le G \le 1 $. Hence, by the explicit form \eqref{eq:K2_explicit} of the $ 2 $-point correlation function $ K_2 $ we obtain that
\[ K_2(t_1,t_2) \ll \sqrt{\dfrac{\left ( \alpha \left ( 1-r^2\right ) - r_1^2 \right )\left ( \alpha \left ( 1-r^2\right ) - r_2^2 \right )}{\left ( 1- r^2\right )^3}}. \]
For $ t_2 - t_1 \ll 1/\sqrt{E} $ we have
\[ r_1(t_1,t_2) = \alpha (t_2-t_1)\left ( 1+O\left ( \sqrt{E}(t_2-t_1) \right ) \right ), \]
\[ r_2(t_1,t_2) = -\alpha (t_2-t_1)\left ( 1+O\left ( \sqrt{E}(t_2-t_1) \right ) \right ). \] Using \eqref{eq:r_asymp} we get that
\begin{equation*}
\begin{split}
\alpha(1-r^2)-r_1^2 &= \alpha^2(t_2-t_1)^2 \left ( 1+O\left ( \sqrt{E}(t_2-t_1) \right ) \right ) \\
&- \alpha^2(t_2-t_1)^2 \left ( 1+O\left ( \sqrt{E}(t_2-t_1) \right )
\right ) = O\left ( E^{5/2}(t_2-t_1)^3 \right )
\end{split}
\end{equation*}
and likewise
\[ \alpha(1-r^2)-r_2^2 = O\left ( E^{5/2}(t_2-t_1)^3 \right ), \]
so
\[ K_2(t_1,t_2) \ll \dfrac{O\left ( E^{5/2}(t_2-t_1)^3 \right )}{\alpha^{3/2}(t_2-t_1)^3\left (1+O\left ( \sqrt{E}(t_2-t_1) \right )\right )} = O(E), \]assuming $ 0<|t_2-t_1|<c_0/\sqrt{E} $ for a sufficiently small constant $ c_0 > 0.$
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:K2=O(r^2 der)}]
Recall from \eqref{eq:K2_explicit} that
\begin{equation}
\label{eq:K2_def_2} K_2(t_1,t_2) = \frac{1}{2\pi}\cdot
\dfrac{1}{\left ( 1- r^2 \right )^{3/2}} \cdot G(\rho) \cdot \mu
\end{equation}
where $ G $, $ \mu $ and $ \rho $ are defined respectively in
\eqref{eq:G_rho}, \eqref{eq: mu def} and \eqref{eq:rho def}. Note
that for every $|\rho|\le 1 $,
\[ G(\rho) = \dfrac{2}{\pi} + O\left ( \rho ^2 \right). \]
For $ r $, $ r_1/\sqrt{E} $, $ r_2/\sqrt{E} $, $ r_{12}/E $ small,
we have
\begin{equation*}
\rho = O\left ( r + r_{12}/E \right ).
\end{equation*} Moreover, since $ |\rho| \le 1 $, this bound holds for every $ (t_1,t_2) $. Thus, for every $ (t_1,t_2) $ we have
\[ G(\rho) = \dfrac{2}{\pi}+ O \left ( r^2 + \left (r_{12}/E \right )^2 \right ). \] For every $(t_1,t_2)$,
\[ \mu = \alpha + E\cdot O \left (r^2 + (r_1/\sqrt{E})^2 + (r_2/\sqrt{E})^2 \right ), \] and for $ r $ bounded away from $ \pm 1$ we have \[ \dfrac{1}{\left ( 1- r^2\right )^{3/2}} = 1 + O\left (r^2\right ). \] Substituting all the expansions in \eqref{eq:K2_def_2}, we get that
\[ K_2(t_1,t_2) = \dfrac{\alpha}{\pi ^ 2} + E \cdot O\left ( r^{2}+(r_{1}/\sqrt{\zeigen})^2 + (r_{2}/\sqrt{\zeigen})^{2}+(r_{12}/\zeigen)^2 \right ). \]
The statement of the lemma now follows, recalling that by
\eqref{eq:K_1}, $ K_1(t) = \sqrt{ \alpha}/\pi$ for all $ t\in
[0,L]$.
\end{proof}
\section{Oscillatory integrals and curvature}
\label{sec:oscillary integrals}
In this section we investigate certain oscillatory integrals on
curves which arise in our work. A key role is played by the
differential geometry of the curve.
\subsection{Differential geometry of \texorpdfstring{$3$}{3}-dimensional curves}
For a smooth curve in ${\mathbb R}^3$, with arc-length parameterization
$\gamma:[0,L]\to \curve \subset {\mathbb R}^3$, so that $T(t) = \gamma'(t)$
is the unit tangent, the curvature of $\gamma$ at $\gamma(t)$ is
$\kappa(t) = ||\gamma''(t)||$. We assume that $\kappa(t)$ never
vanishes, so that $\gamma''(t)=\kappa(t) N(t)$ with $N(t)$ the unit
normal, and under the same assumption the {\em torsion} $\tau(t)$ is
$B'(t) = -\tau(t)N(t)$ where $B = T\times N$ is the binormal vector.
The orthonormal basis $\left (T,N,B\right ) $ is called the
Frenet-Serret frame of the curve. Recall the Frenet-Serret formulas
\begin{equation*}
\begin{matrix}
T'(t) = & & \kappa(t) N(t)& \\ \\
N'(t) =& -\kappa(t)T(t)& &+\tau(t)B(t)\\ \\
B'(t) = &&-\tau(t) N(t) &
\end{matrix}
\end{equation*}
so in particular
$$
T''\left(t\right)=
\kappa'\left(t\right)N\left(t\right)-\kappa^{2}\left(t\right)T\left(t\right)+\kappa\left(t\right)\tau\left(t\right)B\left(t\right).
$$
Let $ K_{\min}$ and $K_{\max}$ the minimal and the maximal curvature of $\mathcal{C} $ respectively.
Since the curvature is assumed to be nowhere vanishing, we have
$$0<K_{\min}\le\kappa\left(t\right)\le K_{\max}.$$
\subsection{Oscillatory integrals}
Recall the classical form of Van der Corput Lemma: let
$\left[a,b\right]$ be a finite interval, $\phi\in
C^{\infty}\left[a,b\right]$ a smooth and real valued phase function,
and $A\in C^{\infty}\left[a,b\right]$ a smooth amplitude. For
$\lambda>0$ define the oscillatory integral
\[
I\left(\lambda\right):=\int_{a}^{b}A\left(t\right)e^{i\lambda\phi\left(t\right)}\mbox{d}t.
\]
\begin{lemma}
\label{lem:Van_Der_Corp}(Van der Corput) For $k\ge2$, if
$\left|\phi^{\left(k\right)}\right|\ge1$, then, as
$\lambda\rightarrow\infty$,
\[
\left|I\left(\lambda\right)\right|\ll\frac{1}{\lambda^{1/k}}\left(\left\Vert
A\right\Vert _{\infty}+\left\Vert A'\right\Vert _{1}\right).
\]
If $\left|\phi'\right|\ge1$ and $\phi'$ is monotone, then
\[
\left|I\left(\lambda\right)\right|\ll\frac{1}{\lambda}\left(\left\Vert
A\right\Vert _{\infty}+\left\Vert A'\right\Vert _{1}\right).
\]
The implied constants are absolute.\end{lemma}
\begin{remark}
If $\left|\phi'\right|\ge 1$ then, independent of the monotonicity
hypothesis on $\phi'$,
\[
\left|I\left(\lambda\right)\right|\ll\frac{b-a+2}{\lambda}\left(\left\Vert
A\right\Vert _{\infty}+\left\Vert A'\right\Vert _{1}\right).
\]
\end{remark}
\subsection{Curves with nowhere vanishing torsion}
\label{sec:nonzero_torsion} Assume that the curve $ \curve $ has
nowhere vanishing torsion, so that
$0<T_{\min}\le|\tau\left(t\right)|\le T_{\max}$, where $ T_{\min}$ and
$T_{\max}$ are the minimal and maximal absolute value of the torsion of $\mathcal{C}$
respectively. Consider a unit vector $\xi\in S^2$, and the phase
function
\begin{equation}
\label{eq:phi phase def} \phi_{\xi}\left(t\right):=\left\langle
\xi,\gamma\left(t\right)\right\rangle
\end{equation}
for $t\in \left [0,L\right ]$. We define an oscillatory integral
\[
I\left(\lambda,\xi\right):=\int_{0}^{L}A\left (t\right
)e^{i\lambda\phi_{\xi}\left(t\right)}\mbox{d}t.
\]
We apply Lemma \ref{lem:Van_Der_Corp}
to give an upper bound (uniform in $\xi$) for
$I\left(\lambda,\xi\right)$:
\begin{proposition}\label{stationary phase 3d}
Let $\curve$ be a smooth curve with nowhere vanishing curvature and
torsion. Then
$$I\left(\lambda,\xi\right)\ll_{\curve}\frac{1}{\lambda^{1/3}}\left ( \left \Vert A\right\Vert_\infty + \left \Vert A' \right\Vert _1 \right ).$$
\end{proposition}
\begin{proof}
We have
\begin{equation}
\label{eq:phi'=<xi,T>} \phi'_{\xi}\left(t\right)=\left\langle
\xi,T\left(t\right)\right\rangle,
\end{equation}
\begin{equation}
\label{eq:phi''=k<xi,N>}
\phi_{\xi}''\left(t\right)=\kappa\left(t\right)\left\langle
\xi,N\left(t\right)\right\rangle,
\end{equation}
and
\[
\phi_{\xi}'''\left(t\right)=\kappa'\left(t\right)\left\langle
\xi,N\left(t\right)\right\rangle
-\kappa^{2}\left(t\right)\left\langle
\xi,T\left(t\right)\right\rangle
+\kappa\left(t\right)\tau\left(t\right)\left\langle
\xi,B\left(t\right)\right\rangle .
\]
Since $\left(T,N,B\right)$ is an orthonormal basis for
$\mathbb{R}^{3}$, we know that
\begin{equation}
\label{eq:1=T,N,B decomp}
\begin{split}
1 & =\left|\xi\right|^{2}=\left|\left\langle
\xi,T\left(t\right)\right\rangle \right|^{2}+\left|\left\langle
\xi,N\left(t\right)\right\rangle \right|^{2}+\left|\left\langle
\xi,B\left(t\right)\right\rangle \right|^{2}.
\end{split}
\end{equation}
Now let
\[
c=\min\left(\frac{1}{3},\frac{K_{\min}^{2}T_{\min}^{2}}{48\left\Vert
\kappa'\right\Vert
_{\infty}^{2}},\frac{K_{\min}^{2}T_{\min}^{2}}{48K_{\max}^{4}}\right)
\]
(if $\left\Vert \kappa'\right\Vert _{\infty}=0$ omit the middle
term). If $\left|\left\langle \xi,T\left(t\right)\right\rangle
\right|^{2}\ge c$, then
$\left|\phi_{\xi}'\left(t\right)\right|\ge\sqrt{c}$. If
$\left|\left\langle \xi,N\left(t\right)\right\rangle \right|^{2}\ge
c$, then
$\left|\phi_{\xi}''\left(t\right)\right|\ge\sqrt{c}K_{\min}$.
Otherwise, i.e. if both $\left|\left\langle
\xi,T\left(t\right)\right\rangle \right|^{2}<c$ and
$\left|\left\langle \xi,N\left(t\right)\right\rangle \right|^{2}<c$,
then necessarily $\left|\left\langle
\xi,B\left(t\right)\right\rangle \right|^{2}\ge\frac{1}{3}$, so
\[
\left|\phi_{\xi}'''\left(t\right)\right|\ge\frac{K_{\min}T_{\min}}{\sqrt{3}}-K_{\max}^{2}\sqrt{c}-\left\Vert
\kappa'\right\Vert
_{\infty}\sqrt{c}\ge\frac{K_{\min}T_{\min}}{2\sqrt{3}}.
\]
Note that $\left\Vert \phi_{\xi}'\right\Vert _{\infty}\le1$,
$\left\Vert \phi_{\xi}''\right\Vert _{\infty}\le K_{\max}$,
\[
\left\Vert \phi_{\xi}'''\right\Vert _{\infty}\le\left(\left\Vert
\kappa'\right\Vert
_{\infty}^{2}+K_{\max}^{4}+K_{\max}^{2}T_{\max}^{2}\right)^{1/2}.
\]
Using again the Frenet-Serret formulas, we can also get an upper
bound in the same fashion for the fourth derivative, say
\[
\left\Vert \phi_{\xi}^{\left(4\right)}\right\Vert_\infty \le
C=C\left(K_{\max},T_{\max},\left\Vert \kappa'\right\Vert
_{\infty},\left\Vert \kappa''\right\Vert _{\infty},\left\Vert
\tau'\right\Vert _{\infty}\right).
\]
Assume now that
$\left|\phi_{\xi}'\left(t_{0}\right)\right|\ge\sqrt{c}$ for some
$t_{0}\in\left[0,L\right]$. Then for every $t$ such that
$\left|t-t_{0}\right|\le\frac{\sqrt{c}}{2K_{\max}}$ we have
\[
\sqrt{c}-\left|\phi_{\xi}'\left(t\right)\right|\le\left|\phi_{\xi}'\left(t_{0}\right)\right|-\left|\phi_{\xi}'\left(t\right)\right|\le\left|\phi_{\xi}'\left(t\right)-\phi_{\xi}'\left(t_{0}\right)\right|\le\left|t-t_{0}\right|\left\Vert
\phi_{\xi}''\right\Vert _{\infty}\le\frac{\sqrt{c}}{2}
\]
so $\left|\phi_{\xi}'\left(t\right)\right|\ge\frac{\sqrt{c}}{2}$.
Similarly, if
$\left|\phi_{\xi}''\left(t_{0}\right)\right|\ge\sqrt{c}K_{\min}$ or
$\left|\phi_{\xi}'''\left(t_{0}\right)\right|\ge\frac{K_{\min}T_{\min}}{2\sqrt{3}}$,
then $\phi_{\xi}''$ or $\phi_{\xi}'''$ is bounded away from zero on
some interval around $t_{0}$, with length independent of $\xi$.
Hence the interval $\left[0,L\right]$ may be divided into a finite,
independent of $\xi$, number of subintervals, such that for every
$\xi$ either $\phi_{\xi}'$, or $\phi_{\xi}''$, or $\phi_{\xi}'''$ is
bounded away from zero on each of the subintervals. We conclude the
proof of the proposition by an application of Lemma
\ref{lem:Van_Der_Corp} and the remark following it.
\end{proof}
\subsection{Real analytic curves}
Assume now that $\mathcal{C}$ is a real analytic, non-planar curve
with nowhere zero curvature. Then the torsion of $\mathcal{C}$ has
finitely many zeros, each of them is of finite order. We have
already treated the case when the torsion is nowhere zero. Assume
now, without loss of generality, that there is exactly one point
$t_{0}\in\left[0,L\right]$ with zero torsion of order $m\ge1$,
namely
\begin{equation}\label{eq:torsion_zero}
\tau\left(t_{0}\right)=\dots=\tau^{\left(m-1\right)}\left(t_{0}\right)=0,\hspace{1em}\tau^{\left(m\right)}\left(t_{0}\right)\ne
0.
\end{equation}
Recall that under the notation of the previous section \[
I\left(\lambda,\xi\right):=\int_{0}^{L}A\left (t\right
)e^{i\lambda\phi_{\xi}\left(t\right)}\mbox{d}t.
\]
We prove the following result.
\begin{proposition}
\label{prop:non_plane_analytic} Let $\curve$ be a non-planar real
analytic curve with nowhere zero curvature, which has exactly one
point with zero torsion of order $ m \ge 1 $. Then
$$I\left(\lambda,\xi\right)\ll_{\curve}\frac{1}{\lambda^{1/(m+3)}}\left ( \left \Vert A\right\Vert_\infty + \left \Vert A' \right\Vert _1 \right ).$$\end{proposition}
\begin{proof}
Using the Frenet-Serret formulas and (\ref{eq:torsion_zero}), we get
that
\begin{alignat*}{1}
\phi_{\xi}^{\left(m+3\right)}\left(t_{0}\right) & =P\left(t_{0}\right)\left\langle \xi,T\left(t_{0}\right)\right\rangle +Q\left(t_{0}\right)\left\langle \xi,N\left(t_{0}\right)\right\rangle \\
& +\kappa\left(t_{0}\right)\tau^{\left(m\right)}\left(t_{0}\right)\left\langle \xi,B\left(t_{0}\right)\right\rangle
\end{alignat*}
where $P,Q$ are polynomials in
$\kappa,\kappa',\dots,\kappa^{\left(m+1\right)}$.
Choose
\[
c=\min\left(\frac{1}{3},\frac{\kappa\left(t_{0}\right)^{2}\left(\tau^{\left(m\right)}\left(t_{0}\right)\right)^{2}}{48P\left(t_{0}\right)^{2}},\frac{\kappa\left(t_{0}\right)^{2}\left(\tau^{\left(m\right)}\left(t_{0}\right)\right)^{2}}{48Q\left(t_{0}\right)^{2}}\right)
\]
(if $P\left(t_{0}\right)=0$ or $Q\left(t_{0}\right)=0$, omit the
corresponding terms). As in the proof of Proposition \ref{stationary
phase 3d}, by the orthonormality of $\left(T,N,B\right)$, either
$\left|\phi_{\xi}'\left(t_{0}\right)\right|\ge\sqrt{c}$,
$\left|\phi_{\xi}''\left(t_{0}\right)\right|\ge\kappa\left(t_{0}\right)\sqrt{c}$,
or (if both
$\phi_{\xi}'\left(t_{0}\right),\phi_{\xi}''\left(t_{0}\right)$ are
small) $\left|\left\langle \xi,B\left(t_{0}\right)\right\rangle
\right|^{2}\ge\frac{1}{3}$. Hence
\begin{alignat*}{1}
\left|\phi_{\xi}^{\left(m+3\right)}\left(t_{0}\right)\right| & \ge\frac{1}{\sqrt{3}}\kappa\left(t_{0}\right)|\tau^{\left(m\right)}\left(t_{0}\right)|-P\left(t_{0}\right)\sqrt{c}-Q\left(t_{0}\right)\sqrt{c}\\
& \ge\frac{1}{2\sqrt{3}}\kappa\left(t_{0}\right)|\tau^{\left(m\right)}\left(t_{0}\right)|.
\end{alignat*}
Since all the derivatives of $\phi_{\xi}$ are bounded from above,
uniformly w.r.t. $\xi$, we conclude that either the first, the
second or the $(m+3)$-th derivative of $\phi_{\xi}$ is bounded away
from zero on an interval around $t_{0}$ of length independent of
$\xi$. Outside that interval the torsion doesn't vanish, so that in
a neighborhood (of length independent of $\xi)$ around any point
outside this interval, either the first, or the second, or the third
derivative of $\phi_{\xi}$ is bounded away from zero. Dividing the
interval $\left[0,L\right]$ to a finite number (independent of
$\xi$) of subintervals, and applying Lemma \ref{lem:Van_Der_Corp} to
each of the subintervals, we finally deduce the statement of
Proposition \ref{prop:non_plane_analytic}.
\end{proof}
\section{Background on sums of three squares}
\label{sec:sums of 3 squares}
A positive integer $\zeigen$ is a sum of three squares if and only
if $\zeigen\neq 4^a(8b+7)$. Let $ \vE(\zeigen)$ be the set of
solutions
\begin{equation*}
\vE(\zeigen) = \{\vec x\in {\mathbb Z}^3: |\vec x|^2=\zeigen\}
\end{equation*}
and denote by $N=N_\zeigen$ the number of solutions
\begin{equation*}
N= N_\zeigen:=\#\vE(\zeigen) \;.
\end{equation*}
Gauss' formula expresses $N_\zeigen$ in terms of class numbers. For
$\zeigen$ square-free, it says that
\begin{equation*}\label{Gauss formula}
\#\vE(\zeigen) = \frac{24
h(d_\zeigen)}{w_\zeigen}\left(1-\leg{d_\zeigen}{2}\right)
\end{equation*}
where $d_\zeigen$, $h(d_\zeigen$) and $w_\zeigen$ are the
discriminant, class number and the number of units in the quadratic
field ${\mathbb Q}(\sqrt{-\zeigen})$. Using Dirichlet's class number formula,
one may then express $\# \vE(\zeigen)$ by means of the special value
$L(1,\chi_{d_\zeigen})$ of the associated quadratic $L$-function: If
$\zeigen\not\equiv 7\bmod 8$ is square-free then
\begin{equation*}\label{N in terms of L(1,chi)}
N_\zeigen=c_\zeigen\sqrt{\zeigen}\cdot L(1,\chi_{d_\zeigen})
\end{equation*}
where $c_\zeigen$ only depends on the remainder of $\zeigen$ modulo
$8$. We may bound the number $\#\vE(\zeigen)$ of such points as
\begin{equation*}
\#\vE(\zeigen)\ll\ \zeigen^{1/2+\epsilon}
\end{equation*}
for all $\epsilon>0$.
The existence of a {\em primitive} lattice point (i.e. $\vec
x=(x_1,x_2,x_3)$ with $\gcd(x_1,x_2,x_3)=1$ and $\|x\|^{2}=\zeigen$)
is equivalent to $\zeigen\not\equiv 0, 4,7 \bmod 8$. If it is indeed
the case, then Siegel's theorem yields a lower bound
\begin{equation}\label{Siegel thm}
\#\vE(\zeigen)\gg \zeigen^{1/2-\epsilon}.
\end{equation}
A fundamental result conjectured by Linnik (established by himself
under the Generalized Riemann Hypothesis), is that for
$\zeigen\not\equiv 0,4,7 \bmod 8$, the points
\begin{equation*}
\^\vE(\zeigen):= \frac 1{\sqrt{\zeigen}}\vE(\zeigen) \subset S^2
\end{equation*}
obtained by projecting to the unit sphere, become equidistributed on
the unit sphere with respect to the normalized Lebesgue measure as
$\zeigen\to \infty$. This was proved unconditionally by Duke
\cite{Duke}, and Golubeva and Fomenko \cite{GF}.
The ``Riesz $s$-energy" of $N$ points $x_1,\dots, x_N$ on $S^2$ is
defined as
\begin{equation}\label{Def of E_s}
E_s(x_1,\dots, x_N):= \sum_{i\neq j} \frac 1{|x_i-x_j|^s}.
\end{equation}
A forthcoming result of Bourgain, Rudnick and Sarnak \cite{BRSsphere} (announced in \cite{BRS} for the electrostatic case $s=1$)
yields a precise asymptotic expression for $E_s(\^\vE(\zeigen))$: for every $0<s<2$ if $\zeigen\to \infty$ such that $\zeigen\neq 0,4,7 \mod 8$,
then there exists some $\delta>0$ so that
\begin{equation}
\label{eq:extremal energy}
E_s(\^\vE(\zeigen)) = I(s) N^2 +O(N^{2-\delta})
\end{equation}
with
\begin{equation*}
I(s)=\int_{S^2} \frac 1{|x - x_0|^s} d\sigma(x) = \frac {2^{1-s}}{2-s},
\end{equation*}
with $x_0\in S^2$ any point on the sphere, and $d\sigma$ the Lebesgue measure, normalized to have unit area
As the details of \eqref{eq:extremal energy} have not appeared at the time of writing, we will prove the following simple bound, which suffices for Theorem \ref{thmvar}:
\begin{proposition}
\label{prop:bd on Riesz2}
Fix $0<s\leq 1$. Then for $\zeigen\not\equiv 0, 4,7 \bmod 8$,
\begin{equation*}
E_s(\^\vE(\zeigen)) \ll N^2\zeigen^\eta,\quad \forall \eta>0\;.
\end{equation*}
\end{proposition}
The proof of Proposition~\ref{prop:bd on Riesz2} will be given in Appendix~\ref{sec:Appendix}.
\section{The second moment of \texorpdfstring{$r$}{r} and its derivatives}
\label{sec:second_moment}
We wish to bound the second moment of the covariance function $r$
and its derivatives. It is here that we need the full arithmetic
input described in \S~\ref{sec:sums of 3 squares}. Recall that
\begin{equation}
\begin{split} \label{eq:cov_def_2}
r(t_1,t_2) &= r_F\left (\gamma(t_1)-\gamma(t_2)\right ) = \dfrac{1}{N} \sum_{\mu\in \mathcal{E}}\cos\left (2\pi \left <\mu ,\gamma(t_1)- \gamma(t_2)\right >\right ) \\
&= \dfrac{1}{N} \sum_{\mu\in \mathcal{E}}e\left (\left <\mu
,\gamma(t_1)- \gamma(t_2)\right >\right ).
\end{split}
\end{equation}
\subsection{Non-planar curves}
Recall that given $\zeigen$ we defined $\Rc_{2}(\zeigen)$ as in
\eqref{eq:r_2}. Proposition \ref{prop:approx Kac-Rice} shows that in
order to bound the nodal intersections variance from above it is
sufficient to bound $\Rc_{2}(\zeigen)$, which is claimed in the
following proposition for the non-planar case.
\begin{proposition}
\label{prop:R2<<1/E^1/3} Assume that the curve $\curve$ is smooth,
with nowhere zero curvature and torsion. Then for every $\eta>0$ we have
\begin{equation}
\label{eq:Rc2<<1/E^1/3}
\Rc_2(\zeigen) \ll \dfrac{1}{E^{1/3-\eta}}.
\end{equation}
\end{proposition}
\begin{remark}
For real-analytic non-planar curves with non-vanishing curvature,
the same argument as below, invoking Proposition
\ref{prop:non_plane_analytic} instead of Proposition \ref{stationary
phase 3d}, yields
\begin{equation}
\label{eq:Rc2<<1/E^delta analytic} \Rc_2 (\zeigen)\ll 1/E^\delta
\end{equation}
for some $\delta=\delta(\curve)>0$.
\end{remark}
\begin{proof}
In what follows we will establish the following bounds on the $2$nd
moment of $r$ and some of its normalized derivatives along $\curve$: for all $\eta>0$
we have
\begin{equation}
\label{eq:cov_equation}
\iint \limits_{[0,L]^2}
r(t_1,t_2)^2\mbox{d}t_1\mbox{d}t_2 \ll \frac 1{\zeigen^{1/3-\eta}},
\end{equation}
\begin{equation}\label{eq:non_planar_ri}
\iint \limits_{[0,L]^2} \left
(r_i(t_1,t_2)/\sqrt{E}\right )^2\mbox{d}t_1\mbox{d}t_2 \ll \frac
1{\zeigen^{1/3-\eta}} \hspace{10pt} (i=1,2),
\end{equation}
and
\begin{equation}
\label{eq:non_planar_r12} \iint \limits_{[0,L]^2} \left
(r_{12}(t_1,t_2)/E\right )^2\mbox{d}t_1\mbox{d}t_2 \ll \frac
1{\zeigen^{1/3-\eta}}.
\end{equation}
The statement \eqref{eq:Rc2<<1/E^1/3} of Proposition
\ref{prop:R2<<1/E^1/3} will follow at once upon substituting
\eqref{eq:cov_equation}, \eqref{eq:non_planar_ri} and
\eqref{eq:non_planar_r12} into the definition \eqref{eq:r_2} of
$\Rc_{2}(\zeigen)$.
First we show \eqref{eq:cov_equation}. Squaring out and integrating
\eqref{eq:cov_def_2}, we find
\begin{equation*}
\begin{split}
\iint \limits_{[0,L]^2} r(t_1,t_2)^2 \, \mbox{d}t_1\mbox{d}t_2 &=
\frac 1{N^2}\sum_{\mu\in \mathcal{E}} \sum_{\mu'\in \mathcal{E}} \,
\, \iint \limits_{[0,L]^2} e(\langle
\mu-\mu',\gamma(t_1)-\gamma(t_2) \rangle) \, \mbox{d}t_1 \mbox{d}t_2 \\
&= \frac{L^2}{N} + \frac 1{N^2} \sum_{ \mu \neq \mu'} \left|
\int_0^L e(\langle\mu-\mu',\gamma(t) \rangle) \, \mbox{d}t
\right|^2.
\end{split}
\end{equation*}
Since $\gamma$ has nowhere vanishing curvature and torsion, we
deduce from Proposition~\ref{stationary phase 3d} that
\begin{equation*}
\int_0^L e(\langle\mu-\mu',\gamma(t) \rangle) \, \mbox{d}t \ll \frac 1{|\mu-\mu'|^{1/3}},
\end{equation*}
which yields
\begin{equation*}
\iint \limits_{[0,L]^2} r(t_1,t_2)^2 \, \mbox{d}t_1\mbox{d}t_2 =
\frac {L^2}{N} + O\left(\frac 1{N^2} \sum_{\mu\neq \mu'} \frac
1{|\mu-\mu'|^{2/3}}\right).
\end{equation*}
The summation inside the error term $O(\cdots)$ is $1/E^{1/3}$ times
the ``Riesz $2/3$-energy" of the set of projected lattice points
$\^\vE(E)=\frac 1{\sqrt{E} }\vE(E)\subset S^2$. By
Proposition \ref{prop:bd on Riesz2},
\begin{equation*}
\sum_{ \substack{\^\mu,\^\mu' \in \^\vE\\ \^\mu\neq \^\mu'}} \frac
1{|\^\mu-\^\mu'|^{2/3}} \ll N^2\cdot E^{\eta},\forall\eta>0,
\end{equation*}
and hence
\begin{equation*}
\frac 1{N^2} \sum_{\mu\neq \mu'} \frac 1{|\mu-\mu'|^{2/3}} = \frac
1{\zeigen^{1/3}} \frac 1{N^2}\sum_{ \substack{\^\mu,\^\mu' \in \^\vE\\
\^\mu\neq \^\mu'}} \frac 1{|\^\mu-\^\mu'|^{2/3}}\ll \frac{1}{\zeigen^{1/3-\eta}},
\end{equation*}
which yields \eqref{eq:cov_equation}.
Now we turn to proving \eqref{eq:non_planar_ri}. We have
\[ \dfrac{1}{2\pi i \sqrt{E}} r_1(t_1,t_2) = \dfrac{1}{N} \sum_{\mu} \left < \dfrac{\mu}{\left \vert \mu \right \vert}, \dot{\gamma}(t_1) \right> e^{2\pi i \left < \mu , \gamma (t_1) - \gamma (t_2) \right > }. \]
Denote
\[ A_{\mu,\mu'}(t) = \left< \dfrac{\mu}{\left \vert \mu \right \vert}, \dot{\gamma}(t)\right> \left< \dfrac{\mu'}{\left \vert \mu' \right \vert}, \dot{\gamma}(t)\right>. \]
Then
\begin{equation} \label{eq:ri_expansion}
\begin{split}
\iint \limits_{[0,L]^2} & \left \vert \dfrac{1}{2\pi \sqrt{E}}
r_1(t_1,t_2) \, \mbox{d}t_1\mbox{d}t_2 \right\vert ^2 =
\dfrac{L}{N^2} \sum_{\mu} \int_0^L A_{\mu,\mu}(t_1) \, \mbox{d}t_1 \\
&+ \dfrac{1}{N^2}\sum_{\substack{\mu,\mu'\in\vE \\
\mu\ne\mu'}}\int_0^L A_{\mu,\mu'}(t_1)e^{2\pi i \left
<\mu-\mu',\gamma(t_1)\right >} \, \mbox{d}t_1 \int_0^L e^{2\pi i
\left <\mu' - \mu,\gamma(t_2)\right >} \, \mbox{d}t_2.
\end{split}
\end{equation}
To evaluate the main term, we use the fact (see \cite[Lemma
2.3]{RW}) that for every $ v\in \mathbb R ^ d $, \[
\dfrac{1}{N}\sum_{\mu\in\vE}\left <\mu,v\right >^2 =
\dfrac{\zeigen}{d} \left \Vert v \right\Vert ^2.
\] Hence, \[ \dfrac{L}{N^2} \sum_{\mu} \int_0^L A_{\mu,\mu}(t_1)\,
\mbox{d}t_1 = \dfrac{L^2}{dN}. \] As for the off-diagonal terms,
note that $ \left\Vert A_{\mu,\mu'} \right\Vert_\infty \le 1$, $
\left\Vert A'_{\mu,\mu'} \right\Vert_\infty \le 2K_{\max}$, so by
Proposition \ref{stationary phase 3d} each of the two last integrals
in \eqref{eq:ri_expansion} is bounded above by $
1/|\mu-\mu'|^{1/3}$. From here we continue as in the proof of
\eqref{eq:cov_equation} to obtain the estimate
\[ \iint \limits_{[0,L]^2} \left (r_1(t_1,t_2)/\sqrt{E}\right )^2 \, \mbox{d}t_1\mbox{d}t_2 \ll \frac 1{\zeigen^{1/3-\eta}},\,\forall\eta>0 \]
and a similar proof yields the same bound for the (normalized)
second moment of $ r_2 $.
Finally we turn to proving \eqref{eq:non_planar_r12}. We have
\begin{equation*}
\begin{split}
\iint \limits_{[0,L]^2} & \left \vert \dfrac{1}{4\pi^2 \zeigen}
r_{12}(t_1,t_2) \, \mbox{d}t_1\mbox{d}t_2 \right\vert ^2 =
\dfrac{1}{N^2} \sum_{\mu} \iint \limits_{[0,L]^2} A_{\mu,\mu}(t_1)A_{\mu,\mu}(t_2) \, \mbox{d}t_1\mbox{d}t_2 \\
&+ \dfrac{1}{N^2}\sum_{\substack{\mu,\mu'\in\vE \\
\mu\ne\mu'}}\left\vert \int_0^L A_{\mu,\mu'}(t)e^{2\pi i \left
<\mu-\mu',\gamma(t)\right >} \right\vert ^2 \, \mbox{d}t,
\end{split}.
\end{equation*}
The diagonal term is bounded above by $ L^2/N $; by Proposition
\ref{stationary phase 3d} the off-diagonal terms are bounded above
by $ 1/|\mu-\mu'|^{1/3}$, so we similarly deduce
\eqref{eq:non_planar_r12}.
\end{proof}
\subsection{Planar curves}
The goal of this section is proving the following estimate on
$\Rc_{2}$ for $\curve$ planar.
\begin{proposition}
\label{prop:R2<<1/E^1/4-o(1) planar} Assume that $\curve$ is a
smooth planar curve, with nowhere zero curvature. Then for all $ \eta>0 $
\begin{equation}
\label{eq:R2<<1/E^1/4-o(1) planar} \Rc_2 (\zeigen )\ll
\dfrac{1}{E^{1/4-\eta}}.
\end{equation}
\end{proposition}
We will now collect a few results needed for Proposition
\ref{prop:R2<<1/E^1/4-o(1) planar}, whose proof is given towards the
end of this section. In this section we assume that $\mathcal{C}$ is
a smooth planar curve with nowhere zero curvature, so that
$\tau\equiv0;$ the binormal vector $B$ is constant in this case. Let
$\epsilon=\epsilon\left(E\right)$ be a small parameter,
$\mu\ne\mu'\in\mathcal{E}\left(E\right)$,
$\lambda=\lambda\left(\mu,\mu'\right)=\left|\mu-\mu'\right|$ and
$\xi=\xi\left(\mu,\mu'\right)=\frac{\mu-\mu'}{\left|\mu-\mu'\right|}$.
We will reuse the definition \eqref{eq:phi phase def} of
$\phi_{\xi}$.
\begin{lemma}
\label{lem:small der 2st 2nd} For all $ \xi \in S^2 $ and for all $t\in [0,L]$ either
$\left|\phi_{\xi}'\left(t\right)\right|\ge\sqrt{\epsilon}$, or
$\left|\phi_{\xi}''\left(t\right)\right|\ge\sqrt{\epsilon}K_{\min}$,
or otherwise
$$
\left|\left\langle \xi,B\right\rangle \right|^{2}>1-2\epsilon.
$$
\end{lemma}
\begin{proof}
This follows from \eqref{eq:1=T,N,B decomp} via
\eqref{eq:phi'=<xi,T>} and \eqref{eq:phi''=k<xi,N>}.
\end{proof}
\begin{lemma}
\label{lem:Geom_Lemma} Let
$\mu,\mu_{1},\mu_{2}$ be distinct points on the sphere $ \sqrt{E}S^2 $, and
assume that \[\left\langle
\frac{\mu_{i}-\mu}{\left|\mu_{i}-\mu\right|},B\right\rangle
>\sqrt{1-2\epsilon}\] for $i=1,2$. Then
$\left|\mu_{2}-\mu_{1}\right|\le 16\sqrt{\epsilon E}.$
\end{lemma}
\begin{proof}
Let $v_{i}=\frac{\mu_{i}-\mu}{\left|\mu_{i}-\mu\right|}-B$
$\left(i=1,2\right)$, so that
$\left|v_{i}\right|^{2}=2-2\left\langle
\frac{\mu_{i}-\mu}{\left|\mu_{i}-\mu\right|},B\right\rangle
\le 4\epsilon$. We write
\[
\mu_{i}=\mu+\left|\mu_{i}-\mu\right|\left(B+v_{i}\right).
\]
Taking norms we get
\[
0=\left|\mu_{i}-\mu\right|^{2}+2\left|\mu_{i}-\mu\right|\left\langle
\mu,B+v_{i}\right\rangle,
\]
so that
\[
\left|\mu_{i}-\mu\right| =-2\left\langle \mu,B\right\rangle -2\left\langle \mu,v_{i}\right\rangle,
\]
and therefore
\[
\mu_{i}=\mu-2\left\langle \mu,B\right\rangle
B-2\left\langle \mu,v_{i}\right\rangle (B + v_i)-2 \left\langle \mu,B\right\rangle v_{i}.
\]
By Cauchy-Schwartz $\left\langle \mu,v_{i}\right\rangle
\le 2\sqrt{\epsilon E}$, and
$
\left\langle \mu,B\right\rangle \le\sqrt{ E},
$
and that implies $$\left|\mu_{2}-\mu_{1}\right|\le 16\sqrt{\epsilon
E}.$$
\end{proof}
We are now in a position to prove Proposition
\ref{prop:R2<<1/E^1/4-o(1) planar}.
\begin{proof}[Proof of Proposition \ref{prop:R2<<1/E^1/4-o(1) planar}.]
In what follows we will establish the following \\
bounds on the $2$nd
moment of $r$ and some of its normalized derivatives along $\curve$
(assumed to be planar, with nowhere zero curvature):
\begin{equation}
\label{eq:planar_r} \iint \limits_{[0,L]^2} r(t_1,t_2)^2 \,
\mbox{d}t_1\mbox{d}t_2=O\left(\frac{1}{E^{1/4-\eta }}\right),
\end{equation}
\begin{equation}
\label{eq:planar_ri} \iint \limits_{[0,L]^2} \left
(r_i(t_1,t_2)/\sqrt{E}\right )^2\mbox{d}t_1\mbox{d}t_2 = O\left(\frac{1}{E^{1/4-\eta}}\right),
\end{equation}
$i=1,2$, and
\begin{equation}
\label{eq:planar_r12} \iint \limits_{[0,L]^2} \left
(r_{12}(t_1,t_2)/E\right )^2\mbox{d}t_1\mbox{d}t_2 =O\left(\frac{1}{E^{1/4-\eta}}\right).
\end{equation}
First we prove \eqref{eq:planar_r}. If $B$ is the constant binormal
to the curve, then
\begin{equation}
\label{eq:int r^2 << sum I^2}
\begin{split}
\iint \limits_{[0,L]^2} r(t_1,t_2)^2\mbox{d}t_1\mbox{d}t_2= &
\frac{L^2}{N}+\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle\xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}\le 1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}\\
+ &\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle \xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}>1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}
\end{split}
\end{equation}
(here $A$, the amplitude involved in $I(\lambda,\xi)$, is
$A(t)\equiv 1$).
To bound the first summation in \eqref{eq:int r^2 << sum I^2} we
observe that for $\mu,\mu'$ with $$\left|\left\langle
\xi\left(\mu,\mu'\right),B\right\rangle \right|^{2}\le 1-2\epsilon$$
we have, thanks to Lemma \ref{lem:small der 2st 2nd}, that for every
$t\in\left[0,L\right]$ either
$\left|\phi_{\xi}'\left(t\right)\right|\ge\sqrt{\epsilon}$ or
$\left|\phi_{\xi}''\left(t\right)\right|\ge\sqrt{\epsilon}K_{\min}$.
Hence Lemma \ref{lem:Van_Der_Corp}, using the same arguments as in
the proof of Proposition \ref{stationary phase 3d}, yields the
following bound, uniform in $\xi$:
\begin{equation}
\begin{split}
I\left(\lambda, \xi
\right)&=\int_{0}^{L}A(t)e^{i\lambda\phi_{\xi}\left(t\right)}\mbox{d}t \\
&\ll_{\mathcal{C}}\min\left\{ \left \Vert A\right\Vert_\infty ,\frac{1}{\left(\sqrt{\epsilon}\lambda\right)^{1/2}}
\left ( \left \Vert A\right\Vert_\infty + \left \Vert A'
\right\Vert _1 \right ) \right\}.
\end{split}
\end{equation}
We may then bound the first summation in \eqref{eq:int r^2 << sum
I^2} as
\begin{equation}
\label{eq:sum I^{2}<<sum mu,mu'}
\begin{split}
&\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle \xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}\le 1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}\\
&\ll\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
|\mu - \mu'|\le 1/\sqrt{\epsilon}
\end{subarray}}1
+\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
|\mu - \mu'|>1/\sqrt{\epsilon}
\end{subarray}}\frac{1}{\sqrt{\epsilon}\left|\mu-\mu'\right|}\\
&\ll
\frac{1}{N^2}\sum_{\mu}\#\left\{ \mu':|\mu - \mu' | \le 1/\sqrt{\epsilon}\right\} +
\frac{1}{\sqrt{\epsilon}N^{2}} \sum_{
\mu\ne\mu'}\frac{1}{\left|\mu-\mu'\right|}.
\end{split}
\end{equation}
The second summation in the r.h.s. of \eqref{eq:sum I^{2}<<sum mu,mu'} is
$1/E^{1/2}$ times the ``Riesz $1$-energy" of the set of projected
lattice points $\^\vE(E)=\frac 1{\sqrt{E} }\vE(E)\subset S^2$, which
by Proposition \ref{prop:bd on Riesz2} is bounded by
\begin{equation*}
\sum_{ \substack{\^\mu,\^\mu' \in \^\vE\\ \^\mu\neq \^\mu'}} \frac
1{|\^\mu-\^\mu'|} \ll N^2\cdot E^{\eta},\, \forall\eta>0,
\end{equation*}
or,
\[
\frac{1}{\sqrt{\epsilon}N^{2}}\sum_{\mu\ne\mu'}\frac{1}{\left|\mu-\mu'\right|}\ll \frac{E^{-1/2+\eta}}{\sqrt{\epsilon}},\, \forall\eta>0.
\]
For every $ \eta >0, $ the number of lattice points in a spherical cap of radius $r$ on the sphere $RS^2$
can be easily shown to be $ O (R^{\eta}\left(1+r\right))$ (e.g. \cite[Lemma 2.2]{BRGAFA}), so that
the total number of solutions to $|\mu - \mu'| \le 1/\sqrt{\epsilon}$ is
$O\left(E^{\eta }/\sqrt{\epsilon}\right)$.
Hence, the first summation in the r.h.s. of \eqref{eq:sum I^{2}<<sum mu,mu'} is $ O\left ( E^{\eta } /\sqrt{\epsilon}N \right) $, and \eqref{eq:sum I^{2}<<sum mu,mu'} is
\begin{equation}
\label{eq:sum mumu'<1-2eps I^2}
\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle \xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2} \le 1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}\ll\frac{E^{\eta - 1/2}}{\sqrt{\epsilon }}.
\end{equation}
For the second summation on the r.h.s. of \eqref{eq:int r^2 << sum
I^2} we bound each of the integrals $I\left(|\mu-\mu'|,\xi\right)$
from above trivially by $1$, yielding
\[
\begin{split}
&\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle \xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}>1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}\\
&\ll\frac{1}{N^{2}}\sum_{\mu}\#\left\{ \mu':\left|\left\langle
\xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}>1-2\epsilon\right\} .
\end{split}
\]
Using Lemma \ref{lem:Geom_Lemma}, we see that all of the lattice
points satisfying \[\left|\left\langle
\xi\left(\mu,\mu'\right),B\right\rangle \right|^{2}>1-2\epsilon\]
are contained in two spherical caps of radius $\ll\sqrt{\epsilon
E}$ so that
the total number of solutions to $\left|\left\langle
\xi\left(\mu,\mu'\right),B\right\rangle \right|^{2}>1-2\epsilon$ is
$O\left(E^{\eta}\left(1+\sqrt{\epsilon E}\right)\right)$.
Therefore
\begin{equation}\label{eq:sum mumu'>1-2eps I^2}
\begin{split}
\frac{1}{N^{2}}\sum_{\begin{subarray}{c}
\mu\ne\mu'\\
\left|\left\langle \xi\left(\mu,\mu'\right),B\right\rangle
\right|^{2}>1-2\epsilon
\end{subarray}}I\left(|\mu-\mu'|,\xi\right)^{2}
&\ll \frac{E^{\eta}\left(1+\sqrt{\epsilon E}\right)}{N} \\
&\ll E^{\eta - 1/2}+\sqrt{\epsilon}E^{\eta}.
\end{split}
\end{equation}
Substituting \eqref{eq:sum mumu'<1-2eps I^2} and \eqref{eq:sum
mumu'>1-2eps I^2} into \eqref{eq:int r^2 << sum I^2} yields the
inequality
\begin{equation}\label{bd 5.11}
\iint \limits_{[0,L]^2} r(t_1,t_2)^2\mbox{d}t_1\mbox{d}t_2=
\frac{L^2}{N}+O\left(\frac{E^{\eta - 1/2}}{\sqrt{\epsilon }}+\sqrt{\epsilon}E^{\eta}\right).
\end{equation}
The estimate \eqref{eq:planar_r} follows by making the optimal
choice $\epsilon=E^{-1/2}$ in \eqref{bd 5.11}.
The proofs of \eqref{eq:planar_ri} and \eqref{eq:planar_r12} are
very similar to the above (cf. the proofs of
\eqref{eq:non_planar_ri} and \eqref{eq:non_planar_r12} within the
proof of Proposition \ref{prop:R2<<1/E^1/3}); we omit the details.
\end{proof}
\begin{remark}
It is possible to slightly improve the exponent of the bound
\eqref{eq:R2<<1/E^1/4-o(1) planar} in Proposition
\ref{prop:R2<<1/E^1/4-o(1) planar} by using a better estimate for
the number of lattice points in spherical caps (\cite[Proposition 1.4]{BRGAFA}).
\end{remark}
\subsection{Concluding the proof of Theorem \ref{thmvar}}
\label{sec:thmvar proof}
\begin{proof}
Use Proposition \ref{prop:approx Kac-Rice} together with either
Proposition \ref{prop:R2<<1/E^1/3} for the\\
nowhere vanishing torsion
case or Proposition \ref{prop:R2<<1/E^1/4-o(1) planar} for the
planar case.
\end{proof}
We finally record that the proof above will give the result
described in Remark~\ref{rem:variance analytic} for $\curve$
non-planar and analytic, by invoking the bound
\eqref{eq:Rc2<<1/E^delta analytic}
instead of Proposition \ref{prop:R2<<1/E^1/3}.
|
\section{Introduction}
\label{intro}
We consider the classical variational inequality problem which is to find a point $x^*\in C$ such that
\begin{equation}\label{vip}
(Ax^*, x-x^*)\geq 0 \quad \forall x\in C,
\end{equation}
where $C$ is a closed convex set in Hilbert space $H$, $(\cdot, \cdot)$ denotes the inner product in $H$, and $A\colon H\r H$ is a some mapping.
We assume that the following conditions hold\\
(C1)\quad The solution set of \eqref{vip}, denoted by $S$, is nonempty.\\
(C2)\quad The mapping $A$ is monotone, i.e.,
$$(Ax-Ay,x-y)\geq 0\quad \forall x,y \in C.$$
(C3)\quad The mapping $A$ is Lipschitz-continuous on $C$ with constant $L>0$, i.e., there exists $L>0$, such that
$$\n{Ax-Ay}\leq L \n{x-y} \quad \forall x,y\in C.$$
Variational inequality theory is an important tool in studying a wide class of obstacle,
unilateral, and equilibrium problems arising in several branches of pure and applied sciences in
a unified and general framework \cite{aubinekeland,baiocchi:84,lions:81,kinderlehrer,konnov2001,nagurney99}. This field is dynamic and is experiencing an explosive growth in both theory and applications. Several numerical methods have been developed for solving variational inequalities
and related optimization problems, see books \cite{baucomb,pang:03,konnov2001} and the references therein.
In order to construct an algorithm which provides strong convergence to a solution of \eqref{vip} we propose the following method \\
\textbf{Algorithm 1}
\begin{equation}\label{our}
\begin{cases}
x_0,z_0\in C,\\
z_{n+1} = P_C(x_n-\lambda Az_n),\\
x_{n+1} = P_{C_n\cap Q_n} x_0,
\end{cases}
\end{equation}
where $P_M$ denotes the metric projection on the set $M$, $\lambda \in (0,1/2L)$, the sets $C_n$ and $Q_n$ are some halfspaces which we will define in the Section \ref{algorithm}.
\section{Relation to the previous work}
There are two general approaches to study variational inequality problem under the conditions above: regularization methods and projection methods. As far as we are concerned, we prioritize the latter; good survey of regularization methods can be found in \cite{bakush,konnov2001}.
The oldest algorithm which provides convergence of the generated sequence under the above assumptions is the extragradient method proposed in 1976 by Korpelevich~\cite{korpel:76} and Antipin~\cite{antipin}. A lot of efficient modifications exist at this moment \cite{apostol12,ceng:10,reich:2011,iusem:00,Iusem:97,khobotov:89,lyashko11,nadezhkina:06,solodov:1999,tseng00,voitova11}. The natural question that arises in the case of infinite dimensional Hilbert space is how to construct an algorithm which provides strong convergence. To answer this question Nadezhkina and Takahashi~\cite{nadezhkina:06} introduced the following method \\
\textbf{Algorithm 2}
\begin{equation}\label{nadezhda}
\begin{cases}
x_0\in C,\\
y_n = P_C(x_n-\lambda Ax_n),\\
z_n = P_C(x_n-\lambda Ay_n),\\
C_n=\{w\in C \colon \n{z_n-w}\leq \n{x_n-w}\},\\
Q_n = \{w\in C \colon (x_n-w, x_0- x_n)\geq 0 \},\\
x_{n+1} = P_{C_n\cap Q_n} x_0,
\end{cases}
\end{equation}
where $\lambda\in (0,1/L)$.
Under the above assumptions (C1)--(C3) they proved that the sequence $(x_n)$, generated by \eqref{nadezhda}, converges strongly to $P_S x_0$. Their method based on the extragradient method and on the hybrid method, proposed by \cite{nakajo03,solodov00}. The computational complexity of \eqref{nadezhda} on every step is three computation of metric projection and two values of $A$.
Inspired by this scheme, Censor, Gibali and Reich \cite{censor_reich:2011} presented the following algorithm \\
\textbf{Algorithm 3}
\begin{equation}\label{censor_strong}
\begin{cases}
x_0\in H,\\
y_n = P_C(x_n-\lambda Ax_n),\\
T_n = \{w\in H \colon (x_n-\lambda Ax_n-y_n,w-y_n)\leq 0\},\\
z_n = \a_n x_n + (1-\a_n) P_{T_n}(x_n-\lambda Ay_n),\\
C_n=\{w\in H \colon \n{z_n-w}\leq \n{x_n-w}\},\\
Q_n = \{w\in H \colon (x_n-w, x_0- x_n)\geq 0 \},\\
x_{n+1} = P_{C_n\cap Q_n} x_0,
\end{cases}
\end{equation}
where $\lambda \in (0,1/L)$ and $(\a_n)\subset [0,\a]$ for some $\a \in [0,1)$.
In contrast to \eqref{nadezhda} the sets $C_n$ and $Q_n$ are halfspaces and hence it is much more simpler to calculate $P_{C_n\cap Q_n}x_0$ than projection onto the general convex set $C$. Therefore in the next schemes we will not take into consideration this projection. Also on the second step it is calculated only a projection onto halfspace $T_n$ but not onto set $C$ like in \eqref{nadezhda}. However on every step of \eqref{censor_strong} we need to calculate two values of $A$ as well as in \eqref{nadezhda}.
Using hybrid method it is not difficult to modify Tseng algorithm \cite{tseng00} \\ \textbf{Algorithm 4}
\begin{equation}\label{tseng_strong}
\begin{cases}
x_0\in H,\\
y_n = P_C(x_n-\lambda Ax_n),\\
z_n =y_n + \lambda (Ax_n-Ay_n),\\
C_n=\{w\in H \colon \n{z_n-w}\leq \n{x_n-w}\},\\
Q_n = \{w\in H \colon (x_n-w, x_0- x_n)\geq 0 \},\\
x_{n+1} = P_{C_n\cap Q_n} x_0,
\end{cases}
\end{equation}
where $\lambda \in (0,1/L)$. In computational sense it is similar to \eqref{censor_strong}.
In this work we show that with some other choice of sets $C_n$ it is possible to throw out in \eqref{nadezhda} or in \eqref{censor_strong} the step of extrapolation which consist in $y_n = P_C(x_n - \lambda Ax_n)$. We emphasize that in order to prove a convergence of Algorithm 1 we use an idea due to Popov~\cite{popov:80} (the refinement of this idea see also in \cite{mal-sem}). It is easy to see that our method~\eqref{our} on every iteration needs only one computation of projection (as in \eqref{censor_strong} or \eqref{tseng_strong}) and only one value of $A$. For example, very often variational inequality problems which arise from optimal control, provide a very complicated operator such that only computation of the latter is a very sophisticate problem. (For more details see \cite{lions71}).
\section{Preliminaries}
\label{pre}
In order to prove our main result, we need the following statements (see books \cite{aubinekeland,baiocchi:84,baucomb,kinderlehrer}). At first, the following well-known properties of the projection mapping will be used throughout this paper.
\begin{lemma}\label{proj}
Let $M$ be nonempty closed convex set in $H$, $x \in H$. Then
\begin{itemize}
\item[i)] $(P_Mx - x, y-P_Mx)\geq 0\quad \forall y\in M$;
\item[ii)] $\n{P_Mx-y}^2\leq \n{x-y}^2 - \n{x-P_Mx}^2 \quad \forall y\in M$.
\end{itemize}
\end{lemma}
Next two lemmas are also well-known.
\begin{lemma}[Minty]\label{minty}
Assume that $A:C\r H$ is a continuous and monotone mapping. Then $x^*$ is a solution of \eqref{vip} iff $x^*$ is a solution of the following problem
$$\text{find}\, x\in C, \quad \text{such that}\quad (Ay,y-x)\geq 0\quad \forall y\in C.$$
\end{lemma}
\begin{remark}
The solution set $S$ of variational inequality \eqref{vip} is closed and convex.
\end{remark}
We write $x_n \rightharpoonup x$ to indicate that the sequence $(x_n)$ converges weakly to $x$. $x_n \to x$ implies that $(x_n)$ converges strongly to $x$.
\begin{lemma}[Kadec--Klee property of a Hilbert space]\label{kadec}
Let $(x_n)$ be a sequence in $H$. Then from $\n{x_n}\r \n{x}$ and $x_n\rr x$ follows that $x_n \r x$.
\end{lemma}
At last, we prove the following result.
\begin{lemma} \label{lema_tech}
Let $(a_n)$, $(b_n)$, $(c_n)$ be nonnegative real sequences, $\a, \b\in \mathbb R$ and for all $n\in \mathbb N$ the following inequality holds
\begin{equation}\label{tech_ineq}
a_n \leq b_n - \a c_{n+1} +\b c_n.
\end{equation}
If $\sum_{n=1}^\infty b_n< +\infty$ and $\a> \b \geq 0$ then $\lim_{n\r \infty} a_n = 0$.
\end{lemma}
\begin{proof}
Using inequality \eqref{tech_ineq} for $n=1, n=2, \dots, n=N$ we obtain
\begin{gather*}
a_1 \leq b_1 - \a c_{2} + \b c_1,\\
a_{2} \leq b_{2} - \a c_{3} +\b c_{2},\\
\dots \\
a_{N} \leq b_N - \a c_{N+1} + \b c_N.
\end{gather*}
Adding all these inequalities yields
\begin{equation*}
\sum_{n=1}^N a_n \leq \sum_{n=1}^N b_n -(\a-\b)\sum_{n=2}^N c_n -\a c_{N+1} + \b c_1 \leq \sum_{n=1}^\infty b_n + \b c_1.
\end{equation*}
Since $N$ is arbitrary, we see that series $\sum_{n=1}^\infty a_n$ is convergent and hence $a_n \to 0$. \qed
\end{proof}
\section{Algorithm and its convergence}
\label{algorithm}
Now we formally state our algorithm.
\begin{algorithm}[Hybrid algorithm without extrapolation step]\label{alg}
\noindent
\begin{enumerate}
\item Choose $x_0, z_0 \in C$ and two parameters $k>0$ and $\lambda > 0$.
\item\label{step} Given the current iterate $x_n$ and $z_n$, compute
\begin{equation}
\label{z_n} z_{n+1} = P_C(x_n-\lambda Az_n).
\end{equation}
If $ z_{n+1}= x_n = z_n$ then stop. Otherwise, construct sets $C_n$ and $Q_n$ as
\begin{equation}\label{C_n}
\begin{split}
C_0 & = H, \\
C_n & = \bigl\{ w\in H\colon \n{z_{n+1}-w}^2 \leq \n{x_n-w}^2+k\n{x_n-x_{n-1}}^2 \\ & - (1-\frac 1 k -\lambda L )\n{z_{n+1}-z_n}^2 + \lambda L \n{z_n-z_{n-1}}^2\bigr\}, \quad n\geq 1, \\
Q_0 &= H, \\
Q_n & = \{w\in H \colon (x_n-w, x_0-x_n)\geq 0\}, \quad n\geq 1,
\end{split}
\end{equation}
and calculate
\begin{equation}
x_{n+1} = P_{C_n\cap Q_n}x_0.
\end{equation}
\item Set $n \leftarrow n+1$ and return to step \ref{step}.
\end{enumerate}
\end{algorithm}
We remark that sets $C_n$ look a slightly complicated in contrast to \eqref{censor_strong}. However it is only for superficial examination, for a computation it does not matter. In \eqref{C_n} and in \eqref{nadezhda} both $C_n$ are some halfspaces.
First we show that the stopping criterion in Algorithm \ref{alg} is valid.
\begin{lemma}
If $z_{n+1} = x_n = z_n $ in Algorithm \ref{alg} then $x_n \in S$.
\end{lemma}
\begin{proof}
From \eqref{z_n} and Lemma \ref{proj} follows
$$\lambda( Ax_n, y - x_n)\geq 0\quad \forall y\in C.$$
Since $\lambda >0$, we have $x_n\in S$.
\qed
\end{proof}
Next lemma is central to our proof of the convergence theorem.
\begin{lemma}\label{main_lemma}
Let $(x_n)$ and $(z_n)$ be two sequences generated by Algorithm \ref{alg} and let $z\in S$. Then
\begin{multline}
\n{z_{n+1}-z}^2 \leq \n{x_n-z}^2+k\n{x_n-x_{n-1}}^2 \\ -(1-\frac 1 k -\lambda L )\n{z_{n+1}-z_n}^2 + \lambda L \n{z_n-z_{n-1}}^2 .
\end{multline}
\end{lemma}
\begin{proof}
By Lemma \ref{proj} we have
\begin{multline}\label{fst}
\n{z_{n+1}-z}^2 \leq \n{x_n-\lambda Az_n -z}^2 - \n{x_n-\lambda Az_n-z_{n+1}}^2 \\ = \n{x_n-z}^2 - \n{x_n-z_{n+1}}^2 -
2\lambda (Az_n,z_{n+1}-z).
\end{multline}
Since $A$ is monotone and $z\in S$, we see that $$(Az_n, z_n- z)\geq 0.$$ Thus, adding $2\lambda(Az_n, z_n- z)$ to the right side of \eqref{fst} we get
\begin{multline}\label{snd}
\n{z_{n+1}-z}^2 \leq \n{x_n-z}^2 - \n{x_n-z_{n+1}}^2 - 2\lambda (Az_n,z_{n+1}-z_n) \\ = \n{x_n-z}^2 - \n{x_n-x_{n-1}}^2 - 2(x_n-x_{n-1}, x_{n-1}-z_{n+1}) \\ - \n{x_{n-1}-z_{n+1}}^2 - 2\lambda (Az_n,z_{n+1}-z_n) = \n{x_n-z}^2 \\ - \n{x_n-x_{n-1}}^2 - 2(x_n-x_{n-1}, x_{n-1}-z_{n+1}) -\n{x_{n-1}-z_n}^2 \\ - \n{z_n-z_{n+1}}^2 - 2(x_{n-1}-z_n,z_n-z_{n+1}) - 2\lambda (Az_n-Az_{n-1},z_{n+1}-z_n) \\ - 2\lambda (Az_{n-1},z_{n+1}-z_n) = \n{x_n-z}^2 - \n{x_n-x_{n-1}}^2 \\ - 2(x_n-x_{n-1}, x_{n-1}-z_{n+1}) - \n{x_{n-1}-z_n}^2 - \n{z_n-z_{n+1}}^2 \\ - 2\lambda(Az_n-Az_{n-1}, z_{n+1}-z_n) + 2(x_{n-1}-\lambda Az_{n-1}- z_n,z_{n+1}-z_{n}).
\end{multline}
As $z_{n} = P_C(x_{n-1}-\lambda Az_{n-1})$ and $z_{n+1}\in C$ we have
\begin{equation}\label{eq:1}
(x_{n-1}-\lambda Az_{n-1}- z_n,z_{n+1}-z_{n}) \leq 0.
\end{equation}
Using the triangle, the Cauchy-Schwarz, and the Cauchy inequalities we obtain
\begin{multline}\label{eq:2}
2(x_n-x_{n-1}, x_{n-1}-z_{n+1})\leq 2\n{x_n-x_{n-1}}\n{x_{n-1}-z_{n}} +2\n{x_{n}-x_{n-1}} \n{z_{n}-z_{n+1}} \\ \leq
\n{x_n-x_{n-1}}^2 + \n{x_{n-1}-z_{n}}^2 + k\n{x_n-x_{n-1}}^2 + \frac 1 k \n{z_{n+1}-z_n}^2.
\end{multline}
Since $A$ is Lipschitz-continuous, we get
\begin{multline}\label{eq:3}
2\lambda (Az_n-Az_{n-1}, z_{n+1}-z_n) \leq 2 \lambda L\n{z_n-z_{n-1}} \n{z_{n+1}-z_{n}} \\ \leq \lambda L(\n{z_{n+1}-z_n}^2 + \n{z_n-z_{n-1}}^2).
\end{multline}
Combining inequalities \eqref{snd}~--~\eqref{eq:3}, we see that
\begin{multline*}
\n{z_{n+1}-z}^2 \leq \n{x_n-z}^2 + k\n{x_n-x_{n-1}}^2 - (1-\frac 1 k -\lambda L)\n{z_{n+1}-z_n}^2 + \lambda L \n{z_{n}-z_{n-1}}^2,
\end{multline*}
which completes the proof.
\qed
\end{proof}
Now we can state and prove our main convergence result.
\begin{theorem}\label{th_1}
Assume that (C1)--(C3) hold and let $\lambda \in (0,\frac 1 {2L})$, $k>\frac{1}{1-2\lambda L}$. Then the sequences $(x_n)$ and $(z_n)$ generated by Algorithm \ref{alg} converge strongly to $P_S x_0$.
\end{theorem}
\begin{proof}
It is evident that sets $C_n$ and $Q_n$ are closed and convex. By Lemma \ref{main_lemma} we have that $S\subseteq C_n$ for all $n\in \mathbb Z^+$. Let us show by mathematical induction that $S\subseteq Q_n$ for all $n\in \mathbb Z^+$. For $n=0$ we have $Q_0 = H$. Suppose that $S \subseteq Q_n$. It is sufficient to show that $S \subseteq Q_{n+1}$. Since $x_{n+1}=P_{C_n\cap Q_n}x_0$ and $S\subseteq C_n\cap Q_n$, it follows that $(x_{n+1}-z, x_0-x_{n+1})\geq 0$ $\forall z\in S$. From this by the definition of $Q_n$ we conclude that $z\in Q_{n+1}$ $\forall z \in S$. Thus, $S\subseteq Q_{n+1}$ and hence $S\subseteq C_n\cap Q_n $ for all $ n\in \mathbb Z^+$. For this reason the sequence $(x_n)$ is defined correctly.
Let $\bar x = P_S x_0$. Since $x_{n+1}\in C_n\cap Q_n$ and $\bar x \in S\subseteq C_n\cap Q_n$, we have
\begin{equation} \label{bounded}
\n{x_{n+1} - x_0}\leq \n{\bar x-x_0}.
\end{equation}
Therefore, $(x_n)$ is bounded. From $x_{n+1}\in C_n\cap Q_n\subseteq Q_n$ and $x_n = P_{Q_n}x_0$, we obtain
\begin{equation*}
\n{x_n-x_0}\leq \n{x_{n+1}-x_0}.
\end{equation*}
Hence, there exists $\lim_{n\r \infty}\n{x_n-x_0}$. In addition, since $x_n = P_{Q_n}x_0$ and $x_{n+1}\in Q_n$ by Lemma \ref{proj} we have
\begin{equation}\label{key}
\n{x_{n+1}-x_n}^2 \leq \n{x_{n+1}-x_0}^2 - \n{x_n-x_0}^2.
\end{equation}
From this it may be concluded that series $\sum_{n=1}^\infty \n{x_{n+1}-x_n}^2$ is convergent. In fact, from \eqref{key} and \eqref{bounded} we obtain
\begin{multline*}
\sum_{n=1}^N \n{x_{n+1}-x_n}^2\leq \sum_{n=1}^N (\n{x_{n+1}-x_0}^2 - \n{x_n-x_0}^2) \leq \n{\bar x-x_0}^2 - \n{x_1-x_0}^2.
\end{multline*}
Since $x_{n+1}\in C_n$, we see that
\begin{multline*}
\n{z_{n+1}-x_{n+1}}^2 \leq \n{x_{n+1}-x_{n}}^2 + k\n{x_n-x_{n-1}}^2 \\ - (1-\frac 1 k -\lambda L)\n{z_{n+1}-z_n}^2 + \lambda L \n{z_{n}-z_{n-1}}^2.
\end{multline*}
Set
\begin{gather*}
a_n = \n{z_{n+1}-x_{n+1}}^2, \quad
b_n = \n{x_{n+1}-x_n}^2 + k\n{x_n-x_{n-1}}^2, \\
c_n =\n{z_{n}-z_{n-1}}^2, \quad
\a = (1-\frac 1 k -\lambda L), \quad
\b = \lambda L.
\end{gather*}
By Lemma \ref{main_lemma} since $\sum_{n=1}^{\infty} b_n <+\infty $ and $\a > \b$,
\begin{equation*
\lim_{n\r \infty}\n{z_{n}-x_{n}} = 0.
\end{equation*}
For this reason $(z_n)$ is bounded and
$$\n{z_{n+1}-z_{n}}\leq \n{z_{n+1}-x_{n+1}} + \n{x_{n+1}-x_n} + \n{x_n-z_n} \r 0.$$
As $(x_n)$ is bounded there exist a subsequence $(x_{n_i})$ of $(x_n)$ such that $(x_{n_i})$ converges weakly to some $x^*\in H$. We show $x^* \in S$.
From \eqref{z_n} by Lemma \ref{proj} it follows that
$$(z_{n_i+1}-x_{n_i} + \lambda Az_{n_i}, y-z_{n_i+1})\geq 0\quad \forall y\in C.$$
This is equivalent to
\begin{multline}\label{weak}
0\leq (z_{n_i+1}-z_{n_i} + z_{n_i}-x_{n_i}, y-z_{n_i+1})+\lambda (Az_{n_i}, y-z_{n_i}) \\ + \lambda(Az_{n_i}, z_{n_i}-z_{n_i+1}) \leq
(z_{n_i+1}-z_{n_i},y-z_{n_i+1}) + (z_{n_i}-x_{n_i}, y-z_{n_i+1}) \\ + \lambda (Ay, y-z_{n_i}) + \lambda(Az_{n_i}, z_{n_i}-z_{n_i+1}) \quad \forall y\in C.
\end{multline}
In the last inequality we used monotonicity of $A$.
Taking the limit as $i\r \infty $ in \eqref{weak} and using that $z_{n_i}\rr x^*$, we obtain
$$0\leq (Ay, y-x^*) \quad \forall y\in C,$$
which implies by Lemma \ref{minty} that $x^* \in S$.
Let us show $x_{n_i}\r x^*$. From $\bar x = P_S x_0$ and $x^*\in S$ it follows that
$$\n{\bar x - x_0}\leq \n{x^* - x_0}\leq \liminf_{i\r \infty}\n{x_{n_i}-x_0} = \lim_{i\r \infty}\n{x_{n_i}-x_0}\leq \n{\bar x - x_0}.$$
Thus, $$\lim_{i\r \infty}\n{x_{n_i}-x_0} = \n{x^* - x_0}.$$
From this and $x_{n_i}-x_0 \rr x^*-x_0$ by Lemma \ref{kadec} we can conclude that $x_{n_i}-x_0 \r x^*-x_0$. Therefore, $x_{n_i}\r x^*$.
Next, we have
$$\n{x_{n_i}-\bar x}^2 = (x_{n_i}-x_0, x_{n_i}-\bar x) + (x_0-\bar x, x_{n_i}-\bar x)\leq (x_0-\bar x, x_{n_i}-\bar x). $$
As $i\r \infty$, we obtain
$$\n{x^*-\bar x}^2 \leq (x_0-\bar x, x^*-\bar x)\leq 0.$$
Hence we have $x^* = \bar x$. Since the subsequence $(x_{n_i})$ was arbitrary, we see that $x_n\r \bar x$. It is clear that $z_n \r \bar x$. \qed
\end{proof}
\section{Computational Experience}
In this section we compare the performance of Algorithms 1, 3 and 4 which provide a strong convergence in Hilbert space. We have already noted that the main advantage of our algorithm is a computation on every step only one value of operator. But we will show that it has competitive performance even on simple examples where the previous reasoning has no impact. Of course all examples are considered in $\mathbb R^m$, therefore there is no sense to use any of algorithms 1,2,3,4 to obtain the solution of variational inequality. However there are many problems that arise in infinite dimensional spaces. In such problems norm convergence is often much more desirable than weak convergence (see \cite{baucomb_weak} and references therein). For this reason strong algorithms 1,2,3,4, can be better than extragradient algorithms that provide weak convergence. Another reason to study their convergence is an academic interest.
Although Algorithm 2 gave birth to all strong algorithms mentioned above, we do not report its results since Algorithm 2 requires a large amount of projections onto a feasible set which are not trivial for our test problems.
We describe the test details below. Computations were performed using Wolfram Mathematica 8.0 on an Intel Core i3-2348M \@2.3GHz running 64-bit Linux Mint 13.
In all tests we take $\lambda = 1/(4L)$, $k= 3$ in Algorithm 1 and $\a_n = 0.5$ in Algorithm 3. The projection onto a feasible set $C$ is performed using the cyclic projection method with error $\varepsilon$. The time is measured in seconds using the intrinsic Mathematica function Timing. The termination criterion is $\n{x-P_C(x-\lambda Ax)} \leq \varepsilon$.
The projection in Algorithms 1, 3, 4 onto the intersection $C_n\cap Q_n$ of two halfspaces is explicit and is given by Haugazeau formula (see \cite{baucomb} for more details).
The first example is very simple. The feasible set $C\subset \mathbb R^5$ is a random polygon with $10$ linear constraints and $Ax = x-u$ where $u$ is a random point in $\mathbb R^5$. It is clear that the solution $x^*$ of variational inequality \eqref{vip} is the point $P_C u$. For this problem we take $L=1$ and $\varepsilon = 0.01$. We study this problem for a different choice of starting points $x_0$. The results for a three different choices of starting point $x_0$ are shown in Table 1.
\begin{center}
\begin{tabular}{cc|c|c|c|c|c|l}
\cline{2-7}
& \multicolumn{2}{|c|}{Ex. 1} & \multicolumn{2}{c|}{Ex. 2} & \multicolumn{2}{c|}{Ex. 3} \\
\cline{2-7} \multicolumn{1}{c|}{} & Iter. & Time & Iter. & Time & Iter. & Time & \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 1} & 307 & 61.1&271 &52.1 &137 &29.4 \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 3} &317 &60.4 & 315& 58.1& 276 &52.5 \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 4} & 518 &97.1 & 312& 58.9 &221 & 42.9 \\ \cline{1-7}
\end{tabular}
\\ \vspace{0.2cm}
\title{Table 1}
\end{center}
Of course this result depends on a feasible set. However our purpose was only to study some examples where Algorithm 1 is effective.
The second problem is the following. We take $Ax = Mx + q$ with the matrix $M$ randomly generated as suggested in \cite{hphard}:
$$M = AA^T + B + D,$$ where every entry of the $n \times n$ matrix $A$ and of the $n\times n$ skew-symmetric matrix $B$ is uniformly generated from $(-5, 5)$, and every diagonal entry of the $n\times n$ diagonal $D$ is uniformly generated from $(0, 0.3)$ (so $M$ is positive definite), with every entry of $q$ uniformly generated from $(-500, 0)$. The feasible set is
$$C = \{x\in \mathbb R^m_+ \mid x_1 + x_2 +\dots + x_m = m\},$$ and the starting point is $x_0 = (1,\dots, 1)$.
For this problem we take $L = \n{M}$. We have generated three random samples with different choice of $M$ and $q$ for both cases $n=2$ and $n=10$, the results are tabulated in Table 2 and 3.
For $n=2$, $\varepsilon = 0.001$
\begin{center}
\begin{tabular}{cc|c|c|c|c|c|l}
\cline{2-7}
& \multicolumn{2}{|c|}{Ex. 1} & \multicolumn{2}{c|}{Ex. 2} & \multicolumn{2}{c|}{Ex. 3} \\
\cline{2-7} \multicolumn{1}{c|}{} & Iter. & Time & Iter. & Time & Iter. & Time & \\ \cline{1-7}
\multicolumn{1}{|c| }{Alg. 1} &245& 4.9 &216 & 34.7 & 385 & 463.7 & \\ \cline{1-7}
\multicolumn{1}{|c| }{Alg. 3} & 278 & 5.7 & 219 &35.1 & 349 &412.8 & \\ \cline{1-7}
\multicolumn{1}{|c| }{Alg. 4} & 291 & 6.0 & 187 & 30.0 & 268& 320.0 &\\ \cline{1-7}
\end{tabular}
\\ \vspace{0.2cm}
\title{Table 2}
\end{center}
And for $n=10$, $\varepsilon = 0.01$
\begin{center}
\begin{tabular}{cc|c|c|c|c|c|l}
\cline{2-7}
& \multicolumn{2}{|c|}{Ex. 1} & \multicolumn{2}{c|}{Ex. 2} & \multicolumn{2}{c|}{Ex. 3} \\
\cline{2-7} \multicolumn{1}{c|}{} & Iter. & Time & Iter. & Time & Iter. & Time & \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 1} &4267 & 123.6 &2728 &32.7 & 554 & 15.3 & \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 3} &2571 &75.9 & 1826& 29.2 & 409 & 12.1 & \\ \cline{1-7}
\multicolumn{1}{ |c| }{Alg. 4} & 2564 &76.4 & 1632& 48.3 & 522 & 16.0&\\ \cline{1-7}
\end{tabular}
\\ \vspace{0.2cm}
\title{Table 3}
\end{center}
As one can see, on our examples Algorithm 1 has competitive performance. We caution, however, that this study is very preliminary.
\bibliographystyle{siam}
|
\section{Introduction}
Measurements of internal oscillation modes of red giant stars by the {\it Kepler} satellite have been used to infer
rapid rotation in the radiative core -- in some cases, an order of magnitude faster than at the surface
\citep{becketal2012,mossetal2012}. This evidence has been interpreted by \cite{CantMBCP2014} in terms of a
decoupling of the rotation of the core from the convective envelope, which is
assumed to rotate as a solid body.
Here we reconsider the role that convection plays in redistributing angular momentum within a star. So far
only a limited understanding has been developed of the interaction between rotation and convection.
The most convincing theoretical calculations involve direct numerical simulations (e.g. \citealt{brunp2009}).
Red giants offer a promising test case, because (i) the influence of the Coriolis force is reduced
(although not entirely eliminated) in comparison with main sequence (MS) stars; (ii) the great depth of
the convective envelope reduces the effect of boundary conditions on the angular velocity profile;
and (iii) feedback from magnetic fields on the rotation profile is less important, especially near the tip of the
red giant and asymptotic giant branches (RGB and AGB), where most of the envelope rotates slowly.
Disentangling the most important processes that control the rotation profile of a
star is, for these reasons, easier in red giants than in the Sun.
The internal rotation of giant stars also has important implications for the magnetism and spin of their
white dwarf (WD) remnants. A basic theme of this paper, and the companion
paper \cite[hereafter Paper II]{KissT2015}
is that one must consider the co-evolution of rotation and magnetic fields if one is to understand either.
The strong flow of material from
the hydrogen-rich envelope, through burning shell(s), and into the core implies that one cannot treat
core and envelope in isolation from each other.
This observation leads us to discard models of the rotation which neglect the influence
of large-scale poloidal magnetic fields that were deposited by previous convective activity. For example,
the Modules for Experiments in Stellar Astrophysics (MESA) code \citep{paxtetal2011} implements a heuristic prescription for angular momentum redistribution
in the radiative parts of a star, which starts with the current-driven (Tayler) instability of a
predominantly toroidal magnetic field, leading to a poloidal component that is then wound back by
differential rotation into the toroidal direction \citep{Spru2002}.
The nearly solid rotation of the Sun's core has been used to infer the presence of a helical magnetic field \citep{GougM1998}.
The introduction of a persistent twist to the field not only facilitates hydromagnetic stability
\citep{brais2004}, but also suppresses differential rotation across poloidal magnetic flux surfaces.
The minimal poloidal field that will transport angular momentum rapidly enough to compensate changes
in the core mass profile on the MS is only $\sim 10^{-6}$ G; but a field $\sim 10^2$-$10^3$ times stronger
is required to compensate a latitude-depenent convective stress that is partly transmitted through
a tachocline layer into the core (Paper II). The ascent of the early red giant branch is
completed over a Gyr, not a great deal shorter than the present age of the Sun, meaning that a similarly
weak poloidal magnetic flux would enforce nearly solid rotation in the core of a subgiant
(e.g. \citealt{maeder14}).
In a $1$-$2M_\odot$ star, the source of this field differs between the early
and later stages of post-MS expansion. The envelope expands in mass during the first dredge-up phase,
so that any remnant field in the outer core would have been deposited by a vigorously convective and
rapidly rotating envelope during the pre-MS contraction, when magnetic fields several orders of
magnitude stronger were present. The minimal poloidal magnetic field that pushes the core toward solid rotation
is well below the threshold of detectability if transported to the surface of the star. We note that our
assumption of solid rotation in radiative material depends on the relative weakness of
additional processes that might source differential rotation, such as the damping of gravity waves
that are excited in adjacent convective zones.
We are led to a simpler model for the angular momentum profile in a post-MS star: nearly solid
rotation in the core, as enforced by large-scale Maxwell stresses, combined with inhomogeneous rotation
in the convective envelope. The details of a dynamo operating around the core-envelope boundary,
the ensuing flow of magnetic helicity into the core, and the conditions under which the dynamo
shuts off and core and envelope decouple, are addressed in Paper II. Here our focus is on the
angular velocity profile of the envelope, especially as it is influenced by the Coriolis force.
Convective Reynolds stresses appear to transport angular momentum inward in accretion flows \citep{BalbHS1996}.
This effect is demonstrated in stars with extended convective envelopes by \cite{brunp2009} using
three-dimensional anelastic simulations. In slowly rotating cases, most relevant to the present
discussion, the specific angular momentum is found to be independent of radius.
The rotation frequency of the solar convective envelope is, by contrast, roughly independent of radius
-- a fact that appears to have influenced treatments of rotation in the post-MS phase. Although
the solar envelope also covers many scale heights, it has a modest aspect ratio (0.7:1 in radius),
implying that thermal boundary conditions have a stronger influence on the rotation profile than in a
star with a deep envelope (e.g. \citealt{browning06}). The Sun is not especially rapidly rotating by the standards of
late-type MS stars, although the Coriolis force still has a significant effect on its rotation profile.
The rotation period $P_{\rm rot} = 2\pi/\Omega$ is about 3 times the local convective time
$\tau_{\rm con} = \ell_P/v_{\rm con}$ (where $v_{\rm con}$ is the speed of a convective eddy and $\ell_P$
the pressure scale height), corresponding to a Coriolis parameter ${\rm Co} \equiv \Omega \tau_{\rm con} \sim 2$.
Dynamo action may also influence the rotation profile \citep{brun04}.
The most important decision to be made here is how to handle the backreaction of the Coriolis force on the
inward pumping of angular momentum. We investigate two simple prescriptions: i) a balance between the Coriolis force and non-radial
pressure gradient forces that result from the entropy difference between convective downflows and upflows; and ii) an upper limit
${\rm Co} \sim 1$ in the inner envelope, representing an approximate balance between the Coriolis and inertial forces.
Both of these prescriptions are compared against asteroseismic fits to {\it Kepler} data for subgiants and
post-core He flash stars. We find that prescription i) provides a very promising quantitative fit.
Measurements of surface rotation in subgiants provide independent constraints on the internal angular momentum distribution in post-MS stars.
\cite{schrijver93} have used stellar models to argue for an inconsistency between these surface measurements and the evolution of
$\sim 1.5\,M_\odot$ subgiants that maintain solid rotation. Adding a more rapidly rotating core, while maintaining solid rotation in the
envelope, was found to have a negligble impact on the result. The discrepency between models and data was therefore interpreted in terms of
an additional magnetic wind torque acting on a star as it develops a convective envelope upon leaving the MS.
Here we show that the same inward pumping of angular momentum in the envelope that reproduces the
{\it Kepler} observations also
causes a factor $\sim 1/2$ reduction in surface rotation speed over the same interval, thereby removing most (but not
necessarily all) of the evidence for a magnetic wind torque. The calculations presented here, and in Paper II,
neglect any rotational effect of surface magnetic fields during the post-MS phase.
We also extend our considerations to the tip of the RGB and AGB. Here the Coriolis force has a reduced effect, due to the
greater expansion of the star ($\gtrsim 10$ times larger than the {\it Kepler} sample analyzed by \citealt{becketal2012}
and \citealt{mossetal2012}). This leads us to consider more inwardly peaked rotation profiles than the one
suggested by fitting {\it Kepler} data.
A final consideration is binary interaction. This is especially important for stars which lose
most of their natal angular momentum to a magnetized wind on the MS, corresponding to an initial mass $\lesssim 1.3\,M_\odot$.
After such a star ascends to the upper RGB or AGB, the inward pumping of angular momentum cannot, by itself, sustain
${\rm Co}\gtrsim 1$ in the convective envelope.
A robust hydromagnetic dynamo is maintained during the greatest expansion of the star only if it encounters an
external source of angular momentum. We consider the interaction of a convecting giant with a companion planet,
calculating the combined orbital evolution of the planet and the change in the angular momentum of the star.
An interesting detail of this evolution is the excitation of orbital eccentricity by the gravitational
quadrupole associated with large-scale convective eddies in the giant. In the case of a planetary companion,
this effect is much more pronounced than when the companion is a neutron star \citep{phinney92}, and we show
that it extends the range of orbital periods within which the planet gives up its orbital angular momentum to the star.
\subsection{Plan of the Paper}
We investigate the rotation profile of a deep convective envelope in Section \ref{s:pump}, with a focus on
the backreaction of the Coriolis force on angular momentum pumping. Various angular velocity profiles are compared
against the {\it Kepler} measurements of core rotation for subgiants and post-He flash stars in Section \ref{s:1.5Msun}.
The effect of angular momentum pumping on the evolution of surface rotation of subgiants is examined in Section
\ref{s:spindown}.
In Section \ref{s:rotation}, we consider the rotational evolution of model stars of initial
mass 1 and $5\,M_\odot$, including the effects of convective pumping, angular momentum loss to a wind, and
planetary interaction.
Details of the evolution of the orbit of a planet around an expanding giant are given in Section \ref{s:orbital_evolution}, including
drag from the tide raised on the star and the excitation of eccentricity by a convective quadrupole.
Some outstanding questions are addressed in the concluding Section \ref{s:conclusions}.
The appendix gives further details of our calculation of eccentricity growth in the orbit of a companion planet.
\section{Pumping of Angular Momentum in Deep Convective Envelopes: Backreaction from the Coriolis Force}\label{s:pump}
The deep convective envelope of a giant star divides into an outer part that is slowly rotating (${\rm Co} \lesssim 1$);
and an inner part where ${\rm Co} \gtrsim 1$ and the rotation profile is influenced by the Coriolis force.
The entire envelope may rotate slowly near maximum expansion, especially if the spin angular momentum of the star
has been strongly depleted by a magnetized wind during the MS phase.
On the other hand, a star is still compact enough during the early stages of post-MS expansion that $\Omega \tau_{\rm con} \gg 1$
throughout much of the convective envelope.
We imagine that entropy is approximately conserved over downflows and upflows that are extended
compared with a local scale height. Some indirect evidence for such a configuration has long been obtained
from one-dimensional models of giant convection, which require a mixing length exceeding a single
scale height \citep{sb91}.
Such a suppression of radial mixing is suggestive of an inwardly peaked rotation profile in the outer, slowly rotating
envelope, corresponding to the conservation of specific angular momentum $j$ by individual convective flows --
as indeed is obtained numerically by \cite{brunp2009}. Rotation
profiles intermediate between $\Omega(r) \propto r^{-2}$ and solid rotation could clearly be obtained from limited mixing,
but for definiteness we focus here on the simple case $\partial j/\partial r \rightarrow 0$ where
${\rm Co} < 1$. Obtaining the angular velocity profile in the inner envelope depends on a more detailed analysis
of the Coriolis force, which we give below.
We consider the evolving partition between fast and slow rotation
in Section \ref{s:partition}, and how it depends on the angular momentum retained (or gained) by the star in Section
\ref{s:rotation}.
Given a flat distribution of specific angular momentum in the outer envelope, an intermediate-mass star generally sustains an
inner zone where ${\rm Co} \gtrsim 1$, even near the tip of the RGB or AGB;
but a star which spins down on the MS must gain angular momentum from a planetary or
stellar companion.
\subsection{Extended Upflows and Downflows with Quasi-Geostrophic Balance}\label{s:extend}
When the inner zone with ${\rm Co} \gtrsim 1$ covers a wide range of radius, we can look for a power-law scaling of
the rotation frequency $\Omega$ with spherical radius. The stellar envelope is nearly adiabatic and spherically stratified.
In this inner zone of fast rotation ($\Omega r\sin\theta \gg v_{\rm con}$) about an axis $\hat z$ ($\theta = 0$),
the steady vorticity equation reads (e.g. \citealt{balbus09})
\begin{equation}\label{eq:vort}
r\sin\theta {\partial\Omega^2\over\partial z} \simeq -{1\over C_p \rho r}{\partial P\over \partial r}
{\partial S\over\partial \theta} = {g(r)\over C_p r} {\partial S\over \partial \theta}.
\end{equation}
Here $P$ and $\rho$ are pressure and density, $g(r) = GM(r)/r^2$ is gravity,
and the specific entropy is $S = [\mu(\gamma~-~1)]^{-1}\ln(P/\rho^\gamma)$,
approximated here as that of an ideal gas with mean atomic weight $\mu$ and specific heat $C_p$ ($C_v$)
at constant pressure (volume).
The `thermal wind' approximation used in Equation (\ref{eq:vort})
holds because $|\partial P/\partial \theta| \ll |\partial P/\partial \ln r|$.
In the upper, slowly rotating part of the envelope, the latitudinal pressure gradient
is sourced by the entropy difference between upflows and downflows,
and is smaller by a factor $(v_{\rm con}/c_s)^2$ than the radial gradient.
Our focus here is on a rapidly rotating solution to Equation (\ref{eq:vort}), where the angular pressure
gradient which sources $\partial S/\partial\theta$ is proportional to $\Omega^2$.
When the gravitating mass in the inner envelope is dominated by the stellar core, gravity $g(r) \propto r^{-2}$.
From Equation (\ref{eq:vort}) we obtain the ansatz
\begin{equation}\label{eq:omsq}
\Omega^2(r,\theta) = \Omega^2(R_c)\left({r\over R_c}\right)^{-3} f(\cos\theta).
\end{equation}
Here $R_c$ is the radius inside of which ${\rm Co} \gtrsim 1$.
One subtlety here is that the stellar mass profile is not as centrally concentrated during the early
subgiant expansion. This generally leads to a shallower scaling of $\Omega$ with $r$:
taking $g(r) \propto r^{-\beta}$, with $\beta < 2$, we find $\Omega(r) \propto r^{-(1+\beta)/2}$. A fit to the gravity profile in
the inner convective envelope over the range of expansion that is probed by the {\it Kepler} asteroseismic data gives
$\beta \sim 1$, corresponding to $\Omega(r) \propto r^{-1}$. The monopolar scaling of $g$ is approached
following the first dredge-up, and is maintained near
the tips of the RGB and AGB.
We imagine that the rotation profile organizes at ${\rm Co} \gg 1$ into
axially symmetric rolls. In other words, the strong latitudinal dependence of the Coriolis force is
hypothesized to translate into a persistent latitudinal entropy gradient. If, furthermore, the entropy
is conserved along radially extended convective plumes, then we can decompose the
entropy profile into low-order spherical harmonics, $S(r,\theta,\phi) = S_0 + \delta S_0 P_\ell(\cos\theta)$.
Substituting this into Equation (\ref{eq:vort}) along with Equation (\ref{eq:omsq}), and normalizing
$f = 1$ at $\theta = 0$, we get
\begin{eqnarray}
f &=& 1\quad (\ell = 2); \quad\quad f = {21\over 20}\cos^2\theta - {1\over 20}\quad (\ell = 4);\mbox{} \nonumber \\ \mbox{}
f &=& {1\over 280}\left(495\cos^4\theta - 234\cos^2\theta + 19\right) \quad (\ell = 6).
\end{eqnarray}
Angular and radial gradients in $\Omega$ are generally comparable in this type of flow (excepting the simplest
case of a quadrupolar entropy pattern). We note that rotation is fastest at the pole for a given
$\ell$ (corresponding to an `anti-solar' rotation profile), but that the opposing gradient can be
obtained by superposing $\ell = 2$ with any higher harmonic. A small anti-rotation is also present in a narrow
latitudinal band when the entropy follows a single spherical harmonic.
A caveat here involves our neglect of the Lorentz force, which may have a considerable influence
near the boundary of a convection zone, but does not obviously compete with the right-hand side of
equation (\ref{eq:vort}) within a deep convective layer. In particular, the magnetorotational instability
(MRI) is ineffective in the outer part of the envelope where ${\rm Co} < 1$, because the growth time of a MRI mode
is longer than the convective time.
\subsection{Rapid Mixing Between Upflows and Downflows}
We now turn to consider the consequences of rapid mixing between upflows and downflows
for the rotation profile in parts of the envelope that reach ${\rm Co} \sim 1$. This provides
a useful comparison with the results of Section \ref{s:extend}, and demonstrates how the
radial angular velocity gradient is tied to the presence of radially extended convective flows.
We show that the downward advection of angular momentum cannot push the flow to ${\rm Co} \gg 1$
in the inner part of an extended envelope, where $\tau_{\rm con}$ {\it decreases} inward. This
result has a simple interpretation: the competition between the Coriolis force and the inertial force
limits ${\rm Co}$ to a value of the order of unity.
In the mixing length approximation, the entropy difference between upflows and downflows
is related to the speed of a convective eddy by (e.g. \citealt{bethe90})
\begin{equation}
\delta S = {\delta h\over k_{\rm B}T} = {C_p\over C_p-C_v}{v_{\rm con}^2 \over k_{\rm B} T}.
\end{equation}
Here $\delta h$ is the perturbation to specific enthalpy and $T$ is temperature. Then the
angular entropy gradient can be estimated as $\partial S/\partial\theta \sim (r/\ell_P)\delta S$.
Because it turns out that ${\rm Co} = O(1)$, we cannot take Equation (\ref{eq:vort}) literally, but
can still use it to estimate the magnitude of the shift of angular velocity across a scale height:
\begin{equation}\label{eq:mix}
\delta(\Omega^2) \sim {g(r)\over \ell_P} {v_{\rm con}^2\mu\over k_{\rm B} T} =
\left({v_{\rm con}\over \ell_P}\right)^2.
\end{equation}
We look for a rotation profile that diverges inward, so that the constant term in $\Omega$ can be neglected.
Then we find
\begin{equation}
\Omega(r) \sim {v_{\rm con}(r)\over \ell_P}.
\end{equation}
As advertised, the Coriolis parameter does not rise much above unity. (Faster rotation can, of course,
still be obtained by superposing uniform rotation on the flow.)
For example, in a nearly adiabatic envelope with $g(r) \propto r^{-2}$,
the speed of an eddy scales as $v_{\rm con} \propto (\rho r^2)^{-1/3}$.
Then $\rho(r) \propto r^{-3/2}$ implies $\Omega(r) \propto r^{-7/6}$ for $\gamma = 5/3$,
which is shallower than the profile (\ref{eq:omsq}).
\subsection{Model $\bar\Omega(r)$ Profiles}
A simple parameterization of the rotation profile which incorporates these lessons may be constructed
as follows. Where the Coriolis force can be neglected, at $r > R_c$, we consider a flat distribution
of specific angular momentum,
\begin{equation}\label{eq:omouter}
\bar\Omega(r) = \bar\Omega(R_c) \left({r\over R_c}\right)^{-2}; \quad r > R_c.
\end{equation}
Here $\bar\Omega$ is the angular velocity averaged over a spherical shell. If the angular momentum of
the star is small enough that $\bar\Omega(r) \tau_{\rm con} < 1$ throughout the envelope, then we identify
$R_c$ with the base of the envelope,
\begin{equation}
R_c = R_{\rm benv}; \quad (\Omega \tau_{\rm con} < 1\;{\rm at} \; R_{\rm benv}).
\end{equation}
When the star has more angular momentum, we maintain the profile (\ref{eq:omouter}) in the
outer part of the convection zone, and consider
\begin{eqnarray}\label{eq:ominner}
({\rm I}) \quad\quad \bar\Omega(r) &=& \bar\Omega(R_c) \left({r\over R_c}\right)^{-\alpha} \mbox{} \nonumber \\ \mbox{}
&=& {{\rm Co}_{\rm trans}\over \tau_{\rm con}(R_c)} \left({r\over R_c}\right)^{-\alpha}; \quad
R_{\rm benv} < r < R_c\mbox{} \nonumber \\ \mbox{}
\end{eqnarray}
in the inner envelope. Here ${\rm Co}_{\rm trans} = O(1)$. As the angular momentum increases $R_c$ moves outward, and may reach the outer
radius of the star, in which case ${\rm Co}$ is larger than unity everywhere in the envelope.
Following the above discussion, we will explore power-law indices $1 < \alpha < 3/2$,
with the preferred value depending on the radial scaling of gravity.
Such a rotation profile corresponds to transport of energy by radially extended, adiabatic plumes.
A second profile, based on mixing length theory, is obtained by setting ${\rm Co}$ to a constant everywhere inside the radius $R_c$,
\begin{equation}\label{eq:ominnerb}
({\rm II}) \quad\quad \bar\Omega(r) =
{{\rm Co}_{\rm trans}\over \tau_{\rm con}(r)} + \Omega_0; \quad R_{\rm benv} < r < R_c.
\end{equation}
The most important difference with profile I is that ${\rm Co}$ remains pegged at ${\rm Co}_{\rm trans} \sim 1$
in the absence of the constant term. Setting $\Omega_0 = 0$
strictly limits the angular momentum which the envelope may contain, a bound which is easily violated
during core helium burning. We also require
that $R_c$ is bounded above by the radius at which $\tau_{\rm con}$ reaches a maximum; this prevents the formation
of a layer below the photosphere in which $\bar\Omega$ increases outward. See Figure
\ref{fig:rotational_profiles_limiting_cases}.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f1a.ps}
\vskip .2in
\plotone{f1b.ps}
\caption{Examples of rotation models I (top panel) and II (bottom panel), both taken from the $5\,M_\odot$ model
near the tip of the AGB. Each panel shows several rotation profiles corresponding to
different total stellar angular momenta. The magenta curve shows, for reference, the profile of $\tau_{\rm con}(r)$.
Black solid lines represent segments where $\Omega(r)\propto r^{-2}$ and red dotted lines $\Omega(r)\propto r^{-1}$.
Top panel: as $J_\star$ is increased, the transition radius $R_c$, for which ${\rm Co}(R_c)=1$, moves outward.
Bottom panel: profiles from bottom to top represent models with increasing $J_\star$, but now when $R_c$
reaches the radius of maximum $\tau_{\rm con}$, we begin to add a uniform $\Omega$ component (the last term in
Equation (\ref{eq:ominnerb})). This prevents the formation of an outer zone of positive $\partial\Omega/\partial r$.
Black curve: constant $j$, reaching ${\rm Co}=1$ at the base of the envelope; green curve: $\Omega(r) = v_{\rm con}/\ell_P$;
blue curves: constant $\Omega$ component that grows with increasing $J_\star$.
{\it Kepler} asteroseismic data favor model I over model II.}
\vskip .2in
\label{fig:rotational_profiles_limiting_cases}
\end{figure}
\subsection{Latitudinal Differential Rotation}\label{s:latdiff}
The latitudinal shift in rotation frequency (e.g. between pole and equator) has an important influence on the
growth of magnetic helicity in the growing core of an RGB or AGB star (Paper II). In rotation model I,
where ${\rm Co} > 1$ is maintained by radially extended convective flows, we deduce a strong latitudinal shift:
$\Delta\Omega/\bar\Omega \equiv (\Omega_{\rm eq}-\Omega_{\rm pole})/\bar\Omega \sim 1$.
A strong shift is also maintained in the mixing length approximation (rotation model II), as long as the envelope
rotates rapidly enough to maintain ${\rm Co} \sim 1$ at the base of the envelope. It is
seen in the more rapidly rotating simulations of \cite{brunp2009}.
We note that weaker latitudinal differential rotation
is expected in a compact convective envelope with $v_{\rm con} \ll \Omega r\sin\theta$, as
is encountered in the Sun, $\Delta\Omega/\bar\Omega \propto {\rm Co}^{-2}$.
\subsection{Matching of Rotation between \\ Core and Convective Envelope}\label{s:couple}
The rotation of core and envelope are easily coupled to each other when the inner envelope retains enough angular
momentum to sustain a large-scale hydromagnetic dynamo. The details of this coupling depend on whether the core
grows, or recedes, as measured in the radial mass coordinate. When the core is growing, the instantaneous state
of the dynamo is the main consideration, because magnetic fields that are amplified near the core-envelope boundary
are advected downward. When the core is receding in mass (as it does when a star first ascends the giant branch during the first dredge-up phase),
the coupling also depends on a prior dynamo process that implanted a magnetic field in the core material.
A key consideration is the magnitude of the
drift speed $v_r - dR_{\rm benv}/dt$ of stellar material with respect to the core-envelope
boundary (averaged if necessary over thermal pulsations, which are present on the terminal AGB). Solid rotation is enforced
in the outer core during downward drift if the radial Alfv\'en speed is larger than this drift speed. This requires
that the large-scale poloidal magnetic field that is generated in the inner envelope be strong enough that
\begin{equation}\label{eq:couple}
{B_r^2\over 4\pi \rho(R_{\rm benv}) v_{\rm con}^2} \gtrsim {(v_r-dR_{\rm benv}/dt)^2\over v_{\rm con}^2}.
\end{equation}
We show in Figure \ref{fig:drift} both the drift speed relative to the core-envelope boundary, and
the convective speed in the inner envelope (as evaluated using mixing-length theory),
during the post-MS evolution of a $1.5\,M_\odot$ model star. The same quantities are also shown
for a $5\,M_\odot$ star near the tip of the AGB. (Further details of these models are described in Sections \ref{s:1.5Msun} and \ref{s:rotation}.)
\begin{figure}[!]
\epsscale{1.0}
\plotone{f2a.ps}
\vskip .2in
\plotone{f2b.ps}
\caption{Flow speed of stellar material with respect to the core-envelope boundary, compared with convective speed in the inner envelope.
Upper panel: $1.5\,M_\odot$ model from the early subgiant expansion to the thermally pulsating AGB.
Bottom panel: $5\,M_\odot$ model at the tip of the AGB.
The logarithmic abscissa in the upper panel is referenced to the time $t_{\rm TAGB}$ of envelope ejection.
The time coordinate in the lower panel has been shifted to focus on the final few hundred years.}
\vskip .2in
\label{fig:drift}
\end{figure}
One sees that during the early subgiant expansion of the $1.5\,M_\odot$ model, the remnant poloidal magnetic field
in the outer core need only retain an energy density that exceeds $\sim 10^{-16}-10^{-15}$ of the instantaneous
convective energy density in order to compensate changes in rotation induced by the evolving mass profile.
Magnetic fields $\sim 10^2$-$10^3$ times stronger are needed to compensate
a latitude-dependent convective stress (Paper II).
The envelope decreases more rapidly in mass near the tip of the RGB and AGB,
and the magnetic energy density that is self-consistently generated during these phases must rise to about
$\sim 10^{-8}-10^{-6}$ of the convective energy density.
We conclude that magnetic decoupling of core from
envelope is most easily achieved during the brief phase where the envelope is expelled.
In Section \ref{s:1.5Msun}
we concentrate on the comparison with {\it Kepler} asteroseismic data, which only covers post-MS stars more
compact than $\sim 10 R_\odot$. These subgiants and core He burning stars evolve on a relatively long timescale ($\sim 10^2$ times
the duration of the thermally pulsating AGB phase). The results shown in
Figure \ref{fig:drift}
strongly suggest that the rotation of core and envelope remain well coupled
in these stars.
We therefore take the core to rotate as a solid with the mean rotation at the base of the envelope,
\begin{equation}\label{eq:omcore}
\Omega_{\rm core} = \Omega_{\rm benv} \equiv \bar\Omega(R_{\rm benv}).
\end{equation}
An explanation for the rotation behavior of the {\it Kepler} sample must be found
in the redistribution of angular momentum within an extended convective envelope.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f3.ps}
\caption{Rotation period at the base of the convective envelope, in a $M_{\rm ZAMS}=1.5M_\odot$ model star, as a function of time during the RGB
and early core He burning phases. Curves correspond to different rotation profiles. Black line: rotation period
$P_{\rm benv}$ corresponding to ${\rm Co_{benv}} = 1$ (with $\tau_{\rm con}$ evaluated
in the mixing-length approximation). Green dashed line: solid rotation. Red dotted line: uniform
specific angular momentum (even where ${\rm Co} > 1$). Blue long-dashed line: rotation model I given
by Equations (\ref{eq:omouter}), (\ref{eq:ominner}) with $\alpha = 1$. Magenta dashed-dotted line: rotation model I with $\alpha = 3/2$. Vertical black lines show the age range probed by the {\it Kepler} subgiant measurements,
with corresponding stellar radius.}
\vskip .2in
\label{fig:rotation_profiles}
\end{figure}
\section{Comparison of $1.5\,M_\odot$ Model Star \\ with {\it Kepler} Data}\label{s:1.5Msun}
We begin by calibrating the rotation models described by Equations (\ref{eq:omouter}) and
(\ref{eq:ominner})-(\ref{eq:omcore}) against the core rotation periods of $1-2\,M_\odot$ subgiants and post-core He
flash stars as measured by \cite{becketal2012} and \cite{mossetal2012}. For ease of comparison, we
use the same stellar model as \cite{CantMBCP2014}, namely a $1.5\,M_\odot$ star of solar
metallicity with zero-age main sequence (ZAMS) equatorial rotation speed 50 km s$^{-1}$.
Our calculation also employs the
1D stellar evolution code MESA (\citealt{paxtetal2011}, version 5527),
but with the built-in prescriptions for angular momentum transport turned off. Instead, the
rotation profile is `patched' onto the sequence of stellar models.
The total spin angular momentum of the
star is evolved self-consistently in response to mass loss, using the prescription described
in Section \ref{s:angloss}. Magnetic wind torques are neglected, for the reasons described
in Section \ref{s:spindown}. The net effect of mass loss is found to be modest up to the onset of core
He burning: a factor $\sim 2$ reduction in angular momentum during the RGB phase.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f4a.ps}
\vskip .2in
\plotone{f4b.ps}
\caption{Rotation period at the base of the convective envelope during the early giant expansion
and the core He burning phase of a $1.5\,M_\odot$ star (upper panel) and a 1 $M_\odot$ star (lower panel).
Rotation model I given by Equations (\ref{eq:omouter}), (\ref{eq:ominner}) with ${\rm Co}_{\rm trans} = 1$ and
varying inner index $\alpha$ (here $\Omega(r) \propto r^{-\alpha}$ inside the radius $R_c$ where
${\rm Co} = {\rm Co}_{\rm trans}$).
The 1.5$ M_\odot$ star is given a ZAMS equatorial rotation speed 50 km s$^{-1}$, while the 1 $M_\odot$
star is given the solar rotational angular momentum.
Diagonal black line: fit to {\it Kepler} subgiant data from \cite{CantMBCP2014}.
The radius at which majority of helium clump stars are found is marked in the upper panel.}
\vskip .2in
\label{fig:Kepler_comparison}
\end{figure}
\begin{figure}[!]
\epsscale{1.0}
\plotone{f5.ps}
\caption{Same as upper panel of Figure \ref{fig:Kepler_comparison}, but now for various initial rotation speeds of the
$1.5\,M_\odot$ star, and two selected scalings for $\Omega(r)$ in the rapidly rotating inner zone.}
\vskip .2in
\label{fig:Kepler_comparison2}
\end{figure}
\begin{figure}[!]
\epsscale{1.0}
\plotone{f6.ps}
\caption{Profile of Coriolis parameter with radius during two phases of the evolution of our $1.5\,M_\odot$ model
with initial $v_{\rm rot} = 50$ km s$^{-1}$. Black curve: subgiant (stellar radius $6\,R_\odot$) with inner rotation
index $\alpha=1$ in Equation (\ref{eq:ominner}). Colored curves: near the tip of the AGB (stellar radius
$100\,R_\odot$) with uniform specific angular momentum in outer envelope, and inner rotation profile corresponding
to either $\alpha = 1$ or 1.5, or to ${\rm Co}$ saturating at unity.}
\vskip .2in
\label{fig:Co_profile}
\end{figure}
We observe from Figure \ref{fig:rotation_profiles} that the inner envelope must rotate rapidly (${\rm Co}_{\rm benv} > 1$),
even in the case of solid rotation, over the range of expansion probed by the {\it Kepler} data. The rotation period
$P_{\rm benv} = 2\pi/\Omega_{\rm benv}$ at the base of the envelope would actually decrease with increasing expansion if the entire envelope
could sustain a profile $\Omega(r) \propto r^{-2}$.
This effect is still present in the less peaked rotation profile I, as given by Equations (\ref{eq:omouter}), (\ref{eq:ominner}) with
$\alpha = 1$ (or 3/2). Now $P_{\rm benv}$ sits close to the observed range of $10-20$ days. This period is identified here with the core rotation
period.
Another important feature is the large jump in rotation period between the red giant and core He burning phases (Figure
\ref{fig:Kepler_comparison}). The latter, slower, rotation results from a retraction of the convective envelope combined with
continuing angular momentum pumping and core-envelope synchronization. (It was previously ascribed by \cite{CantMBCP2014} to the
conservation of angular momentum by an expanding core that has already decoupled rotationally from the inner envelope.)
A closer look at the effect of the inner rotation index $\alpha$, and the initial rotation speed of the star, is provided by Figures
\ref{fig:Kepler_comparison} and \ref{fig:Kepler_comparison2}. Here one sees that an index $\alpha$ around $1-1.1$ provides a close agreement
with the {\it Kepler} measurements for stars more massive than $\sim 1.3\,M_\odot$. There is some degeneracy between the effects of changing $\alpha$ and initial rotation speed. It is interesting to note that this profile agrees well with the analytic scaling derived
for rotation model I in Section \ref{s:extend} in the case (appropriate to the expansion phase probed by the {\it Kepler} data) where $g(r)
\propto r^{-1}$ in the inner envelope. Recall that this rotation model represents radially extended upflows and downflows in a geometrically
deep, adiabatic envelope.
A different conclusion emerges for stars close to solar mass. We see in the bottom panel of Figure \ref{fig:Kepler_comparison} that the core of a $1\,M_\odot$ star without a close planetary companion is predicted to
spin $\sim 5-10$ times more slowly during subgiant expansion than is seen in our $1.5\,M_\odot$ model. On the other hand, the core rotation during the He burning phase is only $\sim 3$ times slower, and indeed remains within the range measured by {\it Kepler} for clump stars. Significant spindown by magnetic wind
torques during the subgiant phase \citep{saders13} would have similar effects.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f7a.ps}
\vskip .2in
\plotone{f7b.ps}
\caption{Profile of convective envelope during RGB expansion (top panel) and AGB expansion (bottom panel)
in rotation model I.
As the aspect ratio of stellar radius to the base of the envelope grows, the domain within which
${\rm Co} \gtrsim 1$ shrinks. The radius $R_c$ marks the transition point between an outer zone
with uniform specific angular momentum, and an inner zone with $\Omega(r) \propto r^{-\alpha}$.
This transition depends only weakly on the inner index $\alpha$, because most of the angular momentum
is stored in the outer envelope.}
\vskip .2in
\label{fig:Rc_profile}
\end{figure}
\subsection{Transition from ${\rm Co} \gtrsim 1$ to ${\rm Co} \lesssim 1$ throughout \\
Bulk of Convective Envelope}\label{s:partition}
The {\it Kepler} sample of subgiants and core He burning stars probes a different rotational regime than is
experienced by most giants near their maximum expansion. The {\it Kepler} stars are compact enough that
we find ${\rm Co} \gtrsim 1$ throughout most of the convective envelope, excepting a relatively thin
layer below the surface. Indeed ${\rm Co}$ reaches a considerable value $\sim 10-30$ at the base of
the envelope (Figure \ref{fig:Co_profile}). The measured core rotation rates probe the angular velocity
profile that is sustained in the envelope in a regime of relatively rapid rotation.
Figure \ref{fig:Rc_profile} shows the evolving partition between an outer zone with uniform specific angular
momentum, and an inner zone with angular velocity parameterized by Equation (\ref{eq:ominner}).
These curves correspond to a star with spin angular momentum close to its birth value, assuming
an equatorial surface rotation speed of $50$ km s$^{-1}$.
Out to a stellar radius $R_\star \sim 10\,R_\odot$, the envelope remains compact enough that ${\rm log}(R_\star/R_c)
\lesssim \log(R_c/R_{\rm benv})$. Beyond this size, the inner rotation rate is controlled by the rotation
profile at ${\rm Co} \lesssim 1$.
The spin and magnetization of the WD remnant, as explored in Paper II,
are strongly influenced by the $\Omega(r)$ profile of a {\it slowly}
rotating convective envelope near the tip of the RGB and AGB.
Then the gravity profile steepens in the inner envelope, and we
maintain consideration of rotation profiles at ${\rm Co} > 1$ that
are somewhat steeper than suggested by the fit to {\it Kepler} data:
$\alpha \sim 3/2$ as opposed to $\alpha \sim 1$ in Equation (\ref{eq:ominner}).
\begin{figure}[!]
\epsscale{1.0}
\plotone{f8.ps}
\caption{Green crosses: observed surface rotation speeds of class IV stars from \cite{schrijver93}, multiplied by
$4/\pi$ to account for unknown inclination, as a function of effective temperature. We show for comparison
our model of a $1.5\,M_\odot$ star on the early red giant branch, with initial rotation speed $50$ km s$^{-1}$.
Curves correspond to different profiles of rotation in the envelope, combined with solid rotation in the core.
Solid rotation is enforced in the envelope when the aspect ratio $R_{\rm benv}/R_\star$ between the base
of the envelope and stellar photosphere is larger than $0.9$ or $0.5$. The $\alpha=1, 1.5$ curves correspond
to rotation model I. In the case $\alpha = 1$ (which best reproduces the asteroseismic fits to
{\it Kepler} data) there is a factor $\sim 1/2$ reduction in surface rotation compared with a solid rotator at
$\log(T_{\rm eff}) < 3.8$.}
\vskip .2in
\label{fig:vsurf}
\end{figure}
\vfill
\subsection{Surface Rotation of Subgiants: \\ Angular Momentum Pumping Versus Magnetized Winds}\label{s:spindown}
The expansion of a star following the MS phase is associated with a drop in surface rotation. The
strength of this drop is sensitive to the distribution of angular momentum within the star, as well as
to external torques associated with magnetic activity in the deepening envelope. \cite{schrijver93}
and \cite{saders13} have
presented evidence for an inconsistency between solid rotation in the envelope and measurements in the
surface rotation of $1.5-2\,M_\odot$ stars: the measured rotation is a factor $\sim 2-3$ slower than is
expected for constant stellar angular momentum.
A similar effect would be imposed by inward pumping of angular momentum. \cite{schrijver93} show
that a fast-spinning core has a negligible effect on the result if the convective envelope maintains solid rotation,
but did not consider envelopes with inward-peaked rotation profiles. The key point here is that the radiative
core becomes very compact early in the expansion onto the red giant branch, in the sense that its
moment of inertia is a very small fraction of $M_{\rm benv}R_{\rm benv}^2$. Therefore
inhomogeneous rotation in the envelope has a much larger effect on surface rotation than does a fast-spinning core.
We show in Figure \ref{fig:vsurf} the evolution of the surface rotation in our $1.5\,M_\odot$ model,
for the same set of inhomogeneous rotation profiles that we have previously explored.
The envelope at high $T_{\rm eff}$ is still relatively thin. We therefore
impose $\partial\Omega/\partial r = 0$, to mimic the solar rotation profile,
when the aspect ratio $R_{\rm benv}/R_\star$
rises above a critical value 0.5 or 0.9. The profile that is favored by the {\it Kepler} core rotation
data, Equations (\ref{eq:omouter}) and (\ref{eq:ominner}) with inner index $\alpha = 1$, shows a factor $\sim 1/2$
reduction in surface rotation compared with the solid rotator.
This removes a significant part of the
discrepency found by \cite{schrijver93} and \cite{saders13}, but perhaps not all.
\section{Evolving Rotation Profile with Mass Loss and Interaction with a Planet} \label{s:rotation}
We now consider the rotational evolution of a star as it approaches the tip of the RGB and AGB.
Here the probability of interacting with a companion increases significantly.
The Coriolis parameter also drops markedly at the base of the convective envelope. We maintain our
focus on a two-layered rotation model.
Two test cases are analyzed in detail: a star with initial mass $M_{\rm ZAMS} = 1\,M_\odot$, which
experiences strong spindown on the MS and interacts with a planet; and an intermediate-mass star
($M_{\rm ZAMS} = 5\,M_\odot$) which remains rapidly rotating at the end of the MS and leaves behind a massive WD.
This second stellar model is also allowed to interact with a companion.
MESA is used to evolve these stellar models (both of solar metallicity)
from the ZAMS to the post-AGB phase.
In the case of the solar-mass star, the interaction with a planet was followed from the
base of the RGB, when the star has expanded to a radius $R_{\star} \sim 10.9R_{\odot}$ (age $t_{10.9} \sim 12.37$ Gyr).
At that point, the spin angular momentum was set equal to that of the present Sun.
Stellar quantities at times intermediate between the outputed MESA models were obtained by linear interpolation.
Details of the orbital integration and the interaction between planet and star are presented in Section \ref{s:orbital_evolution}.
The equatorial rotation speed of the $5\,M_{\odot}$ star is set to 50 km s$^{-1}$ at the ZAMS,
corresponding to the peak in the measured $v \sin i$ distribution \citep{wolfsdv2007}. Both the $1\,M_\odot$ and $5\,M_\odot$
models are taken to rotate as a solid body on the MS. Deviations from solid body rotation develop after the formation
of a deep convective envelope, following the approach of Section \ref{s:pump}.
There are instances around the tip of the AGB when the outer envelope splits into several convection zones,
separated by radiative layers. The entropy gradient in these radiative zones, although positive, remains small.
We impose a matching of specific angular momentum across them.
\subsection{Loss of Angular Momentum on the RGB and AGB}\label{s:angloss}
Magnetic wind braking is not taken into account in our simulations
during the subgiant or giant phases, for the reasons given in
Section \ref{s:spindown}. Then the ejected mass carries the specific
angular momentum of the stellar photosphere,
\begin{equation}\label{eq:djdtml}
\dot{J}_{\star} = \frac{2}{3} \bar\Omega(R_{\star}) R_{\star}^2 \dot{M}_{\star}.
\end{equation}
The rate of mass loss on the RGB is handled using the prescription of \cite{reim1975},
\begin{equation}
\dot{M}_{\star\,\rm Reimers} = 4 \times 10^{-13} \eta_R \frac{\tilde{L}_{\star} \tilde{R}_{\star}}{\tilde{M}_{\star}} \quad M_{\odot}~{\rm yr^{-1}},
\end{equation}
with $\eta_R = 0.5 $. Here (\textasciitilde) denotes stellar parameters measured in solar units. On the AGB we use
\begin{equation}
\dot{M}_{\star\,\rm Bloecker} = 1.93 \times 10^{-21} \eta_B \frac{\tilde{L}_{\star}^{3.7}\tilde R_\star}{\tilde{M}_{\star}^{3.1}}
\quad M_{\odot}~{\rm yr}^{-1}
\end{equation}
from \cite{bloe1995}. Mass loss according to the Bl\"{o}cker formula is turned on when the central H and He
abundances are less than $10^{-2}$ and $10^{-4}$, respectively.
The choice of AGB mass loss parameter $\eta_B$ is motivated by recent cluster data (e.g. \citealt{Dobbetal2009}).
As we discuss further in Paper II, we set $\eta_B = 0.05$, corresponding to a final WD mass $M_{\rm WD} \sim 0.87M_{\odot}$ (from the
$5M_\odot$ progenitor).
The same wind parameters are used for both the $1\,M_\odot$ and $5\,M_\odot$ models. The solar-type star loses $\sim0.25M_{\odot}$ on the
RGB and another $\sim0.21M_{\odot}$ on the AGB for a final WD mass $M_{\rm WD} \sim 0.54M_{\odot}$.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f9.ps}
\caption{Evolution of spin angular momentum of $1\,M_\odot$ model star during RGB, core He burning and final AGB phases. Curves represent
cases described in the text. Red lines ($J_{\rm MS}$): no external source of angular momentum. Black lines ($J_{\rm Jupiter, a_i=1AU}$):
absorption of a Jupiter-mass planet orbiting initially at $a_i = 1$ AU. Short dashed lines ($\Omega=c$): solid
rotation throughout core and
envelope. Other lines correspond to two-layer rotation model, with uniform specific angular momentum in outer envelope,
and varying inner rotation profile. Solid (long-dashed) curves: $\alpha=1$ ($\alpha=1.5$) in Equation (\ref{eq:ominner});
long-dashed-dotted curve: profile (\ref{eq:ominnerb}) with ${\rm Co} = 1$ in inner zone.}
\vskip .2in
\label{fig:J_star_comparison_RGB_AGB}
\end{figure}
\begin{figure}[!]
\epsscale{1.0}
\plotone{f10a.ps}
\vskip .2in
\plotone{f10b.ps}
\caption{Evolution of the Coriolis parameter at the base of the convective envelope. Curves correspond to those in Figure
\ref{fig:J_star_comparison_RGB_AGB}.
Upper panel shows the RGB, core helium burning phase, and AGB. Lower panel shows an expanded view of the tip of the RGB and the beginning of
the core He burning phase.}
\vskip .2in
\label{fig:Co_base_comparison}
\end{figure}
\begin{figure}[!]
\epsscale{1.0}
\plotone{f11a.ps}
\vskip .2in
\plotone{f11b.ps}
\caption{Effect on the 5$M_\odot$ model star of the engulfment of a substellar companion of mass
(3,10,30,100)$M_{\rm Jupiter}$ from an initial orbit $a_i = 2$ AU. Orbital angular momentum of planet is added to
star when it reaches a radius $R_\star = 200R_\odot$ on the AGB, as guided by more detailed tidal evolution calculations.
The physical consequences depend weakly on the epoch of engulfment, if it occurs before the first thermal pulse
(at $t \sim 1.0717\times10^8$ years), when the envelope dynamo begins to contribute to the magnetic field of the
hydrogen-depleted core. Upper panel: time evolution of transition radius $R_c$ between outer slowly rotating zone
(${\rm Co} < 1$) and inner rapidly rotating zone (${\rm Co} > 1$). Bottom panel: change in $\rm Co_{benv}$ due to the
added angular momentum. Rotation profile corresponds to $\alpha=1$ in Equation (\ref{eq:ominner}).
Although $R_c$ increases with increasing angular momentum, $\rm Co_{benv}$ decreases.}
\vskip .2in
\label{fig:R_c_and_Co_benv_for_M_companion}
\end{figure}
\begin{figure}[!]
\epsscale{1.0}
\plotone{f12.ps}
\caption{Same as bottom panel of Figure \ref{fig:R_c_and_Co_benv_for_M_companion}, but now with steeper inner
rotation profile $\alpha=1.5$ in Equation (\ref{eq:ominner}). Now ${\rm Co}_{\rm benv}$ increases monotonically
with the added angular momentum.}
\vskip .2in
\label{fig:R_c_and_Co_benv_for_M_companion2}
\end{figure}
\subsection{Rotation Rate of the Inner Envelope and Core} \label{s:RG_rotation}
There is a simple relation between spin angular momentum and core angular velocity $\Omega_{\rm benv}$
when the envelope is extended and maintains uniform angular momentum per unit mass,
\begin{equation}
j(r) \equiv \frac{dJ(r)}{dM(r)} = \frac{2}{3} r^2 \bar\Omega(r) = {\rm const}.
\end{equation}
Given a net stellar angular momentum $J_\star$, and an effective moment of inertia in the convective envelope,
\begin{equation}
I_{\rm eff} = \frac{2}{3} \int 4\pi r^2 R_{\rm benv}^2 \rho(r) dr = {2\over 3}R_{\rm benv}^2 (M_\star - M_{\rm benv}),
\end{equation}
we obtain
\begin{equation} \label{eq:omega_surface}
\Omega_{\rm benv} = \frac{J_{\star}}{I_{\rm eff} + I_{\rm core}}.
\end{equation}
Excepting near the point where the envelope has been ejected at the end of the AGB,
the moment of inertia of the core can generally be neglected, $I_{\rm core} \ll I_{\rm eff}$.
The angular momentum stored in the core is enhanced by a factor $\sim (R_\star/R_{\rm benv})^2$, as
compared with the case of solid rotation throughout the star.
This simple rotation model breaks down when it predicts ${\rm Co} \gtrsim 1$ at the base of the convective envelope,
and we employ the modified rotation profiles (\ref{eq:ominner}), (\ref{eq:ominnerb}).
Considering first the $1\, M_\odot$ model, we compare several scenarios. The star is either assumed to evolve in isolation;
or to interact with a Jupiter-mass planet that starts in a circular orbit of semi-major axis 1 AU. The internal rotation
profile is either described by a two-layer model (Section \ref{s:RG_rotation}) during its post-MS evolution (`$j=c$' in the outer
envelope, with various inner $\Omega(r)$ profiles); or it is assumed to be solid-body from the ZAMS to the terminal AGB (`$\Omega=c$').
Figure \ref{fig:J_star_comparison_RGB_AGB} shows the evolution of $J_\star$. We see that solid-body rotation
leads to a significant loss of angular momentum through the RGB winds, because little angular momentum is stored deep in the star.
Introducing an outer zone of uniform specific angular momentum allows only ${\rm 1/2}$ of the
angular momentum present at the end of the MS to be lost, with the remainder retained at the onset of core He burning. The engulfment
of the Jupiter greatly augments $J_\star$, but if $\Omega$ is uniform then most of this additional angular momentum is lost to RGB winds.
The most important consideration for us is whether ${\rm Co}_{\rm benv}$ reaches the threshold for dynamo activity. The time
dependence of ${\rm Co}_{\rm benv}$ is shown in Figure \ref{fig:Co_base_comparison}. Only in the case of planet absorption with `$j=c$'
in the outer envelope does
the star comfortably pass this threshold. Enough angular momentum is absorbed from a Jupiter-mass planet that almost any profile
$\Omega(r) \propto r^{-\alpha}$ with $\alpha > 0$ will allow the star to attain ${\rm Co}_{\rm benv} = 1$ on the RGB.
Stronger angular momentum pumping is required near the tip of the AGB to sustain rapid
rotation: uniform rotation combined with absorption of a Jupiter does not suffice.
An isolated $M_{\rm ZAMS} = 1\,M_\odot$ star without a planetary companion does not attain ${\rm Co}_{\rm benv} \sim 1$ in its
inner envelope near the tip of the RGB or AGB, no matter the rotation profile. This dynamo threshold is missed by more than two orders
of magnitude in the case of solid rotation.
Angular momentum could also be transferred to an evolved giant by a tidal interaction with a stellar companion. This happens
inevitably during the evolution to a common envelope and the formation of pre-CV system. For example, the $5M_\odot$ star may easily engulf a
companion during the AGB phase, given the large maximum radius reached by the star ($\sim 600R_\odot$). We consider the
effect of injecting a companion of mass (3,10,30,100)$M_{\rm Jupiter}$ from an orbit with initial semi-major axis
$a_i=2$AU when the star reaches a radius $R_{\star}=200R_\odot$ on the AGB.
We also consider a range of inner rotation profiles, $\alpha = 1$ to $1.5$ in Equation (\ref{eq:ominner}).
Figure \ref{fig:R_c_and_Co_benv_for_M_companion} shows that the added angular momentum pushes $R_c$ outward, as expected;
but when $\alpha = 1$ the relative flatness of the $\tau_{\rm con}(r)$ profile ($\tau_{\rm con}$ increasing roughly in proportion to $r$)
may actually cause $\rm Co_{benv}$ to decrease. In other words, the engulfment of a more massive
companion may result in slower rotation at the base of the convective envelope.
Increasing the rotation index to $\alpha=1.5$ allows $\rm Co_{benv}$ to grow with increasing injected angular momentum
(Figure \ref{fig:R_c_and_Co_benv_for_M_companion2}).
Our inference that angular momentum pumping in the envelopes of subgiants can sustain ${\rm Co}_{\rm benv} \sim 10$
(see Figure \ref{fig:rotation_profiles}) suggests that the rotation profile in the inner envelope may be significantly
steeper than $\alpha=1$ near the tip of the RGB and AGB. This would imply an even stronger magnetization of the WD remnant, as we show in Paper II.
\section{Orbital Evolution of a Planetary \\ Companion to a $1M_{\odot}$ Star} \label{s:orbital_evolution}
Our calculation of the interaction of a planet with an evolving giant star follows the orbit in some detail. We take
into account the dissipation of the tide raised on the star, and gravitational scattering off a fluctuating quadrupole moment.
The second effect, which has been previously considered in the context of massive neutron star binaries \citep{phinney92}, has a
weaker radial dependence than the tidal torque and therefore modestly increases the range of semi-major axis over which a
strong interaction between planet and star can take place.
We evolve the planet's orbit in response to these external forces, considering a range of masses (Earth, Neptune or Jupiter)
along with different initial semi-major axes ($a_i = 0.5,0.75,1,1.5,2$ AU) and eccentricities ($e_i = 0,0.1,0.5$).
The change in orbital angular momentum is deposited in the star and tracked self-consistently, including the effect of stellar mass
loss described by Equation (\ref{eq:djdtml}).
The tidal acceleration is calculated using \citep{hut1981}
\begin{equation}\label{eq:atide}
{d^2{\bf r}\over dt^2} = -3k_2 \frac{G M_p}{r^2} \left(\frac{R_{\star}}{r}\right)^5 \bigg[ (1 + 3\frac{\dot{r}}{r}\tau)\hat{r} -
(\Omega_{\rm env} - \omega) \tau \hat{\phi} \bigg].
\end{equation}
In the above expression, the position of the planet is followed in spherical coordinates, with $\omega$ its orbital angular
velocity and $M_p$ its mass. We set the rotation frequency $\Omega_{\rm env}$ of the stellar envelope\footnote{Equation (\ref{eq:atide}) is
derived assuming solid rotation in the envelope.} to zero. The time lag of the tidal bulge is given by
\begin{equation}
\tau \sim 2\frac{R_{\star}^3}{GM_{\star} \tau_{\rm con}}
\end{equation}
where
\begin{equation} \label{eq:convective_overturn_time}
\tau_{\rm con} \sim \left(\frac{M_{\rm env} R_{\star}^2}{L_{\star}}\right)^{1/3}
\end{equation}
is the global convective overturn time. The Love number $k_2$ is obtained (see Equations (11) of \citealt{zahn1989}
and (17) of \citealt{scha1982}),
\begin{equation}
k_2 = 20.5 \alpha^{4/3} \frac{R_{\star}^{1/3} \tau_{\rm con}}{M_{\rm env}} \int_{x_{\rm benv}}^1 x^{22/3} L_r^{1/3} \rho^{2/3} l_P dx.
\end{equation}
Here $\alpha = 2$ is the convective mixing length parameter, $M_{\rm env} = M_\star - M_{\rm benv}$ is the mass of the convective envelope of
the star, $x = r/R_{\star}$, $x_{\rm benv} = R_{\rm benv} /R_{\star}$, and $L_r$ is the luminosity at a given radius.
The acceleration from the convective quadrupole is calculated using the standard expansion of the
gravitational potential (see \citealt{murrd1999}),
\begin{eqnarray}\label{eq:vquad}
V_{\rm quad} &=& -\frac{G}{2r^5}\Bigl[(I_{yy}+I_{zz}-2I_{xx})x^2 \\ \nonumber
&+& (I_{xx}+I_{zz}-2I_{yy})y^2 + (I_{xx}+I_{yy}-2I_{zz})z^2 \\ \nonumber
&+& 6(I_{yz}\,yz + I_{xz}\,xz + I_{xy}\,xy)\Bigr].
\end{eqnarray}
(In practice we evolve the orbit of the planet in Cartesian coordinates.) The moment of inertia tensor $I_{ij}$
is normalized using the rms convective quadrupole
\begin{equation}\label{eq:Q_conv}
\frac{Q_{\rm con}}{M_{\rm env} R_{\star}^2} \;\simeq\; 7.41 \times 10^{-3} f {L_{\star\,37}^{2/3}R_{\star\,13}^{5/3}\over
(M_\star/M_\odot)(M_{\rm env}/M_{\odot})^{2/3}}.
\end{equation}
Here $f$ is a factor of order unity which depends on the stellar structure. This expression is obtained by integrating the density
perturbation caused by convection inside the star over the envelope; see the Appendix \ref{app:Q_conv} for a derivation.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f13.ps}
\caption{Red line: orbital semi-major axis of a Jupiter-mass planet interacting with a RGB star and
starting from a circular orbit with $a_i = 1$ AU. Black line: stellar radius. Spiral-in of planet just
below the tip of the RGB deposits angular momentum and triggers a dynamo process near the core-envelope boundary
(Paper II).}
\vskip .2in
\label{fig:R_star_and_a_RGB_AGB}
\end{figure}
At time intervals separated by $\tau_{\rm con}$, the diagonal components of the moment of inertia matrix for the quadrupole were chosen randomly,
with a magnitude uniformly distributed in the range [0, $Q_{\rm con}$]. The remaining components were then obtained by a
random rotation of the diagonal matrix. Given the random nature of this forcing effect, we ran ten realizations for each set of
initial conditions to obtain a statistical sample of results.
A planet that is exposed to the tidal acceleration (\ref{eq:atide}) and quadrupolar potential (\ref{eq:vquad}) will evolve
in semi-major axis and eccentricity. It can either collide with the envelope of the star or be ejected from the system.
Here we are interested primarily in initial conditions leading to a collision.
In such an instance, we assume that drag and dynamical
friction act instantaneously, resulting in the planet's engulfment and the transfer of its orbital angular momentum to the giant.
\subsection{Jupiter-mass Planet with $a_i = 1$ {\rm AU}}
Consider a Jupiter-mass planet starting at $a_i = 1$AU and $e_i = 0$. In Figure \ref{fig:R_star_and_a_RGB_AGB} we show one realization of its
orbital evolution, comparing semi-major axis $a$ with $R_\star$. In this case the planet is engulfed shortly before the star reaches the tip of
the RGB. We treat this as our fiducial model for a $M_{\rm ZAMS} = 1\,M_\odot$ star interacting with a planet.
The corresponding eccentricity evolution is shown in Figure \ref{fig:a_and_e_Jupiter_a0=1AU_e0=0}. The stochastic nature of the convective
forcing of orbital parameters is evident here.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f14.ps}
\caption{Expanded view of the orbital parameters of the Jupiter-mass planet
in the same realization shown in Figure \ref{fig:R_star_and_a_RGB_AGB}.
Engulfment of the planet is preceded by a stochastic growth in orbital
eccentricity, due to gravitational scattering off the convective quadrupole.}
\vskip .2in
\label{fig:a_and_e_Jupiter_a0=1AU_e0=0}
\end{figure}
\subsection{Other Planetary Configurations}
Figure \ref{fig:P_engulf_L_orb_0_and_non_0_Q_conv} shows the probability of a planet's engulfment as a function of the initial
(specific) orbital angular momentum. Results for the three planetary masses we consider are shown separately. The separate effect
of shutting off the convective quadrupole ($Q_{\rm con} \rightarrow 0$) is also shown.
The convective forcing has an especially
strong effect on the probability of engulfment for Earth-mass planets. In the case of a Jupiter-mass planet, there is a $\sim 20\%$
enhancement in the range of semi-major axes that results in a merger. These effects arise mainly from the excitation of eccentricity
and a corresponding reduction in periastron.
\begin{figure}[!]
\epsscale{1.0}
\plotone{f15.ps}
\caption{Probability of engulfment of planets of various masses (Earth, Neptune, Jupiter), as a function of the initial specific orbital
angular momentum.
Negative-sloping dash: planet scatters gravitationally off convective quadrupole of giant star. Positive-sloping dash: $Q_{\rm con}=0$.
Probability of engulfment is enhanced by interaction with the quadrupole, especially for lower-mass planets.}
\vskip .2in
\label{fig:P_engulf_L_orb_0_and_non_0_Q_conv}
\end{figure}
\section{Conclusions} \label{s:conclusions}
We have considered the inward pumping of angular momentum in a deep stellar convective envelope, with a
goal of defining a workable rotation model for giant stars.
Independent theoretical and observational lines of argument point to a profile that differs radically from
those obtained from stellar evolution codes such as MESA, in which the implementation of angular momentum transport
typically results in solid rotation within convective layers. The work presented
here and in Paper II also demonstrates
that much improved dynamo models must be implemented in order to capture basic features of stellar rotation, such as the relaxation to solid
rotation in radiative parts of a star, and the rotational coupling between a radiative core and the base of an extended convective envelope.
The rotation of evolving stars cannot be understood even in qualitative terms without including the magnetic field built up
by a hydromagnetic dynamo operating near the boundary between convective and radiative zones. Some features of this dynamo
are special to giant stars. First, the high radiative energy flux leads to a strong broadening
of the tachocline layer situated beneath the core-envelope boundary, which may facilitate mixing of core material
into the envelope. The growth of the core is also accompanied by an inward flux of magnetic helicity. These effects are investigated in
detail in Paper II, where we motivate a strong magnetic coupling between core and inner envelope when ${\rm Co}_{\rm benv} \gtrsim 0.1-0.3$.
When this condition is satisfied, a simple picture emerges: nearly solid rotation in the radiative parts of the
star that previously supported a dynamo, combined with an inhomogeneous rotation profile in a deep convective envelope,
whose slope depends on the local Coriolis parameter. For example, solid rotation in a radiative layer adjacent to
convective zone is facilitated by a tiny helical magnetic field. Differential rotation that is sourced
by a changing stellar mass profile during subgiant expansion can be significantly reduced if the poloidal magnetic
pressure exceeds $\sim 10^{-16}$-$10^{-15}$ of the convective pressure, increasing to $\sim 10^{-8}$
during the thermally pulsating AGB. Magnetic fields $\sim 10^2$-$10^3$ times stronger are required to
compensate a latitude-dependent convective stress that is partly transmitted through a tachocline layer (Paper II).
A magnetic field exceeding this minimal value is plausibly present in the radiative core of a subgiant:
erasure to a lower level would require a very efficient mechanism for eliminating magnetic helicity. Inhomogeneous
rotation could, in principle, be maintained by a supplementary mechanism of angular momentum transport such as
gravity wave damping.
Our results can be summarized as follows.
1. Strong inward pumping of angular momentum is well motivated by measurements of core rotation in stars of radius
$\lesssim 10\,R_\odot$ \citep{becketal2012,mossetal2012}, and by numerical simulation \citep{brunp2009}. A rotational decoupling between
core and envelope (as envisaged e.g. by \citealt{CantMBCP2014}) is difficult to achieve over the long duration of subgiant expansion,
depending on an unrealistically small poloidal magnetic field. We have found that a profile $\Omega(r) \propto r^{-1}$ within parts of the
envelope that are very rapidly rotating (${\rm Co} \gtrsim 1$), combined with solid rotation in the radiative core, provides an excellent fit to the data for stars more massive than $\sim 1.3\,M_\odot$. This corresponds to the profile
that is derived from quasi-geostrophic balance in an adiabatic envelope with heat transported by radially
extended convective upflows and downflows that do not mix significantly with each other over a scale height.
Given a gravity profile $g(r) \propto r^{-\beta}$ with $\beta \simeq 1$, we deduce $\Omega(r) \propto r^{-(1+\beta)/2}
= r^{-1}$.
More focused measurements of individual stars can in principle be used to test this rotation model -- in particular, to
constrain the range of $\Omega$ within the convective envelope -- and also to test for the existence of strong
differential rotation across the core-envelope boundary, as in the model of \cite{CantMBCP2014}. Efforts so far to
measure the full radial profile $\bar\Omega(r)$ have considered separate zones of solid rotation in core and envelope, as
well as smoother profiles; the fits are mainly sensitive to the rotation rate in the core \citep{deheuvels12,deheuvels14}.
Solar-mass stars which spin down
significantly on the MS, and do not absorb a planetary compansion, are predicted to retain much slower core rotation
during subgiant expansion, but to rotate within the range of rates inferred from {\it Kepler} observations of
core He burning stars. Although the {\it Kepler} subgiant sample is mainly composed of stars more massive than $1.3\,M_\odot$, this provides a sharp test of the rotation model presented here. On the other hand, we note that convergence in the core rotation of stars below and above the $1.3\,M_\odot$ threshold during subgiant expansion would be remarkable in almost any rotation model, given the significant
difference in rotation rate at the end of the MS.
2. Independent constraints on internal rotation in subgiants come from measurements of the slowdown in surface
rotation as stars of mass $\gtrsim 1.3\,M_\odot$ first ascend the giant branch. A good part of the discrepency,
found by \cite{schrijver93}, with stellar models which assume solid rotation, is removed by introducing the
same envelope rotation profile that is implied by the {\it Kepler}
asteroseismic data. A separate rapidly rotating core, combined with solid rotation in the envelope, has a negligible
effect on surface rotation because $I_{\rm core}$ is a very small fraction of $M_{\rm benv} R_{\rm benv}^2$ (and of the effective moment of
inertia of the envelope).
3. Stars near the tip of the RGB and AGB (radii $\gtrsim 10^2\,R_\odot$) probe a different regime of slow rotation that
is not directly constrained by the {\it Kepler} observations. Here we have considered steeper rotation profiles,
$j \sim$ const, in the outer parts of the envelope where ${\rm Co} \lesssim 1$. This stronger level of angular momentum pumping
allows rapid rotation to be sustained deep in the envelope even during the final stages of envelope ejection. It also limits the loss of
angular momentum to winds on the RGB and AGB. We find that the conditions
for a hydromagnetic dynamo are favorable in a star which rotates rapidly at the end of the MS (e.g. initial mass
$\gtrsim 1.3\,M_\odot$); or in a solar-mass star which absorbs a Jupiter (or even Neptune) at some point
during its post-MS expansion.
The gravity profile steepens significantly in the inner envelope following the first dredge-up phase, and approaches
$g(r) \propto r^{-2}$ near the tips of the RGB and AGB. If the convective upflows and downflows remain radially extended,
then our scaling solution to the vorticity equation implies that the inner angular velocity profile steepens to
$\Omega(r) \propto r^{-3/2}$.
4. The mass of an individual convective cell in a giant envelope greatly exceeds the mass of Jupiter. This means than the
orbital eccentricity of any planet will be significantly excited by the associated gravitational quadrupole. We have
calculated the orbital evolution of a suit of planets of various masses (Earth, Neptune, Jupiter) in response to this
gravitational stirring as well as tidal drag. The gravitational stirring has a significant effect on the probability
of absorption of an Earth-mass planet starting at a semi-major axis 1 AU.
5. The case of a solar-mass star which spins down on the MS and then {\it does not} regain angular momentum
from a companion is more complicated. The rotation model described here can be applied while the star
remains more compact than $R_\star \sim 10^{12}$ cm; but slow rotation is enountered throughout the convective
envelope near the tip of the giant branches. In these most extended phases of expansion,
the coupling between core and envelope will depend on additional (and perhaps largely hydrodynamic) mixing mechanisms, whose exact nature is
still being debated \citep{maeder00,fuller14}.
More extended simulations of deep convection, including radiative transport and feedback from
magnetic fields, are required to confirm the dependence of
rotation profile on Coriolis parameter. A more centrally peaked profile is plausibly
achieved near maximum expansion on the giant branches, where the effects of rotation are weaker, but
whether this profile really approaches uniform specific angular momentum needs further testing.
We emphasize the degeneracy between the combined effects of
flattening the slope of the rotation profile and increasing the spindown on the MS (both of which
reduce the rotation rate in the inner envelope), and raising the angular momentum that is received from
a sub-stellar companion during post-MS expansion. Although magnetic field amplification in the inner envelope
becomes more difficult as the rotation profile flattens, even a planetary companion can have a significant compensating effect.
\acknowledgements This work was supported by NSERC. We thank the anonymous
referee for detailed comments on the presentation of our results.
\begin{appendix}
\section{Normalization of Convective Quadrupole During Giant Evolution} \label{app:Q_conv}
The internal density, temperature and luminosity profiles of a model giant star
evolve with time in a complicated way. Here we describe a simple prescription
for the net gravitational quadrupole that is induced by convective motions in the giant envelope,
which can be expressed in terms of integral quanties such as stellar luminosity $L_\star$, radius $R_\star$ and core and envelope masses $M_{\rm core}$, $M_{\rm env}$.
We estimate the magnitude of the quadrupole from
\begin{equation}\label{eq:qcon}
Q_{\rm con} \simeq \sqrt{N_{\rm cells}} \left<\delta M r^2\right>
\end{equation}
where $N_{\rm cells}$ is the number of convective cells in the envelope of the star, and
$\delta M$ is the mass perturbation induced by convection in a cell, and the average is over the volume
of the envelope. We assume that the convective cells (of angular size $\Delta\Omega$) penetrate the whole star, so that
\begin{equation}
N_{\rm cells} \sim \frac{4\pi}{\Delta\Omega};\quad\quad
\left<\delta M r^2\right> \sim \Delta\Omega \int_0^{R_{\star}} r^4 \delta \rho dr.
\end{equation}
The density perturbation $\delta \rho \sim \rho v_{\rm con}^2/c_s^2$, where we estimate $\rho v_{\rm con}^3 \simeq L_{\star}/4\pi r^2$.
Therefore
\begin{equation}
\left<\delta M r^2\right> \sim \Delta \Omega \left( \frac{L_{\star}}{4\pi} \right)^{2/3} \int^{R_{\star}}_0 r^{8/3} \frac{\rho^{1/3}}{c_s^2}dr,
\end{equation}
where we take $\Delta\Omega = 1$.
The assumption of a point gravitational potential, in combination with an adiabatic equation of state, gives
a good fit to the actual density profile of a giant. The temperature profile is more sensitive to the mass
distribution in the star. Hence
\begin{equation}
\rho(r) = \frac{4M_{\rm env}}{\pi^2 R_{\star}^3} \left( \frac{R_{\star}}{r} - 1 \right)^{3/2};\quad\quad
c_s^2(r) = \frac{2}{3} \frac{G}{R_{\star}} M(<r) \left( \frac{R_{\star}}{r} - 1 \right) =
{2GM_\star\over 3R_\star} g\left({r\over R_\star},{M_{\rm core}\over M_\star}\right) \left({R_\star\over r}-1\right),
\end{equation}
where $M(<r)$ is the enclosed mass, which is proportional to
\begin{equation}
g(x,M_{\rm core}/M_\star) = {M_{\rm core}\over M_\star} + {16M_{\rm env}\over \pi M_\star} \int_0^x {x^\prime}^2
\left( {1\over x'} - 1 \right)^{3/2} dx^{\prime}.
\end{equation}
The quadrupole (\ref{eq:qcon}) works out to
\begin{equation}
\frac{Q_{\rm con}}{M_{\rm env} R_{\star}^2} \simeq \frac{3}{4^{1/3} \pi^{5/6} G} \frac{\Delta \Omega^{1/2} L_{\star}^{2/3}
R_{\star}^{5/3}}{M_{\rm env}^{2/3} M_{\star}} f; \quad\quad
f \equiv \int_0^1 {x^{8/3}\over g(x,M_{\rm core}/M_\star)} \left( \frac{1}{x} - 1 \right)^{-1/2} dx.
\end{equation}
\end{appendix}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.